PHY411 Lecture notes Part 2

Alice Quillen

October 2, 2018

Contents
1 Canonical Transformations 2
1.1 Poisson Brackets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Canonical transformations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Canonical Transformations are Symplectic . . . . . . . . . . . . . . . . . . . 5
1.4 Generating Functions for Canonical Transformations . . . . . . . . . . . . . 6
1.4.1 Example canonical transformation - action angle coordinates for the
harmonic oscillator . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2 Some Geometry 9
2.1 The Tangent Bundle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.1.1 Differential forms and Wedge product . . . . . . . . . . . . . . . . . 9
2.2 The Symplectic form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.3 Generating Functions Geometrically . . . . . . . . . . . . . . . . . . . . . . 13
2.4 Vectors generate Flows and trajectories . . . . . . . . . . . . . . . . . . . . 15
2.5 The Hamiltonian flow and the symplectic two-form . . . . . . . . . . . . . . 15
2.6 Extended Phase space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.7 Hamiltonian flows preserve the symplectic two-form . . . . . . . . . . . . . 17
2.8 The symplectic two-form and the Poisson bracket . . . . . . . . . . . . . . . 19
2.8.1 On connection to the Lagrangian . . . . . . . . . . . . . . . . . . . . 21
2.8.2 Discretized Systems and Symplectic integrators . . . . . . . . . . . . 21
2.8.3 Surfaces of section . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.9 The Hamiltonian following a canonical transformation . . . . . . . . . . . . 22
2.10 The Lie derivative . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

3 Examples of Canonical transformations 25

3.1 Orbits in the plane of a galaxy or around a massive body . . . . . . . . . . 25
3.2 Epicyclic motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.3 The Jacobi integral . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.4 The Shearing Sheet . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

1
4 Symmetries and Conserved Quantities 37
4.1 Functions that commute with the Hamiltonian . . . . . . . . . . . . . . . . 37
4.2 Noether’s theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
4.3 Integrability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

1 Canonical Transformations
It is straightforward to transfer coordinate systems using the Lagrangian formulation as
minimization of the action can be done in any coordinate system. However, in the Hamil-
tonian formulation, only some coordinate transformations preserve Hamilton’s equations.
Canonical transformations, defined here as those that preserve the Poisson brackets or
equivalently the symplectic 2-form, also preserve Hamilton’s equations. A search for con-
served quantities and symmetries is equivalent to a search for a nice coordinate system
that preserves Hamilton’s equations.

1.1 Poisson Brackets

Consider a function f (q, p, t) and a Hamiltonian H(p, q) where p, q coordinates and mo-
menta. The time dependence of f
df ∂f ∂f ∂f
= q̇ + ṗ +
dt ∂q ∂p ∂t
Using Hamilton’s equations we can write this as
df ∂f ∂H ∂f ∂H ∂f
= − +
dt ∂q ∂p ∂p ∂q ∂t
We can write this short hand with a commutation relation known as the Poisson bracket
df ∂f
= {f, H} + (1)
dt ∂t
with the Poisson bracket for two functions f, g
∂f ∂g ∂f ∂g
{f, g} ≡ −
∂q ∂p ∂p ∂q
For more than one dimension
X  ∂f ∂g ∂f ∂g

{f, g} = −
∂qi ∂pi ∂pi ∂qi
i
P
and using summation notation we neglect the symbol.

2
 What are the Poisson brackets of functions equal to the coordinates and momenta,
f (q, p) = qi and g(q, p) = pi ? We calculate

{pi , pj } = 0 {qi , qj } = 0 {qi , pj } = δij

where 
1 for i = j
δij =
0 6 j
for i =
Coordinates and momenta form a kind of basis.
We can consider functions f (q, p) = q and g(q, p) = p. Inserting these functions into
equation 1 we recover Hamilton’s equations in terms of Poisson brackets
∂H
q̇ = = {q, H}
∂p
∂H
ṗ = − = {p, H}
∂q
The Poisson bracket satisfies the conditions for a Lie algebra. For functions f, g, h,

{{f, g} , h} + {{g, h} , f } + {{h, f } , g} = 0

{f, g} + {g, f } = 0
The first of these is called a Jacobi identity, the second is antisymmetry. In addition they
satisfy another condition known as a Leibnitz type of product rule

{f, gh} = g {f, h} + {f, g} h

A Lie algebra with this extra rule is called a Poisson algebra.

Remark In what contexts are these extra rules important? These relations give the
Poisson bracket Lie bracket-like constraints. The Leibnitz rule makes the Poisson bracket
behave like a derivative. Infinite dimensional continuous systems with equations of motion
corresponding to partial differential equations, a Hamiltonian and a Poisson bracket can
be adopted (obeying the above relations) that give the equations of motion, but we might
lack canonical coordinates.

1.2 Canonical transformations

A canonical transformation is a transformation from one set of coordinates q, p to a new
one Q(q, p), P (q, p) that satisfies the Poisson brackets

{Pi , Pj } = 0 {Qi , Qj } = 0 {Qi , Pj } = δij (2)

The above Poisson brackets are computed using derivatives of p, q.

3
 Using x = (q1 , q2 ...., p1 , p2 , ....pN )

{xi , xj } = ωij

with  
0 I
ω=
−I 0
and I the identity matrix.
Given a Hamiltonian, H(p, q, t), we will show that we can find a new Hamiltonian
K(Q, P, t) such that Hamilton’s equations are obeyed in the new coordinate system.
For a function f (Q(p, q), P(p, q)) using the chain rule and using summation notation

∂f ∂f ∂Qj ∂f ∂Pj
= +
∂qi ∂Qj ∂qi ∂Pj ∂qi
∂f ∂f ∂Qj ∂f ∂Pj
= +
∂pi ∂Qj ∂pi ∂Pj ∂pi

The Poisson bracket of f (Q(p, q), P(p, q)) and g(Q(p, q), P(p, q))

∂f ∂g ∂f ∂g
{f, g} = −
∂q ∂p ∂pi ∂qi
 i i  
∂f ∂Qj ∂f ∂Pj ∂g ∂Qk ∂g ∂Pk
= + +
∂Qj ∂qi ∂Pj ∂qi ∂Qk ∂pi ∂Pk ∂pi
  
∂f ∂Qj ∂f ∂Pj ∂g ∂Qk ∂g ∂Pk
− + +
∂Qj ∂pi ∂Pj ∂pi ∂Qk ∂qi ∂Pk ∂qi
   
∂f ∂g ∂Qj ∂Pk ∂Qj ∂Pk ∂f ∂g ∂Qk ∂Pj ∂Qk ∂Pj
= − − −
∂Qj ∂Pk ∂qi ∂pi ∂pi ∂qi ∂Pj ∂Qk ∂qi ∂pi ∂pi ∂qi
   
∂f ∂g ∂Qj ∂Qk ∂Qj ∂Qk ∂f ∂g ∂Pj ∂Pk ∂Pj ∂Pk
+ − + −
∂Qj ∂Qk ∂qi ∂pi ∂pi ∂qi ∂Pj ∂Pk ∂qi ∂pi ∂pi ∂qi
∂f ∂g ∂f ∂g
= {Qj , Pk } − {Qk , Pj }
∂Qj ∂Pk ∂Pj ∂Qk
∂f ∂g ∂f ∂g
+ {Qj , Qk } + {Pj , Pk }
∂Qj ∂Qk ∂Pj ∂Pk

If the new coordinates obey the Poisson brackets in equation 2 (so that the transformation
is canonical) then we can insert these relations into the above equation.

∂f ∂g ∂f ∂g
{f, g} = δjk − δjk
∂Qj ∂Pk ∂Pj ∂Qk
∂f ∂g ∂f ∂g
= −
∂Qj ∂Pj ∂Pj ∂Qj

4
This is just the definition of the Poisson bracket but with respect to our new coordinates,
P, Q rather than p, q. If the transformation is canonical then we can compute Poisson
brackets using the new coordinates and momenta. If the coordinate transformation is
canonical (the Poisson brackets of equation 2 are obeyed in the new coordinate system)
then the Poisson bracket can be computed in the new coordinate system
{f, g}|pq = {f, g}|P Q
Above we defined canonical transformations without even specifying a Hamiltonian
function. Given H(q, p), Hamilton’s equations give
q̇ = {q, H}|pq ṗ = {p, H}|pq
and using the q, p coordinates. But this is true for any time independent function including
Q(q, p) and P(q, p) so
Q̇ = {Q, H}|pq Ṗ = {P, H}|pq
and the Poisson bracket is computed using the p, q coordinate system. However if the trans-
formation is canonical then the Poisson brackets can be computed using either coordinate
system. So
Q̇ = {Q, H}|P Q Ṗ = {P, H}|P Q
but now we compute the Poisson bracket with the new coordinates P, Q. Thus the new
Hamiltonian is equivalent to the old Hamiltonian but using the new variables;
K(Q, P) = H (q(Q, P), p(Q, P))
You may notice that there is a term missing from this expression. We will discuss time
dependent transformations below.

1.3 Canonical Transformations are Symplectic

A symplectic transformation S, obeys
J = S t JS
where  
0 I
J≡
−I 0
and S t is the transpose of S and I the identity matrix. Consider a canonical transformation
P (p, q), Q(p, q) and the Jacobian matrix
∂Q(q, p) ∂Q(q, p)
 

S=
 ∂q ∂p 
∂P (q, p) ∂P (q, p) 
∂q ∂p

5
Let us compute S t JS
∂Q ∂P ∂Q ∂Q
   
 
∂q ∂q 0 I  ∂q ∂p
S t JS = 
  
∂Q ∂P 
−I 0
 ∂P ∂P 
∂p ∂p ∂q ∂p
∂Q ∂P ∂P ∂P
  

= 
 ∂q ∂q   ∂q ∂p 
∂Q ∂P   ∂Q ∂Q 
−
∂p ∂p ∂q ∂p
∂Q ∂P ∂Q ∂P ∂Q ∂P ∂Q ∂P
 
− −
= 
 ∂q ∂q ∂q ∂q ∂q ∂p ∂p ∂q 
∂Q ∂P ∂Q ∂P ∂Q ∂P ∂Q ∂P 
− −
∂p ∂q ∂q ∂p ∂p ∂p ∂p ∂p
 
0 {Q, P }
=
− {Q, P } 0

If the transformation is canonical and the Poisson brackets are satisfied, then the transfor-
mation is symplectic.

Remark Here we used the Jacobian matrix to see that the infinitesimal transformation is
symplectic. Using an exponential, the coordinate transformation is likely to be symplectic
even when we can’t write it as a matrix. It can be useful in numerical integrations to use
discrete transformations (approximating continuous sytems) that are symplectic.

1.4 Generating Functions for Canonical Transformations

Not every coordinate transformation is canonical. Furthermore, the requirement that Pois-
son brackets are satisfied does not strongly restrict the transformation.
Suppose we take a generating function F2 (q, P ) of old coordinates and new momenta
and define
∂F2 (q, P )
p =
∂q
∂F2 (q, P )
Q = (3)
∂P
Really we describe the generating function in terms of p, q or P, Q so what we mean by
F2 (q, P ) F2 (q(Q, P ), P ). The function F2 has two arguments so

p = ∂1 F2 (q, P )
Q = ∂2 F2 (q, P )

6
where ∂1 is the derivative with respect to the first argument of the function F2 (, ), and ∂2
is the derivative with respect to the second argument. Taking the second equation we can
write
Q = ∂2 F2 (q(Q, P ), P )
∂Q ∂q
= 1 = [∂1 ∂2 F2 ] (4)
∂Q ∂Q
Taking the equation for p we can write
p(Q, P ) = ∂1 F2 (q(Q, P ), P )
∂p ∂q
= [∂1 ∂1 F2 ]
∂Q ∂Q
∂p ∂q
= [∂1 ∂1 F2 ] + ∂1 ∂2 F2
∂P ∂P
∂p ∂p ∂q ∂q
We have computed relations for ∂Q , ∂P , ∂Q . We don’t need to compute a relation for ∂P
because it will cancel out from our next computation.
∂p ∂p
Take the Poisson bracket and insert the relations for ∂Q and ∂P
∂q ∂p ∂q ∂p
{q, p} = −
∂Q ∂P ∂P ∂Q
 
∂q ∂q ∂q ∂q
= ∂1 ∂1 F2 + ∂1 ∂2 F2 − ∂1 ∂1 F2
∂Q ∂P ∂P ∂Q
∂q
= ∂1 ∂2 F2
∂Q
= 1
where the last step uses equation 4.
The coordinate transformation is canonical as long as we define the new coordinate
and momenta using equations 3. Similar choices can be made for generating functions
that depend on old and new coordinates, old and new momenta or old momenta and new
coordinates. Traditionally these are denoted F1 , F2 , F3 , F4 .

1.4.1 Example canonical transformation - action angle coordinates for the

harmonic oscillator
Given coordinates φ, I we consider new coordinates
√ √
q(I, φ) = 2I sin φ p(I, φ) = 2I cos φ (5)
We check the Poisson bracket
∂q ∂p ∂q ∂p
{q, p} = −
∂φ ∂I ∂I ∂φ
= cos φ + sin2 φ = 1
2

7
verifying that this is a canonical transformation. Note that we need a factor of two within
the square root in equation 5 so that the Poisson bracket gives 1 instead of 1/2.
This is a handy canonical transformation for the harmonic oscillator with Hamiltonian
1 2
p + q2


H(p, q) =
2
In the coordinates I, φ the Hamiltonian is particularly simple

K(I, φ) = I

This system is said to be in action angle variables as I (the action) is conserved and φ̇ is
constant.
Can we find a generating function that gives this canonical transformation from p, q to
I, φ? Consider the generating function of old momenta, p, and new (φ) coordinates,

p2
F3 (p, φ) = tan φ
2
with
∂F3
= p tan φ = q
∂p
∂F3 p2
= sec2 φ = I
∂φ 2
From this we find that
tan2 φ = (q/p)2
or
sec2 φ = tan2 φ + 1 = (q/p)2 + 1
and

q/p = atanφ
1 2
(q + p2 ) = I
2
consistent with equation 5.

Remark We defined canonical transformations as transformations that preserve the Pois-

son brackets and below we will show that an equivalent statement is that they preserve
the symplectic two-form. We did not even mention a Hamiltonian in our definition. Often
you see symplectic transformations defined as those that preserved Hamilton’s equations.
However, Arnold gives an example of a transformation that preserves Hamilton’s equations
and does not preserve the Poisson bracket.

8
2 Some Geometry
Why introduce all this mathematical jargon? Dynamics can be reformulated in a coordinate-
free manor. It’s easiers to read applied maths books on dynamics with some familiarity. It
aids in thinking about dynamical problems in a more abstract and general way./
Consider coordinates p as positions or points on a manifold, M. A manifold is a
topological space where each point has a neighborhood that resembles a Euclidean space
of dimension n. A map to the Euclidean space is called a chart. A collection of charts
is called an atlas. Here we assume that our manifold is differentiable which implies that
nearby charts can be smoothly related to one another. Namely, there are transition maps
from one chart to another that are differentiable.

2.1 The Tangent Bundle

A curve is a map from R (parametrized by time t) to the manifold. On each chart (and
near a point p) this gives at each time a vector in RN . If we have a curve c in the manifold
we can consider its image in the local Euclidean space at point x. The tangents to the
curve at the point p are in the tangent space.
The tangent space at position p we can call Tp M. The tangent bundle TM consists of
the manifold M with the collection of all its tangent spaces.
It is convenient to use a coordinate basis to describe vectors
 in the tangent space, for
∂ ∂ ∂
example (ex , ey , ez ) that is sometimes written ∂x , ∂y , ∂z .
Given a curve on the manifold (described by t), the tangent to the curve is q̇. In the
Lagrangian formalism the manifold coordinates are specified by q. The Lagrangian is a
function of q (coordinate on the manifold) and q̇ (in the tangent space) and time t (along
the curve).
In the Hamiltonian formalism the manifold coordinates are specified by phase space
q, p and curves given as a function of the phase space coordinates. Here q̇ and ṗ are both
vectors in the tangent space.
A flow on the manifold is described by velocities which lie in the tangent space at each
point p on the manifold. A vector field on the manifold generates a flow on the manifold.
It is sometimes useful to keep track of the map from the manifold to the Euclidean space
φ(p) which gives a point x in Euclidean space RN . The inverse of this map φ−1 (x → p
gives points on the manifold as a function of positions in RN . A function on the manifold
f (p) then can act effectively in the Euclidean space with f ◦ φ−1 (x).

2.1.1 Differential forms and Wedge product

We define vectors as lying in the tangent space Tp M. We can write vectors as

∂ ∂ ∂
V = Vi + Vj + Vk
∂xi ∂xj ∂xk

9
in three dimensions in terms of coordinates in a chart at a particular position in the
manifold. Tangents to a curve give vectors. Curves are parametrized by a distance along
a curve or time. Another example of a vector is given by a trajectory ṗ, q̇ with
∂ ∂
V = q̇ + ṗ
∂q ∂p
We can define a cotangent space T∗p M dual to the tangent space. An element in the
cotangent space ω is a map Tp M → R. A differential of a function f is in the cotangent
space df ∈ Tp∗ M and acts on a vector V in the tangent space giving a real number

∂f
hdf, Vi = V i ∈R
∂xi
df is called a one form. In three dimensions we can write
∂f i ∂f j ∂f
df = dx + dx + dxk
∂xi ∂xj ∂xk
In a tangent space we have the tangent of a curve. Trajectories on the manifold give
tangent vectors. In the cotangent space we have the gradient of a function.
∂
The directional derivative in the direction of vector A = ai ∂x i of a function f

∂f
ai = hA, df i
∂xi
is the gradient of a function f in the direction of vector A. This can be described as the
Lie derivative of a function
∂f
LA f = ai i . (6)
∂x
One forms are members of the cotangent space at a point p or T∗p M. As the tangent
bundle TM is formed of M and its tangent spaces Tq M, the cotangent bundle T∗ M is
formed of M and its cotangent spaces T∗p M.
Differential forms can be integrated. A one-form can be integrated along a path with
a sum of the values of the form along a bunch of tangent vectors to the path.
A q form maps q vectors in Tp M → R and can be described as sums of products of one
forms, for example
aijk dxi ⊗ dxj ⊗ dxk
For example the two vectors V, W are operated on by the two-form

ω = dx ⊗ dy → ω(V, W) = vx wy

The wedge product, ∧, of one forms is an antisymmetric sum

dxi ∧ dxj = dxi ⊗ dxj − dxj ⊗ dxi

10
 dxi ∧ dxj ∧ dxk = dxi ⊗ dxj ⊗ dxk + dxj ⊗ dxk ⊗ dxi + dxk ⊗ dxi ⊗ dxj
−dxi ⊗ dxk ⊗ dxj − dxj ⊗ dxi ⊗ dxk − dxk ⊗ dxj ⊗ dxi

With
ω = dxi ∧ dxj = dxi ⊗ dxj − dxj ⊗ dxi
we compute
ω(V, W) = vi wj − vj wi .
A differential form is an antisymmetric q form
1
ω= ωµ ...µ dxµ1 ∧ ... ∧ dxµr
r! 1 r
taking an r differential form and giving an r + 1 differential form. Differential forms can be
thought of as volume elements (think volume of a parallelpiped calculated from vectors).
The exterior derivative
 
1 ∂
dω = ωµ ...µ dxν ∧ dxµ1 ∧ ... ∧ dxµr
r! ∂xν 1 r

The exterior derivative gives the boundary of a volume element. As the exterior derivative
is antisymmetric
d2 ω = 0
If ω = wx dx + wy dy then
 
∂wy ∂wx
dω = − dx ∧ dy.
∂x ∂y

This is reminiscent of the cross product.

The geometric formulation is independent of the coordinate system used in the charts.
The coefficients of a vector or a differential-form transform as a tensor.
A form ω is exact if there is a form θ such that dθ = ω.
A form is closed if dω = 0. Every exact form is closed, as d2 θ = 0, but not every closed
form is exact.
The generalized version of Stokes’ theorem relates the integral of a form ω over the
boundary of a region ∂C to the exterior derivative dω and the region C.
Z Z
dω = ω
C ∂C

Compare the above to Stokes’ theorem in three dimensions

Z Z
∇ × F · dA = F · dS
A S

11
where A is a surface bounded by a loop S = ∂A. Stokes’ theorem is also equivalent to
Gaus’ law Z Z
∇ · FdV = F · dA
V A
where V is a volume with boundary A = ∂V . In three dimensions the exterior derivative
gives ∇× or ∇· depending on the dimension of the object that is being integrated. In
three dimensions dS is a one form, dA is related to a two form and dV is related to a three
form.

2.2 The Symplectic form

The one form
θ = qi dpi
has exterior derivative
ω = dθ = dq i ∧ dpi
which is a two form. A manifold with such a two-form (that is not degenerate) is known
as a symplectic manifold. Since ω is a derivative of θ

dω = 0

The symplectic form is non-degenerate and exact. (A form ω is exact if there is a form θ
such that dθ = ω).
What does it mean to be non-degenerate? A two-form maps two vectors, η, ξ to a real
number. For every η 6= 0 there exists a ξ such that ω(η, ξ) 6= 0.
The symplectic form is connected with areas. Consider two vector fields V, W at a
point q, p.
∂ ∂
V = vqi + vpi
∂qi ∂pi
∂ ∂
W = wqi + wpi
∂qi ∂pi

   
∂ ∂ i ∂ ∂
ω(V, W) = vqi + vpi dq wqi + wpi dpi
∂qi ∂pi ∂qi ∂pi
   
∂ ∂ i ∂ ∂
− vqi + vpi dp wqi + wpi dq i
∂qi ∂pi ∂qi ∂pi
= vqi wpi − vpi wqi

For each (qi , pi ) pair we have the area of the parallelogram defined by (vqi , wpi ) and
(wqi , wpi ). The total is the sum of the areas of the n parallelograms. Is this is related

12
to Liouville’s volume theorem? Yes, as we will show the symplectic form is preserved by
Hamiltonian flows.
Consider the two form ω = dq∧dp in a new coordinate system Q, P so that q(P, Q), p(P, Q).
We compute
∂q ∂q
dq = dQ + dP
∂Q ∂P
∂p ∂p
dp = dQ + dP.
∂Q ∂P
Inserting these into ω = dq ∧ dp
   
∂q ∂q ∂p ∂p
ω = dQ + dP ∧ dQ + dP
∂Q ∂P ∂Q ∂P
 
∂q ∂p ∂q ∂p
= − dQ ∧ dP
∂Q ∂P ∂P ∂Q
= {q, p}|P,Q dQ ∧ dP,

where I have written {q, p}|P,Q as the Poisson brackets computed with P, Q. If the coor-
dinate transformation is canonical then the two form can be written

ω = dQ ∧ dP

Canonical transformations preserve the two form ω.

2.3 Generating Functions Geometrically

Consider p, q and P, Q both canonical sets of coordinates. We can look at the one forms

θ1 = pi dqi θ2 = Pi dQi

Because the two sets of coordinates are canonical, the two form

ω = dθ1 = dθ2 = dpi ∧ dqi = dPi ∧ dQi

Consider
θ1 − θ2 = pi dqi − Pi dQi
Because dpi ∧ dqi = dPi ∧ dQi we know that

d(θ1 − θ2 ) = 0

If we can find a function F (the generating function) such that

pi dqi − Pi dQi = dF

13
then we ensure that the coordinate transformation is canonical. By specifying that the
difference θ1 − θ2 is the differential of a function we guarantee that dpi ∧ dqi = dPi ∧ dQi
and so that the transformation is canonical. F serves as our generating function. If we
start with the generating function and then use it to give us the coordinate transformation
then we ensure that the coordinate transformation is canonical.
Let our difference in forms be
∂F1 (q, Q) ∂F1 (q, Q)
pdq − P dQ = dF1 (q, Q) = dq + dQ. (7)
∂q ∂Q
We associate
∂F1 (q, Q) ∂F1 (q, Q)
p= P =−
∂q ∂Q
so equation 7 cancels to zero and the transformation is canonical.
Let
∂F4 (p, P ) ∂F4 (p, P )
qdp − QdP = dF4 (p, P ) = dp + dP.
∂p ∂P
We associate
∂F4 (p, P ) ∂F4 (p, P )
q= Q=−
∂p ∂P
Let
pdq + QdP = dF2 (q, P )
The plus sign here arises because

d(pdq + QdP ) = dp ∧ dq + dQ ∧ dP = dp ∧ dq − dP ∧ dQ

and this we would require is zero for the transformation to be canonical.

∂F2 (q, P ) ∂F2 (q, P )

dF2 (q, P ) = dq + dP
∂q ∂P
giving
∂F2 (q, P ) ∂F2 (q, P )
p= Q=
∂p ∂P
for a canonical transformation.

Remark The sign of the generating functions can be flipped and the transformation is
still canonical.

14
 Given a transformation Q(q, p) and P (q, p) the transformation is canonical (symplectic)
if one of the following forms is exact
σ1 = pdq − P dQ
σ2 = pdq + QdP
σ3 = qdp + P dQ
σ4 = qdp − QdP.
For the different related generating functions
∂F1 (q, Q) ∂F1 (q, Q)
= p = −P
∂q ∂Q
∂F2 (q, P ) ∂F2 (q, P )
= p =Q
∂q ∂P
∂F3 (p, Q) ∂F3 (p, Q)
= q =P
∂p ∂Q
∂F4 (p, P ) ∂F4 (p, P )
= q = −Q.
∂p ∂P

2.4 Vectors generate Flows and trajectories

Consider a vector field, X, and a curve on the manifold, σ(t, x), that is a map R × M → M
such that the tangent vector at each point, x, on the curve is X. An integral curve σ(t, x0 )
that goes through a point x0 on the manifold at t = 0 has tangent

d µ
= Xµ

σ (t, x0 )
dt t=0
I have used local coordinates. The initial condition σ µ (t = 0, x0 ) = xµ0 . Now instead of
requiring σ to only contain a single curve going through x0 we extend σ so that it includes
curves that go through all the points on the manifold. The map σ is known as a flow
generated by the vector field X.
We can define an exponential function exp(Xt) from M to M so that we can move along
the flow from one point to another, σ µ (t, x) = exp(Xt)xµ .
A vector field generates a flow all over the manifold.

2.5 The Hamiltonian flow and the symplectic two-form

The symplectic two-form ω = dq ∧ dp is a map from two vectors to a real number. Consider
two vectors V, W,
∂ ∂
V = vq + vp
∂q ∂p
∂ ∂
W = wq + wp
∂q ∂p

15
recall that
ω(V, W ) = vq wp − wq vp
Now what if we consider the map only using vector V

ω(V, ?) = vq dp − vp dq (8)

This gives a one-form. In this way the symplectic two-form gives us a way to take a vector
(in the tangent space) and generate a one-form (in the cotangent space) from it. Conversely
using the symplectic form ω, we can take a one-form and construct a vector from it, using
equation 8.

Remark By providing an invertible map between vectors and one-forms, ω serves like the
metric tensor in general relativity or Riemannian geometry. However the map is antisym-
metric rather than symmetric and so it is not a metric.

Recall that by specifying a direction, a vector field generates a flow, or given an initial
condition, a trajectory. We can produces trajectories (q̇, ṗ) from a function (H) using this
map.
Consider the one form associated with a Hamiltonian, H(q, p) and use ω to generate a
vector from it.
∂H ∂H
dH = dq + dp
∂q ∂p
= ω(V, ?)
= vq dp − vp dq

On the second line we have used vector V (as written out in equation 8). The V vector
has components
∂H ∂H
vq = vp = −
∂p ∂q
These are associated with q̇ and ṗ for the flow. Let us define (as Arnold) the vector
generated by dH as IdH. The equations of motion (Hamilton’s equations) are equivalent
to ẋ = IdH In this way, a Hamiltonian flow is generated from the Hamiltonian function.

2.6 Extended Phase space

The statements relating the symplectic two-form to Hamiltonian flows can be illustrated
with the one-form (known as the Poincaré-Cartan integral invariant)

σ 1 = pdq − Hdt

in what is known as extended phase space or a R2n+1 space that consists of phase space
(R2n ) with addition of time or (p, q, t). The exterior derivative dσ 1 is a two form. The odd

16
dimension implies that there is always some direction (or vector V ) such that dσ 1 (V, η) = 0
for all vectors η and this direction, V , we associate with the equations of motion. Flows
along this direction are also called vortex lines. The exterior derivative
∂H ∂H
dσ 1 = dp ∧ dq − dq ∧ dt − dp ∧ dt (9)
∂q ∂p
Consider the vector
∂H ∂ ∂H ∂ ∂
V =− + +
∂q ∂p ∂p ∂q ∂t
Let us insert this into dσ 1
∂H ∂H ∂H ∂H
dσ 1 (V, ∗) = − dq − dp + dp + dq
∂q ∂p ∂p ∂q
∂H ∂H ∂H ∂H
− dt + dt
∂q ∂p ∂q ∂p
= 0

So our vector V is in fact the direction defining the vortex lines. But look again at the
vector V
∂H ∂ ∂H ∂ ∂
V = − + +
∂q ∂p ∂p ∂q ∂t
∂ ∂ ∂
→ ṗ + q̇ +
∂p ∂q ∂t
The flow generated by vector V satisfies Hamilton’s equations.

2.7 Hamiltonian flows preserve the symplectic two-form

Here we are going to use the generalized Stokes’ theorem in extended phase space to prove
that the symplectic two-form is conserved by a Hamiltonian flow.
Consider a surface region in (p, q) (call this c) and its Hamiltonian flow along t, from
t = 0 to a time t = τ causing a transformation g τ c (see Figure 1). The boundary of the
surface at t = 0 would be ∂c. Boundary at t = τ would be ∂g τ c. Flow in time from t = 0
to τ is created with an operator J. Every point on c is given a trajectory from t = 0 → τ
using J. These trajectories are vortex lines and so have tangent equal to the vector V
such that dσ(V ) = 0. The total volume of the cylinder is Jc. The sides of the cylinder are
covered by J∂c. The boundary of Jc or ∂(Jc) is the sum of the top and lower faces, c and
g τ c, and the sides, J∂c.
The generalized version of Stokes’ theorem gives
Z Z
dω = ω
c ∂c

17
Figure 1: Vortex lines are shown in red and form the cylinder boundary. The map J takes
base surface c at time t = 0 to top surface g τ c at t = τ .

The integral of a differential form of a boundary of a region in an orientable manifold ∂c

is equivalent to the integral of the exterior derivative of the form dω in the region c.
One consequence of the generalized version of Stokes’ theorem is that the integral of an
exact form over the boundary of a region is zero. In other words we apply Stokes’ theorem
Z Z
dσ = d2 σ = 0
∂c c

where the last step is zero because d2 = 0.

Let us integrate the exterior derivative of dσ 1 over the surface of the cylindrical volume
∂(Jc) or Z
dσ 1
∂(Jc)

Because of Stokes’ theorem Z Z

1
dσ = d2 σ 1 = 0
∂(Jc) Jc

Now we divide ∂(Jc) into three pieces, the bottom of the cylinder c, the top of the
cylinder g t c and the cylindrical surface J∂c. The sum of these three integrals of dσ must
be zero.
Now we perform this sum for the exact two-form dσ 1 with σ 1 = pdq−H. Integrating dσ 1
on the surface J∂c gives zero because the surface is comprised of vortex lines or null vectors.
Consequently integrating dσ 1 on the top and both of the cylinder c and g τ c must give the

18
 ∂
same result. On these surfaces the vectors are perpendicular to ∂t so dσ 1 = ω = dp ∧ dq
t
the symplectic two-form (see equation 9) and we find that g preserves the integral of the
symplectic two-form. Since the symplectic two-form gives the volume in phase space, this
is equivalent to Liouville’s theorem and also implies that a transformations generated by
a Hamiltonian flow are canonical transformations.
Previously we showed that a Hamiltonian flow preserved phase space volume (Liou-
ville’s theorem). Phase space volume conservation can also be written in terms of ω, hence
Liouville’s theorem is equivalent to saying that Hamiltonian flows preserve the two-form
ω. Because ω is preserved by a Hamiltonian flow, the flow also generates canonical trans-
formations between coordinates at any two times.

2.8 The symplectic two-form and the Poisson bracket

We ask, if canonical transformations preserve the Poisson bracket and they preserve the
two-form ω what is the relation between the Poisson bracket and the symplectic two form?
The Poisson bracket takes derivatives of functions.
∂g ∂h ∂g ∂h
{g, h} = −
∂qi ∂pi ∂pi ∂qi
= ∇x g t J∇x h
= ω(∇x g, ∇x h)

where x = (q, p) and we consider ∇x g, ∇x h as vectors. Here

 
0 I
J=
−I 0

and ω = dq ∧ dp is the symplectic two form. The above is true only in a canonical basis
where the two-form provides a direct and trivial way to convert between one-forms and
vectors. Really the gradient of a function should be considered a one-form, not a vector.
What if we write the two-form in a non-canonical basis with

ω = fij dxi ⊗ dxj

and x variables are functions of p, q a canonical set. Here fij an antisymmetric matrix
that effectively gives the wedge product. Now we use the one form to convert between the
∂ i ∂
differential forms of g, h and vectors V, W with V = v i ∂x i and W = w ∂xi .

∂g j
ω(V, ?) = dg = dx
∂xj
∂g j
v i fij dxj = dx
∂xj
∂g
v i fij =
∂xj

19
Likewise we do the same for h but in the second location in the two-form
∂h j
ω(?, W ) = dh = dx
∂xj
∂h i
wj fji dxi = dx
∂xi
∂h
wj fji =
∂xi
Since the symplectic two-form is not degenerate we can invert the matrix fij . We call the
inverse of the matrix F with F ij fjk = δki and both f and F are antisymmetric matrices.
We invert the above relations
∂g jk
F = v i fij F jk
∂xj
= v i δik = v k

∂g jk
vk = F
∂xj
∂h
wi = F ij j
∂x
Now let us compute the two-form on these two vectors
∂g jk ∂h
ω(V, W ) = j
F fki F il l
∂x ∂x
∂g j il ∂h
= δ F
∂xj i ∂xl
∂g il ∂h
= F
∂xi ∂xl
The expression on the right we recognize as similar to a Poisson bracket. So it makes sense
to define
∂g jl ∂h
{g, h} = F
∂xj ∂xl
in the non-canonical basis. In this sense we can consider the two-form like an inverse of
the Poisson bracket.

Remark It is possible to construct a Poisson bracket that cannot be converted into a

two-form. For example in 3 dimensions, the Poisson bracket
∂f ∂g
{f, g} = ijk xk (10)
∂xi ∂xj
cannot be inverted as a three dimensional asymmetric matrix has no inverse.

20
2.8.1 On connection to the Lagrangian
Integrating the one form σ 1 = pdq − Hdt is like integrating the Lagrangian
Ldt = (pq̇ − H)dt
So σ 1 gives us a way to generate the action for the associated Lagrangian.

2.8.2 Discretized Systems and Symplectic integrators

Symplectic integrators provide maps from phase space to phase space separated by a time
∆t. These are symplectic transformations, preserving phase space volume and the sym-
plectic two-form.

2.8.3 Surfaces of section

We can also consider surfaces at an angle in the flow of vortex lines (see Figure 2) and
a two form computed at the p, q on this surface but at different times. We can consider
maps generated from between the times it takes to cross a planar subspace (see Figure 3).
There are a number of ways to generate area or volume preserving maps from Hamiltonian
flows. The key point here is that the vortex lines are flow lines and these are null vectors
with respect to the symplectic form.

Figure 2: Vortex lines going through a tilted plane. A two-form can be constructed with
degrees of freedom lying in the tilted plane. Because the vortex lines give no area, this
two-form is equivalent to one at a single time.

21
Figure 3: Orbits covering a torus. A map can be constructed from the position of the
orbits each time they pass through a plane. This is known as a surface of section and is
another way to create an area preserving map from a Hamiltonian flow in 4-dimensional
phase space (with a time independent Hamiltonian). The flow lines are vortex lines and
so are null vectors with respect to the symplectic two-form in extended phase space. On a
plane that slices the torus one piece of the two form can be set to zero. The remain degrees
of freedom (in that plane) give a two form that is preserved by the map. and the map is
symplectic.

2.9 The Hamiltonian following a canonical transformation

We defined a canonical transformation as one that satisfied the Poisson brackets or equiva-
lently preserved the symplectic 2-form. We did not require that Hamilton’s equations were
preserved, or even required a Hamiltonian to determine whether a coordinate transforma-
tion was canonical. However we showed that Hamilton’s equations were satisfied using a
new Hamiltonian in the old coordinates. This is true as long as coordinate transformation
is time independent.
What happens if the canonical transformation is time-dependent?
For any function f (p, q, t) recall that
∂f
f˙ = {f, H} +
∂t
Following a time dependent canonical transformation Q(p, q, t), P (q, p, t), (satisfying Pois-
son brackets) we can insert Q or P finding
∂Q ∂P
Q̇ = {Q, H} + Ṗ = {P, H} +
∂t ∂t
A new Hamiltonian is required so that the new coordinates satisfy Hamilton’s equations.
We need to find a function K such that

Q̇ = {Q, K} Ṗ = {P, K}

22
 Consider an extended phase space defined by q, p, t. We are adding time as an extra
dimension. We can construct a form

θ = pdq − Hdt + QdP + Kdt (11)

Remember we defined a one form σ 1 = pdq − Hdt such that dσ(V, η) = 0 for V giving the
Hamiltonian flow for all vectors η. Let θ = dF2 where F2 (q, P, t) is a generating function
that now depends on t. Remember we are now working in extended phase space so
∂F2 ∂F2 ∂F2
dF2 (q, P, t) = dq + dP +
∂q ∂P ∂t
We match
∂F2
p =
∂q
∂F2
Q =
∂P
∂F2
K −H =
∂t
With these values our θ = dF2 is exact and the two Poincaré-Cartan invariants are iden-
tical and so have the same vortex lines and the same Hamiltonian flows. Thus our new
Hamiltonian (and one that by definition satisfies Hamilton’s equations) is
∂F2
K=H+
∂t
Following the same procedure for the other classes of generating functions we find
∂F1
K=H+
∂t
and similarly for F3 and F4 .

2.10 The Lie derivative

The Lie derivative evaluates the change of a scalar function, vector field or one-form, along
the flow defined by another vector field. This change is coordinate invariant and so the Lie
derivative is defined on any differentiable manifold. Once the Lie derivative is defined on
a function and vector field, the derivative can be extended so is can be evaluated on any
tensor field.
We express the action of the flow defined by vector field X in terms of coordinates
X = X µ (x) ∂x∂ µ for a local coordinate map x → p where p is a point on the manifold. The
Lie derivative of a function f is the same as the directional derivative
∂f (p(x))
LX f (p) = ∇X f (p(x)) = X µ
∂xµ

23
 How do we compute the Lie derivative of a vector? It may be helpful to compute some
Lie derivatives of functions
 
µ ∂ ν ∂f
LX (LY f ) = X Y (12)
∂xµ ∂xν
 
µ ∂ ν ∂f
LY (LX f ) = Y X (13)
∂xµ ∂xν
∂Y ν ∂ ν ∂
 
µ ∂X
(LX LY − LY LX )f = X µ µ − Y f
∂x ∂xν ∂xµ ∂xν
ν ν
 
µ ∂Y µ ∂X ∂
= X µ
−Y µ
f (14)
∂x ∂x ∂xν

We recognize the last thing as a vector which means we can associate

LX Y = (LX LY − LY LX ) = [X, Y ] = XY − Y X

giving a vector. Here the vectors operate on each other, for example
 
i ∂ j ∂
XY = X Y (15)
∂xi ∂xj
∂Y j ∂ ∂2
= Xi i + X i j
Y (16)
∂x ∂xj ∂xi ∂xj
The Lie derivative can also be defined axiomatically, asserting that it commutes with the
exterior derivative and obeys versions of the Leibniz rule.
Another definition of LX Y is

d t 
LX Y = φX Y 
dt t=0

where φtX is the flow caused by vector X.

The Lie derivative LX Y of the Y vector field along the direction of the X vector field is
a function of two vector fields. It computes the change of one vector along the direction of
the other. The Lie derivative is not the same as a covariant derivative. Using a covariant
derivative you can compute the change of a basis vector in a particle direction. It is not
necessary to have two vector fields, however you do need to know how to transport the
basis vectors of the coordinate system between neighboring regions of the manifold.
One we have a derivative of a vector we can take the Lie derivative of a one form

24
ω = ω i dxi and a vector V

hω, Vi = V i ω i
LX hω, Vi = hLX (ω), Vi + hω, LX Vi
ν ν
 
j ∂ i i µ ∂V µ ∂X ∂


X j
V ω = hLX (ω), Vi + hω, X µ
−V µ
i
∂x ∂x ∂x ∂xν
ν ν
 
ν µ ∂V µ ∂X
= hLX (ω), Vi + ω X −V
∂xµ ∂xµ
∂X ν ∂ω i
hLX (ω), Vi = −ω ν V µ µ − X j V i j (17)
∂x ∂x
We compute

∂ω i j
dω = dx ∧ dxi (18)
∂xj
We can contract a vector with a two form to get a one form
j i
 
i ∂ω j ∂ω
iX dω = hX, dωi = X −X dxj (19)
∂xi ∂xj

Note the contraction notation iV ω = hV, ωi = V i ω i . We can also compute

∂X i i j ∂ω i i j
d(iX ω) = dhX i ω i i = ω dx + j X dx (20)
∂xj ∂x
With some manipulation we can show that the Lie derivative of a one form is consistent
with Cartan’s magic formula or

LX (ω) = iX dω + d(iX ω) (21)

and that
a b
 
b ∂ω b ∂X
LX (ω) = X +ω dxa
∂xb ∂xa

3 Examples of Canonical transformations

3.1 Orbits in the plane of a galaxy or around a massive body
The Keplerian problem of a massless particle in orbit about a massive object of mass M
can be written in polar coordinates and restricted to a plane
1
L(r, θ; ṙ, θ̇) = (ṙ2 + r2 θ̇2 ) − V (r)
2

25
where V (r) = −GM/r. The associated momentum are
∂L
pr = = ṙ
∂ ṙ
∂L
pθ = = r2 θ̇ = L
∂ θ̇
where L is the angular momentum. We find a Hamiltonian
p2r L2
H(r, θ; pr , L) = + 2 + V (r) (22)
2 2r
Because the Hamiltonian is independent of θ, the angular momentum is conserved.
Let us expand this Hamiltonian using y = r − R0 and expand assuming that y  R0 .
p2r L2
H(y, θ, pr , L) = + + V (y + R0 )
2 2(y + R0 )2
p2r L2 y2 3L2
   
0 00
= + y V (R0 ) − 3 + V (R0 ) + 4
2 R0 2 R0

Near a circular orbit the term proportional to y must be zero. L = R02 Ω with Ω = θ̇ and
the angular rotation rate s
V 0 (R0 )
Ω(R0 ) =
R0
We identify
3L2
κ2 (R0 ) = V 00 (R0 ) + = V 00 (R0 ) + 3Ω(R0 )2
R04
as the frequency of radial oscillations or the epicyclic frequency.
The above Hamiltonian (equation 22) is useful to study dynamics of stars in the mid-
plane
p of a galaxy but with a different potential V (r). A circular orbit has velocity vc (r) =
0
rV (r). Many galaxies have nearly flat rotation curves with vc (r) ∼ vc with vc a constant,
corresponding to a logarithmic potential V (r) = vc2 ln r.
Above we used a Lagrangian in cylindrical coordinates to find the Hamiltonian system.
However we could have started with a Hamiltonian in cartesian coordinates
1 p
H(x, px ; y, py ) = (p2x + p2y ) + V ( x2 + y 2 )
2
The transformation is not obviously canonical
px x py y
pr = +
r r
L = xpy − ypx
p
r = x2 + y 2
θ = atan(y/x)

26
however we can check that it is using Poisson brackets. Computing derivatives
∂pr y
= (px y − py x) 3/2
∂x r
∂pr x
= (py x − px y) 3/2
∂y r
We evaluate the Poisson bracket
∂pr ∂L ∂pr ∂L ∂pr ∂L ∂pr ∂L
{pr , L} = − + −
∂x ∂px ∂px ∂x ∂y ∂py ∂py ∂y
y x x y
= (px y − py x) 3/2 (−y) − py + (py x − px y) 3/2 x + px
r r r r
L L
= −
r r
= 0

and likewise for the other brackets.

3.2 Epicyclic motion

Orbits in a disk galaxy are nearly circular. The epicylic approximation assumes that the
orbit can be described by a radial oscillation around a circular orbit. In the Keplerian
setting the radial oscillation period is the same as the orbital period. But in the galactic
setting radial oscillations are faster that the rotation and the orbit does not close (see
Figure 4). By setting pr = 0, we can define a function E(L) giving the energy of a circular
orbit and this can be inverted L(E) to give the angular momentum of a circular orbit with
energy E. We can also define a function rc (L) that gives the radius of a circular orbit with
angular momentum L. These are related by
L2
E(L) = + V (rc (L)) (23)
2rc (L)2

L = rc2 (L)Ω(L)
s
V 0 (rc (L))
Ω(L) = (24)
rc (L)

where Ω(L) is the angular rotation rate θ̇ for the circular orbit with angular momentum L.
Example: For a flat rotation curve what is Ω(L) and rc (L)? For the flat rotation
curve vc = rc (L)Ω(L) is constant. The angular momentum L = rc (L)vc , consequently
L
rc (L) =
vc

27
Figure 4: A rosette orbit. The orbit can be described in terms of radial oscillations or
epicycles about a circular orbit.

28
and
vc v2
Ω(L) = = c
rc (L) L
The energy
L2
E(L) = + V (rc ) = vc2 ln L + constant
2rc (L)2
We can consider orbits with angular momentum and energy near those of the circular
orbit. Consider the following generating function that is a function of old momenta (pr , L)
and new coordinates (θr , θnew )

p2r
F3 (pr , L; θr , θnew ) = cot θr − rc (L)pr − Lθnew
2κ(L)

The canonical transformation gives

∂F3 pr
r = − = rc (L) + cot θr
∂pr κ(L)
∂F3 p2
Jr = = − r sin−2 θr
∂θr 2κ(L)
κ(L) 2
∂F3 drc (L) p
θ = − = θnew − pr + dL2 r cot θr
∂L dr 2κ (L)
∂F3
Lnew = − =L
∂θnew

We can rewrite this so that old coordinates are written in terms of new ones that can be
directly inserted into the Hamiltonian
s
2Jr
r = rc (L) + cos θr
κ(L)
p
pr = − 2Jr κ(L) sin θr
L = Lnew
dκ(L)
drc (L) p Jr
θ = θnew − 2Jr κ(L) sin θr + dL cos(2θr )
dL 2κ(L)

We insert these new coordinates into the Hamiltonian

p2r L2
H(r, θ; pr , L) = + 2 + V (r) (25)
2 2r

29
 q
2Jr
We temporarily define r = rc (L) + δr with δr = κ(L) cos θr and expand to second order
in δr;

L2 L2
 
dV (rc )
H(θr , θnew , Jr , L) = V (rc ) + + − δr
2rc (L)2 dr rc (L)3
 2
3L2
 2
d V (rc ) δr
+ + + Jr κ(L) sin2 θr (26)
dr2 rc (L)4 2

The term dependent on δr cancels due to the definition of rc (L). The term independent of
δr
L2
g0 (L) ≡ V (rc (L)) +
2rc (L)2
dg0 (L)
= Ω(L) (27)
dL
consistent with equation 24 and the definition of the angular rotation rate Ω(L). Using
our definition for g0 (L) and δr the Hamiltonian becomes
 2
3L2

2 Jr d V (rc )
H(θr , θnew ; Jr , L) = g0 (L) + Jr κ(L) sin θr + + cos2 θr (28)
κ(L) dr2 rc (L)4

We retroactively chose κ(L) so as to make the Hamiltonian independent of the epicyclic

angle θr

d2 V (rc ) 3L2 d2 V (rc (L))

κ(L)2 = + = + 3Ω2 (L) (29)
dr2 rc (L)4 dr2
and the Hamiltonian becomes

H(θr , θnew ; Jr , L) = g0 (L) + κ(L)Jr (30)

1
We did this expansion to second order in Jr2 . This is equivalent to or can be described
as the epicyclic approximation. The action variable Jr sets the amplitude of radial os-
cillations and the frequency κ(L) is the epicyclic frequency and governs the frequency of
radial oscillations. It is possible to carry out a higher order expansion in Jr (see George
Contopoulos’s paper: Contopoulos G., 1975, ApJ, 201, 566).
It may be useful to manipulate the definitions for L, rc (L), Ω(L) and κ(L) to show that

drc (L) 2Ω(L)

= 2 (31)
dL κ (L)rc (L)
4Ω2 (L)
 
dΩ(L) 1
= 2 1− 2 (32)
dL rc (L) κ (L)

30
And expansion of g0 (L) about a particular L value, with L = L0 + l gives

4Ω2 l2
  
1
g0 (L0 + l) = g0 (L0 ) + Ω(L0 )l + 2 1 − 2 (33)
rc κ L0 2

The coefficient can also be written

4Ω2 
  
1 Ω dΩ 
1− 2  = (34)
2rc2 κ L0 rc κ2 dr L0

so as to compare to coefficients in the appendix by Contopoulos 1975.

3.3 The Jacobi integral

Consider the Hamiltonian
p2r L2
H(r, θ, pr , L) = + 2 + V (r) + g(θ − Ωb t)
2 2r
where  is small and g(θ − Ωb t) is a perturbation to the potential that is fixed in a frame
rotating with angular rotation rate Ωb . The perturbation is what one would expect for
an oval or bar perturbation, as is found in barred galaxies, which is moving through the
galaxy with a pattern speed or angular rotation rate Ωb . Similarly a planet in a circular
orbit about the Sun can cause a periodic perturbation (with a constant angular rotation
rate) on an asteroid, also in orbit about the Sun. We take a time dependent generating
function of old coordinate θ and new momenta L0

F2 (θ, L0 , t) = (θ − Ωb t)L0

giving new coordinates

∂F2
L = = L0
∂θ
∂F2
θnew = = θ − Ωb t
∂L0
The transformation only involves θ, L so we neglect pr , r in the transformation. Because
the generating function is time dependent we must add
∂F2
= −Ωb L
∂t
to the new Hamiltonian. The new Hamiltonian in the new coordinates
∂F2
K(r.θnew , pr , L0 ) = H −
∂t
2
pr L2
= + 2 − LΩb + V (r) + g(θnew )
2 2r

31
We note that the new Hamiltonian is time independent and so is conserved. This conserved
energy, computed in the rotating frame is called the Jacobi integral. It is equivalent to
the Tisserand relation when written in terms of orbital elements in the context of celestial
mechanics (orbital dynamics in a frame rotating with a planet that is in a circular orbit
around the Sun). The Tisserand relation is used to classify comets and estimate the range
of orbital changes that can be caused by gravitational assists.

3.4 The Shearing Sheet

Figure 5: An illustration of a patch of a rotating disc (on left) and how the shearing box
(on right) approximates it. Arrows are shown with respect to motion in the centre of the
disc patch (on left). In this rotating frame a circular orbit would remain fixed at the black
dot. An orbit with zero epicyclic amplitude and located at the centre of the shearing box
(on right) would also remain fixed. The orientation of our coordinate system is shown on
the right. The shearing sheet is described with Hamiltonian in equation 35 and equations
of motion (equations 37). A central particle remains fixed. Particles with no epicyclic
oscillations (in circular orbits) exhibit shear in their horizontal (x) velocities as a function
of y. Often the shearing sheet is simulated with periodic boundary conditions in both x
and y.

32
 A two-dimensional system in the plane

p2r L2
H(r, θ; pr , L) = + 2 + V (r)
2 2r
For a Keplerian system the potential V (r) = −GM/r. In a disk galaxy with a flat rotation
curve V (r) = vc2 ln r.
We want to consider motion near a particle that is in a circular orbit with radius R0
and has angular rotation rate Ω0 with
 
2 1 dV 
Ω0 =
r dr R0

First let us rescale units so that units of time are in Ω−1

0 and units of distance are R0 .
We want to go into a rotating frame moving with the particle at r = R0 = 1 and has
θ = Ω0 t = t. In our rescaled units we can do this using a generating function using old
coordinates and new momenta

F2 (θ, r; px , py ; t) = (θ − t)(px + 1) + (r − 1)py

giving
∂F2
x = =θ−t
∂px
∂F2
y = =r−1
∂py
∂F2
L = = px + 1
∂θ
∂F2
pr = = py
∂r
∂F2
= −px − 1
∂t
Our new Hamiltonian (neglecting constants)

p2y (px + 1)2

H(x, y; px , py ) = + + V (1 + y) − px
2 2(1 + y)2

The Hamiltonian is independent of x so px is a conserved quantity. The Hamiltonian is

time independent so H is constant. There is a fixed point at x = 0, y = 0, px = 0, py = 0.
For small y we can expand

V (1 + y) = V (1) + V 0 (1)y + V 00 (1)y 2 /2

33
Recall that Ω20 = R0−1 V 0 (R0 ) so with our choice of units V 0 (1) = 1. Below it will be useful
to write κ2 = V 00 (1) + 3 (usually one writes κ20 = V 00 (R0 ) + 3Ω20 but with Ω0 = R0 = 1 this
simplifies). Expanding the Hamiltonian for small y,

p2y (px + 1)2 y2

H(x, y; px , py ) ≈ + (1 − 2y + 3y 2 ) + y + V 00 (1) − px
2 2 2
I have have dropped V (1) as it is a constant. We can also drop terms that are higher than
second order in all coordinates and momenta, (dropping terms ∝ p2x y and ∝ px y 2 ) giving
us
p2y p2 κ2 y 2
H(x, y; px , py ) ≈ + x − 2px y + (35)
2 2 2
and I have used κ2 ≡ V 00 (1) + 3 for the epicyclic frequency.
Hamilton’s equations gives
∂H
= −2y + px = ẋ
∂px
∂H
= py = ẏ
∂py
∂H
= 0 = −ṗx
∂x
∂H
= −2px + κ2 y = −ṗy
∂y
We can solve for
px = ẋ + 2y
and because px is conserved
ẍ = −2ẏ
Inserting px = ẋ + 2y into the expression for ṗy

ÿ = 2ẋ + (4 − κ2 )y

For the Keplerian system κ2 = Ω2 = 1

ẍ = −2ẏ
ÿ = 2ẋ + 3y (36)

With the addition of an additional local potential these are known as Hill’s equations (and
originally derived for orbits near the Moon and usually x and y are interchanged). But in
a system with a different rotation curve (like a disk galaxy)

ẍ = −2ẏ
ÿ = 2ẋ + (4 − κ2 )y (37)

34
 If we added in a local potential (that is a function of distance from the origin) that
could be due to a local mass at the origin then the Hamiltonian would look like this
p2y p2 κ2 y 2 p
H(x, y; px , py ) ≈ + x − 2px y + + W ( x2 + y 2 ) (38)
2 2 2
Going back to the shearing sheet without any extra perturbations
p2y p2 κ2 y 2
H(x, y; px , py ) ≈ + x − 2px y + (39)
2 2 2
The momentum px sets the mean y value about which y oscillates. The oscillation frequency
for y (radial) oscillations is the epicylic frequency κ and it’s independent of the mean value
of y. The angular rotation rate (here ẋ) is set by px and this sets the mean y value. We
can look for solutions that look like y = y0 + A cos(κt + φ0 ) with phase φ0 . Conservation
of px implies that we can take

ẋ = ẋ0 − 2A cos(κt + φ0 ) (40)

with ẋ0 a constant. By subbing into equation 37 this gives

1
ẋ0 = (κ2 − 4)y0 . (41)
2
This is the velocity shear. For κ = 1 (Keplerian rotation) we recover the expected factor
of 3/2.
It may be useful to write the equations of motion

y = y0 + A cos(κt + φ0 ) (42)
1 2A
x = x0 + (κ2 − 4)y0 t − sin(κt + φ0 ) (43)
2 κ
1
ẋ = (κ2 − 4)y0 − 2A cos(κt + φ0 ) (44)
2
ẏ = −Aκ sin(κt + φ0 ). (45)

We can write the above equations in terms of a guiding center xg , yg for the epicyclic motion

y = yg + A cos(κt + φ0 ) (46)
2A
x = xg − sin(κt + φ0 ) (47)
κ
with

yg = y0 (48)
1
xg = x0 + (κ2 − 4)y0 t. (49)
2

35
We can solve for the guiding centers in terms of current positions and velocities
2vx
yg = 4y + (50)
κ2
2vy
xg = x − 2 (51)
κ
with vx = ẋ and vy = ẏ. We can solve for the epicyclic amplitude
A cos(κt + φ0 ) = (y − yg ) (52)
1
A sin(κt + φ0 ) = − κ(x − xg ) (53)
2
As there is a fixed point at x = 0, y = 0, px = 0, py = 0 the Hamiltonian (equation 39)
can be written in the form
1
H = xM x
2
with M the Hessian matrix and ẋ = ωM x and ẍ = (ωM )2 x. As H is second order in
all coordinates, M contains no variables (is just constants), conserved quantities can be
identified from zero value eigenvalues of ωM and frequencies of oscillation from eigenvalues
of (ωM )2 .
The Hamiltonian is independent of x (giving us a conserved quantity px ). An associated
Lagrangian would be independent of x. That means in a simulation we can change x of a
particle and the equations of motion will not change. That means in a simulation, periodic
boundary conditions in x would not change the equation of motion (equations 37).
There is another symmetry in the equations of motion (is this because the Jacobi
integral is conserved in the rotating frame?)
(κ2 − 4)
ẋ → ẋ + 
2
y → y+
This symmetry is exploited in simulations so that the y boundary condition can be periodic
also.
If we are simulating with the Hamiltonian system then we are change momenta and
coordinates rather than velocities and coordinates. Equations of motion are
ẋ = px − 2y
ẏ = py
ṗx = 0
ṗy = 2px − κ2 y
Consider changing y by δy but not affecting any accelerations. To maintain ṗy we require
that
κ2
δpx = δy
2

36
The acceleration in x is ẍ = ṗx − 2py is zero as long as we don’t change py . How does this
transformation affect ẋ?
1
δ ẋ = δpx − 2δy = (κ2 − 4)δy
2
and this is equivalent to the transformation given above. In terms of momenta and coor-
dinates the transformation is

y → y+
κ2
px → px + 
2
As the system has two independent conserved quantities (H, px ), without any additional
perturbations the system is integrable, and this is expected as this system is equivalent to
the epicyclic approximation. The shearing sheet, because it is non-trivial, integrable and
can be simulated with periodic boundary conditions is a nice place for particle integrations
(see recent work by Hanno Rein and collaborators, and including short range interactions
between particles).

4 Symmetries and Conserved Quantities

4.1 Functions that commute with the Hamiltonian
Recall that for a function f (p, q, t)
∂f
f˙ = {f, H} +
∂t
If the Poisson bracket
{f, H} = 0 (54)
and ∂f
∂t = 0 then f is a conserved quantity.
Consider the gradient of a function on p, q or with one-form df = ∂f ∂f
∂q dq + ∂p dq. We can
use the symplectic form to generate a flow or vectors V such that ω(V, ?) = df . In this way
f generates a direction in phase space or a tangent vector. The Hamiltonian generates a
flow. The function f also generates a flow. The function f corresponding to a conserved
quantity is also one that commutes with H and generates a flow that commutes with the
flow generated by H.
Using x = (q, p) and ∇ = (∂q , ∂p ) recall that we could write the equations of motion
as
ẋ = ω∇H
where ω is the matrix made up of a positive and negative identity matrix
 
0 I
ω=
−I 0

37
Using gradients it is possible to rewrite equation 54 as

{f, H} = ∇f t ω∇H = 0

Using the equations of motion

∇f · ẋ = 0
A conserved quantity is one that has a gradient (in phase space) perpendicular to the
direction of motion. If there are many possible such directions (there are many conserved
quantities) then the orbits must be of low dimension.
The Hamiltonian formalism relates flows that commute with the Hamiltonian flow with
conserved quantities.

4.2 Noether’s theorem

Noether’s theorem relates coordinate symmetries of the Lagrangian with a conserved quan-
tity. Consider a Lagrangian L(q, q̇, t) and a transformation of the coordinates

qs = h(s, q)

with qs=0 = h(0, q) and s a continuous parameter. We require that the transformed time
derivative
d
q̇s = qs
dt
so that h transforms both q and q̇ consistently.
The map h is a symmetry of the Lagrangian when

L(qs , q̇s , t) = L(q, q̇, t)

The Lagrangian cannot depend on s so


∂L(qs , q̇s , t) 
=0
∂s 
s=0

Using the chain rule

∂L ∂L ∂qs ∂L ∂ q̇s
= + (55)
∂s ∂q ∂s ∂ q̇ ∂s
We can write
∂ q̇s d ∂qs
= (56)
∂s dt ∂s
and using Lagrange’s equation we can replace ∂L d ∂L
∂q with dt ∂ q̇ . These inserted into equation
55 give    
∂L d ∂L dqs ∂L d ∂qs
= +
∂s dt ∂ q̇ ds ∂ q̇ dt ∂s

38
and   
∂L  d ∂L ∂qs 
= = 0.
∂s s=0 dt ∂ q̇ ∂s s=0
Hence we have a conserved quantity

∂L ∂qs 
I=
∂ q̇ ∂s s=0
or 
∂L ∂qi,s 
I=
∂ q̇i ∂s s=0
using summation notation and in more than one dimension.
Note: Rather than conserving L we should really be thinking about conserving the
action. In this case we should be considering infinitesimals in time and end points.
The Lagrangian formalism relates a symmetry of the Lagrangian to a conserved quan-
tity.
In the Hamiltonian context, we use the symplectic two form and a conserved quantity
(here a function) to generate a vector and so a flow in phase space. This flow commutes
with the Hamiltonian flow.

4.3 Integrability
A time independent Hamiltonian with N degrees of freedom is said to be integrable if N
smooth independent functions Ii can be found that

{Ii , H} = 0

that are conserved quantities and

{Ii , Ij } = 0
are in involution. The reason that the constants should be smooth and independent is that
the equations Ii (q, p) = ci , where the ci ’s are constants, must define N different surfaces
of dimension 2N − 1 in the 2N-dimensional phase space. If the conserved quantities are in
involution then they can be used as canonical momenta.
For example, consider a 4 dimensional phase space with a time independent Hamilto-
nian. An orbit is a trajectory in 4 dimensional phase space. Once we specify the energy
(Hamiltonian) then there is a constraint on the orbit and the orbit must wander in a 3-
dimensional subspace. If we specify an additional conserved quantity then the orbit must
wander in a 2 dimensional subspace. At this point we say the system is integrable. The
orbit could be a lower dimensional object if there are additional conserved quantities. A
circular orbit in the plane of an axisymmetric galaxy is a one dimensional object (points
only a function of angle). Epicyclic motion near the circular orbit covers a 2-dimensional
surface (is a torus). I can describe it in terms of two angles given angular momentum and

39
epicyclic action variable. This system has two conserved quantities (energy and angular
momentum) so is integrable. If the potential is proportional to 1/r (Keplerian setting)
then there is an additional conserved quantity (Runge-Lunz vector) and every orbit is one
dimensional (an ellipse) instead of a 2-dimensional (torus). Systems with extra conserved
quantities (past what is needed to make them integrable) are known as superintegrable.
The Keplerian system is maximally superintegrable as there are 5 conserved quantities for
every orbit in 6 dimensional phase space.
Liouville integrability means there exists a maximal set of Poisson commuting invariants
(i.e., functions on the phase space whose Poisson brackets with the Hamiltonian of the
system, and with each other, vanish).
Locally there could be a complete set of conserved quantities and it might be possible
to construct canonical transformations that give you a Hamiltonian purely in terms of
conserved quantities. But it might not be possible to find a complete set the covers the
entire manifold (in the Liouville sense of integrability).
Often we call a system integrable if the dimension covered by orbits is low. For example
in a 4 dimensional phase space if there are 2 (non-degenerate) conserved quantities (one
of them could be the Hamiltonian itself). Orbits are in a 4 dimensional space, and two
conserved quantities drops the dimension by 2 so orbits cover a surface. They are like tori.
If the system had only a single conserved quantity (like energy) we could make a surface
of section and find area filling orbits that corresponded to orbits filling a 3 dimensional
volume. Periodic orbits would be fixed points in a surface of section and these would be 1
dimensional.

40