Lecture Notes on Nonlinear Dynamics

(A Work in Progress)

Daniel Arovas
Department of Physics
University of California, San Diego
May 28, 2014

Contents
0.1

Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

0 Reference Materials

x
1

0.1

Dynamical Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

0.2

Hamiltonian Mechanics . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

0.3

Differential Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

0.4

Synchronization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

0.5

Nonlinear Wave Equations . . . . . . . . . . . . . . . . . . . . . . . . . . .

0.6

Hydrodynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

0.7

Pattern Formation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

0.8

Biological Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1 Dynamical Systems
1.1

1.2

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1.1.1

Phase space and phase curves . . . . . . . . . . . . . . . . . . . . .

1.1.2

Vector fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1.1.3

Existence / uniqueness / extension theorems . . . . . . . . . . . . .

1.1.4

Linear differential equations . . . . . . . . . . . . . . . . . . . . . .

1.1.5

Lyapunov functions . . . . . . . . . . . . . . . . . . . . . . . . . . .

N = 1 Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1.2.1

Classification of fixed points (N = 1) . . . . . . . . . . . . . . . . .

1.2.2

Logistic equation . . . . . . . . . . . . . . . . . . . . . . . . . . . .

10

ii

CONTENTS

1.2.3

Singular f (u) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

11

1.2.4

Recommended exercises . . . . . . . . . . . . . . . . . . . . . . . . .

12

1.2.5

Non-autonomous ODEs . . . . . . . . . . . . . . . . . . . . . . . . .

12

Flows on the Circle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

13

1.3.1

Nonuniform oscillator . . . . . . . . . . . . . . . . . . . . . . . . . .

13

1.4

Appendix I : Evolution of Phase Space Volumes . . . . . . . . . . . . . . .

15

1.5

Appendix II : Lyapunov Characteristic Exponents . . . . . . . . . . . . . .

15

1.3

2 Bifurcations
2.1

19

Types of Bifurcations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

19

2.1.1

Saddle-node bifurcation . . . . . . . . . . . . . . . . . . . . . . . . .

19

2.1.2

Transcritical bifurcation . . . . . . . . . . . . . . . . . . . . . . . .

20

2.1.3

Pitchfork bifurcation . . . . . . . . . . . . . . . . . . . . . . . . . .

22

2.1.4

Imperfect bifurcation . . . . . . . . . . . . . . . . . . . . . . . . . .

23

Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

25

2.2.1

Population dynamics . . . . . . . . . . . . . . . . . . . . . . . . . .

25

2.3

Appendix I : The Bletch . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

29

2.4

Appendix II : Landau Theory of Phase Transitions . . . . . . . . . . . . . .

30

2.4.1

Landau coefficients from mean field theory . . . . . . . . . . . . . .

32

2.4.2

Magnetization dynamics . . . . . . . . . . . . . . . . . . . . . . . .

34

2.4.3

Cubic terms in Landau theory : first order transitions . . . . . . . .

37

2.4.4

Magnetization dynamics . . . . . . . . . . . . . . . . . . . . . . . .

38

2.4.5

Sixth order Landau theory : tricritical point . . . . . . . . . . . . .

40

2.4.6

Hysteresis for the sextic potential . . . . . . . . . . . . . . . . . . .

42

2.2

3 Two-Dimensional Phase Flows

3.1

47

Harmonic Oscillator and Pendulum . . . . . . . . . . . . . . . . . . . . . .

47

3.1.1

Simple harmonic oscillator . . . . . . . . . . . . . . . . . . . . . . .

47

3.1.2

Pendulum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

49

iii

CONTENTS

3.2

General N = 2 Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

50

3.2.1

The damped driven pendulum . . . . . . . . . . . . . . . . . . . . .

50

3.2.2

Classification of N = 2 fixed points . . . . . . . . . . . . . . . . . .

53

3.2.3

The fixed point zoo . . . . . . . . . . . . . . . . . . . . . . . . . . .

55

3.2.4

Fixed points for N = 3 systems . . . . . . . . . . . . . . . . . . . .

57

3.3

Andronov-Hopf Bifurcation . . . . . . . . . . . . . . . . . . . . . . . . . . .

58

3.4

Population Biology : Lotka-Volterra Models . . . . . . . . . . . . . . . . . .

60

3.4.1

Rabbits and foxes . . . . . . . . . . . . . . . . . . . . . . . . . . . .

60

3.4.2

Rabbits and sheep . . . . . . . . . . . . . . . . . . . . . . . . . . . .

61

3.5

Poincare-Bendixson Theorem . . . . . . . . . . . . . . . . . . . . . . . . . .

64

3.6

Index Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

65

3.6.1

Gauss-Bonnet Theorem . . . . . . . . . . . . . . . . . . . . . . . . .

68

3.6.2

Singularities and topology . . . . . . . . . . . . . . . . . . . . . . .

70

3.7

Appendix : Example Problem

. . . . . . . . . . . . . . . . . . . . . . . . .

4 Nonlinear Oscillators
4.1

72
75

Weakly Perturbed Linear Oscillators . . . . . . . . . . . . . . . . . . . . . .

75

4.1.1

Nave Perturbation theory and its failure . . . . . . . . . . . . . . .

75

4.1.2

Poincare-Lindstedt method . . . . . . . . . . . . . . . . . . . . . . .

77

Multiple Time Scale Method . . . . . . . . . . . . . . . . . . . . . . . . . .

79

4.2.1

Duffing oscillator . . . . . . . . . . . . . . . . . . . . . . . . . . . .

81

4.2.2

Van der Pol oscillator . . . . . . . . . . . . . . . . . . . . . . . . . .

82

Forced Nonlinear Oscillations . . . . . . . . . . . . . . . . . . . . . . . . . .

83

4.3.1

Forced Duffing oscillator . . . . . . . . . . . . . . . . . . . . . . . .

85

4.3.2

Forced van der Pol oscillator . . . . . . . . . . . . . . . . . . . . . .

87

4.4

Synchronization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

91

4.5

Relaxation Oscillations . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

96

4.5.1

Example problem . . . . . . . . . . . . . . . . . . . . . . . . . . . .

98

4.5.2

Multiple limit cycles

4.2

4.3

. . . . . . . . . . . . . . . . . . . . . . . . . .

100

iv

CONTENTS

4.5.3

Example problem . . . . . . . . . . . . . . . . . . . . . . . . . . . .

102

4.6

Appendix I : Multiple Time Scale Analysis to O(2 ) . . . . . . . . . . . . .

103

4.7

Appendix II : MSA and Poincare-Lindstedt Methods . . . . . . . . . . . . .

105

4.7.1

Problem using multiple time scale analysis . . . . . . . . . . . . . .

105

4.7.2

Solution using Poincare-Lindstedt method . . . . . . . . . . . . . .

109

Appendix III : Modified van der Pol Oscillator . . . . . . . . . . . . . . . .

111

4.8

5 Hamiltonian Mechanics

117

5.1

The Hamiltonian . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

117

5.2

Modified Hamiltons Principle . . . . . . . . . . . . . . . . . . . . . . . . .

119

5.3

Phase Flow is Incompressible . . . . . . . . . . . . . . . . . . . . . . . . . .

119

5.4

Poincare Recurrence Theorem . . . . . . . . . . . . . . . . . . . . . . . . .

120

5.5

Poisson Brackets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

121

5.6

Canonical Transformations . . . . . . . . . . . . . . . . . . . . . . . . . . .

122

5.6.1

Point transformations in Lagrangian mechanics . . . . . . . . . . .

122

5.6.2

Canonical transformations in Hamiltonian mechanics . . . . . . . .

124

5.6.3

Hamiltonian evolution

. . . . . . . . . . . . . . . . . . . . . . . . .

124

5.6.4

Symplectic structure . . . . . . . . . . . . . . . . . . . . . . . . . .

125

5.6.5

Generating functions for canonical transformations . . . . . . . . .

126

Hamilton-Jacobi Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

129

5.7.1

The action as a function of coordinates and time . . . . . . . . . . .

129

5.7.2

The Hamilton-Jacobi equation . . . . . . . . . . . . . . . . . . . . .

131

5.7.3

Time-independent Hamiltonians . . . . . . . . . . . . . . . . . . . .

132

5.7.4

Example: one-dimensional motion . . . . . . . . . . . . . . . . . . .

133

5.7.5

Separation of variables . . . . . . . . . . . . . . . . . . . . . . . . .

133

5.7.6

Example #2 : point charge plus electric field . . . . . . . . . . . . .

135

5.7.7

Example #3 : charged particle in a magnetic field . . . . . . . . . .

137

Action-Angle Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

139

5.8.1

139

5.7

5.8

Circular Phase Orbits: Librations and Rotations . . . . . . . . . . .

CONTENTS

5.8.2

Action-Angle Variables . . . . . . . . . . . . . . . . . . . . . . . . .

140

5.8.3

Canonical Transformation to Action-Angle Variables . . . . . . . .

141

5.8.4

Example : Harmonic Oscillator . . . . . . . . . . . . . . . . . . . .

142

5.8.5

Example : Particle in a Box . . . . . . . . . . . . . . . . . . . . . .

143

5.8.6

Kepler Problem in Action-Angle Variables . . . . . . . . . . . . . .

146

5.8.7

Charged Particle in a Magnetic Field . . . . . . . . . . . . . . . . .

147

5.8.8

Motion on Invariant Tori . . . . . . . . . . . . . . . . . . . . . . . .

148

Canonical Perturbation Theory . . . . . . . . . . . . . . . . . . . . . . . . .

149

5.9.1

Canonical Transformations and Perturbation Theory . . . . . . . .

149

5.9.2

Canonical Perturbation Theory for n = 1 Systems . . . . . . . . . .

151

5.9.3

Example : Nonlinear Oscillator . . . . . . . . . . . . . . . . . . . .

154

5.9.4

n > 1 Systems : Degeneracies and Resonances . . . . . . . . . . . .

155

5.9.5

Particle-Wave Interaction . . . . . . . . . . . . . . . . . . . . . . . .

157

5.10 Adiabatic Invariants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

160

5.9

5.10.1

Example: mechanical mirror . . . . . . . . . . . . . . . . . . . . . .

161

5.10.2

Example: magnetic mirror . . . . . . . . . . . . . . . . . . . . . . .

162

5.10.3

Resonances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

164

5.11 Appendix : Canonical Perturbation Theory . . . . . . . . . . . . . . . . . .

164

6 Ergodicity and the Approach to Equilibrium

6.1

6.2

167

Modeling the Approach to Equilibrium . . . . . . . . . . . . . . . . . . . .

167

6.1.1

Equilibrium . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

167

6.1.2

The Master Equation . . . . . . . . . . . . . . . . . . . . . . . . . .

167

6.1.3

Equilibrium distribution and detailed balance . . . . . . . . . . . .

168

6.1.4

Boltzmanns H-theorem . . . . . . . . . . . . . . . . . . . . . . . . .

169

Phase Flows in Classical Mechanics . . . . . . . . . . . . . . . . . . . . . .

170

6.2.1

Hamiltonian evolution

170

6.2.2

Dynamical systems and the evolution of phase space volumes

6.2.3

Liouvilles equation and the microcanonical distribution

. . . . . . . . . . . . . . . . . . . . . . . . .
. . .

171

. . . . . .

175

vi

CONTENTS

6.3

Irreversibility and Poincare Recurrence . . . . . . . . . . . . . . . . . . . .

175

6.3.1

Poincare recurrence theorem . . . . . . . . . . . . . . . . . . . . . .

176

6.3.2

Kac ring model . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

178

Remarks on Ergodic Theory . . . . . . . . . . . . . . . . . . . . . . . . . .

182

6.4.1

Definition of ergodicity . . . . . . . . . . . . . . . . . . . . . . . . .

182

6.4.2

The microcanonical ensemble

. . . . . . . . . . . . . . . . . . . . .

184

6.4.3

Ergodicity and mixing . . . . . . . . . . . . . . . . . . . . . . . . .

184

Thermalization of Quantum Systems . . . . . . . . . . . . . . . . . . . . . .

189

6.5.1

Quantum dephasing . . . . . . . . . . . . . . . . . . . . . . . . . . .

189

6.5.2

Eigenstate thermalization hypothesis . . . . . . . . . . . . . . . . .

190

6.5.3

When is the ETH true? . . . . . . . . . . . . . . . . . . . . . . . . .

191

6.6

Appendix I : Formal Solution of the Master Equation . . . . . . . . . . . .

191

6.7

Appendix II : Radioactive Decay . . . . . . . . . . . . . . . . . . . . . . . .

193

6.8

Appendix III : Canonical Transformations in Hamiltonian Mechanics . . . .

194

6.4

6.5

7 Maps, Strange Attractors, and Chaos

7.1

197

Maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

197

7.1.1

Parametric Oscillator . . . . . . . . . . . . . . . . . . . . . . . . . .

197

7.1.2

Other models of kicked dynamics . . . . . . . . . . . . . . . . . . .

200

One-dimensional Maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

202

7.2.1

Lyapunov Exponents . . . . . . . . . . . . . . . . . . . . . . . . . .

204

7.2.2

Chaos in the logistic map . . . . . . . . . . . . . . . . . . . . . . . .

205

7.2.3

Intermittency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

207

7.3

Attractors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

207

7.4

The Lorenz Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

208

7.4.1

Fixed point analysis . . . . . . . . . . . . . . . . . . . . . . . . . . .

210

7.4.2

Poincare section . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

212

7.4.3

R
ossler System

213

7.2

. . . . . . . . . . . . . . . . . . . . . . . . . . . . .

vii

CONTENTS

8 Front Propagation
8.1

8.2

8.3

217

Reaction-Diffusion Systems . . . . . . . . . . . . . . . . . . . . . . . . . . .

217

8.1.1

Single component systems . . . . . . . . . . . . . . . . . . . . . . .

217

8.1.2

Propagating front solutions . . . . . . . . . . . . . . . . . . . . . . .

218

8.1.3

Stability and zero mode . . . . . . . . . . . . . . . . . . . . . . . . .

221

8.1.4

Fishers equation . . . . . . . . . . . . . . . . . . . . . . . . . . . .

223

8.1.5

Velocity selection and stability . . . . . . . . . . . . . . . . . . . . .

224

Multi-Species Reaction-Diffusion Systems . . . . . . . . . . . . . . . . . . .

225

8.2.1

Propagating front solutions . . . . . . . . . . . . . . . . . . . . . . .

227

Excitable Media . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

230

8.3.1

233

Front propagation in excitable media . . . . . . . . . . . . . . . . .

9 Pattern Formation
9.0.2

239

Reaction-Diffusion Dynamics Revisited . . . . . . . . . . . . . . . .

239

9.1

Turing and Hopf Instabilities . . . . . . . . . . . . . . . . . . . . . . . . . .

241

9.2

The Brusselator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

242

9.2.1

. . . . . . . . . . . . . . . . . . . . . . . .

244

9.3

Rayleigh-Benard Instability . . . . . . . . . . . . . . . . . . . . . . . . . . .

247

9.4

Center Manifold Reduction . . . . . . . . . . . . . . . . . . . . . . . . . . .

249

9.5

Selection and Stability of Spatial Patterns . . . . . . . . . . . . . . . . . . .

251

9.5.1

d=1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

251

9.5.2

Remarks on the amplitude equations for d > 1 . . . . . . . . . . . .

252

9.5.3

d=2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

253

9.5.4

d=3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

260

9.6

Anisotropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

260

9.7

Phase Diffusion : Eckhaus and Zigzag Instabilities . . . . . . . . . . . . . .

261

The amplitude equation

10 Solitons
10.1 The Korteweg-deVries Equation . . . . . . . . . . . . . . . . . . . . . . . .

265
266

viii

CONTENTS

10.1.1

KdV solitons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

267

10.1.2

Periodic solutions : soliton trains . . . . . . . . . . . . . . . . . . .

268

10.1.3

Interlude: primer on elliptic functions . . . . . . . . . . . . . . . . .

269

10.1.4

The soliton lattice . . . . . . . . . . . . . . . . . . . . . . . . . . . .

270

10.1.5

N -soliton solutions to KdV . . . . . . . . . . . . . . . . . . . . . . .

271

10.1.6

Backlund transformations . . . . . . . . . . . . . . . . . . . . . . .

275

10.2 Sine-Gordon Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

278

10.2.1

Tachyon solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . .

280

10.2.2

Hamiltonian formulation . . . . . . . . . . . . . . . . . . . . . . . .

281

10.2.3

Phonons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

281

10.2.4

Mechanical realization . . . . . . . . . . . . . . . . . . . . . . . . .

282

10.2.5

Kinks and antikinks . . . . . . . . . . . . . . . . . . . . . . . . . . .

282

10.2.6

Backlund transformation for the sine-Gordon system . . . . . . . .

284

10.3 Nonlinear Schr

odinger Equation . . . . . . . . . . . . . . . . . . . . . . . .

286

10.3.1

Amplitude-phase representation . . . . . . . . . . . . . . . . . . . .

287

10.3.2

Phonons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

287

10.3.3

Soliton solutions for NLS(+) . . . . . . . . . . . . . . . . . . . . . .

288

10.3.4

Dark solitons for NLS() . . . . . . . . . . . . . . . . . . . . . . . .

289

11 Shock Waves

291

11.1 Nonlinear Continuity Equation . . . . . . . . . . . . . . . . . . . . . . . . .

291

11.2 Shocks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

297

11.3 Internal Shock Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . .

298

11.3.1

Quadratic J() . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

299

11.4 Shock Fitting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

300

11.4.1

An Important Caveat . . . . . . . . . . . . . . . . . . . . . . . . . .

300

11.4.2

Recipe for shock fitting (J () = 0) . . . . . . . . . . . . . . . . . .

301

11.4.3

Example problem . . . . . . . . . . . . . . . . . . . . . . . . . . . .

302

11.5 Long-time Behavior of Shocks

. . . . . . . . . . . . . . . . . . . . . . . . .

305

ix

CONTENTS

11.5.1

Fate of a hump . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

305

11.5.2

N-wave and P-wave . . . . . . . . . . . . . . . . . . . . . . . . . . .

307

11.6 Shock Merging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

308

11.7 Shock Fitting for General J() . . . . . . . . . . . . . . . . . . . . . . . . .

309

11.8 Sources . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

310

11.8.1

Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

311

11.8.2

Moving sources . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

311

11.9 Burgers Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

312

11.9.1

The limit 0 . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

313

11.9.2

Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

315

11.9.3

Confluence of shocks . . . . . . . . . . . . . . . . . . . . . . . . . .

317

11.10 Appendix I : The Method of Characteristics . . . . . . . . . . . . . . . . . .

319

11.10.1 Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

319

11.11 Appendix II : Shock Fitting an Inverted Parabola . . . . . . . . . . . . . .

320

11.12 Appendix III : Red Light, Green Light . . . . . . . . . . . . . . . . . . . . .

322

11.13 Appendix IV : Characteristics for a Periodic Signal . . . . . . . . . . . . . .

324

11.14 Appendix V : Car-Following Models . . . . . . . . . . . . . . . . . . . . . .

331

0.1

CONTENTS

Preface

This is a proto-preface. A more complete preface will be written after these notes are
completed.
These lecture notes are intended to supplement a graduate level course in nonlinear dynamics.

Chapter 0

Reference Materials
No one book contains all the relevant material. Here I list several resources, arranged by
topic. My personal favorites are marked with a diamond ().

0.1

Dynamical Systems

S. Strogatz, Nonlinear Dynamics and Chaos (Addison-Wesley, 1994)

S. Neil Rasband, Chaotic Dynamics of Nonlinear Systems (Wiley, 1990)
J. Guckenheimer and P. Holmes, Nonlinear Oscillations, Dynamical Systems, and Bifurcations of Vector Fields (Springer, 1983)
E. A. Jackson, Perspectives of Nonlinear Dynamics, 2 vols. (Cambridge, 1991)
A. J. Lichtenberg and M. A. Lieberman, Regular and Stochastic Motion (Springer,
1983)
R. Z. Sagdeev, D. A. Usikov, and G. M. Zaslavsky, Nonlinear Physics from the Pendulum to Turbulence and Chaos (Harwood Academic, 1988)
M. Lakshmanan and S. Rajasekar, Nonlinear Dynamics : Integrability, Chaos, and
Patterns (Springer, 2003)
H. G. Schuster and W. Just, Deterministic Chaos, 4th ed. (Wiley-VCH, 2005)

CHAPTER 0. REFERENCE MATERIALS

M. Tabor, Chaos and Integrability in Nonlinear Dynamics (Wiley, 1989)

E. Ott, Chaos in Dynamical Systems, 2nd ed. (Cambridge, 2002)

0.2

Hamiltonian Mechanics

G. M. Zaslavsky, Hamiltonian Chaos and Fractional Dynamics (Oxford, 2005)

J. V. Jose and E. J. Saletan, Classical Dynamics : A Contemporary Approach (Cambridge, 1998)
V. I. Arnold, V. V. Kozlov, and A. I. Neishtadt, Mathematical Aspects of Classical
and Celestial Mechanics (Springer, 2006)
A. L. Fetter and J. D. Walecka, Nonlinear Mechanics (Dover, 2006)
I. Percival and D. Richards, Introduction to Dynamics (Cambridge, 1982)

0.3

Differential Equations

D. Zwillinger, Handbook of Differential Equations, 3rd ed. (Academic Press, 1998)

A. H. Nayfeh, Introduction to Perturbation Techniques (Wiley, 1981)
C. M. Bender and S. A. Orszag, Advanced Mathematical Methods for Scientists and
Engineers (Springer, 1999)
V. I. Arnold, Ordinary Differential Equations (MIT Press, 1973)
V. I. Arnold, Geometrical Methods in the Theory of Ordinary Differential Equations
(Springer, 1988)
L. Perko, Differential Equations and Dynamical Systems, 3rd ed. (Springer, 2001)

0.4. SYNCHRONIZATION

J. A. Sanders and F. Verhulst, Averaging Methods in Nonlinear Dynamical Systems

(Springer, 1985).

0.4

Synchronization

Y. Kuramoto, Chemical Oscillations, Waves, and Turbulence (Dover, 2003)

A. Pikovsky, M. Rosenblum, and J. Kurths, Synchronization : A Universal Concept
in Nonlinear Sciences (Cambridge, 2001)

0.5

Nonlinear Wave Equations

G. B. Whitham, Linear and Nonlinear Waves (Wiley, 1999)

A. Scott, Nonlinear Science, 2nd ed. (Oxford, 2003)
E. Infeld and G. Rowlands, Nonlinear Waves, Solitons, and Chaos (Cambridge, 2000)
R. Rajaraman, Solitons and Instantons (North-Holland, 1987)
H.-K. Rhee, R. Aris, and N. R. Amundson, First-Order Partial Differential Equations
(2 vols.) (Dover, 1986)
L. I. Pismen, Vortices in Nonlinear Fields (Oxford, 1999)

0.6

Hydrodynamics

C. Godreche and P. Manneville (eds.), Hydrodynamics and Nonlinear Instabilities

(Cambridge, 1998)
P. G. Saffman, Vortex dynamics (Cambridge, 1992)

CHAPTER 0. REFERENCE MATERIALS

0.7

Pattern Formation

M. Cross and H. Greenside, Patterm Formation and Dynamics in Nonequilibrium

Systems (Cambridge, 2009)
P. Manneville, Instabilities, Chaos, and Turbulence (Imperial College Press, 2004)
D. Walgraef, Spatio-Temporal Pattern Formation (Springer, 1996)
J. Garca-Ojalvo and J. M. Sancho, Noise in Spatially Extended Systems (Springer,
1999)
P. Manneville, Dissipative Structures and Weak Turbulence (Academic Press, 1990)
L. M. Pismen, Patterns and Interfaces in Dissipative Dynamics (Springer, 2006)
M. I. Rabinovich, A. B. Ezersky, and P. D. Weidman, The Dynamics of Patterns
(World Scientific, 2000)
R. Hoyle, Pattern Formation : An Introduction to Methods (Cambridge, 2006)

0.8

Biological Applications

J. D. Murray, Mathematical Biology, (3rd ed., 2 vols.) (Springer, 2002)

E. M. Izhikevich, Dynamical Systems in Neuroscience (MIT Press, 2007)
L. Edelstein-Keshet, Mathematical Models in Biology (SIAM, 2005)
J. Hofbauer and K. Sigmund, Evolutionary Games and Population Dynamics (Cambridge, 1998)
M. A. Nowak, Evolutionary Dynamics (Harvard, 2006)

Chapter 1

Dynamical Systems
1.1
1.1.1

Introduction
Phase space and phase curves

Dynamics is the study of motion through phase space. The phase space of a given dynamical
system is described as an N -dimensional manifold, M. A (differentiable) manifold M is a
topological space that is locally diffeomorphic to RN .1 Typically in this course M will RN
itself, but other common examples include the circle S1 , the torus T2 , the sphere S2 , etc.
Let gt : M M be a one-parameter family of transformations from M to itself, with
gt=0 = 1, the identity. We call gt the t-advance mapping. It satisfies the composition rule
gt gs = gt+s .

(1.1)

Let us choose a point 0 M. Then we write (t) = gt 0 , which also is in M. The set



gt 0  t R , 0 M is called a phase curve. A graph of the motion (t) in the product
space R M is called an integral curve.

1.1.2

Vector fields

The velocity vector V () is given by the derivative


d 
V () =  gt .
dt t=0

(1.2)

The velocity V () is an element of the tangent space to M at , abbreviated TM .

If M is N -dimensional, then so is each TM (for all p). However, M and TM may
1
A diffeomorphism F : M N is a differentiable map with a differentiable inverse. This is a special
type of homeomorphism, which is a continuous map with a continuous inverse.

CHAPTER 1. DYNAMICAL SYSTEMS

Figure 1.1: An example of a phase curve.

differ topologically. For example, if M = S1 , the circle, the tangent space at any point is
isomorphic to R.
For our purposes, we will take = (1 , . . . , N ) to be an N -tuple, i.e. a point in RN . The
equation of motion is then


d
(t) = V (t) .
(1.3)
dt

Note that any N th order ODE, of the general form



dx
dN 1x
dNx
= F x, , . . . , N 1 ,
dtN
dt
dt

(1.4)

may be represented by the first order system = V (). To see this, define k =
dk1x/dtk1 , with k = 1, . . . , N . Thus, for j < N we have j = j+1 , and N = f .
In other words,

z }|

.1
.
d
.

dt

V ()

}|
{
{ z

2

..

.
=
.

N 1
N


F 1 , . . . , N
N

1.1.3

(1.5)

Existence / uniqueness / extension theorems

Theorem : Given = V () and (0), if each V () is a smooth vector field over some
open set D M, then for (0) D the initial value problem has a solution on some
finite time interval (, + ) and the solution is unique. Furthermore, the solution has a
unique extension forward or backward in time, either indefinitely or until (t) reaches the
boundary of D.
Corollary : Different trajectories never intersect!

1.1. INTRODUCTION

1.1.4

Linear differential equations

A homogeneous linear N th order ODE,

dN 1x
dx
dNx
+ c0 x = 0
+
c
+ . . . + c1
N 1
N
N
1
dt
dt
dt

(1.6)

may be written in matrix form, as

d 2

. =
dt ..
N

Thus,

0
0
..
.

}|
0
1
..
.

1
0
..
.

0
0
..
.

c0 c1 c2 cN 1

1

2
. .
..

(1.7)

= M ,

(1.8)

and if the coefficients ck are time-independent, i.e. the ODE is autonomous, the solution is
obtained by exponentiating the constant matrix Q:
(t) = exp(M t) (0) ;

(1.9)

the exponential of a matrix may be given meaning by its Taylor series expansion. If the
ODE is not autonomous, then M = M (t) is time-dependent, and the solution is given by
the path-ordered exponential,
(t) = P exp

( Zt

dt M (t ) (0) ,

(1.10)

As defined, the equation = V () is autonomous, since gt depends only on t and on no

other time variable. However, by extending the phase space from M to R M, which is of
dimension (N + 1), one can describe arbitrary time-dependent ODEs.

1.1.5

Lyapunov functions

For a general dynamical system = V (), a Lyapunov function L() is a function which
satisfies
L() V () 0 .
(1.11)
There is no simple way to determine whether a Lyapunov function exists for a given dynamical system, or, if it does exist, what the Lyapunov function is. However, if a Lyapunov
function can be found,
then this severely limits the possible behavior of the system. This


is because L (t) must be a monotonic function of time:


d
d
L (t) = L
= L() V () 0 .
dt
dt

(1.12)

CHAPTER 1. DYNAMICAL SYSTEMS

Thus, the system evolves toward a local minimum of the Lyapunov function. In general
this means that oscillations are impossible in systems for which a Lyapunov function exists.
For example, the relaxational dynamics of the magnetization M of a system are sometimes
modeled by the equation
F
dM
=
,
(1.13)
dt
M
where F (M, T ) is the free energy of the system. In this model, assuming constant temper
2
ature T , F = F (M ) M = F (M ) 0. So the free energy F (M ) itself is a Lyapunov
function, and it monotonically decreases during the evolution of the system. We shall meet
up with this example again in the next chapter when we discuss imperfect bifurcations.

1.2

N = 1 Systems

We now study phase flows in a one-dimensional phase space, governed by the equation
du
= f (u) .
dt

(1.14)

Again, the equation u = h(u, t) is first order, but not autonomous, and it corresponds to
the N = 2 system,
  

d u
h(u, t)
=
.
(1.15)
1
dt t
The equation 1.14 is easily integrated:
du
= dt
f (u)

t t0 =

Zu

u0

du
f (u )

(1.16)

This gives t(u); we must then invert this relationship to obtain u(t).
Example : Suppose f (u) = a bu, with a and b constant. Then
dt =

du
= b1 d ln(a bu)
a bu

whence
1
t = ln
b

a b u(0)
a b u(t)

u(t) =

a 
a
exp(bt) .
+ u(0)
b
b

(1.17)

(1.18)

Even if one cannot analytically obtain u(t), the behavior is very simple, and easily obtained
by graphical analysis. Sketch the function f (u). Then note that

f (u) > 0 u > 0 move to right

(1.19)
u = f (u) =
f (u) < 0 u < 0 move to left

f (u) = 0 u = 0 fixed point

1.2. N = 1 SYSTEMS

Figure 1.2: Phase flow for an N = 1 system.

The behavior of N = 1 systems is particularly simple: u(t) flows to the first stable fixed
point encountered, where it then (after a logarithmically infinite time) stops. The motion
is monotonic the velocity u never changes sign. Thus, oscillations never occur for N = 1
phase flows.2

1.2.1

Classification of fixed points (N = 1)

A fixed point u satisfies f (u ) = 0. Generically, f (u ) 6= 0 at a fixed point.3 Suppose

f (u ) < 0. Then to the left of the fixed point, the function f (u < u ) is positive, and the
flow is to the right, i.e. toward u . To the right of the fixed point, the function f (u > u ) is
negative, and the flow is to the left, i.e. again toward u . Thus, when f (u ) < 0 the fixed
point is said to be stable, since the flow in the vicinity of u is to u . Conversely, when
f (u ) > 0, the flow is always away from u , and the fixed point is then said to be unstable.
Indeed, if we linearize about the fixed point, and let u u , then
= f (u ) + 12 f (u ) 2 + O(3 ) ,

(1.20)

and dropping all terms past the first on the RHS gives
h
i
(t) = exp f (u ) t (0) .

(1.21)

The deviation decreases exponentially for f (u ) < 0 and increases exponentially for f (u ) >
0. Note that



1
,
(1.22)
t() = ln
f (u )
(0)
so the approach to a stable fixed point takes a logarithmically infinite time. For the unstable
case, the deviation grows exponentially, until eventually the linearization itself fails.
2
3

When I say never I mean sometimes see the section 1.3.

The system f (u ) = 0 and f (u ) = 0 is overdetermined, with two equations for the single variable u .

10

CHAPTER 1. DYNAMICAL SYSTEMS

Figure 1.3: Flow diagram for the logistic equation.

1.2.2

Logistic equation

This model for population growth was first proposed by Verhulst in 1838. Let N denote
the population in question. The dynamics are modeled by the first order ODE,

N
dN
,
(1.23)
= rN 1
dt
K

where N , r, and K are all positive. For N K the growth rate is r, but as N increases a
quadratic nonlinearity kicks in and the rate vanishes for N = K and is negative for N > K.
The nonlinearity models the effects of competition between the organisms for food, shelter,
or other resources. Or maybe they crap all over each other and get sick. Whatever.
There are two fixed points, one at N = 0, which is unstable (f (0) = r > 0). The other, at
N = K, is stable (f (K) = r). The equation is adimensionalized by defining = N/K
and s = rt, whence
= (1 ) .
(1.24)
Integrating,

 
d
= ds
= d ln
(1 )
1

(s) =

0 + 1 0 exp(s)

(1.25)

As s , (s) = 1 01 1 es + O(e2s ), and the relaxation to equilibrium ( = 1)
is exponential, as usual.
Another application of this model is to a simple autocatalytic reaction, such as
A + X 2X ,

(1.26)

i.e. X catalyses the reaction A X. Assuming a fixed concentration of A, we have

x = + a x x2 ,

(1.27)

where x is the concentration of X, and are the forward and backward reaction rates.

1.2. N = 1 SYSTEMS

1.2.3

11



Figure 1.4: f (u) = A u u  , for > 1 and < 1.

Singular f (u)



Suppose that in the vicinity of a fixed point we have f (u) = A u u  , with A > 0. We
now analyze both sides of the fixed point.

u < u : Let = u u. Then

= A
hence

1
1
= 0
At ,
1
1

h
i 1
1
.
(t) = 1
+
(

1)A
t
0

(1.28)

(1.29)

This, for < 1 the fixed point = 0 is reached in a finite time: (tc ) = 0, with
tc =

1
0
.
(1 ) A

(1.30)

For > 1, we have limt (t) = 0, but (t) > 0 t < .

The fixed point u = u is now half-stable the flow from the left is toward u but from the
right is away from u . Lets analyze the flow on either side of u .
u > u : Let = u u . Then = A , and
h
i 1
1
.
(t) = 1
+
(1

)A
t
0

(1.31)

For < 1, (t) escapes to = only after an infinite time. For > 1, the escape to
infinity takes a finite time: (tc ) = , with
tc =

1
0
.
( 1) A

(1.32)

In both cases, higher order terms in the (nonanalytic) expansion of f (u) about u = u will
eventually come into play.

12

CHAPTER 1. DYNAMICAL SYSTEMS

Figure 1.5: Solutions to = A . Left panel: = u u, with = 1.5 (solid red) and
= 0.5 (dot-dashed blue); A = 1 in both cases. Right panel: = u u , = 1.5 (solid
red) and = 0.5 (dot-dashed blue); A = 4 in both cases

1.2.4

Recommended exercises

It is constructive to sketch the phase flows for the following examples:

v = g

u = A sin(u)

mv = mg v

u = A (u a)(u b)(u c)

mv = mg cv 2 sgn(v)

u = au2 bu3 .

In each case, identify all the fixed points and assess their stability. Assume all constants A,
a, b, c, , etc. are positive.

1.2.5

Non-autonomous ODEs

Non-autonomous ODEs of the form u = h(u, t) are in general impossible to solve by quadratures. One can always go to the computer, but it is worth noting that in the separable case,
h(u, t) = f (u) g(t), one can obtain the solution
du
= g(t) dt
f (u)

Zu

u0

du
=
f (u )

Zt
dt g(t )

(1.33)

which implicitly gives u(t). Note that u may now change sign, and u(t) may even oscillate.
For an explicit example, consider the equation
u = A (u + 1) sin(t) ,

(1.34)

the solution of which is

u(t) = 1 + u0 + 1 exp


A
1 cos(t)

(1.35)

13

1.3. FLOWS ON THE CIRCLE

In general, the non-autonomous case defies analytic solution. Many have been studied, such
as the Riccati equation,
du
= P (t)u2 + Q(t)u + R(t) .
(1.36)
dt
Riccati equations have the special and remarkable property that one can generate all solutions (i.e. with arbitrary boundary condition u(0) = u0 ) from any given solution (i.e. with
any boundary condition).

1.3

Flows on the Circle

We had remarked that oscillations are impossible for the equation u = f (u) because the
flow is to the first stable fixed point encountered. If there are no stable fixed points, the
flow is unbounded. However, suppose phase space itself is bounded, e.g. a circle S1 rather
than the real line R. Thus,
= f () ,
(1.37)
with f ( + 2) = f (). Now if there are no fixed points, (t) endlessly winds around the
circle, and in this sense we can have oscillations.

1.3.1

Nonuniform oscillator

A particularly common example is that of the nonuniform oscillator,

= sin ,

(1.38)

which has applications to electronics, biology, classical mechanics, and condensed matter
physics. Note that the general equation = A sin may be rescaled to the above form.
A simple application is to the dynamics of a driven, overdamped pendulum. The equation
of motion is
I + b + I02 sin = N ,
(1.39)
where I is the moment of inertia, b is the damping parameter, N is the external torque
(presumed constant), and 0 is the frequency of small oscillations when b = N = 0. When
b is large, the inertial term I may be neglected, and after rescaling we arrive at eqn. 1.38.
The book by Strogatz provides a biological example of the nonuniform oscillator: fireflies.
An individual firefly will on its own flash at some frequency f . This can be modeled by the
equation = , where = 2f is the angular frequency. A flash occurs when = 2n for
n Z. When subjected to a periodic stimulus, fireflies will attempt to synchronize their
flash to the flash of the stimulus. Suppose the stimulus is periodic with angular frequency
. The firefly synchronization is then modeled by the equation
= A sin( t) .

(1.40)

Here, A is a measure of the fireflys ability to modify its natural frequency in response to
the stimulus. Note that when 0 < t < , i.e. when the firefly is leading the stimulus,

14

CHAPTER 1. DYNAMICAL SYSTEMS

Figure 1.6: Flow for the nonuniform oscillator = sin for three characteristic values
of .
the dynamics tell the firefly to slow down. Conversely, when < t < 0, the firefly
is lagging the stimulus, the the dynamics tell it to speed up. Now focus on the difference
t. We have
= A sin ,
(1.41)
which is the nonuniform oscillator. We can adimensionalize by defining
s At
yielding

d
ds

,
A

(1.42)

= sin .

Fixed points occur only for < 1, for sin = . To integrate, set z = exp(i), in which
case
z = 21 (z 2 2iz 1)

= 12 (z z )(z z+ ) ,

(1.43)

where = 1 2 and z = i . Note that for 2 < 1, is real, and z(t ) = z .

This equation can easily be integrated, yielding
z(t) =

(z(0) z ) z+ (z(0) z+ ) z exp(t)

,
(z(0) z ) (z(0) z+ ) exp(t)

(1.44)

For 2 > 1, the motion is periodic, with period

T =

Z2
0

The situation is depicted in Fig. 1.6.

2
d
=
.
|| sin
2 1

(1.45)

1.4. APPENDIX I : EVOLUTION OF PHASE SPACE VOLUMES

1.4

15

Appendix I : Evolution of Phase Space Volumes

Recall the general form of a dynamical system, = V (). Usually we are interested in
finding integral curves (t). However, consider for the moment a collection of points in
phase space comprising a region R. As the dynamical system evolves, this region will also
evolve, so that R = R(t). We now ask: how does the volume of R(t),
Z


d ,
(1.46)
vol R(t) =
R(t)

where d = d1 d2 dN is the phase space measure, change with time. We have,

explicitly,
Z


d
vol R(t + dt) =
R(t+dt)

i (t + dt)

d
j (t)

R(t)

n

o
,
d 1 + V dt + O (dt)2

(1.47)

R(t)

since




Vi 
i (t + dt)
dt + O (dt)2 ,
= ij +

j (t)
j (t)

(1.48)

and, using ln detM = Tr ln M ,

det(1 + A) = 1 + TrA + O(2 ) .

(1.49)

Thus,


d
vol R(t) =
dt
=

d V

(1.50)

R(t)

V ,
d n

(1.51)

R(t)

where in the last line we have used Stokes theorem to convert the volume integral over R
to a surface integral over its boundary R.

1.5

Appendix II : Lyapunov Characteristic Exponents

Suppose (t) is an integral curve i.e. a solution of = V (). We now ask: how do
nearby trajectories behave? Do they always remain close to (t) for all t? To answer this,

16

CHAPTER 1. DYNAMICAL SYSTEMS

e (t) + (t), in which case

we write (t)

d
i (t) = Mij (t) j (t) + O 2 ,
dt

Vi 
.
Mij (t) =
j (t)

where

(1.52)

(1.53)

The solution, valid to first order in , is

i (t) = Qij (t, t0 ) j (t0 ) ,

where the matrix Q(t, t0 ) is given by the path ordered exponential,
( Zt
)
Q(t, t0 ) = P exp

dt M (t )

(1.54)

(1.55)

t0






t
t
t
lim 1 +
M (tN 1 ) 1 +
M (t1 ) 1 +
M (t0 ) ,
(1.56)
N
N
N
N
with t = t t0 and tj = t0 + (j/N )t. P is the path ordering operator , which places
earlier times to the right:

A(t) B(t ) if t > t

PA(t) B(t ) =
(1.57)

B(t ) A(t) if t < t .



The distinction is important if A(t), B(t ) 6= 0. Note that Q satisfies the composition
property,
Q(t, t0 ) = Q(t, t1 ) Q(t1 , t0 )
(1.58)
for any t1 [t0 , t]. When M is time-independent, as in the case of a fixed point where
V ( ) = 0, the
path ordered exponential reduces to the ordinary exponential, and Q(t, t0 ) =
exp M (t t0 ) .

Generally it is impossible to analytically compute path-ordered exponentials. However, the

following example may be instructive. Suppose




M1 if t/T 2j, 2j + 1
M (t) =
(1.59)




M2 if t/T 2j + 1, 2j + 2 ,
for all integer j. M (t) is a matrix-valued square wave, with period 2T . Then, integrating
over one period, from t = 0 to t = 2T , we have
( Z2T
)
A exp

AP P exp

= e(M1 +M2 ) T

dt M (t)

(1.60)

( Z2T

dt M (t)

= eM2 T eM1 T .

(1.61)

1.5. APPENDIX II : LYAPUNOV CHARACTERISTIC EXPONENTS

17

In general, A 6= AP , so the path ordering has a nontrivial effect4 .

The Lyapunov exponents are defined in the following manner. Let e be an N -dimensional
unit vector. Define

!
(t)
1

ln
(0 , e) lim lim
,
(1.62)
(t0 )
t b0 t t0

(t0 )=b e

where k k denotes the Euclidean norm of a vector, and where 0 = t0 . A theorem due
to Oseledec guarantees that there are N such values i (0 ), depending on the choice of e,
for a given 0 . Specifically, the theorem guarantees that the matrix
W Qt Q

1/(tt0 )

(1.63)

converges in the limit t for almost all 0 . The eigenvalues i correspond to the different eigenspaces of W . Oseledecs theorem (also called the multiplicative ergodic theorem)
guarantees that the eigenspaces of W either grow (i > 1) or shrink (i < 1) exponenth
tially fast. That

is, the norm any vector lying in the i eigenspace of W will behave as
exp i (t t0 ) , for t .
Note that while W = W t is symmetric by construction, Q is simply a general real-valued
N N matrix. The left and right eigenvectors of a matrix M GL(N, R) will in general
be different. The set of eigenvalues is, however, common to both sets of eigenvectors.
Let { } be the right eigenvectors and { } the left eigenvectors, such that
Mij ,j = ,i
,i Mij = ,j .
We can always choose the left and right eigenvectors to be orthonormal, viz.



 = ,i ,j = .

(1.64)
(1.65)

(1.66)

1
Indeed, we can define the matrix Si = ,i , in which case Sj
= ,j , and

S 1 M S = diag 1 , . . . , N .

(1.67)

The matrix M can always be decomposed into its eigenvectors, as

X
Mij =
,i ,j .

(1.68)

If we expand u in terms of the right eigenvectors,

X
(t) =
C (t) (t) ,

(1.69)

If M1 , M2 = 0 then A = AP .

18

CHAPTER 1. DYNAMICAL SYSTEMS

then upon taking the inner product with , we find that C obeys



C +  C = C .

(1.70)

If = 0, e.g. if M is time-independent, then C (t) = C (0) e t , and

C (0)

z
}|
{
XX

i (t) =
j (0) ,j e t ,i .

(1.71)

Thus, the component of (t) along increases exponentially with time if Re( ) > 0, and
decreases exponentially if Re( ) < 0.

Chapter 2

Bifurcations
2.1

Types of Bifurcations

2.1.1

Saddle-node bifurcation

We remarked above how f (u) is in general nonzero when f (u) itself vanishes, since two
equations in a single unknown is an overdetermined set. However, consider the function
F (x, ), where is a control parameter. If we demand F (x, ) = 0 and x F (x, ) = 0,
we have two equations in two unknowns, and in general there will be a zero-dimensional
solution set consisting of points (xc , c ). The situation is depicted in Fig. 2.1.
Lets expand F (x, ) in the vicinity of such a point (xc , c ):



F 
1 2F 
F 
(x xc )2
(x xc ) +
( c ) +
F (x, ) = F (xc , c ) +
x 

2 x2 
(xc ,c )
(xc ,c )
(xc ,c )


2
2


F 
1 F
+
( c )2 + . . .
(x xc ) ( c ) +
x 
2 2 

(2.1)

(xc ,c )

(xc ,c )

= A ( c ) + B (x xc ) + . . . ,

(2.2)

where we keep terms of lowest order in the deviations x and . If we now rescale u
p
B/A (x xc ), r c , and = At, we have, neglecting the higher order terms, we
obtain the normal form of the saddle-node bifurcation,
du
= r + u2 .
d

(2.3)

The evolution of
For r < 0 there are two fixed points one
the flow is depicted in Fig. 2.2.
stable (u = r) and one unstable (u = + r). At r = 0 these two nodes coalesce and
annihilate each other. (The point u = 0 is half-stable precisely at r = 0.) For r > 0 there
are no longer any fixed points in the vicinity of u = 0. In the left panel of Fig. 2.3 we show
the flow in the extended (r, u) plane. The unstable and stable nodes annihilate at r = 0.
19

20

CHAPTER 2. BIFURCATIONS

Figure 2.1: Evolution of F (x, ) as a function of the control parameter .

2.1.2

Transcritical bifurcation

Another situation which arises frequently is the transcritical bifurcation. Consider the
equation x = f (x) in the vicinity of a fixed point x .
dx
= f (x ) (x x ) + 12 f (x )(x x )2 + . . . .
dt

(2.4)

We rescale u (xx ) with = 12 f (x ) and define r f (x ) as the control parameter,

to obtain, to order u2 ,
du
= ru u2 .
(2.5)
dt
Note that the sign of the u2 term can be reversed relative to the others by sending u u
and r r.
Consider a crude model of a laser threshold. Let n be the number of photons in the laser
cavity, and N the number of excited atoms in the cavity. The dynamics of the laser are
approximated by the equations
n = GN n kn

N = N0 n .

(2.6)
(2.7)

Here G is the gain coefficient and k the photon decay rate. N0 is the pump strength, and
is a numerical factor. The first equation tells us that the number of photons in the cavity
grows with a rate GN k; gain is proportional to the number of excited atoms, and the
loss rate is a constant cavity-dependent quantity (typically through the ends, which are
semi-transparent). The second equation says that the number of excited atoms is equal to
the pump strength minus a term proportional to the number of photons (since the presence

2.1. TYPES OF BIFURCATIONS

21

Figure 2.2: Flow diagrams for the saddle-node bifurcation u = r + u2 (top) and the transcritical bifurcation u = ru u2 (bottom).
of a photon means an excited atom has decayed). Putting them together,
n = (GN0 k) n Gn2 ,

(2.8)

which exhibits a transcritical bifurcation at pump strength N0 = k/G. For N0 < k/G the
system acts as a lamp; for N0 > k/G the system acts as a laser.
What happens in the transcritical bifurcation is an exchange of stability of the fixed points
at u = 0 and u = r as r passes through zero. This is depicted graphically in the bottom

Figure 2.3: Extended phase space (r, u) flow diagrams for the saddle-node bifurcation u =
r + u2 (left) and the transcritical bifurcation u = ru u2 (right).

22

CHAPTER 2. BIFURCATIONS

Figure 2.4: Top: supercritical pitchfork bifurcation u = ru u3 . Bottom: subcritical

pitchfork bifurcation u = ru + u3 .
panel of Fig. 2.2.

2.1.3

Pitchfork bifurcation

The pitchfork bifurcation is commonly encountered in systems in which there is an overall

parity symmetry (u u). There are two classes of pitchfork: supercritical and subcritical.
The normal form of the supercritical bifurcation is
u = ru u3 ,

(2.9)

which has fixed points at u = 0 and u = r. Thus, the situation is as depicted in fig.
2.4 (top panel). For r < 0 there is a single stable fixed point at u = 0. For r > 0, u = 0

is unstable, and flanked by two stable fixed points at u = r.

If we send u u, r r, and t t, we obtain the subcritical pitchfork , depicted in
the bottom panel of fig. 2.4. The normal form of the subcritical pitchfork bifurcation is
u = ru + u3 .

(2.10)

The fixed point structure in both supercritical and subcritical cases is shown in Fig. 2.5.

23

2.1. TYPES OF BIFURCATIONS

Figure 2.5: Extended phase space (r, u) flow diagrams for the supercritical pitchfork bifurcation u = ru u3 (left), and subcritical pitchfork bifurcation u = ru + u3 (right).

2.1.4

Imperfect bifurcation

The imperfect bifurcation occurs when a symmetry-breaking term is added to the pitchfork.
The normal form contains two control parameters:
u = h + ru u3 .

(2.11)

Here, the constant h breaks the parity symmetry if u u.

This equation arises from a crude model of magnetization dynamics. Let M be the magnetization of a sample, and F (M ) the free energy. Assuming M is small, we can expand
F (M ) as
(2.12)
F (M ) = HM + 12 aM 2 + 41 bM 4 + . . . ,
where H is the external magnetic field, and a and b are temperature-dependent constants.
This is called the Landau expansion of the free energy. We assume b > 0 in order that the
minimum of F (M ) not lie at infinity. The dynamics of M (t) are modeled by
F
dM
=
,
dt
M

(2.13)

with > 0. Thus, the magnetization evolves toward a local minimum in the free energy.
Note that the free energy is a decreasing function of time:


dF
F 2
F dM
=
=
.
(2.14)
dt
M dt
M
By rescaling M uM0 with M0 = (b )1/2 and defining r a and h (3 b)1/2 H, we
obtain the normal form
f
u
f (u) = 21 ru2 + 14 u4 hu .
u = h + ru u3 =

(2.15)
(2.16)

24

CHAPTER 2. BIFURCATIONS

Figure 2.6: Left: scaled free energy f (u) = 12 ru2 + 41 u4 hu, with h = 0.2 hc (blue), h = hc
2
r 3/2 . Right: phase diagram for the imperfect
(green), and h = 2 hc (red), where hc = 3
3
bifurcation u = f (u) = h + ru u3 in the (r, h) plane.
Here, f (u) is a scaled version of the free energy.
Fixed points satisfy the equation
u3 ru h = 0 ,

(2.17)

and correspond to extrema in f (u). By the fundamental theorem of algebra, this cubic
polynomial may be uniquely factorized over the complex plane. Since the coefficients are
real, the complex conjugate u
satisfies the same equation as u, hence there are two possibilities for the roots: either (i) all three roots are real, or (ii) one root is real and the other
two are a complex conjugate pair. Clearly for r < 0 we are in situation (ii) since u3 ru
is then monotonically increasing for u R, and therefore takes the value h precisely once
for u real. For r > 0, there is a region h hc (r), hc (r) over which there are three real
roots. To find hc (r), we demand f (u) = 0 as well as f (u) = 0, which says that two roots
2
r 3/2 .
have merged, forming an inflection point. One easily finds hc (r) = 3
3
Examples of the function f (u) for r > 0 are shown in the left panel of Fig. 2.6 for three
different values of h. For |h| < hc (r) there are three extrema satisfying f (u ) = 0: u1 <
u2 < 0 < u3 , assuming (without loss of generality) that h > 0. Clearly u1 is a local
minimum, u2 a local maximum, and u3 the global minimum of the function f (u). The
phase diagram for this system, plotted in the (r, h) control parameter space, is shown in
the right panel of Fig. 2.6.
In Fig. 2.7 we plot the fixed points u (r) for fixed h. A saddle-node bifurcation occurs
3
|h|2/3 . For h = 0 this reduces to the supercritical pitchfork; for finite
at r = rc (h) = 22/3
h the pitchfork is deformed and even changed topologically. Finally, in Fig. 2.7 we show
the behavior of u (h) for fixed r. When r < 0 the curve retraces itself as h is ramped up
and down, but for r > 0 the system exhibits the phenomenon of hysteresis, i.e. there is an
irreversible aspect to the behavior. Fig, 2.7 shows a hysteresis loop when r > 0.

25

2.2. EXAMPLES

Figure 2.7: Top: extended phase space (r, u) flow diagram for the imperfect pitchfork
bifurcation u = h + ru u3 for h = 1. This is in a sense a deformed supercritical pitchfork.
Bottom: extended phase space (h, u) flow diagram for the imperfect pitchfork bifurcation
r = 0.2 (left panel) and r = 1 (right panel). For r < 0 the behavior is completely reversible.
For r > 0, a regime of irreversibility sets in between hc and +hc , where hc = 2(r/3)3/2 .
The system then exhibits the phenomenon of hysteresis. The dotted vertical lines show the
boundaries of the hysteresis loop.

2.2
2.2.1

Examples
Population dynamics

Consider the dynamics of a harvested population,



N

N = rN 1
H(N ) ,
K

(2.18)

where r, K > 0, and where H(N ) is the harvesting rate.

(a) Suppose H(N ) = H0 is a constant. Sketch the phase flow, and identify and classify all
fixed points.

26

CHAPTER 2. BIFURCATIONS

Figure 2.8: Phase flow for the constantly harvested population, = (1)h, for h = 0.30
(left), h = 0.25 (center), and h = 0.20 (right). The critical harvesting rate is hc = 14 .
Solution : We examing N = f (N ) with
f (N ) = rN

r 2
N H0 .
K

(2.19)

Setting f (N ) = 0 yields N = 21 K. f (N ) is a downward-opening parabola whose maximum

value is f 21 K = 14 rK H0 . Thus, if H0 > 14 rK, the harvesting rate is too large and the
population always shrinks. A saddle-node bifurcation occurs at this value of H0 , and for
larger harvesting rates, there are fixed points at
r
4H0
,
(2.20)
N = 12 K 12 K 1
rK
with N unstable and N+ stable. By rescaling the population = N/K, time = rt and
harvesting rate h = H0 /rK, we arrive at the equation
= (1 ) h .

(2.21)

The critical harvesting rate is then hc = 14 . See fig. 2.8.

(b) One defect of the constant harvesting rate model is that N = 0 is not a fixed point. To
remedy this, consider the following model for H(N )1 :
H(N ) =

B N2
,
N 2 + A2

(2.22)

where A and B are (positive) constants. Show that one can rescale (N, t) to (n, ), such
that

n2
n
dn
2
=n 1
,
(2.23)
d
c
n +1
where and c are positive constants. Provide expressions for n, , , and c.
1
This is a model for the dynamics of the spruce budworm population, taken from ch. 1 of J. D. Murray,
Mathematical Biology (2nd edition, Springer, 1993).

27

2.2. EXAMPLES

Figure 2.9: Plot
of h(n) = n/(n2 + 1) (thick black curve). Straight lines show the function
y(n) = 1 nc for different values of c and . The red line is tangent to the inflection

point of h(n) and determines the minimum value c = 3 3 for a bifurcation. The blue lines
show the construction for determining the location of the two bifurcations for c > c (in
this case, c = 9). See the analysis in the text.
Solution : Examining the denominator of H(N ), we must take N = An. Dividing both
sides of N = f (N ) by B, we obtain
A dN
rA 
A 
n2
=
n 1
n 2
,
B dt
B
K
n +1

from which we glean = Bt/A, = rA/B, and c = K/A.

(c) Show that for c sufficiently small that there is a unique asymptotic ( ) value for
the (scaled) population n, for any given value of . Thus, there are no bifurcations as a
function of the control parameter for c fixed and c < c .
(d) Show that for c > c , there are two bifurcations as a function of , and that for 1 < <
2 the asymptotic solution is bistable, i.e. there are two stable values for n( ). Sketch
the solution set phase diagram in the (c, ) plane. Hint: Sketch the functions (1 n/c)
and n/(n2 + 1). The n 6= 0 fixed points are given by the intersections of these two curves.
Determine the boundary of the bistable region in the (c, ) plane parametrically in terms
of n. Find c and 1 (c) = 2 (c).
Solution (c) and (d) : We examine
dn
= g(n) =
d


n
n
2
1
c
n +1

n.

(2.24)

There is an unstable fixed point at n = 0, where g (0) = > 0. The other fixed points
occur when the term in the curvy brackets vanishes. In fig. 2.9 we plot the function
h(n) n/(n2 + 1) versus n. We seek the intersection of this function with a two-parameter

28

CHAPTER 2. BIFURCATIONS

Figure 2.10: Phase diagram for the equation n = (1 n/c) n n2 /(n2 + 1), labeling n 6= 0
fixed points. (The point n = 0 is always unstable.)
family of straight lines, given by y(n) = (1 n/c). The n-intercept is c and the y-intercept
is . Provided c > c is large enough, there are two bifurcations as a function of , which
we call (c). These are shown as the dashed blue lines in figure 2.9 for c = 9.
Both bifurcations are of the saddle-node type. We determine the curves (c) by requiring
that h(n) is tangent to y(n), which gives two equations:S

n
n
= y(n)
=

n2 + 1
c
1 n2

h (n) = 2
= = y (n) .
(n + 1)2
c

h(n) =

(2.25)
(2.26)

Together, these give (c) parametrically, i.e. as (n) and c(n):

(n) =

2n3
+ 1)2

(n2

c(n) =

2n3
.
1)

(n2

(2.27)

Since h(n) is maximized for n = 1, where h(1) = 21 , there is no bifurcation occurring at

values n < 1. If we plot (n) versus c(n) over the allowed range of n, we obtain the phase
diagram in fig. 2.10. The cusp occurs at (c , ), and is determined by the requirement
that the two bifurcations coincide. This supplies a third condition, namely that h (n) = 0,
where
2n (n2 3)
.
(2.28)
h (n) =
(n2 + 1)3

Hence n = 3, which gives c = 3 3 and =

at any value of .

3 3
8 .

For c < c , there are no bifurcations

29

2.3. APPENDIX I : THE BLETCH

2.3

Appendix I : The Bletch

Problem: The bletch is a disgusting animal native to the Forest of Jkroo on the planet
Barney. The bletch population obeys the equation
dN
= aN 2 bN 3 ,
(2.29)
dt
where N is the number of bletches, and a and b are constants. (Bletches reproduce asexually,
but only when another bletch is watching. However, when there are three bletches around,
they beat the @!!*$&* out of each other.)
(a) Sketch the phase flow for N . (Strange as the bletch is, you can still rule out N < 0.)
Identify and classify all fixed points.
(b) The bletch population is now harvested (they make nice shoes). To model this, we
add an extra term to the dynamics:
dN
= hN + aN 2 bN 3 ,
(2.30)
dt
where h is the harvesting rate. Show that the phase flow now depends crucially on h,
in that there are two qualitatively different flows, depending on whether h < hc (a, b)
or h > hc (a, b). Find the critical harvesting rate hc (a, b) and sketch the phase flows
for the two different regimes.
(c) In equilibrium, the rate at which bletches are harvested is R = hN , where N is
the equilibrium bletch population. Suppose we start with h = 0, in which case N
is given by the value of N at the stable fixed point you found in part (a). Now let
h be increased very slowly from zero. As h is increased, the equilibrium population
changes. Sketch R versus h. What value of h achieves the biggest bletch harvest?
What is the corresponding value of Rmax ?
Solution:
(a) Setting the RHS of eqn. 2.29 to zero suggests the rescaling
N=

a
n
b

t=

b
.
a2

(2.31)

This results in

dn
= n2 n3 .
(2.32)
d
The point n = 0 is a (nonlinearly) repulsive fixed point, and n = 1, corresponding to
N = a/b, is attractive. The flow is shown in fig. 2.11.
By the way, the dynamics can be integrated, using the method of partial fractions, to
yield


1
n 1 n0
1
+ ln

= .
(2.33)
n0 n
n0 1 n

30

CHAPTER 2. BIFURCATIONS

Figure 2.11: Phase flow for the scaled bletch population, n = n2 n3 .

(b) Upon rescaling, the harvested bletch dynamics obeys the equation
dn
= n + n2 n3 ,
d

(2.34)

where = bh/a2 is the dimensionless harvesting rate. Setting the RHS to zero yields
n(n2 n + ) = 0, with solutions n = 0 and
q
(2.35)
n = 12 14 .

At = 14 there is a saddle-node bifurcation, and for > 41 the only fixed point (for
real n) is at n = 0 (stable) the bletch population is then overharvested. For > 41 ,
there areq
three solutions: a stable fixed point at n q
= 0, an unstable fixed point at
1
4 , and a stable fixed point
c = 41 , which means hc = a2 /4b.

n = 21

rate is

at n = 12 +

1
4

. The critical harvesting

Figure 2.12: Phase flow for the harvested bletch population, n = n + n2 n3 .

(c) The scaled bletch harvest is given by r = n+ (). Note R = h N+ =
optimal harvest occurs when n is a maximum, which means we set


q
d 1
1
+ 4 = 0 = opt = 92 .
d 2

a3
b2

r. The

(2.36)

4
, meaning R = 4a3 /27 b2 . Note that at = c = 14
Thus, n+ (opt ) = 23 and ropt = 27
1
1
that n+ (c ) = 2 , hence r(c ) = 8 , which is smaller than (opt ) = 32 . The harvest r()
discontinuously drops to zero at = c , since for > c the flow is to the only stable
fixed point at n = 0.

2.4

Appendix II : Landau Theory of Phase Transitions

Landaus theory of phase transitions is based on an expansion of the free energy of a

thermodynamic system in terms of an order parameter , which is nonzero in an ordered
phase and zero in a disordered phase. For example, the magnetization M of a ferromagnet
in zero external field but at finite temperature typically vanishes for temperatures T > Tc ,
where Tc is the critical temperature, also called the Curie temperature in a ferromagnet. A

2.4. APPENDIX II : LANDAU THEORY OF PHASE TRANSITIONS

31

Figure 2.13: Scaled bletch harvest r versus scaled harvesting rate . Optimal harvesting
occurs for opt = 92 . The critical harvesting rate is c = 14 , at which point the harvest
discontinuously drops to zero.
low order expansion in powers of the order parameter is appropriate sufficiently close to Tc ,
i.e. at temperatures such that the order parameter, if nonzero, is still small.
The simplest example is the quartic free energy,
f (m) = f0 + 12 am2 + 41 bm4 ,

(2.37)

where m is a dimensionless measure of the magnetization density, and where f0 , a, and b

are all functions of the dimensionless
temperature , which in a ferromagnet is the ratio
P
= kB T /J , where J = j Jij is a sum over the couplings. Let us assume b > 0, which is
necessary if the free energy is to be bounded from below2 . The equation of state ,
f
= 0 = am + bm3 ,
m

(2.38)

p
has
p three solutions in the complex m plane: (i) m = 0, (ii) m = a/b , and (iii) m =
a/b . The latter two solutions lie along the (physical) real axis if a < 0. We assume
that a( is monotonically increasing, and that there exists a unique temperature c where
a(c ) = 0. Minimizing f , we find
< c

f = f0

> c

f = f0 .

a2
4b

(2.39)
(2.40)

The free energy is continuous at c since a(c ) = 0. The specific heat, however, is discontinuous across the transition, with
c

c+


2

 2
c a (c )
2 
a
= c 2 
.
=
=c 4b
2b(c )

(2.41)

2
It is always the case that f is bounded from below, on physical grounds. Were b negative, wed have to
consider higher order terms in the Landau expansion.

32

CHAPTER 2. BIFURCATIONS

The presence of a magnetic field h breaks the Z2 symmetry of m m. The free energy
becomes
(2.42)
f (m) = f0 + 21 am2 + 14 bm4 hm ,
and the mean field equation is
bm3 + am h = 0 .

(2.43)

This is a cubic equation for m with real coefficients, and as such it can either have three real
solutions or one real solution and two complex solutions related by complex conjugation.
Clearly we must have a < 0 in order to have three real roots, since bm3 +am is monotonically
increasing otherwise. The boundary between these two classes of solution sets occurs when
two roots coincide, which means f (m) = 0 as well as f (m) = 0. Simultaneously solving
these two equations, we find
2 (a)3/2
h (a) = 3/2
,
(2.44)
3
b1/2
or, equivalently,
3
a (h) = 2/3 b1/3 |h|2/3 .
(2.45)
2
If, for fixed h, we have a < a (h), then there will be three real solutions to the mean field
equation f (m) = 0, one of which is a global minimum (the one for which m h > 0). For
a > a (h) there is only a single global minimum, at which m also has the same sign as h.
If we solve the mean field equation perturbatively in h/a, we find
m(a, h) =
=

2.4.1

h b h3
4 + O(h5 )
a
a

(a > 0)

3 b1/2 h2
h

+ O(h3 )
2 |a|
8 |a|5/2

(a < 0) .

(2.46)
(2.47)

Landau coefficients from mean field theory

A simple variational density matrix for the Ising ferromagnet yields the dimensionless free
energy density


1 + m 1 + m 1 m 1 m
2
1
f (m, h, ) = 2 m hm +
ln
+
ln
. (2.48)
2
2
2
2
When m is small, it is appropriate to expand f (m, h, ), obtaining
f (m, h, ) = ln 2 hm + 21 ( 1) m2 +

12

m4 +

30

m6 +

56

m8 + . . . .

(2.49)

Thus, we identify
a() = 1

b() = 13 .

(2.50)

We see that a() = 0 at a critical temperature c = 1.

The free energy of eqn. 2.48 behaves qualitatively just like it does for the simple Landau
expansion, where one stops at order m4 . Consider without loss of generality the case h > 0.

2.4. APPENDIX II : LANDAU THEORY OF PHASE TRANSITIONS

33

Figure 2.14: Phase diagram for the quartic mean field theory f = f0 + 21 am2 + 14 bm4 hm,
with b > 0. There is a first order line at h = 0 extending from a = and terminating in
a critical point at a = 0. For |h| < h (a) (dashed red line) there are three solutions to the
mean field equation, corresponding to one global minimum, one local minimum, and one
local maximum. Insets show behavior of the free energy f (m).

The minimum of the free energy f (m, h, ) then lies at m > 0 for any . At low temperatures,
the double well structure we found in the h = 0 case is tilted so that the right well lies lower
in energy than the left well. This is depicted in fig. 2.15. As the temperature is raised, the
local minimum at m < 0 vanishes, annihilating with the local maximum in a saddle-node
2f
f
= 0 and m
bifurcation. To find where this happens, one sets m
2 = 0 simultaneously,
resulting in

h () =




1+ 1

1 ln
.
2
1 1

(2.51)



The solutions lie at h = h (). For < c = 1 and h h () , +h () , there are three
solutions to the mean field equation. Equivalently we could in principle invert the above
expression to obtain (h). For > (h), there is only a single global minimum in the free
energy f (m) and there is no local minimum. Note (h = 0) = 1.

34

CHAPTER 2. BIFURCATIONS

Figure 2.15: Mean field free energy f (m) at h = 0.1. Temperatures shown: = 1.2 (red),
= 1.0 (dark green), and = 0.7 (blue).

2.4.2

Magnetization dynamics

Dissipative processes drive physical systems to minimum energy states. We can crudely
model the dissipative dynamics of a magnet by writing the phenomenological equation
f
dm
=
.
dt
m
This drives the free energy f to smaller and smaller values:


df
f 2
f dm
=
=
0.
dt
m dt
m

(2.52)

(2.53)

Clearly the fixed point of these dynamics, where m

= 0, is a solution to the mean field
f
equation m
= 0. At the solution to the mean field equation, one has


m+h
f
=0

m = tanh
.
(2.54)
m

The phase flow for the equation m

= f (m) is shown in fig. 2.16. As we have seen,
for any value of h there is a temperature below which the free energy f (m) has two
local minima and one local maximum. When h = 0 the minima are degenerate, but at
finite h one of the minima is a global minimum. Thus, for < (h) there are three
solutions to the mean field equations. In the language of dynamical systems, under the
dynamics of eqn. 2.52, minima of f (m) correspond to attractive fixed points and maxima
to repulsive fixed points. If h > 0, the rightmost of these fixed points corresponds to the

2.4. APPENDIX II : LANDAU THEORY OF PHASE TRANSITIONS

35

Figure 2.16: Dissipative magnetization dynamics m

= f (m). Bottom panel shows h ()
from eqn. 2.51. For (, h) within the blue shaded region, the free energy f (m) has a global
minimum plus a local minimum and a local maximum. Otherwise f (m) has only a single
global maximum. Top panels show an imperfect bifurcation in the magnetization dynamics
at h = 0.0215 , for which = 0.90 Temperatures shown: = 0.80 (blue), = (h) = 0.90
(green), and = 1.2. The rightmost stable fixed point corresponds to the global minimum
of the free energy. The bottom of the middle two upper panels shows h = 0, where both
of the attractive fixed points and the repulsive fixed point coalesce into a single attractive
fixed point (supercritical pitchfork bifurcation).
global minimum of the free energy. As is increased, this fixed point evolves smoothly. At
= , the (metastable) local minimum and the local maximum coalesce and annihilate in
a saddle-note bifurcation. However at h = 0 all three fixed points coalesce at = c and
the bifurcation is a supercritical pitchfork. As a function of t at finite h, the dynamics are
said to exhibit an imperfect bifurcation, which is a deformed supercritical pitchfork.
The solution set for the mean field equation is simply expressed by inverting the tanh
function to obtain h(, m). One readily finds


1+m

m .
h(, m) = ln
2
1m

(2.55)

36

CHAPTER 2. BIFURCATIONS

Figure 2.17: Top panel : hysteresis as a function of ramping the dimensionless magnetic field
h at = 0.40. Dark red arrows below the curve follow evolution of the magnetization on slow
increase of h. Dark grey arrows above the curve follow evolution of the magnetization on
slow decrease of h. Bottom panel : solution set for m(, h) as a function of h at temperatures
= 0.40 (blue), = c = 1.0 (dark green), and t = 1.25 (red).
As we
panel of fig. 2.17, m(h) becomes multivalued for field values
 see in the bottom

h h () , +h () , where h () is given in eqn. 2.51. Now imagine that < c and we
slowly ramp the field h from a large negative value to a large positive value, and then slowly
back down to its original value. On the time scale of the magnetization dynamics, we can
regard h(t) as a constant. Thus, m(t) will flow to the nearest stable fixed point. Initially
the system starts with m = 1 and h large and negative, and there is only one fixed point,
at m 1. As h slowly increases, the fixed point value m also slowly increases. As h
exceeds h (), a saddle-node bifurcation occurs, and two new fixed points are created at
positive m, one stable and one unstable. The global minimum of the free energy still lies at
the fixed point with m < 0. However, when h crosses h = 0, the global minimum of the free
energy lies at the most positive fixed point m . The dynamics, however, keep the system
stuck in what is a metastable phase. This persists until h = +h (), at which point another
saddle-note bifurcation occurs, and the attractive fixed point at m < 0 annihilates with
the repulsive fixed point. The dynamics then act quickly to drive m to the only remaining
fixed point. This process is depicted in the top panel of fig. 2.17. As one can see from

2.4. APPENDIX II : LANDAU THEORY OF PHASE TRANSITIONS

37

the figure, the the system follows a stable fixed point until the fixed point disappears, even
though that fixed point may not always correspond to a global minimum of the free energy.
The resulting m(h) curve is then not reversible as a function of time, and it possesses a
characteristic shape known as a hysteresis loop. Etymologically, the word hysteresis derives
from the Greek , which means lagging behind. Systems which are hysteretic
exhibit a history-dependence to their status, which is not uniquely determined by external
conditions. Hysteresis may be exhibited with respect to changes in applied magnetic field,
changes in temperature, or changes in other externally determined parameters.

2.4.3

Cubic terms in Landau theory : first order transitions

Next, consider a free energy with a cubic term,

f = f0 + 21 am2 31 ym3 + 41 bm4 ,

(2.56)

with b > 0 for stability. Without loss of generality, we may assume y > 0 (else send
m m). Note that we no longer have m m (i.e. Z2 ) symmetry. The cubic term
favors positive m. What is the phase diagram in the (a, y) plane?
Extremizing the free energy with respect to m, we obtain
f
= 0 = am ym2 + bm3 .
m

(2.57)

This cubic equation factorizes into a linear and quadratic piece, and hence may be solved
simply. The three solutions are m = 0 and
r 
y
y 2 a
m = m
.

(2.58)
2b
2b
b
We now see that for y 2 < 4ab there is only one real solution, at m = 0, while for y 2 > 4ab
there are three real solutions. Which solution has lowest free energy? To find out, we
compare the energy f (0) with f (m+ )3 . Thus, we set
f (m) = f (0)

2
1
2 am

31 ym3 + 41 bm4 = 0 ,

(2.59)

and we now have two quadratic equations to solve simultaneously:

0 = a ym + bm2

(2.60)
2

0 = 12 a 13 ym + 41 bm = 0 .

(2.61)

Eliminating the quadratic term gives m = 3a/y. Finally, substituting m = m+ gives us a

relation between a, b, and y:
(2.62)
y 2 = 29 ab .
3

We neednt waste our time considering the m = m solution, since the cubic term prefers positive m.

38

CHAPTER 2. BIFURCATIONS

Figure 2.18: Behavior of the quartic free energy f (m) = 12 am2 13 ym3 + 14 bm4 . A: y 2 < 4ab
; B: 4ab < y 2 < 29 ab ; C and D: y 2 > 92 ab. The thick black line denotes a line of first order
transitions, where the order parameter is discontinuous across the transition.
Thus, we have the following:
y2
4b
y2
2y 2
>a>
4b
9b
2y 2
>a
9b
a>

1 real root m = 0

3 real roots; minimum at m = 0

y
3 real roots; minimum at m =
+
2b

r

y 2 a

2b
b

The solution m = 0 lies at a local minimum of the free energy for a > 0 and at a local
2
2
maximum for a < 0. Over the range y4b > a > 2y
9b , then, there is a global minimum at
m = 0, a local minimum at m = m+ , and a local maximum at m = m , with m+ > m > 0.
2
For 2y
9b > a > 0, there is a local minimum at a = 0, a global minimum at m = m+ , and
a local maximum at m = m , again with m+ > m > 0. For a < 0, there is a local
maximum at m = 0, a local minimum at m = m , and a global minimum at m = m+ , with
m+ > 0 > m . See fig. 2.18.

2.4.4

Magnetization dynamics

Suppose we now impose some dynamics on the system, of the simple relaxational type
f
dm
=
,
dt
m

(2.63)

2.4. APPENDIX II : LANDAU THEORY OF PHASE TRANSITIONS

39

where is a phenomenological kinetic coefficient. Assuming y > 0 and b > 0, it is convenient

to adimensionalize by writing
m

y
u
b

y2
r
b

b
s .
y2

(2.64)

Then we obtain

=
,
s
u
where the dimensionless free energy function is

(2.65)

(u) = 12 ru2 13 u3 + 41 u4 .

(2.66)

We see that there is a single control parameter, r. The fixed points of the dynamics are
then the stationary points of (u), where (u) = 0, with
(u) = u (
r u + u2 ) .

(2.67)

The solutions to (u) = 0 are then given by

u = 0

u =

1
2

1
4

r .

(2.68)

For r > 14 there is one fixed point at u = 0, which is attractive under the dynamics
u = (u) since (0) = r. At r = 41 there occurs a saddle-node bifurcation and a pair of
fixed points is generated, one stable and one unstable. As we see from fig. 2.14, the interior
fixed point is always unstable and the two exterior fixed points are always stable. At r = 0
there is a transcritical bifurcation where two fixed points of opposite stability collide and
bounce off one another (metaphorically speaking).
1
, which
At the saddle-node bifurcation, r = 14 and u = 21 , and we find (u = 12 ; r = 14 ) = 192
is positive. Thus, the thermodynamic state of the system remains at u = 0 until the value
of (u+ ) crosses zero. This occurs when (u) = 0 and (u) = 0, the simultaneous solution
of which yields r = 29 and u = 23 .

Suppose we slowly ramp the control parameter r up and down as a function of the dimensionless time s. Under the dynamics of eqn. 2.65, u(s) flows to the first stable fixed
point encountered this is always the case for a dynamical system with a one-dimensional
phase space. Then as r is further varied, u follows

the position of whatever locally stable
fixed point it initially encountered. Thus, u r(s) evolves smoothly until a bifurcation is
encountered. The situation is depicted by the arrows in fig. 2.19. The equilibrium thermodynamic value for u(
r ) is discontinuous; there is a first order phase transition at r = 29 , as
weve already seen. As r is increased, u(
r ) follows a trajectory indicated by the magenta
arrows. For an negative initial value of u, the evolution as a function of r will be reversible.
However, if u(0) is initially positive, then the system exhibits hysteresis, as shown. Starting
with a large positive value of r, u(s) quickly evolves to u = 0+ , which means a positive
infinitesimal value. Then as r is decreased, the system remains at u = 0+ even through the
first order transition, because u = 0 is an attractive fixed point. However, once r begins to
go negative, the u = 0 fixed point becomes repulsive, and u(s) quickly flows to the stable

40

CHAPTER 2. BIFURCATIONS

Figure 2.19: Fixed points for (u) = 21 ru2 31 u3 + 14 u4 and flow under the dynamics
u = (u). Solid curves represent stable fixed points and dashed curves unstable fixed
points. Magenta arrows show behavior under slowly increasing control parameter r and
dark blue arrows show behavior under slowly decreasing r. For u > 0 there is a hysteresis
loop. The thick black curve shows the equilibrium thermodynamic value of u(
r ), i.e. that
value which minimizes the free energy (u). There is a first order phase transition at r = 29 ,
where the thermodynamic value of u jumps from u = 0 to u = 23 .
q
fixed point u+ = 12 + 14 r. Further decreasing r, the system remains on this branch. If
r is later increased, then u(s) remains on the upper branch past
q r = 0, until the u+ fixed

point annihilates with the unstable fixed point at u =

quickly flows down to u = 0+ again.

2.4.5

1
2

1
4

r, at which time u(s)

Sixth order Landau theory : tricritical point

Finally, consider a model with Z2 symmetry, with the Landau free energy
f = f0 + 21 am2 + 41 bm4 + 16 cm6 ,

(2.69)

with c > 0 for stability. We seek the phase diagram in the (a, b) plane. Extremizing f with
respect to m, we obtain
f
= 0 = m (a + bm2 + cm4 ) ,
(2.70)
m

2.4. APPENDIX II : LANDAU THEORY OF PHASE TRANSITIONS

41

Figure 2.20: Behavior of the sextic free energy f (m) = 12 am2 + 14 bm4 + 61 cm6 . A: a > 0 and

b > 0 ; B: a < 0 and b > 0 ; C: a < 0 and b < 0 ; D: a > 0 and b < 43 ac ; E: a > 0

and 43 ac < b < 2 ac ; F: a > 0 and 2 ac < b < 0. The thick dashed line is a line
of second order transitions, which meets the thick solid line of first order transitions at the
tricritical point, (a, b) = (0, 0).
which is a quintic with five solutions over the complex m plane. One solution is obviously
m = 0. The other four are
v
s 
u
u b
b 2 a
t
(2.71)
.
m=
2c
2c
c
For each symbol in the above equation, there are two options, hence four roots in all.

If a > 0 and b > 0, then four of the roots are imaginary and there is a unique minimum at
m = 0.
For a < 0, there are only three solutions to f (m) = 0 for real m, since the choice for
the sign under the radical leads to imaginary roots. One of the solutions is m = 0. The

42

CHAPTER 2. BIFURCATIONS

other two are

m=

b
+
2c

r

b 2 a
.
2c
c

(2.72)

The most interesting situation is a > 0 and b < 0. If a > 0 and b < 2 ac, all five roots
are real. There must be three minima, separated by two local maxima. Clearly if m is a
solution, then so is m . Thus, the only question is whether the outer minima are of lower
energy than the minimum at m = 0. We assess this by demanding f (m ) = f (0), where
m is the position of the largest root (i.e. the rightmost minimum). This gives a second
quadratic equation,
(2.73)
0 = 12 a + 14 bm2 + 61 cm4 ,
which together with equation 2.70 gives
b = 43

ac .

(2.74)

Thus, we have the following, for fixed a > 0:

b > 2 ac

2 ac > b > 43 ac
43

ac > b

1 real root m = 0

5 real roots; minimum at m = 0

s

b
5 real roots; minima at m = +
2c

r

b 2 a

2c
c

The point (a, b) = (0, 0), which lies at the confluence of a first order line and a second order
line, is known as a tricritical point.

2.4.6

Hysteresis for the sextic potential

Once again, we consider the dissipative dynamics m

= f (m). We adimensionalize by
writing
r
b2
c
|b|
u
,
a
r
,
t
s .
(2.75)
m
c
c
b2
Then we obtain once again the dimensionless equation

u
=
,
s
u

(2.76)

(u) = 12 ru2 14 u4 + 61 u6 .

(2.77)

where
In the above equation, the coefficient of the quartic term is positive if b > 0 and negative
if b < 0. That is, the coefficient is sgn(b). When b > 0 we can ignore the sextic term for
sufficiently small u, and we recover the quartic free energy studied earlier. There is then a
second order transition at r = 0. .

43

2.4. APPENDIX II : LANDAU THEORY OF PHASE TRANSITIONS

Figure 2.21: Free energy (u) = 21 ru2 14 u4 + 16 u6 for several different values of the control
parameter r.
New and interesting behavior occurs for b > 0. The fixed points of the dynamics are
obtained by setting (u) = 0. We have
(u) = 12 ru2 14 u4 + 61 u6

(2.78)

(u) = u (
ru +u ) .

(2.79)

Thus, the equation (u) = 0 factorizes into a linear factor u and a quartic factor u4 u2 + r
which is quadratic in u2 . Thus, we can easily obtain the roots:
r < 0

u =0, u =

0 < r <

1
4

u =0, u =

r >

1
4

u = 0 .

1
2
1
2

+
+

q
q

1
4
1
4

(2.80)

r , u =

1
2

1
4

(2.81)
(2.82)

In fig. 2.22, we plot the fixed points and the hysteresis loops for this system. At r = 14 ,
there are two symmetrically located saddle-node bifurcations at u = 12 . We find (u =

12 , r = 14 ) =

1
48 ,

which is positive, indicating that the stable fixed point u = 0 remains

the thermodynamic minimum for the free energy (u) as r is decreased through r =

Setting (u) = 0 and (u) = 0 simultaneously, we obtain r =

3
16

1
4.

and u = 23 . The

thermodynamic value for u therefore jumps discontinuously from u = 0 to u =

3
; this is a first order transition.
branch) at r = 16

3
2

(either

44

CHAPTER 2. BIFURCATIONS

Figure 2.22: Fixed points (u ) = 0 for the sextic potential (u) = 12 ru2 14 u4 + 61 u6 ,
and corresponding dynamical flow (arrows) under u = (u). Solid curves show stable
fixed points and dashed curves show unstable fixed points. The thick solid black and solid
grey curves indicate the equilibrium thermodynamic values for u; note the overall u u
symmetry. Within the region r [0, 14 ] the dynamics are irreversible and the system exhibits
3
.
the phenomenon of hysteresis. There is a first order phase transition at r = 16

Under the dissipative dynamics considered here, the system exhibits hysteresis, as indicated
in the figure, where the arrows show the evolution of u(s) for very slowly varying r(s). When
the control parameter r is large and positive, the flow is toward the sole fixed point at u = 0.
At r = 41 , two simultaneous saddle-node bifurcations take place at u = 12 ; the outer
branch is stable and the inner branch unstable in both cases. At r = 0 there is a subcritical
pitchfork bifurcation, and the fixed point at u = 0 becomes unstable.
Suppose one starts off with r 41 with some value u > 0. The flow u = (u) then
rapidly results in u 0+ . This is the high temperature phase in which there is no
magnetization. Now let r increase slowly,
using s as the dimensionless time variable. The

scaled magnetization u(s) = u r(s) will remain pinned at the fixed point u = 0+ . As
r passes through r = 14 , two new stable values of u appear, but our system remains at
u = 0+ , since u = 0 is a stable fixed point. But after the subcritical pitchfork, u = 0
becomes unstable. The magnetization u(s) then flows rapidly to the stable fixed point at

1/2
u = 12 , and follows the curve u (
for all r < 0.
r ) = 21 + ( 14 r)1/2
Now suppose we start increasing r (i.e. increasing temperature). The magnetization follows

2.4. APPENDIX II : LANDAU THEORY OF PHASE TRANSITIONS

45

1/2
the stable fixed point u (
r ) = 12 + ( 41 r)1/2
past r = 0, beyond the first order phase
3
, and all the way up to r = 14 , at which point this fixed point is
transition point at r = 16
annihilated at a saddle-node bifurcation. The flow then rapidly takes u u = 0+ , where
it remains as r continues to be increased further.


Within the region r 0, 41 of control parameter space, the dynamics are said to be
irreversible and the behavior of u(s) is said to be hysteretic.

46

CHAPTER 2. BIFURCATIONS

Chapter 3

Two-Dimensional Phase Flows

Weve seen how, for one-dimensional dynamical systems u = f (u), the possibilities in terms
of the behavior of the system are in fact quite limited. Starting from an arbitrary initial
condition u(0), the phase flow is monotonically toward the first stable fixed point encountered. (That point may lie at infinity.) No oscillations are possible1 . For N = 2 phase
flows, a richer set of possibilities arises, as we shall now see.

3.1
3.1.1

Harmonic Oscillator and Pendulum

Simple harmonic oscillator

A one-dimensional harmonic oscillator obeys the equation of motion,

m

d2 x
= kx ,
dt2

(3.1)

where m is the mass and k the force constant (of a spring). If we define v = x,
this may be
written as the N = 2 system,
d
dt
where =
known:

  
  

x
0
1
x
v
=
=
,
v
2 0
v
2 x

(3.2)

k/m has the dimensions of frequency (inverse time). The solution is well
v0
sin(t)

v(t) = v0 cos(t) x0 sin(t) .

x(t) = x0 cos(t) +

(3.3)
(3.4)

1
If phase space itself is multiply connected, e.g. a circle, then the system can oscillate by moving around
the circle.

47

48

CHAPTER 3. TWO-DIMENSIONAL PHASE FLOWS

Figure 3.1: Phase curves for the harmonic oscillator.

The phase curves are ellipses:
x2 (t) + 1 v 2 (t) = C ,

(3.5)

where the constant C = x20 + 1 v02 . A sketch of the phase curves and of the phase flow
is shown in Fig. 3.1. Note that the x and v axes have different dimensions. Note also that
the origin is a fixed point, however, unlike the N = 1 systems studied in the first lecture,
here the phase flow can avoid the fixed points, and oscillations can occur.
Incidentally, eqn. 3.2 is linear, and may be solved by the following method. Write the
equation as = M , with
 


x
0
1
=
and
M=
(3.6)
x
2 0
The formal solution to = M is
(t) = eM t (0) .

(3.7)

What do we mean by the exponential of a matrix? We mean its Taylor series expansion:
eM t = I + M t +

1
2!

M 2 t2 +

1
3!

M 3 t3 + . . . .

(3.8)

Note that




0
1
0
1
M =
2 0
2 0


2
0
=
= 2 I ,
0
2
2

(3.9)

hence
M 2k = ( 2 )k I

M 2k+1 = ( 2 )k M .

(3.10)

49

3.1. HARMONIC OSCILLATOR AND PENDULUM

Thus,
Mt

X
k=0

X
1
1
( 2 t2 )k I +
( 2 t2 )k M t
(2k)!
(2k + 1)!
k=0

= cos(t) I + 1 sin(t) M
=

cos(t)
1 sin(t)
sin(t)
cos(t)

(3.11)

Plugging this into eqn. 3.7, we obtain the desired solution.

For the damped harmonic oscillator, we have
2

x
+ 2 x + x = 0

M=

0
1
2
2

(3.12)

The phase curves then spiral inward to the fixed point at (0, 0).

3.1.2

Pendulum

Next, consider the simple pendulum, composed of a mass point m affixed to a massless rigid
rod of length .

This is equivalent to
d
dt

m2 = mg sin .

(3.13)

  


=
,

2 sin

(3.14)

where = is the angular velocity, and where =

oscillations.

g/ is the natural frequency of small

The phase curves for the pendulum are shown in Fig. 3.2. The small oscillations of the
pendulum are essentially the same as those of a harmonic oscillator. Indeed, within the
small angle approximation, sin , and the pendulum equations of motion are exactly
those of the harmonic oscillator. These oscillations are called librations. They involve
a back-and-forth motion in real space, and the phase space motion is contractable to a
point, in the topological sense. However, if the initial angular velocity is large enough, a
qualitatively different kind of motion is observed, whose phase curves are rotations. In this
case, the pendulum bob keeps swinging around in the same direction, because, as well see
in a later lecture, the total energy is sufficiently large. The phase curve which separates
these two topologically distinct motions is called a separatrix .

50

CHAPTER 3. TWO-DIMENSIONAL PHASE FLOWS

Figure 3.2: Phase curves for the simple pendulum. The separatrix divides phase space into
regions of vibration and libration.

3.2

General N = 2 Systems

The general form to be studied is

d
dt

 
x
=
y

Vx (x, y)
Vy (x, y)

Special cases include autonomous second order ODEs, viz.


  
d x
v
x
= f (x, x)

=
,
=
f (x, v)
dt v

(3.15)

(3.16)

of the type which occur in one-dimensional mechanics.

3.2.1

The damped driven pendulum

Another example is that of the damped and driven harmonic oscillator,

d2
d
+
+ sin = j .
2
ds
ds

(3.17)

This is equivalent to a model of a resistively and capacitively shunted Josephson junction,

depicted in fig. 3.3. If is the superconducting phase difference across the junction, the

3.2. GENERAL N = 2 SYSTEMS

51

Figure 3.3: . The resistively and capacitively shunted Josephson junction. The Josephson
junction is the X element at the bottom of the figure.
current through the junction is given by IJ = Ic sin , where Ic is the critical current.
The current carried by the resistor is IR = V /R from Ohms law, and the current from
Finally, the Josephson relation relates the voltage V across the
the capacitor is IC = Q.
Summing up the parallel
junction to the superconducting phase difference : V = (~/2e) .
currents, we have that the total current I is given by
~
~C
+
+ Ic sin ,
2e
2eR
which, again, is equivalent to a damped, driven pendulum.
I=

(3.18)

This system also has a mechanical analog. Define the potential

U () = Ic cos I .

(3.19)

~C
~
U
.
+
=
2e
2eR

(3.20)

The equation of motion is then

Thus, the combination ~C/2e plays the role of the inertial term (mass, or moment of inertia),
while the combination ~/2eR plays the role of a damping coefficient. The potential U ()
is known as the tilted washboard potential, for obvious reasons. (Though many of you have
perhaps never seen a washboard.)
The model is adimensionalized by defining the Josephson plasma frequency p and the RC
time constant :
r
2eIc
,
RC .
(3.21)
p
~C
The dimensionless combination p then enters the adimensionalized equation as the sole
control parameter:
I
d2
1 d
= 2+
+ sin ,
(3.22)
Ic
ds
p ds

52

CHAPTER 3. TWO-DIMENSIONAL PHASE FLOWS

where s = p t. In the Josephson junction literature, the quantity 2eIc R2 C/~ = (p )2 ,

known as the McCumber-Stewart parameter, is a dimensionless measure of the damping
(large means small damping). In terms of eqn. 3.17, we have = (p )1 and j = I/Ic .
We can write the second order ODE of eqn. 3.17 as two coupled first order ODEs:
d
dt

  


=
,

j sin

(3.23)

Phase space is a cylinder, S1 R1 .

where = .
The quantity p typically ranges from 103 to 103 in Josephson junction applications.
If p is small, then the system is heavily damped, and the inertial term d2/ds2 can be
neglected. One then obtains the N = 1 system

d
= j sin .
ds

(3.24)

If |j| < 1, thenp(s) evolves to the first stable fixed point encountered, where = sin1 (j)
and cos = 1 j 2 . Since (s) is asymptotically a constant, the voltage drop V
This, there is
must then vanish, as a consequence of the Josephson relation V = (~/2e) .
current flowing with no voltage drop!
If |j| > 1, the RHS never vanishes, in which case (s) is monotonic. We then can integrate
the differential equation
~
d
dt =

.
(3.25)
2eR I Ic sin
Asymptotically the motion is periodic, with the period T obtained by integrating over the
interval [0, 2]. One finds
2
~
.
(3.26)
p
T =
2eR
I 2 Ic2
The time-averaged voltage drop is then
hV i =

p
~
~ 2
hi =

= R I 2 Ic2 .
2e
2e T

(3.27)

This is the physics of the current-biased resistively and capacitively shunted Josephson junction in the strong damping limit. It is current-biased because we are specifying the current
I. Note that Ohms law is recovered at large values of I.
For general p , we can still say quite a bit. At a fixed point, both components of the
vector field V (, ) must vanish. This requires = 0 and j = sin . Therefore, there are
two fixed points for |j| < 1, one a saddle point and the other a stable spiral. For |j| > 1
there are no fixed points, and asymptotically the function (t) tends to a periodic limit
cycle LC (t). The flow is sketched for two representative values of j in Fig. 3.4.

3.2. GENERAL N = 2 SYSTEMS

53

Figure 3.4: Phase flows for the equation + 1 + sin = j. Left panel: 0 < j < 1; note
the separatrix (in black), which flows into the stable and unstable fixed points. Right panel:
j > 1. The red curve overlying the thick black dot-dash curve is a limit cycle.

3.2.2

Classification of N = 2 fixed points

Suppose we have solved the fixed point equations Vx (x , y ) = 0 and Vy (x , y ) = 0. Let us

now expand about the fixed point, writing



V
Vx 

x
(3.28)
x =
 (x x ) +
 (y y ) + . . .
x 
y 
(x ,y )
(x ,y )




Vy 
Vy 
(3.29)
y =
 (x x ) +
 (y y ) + . . . .
x 
y 
(x ,y )

(x ,y )

We define

u1 = x x

which, to linear order, satisfy

u2 = y y ,

(3.30)

z }| {
a
b
u1
u1
d
+ O(u2 ) .

=
dt
u2
c
d
u2

(3.31)

u(t) = exp(M t) u(0) ,

(3.32)

The formal solution to u = M u is

54

CHAPTER 3. TWO-DIMENSIONAL PHASE FLOWS

Figure 3.5: Complete classification of fixed points for the N = 2 system.

where exp(M t) =

1
n=0 n!

(M t)n is the exponential of the matrix M t.

The behavior of the system is determined by the eigenvalues of M , which are roots of the
characteristic equation P () = 0, where
P () = det(I M )

= 2 T + D ,

(3.33)

with T = a + d = Tr(M ) and D = ad bc = det(M ). The two eigenvalues are therefore



p
(3.34)
= 21 T T 2 4D .
To see why the eigenvalues control the behavior, let us expand u(0) in terms of the eigenvectors of M . Since M is not necessarily symmetric, we should emphasize that we expand
u(0) in terms of the right eigenvectors of M , which satisfy
M a = a a ,
where the label a runs over the symbols + and , as in (3.34). We write
X
u(t) =
Ca (t) a .

(3.35)

(3.36)

Since (we assume) the eigenvectors are linearly independent, the equation u = M u becomes
C a = a Ca ,

(3.37)

Ca (t) = ea t Ca (0) .

(3.38)

with solution
Thus, the coefficients of the eigenvectors a will grow in magnitude if |a | > 1, and will
shrink if |a | < 1.

3.2. GENERAL N = 2 SYSTEMS

55

Figure 3.6: Fixed point zoo for N = 2 systems. Not shown: unstable versions of node,
spiral, and star (reverse direction of arrows to turn stable into unstable).

3.2.3

The fixed point zoo

Saddles : When D < 0, both eigenvalues are real; one is positive and one is negative,
i.e. + > 0 and < 0. The right eigenvector is thus the stable direction while
+ is the unstable direction.
Nodes : When 0 < D < 41 T 2 , both eigenvalues are real and of the same sign. Thus,
both right eigenvectors correspond to stable or to unstable directions, depending on
whether T < 0 (stable; < + < 0) or T < 0 (unstable; + > > 0). If are
distinct, one can distinguish fast and slow eigendirections, based on the magnitude of
the eigenvalues.
Spirals : When D > 41 T 2 , the discriminant T 2 4D is negative, and the eigenvalues
come in a complex conjugate pair: = + . The real parts are given by Re( ) = 21 T ,
so the motion is stable (i.e. collapsing to the fixed point) if T < 0 and unstable (i.e.
diverging from the fixed point) if T > 0. The motion is easily shown to correspond
to a spiral. One can check that the spiral rotates counterclockwise for a > d and
clockwise for a < d.

Degenerate Cases : When T = 0 we have = D. For D < 0 we have a

saddle, but for D > 0 both eigenvalues are imaginary: = i D. The orbits do
not collapse to a point, nor do they diverge to infinity, in the t limit, as they do

56

CHAPTER 3. TWO-DIMENSIONAL PHASE FLOWS

Figure 3.7: Phase portrait for an N = 2 flow including saddles (A,C), unstable spiral (B),
and limit cycle (D).
in the case of the stable and unstable spiral. The fixed point is called a center , and
it is surrounded by closed trajectories.
When D = 14 T 2 , the discriminant vanishes and the eigenvalues are degenerate. If the
rank of M is two, the fixed point is a stable (T < 0) or unstable (T > 0) star . If M
is degenerate and of rank one, the fixed point is a degenerate node.
When D = 0, one of the eigenvalues vanishes. This indicates a fixed line in phase
space, since any point on that line will not move. The fixed line can be stable or
unstable, depending on whether the remaining eigenvalue is negative (stable, T < 0),
or positive (unstable, T > 0).
Putting it all together, an example of a phase portrait is shown in Fig. 3.7. Note the
presence of an isolated, closed trajectory, which is called a limit cycle. Many self-sustained
physical oscillations, i.e. oscillations with no external forcing, exhibit limit cycle behavior.
Limit cycles, like fixed points, can be stable or unstable, or partially stable. Limit cycles
are inherently nonlinear. While the linear equation = M can have periodic solutions
if M has purely imaginary eigenvalues, these periodic trajectories are not isolated, because
(t) is also a solution. The amplitude of these linear oscillations is fixed by the initial
conditions, whereas for limit cycles, the amplitude is inherent from the dynamics itself, and
the initial conditions are irrelevant (for a stable limit cycle).
In fig. 3.8 we show simple examples of stable, unstable, and half-stable limit cycles. As we
shall see when we study nonlinear oscillations, the Van der Pol oscillator,
x
+ (x2 1) x + x = 0 ,

(3.39)

with > 0 has a stable limit cycle. The physics is easy to apprehend. The coefficient of the
x term in the equation of motion is positive for |x| > 1 and negative for |x| < 1. Interpreting
this as a coefficient of friction, we see that the friction is positive, i.e. dissipating energy,
when |x| > 1 but negative, i.e. accumulating energy, for |x| < 1. Thus, any small motion
with |x| < 1 is amplified due to the negative friction, and would increase without bound

3.2. GENERAL N = 2 SYSTEMS

57

Figure 3.8: Stable, unstable, and half-stable limit cycles.

were it not for the fact that the friction term reverses its sign and becomes dissipative for
|x| > 1. The limit cycle for 1 is shown in fig. 3.9.

3.2.4

Fixed points for N = 3 systems

For an N = 2 system, there are five generic types of fixed points. They are classified
according to the eigenvalues of the linearized dynamics at the fixed point. For a real 2 2
matrix, the eigenvalues must be real or else must be a complex conjugate pair. The five
types of fixed points are then

1 > 0 , 2 > 0

(1) unstable node

1 > 0 , 2 < 0

(2) saddle point

1 < 0 , 2 < 0

(3) stable node

1
1

(4) unstable spiral

(5) stable spiral

Re 1 > 0 , 2 =
Re 1 < 0 , 2 =

How many possible generic fixed points are there for an N = 3 system?
For a general real 3 3 matrix M , the characteristic polynomial P () = det( M ) satisfies
P ( ) = P (). Thus, if is a root then so is . This means that the eigenvalues are either
real or else come in complex conjugate pairs. There are then ten generic possibilities for

58

CHAPTER 3. TWO-DIMENSIONAL PHASE FLOWS

Figure 3.9: Limit cycle of the Van der Pol oscillator for 1. (Source: Wikipedia)
the three eigenvalues:
(1)

unstable node

1 > 2 > 3 > 0

(2)

(+ + ) saddle

1 > 2 > 0 > 3

1 > 0 > 2 > 3

stable note

0 > 1 > 2 > 3

(5)

unstable spiral-node

(6)

unstable spiral-node

1 > Re 2,3 > 0 ; Im 2 = Im 3

(7)

stable spiral-node

(8)

stable spiral-node

(9)

(+ ) spiral-saddle

(3)
(4)

(10)

3.3

(+ ) saddle

(+ + ) spiral-saddle

Re 1,2 > 3 > 0 ; Im 1 = Im 2

0 > 1 > Re 2,3 ; Im 2 = Im 3

0 > Re 1,2 > 3 ; Im 1 = Im 2

1 > 0 > Re 2,3 ; Im 2 = Im 3

Re 1,2 > 0 > 3 ; Im 1 = Im 2 .

Andronov-Hopf Bifurcation

A bifurcation between a spiral and a limit cycle is known as an Andronov-Hopf bifurcation.

As a simple example, consider the N = 2 system,
x = ax by C(x2 + y 2 ) x

y = bx + ay C(x2 + y 2 ) y ,

(3.40)
(3.41)

where a, b, and C are real. Clearly the origin is a fixed point, at which one finds the
eigenvalues = a ib. Thus, the fixed point is a stable spiral if a < 0 and an unstable
spiral if a > 0.

3.3. ANDRONOV-HOPF BIFURCATION

59

Figure 3.10: Hopf bifurcation: for C > 0 the bifurcation is supercritical, between stable
spiral and stable limit cycle. For C < 0 the bifurcation is subcritical, between unstable
spiral and unstable limit cycle. The bifurcation occurs at a = 0 in both cases.

Written in terms of the complex variable z = x + iy, these two equations collapse to the
single equation
z = (a + ib) z C |z|2 z .

(3.42)

The dynamics are also simple in polar coordinates r = |z|, = arg(z):

r = ar Cr 3
= b .

(3.43)
(3.44)

The phase diagram,

pfor fixed b > 0, is depicted in Fig. 3.10. For positive a/C, there is a
limit cycle at r = a/C. In both cases, the limit cycle disappears as a crosses the value
a = 0 and is replaced by a stable (a < 0, C > 0) or unstable (a > 0, C < 0) spiral.
This example also underscores the following interesting point. Adding a small nonlinear
term C has no fundamental effect on the fixed point behavior so long as a 6= 0, when the
fixed point is a stable or unstable spiral. In general, fixed points which are attractors (stable
spirals or nodes), repellers (unstable spirals or nodes), or saddles are robust with respect
to the addition of a small nonlinearity. But the fixed point behavior in the marginal cases
centers, stars, degenerate nodes, and fixed lines is strongly affected by the presence of
even a small nonlinearity. In this example, the FP is a center when a = 0. But as the
(r, ) dynamics shows, a small nonlinearity will destroy the center and turn the FP into an
attractor (C > 0) or a repeller (C < 0).

60

3.4

CHAPTER 3. TWO-DIMENSIONAL PHASE FLOWS

Population Biology : Lotka-Volterra Models

Consider two species with populations N1 and N2 , respectively. We model the evolution of
these populations by the coupled ODEs
dN1
= aN1 + bN1 N2 + cN12
dt

(3.45)

dN2
= dN2 + eN1 N2 + f N22 ,
dt

(3.46)

where {a, b, c, d, e, f } are constants. We can eliminate some constants by rescaling N1,2 .
This results in the following:


x = x r x ky
(3.47)

y = y r y k x ,
(3.48)

where , and can each take on one of three possible values {0, 1}. By rescaling time,
we can eliminate the scale of either of r or r as well. Typically, intra-species competition
guarantees = = +1. The remaining coefficients (r, k, k ) are real may also be of either
sign. The values and especially the signs of the various coefficients have a physical (or
biological) significance. For example, if k < 0 it means that x grows due to the presence of
y. The effect of y on x may be of the same sign (kk > 0) or of opposite sign (kk < 0).

3.4.1

Rabbits and foxes

As an example, consider the model

x = x xy

y = y + xy .

(3.49)
(3.50)

The quantity x might represent the (scaled) population of rabbits and y the population of
foxes in an ecosystem. There are two fixed points: at (0, 0) and at (, 1). Linearizing the
dynamics about these fixed points, one finds that (0, 0) is a saddle while (, 1) is a center.
Lets do this explicitly.
The first step is to find the fixed points (x , y ). To do this, we set x = 0 and y = 0. From
x = x(1 y) = 0 we have that x = 0 or y = 1. Suppose x = 0. The second equation,
y = (x )y then requires y = 0. So P1 = (0, 0) is a fixed point. The other possibility is
that y = 1, which then requires x = . So P2 = (, 1) is the second fixed point. Those are
the only possibilities.
We now compute the linearized dynamics at these fixed points. The linearized dynamics
are given by = M , with

1y
x
x/x

x/y

.
=
(3.51)
M =
y
x
y/x

y/y

3.4. POPULATION BIOLOGY : LOTKA-VOLTERRA MODELS

Evaluating M at P1 and P2 , we find

1
0

M1 =
0

The eigenvalues are easily found:

M2 =

p
P 2 : + = i
p
= i .

P 1 : + = 1
=
Thus P1 is a saddle point and P2 is a center.

61

(3.52)

(3.53)
(3.54)

As we saw earlier, generally speaking we expect nonlinear terms to transform centers to

stable or unstable spirals, possibly with a limit cycle. However for the Lotka-Volterra
system there is a conserved quantity. Consider the general predator-prey system
x = (a b y) x

(3.55)

y = (c d x) y ,

(3.56)

where a, b, c, and d are all positive constants. Now consider the function
H d x + b y c ln x a ln y .
Then

c
H
=d
x
x

Thus, we have x = xy

H
y

and y = xy
q =

H
p

,
H
x .

H
a
=b .
y
y

(3.57)

(3.58)

If we define p ln x and q ln y, then we have

,

p =

H
q

(3.59)

with
H(q, p) = d ep + b eq c p a q .

(3.60)

So the system is a Hamiltonian system in disguise, and we know that for Hamiltonian
systems the only possible fixed points are saddles and centers. The phase curves are level
sets of the function H.

3.4.2

Rabbits and sheep

In the rabbits and foxes model of eqs. 3.49, 3.50, the rabbits are the food for the foxes. This
means k = 1 but k = 1, i.e. the fox population is enhanced by the presence of rabbits,
but the rabbit population is diminished by the presence of foxes. Consider now a model in
which the two species (rabbits and sheep, say) compete for food:
x = x (r x ky)

y = y (1 y k x) ,

(3.61)
(3.62)

62

CHAPTER 3. TWO-DIMENSIONAL PHASE FLOWS

Figure 3.11: Phase flow for the rabbits vs. foxes Lotka-Volterra model of eqs. 3.49, 3.50.
with r, k, and k all positive. Note that when either population x or y vanishes, the
remaining population is governed by the logistic equation, i.e. it will flow to a nonzero fixed
point.
The matrix of derivatives, which is to be evaluated at each fixed point in order to assess its
stability, is

x/x

M=
y/x

x/y

y/y

r 2x ky
k y

kx
1 2y

k x

(3.63)

At each fixed point, we must evaluate D = det(M ) and T = Tr (M ) and apply the classification scheme of Fig. 3.5.
P1 = (0, 0) : This is the trivial statewith no rabbits (x = 0) and no sheep (y = 0).
r 0
The linearized dynamics gives M1 =
, which corresponds to an unstable node.
0 1
P2 = (r,
we have rabbits but no sheep. The linearized dynamics gives
 0) : Here 
r
rk
M2 =
. For rk > 1 this is a stable node; for rk < 1 it is a saddle
0 1 rk
point.
P3 = 
(0, 1) : Here
 we have sheep but no rabbits. The linearized dynamics gives
rk 0
M3 =
. For k > r this is a stable node; for k < r it is a saddle.
k 1
There is one remaining fixed point a nontrivial one where both x and y are nonzero.

3.4. POPULATION BIOLOGY : LOTKA-VOLTERRA MODELS

63

Figure 3.12: Two possible phase flows for the rabbits vs. sheep model of eqs. 3.61, 3.62.
Left panel: k > r > k 1 . Right panel: k < r < k 1 .
To find it, we set x = y = 0, and divide out by x and y respectively, to get
x + ky = r

(3.64)

kx + y = 1 .

(3.65)

This is a simple rank 2 inhomogeneous linear system. If the fixed point P4 is to lie in
the physical quadrant (x > 0, y > 0), then either (i) k > r and k > r 1 or (ii) k < r
and k < r 1 . The solution is



1  
1
rk
1 k
r
.
(3.66)
P4 =
=
k 1
1
1 kk 1 rk
The linearized dynamics then gives

yielding

M4 =
1 kk

T =
D=

kr
k (rk 1)

rk 1 + k r
1 kk

k(k r)
,
rk 1

(k r)(rk 1)
.
1 kk

(3.67)

(3.68)
(3.69)

The classification of this fixed point can vary with parameters. Consider the case r = 1. If
k = k = 2 then both P2 and P3 are stable nodes. At P4 , one finds T = 23 and D = 31 ,
corresponding to a saddle point. In this case it is the fate of one population to die out at the
expense of the other, and which one survives depends on initial conditions. If instead we
took k = k = 12 , then T = 34 and D = 13 , corresponding to a stable node (node D < 41 T 2
in this case). The situation is depicted in Fig. 3.12.

64

CHAPTER 3. TWO-DIMENSIONAL PHASE FLOWS

 
Figure 3.13: Two singularities with index +1. The direction field Vb = V /V  is shown in
both cases.

3.5

Poincar
e-Bendixson Theorem

Although N = 2 systems are much richer than N = 1 systems, they are still ultimately
rather impoverished in terms of their long-time behavior. If an orbit does not flow off
to infinity or asymptotically approach a stable fixed point (node or spiral or nongeneric
example), the only remaining possibility is limit cycle behavior. This is the content of the
Poincare-Bendixson theorem, which states:
IF is a compact (i.e. closed and bounded) subset of phase space,
AND = V () is continuously differentiable on ,
AND contains no fixed points (i.e. V () never vanishes in ),
AND a phase curve (t) is always confined to ,
THEN (t) is either closed or approaches a closed trajectory in the limit t .
Thus, under the conditions of the theorem, must contain a closed orbit.
0 everywhere on
One way to prove that (t) is confined to is to establish that V n
the boundary , which means that the phase flow is always directed inward (or tangent)
along the boundary. Lets analyze an example from the book by Strogatz. Consider the
system
r = r(1 r 2 ) + r cos
= 1 ,

(3.70)
(3.71)

65

3.6. INDEX THEORY

with 0 < < 1. Then define

a
and

1+

(3.72)

n
o

(r, )  a < r < b .

(3.73)

On the boundaries of , we have

r=a
r=b

r = a 1 + cos

r = b 1 cos .

(3.74)
(3.75)

We see that the radial component of the flow is inward along both r = a and r = b. Thus,
any trajectory which starts inside can never escape. The Poincare-Bendixson theorem
tells us that the trajectory will approach a stable limit cycle in the limit t .
It is only with N 3 systems that the interesting possibility of chaotic behavior emerges.

3.6

Index Theory

Consider a smooth two-dimensional vector field V (). The angle that the vector V makes
b1 and
b2 axes is a scalar field,
with respect to the


1 V2 ()
.
(3.76)
() = tan
V1 ()
So long as V has finite length, the angle is well-defined. In particular, we expect that we
can integrate over a closed curve C in phase space to get
I
d = 0 .
(3.77)
C

However, this can fail if V () vanishes (or diverges) at one or more points in the interior
of C. In general, if we define
I
1
WC (V ) =
d ,
(3.78)
2
C

then WC (V ) Z is an integer valued function of C, which is the change in around the

curve C. This must be an integer, because is well-defined only up to multiples of 2.
Note that differential changes of are in general well-defined.
Thus, if V () is finite, meaning neither infinite nor infinitesimal, i.e. V neither diverges
nor vanishes anywhere in int(C), then WC (V ) = 0. Assuming that V never diverges, any
singularities in must arise from points where V = 0, which in general occurs at isolated
points, since it entails two equations in the two variables (1 , 2 ).

66

CHAPTER 3. TWO-DIMENSIONAL PHASE FLOWS

Figure 3.14: Two singularities with index 1.

The index of a two-dimensional vector field V () at a point is the integer-valued winding
of V about that point:
I
1
ind (V ) = lim
d
(3.79)
a0 2
0
Ca (0 )

1
= lim
a0 2

V1 V2 V2 V1
,
V12 + V22

(3.80)

Ca (0 )

where Ca (0 ) is a circle of radius a surrounding the point 0 . The index of a closed curve
C is given by the sum of the indices at all the singularities enclosed by the curve:2
X
WC (V ) =
(3.81)
ind (V ) .
i int(C)

As an example, consider the vector fields plotted in fig. 3.13. We have:

V = (x , y)
V = (y , x)

=
=

=
=+

(3.82)
1
2

(3.83)

The index is the same, +1, in both cases, even though the first corresponds to an unstable
node and the second to a center. Any N = 2 fixed point with det M > 0 has index +1.
2

Technically, we should weight the index at each enclosed singularity by the signed number of times the
curve C encloses that singularity. For simplicity and clarity, we assume that the curve C is homeomorphic
to the circle S1 .

67

3.6. INDEX THEORY

Figure 3.15: Left panel: a singularity with index +2. Right panel: two singularities each
with index +1. Note that the long distance behavior of V is the same in both cases.
Fig. 3.14 shows two vector fields, each with index 1:
V = (x , y)

V = (y , x)

(3.84)

= +

1
2

(3.85)

In both cases, the fixed point is a saddle.

As an example of the content of eqn. 3.81, consider the vector fields in eqn. 3.15. The left
panel shows the vector field V = (x2 y 2 , 2xy), which has a single fixed point, at the origin
(0 , 0), of index +2. The right panel shows the vector field V = (1 + x2 y 2 , x + 2xy),
which has fixed points (x , y ) at (0 , 1) and (0 , 1). The linearized dynamics is given by
the matrix

x
x
2x
2y
x
y

.
M =
(3.86)
=
y
y
1
+
2y
2x
x
y
Thus,

M(0,1) =


0 2
2 0

M(0,1) =

0 2
2 0

(3.87)

At each of these fixed points, we have T = 0 and D = 4, corresponding to a center, with

index +1. If we consider a square-ish curve Caround the periphery of each figure, the vector
field is almost the same along such a curve for both the left and right panels, and the
winding number is WC (V ) = +2.
Finally, consider the vector field shown in fig. 3.16, with V = (x2 y 2 , 2xy). Clearly
= 2, and the index of the singularity at (0 , 0) is 2.
To recapitulate some properties of the index / winding number:

68

CHAPTER 3. TWO-DIMENSIONAL PHASE FLOWS

Figure 3.16: A vector field with index 2.

The index ind (V ) of an N = 2 vector field V at a point 0 is the winding number
0
of V about that point.
The winding number WC (V ) of a curve C is the sum of the indices of the singularities
enclosed by that curve.
Smooth deformations of C do not change its winding number. One must instead
stretch C over a fixed point singularity in order to change WC (V ).
Uniformly rotating each vector in the vector field by an angle has the effect of
sending + ; this leaves all indices and winding numbers invariant.
Nodes and spirals, whether stable or unstable, have index +1 (ss do the special cases
of centers, stars, and degenerate nodes). Saddle points have index 1.
Clearly any closed orbit must lie on a curve C of index +1.

3.6.1

Gauss-Bonnet Theorem

There is a deep result in mathematics, the Gauss-Bonnet theorem, which connects the local
geometry of a two-dimensional manifold to its global topological structure. The content of
the theorem is as follows:
Z
X
(3.88)
dA K = 2 (M) = 2
ind (V ) ,
M

3.6. INDEX THEORY

69

Figure 3.17: Two smooth vector fields on the sphere S2 , which has genus g = 0. Left panel:
two index +1 singularities. Right panel: one index +2 singularity.

where M is a 2-manifold (a topological space locally homeomorphic to R2 ), is the local

Gaussian curvature of M, which is given by K = (R1 R2 )1 , where R1,2 are the principal
radii of curvature at a given point, and dA is the differential area element. The quantity
(M) is called the Euler characteristic of M and is given by (M) = 2 2g, where g is
the genus of M, which is the number of holes (or handles) of M. Furthermore, V () is
any smooth vector field on M, and i are the singularity points of that vector field, which
are fixed points of the dynamics = V ().
To apprehend the content of the Gauss-Bonnet theorem, it is helpful to consider an example.
Let M = S2 be the unit 2-sphere, as depicted in fig. 3.17. At any point on the unit 2sphere, the radii of curvature are degenerate and both equal to R = 1, hence
K = 1. If we
integrate the Gaussian curvature over the sphere, we thus get 4 = 2 S2 , which says
(S2 ) = 2 2g = 2, which agrees with g = 0 for the sphere. Furthermore, the Gauss-Bonnet
theorem says that any smooth vector field on S2 must have a singularity or singularities,
with the total index summed over the singularities equal to +2. The vector field sketched
in the left panel of fig. 3.17 has two index +1 singularities, which could be taken at the
north and south poles, but which could be anywhere. Another possibility, depicted in the
right panel of fig. 3.17, is that there is a one singularity with index +2.
In fig. 3.18 we show examples of manifolds with genii g = 1 and g = 2. The case g = 1 is
the familiar 2-torus, which is topologically equivalent to a product of circles: T2 S1 S1 ,
and is thus coordinatized by two angles 1 and 2 . A smooth vector field pointing in the
direction of
with g = 1

increasing 1 never vanishes, and thus has no singularities, consistent
2
2

and T = 0. Topologically, one can define a torus as the quotient space R /Z2 , or as a
square with opposite sides identified. This is what mathematicians call a flat torus one
with curvature K = 0 everywhere. Of course, such a torus cannot be embedded in three-

70

CHAPTER 3. TWO-DIMENSIONAL PHASE FLOWS

Figure 3.18: Smooth vector fields on the torus T2 (g = 1), and on a 2-manifold M of genus
g = 2.
dimensional Euclidean space; a two-dimensional figure embedded in a three-dimensional
Euclidean space inherits a metric due to the embedding, and for a physical torus, like the
surface of a bagel, the Gaussian curvature is only zero on average.
The g = 2 surface M shown in the right panel of fig. 3.18 has Euler characteristic (M) =
2, which means that any smooth vector field on M must have singularities with indices
totalling 2. One possibility, depicted in the figure, is to have two saddle points with index
1; one of these singularities is shown in the figure (the other would be on the opposite
side).

3.6.2

Singularities and topology

For any N = 1 system x = f (x), we can identify a charge Q with any generic fixed point
x by setting
h
i
Q = sgn f (x ) ,
(3.89)


where f (x ) = 0. The total charge contained in a region x1 , x2 is then
h
i
h
i
Q12 = 12 sgn f (x2 ) 12 sgn f (x1 ) .
(3.90)
It is easy to see that Q12 is the sum of the charges of all the fixed points lying within the
interval x1 , x2 .

In higher dimensions, we have the following general construction. Consider an N -dimensional

(x) be the unit vector field defined by
dynamical system x = V (x), and let n

n(x)
=

V (x)
.
|V (x)|

(3.91)

71

3.6. INDEX THEORY

Figure 3.19: Composition of two circles. The same general construction applies to the
merging of n-spheres Sn , called the wedge sum.
space, which is of dimension (N 1). If we integrate over
Consider now a unit sphere in n
this surface, we obtain
I
2 (N 1)/2

,
(3.92)
N = da na =
N 21

which is the surface area of the unit sphere SN 1 . Thus, 2 = 2, 3 = 4, 4 = 2 2 , etc.

Now consider a change of variables over the surface of the sphere, to the set (1 , . . . , N 1 ).
We then have
I
I
naN
na2
a

(3.93)
N =
da n = dN 1 a1 a na1
N
1
N 1
SN1

The topological charge is then

1
Q=
N

dN 1 a1 a na1
N

naN
na2

1
N 1

(3.94)

The quantity Q is an integer topological invariant which characterizes the map from the
= 1. In mathematical parlance, Q is known as
surface (1 , . . . , N 1 ) to the unit sphere |n|
the Pontrjagin index of this map.
This analytical development recapitulates some basic topology. Let M be a topological
space and consider a map from the circle S1 to M. We can compose two such maps by
merging the two circles, as shown in fig. 3.19. Two maps are said to be homotopic if they
can be smoothly deformed into each other. Any two homotopic maps are said to belong to
the same equivalence class or homotopy class. For general M, the homotopy classes may
be multiplied using the composition law, resulting in a group structure. The group is called
the fundamental group of the manifold M, and is abbreviated 1 (M). If M = S2 , then
any such map can be smoothly contracted to a point on the 2-sphere, which is to say a
trivial map. We then have 1 (M) = 0. If M = S1 , the maps can wind nontrivially, and

72

CHAPTER 3. TWO-DIMENSIONAL PHASE FLOWS

the homotopy classes are labeled by a single integer winding number: 1 (S1 ) = Z. The
winding number of the composition of two such maps is the sum of their individual winding
numbers. If M = T2 , the maps can wind nontrivially around either of the two cycles of the
2-torus. We then have 1 (T2 ) = Z2 , and in general 1 (Tn ) = Zn . This makes good sense,
since an n-torus is topologically equivalent to a product of n circles. In some cases, 1 (M)
can be nonabelian, as is the case when M is the genus g = 2 structure shown in the right
hand panel of fig. 3.18.
In general we define the nth homotopy group n (M) as the group under composition of
maps from Sn to M. For n 2, n (M) is abelian. If dim(M) < n, then n (M) = 0.
In general, n (Sn ) = Z. These nth homotopy classes of the n-sphere are labeled by their
Pontrjagin index Q.
Finally, we ask what is Q in terms of the eigenvalues and eigenvectors of the linearized map

Vi 
.
Mij =
xj x

(3.95)

For simple cases where all the i are nonzero, we have

N
Y

(3.96)

x = 12 x + xy 2x3

(3.97)

y = 25 y + xy y 2 .

(3.98)

Q = sgn

i=1

3.7

Appendix : Example Problem

Consider the two-dimensional phase flow,

(a) Find and classify all fixed points.

Solution : We have
x = x
y = y
The matrix of first derivatives is

M =

x
x

x
y

y
x

y
y

There are six fixed points.

1
2
5
2

+ y 2x2


+xy .

(3.99)
(3.100)

+ y 6x2
y

x
5
2

+ x 2y

(3.101)

73

3.7. APPENDIX : EXAMPLE PROBLEM

Figure 3.20: Sketch of phase flow for x = 21 x + xy 2x3 , y = 25 y + xy y 2 . Fixed point

classifications are in the text.
(x, y) = (0, 0) : The derivative matrix is
M=

1
2

0
5
2

(3.102)

The determinant is D = 45 and the trace is T = 3. Since D < 14 T 2 and T > 0, this is an
unstable node. (Duh! One can read off both eigenvalues are real and positive.) Eigenvalues:
1 = 21 , 2 = 25 .
(x, y) = (0, 25 ) : The derivative matrix is
M=

3
5
2

0
25

(3.103)

1
for which D = 15
2 and T = 2 . The determinant is negative, so this is a saddle. Eigenvalues:
1 = 52 , 2 = 3.

(x, y) = ( 21 , 0) : The derivative matrix is



1 12
M=
,
0
2

(3.104)

for which D = 2 and T = +1. The determinant is negative, so this is a saddle. Eigenvalues:
1 = 1, 2 = 2.
(x, y) = ( 12 , 0) : The derivative matrix is
M=

1 21
0 3

(3.105)

74

CHAPTER 3. TWO-DIMENSIONAL PHASE FLOWS

for which D = 3 and T = +2. The determinant is negative, so this is a saddle. Eigenvalues:
1 = 1, 2 = 3.
(x, y) = ( 32 , 4) : This is one root obtained by setting y = x + 52 and the solving
3 + x 2x2 = 0, giving x = 1 and x = + 32 . The derivative matrix is
M=


9 32
,
4 4

for which D = 30 and T = 13. Since D <

node. Eigenvalues: 1 = 10, 2 = 3.

1
4

1
2

+ y 2x2 =
(3.106)

T 2 and T < 0, this corresponds to a stable

(x, y) = (1, 32 ) : This is the second root obtained by setting y = x + 52 and the solving
1
3
2
2
2 + y 2x = 3 + x 2x = 0, giving x = 1 and x = + 2 . The derivative matrix is
M=

4 1
3
23
2

11
for which D = 15
2 and T = 2 . Since D <
node. Eigenvalues: 1 = 3, 2 = 52 .

1
4

(3.107)

T 2 and T < 0, this corresponds to a stable

(b) Sketch the phase flow.

Solution : The flow is sketched in fig. 3.20. Thanks to Evan Bierman for providing the
Mathematica code.

Chapter 4

Nonlinear Oscillators
4.1

Weakly Perturbed Linear Oscillators

Consider a nonlinear oscillator described by the equation of motion

x
+ 02 x = h(x) .

(4.1)

Here, is a dimensionless parameter, assumed to be small, and h(x) is a nonlinear function

of x. In general, we might consider equations of the form
x
+ 02 x = h(x, x)
,

(4.2)

x
+ (x2 1)x + 02 x = 0 .

(4.3)

such as the van der Pol oscillator,

First, we will focus on nondissipative systems, i.e. where we may write m

x = x V , with
V (x) some potential.
As an example, consider the simple pendulum, which obeys
+ 02 sin = 0 ,

(4.4)

where 02 = g/, with the length of the pendulum. We may rewrite his equation as
+ 02 = 02 ( sin )
=

1
6

02 3

1
120

02 5 + . . .

(4.5)

The RHS above is a nonlinear function of . We can define this to be h(), and take = 1.

4.1.1

Nave Perturbation theory and its failure

Lets assume though that is small, and write a formal power series expansion of the solution
x(t) to equation 4.1 as
x = x0 + x1 + 2 x2 + . . . .
(4.6)
75

76

CHAPTER 4. NONLINEAR OSCILLATORS

We now plug this into 4.1. We need to use Taylors theorem,

h(x0 + ) = h(x0 ) + h (x0 ) + 21 h (x0 ) 2 + . . .

(4.7)

= x1 + 2 x2 + . . . .

(4.8)

with

Working out the resulting expansion in powers of is tedious. One finds

n
h(x) = h(x0 ) + h (x0 ) x1 + 2 h (x0 ) x2 +

1
2

o
h (x0 ) x21 + . . . .

(4.9)

Equating terms of the same order in , we obtain a hierarchical set of equations,

x
0 + 02 x0 = 0

(4.10)

02 x1
02 x2
02 x3

= h(x0 )

(4.11)

= h (x0 ) x1

(4.12)

= h (x0 ) x2 + 12 h (x0 ) x21

(4.13)

x
1 +
x
2 +
x
3 +

et cetera, where prime denotes differentiation with respect to argument. The first of these
is easily solved: x0 (t) = A cos(0 t + ), where A and are constants. This solution then
is plugged in at the next order, to obtain an inhomogeneous equation for x1 (t). Solve for
x1 (t) and insert into the following equation for x2 (t), etc. It looks straightforward enough.
The problem is that resonant forcing terms generally appear in the RHS of each equation
of the hierarchy past the first. Define 0 t + . Then x0 () is an even periodic function


of with period 2, hence so is h(x0 ). We may then expand h x0 () in a Fourier series:

X
hn (A) cos(n) .
h A cos =

(4.14)

n=0

The n = 1 term leads to resonant forcing. Thus, the solution for x1 (t) is

1 X hn (A)
h1 (A)
x1 (t) = 2
cos(n0 t + n) +
t sin(0 t + ) ,
2
1

n
0 n=0
20

(4.15)

(n6=1)

which increases linearly with time. As an example, consider a cubic nonlinearity with
h(x) = r x3 , where r is a constant. Then using
cos3 =
we have h1 =

3
4

rA3 and h3 =

1
4

rA3 .

3
4

cos + 41 cos(3) ,

(4.16)

77

4.1. WEAKLY PERTURBED LINEAR OSCILLATORS

4.1.2

Poincar
e-Lindstedt method

The problem here is that the nonlinear oscillator has a different frequency than its linear
counterpart. Indeed, if we assume the frequency is a function of , with
() = 0 + 1 + 2 2 + . . . ,

(4.17)

then subtracting the unperturbed solution from the perturbed one and expanding in yields
cos(t) cos(0 t) = sin(0 t) ( 0 ) t 12 cos(0 t) ( 0 )2 t2 + . . .
n
o
= sin(0 t) 1 t 2 sin(0 t) 2 t + 12 cos(0 t) 12 t2 + O(3 ) .

(4.18)

What perturbation theory can do for us is to provide a good solution up to a given time,
provided that is sufficiently small. It will not give us a solution that is close to the true
answer for all time. We see above that in order to do that, and to recover the shifted
frequency (), we would have to resum perturbation theory to all orders, which is a
daunting task.
The Poincare-Lindstedt method obviates this difficulty by assuming = () from the
outset. Define a dimensionless time s t and write 4.1 as
2

d2x
+ 02 x = h(x) ,
ds2

(4.19)

where
x = x0 + x1 + 2 x2 + . . .

(4.20)

2 = a0 + a1 + 2 a2 + . . . .

(4.21)

We now plug the above expansions into 4.19:



2

d2x0
d2x1
2
2 d x2
a0 + a1 + a2 + . . .
+ 2 +
+ ...
ds2
ds
ds2


+ 02 x0 + x1 + 2 x2 + . . .
o
n
= h(x0 ) + 2 h (x0 ) x1 + 3 h (x0 ) x2 + 12 h (x0 ) x21 + . . .

(4.22)

Now lets write down equalities at each order in :

a0

d2x0
+ 02 x0 = 0
ds2

(4.23)

a0

d2x1
d2x0
2
+

x
=
h(x
)

a
0 1
0
1
ds2
ds2

(4.24)

a0

d2x0
d2x1
d2x2
2

x
=
h
(x
)
x

a
,
0 2
0
1
2
1
ds2
ds2
ds2

(4.25)

78

CHAPTER 4. NONLINEAR OSCILLATORS

et cetera.
The first equation of the hierarchy is immediately solved by
a0 = 02

x0 (s) = A cos(s + ) .

(4.26)

At O(), then, we have

d2x1
2
+
x
=

h
A
cos(s
+
)
+ 02 a1 A cos(s + ) .
1
0
ds2

(4.27)

The LHS of the above equation has a natural frequency of unity (in terms of the dimensionless time s). We expect h(x0 ) to contain resonant forcing terms, per 4.14. However, we
now have the freedom to adjust the undetermined coefficient a1 to cancel any such resonant
term. Clearly we must choose
h1 (A)
a1 =
.
(4.28)
A
The solution for x1 (s) is then

1 X hn (A)
cos(ns + n) ,
x1 (s) = 2
0 n=0 1 n2

(4.29)

(n6=1)

which is periodic and hence does not increase in magnitude without bound, as does 4.15.
The perturbed frequency is then obtained from
2 = 02

h1 (A)
+ O(2 )
A

() = 0

h1 (A)
2A0

+ O(2 ) .

(4.30)

Note that depends on the amplitude of the oscillations.

As an example, consider an oscillator with a quartic nonlinearity in the potential, i.e.
h(x) = r x3 . Then


(4.31)
h A cos = 43 rA3 cos + 14 rA3 cos(3) .

We then obtain, setting = 1 at the end of the calculation,

= 0

3 rA2
8 0

+ ...

(4.32)

where the remainder is higher order in the amplitude A. In the case of the pendulum,

and with r =

1
6

+ 02 = 16 02 3 + O 5 ,

(4.33)

02 and 0 (t) = 0 sin(t), we find

T (0 ) =

2 n
2
=
1+

1
16

02 + . . .

(4.34)

79

4.2. MULTIPLE TIME SCALE METHOD

One can check that this is correct to lowest nontrivial order in the amplitude, using the
exact result for the period,


4
(4.35)
T (0 ) =
K sin2 12 0 ,
0
where K(x) is the complete elliptic integral.
The procedure can be continued to the next order, where the free parameter a2 is used to
eliminate resonant forcing terms on the RHS.
A good exercise to test ones command of the method is to work out the lowest order
nontrivial corrections to the frequency of an oscillator with a quadratic nonlinearity, such
as h(x) = rx2 . One finds that there are no resonant forcing terms at first order in , hence
one must proceed to second order to find the first nontrivial corrections to the frequency.

4.2

Multiple Time Scale Method

Another method of eliminating secular terms (i.e. driving terms which oscillate at the
resonant frequency of the unperturbed oscillator), and one which has applicability beyond
periodic motion alone, is that of multiple time scale analysis. Consider the equation
x
+ x = h(x, x)
,

(4.36)

where is presumed small, and h(x, x)

is a nonlinear function of position and/or velocity.
We define a hierarchy of time scales: Tn n t. There is a normal time scale T0 = t, slow
time scale T1 = t, a superslow time scale T2 = 2 t, etc. Thus,

d
+
+ 2
+ ...
=
dt
T0
T1
T2

n
=
.
Tn

(4.37)

n=0

Next, we expand
x(t) =

n xn (T0 , T1 , . . .) .

(4.38)

n=0

Thus, we have
X

Tn
n=0
n

2  X

k=0

xk

X
k=0

xk = h

X
k=0

xk ,
Tn
n=0
k

X

k=0

xk

!

80

CHAPTER 4. NONLINEAR OSCILLATORS

We now evaluate this order by order in :

 2


0
O( ) :
+ 1 x0 = 0
T02

 2


x0

2x0
1
+ h x0 ,
O( ) :
+ 1 x1 = 2
T0 T1
T0
T02
 2

2
2
x0
2x0

x1

O(2 ) :
+
1
x
=
2
2
T0 T1
T0 T2
T02
T12



x0
h 
h  x1
+
+
,
x +

x  1 x  T0 T1
{x ,x }
0 0

(4.39)
(4.40)

(4.41)

{x ,x }
0 0

et cetera. The expansion gets more and more tedious with increasing order in .

Lets carry this procedure out to first order in . To order 0 ,

x0 = A cos T0 + ,
(4.42)


where A and are arbitrary (at this point) functions of T1 , T2 , . . . . Now we solve the

next equation in the hierarchy, for x1 . Let T0 + . Then T =

and we have
0

 2

sin + 2 A
cos + h A cos , A sin .
(4.43)
+ 1 x1 = 2
2

T1
T1
Since the arguments of h are periodic under + 2, we may expand h in a Fourier
series:

X

X
h() h A cos , A sin =
k (A) sin(k) +
k (A) cos(k) .
(4.44)
k=1

k=0

The inverse of this relation is

Z2
d
k (A) =
h() sin(k)

(k > 0)

(4.45)

Z2
d
0 (A) =
h()
2

(4.46)

Z2
d
h() cos(k)
k (A) =

(k > 0) .

(4.47)

We now demand that the secular terms on the RHS those terms proportional to cos and
sin must vanish. This means
2
2A

A
+ 1 (A) = 0
T1

(4.48)

+ 1 (A) = 0 .
T1

(4.49)

81

4.2. MULTIPLE TIME SCALE METHOD

These two first order equations require two initial conditions, which is sensible since our
initial equation x
+ x = h(x, x)
is second order in time.
With the secular terms eliminated, we may solve for x1 :


X
k (A)
k (A)
x1 =
sin(k) +
cos(k) + C0 cos + D0 sin .
1 k2
1 k2

(4.50)

k6=1

Note: (i) the k = 1 terms are excluded from the sum, and (ii) an arbitrary solution to the
homogeneous equation, i.e. eqn. 4.43 with the right hand side set to zero, is included. The
constants C0 and D0 are arbitrary functions of T1 , T2 , etc. .
The equations for A and are both first order in T1 . They will therefore involve two
constants of integration call them A0 and 0 . At second order, these constants are taken
as dependent upon the superslow time scale T2 . The method itself may break down at this
order. (See if you can find out why.)
Lets apply this to the nonlinear oscillator x
+sin x = 0, also known as the simple pendulum.
Well expand the sine function to include only the lowest order nonlinear term, and consider
x
+x=

1
6

x3 .

(4.51)

Well assume is small and take = 1 at the end of the calculation. This will work provided
the amplitude of the oscillation is itself small. To zeroth order, we have x0 = A cos(t + ),
as always. At first order, we must solve

 2

(4.52)
sin + 2 A
cos + 16 A2 cos3
+ 1 x1 = 2
2

T1
T1
A

1
=2
A3 cos(3) + 18 A3 cos .
sin + 2 A
cos + 24
T1
T1
We eliminate the secular terms by demanding
A
=0
T1

1
A2 ,
= 16
T1

1
hence A = A0 and = 16
A20 T1 + 0 , and
2
1
16 A0 t
3
1
192 A0 cos 3t

x(t) = A0 cos t

+ 0

3
16

A20 t + 30 + . . . ,

(4.53)

(4.54)

which reproduces the result obtained from the Poincare-Lindstedt method.

4.2.1

Duffing oscillator

Consider the equation

x
+ 2x + x + x3 = 0 .

(4.55)

82

CHAPTER 4. NONLINEAR OSCILLATORS

This describes a damped nonlinear oscillator. Here we assume both the damping coefficient

as well as the nonlinearity both depend linearly on the small parameter . We may
write this equation in our standard form x
+ x = h(x, x),
with h(x, x)
= 2x x3 .
For > 0, which we henceforth assume, it is easy to see that the only fixed point is
(x, x)
= (0, 0). The linearized flow in the vicinity of the fixed point is given by
  
 
d x
0
1
x
=
+ O(x3 ) .
(4.56)
1 2
x
dt x
The determinant is D = 1 and the trace is T = 2. Thus, provided < 1, the fixed
point is a stable spiral; for > 1 the fixed point becomes a stable node.
We employ the multiple time scale method to order . We have x0 = A cos(T0 + ) to zeroth
order, as usual. The nonlinearity is expanded in a Fourier series in = T0 + :


x
h x0 , T0 = 2A sin A3 cos3
0

= 2A sin 34 A3 cos 14 A3 cos 3 .

(4.57)

Thus, 1 (A) = 2A and 1 (A) = 34 A3 . We now solve the first order equations,
A
= 21 1 (A) = A
T1

A(T ) = A0 eT1

(4.58)

as well as
(A)

= 1
= 38 A20 e2T1
T1
2A

(T1 ) = 0 +

3A20
1 e2T1 .
16

(4.59)

After elimination of the secular terms, we may read off

x1 (T0 , T1 ) =
Finally, we have

1 3
32 A (T1 ) cos

3T0 + 3(T1 ) .





3A20
x(t) = A0 et cos t +
1 e2t + 0
16




9A20
1
A30 e3t cos 3t +
1 e2t + 30 .
+ 32
16

4.2.2

(4.60)

(4.61)

Van der Pol oscillator

Lets apply this method to another problem, that of the van der Pol oscillator,
x
+ (x2 1) x + x = 0 ,

(4.62)

with > 0. The nonlinear term acts as a frictional drag for x > 1, and as a negative
friction (i.e. increasing the amplitude) for x < 1. Note that the linearized equation at the
fixed point (x = 0, x = 0) corresponds to an unstable spiral for < 2.

83

4.3. FORCED NONLINEAR OSCILLATIONS

For the van der Pol oscillator, we have h(x, x)

= (1 x2 ) x,
and plugging in the zeroth order
solution x0 = A cos(t + ) gives






x0
h x0 ,
= 1 A2 cos2 A sin
T0


= A + 14 A3 sin + 41 A3 sin(3) ,
(4.63)

with t + . Thus, 1 = A + 14 A3 and 1 = 0, which gives = 0 and

2

A
= A 14 A3 .
T1

(4.64)

The equation for A is easily integrated:





8 dA
2
A
1
1
dT1 =
=

dA = d ln
A (A2 4)
A A2 A+2
A2 4
=
Thus,

A(T1 ) = q
x0 (t) = q

2
1 1

4
A20

exp(T1 )

2 cos(t + 0 )
.


1 1 A42 exp(t)

(4.65)

(4.66)

This behavior describes the approach to the limit cycle 2 cos(t + 0 ). With the elimination
of the secular terms, we have


1
1 3
4 sin 3t + 30
(4.67)
x1 (t) = 32 A sin(3) = 
3/2 .


4
1 1 A2 exp(t)
0

4.3

Forced Nonlinear Oscillations

The forced, damped linear oscillator,

x
+ 2x + x = f0 cos t
has the solution

x(t) = xh (t) + C() cos t + () ,

where
p

xh (t) = A+ e+ t + A e t ,

(4.68)
(4.69)
(4.70)

where = 2 1 are the roots of 2 + 2 + 1 = 0. The susceptibility C and

phase shift are given by


2
1
1
.
(4.71)
,
() = tan
C() = p
1 2
( 2 1)2 + 42 2

84

CHAPTER 4. NONLINEAR OSCILLATORS

The homogeneous solution, xh (t), is a transient and decays exponentially with time, since
Re( ) < 0. The asymptotic behavior is a phase-shifted oscillation at the driving frequency
.
Now lets add a nonlinearity. We study the equation
x
+ x = h(x, x)
+ f0 cos(t + t) .

(4.72)

Note that amplitude of the driving term, f0 cos(t), is assumed to be small, i.e. proportional to , and the driving frequency = 1 + is assumed to be close to resonance.
(The resonance frequency of the unperturbed oscillator is res = 1.) Were the driving frequency far from resonance, it could be dealt with in the same manner as the non-secular
terms encountered thus far. The situation when is close to resonance deserves our special
attention.
At order 0 , we still have x0 = A cos(T0 + ). We write
t = t + t = T0 + T1

(4.73)

where = T0 + (T1 ) as before, and (T1 ) (T1 ) T1 . At order 1 , we must then solve





2
+ 1 x1 = 2A sin + 2A cos + h A cos , A sin + f0 cos( )
2

 

X
=
k sin(k) + k cos(k) + 2A + 1 + f0 sin sin
k6=1



+ 2A + 2A + 1 + f0 cos cos ,

(4.74)

where the prime denotes differentiation with respect to T1 . We thus have the N = 2
dynamical system
dA
= 21 1 (A) 12 f0 sin
dT1

(4.75)

d
(A)
f
= 1
0 cos .
dT1
2A
2A

(4.76)

If we assume that {A, } approaches a fixed point of these dynamics, then at the fixed point
these equations provide a relation between the amplitude A, the detuning parameter ,
and the drive f0 :
h
i2 h
i2
1 (A) + 2A + 1 (A) = f02 .
(4.77)

In general this is a nonlinear equation for A(f0 , ). The linearized (A, ) dynamics in the
vicinity of a fixed point is governed by the matrix

M =

A/A
A/

/A
/

12 1 (A)

(A)
12A

A + 21 1 (A)
(A)
12A

(4.78)

85

4.3. FORCED NONLINEAR OSCILLATIONS

Figure 4.1: Phase diagram for the forced Duffing oscillator.

If the (A, ) dynamics exhibits a stable fixed point (A , ), then one has
x0 (t) = A cos(T0 + T1 + ) = A cos(t + ) .

(4.79)

The oscillators frequency is then the forcing frequency = 1 + , in which case the
oscillator is said to be entrained, or synchronized, with the forcing.

4.3.1

Forced Duffing oscillator

Thus far our approach has been completely general. We now restrict our attention to the
Duffing equation, for which
1 (A) = 2A

1 (A) = 34 A3 ,

(4.80)

which yields the cubic equation

A6

4
16
3 A

2
64
9 (

+ 2 )A2

16 2
9 f0

=0.

(4.81)

Analyzing the cubic is a good exercise. Setting y = A2 , we define

G(y) y 3
and we seek a solution to G(y) =

16 2
9 f0 .

y =

16
3

y2 +

2
64
9 (

+ 2) y ,

(4.82)

Setting G (y) = 0, we find roots at

16
9

8
9

2 32 .

(4.83)

If 2 < 32 the roots are imaginary, which tells us that G(y) is monotonically increasing
2
for real y. There is then a unique solution to G(y) = 16
9 f0 .

86

CHAPTER 4. NONLINEAR OSCILLATORS

Figure 4.2: Amplitude A versus detuning for the forced Duffing oscillator for three values
3/2 . For f > f , there is hysteresis as a
of the drive f0 . The critical drive is f0,c = 316
0
0,c
5/4
function of the detuning.
If 2 > 32 , then the
a local minimum at
cubic G(y) has a local maximum at y = y and
2
y = y+ . For < 3 , we have y < y+ < 0, and since y = A must be positive, this
2
means that once more there is a unique solution to G(y) = 16
9 f0 .

For > 3 , we havep

y+ > y > 0. There are then three solutions for y() for f0
 +

3
f0 , f0 , where f0 = 4 G(y ). If we define /, then
f0

8
9

3/2

3 + 9

2 3 .

(4.84)

The phase diagram is shown in Fig. 4.1. The minimum value for f0 is f0,c =

which occurs at = 3.

16
35/4

3/2 ,

Thus far we have assumed that the (A, ) dynamics evolves to a fixed point. We should
check to make sure that this fixed point is in fact stable. To do so, we evaluate the linearized
dynamics at the fixed point. Writing A = A + A and = + , we have
 
 
d
A
A
=M
,
(4.85)

dT1
with

M =

A
A

12 f0 cos

f0
3
4 A + 2A2

f0
2A

cos

A 83 A3

9
8A A

sin

(4.86)

4.3. FORCED NONLINEAR OSCILLATIONS

87

Figure 4.3: Amplitude versus detuning for the forced Duffing oscillator for ten equally
spaced values of f0 between 3/2 and 10 3/2 . The critical value is f0,c = 4.0525 3/2 . The
red and blue curves are boundaries for the fixed point classification.
One then has T = 2 and




D = 2 + 83 A2 89 A2 .

(4.87)

Setting D = 41 T 2 = 2 sets the boundary between stable spiral and stable node. Setting
D = 0 sets the boundary between stable node and saddle. The fixed point structure is as
shown in Fig. 4.3. Though the amplitude exhibits hysteresis, the oscillator frequency is
always synchronized with the forcing as one varies the detuning.

4.3.2

Forced van der Pol oscillator

Consider now a weakly dissipative, weakly forced van der Pol oscillator, governed by the
equation
x
+ (x2 1) x + x = f0 cos(t + t) ,

(4.88)

where the forcing frequency is = 1 + , which is close to the natural frequency 0 = 1.

We apply the multiple time scale method, with h(x, x)
= (1 x2 ) x.
As usual, the lowest


order solution is x0 = A(T1 ) cos T0 + (T1 ) , where T0 = t and T1 = t. Again, we define
T0 + (T1 ) and (T1 ) (T1 ) T1 . From
h(A cos , A sin ) =

1 3
4A

A sin + 41 A3 sin(3) ,

(4.89)

88

CHAPTER 4. NONLINEAR OSCILLATORS

Figure 4.4: Amplitude versus detuning for the forced van der Pol oscillator. Fixed point
classifications are abbreviated SN (stable node), SS (stable spiral), UN (unstable node), US
(unstable spiral), and SP (saddle point).
we arrive at




x0
2 x0
2
+ h x0 ,
+ 1 x1 = 2
2
T0 T1
T0

1 3
= 4 A A + 2A + f0 sin sin


+ 2A + 2A + f0 cos cos + 41 A3 sin(3) .

(4.90)

We eliminate the secular terms, proportional to sin and cos , by demanding

dA
= 12 A 81 A3 21 f0 sin
dT1

(4.91)

f0
d
=
cos .
dT1
2A

(4.92)

Stationary solutions have A = = 0, hence cos = 2A/f0 , and hence

2
f02 = 4 2A2 + 1 + 14 A2 A2
=

1 6
16 A

12 A4 + (1 + 4 2 )A2 .

(4.93)

For this solution, we have

x0 = A cos(T0 + T1 + ) ,

(4.94)

4.3. FORCED NONLINEAR OSCILLATIONS

89

Figure 4.5: Phase diagram for the weakly forced van der Pol oscillator in the ( 2 , f02 ) plane.
Inset shows detail. Abbreviations for fixed point classifications are as in Fig. 4.4.
and the oscillators frequency is the forcing frequency = 1 + .
To proceed further, let y = A2 , and consider the cubic equation
F (y) =

1 3
16 y

12 y 2 + (1 + 4 2 ) y f02 = 0 .

(4.95)

Setting F (y) = 0, we find the roots of F (y) lie at y = 34 (2 u), where u = (1 12 2 )1/2 .
1
Thus, the roots are complex for 2 > 12
, in which case F (y) is monotonically increasing,
and there is a unique solution to F (y) = 0. Since F (0) = f02 < 0, that solution satisfies
1
y > 0. For 2 < 12
, there are two local extrema at y = y . When Fmin = F (y+ ) <
0 < F (y ) = Fmax , the cubic F (y) has three real, positive roots. This is equivalent to the
condition
32
8 3
8 3
u + 89 u2 < 27
f02 < 27
u + 98 u2 .
(4.96)
27
We can say even more by exploring the behavior of eqs. (4.91) and (4.92) in the vicinity of

90

CHAPTER 4. NONLINEAR OSCILLATORS

Figure 4.6: Forced van der Pol system with = 0.1, = 0.4 for three values of f0 . The
limit entrained solution becomes unstable at f0 = 1.334.
the fixed points. Writing A = A + A and = + , we have


1

3 2
A
1

A
A
A
4
d 2
.
=


dT1
1
1 2

/A
2 1 4A

(4.97)

The eigenvalues of the linearized dynamics at the fixed point are given by = 12 T

T 2 4D , where T and D are the trace and determinant of the linearized equation. Recall

91

4.4. SYNCHRONIZATION

now the classification scheme for fixed points of two-dimensional phase flows, discussed in
chapter 3. To recapitulate, when D < 0, we have < 0 < + and the fixed point is a
saddle. For 0 < 4D < T 2 , both eigenvalues have the same sign, so the fixed point is a node.
For 4D > T 2 , the eigenvalues form a complex conjugate pair, and the fixed point is a spiral.
A node/spiral fixed point is stable if T < 0 and unstable if T > 0. For our forced van der
Pol oscillator, we have
T = 1 12 A 2

D=

1
4

3 4
1 A 2 + 16
A
+ 2 .

(4.98)
(4.99)

From these results we can obtain q

the plot of Fig. 4.4, where amplitude is shown versus


detuning. Note that for 1 < f0 < 32
27 there is a region , + of hysteretic behavior in
which varying the detuning parameter is not a reversible process. The phase diagram in
the ( 2 , f02 ) plane is shown in Fig. 4.5.
Finally, we can make the following statement about the global dynamics (i.e. not simply in
the vicinity of a fixed point). For large A, we have
dA
= 18 A3 + . . .
dT1

d
= + . . . .
dT1

(4.100)

This flow is inward, hence if the flow is not to a stable fixed point, it must be attracted
to a limit cycle. The limit cycle necessarily involves several frequencies. This result the
generation of new frequencies by nonlinearities is called heterodyning.
We can see heterodyning in action in the van der Pol system. In Fig. 4.5, the blue line
which separates stable and unstable spiral solutions is given by f02 = 8 2 + 21 . For example,
if we take = 0.40 then the boundary lies at f0 = 1.334. For f0 < 1.334, we expect
heterodyning, as the entrained solution is unstable. For f > 1.334 the solution is entrained
and oscillates at a fixed frequency. This behavior is exhibited in Fig. 4.6.

4.4

Synchronization

Thus far we have assumed both the nonlinearity as well as the perturbation are weak. In
many systems, we are confronted with a strong nonlinearity which we can perturb weakly.
How does an attractive limit cycle in a strongly nonlinear system respond to weak periodic
forcing? Here we shall follow the nice discussion in the book of Pikovsky et al.
Consider a forced dynamical system,
= V () + f (, t) .

(4.101)

When = 0, we assume that the system has at least one attractive limit cycle 0 (t) =
0 (t + T0 ). All points on the limit cycle are fixed under the T0 -advance map gT , where
0
g (t) = (t + ). The idea is now to parameterize the points along the limit cycle by a

92

CHAPTER 4. NONLINEAR OSCILLATORS

phase angle which runs from 0 to 2 such that (t) increases by 2 with each orbit of the
limit cycle, with increasing uniformly with time, so that = 0 = 2/T0 . Now consider
the action of the T0 -advance map gT on points in the vicinity of the limit cycle. Since
0
each point 0 () on the limit cycle is a fixed point, and since the limit cycle is presumed
to be attractive, we can define the -isochrone as the set of points {} in phase space
which flow to the fixed point 0 () under repeated application of gT . The isochrones are
0
(N 1)-dimensional hypersurfaces.
To illustrate this, we analyze the example in Pikovsky et al. of the complex amplitude
equation (CAE),
dA
= (1 + i) A (1 + i) |A|2 A ,
(4.102)
dt
where A C is a complex number. It is convenient to work in polar coordinates, writing
A = R ei , in which case the real and complex parts of the CAE become
R = (1 R2 ) R
= R2 .

(4.103)
(4.104)

These equations can be integrated to yield the solution

R(t) = p

R0
R02

+ (1 R02 ) e2t



(t) = 0 + ( )t 12 ln R02 + (1 R02 ) e2t

(4.105)
(4.106)

= 0 + ( ) t + ln(R/R0 ) .

As t , we have R(t) 1 and (t)

0 . Thus the limit cycle is the circle R = 1, and
its frequency is 0 = .
We define the isochrones by the relation
(R, ) = ln R .

(4.107)

Then one has

R
= = = 0 .
(4.108)
R
Thus the -isochrone is given by the curve (R) = + ln R, which is a logarithmic spiral.
These isochrones are depicted in fig. 4.7.

At this point we have defined a phase function () as the phase of the fixed point along
the limit cycle to which flows under repeated application of the T0 -advance map gT . Now
0
let us examine the dynamics of for the weakly perturbed system of eqn. 4.101. We have
N
d X dj
=
dt
j dt
j=1

N
X

= 0 +
f (, t) .
j j
j=1

(4.109)

4.4. SYNCHRONIZATION

93

Figure 4.7: Isochrones of the complex amplitude equation A = (1 + i)A (1 + i)|A|2 A,

where A = X + iY .

We will assume that is close to the limit cycle, so that 0 () is small. As an example,
consider once more the complex amplitude equation (4.102), but now adding in a periodic
forcing term.
dA
= (1 + i) A (1 + i) |A|2 A + cos t .
dt

(4.110)

Writing A = X + iY , we have
X = X Y (X Y )(X 2 + Y 2 ) + cos t
Y = Y + X (X + Y )(X 2 + Y 2 ) .

(4.111)
(4.112)

In Cartesian coordinates, the isochrones for the = 0 system are

= tan1 (Y /X) 21 ln(X 2 + Y 2 ) ,

(4.113)

94

CHAPTER 4. NONLINEAR OSCILLATORS

Figure 4.8: Left panel: graphical analysis of the equation = + G(). Right panel:
Synchronization region (gray) as a function of detuning .
hence

d
= 0 +
cos t
dt
X

X + Y
cos t
=
X2 + Y 2
0 ( cos + sin ) cos t
p
= 0 1 + 2 cos( ) cos t .

(4.114)

where = ctn 1 . Note that in the third line above we have invoked R 1, i.e. we
assume that we are close to the limit cycle.
We now define the function

N
X

F (, t)
j 
j=1

fj (0 (), t) .

(4.115)

0 ()

The phase dynamics for are now written as

= 0 + F (, t) .
Now F (, t) is periodic in both its arguments, so we may write
X
F (, t) =
Fkl ei(k+lt) .

(4.116)

(4.117)

k,l

For the unperturbed problem, we have = 0 , hence resonant terms in the above sum are
those for which k0 + l 0. This occurs when pq 0 , where p and q are relatively
prime integers. In this case the resonance condition is satisfied for k = jp and l = jq

95

4.4. SYNCHRONIZATION

for all j Z. We now separate the resonant from the nonresonant terms in the (k, l) sum,
writing
X
= 0 +
Fjp,jq eij(pqt) + NRT ,
(4.118)
j

where NRT denotes nonresonant terms, i.e. those for which (k, l) 6= (jp, jq) for some
integer j. We now average over short time scales to eliminate the nonresonant terms, and
focus on the dynamics of this averaged phase hi.
We define the angle phi qt, which obeys
q
= p hi
= (p0 q) + p
= + G() ,

(4.119)
Fjp,jq eij

P
where q p0 is the detuning and G() = p j Fjp,jq eij is the sum over resonant
terms. Note that the nonresonant terms have been eliminated by the aforementioned averaging procedure. This last equation is a simple N = 1 dynamical system on the circle
a system we have already studied. The dynamics are depicted in fig. 4.8. If the detuning
falls within the range [ Gmin , Gmax ], then flows to a fixed point, and the nonlinear
q . If the detuning
oscillator is synchronized with the periodic external force, with hi
p
is too large and lies outside this region, then there is no synchronization. Rather, (t)
increases on average linearly with time. In this case we have h(t)i = 0 + pq t + 1p (t),
where
Z
d
d
dt =
=
T =
.
(4.120)
G()
G()

Thus, (t) = t + (t), where (t) = (t + T ) is periodic with period T = 2/ . This

leads to heterodyning with a beat frequency (, ).
Why do we here find the general resonance condition = pq 0 , whereas for weakly forced,
weakly nonlinear oscillators resonance could only occur for = 0 ? There are two reasons.
The main reason is that in the latter case, the limit cycle is harmonic to zeroth order, with
x0 (t) = A cos(t + ). There are only two frequencies, then, in the Fourier decomposition
of the limit cycle: 0 = 1. In the strongly nonlinear case, the limit cycle is decomposed
into what is in general a countably infinite set of frequencies which are all multiples of a
fundamental 0 . In addition, if the forcing f (, t) is periodic in t, its Fourier decomposition
in t will involve all integer multiples of some fundamental . Thus, the most general
resonance condition is k0 + l = 0.
Our analysis has been limited to the lowest order in , and we have averaged out the
nonresonant terms. When one systematically accounts for both these features, there are
two main effects. One is that the boundaries of the synchronous region are no longer straight
lines as depicted in the right panel of fig. 4.8. The boundaries themselves can be curved.
Moreover, even if there are no resonant terms in the (k, l) sum to lowest order, they can be
generated by going to higher order in . In such a case, the width of the synchronization

96

CHAPTER 4. NONLINEAR OSCILLATORS

region will be proportional to a higher power of : n , where n is the order of

where resonant forcing terms first appear in the analysis.

4.5

Relaxation Oscillations

We saw how to use multiple time scale analysis to identify the limit cycle of the van der
Pol oscillator when is small. Consider now the opposite limit, where the coefficient of the
damping term is very large. We generalize the van der Pol equation to
x
+ (x) x + x = 0 ,

(4.121)

and suppose 1. Define now the variable

x
y +

Zx
dx (x )
0

x
= + F (x) ,

(4.122)

where F (x) = (x). (y is sometimes called the Li`enard variable, and (x, y) the Li`enard
plane.) Then the original second order equation may be written as two coupled first order
equations:


x = y F (x)
(4.123)
x
(4.124)
y = .

Since 1, the first of these equations is fast and the second one slow . The dynamics
rapidly achieves y F (x), and then slowly evolves along the curve y = F (x), until it is
forced to make a large, fast excursion.
A concrete example is useful. Consider F (x) of the form sketched in Fig. 4.9. This is what
one finds for the van der Pol oscillator, where (x) = x2 1 and F (x) = 31 x3 x. The limit
cycle behavior xLC (t) is sketched in Fig. 4.10. We assume (x) = (x) for simplicity.
Assuming (x) = (x) is symmetric, F (x) is antisymmetric. For the van der Pol oscillator
and other similar cases, F (x) resembles the sketch in fig. 4.9. There are two local extrema:
a local maximum at x = a and a local minimum at x = +a. We define b such that
F (b) = F (a), as shown in the figure; antisymmetry then entails F (b) = F (+a). Starting
from an arbitrary initial condition, the y dynamics are slow, since y = 1 x (we assume
x(0)). So y can be regarded as essentially constant for the fast dynamics of eqn. 4.124,
according to which x(t) flows rapidly to the right if y > F (x) and rapidly to the left if
y < F (x). This fast motion stops when x(t) reaches a point where y = F (x). At this point,
the slow dynamics takes over. Assuming y F (x), we have
y F (x)

y =

x
F (x) x ,

(4.125)

97

4.5. RELAXATION OSCILLATIONS

Figure 4.9: Relaxation oscillations in the so-called Li`enard plane (x, y). The system rapidly
flows to a point on the curve y = F (x), and then crawls slowly along this curve. The slow
motion takes x from b to a, after which the system executes a rapid jump to x = +b,
then a slow retreat to x = +a, followed by a rapid drop to x = b.
which says that
x

x
F (x)

if y F (x)

(4.126)

over the slow segments of the motion, which are the regions x [b, a] and x [a, b]. The
relaxation oscillation is then as follows. Starting at x = b, x(t) increases slowly according
to eqn. 4.126. At x = a, the motion can no longer follow the curve y = F (x), since
y = 1 x is still positive. The motion thus proceeds quickly to x = +b, with


x F (b) F (x)



x a, +b .

(4.127)

After reaching x = +b, the motion once again is slow, and again follows eqn. 4.126,
according to which x(t) now decreases slowly until it reaches x = +a, at which point the
motion is again fast, with


x F (a) F (x)
The cycle then repeats.



x b, +a .

(4.128)

98

CHAPTER 4. NONLINEAR OSCILLATORS

Figure 4.10: A sketch of the limit cycle for the relaxation oscillation studied in this section.
Thus, the limit cycle is given by the following segments:
x [ b, a ]

(x > 0) :

x [ a, b ]

(x > 0) :

x [ a, b ]

(x < 0) :

x [ b, a ]

(x < 0) :

x
F (x)



x F (b) F (x)

x
F (x)



x F (a) F (x) .

(4.129)
(4.130)
(4.131)
(4.132)

A sketch of the limit cycle is given in fig. 4.11, showing the slow and fast portions.
When 1 we can determine approximately the period of the limit cycle. Clearly the
period is twice the time for either of the slow portions, hence
Zb
(x)
T 2 dx
,
x

(4.133)

where F (a) = (a) = 0 and F (b) = F (a). For the van der Pol oscillator, with
(x) = x2 1, we have a = 1, b = 2, and T (3 2 ln 2) .

4.5.1

Example problem

Consider the equation

x
+ |x| 1 x + x = 0 .

(4.134)

99

4.5. RELAXATION OSCILLATIONS

Figure 4.11: Limit cycle for large relaxation oscillations, shown in the phase plane (x, x).

Sketch the trajectory in the Li`enard plane, and find the approximate period of the limit
cycle for 1.
Solution : We define

F (x) = |x| 1

F (x) =

We therefore have


x = y F (x)

1 2

+ 2 x x

21 x2

y =

if x > 0
(4.135)
if x < 0 .

x
,

(4.136)

with y 1 x + F (x).
Setting F (x) = 0 we find x = a, where a = 1 and F (a) = 12 . We also find F (b) =

F (a), where b = 1 + 2. Thus, the limit cycle is as follows: (i) fast motion from x = a
to x = +b, (ii) slow relaxation from x = +b to x = +a, (iii) fast motion from x = +a to
x = b, and (iv) slow relaxation from x = b to x = a. The period is approximately the
time it takes for the slow portions of the cycle. Along these portions, we have y F (x),
and hence y F (x) x.
But y = x/, so
F (x) x

dt =

F (x)
dx ,
x

(4.137)

100

CHAPTER 4. NONLINEAR OSCILLATORS

Figure 4.12: Relaxation oscillations for x
+ |x| 1 x + x = 0 plotted in the Lienard
plane. The solid black curve is y = F (x) = 12 x2 sgn(x) x. The variable y is defined to be
y = 1 x + F (x). Along slow portions of the limit cycle, y F (x).
which we integrate to obtain

1+

Z 2 
Za
1
F (x)
= 2 dx 1
T 2 dx
x
x
1
b
h

i
= 2 2 ln 1 + 2 1.066 .

4.5.2

(4.138)
(4.139)

Multiple limit cycles

For the equation

x
+ F (x) x + x = 0 ,

(4.140)

it is illustrative to consider what sort of F (x) would yield more than one limit cycle. Such
an example is shown in fig. 4.13.
In polar coordinates, it is very easy to construct such examples. Consider, for example, the

101

4.5. RELAXATION OSCILLATIONS

Figure 4.13: Lienard plots for systems with one (left) and two (right) relaxation oscillations.
system
r = sin(r) + cos
= b r ,

(4.141)
(4.142)

with || < 1. First consider the case = 0. Clearly the radial flow is outward for sin(r) > 0
and inward for sin(r) < 0. Thus, we have stable limit cycles at r = 2n + 1 and unstable
limit cycles at r = 2n, for all n Z. With 0 < || < 1, we have


r > 0
for
r 2n + 1 sin1 , 2n + 1 1 sin1
(4.143)
r < 0

for


r 2n + 1 +

sin1 , 2n + 2

sin1

(4.144)

The Poincare-Bendixson theorem then guarantees the existence of stable and unstable limit
cycles. We can put bounds on the radial extent of these limit cycles.


stable limit cycle
: r 2n + 1 1 sin1 , 2n + 1 + 1 sin1
(4.145)
unstable limit cycle


r 2n

sin1 , 2n +

sin1

(4.146)

Note that an unstable limit cycle is a repeller, which is to say that it is stable (an attractor)
if we run the dynamics backwards, sending t t.

102

CHAPTER 4. NONLINEAR OSCILLATORS

Figure 4.14: Three instances of (x).

4.5.3

Example problem

Consider the nonlinear oscillator,

x
+ (x) x + x = 0 ,
with 1. For each case in fig. 4.14, sketch the flow in the Li`enard plane, starting with
a few different initial conditions. For which case(s) do relaxation oscillations occur?
Solution : Recall the general theory of relaxation oscillations. We define
x
y +

Zx
x
dx (x ) = + F (x) ,

in which case the second order ODE for the oscillator may be written as two coupled first
order ODEs:


x
y =
, x = y F (x) .

Since 1, the first of these equations is slow and the second one fast. The dynamics
rapidly achieves y F (x), and then slowly evolves along the curve y = F (x), until it is
forced to make a large, fast excursion.

To explore the dynamics in the Li`enard plane, we plot F (x) versus x, which means we must
integrate (x). This is done for each of the three cases in fig. 4.14.
Note that a fixed point corresponds to x = 0 and x = 0. In the Li`enard plane, this means
x = 0 and y = F (0). Linearizing by setting x = x and y = F (0) + y, we have1
d
dt
1

x
y

y F (0) x
1 x

F (0)
1
0



x
y

We could, of course, linearize about the fixed point in (x, x)

space and obtain the same results.

4.6. APPENDIX I : MULTIPLE TIME SCALE ANALYSIS TO O(2 )

103

Figure 4.15: Phase flows in the Lienard plane for the three examples in fig. 4.14.
The linearized map has trace T = F (0) and determinant D = 1. Since 1 we have
0 < D < 41 T 2 , which means the fixed point is either a stable node, for F (0) > 0, or an
unstable node, for F (0) < 0. In cases (a) and (b) the fixed point is a stable node, while in
case (c) it is unstable. The flow in case (a) always collapses to the stable node. In case (b)
the flow either is unbounded or else it collapses to the stable node. In case (c), all initial
conditions eventually flow to a unique limit cycle exhibiting relaxation oscillations.

4.6

Appendix I : Multiple Time Scale Analysis to O(2 )

Problem : A particle of mass m moves in one dimension subject to the potential

U (x) = 21 m 02 x2 + 13 m 02

x3
,
a

(4.147)

where is a dimensionless parameter.

(a) Find the equation of motion for x. Show that by rescaling x and t you can write this
equation in dimensionless form as
d2 u
+ u = u2 .
ds2

(4.148)

Solution : The equation of motion is

m
x = U (x)
= m02 x m02

(4.149)
x2
a

(4.150)

We now define s 0 t and u x/a, yielding

d2 u
+ u = u2 .
ds2

(4.151)

104

CHAPTER 4. NONLINEAR OSCILLATORS

(b) You are now asked to perform an O 2 multiple time scale analysis of this problem,
writing
T0 = s
,
T1 = s
,
T2 = 2 s ,
and
u = u0 + u1 + 2 u2 + . . . .
This results in a hierarchy of coupled equations for the functions {un }. Derive the first
three equations in the hierarchy.
Solution : We have

+
+ 2
+ ... .
=
ds
T0
T1
T2

(4.152)

Therefore
2 

 


u0 + u1 + 2 u2 + . . . + u0 + u1 + 2 u2 + . . .
+
+ 2
+ ...
T0
T1
T2

2
= u0 + u1 + 2 u2 + . . .
.

(4.153)

Expanding and then collecting terms order by order in , we derive the hierarchy. The first
three levels are
2u0
+ u0 = 0
T02

(4.154)

2u1
2u0
u20
+
u
=
2
1
T0 T1
T02

(4.155)

2u2
2u0
2u0
2u1

2 u0 u1 .
+ u2 = 2
2
2
2
T0 T2
T0 T1
T0
T1

(4.156)

(c) Show that there is no frequency shift to first order in .

Solution : At the lowest (first) level of the hierarchy, the solution is


u0 = A(T1 , T2 ) cos T0 + (T1 , T2 ) .
At the second level, then,

A
2 u1

+ u1 = 2
sin(T0 + ) + 2A
cos(T0 + ) A2 cos2(T0 + ) .
2
T1
T1
T0
We eliminate the resonant forcing terms on the RHS by demanding
A
=0
T1

and

=0.
T1

Thus, we must have A = A(T2 ) and = (T2 ). To O(), then, is a constant, which means
there is no frequency shift at this level of the hierarchy.


4.7. APPENDIX II : MSA AND POINCARE-LINDSTEDT
METHODS

105

(d) Find u0 (s) and u1 (s).

Solution :The equation for u1 is that of a non-resonantly forced harmonic oscillator. The
solution is easily found to be
u1 = 12 A2 + 16 A2 cos(2T0 + 2) .
We now insert this into the RHS of the third equation in the hierarchy:
2 u2
2 u0
2 u0 u1
+ u2 = 2
2
T0 T2
T0
o
n
A

sin(T0 + ) + 2A
cos(T0 + ) 2A cos(T0 + ) 21 A2 + 16 A2 cos(2T0 + 2)
T2
T2



A
sin(T0 + ) + 2A
+ 56 A3 cos(T0 + ) 16 A3 cos(3T0 + 3) .
=2
T2
T2

=2

Setting the coefficients of the resonant terms on the RHS to zero yields

2A

A
=0
T2

A = A0

+ 5 A3 = 0
T2 6

5
= 12
A20 T2 .

Therefore,
u0 (s)

u1 (s)

z
z
}|
}|
{

{




5 2 2
u(s) = A0 cos s 12
A0 s + 16 A20 cos 2s 65 2 A20 s 12 A20 +O 2

4.7
4.7.1

Appendix II : MSA and Poincar

e-Lindstedt Methods
Problem using multiple time scale analysis

Consider the central force law F (r) = k r

2 3

.
2

(a) Show that a stable circular orbit exists at radius r0 = (2 /k)1/ .

Solution : For a circular orbit, the effective radial force must vanish:
Feff (r) =

2
2
k
+ F (r) = 3 3 2 = 0 .
3
r
r
r
2

(4.157)

Solving for r = r0 , we have r0 = (2 /k)1/ . The second derivative of Ueff (r) at this point
is
2 2
k
32

,
(4.158)
Ueff
(r0 ) = Feff
(r0 ) = 4 + ( 2 3) 4 2 =
r0
r04
r
0

106

CHAPTER 4. NONLINEAR OSCILLATORS

which is manifestly positive. Thus, the circular orbit at r = r0 is stable.

(b) Show that the geometric equation for the shape of the orbit may be written
d2 s
+ s = K(s)
d2
where s = 1/r, and
K(s) = s0

s
s0

1 2

(4.159)

(4.160)

with s0 = 1/r0 .
Solution : We have previously derived (e.g. in the notes) the equation
d2 s

+ s = 2 2 F (s1 ) .
2
d
s

(4.161)

From the given F (r), we then have

k
d2 s
2
+ s = 2 s1 K(s) ,
2
d

(4.162)

where s0 (k/2 )1/ = 1/r0 , and where

K(s) = s0

s
s0

1 2

(4.163)

(c) Writing s (1 + u) s0 , show that u satisfies

1 d2 u
+ u = a1 u2 + a2 u3 + . . . .
2 d2

(4.164)

Find a1 and a2 .
Solution : Writing s s0 (1 + u), we have
d2 u
2
+ 1 + u = (1 + u)1
2
d
= 1 + (1 2 ) u + 12 ( 2 )(1 2 ) u2
+ 16 (1 2 )( 2 )(1 2 ) u3 + . . . .
Thus,

1 d2 u
+ u = a1 u2 + a2 u3 + . . . ,
2 d2

(4.165)

(4.166)

where
a1 = 12 (1 2 ) ,

a2 = 16 (1 4 ) .

(4.167)


4.7. APPENDIX II : MSA AND POINCARE-LINDSTEDT
METHODS

107

(d) Now let us associate a power of with each power of the deviation u and write
1 d2 u
+ u = a1 u2 + 2 a2 u3 + . . . ,
2 d2

(4.168)

Solve this equation using the method of multiple scale analysis (MSA). You will have to go
to second order in the multiple scale expansion, writing
X ,
and hence

Y ,

Z 2

(4.169)

1 d

=
+
+ 2
+ ... .
d
X
Y
Z

(4.170)

u = u0 + u1 + 2 u2 + . . . ,

(4.171)

Further writing
derive the equations for the multiple scale analysis, up to second order in .
Solution : We now associate one power of with each additional power of u beyond order
u1 . In this way, a uniform expansion in terms of will turn out to be an expansion in
powers of the amplitude of the oscillations. Well see how this works below. We then have
1 d2 u
+ u = a1 u2 + a2 2 u3 + . . . ,
2 d2

(4.172)

with = 1. We now perform a multiple scale analysis, writing

X ,

Z 2 .

(4.173)

This entails

1 d
=
+
+ 2
+ ... .
d
X
Y
Z
We also expand u in powers of , as
u = u0 + u1 + 2 u2 + . . . .

(4.174)

(4.175)

Thus, we obtain
X + Y + 2 Z + . . .

2

(u0 + u1 + 2 u2 + . . . ) + (u0 + u1 + 2 u2 + . . . )

= a1 (u0 + u1 + 2 u2 + . . . )2 + 2 a2 (u0 + u1 + 2 u2 + . . . )3 + . . . . (4.176)

We now extract a hierarchy of equations, order by order in powers of .
We find, out to order 2 ,
O(0 ) :

2 u0
+ u0 = 0
X 2

(4.177)

O(1 ) :

2 u0
2 u1
+
u
=
2
+ a1 u20
1
X 2
Y X

(4.178)

O(2 ) :

2 u0
2 u0
2 u1
2 u2
+
u
=
2

2
+ 2a1 u0 u1 + a2 u30 .
2
X 2
Z X
Y 2
Z X

(4.179)

108

CHAPTER 4. NONLINEAR OSCILLATORS

(e) Show that there is no shift of the angular period = 2/ if one works only to
leading order in .
Solution : The O(0 ) equation in the hierarchy is solved by writing
u0 = A cos(X + ) ,

(4.180)

where
A = A(Y, Z)

= (Y, Z) .

(4.181)

We define X + (Y, Z), so we may write u0 = A cos . At the next order, we obtain
2 u1

A
sin + 2A
cos + a1 A2 cos
+ u1 = 2
2
Y
Y
=2

A
sin + 2A
cos + 21 a1 A2 + 12 a1 A2 cos 2 .
Y
Y

(4.182)

In order that there be no resonantly forcing terms on the RHS of eqn. 5.37, we demand
A
=0 ,
Y

=0
Y

A = A(Z) ,

= (Z) .

(4.183)

The solution for u1 is then

u1 () = 12 a1 A2 16 a1 A2 cos 2 .

(4.184)

Were we to stop at this order, we could ignore Z = 2 entirely, since it is of order 2 ,

and the solution would be
u() = A0 cos( + 0 ) + 12 a1 A20 61 a1 A20 cos(2 + 20 ) .

(4.185)

The angular period is still = 2/, and, starting from a small amplitude solution at
order 0 we find that to order we must add a constant shift proportional to A20 , as well as
a second harmonic term, also proportional to A20 .
(f) Carrying out the MSA to second order in , show that the shift of the angular period
vanishes only if 2 = 1 or 2 = 4.
Solution : Carrying out the MSA to the next order, O(2 ), we obtain
2 u2
A

+ u2 = 2
sin + 2A
cos + 2a1 A cos
2
Z
Z
=2

A
sin + 2A
cos +
Z
Z

5 2
6 a1

2
1
2 a1 A

16 a1 A2 cos 2 + a2 A3 cos3

+ 43 a2 A3 cos + 16 a21 + 14 a2 A3 cos 3 .

(4.186)

Now in order to make the resonant forcing terms on the RHS vanish, we must choose
A
=0
Z

(4.187)


4.7. APPENDIX II : MSA AND POINCARE-LINDSTEDT
METHODS

109

as well as

=
Z

5 2
12 a1

+ 38 a2 A2

(4.188)

1
( 2 4)( 2 1) .
= 24

(4.189)

The solutions to these equations are trivial:

A(Z) = A0

(Z) = 0

2
1
24 (

1)( 2 4)A20 Z .

(4.190)

With the resonant forcing terms eliminated, we may write

2 u2
+ u2 = 16 a21 + 41 a2 A3 cos 3 ,
2

with solution

u2 =
=

2
1
96 (2a1
1
96

(4.191)

3a2 ) A3 cos 3

2 ( 2 1) A20 cos 3X + 3(Z) .

(4.192)

The full solution to second order in this analysis is then

u() = A0 cos( + 0 ) + 12 a1 A20 61 a1 A20 cos(2 + 20 )
+
with

2
1 2
96 (2a1

3a2 ) A30 cos(3 + 30 ) .

n
= 1

1
24

2 ( 2 1)( 2 4)A20

2 n
2
=
1+

1
24

o
2 ( 2 1)( 2 4)A20 + O(3 ) .

The angular period shifts:

=

(4.193)

(4.194)

(4.195)

Note that there is no shift in the period, for any amplitude, if 2 = 1 (i.e. Kepler potential)
or 2 = 4 (i.e. harmonic oscillator).

4.7.2

Solution using Poincar

e-Lindstedt method

Recall that geometric equation for the shape of the (relative coordinate) orbit for the two
body central force problem is
d2 s
+ s = K(s)
d2
 1 2
s
K(s) = s0
s0

(4.196)
(4.197)

110

CHAPTER 4. NONLINEAR OSCILLATORS

where s = 1/r, s0 = (l2 /k)1/ is the inverse radius of the stable circular orbit, and
2
f (r) = kr 3 is the central force. Expanding about the stable circular orbit, one has
d2 y
+ 2 y = 12 K (s0 ) y 2 + 61 K (s0 ) y 3 + . . . ,
d2

(4.198)

where s = s0 (1 + y), with

K (s) = (1 )

s0
s
2

K (s) = (1 )

 2


s0
s

(4.199)
1+ 2
2

K (s) = (1 ) (1 + )
Thus,

s0
s

(4.200)
2+ 2

(4.201)

d2 y
+ 2 y = a1 y 2 + 2 a2 y 3 ,
d2

(4.202)

a1 = 12 2 (1 2 )

(4.203)

with = 1 and
a2 = + 16 2 (1 2 ) (1 + 2 ) .

(4.204)

Note that we assign one factor of for each order of nonlinearity beyond order y 1 . Note
also that while y here corresponds to u in eqn. 4.166, the constants a1,2 here are a factor
of 2 larger than those defined in eqn. 4.167.
We now apply the Poincare-Lindstedt method, by defining = , with
2 = 20 + 21 + 2 22 + . . .

(4.205)

y() = y0 () + y1 () + 2 y2 () + . . . .

(4.206)

d
d
=
d
d

(4.207)

and
We therefore have
and
20 + 21 +2 22 + . . .

= a1

y0 + y1 + 2 y2 + . . . + 2 y0 + y1 + 2 y2 + . . .
(4.208)




3
2
y0 + y1 + 2 y2 + . . . + 2 a2 y0 + y1 + 2 y2 + . . . .
(4.209)

We now extract equations at successive orders of . The first three in the hierarchy are
20 y0 + 2 y0 = 0
21 y0 + 20 y1 + 2 y1 = a1 y02
22 y0 + 21 y1 + 20 y2 + 2 y2 = 2 a1 y0 y1 + a2 y03 ,

(4.210)
(4.211)
(4.212)

4.8. APPENDIX III : MODIFIED VAN DER POL OSCILLATOR

111

where prime denotes differentiation with respect to .

To order 0 , the solution is 20 = 2 and
y0 () = A cos( + ) ,

(4.213)

where A and are constants.

At order 1 , we have


2 y1 + y1 = 21 y0 + a1 y02

= 21 A cos( + ) + a1 A2 cos2 ( + )

= 21 A cos( + ) + 12 a1 A2 + 21 a1 A2 cos(2 + 2) .

(4.214)

The secular forcing terms on the RHS are eliminated by the choice 21 = 0. The solution is
then
o
a1 A2 n
1
1

y1 () =
cos(2
+
2)
.
(4.215)
3
2 2
At order 2 , then, we have


2 y2 + y2 = 22 y0 21 y1 + 2 a1 y1 y1 + a2 y03
o
a2 A3 n
1 31 cos(2 + 2) cos( + ) + a2 A3 cos2 ( + )
= 22 A cos( + ) + 1 2





2
3
5 a1 A
a21 A3 1
2
3
3
3
= 2 +
+ 4 a2 A A cos( + ) +
+ 4 a2 A cos(3 + 3) .
6 2
6 2
(4.216)
The resonant forcing terms on the RHS are eliminated by the choice


22 = 56 2 a21 + 34 a2 A3
h
i
1 2
= 24
(1 2 ) 5 (1 2 ) + 3 (1 + 2 )
1 2
(1 2 ) (4 2 ) .
= 12

(4.217)

Thus, the frequency shift to this order vanishes whenever 2 = 0, 2 = 1, or 2 = 4. Recall

2
the force law is F (r) = C r 3 , so we see that there is no shift hence no precession
for inverse cube, inverse square, or linear forces.

4.8

Appendix III : Modified van der Pol Oscillator

Consider the nonlinear oscillator

x
+ (x4 1) x + x = 0 .
Analyze this using the same approach we apply to the van der Pol oscillator.

(4.218)

112

CHAPTER 4. NONLINEAR OSCILLATORS

Figure 4.16: Sketch of phase flow and nullclines for the oscillator x
+ (x4 1) x + x = 0.
Red nullclines: v = 0; blue nullcline: x = 0.
(a) Sketch the vector field for this problem. It may prove convenient to first identify the
nullclines, which are the curves along which x = 0 or v = 0 (with v = x).
Argue that a
limit cycle exists.
Solution : There is a single fixed point, at the origin (0, 0), for which the linearized
dynamics obeys
  
 
d x
0 1
x
=
+ O(x4 v) .
(4.219)
1
v
dt v
One finds T = and D = 1 for the trace and determinant, respectively. The origin is an
unstable spiral for 0 < < 2 and an unstable node for > 2.
The nullclines are sketched in Fig. 4.16. One has
x = 0 v = 0

v = 0 v =

1 x
.
1 x4

(4.220)

The flow at large distances from the origin winds once around the origin and spirals in. The
flow close to the origin spirals out ( < 2) or flows radially out ( > 2). Ultimately the flow
must collapse to a limit cycle, as can be seen in the accompanying figures.
(b) In the limit 0 < 1, use multiple time scale analysis to obtain a solution which
reveals the approach to the limit cycle.

113

4.8. APPENDIX III : MODIFIED VAN DER POL OSCILLATOR

Solution : We seek to solve the equation

x
+ x = h(x, x)
,

(4.221)

h(x, x)
= (1 x4 ) x .

(4.222)

with
Employing the multiple time scale analysis to lowest nontrivial order, we write T0 t,
T1 t,
x = x0 + x1 + . . .
(4.223)
and identify terms order by order in . At O(0 ), this yields
2 x0
+ x0 = 0
T02

x0 = A cos(T0 + ) ,

(4.224)

where A = A(T1 ) and = (T1 ). At O(1 ), we have



2 x0
x0
2 x1
+ x1 = 2
+ h x0 ,
T0 T1
T0
T02
=2

A
sin + 2A
cos + h A cos , A sin
T1
T1

with = T0 + (T1 ) as usual. We also have

h(A cos , A sin = A5 sin cos A sin


3
= 18 A5 A sin + 16
A5 sin 3 +

1
16

A5 sin 5 .

(4.225)

(4.226)

To eliminate the resonant terms in eqn. 4.225, we must choose

A
= 12 A
T1

1 5
16 A

=0.
T1

(4.227)

The A equation is similar to the logistic equation. Clearly A = 0 is an unstable fixed point,
and A = 81/4 1.681793 is a stable fixed point. Thus, the amplitude of the oscillations
will asymptotically approach A = 81/4 . (Recall the asymptotic amplitude in the van der
Pol case was A = 2.)
To integrate the A equation, substitute y =
dT1 =

dy
y2
1
d
ln
=
2
y (1 y 2 )
1 y2

1 A2 ,
8

and obtain

y 2 (T1 ) =

1+

(y02

1
.
1) exp(2T1 )

(4.228)

!1/4

(4.229)

We then have
1/4

A(T1 ) = 8

y(T1 ) =

8
4
1 + (8A0 1) exp(2T1 )

114

CHAPTER 4. NONLINEAR OSCILLATORS

(c) In the limit 1, find the period of relaxation oscillations, using Lienard plane
analysis. Sketch the orbit of the relaxation oscillation in the Lienard plane.
Solution : Our nonlinear oscillator may be written in the form
x
+

dF (x)
+x=0 ,
dt

(4.230)

with
F (x) = 51 x5 x .

(4.231)

Note F = (x4 1) x.
Now we define the Lienard variable
y

x
+ F (x) ,

(4.232)

and in terms of (x, y) we have

h
i
x = y F (x)

y =

x
.

(4.233)

As we have seen in the notes, for large the motion in the (x, y) plane is easily analyzed.
x(t) must move quickly over to the curve y = F (x), at which point the motion slows down
and slowly creeps along this curve until it can no longer do so, at which point another big
fast jump occurs. The jumps take place between the local extrema of F (x), which occur
for F (a) = a4 1 = 0, i.e. at a = 1, and points on the curve with the same values of
F (a). Thus, we solve F (1) = 45 = 15 b5 b and find the desired root at b 1.650629. The
period of the relaxation oscillations, for large , is
Zb
ib
h
F (x)
T 2 dx
= 12 x4 2 ln x 2.20935 .
x
a

(4.234)

(d) Numerically integrate the equation (4.218) starting from several different initial conditions.
Solution : The accompanying Mathematica plots show x(t) and v(t) for this system for
two representative values of .

4.8. APPENDIX III : MODIFIED VAN DER POL OSCILLATOR

115

Figure 4.17: Vector field and phase curves for the oscillator x
+ (x4 1) x + x = 0, with
= 1 and starting from (x0 , v0 ) = (1, 1).

Figure 4.18: Solution to the oscillator equation x

+ (x4 1) x + x = 0 with = 1 and initial
conditions (x0 , v0 ) = (1, 3). x(t) is shown in red and v(t) in blue. Note that x(t) resembles
a relaxation oscillation for this moderate value of .

116

CHAPTER 4. NONLINEAR OSCILLATORS

Figure 4.19: Vector field and phase curves for the oscillator x
+ (x4 1) x + x = 0, with
= 0.25 and starting from (x0 , v0 ) = (1, 1). As 0, the limit cycle is a circle of radius
A = 81/4 1.682.

Figure 4.20: Solution to the oscillator equation x

+ (x4 1) x + x = 0 with = 0.25 and
initial conditions (x0 , v0 ) = (1, 3). x(t) is shown in red and v(t) in blue. As 0, the
amplitude of the oscillations tends to A = 81/4 1.682.

Chapter 5

Hamiltonian Mechanics
5.1

The Hamiltonian

Recall that L = L(q, q,

t), and
p =

L
.
q

(5.1)

The Hamiltonian, H(q, p) is obtained by a Legendre transformation,

H(q, p) =

n
X

=1

p q L .

(5.2)

Note that
n 
X

L
L
dH =
p dq + q dp
dq
dq
q
q
=1

n 
X
L
L
dt .
dq
=
q dp
q
t
=1

L
dt
t
(5.3)

Thus, we obtain Hamiltons equations of motion,

H
= q
p
and

H
L
=
= p
q
q

dH
H
L
=
=
.
dt
t
t

(5.4)

(5.5)

Some remarks:
As an example, consider a particle moving in three dimensions, described by spherical
polar coordinates (r, , ). Then


(5.6)
L = 21 m r 2 + r 2 2 + r 2 sin2 2 U (r, , ) .
117

118

CHAPTER 5. HAMILTONIAN MECHANICS

We have
pr =

L
= mr
r

p =

L
= mr 2

p =

L
= mr 2 sin2 ,

(5.7)

and thus
H = pr r + p + p L
=

p2
p2
p2r
+
+
+ U (r, , ) .
2m 2mr 2 2mr 2 sin2

Note that H is time-independent, hence

of the motion.

H
t

dH
dt

(5.8)

= 0, and therefore H is a constant

In order to obtain H(q, p) we must invert the relation p =

L
q

= p (q, q)
to obtain

q (q, p). This is possible if the Hessian,

2L
p
=
q
q q

(5.9)

is nonsingular. This is the content of the inverse function theorem of multivariable

calculus.
Define the rank 2n vector, , by its components,
(
qi
i =
pin

if 1 i n
if n i 2n .

(5.10)

Then we may write Hamiltons equations compactly as

H
i = Jij
,
j
where
J=

Onn Inn
Inn Onn

(5.11)

(5.12)

is a rank 2n matrix. Note that J t = J, i.e. J is antisymmetric, and that J 2 =

I2n2n . We shall utilize this symplectic structure to Hamiltons equations shortly.

119

5.2. MODIFIED HAMILTONS PRINCIPLE

5.2

Modified Hamiltons Principle

We have that
Ztb
Ztb


0 = dt L = dt p q H
ta

(5.13)

ta


Ztb 
H
H
= dt p q + q p
q
p
q p
ta

)



Ztb ( 

tb
H
H
q + q
p + p q  ,
= dt p +
q
p
ta
ta

assuming q (ta ) = q (tb ) = 0. Setting the coefficients of q and p to zero, we recover

Hamiltons equations.

5.3

Phase Flow is Incompressible

A flow for which v = 0 is incompressible we shall see why in a moment. Lets check
that the divergence of the phase space velocity does indeed vanish:
=
=

n 
X
q

=1
2n
X
i=1

p
+
q
p

i X
2H
=
Jij
=0.
i
i j

(5.14)

i,j

Now let (, t) be a distribution on phase space. Continuity implies

=0.
+ ( )
t

(5.15)

D
=
+ = 0 ,
Dt
t

(5.16)

Invoking = 0, we have that

where D/Dt is sometimes

called the convective derivative it is the total derivative of the


function (t), t , evaluated at a point (t) in phase space which moves according to the
dynamics. This says that the density in the comoving frame is locally constant.

120

5.4

CHAPTER 5. HAMILTONIAN MECHANICS

Poincar
e Recurrence Theorem

Let g be the -advance mapping which evolves points in phase space according to Hamiltons equations
H
H
qi = +
,
p i =
(5.17)
pi
qi
for a time interval t = . Consider a region in phase space. Define gn to be the
nth image of under the mapping g . Clearly g is invertible; the inverse is obtained by
integrating the equations of motion backward in time. We denote the inverse of g by g1 .
By Liouvilles theorem, g is volume preserving when acting on regions in phase space, since
the evolution of any given point is Hamiltonian. This follows from the continuity equation
for the phase space density,

+ (u) = 0
(5.18)
t
p}
is the velocity vector in phase space, and Hamiltons equations, which
where u = {q,
say that the phase flow is incompressible, i.e. u = 0:

n 
X
qi pi
u =
+
qi pi
i=1
(



)
n
X
H

H
=0.
(5.19)
=
+

qi pi
pi
qi
i=1

Thus, we have that the convective derivative vanishes, viz.

+ u = 0 ,
Dt
t

(5.20)

which guarantees that the density remains constant in a frame moving with the flow.
The proof of the recurrence theorem is simple. Assume that g is invertible and volumepreserving, as is the case for Hamiltonian flow. Further assume that phase space volume
is finite. Since the energy is preserved in the case of time-independent Hamiltonians, we
simply ask that the volume of phase space at fixed total energy E be finite, i.e.
Z


d E H(q, p) < ,
(5.21)

where d = dq dp is the phase space uniform integration measure.

Theorem: In any finite neighborhood of phase space there exists a point 0 which will
return to after n applications of g , where n is finite.
Proof: Assume the theorem fails; we will show this assumption results in a contradiction.
Consider the set formed from the union of all sets gm for all m:
=

m=0

gm

(5.22)

121

5.5. POISSON BRACKETS

We assume that the set {gm | m Z , m 0} is disjoint. The volume of a union of disjoint
sets is the sum of the individual volumes. Thus,
vol() =

vol(gm )

m=0

= vol()

m=1

1=,

(5.23)

since vol(gm ) = vol() from volume preservation. But clearly is a subset of the entire
phase space, hence we have a contradiction, because by assumption phase space is of finite
volume.
Thus, the assumption that the set {gm | m Z , m 0} is disjoint fails. This means that
there exists some pair of integers k and l, with k 6= l, such that gk gl 6= . Without
loss of generality we may assume k > l. Apply the inverse g1 to this relation l times to get
gkl 6= . Now choose any point gn , where n = k l, and define 0 = gn .
Then by construction both 0 and gn 0 lie within and the theorem is proven.
Each of the two central assumptions invertibility and volume preservation is crucial.
Without either of them, the proof fails. Consider, for example, a volume-preserving map
which is not invertible. An example might be a mapping f : R R which takes any real
number to its fractional part. Thus, f () = 0.14159265 . . .. Let us restrict our attention
to intervals of width less than unity. Clearly f is then volume preserving. The action of f
on the interval [2, 3) is to map it to the interval [0, 1). But [0, 1) remains fixed under the
action of f , so no point within the interval [2, 3) will ever return under repeated iterations
of f . Thus, f does not exhibit Poincare recurrence.
Consider next the case of the damped harmonic oscillator. In this case, phase space volumes
2
contract. For a one-dimensional oscillator obeying x
+2 x+

0 x = 0 one has u = 2 <

0 ( > 0 for damping). Thus the convective derivative obeys Dt = (u) = +2 which
says that the density increases exponentially in the comoving frame, as (t) = e2t (0).
Thus, phase space volumes collapse, and are not preserved by the dynamics. In this case, it
is possible for the set to be of finite volume, even if it is the union of an infinite number of
sets gn , because the volumes of these component sets themselves decrease exponentially,
as vol(gn ) = e2n vol(). A damped pendulum, released from rest at some small angle
0 , will not return arbitrarily close to these initial conditions.

5.5

Poisson Brackets

The time evolution of any function F (q, p) over phase space is given by

n 

F X
F
F
d
F q(t), p(t), t =
+
q +
p
dt
t
q
p
=1

F 
+ F, H ,

(5.24)

122

CHAPTER 5. HAMILTONIAN MECHANICS

where the Poisson bracket { , } is given by




n 
X
A B
A B

A, B
q p
p q

=1
2n
X

i,j=1

Jij

A B
.
i j

(5.25)
(5.26)

Properties of the Poisson bracket:

Antisymmetry:



f, g = g, f .

(5.27)

Bilinearity: if is a constant, and f , g, and h are functions on phase space, then






f + g, h = f, h + {g, h .

(5.28)

Linearity in the second argument follows from this and the antisymmetry condition.
Associativity:







f g, h = f g, h + g f, h .

Jacobi identity:




f, {g, h} + g, {h, f } + h, {f, g} = 0 .

(5.29)

(5.30)

Some other useful properties:

If {A, H} = 0 and

A
t

= 0, then

dA
dt

= 0, i.e. A(q, p) is a constant of the motion.



If {A, H} = 0 and {B, H} = 0, then {A, B}, H = 0. If in addition A and B have
no explicit time dependence, we conclude that {A, B} is a constant of the motion.
It is easily established that
{q , q } = 0 ,

5.6
5.6.1

{p , p } = 0

{q , p } = .

(5.31)

Canonical Transformations
Point transformations in Lagrangian mechanics

In Lagrangian mechanics, we are free to redefine our generalized coordinates, viz.

Q = Q (q1 , . . . , qn , t) .

(5.32)

123

5.6. CANONICAL TRANSFORMATIONS

This is called a point transformation. The transformation is invertible if



Q
det
q

6= 0 .

(5.33)

written as a function of the new coordinates Q and velocThe transformed Lagrangian, L,

is
ities Q,


Q, Q,
t) = L q(Q, t), q(Q,
t) .
L

Q,
(5.34)

Finally, Hamiltons principle,

Ztb

t) = 0
dt L(Q,
Q,

(5.35)

t1

with Q (ta ) = Q (tb ) = 0, still holds, and the form of the Euler-Lagrange equations
remains unchanged:
 

d L
L

=0.
(5.36)
Q
dt Q
The invariance of the equations of motion under a point transformation may be verified
explicitly. We first evaluate

 



d L q
d L q
d L
,
=
=
dt Q
dt q Q
dt q Q

(5.37)

q
q
=

Q
Q

(5.38)

where the relation

follows from
q =

q
q
Q +
.
Q
t

(5.39)

Now we compute

L
L q
L q
=
+
Q
q Q
q Q


L q
2 q
L
2 q

=
+
Q +
q Q
q Q Q
Q t




L d q
d L q
+
=
dt q Q
q dt Q


 
d L q
d L
=
=
,
dt q Q
dt Q
where the last equality is what we obtained earlier in eqn. 5.37.

(5.40)

124

5.6.2

CHAPTER 5. HAMILTONIAN MECHANICS

Canonical transformations in Hamiltonian mechanics

In Hamiltonian mechanics, we will deal with a much broader class of transformations ones
which mix all the q s and p s. The general form for a canonical transformation (CT) is


q = q Q1 , . . . , Qn ; P1 , . . . , Pn ; t
(5.41)


p = p Q1 , . . . , Qn ; P1 , . . . , Pn ; t ,
(5.42)
with {1, . . . , n}. We may also write

i = i 1 , . . . , 2n ; t ,

(5.43)

with i {1, . . . , 2n}. The transformed Hamiltonian is H(Q,

P, t).
What sorts of transformations are allowed? Well, if Hamiltons equations are to remain
invariant, then

H
H
, P =
,
(5.44)
Q =
P
Q
which gives

P
i
Q
+
=0=
.
Q
P
i

(5.45)

I.e. the flow remains incompressible in the new (Q, P ) variables. We will also require that
phase space volumes are preserved by the transformation, i.e.


 
 (Q, P ) 
i
 = 1 .


det
(5.46)
= 
j
(q, p) 

Additional conditions will be discussed below.

5.6.3

Hamiltonian evolution

Hamiltonian evolution itself defines a canonical transformation. Let i = i (t) and i =

H
i (t + dt). Then from the dynamics i = Jij
, we have
j
i (t + dt) = i (t) + Jij
Thus,

Now, using the result

H
dt + O dt2 .
j





i

H
2
dt + O dt
=
+ Jik
j
j i
k


2H
dt + O dt2 .
= ij + Jik
j k


det 1 + M = 1 + Tr M + O(2 ) ,

(5.47)

(5.48)

(5.49)

125

5.6. CANONICAL TRANSFORMATIONS

we have
 
 i 


2H
2


 j  = 1 + Jjk j k dt + O dt


= 1 + O dt2 .

5.6.4

(5.50)
(5.51)

Symplectic structure

We have that
H
.
i = Jij
j

(5.52)

Suppose we make a time-independent canonical transformation to new phase space coordinates, a = a (). We then have
a
a
H
a =
j =
Jjk
.
j
j
k

(5.53)

But if the transformation is canonical, then the equations of motion are preserved, and we
also have

k H
H
= Jab
.
(5.54)
a = Jab
b
b k
Equating these two expressions, we have
Maj Jjk

H
1 H
= Jab Mkb
,
k
k

(5.55)

a
j

(5.56)

where
Maj

is the Jacobian of the transformation. Since the equality must hold for all , we conclude
MJ = J Mt

1

M JM t = J .

(5.57)

A matrix M satisfying M M t = I is of course an orthogonal matrix. A matrix M satisfying

M JM t = J is called symplectic. We write M Sp(2n), i.e. M is an element of the group
of symplectic matrices 1 of rank 2n.
The symplectic property of M guarantees that the Poisson brackets are preserved under a
1

Note that the rank of a symplectic matrix is always even. Note also M JM t = J implies M t JM = J.

126

CHAPTER 5. HAMILTONIAN MECHANICS

canonical transformation:


5.6.5

A, B

A B
i j
A a B b
= Jij
a i b j

A B
t
= Mai Jij Mjb
a b
A B
= Jab
a b


= A, B .
= Jij

(5.58)

Generating functions for canonical transformations

For a transformation to be canonical, we require

Ztb n
Ztb n
o
o

dt p q H(q, p, t) = 0 = dt P Q H(Q,
P, t) .
ta

(5.59)

ta

This is satisfied provided



n
o
dF

,
p q H(q, p, t) = P Q H(Q,
P, t) +
dt

(5.60)

where is a constant. For canonical transformations, = 1.2 Thus,

F
F

Q
H(Q,
P, t) = H(q, p, t) + P Q p q +
q +
q
Q
+

F
F
F
P +
p +
.
p
P
t

(5.61)

Thus, we require
F
= p
q

F
= P
Q

F
=0
p

F
=0.
P

(5.62)

The transformed Hamiltonian is

F

H(Q,
P, t) = H(q, p, t) +
.
t

(5.63)

2
Solutions of eqn. 5.60 with 6= 1 are known as extended canonical transformations. We can always
rescale coordinates and/or momenta to achieve = 1.

127

5.6. CANONICAL TRANSFORMATIONS

There are four possibilities, corresponding to the freedom to make Legendre transformations
with respect to each of the arguments of F (q, Q) :

F1
1
F1 (q, Q, t)
; p = + F
, P = Q
(type I)

F2
2

; p = + F
, Q = + P
(type II)

F2 (q, P, t) P Q
q

F (q, Q, t) =

F3
3

, P = Q
(type III)
; q = F
F3 (p, Q, t) + p q

F4
4
, Q = + P
(type IV)
F4 (p, P, t) + p q P Q ; q = F
p

In each case ( = 1, 2, 3, 4), we have

H(Q,
P, t) = H(q, p, t) +
.
t
Lets work out some examples:

(5.64)

Consider the type-II transformation generated by

F2 (q, P ) = A (q) P ,

(5.65)

where A (q) is an arbitrary function of the {q }. We then have

Q =

F2
= A (q)
P

p =

F2
A
=
P .
q
q

(5.66)

Thus,

q
p .
(5.67)
Q
This is a general point transformation of the kind discussed in eqn. 5.32. For a general
1
, i.e. Q = M q,
linear point transformation, Q = M q , we have P = p M
Q = A (q)

P =

P = p M 1 . If M = , this is the identity transformation. F2 = q1 P3 + q3 P1

interchanges labels 1 and 3, etc.
Consider the type-I transformation generated by
F1 (q, Q) = A (q) Q .

(5.68)

We then have
p =

F1
A
=
Q
q
q

P =

F1
= A (q) .
Q

Note that A (q) = q generates the transformation

 


q
P

.
p
+Q

(5.69)
(5.70)

(5.71)

128

CHAPTER 5. HAMILTONIAN MECHANICS

A mixed transformation is also permitted. For example,

F (q, Q) = q1 Q1 + (q3 Q2 ) P2 + (q2 Q3 ) P3

(5.72)

is of type-I with respect to index = 1 and type-II with respect to indices = 2, 3.

The transformation effected is
Q1 = p 1

Q2 = q 3

Q3 = q 2

(5.73)

P1 = q1

P2 = p3

P3 = p2 .

(5.74)

Consider the harmonic oscillator,

H(q, p) =

p2
+ 12 kq 2 .
2m

(5.75)

If we could find a time-independent canonical transformation such that

r
p
2 f (P )
p = 2mf (P ) cos Q
sin Q ,
,
q=
k

(5.76)

where f (P ) is some function of P , then wed have H(Q,

P ) = f (P ), which is cyclic in
Q. To find this transformation, we take the ratio of p and q to obtain

p = mk q ctn Q ,
(5.77)
which suggests the type-I transformation
F1 (q, Q) =

1
2

mk q 2 ctn Q .

(5.78)

This leads to

F1
p=
= mk q ctn Q
q
Thus,

where =

2P
sin Q
q=
4
mk

mk q 2
F1
P =
=
.
Q
2 sin2 Q

f (P ) =

k
P = P ,
m

(5.79)

(5.80)

k/m is the oscillation frequency. We therefore have

H(Q,
P ) = P ,

(5.81)

whence P = E/. The equations of motion are

H
P =
=0 ,
Q
which yields
Q(t) = t + 0

q(t) =

H
Q =
= ,
P

(5.82)

(5.83)

2E
sin t + 0 .
2
m

129

5.7. HAMILTON-JACOBI THEORY

5.7

Hamilton-Jacobi Theory

Weve stressed the great freedom involved in making canonical transformations. Coordinates and momenta, for example, may be interchanged the distinction between them is
purely a matter of convention! We now ask: is there any specially preferred canonical trans
formation? In this regard, one obvious goal is to make the Hamiltonian H(Q,
P, t) and the
corresponding equations of motion as simple as possible.
Recall the general form of the canonical transformation:
F

,
H(Q,
P ) = H(q, p) +
t

(5.84)

with
F
= p
q

F
=0
p

(5.85)

F
= P
Q

F
=0.
P

(5.86)

We now demand that this transformation result in the simplest Hamiltonian possible, that

is, H(Q,
P, t) = 0. This requires we find a function F such that
F
= H
t

F
q

= p .

(5.87)

The remaining functional dependence may be taken to be either on Q (type I) or on P

(type II). As it turns out, the generating function F we seek is in fact the action, S, which
is the integral of L with respect to time, expressed as a function of its endpoint values.

5.7.1

The action as a function of coordinates and time

We have seen how the action S[( )] is a functional of the path ( ) and a function of the
endpoint values {qa , ta } and {qb , tb }. Let us define the action function S(q, t) as
Zt
S(q, t) = d L , ,
) ,

(5.88)

ta

where ( ) starts at (qa , ta ) and ends at (q, t). We also require that ( ) satisfy the EulerLagrange equations,


d L
L

=0
(5.89)

d
Let us now consider a new path, ( ), also starting at (qa , ta ), but ending at (q + dq, t + dt),

130

CHAPTER 5. HAMILTONIAN MECHANICS

and also satisfying the equations of motion. The differential of S is





dS = S ( ) S ( )
=

t+dt
Z

ta

Zt

Zt
d L(
, , ) d L , ,
)

ta

(5.90)

ta

i
i
L h
L h
( ) ( ) +
( ) ( )

)h

Zt (
i
d L
L
= d
( ) ( )

d
ta
 h
i


L 

(t)

(t)
+ L (t), (t), t dt
+

+ L (t), (t), t dt

= 0 + (t) (t) + L (t), (t),

t dt + O(q dt) ,

where we have defined

=

L
,

(5.91)

(5.92)

and
( ) ( ) ( ) .

(5.93)

Note that the differential dq is given by

dq = (t + dt) (t)

(5.94)

= (t + dt) (t) + (t) (t)

= (t) dt + (t)

= q (t) dt + (t) + O(q dt) .

(5.95)

Thus, with (t) p , we have

dS = p dq + L p q dt
= p dq H dt .

We therefore obtain

S
= p
q

S
= H
t

dS
=L.
dt

(5.96)

(5.97)

What 
about the lower limit at ta ? Clearly there are n + 1 constants associated with this
limit: q1 (ta ), . . . , qn (ta ); ta . Thus, we may write
S = S(q1 , . . . , qn ; 1 , . . . , n , t) + n+1 ,

(5.98)

131

5.7. HAMILTON-JACOBI THEORY

Figure 5.1: A one-parameter family of paths q(s; ).

where our n + 1 constants are {1 , . . . , n+1 }. If we regard S as a mixed generator, which
is type-I in some variables and type-II in others, then each for 1 n may be chosen
to be either Q or P . We will define
(
+Q if = P
S
(5.99)
=
=

P if = Q
For each , the two possibilities = Q or = P are of course rendered equivalent by
a canonical transformation (Q , P ) (P , Q ).

5.7.2

The Hamilton-Jacobi equation

Since the action S(q, , t) has been shown to generate a canonical transformation for which

H(Q,
P ) = 0. This requirement may be written as

S
S  S
H q1 , . . . , qn ,
=0.
(5.100)
,...,
,t +
q1
qn
t

This is the Hamilton-Jacobi equation (HJE). It is a first order partial differential equation
in n + 1 variables, and in general is nonlinear (since kinetic energy is generally a quadratic

function of momenta). Since H(Q,

P, t) = 0, the equations of motion are trivial, and
Q (t) = const.

P (t) = const.

(5.101)

Once the HJE is solved, one must invert the relations = S(q, , t)/ to obtain
q(Q, P, t). This is possible only if


2S
6= 0 ,
(5.102)
det
q

132

CHAPTER 5. HAMILTONIAN MECHANICS

which is known as the Hessian condition.

It is worth noting that the HJE may have several solutions. For example, consider the case
of the free particle, with H(q, p) = p2 /2m. The HJE is
 
1 S 2 S
=0.
(5.103)
+
2m q
t
One solution of the HJE is
S(q, , t) =

m (q )2
.
2t

(5.104)

For this we find

S
= (q ) q(t) = t .

t
m
Here = q(0) is the initial value of q, and = p is minus the momentum.
=

(5.105)

Another equally valid solution to the HJE is

S(q, , t) = q

2m t .

(5.106)

This yields

r
r
2m

S
=q
t q(t) =
(t + ) .
(5.107)
=

2m
For
p this solution, is the energy and may be related to the initial value of q(t) =
/2m.

5.7.3

Time-independent Hamiltonians

When H has no explicit time dependence, we may reduce the order of the HJE by one,
writing
S(q, , t) = W (q, ) + T (, t) .
(5.108)
The HJE becomes



W
T
H q,
.
=
q
t

(5.109)

S(q, , t) = W (q, ) 1 t .

(5.110)

Note that the LHS of the above equation is independent of t, and the RHS is independent
of q. Therefore, each side must only depend on the constants , which is to say that each
side must be a constant, which, without loss of generality, we take to be 1 . Therefore

The function W (q, ) is called Hamiltons characteristic function. The HJE now takes the
form


W
W
,...,
= 1 .
(5.111)
H q1 , . . . , qn ,
q1
qn

Note that adding an arbitrary constant C to S generates the same equation, and simply
shifts the last constant n+1 n+1 + C. This is equivalent to replacing t by t t0 with
t0 = C/1 , i.e. it just redefines the zero of the time variable.

133

5.7. HAMILTON-JACOBI THEORY

5.7.4

Example: one-dimensional motion

As an example of the method, consider the one-dimensional system,

H(q, p) =
The HJE is

(5.112)

 
1 S 2
+ U (q) = .
2m q

(5.113)

q 

2m U (q) ,

(5.114)

which may be recast as

S
=
q
with solution

p2
+ U (q) .
2m

Z
p

S(q, , t) = 2m dq U (q ) t .

We now have

S
=
p=
q

as well as
S
=
=

m
2



2m U (q) ,
q(t)

Thus, the motion q(t) is given by quadrature:

+t=

m
2

dq
t .
U (q )

Zq(t)
dq
p
,
U (q )

(5.115)

(5.116)

(5.117)

(5.118)

where and are constants. The lower limit on the integral is arbitrary and merely shifts
t by another constant. Note that is the total energy.

5.7.5

Separation of variables

It is convenient to first work an example before discussing the general theory. Consider the
following Hamiltonian, written in spherical polar coordinates:
potential U (r,,)

}|
{
 z
p2
B()
1
C()
2
H=
+
p +
+ A(r) + 2 + 2 2 .
2m r r 2 r 2 sin2
r
r sin


p2

(5.119)

We seek a solution with the characteristic function

W (r, , ) = Wr (r) + W () + W () .

(5.120)

134

CHAPTER 5. HAMILTONIAN MECHANICS

The HJE is then







W 2
1 Wr 2
W 2
1
1
+
+
2m r
2mr 2
2mr 2 sin2
+ A(r) +

C()
B()
+ 2 2 = 1 = E .
2
r
r sin

(5.121)

Multiply through by r 2 sin2 to obtain

(
)



2
1 W 2
W
1

+ C() = sin2
+ B()
2m
2m
(
)


1 Wr 2
2
2
r sin
+ A(r) 1 .
2m r

(5.122)

The LHS is independent of (r, ), and the RHS is independent of . Therefore, we may set


1 W 2
+ C() = 2 .
2m

(5.123)

Proceeding, we replace the LHS in eqn. 5.122 with 2 , arriving at

(
)



2
1
W
2
1 W 2
r
= r 2
+ B() +
+ A(r) 1 .
2m
2m r
sin2

(5.124)

The LHS of this equation is independent of r, and the RHS is independent of . Therefore,


1 W 2
2
+ B() +
= 3 .
2m
sin2

(5.125)



3
1 Wr 2
+ A(r) + 2 = 1 .
2m r
r

(5.126)

Were left with

The full solution is therefore

S(q, , t) =

Zr r
3
2m dr 1 A(r ) 2
r
+

(5.127)

Z r
2
2m d 3 B( )
sin2
+

2m

2 C( ) 1 t .

(5.128)

135

5.7. HAMILTON-JACOBI THEORY

We then have
S
=
1 =
1

r(t)

pm
2

dr

t
1 A(r ) 3 r 2
pm
Z (t) p m
Z (t)
d
S
2 d
q
q 2
2 =
+
=
2
sin2 3 B( ) 2 csc2
2 C( )
pm
pm
Z r(t)
Z (t)
S
2 dr
2 d
q
=
+ q
.
3 =
3
r 2 1 A(r ) 3 r 2
3 B( ) 2 csc2
q

(5.129)

(5.130)

(5.131)

The game plan here is as follows. The first of the above trio of equations is inverted to yield
r(t) in terms of t and constants. This solution is then invoked in the last equation (the upper
limit on the first integral on the RHS) in order to obtain an implicit equation for (t), which
is invoked in the second equation to yield an implicit equation for (t). The net result is
the motion of the system in terms of time t and the six constants (1 , 2 , 3 , 1 , 2 , 3 ). A
seventh constant, associated with an overall shift of the zero of t, arises due to the arbitrary
lower limits of the integrals.
In general, the separation of variables method begins with3
W (q, ) =

n
X

W (q , ) .

(5.132)

=1

Each W (q , ) may be regarded as a function of the single variable q , and is obtained by

satisfying an ODE of the form4


dW
= .
(5.133)
H q ,
dq
We then have
p =

W
q

W
+ ,1 t .

(5.134)

Note that while each W depends on only a single q , it may depend on several of the .

5.7.6

Example #2 : point charge plus electric field

Consider a potential of the form

k
Fz ,
(5.135)
r
which corresponds to a charge in the presence of an external point charge plus an external
electric field. This problem is amenable to separation in parabolic coordinates, (, , ):
p
p
x = cos , y = sin , z = 21 ( ) .
(5.136)
U (r) =

Here we assume complete separability. A given system may only be partially separable.
H
` (q , p ) may also depend on several of the . See e.g. eqn. 5.126, which is of the form
Hr r, r Wr , 3 = 1 .
4

136

CHAPTER 5. HAMILTONIAN MECHANICS

Note that
p

p
x2 + y 2 =
p
r = 2 + z 2 = 21 ( + ) .

The kinetic energy is

T = 12 m 2 + 2 2 + z 2
 2


2
1
= 8 m ( + )
+
+ 21 m 2 ,

(5.137)
(5.138)

(5.139)

and hence the Lagrangian is

2 2
L = 18 m ( + )
+

+ 12 m 2

2k
+ 1 F ( ) .
+ 2

(5.140)

Thus, the conjugate momenta are

p =

L
= 41 m ( + )

(5.141)

p =

L
= 14 m ( + )

(5.142)

p =

L
= m ,

(5.143)

and the Hamiltonian is

H = p + p + p
2
=
m

p2 + p2
+

(5.144)
+

p2
2k
+
1 F ( ) .
2m + 2

(5.145)

Notice that H/t = 0, which means dH/dt = 0, i.e. H = E 1 is a constant of the

motion. Also, is cyclic in H, so its conjugate momentum p is a constant of the motion.
We write
S(q, ) = W (q, ) Et

= W (, ) + W (, ) + W (, ) Et .

(5.146)
(5.147)

with E = 1 . Clearly we may take

W (, ) = P ,

(5.148)

137

5.7. HAMILTON-JACOBI THEORY

where P = 2 . Multiplying the Hamilton-Jacobi equation by 21 m ( + ) then gives



dW

2

P2
+ mk 14 F 2 12 mE
4


dW 2 P2 1 2 1
4 F + 2 mE ,

=
d
4

(5.149)

where = 3 is the third constant: = (E, P , ). Thus,

q
Z s
z }| {


P2
mk 1

S , , ; E, P , = d 12 mE +
mF

+
4

4 2
| {z }

Z s
P2

+ d 12 mE 41 mF 2

+ P Et .

5.7.7

(5.150)

Example #3 : charged particle in a magnetic field

The Hamiltonian is

e 2
1 
p A .
2m
c
and we write
We choose the gauge A = Bxy,

(5.151)

H=

S(x, y, P1 , P2 ) = Wx (x, P1 , P2 ) + Wy (y, P1 , P2 ) P1 t .

(5.152)

Note that here we will consider S to be a function of {q } and {P }.

The Hamilton-Jacobi equation is then


Wx
x

2

Wy
eBx

y
c

2

= 2mP1 .

(5.153)

We solve by writing
Wy = P2 y

dWx
dx

2

eBx
+ P2
c

2

= 2mP1 .

(5.154)

This equation suggests the substitution

x=
in which case

cP2
c p
2mP1 sin .
+
eB
eB

c p
x
=
2mP1 cos

eB

(5.155)

(5.156)

138

CHAPTER 5. HAMILTONIAN MECHANICS

and

Wx
1 Wx
Wx
eB
=

=
.
x
x
c 2mP1 cos
Substitution this into eqn. 5.154, we have
2mcP1
Wx
=
cos2 ,

eB

(5.157)

(5.158)

with solution
mcP1
mcP1
+
sin(2) .
eB
2eB

Wx
Wx x p
px =
=
= 2mP1 cos
x

Wx =

We then have

and

py =

Wy
= P2 .
y

(5.159)
(5.160)

(5.161)

The type-II generator we seek is then

mcP1
mcP1
+
sin(2) + P2 y P1 t ,
eB
2eB
where


cP2
eB
1
sin
.
x
=
eB
c 2mP1
Note that, from eqn. 5.155, we may write
c
1
mc
c p

dx =
2mP1 cos d ,
dP2 +
sin dP1 +
eB
eB 2mP1
eB
S(q, P, t) =

(5.162)

(5.163)

(5.164)

from which we derive

tan
=
P1
2P1

1
.
=
P2
2mP1 cos

(5.165)

These results are useful in the calculation of Q1 and Q2 :

Q1 =

S
P1

mcP1
mc
mcP1

mc
+
+
sin(2) +
cos(2)
t
eB
eB P1
2eB
eB
P1
mc
=
t
eB

(5.166)

and
Q2 =

S
P2


mcP1 
1 + cos(2)
eB
P2
p
c
=y
2mP1 cos .
eB

=y+

(5.167)

139

5.8. ACTION-ANGLE VARIABLES

Now since H(P,

Q) = 0, we have that Q = 0, which means that each Q is a constant. We
therefore have the following solution:
x(t) = x0 + A sin(c t + )

(5.168)

y(t) = y0 + A cos(c t + ) ,

(5.169)

where c = eB/mc is the cyclotron frequency, and

x0 =

5.8
5.8.1

cP2
eB

y 0 = Q2

c Q1

A=

c p
2mP1 .
eB

(5.170)

Action-Angle Variables
Circular Phase Orbits: Librations and Rotations

In a completely integrable system, the Hamilton-Jacobi equation may be solved by separation of variables. Each momentum p is a function of only its corresponding coordinate q
plus constants no other coordinates enter:
p =

W
= p (q , ) .
q

(5.171)

The motion satisfies

H (q , p ) = .

(5.172)

The level sets of H are curves C . In general, these curves each depend on all of the
constants , so we write C = C (). The curves C are the projections of the full motion
onto the (q , p ) plane. In general we will assume the motion, and hence the curves C ,
is bounded. In this case, two types of projected motion are possible: librations and rotations. Librations are periodic oscillations about an equilibrium position. Rotations involve
the advancement of an angular variable by 2 during a cycle. This is most conveniently
illustrated in the case of the simple pendulum, for which
H(p , ) =

p2
2I

+ 21 I 2 1 cos .

(5.173)

When E < I 2 , the momentum p vanishes at = cos1 (2E/I 2 ). The system

executes librations between these extreme values of the angle .
When E > I 2 , the kinetic energy is always positive, and the angle advances monotonically, executing rotations.
In a completely integrable system, each C is either a libration or a rotation5 . Both librations
and rotations are closed curves. Thus, each C is in general homotopic to (= can be
5

C may correspond to a separatrix, but this is a nongeneric state of affairs.

140

CHAPTER 5. HAMILTONIAN MECHANICS

Figure 5.2: Phase curves for the simple pendulum, showing librations (in blue), rotations
(in green), and the separatrix (in red). This phase flow is most correctly viewed as taking
place on a cylinder, obtained from the above sketch by identifying the lines = and
= .
continuously distorted to yield) a circle, S1 . For n freedoms, the motion is therefore
confined to an n-torus, Tn :
n times

Tn

}|
{
z
= S1 S1 S1 .

(5.174)

These are called invariant tori (or invariant manifolds). There are many such tori, as there
are many C curves in each of the n two-dimensional submanifolds.
Invariant tori never intersect! This is ruled out by the uniqueness of the solution to the
dynamical system, expressed as a set of coupled ordinary differential equations.
Note also that phase space is of dimension 2n, while the invariant tori are of dimension n.
Phase space is covered by the invariant tori, but it is in general difficult to conceive of how
this happens. Perhaps the most accessible analogy is the n = 1 case, where the 1-tori are
just circles. Two-dimensional phase space is covered noninteracting circular orbits. (The
orbits are topologically equivalent to circles, although geometrically they may be distorted.)
It is challenging to think about the n = 2 case, where a four-dimensional phase space is
filled by nonintersecting 2-tori.

5.8.2

Action-Angle Variables

For a completely integrable system, one can transform canonically from (q, p) to new coordinates (, J) which specify a particular n-torus Tn as well as the location on the torus,
which is specified by n angle variables. The {J } are momentum variables which specify
the torus itself; they are constants of the motion since the tori are invariant. They are

141

5.8. ACTION-ANGLE VARIABLES

called action variables. Since J = 0, we must have

H
J =
=0

H = H(J) .

(5.175)

The { } are the angle variables.

The coordinate describes the projected motion along C , and is normalized by
I
d = 2
(once around C ) .
(5.176)
C

The dynamics of the angle variables are given by

H
(J) .
=
J

(5.177)

Thus,
(t) = (0) + (J) t .
(5.178)


The (J) are frequencies describing the rate at which the C are traversed; T (J) =
2/ (J) is the period.

5.8.3

Canonical Transformation to Action-Angle Variables

The {J } determine the {C }; each q determines a point on C . This suggests a type-II

transformation, with generator F2 (q, J):
p =

F2
q

F2
.
J

(5.179)

Note that6
2 =

d =

which suggests the definition

 I
I
I 
2F2

F2
p dq ,
dq =
=
d
J
J q
J
C

1
J =
2

(5.180)

p dq .

(5.181)

I.e. J is (2)1 times the area enclosed by C .

If, separating variables,
W (q, ) =

W (q , )

(5.182)

` F2
2
2
In general, we should write d J
= J Fq
dq with a sum over . However, in eqn. 5.180 all

coordinates and momenta other than q and p are held fixed. Thus, = is the only term in the sum
which contributes.
6

142

CHAPTER 5. HAMILTONIAN MECHANICS

is Hamiltons characteristic function for the transformation (q, p) (Q, P ), then

I
W
1
dq = J ()
J =
2
q

(5.183)

is a function only of the { } and not the { }. We then invert this relation to obtain
(J), to finally obtain

X


F2 (q, J) = W q, (J) =
W q , (J) .

(5.184)

Thus, the recipe for canonically transforming to action-angle variable is as follows:

(1) Separate and solve the Hamilton-Jacobi equation for W (q, ) =
(2) Find the orbits C the level sets of satisfying H (q , p ) = .
H W
1
(3) Invert the relation J () = 2
q dq to obtain (J).

W (q , ).

(4) F2 (q, J) =

5.8.4

W q , (J) is the desired type-II generator7 .

Example : Harmonic Oscillator

The Hamiltonian is
H=

p2
+ 1 m02 q 2 ,
2m 2

(5.185)

+ m2 02 q 2 = 2m .

(5.186)

hence the Hamilton-Jacobi equation is


Thus,

2

q
dW
p=
= 2m m2 02 q 2 .
dq

We now define
q

2
m02

in which case
1
J=
2
7

dW
dq

1/2

sin

1 2

p dq =
2 0

p=

(5.187)

2m cos ,

(5.188)

Z2

.
d cos2 =
0

(5.189)

= 0, but rather to H
= H(J).

Note that F2 (q, J) is time-independent. I.e. we are not transforming to H

143

5.8. ACTION-ANGLE VARIABLES

Solving the HJE, we write

q dW
dW
=

= 2J cos2 .
d
dq

(5.190)

W = J + 12 J sin 2 ,

(5.191)

Integrating,
up to an irrelevant constant. We then have




W 
1
 .
=
= + 2 sin 2 + J 1 + cos 2
J q
J q
p
To find (/J)q , we differentiate q = 2J/m0 sin :
r

sin

1
2J

dq =
dJ +
tan .
cos d
=

m
J
2J
2m0 J
0
q

(5.192)

(5.193)

Plugging this result into eqn. 5.192, we obtain = . Thus, the full transformation is


p
2J 1/2
sin , p = 2m0 J cos .
q=
(5.194)
m0
The Hamiltonian is

H = 0 J ,
hence =

5.8.5

H
J

(5.195)

= 0 and J = H
= 0, with solution (t) = (0) + 0 t and J(t) = J(0).

Example : Particle in a Box

Consider a particle in an open box of dimensions Lx Ly moving under the influence of

gravity. The bottom of the box lies at z = 0. The Hamiltonian is
H=

p2y
p2
p2x
+
+ z + mgz .
2m 2m 2m

(5.196)

Step one is to solve the Hamilton-Jacobi equation via separation of variables. The HamiltonJacobi equation is written






1 Wx 2
1 Wy 2
1 Wz 2
+
+
+ mgz = E z .
(5.197)
2m x
2m y
2m z
We can solve for Wx,y by inspection:
p
Wx (x) = 2mx x
We then have8

Wy (y) =

2my y .

q


Wz (z) = 2m z x y mgz

3/2
2 2
Wz (z) =
.
z x y mgz
3 mg

Our choice of signs in taking the square roots for Wx , Wy , and Wz is discussed below.

(5.198)

(5.199)
(5.200)

144

CHAPTER 5. HAMILTONIAN MECHANICS

Figure 5.3: The librations Cz and Cx . Not shown is Cy , which is of the same shape as Cx .
p
Step two is to find the C . Clearly px,y = 2mx,y . For fixed px , the x motion proceeds
from x = 0 to x = Lx and back, with corresponding motion for y. For x, we have
q


pz (z) = Wz (z) = 2m z x y mgz ,
(5.201)

and thus Cz is a truncated parabola, with zmax = (z x y )/mg.

Step three is to compute J() and invert to obtain (J). We have
1
Jx =
2

1
px dx =

1
py dy =

(5.202)

ZLy
p
Ly p
dy 2my =
2my

(5.203)

Cx

1
Jy =
2

ZLx p
Lx p
dx 2mx =
2mx

Cy

and
1
Jz =
2

Cz

zZmax
q


1
dx 2m z x y mgz
pz dz =

3/2
2 2
=
z x y
.
3 m g

(5.204)

We now invert to obtain

2
2
2
J
,

=
J2
y
2mL2x x
2mL2y y


2
2
3 m g 2/3 2/3
2

J
+
J2 .
Jz +
z =
x
2
2 y
2mL
2mL
2 2
x
y

x =

(5.205)
(5.206)

145

5.8. ACTION-ANGLE VARIABLES

F2 x, y, z, Jx , Jy , Jz



x
y
2m2/3 g1/3 z 3/2
2/3
=
.
J +
J + Jz
Lx x Ly y
(3)2/3

We now find
x =

F2
x
=
Jx
Lx

and
F2
=
z =
Jz

2m2/3 g1/3 z
(3Jz )

where
zmax (Jz ) =
The momenta are
px =

Jx
F2
=
x
Lx

and

F2
pz =
= 2m
z

y =

2/3

F2
y
=
Jy
Ly

(5.208)

z
,
zmax

(3Jz )2/3
.
2m2/3 g1/3
,

py =

(5.207)

F2
Jy
=
y
Ly

!1/2


3 m g 2/3 2/3

Jz mgz
.
2 2

(5.209)

(5.210)

(5.211)

(5.212)

We note that the angle variables x,y,z seem to be restricted to the range [0, ], which seems
to be at odds with eqn. 5.180. Similarly, the momenta px,y,z all seem to be positive, whereas
we know the momenta reverse sign when the particle bounces off a wall. The origin of the
apparent discrepancy is that when we solved for the functions Wx,y,z , we had to take a
square root in
each case, and we chose a particular branch of the square root. So rather
than Wx (x) = 2mx x, we should have taken
(
2mx x
if px > 0
Wx (x) =
2mx (2Lx x) if px < 0 .

(5.213)

The relation Jx = (Lx /) 2mx is unchanged, hence

(
(x/Lx ) Jx
Wx (x) =
2Jx (x/Lx ) Jx
and
x =

x/Lx
(2Lx x)/Lx

if px > 0
if px < 0 .

if px > 0
if px < 0 .

(5.214)

(5.215)

Now the angle variable x advances by 2 during the cycle Cx . Similar considerations apply
to the y and z sectors.

146

CHAPTER 5. HAMILTONIAN MECHANICS

5.8.6

Kepler Problem in Action-Angle Variables

This is discussed in detail in standard texts, such as Goldstein. The potential is V (r) =
k/r, and the problem is separable. We write9
W (r, , ) = Wr (r) + W () + W () ,

(5.216)

hence






W 2
1 Wr 2
W 2
1
1
+
+
+ V (r) = E r .
2m r
2mr 2
2mr 2 sin2

(5.217)

Separating, we have


1 dW 2
=
2m d

J =

p
dW
= 2 2m .
d

(5.218)

Next we deal with the coordinate:




1 dW 2
=

2m d
sin2
Z0 q
p


J = 4 2m d 1 / csc2

= 2 2m

p


,

(5.219)

where 0 = sin1 ( / ). Finally, we have10



k
1 dWr 2
=E+ 2

2m dr
r
r
s 

I
k
Jr = dr 2m E + 2
r
r
Cr
s
2m
= (J + J ) + k
,
|E|

(5.220)

where weve assumed E < 0, i.e. bound motion.

Thus, we find
H =E=

2 2 mk2
Jr + J + J

Note that the frequencies are completely degenerate:

r,,
9
10

2 .

4 2 mk2
H
=
=

3 =
Jr,,
Jr + J + J

(5.221)

2 mk2
2|E|3

!1/2

We denote the azimuthal angle by to distinguish it from the AA variable .

The details of performing the integral around Cr are discussed in e.g. Goldstein.

(5.222)

147

5.8. ACTION-ANGLE VARIABLES

This threefold degeneracy may be removed by a transformation to new AA variables,

n
o
n
o
(r , Jr ), ( , J ), ( , J ) (1 , J1 ), (2 , J2 ), (3 , J3 ) ,
(5.223)

using the type-II generator

F2 (r , , ; J1 , J2 , J3 ) = ( ) J1 + ( r ) J2 + r J3 ,

(5.224)

which results in
F2
=
J1
F2
= r
2 =
J2
F2
= r
3 =
J3

F2
= J3 J2
r
F2
= J2 J1
J =

F2
J =
= J1 .

1 =

Jr =

The new Hamiltonian is

H(J1 , J2 , J3 ) =

2 2 mk2
,
J32

(5.225)
(5.226)
(5.227)

(5.228)

whence 1 = 2 = 0 and 3 = .

5.8.7

Charged Particle in a Magnetic Field

For the case of the charged particle in a magnetic field, studied above in section 5.7.7, we
found
c p
cP2
2mP1 sin
(5.229)
+
x=
eB
eB
and
p
px = 2mP1 cos
,
py = P2 .
(5.230)
The action variable J is then
J=

Z2
2mcP1
mcP1
px dx =
.
d cos2 =
eB
eB

(5.231)

W = J + 21 J sin(2) + P y ,

(5.232)

We then have
where P P2 . Thus,
=

W
J



= + 12 sin(2) + J 1 + cos(2)
J 

tan
2
1
= + 2 sin(2) + 2J cos
2J
=.

(5.233)

148

CHAPTER 5. HAMILTONIAN MECHANICS

The other canonical pair is (Q, P ), where

W
Q=
=y
P

2cJ
cos .
eB

(5.234)

Therefore, we have
cP
x=
+
eB

and
px =

2cJ
sin
eB
r

y =Q+

2eBJ
cos
c

2cJ
cos
eB

py = P .

(5.235)

(5.236)

The Hamiltonian is
H=
=



p2x
1
eBx 2
+
py
2m 2m
c
eBJ
eBJ
cos2 +
sin2
mc
mc

= c J ,

(5.237)

where c = eB/mc. The equations of motion are

H
=
= c
J
and

H
Q =
=0
P

H
J =
=0

(5.238)

H
P =
=0.
Q

(5.239)

Thus, Q, P , and J are constants, and (t) = 0 + c t.

5.8.8

Motion on Invariant Tori

The angle variables evolve as

(t) = (J) t + (0) .

(5.240)

Thus, they wind around the invariant torus, specified by {J } at constant rates. In general,
while each executed periodic motion around a circle, the motion of the system as a whole
is not periodic, since the frequencies (J) are not, in general, commensurate. In order for
the motion to be periodic, there must exist a set of integers, {l }, such that
n
X

=1

l (J) = 0 .

(5.241)

149

5.9. CANONICAL PERTURBATION THEORY

This means that the ratio of any two frequencies / must be a rational number. On a
given torus, there are several possible orbits, depending on initial conditions (0). However,
since the frequencies are determined by the action variables, which specify the tori, on a
given torus either all orbits are periodic, or none are.
In terms of the original coordinates q, there are two possibilities:
q (t) =

1 =

or

n =

A ei(t)

()

A1 2 n ei1 1 (t) ein n (t)

(libration)

(5.242)

q (t) = q (t) +

B ei(t)

(rotation) .

(5.243)

For rotations, the variable q (t) increased by q = 2 q . R

5.9
5.9.1

Canonical Perturbation Theory

Canonical Transformations and Perturbation Theory

Suppose we have a Hamiltonian

H(, t) = H0 (, t) + H1 (, t) ,

(5.244)

where is a small dimensionless parameter. Lets implement a type-II transformation,

generated by S(q, P, t):11

S(q, P, t) .
H(Q,
P, t) = H(q, p, t) +
t
Lets expand everything in powers of :

(5.245)

q = Q + q1, + 2 q2, + . . .

(5.246)

p = P + p1, + 2 p2, + . . .

(5.247)

=H
+ H
+ 2 H
+ ...
H
0
1
2

(5.248)

S = q|{zP} + S1 + 2 S2 + . . . .

(5.249)

identity
transformation

Then
Q =

11

S
S1
S2
= q +
+ 2
+ ...
P
P
P




S2 2
S1
+ q2, +
+ ...
= Q + q1, +
P
P

Here, S(q, P, t) is not meant to signify Hamiltons principal function.

(5.250)

150

CHAPTER 5. HAMILTONIAN MECHANICS

and
p =

S1
S2
S
= P +
+ 2
+ ...
q
q
q
= P + p1, + 2 p2, + . . . .

(5.251)
(5.252)

We therefore conclude, order by order in ,

qk, =

Sk
P

pk, = +

Sk
.
q

(5.253)

Now lets expand the Hamiltonian:

S

H(Q,
P, t) = H0 (q, p, t) + H1 (q, p, t) +
t
H0
H0
= H0 (Q, P, t) +
(q Q ) +
(p P )
Q
P

+ H1 (Q, P, t) + S1 (Q, P, t) + O(2 )

t

(5.254)

!
H0 S1
S1
H0 S1
+ H1 + O(2 )
+
+
= H0 (Q, P, t) +
Q P
P Q
t



S1
+ O(2 ) .
(5.255)
= H0 (Q, P, t) + H1 + S1 , H0 +
t

In the above expression, we evaluate Hk (q, p, t) and Sk (q, P, t) at q = Q and p = P and

expand in the differences q Q and p P . Thus, we have derived the relation

(Q, P, t) + H
(Q, P, t) + . . .
H(Q,
P, t) = H
0
1

(5.256)

(Q, P, t) = H (Q, P, t)
H
0
0


(Q, P, t) = H + S , H + S1 .
H
1
1
1
0
t

(5.257)

with

(5.258)

The problem, though, is this: we have one equation, eqn, 5.258, for the two unknowns H
1
= 0, which
and S1 . Thus, the problem is underdetermined. Of course, we could choose H
1
basically recapitulates standard Hamilton-Jacobi theory. But we might just as well demand
satisfy some other requirement, such as that H
+ H
being integrable.
that H
1
0
1
Incidentally, this treatment is paralleled by one in quantum mechanics, where a unitary
transformation may be implemented to eliminate a perturbation to lowest order in a small
parameter. Consider the Schr
odinger equation,
i~

= (H0 + H1 ) ,
t

(5.259)

5.9. CANONICAL PERTURBATION THEORY

151

and define by
eiS/~ ,

(5.260)

S = S 1 + 2 S 2 + . . . .

(5.261)

with
As before, the transformation U exp(iS/~) collapses to the identity in the 0 limit.
Now lets write the Schr
odinger equation for . Expanding in powers of , one finds


 S1
1

+ ... H
= H0 + H1 +
S ,H +
(5.262)
i~
t
i~ 1 0
t
where [A, B] = AB BA is the commutator. Note the classical-quantum correspondence,
{A, B}

1
[A, B] .
i~

(5.263)

Again, what should we choose for S1 ? Usually the choice is made to make the O() term
vanish. But this is not the only possible simplifying choice.
in H

5.9.2

Canonical Perturbation Theory for n = 1 Systems

Henceforth we shall assume H(, t) = H() is time-independent, and we write the perturbed
Hamiltonian as
H() = H0 () + H1 () .
(5.264)
Let (0 , J0 ) be the action-angle variables for H0 . Then

We define

( , J ) = H q( , J ), p( , J ) = H
(J ) .
H
0 0 0
0
0 0
0 0
0 0


( , J ) = H q( , J ), p( , J ) .
H
0 0
0 0
1 0 0
1

(5.265)
(5.266)

= H
+ H
is integrable12 , so it, too, possesses action-angle variWe assume that H
0
1
ables, which we denote by (, J)13 . Thus, there must be a canonical transformation taking
(0 , J0 ) (, J), with


(, J), J (, J) K(J) = E(J) .
H
(5.267)
0
0
We solve via a type-II canonical transformation:

S(0 , J) = 0 J + S1 (0 , J) + 2 S2 (0 , J) + . . . ,

(5.268)

where 0 J is the identity transformation. Then

S
S1
S2
=J +
+ 2
+ ...
0
0
0
S2
S1
S
= 0 +
+ 2
+ ... ,
=
J
J
J

J0 =

12
13

This is always true, in fact, for n = 1.

We assume the motion is bounded, so action-angle variables may be used.

(5.269)
(5.270)

152

CHAPTER 5. HAMILTONIAN MECHANICS

and
E(J) = E0 (J) + E1 (J) + 2 E2 (J) + . . .
( , J ) + H
( , J ) .
=H
0

(5.271)
(5.272)

We now expand H(
0 , J0 ) in powers of J0 J:

H(
0 , J0 ) = H0 (0 , J0 ) + H1 (0 , J0 )

(J) + H0 (J J) +
=H
0
0
J

(5.273)

2H

0
(J J)2 + . . .
J 2 0

( , J ) + H1 (J J) + . . .
+ H
1 0 0
0
J

S1 
H
0

(5.274)
= H0 (J) + H1 (0 , J0 ) +
J 0
!
S2 1 2H
S1
 S1 2 H
H
0
1
0
+
+
+
2 + . . . .
J 0
2 J 2 0
J 0
1
2

Equating terms, then,

(J)
E0 (J) = H
0

(5.275)

( , J) + H0 S1
E1 (J) = H
1 0
J 0
S1
 S1 2 H
S2 1 2H
H
1
0
0
+
E2 (J) =
+
.
J 0 2 J 2 0
J 0

(5.276)
(5.277)

How, one might ask, can we be sure that the LHS of each equation in the above hierarchy
depends only on J when each RHS seems to depend on 0 as well? The answer is that we
use the freedom to choose each Sk to make this so. We demand each RHS be independent
of 0 , which means it must be equal to its average, h RHS(0 ) i, where

f 0

Z2


d0
f 0 .
=
2

(5.278)

The average is performed at fixed J and not at fixed J0 . In this regard, we note that holding
J constant and increasing 0 by 2 also returns us to the same starting point. Therefore,
J is a periodic function of 0 . We must then be able to write
Sk (0 , J) =

Sk (J; m) eim0

(5.279)

m=

for each k > 0, in which case



Sk
0


1 
Sk (2) Sk (0) = 0 .
2

(5.280)

153

5.9. CANONICAL PERTURBATION THEORY

(J) is
Lets see how this averaging works to the first two orders of the hierarchy. Since H
0
independent of 0 and since S1 /0 is periodic, we have
this vanishes!


( , J) + H0
E1 (J) = H
1 0
J

and hence S1 must satisfy

z }| {

S1
0

(5.281)

H
S1
1
1
,
=
0
0 (J)

(5.282)

/J. Clearly the RHS of eqn. 5.282 has zero average, and must be a
where 0 (J) = H
0
periodic function of 0 . The solution is S1 = S1 (0 , J) + g(J), where g(J) is an arbitrary
function of J. However, g(J) affects only the difference 0 , changing it by a constant
value g (J). So there is no harm in taking g(J) = 0.
Next, lets go to second order in . We have
this vanishes!

z }| {

 



S1 2
S2
H1 S1
1 0
+2
+ 0 (J)
.
E2 (J) =
J 0
J
1
0
The equation for S2 is then
(


 H

  H
1
H
S2
H
1
1
1
1

= 2
H0
H0
H1
H1 +
0
J
J
J
J
0 (J)

)


1 ln 0
2 

H
12
+
H1 2 H
+2 H
.
1
1
2 J

(5.283)

(5.284)

The expansion for the energy E(J) is then

(J) + H

E(J) = H
0
1

2
+
0 (J)

(

H
1
J



H
1



H1
H1
J

)
1 ln 0 
2
2 
+
+ O(3 ) .
H1 H1
2 J

(5.285)

Note that we dont need S to find E(J)! The perturbed frequencies are
(J) =

E
.
J

(5.286)

Sometimes the frequencies are all that is desired. However, we can of course obtain the full
motion of the system via the succession of canonical transformations,
(, J) (0 , J0 ) (q, p) .

(5.287)

154

CHAPTER 5. HAMILTONIAN MECHANICS

Figure 5.4: Action-angle variables for the harmonic oscillator.

5.9.3

Example : Nonlinear Oscillator

Consider the nonlinear oscillator with Hamiltonian

H0

z
}|
{
p2
2
H(q, p) =
+ 1 m q 2 + 14 q 4 .
2m 2 0

(5.288)

The action-angle variables for the harmonic oscillator Hamiltonian H0 are

0 = tan1 mvq/p)

J0 =

p2
+ 1 m q 2 ,
2m0 2 0

(5.289)

and the relation between (0 , J0 ) and (q, p) is further depicted in fig. 5.4. Note H0 = 0 J0 .
For the full Hamiltonian, we have
r
4
2J0
1

H(0 , J0 ) = 0 J0 + 4
sin 0
m0

= 0 J0 + 2 2 J02 sin4 0 .
(5.290)
m 0
We may now evaluate

E1 (J) = H
1

J 2
= 2 2
m 0

Z2
3J 2
d0
sin4 0 =
.
2
8m2 02

(5.291)

The frequency, to order , is

(J) = 0 +

3 J
.
4m2 02

(5.292)

Now to lowest order in , we may replace J by J0 = 12 m0 A2 , where A is the amplitude of

the q motion. Thus,
3
.
(5.293)
(A) = 0 +
8m0

5.9. CANONICAL PERTURBATION THEORY

155

This result agrees with that obtained via heavier lifting, using the Poincare-Lindstedt
method.
Next, lets evaluate the canonical transformation (0 , J0 ) (, J). We have

S1
J 2 
0

= 2 2 83 sin4 0
0
m 0


J 2
3 + 2 sin2 0 sin 0 cos 0 + O(2 ) .
S(0 , J) = 0 J +
3
2
8m 0

(5.294)

Thus,

S
J
= 0 +
3 + 2 sin2 0 sin 0 cos 0 + O(2 )
3
2
J
4m 0


J 2
S
=J+
J0 =
+ O(2 ) .
4
cos
2

cos
4
0
0
23
8m
0
0
=

(5.295)
(5.296)

Again, to lowest order, we may replace J by J0 in the above, whence

J02
4
cos
2

cos
4
+ O(2 )
0
0
3
8m2 0


J0
= 0 +
3 + 2 sin2 0 sin 20 + O(2 ) .
3
2
8m 0

J = J0

(5.297)
(5.298)

To obtain (q, p) in terms of (, J) is not analytically tractable the relations cannot be

analytically inverted.

5.9.4

n > 1 Systems : Degeneracies and Resonances

Generalizing the procedure we derived for n = 1, we obtain

S
S2
S1
= J + + 2
+ ...

0
0
0
S1
S2
S
= 0 +
+ 2
+ ...
=

J
J
J

J0 =

(5.299)
(5.300)

and
(J)
E0 (J) = H
0

(5.301)

( , J) + (J) S1
E1 (J) = H
0
0
0
0
S2
1 0 S1 S1
H
S1
0
+
.
E2 (J) =
+ 0

J 0
2 J 0 0
0

(5.302)
(5.303)

We now implement the averaging procedure, with


f (J , . . . , J ) =
1

Z2 1
Z2 n


d0
d0

f 10 , . . . , n0 , J 1 , . . . , J n .
2
2
0

(5.304)

156

CHAPTER 5. HAMILTONIAN MECHANICS

The equation for S1 is

0


S1

H
= H
V ei ,
1
1

(5.305)

where = {1 , 2 , . . . , n }, with each an integer, and with 6= 0. The solution is

S1 (0 , J) = i

X V
ei .
0

(5.306)

where 0 = l 0 . When two or more of the frequencies (J) are commensurate, there
exists a set of integers l such that the denominator of D(l) vanishes. But even when the

frequencies are not rationally related, one can approximate the ratios 0 /0 by rational
numbers, and for large enough l the denominator can become arbitrarily small.
Periodic time-dependent perturbations present a similar problem. Consider the system
H(, J, t) = H0 (J) + V (, J, t) ,
where V (t + T ) = V (t). This means we may write
X
V (, J, t) =
Vk (, J) eikt

(5.307)

(5.308)

XX
k

Vk,(J) ei eikt .

(5.309)

by Fourier transforming from both time and angle variables; here = 2/T . Note that
=V
V (, J, t) is real if Vk,
k,l . The equations of motion are
H
J = = i

l Vk,(J) ei eikt

(5.310)

k,

H
= + = 0 (J) +
J

X Vk,(J)
J

k,

ei eikt .

(5.311)

We now expand in :
= 0 + 1 + 2 2 + . . .

J =

J0

J1

J2

+ ... .

To order 0 , J = J0 and 0 = 0 t + 0 . To order 1 ,

X
J1 = i
l Vk,(J0 ) ei(0 k)t ei0

(5.312)
(5.313)

(5.314)

k,l

and

X Vk,(J)

ei(0 k)t ei0 ,
1 = 0 J1 +
J
J
k,

(5.315)

157

5.9. CANONICAL PERTURBATION THEORY

where derivatives are evaluated at J = J0 . The solution is:

X l Vk,(J0 )
ei(0 k)t ei0
k 0
k,
)
(
Vk,(J)
1
0 l Vk,(J0 )

+
ei(0 k)t ei0 .
1 =
J (k 0 )2
J k 0

J1 =

(5.316)

(5.317)

When the resonance condition,

k = 0 (J0 ) ,

(5.318)

holds, the denominators vanish, and the perturbation theory breaks down.

5.9.5

Particle-Wave Interaction

Consider a particle of charge e moving in the presence of a constant magnetic field B = B z

and a space- and time-varying electric field E(x, t), described by the Hamiltonian

2
1
(5.319)
H=
p ec A + eV0 cos(k x + kz z t) ,
2m
from our
where is a dimensionless expansion parameter. Working in the gauge A = Bxy,
earlier discussions in section 5.7.7, we may write
r


2J
p2z
k P
H = c J +
+ eV0 cos kz z +
+ k
sin t .
(5.320)
2m
mc
mc
Here,

r
r
P
2J
2J
+
sin
,
y =Q+
cos ,
(5.321)
x=
mc
mc
mc
with c = eB/mc, the cyclotron frequency. We now make a mixed canonical transformation,
generated by


k P
t K P Q ,
(5.322)
F = J + kz z +
mc


where the new sets of conjugate variables are ( , J ) , (Q , P ) , ( , K ) . We then have
=

F
=
J

Q=
=

k K
F
=
+ Q
P
mc

F
k P
= kz z +
t

K
mc

J=

F
= J

F
=P
Q

(5.324)

F
= kz K .
z

(5.325)

P =
pz =

(5.323)

The transformed Hamiltonian is

F
H = H +
t
r


2J
kz2 2

K K + eV0 cos + k
sin .
= c J +
2m
mc

(5.326)

158

CHAPTER 5. HAMILTONIAN MECHANICS

We will now drop primes and simply write H = H0 + H1 , with

kz2 2
K K
2m
r


2J
H1 = eV0 cos + k
sin .
mc
H0 = c J +

(5.327)
(5.328)

When = 0, the frequencies associated with the and motion are

0 =

H0
= c

0 =

H0
k2 K
= z = kz vz ,

(5.329)

where vz = pz /m is the z-component of the particles velocity. Now let us solve eqn. 5.305:
0

S1
S1
+ 0
= h H1 i H1 .

(5.330)

This yields
S1
+
c

r



kz2 K
S1
2J

= eA0 cos + k
sin
m

mc
!
r

X
2J
Jn k
= eA0
cos( + n) ,
mc
n=

(5.331)

where we have used the result

eiz sin =

Jn (z) ein .

(5.332)

n=

The solution for S1 is

S1 =

X
n

eV0
Jn k

nc kz2 K/m

2J
mc

sin( + n) .

(5.333)

where
We then have new action variables J and K,
S1
+ O(2 )
J = J +

+ S1 + O(2 ) .
K=K

(5.334)
(5.335)

Defining the dimensionless variable

k

2J
,
mc

(5.336)

we obtain the result





X nJn () cos( + n)
mc2
mc2
2

+ O(2 ) ,
=
2
2
2
kz2 K

2eV0 k
2eV0 k
n
n
c

mc

(5.337)

5.9. CANONICAL PERTURBATION THEORY

159

Figure 5.5: Plot of versus for = 0 (Poincare section) for = 30.11 c Top panels are
nonresonant invariant curves calculated to first order. Bottom panels are exact numerical
dynamics, with x symbols marking the initial conditions. Left panels: weak amplitude
(no trapping). Right panels: stronger amplitude (shows trapping). From Lichtenberg and
Lieberman (1983).
=k
where

2J/mc .14

We see that resonances occur whenever

k2 K
z =n,
c mc

(5.338)

for any integer n. Let us consider the case kz = 0, in which the resonance condition is
= nc . We then have
2
2 X n Jn () cos( + n)

2
2

c
n
where
=

E0 ck

B c

(5.339)

(5.340)

14
This arises because we are computing the
Note that the argument of Jn in eqn. 5.337 is and not .
new action J in terms of the old variables (, J) and (, K).

160

CHAPTER 5. HAMILTONIAN MECHANICS

is a dimensionless measure of the strength of the perturbation, with E0 k V0 . In Fig. 5.5

we plot the level sets for the RHS of the above equation () for = 0, for two different
values of the dimensionless amplitude , for /c = 30.11 (i.e. off resonance). Thus, when
the amplitude is small, the level sets are far from a primary resonance, and the analytical and
numerical results are very similar (left panels). When the amplitude is larger, resonances
may occur which are not found in the lowest order perturbation treatment. However, as
is apparent from the plots, the gross features of the phase diagram are reproduced by
perturbation theory. What is missing is the existence of chaotic islands which initially
emerge in the vicinity of the trapping regions.

5.10

Adiabatic Invariants

Adiabatic perturbations are slow, smooth, time-dependent perturbations to a dynamical

system. A classic example: a pendulum with a slowly
varying length l(t). Suppose (t)
is the adiabatic parameter. We write H = H q, p; (t) . All explicit time-dependence to
H comes through (t). Typically, a dimensionless parameter may be associated with the
perturbation:


1  d ln 
,
(5.341)
=
0  dt 

where 0 is the natural frequency of the system when is constant. We require 1 for
adiabaticity. In adiabatic processes, the action variables are conserved to a high degree of
accuracy. These are the adiabatic invariants. For example, for the harmonix oscillator, the
action is J = E/. While E and may vary considerably during the adiabatic process,
their ratio is very nearly fixed. As a consequence, assuming small oscillations,
E = J = 21 mgl 02

0 (l)

2J
3/2 ,
m gl

(5.342)

so 0 () l3/4 .
Suppose that for fixed the Hamiltonian is transformed to action-angle variables via the
generator S(q, J; ). The transformed Hamiltonian is
S d

,
H(,
J, t) = H(, J; ) +
dt

(5.343)

H(, J; ) = H q(, J; ), p(, J; ); ) .

(5.344)

where
We assume n = 1 here. Hamiltons equations are now

H
2S d
= +
= (J; ) +
J
J dt
2S d

H
=
.
J =

dt

(5.345)
(5.346)

161

5.10. ADIABATIC INVARIANTS

Figure 5.6: A mechanical mirror.

The second of these may be Fourier decomposed as
J = i

Sm (J; ) im
e
,

(5.347)

hence
J = J(t = +) J(t = ) = i

X
m

Z
Sm (J; ) d im
m dt
e
.

dt

(5.348)

Since is small, we have (t) = t + , to lowest order. We must therefore evaluate

integrals such as
Z
Sm (J; ) d imt
e
.
(5.349)
Im = dt

dt

The term in curly brackets is a smooth, slowly varying function of t. Call it f (t). We
presume f (t) can be analytically continued off the real t axis, and that its closest singularity
in the complex t plane lies at t = i , in which case I behaves as exp(|m| ). Consider,
for example, the Lorentzian,
C
f (t) =
1 + (t/ )2

Z
dt f (t) eimt = e|m| ,

(5.350)

which is exponentially small in the time scale . Because of this, only m = 1 need be
considered. What this tells us is that the change J may be made arbitrarily small by a
sufficiently slowly varying (t).

5.10.1

Example: mechanical mirror

Consider a two-dimensional version of a mechanical mirror, depicted in fig. 5.6. A particle

bounces between two curves, y = D(x), where |D (x)| << 1. The bounce time is b =
2D/vy . We assume L/vx , where vx,y are the components of the particles velocity, and
L is the total length of the system. There are, therefore, many bounces, which means the
particle gets to sample the curvature in D(x).

162

CHAPTER 5. HAMILTONIAN MECHANICS

The adiabatic invariant is the action,

1
J=
2

ZD
ZD
1
2
dy m vy +
dy m (vy ) = mvy D(x) .
2

(5.351)

Thus,
E = 12 m vx2 + vy2 ) = 21 mvx2 +
or

2 J 2
,
8mD2 (x)
2
.

(5.352)


J
2E

(5.353)
=
m
2mD(x)
The particle is reflected in the throat of the device at horizontal coordinate x , where
vx2

D(x ) =

5.10.2

J
.
8mE

(5.354)

Example: magnetic mirror

Consider a particle of charge e moving in the presence of a uniform magnetic field B = B z.

Recall the basic physics: velocity in the parallel direction vz is conserved, while in the plane
perpendicular to B the particle executes circular cyclotron orbits, satisfying
2
mv
e
= v B

mcv
,
eB

(5.355)

where is the radial coordinate in the plane perpendicular to B. The period of the orbits
is T = 2.v = 2mc/eB, hence their frequency is the cyclotron frequency c = eB/mc.
Now assume that the magnetic field is spatially dependent. Note that a spatially varying
B-field cannot be unidirectional:
Bz
=0.
(5.356)
B = B +
z
The non-collinear nature of B results in the drift of the cyclotron orbits. Nevertheless, if
the field B felt by the particle varies slowly on the time scale T = 2/c , then the system
possesses an adiabatic invariant:
I
I


1
1
J=
(5.357)
p d =
mv + ec A d
2
2
C
C
I
I
m
e
d .
=
v d +
Bn
(5.358)
2
2c
C

int(C)

The last two terms are of opposite sign, and one has
e
m eBz

2 +
Bz 2
2 mc
2c
2c
m2 v
e
eBz 2
=
B (C) =
,
=
2c
2c
2eBz

J =

(5.359)
(5.360)

163

5.10. ADIABATIC INVARIANTS

Figure 5.7: B field lines in a magnetic bottle.

where B (C) is the magnetic flux enclosed by C.
The energy is
2
E = 12 mv
+ 12 mvz2 ,

hence we have
vz =
where
M

2
E MB .
m

e2
e
J=
B (C)
mc
2mc2

(5.361)

(5.362)

(5.363)

is the magnetic moment. Note that vz vanishes when B = Bmax = E/M . When this limit
is reached, the particle turns around. This is a magnetic mirror . A pair of magnetic mirrors
may be used to confine charged particles in a magnetic bottle, depicted in fig. 5.7.
Let vk,0 , v,0 , and Bk,0 be the longitudinal particle velocity, transverse particle velocity,
and longitudinal component of the magnetic field, respectively, at the point of injection.
Our two conservation laws (J and E) guarantee
2
2
2
(z) = vk,0
+ v,0
vk2 (z) + v
2
v,0
v (z)2
=
.
Bk (z)
Bk,0

(5.364)
(5.365)

This leads to reflection at a longitudinal coordinate z , where

v
u
2
u
vk,0

Bk (z ) = Bk,0 t1 + 2 .
v,0
The physics is quite similar to that of the mechanical mirror.

(5.366)

164

5.10.3

CHAPTER 5. HAMILTONIAN MECHANICS

Resonances

When n > 1, we have

J = i
J = i

X
m

Sm (J; ) im
e

Z
Sm (J; ) d imt im
dt
m
e
e
.

dt

(5.367)

(5.368)

Therefore, when m (J) = 0 we have a resonance, and the integral grows linearly with
time a violation of the adiabatic invariance of J .

5.11

Appendix : Canonical Perturbation Theory

Consider the Hamiltonian

H=

q3
p2
+ 21 m 02 q 2 + 13 m 02
,
2m
a

where is a small dimensionless parameter.

(a) Show that the oscillation frequency satisfies (J) = 0 + O(2 ). That is, show that the
first order (in ) frequency shift vanishes.
Solution: It is good to recall the basic formulae
r
p
2J0
q=
sin 0
,
p = 2m 0 J0 cos 0
m0

(5.369)

as well as the results

J0 =
=

S
S2
S1
=J +
+ 2
+ ...
0
0
0

(5.370)

S
S2
S1
= 0 +
+ 2
+ ... ,
J
J
J

(5.371)

and
(J)
E0 (J) = H
0

( , J) + H0 S1
E1 (J) = H
1 0
J 0
 S1 2 H
S1
S2 1 2H
H
0
1
0
+
+
.
E2 (J) =
J 0 2 J 2 0
J 0

(5.372)
(5.373)
(5.374)

5.11. APPENDIX : CANONICAL PERTURBATION THEORY

165

Expressed in action-angle variables,

( , J) = J
H
0 0
0
r
2 20 3/2

H1 (0 , J) =
J
sin3 0 .
3 ma2
Thus, 0 =

H
0
J

(5.375)
(5.376)

= 0 .

Averaging the equation for E1 (J) yields


( , J) = 2
E1 (J) = H
1 0
3

20 3/2
3 
J
sin 0 = 0 .
ma2

(5.377)

(b) Compute the frequency shift (J) to second order in .

Solution : From the equation for E1 , we also obtain
S1
1 
 
H1 H1 .
=
0
0

Inserting this into the equation for E2 (J) and averaging then yields
 





1 H
1 H
1
1

H1 H1
H1
=
E2 (J) =
0 J
0
J
40 J 2
6 
=
sin 0
3ma2

(5.378)

(5.379)
(5.380)

In computing the average of sin6 0 , it is good to recall the binomial theorem, or the Fibonacci tree. The sixth order coefficents are easily found to be {1, 6, 15, 20, 15, 6, 1}, whence
sin6 0 =
=

Thus,

6
1
ei0 ei0
6
(2i)
1
64

(5.381)

2 sin 60 + 12 sin 40 30 sin 20 + 20 .

whence


sin6 0 =

E(J) = 0 J
and
(J) =

5
16

5 2
12

(5.382)
(5.383)

J2
ma2

E
J
= 0 65 2
.
J
ma2

(5.384)
(5.385)

(c) Find q(t) to order . Your result should be finite for all times.
Solution : From the equation for E1 (J), we have
s
2
2J 3
S1
=
sin3 0 .
0
3 m0 a2

(5.386)

166

CHAPTER 5. HAMILTONIAN MECHANICS

Integrating, we obtain
2
S1 (0 , J) =
3

2J 3
cos 0
m0 a2

J 3/2
cos 0
2m0 a2

Thus, with

1
3
1
9

cos3 0

cos 30 .

S(0 , J) = 0 J + S1 (0 , J) + . . . ,

(5.387)
(5.388)

(5.389)

we have


S
3 J 1/2
1
cos

cos
3
= 0 +
0
0
9
J
2 2m0 a2


S
J 3/2
1
J0 =
sin

sin
3
.
=J
0
0
3
0
2m0 a2
=

(5.390)
(5.391)

Inverting, we may write 0 and J0 in terms of and J:

0 = +

3 J 1/2

2 2m0 a2

J0 = J +

J 3/2
2m0 a2

1
9
1
3

cos 3 cos

sin 3 sin .

(5.392)
(5.393)

Thus,
q(t) =
=
=
with

2J0
sin 0
m0

(5.394)




J
2J
+ ...
sin + cos + . . .
sin 1 +
m0
2J

(5.395)

J
2J
sin
1 + 13 cos 2 + O 2 ,
m0
m0 a
(t) = (0) + (J) t .

(5.396)

(5.397)

Chapter 6

Ergodicity and the Approach to

Equilibrium
6.1
6.1.1

Modeling the Approach to Equilibrium

Equilibrium

A thermodynamic system typically consists of an enormously large number of constituent

particles, a typical large number being Avogadros number, NA = 6.021023 . Nevertheless,
in equilibrium, such a system is characterized by a relatively small number of thermodynamic
state variables. Thus,
while a complete description of a (classical) system would require us
to account for O 1023 evolving degrees of freedom, with respect to the physical quantities
in which we are interested, the details of the initial conditions are effectively forgotten over
some microscopic time scale , called the collision time, and over some microscopic distance
scale, , called the mean free path1 . The equilibrium state is time-independent.

6.1.2

The Master Equation

Relaxation to equilibrium is often modeled with something called the master equation. Let
Pi (t) be the probability that the system is in a quantum or classical state i at time t. Then
write


dPi X
Wij Pj Wji Pi .
=
(6.1)
dt
j

Here, Wij is the rate at which j makes a transition to i. Note that we can write this equation
as
X
dPi
=
ij Pj ,
(6.2)
dt
j

1
Exceptions involve quantities which are conserved by collisions, such as overall particle number, momentum, and energy. These quantities relax to equilibrium in a special way called hydrodynamics.

167

168

CHAPTER 6. ERGODICITY AND THE APPROACH TO EQUILIBRIUM

where

(
W
ij = P ij
k Wkj

if i 6= j
if i = j ,

(6.3)

where the prime on the sum indicates that k = j is to be excluded. The constraints on the
Wij are that Wij 0 for all i, j, and we may take Wii 0 (no sum on i). Fermis Golden
Rule of quantum mechanics says that
Wij =

2
2 
h i | V | j i (Ej ) ,
~

(6.4)

 
 
 i = E  i , V is an additional potential which leads to transitions, and (E )
where H
0
i
i
is the density of final states at energy Ei . The fact that Wij 0 means that if each
Pi (t = 0) 0, then Pi (t) 0 for all t 0. To see this, suppose that at some time t > 0 one
of the probabilities
PPi is crossing zero and about to become negative. But then eqn. 6.1

says that Pi (t) = j Wij Pj (t) 0. So Pi (t) can never become negative.

6.1.3

Equilibrium distribution and detailed balance

If the transition rates Wij are themselves time-independent, then we may formally write
Pi (t) = e t

ij

Pj (0) .

(6.5)

Here we have used the Einstein summation convention in which repeated indices are
summed over (in this case, the j index). Note that
X

ij = 0 ,

(6.6)

which says that the total probability

i Pi

is conserved:

X
X X 
d X
ij = 0 .
Pi =
ij Pj =
Pj
dt
i

i,j

(6.7)

~ = (1, 1, . . . , 1) is a left eigenvector of with eigenvalue = 0. The

We conclude that
corresponding right eigenvector, which we write as Pieq , satisfies ij Pjeq = 0, and is a
stationary (i.e. time independent) solution to the master equation. Generally, there is only
one right/left eigenvector pair corresponding to = 0, in which case any initial probability
distribution Pi (0) converges to Pieq as t , as shown in Appendix I (6.6).
In equilibrium, the net rate of transitions into a state | i i is equal to the rate of transitions
out of | i i. If, for each state | j i the transition rate from | i i to | j i is equal to the transition
rate from | j i to | i i, we say that the rates satisfy the condition of detailed balance. In other
words,
Wij Pjeq = Wji Pieq .
(6.8)

6.1. MODELING THE APPROACH TO EQUILIBRIUM

169

Assuming Wij 6= 0 and Pjeq 6= 0, we can divide to obtain

Pjeq
Wji
= eq .
Wij
Pi

(6.9)

Note that detailed balance is a stronger condition than that required for a stationary solution
to the master equation.
If = t is symmetric, then the right eigenvectors and left eigenvectors are transposes of
each other, hence P eq = 1/N , where N is the dimension of . The system then satisfies
the conditions of detailed balance. See Appendix II (6.7) for an example of this formalism
applied to a model of radioactive decay.

6.1.4

Boltzmanns H-theorem

Suppose for the moment that is a symmetric matrix, i.e. ij = ji . Then construct the
function
X
H(t) =
Pi (t) ln Pi (t) .
(6.10)
i

Then
X dP
dH X dPi
i
=
ln Pi
1 + ln Pi ) =
dt
dt
dt
i
i
X
=
ij Pj ln Pi

(6.11)

i,j

X
i,j

where we have used

get

i ij

ij Pj ln Pj ln Pi ,

= 0. Now switch i j in the above sum and add the terms to

1X
dH
ij Pi Pj ln Pi ln Pj .
=
dt
2

(6.12)

i,j

Note that the i = j term does not contribute to the sum. For i 6= j we have ij = Wij 0,
and using the result
(x y) (ln x ln y) 0 ,
(6.13)
we conclude

dH
0.
dt

(6.14)

In equilibrium, Pieq is a constant, independent of i. We write

Pieq =

X
i

H = ln .

(6.15)

170

CHAPTER 6. ERGODICITY AND THE APPROACH TO EQUILIBRIUM

If ij 6= ji , we can still prove a version of the H-theorem. Define a new symmetric matrix
W ij Wij Pjeq = Wji Pieq = W ji ,

(6.16)

and the generalized H-function,

H(t)
Then

X
i

P (t)
Pi (t) ln i eq
Pi

(6.17)


" 


#
Pj
Pj
Pi
Pi
dH
1 X
W ij
=
ln
0.

ln
dt
2
Pieq Pjeq
Pieq
Pjeq

(6.18)

i,j

6.2
6.2.1

Phase Flows in Classical Mechanics

Hamiltonian evolution

The master equation provides us with a semi-phenomenological description of a dynamical

systems relaxation to equilibrium. It explicitly breaks time reversal symmetry. Yet the
microscopic laws of Nature are (approximately) time-reversal symmetric. How can a system
which obeys Hamiltons equations of motion come to equilibrium?
Lets start our investigation by reviewing the basics of Hamiltonian dynamics. Recall the
Lagrangian
t) = T V . The Euler-Lagrange equations of motion for the action

 R L = L(q, q,
S q(t) = dt L are


L
d L
=
,
(6.19)
p =
dt q
q
where p is the canonical momentum conjugate to the generalized coordinate q :
p =

L
.
q

(6.20)

The Hamiltonian, H(q, p) is obtained by a Legendre transformation,

H(q, p) =

r
X

=1

p q L .

(6.21)

Note that
r 
X

L
L
dq
dq
dH =
p dq + q dp
q
q
=1

r 
X
L
L
=
q dp
dq
dt .
q
t
=1

L
dt
t
(6.22)

6.2. PHASE FLOWS IN CLASSICAL MECHANICS

171

Thus, we obtain Hamiltons equations of motion,

H
= q
p
and

H
L
=
= p
q
q

H
L
dH
=
=
.
dt
t
t

Define the rank 2r vector by its components,

if 1 i r
q i
i =

pir if r i 2r .

(6.23)

(6.24)

(6.25)

Then we may write Hamiltons equations compactly as

i = Jij

H
,
j

where
J=

0rr 1rr
1rr 0rr

(6.26)
!

(6.27)

is a rank 2r matrix. Note that J t = J, i.e. J is antisymmetric, and that J 2 = 12r2r .

6.2.2

Dynamical systems and the evolution of phase space volumes

Consider a general dynamical system,

d
= V () ,
dt

(6.28)

where (t) is a point in an n-dimensional phase space. Consider now a compact2 region
R0 in phase space,
its evolution under the dynamics. That is, R0 consists of

 and consider
a set of points | R0 , and if we regard each R0 as an initial condition, we can
define the time-dependent set R(t) as the set of points (t) that were in R0 at time t = 0:



R(t) = (t)  (0) R0 .
(6.29)
Now consider the volume (t) of the set R(t). We have
Z
(t) = d

(6.30)

R(t)

where
d = d1 d2 dn ,
2

Compact in the parlance of mathematical analysis means closed and bounded.

(6.31)

172

CHAPTER 6. ERGODICITY AND THE APPROACH TO EQUILIBRIUM

for an n-dimensional phase space. We then have



Z
Z
 i (t + dt) 


,
(t + dt) = d = d 
j (t) 
R(t)

R(t+dt)

where

(6.32)



 i (t + dt)  (1 , . . . , n )


 (t)  ( , . . . , )
n
1
j

(6.33)

is
which is the Jacobean
of the transformation
from the set of coordinates
 a determinant,



i = i (t) to the coordinates i = i (t + dt) . But according to the dynamics, we have
and therefore

i (t + dt) = i (t) + Vi (t) dt + O(dt2 )
Vi
i (t + dt)
dt + O(dt2 ) .
= ij +
j (t)
j

(6.34)

(6.35)

We now make use of the equality

ln det M = Tr ln M ,

(6.36)

for any matrix M , which gives us3 , for small ,




2




det 1 + A = exp Tr ln 1 + A = 1 + TrA + 12 2 TrA Tr(A2 ) + . . .

Thus,

(t + dt) = (t) +

d V dt + O(dt2 ) ,

(6.37)

(6.38)

R(t)

which says

Z
Z
d
V
= d V = dS n
dt
R(t)

(6.39)

R(t)

Here, the divergence is the phase space divergence,

V =

n
X
Vi
,
i

(6.40)

i=1

and we have used the divergence theorem to convert the volume integral of the divergence to
V , where n
is the surface normal and dS is the differential element
a surface integral of n
of surface area, and R denotes the boundary of the region R. We see that if V = 0
everywhere in phase space, then (t) is a constant, and phase space volumes are preserved
by the evolution of the system.
3
The equality ln det M = Tr ln M is most easily proven by bringing the matrix to diagonal form via a
similarity transformation, and proving the equality for diagonal matrices.

6.2. PHASE FLOWS IN CLASSICAL MECHANICS

173

Figure 6.1: Time evolution of two immiscible fluids. The local density remains constant.
For an alternative derivation, consider a function (, t) which is defined to be the density
of some collection of points in phase space at phase space position and time t. This must
satisfy the continuity equation,

+ (V ) = 0 .
t

(6.41)

This is called the continuity equation. It says that nobody gets lost. If we integrate it over
a region of phase space R, we have
Z
Z
Z
d
(V ) .
(6.42)
d = d (V ) = dS n
dt
R

It is perhaps helpful to think of as a charge density, in which case J = V is the current

density. The above equation then says
Z
dQR
J ,
= dS n
(6.43)
dt
R

where QR is the total charge contained inside the region R. In other words, the rate of
increase or decrease of the charge within the region R is equal to the total integrated current
flowing in or out of R at its boundary.
The Leibniz rule lets us write the continuity equation as

+ V + V = 0 .
t

(6.44)

174

CHAPTER 6. ERGODICITY AND THE APPROACH TO EQUILIBRIUM

But now suppose that the phase flow is divergenceless, i.e. V = 0. Then we have



+ V = 0 .
(6.45)
Dt
t
The combination inside the brackets above is known as the convective derivative. It tells
us the total rate of change of for an observer moving with the phase flow . That is


d
di
(t), t =
+
dt
i dt
t
n
X

D
+
Vi
=
.
=
i
t
Dt

(6.46)

i=1

If D/Dt = 0, the local density remains the same during the evolution of the system. If we
consider the characteristic function
(
1 if R0
(6.47)
(, t = 0) =
0 otherwise
then the vanishing of the convective derivative means that the image of the set R0 under
time evolution will always have the same volume.
Hamiltonian evolution in classical mechanics is volume preserving. The equations of motion
are
H
H
qi = +
,
p i =
(6.48)
pi
qi
A point in phase space is specified by r positions qi and r momenta pi , hence the dimension
of phase space is n = 2r:
 

  
H/p
q
q
=
=
.
(6.49)
,
V =
p
H/q
p
Hamiltons equations of motion guarantee that the phase space flow is divergenceless:

pi
+
V =
qi pi
i=1
(



)
r
X
H

H
=0.
=
+

qi pi
pi
qi
r 
X
qi

(6.50)

i=1

Thus, we have that the convective derivative vanishes, viz.

+ V = 0 ,
Dt
t

(6.51)

for any distribution (, t) on phase space. Thus, the value of the density ((t), t) is
constant, which tells us that the phase flow is incompressible. In particular, phase space
volumes are preserved.

 RECURRENCE
6.3. IRREVERSIBILITY AND POINCARE

6.2.3

175

Liouvilles equation and the microcanonical distribution

Let () = (q, p) be a distribution on phase space. Assuming the evolution is Hamiltonian,

we can write

r 
X

,
qk
= =
+ p k
= iL
(6.52)
t
qk
pk
k=1

is a differential operator known as the Liouvillian:

where L
(
)
r
X
H

H
= i
L

.
pk qk qk pk

(6.53)

k=1

Eqn. 6.52, known as Liouvilles equation, bears an obvious resemblance to the Schr
odinger
equation from quantum mechanics.
Suppose that a () is conserved by the dynamics of the system. Typical conserved quantities include the components of the total linear momentum (if there is translational invariance), the components of the total angular momentum (if there is rotational invariance),
and the Hamiltonian itself (if the Lagrangian is not explicitly time-dependent). Now consider a distribution (, t) = (1 , 2 , . . . , k ) which is a function only of these various
conserved quantities. Then from the chain rule, we have
=

X
a = 0 ,

a
a

(6.54)

since for each a we have

r

X
da
=
dt
=1

a
a
q +
p
q
p

= a = 0 .

(6.55)

We conclude that any distribution (, t) = (1 , 2 , . . . , k ) which is a function solely of

conserved dynamical quantities is a stationary solution to Liouvilles equation.
Clearly the microcanonical distribution,




E H()
E H()

,
=R
E () =
D(E)
d E H()

(6.56)

is a fixed point solution of Liouvilles equation.

6.3

Irreversibility and Poincar

e Recurrence

The dynamics of the master equation describe an approach to equilibrium. These dynamics are irreversible: dH/dt 0, where H is Boltzmanns H-function. However, the
microscopic laws of physics are (almost) time-reversal invariant4 , so how can we understand
4
Actually, the microscopic laws of physics are not time-reversal invariant, but rather are invariant under
the product P CT , where P is parity, C is charge conjugation, and T is time reversal.

176

CHAPTER 6. ERGODICITY AND THE APPROACH TO EQUILIBRIUM

Figure 6.2: Successive images of a set R0 under the -advance mapping g , projected onto
a two-dimensional phase plane. The Poincare recurrence theorem guarantees that if phase
space has finite volume, and g is invertible and volume preserving, then for any set R0
there exists an integer m such that R0 gm R0 6= .
the emergence of irreversibility? Furthermore, any dynamics which are deterministic and
volume-preserving in a finite phase space exhibits the phenomenon of Poincare recurrence,
which guarantees that phase space trajectories are arbitrarily close to periodic if one waits
long enough.

6.3.1

Poincar
e recurrence theorem

The proof of the recurrence theorem is simple. Let g be the -advance mapping which
evolves points in phase space according to Hamiltons equations. Assume that g is invertible
and volume-preserving, as is the case for Hamiltonian flow. Further assume that phase space
volume is finite. Since energy is preserved in the case of time-independent Hamiltonians,
we simply ask that the volume of phase space at fixed total energy E be finite, i.e.
Z

d E H(q, p) < ,

(6.57)

where d = dq dp is the phase space uniform integration measure.

Theorem: In any finite neighborhood R0 of phase space there exists a point 0 which will
return to R0 after m applications of g , where m is finite.
Proof: Assume the theorem fails; we will show this assumption results in a contradiction.

 RECURRENCE
6.3. IRREVERSIBILITY AND POINCARE

177

Consider the set formed from the union of all sets gk R for all m:
=

k=0

gk R0

(6.58)

We assume that the set {gk R0 | k N} is disjoint5 . The volume of a union of disjoint sets
is the sum of the individual volumes. Thus,
vol() =

X
k=0

vol gk R0

= vol(R0 )

X
k=0

(6.59)

1=,

since vol gk R0 = vol R0 from volume preservation. But clearly is a subset of the
entire phase space, hence we have a contradiction, because by assumption phase space is of
finite volume.
Thus, the assumption that the set {gk R0 | k Z+ } is disjoint fails. This means that there
exists some pair of integers k and l, with k 6= l, such that gk R0 gl R0 6= . Without loss
of generality we may assume k < l. Apply the inverse g1 to this relation k times to get
glk R0 R0 6= . Now choose any point 1 gm R0 R0 , where m = l k, and define
0 = gm 1 . Then by construction both 0 and gm 0 lie within R0 and the theorem is
proven.
Poincare recurrence has remarkable implications. Consider a bottle of perfume which is
opened in an otherwise evacuated room, as depicted in fig. 6.3. The perfume molecules
evolve according to Hamiltonian evolution. The positions are bounded because physical
space is finite. The momenta are bounded because the total energy is conserved, hence
no single particle can have a momentum such that T (p) > ETOT , where T (p) is the single particle kinetic energy function6 . Thus, phase space, however large, is still bounded.
Hamiltonian evolution, as we have seen, is invertible and volume preserving, therefore the
system is recurrent. All the molecules must eventually return to the bottle. Whats more,
they all must return with momenta arbitrarily close to their initial momenta!7 In this case,
we could define the region R0 as



R0 = (q1 , . . . , qr , p1 , . . . , pr )  |qi qi0 | q and |pj p0j | p i, j ,
(6.60)
which specifies a hypercube in phase space centered about the point (q0 , p0 ).

Each of the three central assumptions finite phase space, invertibility, and volume preservation is crucial. If any one of these assumptions does not hold, the proof fails. Obviously
if phase space is infinite the flow neednt be recurrent since it can keep moving off in a
5

The natural numbers N is the set of non-negative integers {0, 1, 2, . . .}.

In the nonrelativistic limit, T = p2 /2m. For relativistic particles, we have T = (p2 c2 + m2 c4 )1/2 mc2 .
7
Actually, what the recurrence theorem guarantees is that there is a configuration arbitrarily close to the
initial one which recurs, to within the same degree of closeness.
6

178

CHAPTER 6. ERGODICITY AND THE APPROACH TO EQUILIBRIUM

Figure 6.3: Poincare recurrence guarantees that if we remove the cap from a bottle of
perfume in an otherwise evacuated room, all the perfume molecules will eventually return
to the bottle! (Here H is the Hubble constant.)
particular direction. Consider next a volume-preserving map which is not invertible. An
example might be a mapping f : R R which takes any real number to its fractional part.
Thus, f () = 0.14159265 . . .. Let us restrict our attention to intervals of width less than
unity. Clearly f is then volume preserving. The action of f on the interval [2, 3) is to map
it to the interval [0, 1). But [0, 1) remains fixed under the action of f , so no point within
the interval [2, 3) will ever return under repeated iterations of f . Thus, f does not exhibit
Poincare recurrence.
Consider next the case of the damped harmonic oscillator. In this case, phase space volumes
contract. For a one-dimensional oscillator obeying x
+ 2 x + 02 x = 0 one has V =
2 < 0, since > 0 for physical damping. Thus the convective derivative is Dt =
(V ) = 2 which says that the density increases exponentially in the comoving frame,
as (t) = e2t (0). Thus, phase space volumes collapse: (t) = e22 (0), and are not
preserved by the dynamics. The proof of recurrence therefore fails. In this case, it is possible
for the set to be of finite volume, even if it is the union of an infinite number of sets
gk R0 , because the volumes of these component sets themselves decrease exponentially, as
vol(gn R0 ) = e2n vol(R0 ). A damped pendulum, released from rest at some small angle
0 , will not return arbitrarily close to these initial conditions.

6.3.2

Kac ring model

The implications of the Poincare recurrence theorem are surprising even shocking. If one
takes a bottle of perfume in a sealed, evacuated room and opens it, the perfume molecules
will diffuse throughout the room. The recurrence theorem guarantees that after some finite
time T all the molecules will go back inside the bottle (and arbitrarily close to their initial
velocities as well). The hitch is that this could take a very long time, e.g. much much longer
than the age of the Universe.

 RECURRENCE
6.3. IRREVERSIBILITY AND POINCARE

179

Figure 6.4: Left: A configuration of the Kac ring with N = 16 sites and F = 4 flippers. The
flippers, which live on the links, are represented by blue dots. Right: The ring system after
one time step. Evolution proceeds by clockwise rotation. Spins passing through flippers are
flipped.
On less absurd time scales, we know that most systems come to thermodynamic equilibrium.
But how can a system both exhibit equilibration and Poincare recurrence? The two concepts
seem utterly incompatible!
A beautifully simple model due to Kac shows how a recurrent system can exhibit the
phenomenon of equilibration. Consider a ring with N sites. On each site, place a spin
which can be in one of two states: up or down. Along the N links of the system, F of
them contain flippers. The configuration of the flippers is set at the outset and never
changes. The dynamics of the system are as follows: during each time step, every spin
moves clockwise a distance of one lattice spacing. Spins which pass through flippers reverse
their orientation: up becomes down, and down becomes up.
The phase space for this system consists of 2N discrete configurations. Since each configuration maps onto a unique image under the evolution of the system, phase space volume is
preserved. The evolution is invertible; the inverse is obtained simply by rotating the spins
counterclockwise. Figure 6.4 depicts an example configuration for the system, and its first
iteration under the dynamics.
Suppose the flippers were not fixed, but moved about randomly. In this case, we could focus
on a single spin and determine its configuration probabilistically. Let pn be the probability
that a given spin is in the up configuration at time n. The probability that it is up at time
(n + 1) is then
pn+1 = (1 x) pn + x (1 pn ) ,
(6.61)
where x = F/N is the fraction of flippers in the system. In words: a spin will be up at
time (n + 1) if it was up at time n and did not pass through a flipper, or if it was down
at time n and did pass through a flipper. If the flipper locations are randomized at each

180

CHAPTER 6. ERGODICITY AND THE APPROACH TO EQUILIBRIUM

Figure 6.5: Three simulations of the Kac ring model with N = 2500 sites and three different
concentrations of flippers. The red line shows the magnetization as a function of time,
starting from an initial configuration in which 100% of the spins are up. The blue line shows
the prediction of the Stosszahlansatz , which yields an exponentially decaying magnetization
with time constant .
time step, then the probability of flipping is simply x = F/N . Equation 6.61 can be solved
immediately:
(6.62)
pn = 21 + (1 2x)n (p0 12 ) ,
which decays exponentially to the equilibrium value of peq =
(x) =

1
2

with time scale

1
.
ln |1 2x|

(6.63)

We identify (x) as the microscopic relaxation time over which local equilibrium is established. If we define the magnetization m (N N )/N , then m = 2p 1, so
mn = (1 2x)n m0 . The equilibrium magnetization is meq = 0. Note that for

1
2

<x<1

 RECURRENCE
6.3. IRREVERSIBILITY AND POINCARE

181

Figure 6.6: Simulations of the Kac ring model. Top: N = 2500 sites with F = 201 flippers.
After 2500 iterations, each spin has flipped an odd number of times, so the recurrence time
is 2N . Middle: N = 2500 with F = 2400, resulting in a near-complete reversal of the
population with every iteration. Bottom: N = 25000 with N = 1000, showing long time
equilibration and dramatic resurgence of the spin population.
that the magnetization reverses sign each time step, as well as decreasing exponentially in
magnitude.
The assumption that leads to equation 6.61 is called the Stosszahlansatz 8 , a long German
word meaning, approximately, assumption on the counting of hits. The resulting dynamics
are irreversible: the magnetization inexorably decays to zero. However, the Kac ring model
is purely deterministic, and the Stosszahlansatz can at best be an approximation to the
true dynamics. Clearly the Stosszahlansatz fails to account for correlations such as the
following: if spin i is flipped at time n, then spin i + 1 will have been flipped at time n 1.
8

Unfortunately, many important physicists were German and we have to put up with a legacy of long
German words like Gedankenexperiment, Zitterbewegung, Brehmsstrahlung, Stosszahlansatz , Kartoffelsalat,
etc.

182

CHAPTER 6. ERGODICITY AND THE APPROACH TO EQUILIBRIUM

Also if spin i is flipped at time n, then it also will be flipped at time n + N . Indeed, since
the dynamics of the Kac ring model are invertible and volume preserving, it must exhibit
Poincare recurrence. We see this most vividly in figs. 6.5 and 6.6.
The model is trivial to simulate. The results of such a simulation are shown in figure 6.5 for
a ring of N = 1000 sites, with F = 100 and F = 24 flippers. Note how the magnetization
decays and fluctuates about the equilibrium value meq = 0, but that after N iterations m
recovers its initial value: mN = m0 . The recurrence time for this system is simply N if F is
even, and 2N if F is odd, since every spin will then have flipped an even number of times.
In figure 6.6 we plot two other simulations. The top panel shows what happens when x > 12 ,
so that the magnetization wants to reverse its sign with every iteration. The bottom panel
shows a simulation for a larger ring, with N = 25000 sites. Note that the fluctuations in m
about equilibrium are smaller than in the cases with N = 1000 sites. Why?

6.4
6.4.1

Remarks on Ergodic Theory

Definition of ergodicity

A mechanical system evolves according to Hamiltons equations of motion. We have seen

how such a system is recurrent in the sense of Poincare.
There is a level beyond recurrence called ergodicity. In an ergodic system, time averages
over intervals [0, T ] with T may be replaced by phase space averages. The time
average of a function f () is defined as


1
f () T = lim
T T

ZT


dt f (t) .

(6.64)

For a Hamiltonian system, the phase space average of the same function is defined by
Z
Z






f () S = d f () E H()
d E H() ,
(6.65)
where H() = H(q, p) is the Hamiltonian, and where (x) is the Dirac -function. Thus,


ergodicity
f () T = f () S ,
(6.66)


for all smooth functions f () for which f () S exists and is finite. Note that we do not
average over all of phase space. Rather, we average only over a hypersurface along which
H() = E is fixed, i.e. over one of the level sets of the Hamiltonian function. This is
because the dynamics preserves the energy. Ergodicity means that almost all points will,
upon Hamiltonian evolution, move in such a way as to eventually pass through every finite
neighborhood on the energy surface, and will spend equal time in equal regions of phase
space.

183

6.4. REMARKS ON ERGODIC THEORY

Let R () be the characteristic function of a region R:

(
1
R () =
0

if R
otherwise,

(6.67)

where H() = E for all R. Then


R () = lim
T

time spent in R
T

(6.68)

If the system is ergodic, then


D (E)
R () = P (R) = R
,
T
D(E)

(6.69)

where P (R) is the a priori probability to find R, based solely on the relative volumes
of R and of the entire phase space. The latter is given by
D(E) =

d E H() ,

(6.70)

called the density of states, is the surface area of phase space at energy E, and
DR (E) =

d E H() .

(6.71)

is the density of states for the phase space subset R. Note that
D(E)
=

d E H() =

d
dE

SE

dS
|H|

d(E)
.
d E H() =
dE

(6.72)

(6.73)

Here, dS is the differential surface element, SE is the constant H hypersurface H() = E,

and (E) is the volume of phase space over which H() < E. Note also that we may write
d = dE dE ,
where
dE =
is the the invariant surface element.


dS 
|H| H()=E

(6.74)

(6.75)

184

CHAPTER 6. ERGODICITY AND THE APPROACH TO EQUILIBRIUM

Figure 6.7: Constant phase space velocity at an irrational angle over a toroidal phase space
is ergodic, but not mixing. A circle remains a circle, and a blob remains a blob.

6.4.2

The microcanonical ensemble

The distribution,

E H()
E H()

,
E () =
=R
D(E)
d E H()

(6.76)

defines the microcanonical ensemble (CE) of Gibbs.

We could also write


f () S =

1
D(E)

dE f () ,

(6.77)

SE

integrating over the hypersurface SE rather than the entire phase space.

6.4.3

Ergodicity and mixing

Just because a system is ergodic, it doesnt necessarily mean that

(, t) eq (), for


consider the following motion on the toroidal space = (q, p)  0 q < 1 , 0 p < 1 ,
where we identify opposite edges, i.e. we impose periodic boundary conditions. We also
take q and p to be dimensionless, for simplicity of notation. Let the dynamics be given by
q = 1

p = .

(6.78)

p(t) = p0 + t ,

(6.79)

The solution is
q(t) = q0 + t

hence the phase curves are given by

p = p0 + (q q0 ) .

(6.80)

Now consider the average of some function f (q, p). We can write f (q, p) in terms of its
Fourier transform,
X
fmn e2i(mq+np) .
(6.81)
f (q, p) =
m,n

6.4. REMARKS ON ERGODIC THEORY

185

Figure 6.8: The bakers transformation is a successive stretching, cutting, and restacking.
We have, then,

X
fmn e2i(mq0 +np0) e2i(m+n)t .
f q(t), p(t) =

(6.82)

m,n

We can now perform the time average of f :


f (q, p) T = f00 + lim

= f00

1 X
e2i(m+n)T 1
fmn e2i(mq0 +np0)
T m,n
2i(m + n)

(6.83)

if irrational.

Clearly,


f (q, p) S =

Z1 Z1


dq dp f (q, p) = f00 = f (q, p) T ,
0

(6.84)

so the system is ergodic.

The situation is depicted in fig. 6.7. If we start with the characteristic function of a disc,


(q, p, t = 0) = a2 (q q0 )2 (p p0 )2 ,
(6.85)

then it remains the characteristic function of a disc:

(q, p, t) = a2 (q q0 t)2 (p p0 t)2 ,

(6.86)

A stronger condition one could impose is the following. Let A and B be subsets of SE .
Define the measure
Z
Z
D (E)
,
(6.87)
(A) = dE A ()
dE = A
D(E)

186

CHAPTER 6. ERGODICITY AND THE APPROACH TO EQUILIBRIUM

Figure 6.9: The multiply iterated bakers transformation. The set A covers half the phase
space and its area is preserved under the map. Initially, the fraction of B covered by A is
zero. After many iterations, the fraction of B covered by gnA approaches 12 .
where A () is the characteristic function of A. The measure of a set A is the fraction of
the energy surface SE covered by A. This means (SE ) = 1, since SE is the entire phase
space at energy E. Now let g be a volume-preserving map on phase space. Given two
measurable sets A and B, we say that a system is mixing if


mixing

lim gnA B = (A) (B) .

(6.88)
n

In other words, the fraction of B covered by the nth iterate of A, i.e. gnA, is, as n ,
simply the fraction of SE covered by A. The iterated map gn distorts the region A so
severely that it eventually spreads out evenly over the entire energy hypersurface. Of
course by evenly we mean with respect to any finite length scale, because at the very
smallest scales, the phase space density is still locally constant as one evolves with the
dynamics.
Mixing means that


f () =

d (, t) f ()

d f () E H()

Z

d E H()

h

i

i. h
Tr E H() .
Tr f () E H()

(6.89)

Physically, we can imagine regions of phase space being successively stretched and folded.
During the stretching process, the volume is preserved, so the successive stretch and fold
operations map phase space back onto itself.

6.4. REMARKS ON ERGODIC THEORY

187

Figure 6.10: The Arnold cat map applied to an image of 150 150 pixels. After 300
iterations, the image repeats itself. (Source: Wikipedia)
An example of a mixing system is the bakers transformation, depicted in fig. 6.8. The
baker map is defined by

1

if 0 q < 12
2q , 2 p
g(q, p) =
(6.90)

1
1
1
2q 1 , 2 p + 2
if 2 q < 1 .
Note that g is invertible and volume-preserving. The bakers transformation consists of an
initial stretch in which q is expanded by a factor of two and p is contracted by a factor of
two, which preserves the total volume. The system is then mapped back onto the original
area by cutting and restacking, which we can call a fold. The inverse transformation is
accomplished by stretching first in the vertical (p) direction and squashing in the horizontal
(q) direction, followed by a slicing and restacking. Explicitly,

1

if 0 p < 21
2 q , 2p
(6.91)
g1 (q, p) =

1
1
1
if 2 p < 1 .
2 q + 2 , 2p 1

Another example of a mixing system is Arnolds cat map9

g(q, p) = [q + p] , [q + 2p] ,

(6.92)

where [x] denotes the fractional part of x. One can write this in matrix form as
M

  z }| {  
q
1 1
q
=
mod Z2 .
p
1 2
p
9

The cat map gets its name from its initial application, by Arnold, to the image of a cats face.

(6.93)

188

CHAPTER 6. ERGODICITY AND THE APPROACH TO EQUILIBRIUM

Figure 6.11: The hierarchy of dynamical systems.

The matrix M is very special because it has integer entries and its determinant is det M = 1.
This means that the inverse also has integer entries. The inverse transformation is then

z


M 1

}| {  
 
q
q
2 1
=
mod Z2 .
p
p
1 1

(6.94)

Now for something cool. Suppose that our image consists of a set of discrete points located
at (n1 /k , n2 /k), where the denominator k Z is fixed, and where n1 and n2 range over the
set {1, . . . , k}. Clearly g and its inverse preserve this set, since the entries of M and M 1 are
integers. If there are two possibilities for each pixel (say off and on, or black and white), then
2
there are 2(k ) possible images, and the cat map will map us invertibly from one image to
another. Therefore it must exhibit Poincare recurrence! This phenomenon is demonstrated
vividly in fig. 6.10, which shows a k = 150 pixel (square) image of a cat subjected to the
iterated cat map. The image is stretched and folded with each successive application of the
cat map, but after 300 iterations the image is restored! How can this be if the cat map is
mixing? The point is that only the discrete set of points (n1 /k , n2 /k) is periodic. Points
with different denominators will exhibit a different periodicity, and points with irrational
coordinates will in general never return to their exact initial conditions, although recurrence
says they will come arbitrarily close, given enough iterations. The bakers transformation
is also different in this respect, since the denominator of the p coordinate is doubled upon
each successive iteration.
The student should now contemplate the hierarchy of dynamical systems depicted in fig.
6.11, understanding the characteristic features of each successive refinement10 .
10

There is something beyond mixing, called a K-system. A K-system has positive Kolmogorov-Sinai
entropy. For such a system, closed orbits separate exponentially in time, and consequently the Liouvillian
L has a Lebesgue spectrum with denumerably infinite multiplicity.

189

6.5. THERMALIZATION OF QUANTUM SYSTEMS

6.5
6.5.1

Thermalization of Quantum Systems

Quantum dephasing

Thermalization of quantum systems is fundamentally different from that of classical systems.

Whereas time evolution in classical mechanics is in general a nonlinear dynamical system,
the Schr
odinger equation for time evolution in quantum mechanics is linear:
i~

,
= H
t

(6.95)

is a many-body Hamiltonian. In classical mechanics, the thermal state is conwhere H

structed by time evolution this is the content of the ergodic theorem. In quantum
mechanics, as we shall see, the thermal distribution must be encoded in the eigenstates
themselves.
Let us assume an initial condition at t = 0,
|(0)i =

C | i ,

(6.96)

| i = E | iThe

where | i is an orthonormal eigenbasis for

H
expanP H satisfying
2
sion coefficients satisfy C = h |(0)i and |C | = 1. Normalization requires
X

|C |2 = 1 .

(6.97)

C eiE t/~ | i .

(6.98)

h (0) | (0) i =

The time evolution of |i is then given by

|(t)i =

The energy is distributed according to the time-independent function

X
| (t) i =
P (E) = h (t) | (E H)
|C |2 (E E ) .

(6.99)

Thus, the average energy is time-independent and is given by

| (t) i =
hEi = h (t) | H

Z
X
dE P (E) E =
|C |2 E .

The root mean square fluctuations of the energy are given by

s
2
X
D
X

2 E1/2
|C |2 E .
|C |2 E2
(E)rms = E hEi
=

(6.100)

(6.101)

190

CHAPTER 6. ERGODICITY AND THE APPROACH TO EQUILIBRIUM

Typically we assume that the distribution P (E) is narrowly peaked about hEi, such that
(E)rms E E0 , where E0 is the ground state energy. Note that P (E) = 0 for E < E0 ,
is bounded from below.
i.e. the eigenspectrum of H
Now consider a general quantum observable described by an operator A. We have
X
hA(t)i = h (t) | A | (t) i =
C C ei(E E )t/~ A ,
(6.102)
,

where A = h |A| i. In the limit of large times, we have

1
hAiT lim
T T

ZT
X
dt hA(t)i =
|C |2 A .
0

(6.103)

Note that this implies that all coherence between different eigenstates is lost in the long
time limit, due to dephasing.

6.5.2

Eigenstate thermalization hypothesis

The essential ideas behind the eigenstate thermalization hypothesis (ETH) were described
independently by J. Deutsch (1991) and by M. Srednicki (1994). The argument goes as
follows. If the total energy is the only conserved quantity, and if A is a local, translationallyinvariant, few-body operator, then the time average hAi is given by its microcanonical value,
P
X
2
A (E I)
hAiE ,
(6.104)
hAiT =
|C | A = P
(E I)



where I = E, E + E is an energy interval of width E. So once again, time averages
are micro canonical averages.

But how is it that this is the case? The hypothesis of Deutsch and of Srednicki is that
thermalization in isolated and bounded quantum systems occurs at the level of individual
eigenstates. That is, for all eigenstates | i with E I, one has
A = hAiE .

(6.105)

This means that thermal information is encoded in each eigenstate. This is called the
eigenstate thermalization hypothesis (ETH).
An equivalent version of the ETH is the following scenario. Suppose we have an infinite or
extremely large quantum system U (the universe) fixed in an eigenstate | i. Then form
the projection operator P = | ih |. Projection operators satisfy P 2 = P and their
eigenspectrum consists of one eigenvalue 1 and the rest of the eigenvalues are zero11 . Now
consider a partition of U = W S, where W S. We imagine S to be the system and
11
More generally, we could project onto a K-dimensional subspace, in which case there would be K
eigenvalues of +1 and N K eigenvalues of 0, where N is the dimension of the entire vector space.

6.6. APPENDIX I : FORMAL SOLUTION OF THE MASTER EQUATION

191

W the world. We can always decompose the state | i in a complete product basis for
W and S, viz.
NW NS
X
X
(6.106)
Qpj |pW i |jS i .
| i =
p=1 j=1

Here NW/S is the size of the basis for W/S. The reduced density matrix for S is defined as
S = Tr P =
W

NS
X

j,j =1

NW
X
p=1

Qpj

Qpj

|jS ihjS | .

(6.107)

The claim is that S is a thermal density matrix on S, i.e.

S =

1 H
S ,
e
ZS

(6.108)

is the Hamiltonian restricted to S, and Z = Tre HS , so that Tr = 1 and

where H
S
S
S
S is properly normalized. Here = 1/kB T with T the temperature. The temperature is
) = E (V /V ), where the last factor is a ratio of
fixed by the requirement that Tr(S H
S
S
U
volumes.

6.5.3

When is the ETH true?

There is no rigorous proof of the ETH. Deutsch showed that the ETH holds for the case of
an integrable Hamiltonian weakly perturbed by a single Gaussian random matrix. Horoi et
al. (1995) showed that nuclear shell model wavefunctions reproduce thermodynamic predictions. Recent numerical work by M. Rigol and collaborators has verified the applicability
of the ETH in small interacting boson systems. ETH fails for so-called integrable models,
where there are a large number of conserved quantities, which commute with the Hamiltonian. Integrable models are, however, quite special, and as Deutsch showed, integrability is
spoiled by weak perturbations, in which case ETH then applies.
Note again that in contrast to the classical case, time evolution of a quantum state does
not create the thermal state. Rather, it reveals the thermal distribution which is encoded
in all eigenstates after sufficient time for dephasing to occur, so that correlations between
all the wavefunction expansion coefficients {C } for 6= are all lost.

6.6

Appendix I : Formal Solution of the Master Equation

Recall the master equation Pi = ij Pj . The matrix ij is real but not necessarily sym-

metric. For such a matrix, the left eigenvectors i and the right eigenvectors j are not
the same: general different:
i ij = j
ij j = i .

(6.109)

192

CHAPTER 6. ERGODICITY AND THE APPROACH TO EQUILIBRIUM

Note that the eigenvalue equation for the right eigenvectors is = while that for the
left eigenvectors is t = . The characteristic polynomial is the same in both cases:
F () det ( ) = det ( t ) ,

(6.110)


which means that the left and right eigenvalues are the same. Note also that F () =
F ( ), hence the eigenvalues are either real or appear in complex conjugate pairs. Multiplying the eigenvector equation for on the right by j and summing over j, and multiplying
the eigenvector equation for on the left by i and summing over i, and subtracting the
two results yields




 = 0 ,
(6.111)
where the inner product is

  X
 =
i i .

(6.112)

We can now demand



 = ,

in which case we can write

X 


 
=

ij =

(6.113)
X

i j .

(6.114)

~ = (1, 1, . . . , 1) is a left eigenvector with eigenvalue = 0, since

We have seen that
P
i ij = 0. We do not know a priori the corresponding right eigenvector, which depends
on other details of ij . Now lets expand Pi (t) in the right eigenvectors of , writing
X
C (t) i .
(6.115)
Pi (t) =

Then
X dC
dPi

=
i
dt
dt

= ij Pj =

C ij j

(6.116)

C i .

This allows us to write

dC
= C
dt
Hence, we can write
Pi (t) =

=
X

C (t) = C (0) e t .

C (0) e t i .

(6.117)

(6.118)

It is now easy to see that Re ( ) 0 for all , or else the probabilities will become
negative. For suppose Re ( ) < 0 for some . Then as t , the sum in eqn. 6.118 will be

193

6.7. APPENDIX II : RADIOACTIVE DECAY

dominated by the term for which has the

negative real part; all other contributions
P largest

will be subleading. But we must have i i = 0 since  must be orthogonal to the

~ =0 = (1, 1, . . . , 1). Therefore, at least one component of (i.e. for
left eigenvector
i
some value of i) must have a negative real part, which means a negative probability!12 As
we have already proven that an initial nonnegative distribution {Pi (t = 0)} will remain
nonnegative under the evolution of the master equation, we conclude that Pi (t) Pieq as
t , relaxing to the = 0 right eigenvector, with Re ( ) 0 for all .

6.7

Appendix II : Radioactive Decay

Consider a group of atoms, some of which are in an excited state which can undergo nuclear
decay. Let Pn (t) be the probability that n atoms are excited at some time t. We then model
the decay dynamics by

if m n
0
Wmn = n if m = n 1
(6.119)

0
if m < n 1 .

Here, is the decay rate of an individual atom, which can be determined from quantum
mechanics. The master equation then tells us

dPn
= (n + 1) Pn+1 n Pn .
(6.120)
dt
 
The interpretation here is as follows: let  n denote a state in which n atoms are excited.

2
Then Pn (t) = h (t) | n i . Then Pn (t) will increase due to spontaneous transitions from
| n+1 i to | n i, and will decrease due to spontaneous transitions from | n i to | n1 i.
The average number of particles in the system is
N (t) =

n Pn (t) .

(6.121)

n=0

Note that
i
X h
dN
n (n + 1) Pn+1 n Pn
=
dt
n=0

h
X

n=0

=
12

n=0

n(n 1) Pn n2 Pn

(6.122)

n Pn = N .

Since the probability Pi (t) is real, if the eigenvalue with the smallest (i.e. largest negative) real part
is complex, there will be a corresponding complex conjugate eigenvalue, and summing over all eigenvectors
will result in a real value for Pi (t).

194

CHAPTER 6. ERGODICITY AND THE APPROACH TO EQUILIBRIUM

Thus,
N (t) = N (0) et .

(6.123)

The relaxation time is = 1 , and the equilibrium distribution is

Pneq = n,0 .

(6.124)

Note that this satisfies detailed balance.

We can go a bit farther here. Let us define
P (z, t)

z n Pn (t) .

(6.125)

n=0

This is sometimes called a generating function. Then

i
X h
P
z n (n + 1) Pn+1 n Pn
=
t

n=0

(6.126)

P
P
=
z
.
z
z

Thus,
1 P
P
(1 z)
=0.
t
z

(6.127)

We now see that any function f () satisfies the above equation, where = t ln(1 z).
Thus, we can write


P (z, t) = f t ln(1 z) .
(6.128)


Setting t = 0 we have P (z, 0) = f ln(1 z) , and inverting this result we obtain f (u) =
P (1 eu , 0), i.e.


P (z, t) = P 1 + (z 1) et , 0 .
(6.129)
P
The total probability is P (z = 1, t) =
n=0 Pn , which clearly is conserved: P (1, t) = P (1, 0).
The average particle number is

P 
t
n Pn (t) =
N (t) =
P (1, 0) = N (0) et .
 = e
z
z=1
n=0

6.8

(6.130)

Appendix III : Canonical Transformations in Hamiltonian Mechanics

The Euler-Lagrange equations of motion of classical mechanics are invariant under a redefinition of generalized coordinates,
Q = Q (q1 , . . . , qr , t) ,

(6.131)

6.8. APPENDIX III : CANONICAL TRANSFORMATIONS IN HAMILTONIAN MECHANICS195

called a point transformation. That is, if we express the new Lagrangian in terms of the
new coordinates and their time derivatives, viz.


Q, Q,
t) = L q(Q, t) , q(Q,
t) , t ,
L

Q,
(6.132)
then the equations of motion remain of the form
 

L
d L
.
=
Q
dt Q

Hamiltons equations13 ,
q =

H
p

p =

(6.133)

H
q

(6.134)

are invariant under a much broader class of transformations which mix all the q s and p s,
called canonical transformations. The general form for a canonical transformation is


q = q Q1 , . . . , Qr , P1 , . . . , Pr , t
(6.135)


p = p Q1 , . . . , Qr , P1 , . . . , Pr , t ,
(6.136)

with {1, . . . , r}. We may also write

i = i 1 , . . . , 2r , t ,

with i {1, . . . , 2r}. Here we have

(
qi
if 1 i r
i =
pir if n i 2r

(
Qi
i =
Pir

(6.137)

if 1 i r
if r i 2r .

(6.138)

The transformed Hamiltonian is H(Q,

P, t).
What sorts of transformations are allowed? Well, if Hamiltons equations are to remain
invariant, then

H
H
, P =
,
(6.139)
Q =
P
Q
which gives
P
i
Q
+
=0=
.
(6.140)
Q
P
i
I.e. the flow remains incompressible in the new (Q, P ) variables. We will also require that
phase space volumes are preserved by the transformation, i.e.


 
 (Q, P ) 
i
 = 1 .


det
(6.141)
= 
j
(q, p) 
This last condition guarantees the invariance of the phase space measure
d = hr

r
Y

dq dp ,

=1

where h in the normalization prefactor is Plancks constant.

13

as before.
We revert to using H for the Hamiltonian in this section, rather than H

(6.142)

196

CHAPTER 6. ERGODICITY AND THE APPROACH TO EQUILIBRIUM

Chapter 7

Maps, Strange Attractors, and

Chaos
7.1
7.1.1

Maps
Parametric Oscillator

Consider the equation

x
+ 02 (t) x = 0 ,

(7.1)

where the oscillation frequency is a function of time. Equivalently,

M (t)

(t)

}|
z }|{
z
{ 
  
d x
0
1
x
.
=
02 (t) 0
x
dt x

(7.2)

The formal solution is the path-ordered exponential,

(t) = P exp
Lets consider an example in which

(1 + ) 0
(t) =

(1 ) 0

( Zt

dt M (t ) (0) .

if 2n t (2n + 1)

(7.3)

(7.4)

if (2n + 1) t (2n + 2) .

Define n (2n ). Then

n+1 = exp(M ) exp(M+ ) n ,
197

(7.5)

198

CHAPTER 7. MAPS, STRANGE ATTRACTORS, AND CHAOS

where
M =


0
1
,
2
0

(7.6)

2 I is a multiple of the identity. Evaluating the

with (1 ) 0 . Note that M2 =
Taylor series for the exponential, one finds
!
1
sin
cos

,
(7.7)
exp(M t) =
sin
cos

from which we derive

Q = exp(M ) exp(M+ )
=

cos
sin

1
sin

cos

cos +
+ sin +

!

1
a b
sin +
+
c d
cos +

(7.8)

with
+
sin sin +

1
1
cos sin + +
sin cos +
b=
+

a = cos cos +

(7.9)
(7.10)

c = + cos sin + sin cos +

d = cos cos +

(7.11)

sin sin + .
+

(7.12)

Note that det exp(M ) = 1, hence detQ = 1. Also note that

P () = det Q I = 2 T + ,

(7.13)

where

The eigenvalues of Q are

T = a + d = Tr Q

(7.14)

= ad bc = det Q .

(7.15)

= 12 T

1
2

T 2 4 .

(7.16)

In our case, = 1. There are two cases to consider:

|T | < 2 : + = = ei
|T | > 2 : + =

= cos1 12 T

= e

cosh1 12 |T |

(7.17)
.

(7.18)

When |T | < 2, remains bounded; when |T | > 2, || increases exponentially with time.
Note that phase space volumes are preserved by the dynamics.

199

7.1. MAPS

Figure 7.1: Phase diagram for the parametric oscillator in the (, ) plane. Thick black
lines correspond to T = 2. Blue regions: |T | < 2. Red regions: T > 2. Magenta regions:
T < 2.
To investigate more fully, let 0 . The period of the 0 oscillations is t = 2 , i.e.
pump = / is the frequency at which the system is pumped. We compute the trace of
Q and find
cos(2) 2 cos(2)
1
T
=
.
(7.19)
2
1 2
We are interested in the boundaries in the (, ) plane where |T | = 2. Setting T = +2, we
write = n + , which means 0 npump . Expanding for small and , we obtain the
relation
 1/2
 
2
4 2
.
(7.20)
=
=  
n
Setting T = 2, we write = (n + 12 ) + , i.e. 0 (n + 12 ) pump . This gives
2 = 2

= .

(7.21)

The full phase diagram in the (, ) plane is shown in Fig. 7.1. A physical example is
pumping a swing. By extending your legs periodically, youp
effectively change the length
(t) of the pendulum, resulting in a time-dependent 0 (t) = g/(t).

200

7.1.2

CHAPTER 7. MAPS, STRANGE ATTRACTORS, AND CHAOS

Other models of kicked dynamics

A related model is described by the kicked dynamics of the Hamiltonian

H(t) =

p2
+ V (q) K(t) ,
2m

where
K(t) =

(t n )

(7.22)

(7.23)

n=

is the kicking function. The potential thus winks on and off with period . Note that
lim K(t) = 1 .

(7.24)

In the 0 limit, the system is continuously kicked, and is equivalent to motion in a

time-independent external potential V (q).
The equations of motion are
q =

p
m

p = V (q) K(t) .

(7.25)

Integrating these equations, we obtain the map

p
m n
= pn V (qn+1 ) .

qn+1 = qn +

(7.26)

pn+1

(7.27)

Note that the determinant of Jacobean of the map is unity:




(qn+1 , pn+1 )
1
m
=
.
2
(qn , pn )
V (qn+1 ) 1 m V (qn+1 )

(7.28)

This means that the map preserves phase space volumes.

2

Consider, for example, the Hamiltonian H(t) = L2I V cos() K(t), where L is the angular
momentum conjugate to . This results in the map
n+1 = n + 2 Jn

(7.29)

Jn+1 = Jn sin n+1 ,

(7.30)
p

where Jn = Ln / 2IV and = V /2I. This is known as the standard map 1 . In the
limit 0, we may define x = (xn+1 xn )/ and J = (Jn+1 Jn )/, and we recover the
continuous time dynamics = 2J and J = sin . These dynamics preserve the energy

function E = J 2 cos . There is a separatrix at E = 1, given by J() = 2 cos(/2).
We see from fig. 7.2 that this separatrix is the first structure to be replaced by a chaotic
fuzz as increases from zero to a small finite value.
1
The standard map us usually written in the form xn+1 = xn + Jn and Jn+1
= Jn k sin(2xn+1 ). We
can recover our version by rescaling n = 2xn , Jn k Jn and defining k.

7.1. MAPS

201

Figure 7.2: Top: the standard map, as defined in the text. Four values of the parameter
are shown: = 0.01 (left), = 0.2 (center), and = 0.4 (right). Bottom: details of the
= 0.4 map.
Another well-studied system is the kicked Harper model, for which




2p
2q
H(t) = V1 cos
V2 cos
K(t) .
P
Q

(7.31)

With x = q/Q and y = p/P , Hamiltons equations generate the map

xn+1 = xn + sin(2yn )
(7.32)

(7.33)
yn+1 = yn sin(2xn+1 ) ,

p
p
where = 2 V1 V2 /P Q and = V1 /V2 are dimensionless parameters. In this case, the
conserved energy is
E = 1 cos(2x) cos(2y) .
(7.34)

202

CHAPTER 7. MAPS, STRANGE ATTRACTORS, AND CHAOS

Figure 7.3: The kicked harper map, with = 2, and with = 0.01, 0.125, 0.2, and 5.0
(clockwise from upper left). The phase space here is the unit torus, T2 = [0, 1] [0, 1].
There are then two separatrices, at E = ( 1 ), with equations cos(y) = sin(x)
and sin(y) = cos(x). Again, as is apparent from fig. 7.3, the separatrix is the first
structure to be destroyed at finite . We shall return to discuss this phenomenon below.

7.2

One-dimensional Maps

Consider now an even simpler case of a purely one-dimensional map,

xn+1 = f (xn ) .

(7.35)

A fixed point of the map satisfies x = f (x). Writing the solution as x and expanding about
the fixed point, we write x = x + u and obtain
un+1 = f (x ) un + O(u2 ) .

(7.36)

7.2. ONE-DIMENSIONAL MAPS

203

Figure 7.4: Cobweb diagram showing iterations of the logistic map f (x) = rx(1 x) for
r = 2.8 (upper left), r = 3.4 (upper right), r = 3.5 (lower left), and r = 3.8 (lower right).
Note the single stable fixed point for r = 2.8, the stable two-cycle for r = 3.4, the stable
four-cycle for r = 3.5, and the chaotic behavior for r = 3.8.



Thus, the fixed point is stable if f (x ) < 1, since successive iterates of u then get smaller
and smaller. The fixed point is unstable if f (x ) > 1.

Perhaps the most important and most studied of the one-dimensional maps is the logistic
map, where f (x) = rx(1 x), defined on the interval x [0, 1]. This has a fixed point at
x = 1 r 1 if r > 1. We then have f (x ) = 2 r, so the fixed point is stable if r (1, 3).
What happens for r > 3? We can explore the behavior of the iterated map by drawing a
cobweb diagram, shown in fig. 7.4. We sketch, on the same graph, the curves y = x (in
blue) and y = f (x) (in black). Starting with a point x on the line y = x, we move vertically
until we reach the curve y = f (x). To iterate, we then move horizontally to the line y = x
and repeat the process. We see that for r = 3.4 the fixed point x is unstable, but there is

204

CHAPTER 7. MAPS, STRANGE ATTRACTORS, AND CHAOS

a stable two-cycle, defined by the equations

x2 = rx1 (1 x1 )

(7.37)

x1 = rx2 (1 x2 ) .

(7.38)

The second iterate of f (x) is then

f (2) (x) = f f (x) = r 2 x(1 x) 1 rx + rx2 .

(7.39)

Setting x = f (2) (x), we obtain a cubic equation. Since x x must be a factor, we can
divide out by this monomial and obtain a quadratic equation for x1 and x2 . We find
x1,2 =

1+r

(r + 1)(r 3)
.
2r

(7.40)

How stable is this 2-cycle? We find

d (2)
f (x) = r 2 (1 2x1 )(1 2x2 ) = r 2 + 2r + 4 .
dx

(7.41)

The condition that the 2-cycle be stable is then

At r = 1 +

7.2.1

1 < r 2 2r 4 < 1



r 3, 1 + 6 .

(7.42)

6 = 3.4494897 . . . there is a bifurcation to a 4-cycle, as can be seen in fig. 7.5.

Lyapunov Exponents

The Lyapunov exponent (x) of the iterated map f (x) at point x is defined to be
1
(x) = lim
ln
n n

df (n) (x)
dx

n


1X
ln f (xj ) ,
= lim
n n

(7.43)

j=1

where xj+1 f (xj ). The significance of the Lyapunov exponent is the following. If


Re (x) > 0 then two initial conditions near x will exponentially separate under the
iterated map. For the tent map,
f (x) =

2rx
2r(1 x)

if x <
if x

1
2
1
2

(7.44)

one easily finds (x) = ln(2r) independent of x. Thus, if r > 12 the Lyapunov exponent is
positive, meaning that every neighboring pair of initial conditions will eventually separate
exponentially under repeated application of the map. The Lyapunov exponent for the
logistic map is depicted in fig. 7.6.

7.2. ONE-DIMENSIONAL MAPS

205

Figure 7.5: Iterates of the logistic map f (x) = rx(1 x).

7.2.2

Chaos in the logistic map

What happens in the logistic map for r > 1 + 6 ? At this point, the 2-cycle becomes
unstable and a stable 4-cycle develops. However, this soon goes unstable and is replaced by
a stable 8-cycle, as the right hand panel of fig. 7.5 shows. The first eight values of r where
bifurcations occur are given by
r1 = 3 , r2 = 1 +

6 = 3.4494897 , r3 = 3.544096 , r4 = 3.564407 ,

r5 = 3.568759 , r6 = 3.569692 , r7 = 3.569891 , r8 = 3.569934 , . . .

Feigenbaum noticed that these numbers seemed to be converging exponentially. With the
Ansatz
c
(7.45)
r rk = k ,

206

CHAPTER 7. MAPS, STRANGE ATTRACTORS, AND CHAOS

Figure 7.6: Lyapunov exponent for the logistic map.

one finds
=

rk rk1
,
rk+1 rk

(7.46)

and taking the limit k from the above data one finds
= 4.669202

c = 2.637

r = 3.5699456 .

(7.47)

Theres a very nifty way of thinking about the chaos in the logistic map at the special value
r = 4. If we define xn sin2 n , then we find
n+1 = 2n .
Now let us write
0 =

(7.48)

X
bk
,
2k

(7.49)

k=1

where each bk is either 0 or 1. In other words, the {bk } are the digits in the binary decimal
expansion of 0 /. Now n = 2n 0 , hence
n =

X
bn+k
k=1

2k

(7.50)

We now see that the logistic map has the effect of shifting to the left the binary digits of
n / to yield n+1 /. The last digit falls off the edge of the world, as it were, since it
results in an overall contribution to n+1 which is zero modulo 2. This very emphatically
demonstrates the sensitive dependence on initial conditions which is the hallmark of chaotic
behavior, for eventually two very close initial conditions, differing by 2m , will, after
m iterations of the logistic map, come to differ by O(1).

7.3. ATTRACTORS

207

Figure 7.7: Intermittency in the logistic map in the vicinity of the 3-cycle Top panel:
r = 3.828, showing intermittent behavior. Bottom panel: r = 3.829, showing a stable
3-cycle.

7.2.3

Intermittency

Successive period doubling is one route to chaos, as weve just seen. Another route is
intermittency. Intermittency works like this. At a particular value of our control parameter
r, the map exhibits a stable periodic cycle, such as the stable 3-cycle at r = 3.829, as shown
in the bottom panel of fig. 7.7. If we then vary the control parameter slightly in a certain
direction, the periodic behavior persists for a finite number of iterations, followed by a burst,
which is an interruption of the regular periodicity, followed again by periodic behavior,
ad infinitum. There are three types of intermittent behavior, depending on whether the
Lyapunov exponent goes through Re() = 0 while Im() = 0 (type-I intermittency), or
with Im() = (type-III intermittency, or, as is possible for two-dimensional maps, with
Im() = , a general real number.

7.3

Attractors

An attractor of a dynamical system = V () is the set of values that the system evolves
to after a sufficiently long time. For N = 1 the only possible attractors are stable fixed
points. For N = 2, we have stable nodes and spirals, but also stable limit cycles. For N > 2

208

CHAPTER 7. MAPS, STRANGE ATTRACTORS, AND CHAOS

Figure 7.8: Iterates of the sine map f (x) = r sin(x).

the situation is qualitatively different, and a fundamentally new type of set, the strange
attractor, emerges.
A strange attractor is basically a bounded set on which nearby orbits diverge exponentially
(i.e. there exists at least one positive Lyapunov exponent). To envision such a set, consider
a flat rectangle, like a piece of chewing gum. Now fold the rectangle over, stretch it, and
squash it so that it maintains its original volume. Keep doing this. Two points which
started out nearby to each other will eventually, after a sufficiently large number of folds
and stretches, grow far apart. Formally, a strange attractor is a fractal, and may have
noninteger Hausdorff dimension. (We wont discuss fractals and Hausdorff dimension here.)

7.4

The Lorenz Model

The canonical example of an N = 3 strange attractor is found in the Lorenz model. E.

N. Lorenz, in a seminal paper from the early 1960s, reduced the essential physics of the
coupled partial differential equations describing Rayleigh-Benard convection (a fluid slab of
finite thickness, heated from below in Lorenzs case a model of the atmosphere warmed
by the ocean) to a set of twelve coupled nonlinear ordinary differential equations. Lorenzs
intuition was that his weather model should exhibit recognizable patterns over time. What
he found instead was that in some cases, changing his initial conditions by a part in a
thousand rapidly led to totally different behavior. This sensitive dependence on initial

209

7.4. THE LORENZ MODEL

conditions is a hallmark of chaotic systems.

The essential physics (or mathematics?) of Lorenzs N = 12 system is elicited by the
reduced N = 3 system,
X = X + Y
Y = rX Y XZ
Z = XY bZ ,

(7.51)
(7.52)
(7.53)

where , r, and b are all real and positive. Here t is the familiar time variable (appropriately
scaled), and (X, Y, Z) represent linear combinations of physical fields, such as global wind
current and poleward temperature gradient. These equations possess a symmetry under
(X, Y, Z) (X, Y, Z), but what is most important is the presence of nonlinearities in
the second and third equations.
The Lorenz system is dissipative because phase space volumes contract:
Y
Z
X
+
+
= ( + b + 1) .
X
Y
Z
Thus, volumes contract under the flow. Another property is the following. Let
V =

(7.54)

F (X, Y, Z) = 21 X 2 + 21 Y 2 + 12 (Z r )2 .

(7.55)

Then
F = X X + Y Y + (Z r )Z

= X 2 Y 2 b Z 12 r 12

2

+ 41 b(r + )2 .

(7.56)

Thus, F < 0 outside an ellipsoid, which means that all solutions must remain bounded in
phase space for all times.

Figure 7.9: Evolution of the Lorenz equations for = 10, b = 83 , and r = 15, with initial
conditions (x, y, z) = (0, 1, 0), projected onto the (x, z) plane. The system is attracted by a
stable spiral. (Source: Wikipedia)

210

CHAPTER 7. MAPS, STRANGE ATTRACTORS, AND CHAOS

Figure 7.10: Evolution of the Lorenz equations showing sensitive dependence on initial
conditions. The magenta and green curves differ in their initial X coordinate by 105 .
(Source: Wikipedia)

Figure 7.11: Evolution of the Lorenz equations for = 10, b = 38 , and r = 28, with initial
conditions (X0 , Y0 , Z0 ) = (0, 1, 0), showing the strange attractor. (Source: Wikipedia)

7.4.1

Fixed point analysis

Setting x = y = z = 0, we have three possible solutions. One solution, which is always

present, is x = y = z = 0. If we linearize about this solution, we obtain


X

0
X
d
Y =
r 1 0 Y .
(7.57)
dt
Z
0
0 b
Z

211

7.4. THE LORENZ MODEL

Figure 7.12: The Lorenz attractor, projected onto the (X, Z) plane.
The eigenvalues of the linearized dynamics are found to be
1,2 = 12 (1 + )
3 = b ,

1
2

(1 + )2 + 4(r 1)

(7.58)

and thus if 0 < r < 1 all three eigenvalues are negative, and the fixed point is a stable
node. If, however, r > 1, then 2 > 0 and the fixed point is attractive in two directions but
repulsive in a third, corresponding to a three-dimensional version of a saddle point.
For r > 1, a new pair of solutions emerges, with
X = Y =

b(r 1) ,

Z = r 1 .

(7.59)

Linearizing about either one of these fixed points, we find

X

0
X
d

Y .
1 1 X
Y =
dt

X X
b
Z
Z

(7.60)

The characteristic polynomial of the linearized map is

P () = 3 + (b + + 1) 2 + b( + r) + 2b(r 1) .

(7.61)

Since b, , and r are all positive, P () > 0 for all 0. Since P (0) = 2b(r 1) > 0,
we may conclude that there is always at least one eigenvalue 1 which is real and negative.
The remaining two eigenvalues are either both real and negative, or else they occur as a

212

CHAPTER 7. MAPS, STRANGE ATTRACTORS, AND CHAOS

Figure 7.13: X(t) for the Lorenz equations with = 10, b = 83 , r = 28, and initial conditions
(X0 , Y0 , Z0 ) = (2.7, 3.9, 15.8), and initial conditions (X0 , Y0 , Z0 ) = (2.7001, 3.9, 15.8).
complex conjugate pair: 2,3 = i. The fixed point is stable provided < 0. The
stability boundary lies at = 0. Thus, we set
i
i
h
h
(7.62)
P (i) = 2b(r 1) (b + + 1) 2 + i b( + r) 2 = 0 ,

which results in two equations. Solving these two equations for r(, b), we find
rc =

( + b + 3)
.
b1

(7.63)



The fixed point is stable for r 1, rc . These fixed points correspond to steady convection.
The approach to this fixed point is shown in Fig. 7.9.
The Lorenz system has commonly been studied with = 10 and b = 38 . This means that the
volume collapse is very rapid, since V = 41
3 13.67, leading to a volume contraction
41/3
6
of e
1.16 10 per unit time. For these parameters, one also has rc = 470
19 24.74.
The capture by the strange attractor is shown in Fig. 7.11.
In addition to the new pair of fixed points, a strange attractor appears for r > rs 24.06.
In the narrow interval r [24.06, 24.74] there are then three stable attractors, two of which
correspond to steady convection and the third to chaos. Over this interval, there is also
hysteresis. I.e. starting with a convective state for r < 24.06, the system remains in the
convective state until r = 24.74, when the convective fixed point becomes unstable. The
system is then driven to the strange attractor, corresponding to chaotic dynamics. Reversing
the direction of r, the system remains chaotic until r = 24.06, when the strange attractor
loses its own stability.

7.4.2

Poincar
e section

One method used by Lorenz in analyzing his system was to plot its Poincare section.
This entails placing one constraint on the coordinates (X, Y, Z) to define a two-dimensional

213

7.4. THE LORENZ MODEL

Figure 7.14: Lorenz attractor for b = 83 , = 10, and r = 28. Maxima of Z are depicted by
stars.
surface , and then considering the intersection of this surface with a given phase curve
for the Lorenz system. Lorenz chose to set Z = 0, which yields the surface Z = b1 XY .
Note that since Z = 0, Z(t) takes its maximum and minimum values on this surface; see
fig. 7.14. By plotting the values of the maxima ZN as the integral curve successively passed
through this surface, Lorenz obtained results such as those shown in fig. 7.15, which has
the form of a one-dimensional map and may be analyzed as such. Thus, chaos in the Lorenz
attractor can be related to chaos in a particular one-dimensional map, known as the return
map for the Lorenz system.

7.4.3

R
ossler System

Another simple N = 3 system which possesses a strange attractor is the R

ossler system,
X = Y Z
Y = Z + aY
Z = b + Z(X c) ,

(7.64)
(7.65)
(7.66)

typically studied as a function of c for a = b = 51 . In Fig. 7.17, we present results from

work by Crutchfield et al. (1980). The transition from simple limit cycle to strange attractor
proceeds via a sequence of period-doubling bifurcations, as shown in the figure. A convenient
diagnostic for examining this period-doubling route to chaos is the power spectral density,
or PSD, defined for a function F (t) as
2
 Z

 d

2

it 
F (t) e
(7.67)
F () = 
 = F () .


2

As one sees in Fig. 7.17, as c is increased past each critical value, the PSD exhibits a series
of frequency halvings (i.e. period doublings). All harmonics of the lowest frequency peak

214

CHAPTER 7. MAPS, STRANGE ATTRACTORS, AND CHAOS

Figure 7.15: Plot of relation between successive maxima ZN along the strange attractor for
the Lorenz system.
are present. In the chaotic region, where c > c 4.20, the PSD also includes a noisy
broadband background.

7.4. THE LORENZ MODEL

215

Figure 7.16: Period doubling bifurcations of the R

ossler attractor, projected onto the (x, y)
1
.
plane, for nine values of c, with a = b = 10

Figure 7.17: Period doubling bifurcations of the R

ossler attractor with a = b = 15 , projected
onto the (X,Y) plane, for eight values of c, and corresponding power spectral density for
Z(t). (a) c = 2.6; (b) c = 3.5; (c) c = 4.1; (d) c = 4.18; (e) c = 4.21; (f) c = 4.23; (g)
c = 4.30; (h) c = 4.60.

216

CHAPTER 7. MAPS, STRANGE ATTRACTORS, AND CHAOS

Chapter 8

Front Propagation
8.1

Reaction-Diffusion Systems

Weve studied simple N = 1 dynamical systems of the form

du
= R(u) .
dt

(8.1)

Recall that the dynamics evolves u(t) monotonically toward the first stable fixed point
encountered. Now lets extend the function u(t) to the spatial domain as well, i.e. u(x, t),
and add a diffusion term:
u
= D 2 u + R(u) ,
(8.2)
t
where D is the diffusion constant. This is an example of a reaction-diffusion system. If we
extend u(x, t) to a multicomponent field u(x, t), we obtain the general reaction-diffusion
equation (RDE)
ui
= Dij 2 ui + Ri (u1 , . . . , uN ) .
(8.3)
t
Here, u is interpreted as a vector of reactants, R(u) describes the nonlinear local reaction
kinetics, and Dij is the diffusivity matrix . If diffusion is negligible, this PDE reduces to
decoupled local ODEs of the form u = R(u), which is to say a dynamical system at each
point in space. Thus, any fixed point u of the local reaction dynamics also describes
a spatially homogeneous, time-independent solution to the RDE. These solutions may be
characterized as dynamically stable or unstable, depending on the eigenspectrum of the
Jacobian matrix Mij = i Rj (u ). At a stable fixed point, Re(i ) < 0 for all eigenvalues.

8.1.1

Single component systems

We first consider the single component system,

u
= D2 u + R(u) .
t
217

(8.4)

218

CHAPTER 8. FRONT PROPAGATION

Note that the right hand side can be expressed as the functional derivative of a Lyapunov
functional,


Z
(8.5)
L[u] = ddx 21 D(u)2 U (u) ,
where

Zu
U (u) = du R(u ) .

(8.6)

(The lower limit in the above equation is arbitrary.) Thus, eqn. 8.4 is equivalent to
u
L
=
.
t
u(x, t)

(8.7)

Thus, the Lyapunov functional runs strictly downhill, i.e. L < 0, except where u(x, t) solves
the RDE, at which point L = 0.

8.1.2

Propagating front solutions

Suppose the dynamical system u = R(u) has two or more fixed points. Each such fixed
point represents a static, homogeneous solution to the RDE. We now seek a dynamical,
inhomogeneous solution to the RDE in the form of a propagating front, described by
u(x, t) = u(x V t) ,

(8.8)

where V is the (as yet unknown) front propagation speed. With this Ansatz , the PDE of
eqn. 8.4 is converted to an ODE,
D

du
d2 u
+V
+ R(u) = 0 ,
2
d
d

(8.9)

where = x V t. With R(u) U (u) as in eqn. 8.6, we have the following convenient
interpretation. If we substitute u q, t, D m, and v , this equation describes
the damped motion of a massive particle under friction: m
q + q = U (q). The fixed points

q satisfy U (q ) = 0 and are hence local extrema of U (q). The propagating front solution
we seek therefore resembles the motion of a massive particle rolling between extrema of
the potential U (q). Note that the stable fixed points of the local reaction kinetics have
R (q) = U (q) < 0, corresponding to unstable mechanical equilibria. Conversely, unstable
fixed points of the local reaction kinetics have R (q) = U (q) > 0, corresponding to stable
mechanical equilibria.
A front solution corresponds to a mechanical motion interpolating between two equilibria at
u( = ). If the front propagates to the right then V > 0, corresponding to a positive (i.e.
usual) friction coefficient . Any solution, therefore must start from an unstable equilibrium
point uI and end at another equilibrium uII . The final state, however, may be either a stable
or an unstable equilibrium for the potential U (q). Consider the functions R(u) and U (u)
in the left panels of fig. 8.1. Starting at time = with u = uI = 1, a particle with

219

8.1. REACTION-DIFFUSION SYSTEMS

Figure 8.1: Upper panels : reaction functions R(u) = ru(a u) with r = a = 1 (left) and
R(u) = ru(u a)(u b) with r = 1, a = 1, b = 0.7 (right), along with corresponding
potentials U (u) (bottom panels). Stable fixed points for the local reaction kinetic are shown
with a solid black dot, and unstable fixed points as a hollow black dot. Note that R(u) =
U (u), so stable fixed points for the local reaction kinetics have R (u) = U (u) < 0, and
thus correspond to unstable mechanical equilibria in the potential U (u). Similarly, unstable
fixed points for the local reaction kinetics correspond to stable mechanical equilibria.
positive friction rolls down hill and eventually settles at position uII = 0. If the motion is
underdamped, it will oscillate as it approaches its final value, but there will be a solution
connecting the two fixed points for an entire range of V values. Consider a model where
R(u) = ru(u a)(b u)
r
r
r
U (u) = u4 + (a + b)u3 abu2
4
3
2

(8.10)

with a > 0 > b, which is depicted in the right panels of fig. 8.1. Assuming r > 0, there
are two stable fixed points for the local reaction kinetics: u = a and u = b. Since
1
U (a) U (b) = 12
r(a b)2 (a2 b2 ), the fixed point which is farther from u = 0 has the
higher value. Without loss of generality, let us assume |a| > |b|, as is the case in the figure.

220

CHAPTER 8. FRONT PROPAGATION

One can then roll off the peak at u = a and eventually settle in at the local minimum
u = 0 for a range of c values, provided V is sufficiently large that the motion does not take
u beyond the other fixed point at u = b. If we start at u = b, then a solution interpolating
between this value and u = 0 exists for any positive value of V . As we shall see, this
makes the issue of velocity selection a subtle one, as at this stage it appears a continuous
family of propagating front solutions are possible. At any rate, for this type of front we
have u( = ) = uI and u( = +) = uII , where uI,II correspond to stable and unstable
fixed points of the local dynamics. If we fix x and examine what happens as a function of t,
we have as t , since V > 0, meaning that we start out in the unstable fixed
point and eventually as the front passes over our position we transition to the stable fixed
point. Accordingly, this type of front describes a propagation into an unstable phase. Note
that for V < 0, corresponding to left-moving fronts, we have negative friction, meaning we
move uphill in the potential U (u). Thus, we start at = with u() = 0 and end
up at u(+) = uI,II . But now we have as t , hence once again the stable
phase invades the unstable phase.
Another possibility is that one stable phase invades another. For the potential in the lower
right panel of fig. 8.1, this means starting at the leftmost fixed point where u() = a
and, with V > 0 and positive friction, rolling down hill past u = 0, then back up the other
side, asymptotically coming to a perfect stop at u(+) = b. Clearly this requires that V
be finely tuned to a specific value so that the system dissipates an energy exactly equal to
U (a) U (b) to friction during its motion. If V < 0 we have the time reverse of this motion.
The fact that V is finely tuned to a specific value in order to have a solution means that
we have a velocity selection taking place. Thus, if R(a) = R(b) = 0, then defining
Zb
U = du R(u) = U (b) U (a) ,

(8.11)

we have that u = a invades u = b if U > 0, and u = b invades u = a if U < 0.

The front velocity in either case is fixed by the selection criterion that we asymptotically
approach both local maxima of U (u) as t .
For the equation
Du + V u = ru(u a)(u b) ,
we can find an exact solution of the form
 


ba
a+b
+
tanh(A) .
u() =
2
2
Direct substitution shows this is a solution when
 r 1/2
A = |a b|
8D
r
Dr
V =
(a + b).
2

(8.12)

(8.13)

(8.14)
(8.15)

221

8.1. REACTION-DIFFUSION SYSTEMS

8.1.3

Stability and zero mode

We assume a propagating front solution u(xV t) connecting u() = uL and u(+) = uR .

For || very large, we can linearize about the fixed points uL,R . For we write
R(uL + u) R (uL ) u and the linearized front equation is
D u + V u + R (uL ) u = 0 .

(8.16)

This equation possesses solutions of the form u = A eL , where

D 2L + V L + R (uL ) = 0 .
The solutions are
V

L =
2D

s

V
2D

2

R (uL )
.
D

(8.17)

(8.18)

We assume that uL is a stable fixed point of the local dynamics, which means R (uL ) < 0.
Thus, L, < 0 < L,+ , and the allowed solution, which does not blow up as is
to take L = L + . If we choose L < 0 such that + L << 1, then we can take as initial
conditions u(L ) = uL + AL eL L and u (xL ) = L AL eL L . We can then integrate forward
to = + where u (+ = 0) using the shooting method to fix the propagation speed V .
For +, we again linearize, writing u = uR + u and obtaining
D u + V u + R (uR ) u = 0 .
We find
V

R =
2D

s

V
2D

2

R (uR )
.
D

(8.19)

(8.20)

If we are invading another stable (or metastable) phase, then R (uR ) < 0 and we must
choose R = R, in order for the solution not to blow up as +. If we are invading
an unstable phase, however, then R (uR ) > 0 and there are two possible solutions, with
R, < R,+ < 0. In the limit, the R,+ mode dominates, hence asymptotically we
have R = R,+ , however in general both solutions are present asymptotically, which means
the corresponding boundary value problem is underdetermined.
Next we investigate stability of the front solution. To do this, we must return to our original
reaction-diffusion PDE and linearize, writing
u(x, t) = u(x V t) + u(x, t) ,

(8.21)

where f () is a solution to Du + V u + R(u) = 0. Linearizing in u, we obtain the PDE

2 u
u
=D
+ R u(x V t) u .
2
t
x

(8.22)

While this equation is linear, it is not autonomous, due to the presence of u(x V t) on the
in the argument of R on the right hand side.

222

CHAPTER 8. FRONT PROPAGATION

Lets shift to a moving frame defined by = x V t and s = t. Then

=
+
=
x
x x s

(8.23)

=
+
= V
+
.
t
t
t s
s

(8.24)

So now we have the linear PDE

us = D u + V u + R u() u .

(8.25)

This equation, unlike eqn. 8.22, is linear and autonomous. We can spot one solution
immediately. Let u(, s) = u (), where C is a constant. Then the RHS of the above
equation is
RHS = D u + V u + R (u) u
i
dh
D u + V u + R(u) = 0 ,
=
d

(8.26)

by virtue of the front equation for u() itself. This solution is called the zero mode. It is
easy to understand why such a solution must exist. Due to translation invariance in space
and time, if u(x, t) = u(x V t) is a solution, then so must u(x a, t) = u(x a V t) be
a solution, and for all values of a. Now differentiate with respect to a and set a = 0. The
result is the zero mode, u (x V t).
If we define (writing t = s)
u(, t) = eV /2D (, t) ,

(8.27)

then we obtain a Schr

odinger-like equation for (, t):
t = D + W () ,
where
W () =

V2
R u() .
4D

(8.28)

(8.29)

The Schr
odinger equation is separable, so we can write (, t) = eEt (), obtaining
E = D + W () .

(8.30)

Since the zero mode solution u = u () is an E = 0 solution, and since u () must be

nodeless, the zero mode energy must be the ground state for this system, and all other
eigenvalues must lie at positive energy. This proves that the front solution u(x V t) is
stable, provided the zero mode lies within the eigenspectrum of eqn. 8.30.
For we have u uL with R (uL ) < 0 by assumption. Thus, W () > 0 and all
E > 0 solutions are exponentially decaying. As +, the solutions decay exponentially
if R (uR ) < 0, i.e. if uR is a stable fixed point of the local dynamics. If on the other hand uR

8.1. REACTION-DIFFUSION SYSTEMS

223

is an unstable fixed
ppoint, then R (uR ) > 0. We see immediately that propagation velocities
with V < Vc = 2 DR (uR ) are unstable, since W (+) < 0 in such cases. On the other
hand, for V > Vc , we have that

p
V 2 Vc2 /2D ,
(8.31)
() = eV /2D u() AR exp

which is unnormalizable! Hence, the zero mode is no longer part of the eigenspectrum
translational invariance has been lost! One possible resolution is that V = Vc , where the
threshold of normalizability lies. This is known as the Fisher velocity.

8.1.4

Fishers equation

If we take R(u) = ru(1 u), the local reaction kinetics are those of the logistic equation
u = ru(1 u). With r > 0, this system has an unstable fixed point at u = 0 and a stable
fixed point at u = 1. Rescaling time to eliminate the rate constant r, and space to eliminate
the diffusion constant D, the corresponding one-dimensional RDE is
2u
u
=
+ u(1 u) ,
(8.32)
t
x2
which is known as Fishers equation (1937), originally proposed to describe the spreading of
biological populations. Note that the physical length scale is = (D/r)1/2 and the physical
time scale is = r 1 . Other related RDEs are the Newell-Whitehead Segel equation, for
which R(u) = u(1 u2 ), and the Zeldovich equation, for which R(u) = u(1 u)(a u) with
0 < a < 1.
To study front propagation, we assume u(x, t) = u(x V t), resulting in
du
d2 u
= U (u) ,
+V
2
d
d

(8.33)

U (u) = 13 u3 + 12 u2 .

(8.34)

where
Let v = du/d. Then we have the N = 2 dynamical system
du
=v
d
dv
= u(1 u) V v ,
d
with fixed points at (u , v ) = (0, 0) and (u , v ) = (1, 0). The Jacobian matrix is


0
1
M=
2u 1 V

(8.35)
(8.36)

(8.37)

At (u , v ) = (1, 0), we have det(M ) = 1 and Tr(M ) = V , corresponding to a saddle.

At (u , v ) = (0, 0), we have det(M ) = 1 and Tr(M ) = V , corresponding to a stable node
if V > 2 and a stable spiral if 0 < V < 2. If u(x, t) describes a density, then we must have
u(x, t) 0 for all x and t, and this rules out 0 < V < 2 since the approach to the u = 0
fixed point is oscillating (and damped).

224

CHAPTER 8. FRONT PROPAGATION

Figure 8.2: Evolution of a blip in the Fisher equation. The initial configuration is shown in
red. Progressive configurations are shown in orange, green, blue, and purple. Asymptotically the two fronts move with speed V = 2.

8.1.5

Velocity selection and stability

Is there a preferred velocity V ? According to our analysis thus far, any V Vc = 2 will
yield an acceptable front solution with u(x, t) > 0. However, Kolmogorov and collaborators
proved that starting with the initial conditions u(x, t = 0) = (x), the function u(x, t)
evolves to a traveling wave solution with V = 2, which is the minimum allowed propagation
speed. That is, the system exhibits velocity selection.
We can begin to see why if we assume an asymptotic solution u() = A e as .
Then since u2 u we have the linear equation
u + V u + u = 0

V = + 1 .

(8.38)

Thus, any > 0 yields a solution, with V = V (). Note that the minimum allowed value
is Vmin = 2, achieved at = 1. If < 1, the solution falls off more slowly than for = 1,
and we apparently have a propagating front with V > 2. However, if > 1, the solution
decays more rapidly than for = 1, and the = 1 solution will dominate.
We can make further progress by deriving a formal asymptotic expansion. We start with
the front equation
u + V u + u(1 u) = 0 ,
(8.39)
and we define z = /V , yielding

d2u du
+
+ u(1 u) = 0 ,
dz 2
dz

(8.40)

with = V 2 41 . We now develop a perturbation expansion:

u(z ; ) = u0 (z) + u1 (z) + . . . ,

(8.41)

and isolating terms of equal order in , we obtain a hierarchy. At order O(0 ), we have
u0 + u0 (1 u0 ) = 0 ,

(8.42)

8.2. MULTI-SPECIES REACTION-DIFFUSION SYSTEMS

which is to say

du0
= d ln u1
0 1 = dz .
u0 (u0 1)

Thus,

u0 (z) =

1
,
exp(z a) + 1

225

(8.43)

(8.44)

where a is a constant of integration. At level k of the hierarchy, with k > 1 we have

uk1 + uk + uk

k
X

ul ukl = 0 ,

(8.45)

l=0

which is a first order ODE relating uk at level k to the set {uj } at levels j < k. Separating
out the terms, we can rewrite this as
uk

+ (1 2u0 ) uk =

uk1

k1
X

ul ukl .

(8.46)

l=1

At level k = 1, we have
u1 + (1 2u0 ) u1 = u0 .

(8.47)

Plugging in our solution for u0 (z), this inhomogeneous first order ODE may be solved via
elementary means. The solution is


ln cosh za
2

.
(8.48)
u1 (z) =
2 cosh2 za
2

Here we have adjusted the constant of integration so that u1 (a) 0. Without loss of
generality we may set a = 0, and we obtain
u() =

1
1 ln cosh(/2V )

+ O(V 4 ) .
exp(/V ) + 1 2V 2 cosh2(/2V )

(8.49)

At = 0, where the front is steepest, we have

u (0) =

1
+ O(V 3 ) .
4V

(8.50)

Thus, the slower the front moves, the steeper it gets. Recall that we are assuming V 2
here.

8.2

Multi-Species Reaction-Diffusion Systems

Weve already introduced the general multi-species RDE,

ui
= Dij 2 ui + Ri (u1 , . . . , uN ) .
t

(8.51)

226

CHAPTER 8. FRONT PROPAGATION

We will be interested in stable traveling wave solutions to these coupled nonlinear PDEs.
Well start with a predator-prey model,


2N1
N1
N1
= rN1 1
N1 N2 + D1
t
K
x2

(8.52)

N2
2N2
= N1 N2 N2 + D2
.
t
x2

(8.53)

Rescaling x, t, N1 , and N2 , this seven parameter system can be reduced to one with only
three parameters, all of which are assumed to be positive:
u
2u
= u (1 u v) + D 2
t
x

(8.54)

2v
v
= av (u b) + 2 .
t
x

(8.55)

The interpretation is as follows. According to the local dynamics, species v is parasitic in

that it decays as v = abv in the absence of u. The presence of u increases the growth
rate for v. Species u on the other hand will grow in the absence of v, and the presence of
v decreases its growth rate and can lead to its extinction. Thus, v is the predator and u is
the prey.
Before analyzing this reaction-diffusion system, we take the opportunity to introduce some
notation on partial derivatives. We will use subscripts to denote partial differentiation, e.g.
t =

xx =

2
x2

xxt =

3
x2 t

etc.

(8.56)

Thus, our two-species RDE may be written

ut = u (1 u v) + D uxx

(8.57)

vt = av (u b) + vxx .

(8.58)

We assume 0 < b < 1, in which case there are three fixed points:
empty state: (u , v ) = (0, 0)
prey at capacity: (u , v ) = (1, 0)
coexistence: (u , v ) = (b, 1 b) .
We now compute the Jacobian for the local dynamics:

 

u u u v
1 2u v
u
M=
=
.
v u v v
av
a(u b)
We now examine the three fixed points.

(8.59)

227

8.2. MULTI-SPECIES REACTION-DIFFUSION SYSTEMS

At (u , v ) = (0, 0) we have


1 0
M(0,0) =
0 b

T = 1 b , D = b ,

(8.60)

T = a(1 b) 1 , D = a(1 b) ,

(8.61)

corresponding to a saddle.
At (u , v ) = (1, 0),


1
1
M(1,0) =
0 a(1 b)

which is also a saddle, since 0 < b < 1.

Finally, at (u , v ) = (b, 1 b),


b
b
M(b,1b) =
a(1 b) 0

T = b , D = ab(1 b) .

(8.62)

Since T < 0 and D > 0 this fixed point is stable. For D > 14 T 2 it corresponds
to a spiral, and otherwise a node. In terms of a and b, this transition occurs at
a = b/4(1 b). That is,
stable node: a <

b
4(1 b)

stable spiral: a >

b
.
4(1 b)

The local dynamics has an associated Lyapunov function,

"
"
 #

#
u
u
v
v
L(u, v) = ab
+ (1 b)
.
1 ln
1 ln
b
v
1b
1b

(8.63)

(8.64)

The constants in the above Lyapunov function are selected to take advantage of the relation
x 1 ln x 0; thus, L(u, v) 0, and L(u, v) achieves its minimum L = 0 at the stable
fixed point (b, 1 b). Ignoring diffusion, under the local dynamics we have
dL
= a(u b)2 0 .
dt

8.2.1

(8.65)

Propagating front solutions

We now look for a propagating front solution of the form

u(x, t) = u(x V t)

v(x, t) = v(x V t) .

(8.66)

This results in the coupled ODE system,

D u + V u + u(1 u v) = 0

v + V v + av(u b) = 0 ,

(8.67)
(8.68)

228

CHAPTER 8. FRONT PROPAGATION

where once again the independent variable is = x V t. These two coupled second order
ODEs may be written as an N = 4 system.
We will make a simplifying assumption and take D = D1 /D2 = 0. This is appropriate if one
species diffuses very slowly. An example might be plankton (D1 0) and an herbivorous
species (D2 > 0). We then have D = 0, which results in the N = 3 dynamical system,
du
= V 1 u (1 u v)
d

(8.69)

dv
=w
d

(8.70)

dw
= av (u b) V w ,
d

(8.71)

where w = v . In terms of the N = 3 phase space = (u, v, w), the three fixed points are
(u , v , w ) = (0, 0, 0)

(8.72)

(8.73)

(8.74)

(u , v , w ) = (1, 0, 0)
(u , v , w ) = (b, 1 b, 0) .

The first two are unstable and the third is stable. We will look for solutions where the stable
solution invades one of the two unstable solutions. Since the front is assumed to propagate
to the right, we must have the stable solution at = , i.e. () = (b, 1 b, 0). There
are then two possibilities: either (i) (+) = (0, 0, 0), or (ii) (+) = (1, 0, 0). We will
call the former a type-I front and the latter a type-II front.
For our analysis, we will need to evaluate the Jacobian of the system at the fixed point. In
general, we have

V 1 (1 2u v )
V 1 u
0
M =
0
0
1 .
(8.75)

av
a(u b) V
We now evaluate the behavior at the fixed points.
Lets first look in the vicinity of = (0, 0, 0). The linearized dynamics then give

V 1 0
0
d
0
1 ,
(8.76)
= M
,
M = 0
d
0
ab V
where = + . The eigenvalues are
1 = V 1

We see that 1,2 < 0 while 3 > 0.

2,3 = 21 V

1
2

V 2 + 4ab .

(8.77)

8.2. MULTI-SPECIES REACTION-DIFFUSION SYSTEMS

229

Figure 8.3: Analysis of the characteristic polynomial for the Jacobian of the linearized map
at the fixed point (u , v , w ) = (b, 1 b, 0).
In the vicinity of = (1, 0, 0), we have
d
= M
d

The eigenvalues are

1 = V 1

V
V 1
0
M = 0
0
1 ,
0
a(b 1) V

2,3 = 21 V

1
2

V 2 4a(1 b) .

(8.78)

(8.79)

We now have 1 > 0 and Re(2,3 ) < 0. If we exclude oscillatory solutions, then we
must have
p
(8.80)
V > Vmin = 2 a(1 b) .

Finally, lets examine the structure of the fixed point at = (b, 1 b, 0), where

b V 1
b V 1 0
d
= M
,
M =
0
0
1 ,
(8.81)
d
a (1 b)
0
V
The characteristic polynomial is

P () = det I M

= 3 + V b V 1 2 b + ab(1 b)V 1 .

(8.82)

Although physically we want to restrict a > 0, we will analyze this cubic for general
a, but still 0 < b1 . First, note that P () has extrema when P () = 0, which is to say

1q

2
1
1
V

b
V
3
V b V 1 + 3b .
(8.83)
= = 3

230

CHAPTER 8. FRONT PROPAGATION

Figure 8.4: Sketch of the type-II front. Left panel: 0 < a < a2 , for which the trailing edge
of the front is monotonic. Right panel: a > a2 , for which the trailing edge of the front is
oscillatory. In both cases, 12 < b < 1, and the front propagates to the right.
Note that < 0 < + . Since the sign of the cubic term in P () is positive, we
must have that is a local maximum and + a local minimum. Note furthermore
that both + and are independent of the constant a, and depend only on b and c.
Thus, the situation is as depicted in fig. 8.3. The constant a merely shifts the cubic
P () uniformly up or down. When a = 0, P (0) = 0 and the curve runs through the
origin. There exists an a1 < 0 such that for a = a1 we have P ( ) = 0. Similarly,
there exists an a2 > 0 such that for a = a2 we have P (+ ) = 0. Thus,
a < a1 < 0

Re (1,2 ) < 0 < 3

(8.84)

a1 < a < 0

1 < 2 < 0 < 3

(8.85)

a=0

1 < 2 = 0 < 3

(8.86)

0 < a < a2

1 < 0 < 2 < 3

(8.87)

0 < a2 < a

1 < 0 < Re (2,3 ) .

(8.88)

Since this is the fixed point approached as , we must approach it along one
of its unstable manifolds, i.e. along a direction corresponding to a positive eigenvalue.
Thus, we conclude that if a > a2 that the approach is oscillatory, while for 0 < a < a2
the approach is monotonic.
In fig. 8.4 we sketch the solution for a type-II front, where the stable coexistence phase
invades the unstable prey at capacity phase.

8.3

Excitable Media

Consider the following N = 2 system:

u = f (u, v)

(8.89)

v = g(u, v) ,

(8.90)

231

8.3. EXCITABLE MEDIA

Figure 8.5: Sketch of the nullclines for the dynamical system described in the text.
where 0 < 1. The first equation is fast and the second equation slow. We assume
the nullclines for f = 0 and g = 0 are as depicted in fig. 8.5. As should be clear from the
figure, the origin is a stable fixed point. In the vicinity of the origin, we can write
f (u, v) = au bv + . . .

g(u, v) = +cu dv + . . . ,

(8.91)
(8.92)

where a, b, c, and d are all positive real numbers. The equation for the nullclines in the
vicinity of the origin is then au + bv = 0 for the f = 0 nullcline, and cu dv = 0 for the
g = 0 nullcline. Note that



(f, g) 
a b
=
,
(8.93)
M
c d
(u, v) (0,0)

and detM = ad + b
where c = c and d = d. We then have Tr M = (a + d)
c > 0. Since
the trace is negative and the determinant positive, the fixed point is stable. The boundary
between spiral and node solutions is det M = 14 (TrM )2 , which means

> 2 b
c : stable node
(8.94)
|a d|

< 2 b
|a d|
c : stable spiral .
(8.95)

Although the trivial fixed point (u , v ) = (0, 0) is stable, it is still excitable in the sense that
a large enough perturbation will result in a big excursion. Consider the sketch in fig. 8.6.
Starting from A, v initially increases as u decreases, but eventually both u and v get sucked
into the stable fixed point at O. We call this path the fizzle. Starting from A , however, u
begins to increase rapidly and v increases slowly until the f = 0 nullcline is reached. At
this point the fast dynamics has played itself out. The phase curve follows the nullcline,
since any increase in v is followed by an immediate readjustment of u back to the nullcline.
This state of affairs continues until C is reached, at which point the phase curve makes a
large rapid excursion to D, following which it once again follows the f = 0 nullcline to the

232

CHAPTER 8. FRONT PROPAGATION

Figure 8.6: Sketch of the fizzle, which starts from A, and the burst, which starts from A .

Figure 8.7: Sketch of u(t) and v(t) for the fizzle and burst.

origin O. We call this path a burst. The behavior of u(t) and v(t) during these paths is
depicted in fig. 8.7.
It is also possible for there to be multiple bursts. Consider for example the situation depicted
in fig. 8.8, in which the f = 0 and g = 0 nullclines cross three times. Two of these crossings
correspond to stable (attractive) fixed points, while the third is unstable. There are now
two different large scale excursion paths, depending on which stable fixed point one ends
up at1 .

1
For a more egregious example of a sentence ending in several prepositions: What did you bring that
book I didnt want to be read to out of up around for?

233

8.3. EXCITABLE MEDIA

Figure 8.8: With three nullcline crossings, there are two stable fixed points, and hence two
types of burst. The yellow-centered circles denote stable fixed points, and the blue-centered
star denotes an unstable fixed point.

8.3.1

Front propagation in excitable media

Now lets add in diffusion:

ut = D1 uxx + f (u, v)

(8.96)

vt = D2 vxx + g(u, v) .

(8.97)

ut = u(a u)(u 1) v + D uxx

(8.98)

We will consider a specific model,

vt = b u v .

(8.99)

This is known as the FitzHugh-Nagumo model of nerve conduction (1961-2). It represents a

tractable simplification and reduction of the famous Hodgkin-Huxley model (1952) of electrophysiology. Very briefly, u represents the membrane potential, and v the contributions
to the membrane current from Na+ and K+ ions. We have 0 < a < 1, b > 0, and > 0.
The nullclines for the local dynamics resemble those in fig. 8.5, with the nullcline for the
slow reaction perfectly straight.
We are interested in wave (front) solutions, which might describe wave propagation in
muscle (e.g. heart) tissue. Let us once again define = x V t and assume a propagating
front solution. We then arrive at the coupled ODEs,
D u + V u + h(u) v = 0

V v + bu v = 0 ,

(8.100)
(8.101)

where
h(u) = u(a u)(u 1) .

(8.102)

234

CHAPTER 8. FRONT PROPAGATION

Figure 8.9: Excitation cycle for the FitzHugh-Nagumo model.

Once again, we have an N = 3 dynamical system:
du
=w
d

(8.103)

dv
= b V 1 u + V 1 v
d

(8.104)

dw
= D 1 h(u) + D 1 v V D 1 w ,
d

(8.105)

where w = u .
We assume that b and are both small, so that the v dynamics are slow. Furthermore,
v remains small throughout the motion of the system. Then, assuming an initial value
(u0 , 0, 0), we may approximate
ut D uxx + h(u) .
(8.106)
With D = 0, the points u = 1 and u = 1 are both linearly stable and u = a is linearly
unstable. For finite D there is a wave connecting the two stable fixed points with a unique
speed of propagation.
The equation Du +V u = h(u) may again be interpreted mechanically, with h(u) = U (u).
Then since the cubic term in h(u) has a negative sign, the potential U (u) resembles an
inverted asymmetric double well, with local maxima at u = 0 and u = 1, and a local
minimum somewhere in between at u = a. Since
Z1
U (0) U (1) = du h(u) =
0

hence if V > 0 we must have a <

1
2

1
12 (1

2a) ,

(8.107)

in order for the left maximum to be higher than the

235

8.3. EXCITABLE MEDIA

Figure 8.10: Sketch of the excitation pulse of the FitzHugh-Nagumo model..

right maximum. The constant V is adjusted so as to yield a solution. This entails
Z1
Z
V d u2 = V du u = U (0) U (1) .

(8.108)

The solution makes use of some very special properties of the cubic h(u) and is astonishingly
simple:
V = (D/2)1/2 (1 2a) .
(8.109)
We next must find the speed of propagation on the CD leg of the excursion. There we have
ut D uxx + h(u) vC ,

(8.110)

with u() = uD and u(+) = uC . The speed of propagation is

V = (D/2)1/2 uC 2uP + uD .

(8.111)

We then set V = V to determine the location of C. The excitation pulse is sketched in fig.
8.10
Calculation of the wave speed
Consider the second order ODE,
L(u) D u + V u + A(u u1 )(u2 u)(u u3 ) = 0 .

(8.112)

We assume u1,2,3 are all distinct. Remarkably, a solution may be found. We claim that if
u = (u u1 )(u u2 ) ,

(8.113)

236

CHAPTER 8. FRONT PROPAGATION

then, by suitably adjusting , the solution to eqn. 8.113 also solves eqn. 8.112. To show
this, note that under the assumption of eqn. 8.113 we have
du du

du d
= (2u u1 u2 ) u

u =

= 2 (u u1 )(u u2 )(2u u1 u2 ) .

(8.114)

Thus,
h
i
L(u) = (u u1 )(u u2 ) 2 D(2u u1 u2 ) + V + A(u3 u)
h



i
= (u u1 )(u u2 ) 22 D A u + V + Au3 2 D(u1 + u2 ) .

(8.115)

Therefore, if we choose

A
2D

V =

AD
u1 + u2 2u3
2

(8.116)

with = 1, we obtain a solution to L(u) = 0. Note that the velocity V has been selected.
The integration of eqn. 8.113 is elementary, yielding the kink solution
u2 + u1 exp [(u2 u1 )( 0 )]
1 + exp [(u2 u1 )( 0 )]
h
i
= 21 (u1 + u2 ) + 12 (u1 u2 ) tanh 12 (u2 u1 )( 0 ) ,

u() =

(8.117)

where 0 is a constant which is the location of the center of the kink. This front solution
connects u = u1 and u = u2 . There is also a front solution connecting u1 and u3 :
h
i
(8.118)
u() = 21 (u1 + u3 ) + 12 (u1 u3 ) tanh 12 (u3 u1 )( 0 ) ,
p
with the same value of = A/2D, but with a different velocity of front propagation
V = (AD/2)1/2 (u1 + u3 2u2 ), again with = 1.
It is instructive to consider the analogue mechanical setting of eqn. 8.112. We write D M
and V , and u x, giving
Mx
+ x = A(x x1 )(x x2 )(x x3 ) F (x) ,

(8.119)

where we take x1 < x2 < x3 . Mechanically, the points x = x1,3 are unstable equilibria. A
front solution interpolates between these two stationary states. If = V > 0, the friction
is of the usual sign, and the path starts from the equilibrium at which the potential U (x)
is greatest. Note that
Zx3
U (x1 ) U (x3 ) = dx F (x)
x1

1
12 (x3

i
h
x1 )2 (x2 x1 )2 (x3 x2 )2 .

(8.120)
(8.121)

8.3. EXCITABLE MEDIA

237

Figure 8.11: Mechanical analog for the front solution, showing force F (x) and corresponding
potential U (x).
so if, for example, the integral of F (x) between x1 and x3 is positive, then U (x1 ) > U (x3 ).
For our cubic force F (x), this occurs if x2 > 12 (x1 + x3 ).

238

CHAPTER 8. FRONT PROPAGATION

Chapter 9

Pattern Formation
Patterning is a common occurrence found in a wide variety of physical systems, including
chemically active media, fluids far from equilibrium, liquid crystals, etc. In this chapter we
will touch very briefly on the basic physics of patterning instabilities.

9.0.2

Reaction-Diffusion Dynamics Revisited



Let i (r, t) denote a set of scalar fields satisfying
where

t i + Ji = Ri ,

(9.1)

j
Ji = Dij

(9.2)

is the component of the current density of species i. We assume that the local reaction
kinetics is given by


Ri = Ri {j }, ,
(9.3)
where is a control parameter, or possibly a set of control parameters. Thus,

j + Ri {j }, .
t i = Dij

(9.4)

Let us expand about a homogeneous solution to the local dynamics, R {i , ) = 0, writing
i (r, t) = i + i (r, t) .

We then have


Ri 

+ Dij
j .
t i =
j  j

(9.5)

(9.6)

Assuming Dij
is constant in space, we obtain the linear equation

t (q, t) = Lij (q ; ) j (q, t) ,

239

(9.7)

240

CHAPTER 9. PATTERN FORMATION

Figure 9.1: Instabilities in linear systems = L occur when the eigenvalue with the largest
real part crosses the line Re() = 0. If L is a real operator, its eigenvalues are either real
or come in complex conjugate pairs.
where

Let


Ri 

Lij (q ; ) =
Dij
q q .

j
P () = det I L)

(9.8)

(9.9)

be the characteristic polynomial for L(q ; ). The eigenvalues

a (q ; ) satisfy P (a ) = 0.


If we assume that Lij (q ; ) R is real, then P ( ) = P () , which means that the

eigenvalues a are either purely real or else come in complex conjugate pairs a,1 ia,2 .
The eigenvectors ia (q ; ) need not be real, since L is not necessarily Hermitian. The
general solution is then
X
i (q t) =
Ca ia (q ; ) ea (q ;) t .
(9.10)
a

Modes with Re a > 0 are stabilized by nonlinear terms, e.g. A = rA A3 .

Lets assume the eigenvalues are ordered so that Re (a ) Re (a+1 ), and that Re (1 ) 0
for c .
If 1 (q = 0 ; c ) = 0, we expect a transition between homogeneous (q = 0) states at
= c .
If 1 (q = Q ; c ) = 0, we expect a transition to a spatially modulated structure with
wavevector Q.
If Re 1 (q = 0 ; c ) = 0 but Im 1 (q = 0 ; c ) 6= 0 we expect a Hopf bifurcation and
limit cycle behavior.
If Re 1 (q = Q ; c ) = 0 but Im 1 (q = Q ; c ) 6= 0 we expect a Hopf bifurcation to a
spatiotemporal pattern structure.
In the vicinity of a bifurcation, space and time scales associated with the unstable mode(s)
tend to infinity. This indicates a critical slowing down. If the unstable modes evolve very
slowly, the faster, non-critical modes may be averaged over (i.e. adiabatically eliminated).

241

9.1. TURING AND HOPF INSTABILITIES

For the most unstable mode 1 , we envisage the following possibilities:

= A q2

= A q2 Q

= A q2 Q
2

2 2

2 2

= A q i0

= A q2 Q
where

2 2

homogeneous states ; q = 0 least stable

modulated: q = Q is least stable
B q2 sin2 q

q between q and symmetry axis

i0

Hopf, to modulated state

Hopf, to homogeneous state

9.1

c
.
c

(9.11)

Turing and Hopf Instabilities

As an explicit example in d = 1 dimension, consider the coupled RDEs,

ut = Du uxx + f (u, v)

(9.12)

vt = Dv vxx + g(u, v) .

(9.13)

We linearize about the fixed point u = v = 0, obtaining

L(q)

ut

vt

}|
z

{  
2
f Du q
fv

u
+ ...
= u
gu
gv Dv q 2

u(q, t)

v (q, t)

(9.14)

for the Fourier transforms




Z 
u(x, t) iqx
e
.
= dx
v(x, t)

(9.15)

We now examine the eigenvalues of L. Clearly we have

T Tr(L) = fu + gv (Du + Dv ) q 2
4

(9.16)
2

D det(L) = Du Dv q (Du gv + Dv fu ) q + ,

(9.17)

= fu gv fv gu

(9.18)

where
is the determinant at q = 0. The eigenvalues are
q
1
= 2 T 14 T 2 D .

(9.19)

Recall that in the (T , D) plane, it is the upper left quadrant, with T < 0 and D > 0, where
the fixed point is stable. There are then two instability boundary, both of which are straight

242

CHAPTER 9. PATTERN FORMATION

lines. The first boundary is the positive D axis, i.e. (T = 0 , D > 0), which separates the
stable spiral from the unstable spiral, corresponding to the onset of an oscillatory (Hopf )
instability. Since the q 2 term in T has a negative coefficient1 , this instability first occurs
at q = 0, i.e. in the spatially homogeneous mode. The condition for the Hopf instability is
then
fu + gv = 0 .
(9.20)
The second instability boundary is the half-line (T < 0 , D = 0), which forms the border
with the saddle point region. If the coefficient of the q 2 term in D is negative, i.e. if
(Du gv + Dv fu ) > 0, then the minimum of D(q) occurs at a finite value, q = Q, where
Q2 =

Du gv + Dv fu
.
2Du Dv

(9.21)

In this case, the instability is to a spatially inhomogeneous state. This is the Turing instability. This requires that at least one of fu and gv is positive, or autocatalytic. However, if
both are positive, then the condition for the Hopf instability will already have been satisfied.
So for the Turing instability we must require fu + gv < 0, which says that only one of the
species is autocatalytic. Setting D(Q) = 0, we eliminate Q and obtain the condition
p
Du gv + Dv fu = 2 Du Dv ,

(9.22)

hence, at the threshold of instability, the ordering wavevector is

2

Q =

.
Du Dv

(9.23)

For the Turing instability, we may assume, without loss of generality, that gv < 0 < fu .
The Turing instabilty preempts the Hopf instability when eqn. 9.22 is satisfied before eqn.
9.20. It is therefore a necessary (but not sufficient) condition that Dv > Du . The Turing
instability preempts the Hopf instability when only one species is autocatalytic, and the
autocatalytic species is less diffusive. This requires a slowly diffusing activator and a more
rapidly diffusing inhibitor .

9.2

The Brusselator

Consider the so-called Brusselator model of Prigogine and Lefever (1968). The Brusselator
is a model for two fictitious chemical reactions,
A B

2A + B 3A .
1

We assume both diffusion constants are positive: Du,v > 0.

243

9.2. THE BRUSSELATOR

The species A is assumed to be supplied and removed from the system, in which case, after
adding diffusion, we have two coupled RDEs with
f (u, v) = a (1 + b) u + u2 v

(9.24)

g(u, v) = b u u2 v .

(9.25)

The fixed point f = g = 0 occurs at (u , v ) = (a , b/a). Linearizing the local dynamics

about the fixed point, we obtain


 
fu fv
b 1 a2
.
(9.26)
=
gu gv
b a2
Thus, = a2 > 0. The Hopf instability sets in when fu + gv = 0, i.e. when b = bH , where
bH = 1 + a2 .

(9.27)

For the Turing instability, eqn. 9.22 gives b = bT , where

bT = (1 + c)2 ,

(9.28)

(9.29)

where we have defined

ca

Du
.
Dv

Note that c < a for the Turing instability. These two curves intersect at
p
c = 1 + 1 + a2 .

Note that

Q2 = p

a
Du Dv

Du Q2 = c ,

Dv Q2 =

a2
.
c

(9.30)

(9.31)

Suppose we are close to the Turing instability, and we write b = bT + with || 1. We

first expand the coupled RDEs about the fixed point, writing
u = u + u

(9.32)

(9.33)

v = v + v ,

with u = a and v = ab . Written in terms of u and v, the coupled RDEs take the form


 b
2
2
2
(u) + 2a u v + (u) v
(9.34)
ut = Du uxx u + b u + a v +
a


 b
vt = Dv vxx b u + a2 v
(u)2 + 2a u v + (u)2 v .
(9.35)
a
If we ignore the nonlinear terms, we obtain a linear equation which has a solution
u(x, t) = U11 A(t) cos(Qx)

(9.36)

v(x, t) = V11 A(t) cos(Qx) ,

(9.37)

244

CHAPTER 9. PATTERN FORMATION

Figure 9.2: Instability lines for the Brusselator. The thick blue line denotes a Hopf instability and the thick red line a Turing instability. The dashed light green line is the locus of
points below which the minimum of D lies at Q = 0.
where A(t) = A0 exp(t) is an amplitude, and where the eigenvector (U11 V11 )t satisfies

 


bc1
a2
U11
U11
=
.
(9.38)
2
V11
V11
b
a2 ac
If we set b = bT , and, without loss of generality, take U11 1, we have

c (1 + c)
.
(9.39)
a2
If b > bT , then there exists an eigenvalue which is real and positive, in which case the
amplitude A(t) grows exponentially.
U11 = 1

9.2.1

V11 =

The amplitude equation

The exponential growth of the amplitude A(t) is valid only insofar as the nonlinear terms
in the dynamics are small. Our goal here will be to develop a nonlinear ODE governing
the growth of A(t), assuming |b bT | 1. We follow the treatment of Kessler and Levine
(2009, unpublished).
We assume one Fourier mode will be excited, with q = Q, along with its harmonics. We
therefore write
X

X
Umn Am cos(nQx)
(9.40)
u =
v =

m=1 n=0
X

Vmn Am cos(nQx) .

(9.41)

m=1 n=0

We shall only need the first few terms:

u = U11 A + U31 A3 cos(Qx) + U20 A2 + U22 A2 cos(2Qx) + . . .


v = V11 A + V31 A3 cos(Qx) + V20 A2 + V22 A2 cos(2Qx) + . . . .

(9.42)
(9.43)

245

9.2. THE BRUSSELATOR

Note that we assume U10 = V10 = 0 because the leading behavior is in the U11 and V11
terms. It is through the quadratic nonlinearities that terms with n = 0 are generated.
Note that we also set U21 = V21 = 0, because starting with a solution proportional to
A cos(Qx), the quadratic nonlinearities produce terms with spatial wavevectors q = 0 and
q = 2Q, with amplitude proportional to A2 , corresponding to (m, n) = (2, 0) and (m, n) =
(2, 2), respectively. The cubic nonlinearities generate wavevectors q = Q and q = 3Q, with
amplitude A3 , corresponding to (m, n) = (3, 1) and (3, 3), of which we require only the
former. Starting from A cos(Qx), we never generate terms with (m, n) = (2, 1). We could
include such terms, assigning to them an amplitude B, and derive an amplitude equation
for B, but it would give a subleading contribution to the spatial patterning.
We now undertake the tedious process of working out the RHS of eqns. 9.34 and 9.35 to
order A3 . Throughout our derivation, we shall include only the n = 0, n = 1 and n = 2
harmonics and drop all other terms. We will also assume b = bT whenever it multiplies Am
with m > 1, since = b bT is presumed small, and, as we shall see, the amplitude itself

will be proportional to . Lets roll up our sleeves and get to work!

The first terms we need are all the ones linear in u and v. Thus, we need


Du uxx u = (1 + c) U11 A + U31 A3 cos(Qx) U20 A2

(9.44)

(1 + 4c) U22 A2 cos(2Qx)

Dv vxx =


a2 
4a2
V11 A + V31 A3 cos(Qx)
V A2 cos(2Qx)
c
c 22

(9.45)

as well as
 



3
2
3
cos(Qx)
(9.46)
b u + a v = b U11 A + U31 A + a V11 A + V31 A




+ b U20 + a2 V20 A2 + b U22 + a2 V22 A2 cos(2Qx) .
2

Next, we need the nonlinear terms, starting with

2 2
2 2
(u)2 = 12 U11
A + 21 U11
A cos(2Qx) + U11 2 U20 + U22 A3 cos(Qx) + . . . ,

(9.47)

where the remaining terms are of O(A4 ) or are proportional to cos(3Qx). We also require
2 u v = U11 V11 A2 + U11 V11 A2 cos(2Qx)
(9.48)


+ 2 U11 V20 + 2 V11 U20 + U11 V22 + V11 U22 A3 cos(Qx) + . . . .

Finally, we need

(u)2 v =

3
4

2
U11
V11 A3 cos(Qx) + . . . .

(9.49)

On the left hand side of eqns. 9.34 and 9.35, we have the time derivative terms. Again, as

we shall see, the amplitude A will be proportional to , where = bbT is presumed small.

246

CHAPTER 9. PATTERN FORMATION

Its time derivative At will be proportional to 3/2 . Therefore, terms such as (A2 )t = 2AAt
will be negligible and we shall drop them from the outset. Thus,
ut = U11 At cos(Qx) + . . .

(9.50)

vt = V11 At cos(Qx) + . . . .

(9.51)

We now set to zero the coefficients of cos(nQx) for n = 0, 1, and 2 in each of eqns. 9.34
and 9.35. Setting the n = 0 terms on the RHS of these equations to zero, we obtain
U20 = 0
b 2
U + a U11 V11 + a2 V20 = 0 .
2a 11

(9.52)
(9.53)

Doing the same for the n = 2 terms, we get

b 2
U11 + a U11 V11 + b U22 + a2 1 + 4c1 V22 = 0
2a
4a2
(1 + 4c) U22 +
V =0.
c 22

(9.54)
(9.55)

Solving, we obtain
U20 = 0
V20 =

U22 =
1 c2
2 a3

2 (1 c2 )
9 ac

V22 =

(1 c2 ) (1 + 4c)
.
18 a3

(9.56)
(9.57)

Finally, we need the n = 1 terms. There are three contributions. One comes from the linear
terms, restoring the small differences proportional to = b bT . These terms contribute
a coefficient for cos(Qx) of A in the RHS of eqn. 9.34 and A on the RHS of eqn.
9.35. A second contribution comes from the nonlinear terms. We invoke eqns. 9.47 and
9.48, multiplying the former by ab and the latter by a. The term we seek is proportional to
A3 cos(Qx), with a coefficient


b
(1 + c) (2 + c) (8c2 21c + 4)
U22 + a 2 V20 + V22 + V11 U22 + 43 V11 =
.
a
36 a2 c

We define

(9.58)

(1 + c) (2 + c) (8c2 21c + 4)
.
(9.59)
=
36 a2 c





1
Note that > 0 for c c , c+ , where c = 16
21 313 . Numerically, c 0.20676
and c+ 2.4182. Finally, we have the U31 and V31 terms themselves. Thus, dividing out
the common cos(Qx) factor on both sides of both equations, we have


2
At = A + c (1 + c) U31 + a V31 A3
(9.60)


a2
c (1 + c)
2
At = A (1 + c) U31 +
(1 + c) V31 + A3 .
(9.61)

a2
c


9.3. RAYLEIGH-BENARD
INSTABILITY

247

We can rewrite these equations as a linear system for the coefficients U31 and V31 , viz.

z

}|

c (1 + c)
a2
(1 + c)2 a2

a2
c

{ 

U31
V31

At A + A3
a2 c (1 + c) At + A A3

(9.62)

In order to be able to satisfy the above equation, the RHS must be orthogonal to the
left eigenvector of the matrix M corresponding to the zero eigenvalue. This is called the
solvability condition. It is easy to see that this zero left eigenvector is proportional to


= 1+c c .
(9.63)
Thus, we demand





(1 + c) At A + A3 c a2 c (1 + c) At A + A3 = 0 .

(9.64)

At = A gA3 ,

(9.65)

This, at long last, yields our amplitude equation:

where
=

a2 (b bT )
(1 + c) (a2 c2 )

g=

a2 (2 + c) (8c2 21c + 4)
.
36 a2 c (a2 c2 )

(9.66)

The amplitude equation has a fixed point when At = 0, which says = gA2 . Since c < a,
we have
that is positive for b > bT and negative for b < bT . Furthermore g is positive
for
p


c c , c+ and negative outside this region. Thus, A has a fixed point A = /g (in
addition to the one at A = 0) if both and g are positive, or if both and g are negative.
In the former case, A = 0 is unstable and A = A is stable. In the latter case, A = 0 is
stable and A = A is unstable.

9.3

Rayleigh-B
enard Instability

Consider a layer of fluid between two horizontal plates, as depicted in fig. 9.3. The top plate
is held at temperature T1 and the bottom plate at temperature T2 , with T = T2 T1 > 0.
As the fluid near the bottom plate is heated, it expands, and an upward buoyancy force per
unit volume fbuoy = g T results, where = V1 V
T is the thermal expansion coefficient
and is the fluid density. This buoyancy force is a destabilizing effect, and is opposed
by a stabilizing dissipative force per unit volume fdiss = /d3 , where is the kinematic
viscosity, the thermal diffusivity, and d the distance between the plates. The dimensionless
ratio of these two force densities is known as the Rayleigh number ,
R=

fbuoy
fdiss

g d3 T
.

(9.67)

When R > Rc 1708, the destabilizing effects are sufficient to destroy the homogeneous
state. Due to mass conservation, the entire fluid cannot rise uniformly, hence the instability

248

CHAPTER 9. PATTERN FORMATION

Figure 9.3: Benard convection cells in a fluid heated from below.

occurs at a finite wavelength , owing to the formation of convective rolls known as Benard
cells (see fig. 9.3).
Swift and Hohenberg (1977) showed that the dynamics for this problem reduces to the
following equation for the real field (r, t)2 :



2
2 2
= Q +
3 .
(9.68)
t
Here,

R Rc

(9.69)

x + y y is the in-plane gradient.

measures the distance from the instability, and = x
Distances are measured in units of d, and the maximally unstable wavevector is Q 3.12.
We assume a plane wave disturbance, and we first separate out the oscillating features of
(r, t) so that we may talk of a slowly varying amplitude function A(r, t):
(r, t) = A(r, t) eiQx + A (r, t) eiQx ,

(9.70)

is a unit vector which we henceforth assume to be n

= x.
We then have
where n
3 = A3 e3iQx + 3 |A|2 A eiQx + 3 |A|2 A eiQx + A 3 e3iQx




Q2 + 2 A eiQx = eiQx 2iQ x + x2 + y2 A .

Matching coefficients of eiQx , we find


2 

2
2
A 3 |A|2 A .
t A = 2iQ x + x + y

(9.71)
(9.72)

(9.73)

If we assume that the solution for A is such that x 1/2 and y 1/4 when acting on
A, then the x2 term is subleading relative to Q x and y2 , and we may drop it for || 1
and write



t A =

2iQ x + y2

A 3 |A|2 A .

The field is actually a combination of temperature and vertical velocity fields.

(9.74)

9.4. CENTER MANIFOLD REDUCTION

249

Figure 9.4: Sketch showing separation of frequency scales owing to gap . The fast, stable
modes are to the left, and the slow central modes all have Re a 0.
For > 0 there is a family of stationary solutions of the form

where

A(x, y) = Aq eiqr ei ,

(9.75)

2 1/2
1 
Aq = 2Qqx + qy2
.
3

(9.76)

The instability first occurs at q = 0, at = 0. As we shall see, the nonlinearity severely

limits what multimode structures may form.
The derivation of the Swift-Hohenberg equation utilizes something called the Boussinesq
approximation, in which the density variation of the fluid enters only in the buoyancy
equation. For non-Boussinesq fluids, the symmetry is broken, and one has




2
2 2
t = Q +
+ v 2 3 ,
(9.77)

for which the bifurcation is subcritical. This is called the Haken model.

9.4

Center Manifold Reduction

Consider a dynamical system in the vicinity of a fixed point: i = Lij j P

+ O(2 ). If we
a
expand the variables i in terms of the eigenvectors i of L, writing i = a Aa ia , then
we
 can
treat the Aa as new variables. Furthermore, we assume that the eigenspectrum
a exhibits a gap , as shown in fig. 9.4, which allows us to classify these normal modes
as either stable or central. The stable normal modes have large (Re a < ) negative
1
real parts to their frequencies, and hence relax rapidly. On this time scales <
, the

250

CHAPTER 9. PATTERN FORMATION

central modes are roughly constant. We label these modes as Aca and Asa , respectively. The
dynamics of these modes may be written as
dAca
= Jab Acb + Ma (Ac , As )
dt

(9.78)

dAsa
(9.79)
= Kab Asb + Na (Ac , As ) ,
dt
where Ma and Na are nonlinear. If we assume that the fast, stable modes come to equilibrium, we set A sa = 0 and solve the nonlinear equations Kab Asb + Na (Ac , As ) = 0 to obtain
Asa = Asa (Ac ). Inserting this into the first of the previous sets of equations, we arrive at a
new set of equations for the central modes alone. These new equations, obtained by substituting the solution for the stable modes, which are slaved to the slower central modes, into
the function Ma (Ac , As ), are of the form
dAca
= Lab Acb + Pa (Ac ) .
dt

(9.80)

where Pa is nonlinear.
It is convenient to consider a nonlinear change of variables (Ac , As ) (B c , B s ) so that
the center manifold is described by B s = 0. To this end, we write B = A + F (A), or,
equivalently, A = B + G(B). Note that to linear order the transformation is the identity
A = B because we wish to preserve the identification of the stable and central modes.
As a simple example, consider the system
dA1
= 2A1 A2
dt

(9.81)

dA2
= 5A2 A21 .
(9.82)
dt
We identify A1 as the central mode and A2 as the stable fast mode. We now try a nonlinear
transformation of the form
A1 = B1 + B12 + B1 B2 + B22

A2 = B2 +

B12

+ B1 B2 +

B22

(9.83)
.

We then have
dB1
dB2
dB1
dB2
dB2
dA2
=
+ 2 B1
+ B1
+ B2
+ 2 B2
.
dt
dt
dt
dt
dt
dt

(9.84)

(9.85)

Setting = 0 we obtain the equation

1 + 2 B2

dB2
= 5B2 (1 + 5 ) B12 5 B22 + O(B 3 )
dt

(9.86)

We see that B2 becomes isolated from B1 if we choose = 51 . We are free to choose any
value of we please; for simplicity we choose = 0. The B2 equation is then
dB2
= 5B2 .
dt

(9.87)

9.5. SELECTION AND STABILITY OF SPATIAL PATTERNS

251

The fixed point is B2 = 0. Note that

A2 = B2 15 B12

(9.88)

A2 = 51 A21 + O(A31 ) ,

(9.89)

and so at the fixed point we conclude

which is obvious from inspection of eqn. 9.82 as well.

9.5

Selection and Stability of Spatial Patterns

Consider the spatiotemporal dynamics for a real field (r, t) close to an instability which
lies at |q| = Q. We assume a form
(r, t) =

M 
X
=1

2
n


A (t) eiq r + A (t) eiq r ,

(9.90)

and
= 1. By assuming A = A (t), i.e. with no spatial (r) dependence,
where q = Q n
we are considering a system whose spatial structure is perfect and whose growth rate is
maximum. We now consider the amplitude equations for the A (t). Note that the set of
allowed wavevectors
in the Fourier decomposition
of (r, t) consists of 2M elements, which


we can order q1 , . . . , qM , q1 , . . . , qM . With this ordering, we have q+M = q
and A+M = A , where 1 M . We will use indices i, j, etc. to refer to the entire set:
1 j 2M .

9.5.1

d=1

=x

For systems in one spatial dimension, the amplitude equation is quite limited, since n
is the only possibility, and consequently M = 1. The simplest case is the logistic type
equation,
dA
= A g |A|2 A .
(9.91)
dt
Here, c is a measure of the systems proximity to an instability,
p where is a control
parameter. There are fixed points at A = 0 and, for g > 0, at |A| = /g. Fixed points at
finite A are in fact fixed rings in the Cartesian space (Re A , Im A), since the phase of A is
undetermined and amplitude remains fixed if the phase is varied globally.
With g > 0, eqn. 9.91 describes a supercritical pitchfork bifurcation. The flow is sketched
in the left panel of fig. 9.5. If g < 0, the bifurcation is subcritical, and the finite A fixed
point occurs for < 0 and is unstable. In such a case, we must proceed to the next order
in the amplitude equation,
dA
= A g1 |A|2 A g2 |A|4 A ,
(9.92)
dt
with g1 < 0 and g2 > 0. The bifurcation diagram for this equation is sketched in the right
panel of fig. 9.5.

252

CHAPTER 9. PATTERN FORMATION

Figure 9.5: Flow diagrams for one-dimensional bifurcations A = A g |A|2 A with g > 0
(left) and A = A g1 |A|2 A g2 |A|4 A with g1 < 0 and g2 > 0 (right).

9.5.2

Remarks on the amplitude equations for d > 1

In dimensions d > 1 we have the possibility for nonlinear mixing of K different modes,
provided
K
X
j=1

qj = Q

K
X

j = 0 .
n

(9.93)

j=1

Recall also that Aj is associated with eiqj r and Aj with eiqj r . Under these conditions,
the amplitude equations take the following general form:
X
X
dAi
= Ai + v
ij |Aj |2 Ai
Aj Ak n +n +n , 0 g |Ai |2 Ai g
i
j
k
dt
j
j,k
X

ijkl Aj Ak Al n +n +n +n , 0 + O(A4 ) .
i

(9.94)

j,k,l

The prime on the sum indicates that the term with j = i is excluded. Taking the complex
conjugate, we obtain the equation with index i replaced by i + M . The couplings ij and
ijkl are functions of the relative angles:
i n
j)
ij = (n
i n
j , n
i n
k , . . .) .
ijkl = (n

(9.95)
(9.96)

Note that if we associate Aj with eiqj r we can define Aj Aj , associated with eiqj r .
Also note that v is a constant independent of j and k because the dot products in that case
i n
j = n
j n
k = n
k n
i.
necessarily are all identical: n

9.5. SELECTION AND STABILITY OF SPATIAL PATTERNS

9.5.3

253

d=2

For d = 2 systems all the allowed wavevectors lie on the circle |qj | = Q. Lets consider the
cases M = 1, M = 2, and M = 3.
and q2 = Qx
(up to
For M = 1 we recapitulate the d = 1 case. We have q1 = Qx
2

continuous rotations). The amplitude equation is of the form A = A g |A| A, and the
patterned state is one with bands (stripes). This is not the last word, however, since we
must check its stability with respect to M = 2 and M = 3 patterns.
M = 2 case
For M = 2, we write
(x, y, t) = A1 (t) eiQx + A1 (t) eiQx + A2 (t) eiQy + A2 (t) eiQy .

(9.97)

The amplitude equations are

A 1 = A1 g |A1 |2 A1 g |A2 |2 A1

(9.98)

A 2 = A2 g |A2 |2 A2 g |A1 |2 A2 .

(9.99)

We assume g > 0. There are four possible fixed points (A1 , A2 ):

(A1 , A2 ) = (0 , 0)
p
(A1 , A2 ) = /g (ei1 , 0)
p
(A1 , A2 ) = /g (0 , ei2 )
p
(A1 , A2 ) = /g (1 + ) (ei1 , ei2 )

(I: trivial state)

(II: y-directed bands)
(III: x-directed bands)
(IV: squares)

Note that A1 A1 ei is equivalent to a translation of the pattern in the x-direction

by
x = /Q.
To assess the stability of these fixed points, we write
Aj = Aj (1 + j ) ,

(9.100)

and we find
L

}|
{
z
2
2
2
3g |A1 | g |A2 |
2 g |A2 |
1
1
+ O( 2 ) .
=
2 g |A1 |2
g |A1 |2 3g |A2 |2
2
2

Evaluating L at the four fixed points, we find

2
0

,
LII =
LI =
2 (1 )
0

(9.101)

(9.102)

254

CHAPTER 9. PATTERN FORMATION

Figure 9.6: Phase diagram for the M = 2 system.

and
LIII

(1 ) 2

=
0
2

LIV =

2
1+

(9.103)

Computing T = Tr(L) and D = det(L), we find:

(I) In the trivial phase, we have T = 2 and D = 2 . This fixed point is stable star if
< 0 and an unstable star if > 0.3
(II) In the striped phase, we have T = (1 + ) and D = 2( 1) 2 . Thus, if < 1
the stripes are unstable as the fixed point is a saddle. Now suppose > 1, meaning
D > 0. The fixed point is therefore stable since > 0. If < 0, there is no solution
for the fixed point at all. We next compute D 14 T 2 = 14 ( 3)2 2 which is always
negative, indicating that the fixed point is always a stable node.
(III) Same as case II.
(IV) We find T = 4/( + 1) and D = 42 (1 )/(1 + ). The condition D > 0 is
equivalent to || < 1. For || > 1, this fixed point is a saddle, and is unstable. Note
that |A1,2 |2 = T /4g at the fixed point, so if a solution exists we must have T < 0.
If || < 1 then, the fixed point is stable, and we find D 14 T 2 = 42 2 /(1 + )2 < 0,
indicating a stable node. Note that a fixed point solution exists for < 0 if 1 + < 0,
which means < 1, which is a saddle and unstable.
The phase diagram is shown in fig. 9.6.
3
The star is a nongeneric fixed point, arising here because the eigenspace is degenerate. Note that
D = 41 T 2 for the type I case.

255

9.5. SELECTION AND STABILITY OF SPATIAL PATTERNS

M = 3 case
For M = 3 we write
1 = x

+
2 = 12 x
n

3
2

3 = 12 x

3
2

y ,

(9.104)

j , with n
r+3 = n
. The field is then given by the sum
with qj = Q n
(r, t) = A1 eiq1 r + A2 eiq2 r + A3 eiq3 r + A1 eiq1 r + A2 eiq2 r + A3 eiq3 r .

(9.105)

Lets suppose the symmetry is broken, leaving us with the Haken model,
"
#
2


2
2
= Q +
+ v 2 3 .
(9.106)
t
The resulting amplitude equations are then


dA
= A + v A1 A+1 g |A |2 A g |A1 |2 + |A+1 |2 A ,
dt

(9.107)

Our notation is cyclic in mod 3, so A+1 = A1 when = 3, and A1 = A3 when = 1.

We now convert to amplitude and phase variables, writing
A = R ei .

(9.108)

Plugging this into the amplitude equations and taking the real part, we obtain

where

dR
2
2
+ R+1
= R + vR1 R+1 cos gR3 g R1
R
dt
= 1 + 2 + 3 .

(9.109)

(9.110)

The imaginary part yields the equations

R R
d
= 1 +1 v sin .
dt
R

(9.111)

Adding these last three equations, we obtain a single equation for the total phase angle :
R2 R2 + R22 R32 + R32 R12
d
v sin .
= 1 2
dt
R1 R2 R3
Thus, sin = 0 is a fixed point for these dynamics, and in steady state, we have
v<0
v>0

=
=0.

(9.112)

256

CHAPTER 9. PATTERN FORMATION

Thus, v cos > 0 in steady state, i.e. v cos = |v|. Assuming the phase angle has reached
its steady state value, the amplitude equations are given by


dR1
= R1 + |v|R2 R3 gR13 g R22 + R32 R1
dt


dR2
= R2 + |v|R3 R1 gR23 g R32 + R12 R2
dt

dR3
= R3 + |v|R1 R2 gR33 g R12 + R22 R3 .
dt

(9.113)
(9.114)
(9.115)

Subtracting the first equation from the second gives

"
#






d
R1 R2 = |v|R3 g R12 + R1 R2 + R22 gR32 + gR1 R2 R1 R2 . (9.116)
dt
For sufficiently small , we suspect that the term in brackets is negative, indicating that
the difference R1 R2 tends to zero. As we could have chosen any two distinct indices for
R, we are motivated to consider the symmetric case R1 = R2 = R3 = R. This results in
the equation
dR
= R + |v|R2 g (1 + 2)R3 .
(9.117)
dt
The fixed points R = 0 occur at R = 0 and at R = R , where
R =

|v|

v 2 + 4 g (1 + 2)
.
2g (1 + 2)

(9.118)

This solution describes a hexagonal structure.

To assess the stability of these solutions, it is incorrect to analyze the stability of eqn.
9.117. Rather, we must go back to the full N = 4 system (R1 , R2 , R3 , ). Since the partial
derivatives R / = vR1 R+1 sin all vanish at the fixed point, we only need consider
the N = 3 system (R1 , R2 , R3 ) in eqns. 9.113, 9.114, and 9.115. We must examine the
Jacobian

a b b

R
J =
= b a b
R
b b a
= (a b) I + 3b |ih| ,

(9.119)

where t = 13 (1, 1, 1). Thus, |i is an eigenvector with eigenvalue + = a + 2b, while

the two-dimensional subspace orthogonal to |i has the (doubly degenerate) eigenvalue
= a b. Thus,
+ = + 2|v|R 3(1 + 2)gR2
= 2 |v|R

(9.120)

257

9.5. SELECTION AND STABILITY OF SPATIAL PATTERNS

and
= |v|R 3gR2


2
=
( 1) (2 + ) |v|R .
1 + 2

(9.121)

We have used the fixed point equation

g(1 + 2) R2 |v|R = 0

(9.122)

to eliminate the quadratic terms in R in these expressions for .

Consider the R = R+ fixed point. Plugging the expression for R+ into the expressions for
, the stability criteria < 0 can be investigated. We find that the R = R+ solution is
stable for
(2 + ) v 2
v2
< <
.
(9.123)

4g(1 + 2)
g( 1)2

The lower limit for is set by the condition + = 0, and the upper limit by = 0. The
R = R solution and all other fixed points (such as when v cos < 0) are unstable. For
example, there are so-called mixed modes, where
s
gR12
|v|
,
R2 = R3 =
.
(9.124)
R1 =
g ( 1)
g ( + 1)

These are also unstable.

The hexagonal pattern is described by
h
i
1 r + 1 ) + cos(Q n
2 r + 2 ) cos(Q n
3 r 1 2 ) . (9.125)
(r, t) = 2R(t) cos(Q n

where the upper sign is taken when v > 0 ( = 0), and the lower sign when v < 0 ( = ).
Let us define the primitive reciprocal lattice vectors

b1 = Q x

b2 = Q

2x

With this definition, we have

q2 = b2 b1

q1 = b1
q4 = b1

q5 = b1 b2

3
2

y .

(9.126)

q3 = b2

q6 = b2 .

We also define the primitive direct lattice vectors

4
a1 =
3Q
Note that

2 x

21 y

a b = 2 .

4
a2 =
y .
3Q

(9.127)

(9.128)

258

CHAPTER 9. PATTERN FORMATION

Figure 9.7: Points of high symmetry in the honeycomb lattice. The black dots label the
sublattice, the blue dots the K sublattice, and the red dots the K sublattice.
We can expand r in the basis of primitive direct lattice vectors as
r

1
a1 + 2 a2 .
2
2

(9.129)

Then
h





i
(r, t) = 2R(t) cos b1 r + 1 + cos (b2 b1 ) r + 2 cos b2 r + 1 + 2
h





i
= 2R(t) cos 1 + 1 + cos 2 1 + 2 cos 2 + 1 + 2 .
(9.130)
If we now shift the origin of coordinates, defining
(
1 + 1
if v > 0
1 =
1 + 1 + if v < 0
then we have

2 = 2 + 2 ,

i
h


(r, t) = 2R(t) cos 1 + cos 2 1 + cos 2 ,

(9.131)

(9.132)

where again the upper sign is for v > 0 and the lower sign for v < 0. Consider the case v > 0.
At a fixed time t, the function (r, t) achieves its maxima max = 6R when 1 = 0 modulo
2 and 2 = 0 modulo 2, which is to say when r lies on a triangular lattice of points, which
we call the sublattice. This pattern is known as the H0 hexagonal pattern. The minima lie
on a honeycomb lattice, which can be described as two interpenetrating triangular lattices.
2

The K sublattice is defined by the set of points equivalent to 1 = 4

3 and 2 = 3 , while
4
2

the K sublattice consists of the points equivalent to 1 = 3 and 2 = 3 , again modulo

2 in either component. The sum of the three cosines is then 23 , hence min = 3R. For
v < 0 the roles of minima and maxima are reversed, and the maxima lie on the vertices of
a honeycomb lattice; this is the H structure. See figs. 9.7 and 9.8.

259

9.5. SELECTION AND STABILITY OF SPATIAL PATTERNS

Figure 9.8: Two-dimensional stationary Turing patterns, showing H0 (left), striped (center),
and H (right) structures. White areas correspond to local maxima of the concentration
field. From A. De Wit, Adv. Chem. Phys. 109, 435 (1999).
Hexagons are not the only stable pattern, however. We can find stripe solutions
where
p
R1 = R and R2 = R3 = 0. Nontrivial solutions occur for > 0, where R = /g, as in the
one-dimensional case. The Jacobian at this fixed point is given by

2
0
0
q
0
(1 ) |v| g
.
J =
(9.133)

(1 )
0
|v| g

q
The eigenvalues are 1 = 2 and 2,3 = (1 ) |v| g . Since > 0 in order to have a

nontrivial solution, 1 < 0 and we can focus on 2,3 . If < 1 then 2 > 0, so stripes are
unstable for < 1. If > 1, we have that 3 < 0, and the condition 2 < 0 requires
>

v2
g(1 )2

(9.134)

for the stripes to be stable, along with > 1.

It is convenient to rescale, defining
|v| u
Then we find

4g(1 + 2)

u
RS
g

1 1+
SHEX () =
1 + 2

u2 .

(9.135)

(9.136)

SSTRIPE () =

The upper solution for the hexagons is stable for < < + , whereas the stripe solution
is stable for > s , where
= 1

s =

4 (1 + 2)
( 1)2

These results are plotted in fig. 9.9.

+ =

4 (1 + 2) (2 + )
.
( 1)2

(9.137)

260

CHAPTER 9. PATTERN FORMATION

Figure 9.9: Dimensionless amplitude S versus dimensionless coupling for the M = 3

system discussed in the text, showing stable (solid) and unstable (dashed) solutions.

9.5.4

d=3

We can extend the previous d = 2 results to d = 3 dimensions simply by assuming there

is no variation in the along z, the coordinate in the third direction. Thus, stripes in d = 2
become lamellae in d = 3. New structures emerge as well:
M = 3 : 2M = 6
M = 4 : 2M = 8

simple cubic in reciprocal and real space

BCC reciprocal lattice, FCC direct lattice

M = 6 : 2M = 12 FCC reciprocal lattice, BCC direct lattice .

For a description of the many patterned structures which can form under these periodicities,
see T. K. Callahan and E. Knobloch, Nonlinearity 10, 1179 (1997) and idem, Physica D
132, 339 (1999).

9.6

Anisotropy

Many physical systems exhibit intrinsic spatial anisotropies. To see how these might affect
patterning, consider a modification of the Brusselator with anisotropic diffusion:
ut = Du,k uxx + Du, uyy + a (b + 1) u + u2 v
vt = Dv,k vxx + Dv, vyy + b u u2 v .

(9.138)
(9.139)

The linearized dynamics, from eqn. 9.14, are given by the matrix

e u () q 2
b1D
L(q) =
b

a2
a2

e v () q 2
D

(9.140)

9.7. PHASE DIFFUSION : ECKHAUS AND ZIGZAG INSTABILITIES

261

where
qx = q cos

qy = q sin

(9.141)

2
e () = D cos2 D
D
u
u, sin
u,k

(9.142)

and

2
e v () = D cos2 D
D
v, sin .
v,k

(9.143)

We identify the maximally unstable wavevector for the Turing instability det(L) = 0 as
before, i.e. by minimizing det(L) with respect to q 2 . We then invoke det(L) = 0, which
gives us a second equation. From these two equations we obtain the critical value of b at
the transition and the critical wavevector at the transition:
s
!2
e u ()
D
bT () = 1 + a
(9.144)
e v ()
D
a
.
Q2 () = q
e
e
Du () Dv ()

(9.145)

We have thus described a one-parameter family of Turing instabilities, as a function of the

angle . The earliest (smallest bT () value) of these will preempt the others. Examining
bT (), we find
Du,
Du,k
Du,
Du,k

9.7

>

<

Dv,k
Dv,
Dv,k
Dv,

=0

.
2

(9.146)

(9.147)

Phase Diffusion : Eckhaus and Zigzag Instabilities

Starting from the

equation, the dynamics of a striped configuration, with
 Swift-Hohenberg

(x, y, t) = 2 Re A(x, y, t) eiQx are governed by the Newell-Whitehead-Segel equation,
A
= A +
T

i 2
+
X
2 Y 2

2

A |A|2 A ,

(9.148)

where X, Y , and T are scaled slow variables, X = 0 Qx, Y = |0 |1/2 Qy, T = 4Q2 |0 |t,
and = 4Q2 |0 | , where |0 | 1. The amplitude has also been scaled by a factor of
2 Q2 | |1/2 .
0
3

The optimal pattern is given by the constant solution A = , for > 0, corresponding to

(x) = 2 cos(Qx). However, for > 0 an entire band of wavevectors is linearly unstable,
since (Q2 + 2 )2 acting on any plane wave eiqr will yield a positive growth rate so long

262

CHAPTER 9. PATTERN FORMATION

Figure 9.10: Zigzag, Eckhaus, and Benjamin-Feir instabilities.

as |q 2 Q2 | < . Thus the Newell-Whitehead-Segel (NWS) equation admits a band of

static solutions,
p
(9.149)
A(X, Y ) = k2 eikX ,
for |k| <

. We now investigate the stability of these solutions, writing

A(X, Y, T ) =

p


k2 + (X, Y, T ) eikX ei(X,Y,T ) ,

(9.150)

and linearizing in the amplitude and phase variations and .

We start by defining = kX + . Then
ei

e =
+ ik + i
X
X
X

(9.151)

ei

i
e =
+i
.
Y
Y
Y

(9.152)

Thus,


ei X 2i Y2 ei = ik + X 2i Y2 + iX + 21 Y Y + Y Y + 2i 2Y

(9.153)

p


k2 + , and then keep only terms
We need to square the RHS and then apply it to
up to linear order in , , and their derivatives. Clearly we can drop the last two terms
on the RHS above since Y Y will be nonzero only when acting on or , resulting in a
nonlinear contribution; the last term 2Y is already nonlinear. Even with the reduction from
seven to five terms, squaring is a slightly tedious process, and we skip the intermediate steps.
Multiplying the NWS equation on the left by ei and then collecting real and imaginary

263

9.7. PHASE DIFFUSION : ECKHAUS AND ZIGZAG INSTABILITIES

Figure 9.11: Boundary curves for Eckhaus and zigzag instabilities. Structures within the
shaded region are stable.
terms, we obtain the coupled equations
T = 2( k2 ) 2k

T = p

2k

k2

k2 X + XX + k Y Y +

k2 XY Y 41 Y Y Y Y

1
X + XX + k Y Y p
XY Y 14 Y Y Y Y .
2
2 k

(9.154)
(9.155)

The terms on the RHS of these equations are ordered by increasing powers of derivatives. We
assume a long wavelength disturbance, meaning we can neglect all but the lowest nontrivial
terms. From the RHS of the first equation, we take the first two terms, which yield
k
= p
X .
k2

(9.156)

Note that X , which means that the LHS of the first equation is T XT , which has
one more derivative. Substituting this result into the second equation, we obtain


3k2
XX + k Y Y .
(9.157)
T =
k2
This is an anisotropic diffusion equation. We identify
D,X =

3k2
k2

D,Y = k .

(9.158)

An instability occurs when either diffusion constant is negative. Note that = k2 is the
so-called marginal stability boundary and that no patterned solution exists for k2 > . The

264

CHAPTER 9. PATTERN FORMATION

condition D,X < 0 corresponds to the Eckhaus instability and D,Y < 0 to the zigzag
instability. A sketch of the stability boundaries is provided in fig. 9.11.
These are many other patterning instabilities. For example, the Benjamin-Feir instability
is an analog of the Eckhaus instability which occurs in travelling plane waves. This and
other such examples are discussed in detail in the books by R. Hoyle and by M. Cross and
H. Greenside, both listed in chapter 0 of these notes.

Chapter 10

Solitons
Starting in the 19th century, researchers found that certain nonlinear PDEs admit exact
solutions in the form of solitary waves, known today as solitons. Theres a famous story of
the Scottish engineer, John Scott Russell, who in 1834 observed a hump-shaped disturbance
propagating undiminished down a canal. In 1844, he published this observation1 , writing,
I was observing the motion of a boat which was rapidly drawn along a
narrow channel by a pair of horses, when the boat suddenly stopped - not so the
mass of water in the channel which it had put in motion; it accumulated round
the prow of the vessel in a state of violent agitation, then suddenly leaving it
behind, rolled forward with great velocity, assuming the form of a large solitary
elevation, a rounded, smooth and well-defined heap of water, which continued
its course along the channel apparently without change of form or diminution of
speed. I followed it on horseback, and overtook it still rolling on at a rate of some
eight or nine miles an hour, preserving its original figure some thirty feet long
and a foot to a foot and a half in height. Its height gradually diminished, and
after a chase of one or two miles I lost it in the windings of the channel. Such,
in the month of August 1834, was my first chance interview with that singular
and beautiful phenomenon which I have called the Wave of Translation.
Russell was so taken with this phenomenon that subsequent to his discovery he built a
thirty foot wave tank in his garden to reproduce the effect, which was precipitated by an
initialpsudden displacement of water. Russell found empirically that the velocity obeyed
g(h + um ) , where h is the average depth of the water and um is the maximum
v
vertical displacement of the wave. He also found that a sufficiently large initial displacement
would generate two solitons, and, remarkably, that solitons can pass through one another
undisturbed. It was not until 1890 that Korteweg and deVries published a theory of shallow
water waves and obtained a mathematical description of Russells soliton.
1
J. S. Russell, Report on Waves, 14th Meeting of the British Association for the Advancement of Science,
pp. 311-390.

265

266

CHAPTER 10. SOLITONS

Nonlinear PDEs which admit soliton solutions typically contain two important classes of
terms which feed off each other to produce the effect:
DISPERSION

NONLINEARITY

The effect of dispersion is to spread out pulses, while the effect of nonlinearities is, often,
to draw in the disturbances. We saw this in the case of front propagation, where dispersion
led to spreading and nonlinearity to steepening.
In the 1970s it was realized that several of these nonlinear PDEs yield entire families
of exact solutions, and not just isolated solitons. These families contain solutions with
arbitrary numbers of solitons of varying speeds and amplitudes, and undergoing mutual
collisions. The three most studied systems have been
The Korteweg-deVries equation,
ut + 6uux + uxxx = 0 .

(10.1)

This is a generic equation for long waves in a dispersive, energy-conserving medium,

to lowest order in the nonlinearity.
The Sine-Gordon equation,
tt xx + sin = 0 .

(10.2)

The name is a play on the Klein-Gordon equation, tt xx + = 0. Note that the

Sine-Gordon equation is periodic under + 2.
The nonlinear Schr
odinger equation,
it xx + 2||2 = 0 .

(10.3)

Here, is a complex scalar field. Depending on the sign of the second term, we denote
this equation as either NLS(+) or NLS(), corresponding to the so-called focusing
(+) and defocusing () cases.
Each of these three systems supports soliton solutions, including exact N -soliton solutions,
and nonlinear periodic waves.

10.1

The Korteweg-deVries Equation

Let h0 denote the resting depth of water in a one-dimensional channel, and y(x, t) the
vertical displacement of the waters surface. Let L be a typical horizontal scale of the
wave. When |y| h0 L, and v 0 (speed of propagation small compared with c0 ), the
evolution of an x-directed wave is described by the KdV equation,
yt + c0 yx +

3c0
yy + 1 c h2 y
=0,
2h0 x 6 0 0 xxx

(10.4)

267

10.1. THE KORTEWEG-DEVRIES EQUATION

p
where c0 = gh0 . For small amplitude disturbances, only the first two terms are consequential, and we have
yt + c0 yx 0 ,
(10.5)
the solution to which is
y(x, t) = f (x c0 t) ,

(10.6)

where f () is an arbitrary shape; the disturbance propagates with velocity c0 . When the
dispersion and nonlinearity are included, only a particular pulse shape can propagate in an
undistorted manner; this is the soliton.
It is convenient to shift to a moving frame of reference:
x = x c0 t
hence

=
x
x

Thus,
yt +

t = t ,

= c0 .
t
t
x

3c0
y yx + 61 c0 h20 yx x x = 0 .
2h0

(10.7)

(10.8)

(10.9)

Finally, rescaling position, time, and displacement, and dropping the primes, we arrive at
the KdV equation,
ut + 6uux + uxxx = 0 ,
(10.10)
which is a convenient form.

10.1.1

KdV solitons

We seek a solution to the KdV equation of the form u(x, t) = u(x V t). Then with
x V t, we have x = and t = V when acting on u(x, t) = u(). Thus, we have
V u + 6uu + u = 0 .

(10.11)

V u + 3u2 + u = A ,

(10.12)

Integrating once, we have

where A is a constant. We can integrate once more, obtaining
12 V u2 + u3 + 12 (u )2 = Au + B ,

(10.13)

where now both A and B are constants. We assume that u and all its derivatives vanish in
the limit , which entails A = B = 0. Thus,

du
= u V 2u .
d

(10.14)

u = 21 V sech2 ,

(10.15)

With the substitution

268

CHAPTER 10. SOLITONS

Figure 10.1: Soliton solutions to the KdV equation, with five evenly spaced V values ranging
from V = 2 (blue) to V = 10 (orange). The greater the speed, the narrower the shape.

we find d = 12 V d, hence we have the solution



u(x, t) = 12 V sech2 2V (x V t 0 ) .

(10.16)

Note that the maximum amplitude of the soliton is umax = 21 V , which is proportional to
its velocity V . The KdV equation imposes no limitations on V other than V 0.

10.1.2

Periodic solutions : soliton trains

If we relax the condition A = B = 0, new solutions to the KdV equation arise. Define the
cubic
P (u) = 2u3 V u2 2Au 2B

2(u u1 )(u u2 )(u u3 ) ,

(10.17)

where ui = ui (A, B, V ). We presume that A, B, and V are such that all three roots
u1,2,3 are real and nondegenerate. Without further loss of generality, we may then assume
u1 < u2 < u3 . Then
p
du
= P (u) .
(10.18)
d

Since P (u) < 0 for u2 < u < u3 , we conclude u() must lie within this range. Therefore,
we have

1/2 Z
Zu
2
ds
d
p
0 = p
=
,
(10.19)
u

u
P (s)
1 k2 sin2
3
1
0

u2

where

u u3 (u3 u2 ) sin2

k2

u3 u2
.
u3 u1

(10.20)

269

10.1. THE KORTEWEG-DEVRIES EQUATION

Figure 10.2: The Jacobi elliptic functions sn(, k) (solid) and cn(, k) (dot-dash) versus
/K(k), for k = 0 (blue), k = 12 (green), and k = 0.9 (red).
The solution for u() is then
u() = u3 (u3 u2 ) sn2 (, k) ,
where

u3 u1
0
2
and sn(, k) is the Jacobi elliptic function.
=

10.1.3

(10.21)

(10.22)

Interlude: primer on elliptic functions

We assume 0 k2 1 and we define

Z
(, k) = p
0

d
1 k2 sin2

(10.23)

The sn and cn functions are defined by the relations

sn(, k) = sin
cn(, k) = cos .

(10.24)

Note that sn2 (, k) + cn2 (, k) = 1. One also defines the function dn(, k) from the relation
dn2 (, k) + k2 sn2 (, k) = 1 .

(10.25)

When advances by one period, we have = 2, and therefore = Z, where

Z2
Z= p
0

d
1 k2 sin2

= 4 K(k) ,

(10.26)

270

CHAPTER 10. SOLITONS

Figure 10.3: The cubic function P (u) (left), and the soliton lattice (right) for the case
u1 = 1.5, u2 = 0.5, and u3 = 2.5.
where K(k) is the complete elliptic integral of the first kind. Thus, sn( + Z, k) = sn(, k),
and similarly for the cn function. In fig. 10.2, we sketch the behavior of the elliptic
functions over one quarter of a period. Note that for k = 0 we have sn(, 0) = sin and
cn(, 0) = cos .

10.1.4

The soliton lattice

Getting back to our solution in eqn. 10.21, we see that the solution describes a soliton
lattice with a wavelength

8 K(k)
.
(10.27)
= p
u3 u1

Note that our definition of P (u) entails

V = 2(u1 + u2 + u3 ) .

(10.28)

There is a simple mechanical analogy which merits illumination. Suppose we define

and furthermore E B. Then

W (u) u3 12 V u2 Au ,

(10.29)

 
1 du 2
+ W (u) = E ,
2 d

(10.30)

which takes the form of a one-dimensional Newtonian mechanical system, if we replace

t and interpret u as a velocity. The potential is W (u) and the total energy is E. In
terms of the polynomial P (u), we have P = 2(W E). Accordingly, the motion u() is

271

10.1. THE KORTEWEG-DEVRIES EQUATION

Figure 10.4: The cubic function P (u) (left), and the soliton lattice (right) for the case
u1 = 1.5, u2 = 1.49, and u3 = 2.50.
flattest for the lowest values of u, near u = u2 , which is closest to the local maximum of
the function W (u).
Note that specifying umin = u2 , umax = u3 , and the velocity V specifies all the parameters.
Thus, we have a three parameter family of soliton lattice solutions.

10.1.5

N-soliton solutions to KdV

In 1971, Ryogo Hirota2 showed that exact N -soliton solutions to the KdV equation exist.
Here we discuss the Hirota solution, following the discussion in the book by Whitham.
The KdV equation may be written as


ut + 3u2 + uxx x = 0 ,

(10.31)

which is in the form of the one-dimensional continuity equation ut +jx = 0, where the current
is j = 3u2 + uxx . Let us define u = px . Then our continuity equation reads ptx + jx = 0,
which can be integrated to yield pt + j = C, where C is a constant. Demanding that u and
its derivatives vanish at spatial infinity requires C = 0. Hence, we have
pt + 3p2x + pxxx = 0 .

(10.32)

Now consider the nonlinear transformation

p = 2 (ln F )x =
2

R. Hirota, Phys. Rev. Lett. 27, 1192 (1971).

2Fx
.
F

(10.33)

272

CHAPTER 10. SOLITONS

We then have
pt =

2Fxt 2Fx Ft

F
F2

px =

2Fxx 2Fx2
2
F
F

(10.34)

and
pxx =
pxxx

2Fxxx 6Fx Fxx 4Fx3

+ 3
F
F2
F

2Fxxxx 8Fx Fxxx 6Fx x2 24Fx2 Fxx 12Fx4

=

.
F
F2
F2
F3
F4

(10.35)

When we add up the combination pt + 3p2x + pxxx = 0, we find, remarkably, that the terms
with F 3 and F 4 in the denominator cancel. We are then left with






2
F Ft + Fxxx x Fx Ft + Fxxx + 3 Fxx
Fx Fxxx = 0 .
(10.36)

This equation has the two-parameter family of solutions

F (x, t) = 1 + e(x,t) .

(10.37)

(x, t) = (x b 2 t) ,

(10.38)

where
with and b constants. Note that these solutions are all annihilated by the operator t +x3 ,
and also by the last term in eqn. 10.36 because of the homogeneity of the derivatives.
Converting back to our original field variable u(x, t), we have that these solutions are single
solitons:
2 f
2(F Fxx Fx2 )
(10.39)
=
= 12 2 sech2 ( 21 ) .
u = px =
F2
(1 + f )2
The velocity for these solutions is V = 2 .
If eqn. 10.36 were linear, our job would be done, and we could superpose solutions. We
will meet up with such a felicitous situation when we discuss the Cole-Hopf transformation
for the one-dimensional Burgers equation. But for KdV the situation is significantly more
difficult. We will write
F = 1 + F (1) + F (2) + . . . + F (N ) ,
(10.40)
with
F (1) = f1 + f2 + . . . + fN ,

(10.41)

where
fj (x, t) = ej (x,t)
j (x, t) = j (x 2j t bj ) .
We may then derive a hierarchy of equations, the first two levels of which are


(1)
(1)
Ft + Fxxx
=0
x




(2)
(2)
(1) (1)
(1)
Ft + Fxxx
= 3 Fxx
Fxx Fx(1) Fxxx
.
x

(10.42)

(10.43)

273

10.1. THE KORTEWEG-DEVRIES EQUATION

Lets explore the case N = 2. The equation for F (2) becomes

(2)
(2)

Ft + Fxxx x = 3 1 2 (2 1 )2 f1 f2 ,

with solution

F (2) =

1 2
1 + 2

2

f1 f2 .

(10.44)

(10.45)

Remarkably, this completes the hierarchy for N = 2. Thus,

F = 1 + f1 + f2 +

1 + f1
= det 2
1 2
+ f2
1

1 2
1 + 2

2

2 1 2
1 +2

f1

1 + f2

f1 f2

(10.46)

What Hirota showed, quite amazingly, is that this result generalizes to the N -soliton case,
F = det Q ,
where Q is matrix
Qmn

2 m n
= mn +
f .
m + n m

(10.47)

(10.48)

Note that for N = 2 this agrees with Eqn. 10.46. Thus, N -soliton solutions to the KdV
equation may be written in the form
u(x, t) = 2

2
ln det Q(x, t) .
x2

(10.49)

Here we are free to make a similarity transformation Q S 1 Q S for any nonsingular

e where
matrix S. If we take Smn = en /2 mn , then Q Q,

2 m n p
m /2
+n /2
e
Q=e
Qmn e
= mn +
fm fn
(10.50)
m + n
is symmetric.

Consider the case N = 2. Direct, if tedious, calculations lead to the expression

2
21 f1 f22 + 22 f12 f2
21 f1 + 22 f2 + 2 (1 2 )2 f1 f2 + 1 +2
1
2
.
u=2
i2
h
1 2
2
1 + f1 + f2 + + f1 f2
1

Recall that

(10.51)



fj (x, t) = exp j (xj 2j t bj ) .

(10.52)

Lets consider (x, t) values for which f1 1 is neither large nor small, and investigate what
happens in the limits f2 1 and f2 1. In the former case, we find
u

221 f1
(1 + f1 )2

(f2 1) ,

(10.53)

274

CHAPTER 10. SOLITONS

Figure 10.5: Early and late time configuration of the two soliton solution to the KdV
equation.
which is identical to the single soliton case of eqn. 10.39. In the opposite limit, we have
u

221 g1
(1 + g1 )2

where
g1 =

(f2 1)

1 2
1 + 2

2

f1

(10.54)

(10.55)

But multiplication of fj by a constant C is equivalent to a translation:

C fj (x, t) = fj (x + 1
j ln C , t) fj (x xj , t) .

(10.56)

Thus, depending on whether f2 is large or small, the solution either acquires or does not
acquire a spatial shift x1 , where


 1 + 2 
2
.

(10.57)
ln
xj =
j  1 2 



The function f (x, t) = exp (x 2 t b) is monotonically increasing in x (assuming
> 0). Thus, if at fixed t the spatial coordinate x is such that f 1, this means that the
soliton lies to the right. Conversely, if f 1 the soliton lies to the left. Suppose 1 > 2 ,
in which case soliton #1 is stronger (i.e. greater amplitude) and faster than soliton #2.
The situation is as depicted in figs. 10.5 and 10.6. Starting at early times, the strong
soliton lies to the left of the weak soliton. It moves faster, hence it eventually overtakes the
weak soliton. As the strong soliton passes through the weak one, it is shifted forward, and
the weak soliton is shifted backward. It hardly seems fair that the strong fast soliton gets

10.1. THE KORTEWEG-DEVRIES EQUATION

275

Figure 10.6: Left: Spacetime diagram (x versus t) for the collision of two KdV solitons. The
strong, fast soliton (#1) is shifted forward and the weak slow one (#2) is shifted backward.
The red and blue lines indicate the centers of the two solitons. The yellow shaded circle
is the interaction region where the solution is not simply a sum of the two single soliton
waveforms. Right: t versus x showing the soliton waveforms through the collision. Here,
the velocity of the slow soliton is close to zero. Note that the slow soliton is again shifted
backward. From P. J. Caudrey, Phil. Trans. Roy. Soc. A 28, 1215 (2011).
pushed even further ahead at the expense of the weak slow one, but sometimes life is just
like that3 .

10.1.6

B
acklund transformations

For certain nonlinear PDEs, a given solution may be used as a seed to generate an entire
hierarchy of solutions. This is familiar from the case of Riccati equations, which are nonlinear and nonautonomous ODEs, but for PDEs it is even more special. The general form
of the Backlund transformation (BT) is
u1,t = P (u1 , u0 , u0,t , u0,x )

(10.58)

u1,x = Q(u1 , u0 , u0,t , u0,x ) ,

(10.59)

where u0 (x, t) is the known solution.

A Backlund transformation for the KdV equation was first obtained in 19734 . This provided
a better understanding of the Hirota hierarchy. To proceed, following the discussion in the
book by Scott, we consider the earlier (1968) result of Miura5 , who showed that if v(x, t)
3

See T. Piketty, Capital in the Twenty-First Century (Belknap Press, 2014).

H. D. Wahlquist and F. B. Eastabrook, Phys. Rev. Lett. 31, 1386 (1973).
5
R. M. Miura, J. Math. Phys. 9, 1202 (1968).
4

276

CHAPTER 10. SOLITONS

Figure 10.7: Wrestlings living legend Bob Backlund, in a 1983 match, is subjected to a
devastating camel clutch by the Iron Sheikh. The American Bob Backlund has nothing to
do with the Backlund transformations discussed in the text, which are named for the 19th
century Swedish mathematician Albert Backlund. Note that Bob Backlunds manager has
thrown in the towel (lower right).
satisfies the modified KdV (MKdV) equation6 ,
vt 6v 2 vx + vxxx = 0 ,

(10.60)

u = (v 2 + vx )

(10.61)

then
solves KdV:
ut + 6uux + uxxx = (v 2 + vx )t + 6(v 2 + vx )(v 2 + vx )x (v 2 + vx )xxx




= 2v + x vt 6v 2 vx + vxxx = 0 .

(10.62)

From this result, we have that if

vt 6(v 2 + ) vx + vxxx = 0 ,

(10.63)

u = (v 2 + vx + )

(10.64)

then
solves KdV. The MKdV equation, however, is symmetric under v v, hence
u0 = vx v 2

u1 = +vx v 2

(10.65)

both solve KdV. Now define u0 w0,x and u1 w1,x . Subtracting the above two
equations, we find
u0 u1 = 2vx w0 w1 = 2v .
(10.66)
6
Note that the second term in the MKdV equation is proportional to v 2 vx , as opposed to uux which
appears in KdV.

277

10.1. THE KORTEWEG-DEVRIES EQUATION

Adding the equations instead gives

w0,x + w1,x = 2(v 2 + )
= 12 (w0 w1 )2 + 2 .

(10.67)

Substituting for v = 12 (w0 w1 ) and v 2 + 12 (w0,x + w1,x ) into the MKdV equation, we
have


2
2
(w0 w1 )t 3 w0,x
w1,x
+ (w0 w1 )xxx = 0 .
(10.68)

This last equation is a Backlund transformation (BT), although in a somewhat nonstandard

form, since the RHS of eqn. 10.58, for example, involves only the new solution u1 and
the old solution u0 and its first derivatives. Our equation here involves third derivatives.
However, we can use eqn. 10.67 to express w1,x in terms of w0 , w0,x , and w1 .
Starting from the trivial solution w0 = 0, eqn. 10.67 gives
w1,x = 12 w12 + 2 .

(10.69)

With the choice = 14 2 < 0, we integrate and obtain

w1 (x, t) = tanh

1
2 x

+ (t) ,

(10.70)

were (t) is at this point arbitrary. We fix (t) by invoking eqn. 10.68, which says
2
w1,t = 3w1,x
w1,xxx = 0 .

(10.71)

Invoking w1,x = 21 w12 + and differentiating twice to obtain w1,xxx , we obtain an expression
for the RHS of the above equation. The result is w1,t + 2 w1,x = 0, hence
h
i
w1 (x, t) = tanh 12 (x 2 t b)
h
i
u1 (x, t) = 12 2 sech2 21 (x 2 t b) ,

(10.72)
(10.73)

which recapitulates our earlier result. Of course we would like to do better, so lets try to
insert this solution into the BT and the turn the crank and see what comes out. This is
unfortunately a rather difficult procedure. It becomes tractable if we assume that successive
Backlund transformations commute, which is the case, but which we certainly have not yet
proven. That is, we assume that w12 = w21 , where

w0 1 w1 2 w12

(10.74)

w0 2 w2 1 w21 .

Invoking this result, we find that the Backlund transformation gives

w12 = w21 = w0

4(1 2 )
.
w1 w2

(10.75)

Successive applications of the BT yield Hirotas multiple soliton solutions:

w0 1 w1 2 w12 3 w123 4 .

(10.76)

278

10.2

CHAPTER 10. SOLITONS

Sine-Gordon Model

Consider transverse electromagnetic waves propagating down a superconducting transmission line, shown in fig. 10.8. The transmission line is modeled by a set of inductors,
capacitors, and Josephson junctions such that for a length dx of the transmission line,
the capacitance is dC = C dx, the inductance is dL = L dx, and the critical current is
dI0 = I0 dx. Dividing the differential voltage drop dV and shunt current dI by dx, we
obtain
V
I
= L
x
t

(10.77)

V
I
= C
I0 sin ,
x
t
where is the difference = upper lower in the superconducting phases. The voltage is
related to the rate of change of through the Josephson equation,

2eV
=
,
t
~

(10.78)

2eL

=
I .
x
~

(10.79)

2
1
1 2
2 + 2 sin = 0 ,
2
2
c t
x
J

(10.80)

and therefore
Thus, we arrive at the equation

where c = (L C)1/2 is the Swihart velocity and J = (~/2eLI0 )1/2 is the Josephson length.
We may now rescale lengths by J and times by J /c to arrive at the sine-Gordon equation,
tt xx + sin = 0 .

(10.81)

This equation of motion may be derived from the Lagrangian density

L = 12 2t 21 2x U () .
We then obtain
tt xx =

U
,

(10.82)
(10.83)

and the sine-Gordon equation follows for U () = 1 cos .

Assuming (x, t) = (x V t) we arrive at (1 V 2 ) = U (), and integrating once we
obtain
2
2
1
(10.84)
2 (1 V ) U () = E .
This describes a particle of mass M = 1 V 2 moving in the inverted potential U ().
Assuming V 2 < 1, we may solve for :
2 =

2(E + U ())
,
1V2

(10.85)

279

10.2. SINE-GORDON MODEL

Figure 10.8: A superconducting transmission line is described by a capacitance per unit

length C, an inductance per unit length L, and a critical current per unit length I0 . Based
on fig. 3.4 of A. Scott, Nonlinear Science.
which requires E Umax in order for a solution to exist. For Umax < E < Umin ,
the motion is bounded by turning points. The situation for the sine-Gordon (SG) model is
sketched in fig. 10.9. For the SG model, the turning points are at = cos1 (E + 1),
with 2 < E < 0. We can write
r
E
(10.86)
,
= 2 cos1 .
=
2
This class of solutions describes periodic waves. From

2 d
d

=p
,
2
1V
E + U ()

(10.87)

we have that the spatial period is given by

2
Z

2 (1 V 2 )

d
E + U ()

(10.88)

where [0, ]. If E > Umin , then is always of the same sign, and () is a monotonic
function of . If U () = U ( + 2) is periodic, then the solution is a soliton lattice where
the spatial period of mod 2 is
e=

1V2
2

Z2
p
0

d
.
E + U ()

(10.89)

280

CHAPTER 10. SOLITONS

Figure 10.9: The inverted potential U () = cos 1 for the sine-Gordon problem.
For the sine-Gordon model, with U () = 1 cos , one finds
=

8(1 V 2 ) K

E+2
2

(10.90)

where K(k) is the complete elliptic integral of the first kindd7 and
e=

10.2.1

8 (1 V 2 )
K
E+2

2
E+2

(10.91)

Tachyon solutions

When V 2 > 1, we have

2
1
2 (V

1) 2 + U () = E .

Such solutions are called tachyonic. There are again three possibilities:
E > Umin : no solution.
Umax < E < Umin : periodic () with oscillations about = 0.
E < Umax : tachyon lattice with monotonic ().
It turns out that the tachyon solution is unstable.
7

See the NIST Handbook of Mathematical Functions, ch. 19

(10.92)

281

10.2. SINE-GORDON MODEL

10.2.2

Hamiltonian formulation

The Hamiltonian density is

H = t L ,

(10.93)

L
= t
t

(10.94)

where
=

is the momentum density conjugate to the field . Then

H(, ) = 12 2 + 12 2x + U () .

(10.95)

Note that the total momentum in the field is

Z
Z
Z
P = dx = dx t = V dx x

h
i
= V () () = 2nV ,

(10.96)



where n = () () /2 is the winding number .

10.2.3

Phonons

The Hamiltonian density for the SG system is minimized when U () = 0 everywhere. The
ground states are then classified by an integer n Z, where (x, t) = 2n for ground state
n. Suppose we linearize the SG equation about one of these ground states, writing
(x, t) = 2n + (x, t) ,

(10.97)

and retaining only the first order term in from the nonlinearity. The result is the KleinGordon (KG) equation,
tt xx + = 0 .
(10.98)
This is a linear equation, whose solutions may then be superposed. Fourier transforming
from (x, t) to (k, ), we obtain the equation


2 + k2 + 1 (k, ) = 0 ,
(10.99)
which entails the dispersion relation = (k), where
p
(k) = 1 + k2 .

(10.100)

Thus, the most general solution to the (1 + 1)-dimensional KG equation is

(x, t) =

dk
2




2
2
A(k) eikx ei 1+k t + B(k) eikx ei 1+k t .

(10.101)

282

CHAPTER 10. SOLITONS

For the Helmholtz equation tt xx = 0, the dispersion is (k) = |k|, and the solution may
be written as (x, t) = f (x t) + g(x + t), which is the sum of arbitrary right-moving and
left-moving components. The fact that the Helmholtz equation preserves the shape of the
wave is a consequence of the absence of dispersion, i.e. the phase velocity vp (k) = k is the
same as the group velocity vg (k) =
k . This is not the case for the KG equation, obviously,
since

k
1 + k2

,
vg (k) =
=
,
(10.102)
vp (k) = =
k
k
k
1 + k2
hence vp vg = 1 for KG.

10.2.4

Mechanical realization

The sine-Gordon model can be realized mechanically by a set of pendula elastically coupled.
The kinetic energy T and potential energy U are given by
X
2 2
1
T =
2 m n
n

U=

Xh

1
2

n+1 n

2

+ mg 1 cos n

i

(10.103)

Here is the distance from the hinge to the center-of-mass of the pendulum, and is the
torsional coupling. From the Euler-Lagrange equations we obtain


m2 n = n+1 + n1 2n mg sin n .

(10.104)

Let a be the horizontal spacing between the pendula. Then we can write the above equation
as

n

c2

}|
{
z}|{ z "
 
#
2

1
a
g
n
n
n+1
n1
n = 2
sin n .

m a
a
a

(10.105)

The continuum limit of these coupled ODEs yields the PDE

1
1
tt xx + 2 sin = 0 ,
2
c

(10.106)

which is the sine-Gordon equation, with = (a2 /mg)1/2 .

10.2.5

Kinks and antikinks

Let us return to eqn. 10.84 and this time set E = Umin . With U () = 1 cos , we have
Umin = 0, and thus


d
2
(10.107)
sin 21 .
=
d
1V2

283

10.2. SINE-GORDON MODEL

Figure 10.10: Kink and antikink solutions to the sine-Gordon equation.

This equation may be integrated:

d
d
=
= d ln tan 14 .
1
2
2
sin

1V
2

Thus, the solution is

(x, t) = 4 tan

x V t x0
exp
1V2

(10.108)

(10.109)

where 0 is a constant of integration. This describes either a kink (with d/dx > 0) or
an antikink (with d/dx < 0) propagating with velocity V , instantaneously centered at
x = x0 + V t. Unlike the KdV soliton, the amplitude of the SG soliton is independent of its
velocity. The SG soliton is topological, interpolating between two symmetry-related vacuum
states, namely = 0 and = 2.
Note that the width of the kink and antikink solutions decreases as V increases. This is
a Lorentz contraction, and should have been expected since the SG equation possesses a
Lorentz invariance under transformations
x + vt
x=
1 v2

t + vx
t=
.
1 v2

(10.110)

One then readily finds

t2 x2 = t2 x2 .

(10.111)

The moving soliton solutions may then be obtained by a Lorentz transformation of the
stationary solution,
(x, t) = 4 tan1 e(xx0 ) .

(10.112)

The field itself is a Lorentz scalar, and hence does not change in magnitude under a
Lorentz transformation.

284

10.2.6

CHAPTER 10. SOLITONS

B
acklund transformation for the sine-Gordon system

Recall DAlemberts method of solving the Helmholtz equation, by switching to variables

= 21 (x t)

1
2 (x

x = 21 + 12 t

+ t)

t =

12

1
2 t

(10.113)
.

(10.114)

The DAlembertian operator then becomes

t2 x2 = .

(10.115)

This transforms the Helmholtz equation tt xx = 0 to = 0, with solutions (, ) =

f () + g( ), with f and g arbitrary functions of their arguments. As applied to the SG
equation, we have
= sin .
(10.116)
Suppose we have a solution 0 to this equation. Suppose further that 1 satisfies the pair
of equations,


1 + 0
+ 0,
(10.117)
1, = 2 sin
2


2
1 0
1, = sin
0, .
(10.118)

2
Thus,
1, 0,




1 + 0
= cos
1, + 0,
2




1 + 0
1 0
= 2 cos
sin
2
2

(10.119)

= sin 1 sin 0 .

Thus, if 0, = sin 0 , then 1, = sin 1 as well, and 1 satisfies the SG equation. Eqns.
10.117 and 10.118 constitute a Backlund transformation for the SG system.
Lets give the Backlund crank one turn, starting with the trivial solution 0 = 0. We then
have
1, = 2 sin 12 1

(10.120)

1, = 21 sin 12 1 .

(10.121)

The solution to these equations is easily found by direct integration:

(, ) = 4 tan1 e e / .

(10.122)

In terms of our original independent variables (x, t), we have

+ 1 =

1
2

+ 1 x

1
2

x vt
1 t =
,
1 v2

(10.123)

285

10.2. SINE-GORDON MODEL

where

2 1
v 2
+1

Thus, we generate the kink/antikink solution

1 (x, t) = 4 tan

1+v
=
1v


1/2

x vt
exp
1 v2

(10.124)

(10.125)

As was the case with the KdV system, successive Backlund transformations commute. Thus,

0 1 1 2 12

(10.126)

0 2 2 1 21 ,
with 12 = 21 . This allows one to eliminate the derivatives and write

 



12 0
1 + 2
2 1
tan
=
.
tan
4
1 2
4

(10.127)

We can now create new solutions from individual kink pairs (KK), or kink-antikink pairs
(KK). For KK, taking v1 = v2 = v yields


v sinh(x)
,
(10.128)
KK (x, t) = 4 tan1
cosh(vt)
where is the Lorentz factor,
=

1
.
1 v2

(10.129)

Note that KK (, t) = 2 and KK (0, t) = 0, so there is a phase increase of 2 on each

of the negative and positive half-lines for x, and an overall phase change of +4. For the
KK system, if we take v1 = v2 = v, we obtain the solution


sinh(vt)
1
KK (x, t) = 4 tan
.
(10.130)
v cosh(x)
In this case, analytically continuing to imaginary v with
v=

i
1 2

yields the stationary breather solution,

B (x, t) = 4 tan

1 2

1 2 sin(t)

cosh( 1 2 x)

(10.131)

(10.132)

The breather is a localized solution to the SG system which oscillates in time. By applying
a Lorentz transformation of the spacetime coordinates, one can generate a moving breather
solution as well.

286

CHAPTER 10. SOLITONS

Please note, in contrast to the individual kink/antikink solutions, the solutions KK , KK ,

and B are not functions of a single variable = xV t. Indeed, a given multisoliton solution
may involve components moving at several different velocities. Therefore the total momen

tum P in the field is no longer given by the simple expression P = V () (+) .
However, in cases where the multikink solutions evolve into well-separated solitions, as happens when the individual kink velocities are distinct, the situation simplifies, as we may
consider each isolated soliton as linearly independent. We then have
X
P = 2
ni Vi ,
(10.133)
i

where ni = +1 for kinks and ni = 1 for antikinks.

10.3

Nonlinear Schr
odinger Equation

The Nonlinear Schr

odinger (NLS) equation arises in several physical contexts. One is the
Gross-Pitaevskii action for an interacting bosonic field,


Z Z


~2

S[, ] = dt d x i

U + ,
(10.134)
t
2m


where (x, t) is a complex scalar field. The local interaction U ||2 is taken to be quartic,


(10.135)
U ||2 = 12 g ||4 .
Note that

U ||

|| =

1
2g



2 2
2

.
||
g
2g

(10.136)

Extremizing the action with respect to yields the equation

S
~2 2

+
U + .
=
0
=
i

t
2m

(10.137)

Extremization with respect to yields the complex conjugate equation. In d = 1, we have

it =

~2
xx + U .
2m

(10.138)

We can absorb the chemical potential by making a time-dependent gauge transformation

(x, t) eit (x, t) .

(10.139)

Further rescalings of the field and independent variables yield the generic form
it xx + 2 ||2 = 0 ,

(10.140)

where the + sign pertains for the case g < 0 (attractive interaction), and the sign for
the case g > 0 (repulsive interaction). These cases are known as focusing, or NLS(+), and
defocusing, or NLS(), respectively.


10.3. NONLINEAR SCHRODINGER
EQUATION

10.3.1

287

Amplitude-phase representation

We can decompose the complex scalar into its amplitude and phase:
= A ei .

(10.141)

We then find


t = At + iAt ei


x = Ax + iAx ei

xx = Axx

A2x

(10.142)
i

+ 2iAx x + iAxx e

Multiplying the NLS() equations by ei and taking real and imaginary parts, we obtain
the coupled nonlinear PDEs,


At Axx A2x + 2A3 = 0
(10.143)


At 2Ax x + Axx = 0 .

Note that the second of these equations may be written in the form of a continuity equation,
t + jx = 0 ,

(10.144)

where
= A2

(10.145)

j = 2A2 x .

10.3.2

Phonons

One class of solutions to NLS() is the spatially uniform case

2

0 (x, t) = A0 e2iA0 t ,

(10.146)

with A = A0 and = 2A20 t. Lets linearize about these solutions, writing

A(x, t) = A0 + A(x, t)

(10.147)

(x, t) = 2A20 t + (x, t) .

Our coupled PDEs then yield
4A20 A Axx A0 t = 0

(10.148)

At A0 xx = 0 .

Fourier transforming, we obtain

4A20 k2 iA0
i
A0 k2

!

)
A(k,
)
(k,

=0.

(10.149)

288

CHAPTER 10. SOLITONS

Setting the determinant to zero, we obtain

2 = 4A20 k2 + k4 .

(10.150)

For NLS(), we see that 2 0 for all k, meaning that the initial solution
0 (x, t) is stable.

For NLS(+), however, we see that wavevectors k 2A0 , 2A0 are unstable. This is
known as the Benjamin-Feir instability.

10.3.3

Soliton solutions for NLS(+)

Lets consider moving soliton solutions for NLS(+). We try a two-parameter solution of the
form
A(x, t) = A(x ut)

(10.151)

Axx A 2x + vA x + 2A3 = 0

(10.152)

(x, t) = (x vt) .

This results in the coupled ODEs

A xx + 2Ax x uAx = 0 .

Multiplying the second equation by 2A yields




2x u A2 = 0 =
x

x = 12 u +

P
,
2A2

(10.153)

where P is a constant of integration. Inserting this in the first equation results in

Axx + 2A3 + 14 (2uv u2 )A + 12 (v u)P A1 41 P A3 = 0 .

(10.154)

We may write this as

Axx + W (A) = 0 ,

(10.155)

W (A) = 12 A4 + 81 (2uv u2 )A2 + 12 (v u)P ln A + 81 P A2

(10.156)

where
plays the role of a potential. We can integrate eqn. 10.155 to yield
1 2
2 Ax

+ W (A) = E ,

(10.157)

where E is second constant of integration. This may be analyzed as a one-dimensional

mechanics problem.
The simplest case to consider is P = 0, in which case
W (A) =

1
2

A2 + 12 uv 14 u2 A2 .

(10.158)

If u2 < 2uv, then W (A) is everywhere nonnegative and convex, with a single global minimum at A = 0, where W (0) = 0. The analog mechanics problem tells us that A will
oscillate between A = 0 and A = A , where W (A ) = E > 0. There are no solutions with


10.3. NONLINEAR SCHRODINGER
EQUATION

289

E < 0. If u2 > 2uv, then W (A) has a double well shape8 . If E > 0 then the oscillations
are still between A = 0 and A = A , but if E < 0 then there are two positive solutions to
W (A) = E. In this latter case, we may write






F (A) 2 E W (A) = A2 A20 A21 A2 ,
(10.159)
where A0 < A1 and

E = 21 A20 A21
1 2
4u

1
2 uv

A20

A21

(10.160)
.

The amplitude oscillations are now between A = A0 and A = A1 . The solution is given in
terms of Jacobi elliptic functions:
i
h


(10.161)
(x, t) = A1 exp 2i u (x vt) dn A1 (x ut 0 ) , k ,
where

k2 = 1

A20
.
A21

The simplest case is E = 0, for which A0 = 0. We then obtain

h
i


(x, t) = A exp 2i u (x vt) sech A (x ut 0 ) ,

(10.162)

(10.163)

where 4A 2 = u2 2uv. When u = 0 we obtain the stationary breather solution, for which
the entire function (x, t) oscillates uniformly.

10.3.4

Dark solitons for NLS()

The small oscillations of NLS() are stable, as we found in our phonon calculation. It is
therefore perhaps surprising to note that this system also supports solitons. We write
(x, t) =

0 e2i0 t ei Z(x, t) ,

(10.164)

with an arbitrary constant. This yields

iZt = Zxx + 20 1 |Z|2 Z .

(10.165)

We then write Z = X + iY , yielding

Xt = Yxx + 20 1 X 2 Y 2 Y


Yt = Xxx + 20 1 X 2 Y 2 X .

(10.166)

8
Although we have considered A > 0 to be an amplitude, there is nothing wrong with allowing A < 0.
When A(t) crosses A = 0, the phase (t) jumps by .

290

CHAPTER 10. SOLITONS

We try Y = Y0 , a constant, and set X(x, t) = X(x V t). Then

V X = 20 Y0 1 Y02 X 2

0 = X + 20 1 Y02 X 2 X

Thus,

X =
from which it follows that

20 Y0
1 Y02 X 2
V

40 Y0
XX
V


4 Y 2
82 Y 2
= 0 0 1 Y02 X 2 X = 0 2 0 X .
V
V

(10.167)

(10.168)

X =

(10.169)

Thus, in order to have a solution, we must have

V = 2 0 Y0 .

(10.170)

We now write = x V t, in which case, from Eqn. 10.168,

0 d =

dX
.
1 Y02 X 2

(10.171)

p
e and integrates
From this point, the derivation is elementary. One writes X = 1 Y02 X,
to obtain
q



2
e
X() = tanh
(10.172)
0 1 Y0 ( 0 ) .
We simplify the notation by writing Y0 = sin . Then






i 2i0 t
tanh 0 cos x V t 0 cos + i sin ,
(x, t) = 0 e e

(10.173)

where is a constant and V = 2 0 sin . The density = ||2 is then given by






(x, t) = 0 1 sech2 0 cos x V t 0
.
(10.174)

This is called a dark soliton because the density (x, t) is minimized at the center of
the soliton, where = 0 sin2 , which is smaller than the asymptotic |x| value of
(, t) = 0 .

Chapter 11

Shock Waves
Here we shall follow closely the pellucid discussion in chapter 2 of the book by G. Whitham,
beginning with the simplest possible PDE,
t + c0 x = 0 .

(11.1)

The solution to this equation is an arbitrary right-moving wave (assuming c0 > 0), with
profile
(x, t) = f (x c0 t) ,
(11.2)
where the initial conditions on eqn. 11.1 are (x, t = 0) = f (x). Nothing to see here, so
move along.

11.1

Nonlinear Continuity Equation

The simplest nonlinear PDE is a generalization of eqn. 11.1,

t + c() x = 0 .

(11.3)

This equation arises in a number of contexts. One example comes from the theory of
vehicular traffic flow along a single lane roadway. Starting from the continuity equation,
t + jx = 0 ,

(11.4)

one posits a constitutive relation j = j(), in which case c() = j (). If the individual
vehicles move with a velocity v = v(), then
j() = v()

c() = v() + v () .

(11.5)

It is natural to assume a form v() = c0 (1 a), so that at low densities one has v c0 ,
with v() decreasing monotonically to v = 0 at a critical density = a1 , presumably
corresponding to bumper-to-bumper traffic. The current j() then takes the form of an
291

292

CHAPTER 11. SHOCK WAVES

inverted parabola. Note the difference between the individual vehicle velocity v() and
what turns out to be the group velocity of a traffic
For v() = c0 (1 a), one

 wave, c().
has c() = c0 (1 2a), which is negative for 21 a1 , a1 . For vehicular traffic, we have
c () = j () < 0 but in general j() and thus c() can be taken to be arbitrary.
Another example comes from the study of chromatography, which refers to the spatial
separation of components in a mixture which is forced to flow through an immobile absorbing
bed. Let (x, t) denote the density of the desired component in the fluid phase and n(x, t)
be its density in the solid phase. Then continuity requires
n t + t + V x = 0 ,

(11.6)

where V is the velocity of the flow, which is assumed constant. The net rate at which the
component is deposited from the fluid onto the solid is given by an equation of the form
nt = F (n, ) .

(11.7)

In equilibrium, we then have F (n, ) = 0, which may in principle be inverted to yield

n = neq (). If we assume that the local deposition processes
run to equilibrium on fast time


scales, then we may substitute n(x, t) neq (x, t) into eqn. 11.6 and obtain
t + c() x = 0 ,

c() =

V
.
1 + neq ()

(11.8)

We solve eqn. 11.3 using the method of characteristics. Suppose we have the solution
= (x, t). Consider then the family of curves obeying the ODE


dx
= c (x, t) .
dt

(11.9)

This is a family of curves, rather than a single curve, because it is parameterized by the
initial condition x(0) . Now along any one of these curves we must have
d
dx

=
+
=
+ c()
=0.
dt
t
x dt
t
x

(11.10)

Thus, (x, t) is a constant along each of these curves, which are called characteristics. For
eqn. 11.3, the family of characteristics is a set of straight lines1 ,
x (t) = + c() t .

(11.11)

The initial conditions for the function (x, t) are

(x = , t = 0) = f () ,
1

(11.12)

The existence of straight line characteristics is a special feature of the equation t +c() x = 0. For more
general hyperbolic first order PDEs to which the method of characteristics may be applied, the characteristics
are curves. See the discussion in the Appendix.

293

11.1. NONLINEAR CONTINUITY EQUATION

where f () is arbitrary. Thus, in the (x, t) plane, if the characteristic
curve x(t) intersects
the line t = 0 at x = , then its slope is constant and equal to c f () . We then define


g() c f () .
(11.13)

This is a known function, computed from c() and f () = (x = , t = 0). The equation of
the characteristic x (t) is then
x (t) = + g() t .

(11.14)

Do not confuse the subscript in x (t) for a derivative!

To find (x, t), we follow this prescription:
(i) Given any point in the (x, t) plane, we find the characteristic x (t) on which it lies.
This means we invert the equation x = + g() t to find (x, t).


(ii) The value of (x, t) is then = f (x, t) .

(iii) This procedure yields a unique value for (x, t) provided the characteristics do not
cross, i.e. provided that there is a unique such that x = + g() t. If the characteristics do cross, then (x, t) is either multi-valued, or else the method has otherwise
broken down. As we shall see, the crossing of characteristics, under the conditions
of single-valuedness for (x, t), means that a shock has developed, and that (x, t) is
discontinuous.
We can verify that this procedure yields a solution to the original PDE of eqn. 11.3 in the
following manner. Suppose we invert
x = + g() t

= (x, t) .

(11.15)

t = f () t

(11.16)

We then have


(x, t) = f (x, t)

To find t and x , we invoke x = + g() t, hence

x = f () x



+ g() t x = t + t g () t + g()
t


+ g() t x = x + x g () t 1 ,
0=
x
0=

(11.17)
(11.18)

from which we conclude

f () g()
1 + g () t

(11.19)

f ()
.
1 + g () t

(11.20)

t =
x =

294

CHAPTER 11. SHOCK WAVES

Figure 11.1: Forward and backward breaking waves for the nonlinear continuity equation
t + c() x = 0, with c() = 1 + (top panels) and c() = 2 (bottom panels). The
.
initial conditions are (x, t = 0) = 1/(1 + x2 ), corresponding to a break time of tB = 316
3
Successive (x, t) curves are plotted for t = 0 (thick blue), t = 12 tB (dark blue), t = tB (dark
green), t = 32 tB (orange), and t = 2tB (dark red).

Thus, t + c() x = 0, since c() = g().

As any wave disturbance propagates, different values of propagate with their own velocities. Thus, the solution (x, t) can be constructed by splitting the curve (x, t = 0) into
level sets of constant , and then shifting each such set by a distance c() t. For c() = c0 ,
the entire curve is shifted uniformly. When c() varies, different level sets are shifted by
different amounts.
We see that x diverges when 1 + g () t = 0. At this time, the wave is said to break . The
break time tB is defined to be the smallest value of t for which x = anywhere. Thus,


1
1
.
(11.21)

tB = min

g ()
g (B )

g ()<0

Breaking can only occur when g () < 0, and differentiating g() = c f () , we have that
g () = c (f ) f (). We then conclude
c < 0

need f > 0 to break

c > 0

need f < 0 to break .

Thus, if (x = , t = 0) = f () has a hump profile, then the wave breaks forward (i.e. in the
direction of its motion) if c > 0 and backward (i.e. opposite to the direction of its motion)

295

11.1. NONLINEAR CONTINUITY EQUATION

Figure 11.2: Crossing of characteristics of the nonlinear continuity equation t +c() x = 0,

with c() = 1 + and (x, t = 0) = 1/(1 + x2 ). Within the green hatched region of the
(x, t) plane, the characteristics cross, and the function (x, t) is apparently multivalued.
if c < 0. In fig. 11.1 we sketch the breaking of a wave with (x, t = 0) = 1/(1 + x2 ) for
the cases c = 1 + and c = 2 . Note that it is possible for different regions of a wave to
break at different times, if, say, it has multiple humps.
Wave breaking occurs when neighboring characteristic cross. We can see this by comparing
two neighboring characteristics,
x (t) = + g() t

(11.22)

x+ (t) = + + g( + ) t

= + g() t + 1 + g () t + . . . .

(11.23)

For these characteristics to cross, we demand

x (t) = x+ (t)

t=

1
g ()

(11.24)

Usually, in most physical settings, the function (x, t) is single-valued. In such cases, when

296

CHAPTER 11. SHOCK WAVES

Figure 11.3: Crossing of characteristics of the nonlinear continuity equation t +c() x = 0,


2
with c() = 1 + and (x, t = 0) = x/(1 + x2 ) . The wave now breaks in two places and
is multivalued in both hatched regions. The left hump is the first to break.

the wave breaks, the multivalued solution ceases to be applicable2 . Generally speaking,
this means that some important physics has been left out. For example, if we neglect
viscosity and thermal conductivity , then the equations of gas dynamics have breaking
wave solutions similar to those just discussed. When the gradients are steep just before
breaking the effects of and are no longer negligible, even if these parameters are small.
This is because these parameters enter into the coefficients of higher derivative terms in
the governing PDEs, and even if they are small their effect is magnified in the presence
of steep gradients. In mathematical parlance, they constitute singular perturbations. The
shock wave is then a thin region in which and are crucially important, and the flow
changes rapidly throughout this region. If one is not interested in this small scale physics,
the shock region can be approximated as being infinitely thin, i.e. as a discontinuity in the
inviscid limit of the theory. What remains is a set of shock conditions which govern the
discontinuities of various quantities across the shocks.

2
This is even true for water waves, where one might think that a multivalued height function h(x, t) is
physically possible.

297

11.2. SHOCKS

Figure 11.4: Current conservation in the shock frame yields the shock velocity, vs = j/.

11.2

Shocks

We now show that a solution to eqn. 11.3 exists which is single valued for almost all (x, t),
i.e. everywhere with the exception of a set of zero measure, but which has a discontinuity
along a curve x = xs (t). This discontinuity is the shock wave.
The velocity of the shock is determined by mass conservation, and is most easily obtained in
the frame of the shock. The situation is as depicted in fig. 11.4. If the density and current
are ( , j ) to the left of the shock and (+ , j+ ) to the right of the shock, and if the shock
moves with velocity vs , then making a Galilean transformation to the frame of the shock,
the densities do not change but the currents transform as j j = j v. Thus, in the
frame where the shock is stationary, the current on the left and right are j = j vs .
Current conservation then requires
vs =

j+ j
j
=
.
+

(11.25)

The special case of quadratic j() bears mention. Suppose

j() = 2 + + .

(11.26)

vs = (+ + ) +


= 21 c+ + c .

(11.27)

Then c = 2 + and

So for quadratic j(), the shock velocity is simply the average of the flow velocity on either
side of the shock.
Consider, for example, a model with j() = 2(1 ), for which c() = 2 4. Consider
3
3
+ 18 (), so initially = 1 = 16
for
an initial condition (x = , t = 0) = f () = 16
5
x < 0 and = 2 = 16 for x > 0. The lower density part moves faster, so in order to
avoid multiple-valuedness, a shock must propagate. We find c = 45 and c+ = 34 . The shock
velocity is then vs = 1. Ths situation is depicted in fig. 11.5.

298

CHAPTER 11. SHOCK WAVES

Figure 11.5: A resulting shock wave arising from c = 54 and c+ = 43 . With no shock fitting,
there is a region of (x, t) where the characteristics cross, shown as the hatched region on
the left. With the shock, the solution remains single valued. A quadratic behavior of j()
is assumed, leading to vs = 12 (c+ + c ) = 1.

11.3

Internal Shock Structure

At this point, our model of a shock is a discontinuity which propagates with a finite velocity.
This may be less problematic than a multivalued solution, but it is nevertheless unphysical.
We should at least understand how the discontinuity is resolved in a more complete model.
To this end, consider a model where
j = J(, x ) = J() x .

(11.28)

The J() term contains a nonlinearity which leads to steepening and broadening of regions
dc
dc
> 0 and dx
< 0, respectively. The second term, x , is due to diffusion, and
where dx
recapitulates Ficks law , which says that a diffusion current flows in such a way as to reduce
gradients. The continuity equation then reads
t + c() x = xx ,

(11.29)

with c() = J (). Even if is small, its importance is enhanced in regions where |x | is
large, and indeed x dominates over J() in such regions. Elsewhere, if is small, it
may be neglected, or treated perturbatively.
As we did in our study of front propagation, we seek a solution of the form
(x, t) = () (x vs t)

= x vs t .

(11.30)

Thus, t = vs x and x = , leading to

vs + c() = .

(11.31)

J() vs + A = ,

(11.32)

Integrating once, we have

299

11.3. INTERNAL SHOCK STRUCTURE

where A is a constant. Integrating a second time, we have

Z
0 =
0

d
.
J( ) vs + A

(11.33)

Suppose interpolates between the values 1 and 2 . Then we must have

J(1 ) vs 1 + A = 0

J(2 ) vs 2 + A = 0 ,

(11.34)
(11.35)

which in turn requires

J2 J1
,
(11.36)
2 1
= J(1,2 ), exactly as before! We also conclude that the constant A must be
vs =

where J1,2

A=

11.3.1

1 J2 2 J1
.
2 1

(11.37)

Quadratic J()

For the special case where J() is quadratic, with J() = 2 + + , we may write
J() vs + A = ( 2 )( 1 ) .

(11.38)

We then have vs = (1 + 2 ) + , as well as A = 1 2 . The moving front solution

then obeys


2

d
=
d ln
,
(11.39)
d =
( 2 )( 1 )
(2 1 )
1
which is integrated to yield
h

i
2 + 1 exp (2 1 ) x vs t /
h
(x, t) =
(11.40)

i .
1 + exp (2 1 ) x vs t /
We consider the case > 0 and 1 < 2 . Then (, t) = 1,2 . Note that
(
1 if x vs t
(x, t) =
2 if x vs t ,

(11.41)

where

(11.42)
(2 1 )
is the thickness of the shock region. In the limit 0, the shock is discontinuous. All that
remains is the shock condition,


vs = (1 + 2 ) + = 12 c1 + c2 .
(11.43)
=

We stress that we have limited our attention here to the case where J() is quadratic. It
is worth remarking that for weak shocks where = + is small, we can expand
J()


1
1
2
(
+

),
in
which
case
we
find
v
=
(c
+
c
)
+
O
()
.
about the average 2 +

2 +

300

11.4

CHAPTER 11. SHOCK WAVES

Shock Fitting

When we neglect diffusion currents, we have j = J. We now consider how to fit discontinuous shocks satisfying
J J
(11.44)
vs = +
+
into the continuous solution of eqn. 11.3, which are described by
x = + g() t

(11.45)

= f () ,

(11.46)

with g() = c f () , such that the multivalued parts of the continuous solution are eliminated and replaced with the shock discontinuity. The guiding principle here is number
conservation:
Z
d
dx (x, t) = 0 .
(11.47)
dt

Well first learn how do fit shocks when J() is quadratic, with J() = 2 + + . Well
assume > 0 for the sake of definiteness.

11.4.1

An Important Caveat

Lets multiply the continuity equation t + c() x = 0 by c (). Thus results in

ct + c cx = 0 .

(11.48)

If we define q = 12 c2 , then this takes the form of a continuity equation:

ct + qx = 0 .

(11.49)

Now consider a shock wave. Invoking eqn. 11.25, we would find, mutatis mutandis, a shock
velocity
q q
us = +
(11.50)
= 12 (c+ + c ) .
c+ c
This agrees with the velocity vs = j/ only when j() is quadratic. Something is wrong
there cannot be two velocities for the same shock.
The problem is that eqn. 11.48 is not valid across the shock and cannot be used to determine
the shock velocity. There is no conservation law for c as there is for . One way we can
appreciate the difference is to add diffusion into the mix. Multiplying eqn. 11.29 by c (),
and invoking cxx = c () xx + c () 2x , we obtain
ct + c cx = cxx c () 2x .

(11.51)

301

11.4. SHOCK FITTING

We now see explicitly how nonzero c () leads to a different term on the RHS. When
c () = 0, the above equation is universal, independent of the coefficients in the quadratic
J(), and is known as Burgers equation,
ct + c cx = cxx .

(11.52)

Later on we shall see how this nonlinear PDE may be linearized, and how we can explicitly
solve for shock behavior, including the merging of shocks.

11.4.2

Recipe for shock fitting (J () = 0)

Number conservation means that when we replace the multivalued solution by the discontinuous one, the area under the curve must remain the same. If J() is quadratic, then we
can base our analysis on the equation ct + c cx = 0, since it gives the correct shock velocity
vs = 21 (c+ + c ). We then may then follow the following rules:


(i) Sketch g() = c f () .

(ii) Draw a straight line connecting two points on this curve at and + which obeys
the equal area law, i.e.


 Z+
1
d g() .
2 (+ ) g(+ ) + g( ) =

(11.53)

(iii) This line evolves into the shock front after a time t such that
xs (t) = + g( ) t = + + g(+ ) t .
Thus,
t=

+
.
g(+ ) g( )

(11.54)

(11.55)

Alternatively, we can fix t and solve for . See fig. 11.6 for a graphical description.
(iv) The position of the shock at this time is x = xs (t). The strength of the shock is
c = g( ) g(+ ). Since J() = 2 + + , we have c() = 2 + and hence
the density discontinuity at the shock is = c/2.
(v) The break time, when the shock first forms, is given by finding the steepest chord
satisfying the equal area law. Such a chord is tangent to g() and hence corresponds
to zero net area. The break time is


1
1
.
(11.56)

tB = min

g ()
g(B )

g ()>0

(vi) If g() = g(), the shock strength vanishes as t . If g() > g(+) then
asymptotically the shock strength approaches g = g() g(+).

302

CHAPTER 11. SHOCK WAVES

Figure 11.6: Shock fitting for quadratic J().

11.4.3

Example problem

Suppose the c() and (x, t = 0) are such that the initial profile for c(x, t = 0) is
 x 


|x| ,
c(x, 0) = c0 cos
2

(11.57)

where (s) is the step function, which vanishes identically for negative values of its argument. Thus, c(x, 0) = 0 for |x| .
(a) Find the time tB at which the wave breaks and a shock front develops. Find the position
of the shock xs (tB ) at the moment it forms.
Solution : Breaking first occurs at time
tB = min
x

c (x, 0)

(11.58)

Thus, we look for the maximum negative slope in g(x) c(x, 0), which occurs at x = ,
where c (, 0) = c0 /2. Therefore,
tB =

2
c0

xB = .

(11.59)

(b) Use the shock-fitting equations to derive (t).

Solution : The shock fitting equations are


 Z+
1
d g()
2 (+ ) g(+ ) + g( ) =

(11.60)

303

11.4. SHOCK FITTING

Figure 11.7: Top : crossing characteristics (purple hatched region) in the absence of shock
fitting. Bottom : characteristics in the presence of the shock.
and
t=
Clearly + > , hence g(+ ) = 0 and

+
.
g( ) g(+ )

(

)
Z+
Z/2
2
2 c0

d g() = c0
dz cos z =
1 sin
.

(11.61)

(11.62)

/2

Thus, the first shock fitting equation yields

(


)


2 c0

1
1 sin
.
=
2 (+ ) c0 cos
2

(11.63)

The second shock fitting equation gives

t=

c0 cos

2

Eliminating + , we obtain the relation




4
sin
1 .
=
2
c0 t

(11.64)

(11.65)

304

CHAPTER 11. SHOCK WAVES

Figure 11.8: Evolution of c(x, t) for a series of time values.

Thus,


4
2
1
sin
1
(t) =

c0 t
4 1 sin( /2)

cos( /2)
(
)
r


4
c
t
2
0
sin1
1 +2
1
,
=

c0 t
2

(11.66)

+ (t) = +

(11.67)

where t tB = 2/c0 .
(c) Find the shock motion xs (t).
Solution : The shock position is
xs (t) = + g( ) t


2
4
1 2
=
sin
1 +
1 ,

(11.68)

where = t/tB = c0 t/2, and 1.

(d) Sketch the characteristics for the multivalued solution with no shock fitting, identifying
the region in (x, t) where characteristics cross. Then sketch the characteristics for the
discontinuous shock solution.
Solution : See fig. 11.7.
(e) Find the shock discontinuity c(t).

305

11.5. LONG-TIME BEHAVIOR OF SHOCKS

Solution : The shock discontinuity is



c(t) = g g + = c0 cos
2
s



2
8 c0
1
=
.
1
= 2c0
t
c0 t

(11.69)

(f ) Find the shock velocity vs (t).

Solution : The shock wave velocity is
h



i
cs (t) = 21 g + g + =
=

2 c0
t

2
1
c0 t

1
2

c(t)

(11.70)

(g) Sketch the evolution of the wave, showing the breaking of the wave at t = tB and the
subsequent evolution of the shock front.
Solution : A sketch is provided in Fig. 11.8.

11.5

Long-time Behavior of Shocks

Starting with an initial profile (x, t), almost all the original details are lost in the t
limit. What remains is a set of propagating triangular waves, where only certain gross
features of the original shape, such as its area, are preserved.

11.5.1

Fate of a hump

The late time profile of c(x, t) in fig. 11.8 is that of a triangular wave. This is a general
result. Following Whitham, we consider the late time evolution of a hump profile g(). We
assume g() = c0 for || > L. Shock fitting requires
1
2

Z+


g(+ ) + g( ) 2c0 (+ ) = d g() c0 .
i

(11.71)

Eventually the point + must pass x = L, in which case g(+ ) = c0 . Then

1
2

ZL


g(+ ) c0 (+ ) = d g() c0
i

(11.72)

306

CHAPTER 11. SHOCK WAVES

Figure 11.9: Initial and late time configurations for a hump profile. For late times, the
profile is triangular, and all the details of the initial shape are lost, save for the area A.
and therefore
t=

+
.
g( ) c0

(11.73)

Using this equation to eliminate + , we have

1
2

ZL


t = d g() c0 .

(11.74)

ZL


t d g() c0 A ,

(11.75)

g( ) c0

2

As t we must have L, hence

1
2

g( ) c0

2

where A is the area under the hump to the line c = c0 . Thus,

r
2A
g( ) c0
,
t

(11.76)

and the late time motion of the shock is given by

(11.77)
xs (t) = L + c0 t + 2At
r
A
.
(11.78)
vs (t) = c0 +
2t
p
The shock strength is c = g( ) c0 = 2A/t. Behind the shock, we have c = g() and
x = + g() t, hence
c=

x+L
t

for

L + c0 t < x < L + c0 t +

2At .

(11.79)

As t , the details of the original profile c(x, 0) are lost, and all that remains conserved
is the area A. Both shock velocity and the shock strength decrease as t1/2 at long times,
with vs (t) c0 and c(t) 0.

307

11.5. LONG-TIME BEHAVIOR OF SHOCKS

Figure 11.10: Top panels : An N-wave, showing initial (left) and late time (right) profiles.
As the N-wave propagates, the areas A and B are preserved. Bottom panels : A P-wave.
The area D eventually decreases to zero as the shock amplitude dissipates.

11.5.2

N-wave and P-wave

Consider the initial profile in the top left panel of fig. 11.10. Now there are two propagating shocks, since there are two compression regions where g () < 0. As t , we
have ( , + )A (0 , ) for the A shock, and ( , + )B ( , 0) for the B shock.
Asymptotically, the shock strength




+
c(t) c x
(11.80)
s (t), t c xs (t), t
for the two shocks is given by
A

xs (t) c0 t +
xBs (t) c0 t

2At

2Bt ,

2A
t
r
2B
,
cB
t

cA +

(11.81)
(11.82)

where A and B are the areas associated with the two features This feature is called an
N-wave, for its N (or inverted N) shape.
The initial and late stages of a periodic wave, where g( + ) = g(), are shown in the right
panels of fig. 11.10. In the t limit, we evidently have + = , the wavelength.
Asymptotically the shock strength is given by
c(t) g( ) g(+ ) =

= ,
t
t

(11.83)

308

CHAPTER 11. SHOCK WAVES

Figure 11.11: Merging of two shocks. The shocks initially propagate independently (upper
left), and then merge and propagate as a single shock (upper right). Bottom : characteristics
for the merging shocks.
where we have invoked eqn. 11.55. In this limit, the shock train travels with constant
velocity c0 , which is the spatial average of c(x, 0) over one wavelength:
1
c0 =

d g() .

(11.84)

11.6

Shock Merging

It is possible for several shock waves to develop, and in general these shocks form at different times, have different strengths, and propagate with different velocities. Under such
circumstances, it is quite possible that one shock overtakes another. These two shocks then
merge and propagate on as a single shock. The situation is depicted in fig. 11.11. We label
the shocks by A and B when they are distinct, and the late time single shock by C. We
must have




(11.85)
vsA = 12 g +A + 21 g A

1


B
B
B
1
(11.86)
vs = 2 g + + 2 g .

11.7. SHOCK FITTING FOR GENERAL J()

309

The merging condition requires

+A = B

(11.87)

as well as
+C = +B

C = A .

(11.88)

The merge occurs at time t, where

t=

+

=
.
g() g(+ )
g( ) g()

(11.89)

Thus, the slopes of the A and B shock construction lines are equal when they merge.

11.7

Shock Fitting for General J()

When J() is quadratic, we may analyze the equation ct + c cx , as it is valid across any
shocks in that it yields the correct shock velocity. If J () 6= 0, this is no longer the case,
and we must base our analysis on the original equation t + c() x = 0.
The coordinate transformation
(x, c) (x + ct , c)

(11.90)

preserves areas in the (x, c) plane and also maps lines to lines. However, while
(x, ) (x + c() t , )

(11.91)

does preserve areas in the (x, ) plane, it does not map lines to lines. Thus, the preimage of the shock front in the (x, ) plane is not a simple straight line, and our equal area
construction fails. Still, we can make progress. We once again follow Whitham, 2.9.
Let x(, t) be the inverse of (x, t), with () x(, t = 0). Then
x(, t) = () + c() t .

(11.92)

Note that (x, t) is in general multi-valued. We still have that the shock solution covers the
same area as the multivalued solution (x, t). Let denote the value of just to the right
(+) or left () of the shock. For purposes of illustration, we assume c () > 0, which means
that x < 0 is required for breaking, although the method works equally well for c () < 0.
Assuming a hump-like profile, we then have > + , with the shock breaking to the right.
Area conservation requires
Z
Z


d x(, t) = d () + c() t = ( + ) xs (t) .

(11.93)

310

CHAPTER 11. SHOCK WAVES

Since c() = J (), the above equation may be written as

Z
(J+ J ) t (+ ) = d ()
+

Z
= + + d () .

(11.94)

Now the shock position xs (t) is given by

xs = + c t = + + c+ t ,

(11.95)

hence
h

i c c Z+
+

d () .
(J+ + c+ ) (J c ) =
+

(11.96)

This is a useful result because J , , and c are all functions of , hence what we have
here is a relation between + and . When J() is quadratic, this reduces to our
p earlier

result in eqn. 11.53. For a hump, we still have xs c0 t + 2At and c c0 2A/t as
before, with
Z



A = c (0 ) d () 0 .
(11.97)

11.8

Sources

Consider the continuity equation in the presence of a source term,

t + c x = ,

(11.98)

where c = c(x, t, ) and = (x, t, ). Note that we are allowing for more than just
c = c() here. According to the discussion in the Appendix, the characteristic obey the
coupled ODEs3 ,
d
= (x, t, )
dt
d
= c(x, t, ) .
dt

(11.99)
(11.100)

In general, the characteristics no longer are straight lines.

3

We skip the step where we write dt/ds = 1 since this is immediately integrated to yield s = t.

311

11.8. SOURCES

11.8.1

Examples

Whitham analyzes the equation

ct + c cx = c ,

(11.101)

so that the characteristics obey

dc
= c
dt

dx
=c.
dt

(11.102)

The solution is
c (t) = et g()
x (t) = +

(11.103)

1
1 et g() ,

(11.104)

where = x (0) labels the characteristics. Clearly x (t) is not a straight line. Neighboring
characteristics will cross at time t if
x (t)

=1+

Thus, the break time is

tB = min

t >0
B

"

1
1 et g () = 0 .


#

1
.
ln 1 +

g ()

(11.105)

(11.106)

This requires g () < in order for wave breaking to occur.

For another example, consider
ct + c cx = c2 ,

(11.107)

so that the characteristics obey

dc
= c2
dt

dx
=c.
dt

(11.108)

The solution is now

c (t) =

g()
1 + g() t

x (t) = +

11.8.2

1
ln 1 + g() t .

(11.109)
(11.110)

Moving sources

Consider a source moving with velocity u. We then have

ct + c cx = (x ut) ,

(11.111)

312

CHAPTER 11. SHOCK WAVES

where u is a constant. We seek a moving wave solution c = c() = c(x ut). This leads
immediately to the ODE
(c u) c = () .
(11.112)
This may be integrated to yield
1
2 (u

c )

1
2 (u

Z
c) = d ( ) .
2

(11.113)

Consider the supersonic case where u > c. Then we have a smooth solution,
"

#1/2
Z

c() = u (u c ) 2 d ( )
,
2

(11.114)

provided that the term inside the large rectangular brackets is positive. This is always the
case for < 0. For > 0 we must require
v
u Z
u
u c > u
(11.115)
t 2 d ( )

for all . If () is monotonic, the lower limit on the above integral may be extended to
. Thus, if the source strength is sufficiently small, no shocks are generated. When the
above equation is satisfied as an equality, a shock develops, and transients from the initial
conditions overtake the wave. A complete solution of the problem then requires a detailed
analysis of the transients. What is surprising here is that a supersonic source need not
produce a shock wave, if the source itself is sufficiently weak.

11.9

Burgers Equation

The simplest equation describing both nonlinear wave propagation and diffusion equation
is the one-dimensional Burgers equation,
ct + c cx = cxx .

(11.116)

As weve seen, this follows from the continuity equation t + jx when j = J() x , with
c = J () and c () = 0.
We have already obtained, in 11.3.1, a solution to Burgers equation in the form of a
propagating front. However, we can do much better than this; we can find all the solutions
to the one-dimensional Burgers equation. The trick is to employ a nonlinear transformation
of the field c(x, t), known as the Cole-Hopf transformation, which linearizes the PDE. Once
again, we follow the exceptionally clear discussion in the book by Whitham (ch. 4).

313

11.9. BURGERS EQUATION

The Cole-Hopf transformation is defined as follows:

x
=

c 2

2 ln .

(11.117)

Plugging into Burgers equation, one finds that (x, t) satisfies the linear diffusion equation,
t = xx .

(11.118)

Isnt that just about the coolest thing youve ever heard?
Suppose the initial conditions on (x, t) are
(x, 0) = (x) .

(11.119)

We can then solve the diffusion equation 11.118 by Laplace transform. The result is
1
(x, t) =
4t

Z
2
dx e(xx ) /4t (x ) .

(11.120)

Thus, if c(x, t = 0) = g(x), then the solution for subsequent times is

R

dx (x x ) eH(x,x ,t)/2

(11.121)

Zx
(x x )2
.
H(x, x , t) = dx g(x ) +
2t

(11.122)

c(x, t) =

dx

where

eH(x,x ,t)/2

11.9.1

The limit 0

In the limit 0, the integrals in the numerator and denominator of eqn. 11.121 may be
computed via the method of steepest descents. This means that extremize H(x, x , t) with
respect to x , which entails solving
x x
H

=
g(x
)

.
x
t

(11.123)

Let = (x, t) be a solution to this equation for x , so that

x = + g() t .

(11.124)

We now expand about x = , writing x = + s, in which case

H(x ) = H() + 21 H () s2 + O(s3 ) ,

(11.125)

314

CHAPTER 11. SHOCK WAVES

where the x and t dependence is here implicit. If F (x ) is an arbitrary function which is

slowly varying on distance scales on the order of 1/2 , then we have
s
Z
4

eH()/2 F () .
(11.126)
dx F (x ) eH(x )/2
H ()

Applying this result to eqn. 11.121, we find

c

x
,
t

(11.127)

which is to say
c = g()

(11.128)

x = + g() t .

(11.129)

This is precisely what we found for the characteristics of ct + c cx = 0.

What about multivaluedness? This is obviated by the presence of an additional saddle point
solution. I.e. beyond some critical time, we have a discontinuous change of saddles as a
function of x:
x = + g( ) t = (x, t) .
(11.130)
Then

x
eH( )/2 + + eH(+ )/2
H ( )
H (+ )
1
.
c
H( )/2
t 1
e
+ 1
eH(+ )/2
H ( )

(11.131)

H (+ )

Thus,
H(+ ) > H( )

H(+ ) < H( )

x
t
x +
.
c
t

(11.132)
(11.133)

At the shock, these solutions are degenerate:

H(+ ) = H( )

1
2 (+ ) g(+ ) + g( ) =

Z+
d g() ,

(11.134)

which is again exactly as before. We stress that for small but finite the shock fronts are
smoothed out on a distance scale proportional to .
What does it mean for to be small? The dimensions of are [] = L2 /T , so we must
find some other quantity in the problem with these dimensions. The desired quantity is the
area,
Z


A = dx g(x) c0 ,
(11.135)

315

11.9. BURGERS EQUATION

where c0 = c(x = ). We can now define the dimensionless ratio,

R

A
,
2

(11.136)

which is analogous to the Reynolds number in viscous fluid flow. R is proportional to the
ratio of the nonlinear term (c c0 ) cx to the diffusion term cxx .

11.9.2

Examples

Whitham discusses three examples: diffusion of an initial step, a hump, and an N-wave.
Here we simply reproduce the functional forms of these solutions. For details, see chapter
4 of Whithams book.
For an initial step configuration,
c(x, t = 0) =

c1
c2

if x < 0
if x > 0 .

(11.137)

We are interested in the case c1 > c2 . Using the Cole-Hopf transformation and applying
the appropriate initial conditions to the resulting linear diffusion equation, one obtains the
complete solution,
c(x, t) = c2 +
where
and

c c2
1
 ,
1 + h(x, t) exp (c1 c2 )(x vs t)/2

(11.138)

vs = 12 (c1 + c2 )

(11.139)



xc2 t
erfc 4t
 .

h(x, t) =
xc1 t
erfc + 4t

(11.140)

Zz
2
du eu

(11.141)

Z
2
2
du eu = 1 erf(z) .
erfc(z) =

(11.142)

Recall that erfc(z) is the complimentary error function:

2
erf(z) =

Note the limiting values erfc() = 2, erfc(0) = 1 and erfc() = 0. If c2 < x/t < c1 ,
then h(x, t) 1 as t , in which case the solution resembles a propagating front. It is
convenient to adimensionalize (x, t) (y, ) by writing
s

c1
y
,
r
.
(11.143)
,
t=
x= p
c
c
c2
c1 c2
1 2

316

CHAPTER 11. SHOCK WAVES

Figure 11.12: Evolution of profiles for Burgers equation. Top : a step discontinuity evolving
into a front at times = 0 (blue), = 51 (green), and = 5 (red).. Middle : a narrow hump
c0 + A(x) evolves into a triangular wave. Bottom : dissipation of an N-wave at times = 41
(blue), = 21 (green), and = 1 (red).
We then have

where

c(z, )
2
p
= r 1 +
,
1 + h(z, ) exp(z)
c1 c2



,

z +
z
h(z, ) = erfc
erfc +
2
2

and

1
2

r r 1

zy

1
2

r + r 1 .

(11.144)

(11.145)

(11.146)

The second example involves the evolution of an infinitely thin hump, where
c(x, t = 0) = c0 + A (x) .

(11.147)

317

11.9. BURGERS EQUATION

The solution for subsequent times is

c(x, t) = c0 +

where R = A/2. Defining



R 1) exp xc0 t
(e
4t


 ,

t
t 1 + 1 (eR 1) erfc xc
0
2
4t
xc t
z 0 ,
2At

(11.148)

(11.149)

we have the solution

c = c0 +

2A
t

1/2

1
(eR 1) eRz

.
4R 1 + 12 (eR 1) erfc( R z)

Asymptotically, for t with x/t fixed, we have

(

x/t if 0 < x < 2At

c(x, t) =
0
otherwise .

(11.150)

(11.151)

This recapitulates the triangular wave solution with the two counterpropagating shock fronts
and dissipating shock strengths.
Finally, there is the N-wave. If we take the following solution to the linear diffusion equation,
r
a x2 /4t
(x, t) = 1 +
e
,
(11.152)
t
then we obtain

x
ex /4t
c(x, t) = q
.
t
t
x2 /4t
+
e
a

(11.153)

In terms of dimensionless variables (y, ), where

x=

we have
c=

a y

t = a ,

(11.154)

y
ey /4

.
a
+ ey2 /4

(11.155)

The evolving profiles for these three cases are plotted in fig. 11.12.

11.9.3

Confluence of shocks

The fact that the diffusion equation 11.118 is linear means that we can superpose solutions:
(x, t) = 1 (x, t) + 2 (x, t) + . . . + N (x, t) ,

(11.156)

318

CHAPTER 11. SHOCK WAVES

Figure 11.13: Merging of two shocks for piecewise constant initial data. The (x, t) plane is
broken up into regions labeled by the local value of c(x, t). For the shocks to form, we require
c1 > c2 > c3 . When the function j (x, t) dominates over the others, then c(x, t) cj .
where

j (x, t) = ecj (xbj )/2 e+cj t/4 .

We then have

P
c (x, t)
2x
= Pi i i
.
c(x, t) =

i i (x, t)

(11.157)

(11.158)

Consider the case N = 2, which describes a single shock. If c1 > c2 , then at a fixed time t
we have that 1 dominates as x and 2 as x +. Therefore c(, t) = c1 and
c(+) = c2 . The shock center is defined by 1 = 2 , where x = 12 (c1 + c2 ) t.
Next consider N = 3, where there are two shocks. We assume c1 > c2 > c3 . We identify
regions in the (x, t) plane where 1 , 2 , and 3 are dominant. One finds
1 > 2

x < 12 (c1 + c2 ) t +

b1 c1 b2 c2
c1 c2

(11.159)

1 > 3

x < 12 (c1 + c3 ) t +

b1 c1 b3 c3
c1 c3

(11.160)

2 > 3

x < 12 (c2 + c3 ) t +

b2 c2 b3 c3
.
c2 c3

(11.161)

These curves all meet in a single point at (xm , tm ), as shown in fig. 11.13. The shocks are
the locus of points along which two of the j are equally dominant. We assume that the
intercepts of these lines with the x-axis are ordered as in the figure, with x12 < x13 < x23 ,
where
bi ci bj cj
.
(11.162)
xij
ci cj

319

11.10. APPENDIX I : THE METHOD OF CHARACTERISTICS

When a given i (x, t) dominates over the others, we have from eqn. 11.158 that c ci .
We see that for t < t one has that 1 is dominant for x < x12 , and 3 is dominant for
x > x23 , while 2 dominates in the intermediate regime x12 < x < x23 . The boundaries
between these different regions are the two propagating shocks. After the merge, for t > tm ,
however, 2 never dominates, and hence there is only one shock.

11.10

Appendix I : The Method of Characteristics

Consider the quasilinear PDE

a1 (x, )

+ a2 (x, )
+ . . . + aN (x, )
= b(x, ) .
x1
x2
xN

(11.163)

This PDE is called quasilinear because it is linear in the derivatives /xj . The N
independent variables are the elements of the vector x = (x1 , . . . , xN ). A solution is a
function (x) which satisfies the PDE.
Now consider a curve x(s) parameterized by a single real variable s satisfying


dxj
= aj x, (x) ,
ds

(11.164)

where (x) is a solution of eqn. 11.163. Along such a curve, which is called a characteristic,
the variation of is
N


d X dxj
=
= b x(s), .
(11.165)
ds
xj ds
j=1

Thus, we have converted our PDE into a set of (N + 1) ODEs. To integrate, we must supply
some initial conditions of the form


g x, )
=0.
(11.166)
s=0

This defines an (N 1)-dimensional hypersurface, parameterized by {1 , . . . , N 1 }:

xj (s = 0) = hj (1 , . . . , N 1 )

(s = 0) = f (1 , . . . , N 1 ) .

j = 1, . . . , N

(11.167)
(11.168)

If we can solve for all the characteristic curves, then the solution of the PDE follows. For
every x, we identify the characteristic curve upon which x lies. The characteristics are identified by their parameters (1 , . . . , N 1 ). The value of (x) is then (x) = f (1 , . . . , N 1 ).
If two or more characteristics cross, the solution is multi-valued, or a shock has occurred.

11.10.1

Example

Consider the PDE

t + t2 x = x .

(11.169)

320

CHAPTER 11. SHOCK WAVES

We identify a1 (t, x, ) = 1
and a2 (t, x, ) = t2 , as well as b(t, x, ) = x . The characteristics are curves t(s), x(s) satisfing
dt
=1
ds

dx
= t2 .
ds

(11.170)

The variation of along each of the characteristics is given by

d
= x .
ds

(11.171)

The initial data are expressed parametrically as

t(s = 0) = 0

(11.172)

x(s = 0) =

(11.173)

(s = 0) = f () .

(11.174)

We now solve for the characteristics. We have

It then follows that

dt
=1
ds

t(s, ) = s .

(11.175)

dx
= t2 = s 2
ds

x(s, ) = + 13 s3 .

(11.176)

Finally, we have


d
= x = + 13 s3
ds

(s, ) = f () exp

1 4
12 s

s .

(11.177)

We may now eliminate (, s) in favor of (x, t), writing s = t and = x 13 t3 , yielding the
solution




(11.178)
(x, t) = f x 31 t3 exp 41 t4 xt .

11.11

Appendix II : Shock Fitting an Inverted Parabola

Consider the shock fitting problem for the initial condition



x2
c(x, t = 0) = c0 1 2 (a2 x2 ) ,
a

(11.179)

which is to say a truncated inverted parabola. We assume j () = 0. Clearly cx (x, 0) is

maximized at x = a, where cx (a, 0) = 2c0 /a, hence breaking first occurs at


a 
.
xB , t B = a ,
2c0

(11.180)

11.11. APPENDIX II : SHOCK FITTING AN INVERTED PARABOLA

321

Clearly + > a, hence c+ = 0. Shock fitting then requires

Z+
1
d c()
2 (+ )(c+ + c ) =

(11.181)

=
Since

c0
(2a + )(a )2 .
3a2

c0 2
(a 2 ) ,
a2


a
2
+ = 3 (2a + )
.
a +
c+ + c =

we have

(11.182)

(11.183)
(11.184)

The second shock-fitting equation is

+ = (c c+ ) t .

(11.185)

Eliminating + from the two shock-fitting equations, we have

t=

2a2 2a +

.
3c0 (a + )2

(11.186)

Inverting to find (t), we obtain

(t)
a
a
=
1+
a
3c0 t
3c0 t

1+

6c0 t
.
a

(11.187)

The shock position is then xs (t) = (t) + c (t) t.

It is convenient to rescale lengths by a and times by tB = a/2c0 , defining q and from

x aq and t a /2c0 . Then

2
2
1 + 3 .
=
1+
a
3
3

(11.188)

i
2 h
xs
(1 + 3 )3/2 + 1 .
= 1 +
a
9

(11.189)

q ( ) =
and
qs ( ) =
The shock velocity is
qs =
=

3
4

i 1
2 h
3/2
+ (1 + 3 )1/2
1
+
(1
+
3
)
9 2

( 1) +

81
64

(11.190)

( 1)2 + . . . ,

with vs = 2c0 qs = 12 c if we restore units. Note that qs ( = 1) = 0, so the shock curve

initially rises vertically in the (x, t) plane. Interestingly, vs ( 1) here, while for the
example in 11.4.3, where c(x, 0) had a similar profile, we found vs ( 1)1/2 in eqn.
11.69.

322

11.12

CHAPTER 11. SHOCK WAVES

Appendix III : Red Light, Green Light

Problem : Consider vehicular traffic flow where the flow velocity is given by


2
u() = u0 1 2
.
BB
Solve for the density (x, t) with the initial conditions appropriate to a green light starting
at t = 0, i.e. (x, 0) = BB (x). Solve for the motion of a vehicle in the flow, assuming
initial position x(0) = x0 < 0. How long does it take the car to pass the light?
Solution : Consider the one-parameter family of velocity functions,



u() = u0 1
,
BB
where > 0. The speed of wave propagation is
c() = u() + u ()



.
= u0 1 (1 + )
BB
The characteristics are shown in figure 11.14. At t = 0, the characteristics emanating from
the x-axis have slope 1/(u0 ) for x < 0 and slope +1/u0 for x > 0. This is because
the slope is 1/c (were plotting t on the y-axis), and c(0) = u0 while c(BB ) = u0 .
Interpolating between these two sets is the fan region, shown in blue in the figure. All
characteristics in the fan emanate from x = 0, where the density is discontinuous at t = 0.
Hence, the entire interval [0, BB ] is represented at t = 0.
For the fan region, the characteristics satisfy
x
c() =
t




1
x
=
1
.
BB
1+
u0 t

This is valid throughout the region u0 < x/t < u0 . The motion of a vehicle in this flow
is determined by the equation
u0
1 x
dx
= u() =
+
,
dt
1+ 1+ t
which is obtained by eliminating in terms of the ratio x/t. Thus, we obtain the linear,
non-autonomous, inhomogeneous equation
t

x
u0 t
dx

=
,
dt
1+
1+

whose solution is
x(t) = C t1/(1+) + u0 t ,
where C is a constant, fixed by the initial conditions.

323

11.12. APPENDIX III : RED LIGHT, GREEN LIGHT

Figure 11.14: Characteristics for the green light problem. For x < 0, the density at t = 0
is = BB , and c(x < 0, t = 0) = u0 . For x > 0, the density at t = 0 is = 0 and
c(x > 0, t = 0) = u0 .
What are the initial conditions? Consider a vehicle starting at x(t = 0) = |x0 |. It remains
stationary until a time t = |x0 |/u0 , at which point x/t = u0 , and it enters the fan
region. At this point (see figure 11.15), the trailing edge of the smoothing front region,
which was absolutely sharp and discontinuous at t = 0, passes by the vehicle, the local
density decreases below BB , and it starts to move. We solve for C by setting
x(t ) = |x0 |

1/(1+)

C = (1 + ) /(1+) |x0 |/(1+) u0

Thus,
x(t) = |x0 |

(t < t = |x0 |/u0 )

= u0 t (1 + ) /(1+) |x0 |/(1+) (u0 t)1/(1+)

(t > t ) .

Figure 11.15: Density versus position for the green light problem at three successive times.
The initial discontinuity at x = 0 is smoothed into a straight line interpolating between
= BB and = 0. Because dc/dx > 0 for this problem, the initial profile smooths out.
When dc/dx < 0, the wave steepens.

324

CHAPTER 11. SHOCK WAVES

Finally, we set x(tc ) = 0 to obtain

x(tc ) = 0
For = 2, find tc =

11.13

3 3
2

tc =

1
|x0 |
1
(1 + )1+
.

u0

|x0 |/u0 .

Appendix IV : Characteristics for a Periodic Signal

Problem : Consider traffic flow in the presence of a signal on a single-lane roadway. Assume
initial conditions (x, 0) = 1 . At time t = 0, the signal turns red for an interval Tred = r T ,
after which the signal is green for Tgreen = g T , with r + g = 1. The cycle length is T .
(Traffic engineers call g the green-to-cycle ratio.) It is useful to work with dimensionless
quantities,
/BB

c c/u0

x
x/u0 T

j j/u BB
0

u
u/u0

t/T .

The continuity equation is then

+ c(
)
=0.

Note : Assume j(
) = (1 ), so
u
(
) = 1

c(
) = 1 2
.

(a) During the red phase, two shocks propagate one behind the light (i.e. x

s < 0) and
+
one ahead of the light (i.e. x
s > 0). Find the velocities and discontinuities at these shocks.
Solution : During the red phase, 0 r, a shock is formed behind the light at the
boundary of the = 1 and = 1 regions. The velocity of this shock is
vs =

j(1) j(
1 )
=
1 .
1 1

This shock forms immediately upon the appearance of the red light, hence x

1 .
s ( ) =
Another shock forms ahead of the light, at the boundary between = 0 and = 1 regions.
There,
j(
1 ) j(0)
= 1 1 .
vs =
1 0
1 ) .
Thus, x
+
s ( ) = (1

(b) When the light changes to green, the characteristics develop a fan, and the discontinuity
at x = 0 starts to spread. Let x
> ( ) denote the minimum value of x
for which ( ) = 0,

11.13. APPENDIX IV : CHARACTERISTICS FOR A PERIODIC SIGNAL

325

Figure 11.16: Evolution of traffic density before, during, and after a red light.
and x
< ( ) the maximum value of x
for which ( ) = 1. Show that x
> overtakes x
+
s after a

time + and that x

< overtakes x
s after a time . Compute .
Solution : The situation is depicted in figure 11.16. At = r, the light turns green, and
the queue behind the light is released. The discontinuity at x
= 0 spreads as shown in the
figure. (Recall that regions with dc/dx > 0 spread while those with dc/dx < 0 steepen.)
The boundaries of the spreading region are x
> ( ) = r and x
< ( ) = r , and travel at
+
speeds c> = +1 and c< = 1. Setting x
s (+ ) = x
> (+ ) gives + = r/
1 . Similarly, setting
1

x
s ( ) = x
< ( ) gives = r/(1 1 ). For light traffic (
1 < 2 ), we have < + , while
1
for heavy traffic (
1 > 2 ), > + .
(c) Impose the proper shock conditions at x

s for > , respectively. Show that each

shock obeys the equation
d
xs
=
d

1
2

1 +

x
s
.
2( r)

Apply the proper boundary conditions for the ahead and behind shocks and obtain explicit
expressions for x

s ( ) for all 0. (You will have to break up your solution into different
cases, depending on .)
Solution : For a shock discontinuity between = and = + , we have in general
vs =

j+ j
= 1 + .
+

+
Lets first apply this at the shock x
+
s , we are in a fan region of the
s for > + . For x < x
characteristics, with
x
( ) = c(
)( r)

= (1 2
) ( r) .

Thus,
=

1
2

1
r

326

CHAPTER 11. SHOCK WAVES

Now invoke the shock condition at x

=x
+
s :
d
x+
s
= 1 1
d
=

1
2

1
2

x
+
s
r

x
+
s
,
2( r)

the solution of which is

1/2
x
+
+ (1 2
1 ) ( r) .
s ( ) = A+ ( r)

The constant A+ is determined by setting

x
+
> (+ ) =
s (+ ) = x

1 r
1 1

p
A+ = 2 r 1 (1 1 ) .

Putting this all together, we have

x
+
s ( ) =

(1 1 )

for <

r
1

4r 1 (1 1 ) ( r)1/2 + (1 2
1 ) ( r) for >

r
1

The identical analysis can be applied to the shock at x

s for > . The solution is once

again of the form
1/2
x

+ (1 2
1 ) ( r) ,
s ( ) = A ( r)
now with A chosen to satisfy
x

< ( ) =
s ( ) = x

1 r
1 1

p
A = 2 r 1 (1 1 ) = A+ .

Thus,

s ( )

for <

p
1 ) ( r) for >
4r 1 (1 1 ) ( r)1/2 + (1 2

(d) Compute the discontinuity

( ) at the shocks.
Solution : For the shock at x
+
s , we have
+

1
2

x
+
1 s
r

( ) = 1
p
= r 1 (1 1 ) ( r)1/2 ,

r
1
1
r
1
1

327

11.13. APPENDIX IV : CHARACTERISTICS FOR A PERIODIC SIGNAL

which is valid for > + = r/

1 . For the shock at x

s ,


x

( ) = 2 1
1
r
p
= r 1 (1 1 ) ( r)1/2 ,

which is valid for > = r/(1 1 ). Note

+ ( ) =
( ) for > max(+ , ).

(e) If 1 < 12 , show that the shock which starts out behind the light passes the light after
a finite time . What is the condition that the shock passes the light before the start of
the next red phase? Show that this condition is equivalent to demanding that the number
of cars passing the light during the green phase must be greater than the incoming flux at
x
= , integrated over the cycle length, under the assumption that the shock just barely
manages to pass through before the next red.
Solution : For light traffic with 1 < 21 , the shock behind the light passes the signal at

time , where x

s ( ) = 0, i.e.
p

1 ) ( r) 4r 1 (1 1 ) ( r)1/2 = 0 .
s ( ) = (1 2

Solving for , we have

= r +

r
4r 1 (1 1 )
=
.
2
(1 2
1 )
(1 2
1 )2

Similarly, if 1 > 21 , the shock ahead of the light reverses direction and passes behind the
signal at time = r/(2
1 1)2 .
For the case of light traffic, we see that if r > (1 2
1 )2 , the trailing shock remains behind
the signal when the next green phase starts the shock never clears the signal. In traffic
engineering parlance, the capacity of the intersection is insufficient, and the traffic will back
up. We can derive the result in a very simple way. Until the trailing shock at x

s passes
1
the signal, the density of vehicles at the signal is fixed at = 2 . Hence the flux of vehicles
at the signal is maximum, and equal to j(
= 12 ) = 41 . The maximum number of vehicles
that the signal can clear is therefore 14 g = 14 (1 r), which pertains when the trailing shock
just barely passes the signal before the start of the next red phase. We now require that
the incident flux of cars from x
= , integrated over the entire cycle, is less than this
number:
j(
1 ) 1 = 1 (1 1 ) < 41 (1 r) r < (1 2
1 )2 .
Merging shocks : So far, so good. We now ask: does the
up with the shock at x
+
s ( 1) from the next cycle? Lets

+
x
s ( ) = x
s ( 1) has a solution, which means we set

a 1 r + b ( 1 r)

a r + b ( r) =

(1 1 ) ( 1)

shock at x

s ( ) ever catch
assume that the equation

if > 1 +

r
1

if 1 < < 1 +

r
1

328

CHAPTER 11. SHOCK WAVES

with
a

4r 1 (1 1 ) ,

b 1 2
1 .

We first look for a solution satisfying > 1 + + = 1 + r1 . Let us call this a case I merge.
It is convenient to define s = 21 r, in which case
q
q
1
a s 2 + a s + 21 = b ,
which is solved by

s=

1
4

a2
b2
+
b2
a2

Note s > 21 , so > 1 + r. Solving for , we obtain

= I 1 + r +
r
1 .

We must now check that > 1 +

with
r1 =

a b

b a

2

We have

r
I
1

0 > f (r) 1 +
=

1
4

(1 1 )2 (1 3
1 )
1
(1 2
1 )2
r
+

(1 2
1 )2 1
2 16r 1 (1 1 )
(1 1 )2 (1 3
1 )
(r r1 )(r r2 )
2
(1 2
1 ) 1 r

1 2
1
2 2
1

2

r2 =

(1 2
1 )2
.
4(1 1 )(1 3
1 )

When 1 < 31 , we have r2 > 0 > r1 , and we require 0 < r < r2 . When 1 > 13 , r1 > r2 so
again we require 0 < r < r2 .
Now lets solve for the merging shocks. We write the condition that two characteristics
meet. Thus,
( r) = ( 1 r)
with x
= ( r). Let = r. Then x
= = ( 1), which yields
=

.
1

The combined shock moves at velocity vs , where

vs = x
= 12 ( + ) =

+
.
2 2( 1)

This equation is easily integrated to yield

x
I () = C

( 1) .

329

11.13. APPENDIX IV : CHARACTERISTICS FOR A PERIODIC SIGNAL

Figure 11.17: Phase diagram for the shock merge analysis. For r > (1 2
1 )2 (red line),
the signal is congested, and the trailing shock does not clear the signal before the start of
the following red phase. In case II, the shocks merge when the leading shock is still in its
constant velocity phase. In case I, the shocks merge after the leading shock has already
begun to decelerate.
The constant C is fixed by the initial conditions at time I , when the shocks merge. We
have


b 2
1 a
+
I r = 4
b a
x
(I ) = a
=

1
4

I r + b (I r)

a2
b2

a2
b2

We then have
( 1) =
and we set
1
4

b2
a2

b2
a2

1
16

C=

b.

a2
b2

b2
a2

1
4

2

a2
b2

a2
b2

b,

which simply yields C = b = 1 2

1 .
Finally, consider the case II merge where x

+
s ( ) = x
s ( 1) requires < 1 +
must solve

a r + b ( r) = (1 1 ) ( 1) ,
the solution of which is
II = r + 1

2 1 1
r
.
1

r
1 .

We now

330

CHAPTER 11. SHOCK WAVES

Figure 11.18: Characteristics for the light cycle problem. Red interval signifies red light;
green interval signifies green light. Red characteristics correspond bumper-to-bumper traffic, i.e. = 1. Dotted line characteristics correspond to = 0. Note the merging of shocks.
Congested regions are those with negatively-sloped characteristics, since c < 0 for > 21 .
One can check that II < 1 + r1 is equivalent to r > r2 . In this case, the characteristics to
the right of the shock all have slope 1, corresponding to = 0, and are not part of the fan.
Thus, the shock condition becomes


x

1
x
= 2 1+

with = r as before. The solution here is

x
II = C + .
Again, C is determined by the merging shock condition, x

+
s ( ) = x
s ( 1). We therefore

11.14. APPENDIX V : CAR-FOLLOWING MODELS

331

set
r



2 1 1

2 1 1

1 1
+ 1 r
= (1 1 ) r + 1 r
1 ,
C 1 r
1
1
1

yielding

p
a
C = 4
1 (1 1 ) = .
r

This solution is valid until the characteristics on the right side of the shock extrapolate back
to the x
axis at time = 1 + r, when the fan begins. Thus,

x
= 1 r = 1 = C + ,
which yields = m

1
4
1 (1
1 ) .

p
For > m , we match with the profile C ( 1), which we obtained earlier. Thus,
q
q
p

C
m (m 1) = m 4
1 (1 1 ) m ,
which yields C = b = 1 2
1 , as before.

Putting this all together, we plot the results in figure 11.18. The regions with negativelysloped characteristics are regions of local congestion, since c < 0 implies > 12 . Note that
the diagram is periodic in time, and it is presumed that several cycles have passed in order
for the cycle to equilibrate. I.e. the conditions at t = 0 in the diagram do not satisfy
(x) = 1 for x > 0.

11.14

Appendix V : Car-Following Models

So-called car-following models are defined by equations of motion for individual vehicles.
Let xn (t) be the motion of the nth vehicle in a group, with xn+1 < xn , so that the lead
vehicle (n = 0) is the rightmost vehicle. Three classes of models have traditionally been
considered in the literature:
Traditional car-following model (CFM) Cars accelerate (or decelerate) in an attempt
to match velocity with the car ahead, subject to a delay time :
n
o
x
n (t + ) = x n1 (t) x n (t) .
Optimal velocity model (OVM) Drivers instantaneously accelerate in order to maintain a velocity which is optimally matched to the distance to the next car. (The
distance to the next vehicle is known as the headway in traffic engineering parlance.)
Thus,
o
n
x
n = V (xn1 xn ) x n ,
where V (x) is the optimum velocity function.

332

CHAPTER 11. SHOCK WAVES

Optimal headway model (OHM) Drivers instantaneously accelerate in order to maintain an optimal headway, given their current velocity:
n
o
x
n = H(x n ) (xn1 xn ) .
The optimal headway function is just the inverse of the optimal velocity function of
the OVM: H = V 1 .

(a) The CFM equation above is a linear equation. Solve it by Fourier transforming the
function vn (t) = x n (t). Show that
vn () = r() ei() vn1 () .
Thus, given the velocity v0 (t) of the lead vehicle, the velocity of every other vehicle is
determined, since vn () = r n () ein() v0 (). Derive the stability criterion |r| < 1 in terms
of , , and . Show that, if the motion is stable, no matter how erratic the lead vehicle
moves, for n the variations in vn (t) are damped out and the velocity approaches a
constant.
Solution : We have the linear equation

Solve by Fourier transform:

n
o
v n (t + ) = vn1 (t) vn (t) .

vn (t) =

d
v () eit .
2 n

This yields
vn () = z() vn1 ()
with
z() =

.
i exp(i )

Thus, writing z = r ei , we have

() = tan
and
r() = p

cos
sin

2 + 2 2 sin( )

The velocity of the nth vehicle is now given in terms of that of the lead vehicle, according
to the relation
vn () = r n () ein() v0 () .

11.14. APPENDIX V : CAR-FOLLOWING MODELS

333

For = 0 we have z() = 1, which says that the time-averaged velocity of each vehicle is
the same. For vn () to be bounded requires r() 1, which gives
r() 1

sin
1.

The maximum of the function x1 sin x is unity, for x = 0. This means that the instability
first sets in for infinitesimal . We can ensure that every frequency component is stable by
requiring
< 12 .
The interpretation of this result is straightforward. The motion is stable if the product
of the response time and the sensitivity is sufficiently small. Slow response times and
hypersensitive drivers result in an unstable flow. This is somewhat counterintuitive, as we
expect that increased driver awareness should improve traffic flow.
(b) Linearize the OVM about the steady state solution,
x0n (t) = na + V (a)t ,
where a is the distance between cars. Write xn (t) = x0n (t) + xn (t) and find the linearized
equation for xn (t). Then try a solution of the form
xn (t) = A eikna et ,
and show this solves the linearized dynamics provided = (k). What is the stability
condition in terms of , k, and a? Show that all small perturbations about the steady state
solution are stable provided V (a) < 21 . Interpret this result physically. If you get stuck,
see M. Bando et al., Phys. Rev. E 51, 1035 (1995).
Solution : Writing xn = na + V (a)t + xn , we linearize the OVM and obtain
n
o

xn = V (a) xn1 xn x n + . . . .
We now write

xn (t) = A einka et ,
and obtain the equation
2 + V (a) (1 eika ) = 0 .
This determines the growth rate (k) for each wavelength k. Solving the quadratic equation,
q
(k) = 12 12 2 4V (a) (1 eika ) .
Let us separate into its real and imaginary parts: + i. Then
q
21 + i = 12 2 4V (a) (1 eika )

334

CHAPTER 11. SHOCK WAVES

which, when squared, gives two equations for the real and imaginary parts:
2 + 2 = V (a) (1 cos ka)
2 = V (a) sin ka .

We set = Re = 0 to obtain the stability boundary. We therefore obtain

V (a) <

2 cos2 12 ka

Re (k) > 0 .

The uniform k = 0 mode is the first to go unstable. We therefore have that all k modes are
linearly stable provided
V (a) < 12 .
This says that the stability places a lower limit on the sensitivity exactly the opposite
of what we found for the CFM. The OVM result is more intuitive. Physically, we expect
that the OVM is more realistic as well, since drivers generally find it easier to gauge the
distance to the next car than to gauge the difference in velocity.