This action might not be possible to undo. Are you sure you want to continue?

Welcome to Scribd! Start your free trial and access books, documents and more.Find out more

353J

Nonlinear Dynamics I: Chaos

Daniel H. Rothman

Massachusetts Institute of Technology

Contents

0 Acknowledgements and references 1 Pendulum 1.1 Free oscillator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2 Global view of dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.3 Energy in the plane pendulum . . . . . . . . . . . . . . . . . . . . . . . . . . 2 Stability of solutions to ODEs 2.1 Linear systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2 Nonlinear systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 Conservation of volume in phase space 4 Damped oscillators and dissipative systems 4.1 General remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2 Phase portrait of damped pendulum . . . . . . . . . . . . . . . . . . . . . . 4.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 Forced oscillators and limit cycles 5.1 General remarks . . . . . . . . . . 5.2 Van der Pol equation . . . . . . . 5.3 Energy balance for small α . . . . 5.4 Limit cycle for α large . . . . . . 5.5 A ﬁnal note . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

7

7

8

9

14

14

17

19

22

22

24

26

28

28

28

31

34

36

38

38

39

41

44

45

47

47

48

50

52

54

54

55

56

59

60

63

6 Parametric oscillator 6.1 Mathieu equation . . . . . . . . . . . 6.2 Elements of Floquet Theory . . . . . 6.3 Stability of the parametric pendulum 6.4 Damping . . . . . . . . . . . . . . . . 6.5 Further physical insight . . . . . . . .

7 Fourier transforms 7.1 Continuous Fourier transform . . . . . . . . . . . . . . . . . . . . . . 7.2 Discrete Fourier transform . . . . . . . . . . . . . . . . . . . . . . . . 7.3 Inverse DFT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.4 Autocorrelations, power spectra, and the Wiener-Khintchine theorem 7.5 Power spectrum of a periodic signal . . . . . . . . . . . . . . . . . . . 7.5.1 Sinusoidal signal . . . . . . . . . . . . . . . . . . . . . . . . . 7.5.2 Non-sinusoidal signal . . . . . . . . . . . . . . . . . . . . . . . 7.5.3 tmax /T �= integer . . . . . . . . . . . . . . . . . . . . . . . . . 7.5.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.6 Quasiperiodic signals . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.7 Aperiodic signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

8 Poincar´ sections e 8.1 Construction of Poincar´ sections . . . . . . . . . . . . . . . . . . . e 8.2 Types of Poincar´ sections . . . . . . . . . . . . . . . . . . . . . . . e 8.2.1 Periodic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.2.2 Quasiperiodic ﬂows . . . . . . . . . . . . . . . . . . . . . . . 8.2.3 Aperiodic ﬂows . . . . . . . . . . . . . . . . . . . . . . . . . 8.3 First-return maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.4 1-D ﬂows . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.5 Relation of ﬂows to maps . . . . . . . . . . . . . . . . . . . . . . . . 8.5.1 Example 1: the van der Pol equation . . . . . . . . . . . . . 8.5.2 Example 2: R¨ssler attractor . . . . . . . . . . . . . . . . . o 8.5.3 Example 3: Reconstruction of phase space from experimental 9 Fluid dynamics and Rayleigh-B´nard convection e 9.1 The concept of a continuum . . . . . . . . . . . . 9.2 Mass conservation . . . . . . . . . . . . . . . . . . 9.3 Momentum conservation . . . . . . . . . . . . . . 9.3.1 Substantial derivative . . . . . . . . . . . . 9.3.2 Forces on ﬂuid particle . . . . . . . . . . . 9.4 Nondimensionalization of Navier-Stokes equations 9.5 Rayleigh-B´nard convection . . . . . . . . . . . . e 9.6 Rayleigh-B´nard equations . . . . . . . . . . . . . e 9.6.1 Dimensional form . . . . . . . . . . . . . . 9.6.2 Dimensionless equations . . . . . . . . . . 9.6.3 Bifurcation diagram . . . . . . . . . . . . 9.6.4 Pattern formation . . . . . . . . . . . . . . 9.6.5 Convection in the Earth . . . . . . . . . . 10 Introduction to Strange Attractors 10.1 Dissipation and attraction . . . . . 10.2 Attractors with d = 2 . . . . . . . . 10.3 Aperiodic attractors . . . . . . . . 10.4 Example: R¨ssler attractor . . . . . o 10.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

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

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

68

68

69

69

71

72

73

76

77

77

78

80

82

82

84

85

85

87

89

91

94

94

95

96

97

97

100

100

102

104

106

108

109

109

111

111

113

114

115

120

120

122

11 Lorenz equations 11.1 Physical problem and parameterization 11.2 Equations of motion . . . . . . . . . . 11.2.1 Momentum equation . . . . . . 11.2.2 Temperature equation . . . . . 11.3 Dimensionless equations . . . . . . . . 11.4 Stability . . . . . . . . . . . . . . . . . 11.5 Dissipation . . . . . . . . . . . . . . . 11.6 Numerical solutions . . . . . . . . . . . 11.7 Conclusion . . . . . . . . . . . . . . . . 2

. . . 129 e 13. . . . . . . . .8 Computation of � . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5 Quasiperiodicity . . . . . . . 16. . . . . . .4 Relationship of � to D . . . . . . . . 16 Period doubling route to chaos 16. . . . . . . . . . . . . . . 16. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .9 Linearized doubling operator . . . . . . . . . . . . . . . . . . 17. . . . . . . . . . . . . . . . .6 Smale’s horseshoe attractor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12. .1 Rayleigh-B´nard convection . . 16. . . . . . .1 The H´non map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .10Computation of � . . .3. . . . . . . . . . . . . . . . . 16. . . . . . . . 16. . . . . . . . . . . .2 Logistic map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5 Lyaponov exponents and attractors in 3-D 15. .4 Lyaponov number . . . . 17. . 15. . . . . . . . . . . . . . . . 16. . . . . . . . . . . . . .4 Period doubling bifurcations . . . . . . . . .5. . . . . . . . .3 Example 2: Time-dependent eigenvalues . . .5 Scaling and universality . . . . . . . . . . .2 Computation . . . . . . . . . . . . . . . . . . . . . . . . . 16.11Comparison to experiments . . . . . . . 16. . .7 Doubling operator . . . . . . . . . . . .12 H´non attractor e 12. . . . . . .1 Instability of a limit cycle . . . . . . . . . . . . . . . . . . . . .3. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17. . .2 Dissipation . . . . . . . . . . . . . . 15. . . . . . . . . . . . . . . . .4 Numerical evaluation . . . . . . . . 129 14 Fractals 14. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16. . . . . . . . . . . . . . . . . . 14. . . . . . 17 Intermittency (and quasiperiodicity) 17. . . . . . . . . . . . . . . .1 Deﬁnition . . . . . . . . . 14. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1 Diverging trajectories . . . . .3 Fixed points and stability . . . . . . . . . . . .1 General characteristics of intermittency 17. . . . . . . . . . . . . . . . . . . . . . 17. . . . . . . . . . . .2 Belousov-Zhabotinsky reaction . .3 Correlation dimension � 14. . . . .1 An historical note . . . . . . . . . . . 132 132 133 134 134 136 139 141 141 142 143 144 146 147 150 150 152 154 155 160 163 164 165 166 169 170 173 173 174 177 178 180 180 15 Lyapunov exponents 15. . . . 15. . .3 Average duration of laminar phase . . . . . . . . . . . . 15. 3 . . . . . . . . . . . . . . . . .2 One-dimensional map . . . . . .2 Example 1: M independent of time . . . . . . . . . . . . . . . . .6 Universal limit of iterated rescaled f ’s 16. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1 Deﬁnition . . . . .3 Numerical simulations . . . . . . . . . . . . . 124 124 126 127 13 Experimental attractors 129 13. . . . . . . . . . . . . . . . . . . . . . . . . 14. . . . . . . . . . . . . . . . . . . . . e 12. . . . . . . . . . . . . . . . . . . . . .2 Examples . . . . . . . . . . . 14. . . . .

. . . . . . .2 Ruelle-Takens theory . . . . . . . . . . . . .17. . . . . . 181 4 .5. . .

and my own department (Earth. The conversion followed from LaTeX notes prepared by Kurt Steinkraus. Occasionally these are listed explicitly in the notes. Landau and Lifshitz [8]. Thanks are of course also due to the students themselves. and Palm [9]. and Kadanoﬀ [17]. in rough order of appearance. Greg Lawson.0 Acknowledgements and references Since 1989 I have taught a one-semester course on nonlinear dynamics and chaos to MIT undergraduates. Olav van Genabeek. and Vidal [1]) include the e following: • Systems of ODE’s: Strogatz [2] and Beltrami [3]. Schuster [16]. The construction of the course has been inﬂuenced heavily by the book by Berg´. The lecture notes that follow have undergone frequent revisions. • Fluid dynamics: Tritton [7]. derive directly from their book. • Lorenz equations: Lorenz [11] and Tritton [7]. • Strange attractors: Abraham and Shaw [10]. The TA’s included. Students come from thoughout the institute. particularly in the ﬁrst half of the course. but frequently resulted in new ideas for the course. in the second half of the course. • Liouville’s theorem: Tolman [4]. from partial diﬀerential equations to ordinary diﬀerential equations to maps. and Planetary Sciences). Particular lectures and references used to construct the notes (aside from Berg´. Atmospheric. • Mathieu equation: Landau and Lifshitz [6]. a student in 2003. or references to ﬁgures contained within them are made. • Floquet theory: Bender and Orszag [5]. e and Vidal [1]. the extensive comparison between theory and experiment. e e • Fractals: Barnsley [13] and Grassberger and Procaccia [14]. perhaps most importantly. Alison Cohen. • Van der Pol equation: Strogatz [2]. physics. Kelvin Chan. Although some sections of the notes. An drew Gunstensen. Pomeau. Joshua Weitz. John Olson. There are many other sources of inspiration. Advice and assistance from my TA’s has not only been much appreciated. but predominantly from math ematics. when they were (ﬁnally) converted to LaTeX. but perhaps never more so than in 2004. and David Forney. a greater debt manifests itself in a shared “philosophy” and organization: the introduction by way of oscillators. and. • H´non attrator: H´non [12]. • Period doubling: Feigenbaum [15]. the progression. the emphasis on data analysis. 5 . Pomeau. Since 1998 the course has been jointly sponsored by the Department of Mathematics. Davide Stelitano.

Mechanics. England. Math. 189–208 (1983). (1984). and Lifshitz. Pomeau. Phys. (1976). L. McGraw Hill. D. R. D. Addison-Wesley. M. Dover. [11] Lorenz. J. D. [5] Bender. Weinheim. 2nd edition. 130–141 (1963). and Orszag. S. (1994). San Diego. New York. P. Fluid Mechanics. C. John Wiley & Sons. (1988). and Shaw. (1988). Santa Cruz. (1979). New York. E. N. Physics Today . San Diego. C.References [1] Berg´. E. Germany. Dynamics–The Geometry of Behavior. New York. M. Roads to chaos. Advanced Mathematical Methods for Scientists and Engineers. Fractals Everywhere. Pergamon Press. S. and Procaccia. e 50. E. L. Los Alamos Science 1. J. [6] Landau. Universal behavior in nonlinear systems. [4] Tolman. (1982). E. Pergamon Press. Clarendon Press. [14] Grassberger. C. Order Within Chaos: Toward a Deterministic e Approach to Turbulence. [2] Strogatz. C. Deterministic aperiodic ﬂow. (1959). M. [8] Landau. [10] Abraham. M. R. A. and Vidal. Annual Review of Fluid Mechanics 7. M. Nonlinear dynamics and Chaos. G. Sci. 20. VCH. [15] Feigenbaum. December (1983). D. I. New York. (1988).. Oxford. 69–77 (1976). A two-dimensional mapping with a strange attractor. 4–27 (1980). H. Measuring the strangeness of strange attractors. [13] Barnsley. L. New York. P. Nonlinear thermal convection. [3] Beltrami. [12] H´non. 39–61 (1975). and Lifshitz. Y. Commun. Academic Press. 6 . [16] Schuster.. The Principles of Statistical Mechanics. P. M. (1987). Mathematics for Dynamic Modeling. CA. [17] Kadanoﬀ. Oxford. Physical Fluid Dynamics. Deterministic Chaos: An Introduction. E. Academic Press. Atmos. [7] Tritton. (1978). Aerial Press. [9] Palm. H. Physica D 9.

undamped pendulum.1 1..1 Pendulum Free oscillator To introduce dynamical systems. 7 . a rigid geometry. dt2 l Our pendulum equation is idealized: it assumes. no friction. we consider an unforced. and most importantly. (It acts in the direction opposite to sgn(β)). θ l mg sinθ mg The arc length (displacement) between the pendulum’s current position and rest position (β = 0) is s = lβ Therefore ˙ s = lβ ˙ ¨ s = lβ ¨ From Newton’s 2nd law.g. e. Thus ¨ F = mlβ = −mg sin β or d2 β g + sin β = 0. a point mass. we begin with one of the simplest: a free oscillator. ¨ F = mlβ The restoring force is given by −mg sin β. Speciﬁcally.

Then d2 β g =− β 2 dt l whose solution is �� � g β = β0 cos t+θ l or β = β0 cos(γt + θ) where the angular frequency is � g γ= . 1 0 θ/θ 0 −1 0 1 t/T 2 1. Note that the motion is exactly periodic.The equation is nonlinear. g and β0 and θ come from the initial conditions. the period T is independent of the amplitude β 0. Thus the equation is not easily solved. dt 8 . However for small β � 1 we have sin β ◦ β. l the period is � l T = 2α . Furthermore.2 Global view of dynamics What do we need to know to completely describe the instantaneous state of the pendulum? dβ ˙ The position β and the velocity = β. because of the sin β term.

β) = const. we know that the resulting trajectory must be closed: θ θ In which direction is the ﬂow? What shape does the curve take? To calculate the curve. Because the solution is periodic. we note that it should be characterized by constant energy. for the pendulum. we seek a representation ˙ of the solution in the plane of β and β. As before. s = arc length = lβ. 9 . since no energy is input to the system (it is not driven) and none is dissipated (there is no friction).d.3 Energy in the plane pendulum l cos θ h θ l mg sinθ mg The pendulum’s height above its rest position is h = l − l cos β. 1. β). and expect the trajectories to be ˙ curves of E(β.e.Instead of integrating our o. ˙ Therefore we compute the energy E(β.

eqn. � β0 ˙ E(β0. since E is conserved 10 � . β) is a constant of motion by calculating its time deriva tive: dE 1 ˙¨ ˙ = ml2(2ββ) + mglβ sin β dt 2 � � g 2˙ ¨ = ml β β + sin β l g ¨ = 0 (since the pend. β |λ0 ) = mgl(1 − cos β0 ) 2 Since cos β = 1 − 2 sin2 (β/2). β = − sin β) l So what do these curves look like? Take β0 to be the highest point of motion. θ0 Then and ˙ β(β0) = 0 ˙ E(β0. β) is 1 ˙ ˙ E(β.The kinetic energy T is 1 1 1 ˙ ˙ T = ms2 = m(lβ)2 = ml2β2 ˙ 2 2 2 The potential energy U is U = mgh = mg(l − l cos β) = mgl(1 − cos β) ˙ Therefore the energy E(β. β) = ml2β2 + mgl(1 − cos β) 2 ˙ We check that E(β. β |λ0 ) = 2mgl sin 2 ˙ = E(β. β) in general.

E = 2mgl. the curves are the cosines � � � β g ˙ cos . or � � Thus for small β the curves are circles of radius β0 in the plane of β and � ˙ β/ g/l. For β0 = α. β = ±2 l 2 11 . θ / (g/l)1/2 θ ˙ β � g/l �2 2 + β 2 ◦ β0 What about β0 large? Consider the case β0 = α.Now write T = E − U : � � 1 2 ˙2 β0 β ml β = 2mgl sin2 − sin2 2 2 2 � � g β0 β ˙ sin2 − sin2 β2 = 4 l 2 2 g ˙ β2 ◦ 4 l � 2 β0 β 2 − 4 4 (1) (2) For small β0 such that β � 1. and equation (2) gives ⎡ � �� �α � β g ˙ β2 = 4 sin2 − sin2 l 2 2 � � g β = 4 cos2 l 2 Thus for β0 = α.

0. . are stable equilibrium. 12 . . 3α . . . . or ﬁxed. ˙ The points β = 0. −3α. . −2α. marginally stable). β = . points (actually. . Thus for undampled.(dθ / dt) / (g / l) 1/2 2 0 −2 −2 −1 0 θ/π 1 2 Intuitively. are unstable ﬁxed points. nonlinear pendulum we can construct the following phase portrait: The portrait is periodic. α. . −α. . 2α. β = . . we recognize that this curve separates oscillatory motion (E < 2mgl) from rotary motion (E > 2mgl). ˙ The points β = 0. .

The trajectories appear to cross. then what happens? (The motion stops. but they do not. Why not? (Deterministic trajectories.) 13 . But we shall see that it would take inﬁnite time to arrive at these points. awaiting instability.) If the trajectories actually arrive to these crossing points.

assume that we have the linear system u1 = a11u1 + a12u2 ˙ u2 = a21u1 + a22u2 ˙ 14 . 1. x2 = β. where � � x1 ˙ ˙ ψ= x =ψ 0 x2 ˙ For the pendulum. .e. Then x1 = x 2 ˙ g x2 = − sin x1 ˙ l The equilibrium points. Since sin x1 is periodic. n = 0. l to two ﬁrst order ODE’s. this requires x2 = 0 x1 = ±nα. or ﬁxed points. How may we be more precise? 2. ¨ g β + sin β = 0. i. . are where the trajectories in phase space stop.2 Stability of solutions to ODEs How can we address the question of stability in general? We proceed from the example of the pendulum equation. First. the ﬁrst is stable and the second is not. the only distinct ﬁxed points are � � � � � � � � β β 0 α and ˙ = 0 ˙ = 0 β β Intuitively. We reduce this second order ODE. ˙ Write x1 = β. . 2.1 Linear systems Consider the problem in general.

�2 are eigenvalues of A. (a) If �1 < 0 and �2 < 0. u2 u1 15 . ∞ stable. ˙ Then the only ﬁxed point (where ψ = 0) is ψ = 0.or with ψ (t) = u ˙ ψ = Aψ u u � u1 (t) u2 (t) � and A= � a11 a12 a21 a22 � Assume A has an inverse and that its eigenvalues are distinct. ψ ψ • �1 and �2 are constants (deriving from initial conditions). • c1 . What are the possibilities for stability? 1. is ψ (t) = �1 e�1 t c1 + �2 e�2 t c2 u ψ ψ where • �1 . �1 and �2 are both real. in general. then u(t) ∗ 0 as t ∗ →. u u The solution. c2 are eigenvectors of A.

u2 u1 (c) If �1 < 0 < �2 . u2 u1 16 . u ψ ψ ψ (t) = eθt (λ1 cos qt + λ2 sin qt) u ρ ρ (�1 . 2. ∞ unstable. �1 . Assuming ψ (t) is real. There are three possibilities: (a) Re{�} = ε > 0 =∞ unstable.(b) If �1 > 0 and �2 > 0. then u(t) ∗ → as t ∗ →. u ψ • If ψ (0) is a multiple of c2 . then u(t) ∗ → as t ∗ →. Then � = ε ± iq. • If ψ (0) is a multiple of c1 . u ψ u2 c2 c1 u1 ∞ unstable saddle. then u(t) ∗ 0 as t ∗ →. �2 are both complex. �2 are formed from a linear combination of of A’s eigenvectors and the initial conditions).

x2) ˙ where f1 and f2 are nonlinear functions of x1 and x2.(b) ε < 0 =∞ stable. 2. For a one dimensional function g(x) we would have g(x� + u) � g(x� ) + g � (x� ) · u 17 . i Perform a Taylor expansion around the ﬁxed point. � �� x1 Suppose is a ﬁxed point. Is it stable? x� 2 Deﬁne ui = xi − x� to be a small departure from the ﬁxed point. u2 u1 We leave the case of repeated eigenvalues to Strogatz (pp. u2 u1 (c) ε = 0 =∞ marginally stable. x2) ˙ x2 = f2(x1. 135-6).2 Nonlinear systems We are interested in the qualitative behavior of systems like x1 = f1(x1.

We have x1 = f1(x1.) 0 The question of stability is then addressed just as in the linear case. Also. x2) = − sin x1 ˙ l and A= � 0 1 −g/l 0 � � for x� 1 x� 2 � � = x� 1 x� 2 � � � � 0 = 0 There is a diﬀerent A for the case � α . since we have A is called the Jacobian matrix of f at ψ �. x2) ˙ ˙ Substituting ui = fi (x1. we obtain ˙ ˙ u ψ ◦ Aψ u where ⎪ γf1 γx1 γf2 γx1 γf1 γx2 γf2 γx2 =0 since ﬁxed pt The ﬁrst term vanishes since it is evaluated at the ﬁxed point. x2) = ωfi � � ωfi � � fi(x�. x�) + (x1. (The sign of g/l changes. via calculation of the eigenvalues and eigenvectors. x2) u2 + O(u2) 1 2� � ⎜� ωx1 ωx2 ui = x i − x � i ui = xi = fi(x1. x2) = x2 ˙ g x2 = f2(x1. 18 . x ⎬ A = ⎛ � � � � �� ⎞� � � τ =τ � x x We now apply these results to the pendulum.Here we obtain fi (x1. x2) above. x2) u1 + (x1.

Suppose we have a 3-D phase space. x2. dissipative systems contract volumes. x2. Conversely. we shall see. 19 . x3) ˙ x3 = f3(x1. neither necessarily the same shape nor size: n x V ds x1 x3 V’ x2 Assume the volume V has surface S. where it occupies a volume V ∗ .3 Conservation of volume in phase space We show (via the example of the pendulum) that frictionless systems conserve volumes (or areas) in phase space. x A set of initial conditions enclosed in a volume V ﬂows to another position in phase space. such that x1 = f1(x1. x2. x3) ˙ x2 = f2(x1.” where dψ /dt is the velocity. x3) ˙ or dψ x ψx = f (ψ ) dt The equations describe a “ﬂow.

• ds = a small region of S. ψ • δf = rate of ﬂow of points (trajectories emanating from initial condi tions) through unit area perpendicular to the direction of ﬂow.. a positive ﬂux =∞ a loss of “mass. and • ψ = the unit normal (outward) to ds. but it always contains the same points.” Now we apply the divergence theorem to convert the integral of the vector ψ ﬁeld δf on the surface S to a volume integral: � � ωδ ψ ψ dV = − [� · (δf)]dV ωt V V Letting the volume V shrink.e.Let • δ = density of initial conditions in V . n Then net ﬂux of points out of S = or � � ψ n (δf · ψ )ds S � ωδ ψ n dV = − (δf · ψ )ds ωt V S i. we have ωδ ψ ψ = −� · (δf ) ωt Now follow the motion of V to V ∗ in time νt: V V’ The boundary deforms. 20 .

x2) g x2 = − sin x1 = f2 (x1. Character istic of dissipative systems. We use the example of the pendulum: x1 = x 2 ˙ = f1 (x1. dV ψ ψ −δV � · f = −δ dt or 1 dV ψ ψ =�·f V dt This is called the Lie derivative. by our previous result. We shall next arrive at the following main results by example: ψ ψ • � · f = 0 ∞ volumes in phase space are conserved. dt dt dt Thus.We wish to calculate dδ/dt. x2) ˙ l 21 . which is the rate of change of δ as the volume moves: dδ ωδ ωδ dx1 ωδ dx2 ωδ dx3 = + + + dt ωt ωx1 dt ωx2 dt ωx3 dt ψ ψ ψ ψ = −� · (δf ) + (�δ) · f ψ ψ ψ ψ ψ = −(�δ) · f − δ� · f + (�δ) · f ψ ψ = −δ� · f Note that the number of points in V is N = δV Since points are neither created nor destroyed we must have dN dδ dV = V +δ = 0. ψ ψ • � · f < 0 ∞ dV /dt < 0 ∞ volumes in phase space contract. Characteristic of conservative or Hamiltonian systems.

1 Damped oscillators and dissipative systems General remarks We have seen how conservative systems behave in phase space. What about dissipative systems? What is a fundamental diﬀerence between dissipative systems and conserva tive systems. In conservative systems. • Dissipative systems are not. 22 . aside from volume contraction and energy dissipation? • Conservative systems are invariant under time reversal. they are irreversible. the conservation of volumes in phase space is known as Liouville’s theorem. 4 4. Conservation of areas holds for all conserved systems. This is conventionally derived from Hamiltonian mechanics and the canonical form of equations of motion.Calculate Pictorially ωx1 ωx2 ˙ ˙ ψ ψ �·f = + = 0+0 ωx1 ωx2 x2 x1 Note that the area is conserved.

2 dt Let t ∗ −t and thus ω/ωt ∗ −ω/ωt.. β) = ml β + β 2 l = 1 2 ˙2 ml (β + γ 2 β2) 2 23 � β2 2 � .. the damped pendulum equation is d2 β dβ + ρ + γ2β = 0 dt2 dt where γ 2 = g/l ρ = damping coeﬃcient The sign of θ is chosen so that positive damping is opposite the direction of motion.. i.Consider again the undamped pendulum: d2 β + γ 2 sin β = 0. The fact that most systems are dissipative is obvious if we run a movie backwards (ink drop. Summing the kinetic and potential energies. How does the energy evolve over time? As before.) Formally. we have 1 2 � ˙2 g 2 � ˙ E(β. how may dissipation be represented? Include terms propor tional to odd time derivatives. we calculate kinetic energy = 1 2 ˙2 ml β 2 potential energy = mlg(1 − cos β) ◦ mlg where we have assumed β � 1 in the approximation..e. There is no change—the equation is invariant under the transformation. cigarette smoke. car crash. break time-reversal invariance. In the linear approximation.

Then Conclusion: ρ = 0 ∞ Energy conserved (no friction) ρ > 0 ∞ friction (energy is dissipated) ρ < 0 ∞ energy increases without bound 4.2 Phase portrait of damped pendulum dE ˙ = −ρβ2 dt ˙ Let x = β. y = β.Taking the time derivative. Then ˙ x = β=y ˙ ¨ ˙ y = β = −ρβ − γ 2 β = −ρy − γ 2 x ˙ or � x ˙ y ˙ � = � 0 1 2 −γ −ρ �� x y � The eigenvalues of the system are solutions of (−�)(−ρ − �) + γ 2 = 0 Thus ρ 1� 2 ρ − 4γ 2 �=− ± 2 2 24 . dE ˙ ˙ ˙ = ml2 [β(−ρβ − γ 2 β) + γ 2 ββ] dt ˙ = −ml2 ρβ2 Take ml2 = 1. dE 1 ˙¨ ˙ = ml2 (2ββ + 2γ 2ββ) dt 2 ¨ Substituting the damped pendulum equation for β.

θ θ • for ρ = 0. small enough to allow oscillations). θ θ 25 . trajectories spiral outwards and are unstable.e. θ θ • for ρ > 0. weak damping.Assume ρ 2 � γ 2 (i. trajectories spiral inwards and are stable.. There are three generic cases: • for ρ > 0. Then the square root is complex. trajectories are marginally stable periodic oscillations. and we may approximate � as ρ � = − ± iγ 2 The solutions are therefore exponentially damped oscillations of frequency γ: β(t) = β0e−εt/2 cos(γt + θ) β0 and θ derive from the initial conditions.

we summarize the characteristics of dissipative systems: • Energy not conserved. 26 . • Irreversible. but ρ gives the rate. 1 dV ψ ψ =�·f V dt ωx ωy ˙ ˙ + = 0 − ρ = −ρ < 0 ωx ωy The inequality not only establishes area contraction.3 Summary Finally.It is obvious from the phase portraits that the damped pendulum contracts areas in phase space: θ θ We quantify it using the Lie derivative. • Contraction of areas (volumes) in phase space. which yields 4.

we can also have time i. In 3-D the stretching and thinning can be even stranger! 27 .e. we can have expansion in one dimension and (a greater) contraction in the other..Note that the contraction of areas is not necessarily simple. In a 2-D phase space one might expect θ θ time However.

.e. then so is �β(t). Now.5 5. 5. ρ>0 ρ<0 2 for β2 > β0 2 for β2 < β0 28 . Why? (3) Linearity ∞ if β(t) is a solution.2 Van der Pol equation Consider a damping coeﬃcient ρ(β) such that ρ(β) > 0 ρ(β) < 0 Express this in terms of β 2 : ρ(β) = ρ0 for |β| large for |β| small � β2 2 −1 β0 � where ρ0 > 0 and β0 is some reference amplitude. obviously. βmax < α).1 Forced oscillators and limit cycles General remarks How may we describe a forced oscillator? The linear equation ¨ ˙ β + ρβ + γ 2 β = 0 is in general inadequate. This is incompatible with bounded oscillations (i. � real. We therefore introduce an equation with • a nonlinearity. and • an energy source that compensates viscous damping.

we obtain �� � �2 2 ∗ ∗ β β0 dβ∗ 2d β γ ∗2 β0 + ρ0 −1 γβ0 + γ 2β∗ β0 = 0 ∗ dt β0 dt Divide by γ 2 β0: � dβ∗ d2 β∗ ρ0 � ∗2 + β −1 + β∗ = 0 ∗2 ∗ dt γ dt 29 . • high current =∞ positive (ordinary) resistance. The basic behavior: large oscillations decay and small oscillations grow. Substituting above.Substituting ρ into (3). We deﬁne new units of time and amplitude: • unit of time = 1/γ • unit of amplitude = β0. We transform t ∗ t∗ /γ β ∗ β ∗ β0 where β∗ and t∗ are non-dimensional. It was introduced in the 1920’s as a model of nonlinear electric circuits used in the ﬁrst radios. we get d2 β + ρ0 dt2 � � dβ β2 −1 + γ 2β = 0 2 β0 dt This equation is known as the van der Pol equation. First. and • low current =∞ negative resistance. We shall examine this system in some detail. In van der Pol’s (vaccum tube) circuits. we write it in non- dimensional form.

Now deﬁne the dimensionless control parameter π= Finally. we have 2 π(βmax − 1) < 0 ∞ negative damping ρ0 > 0. drop primes to obtain d2 β dβ + π(β2 − 1) + β = 0. γ (4) Thus trajectories spiral outward: θ θ • But when the amplitude of oscillations is large (βmax > 1). 2 π(βmax − 1) > 0 ∞ positive damping The trajectories spiral inward: θ θ 30 . dt2 dt What can we say about the phase portraits? • When the amplitude of oscillations is small (βmax < 1).

make substantial progress with a simple energy balance argument. circadian (daily) cycles in body temperature. the non-dimensional energy is 1 ˙ ˙ E(β. Small perturbations always return to the standard cycle. Using our previous expression for energy in the pendulum. β) = (β2 + β2 ) 2 31 . the limit cycle is an attractor (and is stable). What can we say about the limit cycle of the van der Pol equation? With the help of various theorems (see Strogatz. Examples of such stable oscilations abound in nature: heartbeats (see Fig ure from Glass). it represents stable oscillations. Ch. 5. This is a new kind of attractor. Instead of representing a single ﬁxed point.Intuitively. For π > 0. and take β small.3 Energy balance for small π Let π ∗ 0. 7) one can prove the existence and stability of the limit cycle. etc. we expect a closed trajectory between these two extreme cases: θ θ This closed trajectory is called a limit cycle. We may. however.

we obtain dE ˙ ˙ = πβ2 − πβ2β2 . πβ2. we obtain dE ˙ ˙ ˙ = πβ2 (1 − β2) − ββ + ββ dt ˙ = πβ2 (1 − β2) (5) (6) Now deﬁne the average of a function f (t) over one period of the oscillation: � tO +2� 1 f∈ f (t)dt. Thus ˙ ˙ πβ2 = πβ2β2 or ˙ ˙ β2 = β2β2 . Substituting this into the expression for dE/dt. We know the approximate solution: β(t) = δ sin t. 2α t0 Then the average energy variation over one period is � t0 +2� dE 1 dE = dt. 32 . dt ˙ In steady state. is exactly compensated by the ˙ dissipation of energy. Now consider the limit π ∗ 0 (from above). πβ2β2. we have ¨ ˙ β = −π(β2 − 1)β − β.The time variation of energy is dE 1 ˙¨ ˙ = (2ββ + 2ββ) dt 2 From the van der Pol equation (4). the production of energy. dt 2α t0 dt Substituting equation (6) for dE/dt.

we have the following approxi mate solution for π � 1: β(t) ◦ 2 sin t. Strogatz) shows that this limit cycle is stable. 2α t0 2 The average rate of energy dissipation is � t0 +2� 1 1 ˙ β2β2 ◦ δ4 sin2 t cos2 tdt = δ4 . simple sinusoidal oscillation of unknown amplitude δ. Graphically. we have a limit cycle with an amplitude of 2 dimensionless units. We thus ﬁnd that.. e. We proceed to calculate δ from the energy balance.e.i. θ 2 2 θ 0 θ −2 0 1 t / 2π 2 Further work (see.. independent of π = ρ0 /γ. 33 .g. That is. 2 8 Therefore δ = 2. The average rate of energy production is � t0 +2� 1 1 ˙ β2 ◦ δ2 cos2 tdt = δ2 . 2α t0 8 The energy balance argument gives 1 2 1 4 δ = δ.

.5. ¨ ˙ Notice that ⎡ � �� 1 3 d x −x . .). First. the phase plane coordinates will not be β and β. Write y = w/π. That is.4 Limit cycle for π large The case of π large requires a diﬀerent analysis. However we make one more change of variable. x+π ˙ x + πx(x2 − 1) = ¨ ˙ 3 dt 1 F (x) = x3 − x 3 w = x + πF (x). ˙ ¨ ˙ Substituting the van der Pol equation (7). ˙ Then. ˙ Recall the van der Pol equation (4). we have w = x + πx(x2 − 1). we introduce an unconventional set of phase plane variables (not x = ˙ ˙ y. y = . 212). using (8) and (9). We follow the argument given in Strogatz (p. this gives w = −x ˙ Now rearrange equation (9) to obtain x = w − πF (x) ˙ (11) (10) (8) (9) (7) Let and We have thus parameterized the system by x and w. but write in terms of x = β: x + π(x2 − 1)x + x = 0. Then (10) and (11) become x = π[y − F (x)] ˙ 1 y = − x ˙ π 34 (12) (13) .

suppose y − F (x) � 1. First.) Eventually the trajectory is so close to y = F (x) such that 1 y − F (x) � 2 π implying that 1 x�y� . x � π � 1 ˙ y � 1/π � 1 ˙ assuming x � 1. Thus the horizontal velocity is large and the vertical velocity is small.. ∞ trajectories move horizontally. that is. The y-nullcline shows that the vertical velocity vanishes for x = 0. 35 .e. and the trajectories crawl ˙ ˙ ˙ slowly along the nullcline. Now imagine a trajectory starting not too close to y = F (x).Now consider a trajectory in the x-y plane. i. ˙ ˙ π Thus the trajectory crosses the nullcline (vertically. since x = 0 on the null˙ cline). This ˙ is the cubic curve y = F (x). Then from the equations of motion (12) and (13). draw the nullcline for x. the curve showing where x = 0. Then x changes sign. we still have x � y � 1/π.

What happens at the knee (the minimum of F (x))? The trajectories jump sideways again. avalanches. as may be inferred from the symmetry x ∗ −x. y ∗ −y. etc. bowed violin strings.) 5. Why? 36 ..) • nerve cells.. heart beats ( large literature in mathematical biology. The trajectory closes to form the limit cycle.5 A ﬁnal note Limit cycles exist only in nonlinear systems. Summary: The dynamics has two widely separated time scales: • The crawls: Γt � π • The jumps: Γt � 1/π (x � 1/π) ˙ (x � π) ˙ A time series of x(t) = β(t) shows a classic relaxation oscillation: Relaxation oscillations are periodic processes with two time scales: a slow buildup is followed by a fast discharge. Examples include • stick-slip friction (earthquakes.

� �= 0. The amplitude of a limit cycle. 37 . That is. but not an isolated x x orbit. x x Thus the amplitude of a periodic cycle in a linear system depends on the initial conditions.˙ A linear system ψ = Aψ can have closed periodic orbits. so is �ψ (t). linearity requires that if ψ (t) is a solution. however. is independent of the initial condi tions.

But is the rest state stable? We investigate the stability of the rest state for a special case: g(t) periodic and sinusoidal: g(t) = g0 + g1 cos(2γt) Substituting into the equation of motion then gives d2 β 2 + γ0 [1 + h cos(2γt)] β = 0 2 dt 38 (14) . Speciﬁcally. imagine subjecting the pivot of a simple frictionless pendulum to an alternating vertical motion: rigid rod This is called a “parametric pendulum.1 Parametric oscillator Mathieu equation We now study a diﬀerent kind of forced pendulum.6 6.” because the motion depends on a time-dependent parameter. that the rest state ˙ β=β=0 is a solution. We know. Consider the parametric forcing to be a time-dependent gravitational ﬁeld: g(t) = g0 + λ(t) The linearized equation of motion is then (in the undamped case) d2β g(t) + β = 0. however. dt2 l The time-dependence of g(t) makes the equation hard to solve.

Any solution β(t) is a linear combination of two linearly independent solutions β1(t) and β2(t): β(t) = Aβ1(t) + Bβ2(t) where A and B come from initial conditions. We begin with two observations: 1. therefore. We pause. because of the time-dependent parametric forcing. Our previous methods are unapplicable. p. then if β(t) is a solution. known as Floquet theory. the natural. unexcited period of the pendulum is Texc = Tnat = 2α γ0 We wish to characterize the stability of the rest state. second-order.where 2 γ0 = g0 /l and Equation (14) is called the Mathieu equation.2 Elements of Floquet Theory Reference: Bender and Orszag. The excitation (forcing) term has amplitude h and period 2α α = 2γ γ On the other hand. however. 2.) (15) (Reason: the system is linear and 39 . 560. 6. If the coeﬃcients are periodic with period T . h = g1 /g0 � 0. so is β(t + T ). We seek to determine the conditions for stability. to consider the theory of linear ODE’s with periodic coeﬃcients. We consider the general case of a second-order linear ODE with periodic coeﬃcients.

Since β1 (t) and β2(t) are solutions, periodicity ∞ β1 (t + T ) and β2(t + T ) are also solutions. Then β1(t + T ) and β2(t + T ) may themselves be represented as linear com binations of β1(t) and β2 (t): β1(t + T ) = �β1 (t) + λβ2(t) β2(t + T ) = ρβ1(t) + νβ2 (t) Thus β(t + T ) = A [�β1 (t) + λβ2(t)] + B [ρβ1(t) + νβ2(t)] = (A� + Bρ)β1(t) + (Aλ + Bν)β2(t) We rewrite the latter expression as β(t + T ) = A∗ β1(t) + B ∗ β2(t) where or � A∗ B∗ � = � � ρ λ ν � A B � � A ψ= a . B �� (16)

ψ ∗ = Mψ , a a � A B �

Now choose value. Then

to be an eigenvector of M , with � the associated eigen-

Any other ρ would have a projection onto one of the eigenvectors. a

A∗ = �A Using (16) and (15), we ﬁnd that

and

B ∗ = �B.

**β(t + T ) = �β(t). Thus β(t) is periodic within a scale factor �. The question of stability then hinges on the magnitude of �.
**

40

Deﬁne µ= Then

ln |�| T

∞

� = eµT .

(We are interested only in growth or decay, not oscillations due to the exponential multiplier.)

β(t + T ) = eµT β(t). Here β is rescaled by eµT each period. Finally, deﬁne P (t) to be periodic such that P (t + T ) = P (t). The rescaling may now be expressed continuously as β(t) = eµt P (t).

[The above � �(t + T ) = eµ(t+T ) P (t + T ) = eµT eµt P (t) = eµT �(t).]

Stability thus rests on the sign of µ. Thus we ﬁnd that the solution to a linear second-order ODE with periodic coeﬃcient—e.g., the Mathieu equation—is of the form (exponential growth or decay) � (periodic function of time) 6.3 Stability of the parametric pendulum

We proceed to determine under what conditions the rest state of the Mathieu equation is unstable, leading to exponentially growing oscillations. We expect the tendency toward instability to be strongest when the excitation frequency is twice the natural frequency, γ0 , of the pendulum.

**This is called parametric resonance.
**

41

**The Mathieu equation (for β � 1) is d2 β 2 + γ0 [1 + h cos(2γt)] β = 0. 2 dt We take the forcing frequency to be 2γ0 + π, π � 1:
**

2 ¨ β + γ0 [1 + h cos(2γ0 + π)t] β = 0.

(17)

We assume that the excitation amplitude h � 1, and seek solutions of the general form � � � � 1 1 β(t) = a(t) cos γ0 + π t + b(t) sin γ0 + π t. 2 2 The frequency γ0 + 1 π is motivated by our resonance argument. Stability will 2 depend on whether a(t) and b(t) exponentially grow or decay. We begin by substituting β(t) into the equation of motion. We will then retain terms that are linear in π and ﬁrst-order in h. The calculation is messy but is aided by recalling trig identities like � ⎡� � � ⎡ � � � ⎡� � � 1 1 1 1 1 cos γ0 + π t cos [(2γ0 + π) t] = cos 3 γ0 + π t + cos γ0 + π t . 2 2 2 2 2 � ⎜� �

higher freq term

The term with frequency 3(γ0 + 1 π) may be shown to be higher order with 2 respect to h. Such higher frequency terms create small amplitude perturba tions of the solution and are neglected. (You shall see on a problem set how such terms arise from nonlinearities.) Also, we shall retain only those terms that are ﬁrst-order in π. Thus we ¨ neglect the O(π2 ) associated with accelerations β. We also assume a � πa, ˙ ˙ b � πb,

**and thereby neglect a, ¨ These assumptions are validated by the ﬁnal result. ¨ b.
**

�

That is, we use cos A cos B = 1 [cos(A + B) + cos(A − B)] and sin A sin B = 1 [sin(A + B) + sin(A − B)]. 2 2

42

The result of the substitution is � � ⎡� � � 1 1 − 2a + bπ + hγ0 b γ0 sin γ0 + π t ˙ 2 2 � � ⎡� � � 1 1 ˙ + 2b − aπ + hγ0 a γ0 cos γ0 + π t = 0 2 2 For this expression to be true for all t. Thus � � �2 � 1 2 1 −µ2 − π − hγ0 =0 4 2 or or µ2 = 1 4 �� 1 hγ0 2 �2 − π2 � (18) Instability (parametric resonance) occurs when µ is real and positive.e. or 1 1 − hγ0 < π < hγ0 . the coeﬃcients of both cos and sin must equal zero. substitute a = µa. 1 2a + bπ + hγ0 b = 0 ˙ 2 1 ˙ 2b − aπ + hγ0 a = 0 2 We seek solutions a(t) ≥ eµt and b(t) ≥ eµt . This will occur when µ2 > 0. ˙ Thus. 2 2 43 . i. b = µb: ˙ � � 1 1 µa + b π + hγ0 = 0 2 2 � � 1 1 a π − hγ0 − µb = 0 2 2 These equations have a solution when the determinant of the coeﬃcients of a and b vanishes.

equation (18)). but by µ − ρ = 0. 6. the less necessary it is for the force to be exactly resonant (γ = 0). The larger the forcing amplitude. For h > 0.4 Damping The damped parametric pendulum is 2 ¨ ˙ β + 2ρβ + γ0 [1 + h cos((2γ0 + π)t)] β = 0 For unforced oscillations (h = 0).This result is summarized in the following phase diagram: ε h ω0 /2 unstable h exact resonance −h ω0 /2 Thus we see that resonance occurs not only for π = 0.. h) is suﬃcient for instability so long as π ∗ 0..e. we expect where the factor eµt results from the periodic forcing. 44 β(t) � e(µ−ε)t × (oscillation) . The instability boundary is therefore no longer given by µ = 0 (i. damping produces solutions like β(t) ◦ e−εt × (oscillation). but for a range of π within the lines ±hγ0 /2.e. inﬁnitesimal forcing (i. Conversely.

we obtain �� � �2 1 1 µ2 = hγ0 − π2 > ρ 2 4 2 or − �� 1 hγ0 2 �2 − 4ρ 2 �1/2 < π < �� 1 hγ0 2 �2 − 4ρ 2 �1/2 Setting π = 0 (exact resonance) shows that instability is only possible when the quantity in the brackets is positive. Using equation (18).5 Further physical insight We have seen that the instability is strongest at exact resonance. That is. the most unstable excitation has period 1 Texc = Tnat 2 corresponding to angular frequencies γexc = 2γ0 This is called a subharmonic instability. 45 . where 4ρ hc = γ0 The phase diagram is therefore modiﬁed accordingly: ε h ω0 /2 unstable hc h − h ω0 /2 6. because instability results at half the frequency of the excitation (i.e. if the natural period of the pendulum is Tnat . γ0 = γexc/2). Thus we require h > h c ..Instability thus occurs for µ2 > ρ 2.

pushing a child on a swing every nth time he/she arrives at maximum height. say.What other periods of excitation would be expected to lead to instability for small h? Any integer multiple of Tnat suﬃces. Therefore (weaker) instabilities will occur for � � 1 Texc = integer × Tnat . Why? Because this would correspond roughly to. 2 46 .

1 Continuous Fourier transform We ﬁrst state the Fourier transform for functions that are continuous with time. idealized cases (such as the linear pendulum). the precise oscillatory nature of an observed time series x(t) may not be identiﬁed from x(t) alone. the inverse Fourier transform is � � 1 f (t) = � F (γ)eiσtdγ. 7. We may ask • How well-deﬁned is the the dominant frequency of oscillation? • How many frequencies of oscillation are present? • What are the relative contributions of all frequencies? The analytic tool for answering these and myriad related questions is the Fourier transform. below. but we save that calculation for the discrete transform. 2α −� That the second relation is the inverse of the ﬁrst may be proven.7 Fourier transforms Except in special. 47 . The Fourier transform of some function f (t) is � � 1 F (γ) = � f (t)e−iσtdt 2α −� Similarly.

x Tmin t We require at least two samples per period. we must know the highest frequency. In modern data. 1. Therefore Γt ∼ 1 Tmin = . Then A practical question concerns the choice of Γt.2 Discrete Fourier transform We are interested in the analysis of experimental (or numerical) data. . To choose it. . The discretization process is pictured as x x(t) Δt j−1 j j+1 t We take the sampling interval—the time between samples—to be Γt. contained in x(t). fmax . . 2. The shortest period of oscillation is Tmin = 1/fmax Pictorially. . one almost always observes a discretized signal xj . n − 1} xj = x(jΓt). 2 2fmax 48 . j = {0.7. Thus we specialize to discrete Fourier transforms. which is almost always discrete.

i. . k oscillations per tmax . within a factor of 1/ n.. . j = 0. • k = 0 ∞ exp(−i2αjk/n) = 1. 1. . . 1. • The frequency indices k = 0. . what is the signiﬁcance of xk for k > n/2? ˆ 49 . . equal to the mean of the xj ’s. . Index k = n/2 then corresponds to �n� � 1 � 1 fmax = = 2 nΓt 2Γt But if n/2 is the highest frequency that the signal can carry. . 1. n/2 correspond to frequencies fk = k/tmax. the period of observation. . consider the range of the exponential multiplier. What would x0 �= 0 mean? ˆ • k = n/2 ∞ exp(−i2αjk/n) = exp(−iαj). . Frequency index n/2 is clearly the highest accessible frequency. . n − 1 is � � n−1 1 � 2αjk xk = � ˆ xj exp −i k = 0. . n − 1 n n j=0 To gain some intuitive understanding. Then 1 � x0 = � ˆ xj n j � Thus x0 is. .The discrete Fourier transform (DFT) of a time series x j . Question: Suppose a seismometer measures ground motion.e. Then 1 � xn/2 = � ˆ xj (−1)j n j (19) (20) = x 0 − x 1 + x2 − x 3 . ˆ This is the “DC” component of the transform. .

We derive this relation as follows. frequency indicies k > n/2 are redundant. z � = a − ib).e. being related by xk = x� ˆ ˆn−k where z � is the complex conjugate of z (i.3 Inverse DFT The inverse DFT is given by � � n−1 1 � 2αjk xj = � xk exp +i ˆ n n k=0 j = 0. n − 1 50 . . . ˆ−k ˆ� ˆ The frequency indicies k > n/2 are therefore sometimes referred to as negative frequencies 7. and the last line again employs the deﬁnition of the DFT.For real xj . Note that we also have the relation x� = xn−k = xk . . From the deﬁnition of the DFT.. 1. . we have � � n−1 2αj(n − k) 1 � x� ˆn−k = � xj exp +i n n j=0 � � n−1 −i2αjk 1 � = � xj exp (i2αj) exp n n j=0 � ⎜� � 1 � � n−1 1 � −i2αjk xj exp = � n n j=0 = xk ˆ where the + in the ﬁrst equation derives from the complex conjugation. if z = a + ib.

except when j ∗ = j. consider j ∗ − j = 1. The sum over the circle always sums to zero. Finally.. n = 4. We begin by substituting the DFT for xk . j ∗ = j 0. i. as illustrated below: x(t) −tmax 0 tmax 2 tmax t 51 . This means that a ﬁnite time series is treated precisely as if it were recurring. so that xj+n = xj . the � � n−1 � 2αk(j ∗ − j) exp −i = e0 + e−i�/2 + e−i� + e−i3�/2 n k=0 = 1+i−1−i = 0.We proceed to demonstrate this inverse relation. j ∗ = j � = 1 (nxj ) n = xj � The third relation derives from the fact that the previous k amounts to a sum over the unit circle in the complex plane. note that the DFT relations imply that xj is periodic in n. For example. The elements of the sum are then just the four points on the unit circle that intersect the real and imaginary axes.e. using dummy variable j ∗ : ˆ ⎭ ⎣ � � � � n−1 n−1 ∗ 1 � 1 �� 2αj k ⎤ 2αkj � xj � exp −i xj = � exp +i n n � n n k=0 j =0 � � n−1 n−1 � 1� 2αk(j ∗ − j) = xj � exp −i n � n j =0 n−1 k=0 1� = xj � × n � j =0 � n.

power spectra.4 Autocorrelations. for example. if mΓt is much less than the dominant period of the data. Then n−1 where 1� 2 η0 = x. m = 0. n j=0 j which is the mean squared value of xj (i.e. Some intuition may be gained as follows: • Consider. and the Wiener-Khintchine theorem Assume that the time series xj has zero mean and that it is periodic. Thus it measures the predictability of the future from the past. |ηm | is relatively small. • Alternatively. i. xj+n = xj .e..7. ηm should not be too much less than η0 . its variance). Deﬁne the autocorrelation function η: 1� ηm = xj xj+m n j=0 ηm = η(mΓt) The autocorrelation function measures the degree to which a signal resembles itself over time. if mΓt is much greater than the dominant period of the data. A typical ηm looks like Ψm n−1 m 52 . • Last..

. k = n − k � = 0. ˆ ˆk ˆ ηm ≥ Fourier transform of the power spectrum xk x� = |xk |2 � � n− 1 1 � 2αmk = xk x� exp −i ˆ ˆk n n k=0 For real time series {xj }. � � � � n− 1 n− 1 n− 1 ∗ ∗ 2αmk � 1 � � 2αj(k + k ) xk xk� exp i ˆ ˆ = 2 exp i n n n j=0 k=0 k � =0 � � ⎜� ∗ = n. � � � � 2 n−1 �� 1 � 2αjk � � 2 | ˆk | = � x xj exp −i � .e. k ∗ = n − k � � n− 1 1 � 2αm(n − k) = xk xn−k exp i ˆ ˆ n n k=0 In the last line we have used the redundancy relation x� = xn−k . i. Substitute the inverse DFT for xj in ηm : � � � �� �� � n−1 n−1 n− 1 ∗ � 1 � � 2αkj 1 2αk (j + m) 1 � � ηm = xk exp i ˆ xk� exp i ˆ n n � n j=0 n n k=0 k =0 autocorrelation ≥ Fourier transform of the power spectrum. We proceed to derive this relation. ˆ ˆ ˆk ˆ ˆ ˆn−k ˆ ˆ 53 . This is called the the Wiener-Khintchine theorem. ˆk ˆ We thus ﬁnd that Of course the inverse relation holds also. n � n � j=0 We proceed to show that for real time series xj . the power spectrum contains redundant information that is similar to that of the Fourier transform but more severe: |xk |2 = xk x� = xk xn−k = x� xn−k = |xn−k |2.Deﬁne the power spectrum to be the magnitude squared of the Fourier transform.

7. . 7.This redundancy results from the fact that neither the autorcorrelation nor the power spectra contain information on any “phase lags” in either x j or its individual frequency components. 1/T. tmax 54 . Thus while the DFT of an n-point time series results in n independent quan tities (2 ×n/2 complex numbers). . x(t) is a sine or cosine.e. the power spectrum yields only n/2 inde pendent quantities. . Consequently a time series cannot be reconstructed from its power spectrum or autocorrelation function..5. but that each time series uniquely deﬁnes its Fourier transform. � � 2αt x(t) = sin . at frequencies 0.1 Sinusoidal signal In the simplest case. . (n − 1)/T. 2/T.5 Power spectrum of a periodic signal Consider a periodic signal � � 2α x(t) = x(t + T ) = x t + γ Consider the extreme case where the period T is equal to the duration of the signal: T = tmax = n�t The Fourier components are separated by 1 Γf = tmax i. i. One may therefore show that there are an inﬁnite number of time series that have the same power spectrum. and vice-versa.e.

beginning with the DFT: � � −i2αjk 1 � xk = � ˆ xj exp n n j � � � � 1 � 2αjΓt −i2αjk = � sin exp n j tmax n ⎡ � � � �� � � 1 � i2αjΓt −i2αjΓt −i2αjk � = exp − exp exp 2i n j tmax tmax n ⎡ � � �� � � ��� 1 � Γt k Γt k � = exp i2αj − − exp −i2αj + tmax n tmax n 2i n j � n ±nΓt = ± when k = 2i tmax Thus |xj |2 = ˆ 7.g. we expect x(t) xk 2 tmax t 1/T k Δ f = k/T We proceed to calculate the power spectrum analytically. a relaxation oscillation as obtained from the van der Pol system. Consider now a non-sinusoidal yet periodic signal. e.2 Non-sinusoidal signal n 4 for k = ±1..5.What is the Fourier tranform? Pictorially. 55 .

but spaced p samples apart on the frequency axis. The non-zero components of the power spectrum must still be at frequencies 1/T. . pictorially. 2/T. Γf = = But since 1 7. . Thus. integer multiples of the fundamental frequency 1/T .. 1 tmax pT the frequency resolution is p times greater. .The non-sinusoidal character of the relaxation oscillation implies that it con tains higher-order harmonics.3 tmax /T �= integer If tmax /T is not an integer. the (eﬀectively periodic) signal looks like x(t) tmax t 56 . .5. Contributions to the power spec trum would remain at integer multiples of the frequency 1/T .e. i. we expect x(t) xk 2 fundamental harmonics tmax t 1/T 2/T 3/T k Δ f = k/T Now suppose tmax = pT . where p is an integer.

We proceed to calculate the power spectrum of such a signal. T n Recall the identity 1 � xk = � ˆ exp (i2αθk j) n j=0 n−1 � n−1 xn − 1 x = x−1 j=0 j Then The power spectrum is 2 1 exp(i2αθk n) − 1 xk = � ˆ n exp(i2αθk ) − 1 xk x � ˆ ˆk 1 = n 1 − cos(2αθk n) 1 − cos(2αθk ) � � 1 sin2 (αθk n) = n sin2 (αθk ) � � |xk | = ˆ nΓt tmax −k = −k T T is the diﬀerence between a DFT index k and the “real” non-integral frequency index tmax /T . Assume the sinusoidal function � � 2αt x(t) = exp i T which yields � � 2αjΓt xj = exp i T The DFT is � � � � n−1 1 � 2αjΓt 2αjk xk = � ˆ exp i exp −i T n n j=0 Set θk = Then Γt k − . nθk = 57 Note that .

so we may also assume αθk � 1. let k0 be the integer closest to tmax /T . as a function of z = nαθk it ˆ appears as � nΓt −k T sin 2 z / z2 1 −3π −2π −π 0 π 2π 3π z=n πφ k The plot gives the kth component of the power spectrum of e i2�t/T as a function of α(tmax /T − k). T Consequently θk � 1. Instead. Then ˆ |xk | 2 1 sin2 (αθk n) ◦ n (αθk )2 sin2 (αθk n) = n (αθk n)2 ≥ sin2 z z2 � =α � tmax −k T � where z = nαθk = α Thus |xk |2 is no longer a simple spike. There are then two extreme cases: 58 .Assume that n is large and k is close to that “real” frequency index such that nθk = nΓt − k � n. To interpret the plot.

T The spectrum is perfectly sharp: xk 2 sin 2z/z k0 k 0 z 2.1. Then tmax 1 − k0 = . T 2 The spectrum is smeared: xk 2 sin 2z/z k0 The smear decays like k 1 (k − tmax /T )2 1 k2 0 z � 7.4 Conclusion The power spectrum of a periodic signal of period T is composed of: 1.5. a peak at the frequency 1/T 2. tmax is an integral multiple of T: tmax − k0 = 0. smears near the harmonics. 59 . tmax /T falls midway between two frequencies. possibly harmonics (integer multiples) of 1/T 4. a smear (sidelobes) near 1/T 3.

t2 . . . γj fj Example: The astronomical position of a point on Earth’s surface changes due to • rotation of Earth about axis (T1 = 24 hours). . ..e. tr ). • we ignore precession and other orbital changes. t2 . Mathematically. . we can conceive of such a function on a 2-D torus T 2 . . . r y is called quasiperiodic if each tj varies with time at a diﬀerent rate (i. . . tr ). . . . j = 1.7. tj . . We have then tj = γj t. .6 Quasiperiodic signals Let y be a function of r independent variables: y = y(t1. if y(t1 . tj + 2α. . r. . The quasiperiodic function y has r fundamental frequencies: γj fj = 2α and r periods 1 2α Tj = = . tr ) = y(t1 . . . . . . t2. T2 T1 60 . y is periodic. . diﬀerent “clocks”). . of period 2α in each argument. . . existing in a 3-D space. j = 1. • revolution of Earth around sun (T2 ◦ 365 days). . . .

For example: x(t) = r � i=1 xi(γit). . . . . f2 = γ2/2α. (m1 . 2 2 and 61 The fundamental frequencies are |f1 − f2| |f1 + f2|. For example. The quasiperiodic signal is a linear combination of independent periodic functions. . . 2. Because the Fourier transform is a linear transformation. . x(t) = sin(2αf1t) sin(2αf2t) = 1 1 cos(|f1 − f2|2αt) − cos(|f1 + f2|2αt). Such behavior can be conceived as a trajectory on the surface of a doughnut or inner tube. f1 f2 What is the power spectrum of a quasiperiodic signal x(t)? There are two possibilities: 1. . the power spec trum of x(t) is a set of peaks at frequencies f1 = γ1/2α. The quasiperiodic signal x(t) depends nonlinearly on periodic functions. m2. . positive integers).Here we think of a disk spinning with period T1 while it revolves along the circular path with period T2. or a torus T2 in R3 . m 2 f2 . . and their harmonics m1 f1 .

� f2 � f1 integer . (21) The nonlinear case requires more attention. m2 positive integers. � f2 � 62 . There are two cases: 1. � f2 � � f2 � f2 Thus the peaks of the spectrum must separated (i. . Then � � � � � f1 � � integer � 1 �m1 + m2 � = � + integer� = integer multiple of . In the rational case. then m1 and m2 may always be chosen so that � � � f1 � �m1 + m2 � is not similarly restricted. = integer f2 m1 . f1/f2 rational ∞ sparse spectrum. for which equation (21) is a speciﬁc example. γ2t) is a peak at f = |m1 f1 + m2 f2 |. The expression above We proceed to specialize to r = 2 frequencies.e.. if f1/f2 is irrational. . with mi positive) mi arbitrary integers. Each nonzero component of the spectrum of x(γ1t. m2 integers . + mr fr |. m1 . and forget about ﬁnite Γf . . derives from m1 f1 ± m2 f2 ± . Alternatively.The harmonics are m1 |f1 −f2| and m2 |f1 +f2 |. if x(t) depends non linearly on r periodic functions. To understand this. rewrite f as � � � f1 � f = f2 �m1 + m2 � . In general. then the harmonics are |m1 f1 + m2 f2 + . 2. . f1/f2 irrational ∞ dense spectrum. sparse).. (This is the most general case.

n2 f2 Since n1. 7. f1 n1 = . and some writers restrict quasiperiodicity to the irrational case.. the signal never exactly repeats itself. it would only be possible to avoid stepping on a crack if the ratio step size crack width were rational. Note further that. In the rational case. the case of a child walking on a sidewalk.7 Aperiodic signals Aperiodic signals are neither periodic nor quasiperiodic. 63 . n2 integers. Then if x(t) were the distance from the closest crack at each step. n1 n2 = f1 f2 the quasiperiodic function is periodic with period T = n 1 T1 = n 2 T2 . All spectral peaks must then be harmonics of the fundamental frequency f0 = 1 f1 f2 = = . as an example.). T n1 n2 Thus the rational quasiperiodic case is in fact periodic. One may consider. in the irrational case. attempting with uniform steps to never step on a crack (and breaking his mother’s back.These distinctions have further implications..

Rather than considering only one white-noise signal. n − 1 64 . . An example is white noise. where the superscipt denotes the particular realization within the ensemble. e. x(1) (t). and therefore uncorrelated to them.. x(t) t Statistically. ﬂat: xk 2 k The ﬂat spectrum of white noise is a consequence of its lack of harmonic structure (i. i. Each realization is independent of the others. . x(2)(t). or dominant fre quency). on average. ..Aperiodic signals appear random.e. j = 0. We proceed to derive the spectrum of a white noise signal x(t).g. . which is a signal that is “new” and unpredictable at each instant. though they may have a deterministic foun dation. . Now discretize each signal so that xj = x(jΓt).. one cannot recognize any particular tone. we consider an ensem ble of such signals.e. each sample of a white-noise signal is independent of the others. . The power spectrum of white noise is. .

e. n j=0 By deﬁnition. assume ⇒x≡ = 0. the mean-square value of the jth sample is ⇒ 2 ≡ xj 1 � (i) 2 = lim (xj ) p�� p i=1 p p Now assume stationarity: ⇒xj ≡ and ⇒ 2≡ are independent of j. Recall the autocorrelation ηm : 1� ηm = xj xj+m . each sample of white noise is uncorrelated with its past and future. We use angle brackets to denote ensemble averages (i. We take these x j 2 mean values to be ⇒x≡ and ⇒x ≡. The ensemble-averaged mean of the jth sample is then 1 � (i) ⇒ j ≡ = lim x xj p�� p i=1 Similarly. Therefore � ⎝ � 1 ⇒ηm ≡ = xj xj+m n j = ⇒x2 ≡νm where νm = � 1 m = 0 0 else n− 1 65 ..We take the signal to have ﬁnite length n but consider the ensemble to contain an inﬁnite number of realizations. respectively. averages taken over the ensemble).

. Thus for white noise. but only beyond some interval Γt. the spectrum is indeed ﬂat. but not constant: xk 2 k In such (red) colored spectra. as previously indicated: xk 2 k A more common case is “colored” noise: a continuous spectrum. The autocorrelation of colored noise is broader.We obtain the power spectrum from the autocorrelation function by the Wiener-Khintchine theorem: � � n−1 � 2αmk 2 ⇒|xk | ≡ = ˆ ⇒ηm ≡ exp i n m=0 2αmk = ⇒x ≡νm exp i n m=0 2 n−1 � � � = ⇒x2 ≡ = constant. Ψm m 66 .g. The signal is still apparently random. there is a relative lack of high frequencies. e.

Thus we turn to Poincar´ sections. e 67 . random systems.Finally. but they cannot distinguish between deterministic systems and statistical. we note a problem: power spectra can recognize a signal that is approximately aperiodic.

8. e 68 . if for no other reason than that the dimensionality of the phase space is too large. The solution is often studied by considering the trajectories of such ﬂows.1 Construction of Poincar´ sections e Suppose we have a 3-D ﬂow �. • The points P0 .8 Poincar´ sections e The dynamical systems we study are of the form d ψ (t) = F (ψ . e. its intersection with a plane (x3 = h): x3 Γ S P0 h x2 P1 P2 x1 • Points of intersection correspond (in this case) to x3 < 0 on �. Thus we seek a geometric depiction of the trajectories in a lower-dimensional space—in essence. t) x x dt Systems of such equations describe a ﬂow in phase space. consider. P2 form the 2-D Poincar´ section. P1 . a view of phase space without all the detail.g. ˙ • Height h of plane S is chosen so that � continually crosses S. Instead of directly studying the ﬂow in 3-D. But the phase trajectory is itself often diﬃcult to determine..

. and aperiodic.The Poincar´ section is a continuous mapping T of the plane S into itself: e Pk+1 = T (Pk ) = T [T (Pk−1)] = T 2 (Pk−1) = .. attractor. etc. P1 determines P2 . e Essentially the Poincar´ section provides a means to visualize an otherwise e messy. we classify three types of ﬂows: periodic. We expect some geometric properties of the ﬂow and the Poincar´ section to e be the same: • Dissipation ∞ areas in the Poincar´ section should contract. However the time interval from point to point is not necessarily con stant. P0 determines P1 . Since the ﬂow is deterministic. e • If the ﬂow has an attractor. we should see it in the Poincar´ section. . quasiperiodic. possibly aperiodic. a limit cycle): P0 69 .1 Periodic The ﬂow is a closed orbit (e.2.2 Types of Poincar´ sections e As we did with power spectra. The Poincar´ section reduces a continuous ﬂow to a discrete-time map e ping. 8. 8.g.

Mij = ωxj �P0 A Taylor expansion about the ﬁxed point yields: � � ωTi � ωTi � � · ν1 + � · ν2.P0 is a ﬁxed point of the Poincar´ map: e P0 = T (P0 ) = T 2(P0 ) = . We proceed to analyze the stability of the ﬁxed point. a Poincar´ map T can be described by a matrix M deﬁned in e the neighborhood of P0 : � ωTi � � . M is called a Floquet matrix. Stability therefore depends on the properties of M . e i = 1. 2 Since T (P0) = P0 . Ti (P0 + ν) ◦ Ti(P0 ) + ωx1 �P0 ωx2 � 0 P T (P0 + ν) ◦ P0 + Mν � In this context. It describes how a point P 0 + ν moves after one intersection of the Poincar´ map. 70 will always be a projection onto an . . To ﬁrst order. T m (P0 + ν) − P0 ◦ M m ν. (There eigenvector. where � is the corresponding eigenvalue.) Then M m ν = �m ν. Therefore T T (P0 + ν) � ◦ T (P0 + Mν) ◦ T (P0) + M 2 ν ◦ P0 + M 2 ν After m interations of the map. Assume that ν is an eigenvector of M. . .

e. the form of the resulting Poincar´ section depends on e the ratio f1 /f2: • Irrational f1/f2.Therefore |�| < 1 ∞ linearly stable |�| > 1 ∞ linearly unstable Conclusion: a periodic map is unstable if one of the eigenvalues of the Floquet matrix crosses the unit circle in the complex plane.g. The closed curve C appears continuous. As with power spectra. The ﬂow is a torus T 2 : The points of intersection of the ﬂow with the plane S appear on a closed curve C. 71 .2. The frequencies are called incommensurate. 8.2 Quasiperiodic ﬂows Consider a 3-D ﬂow with two fundamental frequencies f1 and f2 .

– The curve is not traversed continuously. Thus e Pi = T n1 (Pi ) – Example. – f1 and f2 are frequency locked. – The Poincar´ section is periodic with e period = n1/f1 = n2/f2 – The Poincar´ section contains just n1 points. – Trajectory repeats itself after n1 revolutions and n2 rotations. but rather T (C) = ﬁnite shift along C.2. 72 .3 Aperiodic ﬂows Aperiodic ﬂows may no longer lie on some reasonably simple curve. – There are ﬁnite number of intersections (points) along the curve C. • Rational f1/f2. n1 = 5: P0 P3 P2 P1 P4 8.x2 C x1 – The trajectory on the torus T 2 never repeats itself exactly.

We shall give particular attention to the following quadratic mapping of the unit interval onto itself: xk+1 = 4µxk (1 − xk ).In an extreme case.g. In some cases they create mild departures from a simple curve. Deterministic aperiodic systems often display more order. Such maps are of the form xk+1 = f (xk ). 8. Such cases arise from strong dissipation (and the resulting contraction of areas in phase space). however. We will study these extensively at the end of the course.3 First-return maps First return maps are 1-D reductions of the kind of 2-D Poincar´ maps that e we have been considering. one has just a point cloud: This would be expected for statistical white noise. 73 0 ∼ µ ∼ 1. and to study the iterates of x. It then becomes useful to deﬁne a coordinate x that falls roughly along this curve. . e. This is called a ﬁrst return map.

Thus x = 0 is an unstable ﬁxed point.e. iterates initiated at ¯ x0 = π still converges to x. falls to x = 1. Thus x is a ﬁxed point of f . However we can see graphically that x = 0 is unstable. since f (0) = 0. where it rises to height µ. and has its maximum at x = 1/2. The quadratic rises from x = 0.. which is where the diagonal ¯ (the identity map xk+1 = xk ) intersects f (x). ¯ x = f (¯) ¯ x Another ﬁxed point is x = 0.The mapping is easily described graphically. the case µ = 0.9: 74 . Consider.7: f(x) x k 1 x k+ 1 f(x) f(x0) = x1 x0 x1 = x 1 x Eventually the interations converge to x = x. ¯ What determines stability? Consider graphically the case µ = 0. i. for example. while x = x is stable.

We proceed to x ¯ show this formally. Suppose x� is any ﬁxed point such that x� = f (x�).f(x) 1 1 x We infer that the slope f ∗(¯) determines whether x is stable. Deﬁne xk = x� + πk . our mappings are described by xk+1 = f (xk ). |f ∗ (x�)| < 1 ∞ stability. Then x� + πk+1 = f (x� + πk ) = f (x�) + f ∗(x�)πk + O(π2 ) k Therefore Thus πk+1 ◦ f ∗ (x�)πk . . In general. 75 πk small.

e. ˙ Consider. whereas the magnitude |�| is the relevant quantity for maps. for example. the sign of x = f (x). a particular f (x) that looks like x f(x) x x1 x2 Clearly x1 and x2 are ﬁxed points.. the stability of a 1-D map depends on the magnitude |f ∗ |.4 1-D ﬂows It is instructive to compare the stability of 1-D maps to the stability of the continuous 1-D ﬂow x = f (x). ˙ Thus x1 is stable while x2 is not. the sign of Re(�) determines the stability of ﬂows. i. in the case of mappings. we have called the Floquet matrix). Stability of a ﬁxed point x� is therefore determined as follows: f ∗(x�) < 0 ∞ stable f ∗(x�) > 0 ∞ unstable Whereas the stability of a continuous 1-D ﬂow f depends on the sign of f ∗. 76 .8. In higher dimensions this same distinction holds for the eigenvalues � of the Jacobian (which. That is. Which are stable? The arrows show the direction of ﬂow on the x-axis.

8. 77 .8. There is then some mapping f such that xk+1 = f (xk ). We draw a ray emanating from the origin.1 Example 1: the van der Pol equation Consider again the van der Pol equation d2 β dβ + π(β2 − 1) + β = 0 dt2 dt Recall that for π > 0 the rest position is unstable and that the system has a limit cycle.5.5 Relation of ﬂows to maps We now consider explicitly how ﬂows may be related to maps. and consider two representative trajectories initiating and terminating on it: Let xk be the position of the kth intersection of the trajectory with the ray.

but physical and mathematical reason ing allows us to state some of its properties: • f maps xk to a unique xk+1. Thus x� gives the eﬀective radius of the stable limit cycle.2 Example 2: R¨ssler attractor o Consider the following 3-D ﬂow (the R¨ssler attractor): o x = −y − z ˙ y = x + ay ˙ z = b + z(x − c) ˙ 78 . followed by a gentle upward slope: f(x) xk+1 = x k f(x*) x* x By continuity. there must be a stable ﬁxed point x� characterized by x� = f (x�) and f ∗ (x�) < 1.5. • f ∗(x) > 1 near the origin (divergent spirals).The precise form of f (x) is unknown. • f ∗(x) < 1 far from the origin (convergent spirals). • f ∗(x) > 0 for all x > 0 (since f (x + ν) > f (x)). 8. • f is continuous. The simplest form of f is therefore a curve rising steeply from the origin.

12a. Consider now another Poincar´ section. • The time series. i.10. Numerical solutions yield the time series x(t): Figure IV. we identify this plane with extrema in the time o series x(t). and c are ﬁxed parameters. BPV Conclusions: • The 1-D map. ˙ Consider a sequence xk of such extrema.11. but only when the extremum is a maximum (local peak) of x(t). 79 . xk : Figure IV. however. Then plot xk+1 vs. BPV A Poincar´ section in the x-y plane: e Figure IV. unknown source The time series display great irregularity. in the plane e y + z = 0. From the R¨ssler equations. each intersection of the plane corresponds to x = 0.e. BPV The time series z(t): Figure IV.11.1.. but the Poincar´ section and the e full ﬂow display some order. BPV And a 3-D perspective of the ﬂow: Figure 12. reveals that the e ﬂow contains much order. even more so than the Poincar´ section. b.a. displays no apparent regularity.

5. φ large enough for “independence.. For example: • x(t).g. but it should retain similar geometric properties. the stock market. then. x(t) (where the derivatives are ﬁnite diﬀerences x k − xk−1. beyond an autocorrelation time. it is known as the method of delays.) In most cases it is unlikely that we can specify the equations of motion of the dynamical system that is generating x(t).12. We proceed to show how such Poincar´ sections and 1-D maps can be con e structed from experimental data.This is the essence of deterministic chaos. Such a representation of the attractor is not identical to the “real” phase space. may we visualize the system’s phase space and its attractor? The (heuristic but highly successful) idea. We use the R¨ssler attractor as an example. etc. Crutchﬁeld. is to measure any 3 independent quantities from x(t). How. the weather.3 Example 3: Reconstruction of phase space from experimental data Suppose we measure some signal x(t) (e. ˙ ¨ etc. • x(t). This is the most popular. Farmer. x(t + 2φ ). x(t + φ ).). 8.. and Shaw (1980). o • Comparison of the projection of the trajectories in the x-y plane to the projection of the trajectories in the x-x plane: ˙ Figure IV. originally due to Santa Cruz un dergraduates Packer.” i. BPV 80 . x(t).e.

We shall see that the various representations also yield similar quantitative properties (e. Here we have discussed only qualitative. Schuster With the method of delays. φ is typically the period of the forcing. geometric properties.. measures of Lyaponov exponents). or the period of a characteristic limit cycle.• Example of diﬀerent values for φ using the method of delays: Figure 78.g. 81 .

Later we show e how the RB problem itself may be reduced to the famous Lorenz equations.9 Fluid dynamics and Rayleigh-B´nard convection e In these lectures we derive (mostly) the equations of viscous ﬂuid dynamics. 9. We then show how they may be generalized to the problem of RayleighB´nard convection—the problem of a ﬂuid heated from below. In these lectures we digress. • Reynolds number • Phenomenology of RB convection • Rayleigh number • Equations of RB convection Thus far we have dealt almost exclusively with the temporal behavior of a few variables. Let Lhydro be the characteristic length scale of macroscopic motions. The mean free path κmfp is the characteristic length scale between molecular collisions. and discuss the evolution of a continuum. The highlights of these lectures are as follows: • Navier-Stokes equations of ﬂuid dynamics (mass and momentum conser vation). Fluids may be regarded as continuous ﬁelds if Lhydro � κmpf . 82 .1 The concept of a continuum Real ﬂuids are made of atoms or molecules.

consider a thought experiment in which we measure the density of a ﬂuid over a length scale κ using some particularly sensitive device. and the hydrodynamic scale. We then move the device in the x-direction over a distance of roughly 10κ. where N � ε3 is the average number of molecules in a box of size ε. Thus the continuum hypothesis implies a separation of scales between the molecular scale.e. i. Thus. a rising hot plume). rather than dealing with the motion �1023 molecules and therefore �6 × 1023 ordinary diﬀerential equations of motion (3 equations each for position and momentum). κ � L2. To make this idea clearer. Lhydro. the evolution of the macroscopic ﬁeld may be described by continuum mechanics.. we model the ﬂuid as a continuum. L 1 � κmfp . Our assumption of a continuum implies that there is an intermediate scale. Suppose κ � L1 � κmpf . ) On the other hand.. if κ � Lhydro (see (c)). (Statistics tells us that these ﬂuctuations should decrease like 1/N 1/2 . variations in density should reﬂect density changes due to macroscopic motions (e.When this condition holds. over which ﬂuctuations are small. 83 . partial diﬀerential equations. not merely statistical ﬂuctuations.g. Then we expect the density to vary greatly in space as in Figure (a) below: density (a) x/L1 (b) x/L 2 (c) x/Lhydro We expect that the ﬂuctuations in (a) should decrease as κ increases.

The motion of the continuum is expressed by partial diﬀerential equations for evolution of conserved quantities. and it points in the outward s s normal direction. u s Thus the rate of mass loss may be rewritten as � � � d ωδ − δdv = − dv = + δψ · dψ u s dt V V ωt s Shrinking the volume. ﬁxed in space: dS V u dψ is an element of the surface. ψ is the velocity.2 Let Mass conservation � δ = density ψ = velocity u of a macroscopic ﬂuid particle Consider a volume V of ﬂuid. 9. u The outward mass ﬂux through the element dψ is s Therefore. δψ · dψ. u s V �0 ωt 84 . We begin with the conservation of mass. u s rate of mass loss from V = The total mass in V is � δdv V � s δψ · dψ. we eliminate the integrals and obtain ⎡� � ωδ = − lim δψ · dψ/V . |dψ| is its area.

and • ﬂuid particle can change its momentum by ﬂowing to a place where the velocity is diﬀerent. a net divergence creates a corresponding change in density. We thus obtain ωδ ψ = −� · (δψ u) ωt We see that to conserve mass. There is a conceptual problem: d (particle momentum) cannot be given at a dt ﬁxed location. For incompressible ﬂuids.3 Momentum conservation We seek an expression of Newton’s second law: d (momentum of ﬂuid particle) = force acting on ﬂuid particle dt 9.) Then which is the equation of continuity for incompressible ﬂuids. because • the momentum ﬁeld itself changes with respect to time. is much less than unity. 9. (This result is not an assumption.1 Substantial derivative (22) We ﬁrst focus on the LHS of (22). the square of the ratio of the ﬂuid velocity to the speed of sound. ψ u � · ψ = 0. δ � constant. but instead derives from the assumption that the Mach number. 85 .Recall that the RHS above is the deﬁnition of the divergence operator.3.

νz in the position of the ﬂuid particle: νT = ωT ωT ωT ωT νt + νx + νy + νz ωt ωx ωy ωz Divide by νt to obtain the rate of change: νT ωT ωT νx ωT νy ωT νz = + + + νt ωt ωx νt ωy νt ωz νt νx νy νz ∗ ux . A small change νT is produced by a small changes νt in time and νx. D ω = +ψ ·� u ψ Dt ωt is the substantial derivative or convective derivative operator. consider how a scalar property— the temperature T —of a ﬂuid particle changes in time.To better understand this problem physically. ∗ uz νt νt νt The rate of change of T of a ﬂuid particle is then DT ωT ωT ωT ωT = + ux + uy + uz Dt ωt ωx ωy ωz = where ωT + ψ · �T u ψ ωt In the limit νt ∗ 0. and • the particle can ﬂow to a position where the temperature is diﬀerent (via ψ · �). Thus we see that the temperature of a ﬂuid particle can change because • the temperature ﬁeld changes “in place” (via ω/ωt). νy. ∗ uy . u ψ Note that the same analysis applies to vector ﬁelds such as the velocity ψ : u Dψ u ωψ u = + (ψ · �)ψ u ψ u Dt ωt 86 .

We can now express the rate-of-change of momentum per unit volume (i. Density changes thus mean volume changes. 87 . u u ψ This dual role of velocity is the essential nonlinearity of ﬂuid dynamics and thus the cause of turbulent instabilities. which need not be equal to the rate of change of (momentum per unit volume). which are irrelevant to the momentum change of that particle. LHS of (22)): Dψ u ωψ u δ =δ + δ(ψ · �)ψ u ψ u Dt ωt δ is outside the diﬀerential because a ﬂuid particle does not lose mass.. Above we have written the (rate of change of momentum) per unit volume. 9. we need the RHS of δ Dψ u = Force acting on ﬂuid particle / unit volume.Therefore the velocity ψ enters Dψ /Dt in 2 ways: u u • ψ changes (in place) as the ﬂuid moves (ω/ωt) u • ψ governs how fast that change occurs (ψ · �). gravity) • pressure • viscous friction (internal stresses) ψ Body force.e. Dt These forces are • body force (i.e.3. We represent the externally imposed body force by F .2 Forces on ﬂuid particle To obtain the full dynamical equation..

For example. (A nice analog is Reif’s picture of two mail trains.) In the simple shear ﬂow above. Fluid ﬂows from high to low pressure. This resistance derives from molecular motions.Pressure. They resist relative movements between ﬂuid particles. that trade mailbags. one initially fast and the other initially slow. σ has units of mass/(length × time). there is a ﬂux of x-momentum in the ydirection. this ﬂux. the shear ﬂow y u u x is resisted more by high viscosity ﬂuids than low viscosity ﬂuids. ωxj . Thus pressure force ωp = − unit volume ωx ψ = −�p in 1-D in 3-D Viscous friction. The shear stress can occur at any orientation. which we call Pxy . In Newtonian ﬂuids. Analogous to the 1-D Newto nian condition above. Viscous stresses are the source of dissipation in ﬂuids. is proportional to the gra dient: ωux Pxy = −σ ωy where σ is called the dynamic viscosity. we deﬁne the viscous momentum ﬂux Pij = −σ 88 ωui .

p∗ = U L δU 2 The dynamical equations (without body force) become ψ u �∗ · ψ ∗ = 0 ωψ ∗ u 1 ∗2 ∗ ψ + (ψ ∗ · �∗ )ψ ∗ = −�∗p∗ + u ψ u � ψ u ωt∗ Re 89 . L y∗ = y . Loosely stated. � � ω2 � ω ω(δui) � ψ · Pij = − � Pij = σ ui = −� ωt �viscous ωxj ωx2 j j j We thus ﬁnd that viscous force = σ�2ψ . L z∗ = z L ψ u U p . 9.) Newton’s second law then gives the Navier-Stokes equation for incompressible ﬂuids: ωψ u δ + δ(ψ · �)ψ u ψ u ωt ⎜� � � = ψ −�p + σ�2ψ u � ⎜� � + ψ F �⎜�� (mass per unit vol)×acceleration stresses on ﬂuid element per unit vol body force per unit vol Incompressibility arose from our negelect of compressive forces on ﬂuid ele ments. We obtain the non-dimensional quantities x∗ = ψ∗ = u x . t∗ = t .The conservation of momentum requires that the divergence of the momen tum ﬂux Pij be balanced by a change in the momentum of a ﬂuid particle. u unit volume (A careful derivation requires consideration of the tensorial relationship between viscous stress and the rate of deformation.4 Nondimensionalization of Navier-Stokes equations Deﬁne the characteristic length scale L and velocity scale U .

the equation it solves) does not. the dimensional parameters like U and L can change but the the ﬂow (i.01 cm2/sec � δ water Uswim � Taking L � 100 cm. note the following dimensional quantities: 2 ψ u � δU |δψ · �ψ| u L nonlinearity dissipation |σ�2ψ | � u Their ratio is σU L2 |δψ · �ψ| δUL u ψu � = Reynolds number |σ�2ψ | u σ High Re is associated with turbulence (i. δUL σ The Reynolds number quantiﬁes the relative importance of the nonlinear term to the viscous term.15 cm2/sec and δ � air Urun On the other hand. 104 cm � = 103 cm/sec 10 sec 104 cm � 2 × 102 cm/sec 55 sec � �� σ � � = 0.where Re = Reynolds number = is the dimensionless control parameter.e. This is dynamical similarity. Note that as long as Re remains the same. An example is running vs. nonlinearities).e. swimming: � � � σ � � = 0. To see why.. comparing 100 meter world records. Low Re is asso ciated with laminar or creeping ﬂows dominated by viscous friction. Re(swim) � 2 × 104 and 90 Re(run) � 6 × 103 ..

. well into the turbulent regime. • νT large: convective motion occurs. 9. since the the small size and slow speed of bacteria would correspond to a larger and faster body in a more viscous ﬂuid. Re(swim) is somewhat greater. Surprisingly. Typically. hot ﬂuid rises. Two cases of interest: • νT small: no convective motion. 91 .5 Rayleigh-B´nard convection e In a thermally expansive ﬂuid. heat-conducting plates: T0 T=T 0 g d fluid (cold) pure conduction T=T 0 + δ T temperature T=T 0 + δ T (hot) In the absence of convection—the transport of hot ﬂuid up and cold ﬂuid down—the temperature gradient is constant. due to stabilizing eﬀects of viscous friction. despite the slower speed of swimming. R-B convection concerns the study of the instabilities caused by rising hot ﬂuid and falling cold ﬂuid. Another example: bacteria swimming in water is roughly like us swimming in molasses.Thus for both swimming and running. Re � 104. ﬂuid is conﬁned between two horizontal.

The following ﬂuid properties are important: • viscosity • density • thermal expansivity • thermal diﬀusivity (heat conductivity) Convection is also determined by • d.How large is a “large νT ” ? We seek a non-dimensional formulation. the box size • νT (of course) Consider a small displacement of a cold blob downwards and a hot blob upwards: T=T 0 T=T 0 + δ T Left undisturbed. There are however damping (dissipation) mechanisms: • diﬀusion of heat • viscous friction 92 . buoyancy forces would allow the hot blob to continue rising and cold blob to continue falling.

The rescaled viscosity has units �σ � mass = d (length)2 (time) Dividing the rescaled viscosity by the buoyancy force. The viscous friction between the two blobs diminishes like 1/d (since viscous stresses ≥ velocity gradients). Note units: [g�δ0 νT ] = mass (length)2(time)2 The buoyancy force is resisted by viscous friction between the two blobs separated by � d. Let δ0 = mean density � = expansion coeﬃcient Γδ = −�δ0 ΓT. buoyancy force g�δ0 d νT 93 . we obtain the charac teristic time φm for convective motion: φm � σ/d σ = . Setting ΓT = νT . buoyancy force density = |ψ Γδ| g = g�δ0 νT. which has units length2 [DT ] = time The temperature diﬀerence between the two blobs can therefore be main tained at a characteristic time scale d2 φth � DT We also seek a characteristic time scale for buoyant displacement over the length scale d.Let DT = thermal diﬀusivity.

• small σ.6.Convection (sustained motion) occurs if time for motion < diﬀusion time for temperature diﬀerence φm < φth Thus convection requires or φth > constant φm δ0 g�d3 νT ∈ Ra > constant σDT Ra is the Rayleigh number. �. A detailed stability calculation reveals that the critical constant is 1708. In other words. 9.6 9. Our derivation of the Rayleigh number shows that the convective instability is favored by • large νT . d. convection occurs when the buoyancy force δ 0 g�d3 νT exceeds the dissipative eﬀects of viscous drag and heat diﬀusion. DT . This means that small increases in box size can have a dramatic eﬀect on Ra. Note that box height enters Ra as d3. The momentum equation is therefore the Navier-Stokes equation augmented by the buoyancy force: ωψ u 1ψ + ψ · �ψ = − �p + ��2ψ − ψ u ψu u g�(T − T0 ) ωt δ0 94 .1 Rayleigh-B´nard equations e Dimensional form We employ the Boussinesq approximation: density perturbations aﬀect only the gravitational force. δ0 .

2 Dimensionless equations The equations are nondimensionalized using length scale = d time scale = d2/DT temperature scale = νT /Ra. u ωt 9. An additional dimensionless parameter arises: Pr = Prandtl number = �/DT .6. We shall employ the dimensionless temperature ﬂuctuation β = deviation of dimensionless T from the simple conductive gradient The mass conservation equation is ψ u �·ψ = 0 Momentum conservation yields (ˆ is a unit upward normal) z ⎡ � 1 ωψ u ψ + ψ · �ψ = −�p + βˆ + �2ψ u ψu z u Pr ωt 95 . We now additionally require an equation for the convection and diﬀusion of heat: ωT + (ψ · �)T = DT �2 T.Here we have written the kinematic viscosity � = σ/δ0 The mass conservation equation is again ψ u � · ψ = 0. which is like the ratio of momentum diﬀusion to thermal diﬀusion.

somewhere in the box midway between the top and bottom. there is no convection. a regular structure of convection “rolls” forms. Instabilities are thermally induced.The heat equation becomes ωβ + ψ · �β = Ra(ψ · z ) + �2β u ψ u ˆ ωt Note that there are two nonlinear terms: • ψ · �ψ u ψu • ψ · �β u ψ Their relative importance depends on Pr: • small Pr ∞ ψ · �ψ dominates. For Ra > Rac .3 For Ra < Rac . but not too large. with hot ﬂuid rising and cold ﬂuid falling: T = T0 d T = T0 + δ d Now imagine placing a probe that measures the vertical component v of velocity.6. u ψ Bifurcation diagram 9.” u ψu • large Pr ∞ ψ · �β dominates. Instabilities are “hydrodynamic. A plot of v(Ra) looks like 96 .

12. and Libchaber (1990) • Collective plumes: Zhang et al (1997).4 Pattern formation Rayleigh-B´nard convection makes fascinating patterns. Some examples: e • Figures 22. with the main divisions being the inner core.6. showing quasiperiodic regime. Here the stable states are bold and the unstable states are dashed. Schuster.v v+ convection (stable) conduction (unstable) 0 conduction rest Ra Ra c v− Such a plot is called a bifurcation diagram. Tritton.5 Convection in the Earth The Earth’s radius is about 6378 km. 9. • Plate 1. Note that we cannot know in advance whether the velocity will be up or down. mantle. 9. 97 . Tritton • Plumes in the wind: Zocchi. and crust.6. It is layered.3–8. outer core.) • Plumes: Figure 22. This is called symmetry breaking. Moses. (The 40 sec period is not precise: note details in upper right are not quite periodic.

Recall. However its eﬀective viscosity is the subject of much debate.The Earth’s crust—the outermost layer—is about 30 km thick. Current estimates suggest that 3 × 106 � Ramantle � 109 which corresponds to roughly 103 × Rac � Ramantle � 106Rac The uncertainty derives principally from the viscosity. 26 May 1989 • van der Hilst seismic tomography. showing cold slab descending toward the core-mantle boundary. where d is the “box size”. One might naively think that the huge viscosity would make the Rayleigh number quite small. thorium. At long time scales mantle rock is thought to ﬂow like a ﬂuid. d is nearly 3000 km!!! Consequently Ra is probably quite high. The source of heat is thought to be the radioactive decay of isotopes of uranium.5 Ga. • Zhang and Libchaber (2000) showing ﬂoater-plates. however. • Gurnis (1988) simulation/cartoon showing breakup of continents. and potassium. For the mantle. 98 . The mantle ranges from about 30–2900 km. and its presumed vari ation by a factor of about 300 with depth. The mantle is widely thought to be in a state of thermal convection. that Ra scales like d3. Some pictures illustrate these ideas: • Science cover. Thermal convection is the “engine” that drives plate tectonics and volcanism. Another heat source is related to the heat deriv ing from the gravitational energy dissipated by the formation of the Earth roughly 4.

without convection.It turns out that volcanism is. That is. 99 .) Thus in some sense thermal convection may be said to also sustain life. and thus the source of carbon that is ﬁxed by plants. and therefore we wouldn’t be around to discuss it. there probably would be no CO 2 in the atmo sphere... over the long-term. responsible for the CO 2 in the atmosphere. (Weathering reactions remove C from the atmosphere.

10 Introduction to Strange Attractors Thus far. with attractor A.1 Dissipation and attraction Our studies of oscillators have revealed explicitly how forced systems can reach a stationary (yet dynamic) state characterized by an energy balance: average energy supplied = average energy dissipated An example is a limit cycle: θ θ Initital conditions inside or outside the limit cycle always evolve to the limit cycle. 10. we have studied only classical attractors such as ﬁxed points and limit cycles. x dt ψ ≤ Rn x (23) Assume that the ﬂow ψ (t) is dissipative. More generally. x 100 . Limit cycles are a speciﬁc way in which dissipation ∞ attraction. We emphasize their generic features. we have an n-dimensional ﬂow d ψ x ψ (t) = F [ψ (t)]. In this lecture we begin our study of strange attractors.

Consequence: loss of information on initial conditions. the basin of attraction.. Here d = attractor dimension = 1 n = phase-space dimension = 2. it does not change with time).e. • A is contained within B. Consider.Properties of the attractor A: • A is invariant with ﬂow (i. for example. B is that part of phase space from which all initial conditions lead to A as t ∗ →: A B • A has dimension d < n. 101 . the case of a limit cycle: Γ θ θ The surface � is reduced by the ﬂow to a line segment on the limit cycle (the attractor). This phenomenon is called reduction of dimensionality.

What. by considering the case d = 2. from A to A∗ .. so d < n. 10. dV < 0. dt Consequently. i.We have already quantiﬁed volume contraction. the attractor cannot have n-dimensional volumes. the Lie derivative tells us that V changes as n � ωxi 1 dV ˙ ˙ = �·ψ = x V dt ωxi i As we showed earlier. Given an initial volume V evolving according to the ﬂow (23). and open it again: 102 . is the dimension of the attractor? We proceed by example.e. then. dissipation yields volume contraction.2 Attractors with d = 2 What happens when d (the dimension of the attractor) is 2? Assume a quasiperiodic attractor on a torus T 2 : ω2 ω1 C Cut the torus on a small circle C and open it: B A B’ A’ Finally. cut the long way.

Consider. or each timen around the 1-axis.g. • Rational. ω2 t 2π ’ 2π ω1 t The trajectories densely ﬁll the plane. As usual.ω2 t 2π A A’ B 2π B’ ω1 t Note the parallel trajectories.. the quasiperiodic ﬂows are characterized by two cases: γ1 /γ2 rational or irrational. γ1/γ2 = 1/3: ω2 t 2π ’ 2π ω1 t The trajectory repeats itself exactly every three times around the 2-axis. 103 . • Irrational. e.

Determinism forbids non-parallel trajectories. because they would cross: ω2 t 2π ’ 2π ω1 t Thus a torus T 2 can only be a periodic or quasiperiodic attractor.” We seek. 10. The attractor cannot be aperiodic if d = 2.3 Aperiodic attractors We have already shown that the power spectrum of an aperiodic signal x(t) is continuous: xk 2 k And the autocorrelation �m = ⇒xj xj+m≡ has ﬁnite width: Ψm m The ﬁnite width of �m implies that knowledge of no ﬁnite interval of x(t) allows prediction of all future x(t). 104 . This “unpredictability” is associated with what we call “chaos. a more precise deﬁnition of chaos. however.

e.On an aperiodic attractor.g. i. This phenomenon is called sensitivity to initial conditions (SIC). Note that. still on the attractor. such that trajectories diverge. The conundrum is solved by noting that trajectories converge to the attractor. such that trajectories converge. since phase tractories cannot diverge in two dimensions... It deﬁni tively identiﬁes chaos. but still remain bounded on an attractor? 105 . but diverge on the attractor. Assume n = 3. despite the precision of this deﬁnition. We shall see that the divergence of trajectories is exponential in time. a chaotic attractor. In phase space. Note further that divergence on the attractor implies that the attractor di mension d > 2. Thus we conclude that an aperiodic (chaotic) attractor must have phase space dimension n √ 3. • sensitivity to initial conditions. trajectories on an aperiodic attractor can diverge. How may trajectories converge. small diﬀerences in initial conditions on the attractor lead at later times to large diﬀerences. e. we are left with an apparent conundrum: simultaneously we have • attraction.

e. ˙ Then x = −y = −x + ax ¨ ˙ ˙ 106 .The trajectories are successively stretched (by SIC) and folded (thus re maining bounded).9–10. y plane the system is apprxoximated ˙ by x = −y ˙ y = x + ay. To illustrate these ideas. Abraham and Shaw 10. One dimension can expand while another contracts. Then in the x. see Figures 4.3. we must have volume contraction. Figures 4.g. • Trajectories return to plane (folding).1. back to center of spiral. At the same time.4 Example: R¨ssler attractor o We proceed to consider stretching and folding in more detail. • Trajectories leave plane. Abraham and Shaw • Trajectories diverge in plane by spiralling out (stretching).1. Assume z and z are small. using the R¨ssler o attractor: x = −y − z ˙ y = x + ay ˙ z = b + z(x − c) ˙ where we assume a > 0.

yielding the negatively damped oscillator x − ax + x = 0. Abraham and Shaw Trajectories never close exactly as a surface. y plane.3. from the feedback of z into x. Abraham and Shaw 107 . A sequence of ﬁgures shows how endless divergence can occur in a bounded region: Figures 4.4. • The process repeats.1–4. How is the spreading conﬁned? From the equation for z we see that. x<c ∞ z<0 ˙ x>c ∞ z>0 ˙ Thus we expect trajectories to behave as follows: • Divergence from the origin creates x > c. but more like ﬁlo dough: Figures 4. • Eventually x decreases such that x < c.2–4. for ˙. small b. from the outward spiral. ¨ ˙ Consequently the trajectories spiral out of the origin. Thus we have • stretching. and • folding. • Then x < c ∞ z decreases ∞ back in the x. • x > c ∞ z increases ∞ x decreases.

The “strangeness” arises not so much from each individual property but their combined presence. The combination of these three properties deﬁnes a strange attractor. d < n.5 Conclusion We arrive at the following conclusions: • Aperiodic attractors must have d > 2. where n is the dimension of the phase space. 108 . How can 2 < d < 3? The attractor has a fractal dimension. • Since dissipation contracts volumes. • Suppose n = 3. Then a chaotic attractor must have 2 < d < 3.10. Chaotic attractors have three properties: • Attraction • SIC • Non-integer fractal dimension. Next we study the most celebrated strange attractor—the Lorenz attractor.

i. 2nd ed. Our derivation emphasizes a simple physical setting to which the Lorenz equations apply. 11. This lecture derives from Tritton. and thus point out the impossibility of accurate long-range predictions. an empty circular tube: cold g hot We expect the following possible ﬂows: • Stable pure conduction (no ﬂuid motion) • Steady circulation • Instabilities (unsteady circulation) 109 . One motivation was to demonstrate SIC for weather systems.1 Physical problem and parameterization We consider convection in a vertical loop or torus.. and study their behavior. The equations were ﬁrst derived by writing a severe. 9.11 Lorenz equations In this lecture we derive the Lorenz equations. The derivation is originally due to Malkus and Howard. rather than the mathematics of the low-order truncation. Physical Fluid Dynamics. See Strogatz. Ch. low-order truncation of the equations of R-B convection. for a slightly diﬀerent view.e.

Quantities inside the tube are averaged cross-sectionally: velocity = q = q(θ. t) temperature = T = T (θ. ωt Thus mass conservation. we employ the Boussinesq approximation e (here. Assume that the tube’s inner radius is much smaller than a. roughly like incompressiblity) and therefore assume ωδ = 0. here u gives ωq = 0.The precise setup of the loop: T0−T (external) 1 T0− 2 T z (T0−T ) 3 (T0+T 3) a φ T0+T 2 T0+T 1 (external) θ = position round the loop. (25) ωθ 110 . External temperature TE varies linearly with height: TE = T0 − T1 z/a = T0 + T1 cos θ (24) g q Let a be the radius of the loop. which would give � · ψ in the full problem. t) (inside the loop) As in the Rayleigh-B´nard problem.

1 Equations of motion Momentum equation Recall the Navier-Stokes equation for convection: ωψ u 1ψ + ψ · �ψ = − �p − ψ u ψu g�ΓT + ��2ψ u ωt δ We write the equivalent equation for the loop as ωq 1 ωp =− + g�(T − T0) sin θ − �q. ωt δa ωθ The terms have the following interpretation: • ψ ∗q u • ψ · �ψ ∗ 0 since ωq/ωθ = 0.Thus motions inside the loop are equivalent to a kind of solid-body rotation. such that q = q(t). Here we assume it depends on only two parameters. and • 2T3 between sides at mid-height. u u 111 (27) .2 11. T2 and T3 vary with time: T2 = T2(t).2. T3 = T3(t) (26) 11. T2 and T3. such that T − T0 = T2 cos θ + T3 sin θ. The temperature T (θ) could in reality vary with much complexity. Thus the temperature diﬀerence is • 2T2 between the top and bottom.

ωt 0 The pressure term vanished because � 2� ωp dθ = 0. • � is a generalized friction coeﬃcient. Now substitute the expression for T − T0 (equation (26)) into the momentum equation (27): ωq 1 ωp =− + g�(T2 cos θ + T3 sin θ) sin θ − �q ωt δa ωθ Integrate once round the loop.• �p ∗ 1 γp a γπ by transformation to polar coordinates. there is no net pressure gradient around the loop. corresponding to viscous resistance proportional to velocity. 112 .. with respect to θ. The integrals are easily evaluated: �2� � 2� 1 2 � cos θ sin θ dθ = sin θ� = 0 � 2 0 0 and � 2� 0 sin2 θ dθ = α.e. • A factor of sin θ modiﬁes the buoyancy force F = g�(T − T0 ) to obtain the tangential component: F φ φ Fsinφ The sign is chosen so that hot ﬂuid rises. ωθ 0 i. to eliminate the pressure term: � 2� ωq 2α = g� (T2 cos θ sin θ + T3 sin2 θ)dθ − 2α�q.

no �2T ).e. The full tem perature equation for convection is ωT + ψ · �T = τ�2 T u ψ ωt where τ is the heat diﬀusivity... q is taken to be uncorrelated to ωT /ωθ.Then. not (as it should be) the average of a product. • is the product of averages. T2(t) and T3 (t). the momentum equation is dq g�T3 = −�q + dt 2 where we have written dq/dt instead of ωq/ωt since ωq/ωθ = 0. 11. after dividing by 2α. • Conduction round the loop is negligible (i.e. We approximate the temperature equation by considering only cross-sectional averages within the loop: ωT q ωT + = K(TE − T ) (29) ωt a ωθ Here we have made the following assumptions: • RHS assumes that heat is transferred through the walls at rate K(Texternal − Tinternal). Speciﬁcally: TE = T0 + T1 cos θ T − T0 = T2 cos θ + T3 sin θ 113 .2 Temperature equation We now seek an equation for changes in the temperature T . We also have the external temperature TE varying linearly with height. (28) We see that the motion is driven by the horizontal temperature diﬀerence.2. i. 2T3. q γT a γπ Recall that we parameterized the internal temperature with two time-dependent variables.

which is the diﬀerence between internal and external temperatures at the top and bottom—loosely speaking. we may separate sin θ terms and cos θ terms to obtain sin θ : cos θ : dT3 qT2 − = −KT3 dt a dT2 qT3 + = K(T1 − T2) dt a These two equations. We proceed to simplify by deﬁning T4(t) = T1 − T2(t). dt dt a a Here the partial derivatives of T have become total derivatives since T 2 and T3 vary only with time.” Substitution yields dT3 qT1 qT4 − = −KT3 + dt a a dT4 qT3 = −KT4 + dt a 11.Subtracting the second from the ﬁrst. Since the temperature equation must hold for all θ. Substitute this into the temperature equation (29): dT2 dT3 q q cos θ + sin θ − T2 sin θ + T3 cos θ = K(T1 − T2) cos θ − KT3 sin θ. Y = . aK 2a�K 114 Z= g�T4 2a�K .e. the extent to which the system departs from a “conductive equilibrium.d.’s that govern the dynamics. together with the momentum equation (28).3 Dimensionless equations Deﬁne the nondimensional variables q g�T3 X= . TE − T = (T1 − T2) cos θ − T3 sin θ. are the three o.

For now.Here X = dimensionless velocity Y = dimensionless temperature diﬀerence between up and down currents Z = dimensionless departure from conductive equilibrium Finally. Drop the prime on t to obtain dX = −P X + P Y dt dY = −Y + rX − XZ dt dZ = −Z + XY dt where the dimensionless parameters r and P are g�T1 r = = “Rayleigh number” 2a�K P = � = “Prandtl number” K These three equations are essentially the same as Lorenz’s celebrated system. 115 . deﬁne the dimensionless time t∗ = tK. Lorenz’s system contained a factor b in the last equation: dZ = −bZ + XY dt The parameter b is related to the horizontal wavenumber of the convective motions. but with one diﬀerence. we remain with the loop equations (b = 1).4 Stability We proceed to ﬁnd the ﬁxed points and evaluate their stability. 11.

Let � X ψ θ = ⎛ Y ⎞ . As usual. and a temperature distribution consistent with conductive equilibrium. the convective instability occurs when Ra > Rac . r = Ra/Rac . occur where ˙ ˙ ˙ X = Y = Z = 0.The ﬁxed points.. which corresponds. i. Another steady solution is � X� = Y � = ± r − 1 Z� = r − 1 This solution corresponds to ﬂow around the loop at constant speed. Thus we see that. we determine the stability of the steady-state solutions by deter mining the sign of the eigenvalues of the Jacobian. That sgn(X) = sgn(Y ) implies that hot ﬂuid rises and cold ﬂuid falls. Z ⎪ � X� θ� = ⎛ Y � ⎞ Z� ⎪ The eigenvalues ε are found by equating the following determinant to zero: � � � � −(ε + P ) P 0 � � � r − Z � −(ε + 1) −X � � = 0 � � � � � Y X −(ε + 1) � 116 Then the Jacobian matrix is ⎭ ⎣ � −P +P 0 � ˙ ωθi � � = � r − Z � −1 −X � ⎤ ωθj � � Y� X � −1 π . pure conduction. the ± signs arise because the circulation can be in either sense. or steady solutions.e. Note that the second (convective) solution exists only for r > 1. to a ﬂuid at rest. eﬀectively. respectively. An obvious ﬁxed point is X � = Y � = Z � = 0.

� 1 −(ε + 1) −S � � � S S −(ε + 1) � (Explicitly. � r −(ε + 1) 0 � � � 0 −(ε + 1) � 0 This yields or −(ε + P )(ε + 1)2 + rP (ε + 1) = 0 � � (ε + 1) ε 2 + ε(P + 1) − P (r − 1) = 0. ε1 = −1 � (P + 1)2 + 4P (r − 1) −(P + 1) ± ε2. � S = ± r − 1. Z � = r − 1. ε2.For the steady state without circulation (X � = Y � = Z � = 0). −(ε + p)(ε + 1)2 − P s2 − S 2 (ε + P ) + P (ε + 1) = 0 (ε + 1)[ε 2 + ε(P + 1)] + εS 2 + 2P S 2 = 0. ε2.) 117 . X � = ± r − 1. The eigenvalues ε are now the solution of � � � � −(ε + P ) P 0 � � � = 0. we have � � � � −(ε + P ) P 0 � � � = 0. and ε3} < 0 =∞ stable Re{ε1. � Y � = ± r − 1. There are three roots: As usual.3 = 2 2 Re{ε1. or ε3} > 0 =∞ unstable Therefore X � = Y � = Z � = 0 is stable unstable for for 0<r<1 r>1 � We now calculate the stability of the second ﬁxed point.

Let ε1 be the (negative) real root. or (30) 3 real roots 1 real and 2 complex conjugates Rearranging equation (30). and let ε2. and C are all real and positive.g. B. ε (ε 2 + B) = −Aε 2 − C < 0. e. � ⎜� � � ⎜� � positive real negative real Consequently any real ε < 0.3 = � ± iλ. or • 1 real root and 2 complex conjugate roots. Such an equation has either • 3 real roots. Then and (ε − ε1 )(ε − � − iλ)(ε − � + iλ) = 0 A = −(ε1 + 2�) B = 2�ε1 + �2 + λ 2 C = −ε1(�2 + λ 2) 118 .. and we need only consider the complex roots (since only they may yield Re{ε} > 0).The characteristic equation is cubic: ε 3 + ε 2 (P + 2) + ε(P + r) + 2P (r − 1) = 0 This equation is of the form ε 3 + Aε 2 + Bε + C = 0 where A.

A little trick:

Since � is the real part of both complex roots, we have sgn(Re{ε2,3}) = sgn(�) = sgn(C − AB). Thus instability occurs for C − AB > 0, or 2P (r − 1) − (P + 2)(P + r) > 0, . Rearranging, and we ﬁnd that instability occurs for r > rc = r(2P − P − 2) > 2P + P (P + 2) P (P + 4) . P −2

� � C − AB = 2� (ε1 + �)2 + λ 2 . � ⎜� �

positive real

This condition, which exists only for P > 2, gives the critical value of r for which steady circulation becomes unstable. Loosely speaking, this is analogous to a transition to turbulence. Summary: The rest state, X � = Y � = Z � = 0, is stable unstable for for 0<r<1 r > 1.

� The convective state (steady circulation), X � = Y � = ± r − 1, Z � = r − 1, is stable unstable What happens for r > rc ? Before addressing that interesting question, we ﬁrst look at contraction of volumes in phase space.

119

for for

1 < r < rc r > rc .

11.5

Dissipation

We now study the “full” equations, with the parameter b, such that ˙ Z = −bZ + XY, � ωθi ˙ 1 dV = , V dt ωθi i b > 0.

The rate of volume contraction is given by the Lie derivative i = 1, 2, 3, θ1 = X, θ2 = Y, θ3 = Z.

For the Lorenz equations, ˙ ˙ ˙ ωX ωY ωZ + + = −P − 1 − b. ωX ωY ωZ dV = −(P + 1 + b)V dt which may be solved to yield V (t) = V (0)e−(P +1+b)t. The system is clearly dissipative, since P > 0 and b > 0. The most common choice of parameters is that chosen by Lorenz P = 10 b = 8/3 (corresponding to the ﬁrst wavenumber to go unstable). V (t) = V (0)e− 3 t . Thus after 1 time unit, volumes are reduced by a factor of e− 3 � 10−6. The system is therefore highly dissipative.

41 41

Thus

For these parameters,

11.6

Numerical solutions

**For the full Lorenz system, instability of the convective state occurs for r > rc = P (P + 3 + b) P −1−b
**

120

For P=10, b=8/3, one has rc = 24.74. In the following examples, r = 28. Time series of the phase-space variables are shown in Tritton, Fig 24.2, p. 397 • X(t) represents variation of velocity round the loop.

� � – Oscillations around each ﬁxed point X+ and X− represent variation in speed but the same direction.

– Change in sign represents change in direction. • Y (t) represents the temperature diﬀerence between up and downggoing currents. Intuitively, we expect some correlation between X(t) and Y (t). • Z(t) represents the departure from conductive equilibrium. Intuitively, we may expect that pronounced maxima of Z (i.e., overheating) would foreshadow a change in sign of X and Y , i.e., a destabilization of the sense of rotation. Projection in the Z-Y plane, showing oscillations about the unstable convec tive ﬁxed points, and ﬂips after maxima of Z: BPV, Fig. VI.12 A 3-D perspective, the famous “butterﬂy:” BPV, Fig. VI.14 Note the system is symmetric, being invariant under the transformation X ∗ −X, Y ∗ −Y , Z ∗ Z.

A slice (i.e., a Poincar´ section) through the plane Z = r − 1, which contains

e the convective ﬁxed points:

BPV, Fig. VI.15

121

17 Finally. 122 .18 11. indicating the attractor dimension d ◦ 2. we can construct. as did Lorenz.) The ﬁrst-return map shows that the dynamics can be approximated by a 1-D map. Fig. Fig.7 Conclusion The Lorenz model shows us that the apparent unpredictability of turbulent ﬂuid dynamics is deterministic. the ﬁrst return map zk+1 = f (zk ). VI. The result is BPV. It also reveals the stability properties of the ﬁxed point Z = r − 1: BPV. Fig.16 ˙ (These points intersect the plane XY −bZ = 0. Why? Lorenz’s system is much simpler than the Navier-Stokes equations. Since d ◦ 2. VI. sensitivity to initial conditions is documented by BPV.• The trajectories lie on roughly straight lines. where zk is the kth maximum of Z(t). with (as we shall see) a fractal dimension of 2. but it is essentially contained within them. • These are really closely packed sheets. surely the NavierStokes equations contain suﬃcient complexity to do so also. • d ◦ 2 results from the strong dissipation.06. VI. which corresponds to Z = 0. Because the simpler system exhibits deterministic chaos.

Previous explanations of transitions to turbulence (e. A striking conclusion is that only a few (here.Thus any doubt concerning the deterministic foundation of turbulence.. is now removed. such as assuming that turbulence represents a failure of deterministic equations.g. Landau) had invoked a successive introduction of a large number of degrees of freedom. 123 . three) degrees of freedom are required to exhibit this complexity.

lines of points need not be parallel).12 H´non attractor e The chaotic phenomena of the Lorenz equations may be exhibited by even simpler systems. The e points in R2 are considered to be the the Poincar´ section of a ﬂow in higher 3 dimensions. thus the ﬂow is not restricted by continuity (i. 2-D mapping of the plane into itself. say. 12. We now consider a discrete-time. The restriction that d > 2 for a strange attractor does not apply.e. 2-D mapping of H´non is e 2 Xk+1 = Yk + 1 − �Xk Yk+1 = λXk • � controls the nonlinearity.. because maps generate discrete points. • λ controls the dissipation. Pictorially.1 The H´non map e The discrete time. we may consider a set of initial conditions given by an ellipse: Y X 124 . R .

but preserve the area inside it (we shall soon quantify area preservation): Y’ Map T1 : X∗ = X Y ∗ = 1 − �X 2 + Y X’ Next. ﬂip across the diagonal). reorient along the x axis (i.Now bend the elllipse. Y ’’’ Map T3 : X ∗∗∗ = Y ∗∗ Y ∗∗∗ = X ∗∗ X ’’’ The composite of these maps is T = T3 ∝ T 2 ∝ T 1 . We readily ﬁnd that T is the H´non map: e X ∗∗∗ = 1 − �X 2 + Y Y ∗∗∗ = λX 125 . contract in the x-direction (|λ | < 1) Y ’’ Map T2 : X ∗∗ = λX ∗ Y ∗∗ = Y ∗ X ’’ Finally.e.

∗ Γy c d Γy Geometrically. Yk ) Yk+1 = g(Xk .c)Δ x Δx x 126 x . y0).12. we have. ∗ ΓYk+1 gXk (x0. Yk ) Inﬁnitesimal changes in mapped quantities as a function of inﬁnitesimal changes in inputs follow df = ωf ωf dXk + dYk ωXk ωYk We may approximate. ⎡ � ⎡ ∗ �⎡ � ∗ ΓXk+1 fXk (x0. to ﬁrst order. y0) g k (x0. the increment ΓXk+1 due to small in crements (ΓXk . Δ y) y (b. y0) ΓYk Y Rewrite simply as ⎡ � ⎡ �⎡ � ∗ Γx a b Γx = . ΓYk ) are perturbations about a point (x0. y0) ΓXk = ∗ . this system describes the transformation of a rectangular area determined by the vertex (Γx. Γy) to a parallelogram as follows: y Δy ( Δ x . ΓYk ) as ΓXk+1 ◦ ωf ωf ΓXk + ΓYk ωXk ωYk When (ΓXk .2 Dissipation The rate of dissipation may be quantiﬁed directly from the mapping via the Jacobian. y0) fYk (x0. to ﬁrst order. Δ y’) (a.d)Δ y ’ ( Δ x . We write the map as Xk+1 = f (Xk .

Figure VI. in general. The dissipation is thus considerably less than e the factor of 10−6 in the Lorenz model. An illustration of the attractor is given by BPV. ak = a0 |λ |k . the Jacobian determinant � � � ωXk+1 ωXk+1 � � � � � ωX � ωYk � k � a b � � � � = � J = � � � c d � � � � ωYk+1 ωYk+1 � � � ωXk ωYk (x0 . 12. or.4.3.e.Here we have taken account of transformations like (Γx. e Thus areas are multipled at each iteration by |λ |. d).y0 ) Therefore |J| > 1 =∞ dilation |J| < 1 =∞ contraction For the H´non map.. then the area of the parallelogram is given by the magnitude of the cross product of (a. Γy) ∗ (bΓy.3 Numerical simulations � � � −2�Xk 1 � � = −λ J = � � λ 0 � After k iterations of the map. an initial area a0 becomes H´non chose � = 1. ΓxΓy = 1). 0) ∗ (aΓx.19 127 . cΓx) (0. c) and (b. dΓy) If the original rectangle has unit area (i. λ = 0.

22 At the scale of the attractor we can see the combined eﬀects of stretching and folding: BPV. (We shall soon discuss how this is com puted. Figure VI.26.21 Note the apparent scale-invariance: at each magniﬁcation of scale.23 128 . we see that the upper line is composed of 3 separate lines.Numerical simulations show the basin of attraction to be quite complex.20 The weak dissipation allows one to see the fractal structure induced by the repetitive folding BPV. Sensitivity to initial conditions is conﬁrmed by BPV. Figure VI. Figure VI. The fractal dimension D = 1. Figure VI.) The action of the H´non map near the attractor is evident in the deformation e of a small circle of initial conditions: BPV.

(ΓT ) and (ΓT )∗. inside the convecting system. 13. 13. p. A 3-D phase space is deﬁned by the coordinates ˙ (ΓT ). represent time-dependent thermal gradients measured measured by the refraction of light. Section 8. 129 . (ΓT ).2 Belousov-Zhabotinsky reaction See Strogatz. because it is oscillatory. and (ΓT )∗.3. ˙ The Poincar´ section in the plane (ΓT ). the points are arranged in a “complex but welle deﬁned structure” (BPV. and • pattern formation and nonlinear waves. Figure VI.13 Experimental attractors In this brief lecture we show examples of strange attractors found in experi ments. (ΓT ) is obtained by strobing the e system at the dominant frequency of the ﬂuctuations of one of the dynamical variables. Like the H´non attractor. 137). The B-Z reaction is a particularly well studied chemical reaction that is of interest for • dynamics.1 Rayleigh-B´nard convection e BPV.25 Two dynamical variables.

3 For us the dynamical aspects are of greater interest. can be envisioned as a ﬂux of reactants into a reactor. Section 8. X(t + 2φ ). Plate 1. This chemical reaction.An example of the pattern formation is Strogatz. The periodic regime (limit cycle) is evident in the 2-D projection given by BPV. like many others. and a ﬂux of products out of the reactor: reactants reactor products Nonlinearities arise from chemical reactions like A+B ∗C which yield terms like dC = AB dt The control parameter is typically the reactant ﬂux. The dynamical variable X(t) measures the concentration of a particular chem ical species within the reactor. Phase portraits are obtained in the 3-D phase space formed by X(t). Figure VI. Figure VI.26 The chaotic regime is evident in the 2-D projection given by BPV.27 130 . X(t + φ ).

Figure VI. Plotting the data with a diﬀer ent choice of φ : BPV.32 131 .29a That all the points essentially lie on a simple curve implies deterministic order in the system. Figure VI.31 Nine diﬀerent slices through the system reveal stretching (from 9 to 1) and folding (between 2 and 8): BPV. where Xk is the kth observation of X(t): BPV. Figure VI. Figure VI. but its thinness indicates strong dissipation. as seen in a plot of e intersections with a plane perpendicular to X(t).28 The straightness of the line is insigniﬁcant. Xk .30 Stretching and folding may also be observed.A Poincar´ section reveals evidence of strong dissipation. Due to dissipation. X(t + φ ): BPV. Figure VI. a nice 1-D ﬁrst-return map can be formed by plotting Xk+1 vs.29b Sensitivity to initial conditions is observedby plotting the distribution of points that results after passage close to the same point: BPV. Attraction is demonstrated by inducing a small perturbation and then ob serving the re-establishment of the ﬁrst return map: BPV. Figure VI.

14

Fractals

We now proceed to quantify the “strangeness” of strange attractors. There are two processes of interest, each associated with a measurable quantity: • sensitivity to initial conditions, quantiﬁed by Lyaponov exponents. • repetitive folding of attractors, quantiﬁed by the fractal dimension. Now we consider fractals, and defer Lyaponov exponents to the next lecture. We shall see that the fractal dimension can be associated with the eﬀective number of degrees of freedom that are “excited” by the dynamics, e.g., • the number of independent variables; • the number of oscillatory modes; or • the number of peaks in the power spectrum 14.1 Deﬁnition

**Consider an attractor A formed by a set of points in a p-dimensional space:
**

ε

etc

We contain each point within a (hyper)-cube of linear dimension π. Let N (π) = smallest number of cubes of size π needed to cover A. Then if N(π) = Cπ−D , then D is called the fractal (or Hausdorf ) dimension.

132

as π ∗ 0,

C = const.

Solve for D (in the limit π ∗ 0): D= ln N(π) − ln C . ln(1/π)

Since ln C/ ln(1/π) ∗ 0 as π ∗ 0, we obtain the formal deﬁnition D = lim ln N(π) . ν�0 ln(1/π)

14.2

Examples

**Suppose A is a line segment of length L:
**

L

Then the “boxes” that cover A are just line segments of length π, and it is obvious that N(π) = Lπ−1 =∞ D = 1. Next suppose A is a surface or area S. Then N(π) = Sπ−2 =∞ D = 2. But we have yet to learn anything from D.

Consider instead the Cantor set. Start with a unit line segment:

0 1

**The successively remove the middle third:
**

0 1/9 2/9 etc 1/3 2/3 etc 1

133

Note that the structure is scale-invariant: from far away, you see the middle 1/3 missing; closer up, you see a diﬀerent middle 1/3 missing. The eﬀect is visually similar to that seen in the Lorenz, H´non, and R¨ssler e o attractors. The fractal dimension of the Cantor set is easily calculated from the deﬁnition of D: � � 1 Obviously, N π= = 2 3 � � 1 Then N π= = 4 9 � � 1 N = 8... 27 � 1 N = 2m . m 3 Taking π = 1/3 and using the deﬁnition of D, ln 2m ln 2 ◦ 0.63 D = lim = m�� ln 3m ln 3 14.3 Correlation dimension � Thus �

We proceed now to an alternative procedure for the calculation of the fractal dimension, which oﬀers additional (physical) insight. Rather than calculating the fractal dimension via its deﬁnition, we calculate the correlation dimension �. We shall show that � ∼ D. But ﬁrst we deﬁne it.

14.3.1 Deﬁnition

**Consider a set of points distributed on a plane.
**

134

Let N (r) = number of points located inside a circle of radius r. We expect that N(r) will grow more slowly than r.63 = D. r Now Next. Assume the points are uniformly distributed on a curve like r For r suﬃciently small compared to the curvature of the curve. � = 1. just as before. we have N(r) ≥ r or Now assume the points are uniformly distributed along a surface in two di mensions: N (r) ≥ r� . Indeed. 135 . calculations show that � ◦ 0. reconsider the Cantor set: r N(r) ≥ r2 =∞ � = 2.

14.3.2

Computation

Our implicit deﬁnition of � is clearly generalized by considering • an attractor in a p-dimensional space, and • N (r) = number of points in a p-dimensional hypersphere of radius r. For a time series x(t), we reconstruct a p-dimensional phase space with the coordinates x(t), x(t + φ ), x(t + 2φ ), . . . x(t + (p − 1)φ ) = ψ (t). x Suppose there are m points on the attractor. We quantify the spatial corre lation of these points by deﬁning C(r) = lim More formally, 1 �� C(r) = lim 2 H(r − |ψ i − ψ j |) x x m�� m i j where H(x) = � 1 x>0 0 else.

m m

1 (number of pairs i, j for which |ψ i − ψ j | < r) . x x m�� m2

The summation is performed by centering hyperspheres on each of the m points. In practice, one embeds the signal x(t) in a phase space of dimension p, for p = 2, 3, 4, 5, . . . p is called the embedding dimension. For each p, we calculate C(r). Then, assuming C(r) = r �

136

we plot log C vs. log r and estimating the slope �:

log C(r) slope = ν

log r

Consider the example of white noise. Then x(t) is a series of uncorrelated random numbers, and we expect C(r) ≥ rp , p = embedding dimension.

**Graphically, one expect a series of plots like
**

p= 2 3 4 5 log C(r)

log r

Here �(p) = p, a consequence of the fact that white noise has as many degrees of freedom (i.e., independent “modes”) as there are data points. Consider instead X(t) = periodic function, i.e., a limit cycle, with only one fundamental frequency. Then the attractor looks like

p=2

p=3

137

Graphically, we obtain

**Provided that r is suﬃciently smaller than the curvature of the limit cycle, we expect C(r) ≥ r1, for p = 2, 3, 4, . . .
**

p= 2 log C(r) 3 4 5

log r

and therefore �(p) = 1, independent of p.

We conclude that � measures something related to the “number of degrees of freedom” needed to parameterize an attractor. Speciﬁcally, suppose a dynamical regime has n oscillatory modes. The attractor is then a torus T n , and we expect C(r) ≥ rn . Thus and p ∼ n =∞ C(r) ≥ r p p > n =∞ C(r) ≥ r n , independent of p.

Conclusion: If, for embedding dimensions p √ p0, � is independent of p, then � is the number of degrees of freedom excited by the system. This conclusion provides for an appealing conjecture: since white noise gives �(p) = p,

� independent of p (and reasonably small) implies that the signal is deter ministic, and characterizable by �� variables.

138

Therefore. � m m 1 �� 1 x>0 C(r) = lim 2 H(r − |ψ i − ψ j |).) 14.e. however it can be. In a box of size r. there are on average mpi points. i N(r) 139 � i=1 p2 i . all within the range r. for an attractor containing m points. • r and m must be large enough for reasonable statistics. we have. If the points are uniformly distributed on A. within a factor of O(1) (i. Physica 9D. from Schwartz’s inequality.. using angle brackets to represent mean quantities. the probability that a point falls into the ith hypercube is pi = 1/N(r). By deﬁnition.4 Relationship of � to D (Grassberger and Procaccia. ignoring box boundaries and factors of two arising from counting pairs twice). Suppose we cover an attractor A with N(r) hypercubes of size r. x x H(x) = m�� m 0 else i j C(r) measures the number of pairs of points within a distance r of each other. 1 C(r) = N (r) ⇒p2≡ √ N(r) ⇒pi ≡2 = .(There are some practical limitations: • r must be small compared to the attractor size. We now derive their mathematical relation. N (r) 1 � C(r) ◦ (mpi)2 2 m i=1 N (r) = Then. 183 (1983)) The correlation dimension is not strictly the same as the fractal dimension.

i Thus � < D results from non-uniformity of points on the attractor. Substituting these relations into both sides of the inequality. on the other hand. gives The equality is obtained when ⇒p2 ≡ = ⇒pi ≡2 . 140 . then N (r) ≥ r−D . The deﬁnition of the correlation dimension �. C(r) ≥ r� .If the attractor has fractal dimension D. we see that �∼D r ∗ 0. we ﬁnd r� √ rD Thus as r ∗ 0.

In this lecture we point broadly sketch some of the mathematical issues con cerning Lyaponov exponents. 15.1 Diverging trajectories Lyapunov exponents measure the rate of divergence of trajectories on an attractor.15 Lyapunov exponents Whereas fractals quantify the geometry of strange attractors. their most salient feature. given by dθ ψ ψ = F (θ) dt ψ ψ If instead of initiating the ﬂow at θ(0). We then conclude with a description of a simple model that shows how both fractals and Lyaponov exponents manifest themselves in a simple model. it is initiated at θ(0)+π(0). Lyaponov ex ponents quantify the sensitivity to initial conditions that is. To the ﬁrst order. sensitivity to initial conditions would produce a divergent trajectory: ε( t) ε(0) φ(0) φ( t) Here |ψ| grows with time. π ψ π d(θ + ψ) ψ ψ ◦ F (θ) + M(t) ψ π dt 141 . in eﬀect. We also brieﬂy describe how they are com puted. ψ Consider a ﬂow θ(t) in phase space.

e. Mij (t) = ωθj � (t) τ π dψ π = M(t) ψ. π π 15.where We thus ﬁnd that � ωFi � � . Then M is diagonal and ⎣ e �1 t 0 0 � L(t) = � 0 e 2 t 0 ⎤ 0 0 e �3 t ⎭ where the �i are the eigenvalues of M. e M t = L(t). in the coordinate system of the eigenvectors. let L� be the conjugate (Hermitian) transpose of L. (Recall ˙ that if ρ = Mρ. the eigenvalue with the largest real part dominates the ﬂow ψ(t).) As t increases. β β β β where. i. Y (t) X(t) −b We cannot solve for ψ because of the unknown time dependence of M(t). π dt (31) Consider the example of the Lorenz model.2 Example 1: M independent of time Consider a simple 3-D example in which M is time-independent. L� = Lji . π ψ However one may numerically solve for θ(t). ij 142 . π To express this formally. to obtain (for π mally) ψ(t) = L(t) ψ(0). and thus ψ(t). then ρ(t) = eM t ρ(0). The Jacobian M is given by ⎭ ⎣ −P P 0 M(t) = � −Z(t) + r −1 −X(t) ⎤ . Assume additionally that the phase space coordinates correspond to M ’s eigenvectors.

The solution for νX(t) is νX(t) = νX(0) exp 143 ⎡� t 0 dt A[θ(t )] ∗ ∗ � . νZ(t) in the reference frame of the eigenvectors.3 Example 2: Time-dependent eigenvalues � � 1 � = lim ln Tr[L�(t)L(t)] t�� 2t Now suppose that M(t) varies with time in such a way that only its eigen values. 15. � (� (� Tr[L�(t)L(t)] = e(�1 +�1 )t + e 2 +�2 )t + e 3 +�3 )t Deﬁne � is the largest Lyapunov exponent. C are the time-dependent eigenvalues (assumed to be real). analogous to equation (31). Let ⎣ X(t) ψ θ = � Y (t) ⎤ Z(t) ⎭ and consider small displacements νX(t). ⎭ ˙ ⎣ ⎭ ⎣⎭ ⎣ νX (t) A[θ(t)] 0 0 νX(t) ˙ � νY (t) ⎤ = � ⎤ � νY (t) ⎤ . Its sign is crucial: � < 0 =∞ π(t) decays exponentially � > 0 =∞ π(t) grows exponentially. B. νY (t). but not its eigenvectors. vary. Then. and again assuming that phase space co ordinates correspond to M’s eigenvectors.Also let Tr(L) = diagonal sum = Then � � i=j � Lij . 0 B[θ(t)] 0 ˙ 0 0 C[θ(t)] νZ(t) νZ (t) Here A.

� � 1 � νX(t) � � = lim ln � � νX(0) � t�� t 15.Rearranging and dividing by t. We denote this average by angle brackets: ⇒A≡ = θ-average of A[θ(t)] = time-average of A[θ(t)] 1 = lim t�� t � t 0 ∗ dt∗ A[θ(t )] ⇒A≡ is one of the three Lyapunov exponents for θ(t).4 Numerical evaluation Lyaponov exponents are almost always evaluated numerically. 144 . π π This corresponds to the deﬁnition of ⇒A≡ above. More sophisticated analyses show that the theory sketched above applies to the general case in which both eigenvectors and eigenvalues vary with time. We assume that for suﬃciently long times this average is equivalent to an average of A for all possible ﬂows θ evaluated at the same time. The most obvious method is the one used in the problem sets: For some ψ(0). and then ﬁnd � such that π |ψ(t)| ◦ |ψ(0)|e�t . � � � 1 � νX(t) � 1 t ∗ ∗ � ln � dt A[θ(t )] � νX(0) � = t t 0 The RHS represents the time-average of the eigenvalue A. we assume that the ﬂow is ergodic. π numerically evaluate ψ(t). In other words.

Another way is to reconstruct phase space by. because generally we have only a single time series X(t). Then the comparison of X(t) beyond t2 and t∗2 may yield the largest Lyaponov exponent. the method of delays. say [t1. t2 ] and [t∗1 .A better method avoids saturation at the size of the attractor by successively averaging small changes over the same trajectory: ε( τ) ε(0) φ(0) ε( 2τ) φ( τ) φ( 2τ) Here ψ is renormalized at each step such that π ψ(φ ) = ψ(0)eε1φ π π ψ(2φ ) = π ψ(φ ) ε2 φ π e |ψ(φ )| π The largest Lyaponov exponent is given by the long-time average: n n 1� 1 � � = lim ρi = lim ln |ψ(iφ )| π n�� n n�� nφ i=1 i Experimental data poses greater challenges. Then all trajectories that pass near a certain point may be compared to see the rate at which they diverge. say. 145 . t∗2]. where X(t) is nearly the same on both intervals. One way is to compare two intervals on X(t).

−) (−. all three exponents are negative. −. only two are negative. 0) (−. 0. and the third is zero.15. • There must also be a zero-exponent corresponding to the ﬂow direction. • The smallest exponent must be negative—and of greater magnitude than the largest. 0) (−.5 Lyaponov exponents and attractors in 3-D Consider an attractor in a 3-D phase space. If it is a limit cycle with one frequency. There are 3 Lyaponov exponents. by deﬁnition. 146 . the most interesting is that of a strange attractor: • The largest exponent is. positive. Of the other cases in the table below. since volumes must be contracting. −. Their signs depend on the type of attractor: Type Fixed point Limit cycle Torus T 2 Strange attractor Signs of Lyapunov exponents (−. +) If the attractor is a ﬁxed point. The zero-exponent corresponds to the direction of ﬂow—which can neither be expanding nor contracting. 0.

15.” • Fractal dimensions measure “folding.” Smale’s horseshoe attractor exempliﬁes both.6 Smale’s horseshoe attractor We have seen that • Lyaponov exponents measure “stretching. σ > 1: A’ B’ C’ D’ Now fold like a horseshoe and put back in ABCD: A B 1/2η 1/2η C D Now iterate the process. and allows easy quantiﬁcation. Start with a rectangle: A B C D Stretch by a factor of 2. Stretch and squash: 147 . squash by a factor of 1/(2σ).

the π. To obtain the fractal dimension. ν�0 ln(1/π) (x − stretch) (y − squash) Taking the initial box height to be unity. we use the deﬁnition D = lim ln N(π) .Fold and place back in ABCD: A B 1/(2η) 2 C D Each dimension is successively scaled by its own multiplier. The Lyapunov exponents are �1 = ln �1 �2 = ln �2 Note also that vertical cuts through the attractor appear as the early itera tions of a Cantor set. N pairs for the number N of segments of length π required to cover the attractor is 148 . called a Lyaponov number: �1 = 2 1 �2 = 2σ Area contraction is given by �1�2 = 1/σ.

m 1/(2σ) 2m Therefore the dimension D of the Cantor set is D= ln 2 . meaning that the attractor is nearly squashed to a simple line.. we’ve assumed the box’s width is much greater than its height. D∗ ∗ 2. Note that.. . D∗ ∗ 1.. ln 2σ The dimension D ∗ of the attractor in the plane ABCD is D∗ = 1 + ln 2 . ln 2σ where we have neglected the “bend” in the horseshoe (i. because iterates nearly ﬁll the plane..e. as σ ∗ →..π N 1 1 1/(2σ) 2 1/(2σ)2 4 . as σ ∗ 1. 149 . Conversely.

or a liquid become a gas? We shall see that. x x x x 150 . just as in the study of phase transitions. consider the Poincar´ e section: x0 x1 x2 The periodic regime is linearly unstable if |ψ 1 − ψ 0 | < |ψ 2 − ψ 1| < . 16. . there are universal ways in which systems become chaotic. We ask: How does the transition from periodic to strange attractor occur? The question is analogous to the study of phase transitions: How does a solid become a melt. .16 Period doubling route to chaos We now study the “routes” or “scenarios” towards chaos.1 Instability of a limit cycle To analyze how a periodic regime may lose its stability. There are three universal routes: • Period doubling • Intermittency • Quasiperiodicity We shall focus the majority of our attention on period doubling.

a Poincar´ map T in the neighborhood of ψ 0 is e x described by the Floquet matrix Mij = In a periodic regime.or Recall that. . to ﬁrst order. There are three ways in which |�i | > 1: Im λ i |νψ 1 | < |νψ 2 | < . If |�i | > 1. νψ is ampliﬁed is in the same direction: x x1 x2 x3 x4 This transition is associated with Type 1 intermittency. 151 . x x ωTi . the ﬁxed point is unstable. ωXj Re λ i −1 +1 1. x x But the mapping T sends ψ 0 + νψ ∗ ψ 0 + Mνψ . νψ is ampliﬁed in alternating directions: x x3 x1 x0 x2 This transition is associated with period doubling. 2. π real. . ψ (t + φ ) = ψ (t). � = −(1 + π). � = 1 + π. x x x x Thus stability depends on the 2 (possibly complex) eigenvalues � i of M. π > 0.

This is a period doubling bifurcation. Assume the Poincar´ section goes through x = −0. νψ is rotated: x x 4 3 γ γ γ 2 1 x0 This transition is associated with quasiperiodicity. Let’s look more closely at the second case. 152 . � = � ± iλ = (1 + π)e±iε . |νψ | is ampliﬁed. Then an initial pertur e bation x0 is damped with alternating sign: x1 x3 0 x2 x0 Now vary the control parameter such that � = −1. � = −(1 − π). nonlinear eﬀects eventually cause the instability to saturate. The iterations no longer converge: x1 x3 0 x0 x2 We see that a new cycle has appeared with period twice that of the original cycle through x = 0.3. π > 0. In each of these cases. � ◦ −1.2 Logistic map We now focus on the simplest possible system that exhibits period doubling. 16. Just before the transition.

The term sx2 represents a reduction of natural growth due to crowding and j competition for resources. Let xj = ratio of actual population in jth summer to some reference popula tion. For clarity. Now rescale xj ∗ (r/s)xj . we adopt a biological interpretation. e Thus we look at discrete intervals T. if r > 1 the population grows (exponentially) by a factor r each year. and study the iterates of a transformation on an axis. we set aside n-dimensional (n √ 3) trajectories and focus only on the Poincar´ section and the eigenvector whose eigenvalue crosses (−1). Imagine an island with an insect population that breeds in summer and leaves egges that hatch the following summer. . 153 . This is called the logistic map. Then xj+1 = rxj − rx2. j The term rxj+1 represents natural growth. j Set r = 4µ: xj+1 = 4µxj (1 − xj ). 2T. 3T. We therefore study ﬁrst return maps xk+1 = f (xk ) and shall argue that these maps are highly relevant to n-dimensional ﬂows.In essence. . Assume that next summer’s population is determined by this summer’s pop ulation according to xj+1 = rxj − sx2. .

4µ where the second ﬁxed point exists only for µ > 1/4. x� = 0 and x� = 1 − Recall that stability requires |f ∗ (x�)| < 1 =∞ 154 which yields |4µ(1 − 2x�)| < 1. we consider the range 0 < µ < 1. Re markably.3 Fixed points and stability We seek the long-term dependence of xj on the control parameter µ. Below is a sketch for µ = 0.7: f(x) f(x) f(x0) = x1 0 x0 x1 1−1/(4µ) x k+ 1 = xk 1 1 x The ﬁxed points solve x� = f (x�) = 4µx�(1 − x�).16. we shall see that µ plays a role not unlike that of the Rayleigh number in thermal convection. Recall that we have already discussed the graphical interpretation of such maps. So that 0 < xj < 1. 1 . .

the period of the asymptotic iterates doubles: f(x) 1 x* 1 x* 2 1 x 155 .The stability condition for x� = 0 is therefore µ < 1/4. Just beyond this point. The non-trivial ﬁxed point. The long-time behavior of the insect population x for 0 < µ < 3/4 then looks like 1 x*=1−1/(4µ) x*=0 0 1/4 3/4 µ 16.4 Period doubling bifurcations What happens for µ > 3/4? At µ = 3/4. x� = 1 − 1/(4µ) is marginally stable. is stable for 1/4 < µ < 3/4. x� = 1 − 1/(4µ).

Feigenbaum. f n∗ (x0) = f ∗(x0) f ∗(x1) . Now. x� is also a ﬁxed point of f 2(x). We shall see that period doubling depends on the relationship of the slope of f 2(x�) to the slope of f (x�). • If x� is a ﬁxed point of f (x). in general. a ﬁxed point of f . at µ = 0. x1 = f (x0). . so is f 2(x). • Since f (x) is symmetric about x = 1/2. look � � at both f (x) and f 2(x) = f f (x) just before the transition. f 2 (x0) = ∗ �� d � f f (x) �x0 dx � � = f ∗(x0) f ∗ f (x0) = f ∗(x0) f ∗(x1) Thus. The two slopes are related by the chain rule.Let’s examine this transition more closely. x2 = f (x1) =∞ x2 = f 2(x0). 2.7. First. By deﬁnition. Using the chain rule. . suppose x0 = x�. Fig. f ∗(xn−1). 156 ∗ (32) . Then x1 = x 0 = x � and f 2 (x�) = f ∗(x�) f ∗(x�) = |f ∗(x�)|2.

2 1 ∗ We thus ﬁnd a cycle of period 2. and |f ∗ (x�)| > 1 =∞ |f 2 (x�)| > 1. What happens at the transition. 3. µ = 0. Fig. Moreover. where µ = 3/4? At µ = 3/4. ∗ ∗ 157 . The cycle is stable because |f 2 (x�)| < 1 and |f 2 (x�)| < 1. 1 2 x� = f (x�). the peaks of f 2 increase. Fig. the minimum decreases.7. f ∗ (x�) = −1 =∞ f 2(x�) = 1.785. Feigenbaum. such that 1 2 x� = f (x�). 4. x� and x�. Therefore f 2(x�) is tangent to the identity map. Just after the transition.For the example of µ < 3/4. 1 2 Feigenbaum. where µ > 3/4. if we start at x0 = 1/2. to x = 1/2. |f ∗ (x�)| < 1 =∞ |f 2 (x�)| < 1. f 2 develops 2 new ﬁxed points. then equation (32) shows that f ∗(1/2) = 0 =∞ f 2 (1/2) = 0 =∞ x = 1/2 is an extremum of f 2. ∗ ∗ Equation (32) also shows us that f 2 has an extremum at the x0 that iterates. µ = 0. These inverses of x = 1/2 are indicated on the ﬁgure for µ = 0. the extremum of f . under f .75.

¯ f 2 (x�) = f 2 (x�) = 0. 2. r = 1. 1 1 2 2 1 2 In general. 2. 1 2 This results trivially from equation (32). 1 2 ∗ ∗ At µ = µ2 . n. . since the period-2 oscillation gives f 2 (x�) = f ∗(x�) f ∗(x�) = f ∗ (x�) f ∗(x�) = f 2 (x�). . r = 1. . i. the slopes at the ﬁxed points of f 2 are equal: f 2 (x�) = f 2 (x�). . We thus perceive a sequence of bifurcations at increasing values of µ. . . . . 2. .Importantly. r r and the slopes f n∗ (x�) are all equal: r f n∗ (x�) = f ∗ (x�) f ∗(x�) . µ = µ1 . both x� and x� give rise 1 2 1 2 � � to 2 stable ﬁxed points of f 4(x) = f 2 f 2 (x) as µ increases still further. such that 1 2 n x�+1 = f (x�). . where the 21-cycle is superstable. . .e. r r � and x1 = f (x� ) n then each x� is a ﬁxed point of f n : r x� = f n (x�). . f ∗(x� ). . if x�. . Eventually. n r = 1. . n − 1 ∗ ∗ ∗ ∗ • The period doubling bifurcation derives from the equality of the ﬁxed points—because each ﬁxed point goes unstable for the same µ. r 1 2 n This slope equality is a crucial observation: • Just as the sole ﬁxed point x� of f (x) gives rise to 2 stable ﬁxed points x� and x� of f 2(x) as µ increases past µ = 3/4. the 2-cycle bifurcates to a 22 = 4 cycle. . ¯ 158 . x�. x� is a cycle of period n. there is a transition to a cycle of period 21 .. and is superstable at µ = µ2 . . At µ = µ1 = 3/4.

transition to period 4 (Feigenbaum. • µn = value of µ where 2n cycle is superstable. Fig. ¯ 4 and its doubling to be f = f 2(f 2). superstable 2-cycle ¯ (Feigenbaum. ¯ ¯ where • µn = value of µ at transition to a cycle of period 2n . 5). we consider the fundamental function to be f2. µ = µ2 . Fig. 6). Fig. µ = µ2 . ¯ Note that one of the superstable ﬁxed points is always at x = 1/2. . superstable 4-cycle ¯ (Feigenbaum. Note that in the case µ = µ2 . µ = µ1 . we are concerned with the functional compositions n+1 n� n� f2 = f2 f2 159 .We thus perceive the sequence µ1 < µ 1 < µ 2 < µ 2 < µ 3 < . 7). In general. .

There is geometric convergence: µ� − µn ≥ ν −n for large n. . 160 for large n. however. valid for any unimodal map with quadratic maximum. each increment in µ from one doubling to the next is reduced in size by a factor of 1/ν. The quadratic nature of the maximum means that in a Taylor expansion of f (x) about xmax.5 Scaling and universality The period-doubling bifurcations obey a precise scaling law. 16. Deﬁne µ� = value of µ when the iterates become aperiodic = 0. (obtained numerically. .Cycles of period 2n+1 are always born from the instability of the ﬁxed points of cycles of period 2n . ..669 .. is not the scaling law itself.e. The truly amazing result. . such that νn = µn+1 − µn ∗ν µn+2 − µn+1 ν = 4. “Unimodal” simply means that the map goes up and then down. 2 .e. � f ∗∗ (xmax) = 0. . Period doubling occurs ad inﬁnitum. but that π2 ∗∗ f (xmax) + . is universal.892486 . for the logistic map). f (xmax + π) = f (xmax) + πf ∗ (xmax) + the leading order nonlinearity is quadratic. That is. i. . i.

Therefore an understanding f 2 near its extremum at x = 1/2 will suﬃce to understand the period-doubling cascade.(There is also a relatively technical requirement that the Schwartzian derivative of f must be negative over the entire interval (Schuster)) This is an example of universality: if qualitative properties are present to enable periodic doubling. f 8. f 4.— whose dynamics can be approximated by a unimodal map would undergo period doubling bifurcations in the same quantitative manner. ¯ • superstable ﬁxed points always include x = 1/2. populations. after ¯ n n • reﬂection through x = 1/2. f 16) rapidly converges to a single func tion. oscillators. for f µ2 (x) looks just like ¯ fµ1 (x). Thus we expect that any system—ﬂuids. and • all ﬁxed points have the same slope. y = 1/2. then quantitative properties are predetermined. 2 To see how this works. ¯ ¯ 2 The parabolic curve within the dashed (red) square. f 2. consider fµ1 (x) and fµ2 (x) (top of Figures 5 and 7). etc. How may we understand the foundations of this universal behavior? Recall that • the 2n -cycle generated by f 2 is superstable at µ = µn . Feigenbaum. Figure 8. and • magniﬁcation such that the squares are equal size. 161 . The superposition of the ﬁrst 5 such functions (f.

¯ From one doubling to the next.502 . . The composition of doubled functions therefore has a “stable ﬁxed point” in the space of functions. What is it? The bifurcation diagram looks like x 1 d1 1/2 d2 µ1 µ1 µ2 µ2 µ Deﬁne dn = distance from x = 1/2 to nearest value of x that appears in the superstable 2n cycle (for µ = µn ). in the inﬁnite period-doubling limit. which is the same scale factor for each n (n large). . We shall see that � is also universal: � = 2. The scale reduction is based only on the functional composition n+1 n� n� f2 = f2 f2 This scale factor converges to a constant.Thus as n increases. 162 . this separation is reduced by the same scale factor: dn ◦ −�. dn+1 The negative sign arises because the adjacent ﬁxed point is alternately greater than and less than x = 1/2. a progressively smaller and smaller region near f ’s max imum becomes relevant—so only the order of the maximum matters.

dn ◦ −� =∞ lim (−�)ndn+1 ∈ rn n�� dn+1 2 lim (−�)nfµn+1 (0) ¯ n converges. is g1 for n = 0. n�� must exist. Thus a rescaling of the x-axis describes convergence to the limiting function ⎡ � x n 2 g1(x) = lim (−�)n fµn+1 .. ¯ n−1 For simplicity. ¯ 163 . ¯ n−1 The observation that.e. for n � 1. ¯ n�� (−�)n Here the nth interated function has its argument rescaled by 1/(−�) n and its value magniﬁed by (−�)n . at µ1 . ¯ • Figure 7 (top). is g1 for n = 1. Then 2 dn = fµn (0). Stated diﬀerently. n Thus g1 should be universal for all f ’s with quadratic maximum. shift the x axis so that x = 1/2 ∗ x = 0. n Our superposition of successive plots of f 2 suggests that this result may be generalized to the whole interval. 2 The rescaling of the x-axis shows explicitly that only the behavior of f µn+1 ¯ near x = 0 is important. • Figure 5 (top). when rescaled by �.16. i.6 Universal limit of iterated rescaled f ’s How may we describe the rescaling by the factor �? For µ = µn . ¯ 2 dn = fµn (1/2) − 1/2. at µ2 . dn is the 2n−1 iterate of x = 1/2.

1. .• g1 for n large looks like (Feigenbaum. the location of the elements of the doubled cycles (the circulation squares) is itself universal. ¯ n�� (−�)n Note that gi−1 � x 2 = lim (−�)nfµn+i−1 ¯ n�� (−�)n ⎡ � 1 x n−1+1 2 = lim (−�)(−�)n−1fµn−1+i ¯ n�� (−�) (−�)n−1 n (33) ⎡ Set m = n − 1. 16. Moreover.7 Doubling operator We generalize g1 by introducing a family of functions ⎡ � x n 2 gi = lim (−�)n fµn+i . Then f2 n−1+1 = f2 m+1 = f2 m � f2 m � 164 . . Figure 9) The function g1 is the universal limit of interated and rescaled f ’s. i = 0. .

Thus g is an unstable ﬁxed point of T . T gi . 165 . interates away from g.and ⎡ 1 1 x m 2 2 = lim (−�)(−�)mfµm+i (−�)m fµm+i ¯ ¯ m�� ⎨ (−�)m (−�) (−�)m ⎨ ⎨ ⎨ ⎨ � �⎨ ⎜� ⎧ ⎫ x g i ( −� ) ⎡ � �� x = −�gi gi −� m gi−1 ⎟ ⎨ ⎨ ⎨ ⎠ ⎩ ⎨ ⎨ ⎨ � ⎦ We thus deﬁne the doubling operator T such that ⎡ � �� x gi−1(x) = T gi (x) = −�gi gi −� Taking the limit i ∗ →. We must set a scale. −� n g(x) is the limit. ﬁrst write g(0) = −�g [g(0)] .8 Computation of � To determine �. where i is ﬁnite. and therefore set g(0) = 1 =∞ g(1) = −1/�. of rescaled f 2 . as n ∗ →. evaluated for µ� . 16. we also deﬁne i�� g(x) ∈ lim gi (x) 2 = lim (−�)nfµ� ¯ n�� n ⎡ x (−�)n � (34) We therefore conclude that g is a ﬁxed point of T : ⎡ � �� x g(x) = T g(x) = −�g g . Whereas g is a ﬁxed point of T .

We can however obtain a unique solution for � by specifying the nature (order) of g’s maximum (at zero) and requiring that g(x) be smooth.5028 . and use the short power law expansion g(x) = 1 + bx2.There is no general theory that can solve equation (34) for g. −1 . 16. which is within 10% of Feigenbaum’s � = 2. . .366 (neg root for max at x = 0) 4 and therefore � ◦ 2.73. 166 . Then.9 Linearized doubling operator We shall see that ν determines how quickly we move away from g under application of the doubling operator T . from equation (34).. We thus assume a quadratic maximum. g(x) = 1 + bx 2 � � bx2 = −�g 1 + 2 � � � �� 2 2 bx = −� 1 + b 1 + 2 � 2b2 2 = −�(1 + b) − x + O(x4) � Equating terms. 1+b � = −2b � −2 ± 12 b= ◦ −1. �= which yields. we shall calculate the eigenvalue that corresponds to instability of an unstable ﬁxed point. obtained by using terms up to x14. In essence.

Thus our ﬁrst task will be to linearize the doubling operator T . Therefore n applications of the doubling operator produce A second application of the doubling operator yields � � � � µ ¯ T T (fµ) = T 2fµ� + (¯ − µ� )LT fµ� Lfµ� νf + O (νf )2 . ν will then turn out to be one of its eigenvalues. ¯ ¯ ¯ ¯ � � µ ¯ T n fµ = T n fµ� + (¯ − µ� ) LT n−1fµ� · · · Lfµ� νf + O (νf )2 . −� −� −� (35) The ﬁrst term on the RHS derives from an expansion like g[f (x) + �f (x)] � g[f (x)] + g � [f (x)]�f (x). ¯ We ﬁrst expand fµ (x) around fµ� (x): ¯ ¯ µ ¯ fµ (x) ◦ fµ� (x) + (¯ − µ� ) νf (x). We seek to predict the scaling law µn − µ� ≥ ν −n . ¯ ¯ where the incremental change in function space is given by � � ωfµ(x) � ¯ νf (x) = ωµ �µ� ¯ ¯ Now apply the doubling operator T to fµ and linearize with respect to νf : ¯ ⎡ � �� x T fµ = −�fµ fµ ¯ ¯ ¯ −� ⎡ � � � �� x x ◦ −� [fµ� + (¯ − µ� ) νf ] ∝ fµ� µ ¯ + (¯ − µ� ) νf µ ¯ ¯ ¯ −� −� = T fµ� + (¯ − µ� )Lfµ� νf + O(νf 2 ) µ ¯ ¯ ¯ where Lf is the linearized doubling operator deﬁned by � ⎡ � �� � � ⎡ � ��� x x x ∗ Lf νf = −� f f νf + νf f . ¯ ¯ now expressed in terms of µi rather than µi . ¯ ¯ ¯ ¯ 167 (36) .

that L T fµ� ◦ ¯ Lg .) Thus for large n. a = c1 . . g � � Now assume that only one of �� is greater than one: �1 > 1. n � 1. and h(x) = θ1 . ¯ where ν = �1 . was later proven. ¯ g We simplify by introducing the eigenfunctions θ� and eigenvalues �� of Lg : Lg θ� = �� θ� . g 1 We can now simplify equation (36): µ ¯ T nfµ (x) = g(x) + (¯ − µ� ) · ν n · a · h(x). 2. n�1 n � 1. µ ¯ n � 1. similarly. yielding the approximation Ln νf ◦ �n c1 θ1 . . we expect convergence to the ﬁxed point g(x): ¯ ¯ ⎡ � x n 2 T n fµ� = (−�)nfµ� ◦ g(x). Write νf as a weighted sum of θ� : νf = � � � = 1. ¯ ¯ (−�)n Substituting g(x) into equation (36) and assuming. c� θ� Thus n applications of the linear operator Lg may be written as � Ln νf = � n c � θ� . �1 dominates the sum.For µ = µ� . �� < 1 for � = 1. Now note that when x = 0 and µ = µn . part of the original theory. . ¯ 168 . (37) T nfµ (x) ◦ g(x) + (¯ − µ� ) Ln νf (x). ¯ ¯ µ ¯ T n fµn (0) = g(0) + (¯n − µ� ) · ν n · a · h(0). � (This conjecture.

¯ ¯ n n Recall also that we have scaled g such that g(0) = 1. we obtain −� {g ∗ [g(0)] h(0) + h [g(0)]} = ν · h(0). Note that the approximation h(x) ◦ h(0) =∞ h[g(0)] = h(1) ◦ h(0).2 Recall that x = 0 is a ﬁxed point of fµn (due to the x-shift). Therefore ¯ 2 T n fµn (0) = (−�)nfµn (0) = 0. Now approximate h(x) by h(0). Seting x = 0. (38) 169 .10 Computation of ν Recall that ν is the eigenvalue that corresponds to the eigenfunction h(x). Then applying the linearized doubling operator (35) to h(x) yields � ⎡ � �� � � ⎡ � ��� x x x Lg h(x) = −� g ∗ g h +h g −� −� −� = ν · h(x). We thus obtain the Feigenbaum scaling law: n�� lim (¯n − µ� ) ν n = µ ¯ −1 = constant! a · h(0) 16. recalling that g(0) = 1. −� [g ∗ (1) + 1] = ν. Thus h(0) cancels in each term and. the ﬁrst term in a Taylor expansion about x = 0.

11 Comparison to experiments We have established the universality of � and ν: 170 . . 1979).72.73. we obtain g ∗ (1) = −�.) which is within 1% of the exact value ν = 4. Substituting that above. we obtain ν ◦ 4. −1 ∗ [g (1)g ∗∗(0)] . Then g ∗∗ (0) = Therefore Substituting into equation (38). This result derives from the crude approximation h(0) = h(1). Recall that we previously estimated � ◦ 2. Note that � g ∗ (0) = 0 and g ∗∗ (0) = 0 because we have assumed a quadratic maximum at x = 0.To obtain g ∗ (1). 16.669 .. . diﬀerentiate g(x) twice: ⎡ � �� −x g(x) = −�g g � � ⎡ � �� � � � �� −x −1 ∗ −x g ∗ (x) = −� g ∗ g · g � � � � ⎡ � �� ⎡ � ��2 ⎡ � �� � �� x −x −x −1 −x g ∗∗ (x) = g ∗∗ g g∗ + g∗ g g ∗∗ � −� � � � Substitute x = 0. � ν = �2 − � . Better approximations yield greater accuracy (Feigenbaum.

However we have seen that more complicated systems can also behave as if they depend on only a few degrees of freedom. one may expect that a one-dimensional map is contained.reduced by α reduced by δ These quantitative results hold if a qualitative condition—the maximum of f must be locally quadratic—holds. in a Rayleigh-B´nard system. Due to dissipation. within them. At ﬁrst glance this result may appear to pertain only to mathematical maps. The ﬁrst experimental veriﬁcation of this idea was due to Libchaber. e As the Rayleigh number increases beyond its critical value. a single convection roll develops an oscillatory wave: probe Ra=Ra c Ra>Ra c A probe of temperature X(t) is then oscillatory with frequency f 1 and period 1/f1. so to speak. Successive increases of Ra then yield a sequence of period doubling bifurca 171 .

Figure VIII. The experimental results are shown in BPV. ν = 4. .13b .669 . Identifying Ra with the control parameter µ in Feigenbaum’s theory.13a and VIII. .tions at Rayleigh numbers Ra1 < Ra2 < Ra3 < . Libcha ber found ν ◦ 4. .. Such is the power of scaling and universality! 172 .4 which is amazingly close to Feigenbaum’s prediction. .

Almost all our remarks will be on intermittency. intermittency and quasiperiodicity. • As r ∗ ri from above. near r = 166. • Rayleigh-Benard convection. Figure IX. system displays stable oscillations (e. Deﬁnition: Intermittency is the occurrence of a signal that alternates ran domly between regular (laminar) phases and relatively short irregular bursts. • For r > ri (r − ri small). by Manneville and Pomeau). In the exercises we have already seen examples. and disappear for r < ri. particulary in the Lorenz model (where it was discovered. BPV. Examples: • The Lorenz model.1 General characteristics of intermittency Let r = control parameter.g.b Manneville and Pomeau (1980) 17.17 Intermittency (and quasiperiodicity) In this lecture we discuss the other two generic routes to chaos.9 Figure 1a. 173 . a limit cycle). system in in the intermittent regime: stable oscillations are interrupted by ﬂuctuations. The following summarizes the behavior with respect to r: • For r < ri . we close with a brief de scription of quasiperiodicity. • Only the average intermission time between ﬂuctuations varies. the ﬂuctuations become increasingly rare. not their amplitude nor their duration..

2 One-dimensional map We consider the instability of a Poincar´ map due to the crossing of the unit e circle at (+1) by an eigenvalue of the Floquet matrix. we ﬁnd that the ﬁrst term vanishes: u∗ (u = 0. 17. r = ri ) = 0. r) in the neighborhood of u = 0 and r = ri .” (i. • Scaling law for intermission times.We seek theories for • Linear stability of the limit cycle and “relaminarization. we have u∗ = �(ri)u = u. • Scaling law for Lyaponov exponents. Expand to ﬁrst order in (r − ri ) and second order in u: � � � ∗ ∗ ∗ ωu � 1 2 ω 2u � ωu � ∗ � + u · � + (r − ri ) � u (u. Let u be the coordinate in the plane of the Poincar´ section that points in e the direction of the eigenvector whose eigenvalue � crosses +1.e. (39) Taking �(ri ) = 1 at the intermittency threshold. return to stability after irregular bursts). ri) + u · ωu � i 2 ωr � i ωu2 � i 0.r Evaluating equation (39). 174 .r 0.r 0. This corresponds to the speciﬁc case of Type I intermittency. r) ◦ u∗ (0. The lowest-order approximation of the 1-D map constructed along this line is u∗ = �(r)u. (40) We consider this to be the leading term of a Taylor series expansion of u ∗ (u.

• u− is stable ﬁxed point.From equation (40).e. • u+ is unstable. r = ri . Graphically. 0 u 0 175 . r < ri . u− u’ 0 u+ u • π = 0. we have the following system: u’ • π < 0. i. Finally. • u− = u+ = 0 is marginally stable. we have � ωu∗ � � = �(ri) = 1. where π is now the control parameter. i.r � 1 ω 2u∗ � � = 1 2 � 2 ωu 0.e.ri π ≥ (r − ri ). • u∗ is tangent to indentity map. rescale u such that and set The model now reads u∗ = u + π + u 2 . ωu � i 0.

the itera tions diverge. the iterations look like u’ • u− is an attractor for initial conditions u < u+. 176 .u’ • π > 0. 0 u For π < 0. • no ﬁxed points. r > ri . • For initial conditions u > u+ . i. u− u+ u The situation changes for π > 0. • Iterations beginning at u < 0 drift towards u > 0. r > ri : u’ • No ﬁxed points. the drift for π > 0 corresponds to a ﬂow qualitatively similar to the stable oscillations near u = 0 for π < 0.e. 0 u The ﬁxed points of u∗(u) represent stable oscillations of the continuous ﬂow.e. i. Thus for u ◦ 0.

3 Average duration of laminar phase What can we say about the average duration of the laminar phases? Writing our theoretical model as a map indexed by k. or “relaminarize”? Qualitatively. the picture can look like u’ u Note that the precise timing of the turbulent burst is unpredictable. The iterations eventually run away and become unstable—this is the inter mittent burst of noise. there is no ﬁxed point. we have uk+1 = uk + π + u2 . not a ﬂow.However. 17. when π > 0. How does the laminar phase begin again. Moreover the Lorenz map itself contains a discontinuity. The discontinuity is not inconsistent with the presumed continuity of the underlying equations of motion—this is a map. k 177 . and thus no periodic solution. corresponding to the location of the unstable ﬁxed point.

� α or k = ± π−1/2. uk+1 − uk is no longer small. When k � π−1/2. we can instead write the diﬀerential equation du = π + u2 . π > 0).e. then φ ≥ π−1/2 or φ ≥ (r − ri )−1/2. � tan π � 1/2 k . We thus have u(k) = π We see that u(k) diverges when π1/2k = ± α 2 1/2 � 1/2 (k − k0 ) . to ﬁrst order. and the diﬀerential approxi mation of the diﬀerence equation is no longer valid.4 Lyaponov number We can also predict a scaling law for the Lyaponov number.For uk+1 ◦ uk . dk The general solution of this o. νuk+1 ◦ �1 νuk 178 . is u(k) = π 1/2 tan π Take k0 = 0. 2 The divergence signiﬁes a turbulent burst. (41) Thus the laminar phase lasts increasingly long as the threshold r = r i is approached from above. Thus: if φ = time (≥ number of iterations) needed to traverse the channel. Near the ﬁxed point (u ◦ 0. the step at which iterations traverse the narrowest part of the channel. 17.d. the increment νuk+1 due to an increment uk is.

16–17 179 .) Results from the Lorenz model verify this prediction. Suppose N ◦ the duration of the laminar phase. N i N N φ where the last relation used equation (41). Figure IX. After N iterations. Then �N > 1 and �N −1 ◦ �N −2 ◦ · · · ◦ �1 ◦ 1.where �1 is eigenvalue that passes through (+1). Figure IX. (Recall that ln � = Lyaponov exponent. Figure IX. The “intermittent channel” of the Lorenz map is seen in BPV. The Lyapunov number � is 1 � �N 1 1 � �= �i ◦ ≥ ≥ ≥ π. νuN ◦ �N �N −1�N −2 · · · �1 νu1.15 Behavior qualitatively similar to that predicted by our model has been ob served in the B-Z reaction: BPV.14 and the associated π1/2 scaling of the Lyaponov number is seen in BPV.

5. the latter case results in the addition of a second oscillation.17.. |�| > 1.e.5 Quasiperiodicity Finally. or a torus T 2 .. he envisioned the following sequence of events as Re increases beyond Rec : • Laminar ﬂow (constant velocity) becomes periodic with frequency f 1 by a Hopf bifurcation.1 An historical note In 1944. This is a Hopf bifurcation: the transformation of a limit cycle to a quasiperi odic ﬂow. The route to chaos via quasiperiodicity describes how a torus T 2 (i. we make a few remarks about the third universal route to chaos. 17. Brieﬂy. known as quasiperiodicity. This corresponds to the transition via quasiperiod icity. 180 . As we have seen. • −1 (as we saw in the introduction to period doubling). another Hopf bifurcation. Recall that there are 3 generic ways in which a limit cycle on a Poincar´ map e may become unstable: An eigenvalue � of the Floquet matrix (the Jacobian of the map) crosses the unit circle at • +1 (as in the example of intermittency above).e. • Period ﬂow∗quasiperiodic ﬂow. the Russian physicist Landau proposed a theory for the transition from laminar ﬂow to turbulence as the Reynolds number is increased. i. a quasiperiodic ﬂow) can become a strange attractor. and • � = � ± iλ. The sec ond frequency f2 is incommensurate with f1 .

f4. however. .e. The Ruelle-Takens theory is the quasiperiodic route to chaos. from the work of Lorenz. fr appear in succession (due to more Hopf bifurcations). Recall that we have learned previously that. and • the chaos occurs on a strange attractor. • The torus T 3 can become unstable and be replaced by a strange attrac tor. We now know. dimension of phase space > attractor dimension. • For r large. .. the following sequence of events can occur: • Laminar ﬂow ∗ oscillation with frequency f1 . . As a control parameter is varied. that • 3 degrees of freedom suﬃce to give rise to a chaotic ﬂow. the spectrum appears continuous and the ﬂow (on a torus T r ) is aperiodic (i.• More incommensurate frequencies f3. Thus a consequence of Landau’s theory is that a system must have many degrees of freedom to become chaotic. . which is distinct from a torus (since trajectories diverge on the strange attractor). • A second Hopf bifurcation adds a second (incommensurate) frequency f2 . turbulent). 17. for dissipative ﬂows. • A third Hopf bifurcation adds a third frequency f3 .2 Ruelle-Takens theory Lorenz’s observations were deduced theoretically by Ruelle and Takens in 1971.5. 181 .

21 182 .The transition is demonstrated beautifully in terms of changing power spectra in the Rayleigh-B´nard experiment described by e Libchaber et al. 7: 73-84. S. and C. Fauve. Physica D. Figures VII. such as the Rayleighe B´nard experiment of e BPV.1983.20. Two-parameter study of the routes to chaos. Note that the Rayleigh number of the two spectra varies by less than 1%.. Figure 15 Libchaber. VII. Such a transition can also be seen in Poincar´ sections. Laroche.. A.

- Or-421051- Digital Signal Processing
- 111 Signal Processing
- Lecture 1&2
- @Kul 2 SHM - Circular - Energi
- 11 04 DFT
- Chaotic Dynamics
- Approximate Synthesis of Four Bar Linkages
- Pre IB Syllabus
- Chap2 Solutions Sp07
- me471_lec02
- FFT
- leise
- free_damped_and_forced_oscillations.doc
- Einstein's Mistakes in Deriving Special Theory of Relativity
- Complex Var 1
- StrDyn_0910
- (Chapman & Hall_CRC Applied Mathematics and Nonlinear Science Series) Dean G. Duffy-Mixed Boundary Value Problems-Chapman & Hall_CRC Press (2008)
- 1103.6084
- Tutorial 1
- Finite Differential
- Chapter 2
- An Efficient Technique for the Point Reactor Kinetics Equations With Newtonian Temperature Feedback Effects
- ES361_Spring2014_Homework1
- Sylabus Nda
- Template VSU2013en Final
- Oscillations and Waves
- Article 8
- Nonlinear Vibrations
- Fluid 7835
- A Brief Introduction to SCILAB

Are you sure?

This action might not be possible to undo. Are you sure you want to continue?

We've moved you to where you read on your other device.

Get the full title to continue

Get the full title to continue reading from where you left off, or restart the preview.

scribd