MATHEMATICS 7302 (Analytical Dynamics)

YEAR 2016–2017, TERM 2

HANDOUT #12: THE HAMILTONIAN APPROACH TO MECHANICS

These notes are intended to be read as a supplement to the handout

from Gregory, Classical Mechanics, Chapter 14.

1 The basic set-up

I assume that you have already studied Gregory, Sections 14.1–14.4. The following
is intended only as a succinct summary.

We are considering a system whose equations of motion are written in Hamiltonian

form. This means that:
1. The phase space of the system is parametrized by canonical coordinates
q = (q1 , . . . , qn ) and p = (p1 , . . . , pn ).

2. We are given a Hamiltonian function H(q, p, t).

3. The dynamics of the system is given by Hamilton’s equations of motion

∂H
q̇i = (1a)
∂pi
∂H
ṗi = − (1b)
∂qi
for i = 1, . . . , n.
In these notes we will consider some deeper aspects of Hamiltonian dynamics.

2 Poisson brackets
Let us start by considering an arbitrary function f (q, p, t). Then its time evolution is
given by
n  
df X ∂f ∂f ∂f
= q̇i + ṗi + (2a)
dt i=1
∂qi ∂pi ∂t
n  
X ∂f ∂H ∂f ∂H ∂f
= − + (2b)
i=1
∂qi ∂pi ∂pi ∂qi ∂t

1
where the first equality used the definition of total time derivative together with the chain
rule, and the second equality used Hamilton’s equations of motion.
The formula (2b) suggests that we make a more general definition. Let f (q, p, t) and
g(q, p, t) be any two functions; we then define their Poisson bracket {f, g} to be
n  
def
X ∂f ∂g ∂f ∂g
{f, g} = − . (3)
i=1
∂qi ∂pi ∂pi ∂qi

The time-evolution equation (2) can then be rewritten in compact form as

df ∂f
= {f, H} + . (4)
dt ∂t
But the importance of the Poisson bracket in Hamiltonian mechanics goes far beyond this.

Remark. In the definition (3), the explicit time-dependence, if any, simply goes for
the ride; the important thing is how f and g depend on q and p. So, in discussing
Poisson brackets, we shall often just consider functions f (q, p) and g(q, p) and not
bother to discuss explicit time-dependence.

Let us begin by recording some fundamental properties of the Poisson bracket:

1. Bilinearity. We have
{α1 f1 + α2 f2 , g} = α1 {f1 , g} + α2 {f2 , g} (5)
and likewise for g.
2. Anticommutativity. We have
{f, g} = −{g, f } . (6)
In particular it follows that {f, f } = 0 (why?).
3. Jacobi identity. For any three functions f, g, h we have
{f, {g, h}} + {g, {h, f }} + {h, {f, g}} = 0 (7)
or equivalently (using anticommutativity)
{{f, g}, h} + {{g, h}, f } + {{h, f }, g} = 0 (8)
We will prove the Jacobi identity in the next section.
4. Product identity. For any three functions f, g, h we have
{f g, h} = f {g, h} + g{f, h} . (9)
This is an easy consequence of the product law for partial differentation; you will be
asked to prove it in the next problem set. It basically expresses the fact that the
Poisson bracket {f, g} involves first derivatives of f and of g.

2
 5. Fundamental Poisson brackets. The Poisson brackets among the canonical coor-
dinates q = (q1 , . . . , qn ) and p = (p1 , . . . , pn ) are

{qi , qj } = 0 (10a)
{pi , pj } = 0 (10b)
{qi , pj } = δij (10c)

where δij is the Kronecker delta, i.e.

(
1 if i = j
δij = (11)
0 if i 6= j

The three properties of bilinearity, anticommutativity and the Jacobi identity play
such a fundamental role in many areas of mathematics that they have been given a
name: an algebraic structure involving a “product” that is bilinear, anticommutative
and satisfies the Jacobi identity is called a Lie algebra.1 You already know two other
examples of Lie algebras:

• Vectors in R3 , equipped with the cross product a × b.

• n × n matrices, equipped with the commutator [A, B] = AB − BA.

In both cases the bilinearity and anticommutativity are obvious; I leave it to you to
check the Jacobi identity.

We can now prove an important result in Hamiltonian dynamics:

Total time derivative of a Poisson bracket. For any two functions f (q, p, t)
and g(q, p, t), we have

d n df o n dg o
{f, g} = , g + f, . (12)
dt dt dt

Despite its fairly obvious-looking form, this formula is not obvious; it requires a bit of
calculation.

Proof of (12). From the fundamental time-evolution equation (4) applied to {f, g}, we
have
d ∂
{f, g} = {{f, g}, H} + {f, g} . (13)
dt ∂t
1
After the Norwegian mathematician Sophus Lie (1842–1899), who created the theory of continuous
symmetry — what is now known as the theory of Lie groups and Lie algebras — and applied it to
differential geometry and differential equations. These theories now play a central role in many areas of
mathematics and theoretical physics.

3
The first term on the right-hand side can be transformed using the Jacobi identity and
anticommutativity:

{{f, g}, H} = −{{g, H}, f } − {{H, f }, g} (14a)

= {f, {g, H}} + {{f, H}, g} . (14b)

And for the second term on the right-hand side, we use the fact that ∂/∂t commutes with
the partial derivatives ∂/∂qj and ∂/∂pj occurring in the definition of the Poisson bracket; it
therefore follows that n ∂f o n ∂g o
∂
{f, g} = , g + f, (15)
∂t ∂t ∂t
(you should check the details!). Adding (14) and (15) and using the fundamental time-
evolution equation (4) for f and for g, we obtain (12). 

In particular, if f and g are constants of motion, then so is {f, g}. So this provides a
method for obtaining new constants of motion, given old ones! Of course, these new constants
of motion are not guaranteed to be nontrivial. (For instance, we might have {f, g} = 0.)
But here is one nontrivial example:

Example: Angular momentum. Consider a single particle in Cartesian coordinates,

so that q = (q1 , q2 , q3 ) is the position and that p = (p1 , p2 , p3 ) is the ordinary linear momen-
tum. In the next problem set I will ask you to show that three components of the angular
momentum L = q × p have the Poisson brackets

{L1 , L2 } = L3 (16a)
{L2 , L3 } = L1 (16b)
{L3 , L1 } = L2 (16c)

It follows that if two components of the angular momentum happen to be constants of

motion, then the third component of the angular momentum must also be a constant of
motion.

Note, by contrast, that nothing of the kind follows if only one component of the angular
momentum is a constant of motion. Indeed, we have seen lots of examples of systems
where one component of angular momentum (e.g. the z component) is conserved but
the other two are not.

3 A unified notation for phase space

The key idea of the Hamiltonian formulation of mechanics is the extremely symmetric role
played by the coordinates q and the conjugate momenta p — in contrast to the Lagrangian
formulation, where the coordinates q and the velocities q̇ play very different roles. So it
would be nice to introduce a notation that makes this symmetry between the q and p more
explicit.

4
 This unified notation is defined by the obvious approach of assembling the coordinates
q = (q1 , . . . , qn ) and the conjugate momenta p = (p1 , . . . , pn ) into a single vector X =
(q1 , . . . , qn , p1 , . . . , pn ) of length 2n. That is, we define phase-space coordinates X =
(X1 , . . . , X2n ) by (
qi for 1 ≤ i ≤ n
Xi = (17)
pi−n for n + 1 ≤ i ≤ 2n
We then introduce a 2n × 2n matrix Ω whose n × n blocks look like
!
0 n In
Ω = (18)
−In 0n
where In denotes the n × n identity matrix and 0n denotes the n × n zero matrix; or in more
detail, 
1
 if j = i + n
Ωij = −1 if i = j + n (19)


0 otherwise
Note that the matrix Ω is antisymmetric, and that Ω2 = −I (why?). This matrix is just
the trick we need to get the appropriate minus sign into Hamilton’s equations: namely,
Hamilton’s equations
∂H
q̇i = (20a)
∂pi
∂H
ṗi = − (20b)
∂qi
for i = 1, . . . , n can trivially be rewritten as
2n
X ∂H
Ẋ i = Ωij (21)
j=1
∂Xj

for i = 1, . . . , 2n. (I leave it to you to check that this works: you will simply need to check
separately the cases 1 ≤ i ≤ n and n + 1 ≤ i ≤ 2n.)
Likewise, in this notation the Poisson bracket of two functions f (X, t) and g(X, t) takes
the very simple form
2n
X ∂f ∂g
{f, g} = Ωij . (22)
i,j=1
∂X i ∂X j

(Again, you should check this!) And the fundamental Poisson brackets among the canonical
coordinates are simply
{Xi , Xj } = Ωij . (23)
(You should check this too!)
We see here the fundamental role played by the matrix Ω in defining the “geometry” of
Hamiltonian phase space; it is analogous to the fundamental role played by the identity
matrix I in defining the geometry of Euclidean space.

5
 In what follows it will also be extremely convenient to use Einstein’s summation
convention: namely, whenever an index appears exactly twice in a product — as does the
index j on the right-hand side of (21), and as do both of the indices i and j on the right-
hand side of (22) — then it is automatically considered to be summed from 1 to 2n unless
explicitly stated otherwise. So, for example, we abbreviate (21) by

∂H
Ẋ i = Ωij , (24)
∂Xj

and we abbreviate (22) by

∂f ∂g
{f, g} = Ωij . (25)
∂Xi ∂Xj
This convention saves a lot of writing of summation signs, because in practice repeated indices
are nearly always intended to be summed. (The prototype for this is matrix multiplication.)

Let me now use this unified notation to give the promised proof of the Jacobi identity
for Poisson brackets. Gregory says (p. 416) that Jacobi’s identity
is quite important, but there seems to be no way of proving it apart from crashing it
out, which is very tedious. Unless you can invent a smart method, leave this one alone.

I would like to show you that with the unified notation this proof is a fairly straightforward
calculation. So let us consider three functions f, g, h of the phase-space coordinate X; and
let us prove the Jacobi identity in the form

{f, {g, h}} + {g, {h, f }} + {h, {f, g}} = 0 . (26)

By (22) [and using the summation convention] we have

∂f ∂{g, h}
{f, {g, h}} = Ωij (27a)
∂Xi ∂Xj
 
∂f ∂ ∂g ∂h
= Ωij Ωkl (27b)
∂Xi ∂Xj ∂Xk ∂Xl
 
∂f ∂2g ∂h ∂f ∂2h ∂g
= Ωij Ωkl + . (27c)
∂Xi ∂Xj ∂Xk ∂Xl ∂Xi ∂Xj ∂Xl ∂Xk
| {z } | {z }
“term 1” “term 2”
Similarly, the other two triple brackets {g, {h, f }} and {h, {f, g}} will contains “terms 3–6”
obtained by replacing (f, g, h) by the cyclic permutation (g, h, f ) for terms 3–4 and by (h, f, g)
for terms 5–6. Now I claim that term 1 will cancel term 6, term 3 will cancel term 2, and
term 5 will cancel term 4. Let me show the proof in detail for 1 ↔ 6; the other cases will
obviously follow by cyclic permutation of (f, g, h). We have
 
∂f ∂2g ∂h ∂h ∂ 2 g ∂f
term 1 + term 6 = Ωij Ωkl + (28)
∂Xi ∂Xj ∂Xk ∂Xl ∂Xi ∂Xj ∂Xl ∂Xk

6
In the “term 6” part of this equation, let us interchange the summation indices k and l (both
of them are being summed from 1 to 2n, so we have a right to interchange their names):
since Ωlk = −Ωkl , we have
 
∂f ∂2g ∂h ∂h ∂2g ∂f
term 1 + term 6 = Ωij Ωkl − (29a)
∂Xi ∂Xj ∂Xk ∂Xl ∂Xi ∂Xj ∂Xk ∂Xl
 
∂f ∂h ∂h ∂f ∂2g
= Ωij Ωkl − (29b)
∂Xi ∂Xl ∂Xi ∂Xl ∂Xj ∂Xk
def
= Ωij Ωkl Fijkl (29c)

where Fijkl is symmetric in j, k and antisymmetric in i, l. But we then have

Ωij Ωkl Fijkl = Ωlk Ωji Flkji by renaming dummy indices (30a)
= Ωkl Ωij Flkji by antisymmetry of Ω (used twice) (30b)
= −Ωkl Ωij Fijkl since Fijkl is symmetric in j, k and antisymmetric in i, l
(30c)

But a quantity equal to its own negative must be zero: that is, Ωij Ωkl Fijkl = 0 as claimed.


4 Canonical transformations
When we were studying Lagrangian mechanics, we saw that one of its advantages over the
Newtonian formulation is that it is covariant under arbitrary changes of coordinates: that
is, instead of the original coordinates q = (q1 , . . . , qn ) we could use some new coordinates
q ′ = (q1′ , . . . , qn′ ), defined in an arbitrary way as a function of the old coordinates:

q1′ = q1′ (q1 , . . . , qn , t)

..
. (31)
qn = qn′ (q1 , . . . , qn , t)
′

If we then define the Lagrangian L′ (q ′, q̇ ′ , t) to have the same values as L, i.e.

L′ (q ′ , q̇ ′, t) = L(q, q̇, t) , (32)

it turns out that the Lagrange equations of motion for L′ are equivalent to those for L. (To
prove this directly from the differential equations is a nontrivial calculation; but as pointed
out by Gregory, Chapter 13, pp. 387–388, this is an immediate consequence of the variational
principle.)
One of the outstanding features of the Hamiltonian formalism is that it possesses an
even wider flexibility: not only can we reparametrize coordinate space q to q ′ as in the
Lagrangian formalism (with a corresponding “dual” change from p to p′ ); we can even

7
choose new coordinates that mix q and p! These transformations turn out to be of immense
importance, both theoretical and practical. Here we will only have time to scratch the surface
of the theory of canonical transformations.
Let us consider, as usual, a 2n-dimensional Hamiltonian phase space parametrized by
canonical coordinates q = (q1 , . . . , qn ) and p = (p1 , . . . , pn ). These canonical coordinates
have the fundamental Poisson brackets

{qi , qj } = 0 (33a)
{pi , pj } = 0 (33b)
{qi , pj } = δij (33c)

Now let Q1 , . . . , Qn and P1 , . . . , Pn be arbitrary functions of q and p (and also t if we wish).

We say that Q = (Q1 , . . . , Qn ) and P = (P1 , . . . , Pn ) form new canonical coordinates for
phase space if they have the correct Poisson brackets, i.e.

{Qi , Qj } = 0 (34a)
{Pi , Pj } = 0 (34b)
{Qi , Pj } = δij (34c)

Such a transformation (q, p) 7→ (Q, P ) is called a canonical transformation.

In terms of the unified notation this can be stated even more simply. We start from
canonical coordinates X = (X1 , . . . , X2n ) satisfying the fundamental Poisson brackets

{Xi , Xj } = Ωij . (35)

We then consider new coordinates Y = (Y1 , . . . , Y2n ) that depend in a completely arbitrary
way on X (and on t if we wish). The coordinates Y form new canonical coordinates if their
Poisson brackets are
{Yi , Yj } = Ωij . (36)
In this case the transformation X 7→ Y is called a canonical transformation.
But we can easily work out what this means concretely, by using the definition (22) of
Poisson brackets in the unified notation. We have
2n
X ∂Yi ∂Yj
{Yi , Yj } = Ωkl (37a)
k,l=1
∂Xk ∂Xl

= (JΩJ T )ij (37b)

if we define the Jacobian matrix of the transformation from X to Y ,

def ∂Yi
Jij = . (38)
∂Xj
So the transformation X 7→ Y is a canonical transformation if and only if

JΩJ T = Ω . (39)

8
Note that J is actually a function of X; so what we mean by (39) is that this equation
should hold for all X, i.e. everywhere in phase space.
A 2n × 2n matrix J satisfying (39) is called a symplectic matrix. So a transformation
X 7→ Y is a canonical transformation if and only if its Jacobian at every point of phase
space is a symplectic matrix.

Example 1: Linear transformations. A linear transformation Y = JX is a canonical

transformation if and only if the matrix J (which is indeed the Jacobian matrix of the
transformation) is a symplectic matrix.
One example is the transformation
Qi = pi (40a)
Pi = −qi (40b)
that interchanges the q and p (and makes one compensating sign change). You should figure
out what the matrix J is in this case, and convince yourself that it is indeed symplectic.
(Hint: Use ΩT = −Ω and Ω2 = −I.)

Example 2: Time evolution. Consider a system started at time zero at a phase-space

point X0 = (q0 , p0 ), and let it evolve for a time t. Let us write X(t) for the phase-space
point where the system arrives at time t: this is a function of the initial condition X0 and t.
So, for each value of t, we can consider X(t) as a function of X0 . Let us show that, for each
t, the transformation X0 7→ X(t) is a canonical transformation.
By hypothesis, the function X(X0 , t) satisfies the partial differential equation
∂Xi (t)
= {Xi (t), H} (41)
∂t
(where Poisson brackets are taken with respect to the canonical coordinates X0 ) together
with the initial conditions
X(0) = X0 . (42)
So let us now compute how the Poisson bracket {Xi (t), Xj (t)} varies with time:
∂ n ∂X o ∂Xj
i
{Xi (t), Xj (t)} = , Xj + {Xi , } (43a)
∂t ∂t ∂t
= {{Xi , H}, Xj } + {Xi , {Xj , H}} by the PDE (41) (43b)
= −{Xj , {Xi , H}} + {Xi , {Xj , H}} by antisymmetry (43c)
= {Xj , {H, Xi }} + {Xi , {Xj , H}} again by antisymmetry (43d)
= −{H, {Xi , Xj }} by the Jacobi identity (43e)
= {{Xi , Xj }, H} again by antisymmetry . (43f)
def
So Fij (t) = {Xi (t), Xj (t)} is a function of X0 and t that satisfies the partial differential
equation
∂Fij (t)
= {Fij (t), H} (44)
∂t

9
with initial condition
Fij (t) = {(X0 )i , (X0 )j } = Ωij . (45)
But the solution of this partial differential equation is simply the constant function Ωij !
More precisely, the constant function Ωij does solve this partial differential equation,
since {Ωij , H} = 0 [the Poisson bracket of a constant function with any other function
is zero]. And I am taking for granted that, by general theory, we can prove that the
solution is unique. Therefore, the solution can only be the constant function Ωij .

One property of canonical transformations is that they preserve phase-space volume,

i.e. | det J| = 1. This is, in fact, an immediate consequence of (39): taking determinants of
both sides, we get
(det J)2 (det Ω) = det Ω (46)
and hence (since det Ω 6= 0) det J = +1 or −1. (In fact, with more work one can prove that
det J = +1.)
Applying this in particular to the canonical transformation associated with time evolu-
tion (Example 2 above), we obtain Liouville’s theorem: time evolution under Hamilton’s
equations preserves phase-space volumes.

Let us now look more closely at infinitesimal canonical transformations. That is,
we consider a transformation X 7→ Y that is very close to the identity map, i.e.

Y = X + ǫΨ(X) (47)

for some vector function Ψ(X) — or writing it in components,

Yi = Xi + ǫψi (X) (48)

for some functions ψi (X) (1 ≤ i ≤ 2n). We now attempt to determine conditions on the
{ψi } such that this transformation is canonical through first order in ǫ; we do this by testing
the conditions (39) on the Jacobian matrix

def ∂Yi
Jij = (49)
∂Xj

of the transformation X 7→ Y . We see from (48) that the Jacobian matrix J is

J = I + ǫK (50)

where
def ∂ψi
Kij = (51)
∂Xj
is the Jacobian matrix of the transformation X 7→ Ψ. Substituting (50) into (39) and
keeping only terms through first order in ǫ, we see that the infinitesimal transformation (47)
is canonical if and only if
KΩ + ΩK T = 0 . (52)

10
Since Ω is antisymmetric, we can also write this as

KΩ − ΩT K T = 0 (53)

or in other words
KΩ − (KΩ)T = 0 . (54)
So this says that the matrix KΩ is symmetric, or equivalently that the matrix

Ω(KΩ)ΩT = ΩK (55)

is symmetric. This suggests that we should define a new vector function Φ(X) by

Φ(X) = Ω Ψ(X) (56)

— or in components,
2n
X
ϕi (X) = Ωij ψj (X) (57)
j=1

— so that its Jacobian matrix will be ΩK, i.e.

∂ϕi
= (ΩK)ij . (58)
∂Xj

Then the symmetry of the matrix ΩK says that

∂ϕi ∂ϕj
= (59)
∂Xj ∂Xi

for all pairs i, j. But this is precisely the necessary and sufficient condition for the vector
function Φ(X) to be (locally at least) the gradient of a scalar function F (X). [You know
this in 3 dimensions: a vector field is (locally at least) the gradient of a scalar field if and
only if its curl is zero. But the principle holds true in any number of dimensions.] Thus,
the infinitesimal transformation (47) is canonical if and only if there exists a scalar function
F (X) such that
∂F
ϕi (X) = . (60)
∂Xi
Left-multiplying this by Ω and using the fact that Ψ(X) = −ΩΦ(X) since Ω2 = −I, we get
2n
X ∂F
ψi (X) = − Ωij . (61)
j=1
∂Xj

This is the necessary and sufficient condition for the infinitesimal transformation (47) to
be canonical. It is convenient to get rid of the minus sign by defining G = −F ; we thus
conclude that the infinitesimal transformation (47) is canonical if and only if there exists a
scalar function G(X) such that
∂G
ψi (X) = Ωij (62)
∂Xj

11
(where we are now using the summation convention to lighten the notation). That is, every
infinitesimal canonical transformation is of the form
∂G
Yi = Xi + ǫ Ωij , (63)
∂Xj

and conversely every infinitesimal transformation of this form is canonical. We call G the
generator of this infinitesimal canonical transformation. We also write (63) in the shorthand
form
∂G
δXi = ǫ Ωij . (64)
∂Xj
But by (22) this also has an elegant expression in terms of Poisson brackets, namely

δXi = ǫ {Xi , G} . (65)

So this is one reason why Poisson brackets play such a central role in Hamiltonian mechanics:
they show how to generate infinitesimal canonical transformations.
One important special case of (65) is when the generator G is simply the Hamiltonian H:
then [by (4)] the transformation (65) is simply time evolution (forward by a time ǫ). And
we have already seen in Example 2 above that time evolution is a canonical transformation.
But this second proof, using infinitesimal transformations, is arguably simpler than the first
proof I gave you.
Another important special case is when the generator G is one of the components of
angular momentum L, say Lz . You will show in the next problem set that the (infinitesimal)
canonical transformation generated by Lz is a (infinitesimal) rotation around the z axis.

12