Вы находитесь на странице: 1из 203

E S S E N T I A L S O F H A M I LTO N I A N DY NA M I C S

Classical dynamics is one of the cornerstones of advanced education in physics and


applied mathematics, with applications across engineering, chemistry, and biology.
In this book, the author uses a concise and pedagogical style to cover all the
topics necessary for a graduate-level course in dynamics based on Hamiltonian
methods. Readers are introduced to the impressive advances in the field during
the second half of the twentieth-century, including KAM theory and determinis-
tic chaos. Essential to these developments are some exciting ideas from modern
mathematics, which are introduced carefully and selectively. Core concepts and
techniques are discussed, together with numerous concrete examples to illustrate
key principles. A special feature of the book is the use of computer software to
investigate complex dynamical systems, both analytically and numerically.
This text is ideal for graduate students and advanced undergraduates who
are already familiar with the Newtonian and Lagrangian treatments of classical
mechanics. The book is well suited to a one-semester course, but is easily adapted
to a more concentrated format of one-quarter or a trimester. A solutions man-
ual and introduction to Mathematica are available online at www.cambridge.
org/Lowenstein.

J O H N H . L O W E N S T E I N is Professor Emeritus of Physics at New York


University and has been conducting research in nonlinear dynamics for more than
20 years. Prior to that, his research focus was in quantum field theory with an
emphasis on soluble models and renormalized perturbation theory.
ES S ENT I AL S O F H AMILTONIAN
DY NA MI CS

JOHN H. LOWENSTEIN
Department of Physics, New York University
CAMBRIDGE UNIVERSITY PRESS
Cambridge, New York, Melbourne, Madrid, Cape Town,
Singapore, So Paulo, Delhi, Tokyo, Mexico City
Cambridge University Press
The Edinburgh Building, Cambridge CB2 8RU, UK

Published in the United States of America by Cambridge University Press, New York

www.cambridge.org
Information on this title: www.cambridge.org/9781107005204


c J. H. Lowenstein 2012

This publication is in copyright. Subject to statutory exception


and to the provisions of relevant collective licensing agreements,
no reproduction of any part may take place without the written
permission of Cambridge University Press.

First published 2012

Printed in the United Kingdom at the University Press, Cambridge

A catalog record for this publication is available from the British Library

ISBN 978-1-107-00520-4 Hardback

Additional resources for this publication at www.cambridge.org/Lowenstein.

Cambridge University Press has no responsibility for the persistence or


accuracy of URLs for external or third-party internet websites referred to
in this publication, and does not guarantee that any content on such
websites is, or will remain, accurate or appropriate.
To Marcia
Contents

Preface page xi

1 Fundamentals of classical dynamics 1


1.1 Newtonian mechanics 1
1.2 Configuration space 2
1.2.1 Particle on a sphere 2
1.2.2 Planar double pendulum 2
1.2.3 Independence of constraints 4
1.2.4 Configuration space as a differential manifold 5
1.3 Lagrangian formulation of Newtonian mechanics 7
1.4 Hamiltonian formulation 8
1.4.1 Introduction of the Hamiltonian 8
1.4.2 Phase space 10
1.5 Examples 11
1.5.1 Harmonic oscillator 11
1.5.2 Simple pendulum 11
1.5.3 Exact solution of the pendulum equations 14
1.5.4 A bead on a rotating circle 15
1.5.5 Spherical pendulum 18
1.5.6 Rigid body with one point fixed 21
Exercises 26
2 The Hamiltonian formalism 29
2.1 The Poisson bracket 29
2.2 Canonical transformations 30
2.3 Simple examples 32
2.4 Canonical invariance of Lagrange brackets 32
2.5 Generating functions for canonical transformations 34
2.6 Canonical invariance of phase-space volume 38

vii
viii Contents

2.7 Covariance of Hamiltons equations with explicit time


dependence 39
2.8 Continuous one-parameter groups of canonical transformations 40
2.9 The Hamiltonian formulation of electrodynamics 42
2.10 Normal modes for linear oscillators 44
Exercises 54
3 Integrable systems 56
3.1 The LiouvilleArnold theorem 56
3.2 Fast track for separable systems 57
3.3 System with one degree of freedom 60
3.4 The Kepler problem in spherical polar coordinates 61
3.5 Proof of the LiouvilleArnold theorem 64
3.6 Planar free-particle examples 69
3.6.1 A two-dimensional free particle, model 1 70
3.6.2 A two-dimensional free particle, model 2 71
3.6.3 Circular stadium billiard 73
3.7 Spherical pendulum 78
3.8 The three-particle Toda model 85
Exercises 91
4 Canonical perturbation theory 97
4.1 General approach 97
4.2 Simple pendulum revisited 98
4.3 Two harmonic oscillators with quartic coupling 102
4.4 Gyrating charge in an electrostatic wave 104
4.5 BirkhoffGustavson perturbation theory 109
4.6 A formal second integral for the HnonHeiles model 110
4.7 Integrability analysis of HnonHeiles-like systems 112
Exercises 115
5 Order and chaos in Hamiltonian systems 121
5.1 Nonlinear stability 122
5.2 The KAM theorem 122
5.3 Nonlinear stability of the Lagrange points 127
5.4 Kicked oscillators 129
5.5 Lyapunov exponents 136
5.6 Web-map Lyapunov exponents 138
5.7 A model of a chaotically rotating moon 139
Exercises 142
6 The swing-spring 148
6.1 Two-dimensional motion 149
Contents ix

6.2 Integrable approximations for small oscillations 152


6.3 Three commuting integrals 154
6.4 Dynamics on the level sets 156
6.5 Constraints among the integrals 158
6.6 Flow coordinates and the period lattice 165
6.7 Monodromy 168
6.8 Periodic shift of the swing plane 172
6.9 The swing-spring in molecular modeling 176
Exercises 177
Appendix Mathematica samples 178
A.1 Numerical integration of equations of motion 178
A.2 Hyperion movie 179
A.3 Simple pendulum perturbation theory 181
A.4 BirkhoffGustavson perturbation theory 181
References 184
Index 186
Preface

This is a textbook on classical Hamiltonian dynamics designed primarily for stu-


dents commencing graduate studies in physics. The aim is to cover all essential
topics in a relatively concise format, without sacrificing the intellectual coherence
of the subject, or the conceptual precision which is the sine qua non of advanced
education in physics.
Encouraged by my colleagues at New York University, I have taken it as a ped-
agogical challenge to create a textbook suitable for a twenty-first-century course
of duration no more than one semester (at NYU, the material is covered in about
two-thirds of a semester). To do so, I have chosen to limit the scope of the book in
certain important ways. It is assumed that the student has already had a course in
which Newtonian mechanics, in both F = ma and Lagrangian versions, has been
systematically developed and applied to a standard array of soluble examples: the
harmonic oscillator, the simple pendulum, the Kepler problem, small oscillations
(normal modes), and rigid-body motion. In the present book, the Hamiltonian for-
mulation in phase space is introduced at the outset and applied directly to the same
familiar systems.
Topics usually found in more encyclopedic textbooks, but omitted from the
present treatment, include dissipative systems, nonholonomic constraints, special
and general theories of relativity, continuum mechanics, and classical field theory.
A further choice I have made is to limit the use of advanced differential geom-
etry. Although I consider the concept of a differential manifold to be absolutely
essential for a clear understanding of such concepts as configuration space, phase
space, degrees of freedom, and generalized coordinates, I have found that the use
of differential forms is neither necessary nor desirable for this book.
One of the advantages which a twenty-first-century student enjoys in comparison
with students of earlier generations is a ready access to computers and to software
designed to assist in what would otherwise be long, tedious exercises in algebra and
analysis. In writing this book, I have assumed that the reader has already acquired

xi
xii Preface

some skill and familiarity with elementary scientific programming, and is willing
and able to apply such capabilities in solving some instructive numerical exer-
cises in dynamics. I have found Mathematica (a product of Wolfram Research,
Inc.) to be well suited to the small-scale calculations which I have used to pro-
vide concrete examples, with illustrative graphs and figures, throughout the book.
Obviously the choice of software for the numerical exercises is not unique, and
different instructors will have their own preferences as to how to implement the
numerical component of the course. For those who are interested, I have included
in the appendix some samples of the Mathematica programming which I have used
in the text.
Let us now briefly summarize the content of the book. The opening chapter
contains a rapid review of classical mechanics, from Newtons laws through the
Lagrangian and Hamiltonian formulations, with a number of instructive exam-
ples. This is followed, in Chapter 2, by an introduction to the core concepts of
the Hamiltonian formalism, including phase-space geometry, Hamiltons equations
of motion, Poisson brackets, and canonical transformations. Here we emphasize
an algebraic approach that parallels in certain ways the canonical commutation
algebra of quantum mechanics.
In Chapter 3 we turn to a systematic treatment of an extremely important class
of dynamical systems, namely those which are integrable, in the sense that there
is a complete set of conserved quantities. In the simplest of the integrable systems
discussed in Chapter 3, the method of separation of variables leads directly to a
reformulation in terms of special coordinates (the action-angle variables) for which
each degree of freedom is described as uniform motion on a circle. More generally,
the powerful LiouvilleArnold theorem provides not only important topological
information about the phase space, but also a constructive method for finding an
appropriate set of action-angle variables in models that are not separable in their
original system of coordinates.
While integrable systems are very rare among Hamiltonian systems (rigorously,
the probability of finding one in a random search of function space is zero), it is
often the case that a realistic system can be well modeled in a portion of its phase
space by an integrable one, with small corrections that can be estimated within
the Hamiltonian formalism. This is the realm of canonical perturbation theory, the
topic of Chapter 4.
In canonical perturbation theory, one constructs a formal solution of the equa-
tions of motion, correct to finite order in a small perturbation parameter. Although
in each order the model is in some sense integrable, the perturbative solutions of the
equations of motion do not always converge to exact solutions in the limit of infi-
nite order. Nonetheless, thanks to the remarkable theorem of Kolmogorov, Arnold,
and Moser (KAM), we now know how to set up the perturbation series, namely
Preface xiii

as a finely tuned renormalization process, so that, for a sufficiently small pertur-


bation parameter, most initial conditions do indeed lead to a definite limit. For
the remaining initial conditions, convergence may break down and seemingly ran-
dom (chaotic) orbits are possible. The KAM theory and the fascinating interplay
between order and chaos in Hamiltonian dynamical systems will be explored in
Chapter 5. Examples will be drawn from one of the most important current research
areas in Hamiltonian dynamics, namely the motions, both stable and chaotic, of the
planets and smaller bodies in the Solar System.
In the final chapter of the book, the full power of the concepts and methods
developed in the preceding chapters is brought to bear on a particularly fascinating
dynamical system, the elastic three-dimensional pendulum. This system, known as
a swing-spring, provides an excellent model for certain excitations of the carbon
dioxide molecule, a quantum-mechanical system for which experimentally verifi-
able information can be gleaned from quantization of the classical model. What
makes the swing-spring particularly interesting in this regard is the presence of
nontrivial monodromy, which complicates the task of classifying the quantized
energy levels, but also is associated with an observable phenomenon, namely a kind
of intermittent pendulum-like swinging. We will explore some of these features in
some detail in Chapter 6.
John H. Lowenstein
1
Fundamentals of classical dynamics

In this first chapter, we review the fundamentals and rather quickly introduce the
central analytical and geometrical concepts of the Hamiltonian approach. We then
illustrate these ideas in the context of a number of familiar examples.

1.1 Newtonian mechanics


The physical system we will be dealing with throughout this book consists of N
pointlike particles with masses m k and position vectors xk = (xk , yk , z k ), con-
strained at each time t by C independent equations
f j (x1 , x2 , . . . , x N , t) = 0, j = 1, 2, . . . , C, (1.1)
and moving under the influence of a potential-energy function V (x1 , x2 , . . . , x N , t).
Such a system is said to have n degrees of freedom, where
n = 3N C.
According to the principles of Newtonian mechanics, the instantaneous state of
the system is prescribed by a 3N -dimensional vector X and its time derivative,
X = (x1 , x2 , . . . , x N ), X = (x1 , x2 , . . . , x N ), (1.2)
which evolve in time according to the system of second-order differential equations
(k = 1, 2, . . . , N )
..
m k xk = total force on kth particle
= k V (x1 , x2 , . . . , x N , t) + constraint forces. (1.3)
Given the state of the system at any time t = t0 , the equations (1.3) determine the
state for all later times. We will hardly ever be interested in the explicit representa-
tion of the forces of constraint, so in (1.3) we suppress the details. In fact, we will
very quickly adopt a more efficient dynamical formalism in which the constraints

1
2 Fundamentals of classical dynamics

are completely absent from the equations of motion. In those rare instances in
which we are curious about the constraints, we will always be able to calculate
them by first determining xk (t), k = 1, . . . , N , and then solving (1.3) for the
forces.

1.2 Configuration space


Unless otherwise specified, our constraint functions f j (X, t), X R3N , j =
1, . . . , C, are assumed to be smooth (continuously differentiable), so the equations
(1.1) define an n-dimensional surface in every region of Euclidean 3N -space where
the constraint equations are independent. This surface is the configuration space of
the system. Before discussing more precisely the mathematics, let us consider two
simple examples.

1.2.1 Particle on a sphere


A particle moves on a sphere of time-varying radius a(t) > 0. Its position vector

X = x = (x, y, z)

is constrained by
f (X, t) = x 2 + y 2 + z 2 a(t)2 = 0.

With N = 1 and C = 1, the number of degrees of freedom is

n = 3 1 = 2.

We can conveniently represent the position on the two-dimensional spherical


surface by a pair of spherical polar coordinates and , as shown in Figure 1.1.
These angular coordinates are uniquely defined for in the open interval (0, ),
i.e. everywhere except at the points = 0, .

1.2.2 Planar double pendulum


The system consists of two point masses connected by massless rigid rods to each
other and to a suspension point as shown in Figure 1.2, constrained to move in
a single plane. Now the positions of the two masses are specified by a vector in
six-dimensional Euclidean space,

X = (x1 , y1 , z 1 , x2 , y2 , z 2 ) R6 ,
1.2 Configuration space 3

^r

a(t) ^
y

Figure 1.1 A particle on a sphere of radius a(t).

1 a

b
2

Figure 1.2 A planar double pendulum.

and the constraints are given by

f 1 (X) = y1 = 0,
f 2 (X) = y2 = 0,
f 3 (X) = x12 + z 12 a 2 = 0,
f 4 (X) = (x2 x1 )2 + (z 2 z 1 )2 b2 = 0.

In this example, with N = 2 and C = 4, the number of degrees of freedom is

n = 6 4 = 2,
4 Fundamentals of classical dynamics

and a convenient choice of coordinates for configuration space is the pair of angles
1 and 2 shown in Figure 1.2. The range of the angles is 0 1 , 2 < 2.

1.2.3 Independence of constraints


In determining the configuration space Q of a system, it is of course crucial that
the relevant constraints be nontrivial and independent. In precise terms, this means
that the C gradient vectors f j (X, t) should be non-zero and linearly indepen-
dent at each point X of the configuration space at time t. By continuity, the linear
independence will hold for a neighborhood of Q in Rn .
To check for independence is completely straightforward. One first constructs
the matrix of partial derivatives

f1 f1 f1
X 1 X 2 X 3N

f f2 f2
2


Df(X) = X 1 X 2 X 3N .
. . . .
.. .. .. ..

fC fC fC

X1 X2 X 3N
The gradient vectors are the rows of this matrix, and are linearly independent if, for
each X Q, there exists at least one C-dimensional minor of Df(X) which does
not vanish.
In our particle-on-a-sphere example, the calculation is particularly simple. There
is only one constraint, with gradient vector

f f f
f (X) = , , = (2x, 2y, 2z).
x y z
For all points on the sphere x 2 + y 2 + z 2 = a(t)2 > 0, at least one of the quantities
2x, 2y, or 2z is nonvanishing, so the nontriviality of the constraint is verified.
The planar double pendulum, with its four constraints, requires a somewhat more
elaborate calculation. The matrix of derivatives

0 1 0 0 0 0
0 0 0 0 1 0
Df(X) =


2x1 0 2z 1 0 0 0
2(x1 x2 ) 0 2(z 1 z 2 ) 2(x2 x1 ) 0 2(z 2 z 1 )
has the nontrivial minors
4(z 1 x2 x1 z 2 ), 4z 1 (z 2 z 1 ), 4z 1 (x2 x1 ), 4x1 (z 2 z 1 ), 4x1 (x2 x1 ).
One readily verifies that on Q the sum of the squares of the last four minors is
1.2 Configuration space 5

16(x12 + z 12 )((x1 x2 )2 + (z 2 z 1 )2 ) = 16a 2 b2 > 0,


which establishes the nontriviality and independence of the four constraints in a
neighborhood of Q.

1.2.4 Configuration space as a differential manifold


It is a standard result of differential geometry that a configuration space Q defined
by a set of 3N n nontrivial, independent constraints is an n-dimensional differ-
ential manifold. Roughly speaking, this means that locally Q has the structure of
n-dimensional Euclidean space. More precisely, there exists a finite or countable
set of neighborhoods (bounded, open, simply connected subsets) U j Rn , as well
as diffeomorphisms (smooth mappings with smooth inverses) j from U j into Q

such that the image sets j (U j ) cover Q, i.e. Q j j (U j ). The situation is


sketched in Figure 1.3. Rectangular coordinates in each of the U j endow Q with
an n-dimensional coordinate system on each of the patches j (U j ). The mappings
j are known as charts, and a differential manifold is covered, not necessarily
uniquely, by a complete atlas of charts.
To illustrate the notion of configuration space as a differential manifold, let us
consider once again our simple examples. Figure 1.4 shows the spherical configura-
tion manifold of our first example, together with the chart associated with ordinary
polar coordinates and . There is one departure from conventional usage, how-
ever: in order to conform to the definition of a chart as having an open domain, the
angular coordinates vary over the open rectangle (0, ) (0, 2), with the bound-
ary excluded. The image of this open domain is the sphere with the semicircular
arc = 0, 0 removed, as shown in Figure 1.4.

qn. .q. 3 q2

(U)

q1 Q

Figure 1.3 A sketch of a chart mapping a neighborhood U in Rn into the


n-dimensional configuration manifold Q. This supplies a part of Q with a system
of generalized coordinates q1 , . . . , qn .
6 Fundamentals of classical dynamics

z

=0
2

0
= Q
y

x
0
=

Figure 1.4 A chart mapping the open planar rectangle (0, ) (0, 2 ) into the
sphere. An additional chart is needed to cover the excluded semi-circular arc.

To obtain a complete atlas of charts for the sphere, we must supply at least one
additional chart to cover the excluded semicircle. One choice (among infinitely
many) is to choose another system of spherical polar coordinates, (  ,  )
(0, ) (0, 2) with  = 0 along the y axis and  = 0 in the half-plane
z = 0, x 0. Since the excluded arcs of the two coordinate patches do not
intersect, we have complete coverage of the sphere by the two charts.
In the double-pendulum example, it is natural to introduce the following chart:
x1 = a cos 1 , y1 = 0, z 1 = a sin 1 ,
(1.4)
x2 = b cos 2 , y2 = 0, z 2 = b sin 2 ,
which maps the angular rectangle (0, 2)(0, 2) diffeomorphically into the two-
dimensional configuration space, which consists of a direct product of two circles
(a 2-torus) embedded in R6 . In contrast to our first example, we cannot visualize
this embedding. A convenient representation of the torus is obtained by extending
each of the angular variables 1 and 2 to the full real line, identifying points that
differ by a translation (2m, 2n), where m and n are arbitrary integers. We will
refer to this multi-valued assignment of coordinates as a multichart. Let us use this
representation to construct a complete atlas of genuine charts.
Figure 1.5 shows a small part of the 1 , 2 plane. We note that the configuration
manifold may be identified with the half-open square U = [0, 2) [0, 2) in
the center of the figure, or with any translate of U . For our atlas, we need to cover
U with open neighborhoods. To guarantee invertibility of the corresponding charts,
we choose each neighborhood to be an open subset of a translate of U . Specifically,
we define four open angular rectangles, Ri j , i, j {+, }, such that each Ri j is a
translate of the interior of U and U is properly contained in their union. One such
choice is shown in Figure 1.5. Clearly, the four diffeomorphisms defined by (1.4)
1.3 Lagrangian formulation of Newtonian mechanics 7

R-+
R++

R-- R+-

Figure 1.5 The images of the four squares Ri j cover the configuration manifold
a 2-torus of the planar double pendulum.

with respective domains Ri j , i = +, , j = +, , comprise a complete atlas of


charts for the configuration manifold.

1.3 Lagrangian formulation of Newtonian mechanics


Let us suppose that an N -particle system with C smooth constraints has been
assigned a set of generalized coordinates q1 , . . . , qn , n = 3N C, on a por-
tion of its configuration manifold. As a consequence of Newtons laws (see e.g.
[1]), the state of the system at time t is specified by the instantaneous values of
def
qk , qk = dqk /dt, k = 1, . . . , n. Once the state is known for an initial time t0 , it is
uniquely determined for all other times by the EulerLagrange equations,
d L L
= 0, k = 1, . . . , n, (1.5)
dt qk qk
where the Lagrangian L is defined as

1
N
L(q1 , . . . , qn , q1 , . . . , qn , t) = m i xi (q, q, t)2 V (x1 (q, t), . . . , x N (q, t), t).
2 k=1
(1.6)
8 Fundamentals of classical dynamics

Since
xi xi
xi = qk + ,
k
q k t
we can rewrite (1.6) as
1
L= Akl qk ql + Bk qk U, (1.7)
2
where
xi xi
Akl (q, t) = mi ,
i
qk ql
xi xi
Bk (q, t) = mi ,
i
qk t
1 xi xi
U (q, t) = V (x(q, t), t) mi .
2 i t t
In (1.7), and in other formulas throughout the book, we adopt the convention of
summing over repeated indices without explicitly writing the summation symbol.
The EulerLagrange equations then take the form of a system of second-order
ordinary differential equations (ODEs) for the generalized coordinates as functions
of time:
..
Akl q l +Dklm ql qm + E kl ql + Fk = 0, k = 1, 2, . . . , n, (1.8)
with
Akl 1 Alm Akl Bk Bl Bk U
Dklm = , E kl = + , Fk = + .
qm 2 qk t ql qk t qk

1.4 Hamiltonian formulation


1.4.1 Introduction of the Hamiltonian
For the purpose of numerical integration of the equations of motion, it is standard
practice to recast the latter as a system of 2n first-order ODEs. For example, in
terms of the 2n generalized position and velocity variables qk , vk , k = 1, . . . , n,
(1.8) is clearly equivalent to
q j = v j ,
v j = (A1 ) jk (Dklm vl vm + E kl vl + Fk ),
provided that the matrix A is invertible. But, as the reader is no doubt aware, there
is an even better choice of new variables, namely the generalized positions and
their canonical momenta,
q 1 , q 2 , . . . , q n , p1 , p2 , . . . , pn ,
1.4 Hamiltonian formulation 9

where
def L
pk = . (1.9)
qk
With the same invertibility condition for the matrix A as before, we can solve (1.9)
for qk ,
qk (q, p, t) = (A1 ) jk (q, t) ( pk Bk (q, t)) ,

and introduce the Hamiltonian



H (q, p, t) = pk qk (q, p, t) L(q, q(q, p, t), t). (1.10)
k

Using the defining relation (1.9) and the EulerLagrange equations, rewritten as
L
pk = ,
qk
we obtain the following system of first-order ODEs (Hamiltons equations):
H
qk = , (1.11)
pk
H
pk = . (1.12)
qk
Introducing the 2n-dimensional vector

= (q1 , . . . , qn , p1 , . . . , pn )

and the matrix



0n 1n
= , (1.13)
1n 0n

(here 0n and 1n are the n-dimensional null and unit matrices, respectively)
Hamiltons equations can be written in an elegant form,

=  H. (1.14)

We note that the time evolution of H itself is particlarly simple. Using


Hamiltons equations, we have

H H H H H H H H H
H = qk + pk + = + + = ,
qk pk t qk pk pk qk t t
so that if the Hamiltonian has no explicit time dependence, i.e. H/t = 0, then
H is a conserved quantity.
10 Fundamentals of classical dynamics

If xi has no explicit time dependence, then, from (1.7), the Lagrangian takes the
form
1
L = Akl (q)qk ql V (x(q), t),
2
and the Hamiltonian (1.10) is just the sum of kinetic and potential energies (i.e. the
total energy). The latter is conserved only if the potential energy has no explicit
time dependence.

1.4.2 Phase space


By assigning an n-dimensional momentum space Pq (copy of Rn ) to each configu-
ration point q Q, we obtain a new, 2n-dimensional manifold , the phase space.
This manifold has the structure of a fiber bundle, with base Q and fibers Pq . The
charts on Q extend naturally to charts on
.
When the Hamiltonian is conserved (i.e. H = H (q, p)), the phase space strat-
ifies into distinct constant-H surfaces of dimension 2n 1. A state in the surface
H (q, p) = E evolves in time according to Hamiltons equations (1.14), with the
trajectory always remaining within the same surface. At each point (t) along the
trajectory, the phase-space velocity =  H is a vector tangent to the sur-
face. It is perpendicular to, and equal in magnitude to, the gradient of H ( ). The
situation is sketched schematically in Figure 1.6. We note that the vector-valued
function (vector field) v( ) =  H may be interpreted as a velocity field for an
incompressible flow in phase space whose streamlines are precisely the possible
particle trajectories. The incompressibility is expressed by the condition that the
net flux from each infinitesimal volume element vanishes, i.e.

H 2 H
v( ) = i j = i j = 0,
i j i j

3 2
2n

H
=E
)
H(

= H
1

Figure 1.6 Instantaneous phase-space velocity.


1.5 Examples 11

where we have made use of the commuting of partial derivatives and the antisym-
metry of the matrix .

1.5 Examples
1.5.1 Harmonic oscillator
One of the simplest and most important Hamiltonian systems is the one-
dimensional harmonic oscillator, with Hamiltonian (in suitable units)
1
H (q, p) = ( p 2 + q 2 ). (1.15)
2
Hamiltons equations take the form
   
d q 0 1 q p
= = ,
dt p 1 0 p q

with the well-known solution


  
q(t) cos t sin t q(0)
= .
p(t) sin t cos t p(0)

The time evolution is just a uniform clockwise rotation, with unit angular fre-
quency, in the phase space. On the circular trajectory with Hamiltonian (energy)
H = E, thephase-space velocity is tangent to the circle, with magnitude equal
to | H | = 2E. The phase portrait (plot of typical phase-space trajectories) is
shown in Figure 1.7.

1.5.2 Simple pendulum


A simple pendulum consists of a point particle of mass m confined to a friction-
less circle of radius a in a vertical plane, with the acceleration of gravity acting
downward with magnitude g. We choose as our generalized coordinate the angle
relative to the downward vertical (Figure 1.8), regarded as a variable on the real
line with values differing by an integer multiple of 2 identified. Restricting to
any open interval of length less than or equal to 2 provides a chart for the con-
figuration manifold (an infinite circular cylinder); at least two of these are required
for a complete atlas.
In units where m = a = 1, the Lagrangian takes the form
1
L(, ) = 2 + g cos .
2
12 Fundamentals of classical dynamics

Figure 1.7 The phase portrait of a harmonic oscillator.

Figure 1.8 Simple pendulum.

Solving for in terms of the canonical momentum


L
p = = ,

we construct the Hamiltonian
1 2
H (, p ) = ( p ) p L(, ( p )) = p g cos = total energy
2
1.5 Examples 13

and easily derive the Hamilton equations of motion,

= p ,
p = g sin .

Let us now represent a portion of phase space as a rectangle in the , p plane,


as in Figure 1.9. The phase-space trajectories are represented as curves
1 2
p g cos = E, (1.16)
2
parametrized by the value E of the energy in the range g E < . A sample of
such curves (phase portrait) has been drawn in Figure 1.9. We note that the values
E = g are special: these are the energy values of the stable (E = g) and
unstable (E = +g) equilibrium points of the pendulum, where the particle is at
rest at the lowest and highest points, respectively, of the circle.
For values E<g, the pendulum librates, i.e. oscillates between positive and neg-
ative extreme values, whereas for E>g, it rotates, with increasing or decreasing
without changing sign. For E precisely equal to g, there are, in addition to the equi-
librium solution, also separatrix solutions which approach (but never reach) the
unstable equilibrium for t . The equation for the separatrices is particularly
simple (Exercise 1.1):

p = 2 g cos(/2). (1.17)

On the cylindrical phase space, the separatrix trajectories divide the manifold into
disjoint regions of libration and rotation (see Figure 1.9).
To study the linear stability of the system near its equilibrium points, we can
expand the Hamiltonian in Taylor series, keeping terms up to second degree. Near
= 0, p = 0 we obtain

p

3
2
1
0
1
2
3
6 4 2 0 2 4 6

Figure 1.9 The phase portrait of a simple pendulum. Points whose coordinates
differ by 2 are identified.
14 Fundamentals of classical dynamics
1
H ( p2 + g 2 ),
2
which we recognize as the Hamiltonian of a harmonic oscillator of angular fre-

quency g. The oscillatory motion is typical of a point of stable equilibrium. The
nearly elliptic phase-space orbits are shown in Figure 1.10(a). On the other hand,
the expansion about (0, ) gives
1
H ( p2 g( )2 ),
2
the Hamiltonian of an inverted harmonic oscillator. The particle experiences
a repulsive force asymptotically proportional to the angular distance from an
unstable equilibrium point. The nearly hyperbolic phase-space orbits and linear
separatrices are shown in Figure 1.10(b).

As a final asymptotic limit, consider the pendulum for | p | g, in which case
gravity becomes negligible. Then H 12 p2 , the Hamiltonian of a particle moving
freely on the circle with uniform angular velocity p .

1.5.3 Exact solution of the pendulum equations


The simple pendulum is one of the few nontrivial systems whose initial-value
problem can be solved in closed form in terms of well-known special functions
of mathematical physics [2]. We start with the energy equation
1 2
g cos = E (1.18)
2
and make the substitution

0.10 0.10

0.05 0.05

0.00 0.00

0.05 0.05

0.10 0.10
0.10 0.05 0.00 0.05 0.10 3.05 3.10 3.15 3.20

(a) (b)

Figure 1.10 Zooms of Figure 1.9 near the stable equilibrium point (0, 0) and the
unstable equilibrium point (, 0).
1.5 Examples 15

1 E+g
z = sin , k= .
k 2 2g

Here k is a new parameter, which vanishes at stable equilibrium and assumes the
value 1 on the separatrix. Equation (1.18) then becomes

z 2 = g(1 k 2 z 2 )(1 z 2 ),

or, equivalently,
dz
dt = .
g(1 k 2 z 2 )(1 z 2 )
The last equation can be integrated to get
 z
1 d 1
t t0 = = F(sin1 z|k 2 ),
g 0 (1 )(1 k )
2 2 2 g

where F is the elliptic integral of the first kind [3]. Assuming k 2 [0, 1)
(librational motion), we can invert to obtain a solution of the initial-value problem,

z = sn( g(t t0 ), k 2 ),

where sn is a Jacobi elliptic integral [3]. Thus, if (0) = 0, p (0) = 2E, the
angle at any later time t is

(t) = 2 sin1 (k sn( g t, k 2 )). (1.19)

The right-hand side is a periodic function of t with period


4  
def 4
T = K = F , k2 . (1.20)
g g 2

1.5.4 A bead on a rotating circle


In the preceding example, we had a particle confined to a vertical loop in the pres-
ence of gravity. Now let us suppose that the same loop, attached to an electric
motor, is rotating about its vertical diameter with a time-varying angular veloc-
ity (t) (Figure 1.11). Although the particles trajectory in three dimensions now
ranges over the 2-sphere x x = 1, the systems configuration is still completely
prescribed by the polar angle in the plane of the loop, and hence has only a single
degree of freedom.
To calculate the Lagrangian, we note that the particles velocity x is a vector
lying in the tangent plane to the sphere x x = 1 at the point x. Upon introducing
16 Fundamentals of classical dynamics

(t)

motor

Figure 1.11 A bead on a rotating circle.

the orthogonal pair of unit vectors and pointing in the directions of increasing
latitude and longitude, respectively, we have

x = + sin (t).

The Lagrangian is thus


1
L(, , t) = E kin E pot = ( 2 + (t)2 sin2 ) + g cos .
2
Introducing the canonical momentum
L
p = ,

we construct the Hamiltonian
1 2 1
H (, p , t) = p g cos (t)2 sin2 .
2 2
If the rotation of our loop is uniform, then the Hamiltonian is conserved,
although the total energy E kin + E pot is not: the motor is feeding/extracting energy
at a non-uniform rate.
Hamiltons equations of motion now take the form
H
= = p ,
p
(1.21)
H 1
p = = g sin + (t)2 sin(2 ).
2
1.5 Examples 17

This is equivalent to the single second-order differential equation


.. 1
+g sin (t)2 sin(2 ) = 0.
2
To determine the possible equilibrium states of the bead, we set the right-hand
sides of (1.21) equal to zero, obtaining

(t)2
p = 0, g sin 1 cos = 0. (1.22)
g
We see that the bead has equilibrium points at = 0, , just like the simple pendu-
lum. By expanding about these equilibrium points, we can determine their stability
properties:

g + p2 /2 + 12 (g (t)2 ) 2 + 0, p 0,
H
g + p /2 2 (g + (t) )( ) + , p 0.
2 1 2 2

For (t)2 < g, we have libration about = 0, with a time-varying oscillation


frequency, while = is an unstable equilibrium point. For 2 > g, both equilib-
rium points are unstable. However, (1.22) shows that in the case of constant there
is an additional equilibrium between the other two, at = = cos1 (g/2 ).
Expanding the Hamiltonian about ( , 0),
1 2 4 g 2  
H = H ( , 0) + p + ( )2
+ O ( )3
,
2 22

we see that, for constant and larger than g, the equilibrium points at are
stable. The full story is evident in Figure 1.12, which reveals at least one feature
that could not be inferred from linear stability analysis, namely the existence of
periodic orbits that start out arbitrarily close to the saddle at (0, 0), but librate
around both of the equilibrium points .

3
2
1
0
1

2
3
6 4 2 0 2 4 6

Figure 1.12 The phase portrait for the bead on a wire loop (2 = 1.7, g = 1.0).
18 Fundamentals of classical dynamics

2
1 500
1000 1500 2000
1
2

Figure 1.13 (t) for a bead on a wire loop (2 = 1.57(1 + 12 sin t), g = 1.0).
The time interval extends from t = 0 to t = 2000, with initial values = 1 and
= 0.

There is very little that we can say in general about cases in which (t) has
nontrivial time dependence. However, we can explore a simple example numeri-
cally. In Figure 1.13, we show (t) for a case in which 2 oscillates sinusoidally,
taking values sometimes larger than g, sometimes smaller. The phase-space orbit
appears to be aperiodic, perhaps chaotic: a small difference in initial conditions
can cause two orbits to diverge from one another at an exponential rate. We leave
it to the reader (Exercise 1.9) to check whether this appears to be the case in the
present example. The existence of chaotic orbits turns out to be a generic prop-
erty of nonlinear dynamical systems with one degree of freedom and explicit time
dependence (3/2 degrees of freedom). We will return to this topic much later, in
Chapter 5.

1.5.5 Spherical pendulum


We now consider a system with two degrees of freedom, a particle confined to
a spherical surface in the presence of gravity. The radius of the sphere, a(t), is
allowed to be a prescribed function of time. We adopt the usual spherical polar
coordinates r, , , as shown in Figure 1.1. We note that only and are dynam-
ical variables, the radius being determined as a(t) at any time t. At any point,
the mutually orthogonal unit vectors r, , point in the respective directions of
increasing coordinate values. Using the sketch in Figure 1.1, we can write down
immediately the particles instantaneous velocity vector,

x = a r + a + a sin , (1.23)

so that the Lagrangian is


1 1
L = m x2 V (, t) = m(a 2 + a 2 2 + a 2 sin2 2 ) mga cos .
2 2
The momenta conjugate to the dynamical variables are
L L
p = = ma 2 , p = = ma 2 sin2 .

1.5 Examples 19

Thus, the Hamiltonian is


 
1 p2 1
H (, , p , p , t) = p2 + m a 2 + mga cos ,
2ma 2 sin
2 2
with equations of motion
p p p2 cos
= , = , p = mga sin + , p = 0.
ma 2 ma 2 sin2 ma 2 sin3
The vanishing of p allows us to consider p as a constant of the motion,
p = b, thereby reducing the problem to a single degree of freedom. Once we
have solved the equations of motion for (t), the time evolution of will follow
by simple integration:
 t
ds
(t) = (0) + b .
0 ma(s) sin (s)
2 2

For a rigid sphere (a = 0), the Hamiltonian has no explicit time dependence and
hence is conserved. In terms of and , with mass and length units chosen to make
m = 1 = a, the conservation equation takes the form
1 2 b2
+ + g cos = E. (1.24)
2 2 sin2
Upon introducing a new variable
u = cos , u = sin ,
equation (1.24) takes the simple form
def
u 2 = f (u) = 2(1 u 2 )(E gu) b2 . (1.25)
The function f (u) is a cubic polynomial with the following properties:
f (1) = b2 , f (u) 2gu 3 for u .
The requirements u 2 0 and 1 cos 1 imply that f (u) must be non-
negative in the interval [1, 1]. Figure 1.14 is a sketch of the function f (u) which
embodies all of the constraints. The existence of three real roots u 1 , u 2 , u 3 , the
first two in the interval [1, 1], is generic, with the possibility of a degeneracy
u 1 = u 2 . The motion is thus an oscillation between the endpoints min = cos1 u 2
and max = cos1 u 1 , with the degenerate case corresponding to a constant angle
0 = cos1 u 1 = cos1 u 2 . For the decreasing- part of the oscillation, the elapsed
time at angle can be determined from the differential equation (1.24):
 cos
du
t (max , ) = .
u1 f (u)
The oscillation period is then T = 2t (max , min ).
20 Fundamentals of classical dynamics

f(u)

~2gu3

1 u2 1
u
u1 u3

b2

~2gu3
Figure 1.14 A sketch of the function f (u).

The above analysis assumes the existence of a constraining force, which, acting
radially, always keeps the particle on the spherical surface. An interesting variation
on this problem is the case of a constraint force that can act only radially outward.
For example, consider a pebble set in motion on a ball of ice resting on a plane, with
the idealizations that the ball is a frictionless, impenetrable sphere and the pebble
is a point mass. Obviously the pebble must leave the ice-ball at the point where the
radial component of the constraint force Fc turns positive (see Figure 1.15).
To calculate Fc , we exploit Newtons second law, which tells us
..
Fc = m x Fgrav .

On differentiating (1.23) and using


d r
= + sin ,
dt
d
= r + cos ,
dt
d
= sin r cos ,
dt
we obtain (again using units where m = 1 = a)
..
r x = 2 sin2 2 ,

and so the positivity condition for the radial constraint becomes


..
r Fc = r x +g cos = 3g cos 2E > 0;
1.5 Examples 21

Fc
Fc
spherical pendulum
mo
tio
n
escape (Fc = 0)

pr
oje
cti
le
mo
tion
Figure 1.15 A pebble sliding on, and eventually escaping from, a stationary ball
of ice.

that is,
2E def
u = cos > = u esc .
3g
If the threshold value of u lies between the roots u 1 and u 2 , then the pebble will
leave the ice-ball at the instant when the polar angle has increased to cos1 u esc .
The positivity of this quantity ensures that the take-off will always occur in the
upper hemisphere.

1.5.6 Rigid body with one point fixed


A classical rigid body consists of a large collection of point particles with fixed
relative distances and orientations (e.g. a rigid crystalline array of 1025 atoms).
The system has six degrees of freedom: three to specify the position of a single
point O, say the center of mass, of the body; and two to specify the orientation
(with spherical coordinates , ) of a unit vector r anchored at O and fixed in the
body. With O and r fixed, the only motion remaining is a rotation about the axis r.
Fixing the angle of this rotation completes the specification of the configuration
(see Figure 1.16).
In the following discussion, we shall adopt the following conventions (see
Figure 1.16). The point O will be the origin of a fixed Cartesian (laboratory)
22 Fundamentals of classical dynamics

O y

Figure 1.16 Laboratory and body coordinate systems for a rigid object with one
point fixed.

coordinate system, with unit vectors x, y, z. The orientation of the rigid body will
be given by the angles , , described in the last paragraph. The unit vectors
r, , are fixed in the body: r points along the positive body axis, and, for any
point on that half-axis, () points in the direction of increasing ().
During an infinitesimal time interval t, the motion of a rigid body with one
point O fixed is necessarily a rotation. If the angle and axis of the latter are 
and n, respectively, then the instantaneous angular velocity of the body is defined
to be

= lim n.
t0 t

We can decompose into independent contributions from changes in , , and ,


respectively (since the rotations are infinitesimal, the vectors simply add):

= + z + r.

But z lies in the r, plane, with

z = cos r sin ,

so we get the body-fixed coordinates of ,

= sin + ( + cos )r. (1.26)

With (1.26), we are now in a position to write down the Lagrangian. The latter
is the difference between the rotational kinetic energy and the potential energy:
1.5 Examples 23
1
L = I V (, , , t). (1.27)
2
Here I is the inertia tensor in the body-fixed frame of reference. It is a constant,
symmetric, three-dimensional matrix with positive eigenvalues. Obviously, it is
advantageous to choose our body-fixed coordinate system so that the unit vectors
r, , point along mutually orthogonal principal directions of the inertia tensor.
Upon inserting (1.26), the Lagrangian then takes the form
1 2 
L= I + I sin2 2 + Ir ( + cos )2 V.
2
For the rest of our discussion, we will restrict ourselves to the case of a symmetric
top, i.e. an axially symmetric rigid body,

I = I = I ,

in a uniform gravitational field. The Lagrangian then simplifies to


1 1
L= I ( 2 + sin2 2 ) + Ir ( + cos )2 Mgl cos , (1.28)
2 2
where M is the total mass, g is the acceleration of gravity, and l is the distance
from O to the center of mass along the axis of symmetry.
Before introducing the Hamiltonian and calculating the possible motions of a
symmetric top, let us take a slight detour to review the calculation of the moments
of inertia of an axially symmetric object with uniform mass density. The reader
should be familiar with the case of a disk of infinitesimal thickness r , radius ,
and density d :

M = d 2 r
1 1
I = d 4 + d 2r 2 = M 2 + Mr 2 ,
4 4
1 1
Ir = d = M .
4 2
2 2
The most general density-d axially symmetric body can be decomposed into
infinitesimal disks centered on the r axis with respective radii (r ) (see
Figure 1.17). The total moments of inertia can be calculated by integrating over r :

M = d (r )2 dr,
 
1
I = d (r ) dr + d (r )2r 2 dr,
4
4

1
Ir = d (r )4 dr.
2
24 Fundamentals of classical dynamics

Figure 1.17 Decomposition of an axially symmetric rigid body into infinitesimal


disks.

For a sphere of radius R with O at the center, we can substitute



(r ) = R 2 r 2 , R r R,

and do the integrals to obtain


4 2
M = R 3 d, I = Ir = M R2.
3 5
Another simple case is the cylinder of radius R and height L, centered at O. Here
L L
(r ) = R, r ,
2 2
and we obtain
1 1 1
M = R 2 Ld, I =M R2 + M L 2, Ir = M R 2 .
4 12 2
The dynamical analysis of the symmetric top parallels that of the particle on a
rigid sphere. Introducing the canonical momenta conjugate to , , and ,

p = I ,
p = I sin2 + Ir cos ( + cos ),
p = Ir ( + cos ),
1.5 Examples 25

we construct the Hamiltonian


p2 ( p p cos )2 p2
H= + + + Mgl cos . (1.29)
2I 2I sin2 2Ir
We notice immediately that H has no explicit , , or t dependence, so, according
to Hamiltons equations, the momenta p and p , and the Hamiltonian itself, are
constants of the motion. Setting H = E and p = I thus gives us a first-
order differential equation for the motion. We proceed exactly as in the spherical-
pendulum example, introducing a new variable
p
u = cos , u = sin = sin ,
I
and new constants
2E p2 2Mgl p p
= , = , a= , b= .
I I Ir I I I
The differential equation for u is then
def
u 2 = f (u) = (1 u 2 )( u) (b au)2 . (1.30)

The dynamical equations can thus be summarized as


b au I a (b au)u
u = f (u), = , = . (1.31)
1 u2 Ir 1 u2
Once the first equation has been solved for u(t), the remaining ones can be solved
for (t) and (t) by direct integration of their respective right-hand sides.
As in the previous example, the qualitative features of the tops motion can be
inferred from the equations without explicitly solving them. The cubic polynomial
function f (u) once again has three real roots, u 1 , u 2 , u 3 , with the possibility of the
degeneracy u 1 = u 2 . Physically, the latter case corresponds to uniform precession
and spin rates, with constant. All other motions of the top involve nontrivial
nutation, i.e. oscillatory motion with endpoints 1 and 2 and period
 u2
du
T =2 .
u1 f (u)

From (1.31), we see that the precession velocity is periodic with the same
period T . If 1 < b/a < 1 , then changes sign (i.e. the precession reverses
direction) during each half-cycle. In the special case b/a = 1, the precession
does not reverse direction, but does come to a halt momentarily at one turning
point of the nutation cycle. The three types of precessional motion are illustrated in
Figure 1.18.
26 Fundamentals of classical dynamics

(a) (b) (c)

Figure 1.18 Polar plots of the three types of behavior for a symmetric top. All
the examples have u 1 = 0.2, u 2 = 0.8, and = 1.0. In case (a), u 3 = 1.1 and
b/a = 0.887 876 > u 2 . In case (b), u 3 = 1.4 and b/a = 0.8 = u 2 . In case (c),
u 3 = 4.0 and b/a = 0.651 775 < u 2 .

Exercises
1.1 Derive (1.17). Suppose that a pendulum, at rest at the stable equilibrium point,

is given a kick such that its initial angular velocity is 2 g. Calculate (t) and
describe the asymptotic long-time behavior of the pendulum.
1.2 Plot the simple-pendulum solution (1.19) as a function of t/T for 0 t T ,
with the exact period T given in (1.20).
1.3 Show that the pendulums oscillation period (1.20) tends to that of a harmonic

oscillator with frequency g in the limit of stable equilibrium, and that it
increases without bound as one approaches the separatrix.
1.4 Calculate the exact solution of the pendulum equations, as well as the exact
oscillation period, for E > g. Express your results in terms of elliptic func-
tions and elliptic integrals. Plot the angle as a function of t/T , where T is
the oscillation period. Check that T , as a function of the energy E, has the
expected behavior for E g and E .
1.5 The Morse potential (used to model inter-atomic forces in molecular physics)
is given by
 
V (x) = V0 e2x 2ex .

(a) Using the Mathematica function Plot, plot the Morse potential for vari-
ous values of . Use Manipulate or Animate to display the behavior
over the range of parameter values 0 < 10.
(b) Write the Hamiltonian for a particle on the line interacting with an exter-
nal Morse potential. Use ContourPlot to obtain a phase portrait, i.e. a
picture of a representative set of orbits in phase space.
Exercises 27

(c) Solve the dynamical problem analytically for x(t). (Hint: use energy con-
servation to reduce the number of variables. Elliptic functions are not
needed here.). Distinguish the three energy ranges E > 0, E = 0, and
E < 0.
(d) Solve the ODE of (c) numerically for one energy value in each of the
three ranges. Compare your results with the exact solutions.
(e) Introduce a small oscillation in the parameter :
= 0 (1 +  sin(t)) .
Explore numerically the perturbed orbits near E = 0.
1.6 A mass m is suspended from a spring of force constant k and unstretched
length l0 , and is constrained to move in a vertical plane. Write down the
Lagrangian of this stretchable pendulum and obtain the EulerLagrange equa-
tions of motion. Derive the Hamiltonian and obtain Hamiltons first-order
equations.
1.7 A particle starts at rest and moves along a cycloid whose equation is
 
ay
x = a cos1 + 2ay y 2 .
a
There is a gravitational field of strength g in the negative-y direction. Use the
Mathematica Plot function to graph this constraint equation. Choose a suit-
able generalized coordinate for this problem with one degree of freedom, and
obtain the equations of motion in first-order form. Show numerically that par-
ticles starting from rest at different points on the cycloid will all take the same
amount of time to reach the bottom. Some pointers for using Mathematica to
integrate numerically systems of ODEs may be found in Appendix A.1.
1.8 Using the Mathematica function ContourPlot, obtain phase portraits like
Figure 1.12 for various values of 2 /g. Use Manipulate or Animate to
display the behavior as the parameter ratio ranges over an interval containing
the value 1.
1.9 We want to explore the sensitivity to initial conditions of (t) for the bead on
a loop with 2 /g = (157/100)(1 + 12 sin t). Let us start with ((0), (0)) =
(1, ) and call the corresponding solution  (t). For various small values
of , integrate numerically the equations of motion for  (t) and record the
smallest time t () such that | (t) 0 (t)| exceeds 1. Look for logarithmic
growth of t () as  tends to zero. This is evidence of instability at the cho-
sen point. If this is found also for arbitrarily chosen points elsewhere along
the orbit, we would have good evidence that the orbit is chaotic. To test the
robustness of your results, you may want to increase the numerical precision
of your numerical integrations (see Appendix A.1).
28 Fundamentals of classical dynamics

1.10 A symmetric top is described by the phase-space coordinates , , , p ,


p , p and Hamiltonian (1.29). Show that the three quantities H , p , and p
are independent except on certain sub-manifolds of phase space. Describe the
physical states to which these sub-manifolds correspond.
1.11 Choosing convenient initial conditions, integrate the equations of motion
of the symmetric top and plot (t) and (t) separately for several nutation
periods, using the three sets of parameter values given in the caption of Fig-
ure 1.18. You will find the Mathematica function ListPlot helpful for
plotting lists of data points generated by your numerical integrations.
2
The Hamiltonian formalism

Having introduced the Hamiltonian formulation of classical dynamics, and applied


it within a number of familiar contexts, we now turn to a systematic study of its
salient features. Not least of these is the invariance of the dynamical equations with
respect to the broad class of canonical transformations, whose defining property is
the preservation of a certain antisymmetric, bilinear form, the Poisson bracket.

2.1 The Poisson bracket


For a dynamical system with Hamiltonian

H (q, p, t), q = (q1 , . . . , qn ), p = ( p1 , . . . , pn ),

the instantaneous state at time t, (q(t), p(t)), evolves according to Hamiltons


equations (1.11) and (1.12). These also determine the evolution of any scalar
function F(q, p, t):
d F F F F
F(q(t), p(t), t) = qk + pk + = [F, H ] + ,
dt qk pk t t
where the Poisson bracket [A, B] of arbitrary A and B is defined as
A B A B
[A, B] = .
qk pk pk qk
Here we have adopted the summation convention of summing over repeated indices
from 1 to n. In terms of the 2n-dimensional vector

= (q1 , . . . , qn , p1 , . . . , pn ),

we have
[A, B] = A  B, (2.1)

29
30 The Hamiltonian formalism

where  is given by (1.13). Where more than one coordinate system is present, we
will often write [A, B]q, p or [A, B] in place of [A, B] to avoid ambiguity.
The Poisson bracket is a central concept of the Hamiltonian formalism which
we shall work with in this book. We note that the Poisson bracket is not only a
linear function of each argument, but also antisymmetric under interchange of the
two arguments,
[A, B] = [B, A].

This ensures that the time evolution of the Hamiltonian itself is exceedingly simple,
H H
H = [H, H ] + = ,
t t
a result we obtained earlier.
A more subtle property of the Poisson bracket is the Jacobi identity

[A, [B, C]] + [B, [C, A]] + [C, [A, B]] = 0, (2.2)

which can be verified by direct calculation (Exercise 2.1). An alternative way of


writing the Jacobi identity,

[A, [B, C]] = [[A, B], C] + [B, [A, C]], (2.3)

exhibits an important property: [A, ] is an operator that distributes over [B, C] in


the same way as the differential operator distributes over an ordinary product
of functions, e.g. ( f g) = ( f )g + f g. That the Poisson-bracket operator
distributes in the same way over ordinary products of functions, i.e.

[A, BC] = [A, B]C + B[A, C],

is an easy consequence of the definition.


Of particular importance in what follows are the fundamental Poisson bracket
relations,

[qi , q j ] = 0 = [ pi , p j ] = 0, [qi , p j ] = i j , i, j = 1, 2, . . . , n,

where i j is the Kronecker delta. In terms of the matrix (1.13),

[k , l ] = kl .

2.2 Canonical transformations


Perhaps the most important advantage of the Hamiltonian formulation of classical
mechanics is the existence of a large class of coordinate transformations on a phase
manifold that preserve the form of Hamiltons equations of motion, not just for a
2.2 Canonical transformations 31

particular Hamiltonian function, but for all scalar functions that may serve in that
role. These are the canonical transformations.
There are various ways of defining canonical transformations, but for our pur-
poses the most convenient is the following. A canonical transformation is a change
of phase-space coordinates
(, t), i.e.

Q j (q1 , . . . , qn , p1 , . . . , pn , t), P j (q1 , . . . , qn , p1 , . . . , pn , t),

which preserves the fundamental Poisson brackets:

[Q i , Q j ]q, p = 0 = [Pi , P j ]q, p , [Q i , P j ]q, p = i j , i, j = 1, . . . , n; (2.4)

or, equivalently,
[
i ,
j ] =
i 
j = i j , (2.5)

with the 2n 2n matrix  given by (1.13).


Once the fundamental Poisson-bracket relations have been verified, it will follow
that arbitrary Poisson-bracket relations are preserved as well, i.e.

[F(
, t), G(
, t)]
= [F(
(, t), t), G(
(, t), t)] . (2.6)

The proof is easy. On suppressing the t dependence and summing over repeated
indices,
F
j G
m F G
[F(
( )), G(
( ))] = kl = [
j ,
m ] .

j k
m l
j
m
Upon inserting (2.5), we get
F G
[F(
( )), G(
( ))] =  jm = [F(
), G(
)]
.

j
m
A corollary of this result is the preservation of Hamiltons equations in cases
without explicit time dependence. Suppose that new phase-space coordinates are
introduced via a time-independent mapping
( ). Then
= [
, H ] = [
, K ]
= 
H  ,

where
H  (
, t) = H ((
), t).

In other words, Hamiltons equations preserve their form with the new Hamiltonian
H  . We shall see later on that time-dependent canonical transformations also pre-
serve the form of the equations of motion, but with a more complicated expression
for the new Hamiltonian.
32 The Hamiltonian formalism

2.3 Simple examples


Before proceeding, let us consider some simple examples of canonical transforma-
tions.
Identity. Q k = qk , Pk = pk . This transformation is trivially canonical.
Shift of origin. Q k = qk ak , Pk = pk bk , where ak and bk are constant
offsets. Since all derivatives of the constants vanish, the fundamental Poisson
brackets are immediately seen to be preserved, and hence the transformation
is canonical.
Positionmomentum exchange. Q k = pk , Pk = qk . We must verify the
fundamental Poisson brackets:
[Q k , Q l ] = [ pk , pl ] = 0,
[Pk , Pl ] = [qk , ql ] = [qk , ql ] = 0,
[Q k , Pl ] = [ pk , ql ] = [ql , pk ] = kl .

Linear mapping of configuration space. Q k = Mkl ql , Pk = ( M 1 )kl pl , where


M is a nonsingular real n n matrix and M is its transpose. The Poisson-
bracket relations for the new variables are
[Q k , Q l ] = Mki [qi , q j ] M jl = 0,
[Pk , Pl ] = Mki1 [ pi , p j ]M 1
jl = 0,

[Q k , Pl ] = Mki [ pi , p j ]M 1 1
jl = Mki i j M jl = kl .

Important special cases of linear mappings are orthogonal transformations,


for which the sets {q1 , . . . , qn } and { p1 , . . . , pn } are transformed by the same
matrix M = M 1 , and scale transformations
Q k = qk , Pk = 1 pk ,
where is an arbitrary numerical factor.

2.4 Canonical invariance of Lagrange brackets


Any quantities that are preserved under canonical transformations of phase space
are of great value in Hamiltonian dynamics. One such object is the Lagrange
bracket of a pair of scalar functions, which is defined as
qk pk qk pk
{u, v} p,q = =  . (2.7)
u v v u u v
For the partial derivatives to be meaningful, u and v must be considered as mem-
bers of a set of 2n independent variables. As in the case of Poisson brackets,
2.4 Canonical invariance of Lagrange brackets 33

the subscript will be suppressed unless needed for clarity. The fundamental
Lagrange-bracket relations are

{qk , ql } = 0 = { pk , pl }, {qk , pl } = kl ,

or, more concisely,


{k , l } = kl .

As in the Poisson case, the preservation of arbitrary Lagrange brackets under


canonical transformations follows easily from the preservation of the fundamen-
tal ones (we leave this as an exercise). We want to show that, for any canonical
transformation
(, t), we have

{
k ,
l } = {
k ,
l }
= kl . (2.8)

The proof is a consequence of the fact that at any point of the phase man-
ifold for which both and
coordinates are valid, the matrices (k /
l )
and (
k /l ) are essentially inverses of one another. Thus, if
(, t) is any
nonsingular coordinate transformation (not necessarily canonical), it follows that
l m
j
k l 2
k
{
i ,
j } [
j ,
k ] = lm r s =  = ik .

i
j r s
i ls s
For canonical
(, t), we can insert (2.5) to obtain (2.8).
When checking the canonicity of a particular phase-space coordinate transfor-
mation, it is quite useful to have the option of evaluating the fundamental Lagrange
brackets rather than their Poisson cousins. A prime example of this is the proof that
contact transformations (familiar from Lagrangian dynamics)

Q k = f k (q1 , . . . , qn )

can be made canonical by a suitable transformation of the conjugate momenta,


namely
f j (q)
pk = Pj .
qk
Specifically,

Q k Pk Q k Pk pi pj 2 fk 2 fk
{qi , q j } = = Pk = 0,
qi q j q j qi q j qi q j qi qi q j
Q k Pk Q k Pk f k (q) Pk f k (q) Pk
{ pi , p j } = = = 0,
pi p j p j pi pi p j p j pi

Q k Pk Q k Pk fk
{qi , p j } = = Pk = i j .
qi p j p j qi p j qi
34 The Hamiltonian formalism

v 1 (C)

S

C
k
pk
Ck

u qk

Figure 2.1 The chart assigns coordinates u and v to the surface S bounded by
the closed curve C. The map k , k {1, . . . , n}, takes C into its projection Ck in
the qk , pk subspace of phase space.

The reader will find that this derivation is far more efficient than a direct calculation
of the corresponding Poisson brackets.
We conclude our discussion of Lagrange brackets with a simple proof of the
canonical invariance of one of the so-called Poincar integral invariants. Suppose
u and v are coordinates for a two-dimensional surface S in p, q phase space with a
closed boundary curve C. Let be the chart mapping a bounded region of the u, v
plane into S (see Figure 2.1):
: (u, v)  (u, v) = (q1 (u, v), . . . , qn (u, v), p1 (u, v), . . . , pn (u, v)),
for (u, v) 1 (S).
Now consider the integral of the Lagrange bracket {u, v} over 1 (S):
  
qk pk qk pk
du dv{u, v}q, p = du dv
1 (S) 1 (S) u v v u
 
= dqk dpk = pk dqk .
k Sk k Ck

Here Ck is the boundary of the projection Sk of the surface S onto the qk , pk


plane. The invariance of the Lagrange brackets implies that the sum of pro-
jected areas, one of Poincars integral invariants, is also invariant under canonical
transformations.

2.5 Generating functions for canonical transformations


A wide class of canonical transformations Q(q, p, t), P(q, p, t) can be described
by generating functions of four types:
2.5 Generating functions for canonical transformations 35
F1 F1
type 1: pi = (q, Q, t), Pi = (q, Q, t),
qi Qi
F2 F2
type 2: pi = (q, P, t), Qi = (q, P, t),
qi Pi
F3 F3
type 3: qi = ( p, Q, t), Pi = ( p, Q, t),
pi Qi
F4 F4
type 4: qi = ( p, P, t), Qi = ( p, P, t).
pi Pi
These apply to cases in which at least one of the coordinate sets

(q, Q), (q, P), ( p, Q), ( p, P)

consists of independent variables in phase space. Such cases do not exhaust all
possibilities, but do include virtually all interesting applications. In constructing a
canonical transformation with specific criteria, it is usually advisable to first look
for a suitable generating function. When more than one of the mixed coordinate
sets is suitable in a region of phase space, one can transform among the respective
generating functions by means of Legendre transformations:

F1 (q, Q, t) = F2 (q, P, t) P Q,
F3 ( p, Q, t) = F1 (q, Q, t) p q, (2.9)
F4 ( p, P, t) = F2 (q, P, t) p q.

We verify the first of these equations and leave the remaining ones as an exer-
cise for the reader. Suppose we have a canonical transformation generated by a
type-2 generating function F2 (q, P, t). We wish to prove that F1 (q, Q, t) given
by (2.9) generates the same transformation of coordinates. Here we assume that
the coordinate transformations among the sets (q, p), (q, Q), (q, P), and (Q, P)
are nonsingular for all relevant t. On differentiating F1 (q, Q, t) with respect to qk ,
with the remaining arguments fixed, we have
F1 F2 F2 P j Pj
= + Q j.
qk qk P j qk qk
Upon inserting the derivative formulas for F2 , this reduces to
F1
= pk .
qk
Similarly,
F1 F2 P j Pj
= Q j Pk ,
Qk Pj Qk Qk
36 The Hamiltonian formalism

which reduces to
F1
= Pk .
Qk
This establishes the desired transformation equations. One consequence of the rela-
tions (2.9) is the equality of the partial time derivatives (equality holding wherever
the respective coordinate sets are interchangeable)
F1 F2 F3 F4
(q, Q, t) = (q, P, t) = ( p, Q, t) = ( p, P, t).
t t t t
The proof that the generating functions of the four types produce genuine canon-
ical transformations is a bit tedious, and we treat here only the case of type 1. It
is crucial that one always keep track of which set of independent variables one is
working with. To this end, we write partial derivatives with respect to (Q, P) using
the symbol , while writing those with respect to (q, Q) using . In this notation,
(2.9) becomes
F1 F1
pi = , Pi = . (2.10)
i
q Q i
Our strategy is to calculate the fundamental Lagrange brackets:
 
qk F1
{Q i , Q j } = (i j)
Q i Q j q k
 
qk 2 F1 ql 2 F1
= + (i j)
Q i q k q l Qj k Q j
q
qk P j qk Pi
= + .
k
Q i q k
Q j q
By adding to the right-hand member the vanishing quantity
Q k P j Q k Pi
+
Q i Q k Q j Q k
we get
Pj Pi
{Q i , Q j } = + = 0.
Qi Qj
Turning now to the other fundamental brackets, we have
 
qk F1
{Pi , P j } = (i j)
Pi P j q k
qk 2 F1 ql
= (i j) = 0,
k q
Pi q l Pj
2.5 Generating functions for canonical transformations 37

qk 2 F1 ql qk 2 F1 ql qk 2 F1
{Q i , P j } =
qk q
Qi l Pj Pj k q
q l Qi k Q i
P j q

qk Pi
= = i j .
k
P j q
Thus the fundamental Lagrange brackets are invariant and transformation gen-
erated by F1 is canonical. The cases of F2 , F3 , and F4 can be handled in similar
fashion.
Some simple examples of generating functions and the canonical transforma-
tions they produce are the following:
F2 F2
F2 (q, P) = qk Pk , pi = = Pi , Qi = = qi ,
qi Pi
F1 F1
F1 (q, Q) = qk Q k , pi = = Qi , Pi = = qi ,
qi Qi
F2 f k (q) F2
F2 (q, P) = f k (q)Pk , pi = = Pk , Qi = = f i (q).
qi qi Pi
To illustrate the usefulness of generating functions in constructing canonical
transformations, suppose we are looking for such a mapping from (q, p) to (Q, P)
satisfying P = q cot p. The function Q(q, p) is to be chosen so as to make
the transformation canonical. We note that q can be written as a function of old
and new momenta, which suggests the possibility of a type-4 generating function.
Explicitly, we write
F4 ( p, P)
q = P tan p = ,
p
and integrate directly to obtain the general solution

F4 ( p, P) = P ln cos p + G(P),

giving us the missing transformation equation


F4
Q= = ln cos p + G  (P).
P
The function G may be any sufficiently smooth function. This type of problem, as
well as our chosen solution strategy, will appear again and again in this book (for
example in the discussion of action-angle variables in Chapter 3).
An equally common application is one in which we want to use canonical
invariance to simplify the Hamiltonian, and hence the equations of motion, in a
significant way. For example, suppose we are given

H (q, p) = p +  cos q.
38 The Hamiltonian formalism

We wish to make a type-2 canonical transformation to new variables Q and P,


such that the new Hamiltonian K (Q, P) has the simple form
K (Q, P) = P.
With a generating function F2 (q, P), the transformation equations are
F2 F2
p= , Q=
q P
and the new Hamiltonian is just the substituted version of the old one, so we have
the identity (in the mixed coordinates q, P)
F2
+  cos q = P.
q
A convenient solution of this partial differential equation is
F2 (q, P) = q P  sin q.
The desired transformation is thus
p = P  cos Q, q = Q, K (Q, P) = P.
This type of calculation will be of particular relevance to our discussion of
canonical perturbation theory in Chapter 4.

2.6 Canonical invariance of phase-space volume


Suppose that a canonical transformation Q(q, p), P(q, p) is generated by one of
the four types of generating function, corresponding to a situation in which at least
one of the sets (q, Q), (q, P), ( p, Q), ( p, P) is a set of independent variables.
We shall now show, using an argument from [31], that the Jacobian determinant
(Q, P)/(q, p) is equal to unity, and hence the phase-space volume element is
preserved:
(Q, P) n n
dn Q dn P = d q d p = d n q d n p.
(q, p)
To prove this in (say) the case of intermediate coordinates (q, P) and generating
function F2 , we write the calculus identity
 
(Q, P) (Q, P) (q, p) (Q) ( p)
= = .
(q, p) (q, P) (q, P) (q) (P)
In the last expression, the independent variables are q1 , . . . , qn , P1 , . . . , Pn . Since
Qi 2 F2 pj
= = ,
q j Pi q j Pi
2.7 Covariance of Hamiltons equations with explicit time dependence 39
   
the determinants of the matrices Q i /q j and p j / Pi are equal, and so our
assertion is proved:
(Q, P)
= 1.
(q, p)
The argument is strictly analogous for the other three cases.

2.7 Covariance of Hamiltons equations with explicit time dependence


We now examine the form of Hamiltons equations with respect to the new vari-
ables in cases in which a canonical transformation is generated by a function F
with explicit time dependence. We know from Section 2.2 that, in the case of a
Hamiltonian H (, t), = (q1 , . . . , qn , p1 , . . . , pn ), and canonical transformation

(, t),

k
k H 
k
k = [
k , H ]q, p +

= [
k , H  (
, t)]
+ = kl + ,
t t
l t
where H  (
, t) = H ((
, t), t). The right-hand side does not obviously preserve
the form of Hamiltons equations. The key to restoring covariance is the identity
 

k (, t) F
=
k , . (2.11)
t t
The equations of motion then assume the form

F
=  K (
, t),

K (
, t) = H  (
, t) + . (2.12)
t

The subscript on F/t indicates that the intermediate coordinates appropriate to


the generating function have been substituted in terms of
and t. Thus the form
of Hamiltons equations is preserved, with a new Hamiltonian K (
, t).
To prove (2.11) for a type 2 generating function, we calculate the left- and right-
hand sides of the identity in terms of derivatives of F2 . Once again we use the
partial derivative notation for coordinates q, p, t and for q, P, t:

Qk Q k Q k pl 2 F2 Q k 2 F2
= = ,
t
t pl t Pk
t q
pl t l
 
F2 Q k 2 F2 P j Q k 2 F2 Q k 2 F2 P j
Qk , =

t P j pl
ql t q
pl t l P j ql
pl t

2 F2 Q k 2 F2
= ,
Pk
t q
pl t l
40 The Hamiltonian formalism

Pk Pk Pk pl Pk 2 F2
= = ,
t t pl t pl ql
t
 
F2 Pk 2 F2 P j Pk 2 F2 Pk 2 F2 P j
Pk , =

t P j pl
ql t q
pl t l P j ql
pl t

Pk 2 F2
= .
q
pl t l

The calculation for the remaining types is similar (Exercise 2.4).

2.8 Continuous one-parameter groups of canonical transformations


We generalize the term Hamiltonian flow to include any active transformation
(motion) of phase-space points, (s) = (q(s), p(s)), with s a real parameter,
such that
d(s)
= [(s), G((s), s)],
ds
where the infinitesimal generator G((s), s) is a scalar function on the phase space,
playing the role of a Hamiltonian. Obviously, if G is the actual Hamiltonian, the
differential equation is just that of time evolution. For a scalar function F(, s),
we have
dF F
= [F, G] + .
ds s
It is clear that G generates a continuous group parametrized by s, a subgroup of the
group of canonical transformations (interpreted as active transformations on phase
space).

Example: rotation about the z axis


As a simple example of a one-parameter group of transformations, consider
the three-dimensional phase space with q = (x, y, z) and p = ( px , p y , pz )
and the infinitesimal generator L z = x p y ypx , which is nothing but the z
component of angular momentum. We parametrize the transformations with a
real parameter . The reader will readily verify the differential equations of
evolution,
dq dp
= [q, L z ] = (y, x, 0), = [ p, L z ] = ( p y , px , 0), (2.13)
d d
2.8 Continuous one-parameter groups of canonical transformations 41

with the following solution to the initial-value problem:



x() cos sin 0 x(0)
y() = sin cos 0 y(0) ,
z() 0 0 1 z(0)

px () cos sin 0 px (0)
p y () = sin cos 0 p y (0) .
pz () 0 0 1 pz (0)
We see that the one-parameter group of transformations generated by L z is that
of clockwise rotations about the z axis.

Angular-momentum Poisson-bracket relations


The three functions L z = x p y ypx , L x = ypz zp y , L y = zpx x pz generate
rotations about the three coordinate axes. Their Poisson brackets can be used to
express the rotational invariance of a Hamiltonian of the form
1
H= ( px2 + p 2y + pz2 ) + V ( x 2 + y 2 + z 2 , t),
2m
namely
[H, L x ] = [H, L y ] = [H, L z ] = 0,

which simultaneously represents the conservation of angular momentum,

Lx = Ly = Lz = 0.

We note that the three angular-momentum generators, in spite of being indepen-


dent, do not commute with one another. Their Poisson-bracket relations are in
fact the Lie algebra of the three-dimensional rotation group SO(3) and its covering
group SU(2):

[L x , L y ] = L z , [L y , L z ] = L x , [L z , L x ] = L y .

In quantum mechanics we have an analogous situation. There, the three angular


momenta are self-adjoint linear operators that generate a unitary representation of
the rotation group in Hilbert space. The non-commutation of the operators means
that L x , L y , and L z cannot be simultaneously measured. In classical mechanics
the situation is a little different: in a given state, the three angular momenta do
have well-defined values; however, their non-commutation (in the sense of Poisson
brackets) prevents using two or more of them simultaneously as configuration-
space coordinates.
42 The Hamiltonian formalism

2.9 The Hamiltonian formulation of electrodynamics


The Lorentz force law of electromagnetism can be accommodated in the
Hamiltonian formalism in a simple way. Suppose that a particle of mass m and
charge e is moving under the influence of vector and scalar potentials A(x, t) and
(x, t). The Hamiltonian (with the speed of light set equal to unity)
1
H (x, p, t) = (p eA(x, t))2 + e (x, t)
2m
leads to equations of motion
1
xi = ( pi e Ai (x, t)),
m
A j (x, t)
pi = e( p j e A j (x, t)) e (x, t).
xi
Thus
.. d
m x = p e Ai (x, t) = e(E(x, t) + x B(x, t)),
dt
where

E(x, t) = (x, t) A(x, t), B(x, t) = A(x, t).
t
We now consider an example that illustrates the use of canonical transformations
in a well-known system, namely a charged particle in a uniform magnetic field
directed along the z axis. Specifically, the electromagnetic potentials are
A(x, t) = By x, (x, t) = 0.
Restricting ourselves to motions in the x, y plane, we have the Hamiltonian
1 p 2y
( px + eBy)2 +
H (x, y, px , p y ) = .
2m 2m
The equations of motion, when written in terms of the velocity v = x, take the
simple form
v = c v B,
which corresponds to motion around a circle (gyration) with angular frequency
eB
c = .
m
The radius and center of the circle (X, Y ) are constants of the motion determined
by the initial conditions.
We now take advantage of our knowledge of the solution to simplify the Hamil-
tonian formulation by means of a canonical transformation. We incorporate one
2.9 The Hamiltonian formulation of electrodynamics 43



(X,Y)

Figure 2.2 The charged particle in the x, y plane gyrating in a uniform magnetic
field perpendicular to the plane.

of the conserved quantities, Y , as a new coordinate, and choose for the other
coordinate the azimuthal angle of the particles position on its circular path
(see Figure 2.2). The momentum conjugate to the will turn out to be the con-
served quantity H/c . Thus we seek a canonical transformation (x, y, px , p y )
(, Y, P , PY ) satisfying (see Figure 2.2)
vx = c (y Y ), v y = vx cot ,
i.e.
px = Y mc , p y = mc (y Y )cot .
Since both right-hand sides are functions of the old and new coordinates, we can
write
F1 F1
px = , py = ,
x y
and integrate to get (among other choices)
1
F1 = mc (y Y )2 cot mc Y x.
2
Completing the canonical transformation generated by F1 , we have
F1 mc (y Y )2 mc (vx2 + v 2y )
P = = =
2 sin2 2c
mc 2 H
= = ,
2 c
F1
PY = = mc (y Y ) cot + x.
Y
44 The Hamiltonian formalism

Since the new Hamiltonian,


K = c P ,
has no dependence on either Y or PY , both of these quantities must be constants of
the motion. From the geometrical relations
x = X + cos , y = Y sin ,
we recognize PY as the conserved coordinate X of the center of the circle.
In this problem, the canonical transformation has achieved the following: of
the four new phase-space coordinates, three (Y, P , and PY ) are constants of the
motion, while the fourth is an angular coordinate that increases at the uniform rate
c . In a later chapter we will return to this system, introducing an electrostatic
wave as a perturbation. Having such a convenient set of unperturbed phase-space
coordinates will turn out to be a major advantage for dealing with the perturbed
system.

2.10 Normal modes for linear oscillators


The normal-mode problem for small oscillations about stable equilibrium is a
familiar topic of Lagrangian dynamics. We now show how the problem can be
solved systematically in the canonical formalism.
Given a dynamical system with a known equilibrium point P in phase space, the
question of linear stability at P can be settled by a simple, straightforward proce-
dure. One first expands the Hamiltonian about P in a Taylor series, keeping only
the terms of degree 2 in the dynamical variables (the constant term is irrelevant,
and there are no linear terms, thanks to the equilibrium condition). The quadratic
Hamiltonian can always be written as
1
H ( ) = C , = (q1 , . . . , qn , p1 , . . . , pn ), (2.14)
2
where C is a real 2n 2n symmetric matrix. The linear equations of motion for
small oscillations then take the form
= C , C =  C, (2.15)
where  is the antisymmetric matrix (1.13).
For stable motion in the neighborhood of an equilibrium point, we seek solutions
of the form
(t) = (0)eit ,
so that
= i.
2.10 Normal modes for linear oscillators 45

From (2.15), we see that the desired must be an eigenvector of C, with eigenvalue
= i satisfying the equation

det(C 12n ) = 0. (2.16)

The reality of C implies that the purely imaginary eigenvalues come in complex-
conjugate pairs, j = n+ j , hence

n+ j = j , j = 1, . . . , n.

If we are interested only in the question of linear stability, the construction


of C and the determination of the signs of its eigenvalues is all that we need
be concerned about. On occasion, however, we are also interested in solving the
initial-value problem for small oscillations in terms of so-called normal modes.
With this in mind, we now derive some generalized orthogonality and normaliza-
tion relations for the eigenvectors 1 , . . . , 2n . Using the matrix relation  T =
12n , we have the two identities

j C k = j T  C k = j T j k ,

j C k = j C T  k = j j k ,

and hence
( j + k ) j  k = 0.

We conclude that j C k must vanish except for the complex-conjugate pairs.


In those cases, the normalization is arbitrary,

j C n+ j = N j .

Without loss of generality, we assume N j = 1 for all j. The normal-mode


expansion of an arbitrary solution of the equations of motion then takes the form
n 

(t) = ( j C (0)) j ei j t + ( j C (0)) j ei j t , (2.17)
j=1

def
where j = n+ j .
To make full use of the normal-mode formalism, it is sometimes useful to
introduce new canonical variables:
1 1
a j = j C = ( j q j i p j ),
j 2 j
(2.18)
i 1
a j = j C = ( p i j q j ).
j 2 j
46 The Hamiltonian formalism

The Poisson-bracket relations


[a j , ak ] = 0 = [a j , ak ], [a j , ak ] = jk
are readily verified.
In terms of the new canonical variables, the Hamiltonian and equations of motion
are particularly simple. We expand in terms of the normal-mode basis,

= j (a j j + ia j j ),
j

and substitute into (2.14) to obtain



n

n
H= i j k a j ak ( j C k ) = ik ak ak . (2.19)
j,k=1 k=1

Hamiltons equations for ak and ak are then


ak = [ak , H ] = k ak = ik ak , ak = [ak , H ] = ik ak .
For a one-dimensional simple harmonic oscillator, we have, from (2.18),
1 1
H = ( p 2 + 2 q 2 ) = ( p + iq)( p iq) = ia a,
2 2
and
1 1
a = (q i p), a = ( p iq),
2 2
so the normal-mode Hamiltonian (2.19) is just that of n independent harmonic
oscillators.
One is struck by the similarity of the present formulation of linear oscilla-
tions and the extremely useful quantum-mechanical one in terms of creation and
annihilation operators.

Small oscillations of a collinear triatomic molecule


As a straightforward application of the normal-mode technique, consider three
point masses (m 1 , m 2 , m 3 ), constrained to lie on the x axis and connected by mass-
less springs (unstretched lengths a1 and a2 , force constants k1 and k2 ), as shown in
Figure 2.3.

m1 k1 m2 k2 m3

Figure 2.3 A collinear triatomic molecule.


2.10 Normal modes for linear oscillators 47

In the center-of-mass frame, the particle coordinates x1 , x2 , x3 are constrained by


mi
1 x1 + 2 x2 + 3 x3 = 0, i = , m = m1 + m2 + m3,
m
so the system has two degrees of freedom, q1 and q2 , defined by
x 2 = x 1 + a1 + q 1 , x 3 = x 2 + a2 + q 2 .
On differentiating the above equations and solving for the three velocities, we
obtain
x1 = (2 +3 )q1 3 q2 , x2 = 1 q1 3 q2 , x3 = 1 q1 +(1 +2 )q2 .
With mass units such that m = 1, we can now write the Lagrangian in matrix
notation,
1 1
L(q1 , q2 , q1 , q2 ) = q T q q V q,
2 2
where
 
1 (1 1 ) 1 3 k1 0
T = , V = .
1 3 3 (1 3 ) 0 k2
Following the general prescription (1.10), we next introduce conjugate momenta
L
pi = = Ti j q j , i = 1, 2,
qi
and define the Hamiltonian
1 1
H (q, p) = p T 1 p + q V q.
2m 2
In terms of = (q1 , q2 , p1 , p2 ), we can write the Hamiltonian as

k1 0 0 0
1 0 k2 0 0
H ( ) = C, C =
.

2 0 0 (1 3 )/(
1 2 ) 1/ 2
0 0 1/2 (1 1 )/(2 3 ).
According to our general procedure for constructing normal modes, we need
to diagonalize the matrix C =  C to obtain purely imaginary eigenvalues
1 , 2 , 1 , 2 and the corresponding normalized eigenvectors 1 , 2 , 1 , 2 .
This is a straightforward exercise, which we relegate to Exercise 2.6. One finds that
the eigenvalue equation takes the form of a quadratic equation for the square of the
eigenvalue , namely

1 3 1 1 k1 k2
( ) + k1
2 2
+ k2 2 + = 0. (2.20)
1 2 2 3 1 2 3
48 The Hamiltonian formalism

Having found the j and j , we can easily solve any initial-value problem for
the linear molecule (for example, Exercise 2.7) using (2.17).

Lagrange points of planetary orbits


An important simplified model of celestial mechanics is the restricted three-body
problem with circular orbits. Here one idealizes the Solar System as the Sun and
Jupiter executing coplanar circular orbits about their common center of mass, with
a frequency given by Keplers third law. One then tries to solve the dynamical
problem of an asteroid of negligible mass moving in the same plane under the
gravitational attraction of the massive orbiting bodies.
In 1772, Lagrange [4] discovered that, in the rotating frame of reference in which
the Sun and Jupiter are at rest, there exist a few equilibrium points, which are now
called the Lagrange points L 1 , L 2 , L 3 , L 4 , and L 5 . We now want to study, as a
small-oscillation problem, the linear stability of the L 4 and L 5 points, where the
Sun, Jupiter, and the asteroid lie at the vertices of an equilateral triangle (L 4 being
ahead of Jupiter and L 5 behind it in its orbit), as shown in Figure 2.4. It should be
noted that in 1906 an actual asteroid, Achilles 588, that maintains the L 4 geomet-
rical relation with the Sun and Jupiter was found. As of May, 2011, the numbers of
such observed L 4 and L 5 asteroids were listed as 3116 and 1616, respectively1 .

asteroid
L4

Jupiter
mJ
rJ

o
60
rS
Sun 60o
mS

L5
asteroid

Figure 2.4 The configuration of the Sun, Jupiter, and Trojan asteroids.

1 See http://cfa-www.harvard.edu/iau/lists/JupiterTrojans.html.
2.10 Normal modes for linear oscillators 49

We begin by describing the planar motions of the Sun and Jupiter in an inertial
frame with their center of mass at the origin of a rectangular x, y coordinate system.
If rS and rJ are the respective radii of the circular orbits, then we can take their
respective positions at time t to be
  
xJ cos(t) sin(t) rJ
= ,
yJ sin(t) cos(t) 0
  
xS cos(t) sin(t) rS
= .
yS sin(t) cos(t) 0
The radii are inversely proportional to the masses, so that
rJ = (1 )a, rS = a, m = m J + m S, a = rJ + rS ,
where
mJ
=
m
and the angular frequency is (according to Keplers third law)

Gm
= ,
a3
where G is Newtons gravitational constant.
According to Newtons law of gravitation, the Hamiltonian (i.e. total energy) for
an asteroid of mass m A moving under the influence of the Sun and Jupiter in their
circular orbits is

1 2 1
H = ( px + p y )
2 2
+ ,
2 rAJ rAS
where rAJ and rAS are the asteroidJupiter and asteroidSun distances, and we have
chosen units of mass and length such that m A = 1 and a = 1.
Since the Lagrange points are supposed to be stationary in a non-inertial frame
for which the positions of the Sun and Jupiter remain fixed at (rS , 0) and
(0, rJ ), respectively, we now make a canonical transformation to new phase-space
coordinates X, Y, PX , PY via
  
x cos(t) sin(t) X
= ,
y sin(t) cos(t) Y
  
px cos(t) sin(t) PX
= .
py sin(t) cos(t) PY
The generating function for this transformation is
F2 (x, y, PX , PY , t) = P R(t) x,
50 The Hamiltonian formalism

where P = (PX , PY ), x = (x, y), and R() is the angle- rotation matrix in the
x, y plane. Since F2 is explicitly time-dependent, there is a nontrivial contribution
to the transformed Hamiltonian K (X, Y, PX , PY ), namely
  
F2 sin(t) cos(t) cos(t) sin(t) X
= (PX , PY )
t cos(t) sin(t) sin(t) cos(t) Y
= (PX Y X PY ),

and hence

1 2 1
K (X, Y, PX , PY ) = (PX + PY )
2 2
+ + (PX Y X PY ).
2 rAJ rAS
Hamiltons equations of motion take the form
K K
X = = PX + Y, Y = = PY X,
PX PY

K 2 1
PX = = PY + + , (2.21)
X X rAJ rAS

K mJ mS
PY = = PX + 2 + ,
Y Y rAJ rAS
with

rAJ = (X (1 ))2 + Y 2 , rAS = (X + )2 + Y 2 .

The equilibrium points in the co-rotating frame are obtained by setting the right-
hand sides of the equations (2.21) equal to zero. Intuitively, these points are the
ones where the gravitational pulls of the Sun and Jupiter are exactly balanced by the
fictitious centrifugal and Coriolis forces. To see this explicitly, we can calculate,
using Hamiltons equations, the effective force

Feff = ( X , Y ) = 2 X Veff ,

where the effective potential energy, incorporating both centrifugal and gravita-
tional terms, takes the form
 
2 2 1
Veff (X, Y ) = (X +Y 2 )2 + .
2 (X + )2 + Y 2 (X + 1)2 + Y 2
To find the equilibria, we can ignore the velocity-dependent Coriolis terms and
look for points where the gradient of Veff vanishes. The contour plots in Figure 2.5
suggest that three of these points lie on the X axis, i.e. are collinear with the Sun
and Jupiter. These are saddle points located between the Sun and Jupiter (L 1 ), to the
right of Jupiter (L 2 ), and to the left of the Sun (L 3 ) on the X axis. Upon inserting
2.10 Normal modes for linear oscillators 51

1.5 3

1.0 2

0.5 1

0.0 0

0.5 1

1.0 2

1.5 3

1.5 1.0 0.5 0.0 0.5 1.0 1.5 3 2 1 0 1 2 3

Figure 2.5 Contour plots of the effective potential energy Veff (X, Y ) for
m J /m S = 0.03 (left) and m J /m S = 1 (right). The positions of the L 4 and L 5
equilibrium points are shown as points. The length scale is such that rJ = 1.

1 0.5 0.5 1.5

Figure 2.6 The function f (X ) whose zeros determine the positions of the
Lagrange points (from left to right) L 3 , L 2 , and L 1 . Here = 0.1, but the
behavior is typical of the entire interval 0 < < 1/2.

Y = 0, the equilibrium equations reduce to the following piecewise polynomial


equation for X :

f (X ) = 2 X (X + )2 (X + 1)2 + (1 )sign(X + )(X + 1)2


+ sign(X + 1)(X + )2 .

A typical plot of f (X ) is shown in Figure 2.6. The following inequalities


ensure that f (X ) has a zero in each of the three open intervals, namely
(, ), (, 1 ), and (1 , +):
52 The Hamiltonian formalism

f () f ( ) < 0,
f ( + ) f (1 ) < 0,
f (1 + ) f (+) < 0,
where  is an arbitrarily small positive number.
The remaining Lagrange points appear as local maxima in the contour plots.
They also satisfy a simple geometrical criterion (see [2]): they are equidistant from
the Sun and Jupiter, i.e. rAJ = rAS , hence
1
X +=1 X = .
2
With this condition, the equilibrium conditions simplify to
 32
1
1 +Y 2
= 0,
4
with the solution

3
Y = .
2

One readily verifies that the points (, 0), (1 , 0), and (1/2 , 3/2) are
the vertices of an equilateral triangle. For  1, as is the case for the actual Solar
System, the L 4 and L 5 points lie just outside the orbit of Jupiter.
Although the velocity of the asteroid vanishes at any of the Lagrange points, the
canonical momentum does not. Specifically, one has (PX , PY ) = (Y, X ) at
such points.
We now turn to the question of the linear stability of the Lagrange points. It is
important to emphasize here that the properties of the effective potential energy
Veff , while sufficient to pick out the stationary points, are an unreliable indica-
tor when it comes to stability. The reason, of course, is that the Coriolis force,
despite being negligible for bodies at rest, can play an important role when it
comes to small oscillations about equilibrium. A local maximum of Veff , located
on a ridge separating the gravitational and centrifugal downslopes, might seem an
unlikely place to find stability, but in fact, when one considers the full dynamics,
one finds that it is exactly here that stable oscillations are possible, provided that
the Jupiter/Sun mass ratio is sufficiently small.
Let us concentrate on the L 4 Lagrange point, leaving the other cases for an
exercise at the end of the chapter (Exercise 2.9). We substitute (see Figure 2.7)

1 3
X = + q1 , Y = + q2 ,
2 2

3 1
PX = + p1 , PY = + p2
2 2
2.10 Normal modes for linear oscillators 53

(X + q1, Y + q2)
Y (X, Y)

X
O 1

Figure 2.7 Relative coordinates for theasteroid in the neighborhood of the L 4


Lagrange point, (X 4 , Y4 ) = (1/2 , 3/2).

into the Hamiltonian K (X, Y, PX , PY ), making a quadratic Taylor expansion in


q1 , q2 , p1 , p2 about q1 = q2 = p1 = p2 = 0. Leaving the details as an exercise
(Exercise 2.10), we state the result (omitting an irrelevant constant term):

(2) 1 2 2 1 2 5 2
h (q, p) = ( p1 + p2 ) + (q2 p1 q1 p2 ) +
2
q kq1 q2 q2 , (2.22)
2 8 1 8
where

27
k= (1 2).
16
We can now apply our normal-mode analysis of the small-oscillation problem
for Hamiltonian h (2) (q, p). The resulting eigenvalue equation for the normal-mode
frequencies j = j /i, j = 1, 2, 3, 4, is
27 27
4 + 2 2 + d4 = 0, d= k 2 = (1 ), (2.23)
16 4
with the solution
2
2 = (1 1 4d).
2
For 2 to be real and negative, so that all j are real and the system is stable (i.e.
oscillatory) in the neighborhood of equilibrium, we must have 4d < 1; hence
1
(1 ) < .
27
54 The Hamiltonian formalism
(a) (b)

q2
1.0 0.6 q2
0.4
0.5
0.2
q1 q1
1.5 1.0 0.5 0.5 1.0 1.5 1.0 0.5 0.5 1.0
0.2
0.5
0.4
1.0
0.6

Figure 2.8 The L 4 normal modes for a mass ratio m J /m S of 0.000 954 779,
approximating that of the Trojan asteroids in the actual Solar System: (a) oscil-
lation period 1.003 25 Jovian years; and (b) oscillation period 12.428 Jovian
years.

The bound corresponds to a mass ratio /(1 ) = m J /m S less than 0.040 064 2.
The condition is clearly satisfied by the actual SunJupiter and EarthMoon sys-
tems, for which the measured values of the mass ratio are 0.000 954 779 and
0.012 345 7, respectively. For the SunJupiter case, the periods of the small oscilla-
tions are approximately 1.003 25 and 12.428 Jovian years. Their respective orbits
are shown in Figure 2.8. We note the highly eccentric orbit of the longer-period
mode. As the mass ratio tends to zero, the period of this mode tends to infinity, and
the short/long-axis ratio of the orbit tends to zero. Meanwhile, the shorter-period
mode tends toward synchrony with the orbital motion of the system in the inertial
frame. On the other hand, as the mass ratio tends to the critical value for which
(1 )= 1/27, the orbits of the two modes tend to a common limit, with a
period of 2 Jovian years.
For additional discussion of the Lagrange points and their stability properties,
the reader is referred to [2]. We will return to this system in Chapter 5 for a
discussion of the nonlinear stability of L 4 and L 5 .

Exercises
2.1 Prove the Jacobi identity for the Poisson bracket by direct calculation starting
from the definition. Computer assistance is recommended but not necessary.
2.2 For one degree of freedom, a canonical transformation Q(q, p), P(q, p) has
P = ( p 5)/cos q. Find the most general Q(q, p).
2.3 For one degree of freedom, a canonical transformation Q(q, p), P(q, p) has
Q = q 1/2 et cos p. Find the most general P(q, p).
Exercises 55

2.4 Prove (2.11) for type-1 generating functions. Be sure to indicate explicitly
which set of variables is being used for each of the partial derivatives in your
equations.
2.5 Prove that L 2x + L 2y + L 2z commutes, in the sense of Poisson brackets, with
each of the quantities L x , L y , and L z .
2.6 Obtain (2.20). Show that all four solutions for 2 are negative real quanti-
ties, so that the i are purely imaginary. Calculate the eigenvectors for the
respective eigenvalues. Verify the orthonormality relations for your results.
2.7 Consider a linear triatomic molecule with m 1 = m 3 = 12 m 2 , a1 = a2 , and
k1 = k2 . Simplify the results of Exercise 2.6. Suppose the molecule is initially
at rest with its center of mass at x = 0. At t = 0 an initial velocity v0 is
imparted to particle 1 (via an impulsive force). Calculate xi (t), i = 1, 2, 3.
2.8 Generalize the normal-mode discussion to n degrees of freedom and possibly
degenerate frequencies.
2.9 Investigate the linear stability of the Lagrange points L 1 , L 2 , and L 3 .
2.10 Derive (2.22) and (2.23). You will probably find computer assistance helpful
in dealing with the algebra.
2.11 A linear dynamical system with two degrees of freedom has the Hamiltonian
1
H (q1 , q2 , p1 , p2 ) = ( p12 + p22 + 12 q12 + 22 q22 ) + p1 q2 .
2
For what values of is the equilibrium point at the origin stable?
2.12 A particle A is made to move in a circular orbit of radius R about a fixed ori-
gin with angular frequency . A second particle B, of unit mass, constrained
to move in the plane of As orbit, is attracted to A with a potential-energy
function V (r ) = 12 kr 2 , where r is the interparticle distance.
(a) Write down the Lagrangian of the system in terms of rectangular coor-
dinates x, y and their time derivatives. Calculate the canonical momenta
px , p y and the Hamiltonian function.
(b) Perform a time-dependent canonical transformation, via a generating
function of type 2, to a co-rotating frame of reference, in which particle
A is at rest on one of the coordinate axes.
(c) Calculate the Hamiltonian for Bs motion in the co-rotating frame.
(d) Find all equilibrium points in the co-rotating frame.
(e) Are there parameter values for which any of the equilibrium points
is stable? If so, calculate the frequencies of small oscillations about
equilibrium.
3
Integrable systems

A Hamiltonian system with 2n-dimensional phase space is called integrable if


there exist n independent, mutually commuting, smooth functions F1 , . . . , Fn and
a Hamiltonian H = H (F1 , . . . , Fn ). Owing to the Poisson bracket relations
d Fk
[Fk , H ] = 0 = , k = 1, . . . , n,
dt
the Fk are integrals, i.e. constants of the motion along phase-space orbits. Without
loss of generality, we may assume that one of these, F1 , is H itself.
Among all possible Hamiltonian systems, the integrable ones are exceedingly
rare. Nonetheless they occupy a privileged place in classical dynamics. One reason
is that essentially all systems that are explicitly soluble using the methods of tra-
ditional mathematical physics [2] are integrable. Most of the examples which we
have encountered in this book so far have been in this category. Moreover, typical
attempts to understand non-integrable models involve using a form of perturbation
theory in which the unperturbed system is an integrable one. For these reasons, we
give considerable attention to the construction and analysis of integrable systems
in this book.

3.1 The LiouvilleArnold theorem


What is special about integrable systems? A remarkably detailed answer to this
question is provided by the LiouvilleArnold theorem [6, 7], which states that
a system with independent commuting integrals Fk , k = 1, . . . , n satisfies the
following:
1. The phase space is partitioned into level sets
M f = { : Fk ( ) = f k , k = 1, . . . , n},
each of which is invariant under the Hamiltonian flows generated by the n
functions Fk . This includes the time evolution generated by H = F1 . For

56
3.2 Fast track for separable systems 57

simplicity, we assume that M f is connected; otherwise we consider separately


each connected component of M f .
2. For given f = ( f 1 , . . . , f n ), M f is diffeomorphic to T m Rnm , where T m
is a torus of dimension m, m {0, . . . , n}, and Rnm is a real vector space of
dimension n m.
3. Assume m = n, so that M f is diffeomorphic to an n-torus. In a neighborhood
of M f , canonical coordinates (action-angle variables) 1 , . . . , n , J1 , . . . , Jn ,
may be introduced, such that for each k, k = 1, . . . , n, Jk is a function of
F1 , . . . , Fn , and, moreover, Jk and k evolve in time as
H
Jk (t) = Jk (0), k (t) = k (J )t + k (0), k (J ) = .
Jk
4. Given an initial point (0) on M f , the time-evolved point (t) is a multiperiodic
function of the angles 1 , . . . , n which can be obtained by quadratures, i.e.
solving transformation equations and integrating known functions.

We see that in a neighborhood of a compact, connected level set, the inner work-
ings of an integrable system are revealed by the LiouvilleArnold theorem to be
exceedingly simple. When viewed in a suitable canonical frame of reference, they
consist of n points, each moving uniformly around its own circle. This is sometimes
described as a decomposition into n independent one-dimensional harmonic oscil-
lators, but the analogy is not completely accurate: the respective frequencies of the
constituent oscillators can depend nontrivially on the action coordinates J1 , . . . , Jn ,
whereas those of a simple harmonic oscillator are action-independent.
A convenient feature of Arnolds proof of the theorem is that it is construc-
tive. It gives a detailed prescription for transforming to action-angle coordinates
and integrating the original differential equations of motion by quadratures. The
process is technically somewhat demanding, but worth the effort if less difficult
routes to a solution are unavailable. Fortunately, a simpler approach does exist for
a large class of integrable systems (including all the familiar textbook examples),
namely the separable ones for which the decomposition of each level set into n
independent topological circles is obvious from the start.

3.2 Fast track for separable systems


An integrable system with canonical coordinates (q1 , . . . , qn , p1 , . . . , pn ) and
independent commuting integrals F1 , . . . , Fn is said to be separable on the level
sets M f , f = ( f 1 , . . . , f n ), if each of the equations Fk (q, p) = f k has a solution

pk = pk (qk , f 1 , . . . , f n ),
58 Integrable systems

such that the graph Ck of this function in the qk , pk plane is diffeomorphic to a cir-
cle. Then each point of M f is specified by a unique set of points on the respective
Ck , and so the manifold has the structure of a direct product of n circles, and hence
is diffeomorphic to an n-torus.
One successful solution strategy for separable systems consists of establishing
the existence of a canonical transformation to phase-space coordinates Q, P for
which the new momenta are precisely the integrals, i.e. Pk = Fk , k = 1, . . . , n. In
that case, the new Hamiltonian will be simply P1 and the equations of motion will
reduce to

Pk = 0, k = 1, . . . , n, Q 1 = 1, Q k = 0, k = 2, . . . , n.

The quest for such a canonical transformation falls into a pattern that the reader
will certainly recognize from Chapter 2. The equations

pk = pk (qk , P1 , . . . , Pn )

already supply half the equations needed to specify the coordinate change. To
obtain the rest, we introduce a type-2 generating function,

n
W (q, P) = Wk (qk , P), (3.1)
k=1

by integrating the partial differential equations


Wk
= pk (qk , P1 , . . . , Pn )
qk
to get
 qk
Wk (qk , P) = pk (qk , P)dqk .
q0 (P)

Choosing any convenient initial point q0 (P), p(q0 , P) on the closed curve Ck in
the qk , pk plane, we complete the transformation equations by differentiating the
generating function,
Wk
Qk = , k = 1, . . . , n.
Pk
At this point the dynamical problem is essentially solved. One needs only to
solve the following equations for (q(t), p(t)) in terms of (q(0), p(0)):

Pk (q(t), p(t)) = Pk (q(0), p(0)), k = 1, . . . , n,


Q k (q(t), p(t)) = Q k (q(0), p(0)), k = 2, . . . , n, (3.2)
Q 1 (q(t), p(t)) = Q 1 (q(0), p(0)) + t.
3.2 Fast track for separable systems 59

What we have just derived lies at the heart of the HamiltonJacobi theory
applied to separable Hamiltonian systems (see, for example, Chapter 10 of [1]).
An alternative approach action-angle variables exploits more fully the assumed
torus structure. Here the angle variables will parametrize the independent circles
on the n-torus, while the conjugate action variables denote, up to a factor, the
respective areas of the circles. Specifically, in a neighborhood of M f we define

1 A(Ck )
Jk (F) = pk (qk , F)dqk = ,
2 Ck 2
where the circuit integral is taken around the closed curve Ck in the qk , pk plane and
A(Ck ) is the area enclosed by Ck . Since the Fk are mutually commuting conserved
quantities, so are the Jk . On solving for F(J ) and substituting into the defining
equations for Ck , we get a set of n transformation equations of the form

pk = pk (qk , F(J )).

Once again we can complete the canonical transformation by introducing a type-2


generating function,

n
S(q, J ) = Sk (qk , J ), (3.3)
k=1

by integrating the partial differential equations


Sk
= pk (qk , F(J ))
qk
to get
 qk
Sk (qk , J ) = pk (qk , F(J ))dqk .
q0 (J )

The angle variables k are then obtained by partial differentiation,


S
k = .
Jk
To verify that the k are true angles, we must show 2 periodicity on Ck . This
follows from the fact that Sk , by definition, has periodicity 2 Jk :
   
k 2 Sk Sk
dqk = dqk = dqk = pk dqk = 2.
Ck qk Ck Jk qk Jk Ck qk Jk Ck
The interchange of differentiation and integration can be justified quite generally
(see [7]).
60 Integrable systems

3.3 System with one degree of freedom


For a Hamiltonian system with one degree of freedom, the separability criterion
of the previous section reduces to the assumptions that (i) the Hamiltonian is con-
served and (ii) every curve C of constant energy H (q, p) = E is diffeomorphic to a
circle. For any such system, solution of the dynamical equations using the methods
of HamiltonJacobi and action-angle variables is completely straightforward.
The one-dimensional simple harmonic oscillator provides a simple but important
example. Choosing units for which the mass is unity, the Hamiltonian is
1
H = ( p 2 + 2 q 2 ).
2
The level sets with non-zero energy are concentric ellipses C(E) with equation
H (q, p) = E,
or, alternatively,

p = 2E 2 q 2 . (3.4)
To solve the dynamical problem in the manner of Hamilton and Jacobi, we want to
make a canonical transformation to phase-space variables (Q, E). The generating
function (3.1) takes the form
 q
W (q, E) = dq 2E 2 q 2 ,
a

and so, with the choice a = 2E/,
 q 
W 1 q dq  1
Q= = = cos1 . (3.5)
E a a2 q 2 a
As expected, the equation of motion for Q is nearly trivial,
H
Q = = 1,
E
so that
Q(t) = (t t0 ),
where t0 is a constant. Inverting (3.5) and applying (3.4), we get the familiar
harmonic-oscillator time evolution,
q = A cos[(t t0 )], p = A sin[(t t0 )].
To solve the same problem using action-angle variables, we define J as the area
enclosed by the ellipse C(E) divided by 2 :
A2 E
J= = ,
2
3.4 The Kepler problem in spherical polar coordinates 61

and write down the action-angle generating function


 
2J
S(q, J ) = W (q, E(J )) = dq q 2.

The angle coordinate is then given by

S 1 q
= = cos .
J 2J/
Upon inverting this we obtain

2J
q= cos , p = 2J sin .

To complete the dynamical solution, we note that the action-angle Hamiltonian is
K = E(J ) = J,
so the motion is uniform:
K
= = .
J
Hence, if (0) = 0, we once again get
(t) = (t t0 ).

3.4 The Kepler problem in spherical polar coordinates


A familiar example of a separable system with more than one degree of freedom
is that of a particle moving in a spherically symmetric potential. Using spherical
polar coordinates r, , and , the Hamiltonian takes the form
 
2 2
1 p
p
H= pr2 + 2 + + V (r ).
2m r r 2 sin2
To verify separability, our first task is to find three independent, mutually com-
muting scalar functions Fi (r, , , pr , p , p ), i = 1, 2, 3. As usual, we take
F1 = H . Good candidates for F2 and F3 become obvious if we write H in a
nested format:
  
1 1 p2
F1 = H = pr + 2 p + 2
2 2
+ V (r ), (3.6)
2m r sin
namely

p2
F2 = p2 + , F3 = p . (3.7)
sin2
62 Integrable systems

One readily verifies the independence and vanishing Poisson brackets of the three
Fk , as well as the phase-space orbit projections in the respective coordinate planes

F 2
F2
pr = 2m F1 22 V (r ), p = F22 32 , p = F3 . (3.8)
r sin
This completes the verification of the separability conditions. We next turn to
the solution of the dynamical problem, restricting ourselves to the Newtonian
gravitational potential

V (r ) = , > 0.
r
We insert (3.8) in (3.1) to obtain the type-2 HamiltonJacobi generating function

W (r, , , F1 , F2 , F3 ) = Wr (r, F) + W (, F) + W (, F),

where

W (, F) = P ,


F2
W (, F) = d  sign( p ) F22 23 ,
sin 
 r 
 F22
Wr (r, F) = dr sign( pr ) 2m F1 + .
2mr 2 r 
The integrations are along the closed curves defined by (3.8). The corresponding
canonical transformation

(r, , , pr , p , p ) (Q r , Q , Q , Pr , P , P )

is thus given by

Pr = F1 (r, , , pr , p , p ),
P = F2 (, p , p ),
P = F3 ( p ) = p ,

W W W
Qr = , Q = , Q =
F1 F2 F3
and the new Hamiltonian is
K (Q, P) = Pr .

Thus Pr , P , P , Q , Q are all constants of the motion, while

Q r (t) = Q r (0) + t.
3.4 The Kepler problem in spherical polar coordinates 63

Now let us interpret these equations geometrically, setting Pr equal to the con-
stant total energy E. The Q and Q equations determine the particles orbit. We
simplify by choosing
= 0, p = 0 = P ,
so that = 0 and the orbit (or at least part of it) is restricted to the half-plane
= 0. We further choose = 0 along the direction of perihelion, where r is a
minimum and pr = 0. This fixes the signs in the generating functions and the Q
and Q r equations assume the form
 
1 dr P
Q = + 2
 , (3.9)
2m r E P2 /(2mr 2 ) + /r
 
m dr
Qr =  . (3.10)
2 E P2 /(2mr 2 ) + /r

The integral in the Q equation can be evaluated using the substitution r = 1/u to
give
1 m
= 2 (1 + e cos ( 0 )), 0 = Q (0),
r P
where

2E P2
e= 1+ .
m 2
The orbit is the = 0 part of an ellipse with eccentricity e, perihelion angle 0 and
semi-major axis a = /(2|E|).
The Q r equation now gives the time dependence of the radial motion. The
integral may be calculated with the substitution
r = a(1 e cos ),
to give
  
ma 3 ma 3
t t0 = (1 e cos )d = ( e sin ),

where t0 = Q r (0) is the time of minimum r , i.e. perihelion, with = 0, r =
a(1 e). This is Keplers equation, which, for example, gives the time difference
between any two points on a planets orbit. For = 2, this gives Keplers third
law for the period,

ma 3
T = 2 .

64 Integrable systems

3.5 Proof of the LiouvilleArnold theorem


We now turn to a proof of the LiouvilleArnold theorem, adapted from [7]. The
first step will be to use the n integrals Fi to construct a coordinate system on a
neighborhood of some reference point 0 on the connected manifold M f . The
coordinates of this system are the parameters ti of the Hamiltonian flows generated
by the respective Fi . Using the connectedness of M f , we will show that the coor-
dinate system covers all of the manifold, usually in manyone fashion (it is what
we call a multichart). Using the so-called period lattice, i.e. the coordinate values
corresponding to the same point 0 on the manifold, we will show that M f , which
is assumed to be compact, has the topological structure of an n-torus. In a final step,
we will show that action-angle coordinates can be introduced in a neighborhood of
M f by means of a suitable canonical transformation.

Constructing coordinate curves


The principal strategy of the proof is to use the integrals Fi ( ) to construct for
each M f a multichart, i.e. a manyone smooth mapping from Rn onto M f from
which can be extracted a complete atlas of charts for the manifold. The coordinates
of the multichart will be the parameters ti of a family of coordinate curves Fi (ti )
generated by the Fi :
d Fi
=  Fi ( Fi ). (3.11)
dti
At each point of M f , the vector  Fi ( ) lies in the tangent space and is
perpendicular to all of the gradients F j ( ), since Fi  F j = [Fi , F j ] = 0.
Moreover, the n vector fields  Fi are non-zero and linearly independent at every
point of M f , thanks to the assumed linear independence of the gradients and the
nonvanishing of the determinant of the matrix . Thus, at each point there are n
distinct coordinate directions: the coordinate system generated by the n integrals
has no singular points.
Along the Fi -generated coordinate curves, the coordinates are assigned as fol-
lows (see Figure 3.1). Suppose 0 is an arbitrary point of M f . We define Fi (ti )
to be the solution of the first-order system of ODEs (3.11) with initial condi-
tion Fi (0) = 0 . Taking into account the formal power-series solution in the
neighborhood of ti = 0, we will also use the operator notations

Fi (ti ) = giti 0 = eti [,Fi ] 0 .


def def

Clearly the set {giti 0 : 0 ti T } is a curve in M f passing through the


point . Together, the mappings giti generate an Abelian (i.e. commuting) group
of transformations,
g(t) = g((t1 , . . . , tn )) = g1t1 g2t1 gnt1 .
3.5 Proof of the LiouvilleArnold theorem 65

F = (F1, F2)

v2
v1 0
g2t2 0
g1t1 0

Figure 3.1 The mutually commuting integrals Fi generate n independent coordi-


nate curves on Mf .

The commutativity of the mappings giti is easily seen to be a consequence of the


fact that the Fi commute with one another in the sense of Poisson brackets. We will
only sketch the proof. For asymptotically small ti and t j ,
eti [,Fi ] et j [,F j ] et j [,F j ] eti [,Fi ] =
ti t j ([[, F j ], Fi ] [[, Fi ], F j ]) + O((t)3 ).
By virtue of the Jacobi identity, the second-degree terms vanish, so that the remain-
der is at worst of third degree in ti and t j . For finite ti and t j , we write
ti = N ti and t j = N t j with N large. The difference
t t
giti g jj g jj giti = e N ti [,Fi ] e N t j [,F j ] e N t j [,F j ] e N ti [,Fi ]
t /N
can then be evaluated by O(N 2 ) commutations of the transformations gi i and
t /N
g jj , each one of order N 3 , so that the left-hand side is O(N 1 ). Since N can be
made arbitrarily large, we obtain the result
t t
giti g jj = g jj giti .

Extending the coordinate system


We now want to establish that the coordinate system generated by the n integrals
extends throughout M f , provided that the latter is connected. We note first that
0 M f has a neighborhood U0 that is diffeomorphic, via g(t), to a neighborhood
of t = 0 in Rn . This follows from a general existence theorem for first-order ODEs
(see [8]). Within U0 , each point can be connected to every other by a transformation
in the group.
66 Integrable systems

0 N

Figure 3.2 Any two points on Mf can be linked by a finite number of group
operations.

Since M f is connected, any point of the manifold can be linked to 0 by a finite


continuous curve C. Since C is compact, it can be covered by a finite number of
open sets Ui , i = 0, . . . , N , within each of which every pair of points is connected
by a group element. Without loss of generality, we may assume 0 U0 and =
N U N , and, for all i, Ui Ui+1 = , as pictured in Figure 3.2. Now choose a
point i in each overlap region Vi = Ui Ui+1 , i = 0, . . . , N 1. By construction,
there exist group elements g(t(k) ) such that k+1 = g(t(k) )k , k = 0, . . . , N 1,
and so
= g(t(N 1) ) g(t(1) )g(t(0) )0 .
This establishes a map g from Rn onto M f .
We note that, if M f is compact as well as connected, g cannot be oneone,
since compactness is preserved by a diffeomorphism and Rn is not compact. Thus
the coordinates t do not correspond to a single chart on the manifold, but may be a
useful way to provide a multichart, and from it, by restriction to suitable open sets,
a complete atlas of true charts.

Period lattice
To deal with the redundancy of some or all of the t-space coordinates, we introduce
the concept of the stationary group of a point 0 M f , defined to be the set
 = {t Rn : g t (0 ) = 0 }.
It is easy to see that  is a subgroup of Rn that is independent of the choice of 0 . In
fact, suppose 1 M f , 1 = 0 . Then there exists an s Rn such that 1 = g s 0 ,
and hence, for all t ,
g(t)1 = g(t)g(s)0 = g(s)g(t)0 = g(s)0 = 1 .

We now want to establish that the elements of the stationary group  for con-
nected (but not necessarily compact) M f form a lattice of dimension n. The first
step toward establishing this property, which is of utmost importance in proving
3.5 Proof of the LiouvilleArnold theorem 67

the LiouvilleArnold theorem, is to show that the points of  Rn are uniformly


isolated from one another, i.e. there exists an  > 0 such that no two points of 
are separated by a distance less than . Certainly there exists a ball V of radius /2
surrounding the origin within which g(t) is a diffeomorphism and so  V = {0}.
If s is any non-zero element of  with s + t , t V , and t = 0, then

g(s + t)0 = 0 = g(t)g(s)0 = g(t)0 = 0 ,

and hence t , which is a contradiction. Thus every element of  is separated


from its neighbors by a distance of at least .
We now complete the job of showing that the discrete points of  have the
structure of a lattice, i.e. that there exist k n linearly independent vectors
e1 , e2 , . . . , ek  such that  is exactly the set of their integer linear combinations.
If  = {0}, we are done. Otherwise there exists a non-zero e0 . Let e1 belong
def
to the straight line Re0 = {r e0 : r R}, with e1 the closest point of the line to 0.
Then the integer multiples of e1 are all the members of  lying on Re0 .
If  = Ze1 , we are done, with k = 1. If not, there exists e , e  Re1 .
Let be the distance from e and the nearest point on Re1 , which lies in the finite
segment  = {e1 : m < m + 1, m Z}. To find e2 closest to Re1 , it is
sufficient to search within a cylinder of radius and axis  (see Figure 3.3). Since
there are finitely many elements of  within the cylinder, there exists e2 of minimal
distance.

e
4e1

3e1

2e1

e1

e1

Figure 3.3 The second lattice vector e2 must be within or on the boundary of the
cylinder of radius and axis along e1 .
68 Integrable systems

If  = {m 1 e1 + m 2 e2 : m 1 , m 2 Z} we are done, with k = 2. Otherwise,


choose e3 not on the plane spanned by e1 and e2 but whose perpendicular distance
to that plane is a minimum. The existence of such e3 is guaranteed by an argument
analogous to the cylinder construction of the preceding step.
If  = {m 1 e1 + m 2 e2 + m 3 e3 : m 1 , m 2 , m 3 Z} we are done, with k = 3.
Otherwise, etc., etc. Eventually, for some k n, we have a lattice {{m 1 e1 + m 2 e2
+ + m k ek : m 1 , m 2 , . . . , m k Z}} that comprises all of .
We have now achieved our goal of a multichart for the manifold M f on which
the n integrals of the motion F1 , . . . , Fn take the constant values f 1 , . . . , f n . We
have assumed that M f is connected, but that is not crucial: if it is not connected,
we apply our coordinate construction separately to each of its connected compo-
nents. The topological structure of the connected manifold is clear from the lattice
property of the stationary group.

Topological structure
To see the topological structure of the connected manifold, it is convenient to make
an invertible linear transformation that maps the basis vectors e1 , . . . , en into an
orthogonal basis f1 , . . . , fn , with all fi of length 2 . In this new coordinate system,
it is clear that the manifold is diffeomorphic to the direct product of the compact
k-torus T k and a non-compact real vector space R nk . Clearly, if M f is compact,
we must have k = n and the manifold is an n-torus (see Figure 3.4). For the
remainder of this discussion we assume that this is the case.

Action-angle variables
The final step in the LiouvilleArnold program is to introduce action-angle vari-
ables (, J ) in a neighborhood of M f . The most straightforward way is to define

t2
e2

e1

t1

C1 C2

Figure 3.4 A typical period lattice for n = 2. The basis vectors e1 and e2
correspond to independent circles C1 and C2 on the torus, which cannot be
continuously contracted to points.
3.6 Planar free-particle examples 69

1
Ji (F) = p(q, F) dq,
2 i

where i is the image on the torus of a multichart path from 0 to ei . One can
show that the path can be continuously deformed without changing Ji . In princi-
ple, the defining relation for the Jk can be inverted to give F(J ) and in particular
the action-angle Hamiltonian

K (J ) = F1 (J1 , . . . , Jn ).

To obtain angle variables conjugate to the actions, we use the type-2 generating
function
def
S(q, J ) = W (q, F(J )),

where
 q
W (q, F) = p(q, F) dq.

Clearly
S
pi = ,
qi
and we get, in addition,
S W Fj
i = =
Ji j
F j Ji

with the time evolution


K
k (t) = k (J )t + k (0), k (J ) = .
Jk

3.6 Planar free-particle examples


We now study some simple examples that illustrate the LiouvilleArnold approach
to integrable systems. In these examples our starting point will be a two-degree-
of-freedom Hamiltonian H (q1 , q2 , p1 , p2 ) for which a second integral of the
motion, F(q1 , q2 , p1 , p2 ), is known. From this we will construct Hamiltonian
flows for H and F on a sub-manifold for which both H and F assume def-
inite values. In the process we will discover the topology of the sub-manifold
and endow it with a suitable coordinate system canonically related to the original
one. Where the sub-manifold is connected and compact, this will, by the theorem,
lead naturally to angle coordinates on a 2-torus, together with the conjugate action
variables.
70 Integrable systems

3.6.1 A two-dimensional free particle, model 1


Consider a free particle on the plane, with rectangular coordinates q1 , q2 and
Hamiltonian
1
H = ( p12 + p22 ).
2
We restrict ourselves to the sub-manifold of phase space on which both H and
p1 are constant. Here H can take on any non-negative value and, in order to have
p2 real, the value of p1 must lie in the interval | p1 | 2H , as shown in the
energymomentum diagram of Figure 3.5.
It is easy to verify that H and p1 are independent everywhere in the interior of
the parabola of Figure 3.5, but not on the boundary, where p2 = 0 and so H = 12 p12 .
We will proceed on the assumption p2 = 0.
For p2  = 0, the sub-manifold of fixed H and p1 consists of two planes
 parallel to
the q1 , q2 plane, corresponding to the two possible signs of p2 = 2H p12 . Let
us choose a representative point on one of the planes, say 0 = (q1 , q2 ) = (0, 0) on
the plane with positive p2 . It is easy to see that, starting from (0, 0), we can reach
any point = (q1 , q2 ) on the plane by means of an H motion followed by a p1
motion (see Figure 3.6)
= es[, p1 ] et[,H ] 0 ,
where 
t = q2 2H p12 , s = q1 p1 t.

The correspondence is clearly a diffeomorphism, with a oneone correspondence


between each (s, t) pair and a point on the sub-manifold.

p1

Figure 3.5 The energymomentum diagram for the free particle on a plane with
constant H and p1 .
3.6 Planar free-particle examples 71

q1 t

(0, t) (s, t)
(p1t, q2) s [ . , p1 ] (q1, q2)
e

t[.,H]
e

q1 s
(0, 0) (0, 0)

Figure 3.6 Construction of the s, t coordinate system. Any point q in the plane
can be reached by a temporal displacement of t, generated by H , followed
by a spatial displacement s in the q1 direction, generated by p1 . The left and
right figures show these displacements in the old and new coordinate systems,
respectively.

If we now want to construct coordinates for our sub-manifold that are canon-
ically related to the original phase-space coordinates, we define a generating
function of type 2, designed to give the correct transformation equations for p1
and p2 :
 
S(q1 , q2 , p1 , H ) = ( p1 dq1 + p2 ( p1 , H )dq2 ), p2 = 2H p12 ,

where the path connects (0, 0) to (q1 , q2 ) but is otherwise arbitrary. Trivially,

S(q1 , q2 , p1 , H ) = p1 q1 + 2H p12 q2 .
The new configuration variables are thus
S q2 S p1 q 2
Q2 = = = t, Q1 = = q1 = q1 p1 t = s.
H p2 p1 p2
We see that in this example the flow coordinates are in fact canonically conjugate
to the conserved quantities which we started with.

3.6.2 A two-dimensional free particle, model 2


This is the same system as we just considered, but now we restrict ourselves to a
sub-manifold of constant Hamiltonian H and angular momentum L = q1 p2 p1 q2 .
While H must be non-negative, L can now take on any real value. We restrict
ourselves here to H = 0, L = 0.
We begin with some elementary observations concerning the time evolution
gen-
erated by H . For a given point q = (q1 , q2 ), |q| = q1 + q2 > b , b = L/ 2H ,
2 2 2 2
72 Integrable systems

there are precisely two straight-line orbits passing through the point. Both are
tangent to the circle |q| = b, one moving outward and the other moving inward.
For |q| = b, there is just a single orbit tangent to the circle at that point. Finally,
for |q| < b, there are no points compatible with the specified Hamiltonian and
angular momentum. We note that the outgoing tangent half-lines cover the entire
plane exterior to the circle, and the same holds for the incoming half-lines (see
Figure 3.7). The full sub-manifold M H,L consists of the two planes with central
disks excised, glued together along the circle of radius b. Topologically this is
equivalent to an infinite cylinder.
To make the assignment of flow coordinates s and t for the L and H motions,
respectively,
we choose a representative point 0 on M H,L , say q = (b, 0), p =
(0, 2H ). To get from 0 to any point

= (q1 , q2 , p1 (q, H, L), p2 (q, H, L))

on the outgoing part M+H,L of M H,L we need only rotate through an angle s =

(q) cos1 (b/|q|) and travel forward in time by t = 2H |q|2 L 2 /(2H ) (see
Figure 3.8), so that
= et[,H ] es[,L] 0 .

For the incoming part MH,L of M H,L , we can start again at 0 , rotate by s =

(q) + cos1 (b/|q|), and then travel in time by t = 2H |q|2 L 2 /(2H ). In the
s, t plane, M+H,L (MH,L ) occupies the upper (lower) half-plane, with the common

q2 q2

q1 q1

Figure 3.7 Outward (left) and inward (right) free-particle trajectories of fixed
energy H > 0 and angular momentum L > 0. For negative L, the figures would
be interchanged.
3.6 Planar free-particle examples 73

q
q2
| q| 2
b 2
=
2H
t
|q|

b
s q1

Figure 3.8 Any point in the annulus can be reached by starting at (b, 0), rotating
through an angle s, and then moving as a free particle for time t.

boundary circle occupying the axis t = 0. Note that there is a 2 n ambiguity in


(q), which translates into a 2 n multiplicity in the s direction. A single point on
M H,L is represented by the infinitely many s, t points differing by a multiple of
2 in their s coordinate. This multiplicity corresponds, of course, to the fact that
the manifold is topologically a cylinder. In the language of the LiouvilleArnold
theorem, the stability group (period lattice) is the one-dimensional lattice of points
(2n, 0), n = 0, 1, 2, . . ..
As in the previous example, one can calculate the canonical transformation
generated by
 
b
S(q1 , q2 , L , H ) = L (q) cos1 2H |q|2 L 2 . (3.12)
|q|
Once again the new configuration coordinates are just s and t. The action variable
conjugate to s is, of course, L itself.

3.6.3 Circular stadium billiard


So far in our examples, we have not seen the LiouvilleArnold theorem in full
force, since the invariant manifolds have been unbounded, and hence non-compact.
A small modification of the boundary conditions will change that: we introduce
a circular reflecting stadium of radius a, allowing the particle to move freely
within it, bouncing off the wall like a frictionless billiard ball. The collisions at the
boundary are assumed to be instantaneous and perfectly elastic. Thus, although the
74 Integrable systems

normal component of moment reverses at a collision, the tangential component,


and hence the angular momentum, are conserved.
For an unconfined free particle of energy H , there is no limit to the size of the
angular momentum L: one can always go arbitrarily far from the origin to make
L as large as one likes. Inside a circle of radius a, there is a maximum angular
momentum, namely

L max (H ) = a 2H , (3.13)
corresponding to an orbit circulating around the boundary in the counterclock-
wise direction. Obviously the minimum of L is just L max , corresponding to
a clockwise circular orbit. The energymomentum diagram for this system is
shown in Figure 3.9. Equation (3.13) is an explicit expression of the breakdown
of independence of H and L on the bounding parabola.
For H > 0, |L| < L max , the manifold M H,L consists of two annular sheets,
+
M H,L and M H,L , glued together at the circle |q| = b = L/ 2H . The annu-
lus M+H,L contains the outgoing straight-line orbit segments, with endpoints on the
circle of radius b and the circle of radius a. The latter contains the inner straight-
line orbit segments connecting the circles in the opposite sense (see Figure 3.10).
In phase space, the outer boundaries of the two annuli do not coincide, since the
collisions represent discontinuities in the momenta. However, as we shall see, it is
possible to introduce local coordinates in the neighborhood of a collision in such a
way that the incoming and outgoing line segments together form a single straight-
line segment, i.e. the collision disappears! This is actually nothing more mysterious
than unfolding a folded piece of paper (see Figure 3.11). The result is to convert
M H,L into a 2-torus on which the particle moves without deflection. It is on this
manifold that we apply the LiouvilleArnold technique.

Figure 3.9 The energymomentum diagram for the free particle inside a
reflecting circle.
3.6 Planar free-particle examples 75

q2 q2

q1 q1

Figure 3.10 Outgoing and incoming free-particle orbits inside a reflecting circle.

|q| = a

|q| = a

p = (p|| , p)

p = (p|| , p)

Figure 3.11 In the original phase space, there is a momentum discontinuity for
orbits that reach the outer boundary. If we identify the points on each side of the
discontinuity, we effectively glue together the incoming and outgoing orbits. The
resulting manifold can be endowed with a coordinate chart at |q| = a for which
no discontinuities are present and the orbit is undeflected. The paper-unfolding
analogy shows how this miracle occurs. The resulting manifold is not diffeomor-
phic to a sub-manifold of the original phase space, but let us not worry about that.
It certainly gives a faithful representation, on a torus, of the dynamics for constant
energy and angular momentum.

We proceed as in the unbounded case, selecting our representative point to have


q1 = b and q2 = 0. For q M+H,L , there is a unique outgoing ray joining a
point (b cos , b sin ) on the inner circle with q. For such a point we make the
asssignment

1 b 2H |q|2 L 2
s = = arg(q) cos , t= .
|q| 2H
Similarly, for q MH,L , we have
76 Integrable systems

b 2H |q|2 L 2
s = arg(q) + cos1 , t = .
|q| 2H
As in the unbounded case, we have 2 periodicity in the s direction.
In contrast, we
have defined the temporal coordinate only on the interval |t| 2H a 2 L 2 /(2H ).
However, we can already spot a second periodicity in the s, t plane: (s, t/2) and
(s + s, t/2), where

1 L 2H a 2 L 2
s = 2 cos , t = ,
a 2H H
correspond to the same point on the circle |q| = a. The same multiplicity of the
s, t representation is easily seen to hold for any q: we can always find an orbit
that connects the inner circle to q via a path containing arbitrarily many visits to
the outer circle (see, for example, Figure 3.12). Each innerouterinner excursion
requires an additional rotation s and time interval t.

q2

q
|q| 2
b2
=
2H
t
|q|

s q1
b s 2H
a
t/2
2H
t/2

Figure 3.12 The same phase-space point as in Figure 3.8 can be reached from
(b, 0) by longer paths that visit the outer circle arbitrarily many times before
heading for their target. Shown here is the case of a single bounce.
3.6 Planar free-particle examples 77

2
3e
+
1
4e
e2
s
e1

Figure 3.13 The period lattice of the circular stadium billiard, consisting of the
integer linear combinations of the period vectors e1 and e2 .

Our doubly periodic s, t representation of M H,L , with the period lattice of


Figure 3.13, now differs only by a linear transformation from the standard coor-
dinate assignment for a 2-torus. In particular, if the period vectors are e1 = (2, 0)
and e2 = (s, t), we define angles 1 and 2 such that
 
1 2
(s, t) = e1 + e2 ,
2 2
so that
s 2
1 = s t, 2 = t, (3.14)
t t
To complete the transformation to action-angle variables, we define

1
Jk = ( p1 dq1 + p2 dq2 ), k = 1, 2,
2 k
where the paths 1 and 2 correspond to 0 1 < 2, 2 = 0, and 1 = 0, 0
2 < 2, respectively. The integrals are easily carried out (Exercise 3.8), yielding
J1 = L , J2 = (L s + 2H t)/(2). (3.15)
This gives a transcendental equation for the Hamiltonian as a function of J1 and J2 .
However, the corresponding oscillation frequencies are easily calculated in closed
form (compare with (3.14)):
78 Integrable systems
 
H J2 J2 s H J2 2
1 = = = , 2 = =1 = .
J1 L H t J2 H t
To show that the transformation (q1 , q2 , p1 , p2 ) (1 , 2 , J1 , J2 ) is canonical,
we introduce the generating function of type 2,

S(q1 , q2 , J1 , J2 ) = ( p1 dq1 + p2 dq2 ),

where is any path connecting the reference point (b, 0) to (q1 , q2 ). The evaluation
of the line integral is straightforward (Exercise 3.8), giving us

S(q1 , q2 , J1 , J2 ) = Ls + 2H t, (3.16)

with s, t, L , H expressed in terms of the appropriate variables. Differentiation of


S with respect to J1 , J2 once again produces (3.14).

3.7 Spherical pendulum


In Chapter 1 we introduced a relatively simple model with two degrees of freedom,
namely the spherical pendulum (Figure 3.14): a point mass confined to the surface
of a sphere in the presence of gravity, with Hamiltonian
 
2
1 p
H= p2 + 2 + cos .
2 sin

mg

Figure 3.14 A spherical pendulum.


3.7 Spherical pendulum 79

Here we use spherical polar coordinates and units such that m = g = 1. By


inspection,
[ p , H ] = 0,
so we need only verify the independence of the integrals p and H to establish
integrability. Indeed, the gradient vectors
 
p2 p
p = (0, 0, 0, 1), H = 3 cos sin , 0, p , 2
sin sin
are linearly dependent only for
p = 0, p2 cos + sin4 = 0, 0 < < ,
corresponding to the explicit functional relation
4   
p2 = 2H + H 2 + 3 3 H 2 + H H 2 + 3 , (3.17)
27
as well as for the equilibrium points where p , p , and sin all vanish.
One can readily check, using the positivity of p2 , that (3.17) corresponds to the
maximum allowed value L 2max of p2 for a given H = E. We display the allowed
region of the E, L plane in the energymomentum diagram Figure 3.15. We note
that the condition p = L = 0 (the midline of the diagram) corresponds to one-
dimensional motion of a simple pendulum in a single plane. Since the orbits (with
isolated exceptions) pass through the point = , a singularity of the coordinate
frame, the azimuthal angle has a discontinuity of at that point.

E
2

0
L
2 1 1 2

Figure 3.15 The energymomentum diagram for the spherical pendulum.


80 Integrable systems

We now consider the level sets M E,L , where H = E and p = L. For 0 <
|L| < L max , M E,L is diffeomorphic to the direct product of closed curves in the
, p and , p planes, namely the graphs of

L2
p = 0, p = 2E 2 2 cos .
sin
The , p curves are displayed in Figure 3.16 for three values of the energy. Clearly
the system is separable for this range of L, with M E,L diffeomorphic to a 2-torus.
For L = 0, 1 < E < 1, the level set is again diffeomorphic to a 2-torus, the
discontinuity in the azimuthal angle notwithstanding. The motion is a libration in
an arbitrary plane labeled by and + . For L = 0, E = 1, the level set
contains the unique stable equilibrium point at = . Finally, the level set for
L = 0, E = 1 contains not only the unstable equilibrium point at = 0, but also
the adjoining separatrices of the planar pendulum orbits. The lack of smoothness
at = 0 prevents the level set from being diffeomorphic to a true torus; rather it is
equivalent to what is called a pinched torus (see Appendix B of [9]).
We now restrict ourselves to the cases where M E,L is a 2-torus. Our goal is
to construct on each level set a coordinate system as prescribed in the Liouville
Arnold theorem (for the fast-track approach, see Exercise 3.10). For our system
of flow coordinates s, t, we pick as our reference point 0 on the torus the point
(which always exists) for which = 0, p = 0, and = 1 = cos1 u 1 , where u 1
is the root of the cubic polynomial
f (u) = 2(1 u 2 )(E u) L 2

p 2 2
2
E = 1/2 E=1 E = 3/2

1 1
1

0 0
0

1 1
1

2 2
2
0 1 2 3 0 1 2 3 0 1 2 3

Figure 3.16 Projected orbits in the , p plane for L 0, E = 12 , 1, 32 . Each


closed contour corresponds to a distinct value of L. The projected orbits for L 0
look the same.
3.7 Spherical pendulum 81

furthest to the left on the real axis. Starting at 0 , we can reach any point on the
torus by first applying et [,H ] , then es [, p ] , arriving at s,t . Let us calculate the s
and t values for a given point
  
f (u)
= (, , p , p ) = cos1 u, , , L M E,L .
1 u2
From Chapter 1, we have that the period of the H -generated motion is
 u2
du
T =2 ,
u1 f (u)
and the elapsed time t is given by

" u du 

, even,
T u 1 f (u  )
t = + (3.18)
2 " u 2 du


u , odd,
f (u  )
independently of . As a result of the H -generated motion alone, we end up at the
point
  
t [,H ] 1 f (u)
0,t = e 0 = cos u, , ,L ,
1 u2
where, again according to the treatment of Chapter 1, the excursion during
time t is
"u du 

, even,
u1
(1 u 2 ) f (u  )
(u) = + L (3.19)
2
" u2 du 

u , odd,
(1 u 2 ) f (u  )
where  u2
du
= 2 (u 2 ) = 2L .
u1 (1 u 2 ) f (u)
We reach our final destination, s,t , by means of a pure rotation generated by p :
  
s [, p ] 1 f (u)
s,t = e 0,t = cos u, , ,L ,
1 u2
where
s = (u). (3.20)
We have not yet finished: our phase-space coordinate system is not necessarily
canonically related to the original one, and it is not yet oneone. As in the theorem,
82 Integrable systems

we must construct the so-called period lattice, the lattice of the stationary group of
the flows generated by H and p , and then proceed to construct the action-angle
variables. As discussed in the proof of the theorem, the period lattice (stationary
group) can be identified with the lattice L of points (s, t) in R2 such that s,t = 0 .
In the present case, the calculation of the period lattice is particularly easy. The
points 2k,0 are obviously equal to 0 and hence the points (2k, 0) are on the
lattice. Moreover, we have already studied the Hamiltonian flow generated by H ,
and determined that the orbit starting at 0 returns to 1 for the first time after a
time T with a change in equal to :

0,T = (1 , , 0, L),

which implies
,T = 0 .

Hence ( , T ) L, and in fact

L = { j ( , T ) + k(2, 0) : j, k Z}.

We are now ready to define the action-angle variables for a given E, L torus.
These correspond to a choice of basis vectors e1 = (2, 0) and e2 = ( , T ) for
the period lattice. We define corresponding actions

1
J1 = p , J2 = ( p d + p d),
2
where the integral is along a contour connecting (0, 0) to ( , T ). To evalu-
ate the line integral, we choose the contour specified by (3.18) and (3.20) with
identically equal to zero. Then the evaluation of J2 becomes
  u2
1 f (u)
J2 = p d = 2 du .
2 u1 1 u2
The frequencies governing the s and t evolution of the angle variables 1 and 2
conjugate to the actions J1 and J2 are

E J2 J2
E1 = = = ,
J1 L E T

E J2 2
E2 = =1 = ,
J2 E T
(3.21)
L
L1 = = 1,
J1
L
L2 = = 0.
J2
3.7 Spherical pendulum 83

Thus, as functions on the s, t plane,

k (s, t) = Ek t + k (s, 0) = Lk s + (0, t), k = 1, 2.

Since k (0, 0) = 0, k = 1, 2, we get


2
1 = E1 t + s = t + s, 2 = E2 t = t. (3.22)
T T
Upon inserting the transformation equations, this gives us, for cos1 u 1 < <
cos1 u 2 , p < 0,
 cos 
du L
1 (, ) = + , (3.23)
u1 f (u) T 1 u2

2 cos du
2 (, ) = . (3.24)
T u1 f (u)
The calculation of these formulas for p > 0 is relegated to Exercise 3.11.
On the multichart, any point (s, t) can be expanded in terms of the basis vectors
e1 and e2 . Upon inverting (3.22), one finds that the expansion coefficients are just
the angle variables, divided by 2:
 
1 2
(s, t) = e1 + e2 .
2 2
For those M E,L which are 2-tori, the actions J1 and J2 and their conjugate
angle variables provide a convenient set of canonical coordinates, not only on the
sub-manifold, but also in a finite neighborhood of it. The question of whether a
single global assignment of action-angle variables is possible over the entire phase
space naturally arises. In other words, can the coordinate system introduced in one
small neighborhood be extended continuously everywhere without encountering an
obstruction? What makes this an open question is the fact that the choice of actions
for a given E, L pair is not unique. For any integer k, the quantity J2(k) = J2 +k J1 is
just as suitable a candidate for an action variable (independent of J1 and commuting
with it) as J2 . This is the question of monodromy [10], which is of some relevance
to semi-classical physics, where one would like to have global action variables
in order to carry out a consistent assignment of quantum numbers via Einstein
BrillouinKeller quantization [11, 12, 14, 39]. A readable discussion of this issue,
with specific reference to the spherical pendulum model, can be found in [9].
According to a theorem of Cushman and Duistermaat [15], it is known to
be impossible to assign global action-angle coordinates over any open disk of
the energymomentum diagram that contains the critical point (E, L) = (1, 0), for
which the sub-manifold M E,L degenerates to a pinched torus. We will not study
this in the abstract, but rather illustrate the nontrivial monodromy by choosing a
84 Integrable systems

closed path C surrounding (1, 0) and seeing how the coordinate system varies as
one makes a full circuit. Our choice of path, a unit circle centered at (1, 0), is
depicted in Figure 3.15.
Referring to Figure 3.17, we now start at the point E = 2, L = 0 on the curve C.
Here the lattice is rectangular and our actions J1 and J2 generate horizontal and
vertical translations, respectively. The torus is covered once by the points in the
black rectangle, with the angle coordinates 1 and 2 in the range [0, 2). Moving
continuously around the circle C in the counterclockwise sense, we construct the
period lattice corresponding to each pair E, L. The torus is always covered by
the black lattice cell, which undergoes a monotonic horizontal stretching as one
proceeds. After a full circuit, the top edge of the black cell has shifted by 2 relative
to the bottom edge. The nontrivial monodromy is thus evident.

15 15 15

10 10 10

5 5 5

15 10 5 5 10 15 15 10 5 5 10 15 15 10 5 5 10 15

5 5 5

10 10 10

15 15 15

15 15
E
10 C 10

5 1 5

15 10 5 5 10 15 15 10 5 5 10 15

5
5
-2 1 1 L 2
10
0 10

15 15
1

15 15 15

10 10 10

5 5 5

15 10 5 5 10 15 15 10 5 5 10 15 - 5 10 15
15 10 5

5 5 5

10 10 10

15 15 15

Figure 3.17 The period lattice, tracked around the path C of the energy
momentum diagram of the spherical pendulum. The eight points have E =
1 + cos(k/4), L = sin(k/4), k = 0, 1, . . . , 7.
3.8 The three-particle Toda model 85

3.8 The three-particle Toda model


The N -particle Toda model [16] provides a beautiful example of an integrable sys-
tem which is not separable, but which yields, nonetheless, to the subtler analysis of
the LiouvilleArnold theorem. We will limit ourselves to the simplest nontrivial
case, namely N = 3. The initial stage of our analysis borrows from Section 1.3
of [17].
The system consists of three equal-mass particles on an infinite line, with
Hamiltonian
1
HToda = ( p12 + p22 + p32 ) + e x3 x1 + e x1 x2 + e x2 x3 3.
2
Since the potential energy is translation-invariant, it is natural to introduce the
total momentum as a new phase-space coordinate via the canonical transformation
(Exercise 3.12)
X 1 = x1 x3 , X 2 = x2 x3 , X 3 = x3 , (3.25)
P1 = p1 , P2 = p2 , P3 = p1 + p2 + p3 .
The new Hamiltonian is then
1
K = (P12 + P22 + (P3 P1 P2 )2 ) + eX 1 + e X 1 X 2 + e X 2 3.
2
Recognizing that K has no explicit X 3 dependence, we infer that the total momen-
tum P3 is conserved, and so, by choosing the center-of-mass frame, with P3 = 0,
we reduce the number of degrees of freedom of the system by one:
K = K (X 1 , X 2 , P1 , P2 ).
With an additional canonical transformation (X 1 , X 2 , P1 , P2 ) (x  , y  , px , p y )
generated by
1  
F2 = ( px 3 p y )X 1 + ( px + 3 p y )X 2 ,
4 3
the Hamiltonian simplifies to
1 2
2y  2 3x  2y+2 3x  
( px + p 2
y ) + e + e + e4y 3.
16
A final rescaling of the dynamical variables and the time,

x  = x, y  = y, px = 8 3 px , p y = 8 3 p y , t = / 3,
(3.26)
gives us, in standard form, the Hamiltonian of a particle in a two-dimensional
potential well,
1
H = ( px2 + p 2y ) + V (x, y), (3.27)
2
86 Integrable systems

2
2

1
0
2

1 0

1 1

Figure 3.18 A three-dimensional plot of the Toda potential energy V (x, y).

where
1  2y23x 
V (x, y) = e + e2y+2 3x + e4y 3 .
24
A three-dimensional plot of the potential energy is displayed in Figure 3.18.
The phase-space manifold of constant energy E corresponds to the subset of
R4 = {(x, y, px , p y )} satisfying
1
H (x, y, px , p y ) = ( px2 + p 2y ) + V (x, y) = E.
2
This is three-dimensional, so we can try to visualize it, namely as the set of all
triples (x, y, p y ) for which

px = 2(E V (x, y)) p 2y

is real. This set can be imagined (see Figure 3.19) as a pair of three-dimensional
regions, each bounded by a closed surface ( px = 0) topologically equivalent to a
2-sphere, with the two boundary surfaces identified. The attempt at visualization is
analogous to mapping the northern and southern hemispheres of the Earth on two
disks, with the circular equatorial boundaries identified. It is clear that in our case
we are trying to visualize two halves of a 3-sphere on a pair of three-dimensional
disks sharing a common equatorial 2-sphere.
Next, we want to understand how the constant-energy manifold is organized.
A useful device here is the Poincar section. First we select a convenient plane
for our section, for example x = 0, chosen because it is fairly obvious that all
3.8 The three-particle Toda model 87

x x
px = 0
identify

py py
px > 0 px < 0

y y

Figure 3.19 Representation of a constant-energy sub-manifold of phase space as


two three-dimensional regions sharing a common boundary diffeomorphic to a
2-sphere. The px > 0 midplane, with x = 0, serves as a surface of section.

E=1 E = 256
1.5

20
1

10
0.5

0 0

0.5
10

1
20

0.5 0 0.5 1 1.5 2 1 0 1 2 3 4

Figure 3.20 Toda model Poincar sections for E = 1 and E = 256.

orbits must eventually intersect this plane. We further select as coordinates in this
reference plane y and p y , supplemented by the sign of px . To generate a phase
portrait, we choose a representative sample of initial points (y0 , p y0 ), with px > 0,
and solve the equations of motion numerically, recording all points at which each
orbit returns to the plane x = 0 with x increasing.
High-precision Poincar sections for two values of the energy, E = 1 and E =
256, are shown in Figure 3.20. It is clear that, with the exception of two initial
points, all of the orbits lie on simple closed curves, topologically equivalent to
circles, supporting the conjecture that the constant-E manifold is partitioned into
invariant 2-tori. This is further supported by looking at sections with non-zero x.
88 Integrable systems

For the exceptional initial points, one finds that the orbits are 1-tori. Since the tori
are ubiquitous, it would appear that the system is integrable.
But if the Toda system is integrable, one would expect to find a second conserved
quantity, independent of the Hamiltonian and commuting with it. Such an integral
of the motion in involution with H was discovered by Hnon in 1974 [18]. It is
1 1
K = px ( px2 3 p 2y ) px e4y
2 8
1 2y  
+ e 2 px cosh(2 3x) 2 3 p y sinh(2 3x) . (3.28)
16
We leave it as an exercise to verify the vanishing of the Poisson bracket
(Exercise 3.14). The question of independence is more subtle. Obviously K is not
a global function of H , but independence may break down on lower-dimensional
sub-manifolds. To investigate this, we set up the 2 4 matrix of partial derivatives
H H H H
x y px py
.
K K K K
x y px py
Dependence requires the simultaneous vanishing of all six rank-2 minors of this
matrix, giving us the equations

6dpx + a 3 cp y bcp y = 0,
d + a 3 bd + 6a 2 bcpx p y 12a 2 dp 2y = 0,
acd 2 + 72dpx p y 12bcp 2y = 0,
1 2a 3 b + a 6 b2 36a 2 px2 + 6a 5 cdpx p y + 12a 2 p 2y 12a 5 bp 2y = 0
cd a 3 bcd 72a 2 bpx p y + 12a 2 cdp 2y = 0,
a 3 cdpx + 2 p y 2a 3 bp y 72a 2 px2 p y + 24a 2 p 3y = 0,

where

a = e2y , c = 2 3, b = cosh(cx), d = sinh(cx).

The first equation implies (assuming x = 0)


a 3 cp y + bcp y
px = .
6d
Substituting this into the third equation then gives
d
py = .
ac
3.8 The three-particle Toda model 89

Together,
the two substitutions (together with the identities d 2 = b2 1 and c =
2 3) lead to satisfaction of all the remaining equations. The restriction of the two-
dimensional manifold defined by the two equations to energy E produces a closed
curve whose projection on the x, y plane has the equation

2 cosh(2 3x) = 3e2y 8E + 1 e6y .
Along the curve, K takes the constant value
1  
K (E) = (8E + 1)3/2 1 . (3.29)
16
The projected curves for several values of E and K are displayed in Figure 3.21.
For each sign, the curves for arbitrary E collectively form an infinite two-
dimensional manifold with the topology of a cone, with the vertex at the stable
equilibrium point.
Since K (E) are also the maximum and minimum values assumed by K for
a given E, we see that they define the outer boundary of the energymomentum
diagram for the Toda system, shown in Figure 3.22. In contrast to the spherical
pendulum, there is no unstable equilibrium point and no nontrivial monodromy. It
will be possible, in principle, to define global action-angle variables.

1.5

1.0

0.5

0.0

0.5

1.0

1.5

1.5 1.0 0.5 0.0 0.5 1.0 1.5

Figure 3.21 Maximum-K orbits in the x, y plane for initial points (0, 0.25),
(0, 0.529 246), (0, 0.75), and (0, 1.0), corresponding, respectively, to H, K
values (0.089 659 2, 0.078 149 8), (1, 1.625), (5.729 93, 19.9729), and (41.6887,
382.316).
90 Integrable systems

1.5

0.5

K
4 2 0 2 4

Figure 3.22 The energymomentum diagram for the Toda system. The sub-
manifolds M E,K associated with the interior points of the wedge are all 2-tori.

The three-particle Toda model is an ideal laboratory for testing out the Liouville
Arnold methods, since there does not appear to be any reasonable alternative. Let
us choose particular values E = 1 and K = 1 and a representative point 0 =
(0, 0, 1, 1) M1,1 , and then construct an s, t coordinate system by integrating
numerically the Hamiltonian flow equations
d d
= [, K ], = [, H ].
ds dt
In our usual concise notation, every point of M1,1 can be represented, perhaps
non-uniquely, as
(s, t) = et[,H ] es[,K ] 0 .

From the LiouvilleArnold theorem, we expect to find a double periodicity of


the s, t representation. Unlike the simpler examples studied earlier in the chapter,
we have no a priori idea of what the period vectors should be. To explore the ter-
rain, we calculated, using fourth-order RungeKutta integration, a 100 100 array
of (s, t) values and used the Mathematica function ListInterpolation
to interpolate over the square 0 s, t 5. We then used the function
ContourPlot to plot the contours on which (s, t)i = (0, 0)i , i = 1, 2, 3, 4.
The result is shown in Figure 3.23. The period lattice consists of those (s, t) points
where all four components of return to their initial values. After obtaining the
lattice vectors in this fashion, we improved their accuracy using Newtons method
on the plane, with 100-digit precision. The two vectors (rounded off) for this
torus are
Exercises 91

100

80

60
t 20

40

20

20 40 60 80 100
s 20

Figure 3.23 A plot of the contours where (s, t)i = (0, 0)i , i = 1, 2, 3, 4. The
curves are drawn with the following gray levels: x = black, y = dark gray, px =
medium gray, p y = light gray. The period lattice consists of the mutual intersection
points of all four contours.

e1 = (2.810 355 880 39, 1.857 855 913 92),


e2 = (1.522 737 906 66, 3.150 184 253 86).

Once the period vectors have been determined, the action-angle variables can be
introduced in systematic fashion. The angle i increases uniformly from 0 to 2 as
one traverses the straight-line path from the origin to ei . Moreover, high-precision
numerical values of the corresponding actions Ji can be obtained by evaluating the
contour integrals of px d x + p y dy between the same endpoints.

Exercises
3.1 The Hamiltonian of a freely falling body is H (z, p) = p 2 /(2m) + mgz,
where z is the height of the particle above the ground and m is its mass. Find
92 Integrable systems

a canonical transformation for which the Hamiltonian is the new momentum


variable. What is the new position variable Q? Find Q(t) from Hamiltons
equations in the new canonical frame. Then transform back to get z(t). Does
the result look familiar?
3.2 Derive the Hamiltonian for a particle of mass m attracted to two fixed (New-
tonian) gravitational centers at points (c, 0) in the x, y plane using confocal
elliptical coordinates , given by

x = c cosh cos , y = c sinh sin , < < , 0 2.

Show that the system is not only integrable, but also separable. Calculate the
HamiltonJacobi generating function (3.1) in terms of explicit integrals.
3.3 Given a Hamiltonian system with n degrees of freedom on a phase mani-
fold with coordinates (q, p) = (q1 , . . . , qn , p1 , . . . , pn ), suppose we have
made a canonical transformation to coordinates Q = (Q 1 , . . . , Q n ), P =
(P1 , . . . , Pn ) such that the transformed Hamiltonian is P1 (see, for exam-
ple, Section 3.2). Find a (possibly time-dependent) canonical transformation
(Q, P) ( Q, P) such that the new Hamiltonian H ( Q, P) is identically
zero and all of the coordinates Q 1 , . . . , Q n , P1 , . . . , Pn are constants of the
motion.
3.4 Find the action-angle variables for a particle moving in the one-dimensional
periodic potential
#
U x/ x 0,
V (x) = V (x) = V (x + 2).
U x/ 0 x ,

3.5 A particle of mass m moves in the one-dimensional potential


#  2 
A a (x a)2 0 x 2a, A > 0,
V (x) =
0 x 2a,

and bounces elastically off a wall at x = 0. Sketch the phase portrait, showing
clearly all regions of libration and unbounded motion. For bounded motion,
calculate the action as a function of  = E/(Aa 2 ). Show that the libration
period tends to infinity as ln(1 ) for  1.
3.6 Under what circumstances does the formal power-series representation giti =
eti [,Fi ] have rigorous mathematical validity? Show that when it does, if r and
s are arbitrary non-negative integers, the coefficient of tir t sj in the expansion
t t
of giti g jj g jj giti vanishes by virtue of the Jacobi identity.
3.7 For the free particle in the plane with constant H and L, verify that S of (3.12)
generates a canonical transformation from (q1 , q2 , p1 , p2 ) to (s, t, L , H ),
where s, t are the flow variables discussed in the text.
Exercises 93

3.8 For the free particle inside a reflecting circle, verify (3.15) and (3.16). Show
explicitly that the latter gives the correct expressions for p1 , p2 , 1 , 2 . Study
numerically the behavior of H (J1 , J2 ) defined implicitly by (3.15).
3.9 Draw the analogue of Figure 3.12 for a three-bounce orbit.
3.10 Derive action-angle coordinates for the spherical pendulum using the fast-
track approach of Section 3.2.
3.11 Calculate the transformation formulas analogous to (3.23) and (3.24) for
p > 0.
3.12 Verify that the transformation X (x, p), P(x, p) of (3.25) is canonical. Find
a generating function for it.
3.13 Verify that the transformation (3.26) is not canonical. Why are we nonethe-
less allowed to do it?
3.14 Verify that H and K of the Toda system, given in (3.27) and (3.28), have a
vanishing Poisson bracket. Computer assistance is recommended.
3.15 For the Toda model, calculate numerically the period vectors e1 and e2 , as
well as the actions J1 and J2 for H = 1, K = 1. What are the frequencies
1 and 2 of the 1 and 2 oscillations?
3.16 A two-dimensional harmonic oscillator has Hamiltonian
1
H = ( p12 + p22 + 2 q12 + 2 q22 ).
2
Verify that the angular momentum L = q1 p2 p1 q2 has vanishing Poisson
bracket with H . For which values of E and L are these independent? Study
this model using the method of the LiouvilleArnold theorem proof, leaving
aside what you already know about harmonic oscillators until the very end.
3.17 A nonlinear dynamical system with two degrees of freedom has the Hamilto-
nian

H (1 , 2 , I1 , I2 ) = 2( + )I1 + I2 + (4I12 + 2I1 I2 ) cos(1 22 ).

(a) Verify that F = 2I1 + I2 is conserved by the Hamiltonian flow generated


by H .
(b) Make a canonical transformation (1 , 2 , I1 , I2 ) (1 , 2 , L 1 , L 2 ) such
that
1
1 = 2 , 2 = 2 , L 1 = F.
2
Write the new Hamiltonian, K (1 , 2 , L 1 , L 2 ).
(c) On the manifold M E, f on which H and F take the values E and f ,
calculate L 1 as a function of 1 . Sketch the graph of this function. Check
that the topological structure of M E, f is a 2-torus, as required by the
LiouvilleArnold theorem.
94 Integrable systems

(d) Define the action variables corresponding to the independent circles on


the torus. Calculate the total energy and the two oscillation frequencies
as functions of J1 and J2 . You may express your results in terms of the
integral

d
G(J, ) =
0 + J cos
and its partial derivatives.
3.18 A particle of mass m moves on a frictionless hyperboloid of revolution under
the influence of a uniform vertical gravitational field (acceleration g). The
surface may be represented parametrically in three-dimensional space as

x = a sinh cos ,
y = a sinh sin ,
z = a cosh ,

with 0 < , 0 < 2 and a a positive constant.


(a) In terms of the coordinates and and their time derivatives, write down
the Lagrangian of the system. Simplify the notation by choosing units
such that ma 2 = 1 and writing = g/a.
(b) Calculate the momenta p and p conjugate to and and calculate the
Hamiltonian H of the system.
(c) Show that H and p are constants of the motion. What are the dimen-
sionality and topological structure of a manifold M E,L on which H and
p take on specific values E and L, respectively? You may assume that
the two quantities are independent and that the manifold is bounded and
connected.
(d) Show that on M E,L the motion is governed by a differential equation
of the form
f (u)
u 2 = 2 ,
2u 1
where f (u) is a cubic polynomial in the variable u = cosh . Sketch the
function f (u) in the physical region u 1. Show that all orbits are
bounded.
(e) In terms of the function f (u), calculate the period of the motion and
the amount  by which the azimuthal angle increases during time .
(f) Describe qualitatively the motion of a particle released from rest at time
zero ((0) = 0 > 0, (0) = (0) = (0) = 0). Sketch the graphs of
and as functions of time.
Exercises 95

3.19 A particle moves in three dimensions with Hamiltonian (in spherical polar
coordinates)
pr2 p2
H (r, , , pr , p , p ) = + + mgr cos .
2m 2mr 2 sin2
(a) Derive Hamiltons equations of motion for all six of the phase-space
variables. Verify that , p , and H are constants of the motion.
(b) Show that the radial motion is governed by an effective one-dimensional
potential-energy function of the form
a
Veff (r ) = 2 + br,
r
where a and b are positive functions of the constants , p , and H . Let
E be the energy and let V0 be the minimum value of Veff . Describe the
motion in configuration space and in phase space when (i) E > V0 , (ii)
E < V0 , and (iii) E = V0 .
(c) Prove the independence of the three functions , p , and H everywhere
in the phase space except for a manifold of dimension less than 6. Give
the defining equations for this sub-manifold and describe precisely the
orbits which inhabit it. What are the functional relations among the three
integrals on the exceptional manifold? Relate your results to the results
of part (b).
(d) Consider the sub-manifold M of phase space for which , p , and
H assume specific (non-exceptional) values 0 , L, and E > V0 (0 , L),
respectively. Does M have the topology of a 3-torus, as it must if
it satisfies the integrability criteria of the LiouvilleArnold theorem?
Explain.
3.20 In parabolic coordinates u, v, , the potential energy of a charged particle
interacting with a point charge and, simultaneously, a uniform electric field is
given by
2
V (u, v, ) = + (u v),
u+v
while the kinetic energy takes the form

m u+v 2 u+v 2
T (u, v, , u, v, ) = u + v + 4uv .
2
8 u v
(a) Calculate the canonical momenta pu , pv , and p conjugate to u, v, and
, and show that the Hamiltonian of the system can be written
  
1 2 p2 1 1
H= (upu + vpv ) +
2 2
+ 2 + (u v ) .
2 2
u+v m 2m u v
96 Integrable systems

(b) Show that the function


2upu2 p2
K = + + u 2 H u
m 2mu
2vpv2 p2
= + + v 2 + H v
m 2mv
Poisson-commutes with both p and H , and is independent of them
almost everywhere in phase space.
(c) Show that the system is not only integrable, but also separable in this set
of coordinates.
4
Canonical perturbation theory

Canonical perturbation theory provides a systematic pathway for going beyond the
highly constrained world of complete integrability while retaining the benefits of
the canonical formalism. The use of such techniques in celestial mechanics has
led to an impressive level of predictability in the motion of massive bodies in the
Solar System, and an ability, aided by enormously powerful computers, to simu-
late the history of that system billions of years into the past. Such techniques also
have important applications in atomic and molecular physics, notably in the semi-
classical regime, where the non-zero size of Plancks quantum of action can safely
be ignored.1

4.1 General approach


The perturbative approach begins with approximating a given system by an
integrable one described by action-angle coordinates (1 , . . . , n , J1 , . . . , Jn ) on
n-dimensional tori, with Hamiltonian H0 (J ). We think of the original system as
inhabiting the same phase space, with a Hamiltonian of the form

H (, J ) = H0 (J ) + H (, J ), = (1 , . . . , n ), J = (J1 , . . . , Jn ),

In the perturbative approach, we solve the equations of motion for (t) and J (t)
as formal power series in the parameter  up to some desired order. To do this,
make a succession of canonical transformations,

(, J ) = ( (0) , J (0) )  ( (1) , J (1) )  ( (2) , J (2) )  ,


1 One could argue that the most notable contribution of canonical perturbation theory was its unexpected failure,
within the framework of BohrSommerfeld quantization, to account for the observed energy levels of the
helium atom. According to Heisenberg [19], this was one of the motivating factors in his decision to introduce
the radically different theoretical framework of quantum mechanics. Ironically, a more recent version of semi-
classical quantization [20] doesnt do such a bad job of accounting for the helium spectrum!

97
98 Canonical perturbation theory

implemented by generating functions

F2(k1) ( (k1) , J (k) ) = (k1) J (k) + S (k1) ( (k1) , J (k) )

such that the level-k Hamiltonian has the form

H (k) ( (k) , J (k) ) = H0(k) (J (k) ) + O( k+1 ).

The level-k actions J (k) are then constants of the motion, up to corrections of order
 k+1 , and the level-k angle variables increase at a constant rate (up to corrections of
order  k+1 ). The level-k dynamical problem is thus well approximated by uniform
motion on an n-torus once one has transformed the initial data to the kth level. By
inverting the succession of canonical transformations, one can then extract (t) and
J (t) correct to order  k .
As we shall see, the above construction can almost always be carried out at the
level of formal power series in . Whether the process converges is, of course, a
difficult question, and we know that the answer cannot always be positive. In par-
ticular, the phase space of a typical Hamiltonian system does not decompose into
invariant tori. Some tori, on which the motion is quasiperiodic, may be present,
but quite generally, we expect such motions to be interspersed with less regular,
perhaps chaotic, ones. What is truly remarkable is that, under quite general con-
ditions, if the perturbation is small enough, then the perturbation expansion does
indeed converge to quasiperiodic motions on tori, for most initial conditions. This
is the content of the famous KAM theorem, which we will discuss at some length
in the next chapter.

4.2 Simple pendulum revisited


An instructive illustration of the perturbative method is provided by the familiar
simple pendulum, considered as a perturbed one-dimensional harmonic oscillator.
This choice is somewhat artificial, in the sense that, like all Hamiltonian systems
with a single degree of freedom, it is integrable, and so action-angle variables
can be constructed directly, without the complication of perturbation expansions.
On the other hand, this example reveals much about how the formalism works,
and provides a convenient starting point for more ambitious examples with more
than one degree of freedom. In addition, the convergence of the series in this inte-
grable example will allow us to perform a comparison with exact results obtained
in Chapter 1.
We begin by writing the pendulum Hamiltonian as

1 2 1 1 2
H(q, p) = ( p + q ) 2
cos( q) 1 + q .
1/2
2 2
4.2 Simple pendulum revisited 99

Note that the unperturbed Hamiltonian is just that of a harmonic oscillator, and the
quantity in parentheses is a power series in q 2 :
1 1 1 2 6
H(q, p) = ( p 2 + q 2 ) q 4 +  q + O( 3 ).
2 24 720
To apply canonical perturbation theory, we introduce, via a canonical transforma-
tion, the action-angle variables for the unperturbed oscillator:

q = 2J cos , p = 2J sin , (4.1)
so that the Hamiltonian becomes
1 1
H (, J ) = J  J 2 cos4 +  2 J 3 cos6 + O( 3 ).
6 90
Using trigonometric identities, this reduces to the Fourier series

1 cos(2) cos(4)
H (, J ) = J  J 2
+ + (4.2)
16 12 48

1 cos(2) cos(4) cos(6)
+ 2 J 3 + + + + O( 3 ).
288 192 480 2880
We now make a second canonical transformation, with a type-2 generating
function,
F2(0) (, J1 ) = J1 + S (0) (, J1 ).
With the substitutions
S (0) S (0)
J = J1 + , 1 = + ,
J1
the Hamiltonian becomes

(1) S (0) 1 1 1
H = J1 +  J1
2
+ cos(21 ) + cos(41 ) + O( 2 ).
16 12 48
Since we want the variable J1 to have the status of a first-order action, we choose
S to cancel the first-order oscillatory terms in the Fourier expansion of H (1) , so
(0)

that we will be left with


1
H (1) = J1  J12 + O( 2 ).
16
The generating function S (0) is thus the negative of the indefinite integral of the
O() oscillatory terms of H (1) , with substituted for 1 . This is equivalent to
replacing each cos(n1 ) by (1/n)sin(n), and each sin(n1 ) by (1/n)cos(n).
This gives us

(0) 1 1
S (, J1 ) =  J1 2
sin(2) + sin(4)
24 192
100 Canonical perturbation theory

and the corresponding transformation equations,



S (0) 1 1
= 1 = 1  J1 sin(2) + sin(4) , (4.3)
J1 12 96

S (0) 1 1
J = J1 + = J1 +  J1
2
cos(2) + cos(4) . (4.4)
12 48
Note that the right-hand sides of these formulas involve both 1 and , and so they
must be regarded as implicit equations from which one can generate power-series
expansions in  to any desired order by means of iterative substitution.
We next extend our analysis to second order in , making a canonical transfor-
mation (1 , J1 )  (2 , J2 ) via a generating function

F2(1) (1 , J2 ) = 1 J2 + S (1) (1 , J2 ),

where S (1) is second order in  and specifically tailored to eliminate the second-
order angle dependence of the Hamiltonian. After making the canonical transfor-
mation, the new Hamiltonian will be
1 S (1)
H (2) (2 , J2 ) = J2  J22 +
16 1

1 5 cos(21 ) 43 cos(41 ) cos(61 ) cos(81 )
 J2
2 3
+ + + +
256 576 5760 320 2304
+O( ).
3

Once again we can get term-by-term cancellation by constructing S (1) as the


negative of the integral over the O( 2 ) oscillatory terms in H (2) :

(1) 2 3 5 sin(21 ) 43 sin(41 ) sin(61 ) sin(81 )
S (1 , J2 ) =  J2 + + + .
1152 23 040 1920 18 432

With this choice, H (2) assumes the quasi-action-angle form


1 1 2 3
H (2) (2 , J2 ) = J2 J2  J2 + O( 3 ).
16 2 256
It should now be clear how we can proceed systematically to arbitrarily high
order: once we have calculated the mth Hamiltonian H (m) (m , Jm ), we construct
S (m) (m , Jm+1 ) in order to cancel out the O( m+1 ) oscillatory terms. After making
the substitutions
S (m)
Jm  Jm+1 + (m , Jm+1 )

4.2 Simple pendulum revisited 101

and, recursively,

S (m)
m  m+1 (m , Jm+1 )
Jm+1

in H (m) (m , Jm ), followed by power-series expansion in  and Fourier-series


expansion in m+1 , we arrive at the next Hamiltonian, H (m+1) (m+1 , Jm+1 ). By
means of an efficient Mathematica program, we have automated this process to
calculate E(J ) = H0(n) (J ) up to n = 10:

1 1 2 3 5 3 4 33 63
E(J ) = J J2  J  J 4 J 5 5 J 6
16 256 8192 262 144 2 097 152
527 9387 175 045
6 J 7 7 J 8 8 J 9
67 108 864 4 294 967 296 274 877 906 944
422 565 4 194 753
 9 J 10  10 J 11 . (4.5)
2 199 023 255 552 70 368 744 177 664

The oscillation frequency is then

dE
(E) = (J (E)),
dJ
where J (E) is obtained by perturbative inversion of (4.5). We get

1 5 2 11 3 469 1379
(E) = 1 E E E E4 E5
8 256 2048 262 144 2 097 152
17 223 56 001 11 998 869
E6 E7 E8
67 108 864 4 294 967 296 274 877 906 944
41 064 827 571 915 951
E9 E 10 . (4.6)
2 199 023 255 552 70 368 744 177 664

An advantage of the simple pendulum is that we can calculate the frequency


exactly as an elliptic integral, namely

(E) = K (E/2).
2
Tenth-order Taylor expansion gives precisely (4.6). To exhibit the rate of con-
vergence, we plot in Figure 4.1 the perturbative approximations (n) (E), n =
0, 1, 2, . . . , 10, as well as the exact (E). We note that the rate of convergence
becomes increasingly slow as one approaches the separatrix.
102 Canonical perturbation theory

n=
Simple harmonic oscillator
1.0 0

exac
Angular frequency

0.8 t
1
0.6
2
3
0.4 4
5
6
0.2 7
separatrix 8
9
10
0.5 1.0 1.5 2.0 2.5
Energy E
Figure 4.1 Plots of a pendulums oscillation frequency as a function of its
energy E. Shown are the exact function together with the 11 lowest perturba-
tive approximations. On our dimensionless energy scale, E = 0 at the stable
equilibrium and E = 2 at the unstable equilibrium.

4.3 Two harmonic oscillators with quartic coupling


We consider a pair of harmonic oscillators with a coupling term of degree 4. The
Hamiltonian is
1 1
H = H0 +  H1 = ( p12 + p22 ) + (12 q12 + 22 q22 ) + 12 22 q12 q22 .
2 2
As in the previous example, our first step is a canonical transformation to action-
angle variables for the unperturbed oscillators:

2Ji
qi = cos i , pi = 2Ji i sin i , i = 1, 2, (4.7)
i
so that the Hamiltonian becomes
K (1 , 2 , J1 , J2 ) = 1 J1 + 2 J2 + 41 2 J1 J2 cos2 1 cos2 2 .
Following the general strategy of canonical perturbation theory, we now make a
canonical transformation to new variables (, J ), J(, J ), via a type-2 generating
function
F2 (, J) = 1 J1 + 2 J2 +  S(, J),
4.3 Two harmonic oscillators with quartic coupling 103

and transformation equations


S S
Jk = Jk +  , k = k +  , k = 1, 2,
k Jk
with the aim of annihilating the angle-dependent part of the first-order terms of the
new Hamiltonian,
S S
K = 1 J1 + 2 J2 + 1 + 2
1 2

1
+ 1 2 J1 J2 1 + cos(21 ) + cos(22 ) + cos(21 + 22 )
2

1
+ cos(21 22 ) + O( 2 ).
2
It is fairly obvious how to determine the correct form of S. We simply take
the O() oscillatory part of the Hamiltonian and replace each cos(r 1 + s2 ) by
(1/(r 1 + s2 )) sin(r 1 + s2 ), to obtain

sin(21 ) sin(22 ) sin(21 + 22 )
S(, J ) = 1 2 J1 J2 + +
21 22 4(1 + 2 )

sin(21 22 )
+ .
4(1 2 )
The transformed Hamiltonian retains the non-oscillatory first-order terms,

K = 1 J1 + 2 J2 + 1 2 J1 J2 + O( 2 ).

To first order in , the perturbed oscillation frequencies are thus

K K
1 = = 1 + 1 2 J2 , 2 = = 2 + 1 2 J1 .
J1 J2
We are now in a position to write down the full canonical transformation in first
order
S S
Jk = Jk  (, J ), k = k +  (, J ), k = 1, 2, (4.8)
k Jk
with

cos(2 ) cos(2 + 2 ) cos(2 2 )
J1 = J1 + 1 2 J1 J2
1 1 2 1 2
+ + ,
1 21 + 22 21 22

cos(22 ) cos(21 + 22 ) cos(21 22 )
J2 = J2 + 1 2 J1 J2 + ,
2 21 + 22 21 22
104 Canonical perturbation theory

 sin(21 ) sin(22 ) sin(21 + 22 ) sin(21 22 )
1 = 1 1 2 J2 + + ,
2 1 2 21 + 22 21 22

 sin(21 ) sin(22 ) sin(21 + 22 ) sin(21 22 )
2 = 2 1 2 J1 + + ,
2 1 2 21 + 22 21 22
and thereby solve the initial-value problem.
Suppose we are given q1 (0), q2 (0), p1 (0), and p2 (0), and asked to find
q1 (t), q2 (t), p1 (t), and p2 (t) for arbitrary time t, correct to first order in . We
would
(a) transform the initial data to action-angle variables (0), J (0) by inverting (4.7);
(b) via (4.8), transform to J(0), (0);
(c) use Hamiltons equations for K (, J) to obtain

Jk (t) = Jk (0), k (t) = k (0) + t, k = 1, 2;

(d) solve the four equations (4.8) perturbatively to O() to get (t), J (t);
(e) apply the canonical transformation (4.7) to get, finally, q(t), p(t).
The process is perhaps tedious, but completely straightforward.

4.4 Gyrating charge in an electrostatic wave


In Section 2.10 we studied the motion of a charged particle in a plane perpendicular
to a uniform magnetic field. We found a canonical transformation from the origi-
nal rectangular coordinates x and y, with conjugate momenta px and p y , to new
phase-space variables , Y, P , X , of which all but the first are constants, while
increases at the uniform rate c = eB/m, corresponding to the new Hamiltonian,
K = c P . The physical interpretation of the new coordinates is simple:
the par-
ticle moves in a circular orbit with fixed center (X, Y ), radius = 2P /(mc ),
and uniformly increasing azimuthal angle .
We now imagine that the particle is subjected to an electrostatic wave, i.e. a
scalar potential term in the Hamiltonian of the form

e 0 sin(ky t)

with no accompanying changes in the magnetic vector potential. Studying the


effects of such waves is of some interest in the plasma-physics literature, from
which the current example is drawn [21]. In terms of the new phase-space
coordinates, the Hamiltonian is

K = c P + e 0 sin(kY k sin t).


4.4 Gyrating charge in an electrostatic wave 105

The reader may have noticed that K is already in a form to which we can apply
canonical perturbation theory, with an expectation that it will be of some relevance
for sufficiently small wave amplitudes. In particular, the unperturbed Hamiltonian
is in standard action-angle form, except perhaps for the explicit t dependence. The
latter can be removed by means of the canonical transformation (, Y, P , X ) 
(, , P , P ) generated by

F2 (, Y, P , P , t) = P + (kY t)P ,

which gives
F2 F2
= = , = = kY t,
P P
F2 F2
P = = P , PY = = k P ,
Y
and the new Hamiltonian
F2
K = K + = c P P + e 0 sin( k sin ).
t
The next step follows the general pattern of canonical perturbation theory: we
seek a canonical transformation of type 2, from (, , P , P ) to (1 , 2 , J1 , J2 ),
which will kill the oscillatory (w.r.t. ) part of the perturbation term, leaving the
Hamiltonian, to first order in 0 , as a function of action variables alone. With the
generating function

F2 (, , J1 , J2 ) = J1 + J2 + S(, , J1 , J2 ),

the new Hamiltonian, to first order in 0 and S, will be


   
S S 2J1
c J1 + J2 + + e 0 sin sin .
mc

This will reduce to c J1 J2 if we can choose S to satisfy the first-order


differential equation
 
S S 2J1
c + e 0 sin sin = 0.
mc

As usual, the solution drops out easily once everything is expanded in Fourier series
(using identities in Chapter 9 of [3]):
106 Canonical perturbation theory
 

2J1
e 0 sin sin = e 0 Jm (k )sin( m ),
mc m=

Jm (k )
S = e 0 cos( m ),
m=
+ c

where Jm is the Bessel function of order m.


The transformation equations from the original phase-space coordinates to
action-angle variables can now be written down by differentiating the generating
function. For example, to first order,


S 1 Jm (k )
J1 = P = mc + e 0
2
m sin( m ).
2 m=
+ c

For each value of J1 , there is a functional relation of the form f (, , ) = J1 ,


which should typically be the equation of a torus. We can get an idea of how the
various tori are embedded in phase space by plotting a Poincar section at fixed .
We show the results for two values of the wave amplitude in Figure 4.2. It is instruc-
tive to compare our results of first-order perturbation theory with those obtained by
numerical integration of the exact equations. We use as our Hamiltonian

H = P + c P + e 0 sin( k sin ),

where
1
P = mc 2 .
2
We will use H = E to determine P as a function of , , and , determining
those as functions of t via the Hamilton equations (remember, P is the canonical
variable, not )
H ek
= = c sin cos( k sin ),
P mc
H
= = ,
P
H ek
= (mc )1 = cos cos( k sin ).
mc
On introducing the dimensionless variables

e 0 k 2 c +
r = k, a= , b= , = (4.9)
mc 2
4.4 Gyrating charge in an electrostatic wave 107

50

49

48

k
47

46

45
0 1 2 3 4 5 6

50

49

48

k
47

46

45
0 1 2 3 4 5 6

Figure 4.2 First-order perturbative Poincar sections for frequency ratio /c =


30.11 and dimensionless amplitudes a = 0.01 (upper) and a = 0.1 (lower).

and functions

e(, r, ) = cos( + r sin(2 )),


g(, r, ) = 2(b + (a/r )sin(2 )e(, r, ))1 ,
f (, r, ) = a cos(2 )e(, r, )g(, r, ),
108 Canonical perturbation theory

Figure 4.3 Numerical Poincar sections for frequency ratio /c = 30.11 and
dimensionless amplitudes a = 0.01 (upper) and a = 0.1 (lower).

the differential equations for r and as functions of are


dr d
= f (, r, ), = g(, r, ).
d d
Poincar sections at = 0, for the same values of the amplitude a as used
to generate Figure 4.2, were obtained by selecting representative initial points on
the , r plane, integrating the differential equations, and recording the points for
= 0, 1, 2, . . .. The results are displayed in Figure 4.3. For the relatively small
4.5 BirkhoffGustavson perturbation theory 109

amplitude a = 0.01, the motion appears to be quasiperiodic on tori resembling


those of the perturbative calculation. For the larger amplitude a = 0.1, the situation
is quite different. Here it appears that the tori have survived in some portions of the
planar region, but in others have broken up into island chains or chaotic orbits. This
behavior is actually quite typical of Hamiltonian systems as their parameters are
driven further and further from integrability. The case of two degrees of freedom is
special: the tori in the integrable limit are nested, so, when some of them break up
and chaos emerges, the chaotic orbits remain trapped between surviving tori until
they too eventually break up.

4.5 BirkhoffGustavson perturbation theory


We now describe a powerful scheme, developed by Birkhoff [22], which can
be applied to perturbations of harmonic oscillators of any dimension, provided
that their frequencies are incommensurate, i.e. no integer combination of them
vanishes. Then the Hamiltonian may be written (see Section 2.10) as

n
H (a, a ) =

k ak ak + H (m) (a, a ),
k=1 m=3

where H (m) is homogeneous of degree m in the ai and a j , i, j = 1, 2, . . . , n.


The strategy of the Birkhoff expansion is to choose a succession of canonical
transformations T (m) with generating functions
F2(m) (a, b ) = a b + S (m) (a, b ),
with S (m) a homogeneous polynomial of degree m, such that each T (m) kills
all terms in H (m) except for those which are monomials in the quantities ai ai ,
i = 1, 2, . . . , n. It is easy to see that T (m) cannot affect the lower-degree terms in
H , while typically it will modify those of degree higher than m. This is not a prob-
lem, since these terms will be dealt with later in the process. The result, formally, is
to convert H canonically into a power series in the ak ak . Since iak ak is actually the
action variable Jk , we will end up with a power series in the actions. In practice,
we will terminate the process at some finite order M, ending up with a formally
integrable model that, it is to be hoped, provides useful information about the exact
dynamical system (in the form of n approximately conserved quantities).
It remains to specify the homogeneous polynomials S (m) . We associate with a
monomial M containing (in addition to a complex coefficient) Nk (M) factors ak
and Nk (M) factors ak , k = 1, . . . , n the index

n  
d(M) = k Nk (M) Nk (M) ,
k=1
110 Canonical perturbation theory

which is just the eigenvalue of the differential operator


 
n

D= k ak ak .
k=1
a k ak

We construct S (m) (a, a ) from H (m) (a, a ) by first removing all terms M
with d(M) = 0, and replacing each of the remaining M by M/d(M), so
that DS (m) (a, a ) is precisely equal to the negative of the unwanted terms
in H (m) (a, a ). Under the canonical transformation T (m) , the quadratic terms
$
k k ak ak are replaced by
 
S (m) (b, b ) S
(m)
(b, b )
k bk bk = DS (m) (b, b )
k
b k b
k

plus higher-degree terms. On the other hand, H (m) (a, a ) is transformed into
H (m) (b, b ) plus higher-degree terms, and so the cancellation of unwanted terms
takes place.
It was observed by Gustavson [23] that there actually is no impediment to carry-
ing out the sequence of canonical transformations just described in cases in which
there is degeneracy, i.e. where integer linear relations exist among the oscillation
frequencies. Of course, the condition d(M) = 0 no longer restricts the terms of
the resulting power series as tightly, and the number of conserved quantities is
reduced. For example, if all of the frequencies coincide, d(M) = 0 requires only
that the total number of a factors must be equal to the total number of a factors,
with no constraint on the indices. This corresponds to only one conserved quantity.
For n = 2, this is (formally) enough for integrability if the constant of the motion
is independent of H . Gustavson exploited this in calculating a perturbative second
integral for the HnonHeiles model [24], a project we will describe in the final
section of this chapter.

4.6 A formal second integral for the HnonHeiles model


The HnonHeiles Hamiltonian,
1 1
H(q1 , q2 , p1 , p2 ) = ( p12 + p22 + q12 + q22 ) + q12 q2 q23 ,
2 3
was introduced in [24] as an idealization of an astrophysical model describing the
motion of stars in a galaxy under the influence of an average gravitational potential
with axial symmetry. The cubic potential-energy function is also the lowest-degree
approximation to the three-particle Toda potential energy, but does not share with it
the property of integrability. We shall see that even for moderate values of the total
energy, apparently chaotic orbits share the phase space with the usual invariant tori.
4.6 A formal second integral for the HnonHeiles model 111

In this section, we regard the HnonHeiles Hamiltonian as a perturbed har-


monic oscillator and apply the BirkhoffGustavson method to construct a pertur-
bative integral of the motion that is independent of the Hamiltonian. We begin by
transforming to complex canonical variables ai , ai , i = 1, 2:
H (a1 , a2 , a1 , a2 ) = ia1 a1 + ia2 a2
 3/2 
i 2 1 3
+ (a1 + a1 ) (a2 + a2 ) (a2 + a2 ) .
2 3
As described in the previous section, we are to perform a sequence of canonical
transformations T (m) , m = 3, 4, . . ., each designed to kill those monomials M
in the transformed Hamiltonian of degree m in a1 , a2 , a1 , a2 which do not satisfy
D M = 0, where
 
2

D= i ak ak .
k=1
ak ak
After performing these transformations up to m = N , we have new coordinates
(b, b ) = T (N ) T (N 1) T (3) (a, a )
such that the transformed Hamiltonian satisfies
D H (N ) (b, b ) = 0
and as a consequence
[ib1 b1 + ib2 b2 , H (N ) (b, b )] = 0.
Thus, in the (b, b ) coordinates, we have a conserved quantity that (by direct
check) is independent of H (N ) (b, b ). To rewrite this integral of the motion
in the original (a, a ) coordinates, we simply invert the sequence of canonical
transformations:
 
I = T (3)1 T (N )1 ib1 b1 + ib2 b2 .

A final transformation,
1 1
ak = ((qk pk ) i(qk + pk )), a = ((qk + pk ) i(qk pk )),
2 2
returns us to our original coordinates.
With the help of Mathematica, we have carried out Gustavsons construction to
order N = 10. Rather than writing out the formula for I , which occupies several
printed pages, let us apply it to construct Poincar sections for several values of
the energy. These are displayed in Figure 4.4. There is an increase of complexity
as we increase the energy, but of course no sign of the chaos which we expect to
112 Canonical perturbation theory

0.3 E = 1/24 0.4 E = 1/12

0.2
0.2
0.1

0 0

-0.1
-0.2
-0.2

-0.3 -0.4

-0.3 -0.2 -0.1 0 0.1 0.2 0.3 -0.4 -0.2 0 0.2 0.4

0.4
E = 1/8 0.6 E = 1/6

0.4
0.2
0.2

0 0

-0.2
-0.2

-0.4
-0.4
-0.6

-0.4 -0.2 0 0.2 0.4 0.6 -0.5 -0.25 0 0.25 0.5 0.75 1

Figure 4.4 Poincar sections (x = 0) of the HnonHeiles model, for four values
of the energy, calculated in tenth-order perturbation theory using Gustavsons
method.

find for the exact solutions of the dynamical equations. We have explored the latter
numerically, with the results shown in Figures 4.5 and 4.6.

4.7 Integrability analysis of HnonHeiles-like systems


L. S. Hall [13] investigated the entire family of dynamical systems with Hamilto-
nians of the form (generalizing that of Hnon and Heiles [24])
1 2 1
( p1 + p22 + Aq12 + Bq22 ) + Cq12 q2 + Dq23
2 3
in an attempt to find exact and approximate integrals of the motion of degree less
than or equal to 4. Exact integrals were found in the cases
4.7 Integrability analysis of HnonHeiles-like systems 113

py

x
E = 1/8

e
an
pl
0
x=

Figure 4.5 Some orbits of the HnonHeiles model, with energy E = 1/8, pro-
jected into the x, y, p y 3-space. In the lower right-hand picture, the space has
been sliced at x = 0 to show the Poincar section of a quasiperiodic orbit.

(a) 16A = B, 16C = D,


(b) A = B, C = D,
(c) C = 0,
(d) 6C = D.
Using Mathematica, we investigated numerically two members of the family,
with Hamiltonians
1
H (1) = ( p12 + p22 + q12 + q22 ) + q12 q2 + 2q23 ,
2
1
H (2) = ( p12 + p22 + q12 + q22 ) + q12 q2 q23 .
2
Figure 4.6 A portion of the x = 0 Poincar section of the HnonHeiles model for E = 1/8. The picture on the right is a
closer look at one of the islands in the left-hand picture.
Exercises 115

E = 0.018 E = 0.0185 E = 0.02


0.2 0.2 0.2
0.1 0.1 0.1
0.0 0.0 0.0
0.1 0.1 0.1
0.2 0.2 0.2
0.2 0.2 0.2
0.1 0.1 0.1
0.0 0.0 0.0
0.1 0.1
0.1
0.2
0.2 0.2 0.2 0.1 0.0 0.1 0.2 0.3 0.4 0.5
0.2 0.1 0.0 0.1 0.2 0.3 0.4 0.5 0.2 0.1 0.0 0.1 0.2 0.3 0.4 0.5

Figure 4.7 The change of topology for the constant-energy manifold of the H (2)
model.

From Halls results, we know that the first of these models is integrable, while the
second may, or might not, be. In both cases, if the energy E is less than a cer-
tain threshold value, there exists a compact, connected component of the energy-E
manifold which can be represented as a pair of three-dimensional regions with their
bounding surfaces (topologically equivalent to 2-spheres) identified. This is the
same as in the HnonHeiles model itself. The change of topology as one traverses
the threshold is shown in Figure 4.7 for H (2) .
To search for evidence of bounded but chaotic orbits, we concentrated our
numerical experiments on energies just below the threshold for escape. Using
fourth-order RungeKutta integration, we calculated Poincar sections, plotting
orbit points on the q2 , p2 plane for q1 = 0. As expected, the H (1) model shows only
quasiperiodic motion on tori, whereas in the case of H (2) some chaos, restricted to
rather narrow layers surrounding separatrices, is in evidence.
For comparison, we calculated Poincar sections for the same energy val-
ues using eighth-order Gustavson perturbation theory. In both cases, the sections
through non-chaotic tori (quasiperiodic motion) are accounted for quite well by the
perturbative model, at least if one stays away from the boundary.
Our results are displayed in Figures 4.8 and 4.9.

Exercises
4.1 A perturbed rigid rotor has the Hamiltonian
L2
+ a sin2 sin(t).
H (, L , t) =
2I
(a) Find a canonical transformation to new variables Q, P that produces a
new Hamiltonian of the form K (P) + O(a 2 ), i.e. up to first order in the
perturbation parameter a the Hamiltonian becomes time-independent and
a function only of the new momentum variable. Determine P(, L , t),
and K (P), keeping only terms up to first order in a.
(a) (b)
p2 p2
0.10

0.05
0.05

0.00 0.00

0.05
0.05

0.10
q2 q2
0.15 0.10 0.05 0.00 0.05 0.15 0.10 0.05 0.00 0.05 0.10

Figure 4.8 Poincar sections (x = 0) of the H (1) model, for E = 0.00462, calculated by (a) numerical integration of the ODEs
and (b) eighth-order canonical perturbation theory (Gustavson method).
(a) (b)
0.2
p2 0.3
p2

0.2
0.1

0.1

0.0 0.0

0.1

0.1

0.2

q2 q2
0.2 0.3
0.1 0.0 0.1 0.2 0.3 0.2 0.1 0.0 0.1 0.2 0.3

Figure 4.9 Poincar sections (x = 0) of the H (2) model, for E = 0.01849, calculated by (a) numerical integration of the ODEs
and (b) eighth-order canonical perturbation theory (Gustavson method).
118 Canonical perturbation theory

(b) For what rotation frequencies P/I do the first-order perturbation formu-
las of part (a) blow up due to resonances? If the canonical perturbation
expansion is carried to higher orders, which rotation frequencies P/I are
expected to lead to infinities?
4.2 A particle moves in a double-well potential with Hamiltonian (using simpli-
fied units)
px2
H (x, px ) = x 2 + x 4.
2 2 4
(a) Draw a phase portrait for the system, locating in the two-dimensional
phase space all equilibrium points, both stable and unstable, and sketch-
ing a few of the oscillatory orbits.
(b) Make a canonical transformation from (x, px ) to (q, p), where q is the
position relative to one of the stable equilibrium points. Show that, with
respect to the new coordinates, the Hamiltonian has the form

H (q, p) = H0 (q, p) + H1 (q),

where H0 is a harmonic-oscillator Hamiltonian, and H1 is a polynomial


in q with cubic and quartic terms.
(c) Make a canonical transformation (q, p) (, J ), where and J are the
action-angle variables for the unperturbed motion generated by H0 . Write
the new Hamiltonian in the form

H (, J ) = J + H1 (, J ).

(d) Consider H1 as a nonlinear perturbation on the simple harmonic motion


about stable equilibrium. In canonical perturbation theory, we write H =
J +  H1 and make a canonical transformation (, J ) (, J), which
leads to a new Hamiltonian of the form

H = J +  H1 ( J) + H2 (, J, ),

where H2 = O( 2 ). Without explicitly working out the transformation,


calculate the oscillation frequency as a function of J, correct to first order
in the perturbation.
4.3 A two-dimensional nonlinear oscillator has the Hamiltonian
1 1
H (q1 , q2 , p1 , p2 ) = ( p12 + q12 ) + ( p22 + q22 ) + p12 q22 .
2 2
(a) Introduce, by means of a (complex) canonical transformation, new
variables Q 1 , Q 2 , P1 , P2 such that
Exercises 119

(i)

1 pk
Q k = qk + , k = 1, 2.
2 ik
(ii) Pk is a homogeneous linear function of qk and pk .
(iii) The new Hamiltonian K (Q 1 , Q 2 , P1 , P2 ) takes the form
K = i1 Q 1 P1 + i2 Q 2 P2 + V (Q 1 , Q 2 , P1 , P2 )
Calculate Pk .
(b) We now introduce, via a second canonical transformation, new canonical
variables k , Jk , k = 1, 2, such that
Jk = i Q k Pk , k = 1, 2,
and each k is an angular variable, i.e. a real number in the interval
[0, 2). Express qk , pk , k = 1, 2 in terms of the transformed variables.
Calculate the new Hamiltonian.
(c) By means of a final canonical transformation, introduce the perturbative
action-angle variables k , Jk , k = 1, 2, such that the new Hamiltonian
takes the form
H (1 , 2 , J1 , J2 ) = H0 ( J1 , J2 ) + O( 2 ).
Calculate the oscillation periods to first order in .
(d) It would seem that the technique of part (c) could be repeated again and
again to remove the angular dependence to any desired order. Does this
imply that the model is integrable? Explain briefly.
4.4 After making a complex canonical transformation, the Hamiltonian of a
perturbed one-dimensional harmonic oscillator has the form
H (a, a ) = ia a + (a + a )4 .
Use Birkhoff-type perturbation theory to find a canonical transformation from
a, a to A, A such that the new Hamiltonian has the form
K (A, A ) = i A A + (A A)2 + ,
where the omitted terms on the right-hand side are of degree greater than 4
in A, A . Express A and A as cubic polynomials in a, a , ignoring terms of
higher degree.
4.5 A system with two degrees of freedom has a Hamiltonian


H (q1 , q2 , p1 , p2 ) = 1 p1 + 2 p2 +  h k1 k2 ( p1 , p2 ) ei(k1 q1 +k2 q2 ) ,
k1 ,k2 =
120 Canonical perturbation theory

where 1 and 2 are positive constants. Derive a perturbative canonical


transformation
qi = i +  f i (1 , 2 , J1 , J2 ) + O( 2 ),
pi = Ji + gi (1 , 2 , J1 , J2 ) + O( 2 ),
such that the new Hamiltonian has the form
K (1 , 2 , J1 , J2 ) = 1 J1 + 2 J2 +  K 1 (J1 , J2 ) + O( 2 ).
Obtain the functions f i and gi as Fourier series. For which parameter values
1 and 2 do we run into convergence problems? Calculate the oscillation
frequencies of the system as functions of J1 and J2 , correct to first order in .
5
Order and chaos in Hamiltonian systems

In the preceding chapters, we studied integrable systems and their perturbations.


We noted that integrability is rare among dynamical systems, and that, while
the perturbative approach is quite successful in any finite order, the perturbation
series cannot be counted on to converge in the generic case. As we shall soon
see, the perturbative convergence problem can be overcome if the perturbation
is small enough and certain other hypotheses are satisfied, thanks to the famous
theorem of Kolmogorov, Arnold, and Moser (KAM) [26, 27, 28]. There are sev-
eral approaches (none of them easy!) to the statement and proof of this theorem.
In this chapter we will rely mainly on that of [28]. A helpful discussion of the
theorem, without detailed proofs, can be found in [29].
Perhaps the main message of the KAM theorem is that if we label the invari-
ant n-tori of the unperturbed integrable model by the n oscillation frequencies
1 , . . . , n , and if the perturbation is weak enough, then a fraction, arbitrarily close
to unity, of the tori will be preserved. This is the main result concerning order
in Hamiltonian systems. No comparably strong statement exists concerning what
replaces those tori which break up under the perturbation. Here we rely mainly
on numerical investigations in a variety of models. These suggest certain universal
features, principally island chains and deterministic chaos.
In the present chapter we will introduce the KAM theorem in the context of non-
linear stability of equilibrium states. Having completed our rsum and discussion
of the theorem, we will apply it to a familiar example of nonlinear stability, namely
that of the L 4 and L 5 points in the restricted three-body problem with circular
orbits, with application to the SunJupiterTrojan-asteroid system. The reader will
recall that we have already established the linear stability of the system for realistic
mass ratios.
As an introduction to chaos, we will turn, in the last part of the chapter, to a
simple kicked-oscillator model for which much can be learned analytically as well
as by numerical exploration. In addition, we will study a simple model for one

121
122 Order and chaos in Hamiltonian systems

of the most clear-cut examples of chaos in Solar System dynamics, namely the
erratic rotational motion of Saturns seventh moon, Hyperion. The model turns out
to be too restrictive to accurately model Hyperions motion, but it does provide
a nice example of the complex interplay between regular and chaotic motion in
Hamiltonian systems.

5.1 Nonlinear stability


In the neighborhood of a linearly stable, or elliptic, equilibrium point, the Taylor
series of the Hamiltonian (in relative coordinates) reduces to that of an n-
dimensional harmonic oscillator. It can be shown that this is not sufficient to
guarantee that the system, if perturbed slightly, will always remain close to equilib-
rium. The latter property, called nonlinear stability, or Lyapunov stability, is defined
formally as follows: an elliptic equilibrium point P is called Lyapunov stable if, for
every  > 0, there exists > 0 such that every orbit initially within a distance of
P remains forever within a distance  of P.
For an integrable model with compact constant-energy manifolds, each non-
equilibrium orbit lies on a torus (the LiouvilleArnold theorem), and, in suitable
coordinates, takes the form of a direct product of circles of radii ai , i = 1, 2, . . . , n.
The nonlinear stability is obvious. For non-integrable systems, the stability of ellip-
tic equilibrium points is not at all obvious. For systems with at most two degrees of
freedom, thanks to the analysis of KAM, the question is relatively easy to answer
through calculation of a low-dimensional determinant. For higher-dimensional sys-
tems, the KAM theorem guarantees lots of bounded orbits in the neighborhood of
the equilibrium point, but is not sufficient to establish nonlinear stability.

5.2 The KAM theorem


In this section we study the KAM theorem as restricted to the question of non-
linear stability of elliptic fixed points of a Hamiltonian dynamical system. With no
attempt at completeness, we rely here on V. I. Arnolds masterful 1963 monograph
[28], to which the interested reader is referred for further details. We assume that
our Hamiltonian, via canonical transformations, has already been put in the form
H (, J ) = H (1 , . . . , n , J1 , . . . , Jn ) = H0 (J ) + H1 (, J ),

|Ji /2| /2, 0 i < 2, i = 1, . . . , n,


where

n
n
H0 (J ) = i Ji + i j Ji J j ,
i=1 i, j=1
5.2 The KAM theorem 123

with 1 , . . . , n incommensurate and det(i j ) = 0, and




H1 (, J ) = h k (J )eik ,
k1 ,...,kn =

with
|H1 | < C|J |5/2 , |h k | < Me|k| , |k| = |k1 | + + |kn |.
Note that the Hamiltonian is analytic on the domain
{(, J ) : |Ji | , |Im i | < , i = 1, . . . , n},
consistently with the exponential fall-off of the Fourier amplitudes.
The unperturbed system is integrable, with action-angle variables J and and
Hamiltonian H0 (J ). The phase space decomposes into the equilibrium point at
J = (0, . . . , 0) and tori T(J ) . On the torus labeled by n constant Ji values, the
frequencies are
H0
i (J ) = = i + 2 i j J j .
Ji ij

The nonvanishing of the Hessian determinant


 2
  H0
det i j =
Ji J j
means that we can make a smooth change of coordinates (diffeomorphism) A from
J1 , . . . , Jn to 1 , . . . , n , as sketched in Figure 5.1.
Proceeding as in traditional canonical perturbation theory, we make a canonical
transformation to new variables  and J  which cancels out the oscillatory terms
$n
h k (J ) eiki in the Hamiltonian up to |k| = i=1 |ki | = N 1, producing, one
hopes, a new Hamiltonian
H  (  , J  ) = H0 (J  ) + H1 (  , J  )

J2
T
dA(J)
dJ A

1 J = A1
J1

Figure 5.1 Given a frequency vector , there is a unique n-tuple of actions J =


A1 corresponding to a unique n-torus T . The mapping A is a diffeomorphism.
124 Order and chaos in Hamiltonian systems

with |H1 | < |H1 | for N large enough. In the proof of the KAM theorem, one might
hope to repeat this process infinitely many times, in such a way that the sequence
of tori T(J ) , T  (J  ) , T (J  ) , . . . tends to a limiting torus T for a non-zero fraction
of phase space.
A major obstacle to realizing this program is encountered already in the first
step. The perturbative generating function designed to cancel out the oscillatory
terms up to |k| = N 1 is, from Chapter 4,

F2 (, J  ) = J  + S(, J  ),

where
|k|=N 1
i h k (J  ) ik H0
S(, J  ) = e , i (J ) = , (5.1)
|k|=1
k (J  ) Ji

so that
S S
Ji = Ji + (, J  ), i = i + (, J  )
i Ji
and

H  (  , J  ) = H0 (J  ) + H1 (  , J  ), H0 (J  ) = H0 (J  ) + h 0 (J  ).

In (5.1) we have assumed that the frequency vector (J  ) does not lie on any
of the (n 1)-dimensional hyperplanes k = 0, where k runs over all integer
n-vectors with |k| N 1. Since there are only finitely many of these hyperplanes,
each of measure zero, one might think that they would not be a serious problem.
Unfortunately, that is not the case. In higher orders, all Fourier terms will eventually
come into play, so that all integer n-vectors k lead to singular hyperplanes, and
these are dense in frequency space. Thus, arbitrarily small changes in J will force
(J ) to cross a singular hyperplane.
One of Kolmogorovs key insights was that this small-denominator problem can
be defeated if one modifies the recursive process so that a torus is labeled by a
fixed frequency vector, rather than a fixed J vector, as one proceeds from one
perturbative step to the next. One can then banish from frequency space a small
slab (resonance zone) (see Figure 5.2)
K
|k | ,
|k|n+1
surrounding each rational hyperplane, where K is an arbitrary positive parameter.
If we now choose K small enough, the total volume occupied by the infinitely
many resonance zones can be made as small a fraction of the total volume as we
like (compare Exercise 5.1).
5.2 The KAM theorem 125

k = (2,4)

k = (3,1)

Figure 5.2 A sketch of some resonance zones for n = 2.

To make the above observation more precise, we restrict ourselves to A[0, )n ,
the portion of frequency space diffeomorphic to the n-dimensional cube of side ,
such that the volume occupied by the resonance zone labeled k is bounded above
by the width of the zone multiplied by L n1 for some constant L. Then, the
n-dimensional volume occupied by all of the resonance zones is bounded above by
|k|1 |k|=m
K 2 n K L 1
L n1
= C n
K C n K ,
k
|k| n+1
m=1 k
m n+1
m=1
m 2

where C and C  are constants that are independent of K . Since the total volume of
A[0, )n is also proportional to  n , the fraction occupied by resonance zones can
be made arbitrarily small by choosing K small enough.
In the proof of the KAM theorem, the recursive application of perturbative
canonical transformations must be carefully orchestrated to control the following.
1. The size of the phase-space domain. The latter is reduced by elimination of
level-s resonance zones, as well as the replacement of H0(s1) by H0(s) , which
encroaches on the boundaries by an amount s .
2. The inequalities H1(s) < Ms Ms1
1+
, where is positive.
3. The bounds on the diffeomorphism A(s) , namely

s |d J (s) | < |d A(s) | < s |d J (s) |.


126 Order and chaos in Hamiltonian systems

Each canonical transformation leads to a decrease of s and an increase of s ,


by s .
4. The size of the analyticity domain, |Im i | < s . Each canonical transformation
reduces s by s + s .
According to Arnold, the following scaling of parameters leads to a convergent
process for a non-zero fraction of tori:

1 1
s+1 = s
11/10
, Ms = s ,
T
s = s , s = s ,
3
Ns = ln ,
s Ms
where , K , and 1 must be sufficiently small and T sufficiently large, and
2T /5
moreover  < 1 . See Figure 5.3.
For two degrees of freedom, the KAM result implies nonlinear stability. To see
this, consider an initial point (q1 , q2 , p1 , p2 ) in the domain G  where Ji = 12 ( pi2 +
qi2 ) < , i = 1, 2, with H (q1 , q2 ), p1 , p2 ) = E. It can be shown that the frequency
ratio varies nontrivially in the H = E sub-manifold of G  , in such a way that the
sub-manifold is partitioned by a nested family of preserved tori (as sketched in
Figure 5.4). If our chosen point does not lie on one of them, it lies in a gap between
two nested tori, and its orbit is trapped in the gap forever. Since the bounding tori
lie entirely in G  , and since  can be chosen to be arbitrarily small, the nonlinear
stability follows.
In higher dimensions, an n-dimensional torus is incapable of bounding a region
of non-zero volume in the (2n 1)-dimensional constant-energy manifold. Thus
we are unable to infer that an orbit that starts out near a preserved torus will always
stay close to it.
In the above argument for n = 2, the nonvanishing of the determinant of
the coefficient matrix (i j ) played a crucial role. It can be shown [30] that this
hypothesis can be replaced by the determinantal condition

J2
J = A1

J1

Figure 5.3 A sketch of convergence to a preserved torus. The allowed frequency


vector is in the complement to the infinitely many resonance zones and remains
fixed. By means of a carefully controlled sequence of perturbative canonical trans-
formations, the actions approach limiting values, corresponding to convergence to
a limiting torus.
5.3 Nonlinear stability of the Lagrange points 127

Figure 5.4 A sketch of nesting of tori for fixed energy in the case of two degrees
of freedom.


211 212 1
det 221 222 2 = 0 (5.2)
1 2 0
in the proof of stability.

5.3 Nonlinear stability of the Lagrange points


One of the first applications of KAM theory was to establish the nonlinear stability
of the L 4 and L 5 Lagrange points for a range of mass ratios which include those of
the Sun and Jupiter [31], and of the Earth and Moon.1 We begin by expanding the
Hamiltonian in a Taylor series about the linearly stable equilibrium point, keeping
terms up to degree 4:

h(q1 , q2 , p1 , p2 ) = h (2) (q, p) + h (3) (q, p) + h (4) (q, p),

where h (2) is given in (2.22), and



(3)
7 3 3 2 11k 3 3
h (q, p) = 3 q1 + q1 q2 + q1 q2 + q2 , .
2
36 16 12 16
37 4 25k 3 123 2 2 15k 2 2 15k 3 4
h (4) (q, p) = q1 + q1 q2 q1 q2 q1 q2 q1 q23 q .
128 24 64 8 8 128 2
In Chapter 2 we diagonalized the equations of motion of the quadratic Hamilto-
nian and established linear stability for a Jupiter/Sun mass ratio satisfying

mJ mJ 2
27 < 1+ .
mS mS
1 The EarthMoon case is actually not well approximated by the restricted three-body model. The non-uniform
distribution of the Moons mass leads to significant departures from a spherically symmetric gravitational
field.
128 Order and chaos in Hamiltonian systems

We can thus, by means of a standard canonical transformation, introduce the


harmonic oscillator variables a1 , a2 , a1 = a3 , a2 = a4 , transforming the Hamil-
tonian into

K = 1 a1 a1 + 2 a2 a2 + g1 2 3 4 a11 a22 a13 a24
1 +2 +3 +4 =3

+ h 1 2 3 4 a11 a22 a13 a24 + , (5.3)
1 +2 +3 +4 =4

where g1 2 3 4 and h 1 2 3 4 are k-dependent coefficients.


Following Leontovich [31], we now perform additional canonical transforma-
tions, following the algorithm of Section 4.6, to reduce the displayed terms of K
to Birkhoff normal form

1 b1 b1 + 2 b2 b2 + 11 (b1 b1 )2 + 212 (b1 b1 )(b2 b2 ) + 22 (b2 b2 )2 + . (5.4)

The generating function of the first of these transformations will be of the form

F2(3) (a1 , a2 , b1 , b2 ) = a1 b1 + a2 b2 + S (3) (a1 , a2 , b1 , b2 ),

where S (3) consists of the same sum of cubic terms appearing in the Hamiltonian
K , but with each monomial M replaced by M/d(M), where d(M) is the index
of M defined in Section 4.6:
g1 2 3 4
S (3) (a1 , a2 , b1 , b2 ) = a11 a22 b3 b4 .
1 +2 +3 +4 =3
1 (3 1 ) + 2 (4 2 )

This produces the canonical substitutions

S (3) k g1 2 3 4 a11 a22 3 4


ak  bk + = bk + b b ,
ak
( 1 ) + 2 (4 2 ) ak
=3 1 3
1 +2 +3 +4

S (3) k+2 g1 2 3 4 b3 b4
ak  bk = bk a11 a22 .
bk 1 +2 +3 +4
( 1 ) + 2 (4 2 )
=3 1 3
bk

On applying these rules recursively to K (a1 , a2 , a1 , a2 ), and then discarding all


cubic and quartic terms with non-zero index (since these would be cancelled out
in the second step of the Birkhoff reduction) we obtain a new Hamiltonian in the
normal form (5.4), with
5.4 Kicked oscillators 129
3 3 1
11 = g3000 g0030 g0120 g2001 + g0120 g2001
1 1 21 2
1 1
g1110 g1011 g2100 g0021 + h 2020 ,
2 21 + 2
2 2 2
12 = g0210 g1002 g1200 g0012 g2100 g0021
1 22 1 + 22 21 + 2
2 1 1
g2001 g0120 g2010 g0111 g1101 g1020
21 2 1 1
1 1 1
g0201 g1011 g1110 g0102 + h 1111 ,
2 2 2
3 3 1
22 = g0300 g0003 g0102 g0201 g1002 g0210
2 2 1 22
1 1
g1101 g0111 g1200 g0012 + h 0202 .
1 1 + 22
The determinantal condition (5.2) was shown by Leontovich [31] to be satisfied
for a range of mass-ratio values (with some isolated exceptions), thus establishing
the nonlinear stability of the L 4 and L 5 Lagrange points of the restricted three-body
model.
More recent work on the planar, circular restricted three-body problem has
dealt with the stability of more general asteroidal orbits. In particular, Celletti
and Chierchia [32] developed a method for proving the existence of KAM tori
in manifolds with fixed energy, which they applied to the case of an asteroid whose
heliocentric orbit is approximated by an ellipse with significant eccentricity. Using
a computer-assisted proof, they showed that for any mass ratio m J /m S less than or
equal to 103 , there exist two nested KAM tori that trap the motion of the asteroid,
which is assumed to have unperturbed parameters equal to that of the actual aster-
oid 12 Victoria. For readers interested in further exploring the recent literature on
KAM theory and the n-body problem of celestial mechanics, the introduction of
[32] provides an excellent starting point.

5.4 Kicked oscillators


A particularly instructive example of a Hamiltonian system with chaotic orbits is
a one-dimensional harmonic oscillator that is subjected to instantaneous kicks
times per natural period, with a kick amplitude that varies sinusoidally as a function
of position. In suitable units, the Hamiltonian is
1
H (x, y) = (y 2 + x 2 ) + K cos x (t 2k).
2 k
130 Order and chaos in Hamiltonian systems

Hamiltons equations of motion take the form


H
x = = y, (5.5)
y
H
y = = x + K sin x (t 2k). (5.6)
x k

A typical quasiperiodic orbit (for K = 0.8, = 14 ) is shown in Figure 5.5.


Structurally, the orbit consists of a sequence of momentum shifts K sin xi alternat-
ing with angle-2 circular arcs. The sequence of (xi , yi ) is thus specified by the
recursion relation (in fact, the Poincar map)
   
xk+1 xk cos(2) sin(2) yk + K sin xk
=W = .
yk+1 yk sin(2) cos(2) xk
In the left-hand frame of Figure 5.5, the points (xk , yk ) = W k (x0 , y0 ), k =
0, 1, . . . , 17, are assigned dots. At this stage of the iterative process, no particu-
lar pattern has emerged. However, after ten times as many iterations, such a pattern
is very much in evidence for the Poincar section the dots are filling out a quartet
of closed curves while the continuous orbits reveal little more than the fact that
they will eventually provide a dense covering of an annular region of the plane.
A more interesting (in fact, chaotic) orbit is shown in Figure 5.6. Here we have
chosen an initial point (x0 , y0 ) close to (, 0), an unstable period-4 point of the
Poincar map. Since the kick amplitude is almost zero there, it is not surprising
that the first few points in the left-hand frame (17 kicks) are clustered around the

6
6

4 4

2 2

-6 -4 -2 2 4 -6 -4 -2 2 4 6

-2 -2

-4 -4

-6 -6

Figure 5.5 Quasiperiodic orbit and Poincar section for K = 0.8: 17 kick periods
(left) and 170 kick periods (right).
5.4 Kicked oscillators 131

3
10

2
5
1

-10 -5 5 10
-3 -2 -1 1 2 3

-1 -5

-2
-10

-3

Figure 5.6 Chaotic orbit and Poincar section for K = 0.8: 17 kick periods (left)
and 425 kick periods (right).

saddle point. For larger numbers of iterations, as shown in the right-hand frame
(425 kicks), the orbit moves outward on the plane. Figure 5.7 shows the points
visited by the first 10 000 iterations of the fourth-iterate map W 4 . It appears to be
tracing out, in a quasi-random manner, an infinite web-like region with the discrete
translation invariance of a square crystal. Because of the apparent chaotic nature of
such orbits, the region is known as a stochastic web [33].
To study the geometry and dynamics within the stochastic web, it is helpful
to exploit the crystalline symmetry of the model and identify the points of plane
modulo 2 in both coordinates. This folds the original Poincar map into a map
of the 2-torus onto itself. The action of the map on the fundamental domain,

(x, y) (y + K sin x, x) mod 2 ,

is illustrated in Figure 5.8. A phase portrait of the folded map is given in Figure 5.9.
The partition of the square into regular and chaotic components is not nearly as sim-
ple as one might guess from this picture. This becomes clear if one zooms in on a
small neighborhood of the saddle point (Figure 5.10). One sees that a large number
of chains of islands of stability are embedded in the stochastic web. Similarly,
if one examines the region just outside the stochastic web, one also finds chains
of islands, each wrapped in a thin chaotic layer. The intricate interpenetration of
regularity and chaos in the phase space is perhaps not unexpected given the flavor
of the KAM theorem and its proof.
Chaotic orbits within the stochastic web are characterized by two main features
that resemble random processes. The first is a property of the folded map: initially
132 Order and chaos in Hamiltonian systems

Figure 5.7 Part of the stochastic web visited by the first 10 000 iterations of the
fourth-iterate map.

nearby orbits separate from one another, on the average, as et , where  is a
positive constant known as the Lyapunov exponent. We will return to a quanti-
tative treatment of this property of extreme sensitivity to initial conditions later
in this chapter. It is not difficult to imagine, however, that within a compact phase
space, such as that of the folded kicked-oscillator map, the persistent local instabil-
ity of an orbit with positive  would lead, over long times, to a seemingly random
filling of the allowed phase space (as can be seen, for example, in Figure 5.15
later).
The second manifestation of apparent randomness of the chaotic orbits refers
to a property of the orbits in the plane: every time that an orbit of the fourfold
map has a near-collision with one of the saddle points, it either turns right or turns
5.4 Kicked oscillators 133

f e d a f
1
1

0 b 0 e
origin 2
2 a f c d
a b c
1
web map b 0 e mod 2
2
c d

Figure 5.8 Folding the kicked oscillator Poincar map (web map) into the
fundamental cell.

Figure 5.9 The phase portrait of the folded map, for K = 0.8.

left, with seemingly equal probability. For an asymptotically large number of iter-
ations (time) t, the displacement from its initial point of a typical chaotic orbit in
the stochastic web increases as Dt 1/2 , just like a random walker whose behavior
at each saddle point is governed by the flip of an ideal coin. Numerical experi-
ments [34, 35] involving a large statistical ensemble of initial conditions in the
134 Order and chaos in Hamiltonian systems

0.1

0.05

0.05

0.1

0.15
3 3.05 3.1 3.15 3.2 3.25

Figure 5.10 The neighborhood of the saddle point (, 0). In addition to points of
a single chaotic orbit, several nested quasiperiodic orbits within a chain of islands
have been plotted.

stochastic web have determined the parameter dependence of the diffusion coef-
ficient D. Interestingly, those investigations found what appear to be exceptional
values of K for which the diffusion coefficient blows up: for such values there are
special systems of islands that trap the orbits. The displacement then increases as
t , > 12 , a phenomenon known as superdiffusion.
What is the explanation for the apparent randomness exhibited by orbits in the
stochastic web? The phenomenon is a consequence of the behavior of the invariant
manifolds associated with neighboring unstable fixed points (saddle points) of the
qth-iterate map, for = p/q. Each such point has an unstable direction along
which points asymptotically move directly away from the fixed point, with the
separation of successive points increasing by a factor + > 1 with each iteration
of the map. Such points have the property that their inverse orbit approaches the
fixed point in the limit of infinitely many iterations. The collection of all points
with this property is a one-dimensional manifold. Each saddle point also has a
stable manifold, consisting of points whose forward orbits approach the saddle in
the limit. Asymptotically, successive points of such an orbit have a separation that
decreases by a factor < 1 with each iteration. The stretching/contraction factors
are nothing other than the eigenvalues of the linearized map at the fixed point,
and the corresponding eigenvectors are tangent to the stable and unstable manifolds
at the same point.
5.4 Kicked oscillators 135

3.2
P
3
ble ld

un
sta nifo

sta
ma

ble
3.15
2

ma
nif
old
3.1
1

P
0 3.05
0 1 2 3 4 3.05 3.1 3.15 3.2

Figure 5.11 Heteroclinic intersection of invariant manifolds near a hyperbolic


fixed point. The right-hand panel zooms in to a small neighborhood of the fixed
point.

In Figure 5.11, we have plotted part of the unstable manifold of an unstable fixed
point P. The curve extends down the middle of the stochastic web until it reaches
the vicinity of a neighboring fixed point, P . There it oscillates back and forth, each
peak being magnified (asymptotically) relative to its predecessor by a factor + ,
with the separation between successive peaks shrinking by a factor . It seems
that the manifold, faced with the choice of turning right or turning left, manages to
do both simultaneously!
What about the orbit of a single point on the unstable manifold, at a distance 
from P? Such an orbit will do one of three things. If the initial point (and hence
each of its images under the map) is an intersection point between the unstable
manifold of P and the stable manifold of P , a so-called heteroclinic intersection,
the orbit will tend asymptotically toward P along the stable direction, generating
an infinite sequence of heteroclinic intersections. If the initial point lies between
two intersection points but above (below) the stable manifold of P , it will turn left
(right) once it reaches the vicinity of P . We note that which of these alternatives
occurs depends sensitively on . That is because the stable manifold of P near P
oscillates wildly, the mirror image of the unstable manifold of P near P . Thus the
heteroclinic intersections near P have their separations decreasing like k as one
approaches P, so whether the initial point is above or below the stable manifold
depends extremely sensitively on the distance  from P.
A typical chaotic orbit in the stochastic web will sometimes pass very close
to an unstable fixed point. When it does, the odds of it turning left or right
are approximately even, thus giving rise to the random-walk behavior described
136 Order and chaos in Hamiltonian systems

earlier. As remarked then, there are exceptional values of the parameter for which
island-around-island trapping occurs, leading to anomalous diffusive behavior.

5.5 Lyapunov exponents


It is widely agreed among researchers in nonlinear dynamics that the defining
property of deterministic chaos is the property of extreme sensitivity to initial
conditions, i.e. the distance between initially nearby chaotic orbits increases expo-
nentially with time. While this is correct as a rough definition, it clearly needs to be
made more precise. For example, orbits that originate near an unstable equilibrium
point and lie near its unstable manifold would separate at an exponential rate for
a while, but such local instability would not qualify as chaos unless it persisted,
on the average, over the entire orbit. Even then, there is an obvious loophole: in
an unbounded phase space, nearby orbits might fly apart everywhere without the
increase of complexity that we associate with true chaos. For these reasons we
restrict ourselves to systems with compact phase manifolds, and require that the
extreme sensitivity to initial conditions be an average property over the whole orbit.
A precise notion of average sensitivity to initial conditions is provided by the
concept of the Lyapunov exponent. We first define this for a discrete map F( )
on the (compact) phase space. A point close to can be well approximated by a
point + in the tangent space at , and if we are interested in strictly local
expansion effects, we should calculate how the length of the vector changes
under iterated application of the map. Specifically, given initial 0 and 0 , we have
(see Figure 5.12)

k+1 = F(k ), k+1 = D F(k ) k ,

F
F()

DF() .

Figure 5.12 The discrete map F at accompanied by the linear transformation


D F( ) of the tangent space.
5.5 Lyapunov exponents 137

where D F is the linear map of the tangent space defined by the matrix of partial
derivatives ( Fi / j )( ), i, j = 1, 2, . . . , n. Note that, under iteration of F, the
successive D F(k ) are simply multiplied as matrices:

D F k (0 ) = D F(k1 ) D F(1 ) D F(0 ).

The average expansion rate over the orbit starting at point is defined as

1 | D F k ( ) |
( ) = max lim ln .
k k ||

In many cases the eigenvalues and eigenvectors of D F k ( ) have well-defined limits


for k , so ( ) is just the largest of the eigenvalues. Empirically, one often
finds that ( ) is independent of almost everywhere within a region of phase
space of non-zero measure (in other words, the strength of the chaos in that region
reduces to a single positive number, the Lyapunov exponent).
For a continuous-time dynamical system with equations of motion = f ( ),
the time evolution may be regarded as the t 0 limit of the discrete map
def
x(t + t) = F((t)) = (t) + f ((t))t.

From our discussion of discrete maps, the corresponding evolution of a tangent-


space vector is

(t + t) = D F((t)) = (t) + D f ((t)) (t)t,

which, in the continuous-time limit, yields the differential equation


(t) = D F((t)) (t).

The Lyapunov exponent may then be calculated as



1 | (t) |
( ) = max lim ln .
(0) t t | (0) |
Once again, one typically finds (for Hamiltonian systems sufficiently far from
integrable) chaotic regions of phase space where almost every orbit has the same
positive Lyapunov exponent. Associated with such a Lyapunov exponent is a
Lyapunov time,
1
TLyapunov = ,

during which reasonably reliable predictions are possible. Beyond this time scale,
predictive power is rapidly lost.
138 Order and chaos in Hamiltonian systems

5.6 Web-map Lyapunov exponents


The fourfold stochastic web map (Poincar map of a kicked harmonic oscillator)
introduced in Section 5.4 provides a simple testing ground for the notion of the
Lyapunov exponent as a measure of chaos. In Figures 5.135.17 we see the results
of a few numerical experments for small, moderate, and large values of the parame-
ter a. The last example exhibits the phenomenon of stickiness, whereby the orbit
spends long periods of time trapped near the boundary of a self-similar island-
around-island system. Note that in this case the Lyapunov exponent is likely to
be the same as for the orbit of Figure 5.16, thanks to the presence of the lim-sup
(rather than a true limit) in the definition.

n
6 y
0.00020
5

4 0.00015

3
0.00010

0.00005
1

x n
0
0 1 2 3 4 5 6 2000 4000 6000 8000 10000

Figure 5.13 For a = 0.0008, the orbit initiated at the point (2, 2) fills out a dis-
torted circle quasiperiodically. On the left we plot the first 50 000 points. On the
right is plotted n , the n-iteration approximation to the Lyapunov exponent, for
n = 10, 20, . . . , 100 000. Notice that the graph consists of a sequence of peaks of
decreasing altitude (so that the true Lyapunov exponent vanishes), separated by
short intervals where the eigenvalue is on the unit circle, so that n = 0 there.

0.5 0.5
n n
0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1
n n
0 1000 2000 3000 4000 5000 0 20000 40000 60000 80000 100000

Figure 5.14 For a = 0.8, the orbit initiated at (, 1/100) appears to be chaotic,
as in Figure 5.10. For this orbit, we plot the Lyapunov exponent approximant n
versus n. The graph on the left includes the first 5000 iterations; that on the right
is based on 100 000 iterations. The apparent approach to a positive limiting value
supports our interpretation of the orbit as a chaotic one.
5.7 A model of a chaotically rotating moon 139

5 n
5

4
4
3
y 3
2
2
1
1 n
0 2000 4000 6000 8000 10 000
1 2 3 4 5 6
x

Figure 5.15 For a = 6 and initial point (, 1/100), the phase portrait (left)
appears to consist of a single chaotic sea. The graph of the approximate Lyapunov
exponent n versus n is extremely flat after the first few hundred iterations. Chaos
rules.

6
5
5
4
4
3
y 3 n
2
2
1
1

0 2000 4000 6000 8000 10000


1 2 3 4 5 6 n
x

Figure 5.16 The chaotic orbit for the special value a = 6.349 972 and initial point
(2, 2). Note the small islands of stability near the starting point. Once again, after
a brief transient period, the approximate Lyapunov exponent quickly tends to a
limit near 4.8.

5.7 A model of a chaotically rotating moon


As a final example of a Hamiltonian system exhibiting chaos, we consider a sim-
plified model of the interaction between orbital and rotational degrees of freedom
of a small, irregularly shaped moon in an eccentric orbit about a massive planet.
The model was originally studied by J. Wisdom, S. J. Peale, and F. Mignard [36] as
a first attempt to understand the apparently chaotic rotation of Hyperion, a moon of
Saturn. From visual observations of the Voyager and Cassini spacecraft,2 Hyperion
2 See http://saturn.jpl.nasa.gov/science/moons/hyperion/.
140 Order and chaos in Hamiltonian systems
2.2

5 n
2.0
4
1.8
3
y
1.6 2

1
1.4
n
0 2000 4000 6000 8000 10 000
1.4 1.6 1.8 2.0 2.2
x

Figure 5.17 Sticky behavior near the boundary of one of the islands of stabil-
ity. The orbit remains non-chaotic on the periphery of the island system for
approximately 5000 iterations, then escapes and spreads out chaotically. The
initial point is at x = 182 999 999/100 000 000, y = 182 999 999/100 000 000.

2
1

2 r

r
a ae
O S

Figure 5.18 The configuration of the simplified Hyperion system.

is known to have quite different principal radii (205, 130, and 110 km), and its
orbital eccentricity is known to be about 1/10. The chaotic nature of its motion is
inferred from observations of the intensity of the light reaching us over prolonged
periods of time [37, 38].
The geometry of the planetmoon system is shown in Figure 5.18, where the
moon is replaced by its inertia ellipsoid (with principal moments I1 < I2 < I3 ).
The model assumes that the orbital motion of the moons center of mass is exactly
5.7 A model of a chaotically rotating moon 141

Keplerian, with constant semi-major axis a and eccentricity e, and that the rotation
of the moon has its axis perpendicular to the orbital plane, coincident with the third
principal axis of the inertia ellipsoid. With these simplifying features, the rota-
tional problem involves only one degree of freedom, the angle between the first
principal axis and the major axis of the orbital ellipse. Since the size of the moon
is assumed to be extremely small compared with the average orbital radius, we are
justified in making a multipole expansion of the potential energy, keeping only
the lowest nontrivial term. The effective Hamiltonian for the rotational motion
is thus
p2 GM  
H (, p , t) = + 3 r Q r ,
2I3 r
where G is Newtons gravitational constant, M is the planets mass, r is the unit
vector in the radial direction, and Q is the quadripole moment tensor, which is
related to the inertia tensor by
3 1
Q = I (Tr I)1.
2 2
The scalar combination is most conveniently evaluated in the frame of reference
where Q and I are diagonal and r has coordinates (cos , sin , 0), where =
(see Figure 5.18). Thus,
  3 3 1
r Q r = I1 cos2 + I2 sin2 (I1 + I2 + I3 ).
2 2 2
Although we are regarding the center-of-mass coordinates (r (t), (t)) as known
functions of time, it is still useful to formulate the equations of motion in
autonomous form, namely

= ,
 
3 I2 I1 1 + e cos 3
= 2 sin[2( )],
2 I3 1 e2

1 + e cos 2
= 1 e2 ,
1 e2
where we have written p = I3 .
To view the time evolution stroboscopically, with frequency (Keplers third law)

GM
= ,
a3
we can use the orbital angle as the independent variable, instead of the time t,
ending up with two differential equations involving only dimensionless quantities:
142 Order and chaos in Hamiltonian systems
  2
d 1 e2 1
= ,
d 1 + e cos 1 e2
 
d 3 I2 I1
= (1 + e cos ) (1 e2 )3/2 sin[2( )]. (5.7)
d 2 I3
To see whether our simplified model is at least consistent with the observed
chaos of Hyperion, we inserted (following [36]) the realistic values e = 0.1 and
(I2 I1 )/I3 = 0.264 and numerically integrated the equations of motion (5.7)
over 1000 orbital periods, starting from 30 different initial conditions. Calculating
Poincar sections for 200 values of the phase gave us an animated view of the
Hamiltonian flow in (, /, ) space, which is reproduced in part in Figure 5.19.
The interested reader can find an efficient Mathematica program for reproducing
the full animation in Section A.2 of the appendix.
The most prominent features revealed by the phase portrait are (1) a large swath
of the phase space occupied by chaos and (2) a prominent island of stability embed-
ded in the chaotic sea, synchronized with the orbital period. The prominence of
chaos in the phase portrait relative to the synchronous island can be shown, through
further numerical experimentation [36], to be correlated with the unusually large
values of e and (I2 I1 )/I3 .
Unfortunately, the simple model of Hyperion as a driven system with one degree
of freedom is too simple: the assumption of a constant rotation axis fails when
one takes into account the dynamics in three dimensions. The configurations asso-
ciated with points in the chaotic sea for realistic parameter values are universally
unstable [36]. More recently, numerical integration of the three-dimensional model
has been carried out [39], not only for Hyperion, but also for 33 other planetary
satellites whose size, shape, and orbital eccentricity are known. The authors calcu-
lated positive Lyapunov exponents, corresponding to the possibility of chaotic tum-
bling, for a number of these, including Hyperion. They also considered other fac-
tors that might favor observable chaos (especially the instability of the synchronous
resonance), and came to the conclusion that only three of the satellites, namely
Hyperion, Prometheus, and Pandora (all moons of Saturn) satisfied their criteria.

Exercises
5.1 To help in grasping the KAM approach to solving the small-denominator
problem, it is useful to consider a simpler warmup problem. We recall that
the rational numbers between 0 and 1 are countable (imagine listing them
in order of increasing denominator) and dense (to specify a rational within
10n of any given real number r (0, 1), simply keep the first n digits of
the decimal representation of r ). Now try to imagine covering all rationals in
Exercises 143

0 1 2 3

4 5 6 7

8 9 10 11

12 13 14 15

16 17 18 19

Figure 5.19 A tomographic representation of the Hyperion model phase portrait.


In each frame the horizontal and vertical coordinate ranges are 0 and
0 / 3, respectively.

(0, 1) with small open segments. Show that, on choosing the widths of the
segments appropriately, the sum of their lengths will be less than 1/2.
5.2 We wish to verify numerically the nonlinear stability of the L 4 and L 5
Lagrange points of the SunJupiter system. Using a realistic value for k in
h (4) (q, p), perform the canonical transformation to Birkhoff normal form
144 Order and chaos in Hamiltonian systems

and extract the numerical values of the coefficients in (5.4). Check the
inequality (5.2).
5.3 To generalize the result of Exercise 5.2, regard the coefficients in (5.3) as
abstract variables. Apply the canonical transformation to Birkhoff normal
form and verify the formulas for i, j given in the text. Finally, calculate the
determinant in (5.2) and plot its value versus k. For what mass ratios m J /m S
does this method establish nonlinear stability?
5.4 A simple pendulum is perturbed by a periodic oscillation of the gravitational
acceleration g. Thus

H (q, p, t) = p 2 /2 + 02 (1 + a sin(t))cos q.

For given a, construct a Poincar section by calculating and plotting the


points (q(t), p(t)), t = 0, 2, 4, . . . , 999 for (say) ten representative initial
conditions, with unperturbed energies above, below, and near that of unsta-
ble equilibrium. You should see some evidence of chaos (if not, choose a
larger a). Use Mathematica and fourth-order RungeKutta to integrate the
ODEs, after writing the latter in autonomous form:
dq H dp H d
= (q, p, / ), = (q, p, / ), = .
dt p dt q dt
5.5 The stochastic web-map with fourfold symmetry is defined as

W ((x, y)) = (y + K sin x, x).

(a) Show that, for asymptotically small positive K , the fourth iterate of W
takes the form

Heff Heff
W (x, y) = (x, y) + K
4
, + O(K 2 ),
y x
where
Heff = 2 cos x + 2 cos y.

Thus the discrete map W 4 acts like an infinitesimal time evolution with
time step K .
(b) Find the stable (elliptic) and unstable (hyperbolic) fixed points in the
plane of the Hamiltonian flow generated by Heff . Show that these are also
stable and unstable fixed points of the untruncated map W 4 for arbitrary
parameter K .
(c) Find the linearized map (i.e. the action of W 4 in its tangent space) in
a neighborhood of each fixed point. Determine the eigenvalues of the
Exercises 145

corresponding matrix (Jacobian matrix). Where the eigenvalues are real


(at a hyperbolic point), determine the corresponding eigenvectors.
(d) We wish to construct the unstable manifold at the point P = (, 0), as fol-
lows. Let + be the larger eigenvalue of the Jacobian matrix at P, and e+
the corresponding eigenvector. Choose a small positive number  so that
(, 0)+e+ = W 4 ((, 0)+e+ ) with high precision. Take 1000 equally
spaced points between (, 0) + e+ and (, 0) + e+ and map them
forward with W 4 to generate an approximation to the unstable manifold
of P. Plot the results (you may want to skip the first few iterations before
plotting). You should see the oscillations as the image points approach
a neighboring saddle point. Note: the smaller your choice of  and the
more interpolating points you map, the more accurately you will trace the
unstable manifold. It may be helpful to adjust the numerical precision in
Mathematica, by setting $MinPrecision = n for n-digit precision;
note that exact numbers can be converted to n-digit precision using the
function SetPrecision.
5.6 For initial conditions near (0, ), iterate the fourth iterate of the web map
with = 1/4 and K = 0.8 10 000 times and plot the orbit points as in
Figure 5.7. For orbits that appear to be chaotic, test the sensitivity to initial
conditions as well as the sensitivity to the precision of your floating-point
arithmetic. In Mathematica, the constant $MinPrecision, originally set
to machine precision of 15 or 16 digits, can be set at 30, 50, 100, etc. digits
to see whether the orbit shape is stable.
5.7 Choosing = 1/4, K = 0.8, and 20 different initial conditions (x0 , y0 ),
generate a folded-map phase portrait similar to Figure 5.8. Repeat for K =
0.1, 0.5, 1.0, 2.0, and 6.0.
5.8 In this extended exercise we explore numerically the dynamics of a forced
nonlinear oscillator with one degree of freedom, using methods similar to
those employed in the case of the Hyperion model in Sections 5.7 and A.2.
The laboratory setup (see Figure 5.20) consists of a flexible steel bar with one
end attached to a rigid frame, the other end free to move in a non-uniform
magnetic field. The single coordinate x measures the displacement from the
vertical of the free end of the bar. The field is such that there are three equi-
librium points: an unstable one at x = 0, and stable ones at x = x0 . If the
frame is moved back and forth sinusoidally, there is an effective sinusoidal
force on the bar that will drive its nonlinear oscillations.
In dimensionless units, we assume a potential-energy function,
1 1
V (x) = x 2 + x 4 ,
2 4
146 Order and chaos in Hamiltonian systems

x
N N
S S

Figure 5.20 A forced nonlinear one-dimensional oscillator.

corresponding to equations of motion

x = v, v = x x 3 + a cos t.

It is useful to introduce the phase of the driving term as a new dynami-


cal variable, with trivial time dependence, so that the system of three ODEs
becomes autonomous.

x = v, v = x x 3 + a cos , = 1. (5.8)

At the expense of an extra dimension, the numerical integration is now a


lot simpler. Moreover, the orbits in 3-space have the nice property that they
are non-intersecting.
(a) For a = 0.1, integrate the ODEs of (5.8) for various initial conditions.
Examine the orbits both in two and in three dimensions. Find exam-
ples that are clearly quasiperiodic and others that are chaotic. If you
have trouble finding chaos, you can increase the value of a and see what
happens.
(b) Define a Mathematica function poincare[{x_, v_, phi_}] that takes
as its argument an arbitrary state {x, v, phi}, applies the integration step
n times with time step dt = 2 N[Pi]/n and outputs the resulting state
with the third component reset to its original value.
Exercises 147

For fixed a, say a = 0.1, iterate poincare, starting at appropriate initial


values, all with phi = 0, and ListPlot at least three orbits of the
Poincar map. At least one of these should be chaotic.
(c) Compare the Poincar sections with different choices of phi. If you
are really ambitious, you can generate and animate a sequence that runs
through many (say 200) phi values.
6
The swing-spring

A swing-spring, sketched in Figure 6.1, is a spherical pendulum with an additional


elastic degree of freedom. It consists of a point mass m free to move in space
(coordinates x, y, z) under the influence of a potential energy function
k
V (x, y, z) = mgz + (l0 x 2 + y 2 + z 2 )2 ,
2
where g is the acceleration of gravity and k is the force constant of a spring with
unstretched length l0 . The model provides an excellent review of a number of con-
cepts and methods treated earlier, as well as being relevant to the semi-classical
modeling of the carbon dioxide molecule.
Like the spherical pendulum, the swing-spring has stable and unstable equilibria
on the z axis, namely at z = l = (l0 + mg/k) and z = l0 mg/k, respec-
tively. To simplify the notation, we express lengths in units of l, mass in units of

m, and time in units of l/g, and displace the origin to coincide with the stable
equilibrium point at zero energy. The Hamiltonian then takes the form
 2
1 2 2 1 1
H = ( p x + p y + pz ) + z +
2 2
1 2 x + y + (z 1)
2 2 2 2.
2 2 2
Small oscillations about the stable equilibrium are governed by the quadratic terms
in a Taylor expansion of H , namely
1 1
H2 = ( px2 + p 2y + pz2 ) + (x 2 + y 2 + 2 z 2 ).
2 2
Following the example of numerous authors (for example, [40, 41]), we shall
restrict ourselves to the case = 2 in which the three frequencies of small-
amplitude oscillation are in 1:1:2 resonance. For this choice, the system exhibits
an interesting type of motion with alternating swinging and springing accompanied
by a periodic precession of the swing plane.

148
6.1 Two-dimensional motion 149

Figure 6.1 The swing-spring.

6.1 Two-dimensional motion


The swing-spring is not an integrable system, possessing only two independent,
mutually commuting conserved quantities, the Hamiltonian H and the angular
momentum L = x p y ypx about the z axis. Before getting into the details of the
full model, it is useful to examine a restricted subset of possible motions, namely
those corresponding to L = 0. If we choose initial conditions y = p y = 0, the
orbits will be restricted to the x, z plane, with reduced Hamiltonian
 2
1 2 3 2 1
H = ( p x + pz ) + z + 2
2
x + (z 1) 2 . (6.1)
2 4 8
Near the stable equilibrium point x = z = px = pz = 0, we recognize a familiar
pattern: the phase space decomposes into constant-energy manifolds M H diffeo-
morphic to 3-spheres, each of which can be conveniently visualized as the regions
enclosed by a pair of 2-spheres with the boundaries identified. Specifically, if H is
assigned a specific value, we have from (6.1) that

px = 2H pz2 2V (x, z),

where the potential energy 2V (x, z), whose level lines are shown in Figure 6.2,
behaves near x = z = 0 as
1
V (x, z) = (x 2 + 4z 2 ) + .
2
150 The swing-spring

2.0

1.5

1.0

0.5

0.0

0.5

1.5 1.0 0.5 0.0 0.5 1.0 1.5

Figure 6.2 A contour plot of V (x, z).

Figure 6.3 Boundary surfaces for the allowed region in x, z, px space, for three
values of the energy, H = 0.01, 0.875, 2.875.

The two allowed regions are those for which the argument of the square root is
non-negative, i.e.

pz2 + 2V (x, z) 2H, px 0,

with a common boundary px = 0 diffeomorphic to a 2-sphere. The surfaces for


several values of H are shown in Figure 6.3.
For sufficiently small H , we expect that most orbits will inhabit nested topologi-
cal 2-tori, but, of course, we cannot immediately apply the KAM theorem, since we
have chosen commensurate frequencies for the quadratic part of the Hamiltonian.
Here we will be satisfied with the qualitative description of the phase portrait pro-
vided by a numerically constructed Poincar section. For the latter, we choose
several values of H and the section plane x = 0, px 0. The results, obtained
using fourth-order RungeKutta integration of Hamiltons equations of motion,
are exhibited in Figure 6.4.
6.1 Two-dimensional motion 151
0.15

0.10

0.05

0.00
H = 0.01
0.05

0.10

0.15
0.06 0.04 0.02 0.00 0.02 0.04 0.06

1.0

0.5

H = 0.875 0.0

0.5

1.0

0.6 0.4 0.2 0.0 0.2 0.4 0.6

H = 2.875 0

1.0 0.5 0.0 0.5 1.0 1.5 2.0

Figure 6.4 Phase portraits for the Poincar map for three values of the energy H
and the plane of section x = 0, Px 0.
152 The swing-spring

6.2 Integrable approximations for small oscillations


In the low-energy regime, the swing-spring may be regarded as a perturbed har-
monic oscillator with three degrees of freedom and commensurate frequencies. For
small oscillations about the stable equilibrium at the origin in phase space, canon-
ical perturbation theory, especially the brand of Birkhoff and Gustavson, holds
out the promise of reproducing the main features of the phase portrait. Ideally,
we would like the method to produce fully integrable systems that approximate
the exact model for oscillations of sufficiently small amplitude, but this is hardly
guaranteed by the general theory.
Recall that the BirkhoffGustavson method for the so-called degenerate case of
commensurate unperturbed frequencies is designed to produce a formal third inte-
gral of the motion to complement the Hamiltonian and the angular momentum.
The formal third integral in any finite order is not necessarily a strictly conserved
quantity (typically there are higher-order corrections), nor is it guaranteed to be
L-invariant (only finite-order commutation with H is enforced by the construc-
tion). Surprisingly, the method does much better, at least up to second order,
yielding integrable models with Hamiltonians that are polynomials of degrees 3
and 4 in the transformed variables.
It is instructive, as a review exercise, to work through the first-order transforma-
tion in detail. We begin by making a convenient canonical linear transformation of
the phase-space coordinates:

+ p + p p p
(x, y, z, px , p y , pz ) = , , , , , 2 p . (6.2)
2 2 2 2 2

In terms of the new coordinates, the quadratic Hamiltonian and angular momentum
take the form

1 1 1 1
H2 = ( p2 + 2 ) + ( p2 + 2 ) + ( p2 + 2 ), L = ( p2 + 2 ) ( p2 + 2 ).
2 2 2 2

On expanding the swing-spring Hamiltonian up to degree 3 in the variables


, , , p , p , p , we get

H = H2 + H3 + ,

where
3
H3 = ( 2 + 2 + p2 + p2 + 2 p + 2p ).
4 2
6.2 Integrable approximations for small oscillations 153

To remove unwanted terms in H3 , we make the usual substitutions


1 1 1
a = ( i p ), b = ( i p ), c = ( i p ),
2 2 2

1 1 1
a = ( p i ), b = ( p i), c = ( p i ).
2 2 2
In the resulting polynomial, we assign an index

d(a m 1 bm 2 cm 3 a n 1 bn 2 cn 3 ) = (m 1 + m 2 + 2m 3 n 1 n 2 2n 3 )i

to each term L, deleting it if d(M) = 0, and multiplying it by 1/(M) otherwise.


On transforming back to the coordinates , , etc., this produces (Exercise 6.2) the
lowest-order BirkhoffGustavson generating function,

S3 (, , , p1 , p2 , p3 ) = p1 + p2 + p3
3 
+ ( p1 p2 ) 2( 2 + 2 ) p3 ( p2 + p1 ) p3
4 
2( p12 + p22 ) p3 , (6.3)

where we have set 3 = 3 2/8, and the corresponding transformation equations,
S 3
p = = p1 + ( 4 p3 p2 p3 ), (6.4)
4
S 3
p = = p2 + ( 4 p3 p1 p3 ), (6.5)
4
S 3
p = = p3 + ( p1 p2 ), (6.6)
4
S 3
q1 = =+ ( p2 p3 4 p1 p3 ), (6.7)
p1 4
S 3
q2 = =+ ( p1 p3 4 p2 p3 ), (6.8)
p2 4
S 3
q3 = =+ (2( 2 + 2 + p12 + p22 ) p2 p1 ). (6.9)
p3 4
In the q, p coordinates, the Hamiltonian takes the form

H (q, p) = H2 (q, p) + H3 (q, p) + H4 (q, p) + ,

where
H3 (q, p) = 3 ((q1 q2 q1 q2 ) p3 (q1 p2 + p1 q2 )q3 )
154 The swing-spring

now contains only index-0 monomials. Repeating the process with the quartic
generating function (Exercise 6.2)
S4 (q1 , q2 , q3 , P1 , P2 , P3 ) = q1 P1 + q2 P2 + q3 P3
21   
+ P1 P2 q1 q2 q12 + q22 + P12 + P22
256
69  
+ q1 P1 (P22 q22 ) + q2 P2 (P12 q12 )
1024
1  2
+ q (100q1 q2 9(q1 P1 + q2 P2 ) + 28P1 P2 )
1024 3
+ P32 (412q1 q2 + 153(q1 P1 + q2 P2 ) 340P1 P2 )
+ q3 P3 (141(q12 + q22 ) 520(q1 P2 + q2 P1 )

123(P12 + P22 )) , (6.10)
eliminates the non-zero-index terms in H4 . On performing the canonical transfor-
mation (q, p) (Q, P), and then truncating, we end up with a Hamiltonian
K (Q, P) = K 2 (Q, P) + K 3 (Q, P) + K 4 (Q, P)
that commutes with K 2 (which is the whole point of the construction).

6.3 Three commuting integrals


Following the example of [40], we now introduce coordinates k , k = 1, . . . , 5,
which have particularly simple Poisson brackets and allow a concise representation
of K , K 2 :
k = Pk2 + Q 2k , k = 1, 2, 3,
5 + i4 = (Q 1 i P1 )(Q 2 i P2 )(Q 3 + i P3 ).
Clearly the first three k commute with each other, and one easily checks that
[1 , 4 ] = [2 , 4 ] = [3 , 4 ] = 25 ,
(6.11)
[1 , 5 ] = [2 , 5 ] = [3 , 5 ] = 24 ,
as well as
42 + 52 = 1 2 3 . (6.12)
In terms of these variables, K (Q, P) takes the form (Exercise 6.3)
K = K 2 + 3 4 + 41 (12 + 22 ) + 42 1 2 + 43 (1 + 2 )3 , (6.13)
where
1
K 2 = (1 + 2 + 23 )
2
6.3 Three commuting integrals 155

and
3 51 57
41 = , 42 = , 43 = .
128 256 256
From the commutation relations for the k , the quantity
1
M = (2 1 )
2
is easily seen to commute with both K and K 2 .
The Hamiltonian flows generated by k /2, k = 1, 2, 3, may be viewed
as commuting rotations in the respective Q k , Pk planes, with conjugate angular
variables
k = Arg(Q k i Pk ).

This assignment follows from the commutation relation


i i
[k , k /2] = [ln(Q k i Pk ), Q 2k + Pk2 ] = (i Q k + Pk ) = 1.
2 Q k i Pk
The Hamiltonian flows corresponding to K 2 and M may be represented as a
composition of rotations in the Q k , Pk planes:

e[,K 2 ] = e[,1 ] e[,2 ] e2[,3 ] , e [,M] = e [,1 ] e [,2 ] .

The three mutually commuting functions K , K 2 , and M are independent, by


definition, on those level sets M K ,K 2 ,M where their respective six-dimensional gra-
dients are linearly independent. This can be determined by examining the 3 3
minors of the matrix of partial derivatives. Exploiting the invariance with respect
to transformations generated by K 2 , M, we can restrict ourselves to points where
Q 1 = Q 2 = 0:
 
K , K 2 , M =


3 P2 Q 3 0 0
3 P1 Q 3
0 0

2Q 3 (1 + 43 (1 + 2 )) 2Q 3 0
.
3 P2 P3 + P1 (1 + 441 1 + 242 2 + 243 3 ) P1 P1

3 P1 P3 + P2 (1 + 441 2 + 242 1 + 243 3 ) P2 P2
3 P1 P2 + 2P3 (1 + 43 (1 + 2 )) 2P3 0

Setting all 33 minors equal to 0, we get (Exercise 6.4) the dependence conditions

5 = 0, 4 = 1 2 3 ,
156 The swing-spring

3 (2 3 + 1 3 1 2 ) 2 1 2 3 ((241 + 42 43 )(1 + 2 ) + 243 3 )
= 0. (6.14)
Upon inserting into (6.14) the identities
1 = K 2 M 3 , 2 = K 2 + M 3 ,
we get an equation for 3 that (in principle) can be solved and inserted into K to get
the dependence of K on K 2 and M on the exceptional manifold. Below we will see
that the latter corresponds to the boundary of the admissible region in K , K 2 , M
space

6.4 Dynamics on the level sets


We eventually would like to introduce action-angle variables on all of the non-
degenerate level sets M K ,K 2 ,M . Since the system is not separable in the Q, P coor-
dinates, we will resort once again to the method of the LiouvilleArnold proof.
Before doing that, however, it is very useful to introduce non-canonical coordinates
on the M K ,K 2 ,M , which function very much like the variables u = cos , , in
the symmetric-top model. There, we found that the dynamics reduced neatly to
periodic solutions of an ordinary differential equation for u, which gave directly
not only the nutation angle , but also the precession and spin rates and , all
with the same period. In the present case we will derive analogous roles for the
variables 3 , 1 , and 2 defined in the previous section.
Our first task is to construct a time-evolution equation involving only 3 . We
observe that on M K ,K 2 ,M , all of the variables k , k = 1, . . . , 5, can be expressed
in terms of the constants of the motion, 3 , and the sign of 5 . Specifically,
1 = K 2 M 3 ,
2 = K 2 + M 3 , (6.15)
4 = 1
3 g(3 ),

5 = f (3 ),
where (Exercise 6.5)
g(3 ) = K K 2 + (41 242 )M 2 243 3 (K 2 3 )
(42 + 241 )(K 2 3 )2 , (6.16)
f (3 ) = (K 2 M 3 )(K 2 + M 3 )3 2
3 g(3 ) .
2
(6.17)
Having isolated the variable 3 on the level set, we are in a position to write
down a first-order differential equation for its time evolution:

3 = [3 , K ] = 3 [3 , 4 ] = 23 5 = 23 f (3 ). (6.18)
6.4 Dynamics on the level sets 157

The functions f and g are polynomials in 3 of degree 4 and 2, respectively. The


physically allowed values of K , K 2 , and L are determined by the non-negativity, for
some 3 , of the quantities 1 , 2 , 3 , and f (3 ). In a connected level set M K ,K 2 ,M ,
the variable 3 will oscillate between two positive roots of the polynomial f (3 ),
with the sign of 3 (and hence that of 5 ) changing at the turning points. Since
the polynomial f (3 ) is of degree 4, there may be some values of the integrals for
which there are two disjoint physically allowed ranges bounded by roots of f (3 ).
Once an oscillatory solution of (6.18) has been extracted, the time derivatives of
1 and 2 may be calculated using the Poisson brackets with the Hamiltonian,
4 def
1 = [1 , K ] = 1 + 3 + 441 1 + 242 2 + 243 3 = h 1 (3 ), (6.19)
1
4 def
2 = [2 , K ] = 1 + 3 + 441 2 + 242 1 + 243 3 = h 2 (3 ). (6.20)
2
Here we have used not only (6.11), but also the commutation relations
(Exercise 6.6)
4
[k , 4 ] = , k = 1, 2. (6.21)
k
An arbitrary point on M K ,K 2 ,M is specified by non-canonical coordinates
3 , 1 , 2 and the sign of 5 . Specifically, the six canonical coordinates are given by
 
Q1 cos 1
= K 2 L 3 , (6.22)
P1 sin 1
 
Q2 cos 2
= K 2 + L 3 , (6.23)
P2 sin 2
while Q 3 and Pk can be extracted by linear algebra from

5 = (Q 1 Q 2 P1 P2 )Q 3 + (Q 1 P2 + Q 2 P1 )P3 = f (3 ),

4 = (Q 1 P2 + Q 2 P1 )Q 3 + (Q 1 Q 2 P1 P2 )P3 = 1
3 g(3 ),

namely

Q3 1
=
P3 (K 2 3 )2 L 2
 
cos(1 + 2 ) sin(1 + 2 ) f (3 )
. (6.24)
sin(1 + 2 ) cos(1 + 2 ) 1
3 g(3 )

The use of coordinates 3 , 1 , and 2 on the level sets of K , K 2 , and M has many
advantages, but, as is often the case with polar coordinate systems, one must be on
the lookout for so-called coordinate singularities arising from the non-uniqueness
158 The swing-spring

of the angle 1 (2 ) at points where 1 (2 ) vanishes. Suppose that an orbit arrives


at a point where (say) 1 = 0. At this point, the angle 1 could be incremented
by an arbitrary amount without changing the canonical phase-space coordinates
Q i , Pi , i = 1, 2, 3, and each choice of increment would lead, typically, to a dif-
ferent future evolution of the orbit. Such behavior is, of course, unacceptable in
a dynamical system. Fortunately the failure of predictability can be removed by
supplementing the differential equations with simple rules for incrementing k at
points where k vanishes. We will return to this issue later in the chapter, where
it will turn out to be of some relevance in the discussion of the continuity and
single-valuedness of action-angle coordinate systems.

6.5 Constraints among the integrals


As we have seen, the phase space of our integrable swing-spring decomposes into
compact, connected level sets of the integrals, M K ,K 2 ,M , where the latter are well
defined as long as the quantities k , k = 1, 2, 3 are non-negative and, in addition,
the function f (3 ) is non-negative somewhere on the 3 axis. The triples K , K 2 , M
for which M K ,K 2 ,M is nonempty will be called admissible, and the set E of admis-
sible K , K 2 , M will be seen below to be bounded by precisely the surfaces (6.14)
where independence breaks down. In the language of differential geometry, E is
the base of a fiber bundle whose fibers are the M K ,K 2 ,M . Away from the boundary,
the latter, by virtue of the LiouvilleArnold theorem, are diffeomorphic to three-
dimensional tori. In the literature, the manifold E is described as the image of the
energymomentum map, being analogous to the planar diagrams which we studied
in connection with the integrable models of Chapter 3.
In order to get an intuitive grasp of the manifold structure of E, it is useful
to make use of computer-assisted graphics. The following is a summary of our
explorations using Mathematica. The basic strategy is to plot the function f (3 ) for
sufficiently many values of the integrals. We can restrict our attention to |M| K 2
and the 3 -interval
0 3 K 2 |M|, (6.25)
since, by (6.15), these inequalities are necessary and sufficient for the non-
negativity of 1 , 2 , 3 . It is then only a matter of checking the non-negativity of
f (3 ) on the allowed interval. From (6.17), we see that f (3 ) is a quartic poly-
nomial that tends to for 3 . Moreover, since at the endpoints R of
the allowed interval (6.25) at least one of the k , k = 1, 2, 3, vanishes, we have
f (R ) = 42 0. Hence, on (R , R+ ), f (3 ) has an even number (0, 2, or 4) of
real roots, which we label (if they exist) R0 , R1 , R2 , R3 in order of increasing 3 .
The results of our systematic survey can be summarized rather succinctly. For
all K 2 > 0 and K 2 < M < K 2 , there exists a unique interval
6.5 Constraints among the integrals 159

K (K 2 , M) < K < K + (K 2 , M) (6.26)

such that R0 and R1 exist. For K 2 less than the critical value (Exercise 6.7)
512
K 2 = , (6.27)
361
the constant-K 2 section of E resembles the energymomentum plot of the spherical
pendulum: the real roots R0 and R1 are distinct except on the two boundary curves
K = K (K 2 , M) and at an isolated interior point on the M = 0 axis.
Above the critical value of K 2 , the manifold structure is more complicated. For
fixed K 2 , the allowed region is again bounded by the curves K = K (K 2 , M), on
which the roots R0 and R1 coincide. Now, however, there is a subregion, bounded
by three curves,

K  (K 2 , M) < K < K +

(K 2 , M), = sign(M),

where f has additional real roots R2 and R3 such that f (3 ) is non-negative for
R2 3 R3 . For M, K within that subregion, the admissibility manifold has two
sheets, with a common boundary curve (a piece of K = K  (K 2 , M)) on which
R0  = R1 = R2  = R3 .
The most direct way to picture the manifold structure described in the preceding
two paragraphs is to examine cartoons showing how the function f varies with
K for fixed K 2 and M. Fortunately, Mathematicas interactive functions allow us
to generate these plots with relative ease (see Figure 6.5). In Figure 6.6, we show a

0.5 1.0 1.5 2.0 2.5 3.0

-1

Figure 6.5 An interactive Mathematica panel used to display the function f (3 )


for selected values of K , with K 2 and M fixed.
160 The swing-spring

4 4

2 .9
2
2 .7
2.3 .5
3 3

K= .3
3.1
2.1
2 2

3
3.5
3.7
f f

3.9
1.9

4.1
4.3
1 1

4.5
4.7
0 1.7 0
5
1.
=

1 1
0.0 0.5 1.0 1.5 2.0 2.5 3.0 0.0 0.5 1.0 1.5 2.0 2.5 3.0
3 3
0.10

0.05

K=2.91
f 0.00
2.92
2.93
0.05
2.94
3.01

1
3.0
2.95
0.10
2.0 2.2 2.4 2.6 2.8 3.0
3

Figure 6.6 Plots of the function f (3 ) for K 2 = 3, M = 1/20 and selected values
of K throughout the allowed range.

sequence of plots of the function f (3 ) for parameter values K 2 = 3, M = 1/20,


and K between 3/2 and 11/2. For clarity, we display the sequence in three frames,
the third of which zooms in on the region where f is non-negative on two disjoint
intervals of the 3 axis. With the help of Mathematicas function ContourPlot,
we can obtain a more elegant representation of the constant-K 2 sections. Figure 6.7
shows not only the curvilinear boundaries, but also the contours of the function
R1 (K 2 , M, K ) R0 (K 2 , M, K ) for K 2 = 1 (subcritical) and K 2 = 3 (supercrit-
ical). In the zoom of Figure 6.8, the contours of the two sheets are superposed
for K 2 = 3. We have also combined the data from a number of two-dimensional
sections to obtain a picture of the boundary of E in three dimensions (Figure 6.9).
Having reviewed qualitatively the structure of the admissibility manifold E, it is
time to derive a more precise mathematical description of the nontrivial boundary
set. From our construction, it is clear that the boundary set is defined by the param-
eter values where admissible 3 intervals either are born or die. These are precisely
the values where
f (3 ) = 0, f  (3 ) = 0,
where f  is the derivative with respect to 3 with fixed K , K 2 , and M. From (6.17),
the first equation may be viewed as a formula for 4 :
6.5 Constraints among the integrals 161

2.0 6
K2 = 1 K2 = 3
5
1.5

K 1.0 K
3

0.5
2

0.0 1
1.0 0.5 0.0 0.5 1.0 3 2 1 0 1 2 3
M M

Figure 6.7 Contour plots of R1 R0 for K 2 = 1 (left) and K 2 = 3 (right). The


boundary set includes an isolated point in the first example and a curvilinear arc
in the second. The latter, which is the common boundary of the two sheets of the
manifold, is shown more clearly in Figure 6.8.

3.10

3.00

K2
2.90

2.80

0.3 0.2 0.1 0.0 0.1 0.2 0.3


M
Figure 6.8 A zoom of the central region of the K 2 = 3 example, with contours
of R3 R2 superposed with gray shading to suggest the two-sheeted structure of
the manifold.
162 The swing-spring
10 0
-10

60

40

20

0
5
10

Figure 6.9 A three-dimensional plot of the bounding surfaces of the admissibility


manifold E truncated at K 2 = 14. The two branches are shown superposed (in
reality they share only the lower surface of the inner horn).


4 = 1
3 ((K 2 3 )2 M 2 )3 .

The second equation then gives us, using (6.16),



59 5
332 4K 2 3 + K 22 M 2 23 ((K 2 3 )2 M 2 ) 3 K 2 = 0.
16 2
(6.28)

This equation is easily seen (Exercise 6.9) to be identical to (6.14), allow-


ing us to identify the boundary of admissibility with the manifold where the
6.5 Constraints among the integrals 163

K2 > K2* K2 > K2* K2 K*2


(sheet 1) (sheet 2)

B+ B+
F
A A+ E E+ A G A+
C C+
D C D C+
B
B

A B C D E F G

Figure 6.10 Sketches of typical constant-K 2 sections of the boundary manifold


of E. Each labeled component corresponds to a particular double- or triple-
root scenario for the function f (3 ), shown in the row of plot sketches at the
bottom.

independence of the integrals breaks down. Note that (6.28) can be solved (numer-
ically) for 3 as a function of K 2 and M. When substituted into the formula for
K , this gives the explicit dependence relation K (K 2 , M) on the boundary, and
can be used to obtain accurate numerical and graphical representations of the
manifold.
The organization of the boundary surfaces is summarized succinctly, in terms
of constant-K 2 sections, in Figure 6.10. Each piece of the surface corresponds to
a characteristic behavior of the function f (3 ), which is sketched in the figure.
In each case, there is one value of 3 for which both f (3 ) and f  (3 ) van-
ish, corresponding to a stationary point of the 3 motion. In some cases there
is also separatrix-type motion involving approach to the stationary value over
an infinite time interval. This is analogous to the separatrix motion of a simple
pendulum.
Referring to Figure 6.10, we note (Exercise 6.10) that for cases A we have
M = K 2 and 3 = 0. For A+ , we also have 1 = 0, and for A , we have 2 = 0.
For cases C , the second derivative of f vanishes, i.e. there is a triple root. This is
also true for case G if K 2 = K 2 . Finally, for cases F and G, M = 1 = 2 = 0,
and 3 = K 2 = K .
Figure 6.11, which was obtained by solving the boundary equations numeri-
cally, shows how the corner points A , C , and F evolve as functions of K 2 .
164 The swing-spring

F
12

8
K2 A C C+ A+

0
10 5 0 5 10
M

Figure 6.11 Horn corners, projected on the M, K 2 plane.

12
D F
K2 8 B B+
4
0
0 10 20 30 40 50 60
K

Figure 6.12 The horn profile at M = 0.

The behavior of the C curves just above the bifurcation point at K 2 = K 2 may be
found from an asymptotic analysis of the same equations:

361
M(K 2 ) = (K 2 K 2 )3/2 + O((K 2 K 2 )5/2 ).
1920 6
Moreover, the asymptotic behavior of the curves at infinity is

19
M K2.
59
Another view of the bifurcation and horn-within-horn structure of the boundary
surface is provided by Figure 6.12, where the section at M = 0 is plotted.
6.6 Flow coordinates and the period lattice 165

6.6 Flow coordinates and the period lattice


We now turn to the Arnold construction of flow coordinates conjugate to the inte-
grals. In order to simplify the algebra, it is useful to choose, instead of K 2 and M,
linear combinations that are conjugate to the angles 1 and 2 , namely
1 1
J1 = (K 2 M)/2 = (1 + 3 ), J2 = (K 2 + M)/2 = (2 + 3 ).
2 2
The reader can easily check the Poisson-bracket relations
[Jk , k ] = 1, k = 1, 2.
The notation is intentionally chosen to suggest that the Jk will eventually be seen
to be members of the complete set of action-angle variables on M(K , K 2 , M).
We choose as our flow-coordinate origin 0 on M K ,K 2 ,M the turning point where
3 = R0 and 1 = 2 = /2. At this point,
Q 1 = Q 2 = Q 3 = 0,
and
g(R0 )
P1 = 2J1 R0 , P2 = 2J2 R0 , P3 = .
3 (K 2 R0 )2 L 2
We observe that any point on M K ,K 2 ,M can be connected to 0 by a continuous
path: first, by means of motions generated by J1 and J2 , the angular coordinates
1 and 2 can be reduced continuously to 0 without changing 3 , which must lie
within the allowed interval R0 3 R1 ; then 3 can be reduced continuously
to R0 . This prescription does not depend on the sign of 5 . The level set is thus
connected and compact, and hence has the topology of a 3-torus. As pointed out
earlier, the choice of the root R0 might not be unique. If not, there will be more
than one compact, connected component.
Following Arnolds construction, as described in Section 3.5, we define multi-
chart coordinates t1 , t2 , t3 for each M K ,K 2 ,M by means of the equation
= g(t1 , t2 , t3 )0 = et1 [,J1 ] et2 [,J2 ] et3 [,K ] 0 .
We have already discussed the flow
/2 /2 /2 ]
g(t1 , t2 , 0) = et1 [,J1 ] et2 [,J2 ] = et1 [, 1 ] et2 [, 2 ] e(t1 +t2 )[, 3
which displaces 1 and 2 by t1 and t2 , respectively, while leaving 3 untouched.
On the other hand, the time evolution of all three variables is governed by the
differential equations (6.18)(6.20). Specifically, the round-trip period of the 3
oscillation between roots R0 and R1 is
 R1
1 d
T = , (6.29)
3 R 0 f ()
166 The swing-spring

so, if is the number of half-periods elapsed before time t3 , i.e. = t3 /(T /2),
we have  3

1 d

23 R f () , even,

0
t3 = T /2 +  R1

1 d

, odd.
23 3 f ()
Meanwhile, the angles k , k = 1, 2, have evolved to
 3
1 h k ()d

, even,
23 R0 f ()
k (t3 ) = + Tk /2 +  R1 (6.30)
2
1 h k ()d

, odd,
23 3 f ()
where

1 R1
h k ()d
Tk = k (T ) /2 = , k = 1, 2. (6.31)
3 R0 f ()
The period lattice for M K ,K 2 ,M consists of the set of all 3-vectors (t1 , t2 , t3 ) such
that g(t1 , t2 , t3 )0 = 0 . From the preceding paragraph, it is clear that this will be
the set of all integer linear combinations of the basis vectors
e1 = (2, 0, 0), e2 = (0, 2, 0), e3 = (T1 , T2 , T ). (6.32)
The three-dimensional vector space generated by this basis provides a multichart,
i.e. a many-to-one coordinate system for the 3-torus M K ,K 2 ,M .
Once one has constructed a period lattice, the passage to action-angle variables
on M K ,K 2 ,M is straightforward. This leaves aside questions of continuity under
changes of the integrals, an important issue that we will return to in the next
section. We therefore choose the fundamental parallelepiped of the period lattice
(spanned by the three basis vectors) as a suitable single-valued coordinate domain.
The three angle variables 1 , 2 , 3 are defined to vanish at (t1 , t2 , t3 ) = (0, 0, 0)
and to increase uniformly along the respective directions of e1 , e2 , e3 . The rate of
increase of k is chosen so that the value 2 is attained at the point ek . This gives
us the linear relation
3
i
ei = (t1 , t2 , t3 ),
i=1
2

which can be written as the matrix equation



2 0 T1 1 t1
1
0 2 T2 2 = t2 .
2
0 0 T 3 t3
6.6 Flow coordinates and the period lattice 167

This is easily inverted to give


1 = t1 + 1 t3 , 2 = t2 + 2 t3 , 3 = 3 t3 ,
where the three frequencies are
T1 T2 2
, 1 =
2 = , 3 = .
T T T
To define corresponding action variables, we need to calculate the line integrals

1
P d Q, k = 1, 2, 3,
2 Ck
where Ck is a continuous path, which in (t1 , t2 , t3 ) space starts at (0, 0, 0) and
ends at ek . We exploit the fact that the results are left invariant under continuous
deformations of the paths with the endpoints unchanged to choose the contours of
integration to be as convenient as possible.
For k = 1, we choose C to have constant 3 = R0 and 2 = /2, with 1
increasing from 0 to 2. Along this path, we have, from (6.22)(6.24),
   
Q1 cos 1 Q3 cos 1
= 1 , = 3 ,
P1 sin 1 P3 sin 1
so that
 
1 1 + 3 2
1 + 3
P dQ = sin2 1 d1 = = J1 .
2 C1 2 0 2
Clearly, the analogous integration over C2 gives

1 2 + 3
P dQ = = J2 .
2 C2 2
This establishes the role of J1 and J2 , already anticipated in the notation, as the
action variables conjugate to 1 and 2 . For k = 3, we choose C3 with constant
1 = 2 = /2, and 3 making the circuit from R0 to R1 and back. On this curve,
we have d Q 1 = d Q 2 = 0 and
 
Q3 1 f (3 )
= ,
P3 (2J1 3 )(2J2 3 ) 1
3 g(3 )

so that

def 1
J3 = P dQ
2 C3
  
1 R1
d g() f ()
= .
3 R0 (2J1 )(2J2 ) (2J1 )(2J2 )
168 The swing-spring

As in the case of (6.29), the integral representation of J3 has an integrand that is


singular, but integrable, at the endpoints. This is only a minor difficulty for accurate
numerical integration, but becomes a more serious problem if one wishes to extract
partial derivatives with respect to K , J1 , J2 .

6.7 Monodromy
Thus far we have not worried about the question of continuity and single-
valuedness of the period lattice basis vectors (6.32) as a function of the admissible
integrals. From our experience with the spherical pendulum, we would not be
surprised if a certain class of closed paths in E were to correspond to nontriv-
ial monodromy. To see whether this is in fact the case, let us track numerically the
evolution of the single-period angular increments Tk = k (T ) /2 as we traverse
clockwise the following unit circle in the K 2 = 3 plane:

K 2 = 3, M = cos s, K = 29/10 + sin s, 0 s 2. (6.33)

Like curve in Figure 6.16 later, the path encircles the inner boundary arc. By
integrating numerically the expression for Tk in (6.31), we obtain the plots of
Figure 6.13. It appears that the graph of T1 jumps upward by 2 at a certain param-
eter value, s = s1 , whereas T2 appears to jump downward by 2 at another point,
s = s2 . One can readily check the magnitude and sharpness of the apparent dis-
continuities using high-precision numerical calculations (Exercise 6.11). The latter

5
T1
4

2 T2
1

1 2 3 4 5 6
s
Figure 6.13 Plots of the single-period angle increments T1 (black) and T2 (gray)
as functions of the parameter s of the curve (6.33). One sees apparent 2 jumps
of T1 and T2 .
6.7 Monodromy 169

3
f
2

1.0 1.5 2.0


3
1

Figure 6.14 The graph of f (3 ) at the discontinuity of T1 .

also reveal that, at the discontinuity points, the quantities Tk are not indeterminate,
but rather take on values halfway between the left-hand and right-hand limits.
What mechanism accounts for the discontinuities? An important clue is pro-
vided by the graph of f (3 ) at the site of the discontinuity of T1 (Figure 6.14).
Here the second root R1 (K 2 , M, K ), the upper limit of the 3 oscillation, is located
at the maximum allowed value, namely K 2 M, so 1 vanishes there. Moreover,
the identity (6.12) implies that 4 and 5 are zero at the endpoint as well. The fact
that the integrand for the T1 integral contains a term proportional to 4 /1 sug-
gests a possible mechanism for the discontinuity: if we can show that 4 (R1 ) has
a simple zero at s = s1 , while 1 (R1 ) 4 (R1 )2 , then the ratio of the two is of
order 4 (R1 )1 with opposite signs for s < s1 and s > s1 . This could presumably
account for the observed discontinuity of T1 . A similar argument would hold for T2 .

To explore this idea, we write



1 (R1 ) = K 2 M R1 , 4 (R1 ) = (K 2 L R1 )(K 2 + M R1 )R1
so that, indeed, as we approach s = s1 ,
4 (R1 )2
1 (R1 ) = .
2M(K 2 M) + O(42 )
Furthermore,
(1)
4 (R1 ) = 1 1
3 g(R1 ) = 3 (K K crit (K 2 , M)) + O(4 ),
2

where
(1)
K crit (K 2 , M) = K 2 + 243 K 2 M + (441 243 )M 2 .
(1)
Now the surface K = K crit (K 2 , M) intersects a plane of constant K 2 in a parabolic
curve, and, if our test path intersects this curve transversally at some point, 4 (R1 )
170 The swing-spring

will have a simple zero there and our conjectured scenario will be realized. Detailed
numerical calculations of all the relevant quantities show that this is precisely what
happens in our example for K 2 = 3. The second discontinuity in the example, at
s = s2 , is explained similarly: this time it is 2 that vanishes when the path crosses
(2)
the other critical surface, K = K crit (K 2 , M), with
(2)
K crit (K 2 , M) = K 2 243 K 2 M + (441 243 )M 2 .
From the above discussion, the following general situation, illustrated in
Figures 6.15 and 6.16, emerges. The vanishing of 1 or 2 occurs only at those
values of the integrals lying on one of the critical surfaces, indicated by the dotted
lines in the figures. Each time that a boundary-avoiding test path in E crosses a crit-
(i) (i)
ical surface K = K crit (K 2 , M) in the direction of increasing (decreasing) K K crit ,
the corresponding Ti increases (decreases) discontinuously by 2.
To test for monodromy, we need to construct a set of basis vectors ei for the
period lattice, which vary continuously along any boundary-avoiding path in E.
The vectors in (6.32) clearly do not satisfy this criterion, since e3 suffers disconti-
nuities e1 = (2, 0, 0) or e2 = (0, 2, 0) whenever the test path crosses a
critical surface. But these discontinuities can be erased by suitable lattice transla-
tions. Consider, for instance, the path along circle in Figure 6.16, traversed in the
counterclockwise sense starting at the 12 oclock position. At the initial point we
calculate e3 as prescribed in (6.32), with T1 and T2 calculated according to (6.31).
As we travel along , these quantities evolve continuously according to the differ-
ential equations (6.19) and (6.20), but then T2 suffers a 2 jump when the path
crosses the critical line near the 7 oclock position. But this jump can be nullified
by making a lattice translation: the new basis, continuous with the old one, has lat-
tice vectors (2, 0, 0), (0, 2, 0), and (T1 , T2 2, T ). In the same fashion, e3
evolves continuously across the next critical line at the 5 oclock position, becom-
ing (T1 + 2, T2 2, T ). This expression continues to hold as we arrive back
at 12 oclock. Evidently, the continuously evolving basis vectors are not single-
valued on , so the monodromy is nontrivial. This holds not only for , but for
any closed curve into which can be smoothly deformed without intersecting the
boundary of E. It is not difficult to see that this includes any closed, smooth curve
that surrounds the inner boundary surface.
For a curve such as on the second sheet, we have a different story. From the
geometry, the number of positive traversals of each critical surface is equal to the
number of negative traversals. At each crossing, we increment e3 by a lattice vector
to maintain continuity, but the net increment after going once around the circuit is
zero, and hence the monodromy is trivial. Obviously this property characterizes
any closed smooth path restricted to the second sheet, or any closed smooth path
on the first sheet that does not surround the inner boundary set.
6.7 Monodromy 171

1.4
K2 = 1

1.2

K 1.0

0.8

0.6

-1.0 -0.5 0.0 0.5 1.0


M

12 Sheet 1
K2 = 5
10

8
K
6
Sheet 2

0
-4 -2 0 2 4
M

Figure 6.15 In this plot of the K 2 = 1 and K 2 = 5 sections of the admissibility


manifold E, the critical lines, where 1 or 2 vanishes at the end of the allowed 3
interval, are shown as dotted parabolic arcs. Note that for K 2 = 5 the two sheets of
E are superposed. The portions of the critical lines within the curvilinear triangle
pertain only to the second sheet, while those outside the triangle are relevant only
to the first sheet. Typical non-boundary behavior of the function f (3 ) for (M, K )
on the parabolic arcs is shown at the right.

A tacit assumption in our discussion of monodromy is that the claimed continu-


ity is not destroyed by some peculiar behavior precisely on the critical surface. That
this might happen is suggested by our numerical calculation, in the example with
K 2 = 3 and a closed path like , showing that the value of T1 on the critical surface
172 The swing-spring

sheet 1 K2 K2*

K2 > K2*
sheet 2

Figure 6.16 Boundary-avoiding paths used to test for nontrivial monodromy are
shown in gray, both for the supercritical and for the subcritical cases. In cases
and , the period lattice basis can be modified to remove discontinuities on the
dotted lines, but then it fails to be single-valued on the closed path. All other cases
are either continuous and single-valued (if they do not intersect the dotted lines),
or can be modified to satisfy both continuity and single-valuedness.

(1)
K = K crit , computed using (6.31), differs from the value just below the surface by
, not 2. We cannot compensate for this jump with any lattice translation. But our
calculation of T1 on the critical surface might not be correct, since, as was pointed
out earlier, the critical surface is precisely where one finds a coordinate singularity.
Halfway through the orbit starting at 3 = R0 , 1 = 2 = /2, one lands on a point
with 3 = R1 = K 2 M and 1 = 0. The vanishing of 1 means that the angle 1
is not uniquely defined, and in fact a correct treatment of the dynamics (in a differ-
ent coordinate system) might require that 1 be incremented by a certain amount
as part of the turnaround. Such an increment, in addition to producing the correct
continuation of the orbit, would also increment the accumulated angle T1 by the
same amount. To check this possibility, we performed a high-precision numerical
integration of the orbit in the example, using the coordinates Q i , Pi , i = 1, 2, 3,
and fourth-order RungeKutta integration of Hamiltons equations. The answer is
very clear: the differential equations for 3 , 1 , and 2 need to be supplemented
by an increment of k by (modulo 2 ) whenever the orbit reaches a point with
k = 0. Since Tk gets incremented also on the critical surface, we now have no
problem assuring continuity of the basis along all closed smooth curves.

6.8 Periodic shift of the swing plane


A particularly fascinating aspect of the integrable swing-spring model is the
fact that the mechanism which gives rise to nontrivial monodromy also pro-
duces observable behavior in the laboratory frame of reference, namely a periodic
6.8 Periodic shift of the swing plane 173

alternation of swinging and springing motions, accompanied by precession of


a well-defined swing plane. To achieve really sharp quasi-planar swinging, it is
helpful to choose our integrals as follows:
(1)
0  K K crit (K 2 , M)  M  K 2  1. (6.34)

Let us examine how these inequalities promote the behavior we have described.
With K 2  1, all three of the variables i , i = 1, 2, 3, are small and we are in a
regime of small oscillations about stable equilibrium where the quadratic terms of
the Hamiltonian dominate. In this regime, the springing is only weakly coupled
to the swinging, with the former resembling, for short time scales, the motion of
an isotropic two-dimensional harmonic oscillator, i.e. motion on an origin-centered
ellipse with certain semi-axes a and eccentricity e. These parameters are expected
to change slowly with time, as higher-order effects come into play and energy is
exchanged with the springing mode.
In lowest perturbative order, M is the constant angular momentum. Choosing
this relatively small is responsible for the quasi-planarity of the motion in the labo-
ratory frame (when M is zero, the motion is strictly planar). This corresponds to an
eccentricity (most of the time) close to its maximum value of unity. One can show
(Exercise 6.13) that the precession rate for the instantaneous swing-plane axis is, in
lowest perturbative order, the quantity 2 1 , with the latter given by (6.19) and
(6.20). Since M = (2 1 )/2, the constraint M 0 will keep the swing plane
quasi-stationary most of the time.
Of course, there are exceptions to the last statement, namely where either 1 or
2 is close to zero. This is the reason for requiring that K be close to the critical
line. One can show that the instantaneous eccentricity in the lab frame is given, in
lowest perturbative order, by
 1/4
2 1
e= , = . (6.35)
+ 1 2
(1)
If K K crit (K 2 , M) 0, the eccentricity will remain close to unity for most of
the time (the fourth root helps here), except close to the turning point of the 3
oscillation, where the ratio 1 /2 dives to a minimum value close to zero, dragging
e down dramatically.
Before illustrating the above phenomena in a numerical example let us briefly
justify the claimed formulas for the precession rate and eccentricity of the
instantaneous ellipse. Since we are interested only in lowest-order effects, it
will be sufficient to work with a quasi-laboratory frame of reference related
to the Q 1 , Q 2 , Q 3 , P1 , P2 , P3 frame by the canonical transformation (compare
with (6.2))
174 The swing-spring

Q 1 + P2 Q 2 + P1 Q 3 P1 Q 2 P2 Q 1
(X, Y, Z , PX , PY , PZ ) = , , , , , 2P3 .
2 2 2 2 2
(6.36)
In the quasi-laboratory frame, the swinging motion is approximated by harmonic
oscillation in the X, Y plane with Hamiltonian
h = h X + hY ,
where
1 1
h X = (X 2 + PX2 ), h Y = (Y 2 + PY2 ).
2 2
From Hamiltons equations of motion,
X = PX , Y = PY , PX = X, PY = Y,
one easily verifies the conservation of the quantities
h X , h Y , l = X PY Y PX , m = X Y + PX PY .
The orbit with initial conditions (X 0 , Y0 , PX 0 , PY 0 ) is given by
X = (X, Y ) = (X 0 cos t + PX 0 sin t, Y0 cos t + PY 0 sin t),
which, together with the conservation equations, implies

2h Y m
XMX=l , M= 2
. (6.37)
m 2h X
This is the equation of an ellipse whose principal semi-axes a can be obtained by
means of diagonalizing the matrix M:
M a = a ,
where

= h X + h Y (h X + h Y )2 l 2
and
|a |2 = l 2 .
In terms of the quantities 1 = Q 21 + P12 = h X + h Y l and 2 = Q 22 + P22 =
h X + h Y + l, the eigenvalues and lengths of the semi-axes take the forms
1
= ( 1 2 )2
2
and

l 2
a = |a | = .
| 1 2 |
6.8 Periodic shift of the swing plane 175

The eccentricity of the ellipse, e = 1 a2 /a+2 , is then easily seen to be given by
(6.35). The proof that the polar angle of the vector a+ is given, up to a constant, by
2 1 is left as a straightforward exercise (Exercise 6.13).
To illustrate the existence and periodic advance of the swing plane in the quasi-
laboratory frame, we performed a high-precision numerical integration of the
equations of motion for 3 , 1 , and 2 for the following set of parameters, which
is consistent with the inequalities (6.34):

K 2 = 102 , M = 108 , K = K (1) (K 2 , M) + 1016 .

In Figure 6.17 we show the eccentricity, given by (6.35), and the swing-plane
angle (modulo a constant phase), given by (t) = 2 (t) 1 (t), over the
time interval 0 t 4T . We note how the eccentricity remains very close
to unity (corresponding to planar swinging) except near the turnaround times
T /2 + nT, n = 0, 1, 2, 3, where it dips rapidly to a minimum value close to zero.
Meanwhile, the swing-plane angle shifts by an amount approximately equal to
T2 T1 . The X, Y projection of the motion in the quasi-laboratory frame, shown in
Figure 6.18 for a time interval 17T , confirms our interpretation of the phenomenon.
As a final check, let us perform, with computer assistance, the canonical trans-
formation to the original laboratory frame. This gives, up to terms of degree 3 and
higher,
1
x=X+ (9PX PZ + 6X Z ) + ,
32
1
y = Y + (9PY PZ + 6Y Z ) + ,
32
3
z = Z + (3(PX2 + PY2 ) + 5(X 2 + Y 2 )) + ,
64

1.0
200 400 600 800 1000 1200

2
0.8
4
0.6
e
6

0.4 8

10
0.2
12

200 400 600 800 1000 1200

t t

Figure 6.17 Time evolution over four periods of the eccentricity e and the swing-
plane angle in the quasi-laboratory frame.
176 The swing-spring

0.10

0.05

0.00

0.05

0.10

0.10 0.05 0.00 0.05 0.10

Figure 6.18 The footprint of the projection of the orbit in the X, Y plane over 17
oscillation periods.

3
px = PX (5X PZ + 2Z PX ) + ,
32
3
p y = PY (5Y PZ + 2Z PY ) + ,
32
3
pz = PZ (X PX + Y PY ) + .
16
In our example, the choice K 2 = 102 restricts the size of X, Y, Z , PX , PY , PZ
to be at most of order 101 , so the pointwise relative corrections of degree 2 and
higher can be expected to be at most a few percent. This is in fact borne out by our
numerical calculations. To the eye, Figure 6.18 and the analogous plot obtained by
canonical transformation to the true laboratory frame are indistinguishable.

6.9 The swing-spring in molecular modeling


The dynamics of the swing-spring has attracted a great deal of attention in recent
years, thanks to the fact that it approximates very well the dynamics of certain
oscillations of the CO2 molecule. The latter has six internal degrees of freedom,
two of which, symmetric stretching and doubly degenerate bending, have a reduced
Exercises 177

Hamiltonian (averaging over the remaining rotational and vibrational degrees of


freedom) very similar to the Hamiltonian K we have been studying [41]. In partic-
ular, the quadratic terms, with a 1:1:2 resonance, and cubic terms (so-called Fermi
interaction) are essentially the same. The model is integrable and exhibits nontriv-
ial monodromy and swing-plane switching. Both these properties of the classical
model can be translated into the quantum domain. The nontrivial monodromy
corresponds to an obstruction to a uniform classification scheme for low-lying
quantum states [41], whereas the plane-switching can be seen in the propagation
of suitably constructed wave packets [42].

Exercises
6.1 Obtain Poincar-map phase portraits of the swing-spring constrained to
remain in the x, z plane, for H = 0.5, 1.5, and 4.0.
6.2 Fill in the details of the calculations of S3 and S4 in (6.3) and (6.10).
6.3 Verify (6.13).
6.4 Verify (6.14).
6.5 Verify (6.16) and (6.17).
6.6 Verify (6.21).
6.7 Verify (6.27).
6.8 By means of a Mathematica program using the function Manipulate, ver-
ify the behavior of f (3 ) shown in Figure 6.6. Repeat for K 2 = 1 and K 2 = 5.
Note that the panel shown in Figure 6.5 can be modified to allow user control
of parameters other than K , for example the maximum allowed values of 3
and f .
6.9 Show that (6.14) and (6.28) are identical.
6.10 Verify the statements in the paragraph beginning Referring to Figure 6.10.
6.11 Reproduce Figure 6.13. Show convergence to 2 for the jumps as one
increases the numerical precision. Show that, at the discontinuity points, the
Tk take on values midway between the left- and right-hand limits.
6.12 Obtain analogues of Figure 6.13 for curves , , and in Figure 6.16.
6.13 Show that the precession rate for the instantaneous swing-plane axis in the
quasi-laboratory frame of reference is, in lowest perturbative order, 2 1 .
6.14 Obtain the analogue of Figure 6.18 in the Q, P canonical frame. Describe the
behaviors of the instantaneous swing-plane angle and eccentricity as func-
tions of time. Dont be surprised if you find that the motion looks quite
different in the two canonical frames.
Appendix
Mathematica samples

In this brief appendix, we include a few samples of the Mathematica programming


which we have employed in various examples throughout the book. It is hoped
that they will allow the enterprising reader a chance, once he or she has acquired a
modicum of familiarity with Mathematica, to obtain a deeper understanding of the
principles and methods developed in the text, and also to apply them, with suitable
modifications, in a wide variety of applications.

A.1 Numerical integration of equations of motion


The equations of motion of an autonomous dynamical system take the general form

z = f (z), (A.1)

where z is a d-component array of real (or sometimes complex) numbers, and f


is a d-component array of smooth functions. Here d is the dimension of phase
space, which is equal to twice the number of degrees of freedom for a Hamiltonian
system. Non-autonomous equations z = g(z, t) can always be converted to the
form (A.1) by appending to z a new coordinate , writing

z = g(z, ), = 1.

The fundamental problem of dynamics is to find z(t), given z(t0 ). Numerical


analysis provides a number of methods for solving this problem iteratively with
controllable errors. If one has a simple mapping that takes z(t) into z(t + t),
where t is small, then, by iterating the process T /t times, one can calculate
z(t + T ), with an error that in most cases scales as a positive power of t. For most
of the examples in this book, including the numerical exercises, a reliable choice
for the single-step mapping z(t)  z(t + t) is the fourth-order RungeKutta
map, expressed as the following Mathematica function:

178
A.2 Hyperion movie 179

RK4step[f_, dt_, z_] : = Module[{k1, k2, k3, k4},


k1 = f[z] dt;
k2 = f[z + k1/2] dt;
k3 = f[z + k2/2] dt;
k4 = f[z + k3] dt;
z + (k1 + 2 k2 + 2 k3 + k4)/6]
Here z is a d-component array {z 1 , . . . , z d } and f is a user-specified function on
such objects, while dt is the time step. The quantity RK4step( f, dt, z) reproduces
the Taylor series of z(t + dt) up to terms of degree 4. The reader is invited to verify
this for d = 1 using the differential equation to express multiple time derivatives
of z in terms of derivatives of f (z), and Taylor-expanding RK4step( f, dt, z) using
the built-in Mathematica function Series[]. For d = 1 one could, instead of
using the RungeKutta method, simply use the truncated Taylor series itself. For
larger d, the Taylor expansion becomes unmanageable (because of all the sums and
products of partial derivatives), and the advantage of the RungeKutta method is
considerable.
It is usually a good idea to test the robustness of numerical integrations by
increasing the numerical precision. In Mathematica, one can easily accomplish this
by using the function Set Precision and changing the value of the constant
$MinPrecision.
The simplified simulation of Hyperions rotational motion, which is explored
in the next section, is a good example of RungeKutta numerical integration used
to construct discretized phase-space orbits and, from them, animated sequences of
Poincar sections.

A.2 Hyperion movie


The following Mathematica program was used to generate the sequence of
Poincar sections shown in Figure 5.19. An even better view, which vividly illus-
trates the mixing action within the chaotic layer, can be obtained by animating
a sequence of N sections, one for each RungeKutta time step. The reader is
encouraged to try out different parameter values and initial conditions.

Constants
e = 0.1; ( eccentricity )
Iratio = 0.264;
a = 1/Sqrt[1 - e2];
b = 1.5 a Iratio;
pi = N[Pi];
180 Mathematica samples

Functions
r[eta_] := (1 - e2)/(1 + e Cos[eta]); ( Kepler orbit )
fHyp[{u_, v_, eta_}] := ( r.h.s. of rotational equation )
{a r[eta]2 v, -b Sin[2 u - 2 eta], 1};

Parameters
slices = 200; ( # of sections )
dt = 2 pi/slices; ( step size )
orbits = 30; ( # of orbits )
itns=1000; ( # of Kepler periods )

Initial points
initz = N[{{0, 0, 0}, {0, 1/7, 0}, {0, 9/28, 0}, {0, 3/7, 0},
{0, 15/28, 0},{0, 6/7, 0}, {0, 8/7, 0}, {0, 10/7, 0},
{0, 12/7, 0}, {0, 2, 0}, {0, 16/7, 0}, {0, 5/2, 0},
{0, 37/14, 0}, {0, 39/14, 0}, {0, 81/28, 0},{0, 3, 0},
{pi/2,1/7,0},{pi/2,2/7,0},{pi/2,3/7,0},{pi/2,53/28,0},
{pi/2,57/28,0},{pi/2,29/14,0},{pi/2,18/7,0},{pi/2,20/7,0},
{pi,9/28,0},{pi,3/7,0},{pi,15/28,0},{pi,6/7,0},
{pi,16/7,0},{pi,37/14,0}}];

Poincar sections
For[n = 0, n < orbits, n++,
z = initz[[n + 1]];
tbl[n] = Transpose[Table[
z = RK4step[fHyp,dt, z];
{Mod[z[[1]], 2 pi], z[[2]]},
{itns}, {slices}]]];
Note how the Transpose function converts a list of orbits into a list of
sections, making use of every intermediate point in the RungeKutta integration.

Animation
For[t=1, t<=slices, t++,Print[
ListPlot[Join @@ Table[tbl[n][[t]], n, 0, orbits - 1],
PlotStyle -> Black, PointSize[.003], Axes -> False,
AspectRatio -> 1, PlotRange -> {{0, 2 pi}, {0, 3}} ]]

The sequence of 200 pictures can be animated within Mathematica (select them
and apply menu command Graphics/Rendering/Animate Selected Graphics), or
A.3 Simple pendulum perturbation theory 181

saved and played as a QuickTime movie. The options of the ListPlot can
obviously be modified to suit the users taste.

A.3 Simple pendulum perturbation theory


The following is a Mathematica implementation of the recursive perturbation
algorithm described in Section 4.2. In executing the program, the reader should
first choose the order to be 1 or 2 before embarking on a more time-consuming
higher-order calculation.
order = 10;
hamH = p2/2 + q2/2 +
Sum[(-eps)(n - 1) q(2 n)/Factorial[2n], n, 2,
order + 1];
hamK = Normal[ Series[ TrigReduce[ hamH
/. {q -> Sqrt[2 J] Cos[theta], p -> -Sqrt[2 J ]
Sin[theta]}], {eps, 0, order} ] ];
ham0[0] = J;
For[n = 0, n <= order - 1, n++,
hamK = hamK /. {J1 -> J, theta1 -> theta};
ham = D[hamK, eps, n+1]/Factorial[n+1] /. eps -> 0;
ham0[n+1] = Expand[Integrate[ham, theta, 0,
2 Pi]/(2 Pi)];
If[n == order-1, Break[]];
genS[n] = Integrate[Expand[ham0[n+1] - ham],
theta] /. J -> J1;
hamK = Normal[ Series[hamK
/. J -> J1 + eps(n+1) D[genS[n],theta],
{eps, 0, order}]];
hamK = FixedPoint[TrigReduce[Normal[Series[#
/. theta -> theta1 - eps(n+1) D[genS[n],J1],
{eps, 0, order}]]] &, hamK]];
Sum[ham0[k], {k, 0, order}]

A.4 BirkhoffGustavson perturbation theory


The following Mathematica program was used to implement the Birkhoff
Gustavson perturbation algorithm in the HnonHeiles model in Section 4.6. The
same algorithm, with a different Hamiltonian, was applied to the swing-spring in
Chapter 6.
182 Mathematica samples

Perturbation parameter
L/: L(n_/;n>m):=0
m=10

Canonical transformation
transform[h_,s_,w_]:=
Module[{h1,h2,h3,w1,u1,u2,v1,v2,wu1,wu2,wy1,wy2,zu1,
zu2,zy1,zy2},
h1=ExpandAll[h];
h2=h1/.L(n_/;n>(m-s+2))->0;
h3=h1-h2;
h2=h2/.{x1->u1,x2->u2,y1->v1,y2->v2};
w1=ExpandAll[w]/.{x1->u1,x2->u2};
zu1=D[w1,u1]; zu2=D[w1,u2];
zy1=D[w1,y1]; zy2=D[w1,y2];
subst[mon_]:= ExpandAll[mon/.
{u1 -> x1 - L(s-2) wy1, u2 -> x2 - L(s-2)
wy2,
v1 -> y1 + L(s-2) wu1, v2 -> y2 + L(s-2)
wu2}]
/.{wu1 -> zu1,wu2 -> zu2, wy1 -> zy1,wy2 -> zy2};
Collect[h3 + FixedPoint[subst,#]& /@ h2, L]]

Inverse transformation
inverse[h _ ,s _ ,w _ ]:=
Module [{h1,h2,h3,w1,u1,u2,v1,v2,wx1,wx2,wv1,wv2,zx1,zx2,
zv1,zv2},
h1=ExpandAll[h];
h2=h1/.L(n _ /;n>(m-s+2))->0;
h3=h1-h2;
h2=h2/.{x1->u1,x2->u2,y1->v1,y2->v2};
w1=ExpandAll[w]/.{y1->v1,y2->v2};
zx1=D[w1,x1]; zx2=D[w1,x2];
zv1=D[w1,v1]; zv2=D[w1,v2];
subst[mon_]:=ExpandAll[mon/.
{u1 -> x1 + L(s-2) wv1, u2 -> x2 + L(s-2) wv2,
v1 -> y1 - L(s-2), wx1, v2 -> y2 - L(s-2) wx2}]
/.{wx1 -> zx1, wx2 -> zx2,wv1 -> zv1, wv2 -> zv2};
Collect[h3 + FixedPoint[subst,#]& /@ h2, L]]
A.4 BirkhoffGustavson perturbation theory 183

Generating function
index[mon _ ]:= (q1 D[mon,q1] + q2 D[mon,q2] -
p1 D[mon,p1] - p2 D[mon,p2])/mon/.
{q1->0,q2->0,p1->0,p2->0}
wcalc[h_,s_] := Module[{h1},
h1=ExpandAll[h]/.L(n_/;n!=s)->0;
h1=ExpandAll[h1/.{x1->q1+I p1,x2->q2+I p2,
y1->p1+I q1,y2->p2+I q2}];
h1=If[index[#]==0,0,I mon/index[#]]& /@ h1;
h1=ExpandAll[h1/.{q1->x1-I y1,q2->x2-I y2,
p1->y1-I x1,p2->y2-I x2}]]/.L->1/2]

Calculation of second integral


henon= (1/2) (x12 + x22 + y12 +y22) L2 + (x12 x2
- (1/3) x23) L3;

oscillator= (1/2) ( x12 + x22 + y12 + y22 ) L2;

h = henon;
wlist[3] = wcalc[h,3];
For[s = 4, s <= m, s++,
h = transform[h,s-1,wlist[s-1]];
wlist[s] = wcalc[h,s]];
int2 = oscillator;
For[s = m, s >= 3, s-, int2 = inverse[int2, s, wlist[s]]];
ExpandAll[(int2 - henon)/.L->1]]
References

[1] H. Goldstein, C. Poole, and J. L. Safko, Classical Mechanics (3rd edn.), San
Francisco, CA, Addison-Wesley, 2002
[2] E. T. Whittaker, A Treatise on Analytical Dynamics of Particles and Rigid Bodies,
Cambridge, Cambridge University Press, 1988
[3] M. Abramowitz and I. A. Stegun, Handbook of Mathematical Functions, New York,
Dover, 1965
[4] L. D. Landau and E. M. Lifshitz, Mechanics, Oxford, Pergamon Press, 1960
[5] J. L. Lagrange, uvres, vol. 6, Paris, 1873, pp. 272292
[6] J. Liouville, J. Math. Pures Appl. 20 137138 (1855)
[7] V. I. Arnold, Mathematical Methods of Classical Mechanics, New York, Springer-
Verlag, 1978.
[8] V. I. Arnold, Ordinary Differential Equations, trans. R. A. Silverman, Cambridge,
MA, MIT Press, 1973
[9] K. Efstathiou, M. Joyeux, and D. A. Sadovskii, Phys. Rev. A 69 032504 (2004)
[10] J. J. Duistermaat, Commun. Pure Appl. Math. 33, 687706 (1980)
[11] A. Einstein, Verh. Deutsch. Phys. Ges. 19 82 (1917)
[12] L. Brillouin, J. Phys. Radium 7 353 (1926)
[13] J. B. Keller, Ann. Phys. (N.Y.) 4 180 (1958)
[14] J. B. Keller and S. I. Rubinow, Ann. Phys. (N.Y.) 9 24 (1960)
[15] R. H. Cushman and J. J. Duistermaat, J. Diff. Eqns. 172 42 (2001)
[16] M. Toda, Prog. Theor. Phys. Suppl. 45, 174 (1970)
[17] A. J. Lichtenberg and M. A. Lieberman, Regular and Chaotic Dynamics (2nd edn.),
New York, Springer-Verlag, 1992
[18] M. Hnon, Phys. Rev. B 9 1921 (1974)
[19] W. Heisenberg, Z. Phys. 33 879 (1925)
[20] G. S. Ezra, K. Richter, G. Tanner, and D. Wintgen, J. Phys. B 24 L413 (1991)
[21] C. F. F. Karney and A. Bers, Phys. Rev. Lett. 39 550 (1977)
[22] G. D. Birkhoff, Dynamical Systems, New York, American Mathematical Society,
1927
[23] F. Gustavson, Astron. J. 71 670 (1966)
[24] M. Hnon and C. Heiles, Astron. J. 69 73 (1964)
[25] L. S. Hall, Physica D 8 90116 (1983)
[26] A. N. Kolmogorov, Dokl. Akad. Nauk SSSR 98 527 (1954)
[27] J. K. Moser, Nach. Akad. Wiss. Gttingen II Math.-Phys. Kl. 1 (1962)
[28] V. I. Arnold, Russ. Math. Surveys 18 No. 6, 85191 (1963)

184
References 185

[29] J. V. Jos and E. J. Saletan, Classical Dynamics: a Contemporary Approach,


Cambridge, Cambridge University Press, 1988
[30] V. I. Arnold, Sov. Math. Dokl. 2 247249 (1960)
[31] A. M. Leontovich, Sov. Math. Dokl. 3 425429 (1962)
[32] A. Celletti and L. Chierchia, KAM Stability and Celestial Mechanics, New York,
American Mathematical Society, 2007
[33] G. M. Zaslavskii, R. Z. Sagdeev, D. A. Usikov, and A. A. Chernikov, Weak Chaos
and Quasiregular Patterns, Cambridge, Cambridge University Press, 1991
[34] G. M. Zaslavsky, M. Edelman, and B. A. Niyazov, Chaos 7 159181 (1997)
[35] G. M. Zaslavsky and B. A. Niyazov, Phys. Rep. 283 7393 (1997)
[36] J. Wisdom, S. J. Peale, and F. Mignard, Icarus 58 137152 (1984)
[37] J. J. Klavetter, Astron. J. 97 570579 (1989)
[38] A. V. Devyatkin, D. L. Gorshanov, A. N. Gritsuk, A. V. Melnikov, M. Yu. Sidorov,
and I. I. Shevchenko, Solar System Res. 36 248259 (2002)
[39] V. V. Kouprianov and I. I. Shevchenko, Icarus 176 224234 (2005)
[40] H. Dullin, A. Giacobbe, and R. Cushman, Physica D 190 1537 (2004)
[41] R. H. Cushman, H. R. Dullin, A. Giacobbe, D. D. Holm, M. Joyeux, P. Lynch, D. A.
Sadovskii, and B. I. Zhilinskii, Phys. Rev. Lett. 93 024302 (2004)
[42] M. Sanrey, M. Joyeux, and D. A. Sadovskii, J. Chem. Phys. 124 074318 (2006)
Index

action-angle variables, 59, 60, 166 generating function, 34


generating function, 59 examples, 37
atlas of charts, 5 explicit time dependence, 39
types 14, 34
bead on a rotating circle, 15 gyrating charge
BirkhoffGustavson perturbation theory, 109, 111, in electrostatic wave, 104
119, 128, 144, 152, 181
Hamiltons equations, 9
canonical momentum, 8 canonical invariance, 31
canonical transformation, 30 concise form, 9
conserved Poisson brackets, 31 HamiltonJacobi, 59, 60, 62, 92
continuous groups, 40 generating function, 58
examples, 32 Hamiltonian, 8
explicit time dependence, 39 conserved, 9
generating function, 34 time evolution, 9
preservation of Hamiltons equations, 31 Hamiltonian flow, 10
simple examples, 32 incompressibility, 10
chaos, 108, 115, 130, 131, 135 Hamiltonian formalism, 29
chaotic moon, 139 harmonic oscillator, 11, 60
chart, 5 action-angle approach, 60
configuration space, 2, 5 HamiltonJacobi approach, 60
constraints, 1, 2, 4 HnonHeiles model, 110, 181
heteroclinic intersection, 135
degree of freedom, 1, 15, 18 Hyperion model, 139, 179
diffeomorphism, 5 chaotic rotation, 139
diffusion coefficient, 134 instability in 3D, 142
movie, 179
elastic pendulum, 27, 148 Poincar section, 142
electrodynamics
Hamiltonian formulation, 42 independence of functions, 4, 5, 62, 64, 74, 79, 88, 95,
elliptic fixed point, 122 158, 163
elliptic function, 26 inertia tensor, 23
elliptic integral, 15, 26 integrable systems, 56
energymomentum diagram, 70, 74, 79, 83, 89 examples
energymomentum map, 158, 160 circular stadium billiard, 73
equations of motion, 178 free 2D particle, 70, 71
EulerLagrange equations, 7 free-particle examples, 69
HnonHeiles type, 112
fast track to action-angle variables, 57 one degree of freedom, 60
fiber bundle, 10, 158 nonseparable, 85
flow coordinates, 71, 72, 80, 165 separability, 57

186
Index 187

spherical pendulum, 78 rotation, 13


islands of stability, 131 separatrix, 13, 80, 163
period lattice, 64, 66, 68, 72, 73, 77, 82, 84, 90, 165
Jacobi elliptic integral, 15 perturbation theory
Jacobi identity, 30, 54 BirkhoffGustavson expansion, 109
canonical, 97, 181
KAM theorem, 121, 122 coupled oscillators, 102
nonlinear stability, 122 gyrating charge, 104
Kepler problem, 61 simple pendulum, 98
kicked oscillator, 129 phase portrait, 13
phase space, 10
Lagrange bracket, 32 volume, 38
canonical invariance, 32 phase-space velocity, 10
fundamental, 33 Poincar invariant, 34
inverse of Poisson bracket, 33 Poincar section, 86, 87, 106, 108, 111, 115, 130, 131,
Lagrange points, 48, 127 138, 142, 144, 146, 150, 177, 179
linear stability, 48, 52 Poisson bracket, 29
nonlinear stability, 127 angular momenta, 41
Lagrangian mechanics, 7 antisymmetry, 30
linear oscillator, 44 distribution over a product, 30
linear stability, 44, 48, 52 fundamental, 30
linearized map, 137 Jacobi identity, 30, 54
LiouvilleArnold theorem, 56, 64, 73, 74, 80, 82, 85,
90, 93, 95, 122, 156 quasiperiodic motion, 108
action-angle variables, 68
multichart, 66 resonance zone, 124, 125
period lattice, 66 restricted three-body problem
proof, 64 inertial frame, 48
topological structure, 68 Lagrange points, 50
Lyapunov exponent, 132, 136 linear stability, 52
normal modes, 53
manifold rotating frame, 49
differential, 5 rigid body, 21
Mathematica, 2628, 90, 101, 111, 113, 142, angular coordinates (Euler angles), 21
144146, 158160, 178, 179, 181 angular velocity, 22
matrix , 9 degrees of freedom, 21
monodromy, 83, 170 inertia tensor, 23
spherical pendulum, 83 Lagrangian, 23
swing-spring, 170
multichart, 6, 64, 66, 68, 83 scalar function
time evolution, 29
Newtonian mechanics, 1 sensitivity to initial conditions, 132, 145
normal modes, 44, 45 separable systems, 57
triatomic molecule, 46 HamiltonJacobi theory, 59
variables a, a , 45 one degree of freedom, 60
numerical integration, 178 separatrix, 13, 80, 163
RungeKutta method, 178 small oscillations, 44
triatomic molecule, 46
particle on a sphere, 2 small-denominator problem, 124, 142
pebble on an ice ball, 20 spherical pendulum, 18, 78
pendulum, 11, 181 action-angle variables, 82
double, 2 constraint force, 20
elastic, 148 energymomentum diagram, 79
exact solution, 14 Hamiltonian, 19
libration, 13 Lagrangian, 18
linear stability, 13 monodromy, 83
perturbation theory, 98, 181 reduced equation of motion, 19
phase portrait, 13 stability
188 Index

Lyapunov, 122 small oscillations, 152


nonlinear, 122, 126, 127 swing-plane shift, 172
stable manifold, 134, 135 three commuting integrals, 154
stationary group, 66 two-dimensional model, 149
stochastic web, 130, 131 constant-energy manifold, 149
chaotic orbits, 131 Poincar section, 150
diffusion, 134 symmetric top, 23, 156
superdiffusion, 134 constants of the motion, 25
stochastic web-map equation for u = cos , 25
Lyapunov exponent, 138 Hamiltonian, 25
swing-spring, 148 Lagrangian, 23
action-angle variables, 166 nutation period, 25
admissible integrals, 158 precession, 25
application in molecular modeling, 176 reduced equation of motion, 25
bifurcation point, 164
BirkhoffGustavson expansion, 152 three-body problem
chaos, 149 restricted, 48
constraints among the integrals, 158 Toda model, 85
dynamics on the level sets, 156 constant-energy phase space, 86
flow coordinates, 165 energymomentum diagram, 89
image E of energymomentum map, 158, 160 Hnons second integral, 88
integrable approximations, 152 LiouvilleArnold approach, 90
manifold E, 158 period lattice, 90
monodromy, 170 total energy, 10
period lattice, 165
Poincar section, 2D, 150 unstable manifold, 134, 135