Chaos in Hamiltonian systems

Teemu Laakso
April 26, 2013
Course material:
Chapter 7 from Ott 1993/2002, Chaos in Dynamical Systems, Cambridge
http://matriisi.ee.tut.fi/courses/MAT-35006
Useful reading:
Goldstein 2002, Classical Mechanics, Addison Wesley
Hand & Finch 1998, Analytical Mechanics, Cambridge
Advanced:
Lichtenberg & Lieberman 1992, Regular and Chaotic Dynamics, Springer
Arnold 1989, Mathematical Methods of Classical Mechanics, Springer

Contents
1 Introduction: classical mechanics

2 Symplectic structure

3 Canonical changes of variables

4 Hamiltonian maps

5 Integrable systems

10

6 Perturbations and the KAM theorem

13

7 The fate of resonant tori

15

8 Transition to global chaos

17

Introduction: classical mechanics

Newtonian mechanics
Consider a particle of mass m, subject to a force field F , in a d-dimensional
Euclidean space (one-body system). Newtons second law; m
r = F = a
system of 2d first order ODEs for position r Rd and velocity r Rd . These
are equations of motion in phase space (the space Rd Rd where (r, r)
are
coordinates). The number of degrees of freedom of the mechanical system is
N = d.
Example (harmonic oscillator). Set d = 1, F = kr = r = kr/m
[draw a picture].
dr
= r
dt
dr
k
= r.
dt
m
p

The solution for initial values r(0) = r0 , r(0)

= 0 is r(t) = r0 cos t k/m .

p
[Characteristic equation z 2 + k/m = 0 = z1,2 = i k/m = ih =
r(t) = a cos h t + b sin h t.]

Lagrangian mechanics
For an unconstrained system of n bodies: N = dn (2N ODEs). Under holonomic constraints fj (r1 , r2 , . . . , rn , t) = 0, j = 1, . . . , k we can define generalized coordinates qi , i = 1, . . . , N , where N = dn k, using transformation
equations
r1 = r1 (q1 , q2 , . . . , qN , t)
..
.
rn = rn (q1 , q2 , . . . , qN , t).

(1)

P
Definition. The Lagrangian (function) is L = T V , where T = ni=1 mi (ri
ri )/2 is the kinetic energy, and V = V (r1 , r2 , . . . , rn , t) is the potential energy.

Through the transformation equations (1), we have L = L(q, q,

t). The
Hamiltons principle
Z t2
L(q, q,
t)dt
I = 0, I =
t1

is a variational equation for finding a path q(t) from t1 to t2 for which the line
integral I (action) is stationary [draw a picture]. Solution [e.g., Goldstein]
yields the (Euler-)Lagrange equations of motion:
d L L

= 0,
dt qi qi

i = 1, . . . , N.

[For all possible paths, the system takes the one requiring the least action.]
Example (harmonic oscillator). L = T V = mr 2 /2 kr2 /2,
d L
= m
r,
dt r

L
k
= kr = r = r.
r
m

Hamiltonian mechanics
Definition. The conjugate momenta pi = L/ qi and Hamiltonian H =
qi pi L (Einstein summation convention).
We do a Legendre transformation;
L
L
L
L
dqi +
dqi +
dt = pi dqi + pi dqi +
dt
qi
qi
t
t
L
dH = qi dpi pi dqi
dt
t
dL =

Since we wish H = H(q, p, t),

dH =

H
H
H
dqi +
dpi +
dt.
qi
pi
t

Equating terms, we have the Hamiltons equations of motion:

H
qi
H
qi =
.
pi

pi =

The (q, p) RN RN are canonical phase-space variables.

3

(2)

Example (harmonic oscillator). p = L/ r = mr,

H = mr 2 L = mr 2 /2 +
2
kr /2 = T + V = E.
In general, if the transformation equations (1) do not depend on t explicitly, and the forces are conservative (of the form F = ), Hamiltonian is
the total energy.
For a time-independent Hamiltonian H(p, q) the total energy is conserved;
H
H
H H H H
dH
=
q +
p =

= 0.
dt
q
p
q p
p q
= H(p, q) = E is a constant of motion.

Symplectic structure

The class of Hamiltonian systems is very special. Let x = (p, q)T and

f (x, t) =

H
x

T


0 I
, where =
.
I 0

Now, Hamiltons equations are x = f (x, t). Note that f (x, t) R2N (the
Hamiltonian vector field) is fully determined by H(x, t) R.
Theorem (Liouvilles theorem). Hamiltons equations preserve 2N -dimensional
volumes.
Proof.





H
H

f =

+
= 0,
x
p
q
q
p

I
I
Z 
Z

dx
d
2N
f dS =
d x=
dS =
f d2N x = 0.
dt V
dt
x
S
S
V
[differentiation under integral sign, Hamiltons equations, divergence (Gauss)
theorem].
Corollary. There are no attractors in Hamiltonian systems.

Hamiltons equations are symplectic. Consider three orbits (p, q), (p +

p, q + q), and (p + p0 , q + q 0 ). The differential symplectic area is the sum
of parallelogram areas: [Ott, fig. 7.1]

N 
N
X
pi qi  X
 0

(pi qi0 qi p0i ) = p q 0 q p0 = xT x0 .
pi qi0  =
i=1

i=1

0

dxT
d
T
0
0
T dx
x x =
x + x
dt
dt
dt
T

f
f
x x0 + xT x0
=
x
x
" 
#
T
f
f
= xT
+
x0
x
x
#
"

T
2
2
H
H
+ 2 x0
= xT
2
x
x
"
#

T
2
2
H
H
= xT
T + 2 x0
2
x
x

= 0.
[linearization of f (x + x, t), = I, T = , 2 H/x2 symmetric].
Symplectic condition = conservation of volume.
Definition. The differential symplectic area is a differential form of Poincares
integral invariant
I
N I
X
p dq =
pi dqi ,

i=1

where is a closed path in (p, q) at constant t.

A generalization in extended 2N + 1 phase space (p, q, t) [Ott, fig. 7.2]:
Theorem (Poincare-Cartan integral theorem).
I
I
(p dq Hdt) =
(p dq Hdt),
1

where 1 and 2 are paths around the tube of trajectories.

5

1 and 2 at constant times (dt = 0) = Poincares integral

invariant.
H
If H = H(p, q) and 1 and 2 are on this surface, we have Hdt = 0 and
I
I
p dq =
p dq,
1

where 1 and 2 does not need to be at constant t.

Canonical changes of variables

Let us define a new set of phase-space variables X = (P, Q)T by using a

transformation g : R2N R2N , x 7 X.
Definition. The transformation g is canonical, if it preserves the differential
symplectic area;
p q 0 q p0 = P Q0 Q P 0 .

(3)

Canonical transformations are typically performed by using a generating

function, e.g., S = S(P, q, t);
S
,
P
S
p=
,
q

Q=

which gives (P, Q) implicitly. The transformed Hamiltonian is K(P, Q, t) =

H(p, q, t) + S/t. [Can be derived using the Hamiltons principle; see Goldstein.] In order to check the canonicity of the above one can substitute
2S
2S
P
+
q,
P 2
P q
2S
2S
p =
P + 2 q,
P q
q

Q =

into the condition (3) [homework].

The new variables satisfy Hamiltons equations: dX/dt = (K/X)T .
Hence, the underlying symplectic structure of a given Hamiltonian system
is invariant, and can be represented using any suitable choice of canonical
phase-space variables.
6

Hamiltonian maps

Consider the map Mh : R2N R2N :

Mh (x(t), t) = x(t + h).
Its differential variation [linearize Mh (x + x, t)] is
Mh
x(t) = x(t + h).
x
Symplectic condition =
xT (t)x0 (t) = xT (t + h)x0 (t + h)
T 


Mh 0
Mh
x(t)
x (t)
=
x
x

T 

Mh
Mh
T

x0 (t)
= x (t)
x
x

T 

Mh
Mh

= =
x
x
We say, Mh /x is symplectic. Generally, matrix A is symplectic, if =
AT A. The product of symplectic matrices A and B is symplectic;
(AB)T (AB) = B T (AT A)B = B T B = .
Theorem (Poincare recurrence theorem). Examine a Hamiltonian H(p, q)
with orbits bounded in a finite subset D R2N . Let a ball R0 D have
radius > 0. In time, some of the orbits leaving R0 will return to it.
Proof. Evolve R0 with the volume-preserving Mh , obtaining subsequent regions R1 , R2 , . . .. Conclude that Rr Rs 6= = Rrs R0 6= , r > s
[Arnold, fig. 51].

Poincar
e maps
In general, a Poincare map gives an intersection of an orbit with a lowerdimensional subspace of the phase space, called Poincare section (or surface
of section, SOS). For periodic orbits, plotting subsequent intersections typically draws a figure which reflects deeper characteristics of the dynamical
system.
7

Definition. Poincare section for a Hamiltonian system H(p, q, t) where H

is -periodic in time:
extended (2N + 1) -dimensional phase space (p, q, ), where = t,
additional equation d/dt = 1,
let 0 = mod ,
set 0 = t0 [0, ) as the 2N -dimensional SOS.
Because of the periodicity, M (x, t0 ) = M (x, t0 + k ), k Z. The
surface of section map xn+1 = M (xn ) = M (xn , t0 ) is symplectic.
Example (Standard map). The kicked rotor: bar of moment of inertia I
and length l, frictionless pivot [Ott, fig. 7.3]. Vertical impulse of strength
K/l at times t = 0, , 2, . . . Canonical variables (p , ), Hamiltonian
H(p , , t) =

X
p2
(t n ),
+ K cos
2I
n

where denotes the Dirac delta, and equations of motion (2)

X
dp
= K sin
(t n ),
dt
d
p
= .
dt
I
SOS: (p , ) after each kick. Integration over t (n, (n + 1) ] yields pn+1
pn = K sin n+1 and n+1 n = pn /I. Setting /I = 1, we have
pn+1 = pn + K sin n+1 ,
n+1 = (n + pn ) mod 2.
The mapping (N = 1) preserves differential areas, if
det(xn , x0n ) = det [(xn+1 /xn )(xn , x0n )] ,
where xn = (pn , n )T . Symplecticity can hence be verified by calculating




pn+1 /pn pn+1 /n
1 + K cos n+1 K cos n+1
det
= det
= 1.
n+1 /pn n+1 /n
1
1
8

Definition. Poincare section for a Hamiltonian system H(p, q):

motion in a (2N 1)-dimensional energy surface H(p, q) = E,
by choosing, e.g., q1 = 0, we determine a (2N 2)-dimensional SOS.
For a unique SOS mapping xn+1 = M (xn ), we often need an additional
constraint, e.g., p1 > 0.
Example (Logarithmic potential ). The potential near the centre of a flattened galaxy can be modelled as


y2
1
2
2
(4)
(x, y) = ln x + 2 + b ,
2
a
where a, b R are parameters [draw a picture]. For a star moving in this
potential, the Hamiltonian per unit mass is



1 2
y2
2
2
2
H(px , py , x, y) =
p + py + ln x + 2 + b
.
2 x
a
Equations of motion (2) become
px =
py =

x
x2

y 2 /a2

a2 (x2

+ b2

y
,
+ y 2 /a2 + b2 )

x = px ,
y = py .
Choose y = 0 as the SOS. Now


H(px , py , x, 0) = E = p2y = 2E p2x ln x2 + b2 .
Hence, there are two points py corresponding to y = 0. We choose that
py > 0 on the SOS. In this case, we cannot solve the equations of motion analytically, and points (px , x) on the SOS have to be determined by numerical
integration.

Integrable systems

Consider a Hamiltonian H(p, q). A constant of motion is a quantity that

does not change when the system evolves in time, e.g., H(p, q) = E. In
general, for a function f (p(t), q(t)), we have
df
f
f
f H f H
=
p +
q =

.
dt
p
q
q p
p q
If f (p, q) is a constant of motion, we can write [f, H] = 0, where the Poisson
bracket for functions f and g is defined as
[f, g] :=

f g f g

.
q p p q

Poisson bracket is anticommutative: [f, g] = [g, f ].

Theorem (Liouvilles theorem on integrable systems). The Hamiltonian system H(p, q) is integrable, if it has N independent constants of motion fi (p, q),
i = 1, . . . , N which are in involution, i.e., [fi , fj ] = 0, i, j.
[Proof: see, e.g., Arnold.] The independent constants of motion restrict
an integrable system on an N -dimensional surface in the phase space. Because the constants are in involution, any bounded orbit on this surface is
topologically equivalent to an N -dimensional torus [Ott, fig. 7.4a].

Action-angle variables
The action-angle variables provide an explicit transformation between points
(p, q) and points on the N -torus. Through a canonical change of variables,
it is possible to transform to coordinates (P, Q) in which the new momenta
Pi (p, q), i = 1, . . . , N are constants of motion. In such a case dPi /dt =
H/Qi = 0 = H = H(P ).
A particularly convenient choice is
I
1
p dq,
Pi = Ji :=
2 i
where the paths i wrap around each possible angle direction i = 1, . . . , N
on the N -torus [Ott, fig. 7.4b]. The new variables Ji are actions. They are
constants of motion by the Poincare-Cartan integral theorem.
10

The corresponding conjugate coordinates Qi = i are angles. Suppose

that action-angle variables (J, ) are obtained by the generating function
S(J, q);
S
,
J
S
p=
q
=

By integrating the latter equation around i , we examine the change in S;

I
i S =
p dq = 2Ji .
i

Then,
i =

i S = 2 Ji ,
J
J

or
i j = 2ij ,
where ij = 1, if i = j, and zero otherwise. Therefore, after one circuit
around i , the angle i increases by 2, and the other angles return to their
original values.
The new Hamiltonian is H = H(J), and
dJ
= 0,
dt
d
H
=
=: (J).
dt
J
A solution from initial values (J0 , 0 ) is J(t) = J0 and (t) = 0 + t. The
constants of motion are frequencies which can be interpreted as angular
velocities on the torus.
Definition. An orbit on the torus is quasiperiodic if there is no vector of
integers m Zn such that
m = 0.
A quasiperiodic orbit fills the surface of the torus as t . On the other
hand, if i /j Q, i, j, the orbit is periodic, and closes on itself. The set of
tori with periodic orbits is dense in the phase space, but has a zero Lebesgue
measure. Hence, orbits on a randomly picked torus are quasiperiodic with
probability of one.
11

Example (harmonic oscillator). In general, for N = 1, we have

H(p, q) =

p2
+ V (q),
2m

where p, q R, and V (q) is a potential energy of a general form [Ott, fig.

7.5]. We have
Z
1 q2
{2m[E V (q)]}1/2 dq.
J=
q1
p
For the harmonic oscillator, V (q) = mh2 q 2 /2, where h = k/m is the
angular velocity of the oscillation, and k is the spring constant. From V (q) =
E we have q2 = q1 = [2E/(mh2 )]1/2 , and integration 1 yields J = E/h .
Thus,
H(J) = h J,
and (J) = H/J = h , and is (in this rare case) independent of J. The
angle (t) = 0 + h t. In order to use the generating function S(J, q), we
write H(p, q) = H(J), and substitute p = S/q, obtaining
S
=
q

q
2m(h J mh2 q 2 /2)

Solving for , and integrating 2 , gives

Z

1/2
S
=
= mh
2m(h J mh2 q 2 /2)
dq
J
1/2
Z 
2J
2
dq
q
=
mh
 r

mh
= arcsin q
,
2J
and we have (p, q) as a function of (J, );
r
2J
q=
sin ,
mh
p
p = 2Jmh cos .
R 2
(a x2 )1/2 dx = x2 (a2 x2 )1/2 +
R
dx
2
= arcsin xa
(a2 x2 )1/2
1

a2
2

arcsin xa

12

The trajectory (p(), q()) is an ellipse [Ott, fig. 7.6]. In general, the mapping
(p, q) 7 (J, ) for N = 1 transforms a closed curve into a one-dimensional
torus (a circle).
For N > 1, a general procedure for finding the action-angle variables
(J, ) involves solving the Hamilton-Jacobi equation


S
, q = E.
H
q
If we are lucky, the equation is solvable by separation of variables, and we
have N separation constants which are also constants of motion. The motion
is typically quasiperiodic.

Perturbations and the KAM theorem

The phase space of integrable Hamiltonian systems is foliated by tori. These

appear as nested closed curves on the SOS. The existence of N integrals of
motion allows us to access the tori via action-angle variables.
We know how to solve the Hamilton-Jacobi equation (N > 1) for only a
handful of systems, e.g., harmonic oscillator, isochrone, and general Stackel
potentials. On the other hand, the logarithmic potential, for example, numerically demonstrates the existence of an additional constant of motion (besides
E) almost everywhere on the SOS [computer demo, logarithmic potential].
Motion on a torus is regular, and chaotic orbits are possible only in regions
of phase space where invariant tori do not exist.
In order to see how the integrable tori change under perturbation, we
study
H(p, q) = H0 (p, q) + H1 (p, q),
where H0 is an integrable Hamiltonian, H1 is non-integrable, and R is
small.
We know that under small perturbations some Hamiltonian systems stay
close to integrable (Solar System), but on the other hand, some are globally
chaotic (statistical mechanics).
Kolmogorov (1954), Arnold (1963), and Moser (1973) proved a result that
for sufficiently small , most of the tori survive the perturbation (the KAM
theorem).
13

The actual proof is difficult [see, e.g., Arnold]. Some considerations follow.
In action-angle variables of H0 ;
H(J, ) = H0 (J) + H1 (J, ).
If there are tori in H, we have a new set of action-angle variables (J 0 , 0 ) for
which
H(J, ) = H 0 (J 0 ),
and which are obtained by a generating function S(J 0 , ) such that
S
,

S
0 =
.
J 0
J=

The corresponding Hamilton-Jacobi equation (HJE) is



S
, = H 0 (J 0 ).
H

We look for a solution in the form of a power series:

S = S0 + S1 + 2 S2 + . . . .
We set S0 = J 0 , because = 0 then corresponds to J = J 0 , 0 = .
Substitution to HJE gives




S1
S1
2 S2
0
0
+
+ . . . + H1 J +
+ . . . , = H 0 (J 0 ).
H0 J +

By differentiating at J 0 with respect to the first order -terms we have

H0 (J 0 ) +

H0 S1

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

Next, we write -dependent terms as Fourier series;

X
H1 (J 0 , ) =
H1,m (J 0 ) exp(im ),
m

S1 (J 0 , ) =

S1,m (J 0 ) exp(im ),

14

(5)

where H1 , S1 are real-valued, and the coefficients H1,m , S1,m C, m ZN .

By substituting into (5), and requiring that H 0 is a function of J 0 only, we
have
X H1,m (J 0 )
exp(im ),
S1 = i
m 0 (J 0 )
m
where 0 (J) = H0 (J)/J are the unperturbed frequencies. Similar terms
S2 , S3 , . . . follow, if this technique is applied to higher orders of . Clearly, the
resonant tori for which m 0 = 0 play a special role (the problem of small
denominators). The KAM theorem essentially proves that the series for S
converges for very nonresonant tori with a frequency vector satisfying
|m | > K()|m|(N +1) ,

m ZN \{0},

where K() > 0, and |m| = |m1 | + |m2 | + . . . + |mN |. The Lebesgue measure
of the complement of this set, around resonant tori, goes to zero, as 0.
However, since the unperturbed resonant tori are dense in the phase space,
whenever > 0, arbitrarily near each surviving torus there is a region of
destroyed resonant tori which can host irregular motion.

The fate of resonant tori

Twist map (rn+1 , n+1 ) = M0 (rn , n ) is a model for the SOS map of an
integrable Hamiltonian H0 (J) with N = 2 [Ott, fig. 7.7];
rn+1 = rn ,
n+1 = [n + 2R(rn )]

mod 2,

where R(r) = 1 /2 is the ratio of frequencies for the torus r, and the SOS is
drawn at 2 = constant. On a resonant torus at r = r, R(
r) = j/k, j, k Z
such that k1 j2 = 0, and k applications of the map return to the original
point;
M0k (r, ) = (r, ( + 2j) mod 2) = (r, ).
Hence, each point on r = r is a fixed point of M0k . We assume that R is
smooth and increasing. In such a case there are circles r < r < r+ in such
a way such that under M0k , the points on r and r+ travel clockwise and
counterclockwise, respectively [Ott, fig. 7.8a].

15

When H0 is perturbed by the term H1 , we have a slightly changed map

M ;
rn+1 = rn + g(rn , n ),
n+1 = [n + 2R(rn ) + h(rn , n )]

mod 2.

If is small enough, we still have the (distorted) circles r and r+ where the
points travel to opposite directions under Mk . Thus, there must be a curve
r () in between where Mk moves points only in radial direction [Ott, fig.
7.8b].
When Mk is applied to the points on r (), we obtain another curve r0 ()
[Ott, fig. 7.9]. Since Mk is volume-preserving, the two curves intersect at a
finite and even number of points which are fixed points of Mk . Since we
know the directions where the points travel under Mk with respect to the
curves, we can identify half of the fixed points as elliptic and the other half
as hyperbolic. [Ott, fig. 7.10] This is the Poincare-Birkhoff theorem.
The number of elliptic (and hyperbolic) points is equal to k. An elliptic
orbit on the SOS circles around the elliptic points. [computer demo, logarithmic potential] The neighbourhood of an elliptic point can be further
modelled by another perturbed twist mapping, resulting in a new set of elliptic and hyperbolic points in a smaller scale. If this process is continued ad
infinitum, a fractal-like structure of the phase space is revealed, with subsequently smaller tori winding on bigger ones [Lichtenberg & Lieberman, fig.
3.5].
A hyperbolic point lies at an intersection of an unstable and stable manifold, Wu and Ws , respectively. Consider a repeated SOS mapping (Mk )n . A
point on Ws approaches the hyperbolic point, as n . The same happens
to a point on Wu as n .
The two manifolds Wu and Ws connect to another hyperbolic point in a
way that resembles the phase portrait of a simple pendulum. [Lichtenberg
& Lieberman, fig. 3.3a] For the pendulum, Wu and Ws between the points
coincide, forming a smooth separatrix. However, in the perturbed system,
the Wu and Ws are generally different, and intersect at a homoclinic point.
This point, mapped by (Mk )n must also lie on both Wu and Ws . Hence,
we conclude that there must be an infinite number of homoclinic points as
n .
Since Mk is area-preserving, Wu (Ws ) must oscillate wildly as n
(n ). Near a hyperbolic point, these oscillations overlap, and form a
16

region where points from Wu and Ws are mixed (homoclinic tangle). [Lichtenberg & Lieberman, fig. 3.4b] Orbits in this region show extreme sensitivity
to initial conditions, and hence, are chaotic [computer demo, logarithmic
potential].

Transition to global chaos

In two degrees of freedom, the chain of resonant islands and chaotic regions
surrounding them are confined between surviving KAM-tori. As the perturbation strength increases, chaotic regions grow, and more KAM-tori are
destroyed. When all KAM-tori are destroyed, an can wander anywhere in the
energy surface, and the motion becomes ergodic [computer demo, standard
map].
For the standard map, the rotation number of the last surviving KAMtorus is the golden ratio which is the farthest away from any rational number
in the sense that its representation as a continued fraction is

1
1+ 5
=
.
2
1
1+
1
1+
1 +
For N = 2 the 2-tori can enclose volumes in the three dimensional energy
surface H(p, q) = E. For N > 2 the situation is qualitatively different;
e.g., for N = 3 the 3-tori cannot enclose 5-dimensional volumes much like
lines cannot enclose 3-dimensional volumes. For any > 0, stochastic regions
form a web where orbits can, in principle, reach any part of the energy surface
through Arnold diffusion. However, Nekhoroshev theorem tells us that the
Arnold diffusion happens very slowly; at time scales exp(1/) or even slower.

17