B7.

1 Classical Mechanics

James Sparks, Michaelmas Term 2015

About these notes

These are lecture notes for the B7.1 Classical Mechanics course, which is a third year option in
the mathematics syllabus at the University of Oxford. In putting together the notes I have drawn
freely from the large literature on the subject; notably from the reading list below, but also from
many other books and lecture notes (notably those of Paul Tod and David Tong).
Familiarity with the Oxford first year Dynamics course and second year Calculus of Variations
option will be assumed, although we will essentially develop the theory ab initio. Starred paragraphs are not examinable: this material is often included to make contact with other courses (and
may be safely ignored by students not taking those courses), and/or is more advanced. There are
four problem sheets. Please send any questions/corrections/comments to sparks@maths.ox.ac.uk.

Reading
H. Goldstein, C. P. Poole, J. L. Safko, Classical Mechanics.
L. D. Landau, E. M. Lifshitz, Mechanics (Course of Theoretical Physics, Vol. 1 ).
N. M. J. Woodhouse, Introduction to Analytical Mechanics.
V. I. Arnold, Mathematical Methods of Classical Mechanics.

Introduction
This course is about the Lagrangian and Hamiltonian formulations of classical mechanics. These
were introduced and developed in the late 18th and 19th centuries, and recast Newtons laws in
different mathematical frameworks. The reader might immediately ask what the point of this is:
what mileage do you get out of rewriting Newtonian mechanics in a different language? Moreover,
why is it important to study this subject?
In order to answer these questions, lets begin with the Newtonian theory itself. Newtonian
mechanics is an extremely accurate theory, valid over a vast range of scales, and is applicable to
many classes of dynamical problems. The axioms are also clear and simple to state. Although
the later developments of special relativity, general relativity and quantum mechanics undoubtedly
provide a more accurate description of the real world, these are all much more complex descriptions
1

of Nature. Scientists, even theoretical physicists, will often try to revert to using the techniques
of classical mechanics where possible. Indeed, measuring the differences from Newtonian theory
usually requires very well-designed and delicate experiments.
From a computational point of view Newtonian theory, as summarized for point particles in
section 1, quickly becomes cumbersome and inefficient as the systems become more complicated.
In particular Newtons laws require one to work in an inertial frame in Cartesian coordinates,
which is often inconvenient. The Lagrangian and Hamiltonian formalisms, on the other hand, are
coordinate independent, and provide a more elegant and computationally efficient framework in
which to work. For example, it much easier to solve constrained systems, as there is no need to
introduce constraint forces (such as the tension in a simple pendulum, or the constraint forces for a
bead moving on a wire); rigid body motion is also easier to implement. The Lagrangian formulation
also introduces a new fundamental principle: the principle of least action, also known as Hamiltons
principle. This gives a surprising amount of insight into classical mechanics, for example making
clear the relation between symmetries and conservation laws (via Noethers theorem).
The ideas and techniques developed in the Lagrangian and Hamiltonian formulations of classical
mechanics also generalize to other areas of theoretical physics. For example, there are Lagrangian
and Hamiltonian descriptions of electromagnetism and general relativity, which play an important
role in formulating those theories. The ideas and principles we shall encounter were also key to
the development of quantum mechanics in the 20th century. Those who have studied quantum
mechanics may have noticed that the theory looks nothing like Newtonian mechanics. In fact
the Hamiltonian formulation of the latter is closely related to the quantum description of a nonrelativistic particle you may have already seen. The principle of least action also led to Feynmans
formulation of quantum mechanics as a path integral/sum over histories, which in turn has been
central to the development of particle physics in the second half of the 20th century. Finally,
Lagrangian and Hamiltonian mechanics have also had a significant impact on the development of
various branches of pure mathematics, particularly geometry.
Hopefully this has at least partly addressed the questions posed at the beginning of the introduction.

Contents
1 Newtonian mechanics

1.1

Reference frames . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1.2

Newtons laws . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1.3

Galilean transformations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1.4

Closed systems and Galilean relativity . . . . . . . . . . . . . . . . . . . . . . . . .

1.5

Energy and angular momentum . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

2 Lagrangian mechanics

11

2.1

Generalized coordinates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

11

2.2

The principle of least action . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

14

2.3

Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

18

2.4

Noethers theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

22

2.5

Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

27

2.6

* More on Hamiltons principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

34

3 Small oscillations

36

3.1

Equilibria and quadratic Lagrangians . . . . . . . . . . . . . . . . . . . . . . . . . .

36

3.2

Normal frequencies and normal modes . . . . . . . . . . . . . . . . . . . . . . . . .

37

3.3

Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

39

4 Rigid body dynamics

43

4.1

Rotating frames and angular velocity . . . . . . . . . . . . . . . . . . . . . . . . . .

43

4.2

Motion in a non-inertial frame

. . . . . . . . . . . . . . . . . . . . . . . . . . . . .

46

4.3

Rigid body motion and the inertia tensor . . . . . . . . . . . . . . . . . . . . . . .

47

4.4

Eulers equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

52

4.5

Euler angles and SO(3) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

56

4.6

The Lagrange top . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

59

5 Hamiltonian mechanics

64

5.1

The Legendre transformation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

64

5.2

Hamiltons equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

66

5.3

Poisson brackets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

70

5.4

Canonical transformations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

74

5.5

Generating functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

79

5.6

Liouvilles theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

84

5.7

The Hamilton-Jacobi equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

86

5.8

* Quantum mechanics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

92

A Appendix: Levi-Civita alternating symbol

94

Newtonian mechanics

We begin with an overview of Newtonian mechanics for point particles. Although much of this
material may already be quite familiar, it will be important to have a good grasp of these basic
notions when we come to generalize the results later in the course.

1.1

Reference frames

In classical mechanics space is modelled by R3 , equipped with the usual Euclidean metric. More
precisely, a reference frame S is specified by a choice of origin O, together with a choice of Cartesian
coordinate axes. With respect to this frame one then writes a position vector as r = (x, y, z). The
trajectory of a point particle is represented by a curve r = r(t), parametrized by time t. The
velocity of the particle in the frame S is then r = (x,
y,
z),
where a dot will denote derivative with
respect to time. Similarly, its acceleration is r = (
x, y, z).

'
frame S

'

'

'

'

x(t)

y
x

O
frame S

Figure 1: Relative to a choice of reference frame S, the origin O0 of another frame S 0 is at x(t),
and the coordinate axes of S 0 may also rotate with respect to the axes of S.
Of course the choice of frame is far from unique, and the motion of a particle viewed in two
different frames can look very different. If we fix an initial choice of S, then the origin O0 of any
other frame S 0 will be at some position x(t), measured with respect to the origin O of S. Moreover,
the coordinate axes of S 0 may rotate relative to the axes of S see Figure 1. In particular if S
and S 0 have the same origin (so x(t) = 0 for all t), then the frames differ only by a rotation of
the axes. A position vector r measured in S is then r0 = (x0 , y 0 , z 0 ) = R r as measured in S 0 . Here
R = R(t) is a 3 3 rotation matrix, which in general can depend on time t. Mathematically
the three-dimensional rotation group is O(3) GL(3, R), the set of real 3 3 matrices satisfying
RT R = RRT = 1, where

1 denotes the 3 3 identity matrix. This ensures the transformation

R O(3) between frames preserves Euclidean distances all observers in Newtonian mechanics are
postulated to measure the same distance between any two points. We shall describe the rotation
4

group and rotating frames in much greater detail in section 4 when we come to discuss rigid body
motion.

1.2

Newtons laws

Newtons laws of motion apply to point particles. These are objects whose dimensions may be
neglected, to a good approximation, in describing their motion. For example, this is the case if
the size of the object is small compared to the distances involved in the dynamics, e.g. the motion
of a planet around the Sun. Of course, it is no good treating the Earth as a point particle if you
want to understand the effects of its rotation. On the other hand, as we shall see in section 4 the
centre of mass of an extended rigid body does behave like a point particle. Point particles have a
mass m, and an associated momentum p = mr.
The first of Newtons laws singles out a special equivalence class of reference frames, called
inertial frames. These are defined by
N1: Inertial frames exist. In such a frame an object either remains at rest or moves
with constant momentum (uniform motion in a straight line), unless acted on by an
external force.
Here the force that is acting is understood to have an identifiable source, e.g. gravity, electromagnetism, friction. A non-inertial frame S 0 is accelerating with respect to an inertial frame S. That
is, the origin O0 of S 0 is accelerating with respect to O, or the axes of S 0 are rotating relative to
the axes of S. In the non-inertial frame a particle will appear to be acted on by ficticious forces,
for example the Coriolis force, or centrifugal force (more about this in section 4.2). A frame at
rest on the surface of the Earth is a very good approximation to an inertial frame, ignoring the
rotation of the Earth around its axis and its acceleration about the Sun.1 Compare such a frame S
to someone standing on a roundabout, whose frame S 0 rotates around a fixed vertical axis relative
to S: a stationary object in S will appear to be accelerating to the person on the roundabout.
In an inertial frame the dynamics of a particle is governed by
N2:

F(r, r , t) = p.

Assuming the mass m is constant the right hand side of Newtons second law is p = mr, although
one could also consider variable mass bodies, e.g. a rocket ship that expels the spent fuel. In this
course well assume m is constant. The external force F can in general depend on the particles
position r, its velocity r (e.g. the drag force due to motion through a fluid), and on time t (e.g. a
charged particle moving in a time-dependent electromagnetic field). Newtons second law is then
a second order ODE for r(t). General theorems from the theory of differential equations guarantee
1

The former does give rise to a very small measurable effect. For example, Foucaults pendulum detects the
rotation of the Earth.

that under suitable conditions on the function F(r, r , t), specifying the position r and velocity r
at some initial time t = t0 gives a unique solution for the particle trajectory r(t).
Finally, if we have more than one particle, then
N3: If particle 1 exerts a force F = F21 on particle 2, then particle 2 also exerts a force
F12 = F on particle 1.
In other words, F12 = F21 . This is often paraphrased by saying that every action has an equal
and opposite reaction. There is also a strong form of Newtons third law:
N30 : The force in N3 above acts along the vector connecting particle 1 and particle 2.
That is, F = F21 (r1 r2 ), where rI denotes the position vector of particle I. Such forces
are called central forces. The strong form of Newtons third law is true for the gravitational and
electrostatic forces between particles:

Example (gravitational force between two point masses): According to Newton (and Hooke), the
gravitational force on particle 1 due to particle 2 is given by
F12 = GN

m1 m2 (r1 r2 )
.
|r1 r2 |2 |r1 r2 |

(1.1)

Here mI is the mass of particle I, and GN ' 6.67 1011 N m2 kg2 is Newtons gravitational
constant. The Coulomb electrostatic force between two point charges e1 , e2 takes a similar central
inverse square law form, except unlike masses charges can be both positive and negative, leading
to repulsive as well as attractive forces.

There are examples of forces that do not satisfy the strong form of Newtons third law, notably
the magnetic component of the Lorentz force acting on a charged particle moving in the electromagnetic field generated by another charged particle. However, a proper discussion of this would
take us too far into electromagnetic theory, and in any case requires special relativity (rather than
Galilean relativity) to understand properly, so we will not discuss this further here.

1.3

Galilean transformations

Inertial frames are not unique. Rather there is an equivalence class of inertial frames, related to
each other by Galilean transformations. Specifically, we have the following transformations of an
inertial frame S:

temporal translations, t0 = t s, where s is a constant,

spatial translations, r0 = r x, where x is a constant vector,

constant rotations, r0 = R r, where R O(3) is a constant 3 3 orthogonal matrix,

Galilean boosts, r0 = r vt, where v is a constant velocity.

6

These map uniform motion in S to uniform motion in S 0 . Altogether they generate the 1+3+3+3 =
10-dimensional Galilean group.
Notice that in the first transformation we have been tacitly assuming there is a notion of absolute
time, unique up to what we call time t = 0. In particular in Newtonian mechanics all observers
measure the same time interval between any two events. Of course this is not true in special
relativity (essentially it took until Einstein to realize that assuming there is an absolute notion of
time is an assumption). Galilean boost transformations mean there is no absolute rest frame in
Newtonian mechanics.
The first Galilean transformation is a statement of homogeneity of time (Newtonian physics is
invariant under t t s), while the second and third Galilean transformations say that space is
respectively homogeneous and isotropic. These are symmetries of Galilean spacetime.

1.4

Closed systems and Galilean relativity

A closed system is one in which all forces are internal, acting between the constituents of the
system. To be concrete let us consider a closed system of N point particles. We suppose these
have constant masses mI , I = 1, . . . , N , and are at positions rI (t) as measured in an inertial frame.
Newtons second law N2 for the system may be written
FI (r1 , . . . , rN , r 1 , . . . , r N , t) = p I = mI rI ,

(1.2)

where FI denotes the total force on the Ith particle. For such a closed system we have
Galileos principle of relativity: Newtonian dynamics is invariant under Galilean transformations.
More precisely, if we apply the same Galilean transformation to each of the particle trajectories
{rI (t)} solving (1.2), then the resulting trajectories solve the same system of equations. For
example, if {rI (t)} solves (1.2) then {rI (t s)} must solve the same system, for all s R.
This invariance under temporal translations immediately implies that FI must be independent of
time t, so that we may write FI = FI (r1 , . . . , rN , r 1 , . . . , r N ). Similarly, invariance under spatial
translations and boosts means that the forces depend only on the relative positions (rJ rK )
and relative velocities (rJ r K ) of the particles, respectively. Finally, invariance under a rotation
R O(3) means that FI is further constrained to obey
R FI (r1 , . . . , rN , r 1 , . . . , r N ) = FI (R r1 , . . . , R rN , R r 1 , . . . , R r N ) .

(1.3)

A closed system consisting of a single point particle is not very interesting: a little thought shows
that the above constraints imply F = 0 (a single particle cannot act on itself). Such a particle
moves at constant momentum/velocity by N1. When treating a single point particle subject to
an external force, we have in mind that (a) something else is responsible for producing that force,
7

and (b) we are entirely ignoring the effect the particle has on whatever that something else is (i.e.
we are ignoring its back-reaction). One can imagine such a model arising from a closed system,
in which the particle has been singled out and the external force F is simply the sum of the
forces on the particle from the rest of the system. Such external forces typically wont be Galilean
invariant. Of course, whether or not it is reasonable to neglect the effects of the particle on the
rest of the system depends on the circumstances. It is perhaps worth pointing out though that
closed systems can sometimes be rewritten/reinterpreted as non-closed systems. A good example
is the two-body problem, of two point masses m1 , m2 interacting via the gravitational force (1.1).
This closed system may be rewritten as a problem for a single point particle (with mass given
by the reduced mass =

m1 m2
m1 +m2 )

moving in an external central potential. We shall examine this

example in more detail in section 2.5.

1.5

Energy and angular momentum

Consider a single point particle of mass m, acted on by an external force F. The work done W by
the force F along a path connecting position vectors r(1) , r(2) is the line integral
Z

r(2)

F dr .

(1.4)

r(1)

If we now consider a trajectory of the particle r(t) satisfying Newtons second law mr = F, we
may compute
Z
W

r(2)

t2

F dr =

=
r(1)

t2

F r dt = m
t1

r r dt =
t1

= T (t2 ) T (t1 ) .

1
m
2

t2

t1

d
(r r ) dt
dt
(1.5)

Here the particle trajectory begins at position r(1) = r(t1 ) at time t1 , and ends at r(2) = r(t2 ) at
time t2 , and we have defined the kinetic energy of the particle as
T

1
m|r|2 .
2

(1.6)

Notice that T is not invariant under Galilean boosts, and so in general depends on the choice of
inertial frame S we choose to measure it in. The equality in (1.5) is called the work-energy theorem:
the work done by the force F along the particle trajectory is equal to the change in kinetic energy.
In the remainder of this course we will almost entirely focus on conservative forces. Recall that
these arise from the gradient of a potential function V = V (r) via
F = V (r) .

(1.7)

In Cartesian coordinates r = (x, y, z) this reads F = (x V, y V, z V ). Notice in particular

that V , and hence also F, depends only on the position vector r. Note also that V is defined
only up to an additive constant. It is natural to fix this freedom, where it makes sense to do
8

so, by requiring that the potential is zero when the particle (or more generally distances between
particles) is at infinity, c.f. (1.11) below. Conservative forces have the property that the work
done is independent of the path from r(1) to r(2) . This follows from the fundamental theorem of
calculus:
Z

r(2)

r(2)

V dr = V (r(2) ) + V (r(1) ) .

F dr =
r(1)

(1.8)

r(1)

Combining with the work-energy theorem (1.5) and rearranging we thus deduce the conservation
of energy
T (t1 ) + V (r(t1 )) = T (t2 ) + V (r(t2 )) E .

(1.9)

E T +V ,

(1.10)

Thus the total energy

is constant along the trajectory.

The inverse square law force (1.1), that governs the gravitational or Coulomb interaction between
point masses/charges, is conservative:
V (r) =

F = V =

r,
r3

(1.11)

where r = |r| and is a constant. Non-conservative forces include friction/drag forces that
depend on r . However, the fundamental forces in Nature seem to be conservative. For example,
a body that experiences a frictional force will typically lose energy, but in practice we know what
happens to this energy: it is converted into heat. Modelling this takes us outside the realm of
classical mechanics and into thermodynamics. But at a fundamental level, we expect energy to be
conserved.
Still focusing on the single point particle, recall that its angular momentum (about the origin
O of S) is defined as
L = rp .

(1.12)

We will sometimes write the cross product of two vectors using the Levi-Civita alternating symbol,
discussed in the appendix at the end of these lecture notes. The angular momentum (1.12) is the
moment of the momentum p about O. Similarly, the moment of the force about O is called the
torque
= rF .

(1.13)

Since L = r p + r p = r p (as p = mr is parallel to the velocity r ), the moment of N2 about

the origin gives the equation

= L .
9

(1.14)

In particular central forces have F r, so that the torque = 0. It then follows from (1.14) that
central forces lead to conservation of angular momentum, L = 0.
It is straightforward to extend these concepts to our closed system of N point particles in
section 1.4. If particle J exerts a force FIJ on particle I for J 6= I, then by N3 therefore
FJI = FIJ . Newtons second law for particle I (1.2) then reads
X

FIJ

= FI = p I .

(1.15)

J6=I

We could also add an external force Fexternal

to the left hand side, but the system would then not
I
P
be closed. The total momentum of the system of particles is defined to be P = N
I=1 pI , and by
summing (1.15) over all particles we deduce that
=
P

N X
X

FIJ = 0 .

(1.16)

I=1 J6=I

Here the N (N 1) terms in the sum of forces cancel pairwise due to N3: FIJ = FJI . Thus for
this closed system of point particles we see that the total momentum is conserved.
We may similarly define the total angular momentum about the origin O as
L =

N
X

rI pI .

(1.17)

I=1

As for the single point particle we compute

L =

N
X

rI p I =

I=1

N
X
I=1

rI

FIJ .

(1.18)

J6=I

In particular we have 21 N (N 1) pairs of terms on the right hand side

rI FIJ + rJ FJI

= (rI rJ ) FIJ ,

(1.19)

where we have used N3. Imposing the strong form of Newtons third law N30 means that FIJ
(rI rJ ), and we see that the total angular momentum is conserved, L = 0.
One of the general results we shall derive in Lagrangian mechanics is that conservation of energy,
momentum, and angular momentum for any closed system follow from the symmetries of Galilean
spacetime described in section 1.3.

10

Lagrangian mechanics

In this section we introduce the Lagrangian formulation of classical mechanics. The general theory
is developed in sections 2.1 2.4, with a number of detailed examples presented in section 2.5.
The reader should feel free to dip into the examples while digesting the general theory, especially
on a first reading.

2.1

Generalized coordinates

In order to study the dynamics of a physical system, one first needs to describe its possible
configurations. Any set of independent quantities that specify uniquely the position of the system
at a given time are called generalized coordinates. These label the points of the configuration space
Q. The number of generalized coordinates is called the number of degrees of freedom.
For example, for the point particle discussed in section 1 we used Cartesian coordinates r =
(x, y, z) with respect to an inertial frame S. These coordinates obviously specify uniquely the
position of the particle in R3 at a given time t, so Q = R3 and the number of degrees of freedom
is 3. For N point particles we similarly used N sets of Cartesian coordinates rI = (xI , yI , zI ),
I = 1, . . . , N , giving a configuration space Q = R3N with 3N degrees of freedom.
For the point particle we may also consider changing to non-Cartesian coordinates. Consider
the general coordinate change
qi = qi (x, y, z, t) ,

i = 1, 2, 3 ,

(2.1)

which might depend explicitly on time t (that is, the new coordinate system is moving relative to
the original coordinate system). We thus replace the Cartesian coordinates (x, y, z) (q1 , q2 , q3 ).
Introducing x1 = x, x2 = y, x3 = z we will also have the inverse coordinate transformation
xi = xi (q1 , q2 , q3 , t) ,

i = 1, 2, 3 .

(2.2)

This latter change of coordinates is said to be non-singular at a point (q1 , q2 , q3 ) if the Jacobian
determinant det (xi /qj ) is non-zero at that point. This condition plays a role when writing
differential equations in different coordinate systems, as the Jacobian xi /qj and its inverse
qi /xj enter the chain rule.2
A familiar example is cylindrical polars
x = % cos ,

y = % sin ,

z = z,

(2.3)

where % 0 and [0, 2). Here (q1 , q2 , q3 ) = (%, , z), and the coordinates are adapted to rotation about the z-axis. Notice that % = 0 (the z-axis) is a coordinate singularity, as det (xi /qj ) = %
2

* We may regard a coordinate transformation (x1 , x2 , x3 ) (q1 , q2 , q3 ) at fixed time t as a map x : R3 R3 ,

with components xi = xi (q1 , q2 , q3 ). By the inverse function theorem, if the Jacobian determinant det (xi /qj ) is
non-zero at a point P = (q1 , q2 , q3 ), then there is an open set containing P on which this map is invertible with
differentiable inverse.

11

is zero at % = 0. In practice this causes no problems provided one remembers it is there (for example
it explains why the Laplacian x2 + y2 = %2 + %1 % +

1 2

%2

looks singular at % = 0).

Another natural coordinate system is spherical polars

x = r sin cos ,

y = r sin sin ,

z = r cos ,

(2.4)

where r 0, [0, ], [0, 2). Here (q1 , q2 , q3 ) = (r, , ), with the coordinates adapted to
rotation about the origin r = 0. Again, there are coordinate singularities at r = 0 and at = 0, ,
which are the zeros of the Jacobian determinant det (xi /qj ) = r2 sin .
The notion of generalized coordinates is particularly useful for constrained problems. You have
already met many such problems in Dynamics. Perhaps the simplest example is the simple pendulum. Here a point mass m swings on a rigid (but very light) rod of length l, freely pivoted at
one end to the origin O, and constrained to move in a vertical plane containing the z-axis see
Figure 2. Taking the plane of motion to be the (x, z) plane at y = 0, the rod constrains the point
mass coordinates to satisfy x2 +z 2 = l2 . It is then convenient to describe the position of the system
by the angle given by x = l sin , z = l cos . Thus [, ) is a generalized coordinate for
this problem, q = , which has one degree of freedom. Since = is identified with = ,
notice that the configuration space is a circle, Q = S 1 .

z
z

s = arclength

(s) = (x(s),0,z(s))

(s)

x(s)

mg
Figure 2: A simple pendulum with generalized coordinate q = .

Figure 3: A bead moving on a fixed wire. We

might use arclength s, the x-coordinate, or the
angle as a generalized coordinate.

As another example consider a bead moving on a fixed wire. Here there are several natural
choices of generalized coordinate. For simplicity suppose that the wire lies in the (x, z) plane at
y = 0, as in Figure 3. We can specify the wire as the image of a parametrized curve : [0, 1] R3
given by (s) = (x(s), 0, z(s)). A natural generalized coordinate is then simply the arclength s.
But there are other choices. Provided the projection onto the x coordinate is injective (so that
distinct points on the curve correspond to distinct values of x), we could also use the x coordinate
12

of the bead as a generalized coordinate. Another choice is to use the angle subtended from the
origin, again provided the lines = constant through the origin intersect the curve in at most
one point. For example, the latter would be a natural choice if the curve was a segment of a
circle centred at the origin. This problem again has one degree of freedom, and the configuration
space is an interval (assuming something stops the bead from falling off either end of the wire).
Constrained motion on a surface is similar, with two degrees of freedom.
In section 4 we will consider rigid body motion. A rigid body has six degrees of freedom: three
describe the position of the centre of mass, two determine the orientation in space of some axis
fixed in the body, and there is one more describing rotations about that axis. For generalized
coordinates it is then natural to use a mixture of Cartesian coordinates, describing the centre of
mass motion, and angles, describing the orientation of the rigid body.
In general then a physical system will have n degrees of freedom, described by n real generalized
coordinates qa = qa (t), a = 1, . . . , n. Their time derivatives qa are called generalized velocities, qa
are generalized accelerations, etc. We will use the boldface notation q = (q1 , . . . , qn ), although as
the examples above illustrate this is not to be confused (in general) with a vector in physical space
R3 . The dynamics of the system will trace out a curve q(t) in the configuration space Q, which
one could think of as a quasi-particle moving on Q, whose position at any time determines the
configuration of the whole system.3 However, to avoid any possible confusion we shall generally
refer to q(t) as the system trajectory, or path.
* The intuitive ideas of configuration space Q and generalized coordinates we have described
can be made much more mathematically precise. A more rigorous account of a general
framework for classical mechanics would take Q to be a differentiable manifold. Roughly,
these are topological spaces that look locally like Rn (the open sets are homeomorphic to
open sets in Rn ), but globally they can have much more non-trivial topology. Familiar
examples are smooth curves and surfaces in R3 (with n = 1, n = 2 respectively). The study
of geometry and calculus on these spaces is called differential geometry. However, expanding
on these ideas would take us too far from our main subject, and in any case one doesnt
need the full machinery of differential geometry to understand the fairly simple systems we
shall study in these lectures. Those wishing to see more details might consult chapter 8
of Introduction to Analytical Dynamics by N. M. J. Woodhouse, or the book Mathematical
Methods of Classical Mechanics by V. I. Arnold.
3

For example, the configuration space for two point particles moving in R3 is Q = R6 , which may be coordinatized
by Cartesian coordinates q = (x1 , y1 , z1 , x2 , y2 , z2 ) for the two particles. The position of the quasi-particle in Q then
determines both physical positions of the two actual partices in R3 .

13

2.2

The principle of least action

We are now in a position to introduce a new fundamental principle that governs the dynamics of
physical systems. The principle of least action, or Hamiltons principle, asserts that systems are
characterized by a function L(q1 , . . . , qn , q1 , . . . , qn , t), called the Lagrangian. We write this more
t). The Lagrangian function in turn determines the dynamics of the system (the
briefly as L(q, q,
equations of motion) as a calculus of variations problem.
To describe this, consider all smooth paths q(t) in Q with fixed boundary conditions at the
endpoints. That is, we consider all smooth q(t) satisfying
q(t1 ) = q(1) ,

q(t2 ) = q(2) ,

(2.5)

for some fixed time interval [t1 , t2 ], and with q(i) fixed for i = 1, 2. Then the path followed by the
system between these two positions is such that the action
Z t2

S[q(t)]
L(q(t), q(t),
t) dt ,

(2.6)

t1

is extremized. The action S is a functional that is, a function whose argument is itself a function
and the above extremal problem is of the same type encountered in the second year Calculus of
Variations course.
but not higher derivatives, is related to the fact
That the Lagrangian depends only on q and q,
that one expects the dynamics of a system to be determined uniquely once one specifies q and q
at some initial time t = t0 . Compare to our discussion of Newtons second law in section 1.2. This
expectation is itself sometimes elevated to a principle: the Newton-Laplace determinacy principle.
It also means we expect the equations of motion to be second order differential equations for q(t).

q(2)
q(t)

q(1)

q(t) + q(t)

Figure 4: A trajectory q(t) in configuration space Q, with fixed endpoints q(t1 ) = q(1) and
q(t2 ) = q(2) , together with a variation q(t) + q(t).
Let us determine the differential equations that result from extremizing the action S in (2.6).
We suppose the extremum occurs at a critical function q(t), such that the change S = 0 when
q(t) is replaced by any variation q(t) q(t) + q(t). Here q(t) is a small change in q(t), where
the fixed endpoints of the path require the boundary conditions q(t1 ) = q(t2 ) = 0. We may
14

make this more precise by writing

q(t) = u(t) ,

(2.7)

u(t1 ) = u(t2 ) = 0, and  is a real number. We then compute

Z

t2

t) dt
L(q + u, q +  u,

S[q(t) + q(t)] =
t1

= S[q(t)] + 

n Z
X
a=1

t2

t1


L
L
ua +
u a dt + O( 2 ) .
qa
qa

(2.8)

Here we have simply Taylor expanded around  = 0 using the chain rule. Thus the first order
variation in S is
n Z
X

t2


L
L
S 
ua +
u a dt
qa
qa
a=1 t1
(Z 



t2 )
n
t2
X
L
d L
L
= 

ua dt +
ua
qa dt qa
qa
t1
t1
a=1


n Z t2 
X
d L
L

= 
ua dt ,
qa dt qa
t1

(2.9)

a=1

the last equality holding since u(t1 ) = u(t2 ) = 0. The requirement that q(t) is an extremum of S,
S = 0, for all such variations q(t) = u(t) then means that
 
L
d L

= 0,
dt q
q

(2.10)

or in components
d
dt

L
qa

L
qa

= 0,

a = 1, . . . , n ,

(2.11)

The equations (2.10) are known in classical mechanics as Lagranges equations, or the EulerLagrange equations. They are a set of n second-order differential equations for q(t). Notice that
nothing we have said guarantees that the solution q(t) to (2.10) actually minimizes the action S.
In fact there are examples where it does not i.e. the stationary point solution to S = 0 is not a
minimum of the action (see Problem Sheet 1). It should thus more properly be called the principle
of stationary action.

N.B. In deriving the Lagrange equations (2.10) one should be wary of the first fundamental
confusion of calculus (N. M. J. Woodhouse, Introduction to Analytical Dynamics). The partial
derivatives of L appearing in the second line of (2.8) involve regarding L = L(q, v, t), where v = q
is treated as an independent variable from q. However, in computing the derivative d/dt in the
Lagrange equation (2.10) we mean the derivative of

t),
v (q(t), q(t),

which is a function only of

time t. If you follow the derivation carefully these comments are self-evident, but it is also easy
15

to get confused! The potential confusion arises because we treat q and q as independent variables
on which L depend, but also q = q(t) for the system trajectory.

As mentioned in the introduction, Lagranges equations (2.10) hold in any coordinate system.
This actually follows from the formulation of the principle of least action: it is paths that extremize
S. If we change coordinates then the form of the path will also change, but only due to the
coordinate transformation itself it is the same path in Q. One can check this statement very
directly: if we make a coordinate transformation
q = q(
q, t) ,

(2.12)

then
d
dt

L
qa

qa




n 
X
d L
L qb
=

,
dt qb
qb qa

a = 1, . . . , n .

(2.13)

b=1

Here we use the chain rule to compute

qa =

n
X
qa
b=1

qa
, t) ,
qb +
= qa (
q, q
qb
t

(2.14)

and define
q, q
, t) L(q(
q, q
, t), t) .
L(
q, t), q(

(2.15)

is obtained simply by substituting the coordinate transThat is, the transformed Lagrangian L
formation (2.12), (2.14) into the original Lagrangian. The Lagrangian is said to transform as a

scalar. It is then common to drop the tilde on L, and simply remember that the notation L/ q
, t), and then takes the
means one first substitutes q = q(
q, t) to regard L as a function of (
q, q
, with q
and t held fixed.
partial derivative with respect to q
Going back to (2.13), since for a (non-singular) coordinate transformation the Jacobian matrix
qb / qa is invertible, we see that Lagrange equations (2.10) hold if and only if
!

d L
L

= 0

dt q
q

(2.16)

hold. Verifying (2.13) amounts to a short computation using the chain rule, and is left as an
exercise on Problem Sheet 1.
Notice that if we have two Lagrangians L1 , L2 related by
t) = L1 (q, q,
t) +
L2 (q, q,

d
f (q, t) ,
dt

(2.17)

then these lead to the same Lagrange equations. This follows since the corresponding actions are
related by
S2 = S1 + f (q(2) , t2 ) f (q(1) , t1 ) ,
16

(2.18)

where we have used the fundamental theorem of calculus. The actions thus differ by a term whose
variation is zero. In classical mechanics (2.17) should be regarded as an equivalence relation on
Lagrangians. Another comment is that if we have two Lagrangians L1 , L2 describing decoupled
systems, then the Lagrangian for the combined system is simply L = L1 + L2 . This is reasonable,
as then the equations of motion for both systems follow from that for L.
So far our comments have been very general. As mentioned in the introduction, the principle of
least action (appropriately generalized) actually applies to many physical theories, including for
example electromagnetism and general relativity. However, in classical mechanics the Lagrangian
for a system takes the simple form
L = T V ,

(2.19)

where T is the kinetic energy of the system, and V is its potential energy. We shall see more
precisely which systems this applies to in the next subsection, and why the Lagrange equations are
equivalent to Newtons equations. Our application of Lagrangian mechanics to such systems will
then always begin by identifying T and V , in whichever generalized coordinates we have chosen to
use. Let us first see that this correctly reproduces Newtons second law as the Lagrange equations
for a point particle:

Example (point particle in Cartesian coordinates): As in section 1.5 we consider a single point
particle of mass m, in Cartesian coordinates r = (x1 , x2 , x3 ) for an inertial frame S. The particle
is acted on by an external force F = V , where V = V (r) is the potential. The Lagrangian
(2.19) is thus
L =

1
m|r|2 V (r) .
2

(2.20)

Here the generalized coordinates are qa = xa , a = 1, 2, 3. Since L/q = V /r = F and

L/ q = mr = p, we immediately deduce that the Lagrange equations (2.10) are simply Newtons
second law. It is straightforward to extend this computation to a system of N particles with masses
mI at positions rI (t), interacting through a potential V = V (r1 , . . . , rN ).
We thus see that Hamiltons principle (equivalently the Lagrange equations of motion) is equivalent to Newtons second law for point particles. Lets look at another simple example:

Example (simple pendulum): Let us return to the simple pendulum discussed in the previous
subsection, and shown in Figure 2. Resolving the tension force T in the x and z directions,
Newtons second law gives
m
x = T sin ,

m
z = mg + T cos .

17

(2.21)

Here recall x = l sin , z = l cos . Taking cos times the x equation of motion plus sin times
the z equation of motion we easily derive
g
= sin ,
l

(2.22)

which is the equation of motion for the coordinate . The other linear combination determines the
tension
T = mg cos + ml2 ,

(2.23)

which can be understood as balancing the component of the weight along the rod and the centripetal force for circular motion about the origin O, respectively.
On the other hand the Lagrangian description with generalized coordinate q = leads immediately to (2.22). The kinetic energy is T = 12 m(x 2 + z 2 ) = 21 ml2 2 , while the potential energy is
V = mgz = mgl cos . Thus the Lagrangian is
L = T V =

1 2 2
ml + mgl cos ,
2

(2.24)

and the Lagrange equation (2.10) with q = indeed gives (2.22). This short calculation might
look like a sleight of hand: we have (apparently) entirely ignored the tension force, but still found
the correct dynamics. In fact this is one of the advantages of the Lagrangian formalism, and well
explain why (2.24) correctly describes the dynamics in the next subsection.

2.3

Constraints

In the simple pendulum example just studied the motion of the mass m in the (x, z) plane was
constrained to satisfy x2 + z 2 = l2 , due to the rod pivoted at O. This is an example of a holonomic
constraint. To describe this more generally, we need some more notation. The constrained motion
is by definition embedded inside a larger unconstrained configuration space. Let x1 , . . . , xd denote
(generalized) coordinates on this larger space, denoting x = (x1 , . . . , xd ). For example, for the
constrained motion of a single particle moving in R3 we may take x = (x, y, z) to be Cartesian
coordinates. Then holonomic constraints on the motion constrain the coordinates to satisfy
A = 1, . . . , d n .

fA (x, t) = 0 ,

(2.25)

Here n is the number degrees of freedom satisfying the constraints. For example, for the simple
pendulum the constraint on motion in the (x, z) plane is f (x, z) = 0 where f (x, z) = x2 + z 2 l2 .
This is an example of a scleronomous constraint, meaning it doesnt depend on time t. More
generally these are given by fA (x) = 0. The general time-dependent case in (2.25) are called
rheonomous constraints, e.g. compare the bead moving on the fixed wire in Figure 3 to the case
where the wire is rotating about the x-axis.

18

Let us fix the time t, and consider fA (x, t) = 0, A = 1, . . . , d n, which we assume are d n
independent equations. As such, we expect to be able to eliminate d n of the d variables xi ,
i = 1, . . . , d, in terms of the remaining n variables. Equivalently we can say that there are functions
x = x(q, t) ,

(2.26)

which parametrize the general solution to (2.25). The variables qa , a = 1, . . . , n, are then generalized coordinates for the constrained motion. For example, for the simple pendulum we have
x = l sin , z = l cos , which solve the constraint f (x, z) = 0 where f (x, z) = x2 + z 2 l2 .
A precise definition of the constraints being independent is that the (d n) d matrix fA /xi
has maximal rank d n at every point. It then follows from the implicit function theorem that we
can solve the constraints as in (2.26).4 Moreover, there is always a system of coordinates y1 , . . . , yd
such that the constraints are simply
A = 1, . . . , d n .

yn+A = 0 ,

(2.27)

We may then take ya = qa for a = 1, . . . , n to be our generalized coordinates. Such a coordinate

system is said to be adapted to the constraints {fA = 0}. We will not give a proof of this result,
but it follows from methods similar to those seen in the Part A course Introduction to Manifolds.
As an example, the coordinates y1 = q1 = , y2 = q2 = , y3 = r a are adapted to the constraint
f (x, y, z) = x2 + y 2 + z 2 a2 = 0 for motion on a sphere of radius a.
t) that describes the
Going back to our dynamical problem, there will be a Lagrangian L0 (x, x,
motion without the constraints imposed. For example, for a single particle of mass m moving in
R3 under the influence of a potential V we have Cartesian coordinates x = (x1 , x2 , x3 ) = (x, y, z)
= 21 m(x 21 + x 22 + x 23 ) V (x1 , x2 , x3 ). Provided we can find
and this Lagrangian is L0 = L0 (x, x)
generalized coordinates q that solve the constraints via x = x(q, t), as in (2.26), the Lagrangian
for the constrained problem is simply
t) L0 (x(q, t), x(q,

t), t) .
L(q, q,
q,

(2.28)

In particular here
x i =

n
X
xi
a=1

qa

qa +

xi
,
t

i = 1, . . . , d ,

(2.29)

by the chain rule. This is precisely what we did for the simple pendulum, for example computing
the kinetic energy to be T = 12 m(x 2 + z 2 ) = 21 ml2 2 , where x = l sin , z = l cos solve the
constraint in terms of the single generalized coordinate q = . Solving the constrained problem in
this way, the Lagrange equations for (2.28) are called Lagrange equations of the second kind.
4

Those who have taken the Part A course Introduction to Manifolds may recognize the conditions we just stated.

19

Proof of Hamiltons principle for workless, holonomic constraints

In writing (2.28) we are stating that the Lagrangian for the constrained motion is obtained by
simply substituting the solution to the constraints into the Lagrangian for the unconstrained
motion. While this might seem reasonable, there is also an explanation. This is related to the
Lagrange equations of the first kind where the constraints are not solved first, but are rather
imposed via Lagrange multipliers. Specificially, let us introduce the Lagrangian
= L(x,

, t) = L0 (x, x,
t) +
L
x,

dn
X

A fA (x, t) ,

(2.30)

A=1

where the = (1 , . . . , dn ) are called Lagrange multipliers, which are treated as additional
Rt
dt then
generalized coordinates. The Lagrange equations for an extremum of the action S = 2 L
t1

read
d
dt

L0
x i

L0
xi

dn
X

A=1

fA
,
xi

i = 1, . . . , d ,
A = 1, . . . , d n .

fA (x, t) = 0 ,

(2.31)
(2.32)

Equation (2.31) is the Lagrange equation for xi , while (2.32) follows from the Lagrange equation

for A : notice L/
A = fA , while L is independent of A .
We may now compare the equations (2.31), (2.32) to what we would obtain by applying Newtons laws. Of course (2.32) simply imposes the constraints. The left hand side of (2.31) is the
unconstrained equation of motion for xi . We have already shown that for an unconstrained system
of point particles interacting via a general potential V , this is the same as Newtons second law.
Concretely, for a single particle of mass m moving in R3 in a potential V = V (x) (2.31) gives
V
m
xi +
xi

dn
X
A=1

fA
Ri ,
xi

i = 1, . . . , d .

(2.33)

The right hand side of (2.31)/(2.33) may then be interpreted as the (generalized) constraint force
R = (R1 , . . . , Rd ). An example is the tension for the simple pendulum. By saying that a constraint
force R is workless we mean that the work R x = 0, where x is any displacement tangent to
the constraint space.5 Geometrically this means that the vector R is orthogonal/normal to the
constraint space. Since fA /xi has maximal rank d n, and for fixed A the vector field fA /x is
orthogonal to the constraint space {fA = 0, A = 1, . . . , dn}, it follows that the set of dn vectors
{fA /x, A = 1, . . . , d n} form a basis for the normal space to the constraint space. Hence any
P
fA
constraint force R can be written as a linear combination R = dn
A=1 A x . The Lagrangian
(2.30) then determines the constrained dynamics, with the Lagrange multipliers determining the
constraint forces.
5
Some textbooks refer to such constraint forces as ideal, with R x the virtual work under the displacement x.
It is virtual in the sense that x isnt necessarily an actual displacement of the system obeying the equations of
motion, but rather any potential displacement obeying the constraints.

20


R= f

surface f ( x) = 0

Figure 5: The figure shows a constraint surface defined by the single equation f (x) = 0 in R3 .
A constraint force R is workless if R x = 0 for any displacement vector x tangent to the
constraint space. Since f = f /x is everywhere normal to the constraint space we may hence
write R = f , where is the Lagrange multiplier.
That the Lagrange equations of the first and second kind are equivalent is then straightforward
to see when we use coordinates xi = yi that are adapted to the constraints. Recall that in this
coordinate system the constraints are simply (2.27), with ya = qa , a = 1, . . . , n, being generalized
coordinates for the constrained motion. The constraints (2.32) then simply set yn+A = 0 for each
A = 1, . . . , d n, while the Lagrange equations (2.31) for i = 1, . . . , n read
d
dt

L
q

L
q

dn
X

A=1

fA
= 0.
q

(2.34)

t) is defined by (2.28), and the last equality fA /q = 0 holds since in this

Here L = L(q, q,
coordinate system the constraint functions fA = yn+A are independent of q. Notice that in this
argument we have implicitly used the fact that the Lagrange equations hold in one coordinate
system if and only if they hold in any other coordinate system. The remaining d n equations in
(2.31) determine the Lagrange multipliers A .
Example: Since the above discussion is a little abstract, lets return to our example of the simple
pendulum. The unconstrained coordinates are x1 = x, x2 = z, while the single constraint function
may be taken to be f (x, z) = x2 + z 2 l2 . The Lagrangian for the unconstrained motion is
L0 = 12 m(x 2 + z 2 ) mgz, which is the Lagrangian for a particle of mass m moving in (x, z) space
under gravity. The Lagrange equations of the first kind (2.31), (2.32) hence read
m
x = 2x ,
m
z + mg = 2z ,
x2 + z 2 l2 = 0 .

(2.35)

Here is the Lagrange multiplier. Polar coodinates (%, ), defined by x = % sin , z = % cos ,
are essentially adapted to the constraint. More precisely, adapted coordinates may be taken to be
21

y1 = q = , y2 = %2 l2 , so the constraint is simply y2 = 0. Imposing this of course sets % = l, and

the first two equations in (2.35) become
ml(cos sin 2 ) = 2l sin ,
ml(sin + cos 2 ) + mg = 2l cos .

(2.36)

Taking cos times the first equation plus sin times the second equation then gives the equation of
motion (2.22). Weve already seen that this indeed arises as the Lagrange equation of the second
kind for the system. On the other hand sin times the first equation in (2.36) minus cos times
the second equation instead gives
ml2 mg cos = 2l .

(2.37)

Comparing to (2.23) we thus see that the Lagrange multipler is indeed proportional to the
constraint tension T
=

1
T.
2l

(2.38)

The constraint force vector R = (2x, 2z) = (2l sin , 2l cos ) = (T sin , T cos ) appears
on the right hand side of (2.35), which is a vector of magnitude T that is normal to the constraint
space.

It is also possible to have constrained motion where the constraints cannot be written in the
form (2.25). These are called non-holonomic constraints. Examples include velocity dependent
t) = 0, and constraints defined by inequalities, say f (x, t) 0. Velocity deconstraints f (x, x,
pendent constraints arise naturally in certain problems involving the rolling of rigid bodies, since
the velocity of the point of contact is instantaneously zero. A particle confined to a box may be
defined by inequalities on the coordinates. One can solve both types of problem, but the point is
that our above treatment of holonomic constraints is not directly applicable. There is no general
theory the constraints are implemented on a case-by-case basis. However, it is perhaps worth
noting that some velocity dependent constraints can be integrated to give equivalent holonomic
constraints. For example consider the constraint A(x, t)x + B(x, t) = 0. In general this would not
be holonomic, but if there is a function f (x, t) such that A = f /x and B = f /t then by the
chain rule the constraint may be rewritten as df /dt = 0, so that f (x, t) = constant is an equivalent
holonomic constraint (where we then need to also vary the constant in f (x, t) = constant).

2.4

Noethers theorem

The Lagrangian formulation of classical mechanics leads to a very clear relationship between symmetries and conserved quantities, via Noethers theorem. In this section we give a general account
of Noethers theorem, and then discuss how certain symmetries lead to conservation of energy,
momentum and angular momentum.
22

 t) that is constant when evaluated on a solution to

A conserved quantity is a function F (q, q,
the Lagrange equations of motion, i.e.
n

0 =

d
F X

F (q(t), q(t),
t) =
+
dt
t
a=1

F
F
qa +
qa
qa
qa


,

(2.39)

where q(t) solves the Lagrange equations (2.10).

t). There
We next need to make precise what we mean by a symmetry of a Lagrangian L(q, q,
are various ways to approach this. We shall define a generator of an infinitesimal deformation to
t), which leads to a first order variation in any path q(t) in configuration
be a function = (q, q,
space Q via

q(t) = (q(t), q(t),

t) u(t) .

(2.40)

Notice here that u(t) depends on the path q(t), although for fixed path it is a function only of time
t. Unlike the variations of the action in section 2.2 we do not require u(t) to satisfy any boundary
conditions. Then we say that generates a (infinitesimal) symmetry of L if there exists a function
t) such that for all paths q(t) we have
f (q, q,



=
L(q(t) + u(t), q(t)
+  u(t),
t)
f (q(t), q(t),
t) .

dt
 =0
If one multiplies the left hand side of (2.41) by , this is the first order variation L 

(2.41)
L 
  =0

of L under q q + q. Thus (2.41) simply says that the first order variation in the Lagrangian
is a total time derivative.

Noethers theorem: Suppose we have a symmetry of a Lagrangian L in the above sense. Then
F

t)
= F (q, q,

n
X
L
a f
qa

(2.42)

a=1

is a conserved quantity.

The proof of Noethers theorem is by direct calculation. Evaluating on a solution q(t) to the
Lagrange equations we have
" n
#



n 
X
d X L
d L
L
d

a f
=
ua +
u a f (q(t), q(t),
t)
dt
qa
dt qa
qa
dt
a=1
a=1


n 
X

L
L

=
ua +
u a
L(q(t) + u(t), q(t)
+  u(t),
t)
qa
qa

 =0
a=1

= 0.

(2.43)

Here in the second equality we have used the Lagrange equation on the first term, and (2.41) for
the last term. The final equality simply follows from the chain rule.
23

The simplest case is when f = 0, so that the Lagrangian is actually invariant under the deformation to first order, L = 0. In fact this is the version of Noethers theorem usually given in
textbooks. The more general form we have presented will allow us to treat conservation of energy,
momentum and angular momentum on the same footing. Notice that in deriving the conservation
law (2.43) we only used the definition of the symmetry (2.41) for a solution q(t) to the equations
of motion (rather than a completely general path).

Conservation of energy
t+
Suppose we have a Lagrangian that is invariant under time translations, meaning that L(q, q,
t) is independent of . Hence L/t = 0 and = q generates a symmetry of L since
) = L(q, q,



(t), t)

L(q(t) +  q(t),
q +  q
=
L(q(t), q(t),
t) .
(2.44)

dt
 =0
The reasoning behind this is that the corresponding first order variation in a path arises from
+ O(2 ), so that q(t) =  q(t).

q(t + ) = q(t) +  q(t)

Thus from (2.44) we may simply take f = L
in Noethers theorem, and the conserved quantity is
H =

n
X
L
qa L .
qa

(2.45)

a=1

This is indeed what we would usually call the energy. To see this, let us suppose that L = T V
with V = V (q) and the kinetic energy being quadratic in the generalized velocities q
T

n
1 X
Tab (q) qa qb .
2

(2.46)

a,b=1

For example this is the case if one is describing a point particle where the Cartesian coordinates
xa are related to the generalized coordinates qa via the scleronomic change of variable x = x(q).
P
c xc
Then Tab = m nc=1 x
qa qb . A simple computation then gives
H = 2T L = T + V .

(2.47)

It is worth stressing that although H defined in (2.45) is always conserved when L/t = 0, it is
not always the case that H is simply the sum of kinetic and potential energies T + V , which one
would perhaps think of as the total energy. We might then have called this section conservation
of H, but what one defines as the energy is in any case a matter of convention.
We may also now define the generalized momentum p conjugate to q as
p

L
.
q

(2.48)

For example, for a point particle in R3 this is simply the momentum p = mr, where r = (x1 , x2 , x3 )
P
are Cartesian coordinates. When H = a pa qa L is regarded as a function of (q, p, t), rather
24

 t), it is called the Hamiltonian of the system, and will be the subject of section 5.
than (q, q,
Notice that for a closed system the Lagrangian should indeed be time translation invariant, as this
is one of the Galiliean symmetries discussed in sections 1.3 and 1.4. Conservation of energy follows
directly from this symmetry.

Conservation of momentum
Suppose now that we have a translational symmetry of the configuation space, meaning that = v
is a symmetry of the Lagrangian L with f = 0 and v a fixed vector. The corresponding first order
variation in a path is q(t) = v and u(t) = v. Since u = 0 the symmetry of the Lagrangian is
equivalent to
n
X
L
va = 0 .
qa

(2.49)

a=1

For example, for v = (1, 0, . . . , 0) this means that L/q1 = 0 and the coordinate q1 is an ignorable
coordinate (also sometimes called a cyclic coordinate). Noethers theorem immediately tells us that
the quantity
n
n
X
X
L
va =
p a va
qa
a=1

(2.50)

a=1

is conserved. Of course, this result may also be very simply derived by multiplying the Lagrange
equations (2.10) by va and summing over a. Going back to our example with v = (1, 0, . . . , 0) we
see that it is the conjugate momentum p1 that is conserved. Thus translational symmetry leads
to conservation of momentum.
As an example consider the free motion of a particle in spherical polar coordinates (q1 , q2 , q3 ) =
(r, , ), related to Cartesian coordinates by (2.4). Here the Lagrangian is L = 1 m(r 2 + r2 2 +
2

r2 sin2 2 ). Since L/ = 0 we see that v = (0, 0, 1) generates a symmetry of L, and the

conjugate momentum p3 = p = L/ = mr2 sin2 is conserved.
More generally for a closed system in an inertial frame we have the Galilean spatial translational
symmetry discussed in sections 1.3, 1.4. Consider a system of N particles with masses mI , position
vectors rI , and interacting through a potential V = V ({|rI rJ |}). Here translational symmetry
implies that the potential depends only on the distances between the particles {|rI rJ |}, and the
Lagrangian
N

L =

1X
mI |rI |2 V ({|rI rJ |})
2

(2.51)

I=1

is invariant under the spatial translations rI rI +  k, for any  and fixed vector k. Here the
configuration space has dimension 3N with generalized coordinates (r1 , . . . , rN ) and in the above
notation the 3N -vector is v = (k, . . . , k). Thus Noethers theorem gives us that
N
N
X
X
L
k =
pI k
r I
I=1

I=1

25

(2.52)

is conserved. Since this is true for all k we deduce that the total momentum
N
X

P =

pI

(2.53)

I=1

is conserved. Compare this to our discussion in section 1.5.

Conservation of angular momentum

The Lagrangian (2.51) also has rotational symmetry. By this we mean that L is invariant under
rI R rI for all I = 1, . . . , N , where R O(3) is any rotation matrix. This is simply because
L depends on rI rJ and r I only via their lengths, which are (by definition) invariant under
orthogonal transformations. We shall study rotations in more detail in section 4.1. Consider then
a one-parameter family of rotations R() O(3), where R(0) = 1 is the identity matrix. We may
Taylor expand this in  to write

0
n3 n2

R() = 1 +  n3
0
n1 + O( 2 ) ,
n2 n1
0

(2.54)

where n is a fixed vector describing the first order axis of rotation. Here we have simply used
the fact that dR/d | =0 is an anti-symmetric matrix, which in turn follows from R() being
orthogonal (see the discussion around equation (4.5)). Then rI R() rI is to first order in 
given by
rI

rI +  n rI .

(2.55)

Invariance of the Lagrangian L in (2.51) gives, by Noethers theorem, that

N
N
X
X
L
(n rI ) =
n (rI pI ) = n L
r I
I=1

(2.56)

I=1

is conserved. Here we have used the scalar triple product identity in the second equality, and the
definition of the total angular momentum L in (4.24) in the last equality. Since n is arbitrary
we thus deduce that systems with rotational symmetry have conserved angular momentum. In
comparing this to the more direct analysis in section 1.5 notice that the strong form of Newtons
third law is obeyed by the Lagrangian (2.51), in the sense that for fixed I the force term L/rI
on the Ith particle is a sum of terms FIJ for J 6= I, where FIJ = FJI (rI rJ ) follows since
V depends only on |rI rJ |.
These are the most important symmetries and conservation laws in classical mechanics, but
there are others. For example the Galilean boost symmetry of section 1.3 also leads to a conserved
quantity, which again is not often mentioned in textbooks. The details are left to Problem Sheet 2.
26

There are also conserved quantities that, while they can be derived from the form of Noethers
theorem we have presented, are not obviously related to any geometric symmetry of the system.
These are sometimes called hidden symmetries. The most famous example arises in the Kepler
problem you studied in first year Dynamics. The Lagrangian is L = 12 m|r|2 + r , with a constant.
This is the same potential as (1.11), and for example describes the motion of a planet around the
Sun (more on this two-body problem in the next subsection). Then one can check explicitly that
the Laplace-Runge-Lenz vector
A p L m

r
|r|

(2.57)

is conserved, where p = mr and L = r p are the momentum and angular momentum about
r = 0, respectively. This is a fascinating conserved quantity, with a very interesting history.
Writing r = (x1 , x2 , x3 ), one can show that
a (r, r ) =

1
m [2x a xb xa x b ab r r ] ,
2

(2.58)

where b = 1, 2, 3 is fixed generates a symmetry of the Lagrangian via xa (t) xa (t)+a (r(t), r (t)).
The corresponding Noether conserved quantity in (2.42) (with qa = xa ) is precisely the component
Ab of the vector A = (A1 , A2 , A3 ).
* Pauli used this vector to correctly derive the energy levels of the hydrogen atom before
Schrodinger discovered his equation for quantum mechanics! Essentially the energy levels
and their degeneracies follow from symmetry principles alone one doesnt need to solve any
differential equations. Well come back to this topic again briefly in section 5. In the meantime
there is more on the classical Laplace-Runge-Lenz vector (2.57) on Problem Sheet 1.

2.5

Examples

So far most of our discussion has been fairly abstract, sprinkled with a few very simple examples
to hopefully clarify the basic ideas. However, one really gets to grips with Lagrangian mechanics
by looking at various examples in detail. In this section well focus on the new ideas: finding
generalized coordinates, deriving the Lagrangians, identifying symmetries and conserved quantities.
Typically the Lagrange equations in any case cannot be solved in closed form (there are chaotic
examples, such as the double pendulum). We will study systems analytically near to equilibrium
in section 3, but more generally one will have to settle for a qualitative understanding and/or
numerical simulations of the equations of motion.

Pendulum on a horizontally moving support

Consider a simple pendulum of mass m and length l, attached to a pivot of mass M that is free to
move along the x-axis. As usual we denote by the angle the pendulum makes with the vertical z
direction. If X denotes the position of the pivot mass M , then the pendulum mass m coordinates
in the (x, z) plane are (x, z) = (X + l sin , l cos ) see Figure 6.
27

z
mass M

rod length l

mass m
displacement X
Figure 6: A pendulum on a horizontally moving support. The simple pendulum has mass m and
length l, and the support pivot has mass M and moves freely along the x-axis with displacement
X from the origin. The coordinates of the mass m are (x, z) = (X + l sin , l cos ).
The total kinetic energy of the system is the sum of the kinetic energies of the two masses:
T =

i
1
1
1 h
1
2 + (l sin )
2 .
M X 2 + m(x 2 + z 2 ) = M X 2 + m (X + l cos )
2
2
2
2

(2.59)

The potential energy is V = mgz = mgl cos , precisely as it is for the usual simple pendulum.
Gravity also acts on the pivot mass M , but since it is constrained to z = 0 this force is cancelled by
a reaction force that we dont need to determine in the Lagrangian approach. Thus the Lagrangian
is
L = T V =

1
1
(M + m)X 2 + ml cos X + ml2 2 + mgl cos .
2
2

(2.60)

Notice here that X is an ignorable coordinate, L/X = 0, so its conjugate momentum

pX

L
= (M + m)X + ml cos
X

(2.61)

is conserved (by Noethers theorem or by the Lagrange equation for X). The Lagrange equation
of motion for is
0 =

d
dt

i
L
d h 2
=
ml + ml cos X + ml sin X + mgl sin .

dt

(2.62)

The cross term in X cancels and this reduces to

1
+ g sin = 0 .
+ cos X
l
l

(2.63)

using pX = 0,
term in favour of terms depending only on , ,
One can then eliminate the X
thus obtaining a single second order ODE for (t):
g
(M + m sin2 ) + m sin cos 2 + (M + m) sin = 0 .
l

28

(2.64)

However, since L/t = 0 we have another conserved quantity, namely the energy E. Since T
the argument after equation (2.46) applies and the
is quadratic in the generalized velocities X,
conserved energy is hence
E = T +V =

1
1
(M + m)X 2 + ml cos X + ml2 2 mgl cos .
2
2

(2.65)

We may now substitute for the conserved generalized momentum pX in (2.61) to obtain
1
E (M + m)p2X
2

ml2
(M + m sin2 )2 mgl cos .
2(M + m)

(2.66)

We have thus reduced the problem to a first order ODE for (t). This was possible because of the
existence of the two conserved quantities E and pX . Equation (2.66) may be written as 2 = f (),
p
R
which integrates to t = d/ f (). The solution (t) may then be substituted into the conserved
momentum (2.61), which is a first order ODE for X(t) which may similarly be directly integrated.
We have thus reduced the problem to quadratures. Generally speaking this means writing the
solution to n differential equations as a complete set of n independent definite integrals (generally
using n 1 conservation laws, plus conservation of energy, with a final integration of the remaining
variable after substituting for these conserved quantities). Such systems are also called completely
integrable.

Bead on a circular wire/spherical pendulum

Next we discuss two closely related problems. Consider a small bead of mass m that slides freely on
a circular wire of radius a which rotates about a vertical diameter. Here it is natural to introduce
spherical polar coordinates (2.4) as these are adapted to the constraints. In particular the angle
measures the position of the bead with respect to the vertical axis of rotation, while is the angle
between the plane of the wire and some fixed fiducial (initial) choice of this plane see Figure 7.

Figure 7: A bead of mass m sliding freely on a circular wire of radius a that rotates about a vertical
diameter. One can use the spherical polar angles and as generalized coordinates.
29

The position of the bead in Cartesian coordinates is

r = (x, y, z) = (a sin cos , a sin sin , a cos ) .

(2.67)

One then calculates the kinetic energy

T

1
1
m|r|2 = ma2 (2 + sin2 2 ) .
2
2

(2.68)

The potential energy is (notice that with respect to the simple pendulum we have )
V

= mgz = mga cos .

(2.69)

We may then consider the following two dynamical situations:

1. The wire is forced to rotate at a constant angular velocity = .
2. The wire rotates freely about the vertical diameter.
Case 1 is an example of a time-dependent (rheonomous) holonomic constraint, since (t) is fixed
to be (t) = 0 + t, which leaves just one degree of freedom in the angle . Case 2 on the other
hand has two degrees of freedom , , and is equivalent to a spherical pendulum since the only
constraint is that the bead is distance a from the origin. The two Lagrangians are hence
=
L1 = L1 (, )
)
L2 = L2 (, , ,
=

1
1
ma2 2 + ma2 2 sin2 mga cos ,
2
2
1
2 2
ma ( + sin2 2 ) mga cos .
2

(2.70)
(2.71)

Focusing first on case 1, since L1 /t = 0 we may derive a first order equation from
E1 =

L1
1
L1 = ma2 (2 2 sin2 ) + mga cos

(2.72)

being conserved. Notice this is not the same as taking T + V given by (2.68), (2.69) and putting
= . This gives something very similar to (2.72), but with a minus sign difference (and is not
conserved). Notice also that in (2.70) the kinetic term arising from 12 ma2 sin2 2 with = is
absorbed into an effective potential
1
Veff () = mga cos ma2 2 sin2
2

(2.73)

for the dynamics of (t). We may again in principle integrate the equation E1 = 12 ma2 2 + Veff (),
thus reducing to quadratures. The equation of motion for is


d L1
L1
Veff
=

ma2 =
= ma sin (g + a 2 cos ) ,
dt

(2.74)

and there are hence equilibrium points (see section 3) at = 0 and = , so that the bead is on
p
the axis of rotation, and provided g/a also at
 g 
0 = cos1 2
.
(2.75)
a
30

The stability of these equilibria is examined in Problem Sheet 2 (using the methods of section 3).
Going back to case 2, which is the spherical pendulum, the coordinate is ignorable, so its
conjugate momentum
L2
= ma2 sin2

p =

(2.76)

is conserved. This is simply the component of angular momentum about the vertical axis. Again
L2 /t = 0 so
E2 =
=

L2 L2
+
L2


1
ma2 (2 + sin2 2 ) + mga cos
2

(2.77)

is conserved. This is simply T + V for the bead. One can eliminate using (2.76), thus again
reducing to quadratures.

The double Atwood machine

Next we consider the double Atwood machine, shown in Figure 8. This consists of a cable of length
l passing over a fixed pulley, with a mass m1 tied to one end of the cable and another moveable
pulley tied to the other end. Over this second pulley passes a cable of length l0 , with masses m2 ,
m3 tied to each end. We treat the pulleys and cables as massless, there is no friction, and moreover
we take the radii of the pulleys to be negligible compared to l, l0 .

fixed

l-x

m1
x'

l' - x'

m2
m3

Figure 8: A double Atwood machine. A cable of length l passes over a fixed pulley. Tied to one end
of the cable is a mass m1 , while at the other end is a moveable pulley. A cable of length l0 passes
over this second pulley, with masses m2 , m3 attached to its two ends. The pulleys and cables
are treated as massless, there is no friction, and we assume the radii of the pulleys are negligible
compared to l, l0 . We may use the coordinates x, x0 shown as generalized coordinates.
This system has two degrees of freedom: the vertical displacement x of mass m1 from the fixed
pulley, and the vertical displacement x0 of mass m2 from the moveable pulley. We take these as
31

our generalized coordinates. If we denote the vertical displacements of the three masses from the
fixed pulley by xi , i = 1, 2, 3, and similarly the vertical displacement of the moveable pulley from
the fixed pulley by xp , then
x1 = x ,

x2 = (l x) + x0 ,

x3 = (l x) + (l0 x0 ) ,

xp = l x .

(2.78)

The kinetic energy is hence

T

1
1
1
m1 x 2 + m2 (x 0 x)
2 + m3 (x 0 + x)
2,
2
2
2

(2.79)

while the potential energy is

V

= m1 gx m2 g(l x + x0 ) m3 g(l + l0 x x0 ) .

(2.80)

In general neither coordinate is ignorable, and Lagranges equations for x, x0 read

m1 x
+ m2 (
xx
0 ) + m3 (
x+x
0 ) = (m1 m2 m3 )g ,
m2 (
xx
0 ) + m3 (
x+x
0 ) = (m2 m3 )g .

(2.81)

A little algebra allows us to solve for the accelerations

x
=
x
0 =

m1 (m2 + m3 ) 4m2 m3
g,
m1 (m2 + m3 ) + 4m2 m3
2m1 (m2 m3 )
g.
m1 (m2 + m3 ) + 4m2 m3

(2.82)

In particular notice that if m2 = m3 then x

0 = 0 and x0 moves with constant speed. This also
follows since in this case x0 is an ignorable coordinate, and this statement is conservation of the
conjugate momentum px0 . Also notice that if m1 = m, m2 = m3 =

m
2

then x
=x
0 = 0, and both

px and px0 are conserved.

The two-body problem
As a slightly different example, let us return to the two-body problem. This is the simplest nontrivial closed system, consisting of two particles with Lagrangian
L =

1
1
m1 |r1 |2 + m2 |r2 |2 V (|r1 r2 |) .
2
2

(2.83)

As explained in section 2.4, being a closed system this enjoys conservation of energy, momentum
and angular momentum. Conservation of momentum P = m1 r 1 + m2 r 2 is equivalent to the centre
of mass
R

m 1 r1 + m 2 r2
m1 + m2

(2.84)

moving at constant velocity. Thus we may write (choosing R(0) = 0 without loss of generality)
(0)

R(t) = V0 t ,

where
32

V0 =

(0)

m1 v1 + m2 v2
m1 + m2

(2.85)

(0)

and vI

are the initial velocities of the two particles.

The interesting dynamics is described by the separation vector r = r1 r2 . One computes

m2
r,
m1 + m2
m1
= R
r.
m1 + m2

r1 = R +
r2

(2.86)

Notice we have simply changed coordinates from (r1 , r2 ) (R, r), and the Lagrangian (2.83)
becomes
L =

1
2 + 1 |r|2 V (|r|) ,
(m1 + m2 )|R|
2
2

(2.87)

where we have defined the reduced mass = m1 m2 /(m1 + m2 ). Here R is ignorable, which leads
to conservation of the conjugate momentum pR = P. We thus effectively reduce the problem to a
single particle of mass moving in an external central potential V = V (|r|):
Lreduced =

1
|r|2 V (|r|) .
2

(2.88)

Since V is rotationally invariant angular momentum is conserved, which follows from conservation
of angular momentum for the original two-body problem. Notice we have lost translational
invariance in the reduced problem (translation acts instead on the coordinate R), and hence
also Galilean invariance. The rest of the problem is now familiar from the Dynamics course.
Conservation of angular momentum L = r p implies that motion occurs in the fixed plane
orthogonal to L. Introducing polar coordinates (%, ) in this plane the Lagrangian (2.88) reads
Lreduced =

1
(% 2 + %2 2 ) V (%) ,
2

(2.89)

and since is ignorable (another consequence of angular momentum conservation) we have that
the conjugate momentum p = %2 is conserved. The equation of motion for % is

% =
where Veff (%) = V (%) +

p2
.
2%2

p2
V
V
Veff
+ % 2 =
+ 3 =
,
%
%
%
%

(2.90)

Conservation of energy for the Lagrangian Lradial = 12 % 2 Veff (%)

that gives the equation of motion (2.90) is then

E =

1 2
% + Veff (%) = constant ,
2

(2.91)

which leads to a first order ODE for %(t), and we have reduced the problem to quadratures. Notice
that it is the large number of conserved quantities (energy, momentum, angular momentum) that
have allowed us to reduce the original problem with six degrees of freedom to a single integral
equation for %(t). For V (%) = /%, corresponding to the inverse square law force (1.11), the
solutions are conic sections, as shown in the Dynamics course.

33

2.6

* More on Hamiltons principle

In section 2.2 we introduced a new fundamental principle of classical mechanics: Hamiltons principle of least action. We saw that this is equivalent to Newtons second law for point particles, but
given the elegance and universal nature of the principle it is interesting to ask why it should hold.
The whole of this subsection is understood as starred.
Historically the principle of least action took some time to develop, and Lagranges formulation
generalized earlier related ideas. However, one intuitive way to think about it is that Nature is lazy.
Assuming that energy E = T + V is conserved, then the system is free to transfer energy between
kinetic and potential so long as their sum remains constant. Minimizing the integral of T V then
amounts to minimizing the kinetic energy over the path (i.e. not moving around too much). At
the same time one is then maximizing the potential energy (i.e. maximizing the amount of energy
that may be converted into other forms).6 This is a nice picture, but as we already commented
the action is in general not minimized by the solution to the equations of motion the latter is
just a stationary point.
One can in a sense derive the principle of least action from quantum mechanics, although in
doing so this really just shifts the question elsewhere. Obviously its not appropriate to go into
a detailed discussion of this here, but the basic ideas (due to Feynman) are simple to state. In
the classical principle of least action problem we fixed the endpoints q(1) = q(t1 ), q(2) = q(t2 ) at
fixed times t1 , t2 , and sought a path q(t) which extremizes the action S subject to these boundary
conditions. This is the classical path of the system. In quantum mechanics all possible paths
between these endpoints play a role. More precisely, each path q(t) in the configuration space has
associated to it the complex phase eiS/~ , where S = S[q(t)] is the action of the path (2.6) and
~ ' 1.05 1034 Js is (the reduced) Plancks constant. Notice that both S and ~ have units of
angular momentum, so it makes sense to take the exponential of the dimensionless quantity iS/~.
One then defines the amplitude A to get from (q(1) , t1 ) to (q(2) , t2 ) to be the sum of these phases
over all paths. Given the system is observed at q(1) at time t1 , the probability of finding it at q(2)
at time t2 is given by |A|2 .
As you might imagine, defining these things precisely is not so straightforward. In particular
we have to sum over all paths q(t) connecting q(t1 ) = q(1) and q(t2 ) = q(2) . This is an
infinite-dimensional space, and such sums are called path integrals in quantum mechanics its
an integral/sum over a space of paths. The basic idea is then that at non-stationary points of S
the phases from different paths are very different (S/~ is very large for classical systems, as ~ is
so small), and their contributions to the amplitude tend to cancel. On the other hand near to
a stationary point of S the phases constructively interfere (since by definition near a stationary
point S is constant to first order). Thus the leading contribution to the amplitude A comes from
6

So the next time someone tells you to get out of bed/off the sofa, tell them that youre simply obeying Hamiltons
principle.

34

the classical path, meaning one is most likely to observe the particle near to where it should be
according to the classical equations of motion. Mathematically what we just described is called the
stationary phase approximation to oscillatory integrals, which is part of the subject of asymptotic
analysis.
Hopefully this at least gives a flavour of how the principle of least action fits into quantum
mechanics, and why we observe stationary points of S as classical paths. For those that took Part
A Quantum Mechanics: one can prove that the above formulation of quantum mechanics (and in
particular the probability we defined) is equivalent to the Schrodinger equation and Born interpretation. The interested reader is referred to the book Quantum Mechanics and Path Integrals,
by R. P. Feynman and A. R. Hibbs.

35

Small oscillations

Having set up appropriate generalized coordinates, derived the Lagrangian, and hence Lagrange
equations of motion, the next task in analysing any physical system is to identify its equilibrium
configurations. These are constant solutions q(t) = q(0) to the equations of motion. Of course
there may be no such solutions, but sometimes there can be many, and having found an equilibrium
point the next question is whether or not it is stable. That is, does a small perturbation of the
system tend to return it to the equilibrium point, or drive it away? For stable equilibria we will
show that small oscillations are described by a set of decoupled harmonic oscillators.

3.1

Equilibria and quadratic Lagrangians

Consider a general Lagrangian for a system with n degrees of freedom of the form
L = T V =

n
1 X
Tab (q)qa qb V (q) .
2

(3.1)

a,b=1

Here without loss of generality we take Tab (q) = Tba (q) to be symmetric compare to the discussion
around (2.46). The Lagrange equations of motion are
!
n
n
1 X Tbc
V
d X
Tab qb
qb qc =
,
dt
2
qa
qa

a = 1, . . . , n .

(3.2)

b,c=1

b=1

=0
By definition an equilibrium point has q(t) = q(0) Q independent of time t, so that q = q
throughout the motion. It follows that the left hand side of (3.2) is zero at an equilibrium point,
and hence defining the generalized force as
F

V
,
q

(3.3)

we see this is precisely a critical point of the potential V , or equivalently a point at which the
generalized force is zero.
In order to analyse the dynamics near to such a point we make a Taylor expansion of the
Lagrangian (3.1) around q(0) . Without loss of generality we choose coordinates so that q(0) = 0,
so as to simplify the following formulae. We then expand
T

n
1 X
Tab (0)qa qb + O(q 3 ) ,
2
a,b=1

n
1 X 2V
= V (0) +
(0)qa qb + O(q 3 ) ,
2
qa qb

(3.4)

a,b=1

in a hopefully obvious notation. In particular we assume that both q and q are small, and have
neglected all terms of order 3 and higher. Notice that the linear term in the expansion of V is zero
precisely because we are at a critical point. We may effectively drop the constant potential energy
36

term V (0) as it does not contribute to the Lagrange equations of motion. Thus the dynamics near
to the equilibrium point is governed by the quadratic Lagrangian
Lquadratic =

n
n
1 X
1 X
Tab qa qb
Vab qa qb ,
2
2
a,b=1

where Tab = Tab (0), Vab =

2V
qa qb (0)

(3.5)

a,b=1

are the components of constant n n symmetric matrices.

We write these matrices as T = (Tab ), V = (Vab ). We shall study the solutions to (3.5) in the next
subsection, but we should bear in mind that these solutions will only describe the approximate
dynamics near to equilibrium if they remain small. Otherwise the higher order terms we have
neglected will become relevant.

3.2

Normal frequencies and normal modes

The Lagrange equations of motion for the quadratic Lagrangian (3.5) are
n
X

Tab qb =

b=1

n
X

Vab qb ,

a = 1, . . . , n ,

(3.6)

b=1

which are a set of n linear homogeneous second order ODEs in q(t) with constant coefficients. We
may write these in matrix form
= V q .
Tq
Notice that in this notation the kinetic energy is T =
V =

1 T
2q V

(3.7)
1 T
T
2q

while the potential energy is

q,

as
q. Provided T is invertible we may solve (3.7) in terms of q
= T 1 V q .
q

(3.8)

In physical situations T will be positive definite, and hence invertible, because the kinetic energy
T is positive definite. We thus henceforth assume that T is positive definite.
We may now solve (3.7) using generalized eigenvectors and eigenvalues, as follows. We seek
solutions of the form
q(t) = f (t) ,

(3.9)

where is a constant (non-zero) vector. With this ansatz the equation of motion (3.7) becomes
f(t)T = f (t)V .

(3.10)

Since T is invertible, so T is non-zero, this implies that

f = f ,

where

(T V) = 0 ,

(3.11)

and is a constant. It follows that lies in the kernel of T V, and hence the determinant of
this matrix must be zero.
37

This motivates defining the characteristic equation of the system to be

det(T V) = 0 .

(3.12)

Since this is a polynomial in of degree n, we will have n solutions 1 , . . . , n .7 These are precisely
the eigenvalues of T 1 V. The corresponding eigenvectors are a , a = 1, . . . , n, satisfying
(a T V) a = 0 ,

(3.13)

and the equation of motion (3.7) is solved by q(t) = fa (t)a , where

fa = a fa .

(3.14)

Next we note that the matrix T 1 V is diagonalizable. This follows since it is similar to the matrix
T 1/2 VT 1/2 , which is symmetric as the square root T 1/2 is symmetric8 , and real symmetric
matrices are diagonalizable. There is hence a basis of Rn consisting of eigenvectors {a } of T 1 V.
Hence qa (t) = fa (t)a are n linearly independent solutions to (3.7), and the general solution is a
linear combination of these.9
An important physical fact is that the eigenvalues we just defined are real. This is a simple consequence of the fact that T and V are real and symmetric, and T is positive definite. Indeed, taking
a of a we have
Ta V a = a
Ta T a .
the dot product of (3.13) with the complex conjugate
Ta S a is real, as follows simply by showing it
But for any real symmetric matrix S the quantity
a a )
T T a = 0, and since T is positive definite we
is equal to its complex conjugate. Thus (
a

a = a .
deduce that
If > 0 then the solution to f = f is
f (t) = A cos(t + ) ,

(3.15)

where we have written = 2 , A, are constants of integration, and without loss of generality
we take > 0, A > 0. Thus the corresponding eigenmode executes simple harmonic oscillations,
with amplitude A and angular frequency . In fact we shall simply refer to as the frequency of
the mode. The modulus of the solution (3.15) is bounded by A for all times t.
On the other hand if = 0 then one obtains a linearly growing solution f (t) = A + Bt,

while < 0 leads to exponential growth f (t) = A exp( t) + B exp( t). In particular
in the latter case the mode quickly becomes large and the quadratic approximations we made in
section 3.1 break down. If the system possesses an eigenvalue < 0 we hence say that it has a
Notice that det(1 T 1 V) is the characteristic polynomial of the matrix T 1 V you defined in first year linear
algebra, and 1 , . . . , n are its roots.
8
* For completeness: since T is positive definite and symmetric, so is T 1 . It is hence diagonalizable, T 1 =
PDP T , with P O(n) orthogonal and D = diag(1 , . . . , n ) diagonal with positive eigenvalues a > 0, a = 1, . . . , n.

Then T 1/2 = PD1/2 P T , where D1/2 = diag( 1 , . . . , n ), satisfies (T 1/2 )2 = T 1 and is clearly symmetric.
9
If you recall that q(t) = (q1 (t), . . . , qn (t)) then qa (t) will have two subscripts when written out in components:
qa (t) = (qa,1 (t), . . . , qa,n (t). Dont get confused over the different meanings of these subscripts! One labels the
generalized coordinates, while the other labels different solutions to the equations of motion.
7

38

linear instability. If = 0 then this may be a translational mode (often called a flat direction
of the potential, meaning its derivative is identically zero in that direction), or else one needs to
analyse the third order, or higher order, terms in the expansions (3.4) of section 3.1 to determine
whether or not it is stable.
In particular if all the eigenvalues a > 0, a = 1, . . . , n, then we have a point of stable equilibrium.

The frequencies a = a > 0 are called the normal frequencies, with the corresponding eigenvecP
tors a called normal modes. Since {a } form a basis for Rn , we may write any q = na=1 fa a ,
and the (f1 , . . . , fn ) may then be used as a new set of generalized coordinates, called normal
coordinates. Since each solution fa (t) to the equations of motion satisfies
fa + a2 fa = 0 ,

(3.16)

we see that the coordinate transformation (q1 , . . . , qn ) (f1 , . . . , fn ) essentially diagonalizes the
original quadratic Lagrangian (3.5) into a set of n decoupled harmonic oscillators of frequencies a .

3.3

Examples

Having developed the general theory, we now apply it to some examples.

Mass-spring system
Consider three springs of natural length a and spring constant k lying in a line. One end of the first
spring is fixed at a point A, while the other end is attached to a particle of mass m. This mass is
in turn attached to one end of the second spring, with the other end attached to a second particle
of mass m. Finally this second mass is attached to one end of the third spring, with the other end
fixed at a point B. The distance between A and B is 3a. We denote the horizontal displacement
of the first mass from its equilibrium position by x, and similarly the horizontal displacement of
the second mass by y see Figure 9. The kinetic energy of the system is then
T

1
m x 2 + y 2 ,
2

(3.17)


1  2
k x + (y x)2 + y 2 .
2

(3.18)

while the potential energy is

V

Here we have used the fact that if a spring with spring constant k is displaced by an amount
from its natural length, then the corresponding potential energy is 12 k2 .
Notice that the equilibrium point x V = y V = 0 is at x = y = 0 (as one would expect), and
moreover the Lagrangian L = T V is already quadratic in x, y. We may then write the kinetic

39

fixed

fixed

B
a

Figure 9: The spring-mass system. The upper diagram shows the equilibrium configuration. In
the lower diagram we have shown the horizontal displacements x and y of the two masses from
their equilibrium positions.
and potential energy terms in matrix form as
T
V

1 ( x
2

1 (x
2

y )

m 0
0 m

y)

2k k
k 2k

x
y
!

!
,
x
y

m 2k
k
k
m 2k

(3.19)

(3.20)

Thus
T V =
The solutions to the characteristic equation
0 = det(T V) = (m k)(m 3k) ,
are hence 1 = k/m, 2 = 3k/m, giving normal frequencies
r
r
p
p
k
3k
,
2 =
.
1 =
1 =
2 =
m
m

(3.21)

(3.22)

The corresponding eigenvectors are

1 =

!
1
,
1

2 =

!
1
,
1

(3.23)

which give the normal modes. In particular the lower frequency mode 1 corresponds to the two
masses oscillating in phase, while the higher frequency mode 2 corresponds to the two masses
oscillating out of phase. Of course this is what one might have expected.

40

Double pendulum
A double pendulum consists of a simple pendulum of mass m1 and length l1 pivoted at the origin,
together with another simple pendulum of mass m2 and length l2 , pivoted at the mass m1 . The
whole system moves freely in a vertical plane under gravity. If 1 and 2 denote the angles each
pendulum makes with the vertical, then the Lagrangian is
h
i
1
1
L =
m1 l12 12 + m2 l12 12 + l22 22 + 2l1 l2 cos(1 2 )1 2
2
2
+m1 gl1 cos 1 + m2 g(l2 cos 2 + l1 cos 1 ) .

(3.24)

Deriving this is an exercise on Problem Sheet 1.

For simplicity we shall analyse the normal modes in the case where both pendula have the same
parameters, i.e. l1 = l2 = l and m1 = m2 = m. In particular the potential energy is then
V

= mgl(2 cos 1 + cos 2 ) .

(3.25)

As one would expect 1 = 2 = 0 is a critical point of V , and is a point of stable equilibrium. To

analyse this we Taylor expand V around 1 = 2 = 0:
V



= mgl 2(1 21 12 ) + (1 12 22 ) + fourth order .

(3.26)

Dropping the constant term we thus obtain the quadratic potential

Vquadratic =
=

1
mgl(212 + 22 )
2
!
!
2mgl
0
1
1 ( 1 2 )
.
2
0
mgl
2

(3.27)

Similarly expanding the kinetic term T and dropping cubic and higher order terms we find

1 2  2
ml 21 + 21 2 + 22 ,
Tquadratic =
2


!
!
2ml2 ml2
1
1 1 2
=
.
(3.28)
2
ml2 ml2
2
Thus
T V =

2ml2 2mgl
ml2
ml2
ml2 mgl

!
,

(3.29)

The solutions to the characteristic equation


 
g 2
2
,
(3.30)
0 = det(T V) = m l 2
l

are then 1 = (2 2) gl , 2 = (2 + 2) gl . Both are positive, so the equilibrium is stable, as

2 4

expected. The normal frequencies are hence

r
p
g
1 =
1 =
(2 2) ,
l

2
41

p
=
2 =

r
(2 +

g
2) ,
l

(3.31)

with corresponding normal modes

1 =

!
1

,
2

2 =

!
1

.
2

(3.32)

Again the lower frequency normal mode has the pendula swinging in phase, while for the higher
frequency mode they are out of phase.

42

Rigid body dynamics

In this section we discuss the extension of Lagrangian mechanics for point particles to cover the
dynamics of extended bodies. A rigid body may be defined as any distribution of mass for which
the distance between any two points is fixed. For example, we can consider a finite number of
particles with position vectors rI (I = 1, . . . , N ) as a rigid body, provided that we impose the
constraints |rI rJ | = cIJ = constant. One might imagine these as the positions of atoms in a
solid, with the constraints arising from inter-molecular forces. However, we will more often model
a rigid body as a continuous distribution of matter, which may be regarded as a limit of the point
particle model in which the number of particles tends to infinity. However, before getting into the
details of this, we first need to give a more precise description of rotating frames.

4.1

Rotating frames and angular velocity

both with the

Following on from our discussion in section 1.1, consider two reference frames S, S,
same origin O. Focusing first on the frame S, its coordinate axes correspond to three orthonormal
vectors ei , i = 1, 2, 3. The {ei } form a basis for R3 , and being orthonormal means they have inner
products ei ej = ij (the Kronecker delta symbol ij is defined in the appendix). We may then
write a position vector as r = x1 e1 + x2 e2 + x3 e3 = (x1 , x2 , x3 ) with respect to this basis, where
xi = r ei are the components of r in the frame S.
The frame S will similarly have an orthonormal basis {
ei }, and we may write
i =
e

3
X

Rji ej .

(4.1)

j=1

i . The position vector r

Taking the dot product with ek , this is equivalent to defining Rki = ek e
1 + x
2 + x
3 = (
then has an expansion r = x
1 e
2 e
3 e
x1 , x
2 , x
3 ) in this basis. In section 1.1 we would
i that we
have referred to this as r, as we wanted to emphasize that it is the components x
i = r e

measure in the frame S. However, mathematically r and r are the same vector, just expressed in
different bases. The components of the vector are related as
i =
x
i = r e

3
X

Rji (r ej ) =

j=1

3
X

Rji xj .

(4.2)

j=1

k :
The matrix R is orthogonal, as one sees by taking the dot product of (4.1) with e
i e
k = ik =
e

3
X

Rji Rjk .

(4.3)

j=1

Thus

1 = RT R = RRT , where 1 denotes the 3 3 identity matrix. The matrices in GL(3, R)

satisfying this condition form a subgroup called the orthogonal group O(3). Notice that R O(3)
implies that det R = 1. Using the definition of the scalar triple product as a determinant it
1 (
3 ) = (det R) e1 (e2 e3 ). Since {ei } is orthonormal
is straightforward to check that e
e2 e
43

we must have e1 (e2 e3 ) = 1, and we shall always choose a right-handed frame in which
e1 (e2 e3 ) = +1. In order that a rotation maps a right-handed frame to another right-handed
frame, we thus have det R = 1. The subgroup of O(3) satisfying this condition is called the special
orthogonal group SO(3). (The elements in O(3) that are not in SO(3) involve a reflection.)
In general the rotation matrix R = R(t) will depend on time t, so that the coordinate axes of
S are rotating relative to those of S (and vice versa). Differentiating the orthogonality relation
RRT = 1 with respect to t we obtain
T + RR T
RR

= 0,

(4.4)

T )T implies that the matrix

which since RR T = (RR

T
RR T = RR

0
3 2

= 3
0
1 ,
2 1
0

(4.5)

is anti-symmetric. In the last equality we have introduced functions i = i (t), i = 1, 2, 3,

parametrizing the non-zero entries of this matrix. Notice that we may rewrite (4.5) as
ij

3
X

ijk k ,

(4.6)

k=1

using the Levi-Civita alternating symbol ijk defined in the appendix. The angular velocity of the
P
frame S relative to S is then defined to be the vector = 3 i ei . In general this is a function
i=1

of time = (t).
Now we come to the main point of this subsection. In section 1.1 we simply defined r by
differentiating its components in the frame. However, if one applies this definition in two different
frames which are rotating relative to each other, then the resulting vector will depend on the
choice of frame. In this situation the notation r is hence too ambiguous. In order to be clear,
we will therefore define the time derivative of a vector r = x1 e1 + x2 e2 + x3 e3 with respect to the
orthonormal basis {ei } of a frame S to be the vector
Dr x 1 e1 + x 2 e2 + x 3 e3 .

(4.7)

For example, if r(t) is the position vector of a particle from the origin of S, then Dr and D2 r are
its velocity and acceleration relative to S. We may then derive
of r relative to the two frames S and S
The Coriolis formula: The time derivatives Dr and Dr
are related by

Dr
= Dr + r ,

where is the angular velocity of S relative to S.

44

(4.8)

The proof is just a direct computation. Taking the time derivative of (4.2), in vector notation we
have



x
1
x 1
x1

T
T

2 = R x 2 + R x2 .
x
x
3
x 3
x3
Multiplying on the left by R this reads



x
1
x 1
x1



R x
2 = x 2 + x2 .
x
3
x 3
x3

(4.9)

(4.10)

Given the formula (4.5), the components of the second term on the right hand side are precisely
in the basis {
the components of r in the basis {ei }.10 The components of Dr
ei } for the
frame S are x
1 , x
2 , x
3 , so the entries of the column vector on the left hand side of (4.10) are the
in basis {ei } for the frame S. This proves (4.8).
components of Dr

||

the position vector r of a point P fixed in the frame S changes

Figure 10: As seen in the frame S,
by r = r t in a small time interval t. This is a rotation of r through an angle ||t about
an axis parallel to the vector .
In the application to rigid bodies, we will be interested in applying this formula to the case
where S is an inertial frame (which we regard as the frame in which we are viewing the motion),
while the frame S rotates with the rigid body; that is, S is the rest frame of the body, which in
general will not be an inertial frame. By definition any point P in the body is then at rest in S,
so its position vector r has Dr = 0. But the time derivative of r in the inertial frame S is given by

Dr
= r ,

(4.11)

where is the angular velocity of the rest frame of the rigid body. The formula (4.11) determines
To get some geometric intuition for
the way in which the vector r changes relative to the frame S.
10

The computation is ( r)i =

P3

j=1

ij xj =

P3

j,k=1 ijk k xj

45

P3

j,k=1 ijk j xk

= ( r)i , using (4.6).

this, consider the change r in r in a small time interval t, where we ignore quadratic and higher
order terms. This is r = r t. This vector is orthogonal to both and r, and has modulus
the change in
|||r| sin t, where is the angle between and r see Figure 10. As seen in S,
the position vector r of the point P fixed in the body in the time interval t is hence obtained
by rotating r through an angle ||t about an axis parallel to . The direction of is thus the
instantaneous axis of rotation, while its magnitude || is the rate of rotation.

Example: Consider the special case in which

cos (t) sin (t) 0

R(t) = sin (t) cos (t) 0 ,

0
0
1

(4.12)

1 + sin (t) e
2 , e2 = sin (t) e
1 + cos (t) e
2 , e3 = e
3 is a rotation about
so that e1 = cos (t) e
the third axis. Then one computes = e3 .
As a final comment in this subsection, we note that behaves as you would expect a vector
to do. For example, if the angular velocity of S relative to S is , then the angular velocity of
relative to another frame
S relative to S is . Moreover, if S in turn has angular velocity
relative to S 0 . One can prove these statements either by
S 0 , then S has angular velocity +
using the Coriolis formula (4.8) to characterize , or else by introducing rotation matrices for each
change of basis and using the definition (4.5) (see Problem Sheet 2).

4.2

Motion in a non-inertial frame

In this subsection we take a brief detour from our main topic, and study Newtons equations in a
general non-inertial frame.
and S is another frame whose origin O is at
Suppose that S is an inertial frame with origin O,
If r denotes the position vector of a particle measured from O,
position vector x = x(t) from O.
then
and r is the position of the particle measured from O,
r = r + x .

(4.13)

As in the previous subsection we define the acceleration of the particle relative to S as a = D2 r,

2 r. Taking D
2 of (4.13) and using the Coriolis formula
=D
and similarly relative to S we have a
(4.8) we obtain
2 (r + x) = D(Dr

= D
a
+ r) + A
= a + (D) r + 2 Dr + ( r) + A .

(4.14)

2 x, which is the acceleration of O relative to S,

and it is simple enough
Here we have defined A = D
to check from the definition that D(b c) = (Db) c + b Dc for any two vectors b, c.
46

Since S is an inertial frame, Newtons second law holds and for a particle of mass m we may
write
m
a = F,

(4.15)

where F is the force acting. Substituting from (4.14) we thus have

ma = F m(D) r 2m Dr m ( r) mA .

(4.16)

The additional terms on the right hand side of (4.16) may be interpreted as ficticious forces:
F1 = m(D) r ,

F2 = 2m Dr ,

F3 = m ( r) ,

F4 = mA .

(4.17)

That is, these may be regarded as corrections to the force in F = ma due to the fact that the
frame S is accelerating. The force F1 is known as the Euler force, and arises from the angular
acceleration of S. The Euler force is hence zero for a frame rotating at constant angular velocity,
D = 0. The force F2 is known as the Coriolis force, and is interesting in that it depends on the
velocity Dr of the particle as measured in S (thus far we have only considered velocity-independent
forces). The Coriolis force plays a role in weather, for example being responsible for the circulation
of air around an area of low pressure, which is anti-clockwise in the northern hemisphere. The
force F3 is the centrifugal force. It lies in a plane through r and , is perpendicular to the axis
of rotation , and is directed away from the axis. This is the force you experience standing on a
roundabout. Finally, F4 is simply due to the acceleration of the origin O. For example, this force
effectively cancels the Earths gravitational field in a freely-falling frame.

4.3

Rigid body motion and the inertia tensor

We turn now to our main topic for this section, namely rigid body motion. From the outset we
A particular point in the body then has position
fix a reference inertial frame S with origin O.
We denote this point by O, and take it to be the origin of the rest
vector x = x(t) relative to O.
frame S of the body. Provided the matter distribution is not all along a line, this rest frame is
defined uniquely by the body, up to a constant orthogonal rotation of its axes and a translation
of the origin by a constant vector (relative to S). In particular this latter freedom just shifts the
choice of fixed point in the body. As we shall see shortly, the body itself provides a natural way
to fix both of these freedoms.
As in section 4.1, the rotation of the frame S relative to the inertial frame S is determined by
the time-dependent orthogonal matrix R = R(t). This matrix describes the rotation of the body
about the point O. The angular velocity of the body is defined to be the angular velocity vector
of S defined in (4.5). An important fact is that this is independent of the point O we picked.
To see this note that any other point O0 fixed in the body will have position vector r0 relative to
47

x
^

O
Figure 11: We fix a point O in the rigid body, which is taken to be the origin of the rest frame S
of the body. The frame S has angular velocity , and its origin O has position vector x relative
of an inertial frame S.
Any other point P fixed in the body has position vector r
to the origin O
relative to O.
O, and the axes {e0i } for the corresponding frame S 0 will be fixed relative to the axes {ei }. The
angular velocity 0 of S 0 relative to S is hence the same as the angular velocity of S relative

to S.
Since the rotation of the body is described entirely by the matrix R(t) SO(3), the configuration
space of rotations is hence precisely the special orthogonal group SO(3). A 3 3 matrix has 9
components and RRT = 1 imposes 6 relations, so this rotational configuration space has dimension
3. We will introduce an explicit set of generalized angular coordinates for this space in section 4.5.
Altogether the rigid body has 6 degrees of freedom: the 3 components of the position vector x of
the point O, plus these three angles describing the orientation of the axes of S.
If now r denotes the position vector of a point P fixed in the body, measured from O, then the
velocity of this point as measured in the inertial frame S is
+ x)
v = D(r
= r + vO .

(4.18)

is the velocity of the point O, and we have used the Coriolis formula (4.11).
Here vO = Dx
Momentum and angular momentum
Let us suppose that the distribution of mass in the body is defined by a density (r), so that the
mass m in a small volume V centred at r is m = (r) V . Here r is measured from O. The
total mass of the body is hence
Z
M

(r) dV .

(4.19)

We will also sometimes want to treat two-dimensional bodies, such as a flat disc, or one-dimensional
bodies such as a rigid rod. In this case one replaces by a surface density, or line density,
respectively, and integrates over the surface or curve, respectively.
48

An important quantity for the rigid body is its centre of mass G. With respect to the origin O
this has position vector
rG =

1
M

Z
r (r) dV .

(4.20)

This is the obvious generalization of (2.84) for a collection of point masses, and is an average
position weighted by mass. We will henceforth generally choose the origin O of the rest frame S
of the body to be at the centre of mass, so O = G and rG = 0. This is a natural choice, and as we
will see leads to a number of simplifications in subsequent formulae.
We may now define the total momentum P of the body relative to the inertial frame S by
similarly dividing it into small masses m = (r) V centred at r. Each element has velocity v
meaning that
given by (4.18), again relative to S,
Z
Z
(r)( r + vO ) dV .
(4.21)
v dV =
P =
R

If we now choose O = G then the first term in (4.21) is zero (it is M rG = 0) and we obtain
Z
P =
(r) vG dV = M vG .
(4.22)
R

Thus the total momentum is as if the whole mass M was concentrated at the centre of mass G.
We may examine the total angular momentum L about O in a similar way. We have
Z
Z
L =
r v dV =
(r) r ( r + vO ) dV .
R

(4.23)

If O = G then the second term is zero, and using the vector triple product we may write
Z
Z
L =
(r) r ( r) dV =
(r) [(r r) (r )r] dV .
R

(4.24)

In components of the basis {ei } for S this reads

Li =

3
X

Iij j ,

(4.25)

j=1

where we have introduced the inertia tensor I = (Iij ) with components

Z
Iij =
(r) [(r r)ij ri rj ] dV .

(4.26)

Notice that this is defined with respect to the rest frame S of the body, so is independent of time t,
and is symmetric I = I T . In general the angular momentum L = I is not in the same direction
as the angular velocity , which is what leads to some of the peculiar properties of rigid body
motion. In Cartesian coordinates r = (x, y, z) we have

y2 + z2
xy
zx
Z

I =
(r) xy
z 2 + x2
yz dx dy dz .
R
zx
yz
x2 + y 2
49

(4.27)

Note carefully the form of the terms in this matrix. The diagonal entries are called the moments
of inertia, while the off-diagonal terms are the products of inertia. However, since I is a real
symmetric matrix we may also diagonalize it via an orthogonal transformation. That is, there is
a (constant) orthogonal matrix P O(3) such that PIP T is diagonal. Thus in the new basis

I1 0 0

I = 0 I2 0 .
(4.28)
0 0 I3
The eigenvalues Ii , i = 1, 2, 3, are called the principal moments of inertia. Notice that we have
precisely used the freedom to rotate the rest frame S of the body by a constant orthogonal transformation, so that the axes ei are the eigenvectors of the inertia tensor I, called the principal axes,
with corresponding eigenvalues Ii . These eigenvalues are non-negative: if is an eigenvector of I
with eigenvalue I then
I||

= I =



(r) |r|2 ||2 (r )2 dV 0,

(4.29)

where we have used the Cauchy-Schwarz inequality.

The principal axes may often be identified by symmetry considerations. For example, an axisymmetric body by definition has rotational symmetry about a particular axis. The centre of mass
must then lie on this axis, which is one of the principal axes. One can see this directly from the
formula (4.27), where without loss of generality we take the axis of symmetry to be the e3 direction, and write the integrals in cylindrical polar coordinates (2.3). Thus x = % cos , y = % sin
and = (%, z). In particular one deduces that I1 = I2 , and all off-diagonal products of inertia
are zero. Lets now look at some more explicit examples:

Example (uniform rectangular cuboid): We will mainly focus on uniform distributions of mass,
in which the density = constant. If we take the cuboid to have side lengths 2a, 2b, 2c and mass
M , then = M/(8abc). The centre of mass is the origin of the cuboid, and we take Cartesian axes
aligned with the edges. It is then straightforward to see that the products of inertia in this basis
are zero; for example
I12

M
=
8abc

xy dx dy dz = 0 .
x=a

y=b

(4.30)

z=c

Thus the edge vectors of the cuboid are the principal axes. We next compute


Z a Z b Z c
M 1 3 a
M a2
2
x dx dy dz =
x
2b 2c =
.
8abc 3
3
x=a y=b z=c
a

(4.31)

The integrals involving y 2 and z 2 are of course similar, and we deduce that I has principal moments
of inertia I1 = 31 M (b2 + c2 ), I2 = 13 M (c2 + a2 ), I3 = 13 M (a2 + b2 ).
Example (uniform disc): As a two-dimensional example, consider a uniform disc of radius r and
mass M . Thus = M/(r2 ), and due to the rotational symmetry the centre of mass must be at
50

the origin of the disc. If we take the e3 axis perpendicular to the disc then due to the axisymmetry
we know already that I is diagonal in this basis with I1 = I2 . We compute in polar coordinates
x = % cos , y = % sin
Z
I1 =

y 2 dx dy =

M
r2

%=0

%2 sin2 % d% d =

=0

1
M r2 .
4

(4.32)

Notice here that the integrand is (x2 + y 2 x2 ) = y 2 , as the body is two-dimensional and lies in
the plane z = 0. Since I1 = I2 (which is easy enough to check explicitly) we only have to compute
Z
Z r
M
1
2
2
I3 =
(x + y ) dx dy =
2
%2 % d% = M r2 .
(4.33)
2
r
2
%=0

The inertia matrix I defined by (4.26) in general depends on the choice of origin O. The formulae
above were computed assuming this is at the centre of mass O = G. If we call this I (G) , then the
following result allows us to easily compute the inertia tensor I (P ) about another other point P
fixed in the body:

The parallel axes theorem: Denote the vector from G to P by c = (c1 , c2 , c3 ). Then
(P )

Iij

(G)

= Iij + M [(c c)ij ci cj ] .

(4.34)

Calling this a theorem is perhaps a bit generous. The position vector r0 of a point measured from
P is related to the position vector r measured from G via r0 = r c. In Cartesian coordinates we
then simply compute
Z
Z


(P )
Iij
(r0 r0 )ij ri0 rj0 dV =
[(r c) (r c)ij (ri ci )(rj cj )] dV . (4.35)
R

If one now multiplies out all the brackets and recalls that by the definition of centre of mass we
R
have R r dV = 0, then one obtains (4.34).
Notice that we may rephrase this by saying that the inertia matrix in any rest frame of the body
(with origin at P ) is the sum of two terms: the inertia matrix in a frame with origin at the centre
of mass G with axes parallel to the axes at P , plus the inertia matrix of a single particle of mass
M at c, measured about the centre of mass G.

Kinetic energy
Following the
Finally, let us compute the kinetic energy of the body in the inertial frame S.
arguments we used above for momentum we have
Z
1
|v|2 dV ,
T =
2
R
51

(4.36)

where v is given by (4.18). Thus

Z


1
T =
|vO |2 + 2vO ( r) + | r|2 dV .
2 R

(4.37)

Taking O = G as the centre of mass, the middle term is zero while the first term may be integrated
to give
T

1
1
M |vG |2 +
2
2

Z
( r) ( r) dV .

(4.38)

and
The first term is the kinetic energy due to the motion of the centre of mass relative to O,
again is as though all the mass was concentrated at the centre of mass. The second term is the
rotational kinetic energy of the body. Using the vector identity
( r) ( r) = (r ( r)) ,

(4.39)

together with the formula (4.24) for the total angular momentum L about O we may write
T

1
1
M |vG |2 + L .
2
2

(4.40)

Notice that the rotational kinetic energy may also be written in various ways as
Trot =

4.4

1
L =
2

3
3
3
1X
1 X
1X
i Li =
Iij i j =
Ii i2 .
2
2
2
i=1

i,j=1

(4.41)

i=1

Eulers equations

In this section we consider a free rigid body, meaning that no external forces act on it. From
Newtons laws we thus have that the total momentum P given by (4.22) is constant, and hence
= M v G = 0. One can deduce this from Newtons laws for point particles by considering the
P
PN
body as made up of many particles of small masses mI at positions rI , with M =
I=1 mI
being the total mass. The only forces acting are the constraint forces keeping the masses at fixed
distances apart, |rI rJ | = constant, and summing over the whole body these forces cancel, as in
equation (1.16). We thus deduce that the centre of mass velocity vG is constant. Via a Galilean
boost we may take the inertial frame S to be such that the centre of mass is at rest at the origin,
= G, which we henceforth do.
O
The remaining dynamics is then entirely in the rotation of the basis {ei } for the rest frame S
of the body. Recall from section 4.1 that
ei (t) =

3
X

j ,
Rij (t) e

(4.42)

j=1

Differentiating this with

where now {
ei } is a time-independent basis for the inertial frame S.
respect to time t we have
e i =

3
X
j=1

j =
R ij e

3
X

T )ik ek =
(RR

k=1

3
X
k=1

52

ik ek .

(4.43)

Using (4.6) we may then write this as

e i =

3
X

3
X

ikj j ek =

j,k=1

P3

j=1 j ej . In the second equality in (4.44)

P
ei ej = 3k=1 ijk ek . Choosing the basis {ei }

where recall =
right-handed, so

ej ei j = ei ,

(4.44)

j=1

we have used that the frame is

to be aligned with the principal

axes of the body, we may write the total angular momentum as

3
X

L =

Ii i ei ,

(4.45)

i=1

where recall that Ii are the principal moments of inertia. This is likewise conserved, L = 0, as
with no external forces acting there is no external torque (again, compare to the system of point
particles in section 1.5). Using (4.45) we thus compute
0 = L =

3
X

Ii ( i ei + i e i ) ,

(4.46)

i=1

which combining with (4.44) leads to the three equations

I1 1 (I2 I3 )2 3 = 0 ,
I2 2 (I3 I1 )3 1 = 0 ,
I3 3 (I1 I2 )1 2 = 0 .

(4.47)

This is a set of coupled first order ODEs for the three functions i (t), i = 1, 2, 3, which determine
how the rigid body rotates. The system also depends on the constants Ii , and collectively the
equations are known as Eulers equations. If we include an external torque then the components
i simply appear on the right hand side of (4.47) via L = .
Let us examine (4.47) further. If we multiply the ith equation by i and then sum over i we
find the resulting equation is equivalent to
2T

3
X

Ii i2

(4.48)

i=1

being constant. Of course we immediately recognize T as the rotational kinetic energy (4.41),
which we expect to be conserved for this closed system. Similarly multiplying the ith equation by
Ii i and summing we find
2

LL = L

3
X

Ii2 i2

(4.49)

i=1

is constant. Here L =

P3

i=1 Ii i ei

is the total angular momentum (4.24). Again, on general

grounds we expect this to be conserved.

53

We have thus found two conserved quantities. A little algebra allows us to rearrange the equations as
2I3 T L2 = I1 (I3 I1 )12 + I2 (I3 I2 )22 ,
2I2 T L2 = I1 (I2 I1 )12 + I3 (I2 I3 )32 ,
I12 12 = (I2 I3 )2 22 32 .

(4.50)

We may thus algebraically eliminate 2 , 3 using the first two equations to obtain a first order
ODE for 1 (t) of the form
12 = F (1 ; Ii , T, L2 ) ,

(4.51)

where F is quadratic in 12 . Thus the problem is completely integrable, since we may solve t =

R
d1 / F , and then substitute the solution for 1 (t) into the remaining two equations in (4.50) to
obtain 2 (t), 3 (t) algebraically. In general the integral that arises here is called an elliptic integral
of the first kind, and cannot be written in terms of more elementary functions. However, one can
gain some qualitative understanding of the general solutions to these equations using geometric
methods. In particular notice that equations (4.48) and (4.49), when viewed as constraints on the

angular momentum Li = Ii i , say that L = (L1 , L2 , L3 ) lies on an ellipsoid with semiaxes 2T I1 ,

2T I2 , 2T I3 , and on a sphere of radius L, respectively. For futher discussion see Example 5.7
in the book by Woodhouse, or section 37 of the book by Landau and Lifschitz. We shall instead
examine some special cases:

Example (axisymmetric body): Consider an axisymmetric body (also known as a symmetric top)
with I1 = I2 . Then the third Euler equation in (4.47) implies that 3 = 0, i.e. 3 = constant. We
may then combine the first two equations into the single complex equation
d
(1 + i2 ) = i(1 + i2 ) ,
dt

where

(I3 I1 )
3 .
I1

(4.52)

The solution is hence 1 + i2 = A exp(it), where the constant amplitude A may be taken to be
real by choosing a suitable origin for the time t. Thus
(t) = (A cos t, A sin t, 3 ) .

(4.53)

Since the component 3 along the axis of symmetry e3 is constant, this shows that the vector
rotates uniformly with angular speed about the axis of symmetry, with constant modulus
p
p
|| = A2 + 32 and making a constant angle cos1 (3 / A2 + 32 ). Since the components of
the angular momentum L in this basis are Li = Ii i , the angular momentum vector L has a
similar motion. That is, L rotates uniformly about the axis of symmetry e3 , making a fixed angle
= cos1 (I3 3 /|L|) with it.
Of course this description is as viewed from the rest frame S of the body. Viewed from the
inertial frame S we know that the angular momentum vector is constant. Viewed in this frame
54

it is the axis of symmetry e3 that rotates uniformly about the constant direction of L. We shall
describe this in more detail in section 4.5.

Example: Suppose I1 < I2 , I3 = I1 + I2 and the body is set in rotation with 2 = 0 and
p
3 = (I2 I1 )/(I2 + I1 ) 1 . Then the solution to the Euler equations is 1 (t) = c1 sech(c2 t),
p
p
where c1 = 2T /I2 and c2 = (I2 I1 )/(I2 + I1 ) c1 . Verifying the details is left to Problem
Sheet 3.

Stability
Eulers equations (4.47) have three special solutions where (1 , 2 , 3 ) = (, 0, 0), (0, , 0), or
(0, 0, ), respectively, corresponding to rotation around each of the three principal axes with constant angular speed . When I1 , I2 , I3 are all different, so that without loss of generality we may
assume I1 < I2 < I3 , then the first and last of these solutions are stable, while the second is
unstable.
To see this consider the first solution, and let us perturb it by writing = ( + 1 , 2 , 3 ). Then
ignoring quadratic order terms in i , as in section 3, the Euler equations (4.47) read
I1 1 = 0 ,
I2 2 = (I3 I1 )3 ,
I3 3 = (I1 I2 )2 .

(4.54)

In particular the last two equations imply that

I2 2 = (I3 I1 ) 3 =

(I3 I1 )(I1 I2 ) 2
2 .
I3

Since I1 < I2 < I3 this takes the form 2 = 2 2 where 2 =

(I3 I1 )(I2 I1 ) 2

I2 I3

(4.55)
> 0. We thus find

simple harmonic motion in the 2 and 3 directions, and the solution is stable.
The computation for the solution = (1 , + 2 , 3 ) is very similar, only we now find that
I1 1 =

(I2 I3 )(I1 I2 ) 2
1 ,
I3

(4.56)

2 I1 ) 2
This takes the form 1 = 1 1 , where now 1 = (I3 II21)(I
< 0 and the exponential
I3

solution means this is unstable.

We leave it as an exercise to check that the final case where = (0, 0, ) is stable. One can
demonstrate this very effectively by taking any rigid body with well-separated principal moments
of inertia I1  I2  I3 , and spinning it about each principal axis. For example, we showed
earlier that a rectangular cuboid of mass M and side lengths 2a, 2b, 2c has I1 =

1
2
3 M (b

+ c2 ),

I2 = 13 M (c2 + a2 ), I3 = 13 M (a2 + b2 ). Then a  b  c implies I1  I2  I3 , and rotation about

the middle length axis is the unstable direction. Try it preferably with something that will not
break if you drop it!
55

4.5

Euler angles and SO(3)

As weve already mentioned, the motion of a rigid body may be described by three coordinates for
its centre of mass G, together with three angles which determine the orientation of the axes {ei }
A natural choice for these latter
fixed in the body relative to the axes {
ei } of the inertial frame S.
three generalized coordinates is Euler angles.
of S, S coincide, as we are only interested in
We might as well suppose that the origins O, O
the angles between coordinate axes. The axes of the two frames are then related as in (4.1), where
R SO(3). We then have

Proposition: There exist angles [0, ], , [0, 2) such that

R =

cos cos cos sin sin

cos sin cos + cos sin sin cos

cos cos sin sin cos cos sin sin + cos cos sin sin
sin cos
sin sin
cos

cos sin 0
cos 0 sin
cos sin 0

(4.57)
sin cos 0 0
1
0
sin cos 0 .
0
0
1
sin 0 cos
0
0
1

Moreover, , , are determined uniquely by R, provided that |R33 | =

6 1.
N.B. There are different conventions for Euler angles. Were using the so-called y-convention,
2 axis).
in which the second rotation we describe below is about the y-axis (more precisely, the e
The other commonly used convention is called the x-convention, where one instead rotates about
the x-axis. This amounts to the simple replacement 2 , + 2 .
Proof: There are essentially two approaches to deriving (4.57). One can either proceed algebraically, which is the method used on pages 9 and 10 of the book by Woodhouse, or else one can
proceed geometrically. We shall follow the geometric route, as it gives a better intuition for the
Euler angles.
3 . Each of these defines an orthogonal plane, namely the (e1 , e2 )
Consider the two axes e3 and e
2 ) plane, respectively. If these planes coincide then e3 =
plane and the (
e1 , e
e3 , which is the
special case |R33 | = 1 mentioned above. In this special case the two frames simply differ by an
O(2) rotation of this plane. Otherwise the intersection of the two planes defines a unique line
though the origin, called the line of nodes. We specify its direction by defining the unit vector
3 e3 .
n = e

(4.58)

2 ) plane we may write

Since n is a unit vector lying in both the (e1 , e2 ) plane and the (
e1 , e
n = sin e1 + cos e2 ,

and

56

1 + cos e
2 .
n = sin e

(4.59)

This defines uniquely the two angles , [0, 2). The angle [0, ] is simply the angle
3 . This is probably the clearest description of the Euler angles see Figure 12.
between e3 and e
^
e3

e3

^
e2

e1

^
e1

e2
Figure 12: The Euler angles , , between the two orthonormal bases {
ei } and {ei }. The angle
3 and e3 is , while the vector n = e
3 e3 along the line of nodes is orthogonal to both
between e
2 and n is denoted , while the angle between e2 and n is .
of these vectors. The angle between e
Having given a geometric description of the angles, as suggested by the second line of (4.57) we
may construct the rotation matrix R between the two frames as a sequence of three 22 rotations.
3 onto e3 , but to do this we need to
We start with the {
ei } basis. Clearly wed like to rotate e
rotate about an axis perpendicular to both vectors, i.e. the line of nodes n. Thus we first rotate
3 axis by an angle , which brings the e
2 vector into line with n. Concretely the new
about the e
orthonormal basis is
1 = cos e
1 + sin e
2 ,
e

2 = n = sin e
1 + cos e
2 ,
e

3 = e
3 . (4.60)
e

2 = n, and rotating about this through an angle brings

The line of nodes is the new basis vector e
3 = e
3 into line with e3 . The new orthonormal basis is
e
1 + sin e
2 ) sin e
3 ,
e01 = cos (cos e
1 + sin e
2 ) + cos e
3 .
e03 = sin (cos e

1 + cos e
2 ,
e02 = n = sin e
(4.61)

3 onto e3 , the last step is to ensure that in the final basis {ei } we
Now that we have rotated e
have n = sin e1 + cos e2 , which was the definition of in (4.59). Since currently n = e02 ,
we achieve this by rotating about the e03 = e3 direction by an angle . This hence explains the
sequence of three rotations on the second line of (4.57), of course read right to left as we are
P
i ei = 3j=1 Rij e
j . The case where |R33 | = 1 (equivalently sin = 0) is a coordinate
mapping e
singularity. If one wants to analyse the behaviour near here one should really change coordinates
57

(much like near the origin % = 0 in plane polar coordinates, where the angle is not defined and
one should change to Cartesian coordinates x = % cos , y = % sin ).

Since the Euler angles are generalized coordinates for the rotation group SO(3), we must be able to
write the angular velocity of S relative to S in terms of these variables. Of course here = (t),
= (t), = (t). The brute force method would be to take the rather ugly 3 3 matrix on the
first line of (4.57) and substitute it into the definition of = (1 , 2 , 3 ) in (4.5). This works, but
is is a little tedious by hand.11
There is a much nicer way to get to the result though, again using the decomposition on the
S,
S 0 and S,
second line of (4.57). Let us denote the various frames in the proof above as S,
S S 0 and S 0 S are precisely the three
respectively. Then the rotation matrices from S S,
2 2 rotation matrices

cos sin 0

sin cos 0 ,
0
0
1

cos 0 sin

1
0
0
,
sin 0 cos

cos sin 0

sin cos 0 ,
0
0
1

(4.62)

respectively. Comparing to the example in (4.12), we thus see that the angular velocity of S
relative to S 0 is e3 , the angular velocity of S 0 relative to S is n, and the angular velocity of
3 . Using the comment at the end of section 4.1, the angular velocity of S
S relative to S is e
relative to S is then simply given by the sum of these:
3
= e3 + n + e

(4.63)

= e3 + (sin e1 + cos e2 ) + ( sin cos e1 + sin sin e2 + cos e3 ) .

Thus writing =

P3

i=1 i ei ,

the components i in the frame S are

1 = sin sin cos ,
2 = cos + sin sin ,
3 = + cos .

(4.64)

Example (axisymmetric body): As an application of these results, lets return to the free axisymmetric body discussed in the previous subsection. We showed there that in the body frame
the angular momentum vector L rotated about the e3 axis at a fixed angle , with angular speed
= (I3 I1 )3 /I1 . In the inertial frame S the angular momentum L is constant, and for simplicity
3 direction. From Figure 13 we thus see that = 0 (the angle between
we take it to lie along the e
the axis of symmetry e3 and L is constant) while comparing (4.64) to the solution (4.53) we see
that = and 3 = + cos . Notice here that = is the rate of rotation of the body
11

With a computer algebra package it takes less than a minute to type in this matrix and get the formula (4.64)
for .

58

L
e3

.
=0

e1

n
e2

Figure 13: The precession of an axisymmetric body. The dotted axes are those of the fixed inertial
We take the conserved angular momentum L to lie along the third axis e
3 . The axis of
frame S.

symmetry e3 of the body makes a constant angle with L, so that = 0, and precesses around L
with rate .

about its symmetry axis e3 , which since L is fixed in S explains why L appears to rotate about
e3 with angular speed + in the body frame. We may also deduce that
=

3 +
I3 3
=
.
cos
I1 cos

(4.65)

This is the constant rate of precession of the e3 axis around L, as viewed in the inertial frame S.

4.6

The Lagrange top

Eulers equations (4.47) determine the time-dependence of the angular velocity , but in general
these are not a useful description of the dynamical evolution, especially when we consider rigid
bodies moving under gravity (in (4.47) there are no external forces acting).
Consider a rigid body with rest frame S, with the origin O = G at the centre of mass and the
axes aligned with the principal axes. We may use Cartesian coordinates (x, y, z) for the position
of an inertial frame S.
The orientation of the body may then be
of O with respect to the origin O

specified by giving the Euler angles , , of the frame S relative to the fixed inertial frame S.
Altogether x, y, z, , , are six generalized coordinates for the configuration space. The kinetic
energy of the body is given by (4.40), where the centre of mass velocity is vG = (x,
y,
z).
We may
then use (4.64) to express the rotational kinetic energy (4.41) in terms of Euler angles. Altogether
the kinetic energy of the body relative to S is then
T

1
1
M (x 2 + y 2 + z 2 ) + I1 ( sin sin cos )2
2
2
1
1
+ I2 ( cos + sin sin )2 + I3 ( + cos )2 .
2
2
59

(4.66)

Similarly, if the body is rotating about a general point P fixed both in S and fixed in the body,
then we may take the frame S to be centred at O = P , with axes aligned with the principal axes
of I (P ) , so that
T =

1
1
1
I1 ( sin sin cos )2 + I2 ( cos + sin sin )2 + I3 ( + cos )2 . (4.67)
2
2
2

Here Ii are now the principal moments of inertia about the point P . This follows from the original
expression (4.37) for the kinetic energy, where the middle term in the expression on the right hand
side is now zero since vO = 0 (rather than because O = G). The form (4.67) of the kinetic energy
is relevant for problems where the rigid body is suspended from a particular point P , but may
move freely about this point.
The main application we will address in this section is the dynamics of a rigid body moving
under gravity. An important fact is that if a body is placed in a uniform gravitational field, say of
strength g in the downward z-direction, then the body acts as if all the mass was located at the
centre of mass G. This follows since the total gravitational potential energy of the body is
Z
V =
(r) gz dV = M g zG ,
(4.68)
R

where zG is the z-coordinate of the centre of mass rG in (4.20).

z
e3

zG = l cos
l

P = fixed
Mg
Figure 14: A Lagrange top.
In general this problem is not integrable, but certain special cases are. We will study the
Lagrange top, which is an axisymmetric body where we suspend it from a point P lying on the
axis of symmetry e3 . If we denote by l the fixed distance between P and G along this axis, then
from (4.68) the potential energy is
V

= M gl cos .

(4.69)

In particular notice that and are spherical polar angles for the axis of symmetry e3 . That is,
e3 makes an angle with the upward vertical z-direction, while is the angular speed with which
60

the axis e3 rotates about this vertical. The Lagrangian for the Lagrange top is hence given by
L = T V =

1
1
I1 (2 + 2 sin2 ) + I3 ( + cos )2 M gl cos .
2
2

(4.70)

We stress again that Ii are the principal moments of inertia about P . We immediately see that
both and are ignorable coordinates, with conserved momenta
p =
p =

L
= I3 ( + cos ) = I3 3 ,

L
= I1 sin2 + p cos .

(4.71)

Notice that p is simply the angular momentum of the top about its axis of symmetry, where the
corresponding constant angular velocity 3 is called the spin. Since L/t = 0 the energy E is
conserved, where
2E = I1 2 + I1 2 sin2 + I3 32 + 2M gl cos .

(4.72)

It is then useful to introduce u = cos and rearrange the equations as

p I3 3 u
,
I1 (1 u2 )
(p I3 3 u)u
,
= 3
I1 (1 u2 )
I1 u 2 = F (u) ,
=

(4.73)

where we have introduced

F (u) (2E I3 32 2M glu)(1 u2 )

(p I3 3 u)2
.
I1

(4.74)

We have thus reduced the problem to quadratures, and as for the free Euler equations the uintegral is an elliptic integral. However, one can understand the qualitative behaviour of the
dynamics without solving the equations directly.
Suppose we set the top in motion with starting value = cos1 (u2 ) and = 0. We shall fix the
angular speed 3 > 0 and p , and suppose that uc p /I3 3 satisfies 0 < uc < 1. This latter
condition will lead to various interesting behaviour, as we shall see. Next consider the differential
equation I1 u 2 = F (u), where F (u) is given by (4.74). The constant E in this expression is fixed
by the initial condition
0 = F (u2 ) = (2E I3 32 2M glu2 )(1 u22 )

(p I3 3 u2 )2
.
I1

(4.75)

The function F (u) is a cubic with positive coefficient of u3 , and we also note that F (1) =
(p I3 3 )2 /I1 < 0. The physical range of interest is where 1 u = cos 1 and F (u) 0.
In Figure 15 we have sketched the graph of F (u) as we vary u2 , with all other parameters kept
fixed. Notice that F (u) has two roots u1 u2 (1, 1) in the physical range.
61

F(u)

uc
D
C

-1

B
u0
A

u1

u2

Figure 15: Sketch of the cubic function F (u) for various values of initial condition u2 (where
F (u2 ) = 0). The physical region is where u = cos [1, 1] and F (u) 0. The motion of
u(t) is constrained to lie between two roots u1 u2 of F (u). As we vary u2 the endpoints
F (1) = (p I3 3 )2 /I1 < 0 of the curves are fixed. We have also noted the critical value
uc p /I3 3 (0, 1). There are then 4 distinct behaviours, corresponding to the 4 curves A, B,
C and D.
Notice that to the left of the line u = uc , meaning 1 < u < uc , we see from (4.73) that > 0 is
positive. On the other hand to the right of u = uc , meaning uc < u < 1, we instead have < 0 is
negative. This leads to 4 distinct behaviours for the motion of the top. Notice that in this motion
the axis of symmetry e3 traces out a curve on the unit sphere, with spherical polar coordinates
(, ), centred at P :
(A) Here there is a critical value u0 = u1 = u2 which is a root of both F (u) and F 0 (u). The graph
of F (u) hence touches the u-axis at u = u0 , and the trajectory is simply u = u0 = constant.
The axis thus steadily precesses about the vertical with and > 0 both constant. This
traces out a horizontal circle on the sphere with = 0 = cos1 (u0 ).
(B) Here u2 satisfies u0 < u2 < uc , and u(t) [u1 , u2 ] oscillates between the two roots u1 < u2 of
F (u) on either side of u0 . But in this motion we always have > 0. The axis of symmetry
hence rotates anticlockwise about the vertical (viewed from above the north pole = 0), but
at the same time its angle with the vertical oscillates between the two values corresponding
to the roots of F . This oscillation in is called nutation.
(C) There is a limiting case where u2 = uc . Here = 0 when reaches its minimum, corresponding
to u = uc . While this case might look fine-tuned, its also exactly what happens when you
simply let go of the rigid body from rest, i.e. both = 0 and = 0 initially (of course it can
still spin about its axis, = 3 in this case).
62

(D) Perhaps most interesting is the final case when uc < u2 < 1. In this case is negative for
appropriately small range of , and the curve traced out on the sphere loops back on itself.

(A)

(B)
2
0

P
1

(C)

(D)
2

Figure 16: The distinct behaviours of the Lagrange top corresponding to the curves A, B, C and
D in Figure 15. In each case weve shown the curve traced out by the axis of symmetry on the
unit sphere centred at P .

63

Hamiltonian mechanics

We now leave behind the world of pendulums, springs, pulleys and tops, and develop the Hamiltonian formulation of classical mechanics. This is equivalent to the Lagrangian formulation but
provides a different perspective, giving yet more insight into the structure of the theory. In particular the geometry underlying Hamiltonian mechanics is distinct from that in Lagrangian mechanics.
Mathematically this leads on to the more abstract 20th century subjects of (for example) symplectic
geometry and integrable systems, both still very active areas of research. On the other hand, from
a physics viewpoint Hamiltonian mechanics was very important in the development of quantum
mechanics. For those who have studied the latter topic some of the structures we shall encounter
may look familiar.

5.1

The Legendre transformation

In the Lagrangian formulation of classical mechanics we have the Lagrangian function L(q, v, t),
and the principle of least action leads to Lagranges equations
 
d L
L
dq

= 0,
= v.
dt v
q
dt

(5.1)

Here q = (q1 , . . . , qn ) are generalized coordinates, and we have written the second order Lagrange
equations in a first order form by writing v = q for the generalized velocity.
=
The form of Lagranges equations is invariant under coordinate transformations q q
(q, t), as demonstrated in equation (2.13). Notice in particular that the generalized velocity
q
P
qa
v = (v1 , . . . , vn ) correspondingly transforms as va va = va (q, v, t) = nb=1 qqab vb + t
. However,
=
the form of the equations of motion is not preserved by more general transformations with q
(q, v, t), v
=v
(q, v, t). This is related to the fact that the coordinates q and velocities v = q
q
are treated quite differently in the Lagrangian approach.
The Hamiltonian formulation rewrites the dynamical equations so that the variables appear on
a more equal footing. In fact the form of (5.1) suggests we might achieve this by introducing the
generalized momenta
p

L
.
v

(5.2)

Here p = p(q, v, t), and we saw this definition before in equation (2.48). The Lagrange equations
(5.1) are then
p =

L
,
q

q = v .

(5.3)

We arent quite there yet though, because now we have q, v and p. What wed like to do is
eliminate v in favour of p. This involves regarding L as a function of q and the partial derivative
L/v (and time t). Mathematically the transformation that does this is called the Legendre
transform.
64

The Legendre transformation

For simplicity we focus first on a function of a single variable f (x), so that f : R R. We also
introduce its derivative
s(x)

df
= x f .
dx

(5.4)

Here weve used the partial derivative notation, even though we only have one variable, as this
will generalize below. Wed like to regard f as a function of its derivative s = x f , rather than as
a function of x. In particular this involves inverting the map s : R R. From first year Analysis
we know a sufficient condition to be able to do this: if x s > 0 in some interval [a, b] R then s is
monotonic increasing on [a, b] and will have a differentiable inverse that we call x(s). (Of course
we could replace this by x s < 0 so that s is monotonic decreasing, but it is the case x s > 0 that
arises in classical mechanics, as we shall see shortly.)
We then define the Legendre transform of f to be the function
g(s) s x(s) f (x(s)) .

(5.5)

There are various ways to understand why this is a natural definition. First notice that
s g =

dx df
dx
dg
= x(s) + s

(x(s))
= x(s) .
ds
ds
dx
ds

(5.6)

Thus we have symmetry between s(x) = x f and x(s) = s g. Another important fact is that the
Legendre transform is its own inverse. That is, if we perform another Legendre transformation on
g(s), we simply recover f (x). We may also more informally rewrite (5.5) as
f +g = sx ,

(5.7)

where this is understood to mean either f (x(s)) + g(s) = s x(s), or the equivalent f (x) + g(s(x)) =
s(x) x. Equation (5.7) then has a simple geometric interpretation, shown in Figure 17.
Example: Taking f (x) = x /, one finds that the Legendre transform is g(s) = s /, where
1

= 1 (, > 1).

The multivariable case is similar. Now f = f (x1 , . . . , xn ), and the dual variables are s = x f ,
where s = (s1 , . . . , sn ). Then s = s(x1 , . . . , xn ) will be invertible in a neighbourhood of the point
(x1 , . . . , xn ) Rn provided the Jacobian determinant det (xa xb f ) is non-zero (this is the inverse
function theorem again). In particular this is true if the n n symmetric matrix (xa xb f ), called
the Hessian of f , is positive definite. There is then an inverse x = x(s1 , . . . , sn ) and the Legendre
transform of f = f (x) is
g(s) =

n
X

sa xa (s) f (x(s)) .

a=1

65

(5.8)

graph of f
slope s = s(x)

f(x)
x.s

g(s)
x

Figure 17: The Legendre transform g(s) of the function f (x).

Using the chain rule we compute
s g = x(s) ,

(5.9)

analogously to (5.6).

We may now go back to our classical mechanics problem. We define the Hamiltonian function
H = H(q, p, t) to be the Legendre transform of the Lagrangian with respect to the velocities
Thus
v = q.
H(q, p, t)

n
X
a=1




t) 
pa qa L(q, q,


(5.10)

q(q,p,t)

q=

Comparing to (5.8) we are taking f = L (regarded as a function of v = q the dependence on q

and t goes along for the ride), x = q and hence s = p = q L. In particular in regarding the right

hand side as a function of p, we must invert q = q(q,

p, t). In classical mechanics notice that the
n n symmetric matrix qa qb L is usually positive definite. For example for the Lagrangian (3.1)
this is simply the kinetic energy tensor Tab (whose positivity played a role in small perturbations of

equilibria in section 3). The inverse function theorem then guarantees we can invert q = q(q,
p, t).

5.2

Hamiltons equations

It is now straightforward to derive Hamiltons equations:

p =

H
,
q

q =

H
.
p

The cleanest derivation uses the notation of differentials. We compute


n 
n
X
X
L
L
L
L
dL =
dqa +
dqa +
dt =
(pa dqa + pa dqa ) +
dt ,
qa
qa
t
t
a=1

a=1

66

(5.11)

(5.12)

where we have used Lagranges equations in the form (5.3) for the first term. Thus we find
dH =

n
X

(qa dpa + pa dqa ) dL =

a=1

n
X

(qa dpa pa dqa )

a=1

L
dt .
t

(5.13)

But of course also

dH =

n 
X
H
a=1

H
dqa +
dpa
qa
pa


+

H
dt .
t

(5.14)

Comparing these last two equations we thus deduce (5.11), together with the relation L/t =
H/t.
As an alternative derivation, notice that the second equation in (5.11) is precisely (5.9), where
we identify the Legendre transformed function g = H, and the variables x = q and s = p = q L.
Thus it is this particular property of the Legendre transform that has led to such a simple form for
Hamiltons equations. The first equation in (5.11) is then simply the Lagrange equation written
in the Hamiltonian variables.
Hamiltons equations (5.11) are a set of 2n coupled first order differential equations for the
two sets of functions q(t), p(t). They are equivalent to Lagranges equations, but are much more
symmetric in particular it is the property (5.9) of the Legendre transform that leads to this
symmetry. The coordinates q1 , . . . , qn , p1 , . . . , pn are often referred to as canonical coordinates,
and Hamiltons equations (5.11) are sometimes also called the canonical equations.
Example: Consider a particle of mass m moving in R3 under the influence of a potential V (r).
The Lagrangian is hence
L = L(r, r ) =

1
m|r|2 V (r) .
2

(5.15)

The (generalized) momentum conjugate to r is

p =

L
= mr ,
r

(5.16)

which is the usual linear momentum of the particle. It is particularly trivial to invert this as
r = p/m, so that the Hamiltonian is
H = H(r, p) = p r L =

|p|2
+ V (r) .
2m

(5.17)

Of course this is H = T + V , where the kinetic energy is T = |p|2 /2m. Hamiltons equations (5.11)
read
p =

V
,
r

r =

p
,
m

(5.18)

which we recognize as Newtons second law, together with the standard relation between velocity
and momentum.
67

Example: More generally we might consider Lagrangians of the form

n
n
X
1 X
t) =
L = L(q, q,
Tab (q, t)qa qb +
Aa (q, t)qa V (q, t) ,
2

(5.19)

a=1

a,b=1

where Tab = Tba is symmetric. This is a general quadratic function of the generalized velocities q.
We compute
n
X
L
=
Tab qb + Aa ,
qa

pa =

(5.20)

b=1

and the Hamiltonian is hence

H =

n
X

pa qa L =

a=1

1
2

n
X

n
1 X
Tab qa qb + V
2
a,b=1

(T 1 )ab (pa Aa )(pb Ab ) + V ,

(5.21)

a,b=1

where (T 1 )ab is the matrix inverse of Tab . Notice that the last line of (5.21) is a function of q, p, t,
as it should be.

Phase space
The canonical coordinates q1 , . . . , qn , p1 , . . . , pn parametrize the 2n-dimensional phase space P of
the system. Notice this has twice the dimension of the configuration space Q, which is parametrized
by q1 , . . . , qn . For example, a single particle moving in Rn has phase space P = Rn Rn = R2n ,
with the points labelled by (r, p). In particular a particle moving in one dimension along the x-axis
has phase space R R = R2 , with coordinates (x, p). A more interesting example is the simple
pendulum we studied in section 2. Recall here that the configuration space is Q = S 1 , parametrized
by the angle the pendulum makes with the vertical. Then [, ), with = identified
with = . On the other hand the conjugate momentum p = ml2 can take any real value, so the
phase space is P = S 1 R.
One can think of Hamiltons equations (5.11) as first order equations for a trajectory (t) =
(q1 (t), . . . , qn (t), p1 (t), . . . , pn (t)) in P. If we pick a point (q(0) , p(0) ) P, then general results from
the theory of differential equations guarantee that under suitable conditions on the Hamiltonian
function H we will have a unique solution for the path (t) = (q(t), p(t)) satisfying Hamiltons
equations with initial condition (0) = (q(0) , p(0) ). The set of all trajectories {(t)} is called the
Hamiltonian flow on phase space.
In order to make this more concrete, and to give some examples, lets focus on the case of
dimension n = 1. The phase space is then two-dimensional, parametrized by (q, p). In Figure 18
we have shown parts of the phase spaces and Hamiltonian flow trajectories for 4 systems, where
the spatial variable q is plotted horizontally and the momentum p vertically.
68

(a)

(b)
4

4
2

2
2

4
6

(c)

(d)
4

2
2

4
4

Figure 18: Hamiltonian flows on the phase spaces of a number of one-dimensional systems: (a)
a free particle, (b) the harmonic oscillator, (c) the inverted harmonic oscillator, (d) the simple
pendulum.
(a) For a free particle of mass m the Hamiltonian is simply H =

p2
2m .

If the particle moves on the

x-axis then we may use q = x as a generalized coordinate, and the Hamiltonian flow equations
1
are then (x(t),

p(t))

= (p H, x H) = ( m
p(t), 0). In Figure 18(a) we have plotted some of the

trajectories (setting m = 1). Notice that the entire x-axis at p = 0 is fixed.

(b) Using the same spatial coordinate q = x, the Hamiltonian for the harmonic oscillator is
H =

p2
2m

+ 12 m 2 x2 , where is the frequency. The flow equations are hence (x(t),

p(t))

1
(p H, x H) = ( m
p(t), m 2 x). Some trajectories are plotted in Figure 18(b), where we

have set m = 1 = so that (x,

p)
= (p, x). The origin (x, p) = (0, 0) is a stable equilibrium
point.

69

(c) The inverted harmonic oscillator has minus the potential energy of the harmonic oscillator,
i.e. the Hamiltonian is H =

p2
2m

12 m 2 x2 . With m = 1 = the trajectories now obey

(x,
p)
= (p, x) see Figure 18(c). The origin (x, p) = (0, 0) is now an unstable equilibrium
point.
(d) Finally recall that the simple pendulum has phase space S 1 R, with the angle = q being
the generalized coordinate on S 1 . The Hamiltonian corresponding to the Lagrangian (2.24) is
2
p)
H = p 2 mgl cos , giving trajectories obeying (,
= ( 1 2 p, mgl sin ). Some trajectories
2ml

ml

are shown in Figure 18(d), with m = 1 = l = g. The -axis is horizontal, with the left and
right limits being = , = , respectively, which should be identified. Notice that the
trajectories around the origin (a stable equilibruim point) resemble those for the harmonic
oscillator (as happens near any stable equilibrium point). As |p| increases we begin to see
trajectories that wind around the circle S 1 , i.e. the pendulum swings all the way around. The
point (, p) = (, 0) (, 0) is an unstable equilibrium point.

5.3

Poisson brackets

Suppose we have a function f (q, p, t) on P R (with R the time direction). Then its evolution in
time, when evaluated on a solution (q(t), p(t)) to the Hamilton equations of motion, is computed
as
d
f
dt

n 
X
f

f
t
a=1


n
X
f H
f H
f
=
.

+
qa pa pa qa
t
=

f
qa +
pa
qa
pa

(5.22)

a=1

This motivates defining the Poisson bracket of two functions f , g on phase space as12
{f, g}

n
X
f g
f g

.
qa pa pa qa

(5.23)

a=1

Here f and g may also depend on time t, i.e. f = f (q, p, t), g = g(q, p, t) are more precisely
functions on P R. The Poisson bracket has the following properties, for any functions f, g, h on
P R:
1. anti-symmetry: {f, g} = {g, f },
2. linearity: {f + g, h} = {f, h} + {g, h}, for all constant , R,
3. Leibniz rule: {f g, h} = f {g, h} + {f, h}g,
4. Jacobi identity: {f, {g, h}} + {g, {h, f }} + {h, {f, g}} = 0.
12

Some authors define this with the opposite overall sign.

70

The anti-symmetry and linearity follow almost immediately from the definition, while the Leibniz
rule follows straightforwardly from the usual Leibniz rule of calculus. The Jacobi identity is a little
more tedious to verify you simply substitute from the definition and check that all the terms do
indeed cancel (see Problem Sheet 4). Notice also that {f, c} = 0 for any constant c.
With this definition we may immediately derive some very nice formulae. First, the dynamical
evolution of the funtion f in (5.22) is
d
f
dt

= {f, H} +

f
.
t

(5.24)

In particular when f /t = 0 we have simply f = {f, H}, and we say that the Hamiltonian
H generates the time-evolution of the function (via the Poisson bracket). Two special cases are
Hamiltons equations themselves. Namely putting f = pa or f = qa we have
H
,
qa
H
.
= {qa , H} =
pa

pa = {pa , H} =
qa

(5.25)

The Poisson bracket is clearly quite fundamental! Moreover, it is straightforward to compute the
Poisson brackets between the canonical coordinates themselves:
{qa , qb } = 0 = {pa , pb } ,

{qa , pb } = ab .

(5.26)

These are often called the fundamental Poisson brackets, or canonical Poisson brackets. Those who
have taken any quantum mechanics courses may recognize that these Poisson bracket formulae
look very similar to commutation relations. This is no accident, but well postpone any further
discussion of this to the starred subsection at the end of this section.

Conserved quantities
Lets begin by looking at the Hamiltonian itself. Using (5.24) we compute
d
H
H
H = H = {H, H} +
=
,
dt
t
t

(5.27)

where {f, f } = 0 trivially for any function f due to the anti-symmetry of the Poisson bracket.
Thus if H doesnt depend explicitly on time it is conserved. Of course this is simply the Hamiltonian version of the conservation of energy law we found in Lagrangian mechanics when L/t =
H/t = 0. In terms of the phase space picture we described in the last subsection, conservation
of H means that H is constant along the flow trajectories (t).
More generally a constant of the motion, or conserved quantity, is a function f = f (q, p, t) that
when evaluated on a solution to the Hamilton equations of motion is constant:
d
f
f = f = {f, H} +
= 0.
dt
t
71

(5.28)

In particular if f /t = 0, so that f = f (q, p), then f will be conserved if

f = {f, H} = 0 .

(5.29)

We then say that f Poisson commutes with the Hamiltonian H. An example would be an ignorable
coordinate qa (with the index a fixed). Being ignorable means that H/qa = 0, and from (5.25)
we see that the conjugate momentum pa is conserved, pa = 0.
One consequence of the Jacobi identity is that if we have two functions f, g that Poisson commute
with H, then also their commutator {f, g} Poisson commutes with H. Specifically, putting h = H
in the Jaboci identity above we have
{{f, g}, H} = {f, {g, H}} + {g, {H, f }} = 0 .

(5.30)

We may use this to prove

Poissons theorem: If f, g are two constants of the motion, then {f, g} is also conserved.
Proof: When f = f (q, p), g = g(q, p) dont explicitly depend on time t, this follows immediately
from (5.29) and (5.30). More generally we can use (5.30) to rewrite (5.24) as
d

{f, g} = {{f, g}, H} + {f, g}

dt
t

 

g
f
=
f, {g, H} +
+ {f, H} +
,g
t
t
 


df
dg
=
f,
+
,g ,
dt
dt

(5.31)

from which the result follows. Notice that we have used anti-symmetry of the Poisson bracket on
the second term, and t {f, g} = {t f, g} + {f, t g} where t = /t.
Thus from two conserved quantities we can generate a third conserved quantity! Although notice
that {f, g} might simply turn out to be a constant linear combination of f and g (including perhaps
0), and thus give nothing new. Since conserved quantities are rather special, this is what tends to
happen. If {f, g} = 0 then the functions f and g are said to be in involution.
Example (angular momentum): Consider a particle moving in R3 with angular momentum L =
r p. The Hamiltonian will thus be a function of q = r = (x1 , x2 , x3 ), and p. The components of
L are
Li =

3
X

ijk xj pk .

(5.32)

j,k=1

Then one can compute the following Poisson bracket relations:

{Li , xj } =

3
X

ijk xk ,

{Li , pj } =

k=1

3
X
k=1

72

ijk pk ,

{Li , Lj } =

3
X
k=1

ijk Lk . (5.33)

Since the proofs of these are all quite similar, lets just look at the first identity:
3
X

{Li , xj } =

ikl {xk pl , xj } =

k,l=1
3
X

3
X

ikl (xk {pl , xj } + {xk , xj }pl )

k,l=1

ikl (xk lj + 0) =

k,l=1

3
X

ijk xk .

(5.34)

k=1

In the first equality we have substituted (5.32) and used linearity of the Poisson bracket to bring
the alternating symbol outside of the bracket. In the second equality we have used the Leibniz
rule. The third equality uses the canonical brackets (5.26) and anti-symmetry, while for the final
equality ijk = ikj holds for all i, j, k.
We already know from section 2.4 that angular momentum is conserved for a particle moving
under the influence of a central potential V = V (|r|). One can also derive this using the Poisson
P
bracket formalism above (see Problem Sheet 4). Notice that the bracket {Li , Lj } = 3k=1 ijk Lk
implies that if we know that L1 and L2 are conserved, then it follows from Poissons theorem that
L3 = {L1 , L2 } is conserved.
Example (Laplace-Runge-Lenz vector): Another interesting example is the Laplace-Runge-Lenz
vector, considered at the end of section 2.4. This was defined in (2.57) to be
A p L m

r
,
|r|

(5.35)

where is a constant. Using the brackets (5.33), similar computations to (5.34) lead to the
following Poisson brackets:
{Li , Aj } =

3
X


{Ai , Aj } = 2m

ijk Ak ,

k=1

|p|2

2m
|r|

X
3

ijk Lk .

(5.36)

k=1

In particular for the central potential V = /|r| the Hamiltonian is H = |p|2 /2m /|r|, and
P
the second equation above reads {Ai , Aj } = 2mH 3k=1 ijk Lk . One can also then verify that
{Ai , H} = 0 ,

i = 1, 2, 3 ,

(5.37)

from which we deduce that A is conserved (which you checked rather more directly on Problem
Sheet 1). We then see that the constants of motion L, A and H form a closed algebra under
the Poisson bracket. In particular if we look at solutions with negative energy E = H < 0,
corresponding to the closed elliptic orbits in the Kepler problem, then we may define the new
quantities

1
2



1
A .
L
2mE

(5.38)

From the above expressions one checks these diagonalize the Poisson brackets as

{L
i , Lj } =

3
X

ijk L
k ,

k=1

73

{L+
i , Lj } = 0 .

(5.39)

The elliptical orbits of the planets thus have two sets of Poisson commuting conserved angular momenta L ! Just as the usual angular momentum L is a conserved quantity arising from rotational
symmetry under the group SO(3), the full set of conserved vectors for the elliptic orbits in the
Kepler problem is associated to a hidden action of the the rotation group SO(4) (this actually
follows immediately from (5.39), but to understand that we need to go into more group theory than
we have time for). One can analyse the parabolic and hyperbolic orbits similarly (in particular,
the hidden symmetry of the hyperbolic orbits is coincidentally the Lorentz group SO(3, 1)!).

5.4

Canonical transformations

Recall that we are free to make any choice for the generalized coordinates q (as long as they define
uniquely the configuration of the system at a fixed time). Moreover, Lagranges equations take
the same form under the coordinate transformation13
q Q(q, t) .

(5.40)

The change of variables (5.40) is sometimes called a point transformation. Since Lagranges equations are invariant, so too are Hamiltons equations (5.11). However, the form of Hamiltons
equations is invariant under a much larger class of transformations, called canonical transformations. This fact can be traced to the more symmetric way in which the coordinates and momenta
appear in the Hamiltonian formalism, and is one of the advantages of it. In particular, judicious
choices of transformation may sometimes be used to simplify the form of the Hamiltonian function,
and hence the equations of motion.
There are a number of ways to both describe and construct canonical transformations. We shall
begin with a formulation that is closest to the modern (geometric) viewpoint on the subject. We
first introduce coordinates y = (y1 , . . . , y2n ) = (q1 , . . . , qn , p1 , . . . , pn ) on the whole phase space P,
and label the indices of y by = 1, . . . , 2n. We then define the 2n 2n matrix14
!
0 1

,
1 0

(5.41)

1 is the n n identity matrix, and 0 is the n n matrix with all entries 0. Notice that
2 = 12n2n and T = . In this notation the Poisson bracket between two functions f =
where

f (y, t), g = g(y, t) on P R becomes

{f, g} =

2n
X
f
g

.
y
y

(5.42)

,=1

The transformation
q Q(q, p, t) ,

p P(q, p, t) ,

13

(5.43)

(q, t) in order to avoid a proliferation of tildes in what follows.

We will use the notation Q(q, t) rather than q
This is not to be confused with our use of the symbol in (4.5). The latter use wont appear in the rest of these
lectures.
14

74

is called a canonical transformation if it leaves the Poisson bracket invariant. More precisely, in
terms of the y coordinates the transformation (5.43) may be written
y Y(y, t) ,

(5.44)

and by invariance of the Poisson bracket we mean that

{f, g}y

2n
X
g
f

y
y

,=1

2n
X
g
f

{f, g}Y ,
Y
Y

(5.45)

,=1

should hold for all functions f, g. Since by the chain rule

f
y

2n
X
f Y
,
Y y

(5.46)

2n
X
Y
Y

.
y
y

(5.47)

=1

this is equivalent to
=

,=1

To see this one first uses (5.46) and (5.47) to compute

{f, g}y

2n
X
f
g

y
y

,=1

2n
X

f Y
g Y

Y y
Y y

,,,=1

2n
X
g
f

{f, g}Y ,
Y
Y

(5.48)

,=1

so that (5.45) holds for all f and g. Conversely, if (5.45) holds for all f and g then simply choose
f = Y , g = Y : then (5.45) reads
{Y , Y }y =

2n
X
Y
Y

= {Y , Y }Y = ,
y
y

(5.49)

,=1

which is (5.47). In matrix notation (5.47) reads

= J JT ,

where

Y
.
y

(5.50)

Here J is the Jacobian matrix for the change of coordinates (5.44). 2n 2n matrices J satisfying
= J J T form a subgroup of GL(2n, R) called the symplectic group Sp(2n, R). Thus canonical
transformations are symplectic, in that the Jacobian matrix J Sp(2n, R) holds at each point in
P R.
Notice that equation (5.49) says that a transformation is canonical if and only if it preserves
the canonical Poisson brackets. Indeed, the canonical brackets (5.26) read
{y , y } = ,
75

(5.51)

and we may further rewrite (5.49) as

{Y , Y }y = {y , y }y .

(5.52)

Notice that in the original coordinates on phase space the Jacobian is

!
Qa /qb Qa /pb
J =
,
Pa /qb Pa /pb

(5.53)

where each entry is an n n matrix (a, b = 1, . . . , n), and

J JT

{Qa , Qb } {Qa , Pb }
{Pa , Qb } {Pa , Pb }

!
,

(5.54)

where the Poisson brackets are evaluated in the original q, p coordinates. We thus again see that if
{Qa , Qb } = 0 = {Pa , Pb } and {Qa , Pb } = ab , then J is symplectic, and hence the transformation
is canonical.

Example (point transformation): Of course we expect the point transformation (5.40) to be

canonical. It is interesting to see in detail how this works. Denote the transformed Lagrangian by
t) = L(q(Q, t), q(Q,
t), t). Then the transformed momentum is by definition

L(Q,
Q,
Q,
n
n
X
X

L
L qb
L qb
=
=

qb Qa
Qa
b Qa
b=1
b=1

Pa =

n
X

b=1

pb

qb
.
Qa

(5.55)

Here notice that q = q(Q, t) means

qb =

n
X
qb
qb
Qa +
Qa
t

a=1

qb
qb
=
.
Qa
Q a

(5.56)

In particular (5.55) gives

Pa
pb

qb
.
Qa

(5.57)

If we now use the formula (5.53) for the Jacobian, and noting that Qa /pb = 0, we can directly
compute

J J

= X
n
Pa Qb

pc qc

n
X
Qa Pb

qc pc
=

{Pa , Pb }

c=1

0 1
1 0

!
= ,

(5.58)

c=1

where we use (5.57) in the second equality in (5.58). The remaining fiddly part in proving the
second equality is to show that {Pa , Pb } = 0. To see this we again use (5.57) to compute
n
X
Pa Pb
c=1

qc pc

n
X
Pa qc
.
qc Qb
c=1

76

(5.59)

Here q = q(Q, t) and P = P(q, p, t). Using (5.55) and the chain rule we may write

Pa =
qc

n
X

pd

d,e=1

2 qd Qe
.
Qa Qe qc

(5.60)

Substiting this into the right hand side of (5.59) we thus have
n
X
Pa Pb
c=1

qc pc

n
X

pd

c,d,e=1
n
X

pd

d=1

2 qd Qe qc
Qa Qe qc Qb

2 qd
.
Qa Qb

(5.61)

The right hand side of (5.61) is symmetric in a and b, which thus shows that {Pa , Pb } = 0. This
concludes the proof of (5.58), and hence that point transformations are canonical.

Example (swapping position and momentum): A more interesting canonical transformation is

P = q .

Q = p,

(5.62)

We thus swap the coordinates and momenta, with an appropriate sign. The Jacobian (5.53) for
this transformation is simply itself, and is symplectic since T = T = . Of course
one can also see the invariance of the Poisson bracket (5.23) and Hamiltons equations (5.11) very
simply. What we call position and momentum is therefore somewhat arbitrary. It is thus
better simply to talk of p and q as being canonically conjugate variables, meaning they satisfy the
canonical Poisson bracket relations (5.26).

Next we turn to Hamiltons equations (5.11). In the new notation these read
y =

2n
X

=1

H
.
y

(5.63)

From this equation and the transformation (5.44) we compute

Y =

2n
X
Y
=1

2n
X
=1

y +

2n
X
Y
Y
H
Y
=

+
t
y
y
t
,=1

H
Y
+
,
Y
t

(5.64)

where in the last line we used the symplectic condition (5.47). We have also defined the Hamiltonian
in the new coordinates as

H(Y,
t) H(y(Y, t), t) ,

(5.65)

by inverting (5.44). Similarly, Y /t means more precisely t Y (y, t) |y=y(Y,t) , so that it is a

function of (Y, t).
77

We thus see from (5.64) that for time-independent canonical transformations (so that Y/t =
0), Hamiltons equations take the same form in the transformed coordinates, with the Hamiltonian
transforming simply as a scalar (5.65) compare to the corresponding discussion of the Lagrangian
around equation (2.15). In the more general time-dependent case we can still write the transformed
equations in the form (5.63) by introducing a new Hamiltonian K = K(Y, t). Specifically, one can
show that there exists a function = (Y, t) such that

2n
X

=1

Y
.
t

(5.66)

Given this, we may define the transformed Hamiltonian as

+ ,
K H

(5.67)

so that (5.64) takes the form of Hamiltons equations

Y =

2n
X
=1

K
.
Y

(5.68)

Since the existence of satisfying (5.66) is a little technical, we have relegated it to the starred
paragraph after the next example.

Example (Galilean boost): A simple example in dimension n = 1 that demonstrates the need for
the additional term in (5.67) is a Galilean boost Q(q, t) = q + tv, where v is constant. It follows
that Q = q + v. Taking the Lagrangian for a free particle L = 1 mq2 , the conjugate momentum is
2

p = mq with Hamiltonian H =
=
L

p2 /2m.

The transformed Lagrangian is




2
1
1 2
d
1
2

= mQ
m Qv
mvQ mv t .
2
2
dt
2

(5.69)

The total derivative term doesnt contribute to the Lagrange equations of motion see the comment
after (2.17). So the free particle maps to a free particle, as it should do. We then compute the
Q = m(Q v) = p. Of course this is in accord with our general
conjugate momentum P = L/
formula (5.57) for point transformations. On the other hand the transformed Hamiltonian is
= m(Q v)Q 1 mQ 2 + mv Q 1 mv 2 ,
K = P Q L
2
2
2
1 2 1
P
+ Pv .
=
mQ mv 2 =
+ Pv = H
2
2
2m

(5.70)

) = H(p) |p=P = P 2 /2m. Thus we explicitly find = (Q, P, t) = P v. In this notation

Here H(P
(5.66) reads /Q = P/t and /P = Q/t, which are both seen to hold.

78

* The existence of the function (Y, t) satisfying (5.66) is analogous to the more familiar
fact that if F = 0 then there exists (locally) a function V = V (r) such that F = V .
That is, a vector field with zero curl is (locally) the divergence of a function. In the current
setting our space is 2n-dimensional, and the zero curl condition means

2n
2n
Y
Y
X
X

= 0,
(5.71)
Y
t
Y
t
=1

=1

holds for all , = 1, . . . , 2n. Notice that (5.71) is certainly a necessary condition for (5.66)
to hold, simply using symmetry of the mixed partial derivatives with respect to Y . We shall
assume it is also sufficient (this result goes under the general name of the Poincare Lemma).
We then check explicitly that (5.71) does indeed hold. Using the chain rule one sees that the
left hand side is the anti-symmetric part of the matrix t J J 1 = J t (J 1 ), where
J is the Jacobian matrix in (5.50). But we can also show that the latter is symmetric:
J t J 1

T

t (J 1 )

T

JT

= t J J T = t J J 1 .

(5.72)

Here in the first equality of (5.72) we have used T = , the second equality arises by
differentiating the symplectic relation (J 1 )T = J with respect to time t, and the final
equality again uses the symplectic relation (5.50). This proves (5.71).

5.5

Generating functions

We have so far seen the definition of canonical transformations, and some of their properties, but
how does one actually construct transformations (5.43) that are canonical? In this section we
describe two different approaches that both involve generating functions. The first is somewhat
more mathematical/geometrical, while the second is the more traditional physics approach.

Hamiltonian vector fields

Given any function f = f (y) on phase space, we define the following differential operator
Df

2n
X
f

.
y
y

(5.73)

,=1

This maps functions on phase space to functions on phase space in the obvious way. In fact from
(5.42) notice that if we apply Df to a function g = g(y) we simply obtain the Poisson bracket
between f and g:
Df g = {f, g} .

(5.74)

We shall refer to Df as the Hamiltonian vector field associated to f .15 One can deduce some
immediate properties:
15

This might seem like an odd name, because so far this seems to have nothing to do with either the Hamiltonian
H or vector fields. The relation to the Hamiltonian and Hamiltons equations will become clear below.
P2nThe term
vector field is often used for (or conflated with) a directional derivative, since one can write Df =
=1 V y ,
where V is a vector field on phase space.

79

 Df is linear in f : Df1 +f2 = Df1 + Df2 ,

Df g is linear in g: Df (g1 + g2 ) = Df g1 + Df g2 ,
anti-symmetry of the Poisson bracket and (5.74) implies Df g = Dg f .
Another property of Df is that it is invariant under canonical transformations; that is, for all f
and g we have
(Df )y g {f, g}y = {f, g}Y (Df )Y g .

(5.75)

Said more simply, in (5.73) it doesnt matter whether we evaluate the partial derivatives with
respect to y or Y, provided they are related by a canonical transformation. Given two Hamiltonian
vector fields Df , Dg , we may consider their commutator
[Df , Dg ] Df Dg Dg Df .

(5.76)

The composition of operators on the right hand side works in the obvious way, i.e. applied to
another function h = h(y) we have
(Df Dg )(h) = Df (Dg h) .

(5.77)

Then a calculation shows that (acting on twice differentiable functions on phase space)
[Df , Dg ] = D{f,g} .

(5.78)

That is, the commutator of Hamiltonian vector fields equals the Hamiltonian vector field of the
Poisson bracket. To prove (5.78) one simply applies both sides to an arbitrary function h = h(y):
on the left hand side there are a priori terms involving second derivatives of h, but one checks that
these cancel, and the remaining terms on both sides can be rerranged and shown to be equal we
leave this as an exercise on Problem Sheet 4. The relation (5.78) implies that if f and g Poisson
commute, then their Hamiltonian vector fields commute as differential operators.
Armed with this new technology, let us fix any function f = f (y) and consider
Y(y, s) esDf y

X
(s)n
n=0

n!

(Df )n y .

(5.79)

The operator (Df )n is understood to act as in (5.77), and we shall assume that everything is
suitably differentiable/convergent for the right hand side of (5.79) to make sense. We claim that
Y(y, s) defines a one-parameter family of canonical transformations. To see this, note first that
clearly Y(y, 0) = y, giving the identity transformation on the coordinates y. We next compute

X (s)n1

Y(y, s) =
(Df )n y = Df Y(y, s) = {Y(y, s), f (y)} .
s
(n 1)!
n=1

80

(5.80)

The last equality here is simply (5.74). Writing Y = Y(y, s) we may hence compute

 


{Y , Y } =
Y , Y + Y , Y
s
s
s
= {{Y , f }, Y } + {Y , {Y , f }}
= {{Y , Y }, f } .

(5.81)

Here we have used (5.80) in the second equality, and the Jacobi identity in the last equality. Notice
that this equation takes the same form as (5.80). One can at this point turn things around, and
regard (5.80) as a first order differential equation in s, whose solution is (5.79). We thus deduce
from (5.81) that
{Y (y, s), Y (y, s)} = esDf {Y (y, 0), Y (y, 0)} = esDf = .

(5.82)

Here the second equality follows from Y(y, 0) = y and the canonical Poisson brackets, while the
last equality follows since is constant hence Df = 0. Of course (5.82) is precisely what
we wanted to prove: Y(y, s) defines a canonical transformation for all s R.
It follows that any function f = f (y) on phase space generates an associated one-parameter
family of canonical transformations. The simplest example would be to choose f constant, but
clearly then Df = 0 and Y(y, s) = y for all s (a trivial one-parameter family). The next simplest
example is to consider f linear. Lets focus on n = 1 degree of freedom, and rewrite the Hamiltonian
vector field in terms of the canonically conjugate q and p:
Df

f
f

.
q p p q

(5.83)

Example: Let f (q, p) = p q, with , constant. Then we compute Df = q p , and

hence
Q(q, p, s) esDf q = q + s ,
P (q, p, s) esDf p = p + s .

(5.84)

Notice here that the power series terminate after the first two terms. The resulting linear shifts in
coordinates are clearly canonical transformations, albeit not very interesting ones.

Another interesting application of this formalism is to Hamiltons equations (5.25). These may
be written
y = {y, H} = DH y .

(5.85)

Here the function f = H = H(y) is the Hamiltonian, which we assume here doesnt depend on
time t. But (5.85) is just equation (5.80), with the variable s replaced by time t. We thus deduce
that
y(t) = etDH y
81

(5.86)

solves Hamiltons equations, with initial condition y(0) = y. In this sense, the Hamiltonian
vector field DH generates the solution to Hamiltons equations. An immediate corollary of our
results is that Hamiltonian flow, where we regard y = y(0) y(t) as a one-parameter family of
coordinate transformations on phase space parametrized by time t, is in fact a family of canonical
transformations!

Example: As a more interesting example, consider the one-dimensional harmonic oscillator with
Hamiltonian H(q, p) =

1 2
2 (p

+ q 2 ). The Hamiltonian vector field is DH = qp pq , and one

computes
Q(q, p, t) etDH q = q cos t + p sin t ,
P (q, p, t) etDH p = p cos t q sin t .

(5.87)

Of course these indeed solve Hamiltons equations t Q = P , t P = Q, with initial condition

Q(q, p, 0) = q, P (q, p, 0) = p.

The principle of least action

A different approach, that also involves a notion of generating functions, arises naturally if we
think about the principle of least action in the Hamiltonian formalism. Recall that the action
(2.6) is
Z

t2

t2

t)dt =
L(q, q,

S =
t1

t1

n
X

!
pa qa H

t2

dt =
t1

a=1

n
X

!
pa dqa Hdt

(5.88)

a=1

where we have written the Lagrangian in terms of the Hamiltonian. In fact Hamiltons equations
follow from extremizing (5.88) with respect to both q and p, i.e. we treat them as independent
variables, with S = S[q(t), p(t)]. Indeed, writing q(t) = u(t), p(t) = w(t), as in our derivation
of Lagranges equations we compute the first order variation
n Z
X

t2


H
H
ua dt
wa dt
S = 
wa dqa + pa dua
qa
pa
a=1 t1




n Z t2 
X
H
H
= 
wa dqa
dt ua dpa +
dt
.
pa
qa
t1


(5.89)

a=1

Here we have integrated by parts the second term on the right hand side of the first line, and
used the boundary condition u(t1 ) = u(t2 ) = 0. Hamiltons equations hence follow, much as we
deduced Lagranges equations.
Also as in our discussion of Lagranges equations (see around equation (2.17)), adding a total
time derivative to the integrand in (5.88) does not affect the equations of motion, as such a term
depends only on the boundary data and so has zero variation. If we have two coordinate systems

82

(q, p) and (Q, P) on phase space, with corresponding Hamiltonians H and K, respectively, then
our above discussion means that these will lead to the same equations of motion if
n
X

Pa dQa Kdt =

a=1

n
X

pa dqa Hdt dF1 (q, Q, t) .

(5.90)

a=1

The function F1 (q, Q, t) is called a generating function of the first kind. It generates the change
of coordinates in the sense that (5.90) implies the relations
p =

F1
,
q

P =

F1
,
Q

K = H+

F1
.
t

(5.91)

One must be very careful to read the dependences correctly in these equations. The first equation
in (5.91) determines p as a function of (q, Q, t). One then inverts this (assuming it is invertible)
to find Q as a function of (q, p, t). Similarly, the second equation in (5.91) determines P as a
function of (q, Q, t), which we may then in turn convert to a function of (q, p, t). Notice that the
final equation in (5.91) relates the two Hamiltonians, and may be compared to equation (5.67)
although in the latter equation everything was a function of the new variables (Q, P), and we were

careful to distinguish H(q, p, t) from H(Q,

P, t). In this section we are being less precise, using
the same name for a function and the function composed with a change of variable. This makes
the notation cleaner and easier to read, but one must then be particularly careful when applying
the chain rule. Notice that if the generating function is independent of time, so F1 /t = 0, then
K = H and the new Hamiltonian is obtained by simply substituting for q, p in H their values in
as in (5.65)).
terms of the new variables Q, P (so that more precisely K = H,
By construction, any two Hamiltonian systems related by the change of variables (5.91) will
lead to the same Hamilton equations of motion, and be a canonical transformation. One can
verify this directly by a brute force computation of Poisson brackets, being very careful with the
chain rule, although this is rather long and not very enlightening. A more conceptually interesting
way to think about the invariance of the Poisson brackets {f, g}p,q = {f, g}P,Q is as follows.
Since time t simply appears as a parameter in the Poisson bracket computation, we might as well
consider only time-independent functions. We may then regard the arbitrary function g appearing
in {f, g} as the Hamiltonian of a ficticious dynamical system, so that via Hamiltons equations
for this system {f, g} =

df
dt .

Since Hamiltons equations in the two coordinate systems are the

same, the time-evolution of the function f must be coordinate-independent, and we deduce that
{f, g}p,q = {f, g}P,Q .
Example: The generating function F1 =

Pn

a=1 qa Qa

generates the position and momentum swap

Q = p, P = q.
In (5.90) the generating function F1 = F1 (q, Q, t) is a function of the old and new generalized
coordinates. However, it might be more convenient to express the generating function in terms of
the old coordinate q and new momenta P. The appropriate formulae are obtained via a Legendre
83

transform. That is, we rewrite (5.90) as

dF1 =

n
X

pa dqa

a=1

n
X

Pa dQa + (K H)dt ,

(5.92)

a=1

which in turn may be written

d(F1 +

n
X

Pa Qa ) =

a=1

n
X

pa dqa +

a=1

n
X

Qa dPa + (K H)dt ,

(5.93)

a=1

The argument of the differential on the left hand side may then be regarded as F2 = F2 (q, P, t) =
P
F1 + na=1 Pa Qa . Up to an overall sign, this is a Legendre transform of F1 = F1 (q, Q, t) with
respect to Q, since F1 /Q = P. From (5.93) we obtain the Legendre-transformed relations
p =

F2
,
q

Q =

F2
,
P

K = H+

F2
.
t

(5.94)

We may similarly obtain formulae for generating functions which depend on p and Q, or p and
P, respectively, which are generating functions of the third and fourth kind, respectively.

Example: As a less trivial example, consider the generating function

Z qp
2P x2 dx .
F2 (q, P ) =

(5.95)

It follows from (5.94) that

p =

p
F2
=
2P q 2 ,
q

Q =

F2
q
= arctan p
.
P
2P q 2

(5.96)

The coordinate transformation is obtained by inverting these appropriately. For example, we may
write q, p in terms of Q, P as

q =

2P sin Q ,

p =

2P cos Q .

(5.97)

The new variables Q, P are essentially polar coordinates on the phase space P = R2 . Since F2 is
time-independent, the new Hamiltonian is simply K = H, written in terms of the new variables.
For example, the harmonic oscillator Hamiltonian H(q, p) = 12 (p2 + q 2 ) becomes simply H(Q, P ) =
P in the new variables! In particular Q is an ignorable coordinate, leading to P = 0, while

Q = H/P = 1. Of course this leads to the usual trigonometric solutions q = 2P sin(tt0 ) in the
original variables, but the point is that the transformation essentially trivialises the Hamiltonian
system.

5.6

Liouvilles theorem

The formalism and results weve developed allow us to now prove easily the following result.
Consider a region V P in phase space. Its volume is
Z
vol(V ) =
dq1 dq1 dp1 dpn .
V

84

(5.98)

Now consider a map from phase space to itself which in coordinates is described by q Q =
Q(q, p, t), p P = P(q, p, t). Under this map the region V transforms to another region V with
volume
vol(V ) =

Z
V

dQ1 dQn dP1 dPn .

(5.99)

If the map is a canonical transformation then the volume is preserved, i.e. vol(V ) = vol(V ).
This is straightforward to see using the symplectic property of canonical transformations. Writing
(y1 , . . . , y2n ) = (q1 , . . . , qn , p1 , . . . , pn ), (Y1 , . . . , Y2n ) = (Q1 , . . . , Qn , P1 , . . . , Pn ), the transformed
volume integral may be written as
Z
Z
vol(V ) =
dY1 dY2n =
| det J | dy1 dy2n ,
V

(5.100)

where J = Y /y is the Jacobian matrix. But the symplectic condition (5.50) immediately
implies that (det J )2 = 1, and thus | det J | = 1. We thus conclude that vol(V ) = vol(V ).
Recall that we may regard Hamiltonian flow as a one-parameter family of maps from phase
space to itself, parametrized by time t. Thus (q, p) = (q(0), p(0)) (Q, P) = (q(t), p(t)). If one
chooses a region V = V (0) P at time t = 0 then under Hamiltonian flow this region will change
shape to V (t) P. Since we showed in the last subsection that Hamiltonian flow is a canonical
transformation for all t, we have thus proven

Liouvilles Theorem: The volume of any region in phase space is invariant under Hamiltonian
evolution.
In other words,

d
dt V

= 0.

The physical applications of this result take us in the direction of statistical physics. So far
weve been describing the classical mechanics of a single particle, or more precisely the single
quasi-particle described at the end of section 2.1. This traces out a single curve in phase space
under Hamiltonian evolution. But we might instead want to think about a system with many such
particles, each of which is described by the same Hamiltonian for example, a gas. One is then
not interested in keeping track of the individual behaviours of the particles, but rather wants to
understand the average behaviour. In this situation we may introduce a density function (q, p, t)
on P R, where the number of particles in a small region of phase space V centred at (q, p) P
at time t is N = (q, p, t) V . The total number of particles is
Z
N =
(q, p, t) dq1 dqn dp1 dpn ,

(5.101)

which is fixed, i.e. we assume that particles are not created or destroyed. In fact the latter is a local
property. In particular if we follow the Hamiltonian evolution of a small number of particles N
centred at (q, p) = (q(0), p(0)) at time t = 0, then this will evolve to N = (q(t), p(t), t) V (t)

85

at a later time t. Since particles are not created or destroyed, N must be independent of time t.
But weve also shown that V (t) = V (0) is also independent of time. We thus conclude that
d
dt

= 0,

(5.102)

holds under Hamiltonian evolution. Using (5.24) we may rewrite this as

{, H} +

= 0,

(5.103)

which is often referred to as Liouvilles equation. Equation (5.102) says that the density of states
is constant along every trajectory in phase space.

5.7

The Hamilton-Jacobi equation

The Hamilton-Jacobi equation provides yet another approach to classical mechanics. Again, it
has its own particular advantages for example, this equation can be useful for finding conserved
quantities. On a more conceptual level, the Hamilton-Jacobi equation provides a particularly direct
link with quantum mechanics and the Schrodinger equation.
We begin by reconsidering the action S, which we discussed briefly in the Hamiltonian formalism
in section 5.5, equation (5.88). Recall that the action is extremized over all paths q(t) with fixed
boundary conditions to find the equations of motion. More precisely, lets take the fixed boundary
conditions for the paths to be q(t0 ) = q(0) at time t = t0 , and q(tf ) = qf at time t = tf . The
action for any path q(t) satisfying these boundary conditions is
Z

tf

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

S[q(t)] =

(5.104)

t0

and the classical path taken by the system extremizes this, resulting in a (we assume unique)
solution q(t) = qclassical (t). We now take this classical solution and substitute it back into the
action (5.104). Of course doing this we get a number, but that number depends on the fixed final
time tf and fixed final point qf . We may then consider varying these final boundary conditions,
which defines a function
S(qf , tf ) = S[qclassical (t)] .

(5.105)

Of course this also depends on the initial time t0 and initial point q(0) , but we keep these absolutely
fixed. Physicists usually refer to the action evaluated on the classical solution as the on-shell action
(with the words on-shell really referring to a notion in special relavity, so this is another physics
misnomer).
Now under any variation q(t) of a path we have
(Z 



tf )
n
tf
X
L
d L
L
S =

qa (t) dt +
qa (t)
.
qa dt qa
qa
t0
t0
a=1

86

(5.106)

If we vary the path so that the initial conditions are fixed, meaning q(t0 ) = 0, but at the final
point q(tf ) = qf , then when (5.106) is evaluated on the classical path solving the Lagrange
equations of motion with these boundary conditions we have simply

n
X
L 
q f .
S =
qa t=tf a

(5.107)

a=1

In terms of (5.105) this means that

S
qf

= pf ,

(5.108)

where pf = L/ q |t=tf is the final momentum of the classical path. We may similarly consider
varying the final time tf . From the fundamental theorem of calculus
dS
dtf

= L = L(qclassical (tf ), q classical (tf ), tf ) = L(qf , q f , tf ) ,

(5.109)

where q f is the final velocity. On the other hand via the chain rule
dS
dtf

a=1

a=1

S X S f
S X f f
+
q

=
+
pa qa .
a
tf
t
qaf

(5.110)

Combining these last two equations we find

S
tf

= L(qf , q f , tf )

n
X

pfa qaf = H(qf , pf , tf ) ,

(5.111)

a=1

where H(qf , pf , tf ) is the Hamiltonian evaluated on the final data.

Now that we have finished varying paths, we relabel the final data as qf q, tf t, so that
(5.105) defines a function S(q, t) on Q R. It is called Hamiltons principal function, and weve
just shown that it satisfies
S
+ H(q, p, t) = 0 ,
t

where

p =

S
.
q

In other words S(q, t) satisfies the first order partial differential equation


S
S
S
+ H q1 , . . . , q n ,
,...,
,t
= 0.
t
q1
qn

(5.112)

(5.113)

This is the Hamilton-Jacobi equation.

Since we have a single first order PDE in n + 1 independent variables, general PDE theory says
we should have n + 1 constants of integration in the general solution (called a complete integral ).
Notice that one of these constants of integration is easy to identify: S enters (5.113) only through
its partial derivatives, so if S(q, t) is a solution then so is S(q, t) + c for any constant c. We then
write the general solution to (5.113) as
S = S0 (q1 , . . . , qn , t; 1 , . . . , n ) + A ,
87

(5.114)

where 1 , . . . , n and A are the n + 1 integration constants.

The equations (5.112) may look familiar from our discussion of generating functions in section 5.5. Indeed, consider a generating function of the second kind F2 = F2 (q, P, t), where (q, p)
and (Q, P) are the old and new coordinates, respectively. As discussed around equation (5.94),
this generates the canonical transformation
p =

F2
,
q

Q =

F2
,
P

K = H+

F2
.
t

(5.115)

In particular, if we choose our generating function F2 so that the new Hamiltonian K = 0 is

= 0 = Q.
The transformed
identically zero, then in the new coordinates the dynamics is trivial: P
position and momenta are then conserved quantities, and we may write Q = , and P = ,
where = (1 , . . . , n ), = (1 , . . . , n ) are altogether 2n constants. In particular notice that
F2 = F2 (q, , t). Denoting this generating function by F2 = S, we see that (5.115) are simply the
equations (5.112), while the generating function F2 itself is simply Hamiltons principal function,
with the conserved momenta P = in the new coordinates being interpreted as the integration
constants in the general solution (5.114). Moreover, the second equation in (5.115) tells us that
Q = =

(5.116)

are also conserved quantities.

What weve done might seem miraculous, so its perhaps worth taking stock. Given a Hamiltonian function we construct the Hamilton-Jacobi equation (5.113), which is a first order PDE
for S(q, t) with general solution (5.114) depending on the n + 1 integration constants 1 , . . . , n
and A. We then impose the n algebraic equations (5.116), where are constant. Having done
this we may interpret the general solution (5.114) to the Hamilton-Jacobi equation as a generating function of the second kind. This generates a canonical transformation which trivialises the
Hamiltonian K = 0 in the new coordinates, and where = P are interpreted as the conserved
momenta in the new coordinates. The constants = Q are correspondingly the conserved positions in the new coordinates. Since weve solved Hamiltons equations in the new coordinates,
via the canonical transformation generated by S we have also solved Hamiltons equations in the
original coordinates. Notice we have 2n integration constants in total, as expected.
We can also think of this slightly differently. If one manages to solve the Hamilton-Jacobi
equation to find the complete integral (5.114), essentially one has solved half of the dynamical
problem. Specifically, this determines the momenta
p =

S
q

(5.117)

as a function of q, t and the integration constants . One can then substitute this into the
remaining n Hamilton equations
q =


H 
.
p p=S/q
88

(5.118)

These are now n first order differential equations for q(t), giving rise to another n integration
constants (again, 2n in total). In this approach its less clear that we have in fact solved Hamiltons
equations. But this is easily checked from (5.117):


n
X
d S
2S
2S
pa =
=
+
qb .
dt qa
tqa
qa qb

(5.119)

b=1

On the other hand taking the partial derivative of the Hamilton-Jacobi equation (5.112) with
respect to qa gives
2S
tqa

b=1

b=1

H X H 2 S
H X 2 S

=
qb ,
qa
pb qa qb
qa
qa qb

(5.120)

where in the first equality we used (5.117), and in the second equality we used (5.118). Substituting
this into (5.119) we thus find
p =

H
.
q

(5.121)

When the Hamiltonian is independent of time, so H/t = 0, one can solve the time-dependence
of the Hamilton-Jacobi equation via
S(q, t) = W (q) Et ,

(5.122)

where E is constant and the time-independent function W on configuration space Q is sometimes

called Hamiltons characteristic function. This leads to a reduced Hamilton-Jacobi equation


W
W
,...,
H q1 , . . . , q n ,
= E,
(5.123)
q1
qn
where E is then interpreted as the energy.

Separation of variables
Of course for this approach to be of any practical use, one needs to be able to solve the HamiltonJabobi equation (5.113), or in the time-independent setting the slightly easier (5.123). Either way,
one has to solve a first order PDE. In general one could proceed via the method of characteristics, but this precisely converts the PDE back into ODEs, i.e. effectively Hamiltons equations!
However, in certain cases we may solve the PDE via separation of variables. Of course whether
or not there are separable solutions will depend both on the Hamiltonian and on the particular
coordinates being used: a Hamiltonian may separate in one coordinate system, but not in another.
Theres a whole theory here, but well content ourselves with looking at some examples.
Suppose that the coordinate q1 and derivative term q1 S appear in the Hamilton-Jacobi equation
only through the combination f (q1 , q1 S), for some function f . That is, the equation can be written
in the form
[f (q1 , q1 S), q2 , . . . , qn , q2 S, . . . , qn S, t] = 0 .
89

(5.124)

When then seek a separable solution

2 , . . . , qn , t) + S1 (q1 ) .
S(q1 , . . . , qn , t) = S(q
With this ansatz the Hamilton-Jacobi equation (5.124) becomes
 


dS1
. . . , qn S,
t = 0.
f q1 ,
, q2 , . . . , qn , q2 S,
dq1

(5.125)

(5.126)

Since the dependence on q1 is entirely in the first entry of the function , a solution must satisfy


dS1
f q1 ,
= 1 ,
(5.127)
dq1
where 1 is constant. This is a first order ODE for S1 (q1 ), which may be solved, and substituting
f = 1 back into (5.113) the remaining Hamilton-Jacobi equation has one fewer independent
variables.
In some cases one can successively solve for and eliminate all the variables in this way. In
particular in the time-independent case (5.123) this means one can solve the Hamilton-Jacobi
equation with the ansatz
S(q1 , . . . , qn , t) =

n
X

Sa (qa ) Et ,

(5.128)

a=1

where each Sa = Sa (qa ) is a function of one variable only, leading to n first order ODEs in one
variable. The Hamilton-Jacobi equation is then said to be completely separable. This results
in n integration constants 1 , . . . , n , as in (5.114). Notice that E = E(1 , . . . , n ) is not an
independent integration constant, but rather is determined by substituting the solution W =
Pn
a=1 Sa (qa ) into (5.123).
A special case of the above discussion is an ignorable coordinate, i.e. q1 doesnt appear in the
Hamiltonian, H/q1 = 0, and thus doesnt appear in the Hamilton-Jacobi equation either. This
means that f = f (S/q1 ) above, and (5.127) integrates to S1 (q1 ) = 1 q1 , giving
2 , . . . , qn , t) + 1 q1 .
S(q1 , . . . , qn , t) = S(q

(5.129)

Notice then that p1 = S/q1 = 1 is simply the corresponding conserved momenta. Similarly,
the Et term in (5.128) for a time-independent system corresponds to separation of the ignorable
coordinate t.

Example: We consider the reduced two-body problem Lagrangian

Lreduced =

1
(% 2 + %2 2 ) V (%) ,
2

describing motion in the plane with polar coordinates (%, ). The Hamiltonian is


1
p22
2
H = H(q1 , q2 , p1 , p2 ) =
p1 + 2 + V (q1 ) ,
2
q1
90

(5.130)

(5.131)

where q1 = %, q2 = . The Hamilton-Jacobi equation (5.113) hence reads





S 2
S
S 2
1
1
+
+ V (q1 ) = 0 .
+
t
2 q1
2q12 q2

(5.132)

Since H/t = 0 and H/q2 = 0 our general discussion above implies we can separate variables,
seeking a solution
S(q1 , q2 , t) = S1 (q1 ) + S2 (q2 ) Et .
Substituting this into (5.132) and rearranging slightly gives
"


2 #

1
dS
dS2 2
1
= 2q12 E V (q1 )
.
dq2
2 dq1

(5.133)

(5.134)

Since the left hand side depends only on q2 , while the right hand side depends only on q1 , both
sides must be constant and in particular we may integrate
S2 (q2 ) = 2 q2 ,

(5.135)

with 2 constant (which is the conserved angular momentum), and (5.134) becomes the first order
ODE
dS1
dq1


1/2
2
2E 2V (q1 ) 22
F (q1 ; E, 2 ) .
q1

(5.136)

We have thus reduced the problem to quadratures, with the final solution to the Hamilton-Jacobi
equation being
Z
S(q1 , q2 , t) = Et + 2 q2 +

q1

F (x; E, 2 ) dx .

(5.137)

There is a trivial additive integration constant from integrating the final ODE. The two non-trivial
integration constants are 1 = E and 2 .
Recall that to solve Hamiltons equations we must now impose (5.116). This leads to the
equations
1 =
2 =

Z
S
= t +
E
Z
S
= q2
2

q1

q1

dx ,
F (x; E, 2 )
2
dx .
2
x F (x; E, 2 )

(5.138)
(5.139)

It may not look like it, but we have in fact solved the equations of motion. To see this in a more
familiar guise, recall that in the original variables the energy E is (see equation (2.91))
E =

p2
1 2
% +
+ V (%) ,
2
2%2

(5.140)

The equations of motion may then be written

where 2 = p = %2 .
% =

1
F (%) ,

d
2
= 2
,
dr
% F (%)

which integrate to (5.138) and (5.139).

91

(5.141)

5.8

* Quantum mechanics

We end these lectures with a few further remarks about the relationship between classical mechanics
and quantum mechanics. We already briefly described how the principle of least action enters the
Feynman formulation of quantum mechanics via the path integral (an integral over all paths) in
section 2.6.
Those who have studied quantum mechanics will certainly have recognized the Poisson bracket
structure we encountered in section 5.3. This similarity was first noticed by Dirac, who postulated
that a classical system may be quantized by taking the classical Poisson brackets and mapping
them to commutator brackets of operators
{f, g}

i
[f, g] .
~

(5.142)

The time evolution then maps as

df
dt

= {f, H} i~

df
,
= [f, H]
dt

(5.143)

which is the equation for the time evolution of an operator in the Heisenberg picture. Of course
for all this to make sense to each function f on classical phase space we must associate an operator
f acting on an appropriate Hilbert space H of wave functions, so that (5.142) holds. There
is a beautiful way to do this, under appropriate conditions, called geometric quantization. This
mathematically constructs the Hilbert space directly from the phase space, and provides an explicit
formula for mapping an appropriate class of classical observables f on phase space to operators f
acting on H. The geometric quantization procedure is engineered so that (5.142) and (5.143) then
hold for the operators f. This is all wonderful when the method can be applied. But there is
no general mathematical procedure called quantization that works in all circumstances. On the
other hand, from a physics point of view its not clear why one would expect there to be.
While the Poisson bracket representation of Hamiltonian mechanics relates to the Heisenberg
picture of quantum mechanics, the Hamilton-Jacobi equation relates to the Schrodinger equation.
In fact this was one of Schr
odingers motivations in finding his equation
i~

~2 2
+V ,
2m

(5.144)

describing a quantum particle of mass m moving under the influence of a potential V . In the WKB
approximation one substitutes the ansatz


i
(q, t) = A(q, t, ~) exp S(q, t)
~


,

(5.145)

where A is the real amplitude, and expands in an asymptotic series in ~, as we take ~ 0.

Substituting this into (5.144) the leading term gives


S
S
+ H q, p =
,t
= 0,
t
q
92

(5.146)

where
H(q, p, t) =

|p|2
+ V (q) ,
2m

(5.147)

is precisely the classical Hamiltonian. Thus the leading behaviour of the Schrodinger equation as
~ 0 is governed by the Hamilton-Jacobi equation, with Hamiltons principal function S(q, t)
(which is also the classical action) determining the phase of the wave function.

93

Appendix: Levi-Civita alternating symbol

The Levi-Civita alternating symbol ijk is defined by

ijk

+1 if ijk is an even permutation of 123 ,

1 if ijk is an odd permutation of 123 ,

0
otherwise .

Notice that ijk = kij = jki (cyclically permuting the indices), while ijk = ikj . With this
definition the cross product A B of two vectors A, B has ith component
3
X

(A B)i =

ijk Aj Bk .

j,k=1

A useful identity is

ia ib ic

= det ja jb jc ,
ka kb kc

ijk abc

where the Kronecker delta symbol is defined by

(
ij

1
0

i=j
.
i 6= j

As a special case we have the identity

3
X

ijk ibc = jb kc jc kb .

i=1

As an application, let us compute the scalar quadruple product:

3
X

(A B) (C D) =

ijk Aj Bk ibc Cb Dc

i,j,k,b,c=1
3
X

(jb kc jc kb ) Aj Bk Cb Dc

i,j,k,b,c=1

= (A C)(B D) (A D)(B C) .
Other vector product identities are similarly straightforward to derive (for example the scalar triple
product identity follows immediately from cyclically permuting the indices on ijk ). Although we
wont use the following identity, we present it for interest:
(det M ) ijk =

3
X
a,b,c=1

where M = (Mij ) is any 3 3 matrix.

94

abc Mia Mjb Mkc ,