Anmek

Notes on Analytical Mechanics
Ingemar Bengtsson (2014)
Contents
1 The
1.1
1.2
1.3
1.4
Best of all Possible Worlds

Analytical and Hamiltonian mechanics
The calculus of variations . . . . . . . .
How to solve equations . . . . . . . . .
Phase space . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
page 1
. .
2
. .
6
. .
8
. . 12
2 Lagrangian mechanics
18
2.1 The scope of Lagrangian mechanics . . . . . . . . . . . . . . . 18
2.2 Constrained systems . . . . . . . . . . . . . . . . . . . . . . . . 21
2.3 Symmetries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3 Interlude: Conic sections
32
4 The
4.1
4.2
4.3
4.4
4.5
4.6
36
36
39
40
42
43
46
central force two-body problem

The problem and its formal solution . .
Existence and stability of circular orbits
Keplers First Law . . . . . . . . . . . .
Keplers Third Law . . . . . . . . . . .
Self-similarity and the virial theorem .
The three-body problem . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
5 Small oscillations
51
5.1 Forced oscillations . . . . . . . . . . . . . . . . . . . . . . . . . 51
5.2 Damped and forced oscillations . . . . . . . . . . . . . . . . . . 54
5.3 Several degrees of freedom . . . . . . . . . . . . . . . . . . . . 57
6 Rotation and rigid bodies
6.1 Rotations . . . . . . . . . . .
6.2 Rotating coordinate systems
6.3 The inertia tensor . . . . . .
6.4 Eulers equations . . . . . . .
6.5 The Lagrangian description .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
7 Interlude: Legendre transformations
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
61
61
66
69
72
76
82
ii
8 The
8.1
8.2
8.3
8.4
8.5
8.6
8.7
8.8
Contents
Hamiltonian formulation
Hamiltons equations and Hamiltonian flows
Kets and bras and all that . . . . . . . . . .
The symplectic form . . . . . . . . . . . . . .
The algebraic structure of mechanics . . . . .
The sphere as a phase space . . . . . . . . .
Infinitesimal canonical transformations . . .
The symplectic one-form . . . . . . . . . . .
General transformation theory . . . . . . . .
9 Integrable and chaotic motion

9.1 Can chaos occur? . . . . . . .
9.2 Integrable systems . . . . . . .
9.3 Canonical perturbation theory
9.4 Stability of the Solar System .
Appendix 1 Books
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
85
. 85
. 88
. 92
. 95
. 97
. 99
. 101
. 103
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
108
108
110
113
116
119
1 The Best of all Possible Worlds
Mechanics is the paradise of the mathematical

sciences, because with it one comes to the fruits
of mathematics
Leonardo da Vinci
Sir Isaac Newton was a Master of the Mint. He also formulated three celebrated
laws of mechanics, which we can paraphrase as follows:
1. A particle not subject to any force moves on a straight line at constant
speed.
2. In the presence of a force, the position of a particle obeys the equations of
motion
.
m
xi = Fi (x, x)
(1.1)
3. The force exerted by a particle on another is equal in magnitude, but opposite in direction, to the force exerted by the other particle on the first.
A particle is here thought of as an entity characterized by its mass m, its
location in space, and by nothing else.1 The aim of Newtons mechanics is to
predict the location at arbitrary times, given the position and velocity at some
initial time. This is done by means of a solution of the differential equations
above.
An overdot denotes differentiation with respect to the time parameter t
(this notation, as well as Differential Calculus itself, was invented by Newton),
and xi may denote a vector in three-dimensional space. Sometimes it will be
understood that we are dealing with a set of N particles, and moreover we
often suppress indices. Then the force Fi (x, x)
is a 3N component function
of the 3N variables xi and their 3N derivatives x i . Since the index notation
may be a bit unfamiliar, let me note that whenever indices occur in a formula,
it is understood that they can take any of a specified set of integer values.
1
For further discussion see A. Jenkins, On the title of Moriartys Dynamics of an Asteroid, eprint
arXiv:13012.5855.
The Best of all Possible Worlds
If i {1, 2, ..., n} then eq. (1.1) stands for n separate equations. Throughout
we employ Einsteins summation convention, which means that whenever a
certain index occurs twice in a particular term, a sum over all its allowed
values is understood, eg.
xi yi
n
X
i=1
xi yi =
n
X
xj yj = xj yj .
(1.2)
j=1
It does not matter which letter is being used for a repeated index. To avoid
confusion, the same index never occurs thrice or more in a single term. In
section 8.2 we will introduce index notation in a more sophisticated tensorial
way, but for the time being this is all there is to it. By the way index notation
is not always the best choiceit does not, for instance, make use of any special
properties of three dimensional spacebut it has the advantage that it can be
used for everything, which is why I always use it.
1.1 Analytical and Hamiltonian mechanics

What are we to think of Newtons laws? A physicist might follow Newton in
using them to predict the position of the planets as they go around the sun,
and will conclude that they are very meaningful. A mathematician might say
that they do not say very much, only that the position of a particle is described
by a set of ordinary differential equations. A philosopher might object that
they say nothing at allthe first and second law together seem to state that
a particle moves in a straight line, unless it does something else, in which
case we say that it is subject to a force. But the philosopher Kant valued
Newtons laws highly, and tried to prove that they are somehow necessary
consequences of the way our minds perceive the world, and have a status
similar to Euclids axioms in geometry. Kant overestimated Newtons laws,
but they do have content as they stand. It is a highly non-trivial fact that a
second order differential equation is being postulated, since this means that
the position and the velocity can be chosen arbitrarily at a given instant, but
not the acceleration.
Analytical mechanics is at once more general and more special than Newtons theory. It is more general because it is more abstract. Its equations do
not necessarily describe the positions of particles, but may be applied to much
more general physical systems (such as field theories, including Einsteins general relativity theory). In the version we will study it is more special because
only a restricted set of forces will be allowed in eq. (1.1). Let us see what kind
of restrictions on the function Fi that are of physical interest. Newtons third
law is already a restriction. It can be reformulated as the statement that the
total momentum of a system composed of several particles is conserved:
dPi
d X
mx i = 0 ,
dt
dt particles
(1.3)
where the sum is over all the particles in the system. The existence of such
a conserved vector is clearly an interesting fact. By the way this formulation
is quite superior when we deal with time dependent masses, say with rockets
(see exercise 3). Now consider the function
mx 2
+ V (x) ,
(1.4)
2
where V is some function of x, known as the potential energy. The function T
is called the kinetic energy, while E itself is the energy of the system. Clearly
E =T +V =
E = x i (m
xi + i V (x)) .
(1.5)
It follows that if the force is given by

Fi (x, x)
= Fi (x) = i V (x) ,
(1.6)
then the energy of the system is conserved. Systems for which a conserved
energy function exists are called conservative. In our example, and indeed in
many interesting cases, the energy can be divided into kinetic and potential
parts, and the equation of motion is given by
m
xi = i V (x) .
(1.7)
This move is typical of analytical mechanics, where vectors are usually derived
from scalar functions.
Analytical mechanics devises methods to derive the differential equations describing a given system, strategies for solving them, and ways of describing
the solutions if they cannot be obtained in explicit form.2
We will tentatively restrict ourselves to conservative systems only. If you like
this is a strengthening of the third law, and it is believed that all isolated
systems in Nature are of this type.
What we are trying to do is to find some properties that all the Laws of
Physics, and in particular all allowed equations of motion, have in common.
Now the philosopher Leibnitzwho was the other of the two inventors of
Differential Calculusargued that we live in the best of all possible worlds.
Is it evident from eq. (1.7) that this is so? Indeed it is, as was first realized
half a century after the publication of Newtons Principia. The inspiration
came from optics, and the laws of reflection and refraction. It is known that
the angle of reflection is equal to the angle of incidence, and it was observed
2
As a definition, this is a little vague. M

echanique Analitique was the title of a book by Lagrange.
by the Greeks that this implies that light always travels on the shortest path
available between two points A and B, subject to the restriction that it should
be reflected against the surface. If the angle of reflection were to differ from
the angle of incidence, the distance covered by light in going from A to B
would be greater than it has to be. For refraction, we have Snells Law. Any
medium can be assigned an index of refraction n, and the angle of refraction
is related to the angle of incidence through
n1 sin i = n2 sin r .
(1.8)
Fermat noted that if

c
,
(1.9)
v
where v is the velocity of light in the medium and c is a constant (independent
of the medium), then Snells law can be derived from what is now known as
Fermats principle, namely that the time taken for light to go from A to B is a
minimum. Fermats principle unifies the laws of refraction and reflection, since
it also implies the equality between the angles of incidence and reflection.
More generally the index of refraction may be a function n(x) of position,
say through a dependence on temperature. This is what causes mirages. To
study this mathematically we imagine that we evaluate the curve integral
n=
I=c
dt =
cds
=
v
n(x(s))ds
(1.10)
along an arbitrary path (s) between A and B. Then the path actually taken
by light in going from A to B through the medium is that specific path which
results in the smallest possible value of the integral I. This path may well not
be a straight line. The question is how to do the optimization. We will soon
come to it.
Is there a similar principle underlying mechanics? At least for systems obeying eq. (1.7), there is. Consider two points A and B, and suppose that a particle
starts out at A at time t = t1 , and then moves along an arbitrary path from
A to B with whatever speed that is consistent with the requirement that it
should arrive at B at the time t = t2 . In mathematical terms we are dealing
with a function x(t) such that
x(t1 ) = xA
x(t2 ) = xB ,
(1.11)
but otherwise arbitrary. For any such function x(t) we can evaluate the integral
S[x(t)] =
t2
t1
dt (T V ) =
t2
dt
t1

mx 2
V (x) .
2
(1.12)
S is known as the action. It is a functional, i.e. a function of a functionthe

functional S[x(t)] assigns a real number to any function x(t). Note that S[x(t)]
is not a function of t, hence the square bracket notation. On the other hand
it is a function of xA , xB , t1 , and t2 , but this is rarely written out explicitly.
The statement, to be verified in the next section, is that the action functional (1.12) has an extremum (not necessarily a minimum) for precisely that
function x(t) which obeys the differential equation (1.7). This is known as
Hamiltons Principle, orwith less than perfect historical accuracyas the
Principle of Least Action.
Hamiltonian mechanics deals with those, and only those, equations of motion
which can be derived from Hamiltons Principle, for some choice of the action
functional.
This is a much more general class than that given by eq. (1.7), but it does
exclude some cases of physical interest. Hamiltonian mechanics forms only a
part of analytical mechanicsnamely that part that we will focus on.
Note once again what is going on. The original task of mechanics was to
predict the trajectory of a particle, given a small set of data concerning its
state at some intial time t. We claim that there exists another formulation of
the problem, where we can deduce the trajectory given half as much data at
each of two different times. So there seems to be a local, causal way of looking
at things, and an at first sight quite different global, teleological viewpoint.
The claim begins to look reasonable when we observe that the amount of free
data in the two formulations are the same. Moreover, if the two times t1
and t2 approach each other infinitesimally closely, then what we are in effect
specifying is the position and the velocity at time t1 , just as in the causal
approach.
Why do principles like Fermats and Hamiltons work? In both cases, we are
extremizing a quantity evaluated along a path, and the path actually taken
by matter in nature is the one which makes the quantity in question assume
an extremal value. The point about extremanot only minimais that if the
path is varied slightly away from the extremal path, to a path which differs to
order from the extremal one, then the value of the path dependent quantity
suffers a change which is of order squared. At an extremum the first derivative
vanishes. In the case of optics, we know that the description of light as a bundle
of rays is valid only in the approximation where the wavelength of light is
much less than the distance between A and B. In the wave theory, in a way,
every path between A and B is allowed. If we vary the path slightly, the time
taken by light to arrive from A to B changes, and this means that it arrives
out of phase with the light arriving along the first path. If the wavelength
is very small, phases from light arriving by different paths will be randomly
distributed, and will cancel each other out through destructive interference.
This argument fails precisely for the extremal paths: for them, neighbouring
paths take approximately the same time, light from all neighbouring paths will
arrive with the same phase, and constructive interference takes place. Thus,
whenever the wavelength is negligibly small, it will appear that light always
travels along extremal paths.
Only in the twentieth century was it realized that Hamiltons Principle

works for the same reason that Fermats Principle works. Classical mechanics
is a kind of geometrical optics limit of a wave mechanics of matter, operating
in configuration space. But that is another story.3
1.2 The calculus of variations

Let us now verify the claim made in the first section, namely that Newtons
differential equations, for suitable choices of the dynamical system, are mathematically equivalent to the requirement that a certain functional of all possible
paths of the particles should assume an extremum value at the actual trajectory. First we stare at the definition of the action functional:
S[x(t)] =
t2
dt
t1

mx 2
V (x) .
2
(1.13)
How do we find the extrema of such a function of a function? Let us begin

with some formal considerations. For a function f of an ordinary number x, it
is easy enough to find the extrema. We consider how the function values f (x)
change as we change the number x:
f (x) f (x + x) f (x) = xx f (x) .
(1.14)
We assume that x is so small that second order terms can be ignored. If the
derivative is zero at the point x, the function has a minimum, or a maximum,
or at least an inflection point there. For a function of several variables, the
condition for an extremum (a minimum, a maximum, or a saddle point) is
that
f (x) =
X
i
xi
f
(x1 , . . . , xN ) = 0
xi
(1.15)
for arbitrary choices of the xi , which means that all the N partial derivatives
have to vanish at the extremal points. Now a functional of a function x(t)
can be regarded as a function of an infinite number of variables, say of the
Fourier coefficients of the original function. You can also regard t as a label of
the infinite number of variables on which the functional dependsa kind of
continuous indexand then what we have to do is to replace the sum in eq.
(1.15) with an integral. Like this:
S = S[x(t) + x(t)] S[x(t)] =
t2
t1
dt x(t)
S
(t) .
x
(1.16)
We assume that it is possible to bring S to this form. Then the functional

derivative of S[x] will be defined as the very expression that occurs to the
3
Namely that of quantum mechanics.
1.2 The calculus of variations
right in the integrand. The equations of motion, as obtained from Hamiltons

Principle, then state that the functional derivative of the action is zero, since
the form of the function x(t) is arbitrary.
It remains to be seen if we really can bring S to this formif not, we would
have to conclude that S[x(t)] is not differentiable. First of all, note that we
are all the time evaluating the action between definite integration limits. Then
the extremum, if it exists, will be given by that particular trajectory which
starts at the point x(t1 ) at time t1 , and ends at the point x(t2 ) at time t2 ,
and for which the functional derivative vanishes. We can make this work for
the action functional (1.13). Imagine that we know its value for a particular
function x(t), and ask how this value changes if we evaluate it for a slightly
different function
x
(t) = x(t) + x(t) ,
x(t) = f (t) ,
(1.17)
where f (t) is, for the time being, an arbitrary function while is an infinitesimally small constant. It is important for the following argument that f (t) is
arbitrary, or nearly so. That is infinitesimally small simply means that we
will neglect terms of quadratic and higher orders in in the calculation which
follows:
S = S[
x(t)] S[x(t)] =
=
t2
dt
t1
m
2
t2
t1
dt
m
2

(x + x)
2 V (x + x) S[x(t)]

x 2 + mx
x V (x) xx V (x) + o(2 ) S[x(t)] =
(1.18)
t2
t1
dt (mx
x xx V (x)) + o(2 ) .
The action functional has an extremum at the particular function x(t) for
which this expression vanishes to first order in . What we want to see is
what kind of restrictions this requirement sets on the function. To see this, we
perform a partial integration
S =
t2
dt
t1
x(m
x + x V (x)) +

d
(mxx)
.
dt
(1.19)
Unfortunately this is not quite of the form (1.16), due to the presence of the
total derivative in the integrand. Therefore we impose a restriction on the so
far arbitrary function f (t) that went into the definition of x, so that
x(t1 ) = x(t2 ) = 0 .
(1.20)
This is a way of saying that we are interested only in functions x(t) that
have certain preassigned starting and end points at specified times. With this
restriction, the total derivative in eq. (1.19) goes away. The first term has
to vanish for all allowed choices of the functions x(t). After a moments
reflection, we see that this can happen only if the factor multiplying x in
the integrand is zero! Hence we have proved that the action functional has an
extremum, among all possible functions obeying
x(t1 ) = xA
x(t2 ) = xB ,
(1.21)
for those and only those functions which obey

m
x = x V (x)
(1.22)
at all intermediate points.

So we have proved, in this particular case at least, that Newtons equations
of motion can be derived from the condition that a certain action functional
shall have an extremum value. Note also that the restrictions that we had to
set on the function x(t), eqs. (1.21), make perfect sense. To obtain a definite
trajectory it is not enough to impose the equations of motion. It is also necessary to set some initial values. For differential equations of second order, it
is natural to make a choice of x(0) and x(0).
From the point of view of the

action, it is natural to impose the value of x(t) at two different times, which
is the same amount of information. It should be noted though that whatever
values of x(0) and x(0)
we choose there is always a unique solution for some

range of t, while it is perfectly possible that the equation of motion is such
that there is no solution, or several solutions, for a given pair of x(t1 ) and
x(t2 ).
The rest of this course is an elaboration of the contents of this section. If
you have not understood everything perfectly yet there is still time!
1.3 How to solve equations

It is one thing to be able to set up equations for a physical system, and perhaps
to prove theorems to the effect that a solution always exists and is unique,
given suitable initial conditions. Another issue of obvious interest is how to
solve these equations, or at least how to extract information from them. What
precisely do we mean when we say that a differential equation is soluble?
Consider, as an exercise, a first order differential equation for a single variable:
x = f (x) ,
(1.23)
where f is some function. This can be solved by means of separation of variables:

dx
dt =
f (x)
t(x) =
x
c
dx
,
f (x )
(1.24)
where c is a constant determined by the initial condition. If we do this integral,
and then invert the resulting function t(x) to obtain the function x(t), we
have solved the equation. We will regard eq. (1.24) as an implicit definition
of x(t), and eq. (1.23) is soluble in this sense. This is reasonable, since the
manipulations required to extract t(x) can be easily done on a computer, to any
desired accuracy, even if we cannot express the integral in terms of elementary
functions. But there are some limitations here: It may not be possible to invert
the function t(x) except for small times.
Next consider a second order equation, such as the equation of motion for
a harmonic oscillator:
m
x = ax .
(1.25)
This is a linear equation, and we know how to express the solution in terms of
trigonometric functions, but our third examplea pendulum of length lis
already somewhat worse:
ml2 = gml sin .
(1.26)
Let us therefore approach eq. (1.25) in a systematic fashion, which might

yield results also for the pendulum. As a first step, note that any second order
differential equation can be rewritten as a pair of coupled first order equations:
p = ax
mx = p .
(1.27)
The second equation defines the new variable p. Unfortunately coupled first
order equations are difficult to solve, except in the linear case when they can
be decoupled through a Fourier transformation.
The number of degrees of freedom of a dynamical system is defined to be one
half times the number of first order differential equations needed to describe
the evolution. It will turn out that, for systems whose equations of motion are
derivable from the action principle, the number of first order equations will
always be even, so the number of degrees of freedom is always an integer for
such systems. A system with n degrees of freedom will be described by a set of
2n in general coupled first order equations, and the difficulties one encounters
in trying to solve them will rapidly become severe.
In the cases at hand, with one degree of freedom only, one uses the fact that
these are conservative systems, which will enable us to reduce the problem to
that of solving a single first order equation. For the harmonic oscillator the
conserved quantity is
mx 2 ax2
+
.
2
2
The number E does not depend on t. Equivalently
E=
x 2 =
2E
a
x2 .
m
m
(1.28)
(1.29)
10
Taking a square root we are back to the situation we know, and we proceed
as before:
dt = dx
m
2E ax2
t(x) =
dx
m
.
2E ax2
(1.30)
Inverting the function defined by the integral, we find the solution x(t). The
answer is a trigonometric function, with two arbitrary constants E and c determining its phase and its amplitude. For our purposes the trigonometric
function is defined by this procedure!
We can play the same trick with the non-linear equation for the pendulum,
and we end up with
t() =
d
q
2
(E
ml2
(1.31)
+ gml cos )
We integrate, and we invert. This defines the function (t). We could leave it
at that, but since our example is a famous one, we manipulate the integral a
bit further for the fun of it. Make the substitution
sin
k sin
2
2k cos d
d = p
.
1 k2 sin2
(1.32)
The constant k is undetermined at this stage. The integral becomes

t() =
l
2g
()
2k cos d
q
E
1 k2 sin2 gml
+ 1 2 sin2
(1.33)
The integrand simplifies if we choose the constant k such that

2k2
E
+1 .
gml
(1.34)
One further substitution takes us to our desired standard form;

s Z
.
.
l ()
d
p
t() =
=
sin
x
g c
1 k2 sin2
s Z
l x()
dx
p
=
.
g c
(1 x2 )(1 k2 x2 )
(1.35)
Just as eq. (1.30) can be taken as an implicit definition of a trigonometric

function, this integral implicitly defines the function (t) as an elliptic function.
If you compare it with the previous integral (1.30), you see that an elliptic
function is a fairly natural generalization of a trigonometric, i.e. circular,
function. Since elliptic functions turn up in many contexts they have been
11
Figure 1.1. A Lissajous figure, and how to draw it; x = cos t, y = sin 2t.
studied in depth by mathematicians. Their work remains relevant, even if

Mathematica will plot the solution (t) in no time.
Anyway, the above examples were some of the simplest examples of completely soluble dynamical systems. Just wait till we get to the insoluble ones!
Why did this work at all? The answer is that we had one degree of freedom,
and one constant of the motion, namely E. This reduced the problem to that
of solving a single uncoupled equation. This suggests a general strategy for
solving the equations of motion for a system containing n degrees of freedom,
i.e. solving 2n coupled first order equations: One must find a set of n constants
of the motion with suitable properties, so that the problem reduces to that of
computing n integrals. This idea forms the core of Hamilton-Jacobi theory. It
works sometimes, but not very often. As a result the notion of what it means
to solve a set of differential equations evolved somewhat: a solution might
consist, say, of a convergent power series in t. But frequently this strategy
also fails. A typical Hamiltonian system will exhibit an amount of chaotic
behaviour, and there may not exist any effective procedure to generate the
long term behaviour of the solutions on a computer. What one has to do then
is to find out which questions one can reasonably ask concerning such systems.
Even in situations where one can solve the equations, things may not be
altogether simple. Consider two harmonic oscillators, with the explicit solution
x = a cos (1 t + 1 )
y = b cos (2 t + 2 ) .
(1.36)
The trajectory in the x-y-plane is a Lissajous figure. Fig. 1.1 explains how to
draw them; further examples are readily produced with a computer. If 1 = 2
the trajectory is an ellipse, with circles and straight lines as special cases. More
generally, if there exist integers m and n such that
m1 = n2
(1.37)
12
the trajectory is a closed curve. If there are no such integers the trajectory
eventually fills a rectangle densely, and never closes on itself. Now put yourself into the position of an experimentalist trying to determine by means of
measurements whether the trajectory will be closed or not!
Another example of this type is a particle moving on a straight line on a
plane, but confined to a quadratic box and bouncing elastically from its walls.
Let us ask whether the trajectory is periodic or whether it will eventually come
arbitrarily close to any point in the box. The answer depends on the initial
condition for the direction of motion. If the angle between this direction and
one of the walls is called , the question is whether tan is a rational number
or not. Theoretically this is fine, but for someone who wants to decide the
question by means of measurements of the initial velocity it is not!
1.4 Phase space

It is worthwhile formalizing things a bit further. With the understanding that
every set of ordinary differential equations can be written in first order form,
we write down the general form of N coupled first order equations for N real
variables zi :
zi = fi (z1 , . . . , zN ; t) ,
1iN ,
(1.38)
where the N functions fi are arbitrary. We simplify things by assuming that

there is no explicit dependence on time. We then have the equations that
describe an autonomous dynamical system, namely
zi = fi (z1 , . . . , zN ) .
(1.39)
There are theorems that guarantee the existence and uniqueness of such systems for some range of the parameter t. Thus
zi = zi (z01 , . . . , z0N , t) ,
(1.40)
where z0i are the initial values of zi .

There is no guarantee that such solutions can be obtained in any explicit
form. If we discretize the time variable a computer can easily generate approximative solutions, but it may be practically impossible to produce accurate
solutions over long intervals of time.
We assume that the physical systems we are interested inas far as we
attempt to describe themcan be fully characterized by the N real numbers
zi . We imagine a space whose points are labelled in a one-to-one fashion by
these numbers, and call it phase space.
The set of all possible states of a physical system is in one-to-one correspondence with the points of an N dimensional phase space. The time development
of a system is uniquely determined by its position in phase space.
1.4 Phase space
13
This is the first of several abstract spaces that we will encounter, and you
must get used to the idea of abstract spaces.
A particle moving in space has a 6 dimensional phase space, because its
position (3 numbers) and its velocity (3 numbers) at a given time determine
its position at all times, given Newtons laws. Anything else can either be
computed from these numbersthis is true for its accelerationor else it can
be ignoredthis would be true for how it smells, if it does. The particle also
has a mass, but this number is not included in phase space because it is given
once and for all. Two particles moving in space have a 12 dimensional phase
space, so high dimensional phase spaces are often encountered. We will have
to picture them as best we may.
Now consider time evolution according to eq. (1.39). Because of the theorems
I alluded to, we know that through any point z0 there passes a unique curve
zi (t), with a unique tangent vector zi . These curves never cross each other.
When the system is at a definite point in phase space, it knows where it is
going. The curves are called trajectories, and their tangent vectors define a
vector field on phase space called the phase space flow. Imagine that we can
see such a flow. Then there are some interesting things to be observed. We
say that the flow has a fixed point wherever the tangent vectors vanish. If the
system starts out at a fixed point at t = 0, it stays there forever. There is an
important distinction to be made between stable and unstable fixed points. If
you start out a system close to an unstable fixed point it starts to move away
from it, while in the stable case it will stay close forever. The stable fixed point
may be an attractor, in which case a system that starts out close to the fixed
point will start moving towards it. The region of phase space which is close
enough for this to happen is called the basin of attraction for the attractor.
Consider a one dimensional phase space, with the first order system
z = f (z) .
(1.41)
For generic choices of the function f all fixed points are either stable attractors,
or unstable repellors, but for special choices of f we can have fixed points
that are approached by the flow only on one side. The latter are structurally
unstable, in the sense that the smallest change in f will either turn them into
pairs of attractors and repellors, or cause them to disappear altoghether.
In two dimensions there are more possibilities. We can have sources and
sinks, as well as stable elliptic and unstable hyperbolic fixed points. To see
what the latter two look like, we return to the examples given in section 1.3.
The phase space of the harmonic oscillator is R2 , and it contains one elliptic
fixed point. It is elliptic because it is surrounded by closed trajectories, and
hence it is stable. In the case of the pendulum phase space has a non-trivial
topology: since the coordinate is a periodic angle phase space is the surface
of an infinitely long cylinder. It contains two fixed points. One of them is
elliptic, and the otherthe state where the pendulum is pointing upwardsis
hyperbolic. What is special about the hyperbolic fixed point is that there are
14
Figure 1.2. A one dimensional phase space, containing one stable and one unstable fixed point, as well as one fixed point which is structurally unstable.
Figure 1.3. Fixed points in a two dimensional phase space: a source, a sink, a
limit cycle, an elliptic fixed point, and a hyperbolic fixed point.
two trajectories leading into it, and two leading out of it. The length of the
tangent vectors decrease as the fixed point is approached. Taking the global
structure of phase space into account we see that a trajectory leaving the fixed
point is in fact identical to one of the incoming ones. Hence there are really only
two special trajectories. A striking fact about them is that they divide phase
space into regions with qualitatively different behaviour. One region where
the trajectories go around the elliptic fixed point, and two regions where the
trajectories go around the cylinder. For this reason the special trajectories are
called separatrices, and the regions into which they divide phase space are
called invariant sets by definition an invariant set in phase space is a region
that one cannot leave by following the phase space flow.
It is very important that you see how to relate this abstract discussion of
the phase space of the pendulum to known facts about real pendula. Do this!
It is not by accident that the phase space of the pendulum is free of sources
and sinks. The reason is, as we will see in section 8.1, that only elliptic or
hyperbolic fixed points can occur in Hamiltonian mechanics. Real pendula
tend to have some amount of dissipation present (because they are imperfectly
isolated from the environment), and then the situation changes; see exercise
11. Speaking of Hamiltonian systems it is worthwhile to point out that the
example of the two harmonic oscillators in eq. (1.36) is less frivolous than it
1.4 Phase space
15
may appear. The phase space is four dimensional, but there are two conserved
quantities
2E1 = p21 + 12 x21
2E2 = p22 + 22 x22 .
(1.42)
This means that any given trajectory will be confined to a two dimensional
surface in phase space, labelled by E1 and E2 . This surface is a torus, with
topology S1 S1 . In a sense to be made precise later, non-chaotic motion in
a Hamiltonian system always takes place on a torus in phase space.
Finally we observe that we have the beginnings of a strategy to understand
any given dynamical system. We begin by locating the fixed points of the
phase space flow. Then we try to determine the nature of these fixed points. If
the equations are linear this is straightforward. If not, we can try linearization
of the equations around the fixed points. There is a theorem we can lean on
here:
The Hartman-Grobman theorem: The nature of the fixed points is unchanged
by linearization, as long as the fixed points are isolated and as long as no
elliptic fixed points occur.
The caveat in the statement will be explained presently.
Now consider the pendulum. Its phase space is a cylinder described by the
coordinates (, p ). To see if the phase space flow has any fixed points, you set
1
=
p = 0
p = gml sin = 0 .
(1.43)
ml2
Hence there are fixed points at (, p ) = (0, 0) and (, 0). Linearizing around
them you find the former to be elliptic and the latter to be hyperbolic. If this
remains true for the non-linear equations you can easily draw a qualitatively
correct picture of the phase space flow. No integration is needed.
Were we justified in assuming that the fixed points are elliptic? To see what
can go wrong, consider the non-linear equation
x
+ x2 x + x = 0 .
(1.44)
In the linearised case ( = 0) there is a single elliptic fixed point. In the

non-linear system the flow will actually spiral in or out from the fixed point,
depending on the sign of , so this is an example where the exceptions to
the Hartman-Grobman theorem are important. But in the case of the pure
pendulum we know that the non-linear system is Hamiltonian, and therefore
sources and sinks cannot appearour analysis of the pendulum was therefore
accurate.
Our tentative strategy works very well when the phase space is two dimensional, but if the dimension of phase space exceeds two things can get very
complicated indeed. A famous example is the at first sight innocent looking
Lorenz equations
16
z1 = az1 + az2
z2 = bz1 z2 z1 z3
z3 = cz3 + z1 z2
(1.45)
.
They capture some aspects of thermal convection in a fluid. Lorenz was a

metereologist interested in the long term accuracy of weather prediction, but
he found himself unable to do long term prediction even in this simple model.
The non-linear terms have a dramatic effect, and the slightest change in the
initial data will cause the trajectory to go to completely different regions of
the three dimensional phase space. The behaviour of the Lorenz equations is
chaotic in a technical sense to be explained in chapter 9.
Problem 1.1
Newtons Second Law says that the position and the velocity of
a particle can be freely specified; then the trajectory x(t), and therefore all derivatives
of order higher than one, is determined by the equation of motion. Suppose instead
that either
a) only the position can be freely specified, and that the equation of motion determines
all the derivatives, or
b) position, velocity and acceleration can be freely specified, and that the equation of
motion determines all derivatives of order higher than two.
Discuss these assumptions in the light of Newtons First Law.
Problem 1.2 A particle of mass m1 = 1 kg and a particle of mass m2 = 1

kg (negative mass) interact with each other according to Newtons Law of Gravity.
Describe in qualitative terms the behaviour of the system. Is the energy conserved?
Problem 1.3
Derive the rocket equation

Fi = mv i mu
i,
(1.46)
where Fi is an external force, vi is the velocity of the rocket, and ui is the exhaust
velocity (relative to the rocket).
Problem 1.4
Prove Snells Law of optics, starting from Fermats principle.
Also argue for it using properties of plane waves.
Problem 1.5
An elastic bar extends between x = 0 and x = L. It resists
bending, has a load per unit length given by (x), and is subject to gravity. We may
therefore assume that its energy is given by
E=
dx
0
k
y y (x)y
2
where the slash denotes differentiation with respect to x and k is a constant. The
bar will minimize its energy. Analyse the variational problem to see what equation
determines the equilibrium position, and what conditions one must impose on the end
of the bar in order to obtain a unique solution. Archers want their bows to bend like
circles. Conclude that bows must have a value of k that depends on x.
1.4 Phase space
17
Problem 1.6
Is it true that a once differentiable function x(t) is a solution
of eq. (1.25) if and only if it is a solution of eq. (1.29)? If not, find a non-trivial
counterexample.
Problem 1.7
Using the general solution for the pendulum, eq. (1.35), solve
for (t) in the special case k = 1. Physically, what does this solution correspond to?
Problem 1.8
Use Mathematica to compare the solutions for the pendulum
to those of the harmonic oscillator, for various values of the energy (which you adjust
so that E = 0 corresponds to the stable fixed point in both cases).
Problem 1.9
Consider a projectile that is fired straight up in a gravitational
field (V = GM/r), reaches a maximum height rmax , and falls back again. Prove
that the solution has the parametric form
rmax
r=
(1 cos ) ,
2
rmax
t=
2
rmax
( sin ) .
2GM
Show that the resulting curve in the tr plane is a cycloid, the curve followed by a
point on the perimeter of a circular disk rolling without slipping on the taxis.
Problem 1.10
Using the construction sketched in Fig. 1.1, draw Lissajous
figures for (x, y) = (cos t, cos (2t + )), for = 0, /4, /2. What is the first of
these called?
Problem 1.11 Linearize the pendulum around its fixed points, and then draw
+sin
a careful picture of its phase space. Add a friction term to the equation, +
=
0, and see in qualitative terms what this does to the phase space flow.
Problem 1.12
Give a simple example where linearization around a fixed
point gives an erroneous impression of its nature because the fixed point does not
stay isolated.
Problem 1.13 Show that a non-autonomous dynamical system can always be

rewritten as an autonomous dynamical system by increasing the dimension of its phase
space. Do so in the simplest possible way. Supposing the original non-autonomous
system has trajectories that form circles and figures-of-eight. What do they look like
in the corresponding autonomous system?
2 Lagrangian mechanics
With the agreement that the action integral is an important object, we give
a name also to its integrand, and call it the Lagrangian. In the examples that
we considered so far, and in fact in most cases of interest, the Lagrangian is a
function of a set of n variables qi and their n first order derivatives qi :
S[q(t)] =
t2
dt L(qi , qi ) .
(2.1)
t1
We use q to denote the coordinates because the Lagrangian formalism is very

general, and can be applied to all sorts of systems where the interpretation of
the variables may differ from the interpretation of x as the position of some
particle. The space on which qi are the coordinates is called the configuration
space. Its dimension is one half that of phase space. It is an intrinsic property
of the physical system we are studying, and is a very useful concept. You
should try to think as much as possible in terms of the configuration space
itself, and not in terms of the particular coordinates that we happen to use
(the qs), since the latter can be changed by coordinate transformations. In
fact one of the advantages of the Lagrangian formalism is that it is easy to
perform coordinate transformations directly in the Lagrangian. We will see
examples of this later on. Moreover there are situationssuch as that of a
particle moving on a spherewhen several coordinate systems are needed to
cover the whole configuration space. The important thing is the sphere itself,
not the coordinates that are being used to describe it. Which is not to say
that coordinates are not useful in calculationsthey definitely are!
The pair (q, q)
determines a tangent vector (to some curve) at the point
whose coordinate is q. Taken together q and q are coordinates on the tangent
bundle of configuration space. The dimension of the tangent bundle equals
that of phase space.
2.1 The scope of Lagrangian mechanics

Among all those functions qi (t) for which qi (t1 ) and qi (t2 ) are equal to some arbitrarily prescribed values, the action functional has an extremum for precisely
those functions qi (t) which obey the Euler-Lagrange equations
2.1 The scope of Lagrangian mechanics
L
d
qi dt
L
qi
=0,
19
(2.2)
provided such functions exist. This is straightforward to verify by means of

the calculus of variations; indeed (suppressing indices)

Z t2

L
L
L
d L
d
L
+ q
=
+
q
.
dt q
dtq
q
q
q
dt q
dt
q
t1
t1
(2.3)
The total derivative term gives rise to a boundary term that vanishes because
we are only varying functions whose values at t1 and t2 are kept fixed, so that
q is zero at the boundary. The Euler-Lagrange equations follow as advertized.
The question is to what extent the equations of motion that actually occur in
physics are of this form.
There are some that cannot be brought to quite this form by any means, including some of considerable physical interest; most of them involve dissipation
of energy of some sort. An example is that of a white elephant sliding down a
hillside covered with flowers.1 But then frictional forces are not fundamental
forces. A complete description of the motion of the elephant would involve the
motion of the atoms in the elephant and in the flowers, both being heated by
friction. It is believed that all complete, fundamental equations are derivable
from Hamiltons principle, and hence that they fall within the scope of Lagrangian mechanicsor of quantum mechanics, which is structurally similar
in this regard.
Generally speaking we expect Lagrangian mechanics to be applicable whenever there is no dissipation of energy. For many simple mechanical systems the
Lagrangian equals the difference between the kinetic and the potential energy,
S =
t2
L(x, x)
= T (x)
V (x) .
(2.4)
For instance,
mx 2
L
d L
V
V (x)
=
m
x.
(2.5)
2
x dt x
x
Even in some situations where there is no conservation of energy, analytical
mechanics applies. The simplest examples involve Lagrangians which depend
explicitly on the time t. Dissipation is not involved because we keep careful
track of the way that energy is entering or leaving the system.
Now for an example where the Lagrangian formalism is useful. Suppose we
wish to describe a free particle in spherical polar coordinates
L=
x = r cos sin
y = r sin sin
z = r cos .
(2.6)
and .
This requires
That is to say, we wish to derive the equations for r, ,
1
This problem was first considered by Eddington. See, however, exercise 3.
20
Lagrangian mechanics
an amount of calculation, but the amount shrinks if we perform the change of

variables directly in the Lagrangian:

m 2
m 2
x + y 2 + z 2 =
r + r 2 2 + r 2 sin2 2 .
(2.7)
2
2
Then we obtain the answer as the Euler-Lagrange equations from this Lagrangian. (Do the calculation both ways, and see!) This is often the simplest
way to perform a coordinate transformation even if the Lagrangian is not
known, so that one first has to spend some time in deriving it.
A famous example for which L 6= T V is that of an electrically charged
particle moving in an external electromagnetic field. This example is so important that we will give it in some detail. First of all, external signifies that
we are dealing with an approximation, in which we ignore that the presence of
the electrically charged particle will affect the electromagnetic field in which it
moves. In many situations, this is an excellent approximation. The equations
of motion to be derived are the Lorentz equations
L=
m
xi = e (Ei (x, t) + ijk x j Bk (x, t)) .
(2.8)
The epsilon tensor occurring here may be unfamiliar (but see exercise 1). For
the moment let me just say that the second term on the right hand side means
the cross product of the velocity and the magnetic field. With this hint you
should be able to follow the argument at least in outline, so we proceed. This
example is more tricky than the previous ones, since the force depends not only
on the position but also on the velocity of the particle (as well as explicitly on
time, but this is no big deal). It turns out that in order to derive the Lorentz
equation from a Lagrangian, we need not only one but four potentials, as
follows:
Ei (x, t) = i (x, t) t Ai (x, t)
Bi (x, t) = ijk j Ak (x, t) .
(2.9)
Here is known as the scalar potential and Ai as the vector potential. (They
are both parts of a relativistic four vector.) It is possible to show that the
following action yields the Lorentz equation when varied with respect to x:
S[x(t)] =
dt

mx 2
+ ex i Ai (x, t) e(x, t) .
2
(2.10)
Please verify this!

If we consider a time independent electric field with no magnetic field
present, the Lorentz equation reduces to the more familiar form
m
xi = ei (x) .
(2.11)
This has the same form as Newtons Law of Gravity, if the potential is specified
correctly. The reason why the full Lorentz equation is much more complicated
has to do with the special relativity theory. The magnetic field is a relativistic
2.2 Constrained systems
21
complication. The relativistic version of Newtons law of gravity is yet more

complicated, and is given by Einsteins general relativity theory.
An important difference between gravity and electricity, also in the nonrelativistic case, is that particles couple to gravity through the mass, and all
particles have mass while only some have electric charge. Moreover the mass
serving as charge for gravitational forces is the same as the mass occurring
on the left hand side of Newtons equations.

A strength of the Lagrangian formalism is the way it deals with constrained
systems. An example of a constrained system is the pendulum with a rigid rod,
which featured already in section 1.3. Deriving eq. (1.26) using a Cartesian
coordinate system and Newtons methods is not trivial, since we have to figure
out the constraint force acting in the direction of the rod. Using the Lagrangian
formalism we simply write down the Lagrangian as L = T V , change to polar
coordinates, and quickly arrive at the Euler-Lagrange equation in the form of
eq. (1.26). For a more complicated system, such as the double pendulum in
exercise 5, the latter method wins hands down.
Another problem in the same vein is that of a particle constrained to move
on the surface of a sphere
x2 + y 2 + z 2 = 1 .
(2.12)
The action is still given by

S[x, y, z] =
dt

m 2
x + y 2 + z 2 .
2
(2.13)
This action is a functional of three functions which are constrained to obey

the constraint (2.12), and we are going to extremize it only with respect to
variations that obey the constraint. In this particular case there is an easy way
to proceed. Using spherical polar coordinates the constraint becomes r = 1,
while the angular coordinates can be varied freely. From eq. (2.7) we see immediately that the Lagrangian describing the motion in the angular directions
is

m 2
+ sin2 2 .
(2.14)
2
Of course this depended on the lucky accident that we have a good parametrization of the constraint surface. Correct results would also be obtained if we
simply solve for
L=
z = z(x, y) =
1 x2 y 2 ,
(2.15)
and insert the result back into the action that describes the free particle, ie.
22
S[x, y] =
dt

m 2
x + y 2 + z(x,
y)2 .
2
(2.16)
Now we can vary x and y freely, except that they are not allowed to exceed
one in absolute value. The variations in z are now
z = x
z
z
+ y
,
x
y
(2.17)
and the equations of motion can be derived at the expense of some effort.
There are some unavoidable weaknesses here. From eq. (2.16) it appears as
if the configuration space were the unit disk in the plane, since x and y are not
allowed to take values outside this disk. Or perhaps the configuration space is
two copies of the unit disk, since there are two branches of the square root?
But the true configurations space is a sphere. What we see is a reflection of
the known fact that it is impossible to cover a sphere with a single coordinate
systemour equations have only a local validity. This kind of difficulties
will become more pronounced in the general problem we are heading for: Consider a Lagrangian L0 defined on an n dimensional configuration space, with
coordinates q1 , . . . , qn , and suppose that the system is confined to live in the
(n m) dimensional submanifold defined by the m conditions
I (q1 , . . . , qn ) = 0 ,
1Im.
(2.18)
Derive equations of motion consistent with this requirement. One way to do

this is to solve for m of the qs, q1 , . . . , qm say, by means of the m conditions
(2.18), and insert the result in the action. In general this will be a lot of hard
work, and the difficulties we had with coordinatizing the sphere will recur with
a vengeance.
The fact that the procedure avoids dealing with the constraint forces is a
weakness too. If we try to design a pendulum in such a way that the approximation of a totally rigid rod holds for the kind of motion the pendulum will
be subject to, we will want to know how strong the constraint force actually is. The method of Lagrange multipliers solves this problem, and at the
same time has the advantage that the difficulties with coordinatizing the constraint surface are postponed to a later stage. The claim we will verify is this:
Extremizing the action
S[q] =
dt L0 (q, q)
(2.19)
using only variations consistent with the constraints (2.18) is equivalent to

extremizing the action
S[q, ] =
dt L0 (q, q)
+ 1 1 (q) + + m m (q)
(2.20)
23
under arbitrary variations of the functions q and . The s are the Lagrange
multipliers, and are treated as new dynamical variables.
Indeed, when the action (2.20) is varied with respect to the s we obtain the
constraints (2.18) as equations of motion. When we vary with respect to the
qs the resulting equations will contain the otherwise undetermined Lagrange
multipliers, and it not obvious that these equations have anything to do with
the problem we wanted to consider. But they do. Consider the analogous
problem encountered in trying to find the extrema of an ordinary function
f (q) of the n variables q, subject to the m conditions (q) = 0. (Remember
suppression of indices!) First suppose that we use the constraints to solve
for m of the qsit will not matter which onesand call them y, leaving
n m independent variables x. The extrema of f (q) may be found through
the equations
0 = f = xx f + yy f ,
(2.21)
where, however, the variations y are not independent variations, but have to
be consistent with the constraints. In fact they are linear function of the xs,
given by the conditions
0 = = xx + yy .
(2.22)
This equation has to be solved for y and the result inserted into eq. (2.21),
which is therefore really an expression of the form x(x f + something else) =
0. It does not imply x f = 0.
Since = 0 for the variations we consider, nothing prevents us from rewriting eq. (2.21) in the form
0 = f = f + = x(x f + x ) + y(y f + y ) ,
(2.23)
where the s are arbitrary functions. The ys are still given in terms of the xs,
so it would seem at first sight that we cannot conclude that x f + x = 0,
butand here comes the punch linein fact we can, provided we choose the
so far arbitrary functions in such a way that y f + y = 0. Since the
division of the qs into xs and ys was arbitrary, we see that the restricted way
of finding the extremamaking variations consistent with the constraintsis
equivalent to solving the n + m equations
(q) = 0
q f + q = 0
(2.24)
for q and . But these are precisely the equations that we obtain from the Lagrange multiplier method, in which we do not care about the constraints while
varying the action! In all fairness though, we have not solved the equations,
we have just derived them in a convenient way.
As long as the constraints depend only on q (and not on q)
it is straightforward to generalize the argument from functions to functionals. From the
action
24
S[q, ] = S0 [q] +
dt (q)
(2.25)
we rederive the constraints, together with the equations of motion

S
S0
=
+ q = 0 .
q
q
(2.26)
This is the analogue of the second equation (2.24). Written out, if L = L(q, q)
and if there is only one constraint, this is

d L
L
=
+
dt qi
qi
qi
(2.27)
(q) = 0 .
(2.28)
These equations have a simple interpretation. The constraint defines a surface

in configuration space. We have modified the unconstrained systems by adding
an extra force term i to the equations. This force is directed along the
gradient of the constraint function, which means that it acts in a direction
orthogonal to the constraint surface. To ensure that the trajectory is confined
to the surface we must choose the strength of the force (given by ) in such a
way that this is ensured. In other words, once we have solved these equations
we know the strength of the constraint forces.
We have also refrained from committing us to any coordinate system adapted
to the specific form of the constraint surface. That this is an advantage becomes evident when we return to the problem of the particle on a sphere,
starting from the Lagrangian
L=

m 2
x + y 2 + z 2 + x2 + y 2 + z 2 1 .
2
(2.29)
The equations of motion (in these inertial coordinates) are

m
x = 2x
m
y = 2y
m
z = 2z .
(2.30)
By inspection we see that there are three constants of the motion,

Jx = y z z y
Jy = z x xz
Jz = xy y x .
(2.31)
At this point we go over to polar coordinates, using eqs. (2.6) with r = 1. The
constants of the motion become
Jx = sin cos cos sin
Jy = cos sin cos sin

(2.32)
2
Jz = sin .
2.3 Symmetries
25
It is possible to check directly, using the equations of motion for and , that
these are constants of the motionbut only Jz is obviously conserved. The
coordinate system (, ) somehow hides the others.
By the way, the kinetic energy can be expressed as
T =
m

m 2
+ 2 sin2 =
Jx2 + Jy2 + Jz2 .
2
2
(2.33)
This is the angular momentum squared.

So far we have dealt only with holonomic constraints, that is constraints
involving the configuration space variables only. But consider a ball moving
without friction across a plane. The configuration space has five dimensions:
the position (x, y) of the center of mass, and three angular coordinates describing the orientation of the ball. Now suppose instead that the ball rolls without
slipping. If we are given the position of the center of mass as a function of time
then the motion of the ball is fully determined, which suggests that x and y are
the true degrees of freedom, and that the constrained configuration space is
two dimensional. But the situation is more complicated than that (and cannot be described by holonomic constraints). It is impossible to solve for the
angular coordinates in terms of x and y. Indeed from our experience with such
things we know that the orientation of the ball at a given point depends on
how it got there. Mathematically there is a constraint relating the velocity of
the center of mass to the angular velocity; the ball spins in a way determined
by the motion of its center of mass. Constraints that cannot be expressed as
conditions on the configuration space are called anholonomic. The Lagrange
multiplier method can be generalized so that it applies to this case as well,
but we will not do so here.
2.3 Symmetries
Let us return to Newtons Third Law. It amounts to a restriction on the kind
of forces that are allowed in the second law, and implies that there exist a
set of constants of the motion, namely the momenta. (The terminology is a
little unfortunate, since we will soon introduce something called canonical
momenta. They are indeed identical with the conserved momenta in simple
cases, but logically there need be no connection.) Constants of the motion
are useful when trying to solve the equations of motion, and Emmy Noether
proved a theorem explaining when and why they exist. We present the proof
for a Lagrangian of the general form L = L(q, q),
and afterwards we discuss a
simple example. Let us say at the outset that the argument is quite subtle.
Consider first an arbitrary variation of the action. According to eq. (2.3)
the result is
S =
t2
t1
dt q
L
d L
q
dt q

t
L 2
+ q
.
q t1
(2.34)
26
In deriving the equations of motion the variations q(t) were restricted in such
a way that the boundary terms vanish. This time we do something different.
The variations are left unrestricted, but we assume that the function q(t)
that we vary around obeys the Euler-Lagrange equations. Then the only nonvanishing term is the boundary term, and
S = (Q(t2 ) Q(t1 )) ,
Q(t) q
L
.
q
(2.35)
Here and in the following is the constant occuring in q = f , where f is an

arbitrary function of t. The point is to ensure that there is nothing infinitesimal
about Q.
So far nothing has been assumed about the variations. Now suppose that,
for the given Lagrangian, there exists a set of variations q of some specified
form
q = q(q, q)
,
(2.36)
such that for these special variations

S = 0 .
(2.37)
It is understood that the Lagrangian is such that eq. (2.37) holds as an identity,
regardless of the choice of q(t), for the special variations q. (Note that, given a
Lagrangian, it is not always the case that such variations exist. But sometimes
they do.)
Next comes the crux of the argument. Consider variations of the particular
kind that makes eq. (2.37) hold as an identityso that q = f is a known
functionand restrict attention to q(t)s that obey the equations of motion.
With both these restrictions in force, we can combine eqs. (2.37) and (2.35)
to conclude that
0 = S = (Q(t2 ) Q(t1 )) .
(2.38)
The times t1 and t2 are arbitrary, and therefore we can conclude that Q =
Q(q, q)
is a constant of the motion.
What this theorem does for us is to transform the problem of looking for
constants of the motion to the problem of looking for variations under which
the variation of the action is identically zero. Before we turn to examples
we generalize the argument slightly, and state the theorem properly. Thus,
suppose that there exists a special form of q, such that
S =
t2
dt
t1
d
(q, q)
.
dt
(2.39)
Here can be any functionthe important and unusual thing is that the
integrand is a total time derivative. Then the quantity Q, defined by
2.3 Symmetries
Q(q, q)
= qi
L
(q, q)
qi
27
(2.40)
is a constant of the motion. This is easy to see along the lines we followed
above.
The theorem can now be stated as follows:
Noethers theorem: To any variation for which S takes the form (2.39), there
corresponds a constant of the motion given by eq. (2.40).
We will have to investigate whether Lagrangians can be found for which such
variations exist, otherwise the theorem is empty. Fortunately it is by no means
empty, indeed eventually we will see that all useful constants of the motion
arise in this way.
For now, one examplebut one that has many symmetrieswill have to
suffice. Consider a free particle described by
m
x i x i .
2
Since only x appears in the Lagrangian, we can choose
L=
xi = i ,
(2.41)
(2.42)
where i is independent of time. Then the variation of the action is automatically zero, Noethers theorem applies, and we obtain a vectors worth of
conserved charges
Pi =
L
= mx i .
x i
(2.43)
We use the letter P rather than Q because this is the familiar conserved
momentum vector whose presence is postulated in Newtons Third Law.
Another set of three conserved charges can be found easily, since
xi = ijk j xk
S = 0 .
(2.44)
Here i is again independent of t, and ijk is the totally anti-symmetric epsilon tensor. Noethers theorem now implies the existence of another conserved
vector, namely
Li = ijk xj x k .
(2.45)
This is the angular momentum vector.

We know that there is at least one more conserved quantity, namely the
kinetic energy. Actually there are several, but the story now becomes a bit more
complicated because we have to deal with variations for which the variation
of the Lagrangian is a total derivative, as in eq. (2.39), rather than zero. Thus
28
xi = x i
S =
dt
d m 2
x
.
dt 2
(2.46)
Using eq. (2.40) we obtain the constant of the motion
m
x i x i .
(2.47)
2
This is the conserved energy of the particle. There is yet another conserved
quantity that differs from the others in being an explicit function of timebut
its total time derivative vanishes since it also depends on the time dependent
dynamical variables. Thus
E=
xi = i t
S =
dt
d
(mi xi ) .
dt
(2.48)
Eq. (2.40) gives the conserved charge

Qi = mxi tmx i ,
(2.49)
and it is easy to check that its total time derivative vanishes as a consequence
of the equations of motion. Our analysis of the free particle ends here, but we
will return to it in a moment, to show that the conserved quantities have a
clear physical meaing.
What does it all mean? What does it mean for an action S[q(t)] to admit
variations q(t) leaving the action unaffected? To see this, select a solution
q(t) of the equations of motion. We know that this gives an extremum of
the action. Then consider q (t) = q(t) + q(t), where the variation is of the
special kind that leaves the value of the action unchanged. Obviously then
S[q (t)] = S[q(t)], so that the extremum is not an isolated point in the space
of all qs, but rather occurs for a set of qs that can be reached from each other
by means of iteration of the special variation q(t), In other words, given a
particular solution of the equations of motion, we can get a whole set of new
solutions if we apply the special variation, without going through the work of
solving the equations of motion again. This leads to an important definition:
A symmetry transformation is any transformation of the space of functions
q(t) having the property that it maps solutions of the equations of motion to
other solutions.
This is not a property of the individual solutions, but of the set of all solutions.
The special variations occurring in the statement of Noethers theorem are
examples of symmetry transformations. Given the converse of the statement
that we proved, namely that any constant of the motion gives rise to a special
variation of the kind considered by Noether, we observe that any constant of
the motion arises because of the presence of a symmetry. (There is a converse
of this statement that we will come to in section 8.6.)
2.3 Symmetries
29
Let us interpret the symmetry transformations that we found for the free
particle, beginning with eq. (2.42). This is clearly a translation in space. Therefore momentum conservation is a consequence of translation invariance. It is
immediate that we can iterate the infinitesimal translations used in Noethers
theorem to obtain finite translations, and the statement is that given a solution
to the equations of motion all trajectories that can be obtained by translating
this solution are solutions, too. To be definite, given that (vt, 0, 0) is a solution for constant v, (a + vt, b, c) is a solution too, for all real values of (a, b, c).
Translation invariance acquires more content when used in the fashion of Newtons third law, which we can restate as the action for a set of particles has
translation symmetry. For free particles this is automatic. When interactions
between two particles are added, the law becomes a restriction on the kind of
potentials that are admitted in
m1 2 m2 2
x1 +
x V (x1 , x2 ) .
(2.50)
2
2 2
Indeed invariance under (2.42) requires that V (x1 , x2 ) = V (x1 x2 ), which is
a strong restriction.
Eq. (2.44) expresses the fact that the Lagrangian has rotation symmetry,
while eq. (2.46) is an infinitesimal translation in time: Given a solution x(t),
the function
L=
x (t) = x(t + t0 ) = x(t) + t0 x(t)

+ o(t20 )
(2.51)
is a solution too. So we can make the elegant summary that conservation of

momentum, angular momentum and energy are consequences of symmetries
under translations and rotations in space, together with translations in time.
Eq. (2.48) expresses invariance under boosts, since it changes all velocities
by a constant amount. The free particle is exceptional because we can reach
any solution by a symmetry transformation, starting from any given solution.
Having said all this, it is not true that every symmetry gives rise to a
constant of the motion. Discrete symmetries like reflections, that do not arise
by iterating an infinitesimal symmetry, are counterexamples, and we will see
another in section 4.5.
To sum up, symmetries are important from two quite different points of
view. Given the equations they facilitate the search for solutions, but they
also facilitate the search for the correct equations (if we believe that they
should exhibit a certain symmetry). Noethers theorem is a tool for discovering symmetries, as well as for deducing their corresponding constants of the
motion.
Problem 2.1
Prove the epsilon-delta identity

ijk kmn = im jn in jm .
What does it look like in the cross product notation?
30
Problem 2.2
Derive Lorentz equation from the action (2.10).
Problem 2.3
Prove that the equation

m
x + x + kx = 0
cannot be derived from an autonomous Lagrangian, that is to say a Lagrangian that

does not explicitly depend on time. Then derive the equation from a Lagrangian of
the form L = L(x, x,
t).
Problem 2.4
Consider the Lagrangian
1
Mij (q)qi qj ,
2
where the matrix elements of Mij depend on the configuration space coordinates,
and the matrix is assumed to have an inverse Mij1 . Write down the Euler-Lagrange
equations and solve for the accelerations.
L=
Problem 2.5
Write down the Lagrangian for a double pendulum. (The rod
of the second is attached to the bob of the first. Bobs are heavy, the rigid rods not.)
How many constants of the motion can you find?
Problem 2.6 Take N positive numbers summing to one, p1 +p2 + +pN = 1.

Their geometric mean is defined as (p1 p2 pN )1/N . What is the maximum of the
geometric mean?
Problem 2.7
Check that the result from eq. (2.17) is the same as that
obtained from the recipe in eq. (2.22). Beware of changes in notation!
Problem 2.8
Consider a particle with kinetic energy T = m(x 2 + y 2 z 2 )/2,
and constrain it to the hyperboloid x2 + y 2 z 2 = 1, z > 0. Treat this both with
coordinates adapted to the hyperboloid and with the Lagrange multiplier method.
Show that the kinetic energy is positive and find three constants of the motion.
Problem 2.9
In the brachistrone problem one considers a particle sliding
along a curve in the x-z-plane (z is vertical) under the influence of gravity. Choose
this curve so that the time of descent from (x, z) = (x0 , z0 ) to the origin is minimal.
Problem 2.10
Consider the Lagrangian
1 2
q q n ,
2
where is a real number and n is an integer. Determine those values of n for which
the Lagrangian transforms into a total derivative under
L=

q
q = tq
.
2
This is known as conformal symmetry.
Problem 2.11
For a free particle, consider the action integral
2.3 Symmetries
S=
t2
t1
31
m 2
x dt .
2
Evaluate this integral for an x(t) that solves the equations of motion, and express the
answer as a function of t1 , t2 , and the initial and final positions x1 and x2 . Repeat
the exercise for a harmonic oscillator.
3 Interlude: Conic sections
The theory of conic sections was one of the crowning achievements of the
Greeks. Their results will be important in the gravitational two body problem,
but the theory is no longer as well known as it deserves to be, so here is a brief
account.
By definition a conic section is the intersection of a circular cone with a
a plane. The straight lines running through the apex of the cone are called
its generatorsand we will consider a cone that extends in both directions
from its apex. If you like, it is the set of one dimensional subspaces in a three
dimensional vector space. Generically, the plane will intersect the cone in such
a way that every generator crosses the plane once, or in such a way that exactly
two of the generators miss the plane. Apollonius proved that the intersection is
an ellipse in the first case, and a hyperbola in the second. There is a borderline
case when exactly one generator is missing. Then the intersection is a parabola.
We ignore the uninteresting case when the plane goes through the apex of the
cone.
This is all very easy if we use the machinery of analytic geometry. For
simplicity, choose a cone with circular base, symmetry axis orthogonal to the
base, and opening angle 90 degrees. It consists of all points obeying
x2 + y 2 z 2 = 0 .
(3.1)
Without loss of generality, the plane can be described by

cx + z = d .
(3.2)
Inserting the solution for z in the equation it is easy to see that the intersection
is either an ellipse, a hyperbola, or a parabolaprovided you recognize their
equations, as I assume. The section is a circle if c = 0 and a parabola if c = 1.
It is an interesting exercise to prove this in the style of Apollonius. Let the
cone have arbitrary opening angle. Take the case when the plane intersects
every generator once in the upper half of the cone. Place two spheres inside
the cone, one above and one below the plane, and let them grow until each
touches the plane in a point and the cone in a circle. (See Fig. 3.1.) This
clearly defines the spheres uniquely. Denote the points by F1 and F2 , and the
circles by C1 and C2 . Now consider a point P in the intersection of the cone
Interlude: Conic sections
33
Figure 3.1. A vertical cross section through Apollonius proofbut to understand the proof you have to think in three dimensions.
and the plane. The generator passing through P intersects the circles C1 and
C2 in the points Q1 and Q2 . Now the trick is to prove that the distance P F1
equals the distance P Q1 , and similarly the distance P F2 equals the distance
P Q2 . This is true because the distances measure the lengths of two tangents
to the sphere, meeting at the same point. It then follows that the sum of the
distances P F1 and P F2 is constant and equal to the length of the segment of
the generator between the circles C1 and C2 , independently of which point P
on the intersection we choose. This property defines the ellipse. This is the
proof that the intersection between the cone and the plane is an ellipse with
its foci at F1 and F2 . If you are unable to see this, consult an old fashioned
geometry book.
Figure 3.2. An ellipse, with eccentricity 0.58, semi-major axis a, semi-minor

axis b, and latus rectum p. The directrix appears on the right.
34
For our purposes it is convenient to define the ellipse somewhat differently.

An ellipse of eccentricity e < 1 can be defined as the set of points whose
distance from a given point, called a focus, is e times the distance to a straight
line, called a directrix. For the circle e = 0, and the directrix is at infinity.
The latus rectum of an ellipse is a chord through the focus parallel to the
directrix, and has length 2p. Now place the origin of a coordinate system at
that focus, with the x-axis pointing towards the directrix. The distance of a
point on the ellipse to the focus is
r = e(distance to the directrix) = e
p

x .
(3.3)
e
(To see this, note that the distance from the focus to the directrix is p/e.)
Otherwise expressed
p
= 1 + e cos ,
(3.4)
r
where = 0 gives the point closest to the directrix. For a general point on the
ellipse we find
r = p er cos
x2 + y 2 = r 2 = (p ex)2
(x + ea)2 y 2
+ 2 =1,
a2
b
(3.5)
where
p
b2 = pa = (1 e2 )a2 .
(3.6)
1 e2
The major axis of the ellipse has length 2a, and the minor axis has length 2b.
The ellipticity is given in terms of these by
a
a2 b2
.
(3.7)
a2
Finally the distance between the center and the focus equals ea. To see this
we set = 0 and = in eq. (3.7), and calculate
e2 =
r() r(0)
p
=
2
2
1
1
1e 1+e
= ea .
(3.8)
For some further information consult exercise 1.

Similar treatments can be given for the hyperbola, for which e > 1, and for
the parabola, for which e = 1.
In our study of the two-body problem we will find it interesting to relate
an ellipse centred at a focus to an ellipse centred at the origin. The latter is
described in Cartesian coordinates by the complex trajectory
w(t) = a cos t + ib sin t .
(3.9)
The parameter t must not be confused with the angle between the radius
35
vector and the x-axis. Surprisingly, if we square this ellipse we obtain an ellipse
with its focus at the origin. First we see that
a2 + b2
a2 b2
cos 2t + iab sin 2t +
.
2
2
Using eq. (3.7) we see that the eccentricity E of the new ellipse is
Z(t) = w2 =
a2 b2
.
a2 + b2
So we can rewrite the equation for the new ellipse as
E=
Z(t) = A cos 2t + iB sin 2t + EA ,
(3.10)
(3.11)
(3.12)
where
a2 + b2
,
B ab ,
EA = A2 B 2 .
(3.13)
2
E is the eccentricity of an ellipse with semi-major axis A and semi-minor axis
B, and consequently EA is the distance between its focus and its centre. This
is again an ellipse, but shifted to be centred at one of its fociin the sense
that the angle is now seen from a focusand traversed twice as the original
ellipse is traversed once.
This trick was introduced by Karl Bohlin, working at the Pulkovoobservatory in Russia in 1911. His point was that the transformation w = Z is not
analytic at the origin. This enabled him to deal with collisions between point
particles, thought of as limiting cases of elliptical orbits whose eccentricity
approaches 1. At the collision the particle presumably reverses its direction,
but this is a rather singular occurence. In terms of the variable w it is an
undramatic event.
A
Problem 3.1
Using the definition in terms of the directrix, prove that the
sum of the distances from the two foci to a point on an ellipse is constant. Also prove
that the area of an ellipse equals ab.
Problem 3.2 Place a lamp at one focus of an ellipse, and let the circumference
of the ellipse act as a mirror. Prove that all planar light rays reconverge at the same
time at the other focus. Try to do this in two different ways: by means of a calculation,
and by means of an argument that makes it all obvious.
Problem 3.3
Consider a family of highly eccentric Kepler ellipses Z(t), and
take the limit representing colliding particles. Exactly
what happens at the collision
p
when it is described by the Hooke ellipse w(t) = Z(t)?
4 The central force two-body problem
Johannes Kepler spent his life pondering the observations of the solar system
made by Tycho Brahe, and found that the motion of the planets around the
sun follows three simple rules:
1. A planet moves along an ellipse with the sun in one of the foci.
2. The radius vector covers equal areas in equal times.
3. The square of the period of all the planets is proportional to the cube of
their major axes.
To appreciate Keplers work fully, note that there are important facts about
the solar system (such as what the distances are) that do not follow simple
rules. Moreover the observational data gave the planetary orbits projected on
a sphere centred at a point which itself moves along an ellipse around the sun,
so it was not obvious that they admitted of a simple description at all.
Newton derived Keplers laws from his own Laws, with the necessary assumption that the force between the planets and the sun is directed along
the radius vector (the force is central) and is inversely proportional to the
square of the distance. This remains the number one success story of physics,
so we should be clear about why this is so. Naively Keplers laws may seem
simpler than Newtons, but this is not so, for at least two reasons. One is that
Newtons laws unify a large body of phenomena, from the motion of planets
to the falling of stones close to the Earth. The other reason is that improved
observations reveal that Keplers laws are not quite exact, and the corrections
can be worked out mathematically from Newtons laws.
For Mercury (which is hard to observe) the eccentricity e = 0.21, for the
Earth e = 0.02, and for Mars e = 0.09. Keplers main concern was with Mars.
If e = 0 the ellipse becomes a circle. Unbound motion through the solar system
is described by hyperbolas with e > 1, but Kepler did not know this.
4.1 The problem and its formal solution

We want to derive Keplers laws. As an approximation we assume that it
suffices to treat the planets independently of each other. Moreover we assume
that the precise shapes of the sun and the planets are unimportant and that
4.1 The problem and its formal solution
37
they can be approximated as being pointlike. (In the Principia, Newton proved
from properties of the inverse square law that this approximation is exact for
spherical bodies.) Later on, we can go back to these assumptions and see if we
can relax themthis will give the corrections referred to above.
So we have decided that the configuration space of our problem has six
dimensions, spanned by the positions of the sun (X) and one planet (xP ), and
we try the Lagrangian
M
mp
Xi Xi +
x P i x P i V (X, xP ) .
(4.1)
2
2
Depending on the form of the function V we may have to exclude the points
X = xP from the configuration spacewe insist that the function V takes
finite values only as a function on configuration space. Anyway this will give six
coupled second order differential equations. In general they will not admit any
simple solutions. Keplers work implies that the solutions should be simple,
so we try to build some symmetries into the problem. We aim for at least
six conserved quantitities, one for each degree of freedom, since this should
result in a soluble problem. Conservation of momentum and energy is already
postulated, so we need two more. The answer is rotational symmetry. It is not
an objection that an ellipse is not symmetric under rotations. All we need is
that if a particular ellipse is a solution, then any ellipse which can be obtained
from it by means of rotations is a solution too, even if there is no planet moving
along it due to the choice of initial conditions. It might seem that rotational
symmetry is overdoing it, since it will yield three conserved quantitities, but
for reasons fully explained by Hamilton-Jacobi theoryonly two of these are
really useful.
With translational and rotational symmetry in place, we find
L=
V (X, xP ) = V (X xP ) = V (|X xP |) .
(4.2)
This works for any function V of one variable. Next we introduce coordinates
xi = xP i Xi which are invariant under translations, together with coordinates
describing the centre of mass. Then the centre of mass coordinates decouple,
and their equations can be solved and set aside. There remains a Lagrangian
for a one-body problem, involving only three degrees of freedom:
m 2
x V (|x|) .
(4.3)
2 i
Here m is the reduced mass, almost equal to the mass of the planet since the sun
is very heavy in comparison. The coordinate xi vanishes at the centre of mass
of the system, which is well inside the sun, and can be approximately identified
with the centre of the sun. This maneouvre should be familiar from elementary
mechanics, I just want to emphasize that it is translational symmetry in action.
Rotational symmetry implies the existence of a conserved vector
L=
Li = mijk xj x k .
(4.4)
The central force two-body problem
38
This vector is orthogonal to both x and x,

which means that the motion
is confined to a plane orthogonal to the angular momentum vector. We are
down to a two dimensional configuration space. To take maximal advantage of
spherical symmetry we introduce spherical polar coordinates, chosen so that
the plane containing the orbit is at = /2. The Lagrangian simplifies to
m 2
(r + r 2 2 ) V (r) .
(4.5)
2
We have used the constant direction of the angular momentum vector. But
its magnitude is constant too. This happens because the Lagrangian (4.5) is
invariant under translations in the angle . Using Noethers theorem we find
the constant of the motion
L=
l=
L
= mr 2 .
(4.6)
This equation is of considerable interest in itself. It says that

r 2
l
A =
=
= constant ,
2
2m
(4.7)
where A is the area covered by the radius vector per unit time. But this is
Keplers Second Law, which therefore holds for all central forces. We are on
the right track!
Together with Keplers second law, energy conservation is enough to solve
the problem. Using eq. (4.6) the conserved energy is

m 2
mr 2
l2
+ V (r) = constant .
r + r 2 2 + V (r) =
+
2
2
2mr 2
This gives the formal solution
E=
dt = q
dr
2
m
l2
2mr 2
.
V (r)
(4.8)
(4.9)
It may or may not, depending on our choice of V (r), be possible to do the

integral in terms of elementary functions, but anyway this equation determines
the function r(t), and hence solves the problem. To find (t) we combine eqs.
(4.6) and (4.9), and get
d =
ldt
ldr
= q
.
2
mr
l2
r 2 2m E 2mr
V
(r)
2
(4.10)
This equation determines (r(t)), and the central force problem is thereby
fully solved at the formal level. If we are only interested in the form of the
orbits, and not the time development, eq. (4.10) is all we needit will give us
(r), and after inversion r(), which is the equation for the form of the orbit.
4.2 Existence and stability of circular orbits
39
4.2 Existence and stability of circular orbits

There is a special kind of solution that we can look for directly, with no great
expense of effort, namely circular orbits. From the expression (4.8) we see
that the two body central force problem has been reduced to one dimensional
motion in the effective potential
l2
+ V (r) .
(4.11)
2mr 2
A simple case is V (r) = 0, i.e. no force at all. In the effective one dimensional
problem this corresponds to a repulsive Vef f the particle comes in from infinity, reaches a minimum value of r, and then disappears to infinity again. If
we want a bound orbit the potential V (r) must be attractive.
A circular orbit is one for which r = 0 identically, which means that the
particle is sitting at the bottom of the effective potentialif it does have a
bottom. The radius r of the circular orbit must obey
Vef f (r) =
Vef f (r) = 0 .
(4.12)
If this happens at a local maximum of Vef f the solution is unstable, and

unlikely to be realized in Nature. The orbit is stable under small perturbations
if and only if
Vef f > 0
(4.13)
at the value of r for which Vef f = 0.

We look into these equations for the special choice
l2
kr ,
2mr 2
with arbitrary. The radius of the circular orbit is found to be
V (r) = kr
r=
Vef f (r) =
l2
km
1
+2
(4.14)
(4.15)
This makes sense only if l2 6= 0a question of initial conditionsand k < 0.

(The case = 2 is obviously special.) Then the second derivative will be
positive if and only if
> 2 .
(4.16)
Thus stability of the circular orbit requires that the force does not fall off too
quickly with distance.
There is still the question whether small departures from the circular orbit
will give rise to ellipses, or to something more complicated. This is really a
question about the ratio between the time it takes for the planet to complete
40
a full revolution in , to the time it takes to complete a full oscillation in r. If

the orbit is an ellipse centred at a focus these times must be equal, and this
is likely to happen only for a very special V (r). Compare to the discussion of
Lissajous figures in section 1.3. For general bounded motion the amount by
which the perihelion precesses during one period of the radial motion follows
from eq. (4.10). It is
= 2
rmax
rmin
r2
ldr
.
2m (E Vef f (r))
(4.17)
For the planets, Keplers first law requires that = 2.

A final comment: if we do choose the exponent = 1 we have the problem
that the energy is unbounded from below. We now see that this problem cannot
be too serious, because
l2
k
(4.18)
2
2mr
r
is in fact bounded from below whenever l 6= 0. The case when l = 0 is indeed
troublesome from a physical point of view: the two bodies will collide, and we
do not have a prescription for what is to happen after the collision. The case
of coinciding particles is not included in our configuration space.
Vef f (r) =
4.3 Keplers First Law

What force laws are consistent with Keplers First Law? Hookes law does give
elliptical orbits. This is most easily seen by transforming the Lagrangian back
to Cartesian coordinates:
m 2
m
(r + r 2 2 ) kr 2 = (x 2 + y 2 ) k(x2 + y 2 ) .
(4.19)
2
2
This is two harmonic oscillators of equal frequencies, and the corresponding
Lissajous figures are indeed ellipses. They are not the right kind of ellipses
however, since they are centred at the origin. Let us call them Hooke ellipses.
They can be related to Kepler ellipses (centred at a focus) by means of Bohlins
trick (chapter 3). If
L=
w = |w|ei = a cos t + ib sin t
(4.20)
is a Hooke ellipse, then

a2 b2 a2 + b2
+
cos 2t + iab sin 2t .
(4.21)
2
2
This is indeed an ellipse centered at a focus, as we saw in chapter 3. For the
Hooke ellipse w(t) we know that the force law is
Z = w2 = |w|2 e2i =
4.3 Keplers First Law
w
= w
|w|
2 + |w|2 = 2 .
41
(4.22)
Moreover Keplers second law holds, so that

|w|2
d
= constant .
dt
(4.23)
This relates the parameter t to the angle between the radius vector and the
major axis. The idea now is to introduce a new time , related to t in such a
way that Keplers second law holds also for the ellipse we get when we square
the Hooke ellipse. Thus, remembering that the phase of Z is 2, we require
2|Z|2
d
= constant .
d
(4.24)
A suitable choice of the two constants gives the desired relation

d
|Z|2
=
= |w|2
dt
|w|2
d
1 d
=
.
d
|w|2 dt
(4.25)
At this point then we have obtained a function Z( ) that describes motion in

accordance with Keplers First and Second Laws. It only remains to investigate what differential equation it obeys. But using eq. (4.22) this is a simple
calculation:
d2 Z
1 d
=
d 2
|w|2 dt
1 dw2
|w|2 dt
2 d
|w|2 dt

w
Z
= = 4 3 ,
w
|Z|
(4.26)
where is the constant energy of the Hooke ellipse. This is precisely Newtons
force law for gravity. So we conclude that Keplers First and Second Laws
together imply the inverse square law, with the potential
V (r) =
k
.
r
(4.27)
There is no other solution. Keplers First and Second Laws hold if and only if
the inverse square law holds for the force. The argument is watertight because
every Kepler ellipse can be obtained from a Hooke ellipse using Bohlins trick,
and a Hooke ellipse arises only in the harmonic oscillator potential.
To confirm our conclusion let us go back to eq. (4.10), which gives a formal
solution for the form of the orbit. We choose eq. (4.27) for V (r), and we also
perform the substitution
u=
The result is
1
r
du =
dr
.
r2
(4.28)
42
ldu
d =
.
2mE l2 u2 + 2mku
(4.29)
This defines u as a trigonometric function of , with the energy E < 0 and

the phase 0 as undetermined integation constants. In fact
l2 u
mk
= 0 arccos q
1+
2El2
mk2
(4.30)
Inverting this, and cleaning up the answer a little, we obtain

r
l2 u
2El2
l2
=
=1+ 1+
cos ( 0 ) .
(4.31)
mkr
mk
mk2
The constant 0 is the value of the coordinate for which the planet is at
its perihelion, that is when it is closest to the sun. Comparing to eq. (3.7) we
read off that the eccentricity of the ellipse is
e=
1+
2El2
.
mk2
(4.32)
The semi-major axis of the ellipse is

a=
p
l2 mk2
k
=
=
.
2
2
1e
mk 2|E|l
2|E|
(4.33)
The solution remains valid also for E > 0, in which case it describes a hyperbola with e > 1. Physically this is an unbound trajectory, like that of a
spaceship heading for the stars.
4.4 Keplers Third Law

Keplers Third Law awaits proof. If you recall the basic facts about ellipses
the following is easy: since the areal velocity is constant the period T is simply
related to the area of the ellipse. Starting with eq. (4.7) and recalling exercise
3.1 we obtain
l
T = A = ab .
(4.34)
2m
Remembering that b2 = ap, and using eq. (4.31) to identify the latus rectum
p, gives
4 2 m2 a3 p
m
= 4 2 a3 .
(4.35)
2
l
k
This is Keplers Third Law, since k = GmM so that the proportionality
T2 =
4.5 Self-similarity and the virial theorem
43
constant is independent of what planet we are looking at, to the approximation

that we can take M equal to the mass of the Sun.
Here is a more involved proof, using the full force of our solution (4.9). We
rewrite it using our expressions for a and e:
dt =
ma
q
k
rdr
l2
2m|E|
+ 2ar r 2
ma
rdr
p
.
2
2
k
a e (r a)2
(4.36)
We can do the integral if we can find a substitution that simplifies the integral
Z
rdr
,
1 (r 1)2
(4.37)
and such a substitution is readily found. Reinserting the constants we set

r = a + ae cos
(4.38)
Now we can do the integral. With a suitable choice of the integration constant
we obtain
t=
ma3
( + e sin ) .
k
(4.39)
Taken together, eqs. (4.38-4.39) provide a parametric representation of the

orbit, and we can read off its period
T = 2
ma3
.
k
(4.40)
This is Keplers Third Law once again.

Keplers Third Law says that, given a solution, one can simply enlarge it to get
another solutionprovided one also slows down the rate at which things are
happening. It is really a consequence of mechanical similarity or self-similarity,
a kind of symmetry not covered by Noethers theorem. It arises as follows. Take
the Lagrangian
m
L=
2
dq
dt
2
V (q) ,
(4.41)
where we do not use the dot notation because we will soon have two different
time parameters to reckon with. Assume that the potential is homogeneous
of degree , meaning that there exists a real number such that for any real
non-zero number
44
V (q) = V (q) .
(4.42)
There could be several variables qi . For simplicity I write only q. Let us also
change the time scale, and define a new function q by
q(t) q (t ) = q(t) ,
t =
2
2
t.
(4.43)
It follows that
dq
dq
dt d
.
= (q(t)) = 2
dt
dt dt
dt
(4.44)
We can now check that our rescalings represent a symmetry because, under
this transformation,

dq
L q,
dt
dq
L q,
d

dq
= L q,
dt
(4.45)
This has the effect of changing the value of the action with a constant factor,
and it follows that q (t ) is an extremum of S[q (t )] if q(t) is an extremum of
S[q(t)]. In this sense rescaling is a symmetry of the action. (Compare problem
2.10. If you find the argument difficult, you can check directly that q (t ) is a
solution whenever q(t) is.)
The harmonic oscillator has a potential V q 2 , homogeneous with = 2.
Scaling symmetry is present with t = t. Given a solution q(t) there is another
solution that is a blown up version of this, with amplitude a factor of larger.
Because t = t the period of the oscillations are unaffected by the scaling,
and we seewithout looking at any explicit solutionsthat the period of
the oscillations are independent of their amplitudes. Galilei first made this
observation while celebrating mass in the cathedral of Pisa.
Newtons law of gravity uses a homogeneous potential with = 1, so
similarity holds with t t = 3/2 t. Two ellipses with the same shape (and
the planetary orbits are all close to circular) will therefore have their periods
and their axes related by
R R = R
T T = 3/2 T
T 2
R3
=
.
T2
R3
(4.46)
This is Keplers Third Law for the third time.

Another dramatic theorem can be proved for self-similar systems. It is called
the Virial Theorem, and relates the time averages of the kinetic and potential
energies to each other. If it exists, the time average of a function f (t) is defined
by
1
hf i lim
t t t0
dt f (t ) .
t0
(4.47)
45
For the argument to follow it is important that the time average of the derivative of a bounded function is zero, i.e.

df
dt
1
(f (t) f (t0 )) = 0
t t
= lim
(4.48)
whenever f (t) < for all t.

We are ready to study the time average of the kinetic energy, given the
assumptions that the system obeys Newtons law
m
xi = i V (x) ,
(4.49)
that the potential is homogeneous of degree , that the motion is bounded

in space, and that the velocities are everywhere finite. On the other hand we
are not restricting the index i. It could run between 1 i 3N , in which
case we are actually studying an N -body problem; this could be a cluster of
galaxies under the tentative assumption that the cluster is a bound system,
or it could be 1023 atoms confined in a box. For the argument we will need
Eulers theorem on homogeneous functions states. Regardless of the number
of variables it states that
V (x) = V (x)
xi i V (x) = V (x) .
(4.50)
(Proof: Take the derivative with respect to , and then set = 1.)
Now the calculation is easy:
2 hT i = mx
d
(mxi x i ) mxi x
i
dt
= hxi m
xi i =
(4.51)
= hxi i V (x)i = hV (x)i .

This is the conclusion we were after.
For bounded motion in homogeneous potentials
2 hT i = hV i ,
(4.52)
where is the degree of homogeneity of V (x).

For the inverse square law the virial theorem implies that
h2T + V i = 0
hEi = hT + V i = hT i 0 .
(4.53)
This is the familiar fact that motion bounded by gravity can take place only
if the total energy is negative. For the harmonic oscillator we deduce that the
time averages hT i and hV i are equal.
The virial theorem has been used by astronomers to estimate the masses
46
of clusters of stars and clusters of galaxies, assuming that they are gravitationally bound. This led to the first evidence for dark matter.1 Note also the
counterintuitive fact that if energy leaves a self-gravitating cluster of particles
its energy hEi becomes more negative, which means that its average kinetic
energy hT i grows. In some sense the gas becomes hotter when energy is lost.
The calculation in eq. (4.51) is of interest even for non-potential forces, if
we break it off after the first line:
2 hT i = hxi Fi i .
(4.54)
If the forces are the constraint forces keeping an ideal gas contained inside a
box, we can use this relation to deduce the ideal gas law. We turn the sum
into an integral, recall the definition of the pressure P as force per unit area,
and apply Gauss law to the result:
2 hT i = P
dAi xi = P
dV i xi = 3P V .
(4.55)
If we are willing to identify hT i with (a factor times) the temperature T we

have Boyles Law,
P V = RT .
(4.56)
4.6 The three-body problem

The three-body problemthree masses interacting according to Newtons Law
of Gravityis not soluble in the sense that the two-body problem is. The
number of conserved quantities is the same in both problems, and for the
nine degrees of freedom in the three-body problem this is not enough. But
the three-body problem is very important, and in fact motivated many of the
developments that we will come to later on.
A natural first step is to look for special exact solutions, which may be used
as starting points for perturbation theory, or in other ways. An interesting example was found by Lagrange. Let us begin by assuming that the motion takes
place in a plane, and use complex numbers zi (t) to denote the trajectories. The
equations are
z1 z2
z1 z3
m3
3
|z1 z2 |
|z1 z3 |3
z2 z3
z2 z1
z2 = m3
m1
3
|z2 z3 |
|z2 z1 |3
z3 z1
z3 z2
z3 = m1
m2
.
3
|z3 z1 |
|z3 z2 |2
z1 = m2
(4.57)
The pioneering work on the Coma cluster is described, very readably, in F. Zwicky, On the Masses
of Nebulae, and Clusters of Nebulae, Astrophys. J. 86 (1937) 217.
47
We assume that the center of mass is at rest,

m1 z1 + m2 z2 + m3 z3 = 0 .
(4.58)
The particles form a triangle, with sides represented by

w1 = z3 z2 ,
w2 = z1 z3 ,
w3 = z2 z1 .
(4.59)
It turns out to be convenient to rewrite the equations of motion in terms of

these variables. A small calculation shows them to take the form
w1
+ m1 a
|w1 |3
w2
w
2 = m
+ m2 a
|w2 |3
w3
w
3 = m
+ m3 a ,
|w3 |3
w
1 = m
(4.60)
where m = m1 + m2 + m3 and
a=
w1
w2
w3
+
+
.
|w1 |3 |w2 |3 |w3 |3
(4.61)
This time we are looking for a special solution, not at the general case. So let
us assume that the triangle is an equilateral one,
w2 = e2i/3 w1 ,
w3 = e4i/3 w1 .
(4.62)
A glance at eqs. (4.60) shows that this property can be preserved in time.
Then we have that a = 0, and the only equation we need to solve is
w
1 = m
This we know how to do.
To interpret the solution, solve for
mz1 = m3 w2 m2 w3
w1
.
|w1 |3
(4.63)
m2 |z1 |2 = (m22 + m2 m3 + m23 )|w1 |2
(4.64)
and so on. A small calculation then shows that

3
z1 =
(m22 + m2 m3 + m23 ) 2 z1
,
m2
|z1 |3
(4.65)
and similarly for the other two particles. Hence the particles are all being
accelerated towards their common center of mass, with effective masses that
take an unexpected form. Each individual particle travels on an ellipse, but
the tree of them do so in unison, in such a way that they always span an
equilateral triangle.
48
Figure 4.1. The two stable Lagrange points in Jupiters orbit. The Greek and
Trojan asteroids lie within roughly one astronomical unit from the Lagrange
points.
A special case of this solution is of considerable physical interest. Let one

of the particles have negligibly small mass. Then the remaining pair trace out
the same orbits that they would follow in the absence of the third member.
Nevertheless the three particles span an equilateral triangle. This is the origin
of the two Lagrange points on the orbit of a planet, where small bodies may sit.
To draw this conclusion we should also investigate whether the exact solution
is stable under small perturbations. This turns out to be the case.
The Lagrange points we have found are called L4 and L5 , since there is
another set of three equilibria on the axis through the two bodiesalthough
they are of less interest since they are unstable. About a thousand asteroids
have been found close to the Lagrange points L4 and L5 on the orbit of Jupiter.
They are known as the Greek and Trojan asteroids (with names taken from
Homeros). It has been observed that the Earths Lagrange points are suitable
places where an alien civilisation could place a satellite surveying the Earth;
however, when the STEREO spacecrafts passed through (in 2009) they found
nothing of the sort.
But what can we say about the three-body problem in general? To celebrate
the sixtieth anniversary of King Oscar II of Sweden and Norway a large prize
was offered for a solution to the following problem: For a system of arbitrarily many mass points that attract each other according to Newtons laws,
assuming that no two points ever collide, give the coordinates of the individual
points for all time as the sum of a uniformly convergent series whose terms
are made up of known functions. The prize was awarded to Henri Poincare,
who did not solve the problem but whose contribution laid the foundations of
the modern theory of possibly chaotic dynamical systems. For the three-body
problem a solution was in fact found by Karl F. Sundman in 1912. He did
express a generic solution as a uniformly convergent power series in t1/3 . The
catch is that the series converges very slowly. It is estimated that, in order to
get useful information, one would have to sum the first 108000000 terms. Hence
the interest in the exact general solution dwindled from that point in.
49
With the advent of the computer it has become possible to follow a large
number of solutions to the three body problem on the screen, with no special
effort. The zoo of solutions include ones where the third body escapes from
the system, leaving the remaining pair more tightly bound than before.
Problem 4.1
The conservation of angular momentum is used in the gravitational two body problem to show that the trajectory is confined to a plane. Now
consider an electrically charged particle moving in the electromagnetic field of a magnetic monopole, that is to say a hypothetical particle which, if placed at the origin,
gives rise to the magnetic field
xi
.
r3
Here b is the magnetic charge of the monopole, and the electrically charged particle
obeys Lorentz Law. Compute the time derivative of the particles angular momentum.
Find a conserved quantity which is a modified version of the angular momentum, and
use the existence of this quantity to draw a qualitative conclusion about the trajectory
of the particle.
Bi = b
Problem 4.2
Consider the Earth-Moon system. Because of the tides some
dissipation of energy takes place. How does this affect the distance of the moon from
the earth?
Problem 4.3
Consider the Yukawa potential
er
,
k>0, >0.
r
What can you say about the existence and stability of circular orbits?
V (r) = k
Problem 4.4
In the theory of black holes one encounters the following
equation for particles orbiting the black hole,

2

2m
L
r 2 + 1
+
1
= E2 ,
r
r2
where r(t) is related to the distance to the event horizon at r = 2m, t is related
to time, m is the mass of the black hole, E is the energy of the particle, and L its
angular momentum. The equations make sense only if r > 2m. You can choose L2 and
E 2 freely. Compute the smallest possible value of r for which a (marginally) stable
circular orbit (with r = constant) exists.
Problem 4.5
If the Sun is flattened at its poles it will have a quadrupole
moment, and the potential is
q
k
+ 3 .
r
r
The orbits will no longer be closed. Compute the angle by which the perihelion moves
during one revolution, to first order in q.
V (r) =
50
Problem 4.6
For the Newtonian potential (4.27) the two body problem
admits an additional conserved vector
k
xi .
r
This is known as the Runge-Lenz vector. Check that it is indeed conserved. In what
direction does it point?
Mi = ijk x j Lk
Problem 4.7
The mass of Jupiter is 2 1027 kg and its distance from the Sun
is 8 1011 m. Suppose the Sun suddenly disappears. How long would it take for the
Trojan asteroids to crash onto Jupiter?
Problem 4.8
Let
w
= w|w|a1 ,
Z = w .
Choose a time parameter = (t) so that Keplers Second Law holds for Z( ), and
prove that
d2 Z
= cZ|Z|A1 ,
d 2
where c is a constant and
=
a+3
,
2
(a + 3)(A + 3) = 4 .
5 Small oscillations
An important class of equations that we can actually solve are the linear
ones. Their importance stems from the fact that departures from equilibrium
are described by equations that are linear to first ordernear a minimum,
most smooth potentials look like a collection of harmonic oscillators. When
left alone they are too simple to be interesting, but once we couple them to
external forces many delightful things happen. Or harmful things, depending
on your point of view. Either way the subject is of interest to engineers and
physicists alike.
It is perhaps worth remarking that once we allow the external forces to
depend on time the phase space picture is affected in a significant way
the phase space trajectories are now allowed to cross themselves, because the
external conditions may have changed by the time the trajectory returns to
the initial point.
5.1 Forced oscillations

We begin with a single degree of freedom obeying the equation
m
x + kx = 0
x
+ 2 x = 0 ,
k
.
m
(5.1)
The assumption that the string constant k is positive was slipped in when we
wrote 2 = k/m. The general solution is
where

x(t) = a1 cos t + a2 sin t = A cos (t + ) = Re eit ,
= Aei ,
A=
a21 + a22 ,
tan =
a2
.
a1
(5.2)
(5.3)
Observe that is a complex constant carrying information about both the

amplitude A and the phase . It is in fact highly convenient to write the
solution as the real part of a complex solution. This is a trick that works
because the equation is linear: two solutions can be added, and the result is
52
Small oscillations
still a solution. It does not matter if we take the real part before or after the
additionbut care must be exercised if we multiply two solutions together.
The total energy of the oscillator is
m 2 k 2 m 2 2
x + x = A .
(5.4)
2
2
2
It is manifestly independent of time. And the subject is exhausted.
To make matters more interesting we introduce an external force, and consider the forced oscillator
E=
1
F (t) .
(5.5)
m
To solve this equation, with some specified function F (t), we appeal to the
general theory of ordinary differential equations: it is enough to find one particular solution and then add the general solution of the homogeneous equation
(the one with F = 0).
It goes without saying that some forces are more important than others. An
interesting example is the periodically varying force
m
x + kx = F (t)
x
+ 2x =
F (t) = f cos (t + ) ,
(5.6)
which we will solve under the assumption that 6= . For a particular solution
we try the Ansatz
xpart (t) = B cos (t + ) .
(5.7)
Plugging this into the equation will determine B. Adding the general homogeneous solution we find the noteworthy general solution
x(t) = A cos (t + ) +
f
cos (t + ) .
2 )
m( 2
(5.8)
The remarkable thing is that the amplitude will get very large if the system is
driven by a periodic force whose frequency is close to the natural frequency of
the system. This phenomenon is called resonance. In fact, it may well be that
resonance drives the system out of the regime in which the harmonic oscillator
approximation is valid. (For the special case = , do exercise 2.)
The solution is a superposition of two harmonics with different frequencies.
Close to resonance the frequencies of the two harmonics almost coincide, and
we will observe the phenomenon of beats. Recall that
(1 2 )t
(1 + 2 )t
cos
.
(5.9)
2
2
If 1 2 this can be regarded as a vibration with frequency 1 2 ,
but with an amplitude modulated by a sine-wave of very low frequency. Our
case is a bit more complicated because of the differing amplitudes and phases
cos 1 t + cos 2 t = 2 cos
5.1 Forced oscillations
53
of the harmonics that we superpose. Let us assume that = +, and rewrite

the general solution as the real part of the complex amplitude

x(t) = eit + ei(+)t = + eit eit .
(5.10)
| + eit |2 = ||2 + ||2 + 2|||| cos (t + a phase) .
(5.11)
Provided is small compared to it makes sense to look at this as if the

system were oscillating at frequency , but modulated by a slowly time varying
amplitude whose square is given by
Thus the amplitude varies between || || and || + ||. If is small we see

beats.
To deal with a force of a quite arbitrary form we call on the complex numbers
to do a trick. Observe first that
d
(x + ix) i(x + ix) .
dt
Hence we define the complex variable
x
+ 2x =
(5.12)
z = x + ix ,
(5.13)
whichif you recall your quantum mechanicsis the annihilation operator

in slight classical disguise. The original equation becomes
1
F (t) .
(5.14)
m
Since the variable is now complex this is actually a pair of real first order
equations, so we are not violating any rules concerning how to count the
degrees of freedom. Once we have solved for z(t) we recover x(t) through
z iz =
1
Im [z(t)] .
(5.15)
Solving the first order system is straightforward. A particular solution to eq.

(5.14) can be written in integral form. Adding the general solution of the
homogeneous equation we obtain
x(t) =
it
z(t) = e
1
z0 +
m
it
F (t )e
dt
(5.16)
The particular solution is known as Duhamels integral.

Let us take a look at the energy budget. Assume that z() = 0, so that
the system starts out at rest. The energy we end up with is then
Z
2

m
1
2
it
E() = |z()| =
F
(t)e
dt
.
2
2m
(5.17)
54
Small oscillations
There is a net transfer of energy if the external force has a Fourier component
corresponding to the intrinsic frequency of the oscillator. If the force acts for
a short timeas compared with the natural time scale in the problem, which
is set by the exponential in the integrand can be ignored, and the energy
imparted
to the system is the kinetic energy associated to the momentum
R
F dt.
5.2 Damped and forced oscillations
A particle encountering air resistance, or more generally a particle moving in
a viscous medium, is really a system with very many degrees of freedom. It
often happens that the frequencies associated with the environment are very
much larger than the frequencies associated with the system of interest. In
many such cases we can introduce friction as an effective force. The coupling
is expected to grow with velocity, so we try the equation
m
x+ x+kx
=0
2
x
+2x+
0x = 0 ,
02 =
k
,
m
. (5.18)
2m
This is the damped oscillator.

As an aside, note that the damped oscillator is easily realized in electrical
circuit theory. A capacitor leads to a voltage difference VC = q/C, where
q is the charge and C the capacitance. Across an inductor there is a voltage
where I = q is the current and L the inductance. Coupling
difference VL = LI,
the capacitor, the inductor, and a resistor with resistance R in a series we
obtain the equation
L
q + Rq +
1
q=0.
C
(5.19)
Thus 0 = 1/ LC. There are numerous useful analogies between electrical,

acoustical, and mechanical systemsbut this is by the way.
To find the solution of the damped oscillator we try the Ansatz x = eit .
Inserting this, and cancelling the exponential, leads to the equation
2 + 2i + 02 = 0
Hence the general solution is
= i
02 2 .
2 2
2 2
x(t) = a1 et+i 0 t + a2 eti 0 t .
(5.20)
(5.21)
There are two qualitatively distinct cases to consider. If the damping is weak,
that is if 02 > 2 , the general solution is a damped oscillation

x(t) = et a1 eieff t + a2 eieff t .
(5.22)
5.2 Damped and forced oscillations
55
Note that the effective frequency eff is always smaller than the bare frequency 0 , as is reasonable. If the damping is strong enough so that 02 < 2
the solution decays exponentially without oscillations.
It is instructive to consider the critical case 02 = 2 . In the limit we obtain
the solution aet , but this cannot be the general solution since it contains
only one integration constant. In fact the general solution is
x(t) = (a + bt)et .
(5.23)
This must be the general solution because it contains two integration constants.
Finally we come to the forced and damped oscillator
m
x + x + kx = F (t) = f cos (t) ,
(5.24)
where we restrict ourselves to the important case of a periodic external force.

We rewrite thisin its complex formulationas
x
+ 2x + 02 x =
f it
e ,
m
(5.25)
make the Ansatz

xpart (t) = Beit ,
(5.26)
insert in the equation, cancel the exponential, and solve for B. The result is
(2 + 2i + 02 )B =
f
m
B=
1
f
.
2
2
m 0 + 2i
(5.27)
The amplitude is the absolute value of B. In fact

B = bei ,
b=
1
f
p
,
m (02 2 )2 + 42 2
tan =
2
.
02
(5.28)
The real solution is then
x(t) = b cos (t + ) .
(5.29)
Of course we could add the general homogeneous solution, but since this decays
to zero we ignore it. When the force has acted for some time the transient part
will be effectively zero, and a steady state described by the above solution sets
in.
Because of the damping the amplitude no longer goes to infinity at resonance, but it still has a pronounced maximum. The phase shows that the
system does not oscillate in phase with the external force. When is very
small so is , meaning that the oscillations follow the external force with no
phase shift. The sign of is negative, but its tangent switches sign at = 0 .
Small oscillations
56
Indeed always lies between 0 and , meaning that the system always lags
behind the force. For very large the phase shift is close to , because the
accelerations are large, and the acceleration of an oscillator is 180 out of phase
with its displacement.
We should look at the energy budget of the damped oscillator. We begin
with the damped but free oscillator (F = 0). Then the time derivative of the
energy function must be negative. Indeed
d
E =
dt
m 2 k 2
x + x
2
2
= x(m
x + kx) = x 2 = 2mx 2 .
(5.30)
Now consider the forced and damped oscillator in steady state. It is still losing
energy to the frictional forces at exactly this rate. This energy must then be
supplied as work by the external force, which is something we may want to
know about. Inserting the solution (5.29) we obtain the energy supplied by
the external force per unit time as
= 2mb2 2 sin2 (t + ) .
|E|
(5.31)
For most purposes it will be enough to know the time average of this quantity.
This time average is a function of the frequency ; recalling from (5.28) how
the amplitude b depends on we obtain
2
2
=f
I() h|E|i
.
m (02 2 )2 + 42 2
(5.32)
This has a maximum at = 0 . Let us suppose that = 0 + , and that

both and the damping are small compared to 0 . In this approximation
we can replace I() with the function
I() =
f 2 1
.
4m 2 + 2
(5.33)
This is the famous Lorentzian line shape function, giving the sharp resonant
response of a low-loss system. Among many other applications it expains the
intrinsic width of spectral absorption (and emission) linesalthough the dominating effect when astronomers observe these is the broadening due to the
Doppler shift, since the atoms are in thermal motion.
We see that the width of the Lorentzian line shape function grows with
the damping . On the other hand the total area under the curve is (almost)
independent of , as follows from the calculation
Z
f
I()d =
I()d
I()d =
4m
0
1 d
f
=
. (5.34)
2
4m
1 + 2
Which is mildly surprising.

We have only scratched the surface here. The feed-back of the environment
5.3 Several degrees of freedom
57
on an oscillating system can happen in many other ways. Equations of interest

include
x
+ 2 (t)x = 0
(5.35)
with some prescribed function 2 (t). This leads to the theory of parametric
resonancea child setting a swing in motion belongs to this class. Another
interesting class is given by
x
(t) + 2 x(t) + gx(t c) = 0 .
(5.36)
Provided that 0 < c < the force is in phase with the velocity, and any initial
oscillation tends to grow. Energy is being absorbed from the environment. In
the limit of small c this is a negatively damped oscillator. See ex. 5. The
Tacoma Narrows bridgea famous example in the theory of small, and not so
small, oscillationsbelongs here.1

Let us now consider N degrees of freedom. The variables are xi , and the
equations to be solved are
mij x
j + ij x j + kij xj = Fi (t) ,
1 i, j N .
(5.37)
When ij = 0 they follow from the Lagrangian

1
1
L = mij x i x j kij xi xj + xi Fi ,
(5.38)
2
2
and it follows that mij and kij are symmetric when considered as matrices. It
is less obvious that
ij = ji .
(5.39)
Nevertheless it is true, because of a subtle argument from statistical physics

that we take on trust here.
To start with we now try to solve the equations for a set of harmonic oscillators that interact among themselves, but not with the outside world. That
is
j + kij xj = 0 .
mij x
(5.40)
We know that the general solution must contain 2N integration constants, or

equivalently N complex integration constants. So we try the Ansatz
1
For a nice account, with many historical glimpses, see A. Jenkins, Self-oscillation, Phys. Rep. 525
(2013) 167.
58
Small oscillations
Figure 5.1. Two identical pendula connected by an elastic spring.
xi = i eit .
(5.41)
Inserting this into the equation, and then cancelling the exponential, gives the
matrix equation
( 2 mij + kij )j = 0 .
(5.42)
There is a non-zero solution for the vector i if and only if the determinant of
the matrix vanishes, that is if and only if

kij 2 mij = 0 .
(5.43)
This is an eigenvalue equation for 2 . It is also a polynomial equation of order

N , which means that it will have N not necessarily distinct roots. On physical
grounds all the roots must be real and positive. If not there would be exponentially growing or decaying solutions, and the system would be unstable.
(Note the logic: We try to model an actually existing system. Therefore the
matrices mij and kij must be special in various ways.)
The conclusion so far is that we have found N in general distinct frequencies,
called normal frequencies. The next step is to actually solve eq. (5.42) for the
eigenvector i . It will be determined only up to an overall complex constant.
Doing so for each normal frequency we end up with N particular solutions, each
coming with one undetermined complex number. Adding all these particular
solutions together gives the general solution to our problem.
A simple example of a system admitting two normal modes is given by two
identical pendula connected by an elastic spring. Assuming that the oscillations are small we use the Lagrangian

1 2 2
1 + 2 12 22 k(1 2 )2 .
(5.44)
2
To find the normal modes we must diagonalize T and V . This is achieved by
L=T V =
1
q1 = (1 + 2 ) ,
2
The Lagrangian becomes
1
q2 = (1 2 ) .
2
(5.45)
59

1 2
q1 + q22 q12 2 q22 ,
2 = 1 + 2k .
(5.46)
2
The normal modes have a simple interpretation. If q2 = 0 the two pendula
move in phase with the original frequency (equal to 1), if q1 = 0 the pendula
move in opposite phase with a frequency increased by the spring. As an afterthought we observe that, in this case, the two normal modes could have
been identified from the outset using only symmetry considerations.
The general solution for 1 , 2 is thus a sum of two harmonicsand if the
difference in frequency between the two is small we should be able to observe
the interesting phenomenon of beats, as in eq. (5.11). See exercise 6.
When we couple a set of harmonic oscillators to an external force it is the
normal modes that count. As the simplest example, consider
L=
x
1 + 02 x1 + k(x1 x2 ) = f cos t
x
2 + 02 x2 k(x1 x2 ) = 0
(5.47)
.
In terms of the normal modes q1 = x1 + x2 , q2 = x1 x2 this becomes

q1 + 02 q1 = f cos t
(5.48)
q2 + (02 + 2k)q2 = f cos t
Each normal mode gives rise to its own resonance frequency. The mathematical problem has been reduced to that of studying twoor in general N
independent forced oscillators.
Finally, let us consider eq. (5.40) for N degrees of freedom. We can decouple
the degrees of freedom in three steps: First we diagonalise the symmetric matrix mij , using rotations. Its eigenvalues must be positive (to prevent exercise
1.2 from becoming relevant). In the next step we rescale its eigenvectors so
that the diagonalized matrix becomes the identity matrix. These coordinate
changes are conveniently described on the Lagrangian level,
L=
1X
1X
(x i mij x j xi kij yj ) =
(y i i ij y j yi kij
yj ) =
2 i,j
2 i,j
=
1X
1 1
p zj
zi ij zj zi kij
2 i,j
i
j
(5.49)
.
For once we did not use the Einstein summation convention, since it would not
work here. Note that kij

is not the same matrix as kij , since the coordinate
system was rotated. In the final step we define the symmetrix matrix
1 1
p ,
kij
= kij
i
j
(5.50)
60
Small oscillations
we diagonalize itby means of a rotation that leaves ij invariantand arrive

at the Lagrangian
L=
1X
(u i u i i2 ui ui ) .
2 i
(5.51)
The normal modes ui are now decoupled, and the equations of motion trivial
to solve.
If you remember your linear algebra, you may be a little surprised by our
success. In general the matrices mij and kij do not commute, so how could we
bring both of them to diagonal form? The answer of course lies in the rescaling
of the eigenvectors. We did bring both matrices to diagonal form, but not by
means of rotations only.
Problem 5.1
Verify eqs. (5.3). By the way you are supposed to verify all
equations as simple to derive as those two.
Problem 5.2
Problem 5.3
integral.
Solve eq. (5.5) when F (t) = f cos (t + ).

Derive eq. (5.8) by inserting the relevant force into Duhamels
Problem 5.4
Consider a bouncing ball, obeying z = g but with a floor
at z = 0. At each bounce the absolute value of its velocity decreases by a factor ,
0 < < 1. Assume it left the floor with velocity v0 at time t = 0 and that bounces
are instantaneous. At what time does the motion stop? Sketch the motion in the zt
plane.
Problem 5.5
Consider the self-oscillator described by eq. (5.36). Assume the
time lag c is small. Expand the last term to first order in c and check that you get a
negatively damped oscillator.
Problem 5.6
Consider the two pendula connected by the spring, and let the
spring be very weak (k << 1). Choose initial data 1 = 2 = 0, 1 = v, 2 = 0. Show
that the amplitudes of the two pendula are modulated in such a way that after some
time T the first pendulum is stationary and all the energy has gone to the second.
Problem 5.7
Consider a particle in a plane, fastened to the corners of an
equilateral triangle by means of identical springs (using Hookes law). How will it
move?
Problem 5.8 You can think of a metal as N independent harmonic oscillators

(Einstein), or as a gas of sound waves with a certain number of allowed frequencies
(Debye). Show that the number of allowed frequencies equals the number of degrees
of freedom in the oscillators.
6 Rotation and rigid bodies
We have all played with spinning tops, and know that their dynamics is very
rich. The subject is of considerable practical importance, say to spacecraft
designers. To understand it we must understand the Lie group of rotations in
three dimensional space, and this is where we begin the story.
6.1 Rotations
In the plane, the mathematics of rotations is fairly trivial. Any rotation takes
place around a fixed point, and is uniquely characterized by its angle of rotation
. Its action on the coordinates is

x
y
cos sin
sin cos

x
y
(6.1)
What does this mean? Actually it can mean two quite different things: A
passive coordinate transformation changing the coordinates used to describe
a given point, or an active rotation moving a given point to another point
described by different values of the coordinates. These are two very different
operations, although the formul describing them are mostly identical. One
must stay awake during calculations.
This attended to, we observe that rotations form a group. A group is a set of
objects g1 , g2 , etc, together with a rule for combining them, so that g1 g2 = g3
is again a member of the set. This rule must be associative. One of the elements
is the identity element e, and has the property that e g = g e = g for any
other element g, and finally every element g possesses an inverse g1 such that
g g1 = g1 g = e. In general it may or may not be true that g1 g2 = g2 g1 .
For the two dimensional rotation group the elements can be written g(), so
the number of elements are continuously infinite. The set of all group elements
thus forms an abstract space, called the group manifold. In this case the group
manifold is a circle, coordinatized by . It is also a Lie group. This means that
the continuous parameters of a product are given by a analytic functions of
the parameters of its factors. Thus
g(1 ) g(2 ) = g(3 (1 , 2 )) .
(6.2)
62
Rotation and rigid bodies
where 3 is an analytic function of its arguments. In our example

3 = 1 + 2 ,
(6.3)
which is certainly analytic. The surprising thing about Lie groups is that they
can be almost completely understood through a careful study of their properties in a small neighbourhood of the identity element. All rotation groups are
Lie groups.
Rotations in three dimensional space are hard to understand, in the first
place because rotations do not commute in general. Rotations can be represented by matrices acting on vectors,
xi = Rij xj .
(6.4)
By definition this is a rotation if the lengths of all vectors are preserved,

xi xi = xi xi .
(6.5)
Hence the matrix Rij must be subject to some restrictions. On the other hand
it is immediate from the definition that rotations form a group. (Why?) To
see how the matrix is restricted, we observe that scalar products must be
preserved too, and then we perform a little calculation:
xi yi = Rki xi Rkj yj = xi Rki Rkj yj = xi yi
xi (Rki Rkj ij )yj = 0 . (6.6)
Here ij is the Kronecker delta. Since the vectors are arbitrary this is equivalent
to
Rki Rkj = ij .
(6.7)
RT R = RRT = 1 .
(6.8)
In matrix notation this is
The group properties can be now be checked on the level of matrices. The
columns, and the rows, of the matrix R must form an orthonormal triplet of
vectors. Such matrices are called orthogonal. Matrix multiplication is associative, and moreover
R1 R2 = R3
R3T R3 = (R1 R2 )T R1 R2 = R2T R1T R1 R2 = 1 .
(6.9)
This confirms that R3 is a group element. We observe that

1 = det RRT = det R det RT = (det R)2
det R = 1 .
(6.10)
The group can be restricted by insisting that det R = 1. This restriction can
be formulated in terms of the epsilon tensor:
6.1 Rotations
63
Figure 6.1. To coordinatize the set of all rotations (the group manifold of the
rotation group), choose a rotation axis and a number along it. Note that the
two endpoints of the axis represent the same rotation.
Rim Rjn Rkp mnp = ijk det R = ijk ,
(6.11)
where the definition (sic!) of the determinant was used. In N dimensions the restricted rotation group is called the special orthogonal group, denoted SO(N ).
Special refers to the restriction that the determinant equals one. If matrices with determinant 1 are admitted the group is called O(N ), and includes
reflections.
Eulers theorem states that any rotation in 3-space is a rotation around a
fixed axis. To prove it, note that an orthogonal matrix is unitary, and the
eigenvalues of a unitary matrix always take the form
= ei
(6.12)
for some angle . One can prove this by diagonalizing the unitary matrix. But
an orthogonal matrix is a real matrix too, which means that
det (R 1) = 0
(det (R 1)) = det (R 1) = 0 .
(6.13)
Hence complex eigenvalues, if they occur, must occur in pairs because their
complex conjugates are eigenvalues too. Since the number of eigenvalues of an
SO(3) matrix is odd, one of them must be real, and in fact it must equal one
because the determinant does. The corresponding eigenvector is the fixed axis
of rotation. Note that rotations in even dimensions (like two dimensions) work
differently. Note also that rotations can be hard to grasp: if the rotation axes
of R1 and R2 are known, what is the rotation axis of R1 R2 ?
The group manifold of SO(3), that is to say the set of all rotations, is easy
to visualize. The set of all rotation axes can be identified with the surface of
a sphere, or more precisely with the pairs of antipodal points where the axis
cuts the sphere. An arbitrary rotation is determined by its axis and an angle
, hence we can think of the set of all rotations as a solid ball with the identity
matrix at its center, with as a radial coordinate, and with antipodal points on
its surface identified because of the periodicity of . It sounds simple, but there
64
are some subtleties. The topology of this group manifold is such that there
are closed curves, starting and ending at the identity element, that cannot be
shrunk to zero in a continuous way. There is a famous trick one can play with
a pair of scissors sliding along a belt, to verify that this property has tangible
consequences.
We will need a coordinate system on the group manifold, and we choose
the Euler angles for this purpose. From our present perspective it is a little
difficult to make them appear natural. We introduce them by brute force, as
follows: Define
cos sin 0
R = sin cos 0
0
0
1
Compute
1
0
0
R = 0 cos sin
0 sin cos
cos sin 0
R = sin cos 0
0
0
1
R(, , ) = R R R =
cos cos cos sin sin

= sin cos cos sin cos
sin sin
(6.14)
(6.15)
cos sin + cos cos sin

sin sin + cos cos cos
sin cos
sin sin
cos sin .
cos
In the absence of the argument that makes this construction appear natural,
how do we know that this is correct in the sense that an arbitrary rotation
matrix can be expressed in terms of the Euler angles? Recall that every orthogonal matrix can be thought of as three orthonormal column vectors. By
inspection we see that and can be chosen so that the third column agrees
with any unit vector, and further inspection of the remaining columns shows
that they are restricted only to the extent needed for them to fill out a right
handed orthonormal triad. All this is true provided that
0 < 2 ,
0 < 2 ,
0 .
(6.16)
We accept this, and now we have a coordinate system on the group manifold
SO(3) available whenever we need one.
In writing eq. (6.15) we proved that an arbitrary rotation can be effected by
first rotating around the z-axis, then around the x-axis, and finally around the
z-axis again. There are 3 2 2 = 12 different ways of choosing the axes here,
leading to 12 different ways of parametrizing an arbitrary rotation matrix.
This has the consequence that whenever Euler angles are encountered in the
literature, one must check which of the 12 possible definitions that was used.
6.1 Rotations
65
A final point: I advertized that Lie groups can be understood through a

study of what they look like close to the origin. To see how for the orthogonal
groups, note that an orthogonal matrix can be written in the form
R = eA 1 + A +
1 2
A + ... .
2!
(6.17)
This is an orthogonal matrix if and only if the matrix A is anti-symmetric,

AT = A .
(6.18)
We can define a special class of curves in the group by

R(t) = etA
R(0) = 1 .
(6.19)
We would like to claim that every rotation group element can be written as
etA , for some choice of anti-symmetric matrix A. This is actually so for any
SO(3) rotation, but not for the additional reflections present in O(3). To see
how it works, we begin by using the definition (6.17) to conclude that

0 1
cos t sin t
exp t
=
.
1 0
sin t cos t
(6.20)
So the statement is true for any two dimensional rotation matrix with unit
determinant. But we can use Eulers theorem to reduce the three dimensional
case to the two dimensional onewe simply adapt our coordinates so that one
of the axes points along the eigenvector of the given but otherwise arbitrary
rotation matrix.
We can now see by means of a second order Taylor expansion what the noncommutativity means in terms of what goes on close to the identity element:
R1 (t1 )R2 (t2 )R11 (t1 )R21 (t2 )
(1 + t1 A1 +
t21 2
t2
t2
t2
A1 )(1 + t2 A2 + 2 A22 )(1 t1 A1 + 1 A21 )(1 t2 A2 + 2 A22 )
2
2
2
2
(6.21)
1 + t1 t2 (A1 A2 A2 A1 ) .
The group elements commute only if the commutator [A1 , A2 ] vanishes.

Anti-symmetric matrices can be added and multiplied with real numbers,
while staying anti-symmetric. Therefore they form a vector space, known as
the tangent space of the group manifold at the identity element. This vector
space is also known as the Lie algebra of the group, because the commutator
[A1 , A2 ] provides us with a peculiar way to obtain a third vector from any
given pair of vectors. For SO(3) any anti-symmetric matrix can be expressed
as a linear combination of the three matrices
66
0 0
A1 = 0 0
0 1
0
1
0
0 0
A2 = 0 0
1 0
1
0
0
0
A3 = 1
0
1 0
0 0 . (6.22)
0 0
They obey
[A1 , A2 ] = A3
[A2 , A3 ] = A1
[A3 , A1 ] = A2 .
(6.23)
They form a basis for the Lie algebraand would perhaps look more familiar
had we multiplied the Lie algebra elements Ai with an imaginary factor of
i, to make them Hermitian matrices. The idea here (a grand one!) is that it
works backwards toofrom a Lie algebra of commutators one can reconstruct
a Lie group.
In index notation an arbitrary element of the SO(3) Lie algebra can be
written as
Aij = ikj k ,
(6.24)
where i is a vector. If a physical system is subject to a time dependent

rotation R(t) such that
R ij (0) = ikj k ,
(6.25)
then i is referred to as the angular velocity vector at time t = 0. In space it is

pointing along Eulers fixed axis of rotation. If we return to the picture of the
group manifold as a solid ball with antipodal points on its surface identified,
we recognize the angular velocity vectors as a set of tangent vectors sitting at
the identity element, that is at the center of the ball.

Newtons Laws take a simple form only in inertial coordinate systems, that is
in coordinate systems adapted to the straight lines referred to in his First Law
(or to Absolute Space, as Newton himself expressed things). If the coordinate
system is not inertial more work is neededit would not have been easy to
figure out the laws if all experiments had been carried out on roundabouts.
Let Xi (t) denote the trajectory of a particle relative to an inertial system, and
let xi (t) denote the same trajectory described relative to a rotating coordinate
system. Assume that the coordinates agree at t = 0. Then there exists a
t-dependent rotation matrix R such that

Xi (t) = Rij (t)xj (t) = ij + ikj k t + o(t2 ) xj (t) .
(6.26)
At time t = 0 the inertial and rotating coordinate systems agree, and we
67
assume that the angular velocity vector i is constant. The time derivatives
are related by
X i = Rij x j + R ij xj .
(6.27)
At t = 0 (a point in time picked by convention) our assumption gives

X i = x i + ikj k xj .
(6.28)
The dot-notation can get confusing at this point, so it may be helpful to write
this as an operator relation
X i = Dtij xj =

d
ij + ikj k xj ,
dt
(6.29)
If we differentiate twice we obtain Newtons second law in the form

i = mDtij Dtjk xk = m
Fi = mX
xi + 2mijk j x k + mijk kmn j m xn . (6.30)
By the --identity, or equivalently by means of the formula for repeated cross
products, this is

m
xi = Fi 2mijk j x k + m 2 ij i j xj .
(6.31)
This is what Newtons second law looks like on the roundabout. The extra
terms on the right hand side are known as inertial or fictitious forcesbut
they feel real enough.
The third term on the right hand side of eq. (6.31) is known as the centrifugal force. It is responsible for the repulsive part of the effective potential
(4.8) in the two body problem. The second term is the Coriolis force. It is perpendicular both to the velocity x i and to the angular velocity i . To see that
such a term must be there, consider a free particle moving out from the center
on a frictionless rotating disk, and anchor the coordinate system to the disk.
Alternatively, consider the Foucault pendulum somewhere on earth. Choose a
coordinate system with a vertical z-axis, and let the pendulum perform small
oscillations in the x-y-plane. We ignore all terms of second order in the angular
velocity i of the earth. In particular we ignore the centrifugal force, but the
angular velocity enters the equations through the Coriolis term:
x
= k2 x + 2z y
y = k2 y 2z x .
(6.32)
Here z = || sin , and is the latitude of the pendulum (59 degrees 21

minutes, if it is in Stockholm). To solve the equations, introduce the complex
variable w = x + iy and make the Ansatz w = et :
w
+ k2 w + 2iz w = 0
2 + k2 + 2iz = 0 .
(6.33)
68
Figure 6.2. This picture illustrates Einsteins article on Beers Law, so presumably this is what the great mans tea cup looked like.
The solutions are

p
2z
2
2
= iz i k + z iz i k +
iz ik .
2k
(6.34)
Therefore
w = eiz t c1 eikt + c2 eikt
(6.35)
At the equator, where z = 0, we can arrange the initial conditions so that

the pendulum swings in a plane with y = 0 (say), but at non-zero latitudes
the plane will turn at a rate given by sin times the rate of rotation of the
earth. At the Poles it will make one revolution in 24 hours.
It is interesting to see how the fictitious forces shape the face of the earth.
They explain the erosion of river beds, as shown by Albert Einstein. He began
by looking at a simpler problem: Take a cup of tea, with some tea leaves in
it, and set the tea into rapid rotation by means of a spoon. You will see the
tea leaves gather into a little clump at the bottom, just at the center of the
cup. This happens because the flow velocity of the tea is zero at the edges
of the cup, and rises gradually from zero as one moves away from the edges.
Therefore the angular velocity will be greater at the tea surface, and lower
close to the bottom. Therefore the outwards directed centrifugal force will be
greater at the surface, and as a result a vertical velocity profile will be set
up in the cup. Therefore the tea leavesstaying close to the bottom due to
their specific weightwill be carried to the center of the cup, as was to be
explained.
A very similar story can be told about rivers, if there is a river bend. The
vertical velocity profile will carry fast moving surface water to the outer river
bank, and will cause erosion there. The latter will therefore suffer from more
severe erosion than the inner bank, and the net result is that the whole river
will start to move outwardsit is meandering. The Coriolis force is also of
importance to rivers, provided the rivers are large and do not bend too much.
In fact it causes Beers Law, an observation due to geographers according to
which large rivers on the Northern hemisphere undermine their right banks,
and large rivers on the Southern hemisphere undermine their left banks. Beers
6.3 The inertia tensor
69
law does not apply to small and bending rivers, where the centrifugal force
dominates.
Even more dramatically, fix a coordinate system in the earth so that one
of its axes point at the sun. It will make one revolution per year, and create
a centrifugal force just balancing the gravitational attraction felt by the sun.
This enables the sun to stand still. Now let the coordinate system follow the
earth in its daily rotation. As a result a huge centrifugal force will act on
the sun, completely overwhelming the gravitational force. In this coordinate
system the sun appears to be in rapid motion (x i 6= 0), and the Coriolis force
steps in to the rescue and prevents the sun from disappearing into outer space.

A rigid body is defined as a number of particles at fixed distances from each
other, or equivalently as a continuous mass distribution whose shape is fixed.
In the latter case the mass points m are replaced by mass elements dV . The
center of mass of the body can move, and its overall orientation can change
by means of a rotation. Any rotation. Hence the configuration space of a rigid
body is R3 SO(3). This is not only 6 dimensional, it has a non-trivial topology
since the topology of the group manifold SO(3) is non-trivial.
We introduce an internal coordinate system whose origin is fixed at some
point O within the body. These coordinates are called xi . We also introduce
an inertial coordinate system so that O has coordinates Xi . The velocity of
any given point xi within the body with respect to the inertial coordinate
system is denoted vi , and the velocity of the point O is denoted Vi . The body
rotates around O with angular velocity i , but the internal coordinate system
(with origin O) is not rotating. Make sure to note that the velocity vi is not
the time derivative of xi , and that we will not be bothered by fictitious forces.
The various quantities we introduced are related by
vi = Vi + x i = Vi + ikj k xj ,
(6.36)
To what extent does it matter where the origin O is placed? If we translate it

so that
xi xi = xi + ai ,
(6.37)
we see that the instantaneous velocities transform according to

vi = Vi + ikj k xj = Vi ikj k aj + ikj k xj = Vi + ikj k xj .
(6.38)
We observe that the angular velocity is unchanged, so we can talk of the

angular velocity of the body without caring about the point relative to which
we compute it.
The kinetic energy of a rigid body is
70
T =
Xm
2
(Vi + ikj k xj )2 =
(6.39)

m
=
V 2 + mikj Vi k xj + (ikj k xj )2 .
2
2
(The sum runs over all the particles in the body, even though no explicit
indices have been placed on them.) This can be rewritten in terms of the total
mass M as
X m
X
M 2 1X
mxj .
(6.40)
V +
m(x2 ij xi xj )i j + ikj Vi k
2
2
If O, the origin within the body, is placed at the centre of mass the last term
vanishes. It also vanishes if the origin coincides with a point that is fixed in
space (when Vi = 0). Now define the inertia tensor with respect to O,
T =
Iij =
m(x2 ij xi xj ) .
(6.41)
Again the sum runs over all the particles in the body. We observe that the
angular momentum with respect to O is
Li =
ijk xj mx k =
ijk kmn mxj m xn = Iij j .
(6.42)
Unlike the angular velocity, both the angular momentum and the inertia tensor
change if we shift the position of O. Let us assume that O is placed at the
center of mass. Then
1
1
1
1
T = M V 2 + Iij i j = M 1 Pi Pi + Iij1 Li Lj .
(6.43)
2
2
2
2
The inertia tensor describes the resistance of the body to changes of its rotation, just as the mass describes its resistance to changes of its translational
state. But the former is a tensor, not a scalar, and hence much harder to grasp.
Written out, the inertia tensor is
P
m(x22 + x23 )
P
I=
mx2 x1
P
mx3 x1
mx1 x2
mx3 x2
m(x21
x23 )
P
P
mx1 x3
mx2 x3
m(x21 + x22 )
(6.44)
This is a symmetric matrix, hence it can be diagonalized by rotations. This

means that we can rotate the internal coordinate system so that
I1 0 0
I = 0 I2 0 .
0 0 I3
(6.45)
71
The eigenvalues are known as moments of inertia. The eigenvectors are known
as principal axes of the body, and are orthogonal to each other.
The moments of inertia are all positive since, for an arbitrary vector ni ,
ni Iij nj =

m n2 x2 (ni xi )2 0 .
(6.46)
Indeed the individual terms are all positive. A vanishing moment of inertia can
occur only if all the particles lie on a line (because it would be necessary that
(nx)2 = n2 x2 for each individual particlerecall that the notation suppresses
the sum over all particles in the body, but in this case all the individual terms
are positive or zero.) Once we have adapted the coordinates so that the inertia
tensor is diagonal we see immediately that
I1 + I2 =
m(x21 + x22 + 2x23 )
m(x21 + x22 ) = I3 .
(6.47)
Hence no moment of inertia can exceed the sum of the other two. Equality
happens if and only if x3 = 0 for all the particles, that is for a planar body.
Some general facts about the inertia tensor can be stated:
If the body is symmetric under reflection in a plane, two of the principal
axes lie in it.
This must be so because the reflection must preserve the principal axes. The
only way to arrange this is to let two of them lie in the plane. The third axis
is then orthogonal to the plane and is taken into itself by the reflection. (Note
that this theorem is enough to identify the principal axes of an ellipsoid.)
The body can also be symmetric under rotations around some axis through
an angle which is some fixed fraction of 2:
If the body has a symmetry axis of order higher than two this axis must be
a principal axis, and the other two moments of inertia are equal (because
the corresponding eigenvectors cannot be unique).
A body with two equal moments of inertia is called a symmetrical top. If there
are at least three higher order symmetry axesthis is true for a cube, sayit
follows that all the eigenvalues are equal, so in fact every axis is a principal
axis.
The shape of the body is reflected in the inertia tensor. In the gravitational
N -body problem Newton proved that if all bodies are spherical, they can be
regarded as mass points. Now we see that if a body is rigid it can be regarded as
a homogeneous ellipsoid, since every tensor of intertia can be thus reproduced.
Nothing else matters as far as the response to arbitrary forces is concerned.
The inertia tensor depends on the point with respect to which it is computed:
Steiners theorem: For a body of total mass M , let Iij be its inertia tensor
relative to the centre of mass, and Iij its inertia tensor relative to a point
translated from the centre of mass by the vector ai . Then
72
Iij = Iij + M (a2 ij ai aj ) .
(6.48)
Roughly speaking it is easiest to rotate the body around its centre of mass.
The proof is easy, once we recall that
X
mai xi = ai
mxi = 0 ,
(6.49)
where xi is position relative to the centre of mass.

It is instructive to prove that any inertia tensor can be reproduced by placing
four particles of equal mass at appropriate distances from each other. To do
so, let the coordinates of the particles be
(xi , yi , zi ) ,
1i4.
(6.50)
This permits us to define three vectors x, y, z in R4 . We define a fourth auxiliary vector whose components are equal,
vT = (1, 1, 1, 1) .
(6.51)
That is to say that we are looking at the matrix
x1
x2
[x y z v] =
x3
x4
y1
y2
y3
y4
z1
z2
z3
z4
1
1
,
1
1
(6.52)
where the row index labels the four particles. Now we insist that the inertia
tensor be diagonal and that the centre of mass is at the origin. This translates
into the six conditions
xy =xz=yz=vx=vy =vz=0 .
(6.53)
In other words the four vectors must be mutually orthogonal, which is easily
arranged in R4 . It remains to check that we can reproduce the most general
diagonal inertia tensor; the solution is given in exercise 4.
6.4 Eulers equations

The following picture emerges: Associated with any rigid body there are three
kinds of axes in space, none of which coincide in general. Its principal axes
are associated with its shape andwith a little practicethey can be literally
seen. Then its angular velocity vector, also visible to the practiced eye. But
the axis that matters most to the motion of the body is its invisible angular
momentum vector, because this is the axis that is conserved in force free
motion. The visible axes will rotate around it as if by magic.
73
When its centre of mass is fixed, the equation of motion for a rigid body
subject to a torque i is
L i =
mijk xj Fk = i ,
(6.54)
where both the angular momentum and the torque are computed with respect
to the centre of mass. The equation is complicated by the fact that the inertia
tensor is time dependent,
L i = Iij j + Iij j .
(6.55)
This is awkward, whether there is a torque or no.

For the rest of this section we confine ourselves to torque free motion, with
the centre of mass kept fixed. (Even if we are in the gravitational field of the
earth it is possible to manufacture tops which realize torque free motion because they are suspended at their centres of mass.) Theoretically, the guiding
idea will be to formulate the equations of motion with respect to a coordinate
system fixed within the body, because then the inertia tensor is time independent. The direction of the angular momentum vector will then change with
time because the coordinate system rotates. We denote this vector by Ji . We
proceed in analogy with eq. (6.28), that is to say we assume that the inertial
and body fixed coordinate systems agree at t = 0, and get for torque free
motion
Ji + ikj k Jj = Iij j + ikj k Jj = 0 .
(6.56)
Being moderately intelligent we adapt the rotating coordinate system so that

its axes coincide with the principal axes, and arrive at Eulers equations
1 + (I3 I2 )2 3 = 0
I1
2 + (I1 I3 )3 1 = 0
I2
3 + (I2 I1 )1 2 = 0
I3
(6.57)
.
These equations can be solved exactly in terms of elliptic functions. In the

absence of torque the spinning top is an integrable system.
Much of the relevant information can be easily extracted. By inspection we
see that the body can execute constant rotation around any principal axis, i.e.
a solution is 1 = 0 = constant, 2 = 3 = 0. Is this solution stable? Set
1 = 0 + 1
2 = 2
3 = 3 ,
(6.58)
where 1 , 2 , 3 are time dependent and assumed to be small. Then expand

the equations to first order in the unknowns:
I1 1 = 0
I2 2 + (I1 I3 )0 3 = 0
I3 3 + (I2 I1 )0 2 = 0
I2
2 +
(I1 I3 )(I1 I2 ) 2
0 2
I3
=0
I3
3 +
(I2 I1 )(I3 I1 ) 2
0 3
I2
=0
(6.59)
.
74
Figure 6.3. The inertia ellipsoid, and its intersections (as curves) with spheres
of three different sizes.
The perturbation will grow, and the solution will be unstable, unless
I1 > I2 , I3
or
I1 < I2 , I3 .
(6.60)
We conclude that rotation around the largest and smallest of the principal
axes is stable, rotation around the remaining principal axis is unstable.
This can be seen more elegantly. The energy surface is the ellipsoid
J12 J22 J32
+
+
= 2E .
I1
I2
I3
(6.61)
The angular momentum vector itself is changing (in this coordinate system),
but its magnitude remains constant:
J12 + J22 + J32 = constant .
(6.62)
This is a sphere, intersecting the energy surface along one dimensional curves.
The system must move along them. It is seen that there are elliptic fixed points
at the major and minor axes, and hyperbolic fixed points at the middle axis.
Although it says nothing about the speed of the motion, this analysis does say
more than the perturbative calculation since it describes all solutions exactly.
Let us think a little bit more about this. The Euler equations define a dynamical system whose trajectories are confined to a three dimensional space,
for which we can use the coordinates J1 , J2 , J3 . A solution of the Euler equations is a curve in this three dimensional space, and curves in a three dimensional space can be very unruly indeed. But the Euler equations are special
because of the existence of two well behaved constants of the motion. Each of
them defines a set of easy-to-describe surfaces that fill phase space, and any
solution must lie where two such surfaces intersect. Hence the solutions are
also easy to describe. But this is in fact a quite special property of the Euler equations. The Lorenz equations (1.45) behave quite differently: for them
there are no well behaved constants of the motion, the solutions are unruly,
75
and the system is chaotic. We will come back to this kind of issues in chapter
9.
We still lack a complete description of the motion in space. The Poinsot
construction provides this. It describes how the angular velocity vector moves,
relative to the body, and relative to absolute space. The argument may be a bit
hard to follow, so please begin by looking at Fig. 6.3. From the point of view
of the top the angular momentum vector is moving along one of the curves
shown, but from the point of view of Absolute Space the angular momentum
vector is fixed while the topand its inertia ellipsoidis moving. We want to
turn this simple observation into a more detailed argument.
Using inertial coordinates we define the ellipsoid of inertia
Xi Iij Xj = 2E .
(6.63)
Think of it as a massless shell surrounding the body. If we compute the inertia

tensor with respect to Absolute Space we will find that it moves. The numerical
value 2E is at the moment just a convention, but since the kinetic energy E is
constant the convention guarantees that the angular velocity vector i moves
on this ellipsoid. There it traces out a curve known as the polhode (meaning
axis path). In absolute space it traces out a curve known as the herpolhode.
The angular velocity vector also obeys
Li i = 2E .
(6.64)
Since the angular momentum vector is fixed in absolute space we can assume
that it points along the X3 -axis, and then 3 is a constant, which means that
the herpolhode is a plane curve. The plane to which it is confined is known
as the invariable plane, and is orthogonal to Li . Now the normal vector of the
ellipsoid of inertia is
ni (X) = 2Iij Xj .
(6.65)
Evaluated at the point i this is

ni () = 2Iij j = 2Li .
(6.66)
This coincides with the normal of the invariable plane. Because of eq. (6.64)
the distance from the center of the ellipsoid of inertia to the invariable plane
is constant, and having placed the latter appropriately we conclude that the
ellipsoid of inertia rolls on the invariable plane. It rolls without slipping because
the point of tangency lies on the instantaneous rotation axis, so its velocity
equals zero. Hence the polhode rolls without slipping on the herpolhode. The
polhode is always a closed curve, but the herpolhode need not bewhen the
point of tangency has made one full revolution on the ellipsoid, the body will
have turned through some angle around the X3 -axis. Hence there are two
frequencies involved, and if they are not commensurable in the sense of eq.
76
(1.37) the herpolhode never closes. This is quite reminiscent of the Lissajous
figures.
Things simplify for a symmetrical top because then both the polhode and
the herpolhode are circles. Moreover the symmetry axis of the top coincides
with a principal axis of the ellipsoid of inertia, which means that the tip of the
symmetry axis also traces out a circle around Li . This motion is called regular
precession, and is not to be confused with the precession of a top placed in
a gravitational field, or with the precession of the equinoxes caused by the
spinning earth. Note that Eulers equations (6.57) for a symmetrical top are
easily integrated in terms of trigonometric functions.
The Earth is a symmetrical top, with I1 6= I2 = I3 and
1
I1 I3
.
I1
305
(6.67)
This is the ellipticity of the Earth. Moreover its rotation axis differs slightly
from its geometrical symmetry axis. Judging from Eulers equations, in particular eqs. (6.59) with 0 = 2/day, we expect the rotation axis to move at the
rate of one revolution per 305 days, and the polhode to be a circle surrounding
the geometrical North Pole. Although smallthe polhode is only about 15 meters acrosssuch a motion is indeed observed, and is known as the Chandler
wobble. However, the polhode is not a circle, and there is a period of close to
14 months as well as an annual periodicity. The latter is presumably caused
by metereological disturbances, while the longer period is due to the effect we
have discussed. The discrepancy with our prediction arises because the Earth
is not perfectly rigid, and can be explained if the Earth has an elasticity approximating that of steel. Poincare produced arguments showing that a fluid
core inside a rigid shell need not invalidate the argument.
Incidentally, from this example one can see how difficult it would be to
devise an experimental test of rigid body dynamics. What experiments can do
is to estimate the importance of friction, the departure from rigidity, and so
on, and to guide mathematical modelling of such effects.
6.5 The Lagrangian description

Evidently the Lagrangian of the rigid body must be L = T V , but going back
to our expression (6.43) for the kinetic energy of the body we are momentarily
confused. We recognize neither q nor q.
They are still there though. Recall
that the configuration space is the group SO(3), whose tangent vectors are
Lie algebra elements. According to the discussion is section 6.1 an arbitrary
rotation can be obtained by exponentiating an anti-symmetric matrix, of the
general form
Aij = ikj k
A = 1 A1 + 2 A2 + 3 A3 ,
(6.68)
77
where we use the basis introduced in eqs. (6.22) and i is the angular velocity
vector. Furthermore
R = etA
1 .
A = RR
(6.69)
A minus sign was used because we are now considering an active rotation of the
body, rather than a passive change to a rotating coordinate system. Once we
have A we can read off the angular velocity vector by comparing to eq. (6.68).
And at the expense of a minor amount of work we can express A in terms of
the Euler angles, introduced in section 6.1 as an example of an explicit set of
coordinates on the group manifold of SO(3). This is how we will express the
Lagrangian of the top in terms of coordinates. The calculation goes as follows:
1 =
A = RR
d
(R R R )(R R R )1 =
dt
= R R1 + R R R1 R1 + R R R R1 R1 R1 =
(6.70)
3 (cos
= A
A1 + sin A2 ) (sin
(sin A1 cos A2 ) + cos A3 ) .
In the final step we fell back on the explicit formul (6.14). Comparing to eq.
(6.68) we read off that
1 = cos + sin sin
2 = sin sin cos
3 = + cos
(6.71)
.
This we can insert in a Lagrangian.

We will choose a Lagrangian to describe a symmetrical top with one point
fixed, subject to a gravitational force. The restriction to a symmetrical top
implies that two of the eigenvalues of the inertia tensor are equal, say I1 = I2 ;
we make this restriction for the pragmatic reason that the general case is not
soluble. To connect the Euler angles to the orientation of the top we place
the origin at the fixed point, and let the identity matrix correspond to a top
whose symmetry axis is vertical. When the symmetry axis is inclined the angle
describes its rotation around the vertical, the angle its inclination, and the
angle the rotation of the body around its symmetry axis. Hence these angles
are referred to as the angle of precession, nutation, and spin, respectively (or
yaw, pitch, and roll, if you are a spacecraft designer). In these coordinates the
gravitational potential energy is
V = V () = M gl cos .
Now we can write the Lagrangian. We obtain
(6.72)
78
L=
I3
I1
21 + 22 + 23 V () =
2
2
(6.73)

I3
I1 2 2 2
=
+ sin + ( + cos )2 M gl cos .
2
2
For the symmetrical top neither nor appear in the Lagrangian, which
means thatenergy includedwe will get three constants of the motion,
enough to solve the equations of motion explicitly.
To be precise, we find the constants of motion
Lz =
L
= I1 sin2 + I3 cos ( + cos )

(6.74)
L
= I3 ( + cos ) .

(6.75)
L3 =
They are the angular momentum along the vertical axis (in absolute space) and
along the symmetry axis of the top, respectively. We can solve these equations
for the velocities:
L3
+ cos =
I3
(6.76)
Lz L3 cos
=
.
I1 sin2
(6.77)
Next we turn to the energy, which is

E=
I1 2
I3
( + 2 sin2 ) + ( + cos )2 + M gl cos =
2
2
(6.78)
2
L23
I1 2 (Lz L3 cos )
+ M gl cos +
.
+
2
2I3
2I1 sin2
This leads to a first order differential equation which can be explicitly solved
in terms of elliptic functions. Inserting the result successively in eqs. (6.77)
and (6.76) will lead us to the complete solution for the heavy symmetrical
top.
As usual one can go a long way with qualitative arguments. Eq. (6.78),
which governs the nutation of the top, can be described as one dimensional
motion in an effective potential. It can be written on the form
(b a cos )2
= 2 +
+ cos ,
sin2
where
(6.79)
79
Figure 6.4. Precession might possibly turn retrograde.
2EI3 L23
,
I1 I3
=
b=
Lz
,
I1
a=
L3
,
I1
2M gl
>0.
I1
(6.80)
(6.81)
If we introduce the variable u = cos the effective potential becomes a cubic

polynomial,
u 2 = u3 ( + a2 )u2 + (2ab )u + b2 .
(6.82)
The velocity can change sign only when the right hand side vanishes. The
polynomial on the right hand side goes from to , and
u = 1
u 2 = (b a)2 .
(6.83)
On physical grounds u 2 must be positive somewhere within the physically

relevant interval 1 u 1. Except for the special case b = a it then follows
that the cubic polynomial has two roots in the physically relevant interval. This
observation is enough to reveal the key qualitative features of the nutation of
the topclearly its axis is nutating between a maximal and a minimal value
of .
For the precession, described by eq. (6.77), an interesting qualitative question is whether the derivative changes its sign as evolves. The answer
depends on the relative size of the initial data Lz and L3 ; see Fig. 6.4 for the
possibilities.
To see why there is precession in the first place, take a top whose point
of contact with the ground is fixed but which is otherwise not subject to
forces. Let it rotate around its angular momentum vector. Then grab hold of
it, trying to increase the angle its axis makes with the vertical axis. If you
pull hard enough in a downwards direction you will succeed to some extent,
but there will be no z-component of the torque, and hence Lz must remain
constant. The top manages to keep Lz constant because it begins to precess,
thus adding an extra precessional contribution to Lz .
80
These considerations are applicable to the Earth, which is an oblate symmetrical top subject to tidal forces (primarily from the Moon) trying to decrease the angle between its axis and the normal of the ecliptic. The monotonic
precession of the Earth was known to the Greeks. It has a period of 26 000
yearswhich incidentally means that the position of the Sun relative to the
Zodiac has drifted noticeably since the Greeks determined it, with no apparent
consequences to astrology. See ex. 10. The nutation of the Earth was reported
by Bradley in the 18th century, who first observed it for a complete period,
18.7 years. A lesson, perhaps, for modern astronomers. It is a small effect, but
significantly greater than the Chandler wobble.
Problem 6.1 You are given an SO(3) matrix explicitly. You know it describes
a rotation by an angle through some axis, but you are not told what axis. What is
the quickest way to compute ?
Problem 6.2
Compute the inertia tensor with respect to the centre of mass
for a sphere, a cube, a circular cylinder and a circular cone, all of them having constant
density.
Problem 6.3
Place four equal masses at the corners of a regular tetrahedron
and compute the inertia tensor with respect to their centre of mass. To what extent
is the result obvious?
Problem 6.4
Complete the proof that the inertia tensor of any body can be
reproduced by placing four equal masses at appropriate distances from each other.
Problem 6.5
The mass of the Sun is 2 1030 kg, its equatorial radius is 7 108
m, and its sidereal rotation period is 25 days. Approximate the Sun as a homogeneous
sphere and compute its angular momentum. The mass of Jupiter is 21027 kg, its semimajor axis is 8 1011 m, and its orbital period is 4332 days. Approximate Jupiter as a
point mass in circular orbit and compute its orbital angular momentum. Comment?
Repeat the same exercise for the Earth-Moon system. The mass of the Earth is 61024
kg and its equatorial radius is 6 106 m, the mass of the Moon is 7 1022 kg and its
semi-major axis is 4 108 m.
Problem 6.6
Suspend a large key-ring in a twisted thread and let go. What
happens? Analyze the situation using Eulers equations.
Problem 6.7
Solve Eulers equations for a symmetrical top, I2 = I3 , and
draw the analogue of Fig. 6.3.
Problem 6.8
valid definition
Repeat the derivation leading to eqs. (6.71), but use the equally
A = R1 R .
What difference does it make?
Problem 6.9
For the symmetrical top in a gravitational field there must be
81
a solution for a sleeping top, that is a top spinning around the vertical axis. Show
that this solution is stable if and only if
L2z > 4M glI1 .
(6.84)
The top wakes up when friction has diminished its spin so that this bound is
violated.
Problem 6.10
Around 150 B.C. Hipparchos established the dates when the
Sun is in Capricorn. Given that the period of the precession is around 26 000 years,
use your understanding of the symmetrical spinning top to establish the direction in
which this assignment has been drifting since then, i.e. establish in which sign the
Sun actually is when it says in the astrology column that the Sun is in Capricorn.
7 Interlude: Legendre transformations
As is known, it is frequently a good idea to perform a transformation of some

functions that one may be working on. The Fourier transform is an example
of this idea. It turns out that the kind of manipulations that one may wish
to perform on the original function are easier to do on its Fourier transform,
and once this has been done one may revert to the original kind of function
by means of an inverse Fourier transform.
Another example of this is the Legendre transformation of a function f =
f (x). We will use it to pass from the Lagrangian to the Hamiltonian formulation of the equations of motion, and later on to obtain an interesting class
of canonical transformations of phase space. The Legendre transformation
is also used to great effect in thermodynamics. In this interlude we discuss it
from a purely mathematical point of view, assuming for simplicity that the
function depends on one variable only.
We will assume that the function to be transformed is twice differentiable,
and obeys
d2 f
>0.
dx2
(7.1)
This is not strictly necessary (see exercise 1) but convenient for our purposes.
A function obeying this condition is said to be convex. The Legendre transform
of the convex function f is a function g defined as

g(p) = maxx xp f (x) .
(7.2)
The maximum that we ask for is achieved if and only if

p=
df
.
dx
(7.3)
Precisely because we imposed condition (7.1) the derivative of f is a monotoneously increasing function of x, which means that eq. (7.3) can be solved to
give the unique solution x = x(p). Inserting this into the right hand side of
eq. (7.2) defines the function g(p) uniquely.
A function and its Legendre transform are related by
Interlude: Legendre transformations
83
Figure 7.1. The Legendre transformation of a function f (x). The connection

to eqs. (7.2) and (7.4) should be evident.
f (x) + g(p) = xp .
(7.4)
If eq. (7.3) holds this equation defines the Legendre transform g(p) once the
original function f (x) is known. But the equation is fully symmetric between
f and g, which means that if
x=
dg
dp
(7.5)
then eq. (7.4) defines f (x) as the Legendre transform of the function g(p). To
make sure that this claim is valid one must show that the function g is convex
because the function f is. But this is so because
d2 g
dx
=
=
2
dp
dp
dp
dx
1
d2 f
dx2
1
>0.
(7.6)
Hence eq. (7.5) can be solved for p = p(x). We may conclude that all the
information in the function f is present in its Legendre transform g, and also
thatunlike the Fourier transformation, saythe Legendre transformation is
its own inverse. Generalization to an arbitrary number of variables is quite
straightforward.
84
Interlude: Legendre transformations
There is an interesting geometric interpretation of the Legendre transformation, and I hope this can be deciphered from Fig. 7.1: to find out the value
of the function g at p0 , start by drawing the straight line p0 x, and then use the
definition of the Legendre transform to find g(p0 ) as an intercept of a tangent
of the curve with the ordinate axis. If the original function describes a curve
by giving the set of points on the curve, the Legendre transform describes it
by giving the set of lines that are tangent to it. This is a rather deep idea, but
one that belongs to another story.
Problem 7.1
Consider the function
1x
f (x) =
0
x2
, x<1
, 1x2
, 2<x
Compute the Legendre transform using the definition (7.2).
(7.7)
8 The Hamiltonian formulation
Any set of ordinary differential equations can be written in first order form,
provided we introduce enough extra variables. The Hamiltonian formulation is
a special way of doing this to the Euler-Lagrange equations, and reveals that
the latter have a very special form. The central features of mechanicsthose
that classical and quantum mechanics have in commonare brought out very
clearly by the Hamiltonian formulation.
8.1 Hamiltons equations and Hamiltonian flows

For a Lagrangian of the form L(q, q),
which is general enough for our purposes,
the Euler-Lagrange equations are
d L
L
=
.
dt qi
qi
(8.1)
These are second order ODEs. One obvious way to turn them into first order
equations is to define the canonical momenta
pi
L
.
qi
(8.2)
The expression on the right hand side will be equal to mqi in simple cases, and
in general it will be some function of q and q. We already know that the right
hand side is of some importance; it occurs when one sets boundary conditions
in the variational principle, and in connection with Noethers theorem. See
eq. (2.34). So the canonical momenta are the extra variables to be used in
turning the Euler-Lagrange equations into first order form. The use of q for
generalized coordinates and p for their momenta goes back to Jacobi, and
was solidified by Whittaker in a famous textbook.
Eqs. (8.1) now take the first order form
pi =
L
.
qi
(8.3)
But we need equations for q too. To this end we assume that eqs. (8.2) can be
inverted, that is
86
The Hamiltonian formulation
pi = pi (q, q)
qi = qi (q, p) .
(8.4)
Now we do have a first order system equivalent to the original Euler-Lagrange

equations. As a matter of fact it can happen that it is impossible to solve eqs.
(8.2) for the velocities; exactly how to deal with every possibility was explained
by Dirac when he developed his theory of constrained Hamiltonian systems.
Although this is interesting, from now on we assume for simplicity that the
equations are invertible.
Now form the function
H(q, p) = qi pi L(q, q)
.
(8.5)
Thus H and L are Legendre transformations of each other, in the sense of

chapter 7. It is a partial Legendre transformation, since only the q are involved
while the coordinates q are left untouched. We see that
qi =
H
pi
pi =
H
.
qi
(8.6)
These are Hamiltons equations, and the function H is known as the Hamiltonian. The variables pi are known as the canonical momenta. By means of the
Legendre transformation we have traded the coordinate q for the coordinate
p.
In practice, the Legendre transformation is usually easy to perform. We
know that
m 2
1 2
q
H=
p .
(8.7)
2
2m
An apparently more complicated case, involving many degrees of freedom, is
L=
1
(8.8)
L = Mij (q)qi qj V (q) .
2
Here the matrix M depends on the configuration space variables. See exercise
2.4, or for a concrete example see the Lagrangian for a spinning top in section
6.5. Using matrix notation the Hamiltonian is immediately found to be
1
H = pj Mij1 pj + V (q) .
(8.9)
2
We only have to check that the matrix M is invertible everywhere in configuration space. This in fact amounts to checking that the Lagrangian is a convex
function of the variables qi .
Our phase space is spanned by the 2n variables qi , pi . If we compare to
the general phase spaces discussed in section 1.4 we see that this is already
a restriction: the phase space of a Hamiltonian system is always even dimensional. But there is another and more dramatic difference. In section 1.4 time
8.1 Hamiltons equations and Hamiltonian flows
87
evolution was described by the equations zi = fi (z), so that the general case
is obtained by choosing 2n independent functions fi . In the Hamiltonian case
the time evolution is determined by a single function H(q, p). This is indeed
a very strong restriction, but it is one that Nature seems to respect for all her
fundamental equations.
There are consequences. One of them is that
H
H
H
H H
H H
H
H
=
=
= 0 . (8.10)
+ pi
+
+
H = qi
qi
pi
t
pi qi
qi pi
t
t
Hence the Hamiltonian is always a conserved quantity, unless it depends explicitly on time. The set of points in phase space obeying
H(q, p) = E
(8.11)
is called the energy surface. Time evolution takes place within the energy
surface. In two dimensions (for one degree of freedom) the energy surface
coincides with the one dimensional trajectories; hence Hamiltonian systems
with one degree of freedom are always soluble.
Next let us imagine time evolution as the flow of a fluid. The flow lines
are defined by the little arrows in phase space. The fluid is incompressible
if, for any fixed finite volume in phase space, the amount of fluid going out
through its surface equals the amount that is going in. By Gauss theorem the
difference between them is equal to the integral of the divergence of the flow
over the volume. Since this must vanish for every volume we conclude that the
flow is that of an incompressible fluid if and only if its divergence vanishes.
For Hamiltonian time evolution this is indeed so:
divz = z = qi qi (q, p) + pi pi (q, p) = qi pi H + pi (qi H) = 0 . (8.12)
This conclusion is worth stating as a theorem.
Liouvilles theorem: In Hamiltonian mechanics the phase space flow preserves
volume.
Liouvilles theorem is the origin of the claimmade in section 1.4that
sources and sinks do not occur in Hamiltonian systems. In two dimensions
the only fixed points that occur are either elliptic or hyperbolic.
On reflection one sees that the behaviour allowed by Liouvilles theorem
can still be very complex. Already in two dimensions the shape of a small
piece of the phase space fluid passing close to a hyperbolic fixed point will
be squeezed in one direction and stretched in another. In the end the original
volume element can acquire a very involved shape, and for an observer with
limited resolution it may in fact seem as if it has been smeared all over the
energy surface, even though on microscopic scales it does preserve its volume.
88
Figure 8.1. Liouvilles theorem in action: a volume element in phase space

here shown at three successive momentspreserves its volume, but may become
very distortedhere because it approaches a hyperbolic fixed point.
8.2 Kets and bras and all that

There is an intriguing geometrical structure behind the Hamiltonian equations
of motion. To appreciate it, we begin by recalling the kets and bras notation
in quantum mechanics. In quantum mechanics the state space is a vector space.
Vectors are called kets and denoted by |i. They form a vector space because
one can take linear combinations of kets: if |1 i and |2 i are kets and a1 , a2
are numbers then |3 i is a ket too, where
|3 i = a1 |1 i + a2 |2 i .
(8.13)
In quantum mechanics the numbers a1 , a2 are complex, but from now on we

assume that all such numbers are real. It is important to realize that, in
quantum mechanics, one never takes the scalar product of two vectorsthat
is to say of two kets. What one does is to introduce another vector space
whose vectors are called bras, and denoted by h|. By definition a bra is a
linear function from the vector space of kets to the real numbers. Thus, given
a bra h|,
h|i = a
(8.14)
is defined for all kets |i. Dirac calls this a bracket, which explains the names
kets and bras. (It is believed that Dirac did not know that bra already
had a meaning in English.) An added complication in quantum mechanics is
that the vector spaces are often infinite dimensional function spaces, but we
do not bother with this here. We do need to check that the bras form a linear
space. But this is so because the map to the real numbers is linear,
(a1 h1 | + a2 h2 |)|i = a1 h1 |i + a2 h2 |i .
Hence h3 | = a1 h1 | + a2 h2 | is a bra.
(8.15)
89
At this point we have two vector spaces, and a priori they are different. The
equation h| = |i makes no sense at all. Apples and pears are never equal.
However, there is a final twist to the story, because there does exist a one-toone correspondence between the kets and the bras. If you have a ket written out
as a column vector with respect to some basis, there is a unique bra obtained
by transposing the column to a row (and taking the complex conjugates of
all the numbers, but this is irrelevant here since all our components are real).
This is an extra piece of structure, giving rise to a one-to-one correspondence
h| |i. It means that we can associate a unique number to any ket, namely
the real number h|i. This number is the length squared of the ket |i.
How can we make the distinction between bras and kets in the index notation? The answer is simple. Let us denote all the original vectors, the kets,
by x . We are using Greek rather than Latin indices, and will continue to do
so whenever it is our intention to use the indices in a correct tensorial way
(that is, in the way I am just going to explain). When the indices are only
used to label a set of objects we continue to use Latin indices. The important thing is that we place the index on a ket upstairs. Such a vector is
called contravariant. A bra will be denoted by u , that is to say with its index
downstairs, and is called a covariant vector. This should represent a linear
map from the space of all kets to the real numbers, and we define this map as
u(x) = u x = a ,
(8.16)
where Einsteins summation convention is understood. It is important to note

that in the upstairs-downstairs notation, the expressions x y and u v are
simply not allowed, so that we cannot define the length squared of a vector
x as x x . It is forbidden. There is no notion of length squared until we have
introduced further structure. With these new rules, many of the formul in,
say, section 6.1, are simply forbidden. We still want to use them however, so
we have to modify them in a suitable way.
Let us agree that x simply stands for the components of a ket relative to
some basis in the ket vector space. That is to say that
|i = |e ix ,
(8.17)
where the vectors |e i form a basis (and must not be confused with the components u of a bra vector!). A matrix operating on the ket vector space will
be written as a mixed tensor A . Then the equations
y = A x
and
v = u B
(8.18)
do make sensegiven that we always use the Einstein summation convention

for repeated indices in upstairs-downstairs position. As a matter of fact these
equations make two different kinds of sense. The vector
| i = |e iy = |e iA x = A|i
(8.19)
90
is a new vector, obtained by transforming the old vector |i. This is an active
transformation of the vector. On the other hand the vector
|i = |e ix = |e i(A1 ) A x = |e iy
(8.20)
is just the original vector |i in a new coordinate system. This is a passive

transformation of the vector.
In a passive transformation the components and the basis vectors transform
in opposite ways. This is an important idea. We will insist that, whatever linear
transformation we do of the bra and ket vector spaces, the numbers u x
should remain unchanged. We can arrange this by insisting that whenever we
transform the kets, we also transform the bras in the opposite way. That is to
say
x A x
u u (A1 ) .
(8.21)
This is part of the origin of the names contravariant and covariant vectors
they transform in opposite directions, to ensure that u x remains unchanged.
We can go on to define tensors with more than one index in the same way.
A tensor with k indices running from 1 to n is defined as a collection of kn
components transforming in specific way under changes of basis. Examples
include
T A A A T ,
S S (A1 ) (A1 ) (A1 ) . (8.22)
Students are usually disturbed by the fact that the components of a tensor
with more than two indices cannot be displayed as a matrix, but really the
definition does not require this.
Still something is missing. The whole point about chapter 6 was to discuss
the special matrices that preserve the length of the vectors. To do so here we
need to set up a one-to-one correspondence between the set of x and the set
of u , and then define the analogue of h|i.
The Kronecker delta is the key. We write it as a covariant tensorwith both
indices downstairsas . In fact we can be a bit more general. We introduce
a metric tensor. By definition this is any covariant symmetric tensor with two
indices downstairs,
g = g ,
(8.23)
which has a contravariant inverse g , in the sense that

g g = .
(8.24)
Here the left hand side defines an operation which acts as the identity on both
our vector spaces,
x = x
and
u = u .
(8.25)
91
Figure 8.2. Bras and kets: kets are arrows, bras collections of parallel planes
(or measuring tapes). Their lengths are undefined but one can multiply them
with real numbers, and the quantity u x has a clear meaning.
Thus is a special example of a metric tensor, while is an operator that

acts on ketseven though, when you write them out as matrices, they look
the same.
Now we have the means to set up a one-to-one correspondence between bras
and kets:
x x = g x
u u = g u .
(8.26)
We define the length squared of a vector x as x g x . If g = we recover

the naive definition, used implicitly in section 6.1. Finally, let us consider eq.
(6.7), which defined those special matrices that correspond to rotations, that
is to say those matrices that transform the vectors in such a way that their
lengths are preserved. It is not hard to rewrite this equation in such a way
that they are allowed by our new conventions. The result is
R R g = g .
(8.27)
If the metric is the Kronecker delta, this equation says that the rotation is done
by means of an orthogonal matrix. In this way we recover the full content of
section 6.1.
There is a simple geometrical picture of kets and bras that may be helpful.
Represent the kets as arrows pointing from the origin. If the ket x is multiplied
with (say) 2, the length of the arrow in the picture doubles. Represent the bras
with measuring tapes through the origin, or more accurately as a set of parallel
hyperplanes with constant spacing (level curves of a linearly rising function).
The bras transform oppositely to the kets, so if the ket is multiplied by 2 the
bra must be multiplied with 1/2. The number u x should stay unchanged
92
under this transformation. If we interpret this number as the number of level

surfaces pierced by the arrow, we see that multiplying the ket with 2 means
that the spacing between the level surfaces doublesso the function giving
rise to the level surfaces has been multiplied by 1/2. We have arrived at a
consistent geometrical picture.
The true length of x is not defined yet. You might think that it could be
defined by choosing some special measuring tape, rotating this tape until it
is aligned with the arrow, and afterwards compute u x . But this assumes
that the measuring tape is unchanged by the rotation, and the catch is that
rotation is defined as a transformation that preserves the lengths of all
arrows, or the spacing of all measuring tapes. We are caught in a circular
definition, and forced to introduce some further structurea metric!before
we can define length.
The reason why we did not bother about these things in chapter 6 is that
there was simply no need for them. Because the metric tensor was so simple
equal to the Kronecker deltathe one-to-one correspondence between bras and
kets was so obvious that we simply identified them from the outset, without
comment. But the more careful treatment here is useful for various generalisations. In particular it is useful to understand the Hamiltonian formulation
of the equations of motion. There may even be a moral here. Forbidding certain things (in this case summing over repeated indices in any way we please)
sometimes gives much more freedom.
8.3 The symplectic form

To begin with we assume that phase space is a vector space. More involved
casessuch as the phase space of a rigid bodycan wait. If we have a single vector x , how can we assign a number to it? We know the answer. We
introduce a metric tensor, and define
||x||2 = x g x .
(8.28)
The simple choice g = is especially interesting, but in principle any

choice of g is allowed, as long as it is invertible and symmetric. Once we
have a metric we can also assign a real number to any pair of vectors, namely
the scalar product
y g x = ||x|| ||y|| cos .
(8.29)
But geometrically it is obvious that there is another number that we can assign
to the pair, namely the (oriented) area A that they span. Let us adapt our
coordinates so that the two vectors have only two non-zero components each.
Then A is given by a determinant
8.3 The symplectic form
93
Figure 8.3. Two ways of associating a number to a pair of vectors: the angle
they subtend, and the area they span.
We can write this as
1 1
x y
A = 2 2 .
x y
(8.30)
A = y x ,
(8.31)
where
=
0 1
1 0
(8.32)
Like the scalar product the area is a bilinear function of the vectors.
The next step is to liberate eq. (8.31) from its origins. Any anti-symmetric
and invertible tensor can serve as a symplectic form on a vector space, just
as any symmetric invertible tensor can serve as a metric. In other words eq.
(8.31) will be taken to be a meaningful number associated to any pair of
vectors, regardless of the dimension of the vector space, provided only that
the symplectic form obeys
= ,
(8.33)
and provided that there exists an inverse matrix so that

= .
(8.34)
The latter equation is analogous to equation (8.24) for the metric. But it
happens that an anti-symmetric N N matrix has determinant zero if N
is odd. If the determinant is zero its inverse cannot exist. Hence symplectic
forms exist only on even dimensional vector spaces, while metrics exist on
vector spaces of all dimensions. (This fact gave us Eulers theorem in section
6.1the antisymmetric matrix defining an infinitesimal rotation in 3-space
94
necessarily has a zero eigenvalue, and hence the rotation it generates has a
fixed axis.)
Just as we may wish to define metrics on curved spaces, not only on vector
spaces, so we may wish to define symplectic forms on more general phase
spaces. A symplectic form on an arbitrary phase space will be represented by an
invertible and anti-symmetric tensor , but in principle its components may
depend on the particular point of phase space where it sits. This cannot happen
in a quite arbitrary fashion though, in fact the third and final requirement on
a symplectic form is
+ + = 0 .
(8.35)
We will mostly be interested in the case when the components of are constant, and then this extra requirement is trivial. For now we only remark that
the equation is recognisable as one of Maxwells equations in electrodynamics.
There it guarantees the existence of the vector potential. We will see that a
similar object arises in Hamiltonian mechanics, once we come to section 8.8.
The most common situation is that the symplectic form is constant and
block diagonal, that is to say it has the form
0 1 0 0 . . .
1 0 0 0 ...
0 0 0 1 . . .
0 0 1 0 ...
..
..
.
.
0 0 0 0 ...
0 0 0 0 ...
0
0
0
0
..
.
0
0
0
0
..
.
0 1
1 0
(8.36)
In exercise 4 you will show that, on a vector space, one can always choose
the coordinates in such a way that any symplectic form takes this form. In
fact there is a stronger statement known as Darboux theorem: On any phase
space with topology R2n , coordinates can always be introduced so that the
symplectic form takes this standard form. Although we do not need to go into
it now, it is interesting to know that metric tensors behave in a completely
different way in this respect.
Because the symplectic form has an inverse, it can be used to relate contravariant vectors with covariant ones in a unique manner, just as a metric
can. The archetypical contravariant vector is the little arrow z that gives
the time evolution of a system, while the archetypical covariant vector is the
gradient of a function, say H. Let us relate them:
z = H .
(8.37)
Let us further suppose that the symplectic form takes the standard form (8.36),
and let us adapt the description of the coordinates according to
8.4 The algebraic structure of mechanics
z1
z2
z3
z4
..
.
z =
2n1
z
z 2n
q1
p1
q2
p2
..
.

=

qn
pn
95
(8.38)
Remembering that we agreed to be careless about the position of Latin indices

we then find that
z = H
qi =
H
pi
(8.39)
pi = H
.
qi
This is indeed Hamiltons equations of motionhereby revealed as a topic in

symplectic geometry.
8.4 The algebraic structure of mechanics

In Hamiltonian mechanics a single function H(z) = H(q, p) on phase space
is used to specify time evolution. This is reminiscent of quantum mechanics.
Let us recall the basic facts about the latter. Time evolution is specified by
In the Heisenberg picture the time
choosing a definite Hermitian operator H.
evolution of an arbitrary operator is determined by the equation
dA
H]
.
= i~[A,
dt
(8.40)
Here the bracket represents a commutator. It obeys the algebraic relations

= a1 [A1 , B]
+ a2 [A2 , B]
,
[a1 A1 + a2 A2 , B]
(8.41)
B]
= [B,
A]
,
[A,
(8.42)
[B,
C]]
+ [C,
[A,
B]]
+ [B,
[C,
A]]
=0,
[A,
(8.43)
where a1 , a2 are arbitrary numbers. These relations also characterize the Lie
bracket occurring in the study of Lie groups (see section 6.1); the last relation
is known as the Jacobi identity.
In classical mechanics all functions commute, so at first sight there does not
seem to be a classical analogue of the commutator. But there is one! Using the
96
symplectic form we define the Poisson bracket of two arbitrary phase space
functions as
{A, B} A B .
(8.44)
By inspection we see that the Poisson bracket obeys

{a1 A1 + a2 A2 , B} = a1 {A1 , B} + a2 {A2 , B}
(8.45)
{A, B} = {B, A} .
(8.46)
It is less obvious, but true (see problem 5) that it also obeys the Jacobi identity
{A, {B, C}} + {C, {A, B}} + {B, {C, A}} = 0 .
(8.47)
Granted this, the analogy to the quantum mechanical commutator is clear:

phase space comes equipped with a Lie bracket. Eq. (8.39) can be written in
the form
z = H = {z , H} .
(8.48)
The time derivative of an arbitrary phase space function is

A = {A, H} ,
(8.49)
The classical and the quantum mechanical Hamiltonians are analogous too.
These observations led Dirac, in the course of a Sunday walk in Cambridge,
to the belief that any classical system can be quantized by setting up a
correspondence between functions on phase space and operators on a (complex) linear space, and by replacing the Poisson brackets with commutators
according to the rule
{A, B}
1
[A, B] .
i~
(8.50)
On the whole, although various complications do arise, Diracs idea has proved
to be correct.
If the symplectic form takes the standard form (8.36) we can use eq. (8.38)
to split the 2n coordinates z into n pairs (qi , pi ). Then the Poisson bracket
reads
{A, B} =
A B
A B
.
qi pi pi qi
(8.51)
(Yes, as always a sum over repeated indices is understood.) Moreover eq. (8.48)
becomes
8.5 The sphere as a phase space
q = {q, H}
p = {p, H} .
97
(8.52)
This form of Hamiltons equations is frequently seen.

For the explicit evaluation of Poisson brackets we look carefully at the definition, and observe that
{A, F } = A F = (q Ap p Aq )F .
(8.53)
Hence we can think of {A, } as a differential operator acting on phase space

functions. It obeys Leibniz rule
{A, BC} = B{A, C} + {A, B}C .
(8.54)
In this sensewhich is a very important oneit acts like a kind of derivative. Using this the evaluation of any Poisson bracket can ultimately be made
starting from the fundamental Poisson brackets
{qi , pj } = ij ,
{qi , qj } = {pi , pj } = 0 .
(8.55)
We can now confirm that the Poisson bracket acts like a derivative in the
sense that
{qi , A} = pi A
{pi , A} = qi A ,
(8.56)
where A = A(q, p) is any function on phase space. Note that the quantum
mechanical commutator behaves in a similar way, except that in quantum
mechanics we have to care about the order in which we place things. The
ordering on the right hand side of eq. (8.54) would be the correct one to use
there, while ordering does not matter in classical mechanics.
8.5 The sphere as a phase space

A set of equations that at first sight cannot be cast into Hamiltonian form
is Eulers equations for a spinning top, eqs. (6.57), for the simple reason that
phase space is three dimensional. This is a bit deceptive though. Recall that
J 2 = J12 + J22 + J32 is constant. Let us therefore define angular coordinates ,
through
J1 = J cos sin ,
J2 = J sin sin ,
J3 = J cos .
(8.57)
If we write out Eulers equations in terms of Ji = Ii i we find, after a calculation, that they become
98
= J
1
1
I1 I2
= J
cos sin sin

(8.58)
1
1
I1 I3
cos2 cos J
1
1
I3 I2
sin2 cos .
Now stare at this. Then define a function on the sphere,

J2
J2
J12
+ 2 + 3 .
2I1 2I2 2I3
H(, ) =
(8.59)
We also define the symplectic form

= J
0
sin
sin
0
1
=
J sin
0 1
1 0
(8.60)
(The indexing here is such that = J sin .) A calculation now verifies that
Eulers equations for fixed angular momentum squared take the Hamiltonian
form
= H ,
= H .
(8.61)
Moreover it is easily checked that we have the Poisson bracket algebra

{J1 , J2 } = J3 ,
{J2 , J3 } = J1 ,
{J3 , J1 } = J2 .
(8.62)
This follows directly from the definition (8.44).

This is an example where the phase space is a sphere, and where no natural split of the coordinates ( and ) into a canonical pair (q, p) exists. But
according to Darboux theorem canonical coordinates must exist on any open
region of the sphere. Indeed it is easy to check that
q = cos ,
p=,
(8.63)
form a canonical pair, and they are good coordinates on the sphere everywhere except at the poles. In these coordinates the area element on the sphere
becomes
dA = J sin dd = Jdqdp .
(8.64)
Hence the canonical coordinates are suitable if you want a map of the sphere to
display areas correctly. (By all means draw a picture to illustrate this result!)
Closer examination of this example reveals that generalization to any evendimensional sphere is non-trivial. The point is that we cannot use any antisymmetric non-degenerate tensor as a symplectic form. According to the definition it must also solve eq. (8.35). In the two-dimensional case this is a trivial
8.6 Infinitesimal canonical transformations
99
point, but as it turns out there is simply no solution for any higher dimensional sphere. Other, non-spherical, examples of phase spaces with non-trivial
topologies do exist in higher dimensions.
8.6 Infinitesimal canonical transformations

We have yet to discuss the classical analogue of unitary transformations. Unitary transformations preserve commutators, so we define canonical transformations as transformations that preserve Poisson brackets. This is the same
thing as saying that canonical transformations leave the symplectic form invariant, and is analogous to the definition of rotations as those transformations
that leave a specific metric invariant, eq. (8.27). As always, we can regard the
transformations as active, giving rise to a movement in phase space (or Hilbert
space), or as passive transformations from one coordinate description to another. Both viewpoints are useful, but for the general discussion the active
viewpoint is the better.
Anyway, if we are given a transformation on phase space
q q = q (q, p) ,
p p = p (q, p) ,
(8.65)
then phase space functions are transformed according to

A (q , p ) = A(q, p) .
(8.66)
(The new function takes the same value at the new point as the old function
does at the old point.) The transformation is canonical if
{A, B} = C
{A , B } = C .
(8.67)
How do we find such transformations?

We begin by looking for infinitesimal canonical transformations. Let us express the function A as a function of q and p, and consider
A (q, p) = A(q, p) + A(q, p) .
(8.68)
The function A must be chosen so that eq. (8.67) holds, and is assumed to be
small so that terms of second order in A can be ignored when this property is
checked. To obtain a candidate A we choose a function F = F (q, p) on phase
space, and set
A = {A, F } .
(8.69)
Hence, to first order in ,

A = A + {A, F } ,
B = B + {B, F } ,
(8.70)
100
and so on. Computing to first order in and using the Jacobi identity we find
that
{A , B } = {A, B} + ( {{A, F }, B} + {A, {B, F }} ) =
= C {F, {A, B}} = C + {C, F } = C + C = C
(8.71)
(to that order). This works for any phase space function F . In principle it is
possible to integrate eq. (8.69) to obtain a finite canonical transformation
the equation has the same form as Hamiltons equations of motion, and the
only catch is be that it may be difficult to do the integration explicitly.
Again, note the similarity to quantum mechanics: in quantum mechanics
any Hermitian operator generates a unitary transformation, in classical mechanics any phase space function generates a canonical transformation. At the
same time there is an interesting difference between canonical and unitary
transformations: there are many more phase space functions than Hermitian
operators, hence there are many more canonical transformations. This has
to do with the fact that the unitary transformations also preserve a scalar
product, and in this sense they are analogous to rotations. If the space has
a finite dimension, there are indeed rather few rotations, but the set of
canonical transformations is always infinite dimensional because the space of
all functions has infinite dimensions. (The discerning reader may think that
the comparison is unfair, since rotations take place in a vector space, while
the kind of transformation we now allow are more general than that. However,
it can be shown that the set of transformations leaving a given metric tensor
invariant is at most as large as the set of translations and rotations in a vector
space, and indeed often smaller than that. So the objection has no force.)
Let us revisit the discussion of Noethers theorem in section 2.3. We used
it already to motivate the definition of the canonical momenta, but there is
more to say about it. Let us choose spatial rotation, eq. (2.44), as an example
of a transformation that we will want to make. It is easy to see that
Li = ijk xj pk
{xi , j Lj } = ijk j xk = xi .
(8.72)
The Noether charge generates the transformation via the Poisson bracket!
More is true. The -tensor obeys the identity
ijk kmn + ink kjm + imk knj = 0 .
(8.73)
By means of this identity it is easy to show that

{Li , Lj } = ijk Lk .
(8.74)
The educated way of describing this result is to say that the rotational Noether
charges form a Poisson bracket representation of the Lie algebra of the rotation
group. In quantum mechanics, we have a commutator representation of the
same Lie algebra.
8.7 The symplectic one-form
101
These results concerning rotations can be generalized to any symmetry and

any Noether charges. Let us just mention time translation invariance, eq.
(2.46). Here the Noether charge is the Hamiltonian itself. In perfect analogy
to our discussion of rotations, we find
xi = x i = {xi , H} .
(8.75)
Again the Noether charge generates the transformation to which it owes its
existence.
8.7 The symplectic one-form

To proceed further with canonical transformations we need to develop some
more machinery. First a little discussion of vector fields. A contravariant vector
field (with indices upstairs) is an assignment of a little arrow to each point in
a space. These arrows can always be regarded as the tangent vectors of a set
of curvesthe flowlines of the vector field. Hence the general definition: A
dynamical system is a vector field on a manifold. This is a mathematical way
of saying that we are dealing with equations of the form z = f (z), for some
functions f .
In vector spaces we thought of covariant vectors (with indices downstairs)
as assigning a set of parallel planes, as shown in Fig. 8.2. So we would expect
that a covariant vector field is a set of level surfaces of some function. But this
is simply not true in general. It is true if and only if the covariant vector field
is the gradient of some function, u = f . And it can be shown that this is
so if and only if
u u = 0 .
(8.76)
If this condition fails the level planes that exist at each point do not fit together
to form level surfaces extending over all space. Actually this phenomenon is
well known from the theory of curve integrals. A covariant vector field is just
what we need in the integrand of such an integral. Suppose that a curve
is defined explicitly by the functions z = z (t), where t is some parameter
along the curve. Then
Z
dz u (z) =
t2
dt
t1
dz
u (z(t)) .
dt
(8.77)
It is known that this integral is independent of the path and defines a function
f (z(t)), if and only if condition (8.76) holds. But the curve integral is well
defined regardless of whether this is true or not.
It makes sense: in a vector space a covariant vector is a map from the set
of all arrows to the real numbers, and in general a covariant vector field is a
map from the set of all flowlines to the real numbers.
This leads to a useful piece of notation. If the coordinates on our space
102
are denoted by z and if u is a covariant vector field, then we define the

differential one-form du as
du = dz u .
(8.78)
That is, the coordinate differentials dz are chosen as the basis in which we
expand the covariant vector. The gradient of a function f defines a one-form
of a special kind, namely
df = dx f .
(8.79)
The notation is useful since it gives the correct behaviour of u (x) under
coordinate transformations. Let x = x (x). Then
du = dx u = dx
x
u = dx u
x
u (x ) =
x
u (x) .
x
(8.80)
Notice that everything happens automatically.

We brought this up because there is a symplectic one-form lurking behind
the symplectic two-form . To see why we come back to eq. (8.35), which
was part of the definition of the latter. If the components of are constants
it obviously holds, but in general it is something that must be checked. The
same equation occurs in electrodynamics when written in tensor form. It is
this equation that implies that the vector potential exists:
+ + = 0
= .
(8.81)
Here is some covariant vector field, known in symplectic geometry as the

symplectic one-form. (If you like, one-form is simply another name for covariant vector field. The if and only if statement is true only if the topology
is trivial.)
Although I have stated it without proof, this result is extremely important.
When electrodynamics is written in tensor notation one finds that the electric
and magnetic fields can be written as a two-form
F =
0
Ei
Ei ijk Bk
0 E1 E2 E3
E1
0
B3 B2
.
=
E2 B3
0
B1
E3 B2 B1
0
(8.82)
Eq. (8.81) are part of Maxwells equations; they guarantee the existence of
the vector potential, which is a one-form. There is also a second tensor equation relating the two-form to the electric current, but this has no analogue in
symplectic geometry.
What does the symplectic one-form look like if we use our standard coordinates (qi , pi ) on phase space, so that the symplectic two-form takes the form
8.8 General transformation theory
103
(8.36)? There is no unique answer to this question since the previous discussion
implies that if
= +
(8.83)
then and give rise to the same symplectic two-form. However, a possible
choice of symplectic one-form is
= pi dqi .
(8.84)
This will be our standard choice. To make sure that you see how it works,
consider a two dimensional phase space with
z =
It is simple to verify that
=
q
p
0
q 2 p 1
p 1 q 2
0
p, 0
0 1
1 0
(8.85)
{q, p} = 1 . (8.86)
More generally, using eq. (8.38) we recover our standard symplectic two-form
(8.36) in any dimension.

So far we have discussed infinitesimal canonical transformations only, and we
must now turn to finite ones. We will give the general transformation theory for
the special case of two dimensional phase spaces only; however, every detail of
the following arguments does have a natural and straightforward generalization
to the higher dimensional cases, so the loss of generality is only apparent.
First of all we adopt a practice introduced by Whittaker, who denoted the
transformed coordinates by capital letters rather than primes. Thusin the
active interpretationthe point labelled by (q, p) is transformed to the point
labelled (Q, P ),
Q = Q(q, p)
P = P (q, p) .
(8.87)
This is a canonical transformation if the Poisson brackets obey

{Q, P } = {q, p} .
(8.88)
It is more convenient to study the transformation of the symplectic one-form.

The transformation is canonical if it leaves the symplectic one-form unchanged.
In fact we can be a little more general, because if we add a gradient to the
symplectic one-form the symplectic two-form, and hence the Poisson brackets,
are unchanged. Hence the transformation is canonical if and only if
104
pi dqi = Pi dQi + dF ,
(8.89)
where F is some function F = F (Q, P ).

Let us assume that the canonical transformation we are about to construct
is such that q and Q together also can serve as coordinates on phase space.
(See exercise 10.) Denote the function F by F1 = F1 (q, Q). Then we derive
that
(p q F1 ) dq (P + Q F1 ) dQ = 0 .
(8.90)
Since q and Q serve as coordinatessince dq and dQ are linearly independent

one-formswe can conclude from this that
p = q F1 (q, Q)
P = Q F1 (q, Q) .
(8.91)
We assume that these equations are invertible, that is to say that they can be
used to derive the explicit formulas
Q = Q(q, p)
P = P (q, p)
q = q(Q, P )
p = p(Q, P )
(8.92)
.
This is the canonical transformation generated by the generating function

F1 (q, Q).
You may have wondered how I knew that I could use q and Q as coordinates
on phase space. The answer is that I did not know, but in a concrete situation
I would have noticed whether anything was wrong once I tried to invert eqs.
(8.91) in order to obtain the explicit canonical transformation. If you think
about it, it is clear that there cannot exist a generating function F1 (q, Q)
generating the simplest canonical transformation we can think of, namely the
identity transformation
Q=q
P =p.
(8.93)
Evidently the pair q and Q does not coordinatize phase space. To circumvent
this difficulty we define a new generating function by means of a Legendre
transformation from the first:
F2 = F1 (q, Q) + QP .
(8.94)
The sign conventions here are a little odd. Still the relation between the functions F2 and F1 is recognisably a Legendre transformation, and it followsby
virtue of eqs. (8.91)that F2 is a function of q and P only. If we assume that
q and P together coordinatize phase spacewhich is certainly true for the
identity transformationwe can use F2 to generate canonical transformations
by repeating the previous logic:
105
pdq = d(P Q)QdP +F1 = QdP +dF2 = QdP +q F2 dq+P F2 dP . (8.95)

Comparing coefficients in front of the linearly independent one-forms, in this
case dq and dP , we conclude that
p = q F2
Q = P F2 .
(8.96)
Again we assume that these equations can be inverted to yield eqs. (8.92) in
explicit form. The identity transformation is generated by
F2 (q, P ) = qP .
(8.97)
We can go on to define the generating functions F3 = F3 (p, Q) and F4 =

F4 (p, P ) by means of further Legendre transformations, in the obvious way.
We have recovered the conclusion from section 8.6, namely that every function on phase space generates a canonical transformation, but now we have a
recipe for how to find such transformations in finitenot only infinitesimal
form. Hamiltons equations retain their form under all canonical transformations, with a new Hamiltonian defined as the function
K(Q, P ) = H(q, p) = H(q(Q, P ), p(Q, P )) .
(8.98)
In chapter 9 we will regard canonical transformations as passive coordinate

transformations, and we will try to design a canonical transformation to ensure that the new Hamiltonian takes a very simple standard form, so that
the equations of motion are always soluble once we have arrived at the new
coordinates Q and P .
By the way, notice that the mathematical manipulations we are doing resemble those of thermodynamics. In fact there is a close relation between
symplectic geometry and thermodynamics, although in the latter case the situation is a little differentin thermodynamics one always starts out in an
odd dimensional space, with coordinates U, (T, S), (P, V ), and so on. In phase
space there is no analogue of the extra coordinate U .
The setup can be generalized somewhat, by allowing the generating transformations to depend on time. Hamiltons equations will again retain their
form, but the numerical value of the Hamiltonian will change. In fact, at least
on the level of Gedanken calculations, it is possible to choose a time dependent
generating function ensuring that the transformed Hamiltonian vanishes, so
that the resulting equations of motion solve themselves. This is the subject of
Hamilton-Jacobi theory, but we will not go into this.
Problem 8.1
Derive the Hamiltonian corresponding to the Lagrangian for a
charged particle in an external field, eq. (2.10).
Problem 8.2
The Lagrangian for a relativistic particle is
106
p
L = mc c2 x 2 ,
x 2 = x i x i .
(8.99)
Show that this reduces to the ordinary free particle when c is large compared to the
velocity. Then derive the Hamiltonian for the relativistic particle.
Problem 8.3
Let the coordinates of a space transform according to x
A x . Prove that this implies that the gradient of a function transforms like a covariant vector, f f (A1 ) .
Problem 8.4
Show that, given any metric g (with positive eigenvalues)
on a vector space, one can always change the coordinates x to new coordinates
X = A x , in such a way that
x g y = X G Y ,
(8.100)
and such that the new metric tensor G becomes equal to a Kronecker delta. Show
that if we instead require
y x = Y X
(8.101)
then the symplectic form in the new coordinates can always be made to assume the
standard form (8.36).
Problem 8.5
Use eq. (8.35) to prove the Jacobi identity (8.47) for an arbitrary symplectic form (i.e. with components that may be non-constant phase space
functions).
Problem 8.6
Replace Hamiltons equations with the seemingly equally nat-
ural
q = q H
p = p H .
(8.102)
Is there an analogue of Liouvilles theorem? If you choose H to be the harmonic

oscillator Hamiltonian, what would the phase space flow look like?
Problem 8.7
Consider Eulers equations for a spinning top. Write them in
terms of Ji . There are two conserved quantities of the form G = G(J1 , J2 , J3 ) and
H = H(J1 , J2 , J3 ). Prove that the time development of any function F = F (J1 , J2 , J3 )
is given by
F = ijk i F j Gk H ,
(8.103)
where ijk is totally anti-symmetric and the partial derivatives are with respect to Ji .
Indices run from 1 to 3. For your information, this is called Nambu mechanics.
Problem 8.8
Evaluate the mutual Poisson brackets enjoyed by the three
conserved charges (2.32) of the particle on the sphere, using polar coordinates for
the calculation. Do the same for the three conserved charges of the particle on the
hyperboloid, considered in problem 2.8.
Problem 8.9
Consider two Lagrangians related by a total derivative,
107
d
1
d
(qi ) = qi qi V (qi ) + (qi ) ,
(8.104)
dt
2
dt
for some function . Derive the Hamiltonian formulations of these two Lagrangians,
and show that there is a canonical transformation relating them. Show explicitly that
Hamiltons equations are equivalent in the two cases. (Have you seen this Lagrangian
before?)
L (qi , qi ) = L(qi , qi ) +
Problem 8.10
Transform from Cartesian coordinates (x, y) to polar coordinates (r, ). Do the functions (x, r) coordinatize the plane? The upper half plane?
Problem 8.11
Show that the transformation

Q = ln
sin p
q
P = q cot p
(8.105)
is canonical, and determine the generating functions F1 and F2 .
Problem 8.12
Find the canonical transformation generated by

F1 = kq 2 cot Q ,
where k is a constant.
(8.106)
9 Integrable and chaotic motion
We have arrived at the Hamiltonian form of the equations of motion, as well as

the most general transformations in phase space leaving that form invariant.
Can these results be used to describe the properties of the solutions with any
generality? Partial answers to this question can in fact be given, and this is
the topic of this chapterwhich will, however, end in chaos.
9.1 Can chaos occur?

Although we know that the solution to a generic dynamical systema system
of first order coupled ordinary differential equationsexists and is unique,
given the initial data, it is not clear that we can expect to actually find these
solutions with a reasonable expense of effort. This depends to a large extent
on the dimension of phase space. If the dimension is two it is clear that the
solutions cannot look too bad. As long as the system is autonomousi.e.
when the right hand side of eq. (1.39) does not depend explicitly on time
the key observation is that the flow lines in phase space never cross. If we fill
a two dimensional plane (or some other two dimensional surface) with nonintersecting flow lines we necessarily get a fairly orderly pattern, and hence
the solutions of the equations of motion will be orderly too. For a Hamiltonian
system, with a Hamiltonian of the form
H=
p2
+ V (q) ,
2
(9.1)
we obtain a differential equation that determines the shape of the flowlines:
q = p
p = q V
dp
q V
=
.
dq
p
(9.2)
The solution for the curves p = p(q) is found to be

p=
2(E V (q)) ,
(9.3)
9.1 Can chaos occur?
109
where the energy E is an integration constant that labels the individual flow
lines. And this is almost the end of the story.
Note that for non-autonomous systems the evolution equations depend explicitly on t, as in eq. (1.38). This includes systems that are driven by some
time-dependent external force. For them it is no longer true that the phase
space flow lines are non-intersecting. In this case we expect that trouble will
arise already in two dimensional phase spaces.
If the dimension of phase space is three things can become much more
complex. Once the flow lines can move in a third dimension there is absolutely
no guarantee that they form an orderly pattern. It is instructive to compare
two three dimensional systems that we have encountered, namely the Lorenz
and Euler equations. In the Euler equations (6.57) the phase space coordinates
can be taken to be J1 , J2 , J3 . We can solve the equations because any solution
Ji (t) is a curve that lies on two well behaved surfaces
J12 + J22 + J32 = constant ,
J12 J22 J32

+
+
= 2E .
I1
I2
I3
(9.4)
Hence any solution curve is the intersection of a sphere and an ellipsoid, and
such a curve is not difficult to describe, nor is the set of all such curves hard
to understand. The lesson is that the solutions of the Euler equations are easy
to describe precisely because there exist two constants of the motion, each
describing a well behaved surface in phase space. With the Lorenz equations
(1.45) things are very different: in this case there simply does not exist any
well behaved constants of the motion, and hence there is no reason to believe
that the set of all solutions can be described in any easy and comprehensive
manner. And indeed the solutions are chaotic, in particular they turn out to
be chaotic in the sense that the smallest change in the initial data will change
the long time behaviour of the solution in a dramatic fashion. The point here
is simply that one expects this to happen if the dimension of phase space is
three or moreunless there are special reasons to think otherwise.
Still, the Lorenz equations are not Hamiltonian. They cannot be, because
their phase space has odd dimension. Now Hamiltonian phase space flows have
some rather special properties, such as that uncovered by Liouvilles theorem
they are like the flow of an incompressible fluid, and certain special kinds of
fixed points cannot occur. Hence there is still some hope that the long time
behaviour of Hamiltonian systems may be significantly simpler than that of
a typical dynamical system. This is what we will look into next. (Eventually
our hope of understanding the detailed behaviour of the general solution of a
general Hamiltonian system will be completely dashed, but there will be many
compensations.)
To study the kind of phase space flows that can occur we must be able to
picture them effectively. If the space to which the flow is confined is three
dimensionaleither because the phase space has three dimensions or because
the flow is confined to a three dimensional energy surface in a four dimensional
phase spacethe Poincare section solves the problem for us. The idea is to
110
Integrable and chaotic motion
Figure 9.1. A Poincare section displaying an orbit in three dimensions as a set

of dots on a plane; a periodic orbit giving rise to just one dot occurs in the
corner of the plane.
study a two dimensional cross section of the three dimensional space, through
which the orbits pass in some definite direction. Each time the orbit passes
out through the section one marks the corresponding point with a dot. If the
orbit is periodic, the number of dots one obtains is finite. If the orbit is highly
irregular, the pattern of dots will be highly irregular too.
9.2 Integrable systems

The harmonic oscillator is a very simple example of a soluble system. We
will now discuss it using the whole apparatus of canonical transformations, in
the hope that this will lead to a general strategy for how to deal with more
complicated examples. The phase space flow takes place on circles surrounding
a single fixed point, and the idea is to introduce new canonical coordinates
(Q, P ) such that P labels the flow line we are on, while Q tells us where we are
on the given flow line. We insist that the transformation to these coordinates
should be canonical because we want to make maximum use of the special
structure of Hamiltonian dynamics.
The Hamiltonian of the harmonic oscillator is
2 2
1
q ,
H = p2 +
2
2
(9.5)
and the explicit solution of Hamiltons equations is given by

q = A sin (t + )
p = A cos (t + ) ,
(9.6)
9.2 Integrable systems
111
where the amplitude A is given in terms of the energy E by
2 2
2E
E=
A=
A
.
(9.7)
2
We define the transformation to new phase space coordinates (Q, P ) by

q=
2P
sin Q
p=
2P cos Q .
(9.8)
It follows that
E = H = P ,
(9.9)
so indeedup to a constant factorone of the new coordinates is just the

energy, and we already know that the energy serves to label the flow lines in
phase space. Moreover, using the tricks introduced in section 8.4, a calculation
verifies that
{q, p} =
q p
q p
= = 1 = {Q, P } .
Q P
P Q
(9.10)
The transformation is indeed canonical. After the transformation Hamiltons

equations take the simple form
Q = {Q, H} =
P = {P, H} = 0 .
(9.11)
In these coordinates the solution is trivial.

This simple example suggests a general strategy for solving Hamiltons equations of motion for n degrees of freedom. In the first step we must find n
constants of the motion Ii , such that
{Ii , Ij } = 0 .
(9.12)
The idea is that each of these constants of the motion will serve as a member
of a canonical pair, whose other member will be called i . In the second step
we must devise a canonical transformation
qi = qi (, I)
pi = pi (, I) ,
(9.13)
such that the Hamiltonian has no dependence on the ,

H = H(I) .
(9.14)
If we succeed in this Hamiltons equations will take the soluble form

i = Ii H(I)
Ii = i H(I) = 0 .
(9.15)
The new canonical variables Ii and i are called action-angle variables. If such
112
variables exist the Hamiltonian system is said to be integrable. Since every i

is constant the motion is associated to n constant frequencies i .
In an integrable system every trajectory in the 2n dimensional phase space
is confined to an n dimensional submanifold, defined by the n conditions Ii =
constant. We also have available n vector fields
Ii = Ii .
(9.16)
The normal vector of any hypersurface defined by the equation Ii = constant

is
ni = Ii ,
(9.17)
and it is clear that

{Ii , Ij } = 0
ni Ij = 0 .
(9.18)
Thus these vector fields are tangential to all the n hypersurfaces, and therefore
tangential to the n-dimensional submanifold. Moreover they must be everywhere non-vanishing, because
1 I1 = 1 I1 = {1 , I1 } = 1 6= 0
(9.19)
and so on for all the n vector fields. We have found n everywhere non-vanishing
vector fields pointing along the surface of the n dimensional submanifold defined by the n equations Ii = 0. Let us assume for simplicity that these submanifolds are closed and bounded. Then our conclusion is rather remarkable,
because very few closed and bounded manifolds admit even a single everywhere non-vanishing vector field. (The circle does, but the sphere does not
you cannot comb a sphere.) Closer inspection shows that eq. (9.12) means
that an additional technical condition on these vector fields is obeyed, namely
that the vector fields commute. Without bothering too much about what this
means, we can then rely on the following mathematical theorem:
The only closed and bounded n dimensional manifold admitting n everywhere
non-vanishing commuting vector fields is the n dimensional torus.
A one dimensional torus is a circle, and a two dimensional torus is an ordinary
torus, or more abstractly it is like a square with periodic boundary conditions.
A three dimensional torus is like a cube with periodic boundary conditions,
and so on. And the conclusion is that in an integrable Hamiltonian system all
trajectories are confined to tori with half the dimension of the entire phase
space.
The motion on a torus depends on the n frequencies that characterize the
motion on the given torus. The trajectories will be open or closed depending
on whether the frequencies are rationally related (as in the Kepler problem),
or not (as in the case of a Lissajous figure that never repeats, see section 1.3).
113
Meanwhile, our emphasis on conserved quantities is now understandable. We

also see why the three components of the conserved angular momentum vector
really contributed only to the decoupling of two of the degrees of freedom in
the central force two body problem. What we need for the general strategy are
constants of the motion that Poisson commute, as in eq. (9.12). From the
angular momentum vector we can construct two such constants of the motion,
namely (say) L3 and L2 . We know that
{Li , Lj } = ijk Lk
{L3 , L2 } = 0 .
(9.20)
Therefore L3 and L2 can serve as action variables. Clearly this is reminiscent

of quantum mechanics, with its emphasis on commuting operators.

The canonical transformation to action-angle variables will be difficult to find
in general, and we will have to resort to canonical perturbation theory to construct it order by order in some parameter. There is a well developed technology for doing this. In 1969 it took us to the Moon.
The first step then is to split the Hamiltonian function into two parts,
H(, I) = H0 (I) + Hint (, I) ,
(9.21)
where the split is defined in such a way that the equations of motion coming
from H0 alone are integrable. It is assumed that is some small parameter,
and the full problem is to be solved as a power series expansion in .
Already in two dimensions we can see some problems with this idea. Consider the case of a pendulum, with the Hamiltonian
1
2
H = p2
cos q .
(9.22)
2
2
The problem is not the non-linearity of the equations of motion as such. The
problem is that, as discussed in section 1.4, phase space splits into three regions separated by separatrices. The two separatrices, and the hyperbolic fixed
point, all have the same energy E, so the conserved energy will not do as a coordinate uniquely labelling the orbits. A related problem is that motion along
a separatrix is not periodic; formally it corresponds to zero frequency, because
it takes an infinite amount of time to traverse it.
When the energy is large the effect of gravity is small (the pendulum is
rotating almost freely). Thus we can choose
1
(9.23)
H 0 = p2 ,
2
and treat as a small parameter. For small oscillations on the other hand we
can set
114
Figure 9.2. The Kirkwood gaps. The asteroids clustered around 1 : 1 are not a
contradictionthey are the Greek and Trojan asteroids at the Lagrange points.
1
2
H 0 = p2 + q 2 .
2
2
(9.24)
The unperturbed system behaves as a harmonic oscillator. Supposing that we

obtain the solution as two different power series based on these two different
ways of splitting the Hamiltonian, how can we join them to obtain the
behaviour at the separatrix? Actually what will happen is that the question of
the convergence of the series will raise its ugly head, and the complete solution
for the pendulum will escape us.
In a two dimensional phase space there is trouble only close to the separatrix,
where the frequency vanishes. What can go wrong in higher dimensions? In
four dimensions it turns out that there will be trouble in the neighbourhood
of those tori where the frequencies obey
n1 1 + n2 2 = 0
(9.25)
for some integers n1 , n2 . Moreover the trouble is most severe if these integers
are small.
So are almost all Hamiltonian systems integrable? No. But if the Hamiltonian is of the form (9.21), with an that is not too large, then large regions
of phase space will still be filled with tori. In between there will be regions
where the tori have been destroyed and the motion is chaotic. The chaotic
regions grow in size with , and the tori that disappear first are those tori for
which the unperturbed motion obeys eq. (9.25) for small integers n1 , n2 . The
full story here is known as the KAM theorem, for Kolmogorov, Arnold, and
Moser.
A beautiful example of the KAM theorem in action is provided by the distribution of asteroids inside Jupiters orbit. Presumably the asteroid belt was
originally created in such a way that the number of asteroids as a function
of their angular frequencies could be approximated by a fairly smooth function. Then, as time goes on, an asteroid whose trajectory was on a torus in
115
phase space will remain there, while an asteroid on a chaotic orbit sooner or
later changes its angular frequency significantly. Let 1 be the frequency of
the asteroids unperturbed motion, and 2 that of Jupiters. The theory then
suggests that there will be gaps in the asteroid distribution, so that asteroids with an 1 obeying eq. (9.25) for small values of n1 and n2 are missing.
Observation bears this out, and the gaps are known as the Kirkwood gaps for
their discoverer. An exception is that the number of asteroids with 1 = 2
is particularly largebut these are not an exception to the KAM theorem,
rather they are the Trojan asteroids that we discussed in section 4.6.
Figure 9.3. Poincare sections, in the (y, y)

plane, for the Henon-Heiles Hamiltonian. To the left you see the prediction from eighth order perturbation theory,
to the right numerical results. The chaotic dots in the lower two figures come
from a single orbit.
Another beautiful illustration is provided by the Henon-Heiles Hamiltonian

1
1
1
1
y3
H = (p2x + p2y ) + V (x, y) = (p2x + p2y ) + x2 + y 2 + x2 y
. (9.26)
2
2
2
2
3
This toy model was originally inspired by the study of stars moving in the
116
gravitational field of a galaxy. The potential is not bounded from below, but
there is a potential well near the origin. Note that
1
1
V (x, y) =
6
3

1
y+
(y 1 3x)(y 1 + 3x) .
2
(9.27)
Setting this to zero defines three straight lines that bound a triangular well.
Provided the energy does not exceed E = 1/6 motion can be confined inside
this triangle, and the energy surface is bounded. The influence of the cubic
terms in the potential becomes more pronounced as the energy goes up.
To picture the dynamics it is convenient to study a Poincare section based
on some two dimensional cross section of phase space, say the plane spanned by
the coordinates y and py . When studying a trajectory (presumably generated
by a computer) one makes a dot on the two dimensional plane whenever the
trajectory passes through it. If the trajectory is confined to a torus in phase
space one should then see these dots lining up along some one dimensional
curve in the resulting picture. If this does not happen one concludes that the
toriwhich would be present if the cubic terms in the Hamiltonian could be
ignoredhave been destroyed by the perturbation.
What Henon and Heiles found illustrates the KAM theorem quite well. For
E = 1/24 and E = 1/12 the actual trajectories do give rise to closed curves on
the Poincare section, and these curves lie exactly where they should lie according to an eighth order canonical perturbation theory calculation performed by
Gustavson. Above E = 1/9, when the importance of the cubic terms is larger,
one finds orbits that definitely do not lie on tori since they fill large parts of the
picture with chaotic dots. But there will also be orbits (even when E > 1/6)
that do lie on tori, so the picture becomes that of a mixture of order and chaos.
9.4 Stability of the Solar System

Since Newton many of these developments were driven by the wish to know
whether the Solar System is stable in the long run. Its Hamiltonian includes
eight planets and the Sun, mutually attracting each other according to Newtons law of gravity, as well as numerous other bodies whose masses were not
well known to the pioneers. A brief account of what has been learned seems
an appropriate way to end this story.
To lowest order we simply ignore the attractions between planets, and the
system is integrable as we have seen. However, the planets in fact do not move
exactly on Kepler ellipses. Indeed it was realised early onby Halley, using
observational data going back to 228 B.C.that Jupiter is accelerating and
Saturn is decelerating. Moreover the effect is not small. If it were extrapolated
linearly one would conclude that these two planets were at the same distance
from the Sun six million years ago, raising the issue of a possible collision
between them. If one tries to include the interplanetary forces in perturbation theory the question becomes whether this is a periodic perturbation, or
9.4 Stability of the Solar System
117
whether it is proportional to time, in which case it is said to be a secular

perturbation. Laplace answered this question to lowest order in perturbation
theory: the observed variations in the orbits of Jupiter and Saturn are periodic, with a period of about 900 years. Thus they do not threaten the long
term stability of the solar system.
In the nineteenth century it was realised that higher order perturbation
theory leads to terms containing expressions such as ni 1 + n2 2 in their
denominators. This means that higher order terms can become very large when
the frequencies involved are in resonance. Eventually Poincare dashed all hopes
that the series will converge for arbitrary intitial data. The KAM theorem on
the other hand clearly suggests the possibility that planets nevertheless lie on
regular orbits, since there will be regions in phase space with this behaviour.
In the twentieth century it became clear that the question is relevant only
over a timescale of about 109 years, since this is the expected lifetime of the
Sun. At the end of the twentieth century extensive calculations were made on
computers especially designed for the purpose. (Such a computer was known
as a digital orreryorreries were first constructed using cogwheels, but the
clockwork metaphor of mechanics is by now out of fashion.) What one finds is
that trajectories arising from near-by initial data may diverge exponentially
in time, with an e-folding time of the order of 5 106 years. The conclusion
therefore seems to be that accurate prediction in the form of ephemeridis tables
is possible over timescales of 107 years, but not possibleat least not for the
inner planetsover 108 years. On the other hand the qualitative properties of
the Solar System do seem to be stable over much longer time scales.
Qualitative questions one can ask concern, for instance, the bounds within
the eccentricities of the orbits are likely to vary. If the eccentricity of the orbit
of Mars reaches 0.3 and that of the Earth reaches 0.1 a collision might ensue.
The most recent results are not altogether reassuring.1 Over a timescale of
1010 years variations in the eccentricities can be ignored as far as the four
major planets are concerned, but there is a potential problem with Mercury.
The Newtonian theory gives a 60 % probability that its eccentricity will grow
to such an extent that it causes a dramatic rearrangement of the inner Solar
System sometime during the next 5109 years. Mercury may collide with Venus,
and Mars may be ejected from the system. Fortunately this probability drops
to about 1 % if corrections from General Relativity are taken into account,
because the increase in the perihelion precession of Mercury turns out to be
helpful. There is no room for the addition of a fifth inner planetif there ever
was one it was ejected from the system or disappeared in a collision (perhaps
giving rise to the Moon). The Solar System seems to be at most marginally
stable over this timescale.
The Solar System would not be stable if the mass of its Sun were significantly
less than its actual mass. Hence the Solar System as we know it is unlikely
to survive the post-main-sequence mass loss of the Sun, which is expected to
occur in about 5 109 years. And this is the end of the story.
1
J. Laskar, Is the Solar System Stable?, Poincar

e S
eminar, Paris 2010.
118
Problem 9.1
Consider the Hamiltonian (9.1), and let V (q) be a fourth order
polynomial. Draw the phase space flow for all qualitatively different choices of this
polynomial. Pay special attention to the transition cases, when a maximum and a
minimum merge to produce an inflection point on the graph of V (q).
Problem 9.2
Solve the evolution equation

x = x + x2
(9.28)
as a series expansion in , to second order in . Then solve the equation exactly, and
compare the results. As initial condition, set x(0) = k.
Problem 9.3
For the purposes of canonical perturbation theory it is convenient to write the Henon-Heiles Hamiltonian as
1 2
y3
(px + p2y + x2 + y 2 ) + (x2 y ) .
(9.29)
2
3
Explain why small coupling constants correspond to small values of the total energy
in eq. (9.26). What is the nature of the energy surface when is small? When it is
large?
H=
Problem 9.4
Once we know that integrable motion takes place on tori on
phase space we become interested in doubly periodic functions. Show that the elliptic
function = (t) which solves the equation for the pendulum has this property. Do
this indirectly by showing that the equations make sense also if we make t purely
imaginary. Then let the physics tell you that if t is taken to be a complex variable
(t) must be periodic under purely real and purely imaginary shifts in t.
Appendix 1 Books
I recommend the following books:

V. I. Arnold: Mathematical Methods of Classical Mechanics, Springer 1978.
Written by the person who understood the subject best.
C. Lanczos: The Variational Principles of Mechanics, Dover 1986. A beautiful book with a philosophical slant.
L. D. Landau and E. M. Lifshitz: Mechanics; Pergamon 1960. Traditional,
and perfectly clear.
I. Percival and D. Richards: Introduction to Dynamics, Cambridge 1982.
Dynamical systems point of view; selective coverage.
J. R. Taylor: Classical Mechanics, University Science Books, 2005. Helpful
as an introduction at a somewhat lower level.
You may also enjoy:
Max Born, Cause, Purpose and Economy in Natural Laws, in: Physics in My
Generation, Springer 1969. The text of a lecture on variational principles.
Julian Barbour: The End of Time, Weidenfeld and Nicolson 1999. A stimulating account of philosophical issues raised by Newtons laws.
Ernst Mach: Die Mechanik in ihrer Entwickling, many editions (avoid the
ninth!). For the prehistory of mechanics.
A. Dahan Dalmedico et al.: Chaos et determinisme, Editions

du Seuil 1992.
An interesting collection of articles about chaos.

Anmek

Uploaded by

Copyright:

Available Formats

Anmek

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Anmek

Uploaded by

Copyright:

Available Formats

Notes on Analytical Mechanics

Ingemar Bengtsson (2014)

Best of all Possible Worlds

central force two-body problem

7 Interlude: Legendre transformations

9 Integrable and chaotic motion

1 The Best of all Possible Worlds

Mechanics is the paradise of the mathematical

The Best of all Possible Worlds

1.1 Analytical and Hamiltonian mechanics

1.1 Analytical and Hamiltonian mechanics

It follows that if the force is given by

As a definition, this is a little vague. M

The Best of all Possible Worlds

Fermat noted that if

S is known as the action. It is a functional, i.e. a function of a functionthe

1.1 Analytical and Hamiltonian mechanics

The Best of all Possible Worlds

Only in the twentieth century was it realized that Hamiltons Principle

1.2 The calculus of variations

How do we find the extrema of such a function of a function? Let us begin

We assume that it is possible to bring S to this form. Then the functional

Namely that of quantum mechanics.

1.2 The calculus of variations

right in the integrand. The equations of motion, as obtained from Hamiltons

The Best of all Possible Worlds

for those and only those functions which obey

at all intermediate points.

From the point of view of the

we choose there is always a unique solution for some

1.3 How to solve equations

where f is some function. This can be solved by means of separation of variables:

where c is a constant determined by the initial condition. If we do this integral,

1.3 How to solve equations

Let us therefore approach eq. (1.25) in a systematic fashion, which might

The Best of all Possible Worlds

The constant k is undetermined at this stage. The integral becomes

The integrand simplifies if we choose the constant k such that

One further substitution takes us to our desired standard form;

Just as eq. (1.30) can be taken as an implicit definition of a trigonometric

1.3 How to solve equations

studied in depth by mathematicians. Their work remains relevant, even if

The Best of all Possible Worlds

1.4 Phase space

where the N functions fi are arbitrary. We simplify things by assuming that

where z0i are the initial values of zi .

1.4 Phase space

The Best of all Possible Worlds

1.4 Phase space

2E2 = p22 + 22 x22 .

In the linearised case ( = 0) there is a single elliptic fixed point. In the

The Best of all Possible Worlds

They capture some aspects of thermal convection in a fluid. Lorenz was a

Problem 1.2 A particle of mass m1 = 1 kg and a particle of mass m2 = 1

Derive the rocket equation

1.4 Phase space

Problem 1.13 Show that a non-autonomous dynamical system can always be

We use q to denote the coordinates because the Lagrangian formalism is very