Classical Mechanics and Symplectic

Integration
1 0.5 0 0.5 1 1.5
1.5
1
0.5
0
0.5
1
1.5
y
p
y
Lecture Notes
Poul G. Hjorth Nikolaj Nordkvist
Contents
1 Preliminaries 1
2 Lagrangian Mechanics 5
2.1 Calculus of Variations . . . . . . . . . . . . . . . . . . . . . . 5
2.2 Lagranges Equations . . . . . . . . . . . . . . . . . . . . . . . 10
2.3 Symmetries and Conservation Laws . . . . . . . . . . . . . . . 13
3 Hamiltonian Mechanics 15
3.1 Hamiltons Equations . . . . . . . . . . . . . . . . . . . . . . . 15
3.2 Canonical Transformations . . . . . . . . . . . . . . . . . . . . 19
3.3 Integrable Systems . . . . . . . . . . . . . . . . . . . . . . . . 26
3.4 Nearly Integrable Systems . . . . . . . . . . . . . . . . . . . . 29
3.4.1 The KAM Theorem . . . . . . . . . . . . . . . . . . . . 33
3.4.2 Nekhoroshevs Theorem . . . . . . . . . . . . . . . . . 34
3.5 Symplectic Integrators . . . . . . . . . . . . . . . . . . . . . . 36
Chapter 1
Preliminaries
We start by recaling some usefull theory, that is some results from multivari-
able calculus and the concept of a Poincar map.
Elements of Multivariable Calculus
In the theory of analytical mechanics two main theorems are frequently used,
namely the chain rule and the inverse function theorem. Therefore we start
be recalling these basic mathematical facts.
Throughout this text we will denote a vector x R
n
by boldface and its
elements by x
i
, i.e. x = (x
1
, . . . , x
n
). For a smooth function f : R
n
R
m
,
f = f(x), that is f C

(R
n
, R
m
), the Jacobian matrix
f
x
is the m n
matrix given by
1
f
x
=

f
1
x
1

f
1
x
n
.
.
.
.
.
.
.
.
.
f
m
x
1
. . .
f
m
x
n

With this the chain rule can be given as

Theorem 1 (The chain rule). For the smooth functions f : R
n
R
m
,
f = f(x), and g : R
m
R
k
, g = g(y), the Jacobian of g f : R
n
R
k
is
given by
g f
x
(x) =
g
y
(f(x))
f
x
(x)
Example 1. Let f : R R
n
, f = f(t), and g : R
n
R, g = g(x) then we
1
Often, when there is no risk of confusion, we will for functions f : R
n
R also use
the notation
f
x
for its transpose, i.e. the gradient of f.
2 Preliminaries
have by the chain rule that
d
dt
g f(t) =
g
x
(f(t))
df
dt
(t)
=
n

i=1
g
x
i
(f(t))

f
i
(t)
= g(f(t))

f(t)
where the dot is the usual Euclidian scalar product.
If a function f : R
n
R
n
is bijective (one-to-one and onto) with inverse
f
1
, and if both f and f
1
are smooth, then f is said to be a dieomorphism.
A dieomorphism is also sometimes refered to as a coordinate change, since
it can be used to dene new coordinates in which every function will be as
many times dierentiable as in the old coordinates. We will often specify a
coordinate transformation : R
n
R
n
by x y to state explicitly that
it takes the x-coordinates and transforms to the y-coordinates, and
1
vice
versa. The inverse function gives sucient conditions on a map f : R
n
R
n
to be a dieomorphism.
Theorem 2 (Inverse function theorem). Let U R
n
be open, and let f :
U R
n
, f = f(x), be smooth. If
f
x
is non-singular at x
0
U, i.e.
det

f
x
(x
0
)

= 0
then there exists an open set V with x
0
V U such that f[
V
is a dieo-
morphism.
Example 2. If g : R
2n
R, g = g(x, y), x, y R
n
, is smooth and if
det

2
g
yy

= 0
where

2
g
yy
is the Jacobian matrix of
g
y
, then we know by the inverse
function theorem, that it is possible to dene a coordinate transformation
: R
2n
R
2n
, (x, y) ( x, y), as
( x, y) =

x,
g
y
(x, y)

This transformation is important in the theory of analytical mechanics and

is called the Legendre transform.
A more general result is the following
3
Theorem 3 (Implicit function theorem). Let U R
nm
R
m
be open, and
let f : U R
m
be smooth. Denote the Cartesian coordinate system on
R
nm
R
m
by (x, y) = (x
1
, . . . , x
nm
, y
1
, . . . , y
m
). Suppose that at the point
(x
0
, y
0
) U
f(x
0
, y
0
) = 0
and the matrix
f
y
is nonsingular, i.e.
det

f
y
(x
0
, y
0
)

= 0
Then there exists an open neighborhood V of x
0
in R
nm
and an open neigh-
borhood W of y
0
in R
m
such that V W U, and there exists a smooth
map g : V W such that for each (x, y) V W
f(x, y) = 0 y = g(x)
The inverse function theorem can be deduced from the implicit function
theorem as a corrolary.
Poincar Maps
For a general autonomous dierential equation x = f(x), x R
n
, a surface
of section S is the image of a map g : U R
n
, U R
n1
, e.i. S = g(U),
2
such that f is transversal to it, that is the ow f(x) intersects S only in
points, not lines. Let
t
be the ow of f, =
[0,)
(p) the trajectory from
p R
n
assuming that
t
(p) is dened for all t [0, ] and x
k
the
k-th intersection, in a particular sence (e.g. from right to left) of and S,
x
k+1
=

(x
k
) for some (0, ), then the Poincar map : S S is
dened as the mapping giving
(x
k
) = x
k+1
Studying the Poincar map of a dierential equation can simplify some
of the qualitative analysis. Consider a dierential equation in R
3
and the
Poincar map given by considering succesive intersections with a plane,
from one specic side to the other. If has a xed point we know that the
dierential equation has a closed orbit, and if x
1
, x
2
= (x
1
), x
3
=
2
(x
1
), . . .
lies on a closed curve we know that near the plane of section the motion takes
place on something similar to a cylinder.
2
This dinition slightly diers from the standerd one, where S is a n 1 dimensional
submanifold of R
n
. So we see that our denition is contained in the correct one.
4 Preliminaries
Chapter 2
Lagrangian Mechanics
2.1 Calculus of Variations
A functional is a mapping from a vector space to the real numbers. Let
L : R
n+1
R be a smooth function, called the Lagrangian, and consider a
smooth curve : [t
0
, t
1
] R
n
, then we dene the functional I as
I() =

t
1
t
0
L((t), (t), t)dt
The number n is refered to as the number of degrees of freedom, for reasons
which will become clear later. The variation I of I is for a smooth curve
: [t
0
, t
1
] R
n
dened as
I(, ) =
d
ds

s=0
I( + s) (2.1)
We see that a necessary condition for to be an ekstremum of I, i.e. I()
lim
s0
I(+s) or I() lim
s0
I(+s) for all curves , is that I(, ) = 0.
If we wish to nd the curves starting at q
0
and ending at q
1
that are extremals
of I we therefore calculate 2.1 with the condition on that (t
0
) = (t
1
) = 0,
which gives
I(, ) =

t
1
t
0

L
q
((t), (t), t) (t) +
L
q
((t), (t), t) (t)

dt
=

t
1
t
0

L
q
(, (t), t)
d
dt
L
q
((t), (t), t)

(t)dt +

L
q

t
1
t
0
=

t
1
t
0

L
q
((t), (t), t)
d
dt
L
q
((t), (t), t)

(t)dt (2.2)
where q = (q
1
, . . . , q
n
), q = ( q
1
, . . . , q
n
) are Euclidian coordinates and L =
L(q, q, t). In order to proceed we need the so called fundamental lemma of
the calculus of variations
6 Lagrangian Mechanics
Lemma 1. Let g C
0
([t
0
, t
1
], R
n
), then

t
1
t
0
g(t) h(t)dt = 0 h C

([t
0
, t
1
], R
n
)
if and only if
g = 0
Proof. The if part is obvious. The only if part we show by contradiction, so
assume that

t
1
t
0
g(t) h(t)dt =

t
1
t
0

n
i=1
g
i
(t)h
i
(t)dt = 0 for some g
i
0
= 0.
Then there exists (t
0
, t
1
) such that g
i
0
() = 0. Since g
i
0
is continous
there exists > 0 such that g
i
0
(t) >
1
2
g
i
0
() for t ( , + ). Then
choose h smooth such that
h
i
0
(t) = 0 for t [t
0
, ] [ + , t
1
]
h
i
0
(t) > 0 for t ( , + )
h
i
(t) = 0 for i = i
0
, t [t
0
, t
1
]
But then we get

t
1
t
0
n

i=1
g
i
(t)h
i
(t)dt =

g
i
0
(t)h
i
0
(t)dt > g
i
0
()
which is a contradiction.
Using this lemma we thus get the following
Theorem 4. If a curve C

([t
0
, t
1
], R
n
) with (t
0
) = q
0
and (t
1
) = q
1
is an extremum of I among the curves satisfying these boundary conditions
then it satises the equations
d
dt
L
q
(q, q, t)
L
q
(q, q, t) = 0
These equations are called the Euler-Lagrange equations. If we use
the chain rule the Euler-Lagrange equations become
n

j=1

2
L
q
i
q
j
q
j
+
n

j=1

2
L
q
i
q
j
q
j
+

2
L
q
i
t

L
q
i
= 0 i = 1, . . . , n
So we see that the Euler-Lagrange equations are a set of coupled second
order dierential equations. From the classical theorem on the existence
and uniqueness of solutions to a dierential equation we see that a sucient
condition to ensure the existence and uniqueness of a solution to the Euler-
Lagrange equations is
det

2
L
q q

= 0
2.1 Calculus of Variations 7
Example 3. In 1696 John Bernoulli posed the following problem, which was
later solved by John Bernoulli, James Bernoulli, Newton, and LHospital, and
played an important part in the development of the theory of the calculus of
variations.
Let A and B be xed points in a vertical plane, and assume that B is
lower than A. Let a particle slide without friction along a curve joining A
and B, then the time it takes to reach B from A is a functional of the curve,
and the curve which takes the least time is called the brachistochrone, and
can be calculated using the Euler-Lagrange equations as follows:
We choose normal Euclidian x-y coordinates in the plane, where gravity
acts in the positive direction of the y-axis, and assume for simplicity that
the particle starts at rest and that A is the origin of coordinates. Letting the
curve be a function of x we get
v =

1 + (y

)
2
dx
dt
and due to conservation of energy we have that
v =

2gy
where g is the gravitational accelleration. The transit time T of a curve y is
therefore given by
T(y) =

t
end
t
start
dt =

x
1
0

1 + (y

)
2

2gy
dx
So the Lagrangian is given by
L(y, y

) =

1 + (y

)
2

2gy
and the Euler-Lagrange equation for this problem becomes
2y

(x)y(x) + (y

(x))
2
+ 1 = 0
We notice that this equation is independent of the values of g and the mass
of the particle.
It turns out that the solution to the Euler-Lagrange equation is a family
of cycloids
x = r( sin ) y = r(1 cos )
parametrized by .
8 Lagrangian Mechanics
Geodesic Curves
Consider a curve in U R
n
. The length L of a curve in U is the functional
given by
L() =

t
1
t
0
| (t)|dt
The curves which minimize length are called geodesic curves.
If we dene the functional c as
c() =
1
2

t
1
t
0
| (t)|
2
dt
we get using Schwartz inequality
L() =

t
1
t
0
1 | |dt

t
1
t
0
dt

1/2

t
1
t
0
| (t)|
2
dt

1/2

2(t
1
t
0
)

c()
where we have equality if and only if | (t)| = constant. So when minimizing
c we nd a geodesic curve with constant speed.
Example 4. Consider U = R
n
with coordinates x = (x
1
, . . . , x
n
). We have
that the Lagrangian for c is
L(x, x) =
1
2
| x|
2
=
1
2
( x
2
1
+ . . . + x
2
n
)
So according to the Euler-Lagrange equations we get that the geodesic curves
satises
x = 0
which means that the curves in R
n
with minimal length are
(t) = at +b
i.e. straight lines, a well known fact.
Another instructive example is the following
Example 5. Consider the sphere S
2
R
3
. In R
3
we can choose spherical
coordinates (, , r) and in these coordinates S
2
will be given by
x = sin cos
y = sin sin
z = cos
2.1 Calculus of Variations 9
Therefore on S
2
we have that the Lagrangian for the energy is
L(, ,

,

) =
1
2
| x[
S
2|
2
=
1
2

x[
2
S
2 + y[
2
S
2 + z[
2
S
2

=
1
2

cos cos

sin sin )
2
+ (

cos sin +

sin cos )
2
+ (

sin )
2

=
1
2
(

2
+

2
sin
2
)
and this is thus the Lagrangian determining the geodesic curves on S
2
via
the Euler-Lagrange equations.
If we have a smooth coordinate transformation : R
n
R
n
, q q(q),
then a smooth curve in the q-coordinates must be a smooth curve in the q-
coordinates and vice versa. Thus it seems that if a curve in the q-coordinates
is an ekstremum for I then it must also be so in the q-coordinates. The
following result would be a direct consequence of such a result.
Proposition 1. The Euler-Lagrange equations are invariant under a smooth
coordinate transformation : R
n
R
n
, q q(q). That is if : [t
0
, t
1
]
R
n
, (t) = ((t)), and L =

L then we have
d
dt
L
q
((t), (t), t)
L
q
((t), (t), t) = 0
if and only if
d
dt

q
( (t),

(t), t)

L
q
( (t),

(t), t) = 0
Proof. From the chain rule we have

q
j
=
n

k=1
q
j
q
k
q
k



q
j
q
i
=
q
j
q
i
Thus we get using the chain rule on L(q, q, t) =

L( q(q),

q(q, q), t)
L
q
i
=
n

j=1

q
j

q
j
q
i
=
n

j=1

q
j
q
j
q
i
giving
d
dt

L
q
i

=
n

j=1

d
dt

q
j

q
j
q
i
+
n

j=1

q
j
n

k=1

2
q
j
q
i
q
k
q
k
10 Lagrangian Mechanics
The second part of the Euler-Lagrange equations is similarly calculated using
the chain rule
L
q
i
=
n

j=1

L
q
j
q
j
q
i
+
n

j=1

q
j

q
j
q
i
=
n

j=1

L
q
j
q
j
q
i
+
n

j=1

q
j
n

k=1

2
q
j
q
i
q
k
q
k
Combining these we get
d
dt
L
q
i

L
q
i
=
q
j
q
i

d
dt

q
j

L
q
j

We notice that
q
j
q
i
are the elements of the Jacobian matrix
q
q
for and we
get
d
dt
L
q

L
q
=
q
q
T

d
dt

L
q

(2.3)
Since is a coordinate transformation
q
q
is nonsingular and therefore we
get the desired result.
Problems that are invariant under smooth coordinate transformations can
typically be stated in a dierential geometric setting which makes it possible
to work with the problem in a coordinate free way giving many powerfull
tools for analysis.
2.2 Lagranges Equations
Consider a mechanical system of k point masses m
i
. If we by x
j
= (x
j
, y
j
, z
j
)
denote the Cartesian coordinates of the i-th point mass, the entire system can
be described by (x
1
, . . . , x
k
) R
3k
, i.e. by 3k coordinates. Throughout this
monograph we will only consider conservative systems which are mechanical
systems for which the total force on the elements in the system can be derived
from a potential V : R
3k
R according to
F
i
=
V
x
i
so newtons second law gives
m
i
x
i
=
V
x
i
2.2 Lagranges Equations 11
If we for a mechanical system of point masses dene the Lagrangian as
L =
k

i=1
1
2
m
i
| x
i
|
2
V (x
1
, . . . , x
k
)
that is the kinetic energy K minus the potential energy V , the Euler-Lagrange
equations for this system becomes
m
i
x
i
=
V
x
i
so we see that Newtons laws and Lagranges equations with L = K V for
k point masses are equivalent.
Due to proposition 1 we know that we can choose arbitrary coordinates
for the analysis of the system. Consider a mechanical system of point masses
described locally in V R
d
by the coordinates z = (z
1
, . . . , z
d
), not neces-
sarily Cartesian coordinates, and assume that the potential is given by
V = V
1
(z
1
, . . . , z
d
) +
1
2
1

(z
1
z
d
)
2
for some > 0. Since the solution to a dierential equation depending
continously on a parameter will depend continously on that parameter the
solution of Euler-Lagranges equation for this potential will depend contin-
uosly on . For very small we will have z
d
z
1
, when z
d
(0) = z
1
(0), so
eectively in the equations the z
d
coordinate can be considered a constant
and the Euler-Lagrange equations will still describe the motion. So if we
examine a mechanical system of point masses with a potential which to a
good approximation gives the constraints
z
dc+1
= 0 . . . z
d
= 0
then we expect the motion of the system still to be expressed by the Euler-
Lagrange equations, where z
dc+1
, . . . , z
d
are just constants. The number
n = c d is called the degrees of freedom of the system since this is the least
number of coordinates which is needed to specify the conguration of the
system. The coordinates q = (q
1
, . . . , q
n
) = (z
1
, . . . , z
n
) U R
n
are called
generalised coordinates, and with
L(q, q, t) = L(q
1
, . . . , q
n
, z
dc+1
, . . . , z
d
, q
1
, . . . , q
n
, 0, . . . , 0)
the motion of the system is thus determined by
d
dt
L
q
(q, q, t)
L
q
(q, q, t) = 0
These equations are called Lagranges equations so basically Lagranges
equations are the Euler-Lagrange equations for a mechanical system with
Lagrangian equal to the kinetic- minus the potential energy.
12 Lagrangian Mechanics
m
g
x
y

r
Figure 2.1: The pendulum.
Example 6. For a pendulum, see gure 2.1, we have, due to the fact that
the length of the pendulum approximative is constant, that in cartesian
coordinates (x, y)
x
2
+ y
2
=
2
If we instead choose polar coordinates (r, ), where = 0 in the vertical
downward position, we have r = 0 and the kinetic energy K and the potential
energy V is given by
K =
1
2
m
2

2
V = mg cos()
where m is the mass and g is the gravitational accelleration. From this the
motion is determined by Lagranges equations
d
dt
L

= 0
giving

= g
1
sin()
This equation would in the Newtonian framework have been deduced from
angular momentum considerations and not Newtons second law directly,
but in the Lagrangian framework everything is deduced from Lagranges
equations alone.
2.3 Symmetries and Conservation Laws 13
2.3 Symmetries and Conservation Laws
If the Lagrangian L is independent of q
i
then q
i
is called a cyclic coordinate,
and from the Euler-Lagrange equations we get
d
dt
L
q
i
=
L
q
i
so if q
i
is a cyclic coordinate, then
L
q
i
is a constant of the motion.
The energy E : R
2n+1
R is dened, using the Lagrangian, as
E(q, q, t) :=
L
q
(q, q, t) q L(q, q, t)
Letting (t) be a solution to the Euler-Lagrange equations we get using the
chain rule and the Euler-Lagrange equations
d
dt
E((t), (t), t) =
d
dt
L
q
((t), (t), t) (t) +
L
q
((t), (t), t) (t)

d
dt
L((t), (t), t)
=
L
q
((t), (t), t) (t) +
L
q
((t), (t), t) (t)

L
q
((t), (t), t) (t) +
L
q
((t), (t), t) (t)
+
L
t
((t), (t), t)

=
L
t
((t), (t), t)
So if the Lagrangian is independent of t the energy is a constant of the
motion.
The fact that conservation of momentum or angular momentum of a
mechanical system follows from invariance of the problem with respect to
translation or rotation respectively is a consequence of a more general result
originally showed by E. Noether.
Theorem 5 (Noethers theorem). Let L : R
2n+1
R be smooth and assume
there exists a one-parameter family of smooth maps h
s
: R
n
R
n
smooth in
s (, ), > 0, and with h
0
= id. If
L

h
s
((t)),
d
dt
h
s
((t)), t)

= L((t), (t), t)
for all s (, ) and all smooth curves : [t
0
, t
1
] R
n
. Then for any
solution of the Euler-Lagrange equations the function
F(q, q, t) =
L
q
(q, q, t)
dh
s
ds

s=0
(q)
is a constant of the motion.
14 Lagrangian Mechanics
Proof. By assumption we have that
0 =
d
ds

s=0
L

h
s
((t)),
d
dt
h
s
((t)), t)

=
L
q
((t), (t), t)
dh
s
ds

s=0
((t)) +
L
q
((t), (t), t)
d
dt
dh
s
ds

s=0
((t))
Let (t) be a solution to the Euler-Lagrange equations, i.e. we have
d
dt
L
q
((t), (t), t)
L
q
((t), (t), t) = 0
Inserting this we get
0 =
d
dt
L
q
((t), (t), t)
dh
s
ds

s=0
((t)) +
L
q
((t), (t), t)
d
dt
dh
s
ds

s=0
((t))
=
d
dt

L
q
((t), (t), t)
dh
s
ds

s=0
((t))

which shows that F is a constant of the motion.

Example 7. Consider a mechanical system of n point masses m
j
and coor-
dinates q
j
= (x
j
, y
j
, z
j
)
L =
n

j=1
1
2
m
j
| q
j
|
2
V (q
1
, . . . , q
n
)
Assume that the system is invariant under a translation along the x-axis, i.e.
h
s
(q
1
, . . . , q
n
) = (q
1
+ se
x
, . . . , q
n
+ se
x
)
where e
x
is a unit vector in the x-direction. Then we get that
n

j=1
m
j
x
j
which is the total momentum along the x-axis, is conserved by the ow. From
proposition 1 we thus know that a mechanical system of point masses which
is invariant under translation in some direction has total momentum along
this axis as a constant of the motion.
A system consisting of two point masses with a potential given by
V (q
1
, q
2
) = V (|q
1
q
2
|)
is invariant under a translation along any axis and therefore the total mo-
mentum in any direction is conserved. Examples of such systems include two
masses interacting gravitationally, e.g. the sun and a planet, and two charges
interacting due to Coulombs law.
Chapter 3
Hamiltonian Mechanics
3.1 Hamiltons Equations
Given a smooth Lagrangian L : R
2n+1
R the smooth map (q, q) R
n
given by
p =
L
q
gives the Legendre transformation (q, q) (q, p) if
det

2
L
q q

= 0
which insures according to the inverse function theorem that (q, q) (q, p)
indeed locally denes a smooth coordinate transformation.
1
Remark 1. Often for mechanical systems the potential energy is a function
of the conguration only and the kinetic energy K is given by a quadratic
form, i.e.
K =
1
2
q
T
G(q) q =
1
2
n

i,j=1
g
ij
(q) q
i
q
j
where G is a positive denite and symmetric nn matrix, possibly depending
on q, with elements g
ij
. Since G is positive denite its determinant is strictly
positive and we get
det

2
L
q q

= det

2
K
q q

= det(G) = 0
So for these systems the Legendre transformation does exist.
1
If L depends on t we know from the implicit function theorem that the Legendre
transformation is smooth in t.
16 Hamiltonian Mechanics
The Legendre transformation makes it possible to dene the Hamilto-
nian which is the smooth function H : R
2n+1
R given by
H(p, q, t) =
n

j=1
p
j
q
j
L(q, q(q, p, t), t)
Remark 2. Consider a mechanical system with potential energy a function
of the conguration only and the kinetic energy K given by a quadratic form
G
K =
1
2
q
T
G(q) q =
1
2
n

i,j=1
g
ij
(q) q
i
q
j
The Legendre transformation gives
p = G(q) q q = G
1
(q)p
p
i
=
n

j=1
g
ij
(q) q
j
q
i
=
n

j=1
g
ij
(q)p
j
where g
ij
are the elements of the matrix G
1
, which is also positive denite
and symmetric. Then,
K =
1
2
p
T
G
T
(q)G(q)G
1
(q)p
=
1
2
p
T
G
1
(q)p
and
H = p
T
G
1
(q)p
1
2
p
T
G
1
(q)p + V (q)
=
1
2
p
T
G
1
(q)p + V (q)
= K(p, q) + V (q)
Thus we see that for such a system the Hamiltonian is equal to the total
energy of the system.
How the dynamics of the Euler-Lagrange equations is expressed in the
coordianates q and p is given by the following
Theorem 6. The Euler-Lagrange equations are equivalent to the following
equations
p =
H
q
q =
H
p
These equations are called Hamiltons equations.
3.1 Hamiltons Equations 17
Proof. The dierential of H is given by
dH =
n

j=1

H
p
j
dp
j
+
H
q
j
dq
j

+
H
t
dt (3.1)
If instead H is considered as a function in q and q we get the following
dierential
dH =
n

j=1

p
j
d q
j
+ q
j
dp
j

L
q
i
dq
i

L
q
j
d q
j

+
H
t
dt
=
n

j=1

q
j
dp
j

L
q
i
dq
i

+
H
t
dt (3.2)
The Euler-Lagrange equations is in the coordinates q and p given by
d
dt
p =
L
q
Inserting this in 3.2 gives
dH =
n

j=1
( q
j
dp
j
p
j
dq
j
) +
H
t
dt (3.3)
Comparing equation 3.3 and 3.1 gives Hamiltons equations.
Example 8. Consider the pendulum. We have that the kinetic energy K
and the potential energy V are given by
K =
1
2
m
2

2
V = mg cos()
According to remark 2 we thus have
p = m
2

and
H =
1
2
p
2
m
2
mg cos()
Therefore Hamiltons equations for the pendulum are
p = mg sin()

=
p
m
2
18 Hamiltonian Mechanics
We dene the 2n 2n matrix J as
J =

0 I
I 0

J
1
= J
T
= J
where I is the n n identity matrix. This Matrix is called the symplectic
identity. Hamiltons equations can then be written as

p
q

0 I
I 0

H
p
H
q

We dene y = (p, q) which gives the equations in the compact form

y = J
1
H(y, t)
We will denote by
H
t
the ow of Hamiltons equations y = J
1
H(y, t).
Proposition 2. We have
d
dt
H(
H
t
(y
0
, t
0
), t) =
H
t
(
H
t
(y
0
, t
0
), t)
So if H is independent of t, H is a constant of the motion for Hamiltons
equations y = J
1
H(y, t).
Proof. This follows from the direct calculation using the chain rule and
Hamiltons equations. Take (p(t), q(t)) =
H
t
(y
0
, t
0
)
d
dt
H(p(t), q(t), t) =
H
p
(p(t), q(t), t) p(t) +
H
q
(p(t), q(t), t) q(t)
+
H
t
(p(t), q(t), t)
=
H
p
(p(t), q(t), t)

H
q
(p(t), q(t), t)

+
H
q
(p(t), q(t), t)
H
p
(p(t), q(t), t) +
H
t
(p(t), q(t), t)
=
H
t
(p(t), q(t), t)
If a system does indeed depend explicitly on t, that is H = H(p, q, t),
then we can dene a new Hamiltonian H : R
2n+2
R as
H(p, p
0
, q, q
0
) = p
0
+ H(p, q, q
0
)
Hamiltons equation for this system then are
p =
H
q
p
0
=
H
q
0
q =
H
p
q
0
= 1
Thus every time dependent system with n degrees of freedom can be written
as a system independent of time with n + 1 degrees of freedom. For this
reason from now on only time independent systems will be considered.
3.2 Canonical Transformations 19
Poincar Maps for Hamiltonian Systems with 2 Degrees of Freedom
When you have a time independent Hamiltonian system with 2 degrees of
freedom with Hamiltonian H, we know from proposition 2 that the Hamilto-
nian will be a constant of the motion. Thus we know that a trajectory will be
conned to move on a set W
E
= (p
1
, p
2
, q
1
, q
2
) R
4
[ H(p
1
, p
2
, q
1
, q
2
) = E.
On W
E
we can locally write p
2
= p
2
(p
1
, q
1
, q
2
) due to the implicit function
theorem, so we can regard W
E
as a subset of R
3
. Therefore instead of in-
vestigating a Poincar map for a specic surface of section in R
4
, that is
a volume, we can simplify matters and only consider the trajectory in the
p
1
, q
1
, and q
2
coordinates, i.e. in W
E
R
3
, and in these coordinates examine
the Poincar map for a surface of section, now indeed a surface, in R
3
, e.g.
the plane given by q
2
= 0.
3.2 Canonical Transformations
The concept of a canonical coordinate transformation is very important since
this is the class of coordinate transformations which leave invariant the form
of Hamiltons equations.
q
p
q
p

H
t

H
t

Figure 3.1: Coordinate transformation. It is canonical if H =

H .
Denition 1. The smooth coordinate transformation : R
2n
R
2n
, ( p, q)
(p, q) is said to be canonical if for any Hamiltonian H Hamiltonss equations
are equivalent to

p =

H
q

q =


H
p
where

H = H .
We then have that is a canonical transformation if and only if

H
t
=
e
H
t
20 Hamiltonian Mechanics
i.e. solutions are mapped into solutions. Denoting by
y
y
the Jacobian of
we have the following result giving an easy characterisation of canonical
transformations.
Proposition 3. The transformation : R
2n
R
2n
, y y, is canonical if
and only if its Jacobian
y
y
satises the relation
y
y
T
J
y
y
= J
Proof. Since with y = y( y) we have
y =
y
y

y
we get that two general dierential equations
y = Y (y)

y =

Y ( y)
are equal if and only if
Y (y( y)) =
y
y

Y ( y)
using this we get that the transformation is canonical if and only if
J
1
H(y( y)) =
y
y
J
1

H( y)

=
y
y
J
1
(H(y( y))
=
y
y
J
1
y
y
T
H(y( y))
giving the relation
y
y
J
1
y
y
T
= J
1
Transposing this expression and using J
1
= J
T
= J gives the desired
relation.
This immediately gives that a composition of canonical transformations
is a canonical transformation.
A 2n 2n matrix A is said to be symplectic if it satises the relation
A
T
JA = J
Therefore a smooth map, which is a canonical transformation, is also said to
be symplectic.
Proposition 3 can be used to show the following theorem, which is inter-
esting in its own right, but which also encourage us to examine symplectic
integrators, i.e. integrators that are symplectic mappings.
3.2 Canonical Transformations 21
Proposition 4. For xed t R the ow
H
t
: R
2n
R
2n
of Hamiltons
equations y = J
1
H(y) is a canonical transformation.
Proof. Since H is assumed smooth the ow
H
t
(y) is a smooth function of t
and y and since

H
t

1
=
H
t
it denes a smooth coordinate transformation.
Denoting by
2
H the Hessian matrix of H we get from

d
dt

H
t
(y)

J
1
H(
H
t
(y))

that
d
dt

t
y
(y) = J
1

2
H(
H
t
(y))

t
y
(y)
We use this, the fact that
2
H is symmetric, and the relations JJ = I,
J
1
= J, and J
T
= J, to calculate the following
d
dt

H
t
y
T
J

H
t
y

d
dt

H
t
y

T
J

H
t
y
+

H
t
y
T
J

d
dt

H
t
y

=

H
t
y
T

2
H(
H
t
(y))

T
J
T
J

H
t
y
+

H
t
y
T
JJ
1

2
H(
H
t
(y))

H
t
y
=

H
t
y

2
H(
H
t
(y))

H
t
y
+

H
t
y
T

2
H(
H
t
(y))

H
t
y
= 0
Since
0
= id we have

H
0
y
T
J

H
0
y
= J
so
H
0
must be a canonical transformation and because of the above
H
t
must
be a canonical transformation for all t.
This proposition can also be used to search for a Hamiltonian such that
the canonical transformation

t
, t xed, gives the Hamiltonian system under
consideration a Hamiltonian which is simpler in some sense.
From the fact that the ow
H
t
of Hamiltons equations is a canonical
transformation we get for xed t
det

H
t
y
T
J

H
t
y

det

H
t
y

2
det(J)
= det(J)
22 Hamiltonian Mechanics
3 2 1 0 1 2 3
3
2
1
0
1
2
3

p
Figure 3.2: The pendulum. The vector eld and some solution curves. Two
disks of initial conditions and their appearence after time t = 1.5 and t = 3.
The volume of the disks are unchanged.
Since det(J) = 0,
H
t
is a smooth function of y, and
H
0
= id we thus get
det

H
t
y

= 1
The volume of a subset U of R
2n
, Vol(U) =

U
dy
1
. . . dy
2n
, when mapped by
the ow
H
t
for xed t, can be calculated using the above and the change of
variables formula for integrals as
Vol(
H
t
(U)) =

H
t
(U)
dy
1
. . . dy
2n
=

det

H
t
y

dy
1
. . . dy
2n
=

U
dy
1
. . . dy
2n
= Vol(U)
Thus we see that the ow of Hamiltons equations preserves the volume.
This leads to several important facts about the ow of Hamiltons equations,
e.g. it is not possible for the ow to have asymptotically stable equilibrium
positions or asymptotically stable limit cycles.
A canonical transformation : R
2n
R
2n
, y y, can very well mix
the p and q coordinates. This is very dierent from the Lagrangian setting
where only point transformations q q are considered.
The following theorem gives a very usefull way of constructing canonical
transformations.
3.2 Canonical Transformations 23
Proposition 5. Given a smooth function S
1
: R
2n
R, S
1
= S
1
( q, q),
q, q R
n
, such that
det

2
S
1
qq

= 0
the equations
p =
S
1
q
p =
S
1
q
implicitly dene a local canonical coordinate transformation (p, q) = ( p, q).
Proof. If we dene p = p( q, q) and p = p( q, q) by
p =
S
1
q
p =
S
1
q
(3.4)
then the condition
det

2
S
1
qq

= 0
gives according to the implicit function theorem that q = q( p, q) and thus
p = p( p, q). We can therefore calculate
(p, q)
( p, q)
=

p
p
p
q
q
p
q
q

p
p

2
S
1
qq

2
S
1
q q

1
q
q

which is nonsingular so the inverse function theorem insures that 3.4 implic-
itly denes a coordinate transformation (p, q) = ( p, q).
Hamiltons equations with Hamiltonian H : R
2n
R, H = H(p, q), are
the Euler-Lagrange equations with Lagrangian L : R
4n
R given by
L(q, p, q, p) =
n

i=1
p
i
q
i
H(p, q)
For L : R
2m
R and S : R
m
R smooth, the functionals
I
1
() =

t
1
t
0
L((t), (t))dt
I
2
() =

t
1
t
0

L((t), (t)) +
d
dt
S((t))

dt
24 Hamiltonian Mechanics
where is a smooth curve : [t
0
, t
1
] R
m
, leads to the same Euler-Lagrange
equations. Combining these two facts and proposition 1 we see that if the
canonical coordinates (p, q) and the general coordinates ( p, q) with (p, q) =
( p, q) satises
n

i=1
p
i
q
i
H(p, q) =
n

i=1
p
i

q
i

H( p, q) +
d
dt
S( p, q)
where

H = H, then must be a canonical transformation. This sucient
condition is equivalent to
dS =
n

i=1
p
i
d q
i
+
n

i=1
p
i
dq
i
+ (

H H)dt
=
n

i=1
p
i
d q
i
+
n

i=1
p
i
dq
i
(3.5)
Thus when comparing
dS
1
=
n

i=1
S
1
q
i
d q
i
+
n

i=1
S
1
q
i
dq
i
with the sucient condition on canonicity, that is equation 3.5, we see that
the coordinate transformation (p, q) = ( p, q) given by 3.4 is canonical.
A similar constructive proposition for nding canonical transformations
is the following regarding a function of p and q
Proposition 6. Given a smooth function S
2
: R
2n
R, S
2
= S
2
( p, q),
p, q R
n
, such that
det

2
S
2
pq

= 0
the equations
p =
S
2
q
q =
S
2
p
implicitly dene a local canonical coordinate transformation (p, q) = ( p, q).
Proof. The argumentation that this indeed does implicitly dene a coordinate
transformation is completely similar to that in proposition 5.
If we in proposition 5 use
S( p, q) =

S( p, q)
n

i=1
p
i
q
i
3.2 Canonical Transformations 25
The sucient condition on canonicity, that is equation 3.5, is seen to be
equivalent to
d

S =
n

i=1
p
i
dq
i
+
n

i=1
q
i
d p
i
Since
dS
2
=
n

i=1
S
2
p
i
d p
i
+
n

i=1
S
2
q
i
dq
i
=
n

i=1
q
i
d p
i
+
n

i=1
p
i
dq
i
we therefore see that this transformation is canonical.
Similar theorems can be made for functions S
3
(p, q) and S
4
(p, p). The
functions S
1
, S
2
, S
3
, and S
4
are called generating functions. The sign
when dierentiating on of the generating functions can be read from the
mnemonic gure 3.3.
q q
p p
S
3
S
1
S
4
S
2
Figure 3.3: Canonical transformations given by a generating function. Go-
ing along an arrow, in any direction, gives the new variable as the end point.
Going in the arrows direction corresponds to dierentiating, w.r.t. the vari-
able at the starting point, the generating function at the starting point and
nothing else whereas going opposite the arrows direction corresponds to dif-
ferentiating, and changing the sign.
Example 9. Consider the generating function S
2
given by
S
2
( p, q) = p g(q)
which gives the canonical transformation
p =

g
q
(q)

T
p q = g(q)
26 Hamiltonian Mechanics
This is seen to give the way in which a transformation of the qs extends
canonically to the ps. Such a transformation is called a contact transforma-
tion. For g(q) = q this is the identity transformation.
These propositions give a constructive way of searching for canonical
transformations which makes the Hamiltonian in the new coordinates much
simpler, e.g.

H( p, q) =

H( q). This is accomplished if for example
H

S
1
q

q=q( p, q)
, q( p, q)

=

H( q)
so we get a partial dierential equation which, if it has a solution, can be used
to nd such a generating function. The above partial dierential equation,
along with similar equations for the other generating functions, is called the
Hamilton-Jacobi equation.
3.3 Integrable Systems
For two smooth functions f, g : R
2n
R their Poisson bracket f, g is
the smooth function given by
f, g =

f(y)

T
J
1
g(y)
=
n

j=1

f
q
j
g
p
j

f
p
j
g
q
j

The following proposition is an immediate consequense of the denition.

Proposition 7. The Poisson bracket , : C

(R
2n
, R) C

(R
2n
, R)
C

(R
2n
, R), given by f, g =

f(y)

T
J
1
g(y), f, g, h C

(R
2n
, R) ,
satises:
1. f, g is bilinear in f and g.
2. f, g = g, f, skew-symmetry.
3. fg, h = fg, h + gf, h, derivation in each argument.
4. f, g, h +h, f, g +g, h, f = 0, Jacobis identity.
We then have
d
dt
F(
H
t
(y)) =
n

i=1
F
y
i
(
H
t
(y))

H
t
(y)

i
=

F(
H
t
(y)

T
J
1
H(
H
t
(y))
= F, H(
H
t
(y))
3.3 Integrable Systems 27
Thus we see that F is a constant of the motion of Hamiltons equations with
Hamiltonian H if and only if F, H = 0, and therefore we also immediately
get that H is a constant of the motion since H, H = 0 due to the skew-
symmetry of the bracket. Since we get from Jacobis identity
F
1
, H = 0 , F
2
, H = 0 F
1
, F
2
, H = 0
we have that if F
1
and F
2
are constants of the motion of Hamiltons equations
then so is F
1
, F
2
.
If F
1
, . . . , F
n
are smooth and linearly independent, i.e.
Rank

F
y

= n
where F : R
2n
R
n
, y [F
1
(y) . . . F
n
(y)]
T
, then we know from the
implicit function theorem that F = c = constant denes an n dimensional
subset of R
2n
for which it is possible to assign coordinates. If furthermore
F
i
, F
j
= 0, for all i, j, we have since
F
i
, F
j
= 0
F
i
t

F
j
s
=
F
j
s

F
i
t
that we can take s
1
, . . . , s
n
as local coordinates on
c
= y R
2n
[F(y) = c
dened by

F
1
s
1
. . .
F
n
s
n
(y
0
) : R
n

c
where y
0

c
. These considerations are of importance in the proof of the
following theorem
Theorem 7 (Liouville-Arnold theorem). Assume there exist n smooth func-
tions F
1
, . . . , F
n
: R
2n
R, such that F
i
, F
j
= 0 for all i, j. Assume
F
1
, . . . , F
n
are linearly independent on a level set
c
. Furthermore assume
that
c
is compact, then:
1.
c
is dieomorphic to the n-dimensional torus T
n
= R
n
/Z
n
2. There exists a neighborhood U R
n
of c such that the set
D
U
=

is dieomorphic to U
c
and in D
U
there exists a canonical
transformation (p, q) = (I, ), dened for I B R
n
and T
n
,
D
U
= (B T
n
), such that I is constant on tori and, conversely,

F = F depend only on I, i.e.

F(I, ) =

F(I).
Proof. See [1]
From this we immediately get
28 Hamiltonian Mechanics
B
R
n
T
n
Figure 3.4: Integrable system. The foliation of B R
n
with tori with quasi
periodic motion.
Corollary 1. Consider Hamiltons equations y = J
1
H(y). Let F
1
= H
and assume that there exists n 1 smooth functions F
j
: R
2n
R, j =
2, . . . , n, such that F
i
, F
j
= 0 for all i, j = 1, . . . , n, i.e. F
j
, j = 1, ..., n,
are constants of the motion of the ow of y = J
1
F
1
(y). If F
1
, . . . , F
n
are
linearly independent on the level sets
c
of the n constants of the motion
and if these level sets are compact, then each of them is dieomorphic to
the n-dimensional torus T
n
. In a neighborhood of
c
there exist a canonical
coordinate transformation : R
2n
R
2n
, (p, q) = (I, ), such that the
new hamiltonian

H = H is given by

H(I, ) = h(I)
The I-coordinates are called action variables and the coordinates are
called the angle variables, and together (I, ) are called action-angle vari-
ables. Hamiltonian systems which satisfy the assumptions of the Liouville-
Arnold theorem are said to be integrable. Thus for integrable systems the
dynamics is given by

I = 0
= (I)
where (I) =
h
I
(I). So the motion takes place on a torus T
n
, one for each
value of I, with constant angular velocity. Such a motion is said to be quasi
periodic with frequency (I).
Poincar Maps for integrable Systems with 2 Degrees of Freedom
If we follow the procedure explained earlier about Poincar maps for Hamil-
tonian systems with 2 degrees of freedom, we know according to the Liouville-
Arnold theorem that for every initial condition the trajectory will lie on a
torus T
2
imbedded in W
E
R
3
. Thus when making a Poincar map for
3.4 Nearly Integrable Systems 29
such a system each point will lie on on a closed curve, and if we change the
denition of direction in the map the points of this map will lie on another
closed curve. Therefore a Poincar map can immediately show if a system
with 2 degrees of freedom is not integrable.
3.4 Nearly Integrable Systems
Two cornerstones in pertubation theory for integrable Hamiltonian systems
are the KAM theorem and the Nekhoroshev theorem. The KAM theorem
gives sucent conditions for a perturbed integrable system to behave qual-
itatively as the integrable system. The Nekhoroshev theorem instead gives
conditions on the system for the action variables to be almost conserved quan-
tities of the motion. Here we will only state the theorems without proofs since
these are very long and technical.
A nearly integrable Hamiltonian system is a system with Hamilto-
nian H : B T
n
R, B R
n
, of the form
H(I, ) = h(I) + f(I, ) > 0 (3.6)
Suppose we can nd a canonical transformation (I, ) = (

I, ) which
simplies the Hamiltonian as

H =

h(

I) + g(

I) +
2

f(

I, )
Such a generation function must depend on and for = 0 it should be
the identity. A near the identity canonical transformation is a canonical
transformation =

, small, with
lim
0

= id
Therefore we need a near the identity canonical transformation to simplify
the Hamiltonian. It can be proven that any near the identity canonical
transformation can be constructed using a generating function of the form
S

2
(

I, ) =

I

I, , )
Since we will need only the rst order part in we will use
S
2
(

I, ) =

I (

I, )
According to proposition 6 we have
I =

I

I, )
=

I
(

I, )
30 Hamiltonian Mechanics
Inverting these we get
I =

I

I, ) +O(
2
)
= +O()
and the hamiltonian transforms to

H(

I, ) = h(

I) +

I)


(

I, ) + f(

I, )

+O(
2
)
where =
h
I
. Thus, when dropping the tilde, we wish to nd a function
such that
g(I) = (I)

(I, ) + f(I, ) (3.7)

for some function g. Since is a function on B T
n
it can be written
(I, ) =

Z
n

(I)e
i
and therefore

= 0
where '` =
1
(2)
n

2
0
(I, )d
1
. . . d
n
denotes the average in the angle
variables. Equation 3.7 therefore gives g(I) = 'f` and simplies to
(I)

(I, ) = f(I, ) 'f` (3.8)

This equation however can generally not be solved, a fact which is known as
the Poincar diculty, and can be stated precisely as follows
Proposition 8. Let the Hamiltonian be given by
H(I, ) = h(I) + f(I, ) (I, ) B T
n
and assume:
1. h is non degenerate, i.e.
det

2
h
II

= 0
in an open subset B
0
B.
3.4 Nearly Integrable Systems 31
2. f has essentially full Fourier series, more precisely, denoting
f(I, ) =

Z
n
f

(I)e
i
for any Z
n
there exits

parallel to such that f

(I) = 0 in B
0
.
Then in B
0
there doesnt exist a function solving equation 3.8
Proof. Since
(I)

Z
n
i (I)

(I)e
i
we see that in order for 3.8 to be satised in B
0
it would be necessary that
i (I)

(I) = f

(I) Z
n
` 0
in B
0
. This can formally be solved as

(I) =
f

(I)
i (I)
According to assumption 1 is a dieomorphism and since Q
n
is dense in
R
n
we will have (I) = 0 in a dense subset B

0
=
1
(Q
n
(B
0
)) B
0
,
and therefore for all

parallel to , but f

= 0 for at least one

due to
assumption 2, which proves the proposition.
From the proof of this we see that in order to proceed we need to dene
sucient conditions for

(I, ) =

Z
n

(I)e
i

(I) =
f

(I)
i (I)
Z
n
` 0
to be solvable. We see that it is necesarry that is nonresonant, i.e.
(I) = 0 Z
n
It turns out however that it is not sucient that is nonresonant.
The proper suciency condition on a frequency R
n
is that there
must exist real constants > 0 and > n 1 such that
[ [

[[

Z
n
` 0
where [[ = [
1
[ + . . . [
n
[. Then is said to be Diophantine, or more
precisely (, )-Diophantine. The set
,
consisting of all (, )-Diophantine
frequencies in a ball K in R
n
, has the counter intuitive property that it has
large measure, but its complement K `
,
is open and dense in K.
32 Hamiltonian Mechanics
Example 10. Consider the interval [0, 1]. Then the set [0, 1] Q of rational
numbers in [0, 1] is countable and denote by a
i
the i-th element. Denote by
W and V the sets
W =

iN

a
i

1
2

1
4

i
, a
i
+
1
2

1
4

[0, 1] V = [0, 1] ` W
Then we have that W is open and dense in [0, 1], since the rational numbers
are dense in the real numbers. Furthermore we have
1 Measure(V ) = 1 Measure(W) 1

i=1

1
4

i
=
2
3
So the set V has the same peculiar proporties as
,
.
In the dierent theorems on nearly integrable systems some denitions
will be needed:
1. = (
I
,

) R
+
R
+
.
2. | | denotes the Euclidian norm in R
n
.
3. The strip o

= C
n
: [Im
j
[ <

, j = 1, . . . , n.
4. The supremum norm | |

of a function F : o

C
|F|

= sup
S

|F()|
5. The Fourier norm | |

of a function F : o

C, periodic of real
period 2 in each argument
|F|

Z
n
[F

[e

||
6. For a ball B R
n
B

IB

(I) ,

(I) =

I C
n
: [

I
j
I
j
[ <
I
, j = 1, . . . , n
7. T

= B

8. The Fourier norm ||

of a function F : T

C, f = F(I, ), periodic
of real period 2 in
i
|F|

= sup
IB

|F(I, )|

A complex function F : U C, U C
n
open, is said to be analytic if it is
dierentiable in every point in U or equivalently if its Taylor series in around
every point in U converges in a neighbourhood of that point (see e.g. [14]).
3.4 Nearly Integrable Systems 33
3.4.1 The KAM Theorem
This theorem is named after A. N. Kolmogorov who rst proved it (in 1954)
and V. I. Arnold, and J. K. Moser who shortly after proved some slightly
dierent theorems about esentially the same subject. Here we give the the-
orem in the form originally given by Kolmogorov.
Consider the Nearly integrable system with Hamiltonian
H(I, ) = h(I) + f(I, ) (I, ) B T
n
Consider a xed I

B and dene
H

(J, ) = H(I

+J, )
Then a Taylor expansion gives
H

(J, ) = h(I

) +

J +
1
2
J J
+

A() + B() J +
1
2
C()J J

+O(|J|
3
)
where

= (I

) =

2
h
II
(I

)
A() = f(I

, ) B() =
f
I
(I

, ) C() =

2
f
II
(I

, )
Kolmogorovs theorem then asserts that under certain conditions it is possible
to construct a canonical transformation such that A and B vanish the
resulting Hamiltonian is then said to be in Kolmogorov weak normal form.
?
R
n
T
n
B
Figure 3.5: Nearly integrable system. The foliation of a large subset, with
respect to measure, of B R
n
with tori with quasi periodic motion.
34 Hamiltonian Mechanics
Theorem 8 (Kolmogorovs theorem). Consider the Hamiltonian system with
Hamiltonian
H(I, ) = h(I) + f(I, ) (I, ) B T
n
and suppose it is analytic in a complex neighborhood T

of BT
n
with some
= (
I
,

). Let I

B be such that

= (I

) is (, )-Diophantine and
=

2
h
II
(I

) is invertible, and dene H

(J, ) = H(I

+J, ).
Then there exist a positive constant c, depending on n, , , |

|, and
|
1
|, such that if |f|

< c, then there exists a neighborhood B

0
of J = 0,
and a canonical transformation (J, ) = (J

), dened in B
0
T
n
, such
that the new Hamiltonian H

= H

is in Kolmogorov weak normal form

H

(J

) =

+
1
2

+O(|J

|
2
)

being a constant matrix close to .

Proof. See [4]
From this it follows directly
Corollary 2. Under the conditions as in Kolmogorovs theorem. For each
Diophantine

, if the norm of the pertubation is suciently small, the origi-

nal Hamiltonian system has an invariant torus, close to the unpertubed torus
I

T
n
, and on this torus the motion is quasi periodic with frequency

.
It can be proved that the set of invariant tori has relative measure diering
from 1 by at most quantities of order O(
1/4
).
Poincar Maps for Nearly Integrable Systems with 2 Degrees of
Freedom
If we investigate a Poincar map for a nearly integrable system with 2 degrees
of freedom we know that, if the conditions of Kolmogorovs theorem are
satised, then the trajectory is most likely lying on a torus T
2
imbedded in
W
E
R
3
. If we for such a system plot several Poincar maps in W
E
, E
being the same for all of them, we then have that most of what we will see
are closed curves lying inside closed curves, and since two tori T
2
in R
3
will
conne trajectories in between, a motion, which is no longer lying on a torus,
will still be conned to lie in a bounded region.
3.4.2 Nekhoroshevs Theorem
The KAM theorem was concerned with which tori persist under a pertu-
bation of the system. The Nekhoroshev theory presents a rather dierent
perspective to nearly integrable systems, since it is concerned with sucient
conditions on the system to insure that the action variables change only a
3.4 Nearly Integrable Systems 35
little along the motion.
The Nekhoroshev theorem can be stated as follows
Theorem 9 (Nekhoroshev). Consider a nearly integrable Hamiltonian sys-
tem
H(I, ) = h(I) + f(I, ) (I, ) B T
n
and assume:
1. H is analytic in the complex neighborhood T

of B T
n
.
2. h is (l, m)-quasi-convex in B, namely the system
[(I) v[ < l|v| [

I
(I)v v[ < m|v|
2
has no solution for I B.
Denote
0
= inf
IB
|(I)|. Then there exist positive constants

, c
1
,
c
2
depending on , l, m,
0
, |h|

, |f|

, and positive constants a and b

depending only on n, such that any motion (I(t), (t) with initial values
(I
0
,
0
) B T
n
satises
|I(t) I
0
| < c
1

b
for [t[ < c
2
e
(

/)
a
Proof. See [6]
The class of Hamiltonian systems with Hamiltonian H(I, ) = I,
R
n
, which is clearly seen to be integrable, is called isochronous systems.
A theorem regarding the stability of the actions for a pertubed isochronous
system is the following, which is considerably easier to prove than Nekhoro-
shevs theorem
Theorem 10. Consider a pertubed isochronous system
H(I, ) = I + f(I, ) (I, ) B T
n
Assume:
1. f is analytic in the complex neighborhood T

of B T
n
.
2. is (, n)-Diophantine.
3. <

= cc
3
/|f|

, where c
3
=
I

and c is a specic positive

constant depending on n.
Then for initial values (I(0), (0)) = (I
0
,
0
) BT
n
we have, with c
4
=
c

and a =
1
n+1
, that
|I(t) I
0
| 3

I
for [t[ < c
4
e
(

/)
a
Proof. See [6]
These two theorems dealing with the stability of the actions of nearly
integrable systems are some of the strongest theorems in this area.
36 Hamiltonian Mechanics
3.5 Symplectic Integrators
As described in [10] a symplectic integrator
h
: R
n
R
n
is an integrator
which is also a symplectic map. Examples include the symplectic Euler
method, the Strmer/Verlet method and the splitting method

h
=
H
1
h/2

H
2
h

H
1
h/2
The motivation for studying symplectic integrators comes from the fact that
the exact ow of Hamiltons equations is a symplectic map, and therefore
when studying Hamiltons equations numerically a better performance could
be anticipated if using a symplectic integrator.
Before going in to details with the important theorems on symplectic in-
tegrators, we will rst describe a very usefull way of writting the Taylor series
for the ow of a dierential equation, and give some necessary denitions .
A function h : U U,U R
d
or U C
d
, is said to be analytic in
z
0
C
d
if the Taylor series for h in a neighborhood of z
0
converges. If we
consider a dierential equation given by the vector eld f : R
d
R
d
x = f(x)
then if f is analytic the solution
t
must be analytic (see e.g. [7]), therefore
we have in a neighborhood of t = 0

t
(x
0
) =

k=0
t
k
k!
d
k
dt
k

t=0

t
(x
0
)
Since for a dierentiable function g : R
d
R we have
d
dt
g(
t
(x
0
)) =
g
x
(
t
(x
0
))
t
(x
0
)
=
g
x
(
t
(x
0
)) f(
t
(x
0
))
The Lie derivative with respect to the vector eld f, L
f
, is for a dierentiable
function g : R
d
R dened as
L
f
g =
g
x
(x
0
) f(x
0
)
and the Lie derivative with respect to the vector eld f, for a dierentiable
function g : R
d
R
c
, g = (g
1
, . . . , g
c
), is dened as L
f
g = (L
f
g
1
, . . . , L
f
g
c
).
Using these expressions we have by induction
d
k
dt
k

t=0
g(
t
(x
0
)) =

L
k
f
g

(x
0
)
3.5 Symplectic Integrators 37
With this we get for g and f analytic in a neighborhood of t = 0
g(
t
(x
0
)) =

k=0
t
k
k!
d
k
dt
k

t=0
g(
t
(x
0
))
=

k=0
t
k
k!
L
k
f
g

(x
0
)
=

exp(tL
f
)g

(x
0
)
From this we see that for f analytic we have in a neighborhood of t = 0 that

t
(x
0
) =

exp(tL
f
)x

(x
0
)
Two more denitions which are quite similar to some of the denitions
necessary for the theory on nearly integrable systems are needed. Let
U R
d
and R
d
,
i
> 0, then dene
|

xU
z C
d
: [z
i
x
i
[
i
, i = 1, ..., d
For a function w : |

C
d
we dene the norm | |

as
|w|

= max
i{1,...,d}
sup
xU

[w
i
(x)[

i
With this in hand we can state the main theorem regarding symplectic inte-
grators as follows.
Theorem 11. Consider the mapping

: U U, U R
d

(x) = x +

k=1

k
(x) 0
Assume the functions
k
, as extensions of functions to |

, are real analytic

in |

and satisfy there the estimates

|
k
|

k1

for some positive constants and . Denote

= 4 max(, )
Then there exists a formal series of vector elds
f

= f
1
+ f
2
+
2
f
3
+ . . .
analytic in |

such that:
38 Hamiltonian Mechanics
1. One has formally

= exp

L
f

x
2. The vector elds f
k
satises the estimates
|f
1
|

|f
k
|
/2
<
1
2
k
k1

k1
, k 2
3. Let

=
1
2e
, then the ow
t
of the vector eld
f
r

= f
1
+ f
2
+ . . . +
[

/]1
f
[

/]
satises
|

|
/4
< 3e
[

/]
where [] gives the integer part.
4. If

is a symplectic map, then all the vector elds f

1
, f
2
, . . . are
Hamiltonian, i.e. f
1
= J
1

H
1
, f
2
= J
1

H
2
, . . ., for some Hamil-
tonians

H
1
,

H
2
, . . ..
Proof. See [5]
The meaning of formal in this theorem simply means that f

can be
written in the stated way, but this series doesnt necessarily converge. Since

is analytical in = 0 its Taylor series converge, and item 1 in the theorem

gives the way to determine the Taylor coecients of

even though the

series f

may not converge.

Thus according to 1 of this theorem we see that formally the map

is
the ow for time of the formal vector eld f

.
If we consider an analytic symplectic integrator
h
of order m approxi-
mating a solution to Hamiltons equations y = J
1
H, with analytic Hamil-
tonian, we then get

H
h

h
= exp (hL
J
1
H
) y exp

hL
(J
1

b
H
1
+hJ
1

b
H
2
+...)

y
=

id + hL
J
1
H
+ . . . +
h
m
m!
L
m
J
1
H
+ . . .

id + hL
(J
1

b
H
1
+hJ
1

b
H
2
+...)
+ . . . +
h
m
m!
L
m
(J
1

b
H
1
+hJ
1

b
H
2
+...)
+ . . .

y
3.5 Symplectic Integrators 39
but due to the order

H
h

h
= O(h
m+1
)
we get when collecting terms, using L
hf
= hL
f
, that

H
1
= H

H
k
= 0 k = 2, . . . , m
Combining this and theorem 11 we get
Corollary 3. Let
h
: U U, U R
2n
, be a symplectic integrator of order
m approximating a solution to Hamiltons equations y = J
1
H, with ana-
lytic Hamiltonian, and assume that
h
satises the assumptions of theorem
11 and dene , h

and according to this theorem. Then there exist a

Hamiltonian

H = H + h
m

H
m+1
+ . . . + h
[h

/h]1

H
[h

/h]
such that
|
b
H
h

h
|
/4
< 3he
[h

/h]
The Hamiltonian

H is called the modied Hamiltonian of the sym-
plectic integrator.
This corollary is very important in connection with numerical calcula-
tions using symplectic integrators, since a computer only has a nite pre-
cision. Therefore eectively a symplectic integrator can be constructed, by
choosing the timestep h small enough, such that the numerical algorithm
exactly, that is up to machine accuracy, solves a Hamiltonian system which
is a pertubed version of the Hamiltonian system under consideration. This
corollary thus gives a very usefull connection between symplectic integrators
and pertubation theory for Hamiltonian systems.
Theorem 11 and corollary 3 is used to show the following proposition.
Proposition 9. Let
h
, m, H,

H, , h

, and be as in corollary 3, and

suppose
h
(y
0
) stays within a compacs set K, where y
0
U. Then there
exist a positive constant such that

H(
i
h
(y
0
))

H(y
0
)

3e

[h

/h]
2

H(
i
h
(y
0
)) H(y
0
)

= O(h
m
)
for ih e
[h

/h]
2
.
Proof. Due to theorem 11 item 2 we know that there exist a Lipshitz constant
such that for y
1
, y
2
U we have
[

H(y
2
)

H(y
1
)[ |y
2
y
1
|
/4
40 Hamiltonian Mechanics
and thus
[

H(
i
h
(y
0
))

H(y
0
)[
i

j=1
[

j
h
(y
0
)

b
H
h
(
j1
(y
0
))

[
i|
h

b
H
h
|
/4
3e

[h

/h]
2
.
Since a continous function on a compact set attains a maximum we get
using corollary 3

H(
i
h
(y
0
)) H(y
0
)

H(
i
h
(y
0
))

H(y
0
) + h
m

H
m+1
(y
0
) h
m

H
m+1
(
i
h
(y
0
))
+. . . + h
[h

/h]1

H
[h

/h]
(y
0
) h
[h

/h]1

H
[h

/h]
(
i
h
(y
0
))

H(
i
h
(y
0
))

H(y
0
)

+ 2h
m
max
yK

H
m+1
(y)

+ . . . + 2h
[h

/h]1
max
yK

H
[h

/h]
(y)

2h
m
M
m+1
+ . . . + 2h
[h

/h]1
M
[h

/h]
+ 3e

[h

/h]
2
So the dominating term in this expression is seen to be of order O(h
m
).
Remark 3. Corollary 3 and proposition 9 shows why standard step size
control cannot be immediately applied to a symplectic integrator without
aecting it in a negative way. This comes from the fact that by regulating the
step size the modied Hamiltonian will become time dependent and a time
dependent Hamiltonian has not conservation of the Hamiltonian. Therefore
a symplectic integrator with standard step size control can not be expected
to have good properties when it comes to conservation of the Hamiltonian.
If we instead consider a completely general method
h
of order m we get
when performing the same estimates as in this proposition and again using
theorem 11

H(
i
h
(y
0
)) H(y
0
)

j=1

H(
j
h
(y
0
)) H(
j1
h
(y
0
))

j=1
|
j
h
(y
0
)
j1
h
(y
0
)|
/4
Since |
j
h
(y
0
))
j1
h
(y
0
)|
/4
= O(h
m+1
) we get

H(
i
h
(y
0
)) H(y
0
)

= O(th
m
)
where t = ih. Thus for a completely general nonsymplectic method we see
that the error in conservation of energy grows up to linearly in time, so
when it comes to energy properties symplectic integrators therefore behave
superior to nonsymplectic methods.
3.5 Symplectic Integrators 41
Modied Hamiltonians for Symplectic Splitting Methods
The easiest methods to calculate modied Hamiltonians for are the splitting
methods, for which the BCH theorem (see e.g. [10]) gives an elegant way
of expressing them using the Poisson bracket. This also gives that for a
symmetric splitting method the modied Hamiltonian only consists of even
terms in h. Consider the Hamiltonian system with Hamiltonian
q
4 2
-1
p
2
6
0
-4
1
-2
-2 -6 0
Figure 3.6: The pendulum. The solid curves are level curves for the pen-
dulum Hamiltonian and the dashed curves are level curves of its modied
Hamiltonian for a 1st order splitting method.
H = H
1
+ H
2
where H
1
and H
2
are analytic, and the two symplectic methods for solving
Hamiltons equations y = J
1
H

h
=
H
1
h

H
2
h

S
h
=
H
1
h/2

H
2
h

H
1
h/2
of order 1 and 2 respectively. Then the modied Hamiltonian

H = H +
h

H
2
+ h
2

H
3
+O(h
3
) for
h
is given by

H
2
=
1
2
H
1
, H
2

H
3
=
1
12
(H
1
, H
2
, H
2
+H
2
, H
1
, H
1
)
whereas the modied Hamiltonian

H
S
= H + h
2

H
S
3
+O(h
4
) for
S
h
is given
by

H
S
3
=
1
24
H
2
, H
1
, H
1
+
1
12
H
1
, H
2
, H
2

Higher order terms are likewise composed of Poisson brackets of Poisson

brackets and so forth of H
1
and H
2
.
42 Hamiltonian Mechanics
Example 11. Consider the pendulum which has Hamiltonian
H =
1
2
p
2
cos(q)
If the Hamiltonian is split into its kinetic energy H
1
=
1
2
p
2
and potential
energy H
2
= cos(q) then we get that the modied Hamiltonian for the rst
order symplectic splitting method
h
=
H
1
h

H
2
h
has rst order element

H
2
=
1
2
H
1
, H
2

=
1
2
[p 0]

0 1
1 0

0
sin(q)

=
1
2
p sin(q)
Level curves for H and H + h

H
2
are shown in gure 3.6 with h =
1
2
.
Bibliography
[1] V. I. Arnold, Mathematical Methods of Classical Mechanics, Springer
(1989).
[2] G. Benettin, The Elements of Hamiltonian Pertubation The-
ory, Unpublished lecture notes (2001). Can be downloaded from
www.math.unipd.it/benettin/
[3] G. Benettin, A. M. Cherubini, and F. Fass, A Changing-chart Sym-
plectic Algorithm for Rigid Bodies and other Hamiltonian Systems on
Manifolds, SIAM J. Sci. Comput. 23, 1189-1203 (2001).
[4] G. Benettin, L. Galgani, A. Giorgilli, and J. -M. Strelcyn A Proo of
Kolmogorovs Theorem on invariant Tori Using Canonical Transforma-
tions Dened by the Lie Method, Nuovo Cimento B 79, 201-223 (1984).
[5] G. Benettin, and A. Giorgilli, On the Hamiltonian Interpolation of Near-
to-the-Identity Symlectic Mappings with Application to Symplectic Inte-
gration Algorithms, Journal of Statistical Physics 74, 1117-1143 (1994).
[6] A. M. Bloch, J. Baillieul, P. Crouch and J. E. Marsden, Nonholonomic
Mechanics and Control, Springer (2003).
[7] E. A. Coddington and N. Levinson, Theory of Ordinary Dierential
Equations, McGraw-Hill (1955).
[8] I. M. Gelfand and S. V. Fomin, Calculus of Variations, Dover (2000).
[9] E. Hairer, C. Lubich, and G. Wanner, Geometric Numerical Integra-
tion: Structure-Preserving Algorithms for Ordinary Dierential Equa-
tions, Springer (2002).
[10] E. Hairer, C. Lubich, and G. Wanner, Geometric Numerical Integration
Illustrated by the Strmer/Verlet Method, Acta Numerica 12, 399-450
(2003).
[11] A. J. Lichtenberg, and M. A. Lieberman, Regular and Stochastic Motion,
Springer (1983).
44 BIBLIOGRAPHY
[12] J. Jost, and X. Li-Jost, Calculus of Variations, Cambridge University
Press (1999).
[13] J.E. Marsden, and T. S. Ratiu, Introduction to Mechanics and Symme-
try, Springer (1994).
[14] E. B. Sa, and A. D. Snider, Fundamentals of Complex Analysis for
Mathematics, Science, and Engineering, Prentice Hall (1993).
[15] F. W. Warner, Foundations of Dierentiable Manifolds and Lie Groups,
Springer (1983).