Professional Documents
Culture Documents
Essentials of Hamiltonian Dynamics - John H.lowenstein
Essentials of Hamiltonian Dynamics - John H.lowenstein
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
A catalog record for this publication is available from the British Library
Preface page xi
vii
viii Contents
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
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
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.
X = x = (x, y, z)
is constrained by
f (X, t) = x 2 + y 2 + z 2 a(t)2 = 0.
n = 3 1 = 2.
X = (x1 , y1 , z 1 , x2 , y2 , z 2 ) R6 ,
1.2 Configuration space 3
^r
a(t) ^
y
1 a
b
2
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.
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.
qn. .q. 3 q2
(U)
q1 Q
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.
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
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)) ,
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 )
(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)
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.
3 2
2n
H
=E
)
H(
= H
1
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
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.
= p ,
p = g sin .
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 .
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)
(t)
motor
the orthogonal pair of unit vectors and pointing in the directions of increasing
latitude and longitude, respectively, we have
x = + sin (t).
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.
x = a r + a + a sin , (1.23)
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 .
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.
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
= + z + r.
z = cos r sin ,
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 ,
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
p = I ,
p = I sin2 + Ir cos ( + cos ),
p = Ir ( + cos ),
1.5 Examples 25
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
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
= (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)
[qi , q j ] = 0 = [ pi , p j ] = 0, [qi , p j ] = i j , i, j = 1, 2, . . . , n,
[k , l ] = kl .
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.
or, equivalently,
[
i ,
j ] =
i
j = i j , (2.5)
[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
[Q k , Pl ] = Mki [ pi , p j ]M 1 1
jl = Mki i j M jl = kl .
the subscript will be suppressed unless needed for clarity. The fundamental
Lagrange-bracket relations are
{qk , ql } = 0 = { pk , pl }, {qk , pl } = kl ,
{
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 )
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
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),
H (q, p) = p + cos q.
38 The Hamiltonian formalism
(, 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
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
Lx = Ly = Lz = 0.
[L x , L y ] = L z , [L y , L z ] = L x , [L z , L x ] = L y .
(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
From (2.15), we see that the desired must be an eigenvector of C, with eigenvalue
= i satisfying the equation
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.
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.
j C n+ j = N j .
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
m1 k1 m2 k2 m3
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).
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
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.
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.
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
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
56
3.2 Fast track for separable systems 57
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.
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
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)):
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
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
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 .
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
F = (F1, F2)
v2
v1 0
g2t2 0
g1t1 0
0 N
Figure 3.2 Any two points on Mf can be linked by a finite number of group
operations.
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
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
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
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.
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
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.
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.
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 .
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)
mg
E
2
0
L
2 1 1 2
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
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 .
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
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
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
E=1 E = 256
1.5
20
1
10
0.5
0 0
0.5
10
1
20
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).
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
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 .
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.
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
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 ,
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
x = a sinh cos ,
y = a sinh sin ,
z = a cosh ,
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
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
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,
97
98 Canonical perturbation theory
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.
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)
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
and, recursively,
S (m)
m m+1 (m , Jm+1 )
Jm+1
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
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
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.
K = 1 J1 + 2 J2 + 1 2 J1 J2 + O( 2 ).
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
(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.
e 0 sin(ky t)
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 ),
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
S 1 Jm (k )
J1 = P = mc + e 0
2
m sin( m ).
2 m=
+ c
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
and functions
Figure 4.3 Numerical Poincar sections for frequency ratio /c = 30.11 and
dimensionless amplitudes a = 0.01 (upper) and a = 0.1 (lower).
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.
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
0.2
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.
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.
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 (, J ) = J + H1 (, J ).
H = J + H1 ( J) + H2 (, J, ),
(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
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.
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
J2
T
dA(J)
dJ A
1 J = A1
J1
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
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)
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
J2
J = A1
J1
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.
The generating function of the first of these transformations will be of the form
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 )
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
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,
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
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.
F
F()
DF() .
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:
The average expansion rate over the orbit starting at point is defined as
1 | D F k ( ) |
( ) = max lim ln .
k k ||
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
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 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
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
(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.
(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
x
N N
S S
x = v, v = x x 3 + a cos t.
x = v, v = x x 3 + a cos , = 1. (5.8)
148
6.1 Two-dimensional motion 149
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
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.
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
H = 2.875 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
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
H = H2 + H3 + ,
where
3
H3 = ( 2 + 2 + p2 + p2 + 2 p + 2p ).
4 2
6.2 Integrable approximations for small oscillations 153
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
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
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).
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 ).
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
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
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
-1
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.
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
3.10
3.00
K2
2.90
2.80
60
40
20
0
5
10
4 = 1
3 ((K 2 3 )2 M 2 )3 .
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
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
12
D F
K2 8 B B+
4
0
0 10 20 30 40 50 60
K
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
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
so that
def 1
J3 = P dQ
2 C3
1 R1
d g() f ()
= .
3 R0 (2J1 )(2J2 ) (2J1 )(2J2 )
168 The swing-spring
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:
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
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 .
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
12 Sheet 1
K2 = 5
10
8
K
6
Sheet 2
0
-4 -2 0 2 4
M
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.
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):
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
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
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.
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
z = f (z), (A.1)
z = g(z, ), = 1.
178
A.2 Hyperion movie 179
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.
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]
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
186
Index 187