156 CHAPTER 6.

HAMILTON’S EQUATIONS

later subdivide these into coordinates and velocities. We will take the space
in which x takes values to be some general n-dimensional space we call M,
which might be ordinary Euclidean space but might be something else, like
the surface of a sphere1 . Given a differentiable function f of n independent
variables xi , the differential is
Chapter 6 df =
n
X ∂f
dxi . (6.1)
i=1 ∂x i

Hamilton’s Equations What does that mean? As an approximate statement, this can be regarded
as saying
n
X ∂f
df ≈ ∆f ≡ f (xi + ∆xi ) − f (xi ) = ∆xi + O(∆xi ∆xj ),
We discussed the generalized momenta i=1 ∂xi

∂L(q, q̇, t) with some statement about the ∆xi being small, followed by the dropping of
pi = , the “order (∆x)2 ” terms. Notice that df is a function not only of the point
∂ q̇i
x ∈ M, but also of the small displacements ∆xi . A very useful mathematical
and how the canonical variables {qi , pj } describe phase space. One can use language emerges if we formalize the definition of df , extending its definition
phase space rather than {qi , q̇j } to describe the state of a system at any to arbitrary ∆xi , even when the ∆xi are not small. Of course, for large
moment. In this chapter we will explore the tools which stem from this ∆xi they can no longer be thought of as the difference of two positions in
phase space approach to dynamics. M and df no longer has the meaning of the difference of values of f at
two different points. Our formal df is now defined as a linear function of
these ∆xi variables, which we therefore consider to be a vector ~v lying in
6.1 Legendre transforms an n-dimensional vector space Rn . Thus df : M × Rn → R is a real-valued
function with two arguments, one in M and one in a vector space. The dxi
The important object for determining the motion of a system using the La- which appear in (6.1) can be thought of as operators acting on this vector
grangian approach is not the Lagrangian itself but its variation, under arbi- space argument to extract the i0 th component, and the action of df on the
trary changes in the variables q and q̇, treated as independent variables. It P
argument (x, ~v ) is df (x, ~v ) = i (∂f /∂xi )vi .
is the vanishing of the variation of the action under such variations which This differential is a special case of a 1-form, as is each of the operators
determines the dynamical equations. In the phase space approach, we want dxi . All n of these dxi form a basis of 1-forms, which are more generally
to change variables q̇ → p, where the pi are components of the gradient of
X
the Lagrangian with respect to the velocities. This is an example of a general ω= ωi (x)dxi ,
procedure called the Legendre transformation. We will discuss it in terms of i
the mathematical concept of a differential form.
Because it is the variation of L which is important, we need to focus where the ωi (x) are n functions on the manifold M. If there exists an ordi-
our attention on the differential dL rather than on L itself. We first want nary function f (x) such that ω = df , then ω is said to be an exact 1-form.
to give a formal definition of the differential, which we will do first for a 1
Mathematically, M is a manifold, but we will not carefully define that here. The
function f (x1 , ..., xn ) of n variables, although for the Lagrangian we will precise definition is available in Ref. [16].

155
6.1. LEGENDRE TRANSFORMS 157 158 CHAPTER 6. HAMILTON’S EQUATIONS

Consider L(qi , vj , t), where vi = q̇i . At a given time we consider q and v where d̄Q is not an exact differential, and the heat Q is not a well defined
as independant variables. The differential of L on the space of coordinates system variable. Though Q is not a well defined state function, the differential
and velocities, at a fixed time, is d̄Q is a well defined 1-form on the manifold of possible states of the system.
X ∂L X ∂L X ∂L X It is not, however, an exact 1-form, which is why Q is not a function on that
dL = dqi + dvi = dqi + pi dvi . manifold. We can express d̄Q by defining the entropy and temperature, in
i ∂qi i ∂vi i ∂qi i terms of which d̄Q = T dS, and the entropy S and temperature T are well
If we wish to describe physics in phase space (qi , pi ), we are making a change defined state functions. Thus the state of the gas can be described by the
of variables from vi to the gradient with respect to these variables, pi = two variables S and V , and changes involve an energy change
∂L/∂vi , where we focus now on the variables being transformed and ignore
P
the fixed qi variables. So dL = i pi dvi , and the pi are functions of the vj dE = T dS − pdV.
determined by the function L(vi ). Is there a function g(pi ) which reverses
P We see that the temperature is T = ∂E/∂S|V . If we wish to find quantities
the roles of v and p, for which dg = i vi dpi ? If we can invert the functions
P appropriate for describing the gas as a function of T rather than S, we define
p(v), we can define g(pi ) = i pi vi (pj ) − L(vi (pj )), which has a differential
X X X X X the free energy F by −F = T S − E so dF = −SdT − pdV , and we treat
dg = pi dvi + vi dpi − dL = pi dvi + vi dpi − pi dvi F as a function F (T, V ). Alternatively, to use the pressure p rather than V ,
i i i i i
X we define the enthalpy X(p, S) = V p + E, dX = V dp + T dS. To make both
= vi dpi changes, and use (T, p) to describe the state of the gas, we use the Gibbs
i
free energy G(T, p) = X − T S = E + V p − T S, dG = V dp − SdT . Each of
as requested, and which also determines the relationship between v and p, these involves a Legendre transformation starting with E(S, V ).
∂g Unlike Q, E is a well defined property of the gas when it is in a volume
vi = = vi (pj ), V if its entropy is S, so E = E(S, V ), and
∂pi
giving the inverse relation to pk (v` ). This particular form of changing vari- ∂E ∂E
ables is called a Legendre transformation. In the case of interest here, T = , p= .
∂S V
∂V S
the function g is called H(qi , pj , t), the Hamiltonian,
X
H(qi , pj , t) = pk q̇k (qi , pj , t) − L(qi , q̇j (q` , pm , t), t). (6.2) ∂ 2E ∂2E ∂T ∂p
As = we can conclude that = . We may also
k ∂S∂V ∂V ∂S ∂V S ∂S V
Then for fixed time, consider the state of the gas to be described by T and V , so
X X
dH = (dpk q̇k + pk dq̇k ) − dL = (q̇k dpk − pk dqk ) , ∂E ∂E
dE = dT + dV
k k ∂T ∂V
V T
∂H ∂H " !#
= q̇k , = −pk . (6.3) 1 p 1 ∂E 1 ∂E
∂pk q,t
∂qk p,t dS = dE + dV = dT + p+ dV,
T T T ∂T V
T ∂V T
Other examples of Legendre transformations occur in thermodynamics.
The energy change of a gas in a variable container with heat flow is sometimes from which we can conclude
written ! " !#
∂ 1 ∂E ∂ 1 ∂E
= p+ ,
dE = d̄Q − pdV, ∂V T ∂T V T
∂T T ∂V T V
6.1. LEGENDRE TRANSFORMS 159 160 CHAPTER 6. HAMILTON’S EQUATIONS

and therefore In Section 2.1 we discussed how the Lagrangian is unchanged by a change
∂p ∂E of generalized coordinates used to describe the physical situation. More pre-
T −p= .
∂T V
∂V T cisely, the Lagrangian transforms as a scalar under such point transforma-
This is a useful relation in thermodynamics. tions, taking on the same value at the same physical point, described in the
Let us get back to mechanics. Most Lagrangians we encounter have the new coordinates. There is no unique set of generalized coordinates which
decomposition L = L2 +L1 +L0 into terms quadratic, linear, and independent describes the physics. But in transforming to the Hamiltonian language,
of velocities, as considered in 2.4.2. Then the momenta are linear in velocities, different generalized coordinates may give different momenta and different
P Hamiltonians. An nice example is given in Goldstein, a mass on a spring at-
pi = j Mij q̇j + ai , or in matrix form p = M · q̇ + a, which has the inverse
relation q̇ = M −1 · (p − a). As H = L2 − L0 , H = 12 (p − a) · M −1 · (p − a) − L0 . tached to a “fixed point” which is on a truck moving at uniform velocity vT ,
As a simple example, with a = 0 and a diagonal matrix M , consider spherical relative to the Earth. If we use the Earth coordinate x to describe the mass,
coordinates, in which the kinetic energy is the equilibrium position of the spring is moving in time, xeq = vT t, ignoring a
negligible initial position. Thus U = 12 k(x − vT t)2 , while T = 12 mẋ2 as usual,
!
m 2  1 p2 p2 and L = 12 mẋ2 − 12 k(x − vT t)2 , p = mẋ, H = p2 /2m + 12 k(x − vT t)2 . The
T = ṙ + r2 θ̇2 + r2 sin2 θφ̇2 = p2r + 2θ + 2 φ 2 . equation of motion is ṗ = mẍ = −∂H/∂x = −k(x − vT t), of course. This
2 2m r r sin θ
shows that H is not conserved, dH/dt = (p/m)dp/dt + k(ẋ − vT )(x − vT t) =
Note that the generalized momenta are not normalized components of the −(kp/m)(x − vT t) + (kp/m − kvT )(x − vT t) = −kvT (x − vT t) 6= 0. Alterna-
ordinary momentum, as pθ 6= p~ · êθ , in fact it doesn’t even have the same tively, dH/dt = −∂L/∂t = −kvT (x − vT t) 6= 0. This is not surprising; the
units. spring exerts a force on the truck and the truck is doing work to keep the
The equations of motion in Hamiltonian form (6.3), fixed point moving at constant velocity.
On the other hand, if we use the truck coordinate x0 = x − vT t, we
q̇k =
∂H
, ṗk = −
∂H
, may describe the motion in this frame with T 0 = 12 mẋ0 2 , U 0 = 12 kx0 2 , L0 =
∂pk q,t
∂qk p,t
1
2
mẋ0 2 − 12 kx0 2 , giving the correct equations of motion p0 = mẋ0 , ṗ0 = mẍ0 =
−∂L0 /∂x0 = −kx0 . With this set of coordinates, the Hamiltonian is H 0 =
are almost symmetric in their treatment of q and p. If we define a 2N ẋ0 p0 − L0 = p0 2 /2m + 12 kx0 2 , which is conserved. From the correspondence
dimensional coordinate η for phase space, between the two sets of variables, x0 = x − vT t, and p0 = p − mvT , we see that
 the Hamiltonians at corresponding points in phase space differ, H(x, p) −
ηi = qi
for 1 ≤ i ≤ N, H 0 (x0 , p0 ) = (p2 − p02 )/2m = 2mvT p − 12 mvT2 6= 0.
ηN +i = pi
Thus the Hamiltonian is not invariant, or a scalar, under change of gen-
we can write Hamilton’s equation in terms of a particular matrix J, eralized coordinates, or point transformations. We shall see, however, that
it is invariant under time independent transformations that are more general
2N
X  
∂H 0 1IN ×N than point transformations, mixing coordinates and momenta.
η̇j = Jjk , where J = .
k=1 ∂ηk −1IN ×N 0

J is like a multidimensional version of the iσy which we meet in quantum- 6.2 Variations on phase curves
mechanical descriptions of spin 1/2 particles. It is real, antisymmetric, and
because J 2 = −1I, it is orthogonal. Mathematicians would say that J de- In applying Hamilton’s Principle to derive Lagrange’s Equations, we consid-
scribes the complex structure on phase space, also called the symplec- ered variations in which δqi (t) was arbitrary except at the initial and final
tic structure. times, but the velocities were fixed in terms of these, δ q̇i (t) = (d/dt)δqi (t).
6.3. CANONICAL TRANSFORMATIONS 161 162 CHAPTER 6. HAMILTON’S EQUATIONS

In discussing dynamics in terms of phase space, this is not the most natural became simpler. We switched from cartesian to center-of-mass spherical coor-
variation, because this means that the momenta are not varied indepen- dinates to discuss planetary motion, for example, or from the Earth frame to
dently. Here we will show that Hamilton’s equations follow from a modified the truck frame in the example in which we found how Hamiltonians depend
Hamilton’s Principle, in which the momenta are freely varied. on coordinate choices. In all these cases we considered a change of coordinates
We write the action in terms of the Hamiltonian, q → Q, where each Qi is a function of all the qj and possibly time, but not of
Z tf "X # the momenta or velocities. This is called a point transformation. But we
I= pi q̇i − H(qj , pj , t) dt, have seen that we can work in phase space where coordinates and momenta
ti i enter together in similar ways, and we might ask ourselves what happens if
we make a change of variables on phase space, to new variables Qi (q, p, t),
and consider its variation under arbitrary variation of the path in phase Pi (q, p, t), which also cover phase space, so there is a relation (q, p) ↔ (Q, P )
space, (qi (t), pi (t)). The q̇i (t) is still dqi /dt, but the momentum is varied free given by the two descriptions of the same point in phase space. We should
of any connection to q̇i . Then not expect the Hamiltonian to be the same either in form or in value, as we
Z tf "X !
X
!#
X tf saw even for point transformations, but there must be a new Hamiltonian
∂H ∂H
δI = δpi q̇i − − δqi ṗi + dt + pi δqi , K(Q, P, t) from which we can derive the correct equations of motion,
ti i ∂pi i ∂qi i ti
RP ∂K ∂K
where we have integrated the pi dδqi /dt term by parts. Note that in order Q̇i = , Ṗi = − .
∂Pi ∂Qi
to relate stationarity of the action to Hamilton Equations of Motion, it is
 
necessary only to constrain the qi (t) at the initial and final times, without Q
The analog of η for our new variables will be ζ = , and the relation
imposing any limitations on the variation of pi (t), either at the endpoints, as P
we did for qi (t), or in the interior (ti , tf ), where we had previously related pi η ↔ ζ means each can be viewed as a function of the other. Hamilton’s
∂K
and q̇j . The relation between q̇i and pj emerges instead among the equations equations for ζ are ζ̇ = J · . If this exists, we say the new variables
of motion. ∂ζ
The q̇i seems a bit out of place in a variational principle over phase space, (Q, P ) are canonical variables and the transformation (q, p) → (Q, P ) is a
and indeed we can rewrite the action integral as an integral of a 1-form over canonical transformation. Note that the functions Qi and Pi may depend
a path in extended phase space, on time as well as on q and p.
These new Hamiltonian equations are related to the old ones, η̇ = J ·
Z !
X ∂H/∂η, by the function which gives the new coordinates and momenta in
I= pi dqi − H(q, p, t)dt . terms of the old, ζ = ζ(η, t). Then
i

We will see, in section 6.6, that the first term of the integrand leads to a very dζi X ∂ζi ∂ζi
ζ̇i = = η̇j + .
important form on phase space, and that the whole integrand is an important dt j ∂ηj ∂t
1-form on extended phase space.
Let us write the Jacobian matrix Mij := ∂ζi /∂ηj . In general, M will not be a
constant but a function on phase space. The above relation for the velocities
6.3 Canonical transformations now reads

We have seen that it is often useful to switch from the original set of coor- ∂ζ
ζ̇ = M · η̇ + .
dinates in which a problem appeared to a different set in which the problem ∂t η
6.3. CANONICAL TRANSFORMATIONS 163 164 CHAPTER 6. HAMILTON’S EQUATIONS

The gradients in phase space are also related, also be continuous, or depending on a parameter. For example, with rotating
coordinates we considered x = r cos(θ + φ), y = r sin(θ + φ), which can be
∂ X ∂ζj ∂
= , or ∇η = M T · ∇ζ . viewed as a set of possible discrete transformations (r, θ) ↔ (x, y) depending
∂ηi t,η j ∂ηi t,η
∂ζj t,ζ on a parameter φ, which might even be a function of time, φ(t) = Ωt. For each
fixed φ this transformation Pφ : (x, y) 7→ (r, θ) is an acceptible point trans-
Thus we have formation, which can be used to describe phase space at any time t. Thus we
−1
∂ζ ∂ζ ∂ζ may also consider a composition, the point tranformation Pφ(t+∆t) Pφ(t) from
ζ̇ = M · η̇ + = M · J · ∇η H + = M · J · M T · ∇ζ H + the polar coordinate system at time t to that at t + ∆t. This is an infinites-
∂t ∂t ∂t
= J · ∇ζ K. imal point transformation, and is what we used in discussing the angular
velocity in rigid body motion.
Let us first consider a canonical transformation which does not depend Just as for orthogonal transformations, symplectic transformations can be
on time, so ∂ζ/∂t|η = 0. We see that we can choose the new Hamiltonian to divided into those which can be generated by infinitesimal transformations
be the same as the old, K = H (i.e. K(ζ, t) := H(η(ζ), t)), and get correct (which are connected to the identity) and those which can not. Consider a
mechanics, provided transformation M which is almost the identity, Mij = δij + Gij , or M =
1I + G, where  is considered some infinitesimal parameter while G is a finite
M · J · M T = J. (6.4) matrix. As M is symplectic, (1 + G) · J · (1 + GT ) = J, which tells us that
to lowest order in , GJ + JGT = 0. Comparing this to the condition for
We will require this condition even when ζ does depend on t, but then we the generator of an infinitesimal rotation, Ω = −ΩT , we see that it is similar
need to revisit the question of finding K. except for the appearence of J on opposite sides, changing orthogonality to
The condition (6.4) on M is similar to, and a generalization of, the condi- symplecticity. The new variables under such a canonical transformation are
tion for orthogonality of a matrix, OOT = 1I, which is of the same form with ζ = η + G · η.
J replaced by 1I. Another example of this kind of relation in physics occurs The condition (6.4) for a transformation η → ζ to be canonical does not
in special relativity, where a Lorentz transformation Lµν gives the relation involve time — each canonical transformation is a fixed map of phase-space
P
between two coordinates, x0µ = ν Lµν xν , with xν a four dimensional vector onto itself, and could be used at any t. We might consider a set of such
with x4 = ct. Then the condition which makes L a Lorentz transformation maps, one for each time, giving a time dependant map g(t) : η → ζ. Each
is such map could be used to transform the trajectory of the system at any
  time. In particular, consider the set of maps g(t, t0 ) which maps each point
1 0 0 0
0 0 0  η at which a system can be at time t0 into the point to which it will evolve
 1 
L · g · LT = g, with g =  . at time t. That is, g(t, t0 ) : η(t0 ) 7→ η(t). If we consider t = t0 + ∆t for
0 0 1 0 
0 0 0 −1 infinitesimal ∆t, this is an infinitesimal transformation. As ζi = ηi + ∆tη̇i =
P P
ηi + ∆t k Jik ∂H/∂ηk , we have Mij = ∂ζi /∂ηj = δij + ∆t k Jik ∂ 2 H/∂ηj ∂ηk ,
P
The matrix g in relativity is known as the indefinite metric, and the condition so Gij = k Jik ∂ 2 H/∂ηj ∂ηk ,
on L is known as pseudo-orthogonality. In our current discussion, however, !
J is not a metric, as it is antisymmetric rather than symmetric, and the word T
X ∂2H ∂2H
(GJ + JG )ij = Jik J`j + Ji` Jjk
which describes M is symplectic. k` ∂η` ∂ηk ∂η` ∂ηk
So far the only examples of canonical transformation which we have dis- X ∂2H
cussed have been point transformations. Before we continue, let us recall that = (Jik J`j + Ji` Jjk ) .
k` ∂η` ∂ηk
such transformations can be discrete, e.g. (x, y, z) → (r, θ, φ), but they can
6.4. POISSON BRACKETS 165 166 CHAPTER 6. HAMILTON’S EQUATIONS

The factor in parentheses in the last line is (−Jik Jj` + Ji` Jjk ) which is anti- The Poisson bracket is a fundamental property of the phase space. In
symmetric under k ↔ `, and as it is contracted into the second derivative, symplectic language,
which is symmetric under k ↔ `, we see that (GJ + JGT )ij = 0 and we
X ∂u ∂v
have an infinitesimal canonical transformation. Thus the infinitesimal flow [u, v] = Jij = (∇η u)T · J · ∇η v. (6.8)
of phase space points by the velocity function is canonical. As compositions ij ∂ηi ∂ηj
of canonical transformations are also canonical2 , the map g(t, t0 ) which takes
If we describe the system in terms of a different set of canonical variables
η(t0 ) into η(t), the point it will evolve into after a finite time increment t−t0 ,
ζ, we should still find the function f (t) changing at the same rate. We may
is also a canonical transformation.
think of u and v as functions of ζ as easily as of η. Really we are thinking
Notice that the relationship ensuring Hamilton’s equations exist,
of u and v as functions of points in phase space, represented by u(η) = ũ(ζ)
∂ζ and we may ask whether [ũ, ṽ]ζ is the same as [u, v]η . Using ∇η = M T · ∇ζ ,
M · J · M T · ∇ζ H + = J · ∇ζ K,
∂t we have
 T
with the symplectic condition M ·J·M T = J, implies ∇ζ (K−H) = −J·∂ζ/∂t, [u, v]η = M T · ∇ζ ũ · J · M T ∇ζ ṽ = (∇ζ ũ)T · M · J · M T ∇ζ ṽ
so K differs from H here. This discussion holds as long as M is symplectic,
even if it is not an infinitesimal transformation. = (∇ζ ũ)T · J∇ζ ṽ = [ũ, ṽ]ζ ,

so we see that the Poisson bracket is independent of the coordinatization

6.4 Poisson Brackets used to describe phase space, as long as it is canonical.
The Poisson bracket plays such an important role in classical mechanics,
Suppose I have some function f (q, p, t) on phase space and I want to ask how and an even more important role in quantum mechanics, that it is worthwhile
f , evaluated on a dynamical system, changes as the system evolves through to discuss some of its abstract properties. First of all, from the definition it
phase space with time. Then is obvious that it is antisymmetric:
X ∂f X ∂f
df
= q̇i + ṗi +
∂f [u, v] = −[v, u]. (6.9)
dt i ∂q i i ∂p i ∂t
X ∂f ∂H X ∂f ∂H ∂f It is a linear operator on each function over constant linear combinations,
= − + . (6.5) but is satisfies a Leibnitz rule for non-constant multiples,
i ∂qi ∂pi i ∂pi ∂qi ∂t
The structure of the first two terms is that of a Poisson bracket, a bilinear [uv, w] = [u, w]v + u[v, w], (6.10)
operation of functions on phase space defined by which follows immediately from the definition, using Leibnitz’ rule on the
X ∂u ∂v X ∂u ∂v partial derivatives. A very special relation is the Jacobi identity,
[u, v] := − . (6.6)
i ∂qi ∂pi i ∂pi ∂qi
[u, [v, w]] + [v, [w, u]] + [w, [u, v]] = 0. (6.11)
Thus Eq. 6.5 may be rewritten as
We need to prove that this is true. To simplify the presentation, we introduce
df ∂f some abbreviated notation. We use a subscript ,i to indicate partial derivative
= [f, H] + . (6.7)
dt ∂t with respect to ηi , so u,i means ∂u/∂ηi , and u,i,j means ∂(∂u/∂ηi )/∂ηj . We
2
If M = M1 · M2 and M1 · J · M1T = J, M2 · J · M2T = J, then M · J · M T = will assume all our functions on phase space are suitably differentiable, so
(M1 · M2 ) · J(·M2T · M1T ) = M1 · (M2 · J · M2T ) · M1T = M1 · J · M1T = J, so M is canonical. u,i,j = u,j,i . We will also use the summation convention, that any index
6.4. POISSON BRACKETS 167 168 CHAPTER 6. HAMILTON’S EQUATIONS

which appears twice in a term is assumed to be summed over3 . Then [v, w] = and we see that this combination of Poisson brackets involves the commutator
v,i Jij w,j , and of differential operators. But such a commutator is always a linear differential
operator itself,
[u, [v, w]] = [u, v,i Jij w,j ]
= [u, v,i ]Jij w,j + v,i Jij [u, w,j ]
X ∂ ∂ X ∂fj ∂ X ∂2
Dh Df = hi fj = hi + hi fj
ij ∂ηi ∂ηj ij ∂ηi ∂ηj ij ∂ηi ∂ηj
= u,k Jk` v,i,` Jij w,j + v,i Jij u,k Jk` w,j,` .
X ∂ ∂ X ∂hi ∂ X ∂2
In the Jacobi identity, there are two other terms like this, one with the Df Dh = fj hi = fj + hi fj
ij ∂ηj ∂ηi ij ∂ηj ∂ηi ij ∂ηi ∂ηj
substitution u → v → w → u and the other with u → w → v → u, giving
a sum of six terms. The only ones involving second derivatives of v are the
first term above and the one found from applying u → w → v → u to the so in the commutator, the second derivative terms cancel, and
second, u,i Jij w,k Jk` v,j,` . The indices are all dummy indices, summed over, so X ∂fj ∂ X ∂hi ∂
their names can be changed, by i → k → j → ` → i, converting this second Dh Df − Df Dh = hi − fj
ij ∂ηi ∂ηj ij ∂ηj ∂ηi
term to u,k Jk` w,j Jji v,`,i . Adding the original term u,k Jk` v,i,` Jij w,j , and using !
X ∂fj ∂hj ∂
v,`,i = v,i,` , gives u,k Jk` w,j (Jji +Jij )v,`,i = 0 because J is antisymmetric. Thus = hi − fi .
the terms in the Jacobi identity involving second derivatives of v vanish, but ij ∂ηi ∂ηi ∂ηj
the same argument applies in pairs to the other terms, involving second
derivatives of u or of w, so they all vanish, and the Jacobi identity is proven. This is just another first order differential operator, so there are no second
This argument can be made more elegantly if we recognize that for each derivatives of g left in (6.12). In fact, the identity tells us that this combina-
function f on phase space, we may view [f, ·] as a differential operator on tion is
functions g on phase space, mapping g → [f, g]. Calling this operator Df , Dh Df − Df Dh = D[h,f ] (6.13)
we see that
! An antisymmetric product which obeys the Jacobi identity is what makes
X X ∂f ∂
Df = Jij , a Lie algebra. Lie algebras are the infinitesimal generators of Lie groups,
j i ∂ηi ∂ηj or continuous groups, one example of which is the group of rotations SO(3)
which we have already considered. Notice that the “product” here is not as-
which is of the general form that a differential operator has,
sociative, [u, [v, w]] 6= [[u, v], w]. In fact, the difference [u, [v, w]] − [[u, v], w] =
X ∂ [u, [v, w]] + [w, [u, v]] = −[v, [w, u]] by the Jacobi identity, so the Jacobi iden-
Df = fj ,
j ∂ηj tity replaces the law of associativity in a Lie algebra. Lie groups play a
major role in quantum mechanics and quantum field theory, and their more
where fj are an arbitrary set of functions on phase space. For the Poisson extensive study is highly recommended for any physicist. Here we will only
bracket, the functions fj are linear combinations of the f,j , but fj 6= f,j . mention that infinitesimal rotations, represented either by the ω∆t or Ω∆t
With this interpretation, [f, g] = Df g, and [h, [f, g]] = Dh Df g. Thus of Chapter 4, constitute the three dimensional Lie algebra of the rotation
group (in three dimensions).
[h, [f, g]] + [f, [g, h]] = [h, [f, g]] − [f, [h, g]] = Dh Df g − Df Dh g Recall that the rate at which a function on phase space, evaluated on the
= (Dh Df − Df Dh )g, (6.12) system as it evolves, changes with time is
3
This convention of understood summation was invented by Einstein, who called it the df ∂f
“greatest contribution of my life”. = −[H, f ] + , (6.14)
dt ∂t
6.4. POISSON BRACKETS 169 170 CHAPTER 6. HAMILTON’S EQUATIONS

where H is the Hamiltonian. The function [f, g] on phase space also evolves 6.5 Higher Differential Forms
that way, of course, so
In section 6.1 we discussed a reinterpretation of the differential df as an
d[f, g] ∂[f, g] example of a more general differential 1-form, a map ω : M × Rn → R. We
= −[H, [f, g]] + saw that the {dxi } provide a basis for these forms, so the general 1-form can
dt ∂t " # " # P
∂f ∂g be written as ω = i ωi (x) dxi . The differential df gave an example. We
= [f, [g, H]] + [g, [H, f ]] + , g + f, defined an exact 1-form as one which is a differential of some well-defined
∂t ∂t P
" !# " !# function f . What is the condition for a 1-form to be exact? If ω = ωi dxi
∂g ∂f
= f, −[H, g] + + g, [H, f ] − is df , then ωi = ∂f /∂xi = f,i , and
∂t ∂t
" # " #
dg df ∂ωi ∂2f ∂2f
= f, − g, . ωi,j =
∂xj
=
∂xi ∂xj
=
∂xj ∂xi
= ωj,i .
dt dt
Thus one necessary condition for ω to be exact is that the combination ωj,i −
If f and g are conserved quantities, df /dt = dg/dt = 0, and we have the ωi,j = 0. We will define a 2-form to be the set of these objects which must
important consequence that d[f, g]/dt = 0. This proves Poisson’s theorem: vanish. In fact, we define a differential k-form to be a map
The Poisson bracket of two conserved quantities is a conserved quantity.
We will now show an important theorem, known as Liouville’s theo- ω (k) : M × |Rn × ·{z
· · × Rn} → R
rem, that the volume of a region of phase space is invariant under canonical k times
transformations. This is not a volume in ordinary space, but a 2n dimen-
Q which is linear in its action on each of the Rn and totally antisymmetric in
sional volume, given by integrating the volume element 2n i=1 dηi in the old
coordinates, and by its action on the k copies, and is a smooth function of x ∈ M. At a given
point, a basis of the k-forms is4
2n
Y ∂ζi Y2n Y2n X
dζi = det dηi = |det M | dηi dxi1 ∧ dxi2 ∧ · · · ∧ dxik := (−1)P dxiP 1 ⊗ dxiP 2 ⊗ · · · ⊗ dxiP k .
i=1 ∂ηj i=1 i=1 P ∈Sk

in the new, where we have used the fact that the change of variables requires For example, in three dimensions there are three independent 2-forms at a
a Jacobian in the volume element. But because J = M · J · M T , det J = point, dx1 ∧dx2 , dx1 ∧dx3 , and dx2 ∧dx3 , where dx1 ∧dx2 = dx1 ⊗dx2 −dx2 ⊗
det M det J det M T = (det M )2 det J, and J is nonsingular, so det M = ±1, dx1 , which means that, acting on ~u and ~v , dx1 ∧ dx2 (~u, ~v ) = u1 v2 − u2 v1 . The
and the volume element is unchanged. product ∧ is called the wedge product or exterior product, and can be
extended to act between k1 - and k2 -forms so that it becomes an associative
In statistical mechanics, we generally do not know the actual state of a
distributive product. Note that this definition of a k-form agrees, for k = 1,
system, but know something about the probability that the system is in a
particular region of phase space. As the transformation which maps possible 4
Some explanation of the mathematical symbols might be in order here. Sk is the group
values of η(t1 ) to the values into which they will evolve at time t2 is a canon- of permutations on k objects, and (−1)P is the sign of the permutation P , which is plus
ical transformation, this means that the volume of a region in phase space or minus one if the permutation can be built from an even or an odd number, respectively,
of transpositions of two of the elements. The tensor product ⊗ of two linear operators into
does not change with time, although the region itself changes. Thus the prob- a field is a linear operator which acts on the product space, or in other words a bilinear
n n
ability density, specifying the likelihood that the system is near a particular operator with two arguments. Here dxi ⊗ dxj is an operator on R × R which maps the
point of phase space, is invariant as we move along with the system. pair of vectors (~u, ~v ) to ui vj .
6.5. HIGHER DIFFERENTIAL FORMS 171 172 CHAPTER 6. HAMILTON’S EQUATIONS

with our previous definition, and for k = 0 tells us a 0-form is simply a k = 3: There is only one basis 3-form available in three dimensions, dx1 ∧
function on M. The general expression for a k-form is dx2 ∧ dx3 . Any other 3-form is proportional to this one, though the
X proportionality can be a function of {xi }. In particular dxi ∧dxj ∧dxk =
ω (k) = ωi1 ...ik (x)dxi1 ∧ · · · ∧ dxik . ijk dx1 ∧ dx2 ∧ dx3 . The most general 3-form C is simply specified by
i1 <...<ik an ordinary function C(x), which multiplies dx1 ∧ dx2 ∧ dx3 .

Let us consider some examples in three dimensional Euclidean5 space E 3 , Having established, in three dimensions, a correspondance between vec-
where there is a correspondance we can make between vectors and 1- and tors and 1- and 2-forms, and between functions and 0- and 3-forms, we can
2-forms. In this discussion we will not be considering how the objects change ask to what the wedge product corresponds in terms of these vectors. If
under changes in the coordinates of E 3 , to which we will return later. ~ and C
A ~ are two vectors corresponding to the 1-forms A = P Ai dxi and
P
C = Ci dxi , and if B = A ∧ C, then
k = 0: As always, 0-forms are simply functions, f (x), x ∈ E 3 . X X X
B= Ai Cj dxi ∧ dxj = (Ai Cj − Aj Ci )dxi ⊗ dxj = Bij dxi ⊗ dxj ,
P ij ij ij
k = 1: A 1-form ω = ωi dxi can be thought of, or associated with, a
~ P
vector field A(x) = ωi (x)êi . Note that if ω = df , ωi = ∂f /∂xi , so Bij = Ai Cj − Aj Ci , and
so A = ∇f . The 1-form is not actually the vector field, as it is a
~ ~
1X 1X 1X X
function that depends on two arguments, x and ~v . If the 1-form ω is Bk = kij Bij = kij Ai Cj − kij Aj Ci = kij Ai Cj ,
associated with the vector field A ~ = P Ai (x)êi , acting on the vector 2 2 2
~ = P Bi (x)êi , we have ω(x, B)
field B ~ = P Ai Bi = A ~ · B.
~ so
B ~ × C,
~ =A ~
k = 2: A general two form is a sum over the three independent wedge prod- and the wedge product of two 1-forms is the cross product of their vectors.
ucts with independent functions B12 (x), B13 (x), B23 (x). Let us extend If A is a 1-form and B is a 2-form, the wedge product C = A ∧ B =
the definition of Bij to make it an antisymmetric matrix, so C(x)dx1 ∧ dx2 ∧ dx3 is given by
X X XX
B= Bij dxi ∧ dxj = Bij dxi ⊗ dxj . C = A∧B = Ai Bjk dxi ∧ dxj ∧ dxk
i j<k
|{z} | {z }
i<j i,j
jk` B` ijk dx1 ∧ dx2 ∧ dx3
X X
As we did for the angular velocity matrix Ω in (4.2), we can condense = Ai B` jk` ijk dx1 ∧ dx2 ∧ dx3
i` j<k
| {z }
the information in the antisymmetric matrix Bij into a vector field symmetric under j ↔ k
~ = P Bi êi , with Bij = P ijk Bk . Note that this step requires that
B 1X X X
we are working in E 3 rather than some other dimension. Thus B = = Ai B` jk` ijk dx1 ∧ dx2 ∧ dx3 = Ai B` δi` dx1 ∧ dx2 ∧ dx3
P 1 P 1 P 2 i` jk i`
ijk ijk Bk dxi ⊗ dxj . Also note 2 ij ijk Bij = 2 ij` ijk ij` B` = Bk .
The 2-form B takes two vector arguments, so acting on vectors A ~ = ~·B
= A ~ dx1 ∧ dx2 ∧ dx3 ,
P ~ P ~ ~ P
Ai (x)êi and C = Ci (x)êi , we have B(x, A, B) = Bij Ai Cj =
P so we see that the wedge product of a 1-form and a 2-form gives the dot
~
ijk Bk Ai Cj = (A(x) × C(x))
~ · B(x).
~
product of their vectors.
5
Forms are especially useful in discussing more general manifolds, such as occur in If A and B are both 2-forms, the wedge product C = A ∧ B must be a
general relativity. Then one must distinguish between covariant and contravariant vectors, 4-form, but there cannot be an antisymmetric function of four dxi ’s in three
a complication we avoid here by treating only Euclidean space. dimensions, so C = 0.
6.5. HIGHER DIFFERENTIAL FORMS 173 174 CHAPTER 6. HAMILTON’S EQUATIONS

The exterior derivative We can summarize the action of the exterior derivative in three dimensions
in this diagram:
We defined the differential of a function f , which we now call a 0-form, giving
P
a 1-form df = f,i dxi . Now we want to generalize the notion of differential d d d d
so that d can act on k-forms for arbitrary k. This generalized differential f - ω (1) ∼ A
~ - ω (2) ∼ B
~ - ω (3) - 0
∇f ∇×A ∇·B
d : k-forms → (k + 1)-forms
Now that we have d operating on all k-forms, we can ask what happens
is called the exterior derivative. It is defined to be linear and to act on if we apply it twice. Looking first in three dimenions, on a 0-form we get
one term in the sum over basis elements by d2 f = dA for A ~ ∼ ∇f , and dA ∼ ∇ × A, so d2 f ∼ ∇ × ∇f . But the curl of a
gradient is zero, so d2 = 0 in this case. On a one form d2 A = dB, B
~ ∼ ∇×A ~
d (fi1 ...ik (x)dxi1 ∧ · · · ∧ dxik ) = (dfi1 ...ik (x)) ∧ dxi1 ∧ · · · ∧ dxik and dB ∼ ∇ · B = ∇ · (∇ × A). Now we have the divergence of a curl,
~
X
= fi1 ...ik ,j dxj ∧ dxi1 ∧ · · · ∧ dxik . which is also zero. For higher forms in three dimensions we can only get zero
j because the degree of the form would be greater than three. Thus we have
a strong hint that d2 might vanish in general. To verify this, we apply d2 to
P
Clearly some examples are called for, so let us look again at three dimen- ω (k) = ωi1 ...ik dxi1 ∧ · · · ∧ dxik . Then
sional Euclidean space. X X
P dω = (∂j ωi1 ...ik ) dxj ∧ dxi1 ∧ · · · ∧ dxik
k = 0: For a 0-form f , df = f,i dxi , as we defined earlier. In terms of j i1 <i2 <···<ik
X X
vectors, df ∼ ∇f
~ . d(dω) = ( ∂` ∂j ωi1 ...ik ) dx` ∧ dxj ∧dxi1 ∧ · · · ∧ dxik
`j i1 <i2 <···<ik
| {z } | {z }
P P P symmetric antisymmetric
k = 1: For a 1-form ω = ωi dxi , dω = i dωi ∧ dxi = ij ωi,j dxj ∧ dxi =
P = 0.
ij (ωj,i − ωi,j ) dxi ⊗dxj , corresponding to a two form with Bij = ωj,i −
ωi,j . These Bij are exactly the things which must vanish if ω is to be This is a very important result. A k-form which is the exterior derivative of
exact. In three dimensional Euclidean space, we have a vector B ~ with some (k − 1)-form is called exact, while a k-form whose exterior derivative
1 P P
components Bk = 2 kij (ωj,i − ωi,j ) = kij ∂i ωj = (∇ × ω ~ ~ )k , so here vanishes is called closed, and we have just proven that all exact k-forms are
the exterior derivative of a 1-form gives a curl, B = ∇ × ω
~ ~ ~. closed.
P The converse is a more subtle question. In general, there are k-forms
k = 2: On a two form B = i<j Bij dxi ∧ dxj , the exterior derivative gives which are closed but not exact, given by harmonic functions on the manifold
P P
a 3-form C = dB = k i<j Bij,k dxk ∧ dxi ∧ dxj . In three-dimensional M, which form what is known as the cohomology of M. This has to do with
Euclidean space, this reduces to global properties of the space, however, and locally every closed form can be
XX X written as an exact one.6 The precisely stated theorem, known as Poincaré’s
C= (∂k ij` B` ) kij dx1 ∧ dx2 ∧ dx3 = ∂k Bk dx1 ∧ dx2 ∧ dx3 ,
6
k` i<j k An example may be useful. In two dimensions, irrotational vortex flow can be repre-
sented by the 1-form ω = −yr−2 dx + xr−2 dy, which satisfies dω = 0 wherever it is well
defined, but it is not well defined at the origin. Locally, we can write ω = dθ, where θ is
so C(x) = ∇ ~ · B,
~ and the exterior derivative on a 2-form gives the
the polar coordinate. But θ is not, strictly speaking, a function on the plane, even on the
divergence of the corresponding vector. plane with the origin removed, because it is not single-valued. It is a well defined function
on the plane with a half axis removed, which leaves a simply-connected region, a region
k = 3: If C is a 3-form, dC is a 4-form. In three dimensions there cannot with no holes. In fact, this is the general condition for the exactness of a 1-form — a
be any 4-forms, so dC = 0 for all such forms. closed 1-form on a simply connected manifold is exact.
6.5. HIGHER DIFFERENTIAL FORMS 175 176 CHAPTER 6. HAMILTON’S EQUATIONS

Lemma, is that if ω is a closed k-form on a coordinate neighborhood U of a so

manifold M , and if U is contractible to a point, then ω is exact on U . We will
X ∂ f˜ X ∂f ∂xi ∂yk
ignore the possibility of global obstructions and assume that we can write df˜ = dyk = dxj
closed k-forms in terms of an exterior derivative acting on a (k − 1)-form. k ∂yk ijk ∂xi ∂yk ∂xj
X ∂f X
= δij dxj = f,i dxi = df.
Coordinate independence of k-forms ij ∂xi i

We have introduced forms in a way which makes them appear dependent We impose this transformation law in general on the coefficients in our k-
on the coordinates xi used to describe the space M. This is not what we forms, to make the k-form invariant, which means that the coefficients are
want at all7 . We want to be able to describe physical quantities that have covariant,
intrinsic meaning independent of a coordinate system. If we are presented
with another set of coordinates yj describing the same physical space, the X ∂xi
ω̃j = ωi
points in this space set up a mapping, ideally an isomorphism, from one i ∂yj
coordinate system to the other, ~y = ~y (~x). If a function represents a physical k
!
X Y ∂xi`
field independent of coordinates, the actual function f (x) used with the x ω̃j1 ...jk = ωi1 ...ik .
coordinates must be replaced by another function f˜(y) when using the y i1 ,i2 ,...,ik `=1 ∂yjl

coordinates. That they both describe the physical value at a given physical
point requires f (x) = f˜(y) when y = y(x), or more precisely8 f (x) = f˜(y(x)). Integration of k-forms
This associated function and coordinate system is called a scalar field. Suppose we have a k-dimensional smooth “surface” S in M, parameterized
If we think of the differential df as the change in f corresponding to by coordinates (u1 , · · · , uk ). We define the integral of a k-form
an infinitesimal change dx, then clearly df˜ is the same thing in different X
coordinates, provided we understand the dyi to represent the same physical ω (k) = ωi1 ...ik dxi1 ∧ · · · ∧ dxik
displacement as dx does. That means i1 <...<ik

X ∂yk over S by
dyk = dxj . !
j ∂xj Z Z X k
Y
(k) ∂xi`
ω = ωi1 ...ik (x(u)) du1 du2 · · · duk .
As f (x) = f˜(y(x)) and f˜(y) = f (x(y)), the chain rule gives
S i1 ,i2 ,...,ik `=1 ∂u`

X ∂ f˜ ∂yj We had better give some examples. For k = 1, the “surface” is actually
∂f ∂ f˜ X ∂f ∂xi
= , = , a path Γ : u 7→ x(u), and
∂xi j ∂yj ∂xi ∂yj i ∂xi ∂yj
Z X Z umax X
∂xi
7 ωi dxi = ωi (x(u)) du,
Indeed, most mathematical texts will first define an abstract notion of a vector in
the tangent space as a directional derivative operator, specified by equivalence classes of
Γ umin ∂u
parameterized paths on M. Then 1-forms are defined as duals to these vectors. In the R
~ · d~r, the path integral of
which seems obvious. In vector notation this is Γ A
first step any coordinatization of M is tied to the corresponding basis of the vector space
Rn . While this provides an elegant coordinate-independent way of defining the forms, the ~
the vector A.
abstract nature of this definition of vectors can be unsettling to a physicist. For k = 2,
8
More elegantly, giving the map x → y the name φ, so y = φ(x), we can state the Z Z
∂xi ∂xj
relation as f = f˜ ◦ φ. ω (2) = Bij dudv.
S ∂u ∂v
6.5. HIGHER DIFFERENTIAL FORMS 177 178 CHAPTER 6. HAMILTON’S EQUATIONS

In three dimensions, the parallelogram of M, with ∂C its boundary. Let ω be a (k − 1)-form. Then Stokes’ theorem
which is the image of the rectangle [u, u + says
du] × [v, v + dv] has edges (∂~x/∂u)du and v Z Z
(∂~x/∂v)dv, which has an area equal to the dω = ω. (6.15)
magnitude of C ∂C
u
! This elegant jewel is actually familiar in several contexts in three dimen-
∂~x ∂~x
~ =
“dS” × dudv sions. If k = 2, C is a surface, usually called S, Rbounded by a closed path
∂u ∂v Γ = ∂S. If ω is a 1-form associated with ~ then ω = R A
A, ~ · d~`. Now dω is
  Γ Γ
~
and a normal in the direction of “dS”. Writing Bij in terms of the corre- the 2-form ∼ ∇ ~ and R dω = R ∇
~ × A, ~ ×A ~ · dS,
~ so we see that this Stokes’
S S
~ Bij = ijk Bk , so
sponding vector B, theorem includes the one we first learned by that name. But it also includes
other possibilities. We can try k = 3, where
R
C =R V is a volume with surface
Z Z ! !
(2) ∂~x ∂~x S = ∂V . Then if ω ∼ B is a two form, S ω = S B
~ ~ · dS,
~ while dω ∼ ∇ ~ · B,
~
ω = ijk Bk dudv R R
S S ∂u ∂v so V dω = ∇ ~ · BdV
~ , so here Stokes’ general theorem gives Gauss’s theo-
i j
Z ! Z rem. Finally, we could consider k = 1, C = Γ, which has a boundary ∂C
∂~x ∂~x
= Bk × dudv = ~ · dS,
B ~ consisting of two points,Rsay A and B. Our 0-form ω = f is a function, and
S ∂u ∂v k S
Stokes’ theorem gives10 Γ df = f (B) − f (A), the “fundamental theorem of
R calculus”.
so ω (2) gives the flux of B
~ through the surface.
Similarly for k = 3 in three dimensions,
X
! ! ! 6.6 The natural symplectic 2-form
∂~x ∂~x ∂~x
ijk dudvdw
∂u i
∂v j
∂w k
We now turn our attention back to phase space, with a set of canonical
coordinates (qi , pi ). Using these coordinates we can define a particular 1-
P
is the volume of the parallelopiped which is the image of [u, u + du] × [v, v + form ω1 = i pi dqi . For a point transformation Qi = Qi (q1 , . . . , qn , t) we
dv] × [w, w + dw]. As ωijk = ω123 ijk , this is exactly what appears: may use the same Lagrangian, reexpressed in the new variables, of course.
Z Z X Z Here the Qi are independent of the velocities q̇j , so on phase space11 dQi =
∂xi ∂xj ∂xk P
ω (3) = ijk ω123 dudvdw = ω123 (x)dV. j (∂Qi /∂qj )dqj . The new velocities are given by
∂u ∂v ∂w
X ∂Qi ∂Qi ∂ Q̇i ∂Qi
Notice that we have only defined the integration of k-forms over subman- Q̇i = q̇j + , so = .
ifolds of dimension k, not over other-dimensional submanifolds. These are j ∂qj ∂t ∂ q̇j ∂qj
the only integrals which have coordinate invariant meanings. Also note that 10
Note that there is a direction associated with the boundary, which is induced by a
the integrals do not depend on how the surface is coordinatized. direction associated with C itself. This gives an ambiguity in what we have stated, for
We state9 a marvelous theorem, special cases of which you have seen often example how the direction of an open surface induces a direction on Rthe closed loop which
bounds it. Changing this direction would clearly reverse the sign of A ~ · d~`. We have not
before, known as Stokes’ Theorem. Let C be a k-dimensional submanifold
worried about this ambiguity, but we cannot avoid noticing the appearence of the sign in
9 this last example.
For a proof and for a more precise explanation of its meaning, we refer the reader to
the mathematical literature. In particular [14] and [3] are advanced calculus texts which 11
We have not included a term ∂Q i
∂t dt which would be necessary if we were considering
give elementary discussions in Euclidean 3-dimensional space. A more general treatment a form in the 2n + 1 dimensional extended phase space which includes time as one of its
is (possibly???) given in [16]. coordinates.
6.6. THE NATURAL SYMPLECTIC 2-FORM 179 180 CHAPTER 6. HAMILTON’S EQUATIONS

Thus the old canonical momenta, also know M is invertible and that J 2 = −1, so if we multiply this known
equation from the left by −J · M −1 and from the right by J · M , we learn
∂L(q, q̇, t) X ∂L(Q, Q̇, t) ∂ Q̇j X ∂Qj
pi = = = Pj . that
∂ q̇i q,t j ∂ Q̇j q,t
∂ q̇i q,t j ∂qi
−J · M −1 · M · J · M T · J · M = −J · M −1 · J · J · M
Thus the form ω1 may be written = J · M −1 · M = J
XX ∂Qj X = −J · J · M · J · M = M T · J · M,
T
ω1 = Pj dqi = Pj dQj ,
i j ∂qi j
which is what we wanted to prove. Thus we have shown that the 2-form ω2
so the form of ω1 is invariant under point transformations. This is too lim- is form-invariant under canonical transformations, and deserves its name.
ited, however, for our current goals of considering general canonical transfor- One important property of the 2-form ω2 on phase space is that it is
mations on phase space, under which ω1 will not be invariant. However, its non-degenerate. A 2-form has two slots to insert vectors — inserting one
exterior derivative X
leaves a 1-form. Non-degenerate means there is no non-zero vector ~v on phase
ω2 := dω1 = dpi ∧ dqi space such that ω2 (·, ~v ) = 0, that is, such that ω2 (~u, ~v ) = 0, for all ~u on phase
i space. This follows simply from the fact that the matrix Jij is non-singular.
is invariant under all canonical transformations, as we shall show momen-
tarily. This makes it special, the natural symplectic structure on phase Extended phase space
space. We can reexpress ω2 in terms of our combined coordinate notation ηi ,
because One way of looking at the evolution of a system is in phase space, where
X X X a given system corresponds to a point moving with time, and the general
− Jij dηi ∧ dηj = − dqi ∧ dpi = dpi ∧ dqi = ω2 . equations of motion corresponds to a velocity field. Another way is to con-
i<j i i
sider extended phase space, a 2n + 1 dimensional space with coordinates
We must now show that the natural symplectic structure is indeed form (qi , pi , t), for which a system’s motion is a path, monotone in t. By the modi-
invariant under canonical transformation. Thus if Qi , Pi are a new set of fied Hamilton’sRprinciple, the path of a system in this space is an extremum of
t P
canonical coordinates, combined into ζj , we expect the corresponding object the action I = tif pi dqi − H(q, p, t)dt, which is the integral of the one-form
P
formed from them, ω20 = − ij Jij dζi ⊗ dζj , to reduce to the same 2-form, ω2 . X
We first note that ω3 = pi dqi − H(q, p, t)dt.
X ∂ζi X
dζi = dηj = Mij dηj , The exterior derivative of this form involves the symplectic structure, ω2 ,
j ∂ηj j
as dω3 = ω2 − dH ∧ dt. The 2-form ω2 on phase space is nondegenerate,
with the same Jacobian matrix M we met in section 6.3. Thus and every vector in phase space is also in extended phase space. On such a
X X X X vector, on which dt gives zero, the extra term gives only something in the dt
ω20 = − Jij dζi ⊗ dζj = − Jij Mik dηk ⊗ Mj` dη`
ij ij k `
direction, so there are still no vectors in this subspace which are annihilated
X  by dω3 . Thus there is at most one direction in extended phase space which
= − MT · J · M dηk ⊗ dη` .
k` is annihilated by dω3 . But any 2-form in an odd number of dimensions
k`
must annihilate some vector, because in a given basis it corresponds to an
Things will work out if we can show M T · J · M = J, whereas what we know antisymmetric matrix Bij , and in an odd number of dimensions det B =
for canonical transformations from Eq. (6.4) is that M · J · M T = J. We det B T = det(−B) = (−1)2n+1 det B = − det B, so det B = 0 and the matrix
6.6. THE NATURAL SYMPLECTIC 2-FORM 181 182 CHAPTER 6. HAMILTON’S EQUATIONS

is singular, annihilating some vector ξ. In fact, for dω3 this annihilated vector 0. In addition, dω3 (~v , ~v ) = 0, of course, so dω3 (·, ~v ) vanishes on a complete
ξ is the tangent to the path the system takes through extended phase space. basis of vectors and is therefore zero.
One way to see this is to simply work out what dω3 is and apply it to
the vector ξ, which is proportional to ~v = (q̇i , ṗi , 1). So we wish to show 6.6.1 Generating Functions
dω3 (·, ~v ) = 0. Evaluating
Consider a canonical transformation (q, p) → (Q, P ), and the two 1-forms
X X X X X P P
dpi ∧ dqi (·, ~v ) = dpi dqi (~v ) − dqi dpi (~v ) = dpi q̇i − dqi ṗi ω1 = i pi dqi and ω10 = i Pi dQi . We have mentioned that the difference of
dH ∧ dt(·, ~v ) = dH dt(~v ) − dt dH(~v ) these will not vanish in general, but the exterior derivative of this difference,
X ∂H X ∂H
! d(ω1 − ω10 ) = ω2 − ω20 = 0, so ω1 − ω10 is an closed 1-form. Thus it is exact13 ,
∂H
= dqi + dpi + dt 1 and there must be a function F on phase space such that ω1 − ω10 = dF .
∂qi ∂pi ∂t We call F the generating function of the canonical transformation14
!
X ∂H X ∂H ∂H If the transformation (q, p) → (Q, P ) is such that the old q’s alone, without
−dt q̇i + ṗi +
∂qi ∂pi ∂t information about the old p’s, do not impose any restrictions on the new Q’s,
!
X ∂H X ∂H X ∂H ∂H then the dq and dQ are independent, and we can use q and Q to parameter-
= dqi + dpi − dt q̇i + ṗi ize phase space15 . Then knowledge of the function F (q, Q) determines the
∂qi ∂pi ∂qi ∂pi
! ! transformation, as
X ∂H ∂H
dω3 (·, ~v ) = q̇i − dpi − ṗi + dqi  
∂pi ∂qi X X ∂F ∂F
! ω1 − ω10 = (pi dqi − Pi dQi ) = dF =  dqi + dQi 
X ∂H ∂H ∂qi ∂Qi
+ q̇i + ṗi dt i i Q q
∂qi ∂pi
∂F ∂F
= 0 =⇒ pi = , −Pi = .
∂qi Q
∂Qi q
where the vanishing is due to the Hamilton equations of motion.
There is a more abstract way of understanding why dω3 (·, ~v ) If the canonical transformation depends on time, the function F will
vanishes, from the modified Hamilton’s principle, which states also depend on time. Now if we consider the motion in extended phase
δη v dt space, we know the phase trajectory that the system takes through extended
that if the path taken were infinitesimally varied from the physi-
cal path, there would be no change in the action. But this change phase space is determined by Hamilton’s equations, which could be written
is the integral of ω3 along a loop, forwards in time along the first in any set of canonical coordinates, so in particular there is some Hamiltonian
trajectory and backwards along the second. From Stokes’ the- K(Q, P, t) such that the tangent to the phase trajectory, ~v , is annihilated by
P
orem this means the integral of dω3 over a surface connecting dω30 , where ω30 = Pi dQi − K(Q, P, t)dt. Now in general knowing that two
these two paths vanishes. But this surface is a sum over infinitesimal parallel- 2-forms both annihilate the same vector would not be sufficient to identify
ograms one side of which is ~v ∆t and the other side of which12 is (δ~q(t), δ~p(t), 0). them, but in this case we also know that restricting dω3 and dω30 to their
As this latter vector is an arbitrary function of t, each parallelogram must in- action on the dt = 0 subspace gives the same 2-form ω2 . That is to say, if
dependently give 0, so that its contribution to the integral, dω3 ((δ~q, δ~p, 0), ~v )∆t = 13
We are assuming phase space is simply connected, or else we are ignoring any compli-
cations which might ensue from F not being globally well defined.
12 14
It is slightly more elegant to consider the path parameterized independently of time, This is not an infinitesimal generator in the sense we have in Lie algebras — this
and consider arbitrary variations (δq, δp, δt), because the integral involved in the action, generates a finite canonical transformation for finite F .
15
being the integral of a 1-form, is independent of the parameterization. With this approach Note that this is the opposite extreme from a point transformation, which is a canonical
we find immediately that dω3 (·, ~v ) vanishes on all vectors. transformation for which the Q’s depend only on the q’s, independent of the p’s.
6.6. THE NATURAL SYMPLECTIC 2-FORM 183 184 CHAPTER 6. HAMILTON’S EQUATIONS

~u and ~u 0 are two vectors with time components zero, we know that (dω3 − in momenta to allow [Qi , Pj ] = δij to remain unchanged. This also doesn’t
dω30 )(~u,~u 0 ) = 0. Any vector can be expressed as a multiple of ~v and some change H. For λ = 1, this is the identity transformation, for which F = 0,
vector ~u with time component zero, and as both dω3 and dω30 annihilate ~v , of course.
we see that dω3 − dω30 vanishes on all pairs of vectors, and is therefore zero. Placing point transformations in this language provides another example.
Thus ω3 − ω30 is a closed 1-form, which must be at least locally exact, and For a point transformation, Qi = fi (q1 , . . . , qn , t), which is what one gets with
indeed ω3 − ω30 = dF , where F is the generating function we found above16 . a generating function
P P
Thus dF = pdq − P dQ + (K − H)dt, or X
F2 = fi (q1 , . . . , qn , t)Pi .
i
∂F
K=H+ . Note that
∂t ∂F2 X ∂fj
pi = = Pj
The function F (q, Q, t) is what Goldstein calls F1 . The existence of F ∂qi j ∂qi
as a function on extended phase space holds even if the Q and q are not is at any point ~q a linear transformation of the momenta, required to preserve
independent, but in this case F will need to be expressed as a function of the canonical Poisson bracket, but this transformation is ~q dependent, so
other coordinates. Suppose the new P ’s and the old q’s are independent, so ~ is a function of ~q and t only, independent of p~, P~ (q, p, t) will in general
P while Q
we can write F (q, P, t). Then define F2 = Qi Pi + F . Then have a nontrivial dependence on coordinates as well as a linear dependence
X X X X on the old momenta.
dF2 = Qi dPi + Pi dQi + pi dqi − Pi dQi + (K − H)dt
X X For a harmonic oscillator, a simple scaling gives
= Qi dPi + pi dqi + (K − H)dt,
p2 k 1q  
so H= + q2 = k/m P 2 + Q2 ,
2m 2 2
∂F2 ∂F2 ∂F2
Qi = , pi = , K(Q, P, t) = H(q, p, t) + .
∂Pi ∂qi ∂t where Q = (km)1/4 q, P = (km)−1/4 p. In this form, thinking P
θ
The generating function can be a function of old momenta rather than the of phase space as just some two-dimensional space, we seem to
old coordinates. Making one choice for the old coordinates and one for the be encouraged to consider a second canonical transformation
Q
new, there are four kinds of generating functions as described by Goldstein. Q, P −→ θ, P, generated by F1 (Q, θ), to a new, polar, coordi-
P F1
Let us consider some examples. The function F1 = i qi Qi generates an nate system with θ = tan−1 Q/P as the new coordinate, and
interchange of p and q, we might hope to have the radial coordinate related to the new
momentum, P = − ∂F1 /∂θ|Q . As P = ∂F1 /∂Q|θ is also Q cot θ,
Qi = p i , Pi = −qi , we can take F1 = 12 Q2 cot θ, so
which leaves the Hamiltonian unchanged. We saw this clearly leaves the 1 1 1
P = − Q2 (− csc2 θ) = Q2 (1 + P 2 /Q2 ) = (Q2 + P 2 ) = H/ω.
form of Hamilton’s equations unchanged. An interesting generator of the 2 2 2
P
second type is F2 = i λi qi Pi , which gives Qi = λi qi , Pi = λ−1
i pi , a simple
Note as F1 is not time dependent, K = H and is independent of θ, which is
change in scale of the coordinates with a corresponding inverse scale change therefore an ignorable coordinate, so its conjugate momentum P is conserved.
16 q P differs from the conserved Hamiltonian H only by the factor
Of course
From its definition in that context, we found that in phase space, dF = ω1 − ω10 ,
which is the part of ω3 − ω30 not in the time direction. Thus if ω3 − ω30 = dF 0 for some ω = k/m, so this is not unexpected. With H now linear in the new
other function F 0 , we know dF 0 − dF = (K 0 − K)dt for some new Hamiltonian function momentum P, the conjugate coordinate θ grows linearly with time at the
K 0 (Q, P, t), so this corresponds to an ambiguity in K. fixed rate θ̇ = ∂H/∂P = ω.
6.6. THE NATURAL SYMPLECTIC 2-FORM 185 186 CHAPTER 6. HAMILTON’S EQUATIONS

Infinitesimal generators, redux through phase space rather than as a change in coordinates. More generally,
we may view a canonical transformation as a diffeomorphism18 of phase
Let us return to the infinitesimal canonical transformation
space onto itself, g : M → M with g(q, p) = (Q, P ).
ζi = ηi + gi (ηj ). For an infinitesimal canonical transformation, this active interpretation
gives us a small displacement δη = [η, G] for every point η in phase space,
Mij = ∂ζi /∂ηj = δij + ∂gi /∂ηj needs to be symplectic, and so Gij = ∂gi /∂ηj so we can view G and its associated infinitesimal canonical transformation
satisfies the appropriate condition for the generator of a symplectic matrix, as producing a flow on phase space. G also builds a finite transformation by
G · J = −J · GT . For the generator of the canonical transformation, we repeated application, so that we get a sequence of canonical transformations
need a perturbation of the generator for the identity transformation, which g λ parameterized by λ = n∆λ. This sequence maps an initial η0 into a se-
can’t be in F1 form (as (q, Q) are not independent), but is easily done in F2 quence of points g λ η0 , each generated from the previous one by the infinites-
P imal transformation ∆λG, so g λ+∆λ η0 − g λ η0 = ∆λ[g λ η0 , G]. In the limit
form, F2 (q, P ) = i qi Pi + G(q, P, t), with pi = ∂F2 /∂qi = Pi + ∂G/∂qi ,
Qi = ∂F2 /∂Pi = qi + ∂G/∂Pi , or ∆λ → 0, with n allowed to grow so that we consider a finite range of λ, we
have a one (continuous) parameter family of transformations g λ : M → M,
      
Qi qi 0 1I ∂G/∂qi satisfying the differential equation
ζ= = + = η + J · ∇G,
Pi pi −1I 0 ∂G/∂pi
dg λ (η) h λ i
where we have ignored higher order terms in  in inverting the q → Q relation = g η, G .
dλ
and in replacing ∂G/∂Pi with ∂G/∂pi .
The change due to the infinitesimal transformation may be written in This differential equation defines a phase flow on phase space. If G is not
terms of Poisson bracket with the coordinates themselves: a function of λ, this has the form of a differential equation solved by an
exponential,
δη = ζ − η = J · ∇G = [η, G]. g λ (η) = eλ[·,G] η,
which means
In the case of an infinitesimal transformation due to time evolution, the small 1
parameter can be taken to be ∆t, and δη = ∆t η̇ = ∆t[H, η], so we see that g λ (η) = η + λ[η, G] + λ2 [[η, G], G] + ....
2
the Hamiltonian acts as the generator of time translations, in the sense that
it maps the coordinate η of a system in phase space into the coordinates the In the case that the generating function is the Hamiltonian, G = H, this
system will have, due to its equations of motion, at a slightly later time. phase flow gives the evolution through time, λ is t, and the velocity field on
This last example encourages us to find another interpretation of canoni- phase space is given by [η, H]. If the Hamiltonian is time independent, the
cal transformations. Up to now we have taken a passive view of the transfor- velocity field is fixed, and the solution is formally an exponential.
mation, as a change of variables describing an unchanged physical situation, Let us review changes due to a generating function considered in the
just as the passive view of a rotation is to view it as a change in the descrip- passive and alternately in the active view. In the passive picture, we view η
tion of an unchanged physical point in terms of a rotated set of coordinates. and ζ = η+δη as alternative coordinatizations of the same physical point A in
But rotations are also used to describe changes in the physical situation with P
phase space. For an infinitesimal generator F2 = i qi Pi + G, δη = J∇G =
regards to a fixed coordinate system17 , and similarly in the case of motion [η, G]. A physical scalar defined by a function u(η) changes its functional
through phase space, it is natural to think of the canonical transformation
18
generated by the Hamiltonian as describing the actual motion of a system An isomorphism g : M → N is a 1-1 map with an image including all of N (onto),
which is therefore invertible to form g −1 : N → M. A diffeomorphism is an isomorphism
17
We leave to Mach and others the question of whether this distinction is real. g for which both g and g −1 are differentiable.
6.6. THE NATURAL SYMPLECTIC 2-FORM 187 188 CHAPTER 6. HAMILTON’S EQUATIONS

form to ũ, but not its value at a given physical point, so ũ(ζA ) = u(ηA ). For ζ 1 (A) η1 (B) = ζ1 (A)
the Hamiltonian, there is a change in value as well, for H̃ or K̃ may not be η2 η2
the same as H, even at the corresponding point,
ζ 2 (A) δη
A A η (A)
η2(A)
B η2 (B) = ζ (A)
∂F2 ∂G 2
K̃(ζA ) = H(ηA ) + = H(ηA ) +  . 2
∂t ∂t A η1 η1

η1(A) η1(A)
Now consider an active view. Here a canonical transformation is thought
of as moving the point in phase space, and at the same time changing the Passive view of the canonical trans- Active view of the transformation,
functions u → ũ, H → K̃, where we are focusing on the form of these func- formation. Point A is the same moving from point A in phase space
tions, on how they depend on their arguments. We think of ζ as representing point, whether expressed in coor- to point B. So ηB = ζA . The func-
the η coordinates of a different point B of phase space, although the coordi- dinates ηj or ζj , and scalar func- tion ũ then differs from u when eval-
nates ηB = ζA . We ask how ũ and K̃ differ from u and H at B, evaluated tions take the same value there, so uated on the same coordinates η.
at the same values of their arguments, not at what we considered the same u(ηA ) = ũ(ζA ).
physical point in the passive view. Let19 Note that if the generator of the transformation is a conserved quantity,
the Hamiltonian is unchanged, in that it is the same function after the trans-
formation as it was before. That is, the Hamiltonian is form invariant.
∂u So we see that conserved quantities are generators of symmetries of the
∆u = ũ(ηB ) − u(ηB ) = ũ(ζA ) − u(ζA ) = u(ηA ) − u(ζA ) = −δηi
∂ηi problem, transformations which can be made without changing the Hamil-
X ∂u tonian. We saw that the symmetry generators form a closed algebra under
= − [ηi , G] = −[u, G]
i ∂ηi Poisson bracket, and that finite symmetry transformations result from expo-
nentiating the generators. Let us discuss the more common conserved quan-
tities in detail, showing how they generate symmetries. We have already seen
that ignorable coordinates lead to conservation of the corresponding momen-
tum. Now the reverse comes if we assume one of the momenta, say pI , is
conserved. Then from our discussion we know that the generator G = pI
∆H = K̃(ηB ) − H(ηB ) = K̃(ζA ) − H(ηB ) will generate canonical transformations which are symmetries of the system.
!
∂G ∂G Those transformations are
= H(ηA ) +  − H(ηA ) − δη · ∇η H =  − [H, G]
∂t A ∂t
δqj = [qj , pI ] = δjI , δpj = [pj , pI ] = 0.
dG
=  .
dt Thus the transformation just changes the one coordinate qI and leaves all
the other coordinates and all momenta unchanged. In other words, it is a
translation of qI . As the Hamiltonian is unchanged, it must be independent
19
We differ by a sign from Goldstein in the definition of ∆u. of qI , and qI is an ignorable coordinate.
6.6. THE NATURAL SYMPLECTIC 2-FORM 189 190 CHAPTER 6. HAMILTON’S EQUATIONS

Second, consider the angular momentum component ω ~ ·L

~ = ijk ωi rj pk The use of the Levi-Civita ijk to write L as a vector is peculiar to three
for a point particle with q = ~r. As a generator, ~ω · L
~ produces changes dimensions; in other dimensions d 6= 3 there is no -symbol to make a vector
out of L, but the angular momentum can always be treated as an antisym-
δr` = [r` , ijk ωi rj pk ] = ijk ωi rj [r` , pk ] = ijk ωi rj δ`k = ij` ωi rj
metric tensor, Lij = xi pj − xj pi . There are D(D − 1)/2 components, and the
= (~ω × ~r)` , Lie algebra again closes
which is how the point moves under a rotation about the axis ω
~ . The mo-
mentum also changes, [Lij , Lk` ] = δjk Li` − δik Lj` − δj` Lik + δi` Ljk .

δp` = [p` , ijk ωi rj pk ] = ijk ωi pk [p` , rj ] = ijk ωi pk (−δ`j ) = −i`k ωi pk We have related conserved quantities to generators of infinitesimal canon-
= (~ω × p~)` , ical transformation, but these infinitesimals can be integrated to produce fi-
nite transformations as well. As we mentioned earlier, from an infinitesimal
so p~ also rotates in the same way. generator G we can exponentiate to form a one-parameter set of transforma-
By Poisson’s theorem, the set of constants of the motion is closed un- tions
der Poisson bracket, and given two such generators, the bracket is also a
symmetry, so the symmetries form a Lie algebra under Poisson bracket. ζα (η) = eα[·,G] η
For a free particle, p~ and L ~ are both symmetries, and we have just seen  
1
that [p` , Li ] = ik` pk , a linear combination of symmetries, while of course = 1 + α[·, G] + α2 [[·, G], G] + ... η
2
[pi , pj ] = 0 generates the identity transformation and is in the algebra. What 1 2
about [Li , Lj ]? As Li = ik` rk p` , = η + α[η, G] + α [[η, G], G] + ....
2
[Li , Lj ] = [ik` rk p` , Lj ] In this fashion, any Lie algebra, and in particular the Lie algebra formed
= ik` rk [p` , Lj ] + ik` [rk , Lj ]p` by the Poisson brackets of generators of symmetry transformations, can be
= −ik` rk j`m pm + ik` jmk rm p` exponentiated to form a continuous group of finite transformations, called a
Lie Group. In the case of angular momentum, the three components of L ~
= (δij δkm − δim δjk ) rk pm − (δij δm` − δim δj` ) rm p`
form a three-dimensional Lie algebra, and the exponentials of these form a
= (δia δjb − δib δja ) ra pb
three-dimensional Lie group which is SO(3), the rotation group.
= kij kab ra pb = ijk Lk . (6.16)
We see that we get back the third component of L, ~ so we do not get a new
kind of conserved quantity, but instead we see that the algebra closes on the
6.7 Hamilton–Jacobi Theory
space spanned by the momenta and angular momenta. We also note that We have mentioned the time dependent canonical transformation that maps
it is impossible to have two components of L ~ conserved without the third
the coordinates of a system at a given fixed time t0 into their values at
component also being conserved. Note also that ω ~ ·L
~ does a rotation the
a later time t. Now let us consider the reverse transformation, mapping
~
same way on the three vectors ~r, p~, and L. Indeed it will do so on any vector (q(t), p(t)) → (Q = q0 , P = p0 ). But then Q̇ = 0, Ṗ = 0, and the Hamilto-
composed from ~r, and p~, rotating all of the physical system20 . nian which generates these trivial equations of motion is K = 0. We denote
20 by S(q, P, t) the generating function of type 2 which generates this transfor-
If there is some rotationally non-invariant property of a particle which is not built
~ = ~r × p~, in which case L
out of ~r and p~, it will not be suitably rotated by L ~ is not the mation. It satisfies
full angular momentum but only the orbital angular momentum. The generator of
a rotation of all of the physics, the full angular momentum J, ~ is then the sum of L~ and ∂S ∂S
another piece, called the intrinsic spin of the particle. K = H(q, p, t) + = 0, with pi = ,
∂t ∂qi
6.7. HAMILTON–JACOBI THEORY 191 192 CHAPTER 6. HAMILTON’S EQUATIONS

so S is determined by the differential equation which can be easily integrated

! Z qq
∂S ∂S
H q, ,t + = 0, (6.17) W = 2mα − kmq 2 dq + f (α)
∂q ∂t 0
 
α 1
which we can think of as a partial differential equation in n + 1 variables q, t, = f (α) + θ + sin 2θ , (6.20)
ω 2
thinking of P as fixed and understood. If H is independent of time, we can
q q
solve by separating the t from the q dependence, we may write S(q, P, t) = where we have made a substitution sin θ = q k/2α, used ω = k/m, and
W (q, P ) − αt, where α is the separation constant independent of q and t, but
made explicit note that the constant (in q) of integration, f (α), may depend
not necessarily of P . We get a time-independent equation
!
on α. For other hamiltonians, we will still have the solution to the partial
∂W differential equation for S given by separation of variables S = W (q, P ) − αt,
H q, = α. (6.18) because H was assumed time-independent, but the integral for W may not
∂q
be doable analytically.
The function S is known as Hamilton’s principal function, while the As S is a type 2 generating function,
function W is called Hamilton’s characteristic function, and the equa-
tions (6.17) and (6.18) are both known as the Hamilton-Jacobi equation. ∂F2 ∂W
p= = .
They are still partial differential equations in many variables, but under some ∂q ∂q
circumstances further separation of variables may be possible. We consider
first a system with one degree of freedom, with a conserved H, which we will For our harmonic oscillator, this gives
sometimes specify even further to the particular case of a harmonic oscillator. ,
∂W ∂q α 1 + cos 2θ √
Then we we treat a separable system with two degrees of freedom. p= = q = 2αm cos θ.
We are looking for new coordinates (Q, P ) which are time independent, ∂θ ∂θ ω 2α/k cos θ
and have the differential equation for Hamilton’s principal function S(q, P, t):
! Plugging into the Hamiltonian, we have
∂S ∂S
H q, + = 0.
∂q ∂t H = α(cos2 θ + sin2 θ) = α,
For a harmonic oscillator with H = p2 /2m + 12 kq 2 , this equation is which will always be the case when (6.18) holds.
!2 We have not spelled out what our new momentum P is, except that it is
∂S ∂S
+ kmq 2 + 2m = 0. (6.19) conserved, and we can take it to be α. The new coordinate Q = ∂S/∂P =
∂q ∂t
∂W/∂α|q − t. But Q is, by hypothesis, time independent, so
For any conserved Hamiltonian, we can certainly find a separated solution of
the form ∂W
= t + Q.
S = W (q, P ) − α(P )t, ∂α q
and then the terms in (6.17) from the Hamiltonian are independent of t. For
the harmonic oscillator, we have an ordinary differential equation, For the harmonic oscillator calculation (6.20),
!2
dW ∂W 1 1 α ∂θ θ
= 2mα − kmq 2 , = f 0 (α) + (θ + sin 2θ) + (1 + cos 2θ) = f 0 (α) +
dq ∂α q
ω 2 ω ∂α q ω
6.7. HAMILTON–JACOBI THEORY 193 194 CHAPTER 6. HAMILTON’S EQUATIONS
q
where for the second equality we used sin θ = q k/2α to evaluate and s s !
1 4c2 − η 2 ξ 2 − 4c2
s ẏ = ξ ξ˙ 2 − η η̇ .
∂θ −q k 1 4c ξ − 4c2 4c2 − η 2
= = − tan θ.
∂α q
2α cos θ 2α 2α Squaring, adding in the x contribution, and simplifying then shows that
!
m ξ 2 − η2 2 ξ 2 − η 2 ˙2
Thus θ = ωt + δ, for δ some constant. T = η̇ + ξ .
As an example of a nontrivial problem with two degrees of freedom which 8 4c2 − η 2 ξ 2 − 4c2
is nonetheless separable and therefore solvable using the Hamilton-Jacobi Note that there are no crossed terms ∝ ξ˙η̇, a manifestation of the orthogo-
method, we consider the motion of a particle of mass m attracted by Newto- nality of the curvilinear coordinates ξ and η. The potential energy becomes
nian gravity to two equal masses fixed in space. For simplicity we consider   !
1 1 2 2 −4Kξ
only motion in a plane containing the two masses, which we take to be at U = −K + = −K + = .
(±c, 0) in cartesian coordinates x, y. If r1 and r2 are the distances from the r1 r 2 ξ+η ξ−η ξ 2 − η2
particle to the two fixed masses respectively, the gravitational potential is In terms of the new coordinates ξ and η and their conjugate momenta, we
U = −K(r1−1 + r2−1 ), while the kinetic energy is simple in terms of x and y, see that
T = 12 m(ẋ2 + ẏ 2 ). The relation between these is, of course, 2/m  2 2 
2 2 2 2
H= 2 p ξ (ξ − 4c ) + p η (4c − η ) − 2mKξ .
ξ − η2
r12 = (x + c)2 + y 2
Then the Hamilton-Jacobi equation for Hamilton’s characteristic function is
r22 = (x − c)2 + y 2
y
 !2 !2 
2/m  2 ∂W ∂W
r1 r2 (ξ − 4c2 ) + (4c2 − η 2 ) − 2mKξ  = α,
ξ −η
2 2 ∂ξ ∂η
Considering both the kinetic and
or
potential energies, the problem will !2
c c x ∂W 1
not separate either in 2 2
(ξ − 4c ) − 2mKξ − mαξ 2
terms of (x, y) or in terms of (r1 , r2 ), but it does separate in terms of elliptical ∂ξ 2
coordinates !2
2 ∂W 2 1
+(4c − η ) + αmη 2 = 0.
ξ = r1 + r 2 ∂η 2
η = r1 − r 2 If W is to separate into a ξ dependent piece and an η dependent one, the
first line will depend only on ξ, and the second only on η, so they must each
From r12 − r22 = 4cx = ξη we find a fairly simple expression ẋ = (ξ η̇ + ξη)/4c.
˙ be constant, with W (ξ, η) = Wξ (ξ) + Wη (η), and
The expression for y is more difficult, but can be found from observing that !2
1 2
(r + r22 ) = x2 + y 2 + c2 = (ξ 2 + η 2 )/4, so
2 1 2 2
(ξ − 4c )
dWξ (ξ) 1
− 2mKξ − αmξ 2 = β
!2 dξ 2
ξ 2 + η2 ξη (ξ 2 − 4c2 )(4c2 − η 2 ) !2
y2 = − − c2 = , 2
(4c − η )
dWη (η)
2 1
+ αmη 2 = −β.
4 4c 16c2 dη 2
or
1q 2 q These are now reduced to integrals for Wi , which can in fact be integrated
y= ξ − 4c2 4c2 − η 2 to give an explicit expression in terms of elliptic integrals.
4c
6.8. ACTION-ANGLE VARIABLES 195 196 CHAPTER 6. HAMILTON’S EQUATIONS
q
6.8 Action-Angle Variables so J is just a constant 1/ k/m times the old canonical momentum α, and
q q
thus its conjugate φ = k/m(t + β), so ω = k/m as we expect. The
Consider again a general system with one degree of freedom and a conserved
important thing here is that ∆φ = 2π, even if the problem itself is not
Hamiltonian. Suppose the system undergoes periodic behavior, with p(t)
solvable.
and q̇(t) periodic with period τ . We don’t require q itself to be periodic as
it might be an angular variable which might not return to the same value
when the system returns to the same physical point, as, for example, the Exercises
angle which describes a rotation.
If we define an integral over one full period, 6.1 In Exercise 2.7, we discussed the connection between two Lagrangians, L1 and
L2 , which differed by a total time derivative of a function on extended configuration
1 Z t+τ space,
J(t) = p dq, d
2π t L1 ({qi }, {q̇j }, t) = L2 ({qi }, {q̇j }, t) + Φ(q1 , ..., qn , t).
dt
it will be time independent. As p = ∂W/∂q = p(q, α), the integral
R
can be You found that these gave the same equations of motion, but differing momenta
defined without reference to time, just as the integral 2πJ = pdq over one (1) (2)
pi and pi . Find the relationship between the two Hamiltonians, H1 and H2 ,
orbit of q, holding α fixed. Then J becomes a function of α alone, and if we and show that these lead to equivalent equations of motion.
assume this function to be invertible, H = α = α(J). We can take J to be
our canonical momentum P . Using Hamilton’s Principal Function S as the 6.2 A uniform static magnetic field can be described by a static vector potential
generator, we find Q = ∂S/∂J = ∂W (q, J)/∂J −(dα/dJ)t. Alternatively, we A = 12 B
~ ~ × ~r. A particle of mass m and charge q moves under the influence of this
might use Hamilton’s Characteristic Function W by itself as the generator, field.
to define the conjugate variable φ = ∂W (q, J)/∂J, which is simply related (a) Find the Hamiltonian, using inertial cartesian coordinates.
to Q = φ − (dα/dJ)t. Note that φ and Q are both canonically conjugate (b) Find the Hamiltonian, using coordinates of a rotating system with angular
to J, differing at any instant only by a function of J. As the Hamilton- ~ = −q B/2mc.
velocity ω ~
Jacobi Q is time independent, we see that φ̇ = dα/dJ = dH/dJ = ω(J),
6.3 Consider a symmetric top with one point on the symmetry axis fixed in
which is a constant, because while it is a function of J, J is a constant in
space, as we did at the end of chapter 4. Write the Hamiltonian for the top.
time. We could also derive φ̇ from Hamilton’s equations considering W as a
Noting the cyclic (ignorable) coordinates, explain how this becomes an effective
genenerator, for W is time independent, the therefore the new Hamiltonian one-dimensional system.
is unchanged, and the equation of motion for φ is simply φ̇ = ∂H/∂J. Either
way, we have φ = ωt + δ. The coordinates (J, φ) are called action-angle 6.4 (a) Show that a particle under a central force with an attractive potential
variables. Consider the change in φ during one cycle. inversely proportional to the distance squared has a conserved quantity D = 12 ~r ·
I I ! p~ − Ht.
∂φ ∂ ∂W d I d (b) Show that the infinitesimal transformation generated by G := 12 ~r · p~ scales
∆φ = dq = dq = pdq = 2πJ = 2π.
∂q ∂q ∂J dJ dJ ~r and p~ by opposite infinitesimal amounts, Q ~ = (1 +  )~r, P~ = (1 −  )~
2 2 p, or for a
~ ~ −1
finite transformation Q = λ~r, P = λ p~. Show that if we describe the motion in
Thus we see that in one period τ , ∆φ = 2π = ωτ , so ω = 1/τ .
terms of a scaled time T = λ2 t, the equations of motion are invariant under this
For our harmonic oscillator, of course, combined transformation (~r, p~, t) → (Q,
~ P~ , T ).
s
I √ 2α Z 2π 2απ
2πJ = pdq = 2αm cos2 θdθ = q 6.5 We saw that the Poisson bracket associates with every differentiable function
k 0 k/m f on phase space a differential operator Df := [f, ·] which acts on functions g
6.8. ACTION-ANGLE VARIABLES 197 198 CHAPTER 6. HAMILTON’S EQUATIONS

on phase space by Df g = [f, g]. We also saw that every differential operator is Find as many relations as you can, expressible without coordinates, among these
associated with a vector, which in a particular coordinate system has components forms. Consider using the exterior derivative and the wedge product.
fi , where
X ∂
Df = fi . 6.8 We have considered k-forms in 3-D Euclidean space and their relation to
∂ηi
vectors expressed in cartesian basis vectors. We have seen that k-forms are in-
A 1-form acts on such a vector by variant under change of coordinatization of M, so we can use them to examine
dxj (Df ) = fj . the forms of the gradient, curl, divergence and laplacian in general coordinates in
three dimensional space. We will restrict our treatment to orthogonal curvilinear
Show that for the natural symplectic structure ω2 , acting on the differential oper- coordinates (q1 , q2 , q3 ), for which we have, at each point p ∈ M, a set of or-
ator coming from the Poisson bracket as its first argument, thonormal basis vectors êi directed along the corresponding coordinate, so that
dqi (êj ) = 0 for i 6= j. We assume they are right handed, so êi · êj = δij and
ω2 (Df , ·) = df, P
êi × êj = k ijk êk . The dqi are not normalized measures of distance, so we define
which indicates the connection between ω2 and the Poisson bracket. hi (p) so that dqi (êj ) = h−1
i δij (no sum).
P
6.6 Give a complete discussion of the relation of forms in cartesian coordinates (a) For a function f (q1 , q2 , q3 ) and a vector ~v = vi êi , we know that df (~v ) = ~v · ∇f
~ .
in four dimensions to functions, vector fields, and antisymmetric matrix (tensor) Use this to find the expression for ∇f in the basis êi .
~
P
fields, and what wedge products and exterior derivatives of the forms correspond (b) Use this to get the general relation of a 1-form ωi dqi to its associated vector
P
to in each case. This discussion should parallel what is done in section 6.5 for ~v = vi êi .
three dimensions. [Note that two different antisymmetric tensors, Bµν and B̃µν = (c) If a 1-form ω (a) is associated with ~v (a) and 1-form ω (b) is associated with ~v (b) ,
1P we know the 2-form ω (a) ∧ ω (b) is associated with ~v (a) × ~v (b) . Use this to find the
2 ρσ µνρσ Bρσ , can be related to the same 2-form, in differing fashions. They
are related to each other with the four dimensional jk`m , which you will need to general association of a 2-form with a vector.
define, and are called duals of each other. Using one fashion, the two different (d) We know that if a 1-form ω is associated with a vector ~v , then dω is associated
2-forms associated with these two matrices are also called duals. with ∇ ~ × ~v . Use this to find the expression for ∇ ~ × ~v in orthogonal curvilinear
(b) Let Fµν be a 4 × 4 matrix defined over a four dimensional space (x, y, z, ict), coordinates.
with matrix elements Fjk = jk` B` , for j, k, ` each 1, 2, 3, and F4j = iEj = −Fj4 . (e) If the 1-form ω is associated with ~v and the 2-form Ω is associated with F~ , we
Show that the statement that F corresponds, by one of the two fashions, to a know that ω ∧ Ω is associated with the scalar ~v · F~ . Use this to find the general
closed 2-form F, constitutes two of Maxwell’s equations, and explain how this association of a 3-form with a scalar.
implies that 2-form is the exterior derivative of a 1-form, and what that 1-form is (f) If the 2-form Ω is associated with ~v , we know that dΩ is associated with the
in terms of electromagnetic theory described in 3-dimensional language. divergence of ~v . Use this to find the expression for ∇ ~ · ~v in orthogonal curvilinear
(c) Find the 3-form associated with the exterior derivative of the 2-form dual to F, coordinates.
and show that it is associated with the 4-vector charge current density J = (~j, icρ), (g) Use (a) and (f) to find the expression for the laplacian of a scalar, ∇2 f = ∇· ~ ∇f ~ ,
where ~j is the usual current density and ρ the usual charge density. in orthogonal curvilinear coordinates.

6.7 Consider the following differential forms:

6.9 Consider the unusual Hamiltonian for a one-dimensional problem
A = y dx + x dy + dz
B = y 2 dx + x2 dy + dz H = ω(x2 + 1)p,
C = xy(y − x) dx ∧ dy + y(y − 1) dx ∧ dz + x(x − 1) dy ∧ dz
where ω is a constant.
D = 2(x − y) dx ∧ dy ∧ dz
E = 2(x − y) dx ∧ dy (a) Find the equations of motion, and solve for x(t).