Anmek Analytic Mech

Notes on Analytical Mechanics
Ingemar Bengtsson (2012)

Contents
1 The Best of all Possible Worlds page 1
1.1 Analytical and Hamiltonian mechanics . . . . . . . . . . . . . 2
1.2 The calculus of variations . . . . . . . . . . . . . . . . . . . . . 5
1.3 How to solve equations . . . . . . . . . . . . . . . . . . . . . . 8
1.4 Phase space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2 Lagrangian mechanics 17
2.1 The scope of Lagrangian mechanics . . . . . . . . . . . . . . . 17
2.2 Constrained systems . . . . . . . . . . . . . . . . . . . . . . . . 20
2.3 Symmetries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3 Interlude: Conic sections 30
4 The central force two-body problem 34
4.1 The problem and its formal solution . . . . . . . . . . . . . . . 34
4.2 Existence and stability of circular orbits . . . . . . . . . . . . . 37
4.3 Keplers First Law . . . . . . . . . . . . . . . . . . . . . . . . . 38
4.4 Keplers Third Law . . . . . . . . . . . . . . . . . . . . . . . . 40
4.5 Self-similarity and the virial theorem . . . . . . . . . . . . . . 41
4.6 The three-body problem . . . . . . . . . . . . . . . . . . . . . . 44
5 Small oscillations 49
5.1 Forced oscillations . . . . . . . . . . . . . . . . . . . . . . . . . 49
5.2 Damped and forced oscillations . . . . . . . . . . . . . . . . . . 52
5.3 Several degrees of freedom . . . . . . . . . . . . . . . . . . . . 54
6 Rotation and rigid bodies 58
6.1 Rotations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
6.2 Rotating coordinate systems . . . . . . . . . . . . . . . . . . . 63
6.3 The inertia tensor . . . . . . . . . . . . . . . . . . . . . . . . . 66
6.4 Eulers equations . . . . . . . . . . . . . . . . . . . . . . . . . . 69
6.5 The Lagrangian description . . . . . . . . . . . . . . . . . . . . 73
7 Interlude: Legendre transformations 78
ii Contents
8 The Hamiltonian formulation 81
8.1 Hamiltons equations and Hamiltonian ows . . . . . . . . . . 81
8.2 Kets and bras and all that . . . . . . . . . . . . . . . . . . . . 84
8.3 The symplectic form . . . . . . . . . . . . . . . . . . . . . . . . 88
8.4 The algebraic structure of mechanics . . . . . . . . . . . . . . . 91
8.5 Innitesimal canonical transformations . . . . . . . . . . . . . 93
8.6 The symplectic one-form . . . . . . . . . . . . . . . . . . . . . 95
8.7 General transformation theory . . . . . . . . . . . . . . . . . . 97
9 Integrable and chaotic motion 102
9.1 Can chaos occur? . . . . . . . . . . . . . . . . . . . . . . . . . 102
9.2 Integrable systems . . . . . . . . . . . . . . . . . . . . . . . . . 104
9.3 Canonical perturbation theory . . . . . . . . . . . . . . . . . . 107
Appendix 1 Books 110
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 x
i
= F
i
(x, x) . (1.1)
3. The force exerted by a particle on another is equal in magnitude, but op-
posite in direction, to the force exerted by the other particle on the rst.
A particle is here thought of as an entity characterized by its mass m, its
location in space x(t), and by nothing else. 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 dierential
equations above.
An overdot denotes dierentiation with respect to the time parameter t (this
notation, as well as Dierential Calculus itself, was invented by Newton), and
x
i
may denote a vector in three-dimensional space. Sometimes it will be un-
derstood that we are dealing with a set of N particles, and moreover we often
suppress indices. Then the force F
i
(x, x) is a 3N component function of the
3N variables x
i
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 specied set of integer values. If
i 1, 2, 3 eq. (1.1) stands for three separate equations. Throughout we em-
ploy 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.
x
i
y
i

3
i=1
x
i
y
i
=
3
i=1
x
j
y
j
= x
j
y
j
. (1.2)
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 choice, but 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 dierential equations. A philosopher might object that
they say nothing at allthe rst 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 dierential 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 New-
tons 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 eld theories, including Einsteins gen-
eral relativity theory). In the version we will study it is more special because
only a restricted set of forces are allowed in eq. (1.1). Let us see what kind
of restrictions on the function F
i
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:
dP
i
dt

d
dt
particles
m x
i
= 0 , (1.3)
where the sum is over all the particles in the system. The existence of such a
conserved vector is clearly an interesting factand by the way this formulation
1.1 Analytical and Hamiltonian mechanics 3
allows us to deal with time dependent masses, say with rockets. Now consider
the function
E = T +V =
m x
2
2
+V (x) , (1.4)
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 = x
i
(m x
i
+
i
V (x)) . (1.5)
It follows that if the force is given by
F
i
(x, x) = F
i
(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 x
i
=
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 dierential equations de-
scribing a given system, strategies for solving them, and ways of describing
the solutions if they cannot be obtained in explicit form.
1
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 nd 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
Dierential 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 rst realized
half a century after the publication of Newtons Principia. The inspiration
came from optics, and the laws of reection and refraction. It is known that
the angle of reection is equal to the angle of incidence, and it was observed
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 reected against the surface. If the angle of reection were to dier 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
1
As a denition, this is a little vague. Mechanique Analitique was the title of a book by Lagrange.
medium can be assigned an index of refraction n, and the angle of refraction
is related to the angle of incidence through
n
1
sin
i
= n
2
sin
r
. (1.8)
Fermat noted that if
n =
c
v
, (1.9)
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 unies the laws of refraction and reection, since
it also implies the equality between the angles of incidence and reection.
Is there a similar principle underlying mechanics? At least for systems obey-
ing eq. (1.7), there is. Consider two points A and B, and suppose that a particle
starts out at A at time t = t
1
, 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 = t
2
. In mathematical terms we are dealing
with a function x(t) such that
x(t
1
) = x
A
x(t
2
) = x
B
, (1.10)
but otherwise arbitrary. For any such function x(t) we can evaluate the integral
S[x(t)] =
_
t2
t1
dt (T V ) =
_
t2
t1
dt
_
m x
2
2
V (x)
_
. (1.11)
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 x
A
, x
B
, t
1
, and t
2
, but this is rarely written out explicitly.
The statement, to be veried in the next section, is that the action func-
tional (1.11) has an extremum (not necessarily a minimum) for precisely that
function x(t) which obeys the dierential 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, and forms only a part of analytical
mechanics.
Note once again what is going on. The original task of mechanics was to
1.2 The calculus of variations 5
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 dierent times. So there seems to be a local, causal way of looking
at things, and an at rst sight quite dierent 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 t
1
and t
2
approach each other innitesimally closely, then what we are in eect
specifying is the position and the velocity at time t
1
, 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 diers to
order from the extremal one, then the value of the path dependent quantity
suers a change which is of order squared. At an extremum the rst 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 rst path. If the wavelength
is very small, phases from light arriving by dierent 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 conguration space. But that is another story.
1.2 The calculus of variations
Let us now verify the claim made in the rst section, namely that Newtons
dierential equations, for suitable choices of the dynamical system, are mathe-
matically equivalent to the requirement that a certain functional of all possible
paths of the particles should assume an extremum value at the actual trajec-
tory. First we stare at the denition of the action functional:
S[x(t)] =
_
t2
t1
dt
_
m x
2
2
V (x)
_
. (1.12)
How do we nd 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 nd the extrema. We consider how the function values f(x)
change as we change the number x:
f(x) f(x +x) f(x) = x
x
f(x) . (1.13)
If the derivative is zero at the point x, the function has a minimum, or a max-
imum, or at least an inection 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) =
i
x
i
f
x
i
(x
1
, . . . , x
N
) = 0 (1.14)
for arbitrary choices of the x
i
, 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 innite number of variables, say of the
Fourier coecients of the original function. You can also regard t as a label of
the innite 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.
(1.14) with an integral. Like this:
S = S[x(t) +x(t)] S[x(t)] =
_
t2
t1
dt x(t)
S
x
(t) . (1.15)
We assume that it is possible to bring S to this form. Then the functional
derivative of S[x] will be dened as the very expression that occurs to the
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 dierentiable. First of all, note that we
are all the time evaluating the action between denite integration limits. Then
the extremum, if it exists, will be given by that particular trajectory which
starts at the point x(t
1
) at time t
1
, and ends at the point x(t
2
) at time t
2
,
and for which the functional derivative vanishes. We can make this work for
the action functional (1.12). 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
dierent function
x(t) = x(t) +x(t) , x(t) = f(t) , (1.16)
where f(t) is, for the time being, an arbitrary function while is an innites-
imally small constant. It is important for the following argument that f(t) is
arbitrary, or nearly so. That is innitesimally small simply means that we
1.2 The calculus of variations 7
will neglect terms of quadratic and higher orders in in the calculation which
follows:
S = S[ x(t)] S[x(t)] =
_
t2
t1
dt
_
m
2
( x + x)
2
V (x +x)
_
S[x(t)]
=
_
t2
t1
dt
_
m
2
x
2
+m x x V (x) x
x
V (x) +o(
2
)
_
S[x(t)] = (1.17)
=
_
t2
t1
dt (m x x x
x
V (x)) +o(
2
) .
The action functional has an extremum at the particular function x(t) for
which this expression vanishes to rst 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
t1
dt
_
x(m x +
x
V (x)) +
d
dt
(mx x)
_
. (1.18)
Unfortunately this is not quite of the form (1.15), 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 denition of x, so that
x(t
1
) = x(t
2
) = 0 . (1.19)
This is a way of saying that we are interested only in functions x(t) that
have certain preassigned starting and end points at specied times. With this
restriction, the total derivative in eq. (1.18) goes away. The rst term has
to vanish for all allowed choices of the functions x(t). After a moments
reection, 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(t
1
) = x
A
x(t
2
) = x
B
, (1.20)
for those and only those functions which obey
m x =
x
V (x) (1.21)
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. (1.20), make perfect sense. To obtain a denite
trajectory it is not enough to impose the equations of motion. It is also nec-
essary to set some initial values. For dierential 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 dierent times, which
is the same amount of information. Note, however, 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(t
1
) and x(t
2
).
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 eect 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 dierential equation is soluble?
Consider, as an exercise, a rst order dierential equation for a single variable:
x = f(x) , (1.22)
where f is some function. This can be solved by means of separation of vari-
ables:
dt =
dx
f(x)
t(x) =
_
x
c
dx
f(x
)
, (1.23)
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. (1.23) as an implicit denition
of x(t), and eq. (1.22) 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 = kx . (1.24)
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:
ml
2
= gml sin . (1.25)

Let us therefore approach eq. (1.24) in a systematic fashion, which might
1.3 How to solve equations 9
yield results also for the pendulum. As a rst step, note that any second order
dierential equation can be rewritten as a pair of coupled rst order equations:
p = kx m x = p . (1.26)
The second equation denes the new variable p. Unfortunately coupled rst
order equations are dicult 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 dened to be one
half times the number of rst order dierential 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 rst 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, in general, be described
by a set of 2n coupled rst order equations, and the diculties 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 rst order equation. For the harmonic oscillator the
conserved quantity is
E =
m x
2
2
+
kx
2
2
. (1.27)
The number E does not depend on t. Equivalently
x
2
=
2E
m

k
m
x
2
. (1.28)
Taking a square root we are back to the situation we know, and we proceed
as before:
dt = dx
_
m
2E kx
2
t(x) =
_
x
c
dx
_
m
2E kx
2
. (1.29)
Inverting the function dened by the integral, we nd the solution x(t). The
answer is a trigonometric function, with two arbitrary constants E and c, which
determine the phase and the amplitude. For our purposes the trigonometric
function is dened by this procedure!
We can play the same trick with the non-linear equation for the pendulum,
and we end up with
t() =
_

c
d
_
2
ml
2
(E +gml cos
)
. (1.30)
We integrate, and we invert. This denes 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

2
k sin
=
2k cos
_
1 k
2
sin
2
. (1.31)
This converts the integral to
t() =
l
2g
_
()
c
2k cos
_
1 k
2
sin
2
_
E
gml
+ 1 2 sin
2
2
. (1.32)
Now we choose the so far undetermined constant k by
2k
2
E
gml
+ 1 . (1.33)
The integrand then simplies, and one further substitution takes us to our
desired standard form;
t() =
l
g
_
()
c
d
_
1 k
2
sin
2
=
_
sin
_
=
l
g
_
x()
c
dx
_
(1 x
2
)(1 k
2
x
2
)
. (1.34)
Just as eq. (1.29) can be taken as an implicit denition of a trigonometric
function, this integral implicitly denes the function (t) as an elliptic function.
If you compare it with the previous integral (1.29), 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
studied in depth by mathematicians.
Anyway, the above examples were some of the simplest examples of com-
pletely 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 rst order equations: One must nd 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: a typical Hamiltonian system will exhibit
an amount of chaotic behaviour. As a result the notion of what it means
to solve a set of dierential equations evolved somewhat: a solution might
consist, say, of a convergent power series in t. But frequently this strategy also
fails, and there may not exist any eective procedure to generate the long term
behaviour of the solutions on a computer. What one has to do then is to nd
out which questions one can reasonably ask concerning such systems.
1.4 Phase space 11
Figure 1.1. A Lissajous gure, and how to draw it; x = cos t, y = sin 2t.
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
) . (1.35)
The trajectory in the x-y-plane is a Lissajous gure. 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
m
1
= n
2
(1.36)
the trajectory is a closed curve. If there are no such integers the trajectory
eventually lls a rectangle densely, and never closes on itself. Put yourself into
the position of an experimentalist trying to determine by means of measure-
ments whether the trajectory will be closed or not!
1.4 Phase space
It is worthwhile formalizing things a bit further. With the understanding that
every set of ordinary dierential equations can be written in rst order form,
we write down the general form of N coupled rst order equations for N real
variables z
i
:
z
i
= f
i
(z
1
, . . . , z
N
; t) , 1 i N , (1.37)
where the N functions f
i
are arbitrary. We simplify things by assuming that
there is no explicit dependence on time,
z
i
= f
i
(z
1
, . . . , z
N
) . (1.38)
There are theorems that guarantee the existence and uniqueness of such sys-
tems for some range of the parameter t. Thus
z
i
= z
i
(z
01
, . . . , z
0N
, t) , (1.39)
where z
0i
are the initial values of z
i
. (There is no guarantee that such solutions
can be obtained in any explicit form.) 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 z
i
. 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 correspon-
dence with the points of an N dimensional phase space. The time development
of a system is uniquely determined by its position in phase space.
This is the rst 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. (1.38). Because of the theorems
I alluded to, we know that through any point z
0
there passes a unique curve
z(t), with a unique tangent vector z. These curves never cross each other.
When the system is at a denite point in phase space, it knows where it is
going. The curves are called trajectories, and their tangent vectors dene a
vector eld on phase space called the phase space ow. Imagine that we can
see such a ow. Then there are some interesting things to be observed. We
say that the ow has a xed point wherever the tangent vectors vanish. If the
system starts out at a xed point at t = 0, it stays there forever. There is an
important distinction to be made between stable and unstable xed points. If
you start out a system close to an unstable xed point it starts to move away
from it, while in the stable case it will stay close forever. The stable xed point
may be an attractor, in which case a system that starts out close to the xed
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 rst order system
1.4 Phase space 13
Figure 1.2. A one dimensional phase space, containing one stable and one un-
stable xed point, as well as one xed point which is structurally unstable.
z = f(z) . (1.40)
For generic choices of the function f all xed points are either stable attractors,
or unstable repellors, but for special choices of f we can have xed points
that are approached by the ow 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 xed 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 R
2
, and it contains one elliptic
xed 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 one of the coordinates is a periodic angle phase space is the
surface of an innitely long cylinder. It contains two xed points. One of them
is elliptic, and the otherthe state where the pendulum is pointing upwards
is hyperbolic. What is special about the hyperbolic xed point is that there
are two trajectories leading into it, and two leading out of it. The length of the
tangent vectors

decrease as the xed point is approached. Taking the global
structure of phase space into account we see that a trajectory leaving the xed
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 dierent behaviour. One region where
the trajectories go around the elliptic xed 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 denition an invariant set in phase space is a region
that one cannot leave by following the phase space ow.
Figure 1.3. Fixed points in a two dimensional phase space: a source, a sink, a
limit cycle, an elliptic xed point, and a hyperbolic xed point.
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 xed 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
8. Speaking of Hamiltonian systems it is worthwhile to point out that the
example of the two harmonic oscillators in eq. (1.35) is less frivolous than it
may appear. The phase space is four dimensional, but there are two conserved
quantities
2E
1
= p
2
1
+
2
1
x
2
1
2E
2
= p
2
2
+
2
2
x
2
2
. (1.41)
This means that any given trajectory will be conned to a two dimensional
surface in phase space, labelled by E
1
and E
2
. This surface is a torus, with
topology S
1
S
1
. In a sense to be made clear 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 xed points of the
phase space ow. Then we try to determine the nature of these xed points. If
the equations are linear this is straightforward. If not, we can try linearisation
of the equations around the xed points. There is a theorem that we can lean
on here:
The Hartman-Grobman theorem: The nature of the xed points is unchanged
by linearisation, as long as the xed points are isolated and as long as no
elliptic xed points occur.
Thus, consider the pendulum. You know that its phase space is a cylinder
described by the coordinates (, p
). To see if the phase space ow has any

xed points, you set
=
1
ml
2
p
= 0 p
= gml sin = 0 . (1.42)

Hence there are xed points at (, p
) = (0, 0) and (, 0). Linearizing around

1.4 Phase space 15
them you nd 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 ow. No integration is needed.
Were we justied in assuming that the xed points are elliptic? To see what
can go wrong, consider the non-linear equation
x +x
2
x +x = 0 . (1.43)
In the linearised case ( = 0) there is a single elliptic xed point. In the
non-linear system the ow will actually spiral in or out from the xed 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 dimen-
sional, but if the dimension of phase space exceeds two things can get very
complicated indeed. A famous example is the at rst sight innocent looking
Lorenz equations
z
1
= az
1
+az
2
z
2
= bz
1
z
2
z
1
z
3
z
3
= cz
3
+z
1
z
2
.
(1.44)
They capture some aspects of thermal convection in a uid. 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 eect, and the slightest change in the
initial data will cause the trajectory to go to completely dierent 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 specied; 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 specied, and that the equation of motion determines
all the derivatives, or
b) position, velocity and acceleration can be freely specied, 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 m
1
= 1 kg and a particle of mass m
2
= 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 Prove Snells Law of optics, starting from Fermats principle.
Problem 1.4 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 =
_
L
0
dx
_
k
2
y
(x)y
_
,
where the slash denotes dierentiation with respect to x and k is a constant. The
bar will minimise 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.
Problem 1.5 Is it true that a once dierentiable function x(t) is a solution
of eq. (1.24) if and only if it is a solution of eq. (1.28)? If not, nd a non-trivial
counterexample.
Problem 1.6 Using the general solution for the pendulum, eq. (1.34), solve
for (t) in the special case that k = 1. Physically, what does this solution correspond
to?
Problem 1.7 Consider a projectile that is red straight up in a gravitational
eld (V = GM/r), reaches a maximum height r
max
, and falls back again. Prove
that the solution has the parametric form
r =
r
max
2
(1 cos ) , t =
r
max
2
_
r
max
2GM
( sin ) .
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 withot slipping on the taxis.
Problem 1.8 Linearize the pendulum around its xed points, and then drawa
careful picture of its phase space. Add a friction term to the equation,

+

+sin = 0,
and see in qualitative terms what this does to the phase space ow.
Problem 1.9 Give a simple example where linearisation around a xed point
gives an erroneous impression of its nature because the xed point does not stay
isolated.
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 q
i
and their n rst order derivatives q
i
:
S[q(t)] =
_
t2
t1
dt L(q
i
, q
i
) . (2.1)
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 interpreta-
tion of the variables may dier from the interpretation of x as the position
of some particle. The space on which q
i
are the coordinates is called the con-
guration space, while q
i
are its the tangent vectors. 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 conguration 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
one needs several coordinate systems to cover the whole conguration space.
In this case one sees clearly that 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 denitely are!
2.1 The scope of Lagrangian mechanics
Among all those functions q
i
(t) for which q
i
(t
1
) and q
i
(t
2
) are equal to some ar-
bitrarily prescribed values, the action functional has an extremum for precisely
those functions q
i
(t) which obey the Euler-Lagrange equations
L
q
i
d
dt
_
L
q
i
_
= 0 , (2.2)
provided such functions exist. This is straightforward to verify by means of
the calculus of variations; indeed (suppressing indices)
S =
_
t2
t1
dt
_
q
L
q
+ q
L
q
_
=
_
t2
t1
dtq
_
L
q

d
dt
L
q
_
+
d
dt
_
q
L
q
_
.
(2.3)
The total derivative term gives rise to a boundary term that vanishes because
we are only varying the functions whose values at t
1
and t
2
are kept xed, so
that q is zero at the boundary. The Euler-Lagrange equations follow as ad-
vertized. 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 this form by any means, includ-
ing some of considerable physical interest; most of them involve dissipation of
energy of some sort. An example is that of white elephant sliding down a hill-
side covered with owers. However, in some sense the frictional force involved
here is not a fundamental force. A complete description of the motion of the
elephant would involve the motion of the atoms in the elephant and in the
owers, both being heated by friction. It is believed that all complete, fun-
damental 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 when-
ever there is no dissipation of energy. For many simple mechanical systems the
Lagrangian equals the dierence between the kinetic and the potential energy,
L(x, x) = T( x) V (x) . (2.4)
For instance,
L =
m x
2
2
V (x) 0 =
L
x

d
dt
L
x
=
V
x
m x . (2.5)
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
x = r cos sin y = r sin sin z = r cos . (2.6)
That is to say, we wish to derive the equations for r,

, and

. This requires a
certain amount of calculation. The amount shrinks if we perform the change
of variables directly in the Lagrangian:
L =
m
2
_
x
2
+ y
2
+ z
2
_
=
m
2
_
r
2
+r
2

2
+r
2
sin
2
2
_
. (2.7)
2.1 The scope of Lagrangian mechanics 19
Then we obtain the answer as the Euler-Lagrange equations from this La-
grangian. (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 rst has to spend some time in deriving it.
A famous example for which L ,= T V is that of an electrically charged
particle moving in an external electromagnetic eld. This example is so im-
portant that we will give it in some detail. First of all, external signies that
we are dealing with an approximation, in which we ignore that the presence of
the electrically charged particle will aect the electromagnetic eld in which it
moves. In many situations, this is an excellent approximation. The equations
of motion to be derived are the Lorentz equations
m x
i
= e (E
i
(x, t) +
ijk
x
j
B
k
(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 eld. 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:
E
i
(x, t) =
i
(x, t)
t
A
i
(x, t) B
i
(x, t) =
ijk
j
A
k
(x, t) . (2.9)
Here is known as the scalar potential and A
i
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
_
m x
2
2
+e x
i
A
i
(x, t) e(x, t)
_
. (2.10)
Please verify this!
If we consider a time independent electric eld with no magnetic eld
present, the Lorentz equation reduces to the more familiar form
m x
i
= e
i
(x) . (2.11)
This has the same form as Newtons Law of Gravity, if the potential is specied
correctly. The reason why the full Lorentz equation is much more complicated
has to do with the special relativity theory. The magnetic eld is a relativistic
eect. The relativistic version of Newtons law of gravity is yet more compli-
cated, and is given by Einsteins general relativity theory.
An important dierence between gravity and electricity, also in the non-
relativistic 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.
2.2 Constrained systems
There is a delightful trick, due to Lagrange, which illustrates the suppleness of
variational calculus very nicely. The type of problem to be considered is this: A
particle is moving in space but, one way or another, it is constrained to move on
a two dimensional surface. The Lagrangian formalism is well suited to derive
equations of motion consistent with this requirement. For deniteness, let the
surface be a sphere dened by the equation
x
2
1
+x
2
2
+x
2
3
= 1 . (2.12)
The rst idea that springs to mind is that one should solve this for
x
3
= x
3
(x
1
, x
2
) =
_
1 x
2
1
x
2
2
, (2.13)
and insert the result back into the action that describes the free particle, ie.
S[x
1
, x
2
] =
_
dt
m
2
_
x
2
1
+ x
2
2
+ x
3
(x
1
, x
2
)
2
_
. (2.14)
This can be done, and will indeed result in a set of Euler-Lagrange equations
describing a particle conned to the surface of a sphere, but there are draw-
backs. From eq. (2.14) it appears at rst sight as if the conguration space
is the unit disk in the plane, since x
1
and x
2
are not allowed to take values
outside this disk. At second sight it appears that the conguration space is
two copies of the unit disk, since there are two branches of the square root.
But the true congurations space is a sphere. What we see is a reection of
the known fact that it is impossible to cover a sphere with a single coordi-
nate systemour equations have only a local validity. This is in a way an
unavoidable problem, but the procedure contains some arbitrariness, and it is
rather clumsy.
A better way is to transform to polar coordinates, after which the constraint
will inform us to keep the value of r xed. But this procedure has some of the
same drawbacks, and moreover such a natural coordinate system exists only
in very special cases (such as the sphere), while we are heading for the general
problem: Consider a Lagrangian L
0
dened on an n dimensional conguration
space, with coordinates q
1
, . . . , q
n
, and suppose that the system is conned to
live in the (n m) dimensional submanifold dened by the m conditions
I
(q
1
, . . . , q
n
) = 0 , 1 I m . (2.15)
Derive equations of motion consistent with this requirement. One way to do
this is to solve for m of the qs, q
1
, . . . , q
m
say, by means of the m conditions
2.2 Constrained systems 21
(2.15), and insert the result in the action. In general this will be a lot of hard
work, and the diculties we had with coordinatizing the sphere will recur with
a vengeance.
The method of Lagrange multipliers is more convenient. It does not avoid the
diculties with coordinatizing, but postpones them to a later stage. The claim
that we will verify is that the following action is equivalent to the previous
one:
S[q, ] =
_
dt L
0
(q, q) +
1
1
(q) + +
m
m
(q) . (2.16)
The s are the Lagrange multipliers, to be treated as new dynamical variables.
When the action (2.16) is varied with respect to the s we obtain the con-
straints (2.15) as equations of motion. When we vary with respect to the qs the
resulting equations will contain the otherwise undertermined Lagrange multi-
pliers, 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 nd 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 qs
it 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 = x
x
f +y
y
f , (2.17)
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 = = x
x
+y
y
. (2.18)
This equation has to be solved for y and the result inserted into eq. (2.17),
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 rewrit-
ing eq. (2.17) in the form
0 = f = f + = x(
x
f +
x
) +y(
y
f +
y
) , (2.19)
where the s are arbitrary functions. The ys are still given in terms of the xs,
so it would seem at rst 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 nding the extremamaking variations consistent with the constraintsis
equivalent to solving the n +m equations
(q) = 0
q
f +
q
= 0 (2.20)
for q and . But these are precisely the equations that we obtain from the La-
grange 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 straight-
forward to generalize the argument from functions to functionals. From the
action
S[q, ] = S
0
[q] +
_
dt (q) (2.21)
we rederive the constraints, together with the equations of motion
S
q
=
S
0
q
+
q
= 0 . (2.22)
This is the analogue of the second equation (2.20). Written out, if L = L(q, q)
and if there is only one constraint, this is
d
dt
L
q
i
=
L
q
i
+
q
i
(2.23)
(q) = 0 . (2.24)
These equations have a simple interpretation. The constraint denes a surface
in conguration space. We have modied 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 conned
to the surface we must choose the strength of the force (given by ) in such a
way that this is ensured.
Let us do a simple example in detail. We choose a free particle in three
dimensions, with the unconstrained Lagrangian
L =
m
2
(

X
2
+

Y
2

Z
2
) . (2.25)
The minus sign here is unusual, indeed in its unconstrained form this La-
grangian is pathological. (Why?) We constrain the particle to the hyperboloid
X
2
+Y
2
Z
2
= 1 , Z > 0 . (2.26)
There are two physical degrees of freedom. One way, among many, to see
this is to observe that any point on the hyperboloid can be described by the
coordinates r and , where
2.2 Constrained systems 23
X = sinhr cos Y = sinhr sin Z = cosh r . (2.27)
You can convince yourself that there is a one-to-one correspondence between
the pairs (r, ) and the points (X, Y, Z) on the hyperboloid. Actually this is
an exceptional situation: it is often impossible to nd a set of coordinates that
cover the entire constraint surface. If we use eqs. (2.27) in the Lagrangian we
obtain
L =
m
2
( r
2
+ sinh
2
r

2
) . (2.28)
The Lagrangian is equal to the kinetic energy T of the particle. It can be
checked that it is a conserved quantity. Here we see an advantage with using
the physical degrees of freedom only, because it is now evident that the kinetic
energy is bounded from below, a physically important property that was not
at all obvious in the original Lagrangian.
The advantages of the Lagrange multiplier method make themselves felt
when we look for further constants of the motion. From
L =
m
2
(

X
2
+

Y
2

Z
2
) + (X
2
+Y
2
Z
2
+ 1) (2.29)
we get the equations of motion
m

X = 2X m
Y = 2Y m
Z = 2Z . (2.30)
It is then evident that we have three constants of the motion
J
X
= Z

Y Y

Z J
Y
= Z

X X

Z J
Z
= X

Y Y

X . (2.31)
Using eqs. (2.27) we can express these constants of the motion as
J
X
= r sin +

cosh r sinhr cos J
Y
= r cos

cosh r sinh r sin
(2.32)
J
Z
=

sinh
2
r .
It is possible to check directly, using the equations of motion for r and , that
these are constants of the motionbut only J
Z
is obviously conserved. The
coordinate system (r, ) somehow hides the others. By the way, the kinetic
energy can be expressed as
T =
m
2
_
J
2
X
+J
2
Y
J
2
Z
_
. (2.33)
This example gives a little bit of the avour of constrained systems. It is not
taken quite out of the blue, since it describes motion of a particle on a surface
with constant negative curvature, as opposed to a sphere which has constant
positive curvature. But this is by the way.
The Lagrange multiplier method works only for holonomic constraints, that
is constraints that involve the conguration space variables only. But consider
a ball moving without friction across a plane. The conguration space has
ve 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 conguration 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 conguration space are called
anholonomic; we do not deal with them 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 rst an arbitrary variation of the action. According to eq. (2.3)
the result is
S =
_
t2
t1
dt q
_
L
q

d
dt
L
q
_
+
_
q
L
q
_
t2
t1
. (2.34)
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 dierent.
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 non-
vanishing term is the boundary term, and
S = (Q(t
2
) Q(t
1
)) , Q(t) q
L
q
. (2.35)
Here and in the following is the constant occuring in q = f, where f is an
2.3 Symmetries 25
arbitrary function of t. The point is to ensure that there is nothing innitesimal
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 specied
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(t
2
) Q(t
1
)) . (2.38)
The times t
1
and t
2
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
t1
dt
d
dt
(q, q) . (2.39)
Here can be any functionthe important and unusual thing is that the
integrand is a total time derivative. Then the quantity Q, dened by
Q(q, q) = q
i
L
q
i
(q, q) (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
suce. Consider a free particle described by
L =
m
2
x
i
x
i
. (2.41)
Since only x appears in the Lagrangian, we can choose
x
i
=
i
, (2.42)
where
i
is independent of time. Then the variation of the action is auto-
matically zero, Noethers theorem applies, and we obtain a vectors worth of
conserved charges
P
i
=
L
x
i
= m 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
x
i
=
ijk
j
x
k
S = 0 . (2.44)
Here
i
is again independent of t, and
ijk
is the totally anti-symmetric ep-
silon tensor. Noethers theorem now implies the existence of another conserved
vector, namely
L
i
=
ijk
x
j
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
x
i
= x
i
S =
_
dt
d
dt
_
m
2
x
2
_
. (2.46)
Using eq. (2.40) we obtain the constant of the motion
E =
m
2
x
i
x
i
. (2.47)
This is the conserved energy of the particle. There is yet another conserved
quantity that diers from the others in being an explicit function of timebut
2.3 Symmetries 27
its total time derivative vanishes since it also depends on the time dependent
dynamical variables. Thus
x
i
=
i
t S =
_
dt
d
dt
(m
i
x
i
) . (2.48)
Eq. (2.40) gives the conserved charge
Q
i
= mx
i
tm x
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 unaected? 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 denition:
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 (which is also true), 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.
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. There-
fore momentum conservation is a consequence of translation invariance. It is
immediate that we can iterate the innitesimal translations used in Noethers
theorem to obtain nite 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 denite, given that (vt, 0, 0) is a solu-
tion 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 New-
tons 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
L =
m
1
2
x
1
2
+
m
2
2
x
2
2
V (x
1
, x
2
) . (2.50)
Indeed invariance under (2.42) requires that V (x
1
, x
2
) = V (x
1
x
2
), which is
a strong restriction.
Eq. (2.44) expresses the fact that the Lagrangian has rotation symmetry,
while eq. (2.46) is an innitesimal translation in time: Given a solution x(t),
the function
x
(t) = x(t +t
0
) = x(t) +t
0
x(t) +o(t
2
0
) (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 reections, that do not arise
by iterating an innitesimal symmetry, are counterexamples, and we will see
another in section 4.5.
To sum up, symmetries are important from two quite dierent 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 discover-
ing 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?
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).
2.3 Symmetries 29
Problem 2.4 Consider the Lagrangian
L =
1
2
M
ij
(q) q
i
q
j
,
where the matrix elements of M
ij
depend on the conguration space coordinates,
and the matrix is assumed to have an inverse M
1
ij
. Write down the Euler-Lagrange
equations and solve for the accelerations.
Problem 2.5 Take N positive numbers summing to one, p
1
+p
2
+ +p
N
= 1.
Their geometric mean is dened as (p
1
p
2
p
N
)
1/N
. What is the maximum of the
geometric mean?
Problem 2.6 Consider the Lagrangian
L =
1
2
q
2
q
n
,
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
q =
_
t q
q
2
_
.
This is known as conformal symmetry.
Problem 2.7 Write down the Lagrangian for a double pendulum. (The rod
of the second is attached to the bob of the rst. Bobs are heavy, rods not.) How many
constants of the motion can you nd?
Problem 2.8 For a free particle, consider the action integral
S =
_
t2
t1
m
2
x
2
dt .
Evaluate this integral for an x(t) that solves the equations of motion, and express the
answer as a function of t
1
, t
2
, and the initial and nal positions x
1
and x
2
. 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 denition 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 rst 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
x
2
+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, let us say 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 denes the spheres uniquely. Denote the points by F
1
and F
2
, and the
circles by C
1
and C
2
. Now consider a point P in the intersection of the cone
Interlude: Conic sections 31
Figure 3.1. A vertical cross section through Apollonius proofbut to under-
stand the proof you have to think in three dimensions.
and the plane. The generator passing through P intersects the circles C
1
and
C
2
in the points Q
1
and Q
2
. Now the trick is to prove that the distance PF
1
equals the distance PQ
1
, and similarly the distance PF
2
equals the distance
PQ
2
. 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 PF
1
and PF
2
is constant and equal to the length of the segment of
the generator between the circles C
1
and C
2
, independently of which point P
on the intersection we choose. This property denes the ellipse. This is the
proof that the intersection between the cone and the plane is an ellipse with
its foci at F
1
and F
2
. 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.
32 Interlude: Conic sections
For our purposes it is convenient to dene the ellipse somewhat dierently.
An ellipse of eccentricity e < 1 can be dened 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 innity.
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
e
x
_
. (3.3)
(To see this, note that the distance from the focus to the directrix is p/e.)
Otherwise expressed
r = p er cos
p
r
= 1 +e cos , (3.4)
where = 0 gives the point closest to the directrix. For a general point on the
ellipse we nd
x
2
+y
2
= r
2
= (p ex)
2
(x +ea)
2
a
2
+
y
2
b
2
= 1 , (3.5)
where
a
p
1 e
2
b
2
= pa = (1 e
2
)a
2
. (3.6)
The major axis of the ellipse has length 2a, and the minor axis has length 2b.
Finally the distance between the center and the focus equals ea. To see this
we set = 0 and = in eq. (3.4), and calculate
r() r(0)
2
=
p
2
_
1
1 e

1
1 +e
_
= ea . (3.7)
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 nd it interesting to relate
an ellipse centered at a focus to an ellipse centered at the origin. The latter is
described in Cartesian coordinates by the complex trajectory
w(t) = a cos t +ib sin t . (3.8)
The parameter t must not be confused with the the angle between the radius
vector and the x-axis. Surprisingly, if we square this ellipse we obtain an ellipse
with its focus at the origin:
Interlude: Conic sections 33
Z(t) = w
2
=
a
2
+b
2
2
cos 2t +iab sin 2t +
a
2
b
2
2

(3.9)
Acos 2t +iBsin 2t +EA ,
where
EA =
A
2
B
2
, (3.10)
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 center. This
is again an ellipse, but centered at one of its foci and traversed twice as the
original ellipse is traversed once.
This trick was introduced by Karl Bohlin, working at the Pulkovo observa-
tory 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.
Problem 3.1 Using the denition 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 dierent 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
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 centered 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 as-
sumption 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
suces 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 35
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 conguration space of our problem has six
dimensions, spanned by the positions of the sun (X) and one planet (x
P
), and
we try the Lagrangian
L =
M
2
X
i

X
i
+
m
p
2
x
Pi
x
Pi
V (X, x
P
) . (4.1)
Depending on the form of the function V we may have to exclude the points
X = x
P
from the conguration spacewe insist that the function V takes
nite values only as a function on conguration space. Anyway this will give six
coupled second order dierential 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 nd
V (X, x
P
) = V (X x
P
) = V ([X x
P
[) . (4.2)
This works for any function V of one variable. Next we introduce coordinates
x
i
= x
Pi
X
i
which are invariant under translations, together with coordinates
describing the center of mass. Then the center 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:
L =
m
2
x
2
i
V ([x[) . (4.3)
Here m is the reduced mass, almost equal to the mass of the planet since the sun
is very heavy in comparison. The coordinate x
i
vanishes at the center of mass
of the system, which is well inside the sun, and can be approximately identied
with the center 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
i
= m
ijk
x
j
x
k
. (4.4)
This vector is orthogonal to both x and x, which means that the motion
is conned to a plane orthogonal to the angular momentum vector. We are
down to a two dimensional conguration 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 simplies to
L =
m
2
( r
2
+r
2

2
) V (r) . (4.5)
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 nd
the constant of the motion
l =
L
= mr
2

. (4.6)
This equation is of considerable interest in itself. It says that
A =
r
2

2
=
l
2m
= constant , (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
E =
m
2
_
r
2
+r
2

2
_
+V (r) =
m r
2
2
+
l
2
2mr
2
+V (r) = constant . (4.8)
This gives the formal solution
dt =
dr
_
2
m
_
E
l
2
2mr
2
V (r)
_
. (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 nd (t) we combine eqs.
(4.6) and (4.9), and get
d =
ldt
mr
2
=
ldr
r
2
_
2m
_
E
l
2
2mr
2
V (r)
_
. (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 37
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 eort, 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 eective potential
V
eff
(r) =
l
2
2mr
2
+V (r) . (4.11)
A simple case is V (r) = 0, i.e. no force at all. In the eective one dimensional
problem this corresponds to a repulsive V
eff
the particle comes in from in-
nity, reaches a minimum value of r, and then disappears to innity 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 eective potentialif it does have a
bottom. The radius r of the circular orbit must obey
V

eff
(r) = 0 . (4.12)
If this happens at a local maximum of V
eff
the solution is unstable, and
unlikely to be realized in Nature. The orbit is stable under small perturbations
if and only if
V

eff
> 0 (4.13)
at the value of r for which V
eff
= 0.
We look into these equations for the special choice
V (r) = kr
V
eff
(r) =
l
2
2mr
2
kr
, (4.14)
with arbitrary. The radius of the circular orbit is found to be
r =
_
l
2
km
_ 1
+2
. (4.15)
This makes sense only if l
2
,= 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 o 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
a full revolution in , to the time it takes to complete a full oscillation in r. If
the orbit is an ellipse centered 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 gures 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
ldr
r
2
_
2m(E V
eff
(r))
. (4.17)
For the planets, Keplers rst law requires that = 2.
A nal comment: if we do choose = 1 we have the problem that the
energy is unbounded from below. We now see that this problem cannot be too
serious, because
V
eff
(r) =
l
2
2mr
2

k
r
(4.18)
is in fact bounded from below whenever l ,= 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 conguration space.
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:
L =
m
2
( r
2
+r
2

2
) kr
2
=
m
2
( x
2
+ y
2
) k(x
2
+y
2
) . (4.19)
This is two harmonic oscillators of equal frequencies, and the corresponding
Lissajous gures are indeed ellipses. They are not the right kind of ellipses
however, since they are centered at the origin. Let us call them Hooke el-
lipses. They can be related to Kepler ellipses (centered at a focus) by means
of Bohlins trick (chapter 3). If
w = [w[e
i
= a cos t +ib sin t (4.20)
is a Hooke ellipse, then
Z = w
2
= [w[
2
e
2i
=
a
2
b
2
2
+
a
2
+b
2
2
cos 2t +iab sin 2t . (4.21)
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
4.3 Keplers First Law 39
w = w [ w[
2
+[w[
2
= 2 . (4.22)
Moreover Keplers second law holds, so that
[w[
2
d
dt
= constant . (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
d
= constant . (4.24)
A suitable choice of the two constants gives the desired relation
d
dt
=
[Z[
2
[w[
2
= [w[
2
d
d
=
1
[w[
2
d
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 investi-
gate what dierential equation it obeys. But using eq. (4.22) this is a simple
calculation:
d
2
Z
d
2
=
1
[w[
2
d
dt
_
1
[w[
2
dw
2
dt
_
=
2
[w[
2
d
dt
_
w
w
_
= = 4
Z
[Z[
3
, (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. The argument is water tight 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 conrm 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 =
1
r
du =
dr
r
2
. (4.28)
The result is
d =
ldu
2mE l
2
u
2
+ 2mku
. (4.29)
This denes u as a trigonometric function of , with the energy E < 0 and
the phase
0
as undetermined integation constants. In fact
=
0
arccos
l
2
u
mk
1
_
1 +
2El
2
mk
2
. (4.30)
Inverting this, and cleaning up the answer a little, we obtain
l
2
mkr
=
l
2
u
mk
= 1 +
_
1 +
2El
2
mk
2
cos (
0
) . (4.31)
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. (3.4) we
read o that the eccentricity of the ellipse is
e =
_
1 +
2El
2
mk
2
. (4.32)
The semi-major axis of the ellipse is
a =
p
1 e
2
=
l
2
mk
mk
2
2[E[l
2
=
k
2[E[
. (4.33)
The solution remains valid also for E > 0, in which case it describes a hy-
perbola 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. Here is a simple one: since the areal velocity
is constant, the period T is simply related to the area of the ellipse. For the
special case we are looking at eq. (4.7) together with exercise 3.1 imply that
l
2m
T = A = ab . (4.34)
Remembering that b
2
= ap, and using eq. (4.31) to identify the latus rectum
p, gives
T
2
=
4
2
m
2
a
3
p
l
2
= 4
2
m
k
a
3
. (4.35)
This is Keplers Third Law. We see that it holds only to the extent that we
can regard the reduced masses m of all the planets to be the same.
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:
4.5 Self-similarity and the virial theorem 41
dt =
_
ma
k
rdr
_
l
2
2m|E|
+ 2ar r
2
=
_
ma
k
rdr
_
a
2
e
2
(r a)
2
. (4.36)
We can do the integral if we can nd a substitution that simplies the integral
_
r
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 =
_
ma
3
k
( +e sin ) . (4.39)
Taken together, eqs. (4.38-4.39) provide a parametric representation of the
orbit, and we can read o its period
T = 2
_
ma
3
k
. (4.40)
This is Keplers Third Law once again.
4.5 Self-similarity and the virial theorem
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
L =
m
2
_
dq
dt
_
2
V (q) , (4.41)
where for once we do not use the dot notation because we will soon have
two dierent 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
V (q) =
V (q) . (4.42)
There could be several variables q
i
. For simplicity I write only q. Let us also
change the time scale, and dene a new function q
by
q(t) q
(t
) = q(t) , t
=
2
2
t . (4.43)
It follows that
dq
dt
=
dt
dt
d
dt
(q(t)) =
2
dq
dt
. (4.44)
We can now check that our rescalings represent a symmetry because, under
this transformation,
L
_
q,
dq
dt
_
L
_
q
,
dq
d
_
=
L
_
q,
dq
dt
_
. (4.45)
This has the eect 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 it is a symmetry of the action. (Compare problem 2.6. If
you nd the argument dicult, 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 unaected by the scaling,

and we seewithout looking at any explicit solutionsthat the period of
the oscillations are independent of their amplitudes. Galilei rst 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
T
2
=
R
3
R
3
. (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 dened
by
f) lim
t
1
t t
0
_
t
t0
dt
f(t
) . (4.47)
For the argument to follow it is important that the time average of the deriva-
tive of a bounded function is zero, i.e.
_
df
dt
_
= lim
t
1
t
(f(t) f(t
0
)) = 0 (4.48)
4.5 Self-similarity and the virial theorem 43
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 x
i
=
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 nite. 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 10
23
atoms conned in a box. Regardless of the number of variables Eulers
theorem on homogeneous functions states that
V (x) =
V (x) x
i
i
V (x) = V (x) . (4.50)
Then
2 T) =
m x
2
_
=
_
d
dt
(mx
i
x
i
) mx
i
x
i
_
= x
i
m x
i
) =
(4.51)
= x
i
i
V (x)) = V (x)) .
This is the conclusion we were after.
For bounded motion in homogeneous potentials
2 T) = V ) , (4.52)
where is the degree of homogeneity of V (x).
For the inverse square law the virial theorem implies that
2T +V ) = 0 E) = T +V ) = T) 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 T) and V ) are equal.
The virial theorem has been used by astronomers to estimate the masses of
clusters of stars and clusters of galaxies, assuming that they are gravitationally
bound. This led to the rst evidence for dark matter.
The calculation in eq. (4.51) is of interest even for non-potential forces, if
we break it o after the rst line:
2 T) = x
i
F
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 denition of the pressure P as force per unit area,
and apply Gauss law to the result:
2 T) = P
_
dA
i
x
i
= P
_
dV
i
x
i
= 3PV . (4.55)
If we are willing to identify T) with (a factor times) the temperature T we
have Boyles Law,
PV = 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 rst step is to look for special exact solutions, which may be used
as starting points for perturbation theory, or in other ways. An interesting ex-
ample was found by Lagrange. Let us begin by assuming that the motion takes
place in a plane, and use complex numbers z
i
(t) to denote the trajectories. The
equations are
z
1
= m
2
z
1
z
2
[z
1
z
2
[
3
m
3
z
1
z
3
[z
1
z
3
[
3
z
2
= m
3
z
2
z
3
[z
2
z
3
[
3
m
1
z
2
z
1
[z
2
z
1
[
3
(4.57)
z
3
= m
1
z
3
z
1
[z
3
z
1
[
3
m
2
z
3
z
2
[z
3
z
2
[
2
.
We assume that the center of mass is at rest,
m
1
z
1
+m
2
z
2
+m
3
z
3
= 0 . (4.58)
The particles form a triangle, with sides represented by
w
1
= z
3
z
2
, w
2
= z
1
z
3
, w
3
= z
2
z
1
. (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
4.6 The three-body problem 45
w
1
= m
w
1
[w
1
[
3
+m
1
a
w
2
= m
w
2
[w
2
[
3
+m
2
a (4.60)
w
3
= m
w
3
[w
3
[
3
+m
3
a , (4.61)
where m = m
1
+m
2
+m
3
and
a =
w
1
[w
1
[
3
+
w
2
[w
2
[
3
+
w
3
[w
3
[
3
. (4.62)
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,
w
2
= e
2i/3
w
1
, w
3
= e
4i/3
w
1
. (4.63)
Then a = 0, and the only equation we need to solve is
w
1
= m
w
1
[w
1
[
3
. (4.64)
This we know how to do.
To interpret the solution, solve for
mz
1
= m
3
w
2
m
2
w
3
m
2
[z
1
[
2
= (m
2
2
+m
2
m
3
+m
2
3
)[w
1
[
2
(4.65)
and so on. A small calculation then shows that
z
1
=
(m
2
2
+m
2
m
3
+m
2
3
)
3
2
m
2
z
1
[z
1
[
3
, (4.66)
and similarly for the other two particles. Hence the particles are all being
accelerated towards their common center of mass, with eective 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.
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 L
4
and L
5
, 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
Figure 4.1. The two stable Lagrange points in Jupiters orbit.
have been found close to the Lagrange points L
4
and L
5
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 oered for a solution to the following problem: For a system of arbi-
trarily 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 t
1/3
. 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 rst 10
8000000
terms. Hence
the interest in the exact general solution dwindled from that point in.
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
eort. 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 gravi-
tational two body problem to show that the trajectory is conned to a plane. Now
consider an electrically charged particle moving in the electromagnetic eld of a mag-
netic monopole, that is to say a hypothetical particle which, if placed at the origin,
gives rise to the magnetic eld
4.6 The three-body problem 47
B
i
= b
x
i
r
3
.
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 modied version of the angular momentum, and
use the existence of this quantity to draw a qualitative conclusion about the trajectory
of the particle.
Problem 4.2 Consider the Earth-Moon system. Because of the tides some
dissipation of energy takes place. How does this aect the distance of the moon from
the earth?
Problem 4.3 Consider the Yukawa potential
V (r) = k
e
r
r
, k > 0 , > 0 .
What can you say about the existence and stability of circular orbits?
Problem 4.4 If the Sun is attened at its poles it will have a quadrupole
moment, and the potential is
V (r) =
k
r
+
q
r
3
.
The orbits will no longer be closed. Compute the angle by which the perihelion moves
during one revolution, to rst order in q.
Problem 4.5 For the Newtonian potential (4.27) the two body problem
admits an additional conserved vector
M
i
=
ijk
x
j
L
k
k
r
x
i
.
This is known as the Runge-Lenz vector. Check that it is indeed conserved. In what
direction does it point?
Problem 4.6 The mass of Jupiter is 2 10
27
kg and its distance from the Sun
is 8 10
11
m. Suppose the Sun suddenly disappears. How long would it take for the
Trojan asteroids to crash onto Jupiter?
Problem 4.7 Let
w = w[w[
a1
, Z = w
.
Choose a time parameter = (t) so that Keplers Second Law holds for Z(), and
prove that
d
2
Z
d
2
= cZ[Z[
A1
,
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 rst 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 aected in a signicant 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
x(t) = a
1
cos t +a
2
sint = Acos (t +) = Re
_
e
it
, (5.2)
where
A =
_
a
2
1
+a
2
2
, tan =
a
2
a
1
, (5.3)
and 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 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
E =
m
2
x
2
+
k
2
x
2
=
m
2

2
A
2
. (5.4)
It is manifestly independent of time. And the subject is exhausted.
To make matters more interesting we introduce an external force, and con-
sider the forced oscillator
m x +kx = F(t) x +
2
x =
1
m
F(t) . (5.5)
To solve this equation, with some specied function F(t), we appeal to the
general theory of ordinary dierential equations: it is enough to nd one par-
ticular 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
F(t) = f cos (t +) , (5.6)
which we will solve under the assumption that ,= . For a particular solution
we try the Ansatz
x
part
(t) = Bcos (t +) . (5.7)
Plugging this into the equation will determine B. Adding the general homo-
geneous solution we nd the noteworthy general solution
x(t) = Acos (t +) +
f
m(
2
2
)
cos (t +) . (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 dierent frequencies.
Close to resonance the frequencies of the two harmonics almost coincide, and
we will observe the phenomenon of beats. Recall that
cos
1
t + cos
2
t = 2 cos
(
1
2
)t
2
cos
(
1
+
2
)t
2
. (5.9)
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 diering amplitudes and phases
5.1 Forced oscillations 51
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) = e
it
+e
i(+)t
=
_
+e
it
_
e
it
. (5.10)
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
[ +e
it
[
2
= [[
2
+[[
2
+ 2[[[[ cos (t + a phase) . (5.11)
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 rst that
x +
2
x =
d
dt
( x +ix) i( x +ix) . (5.12)
Hence we dene the complex variable
z = x +ix , (5.13)
in terms of which the original equation becomes the rst order equation
z iz =
1
m
F(t) . (5.14)
Since the variable is now complex this is actually a pair of real rst 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
x(t) =
1
Im[z(t)] . (5.15)
Solving the rst 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
z(t) = e
it
_
z
0
+
1
m
_
t
F(t
)e
it
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
E() =
m
2
[z()[
2
=
1
2m
F(t)e
it
dt
2
. (5.17)
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
_
Fdt.
5.2 Damped and forced oscillations
Friction is a particular kind of coupling to the environment. In many situations
the strength of the coupling grows with velocity. Examples include particles
encountering air resistance, or more generally particles moving in a viscous
medium. Hence we try the equation
m x+ x+kx = 0 x+2 x+
2
0
x = 0 ,
2
0
=
k
m
, =

2m
. (5.18)
This is the damped oscillator.
As an aside, note that the damped oscillator is realised in electical circuit
theory. A capacitor leads to a voltage dierence V
C
= q/C, where q is the
charge and C the capacitance. Across an inductor there is a voltage dierence
V
L
= L

I, where I = q is the current and L the inductance. Coupling the
capacitor, the inductor, and a resistor with resistance R in a series we obtain
the equation
L q +R q +
1
C
q = 0 . (5.19)
Thus
0
= 1/
LC. There are numerous useful analogies between electrical,

acoustical, and mechanical systemsbut this is by the way.
To nd the solution of the damped oscillator we try the Ansatz x = e
it
.
Inserting this, and cancelling the exponential, leads to the equation
2
+ 2i +
2
0
= 0 = i
_
2
0
2
. (5.20)
Hence the general solution is
x(t) = a
1
e
t+i
2
0
2
+a
2
e
ti
2
0
2
. (5.21)
There are two qualitatively distinct cases to consider. If the damping is weak,
that is if
2
0
>
2
, the general solution is a damped oscillation
x(t) = e
t
_
a
1
e
i
eff
t
+a
2
e
i
eff
t
_
. (5.22)
Note that the eective frequency
e
is always smaller than the bare fre-
quency
0
. If, on the other hand, the damping is strong enough so that
2
0
<
2
the general solution decays exponentially without oscillations.
It is instructive to consider the critical case
2
0
=
2
. In the limit we obtain
5.2 Damped and forced oscillations 53
the solution ae
t
, but this cannot be the general solution since it contains
only one integration constant. In fact the general solution is
x(t) = (a +bt)e
t
. (5.23)
This must be the general solution because it contains two integration con-
stants.
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 + 2 x +
2
0
x =
f
m
e
it
, (5.25)
make the Ansatz
x
part
(t) = Be
it
, (5.26)
insert in the equation, cancel the exponential, and solve for B. The result is
(
2
+ 2i +
2
0
)B =
f
m
B =
1
m
f
2
0
2
+ 2i
. (5.27)
The amplitude is the absolute value of B. In fact
B = be
i
, b =
1
m
f
_
(
2
0
)
2
+ 4
2
2
, tan =
2
2
0
. (5.28)
The real solution is then
x(t) = b cos (t +) . (5.29)
Of course we should 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 eectively zero, and a steady state described by the
above solution sets in.
Because of the damping the amplitude no longer goes to innity at reso-
nance, 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
.
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 is
E =
d
dt
_
m
2
x
2
+
k
2
x
2
_
= x(m x +kx) = x
2
= 2m x
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
E = 2mb
2
2
sin
2
(t +) . (5.31)
For most purposes it will be enough to know the time average of this quantity.
Since the amplitude b depends on this time average is a function of the
frequency . Making use of a relation from (5.28) we obtain
I()

E) =
f
2
2
(
2
0
2
)
2
+ 4
2
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
4m
1
2
+
2
. (5.33)
This is the famous Lorentzian line shape function, giving the intrinsic width
of spectral absorption (and emission) linesalthough the dominating eect
when astronomers observe them 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
_

0
I()d =
_

0
I()d
_

I()d =
f
4m
_

1
1 +

2
2
d
=
f
4m
. (5.34)
5.3 Several degrees of freedom
Let us now consider N degrees of freedom. The variables are x
i
, and the
equations to be solved are
m
ij
x
j
+
ij
x
j
+k
ij
x
j
= F
i
(t) , 1 i, j N . (5.35)
5.3 Several degrees of freedom 55
When
ij
= 0 they follow from the Lagrangian
L =
1
2
m
ij
x
i
x
j
1
2
k
ij
x
i
x
j
+x
i
F
i
, (5.36)
and it follows that m
ij
and k
ij
are symmetric when considered as matrices. It
is less obvious that
ij
=
ji
. (5.37)
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 oscil-
lators that interact among themselves, but not with the outside world. That
is
m
ij
x
j
+k
ij
x
j
= 0 . (5.38)
We know that the general solution must contain 2N integration constants, or
equivalently N complex integration constants. So we try the Ansatz
x
i
=
i
e
it
. (5.39)
Inserting this into the equation, and cancelling the exponential, gives the ma-
trix equation
(
2
m
ij
+k
ij
)
j
= 0 . (5.40)
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
k
ij

2
m
ij
= 0 . (5.41)
This is an eigenvalue equation for . 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 expo-
nentially growing or decaying solutions, and the system would be unstable.
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. (5.40) 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 oscilla-
tions are small we use the Lagrangian
Figure 5.1. Two identical pendula connected by an elastic spring.
L = T V =
1
2
_
2
1
+

2
2
2
1
2
2
k(
1
2
)
2
_
. (5.42)
To nd the normal modes we must diagonalise T and V . This is achieved by
q
1
=
1
2
(
1
+
2
) , q
2
=
1
2
(
1
2
) . (5.43)
The Lagrangian becomes
L =
1
2
_
q
2
1
+ q
2
2
q
2
1
2
q
2
2
_
,
2
= 1 + 2k . (5.44)
The normal modes have a simple interpretation. If q
2
= 0 the two pendula
move in phase with the original frequency (equal to 1), if q
1
= 0 the pendula
move in opposite phase with a frequency increased by the spring. As an af-
terthought we observe that, in this case, the two normal modes could have
been identied from the outset using only symmetry considerations.
The general solution for
1
,
2
is thus a sum of two harmonicsand if the
dierence 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 5.
When we couple a set of harmonic oscillators to an external force it is the
normal modes that count. As the simplest example, consider
x
1
+
2
0
x
1
+k(x
1
x
2
) = f cos t
x
2
+
2
0
k(x
1
x
2
) = 0 .
(5.45)
In terms of the normal modes q
1
= x
1
+x
2
, q
2
= x
1
x
2
this becomes
q
1
+
2
0
q
1
= f cos t
q
2
+ (
2
0
+ 2k)q
2
= f cos t .
(5.46)
Each normal mode gives rise to its own resonance frequency. The mathemat-
ical problem has been reduced to that of studying twoor in general N
independent harmonic oscillators.
Finally, let us consider eq. (5.38) for N degrees of freedom. We can decouple
5.3 Several degrees of freedom 57
the degrees of freedom in three steps: First we diagonalise the symmetric
matrix k
ij
, 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 diagonalised matrix becomes the identity matrix. These coordinate
changes are conveniently described on the Lagrangian level,
L =
1
2
i,j
( x
i
k
ij
x
j
x
i
m
ij
y
j
) =
1
2
i,j
( y
i
ij
y
j
y
i
m
ij
y
j
) =
(5.47)
=
1
2
i,j
_
z
i
ij
z
j
z
i
1
i
m
ij
1
_
j
z
j
_
.
For once we did not use the Einstein summation convention, since it would not
work here. Note that m
ij
is not the same matrix as m
ij
, since the coordinate
system was rotated. In the nal step we dene the symmetrix matrix
m
ij
=
1
i
m
ij
1
_
j
, (5.48)
we diagonalise it, and arrive at the Lagrangian
L =
1
2
i
( u
i
u
i
2
i
u
i
u
i
) . (5.49)
The normal modes u
i
are now decoupled, and the equations of motion trivial
to solve.
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 Solve eq. (5.5) when F(t) = f cos (t +).
Problem 5.3 Derive eq. (5.8) by inserting the relevant force into Duhamels
integral.
Problem 5.4 Consider a bouncing ball, obeying z = g but with a oor
at z = 0. At each bounce the absolute value of its velocity decreases by a factor ,
0 < < 1. Assume it left the oor with velocity v
0
at time t = 0. At what time does
the motion stop? Sketch the motion in the zt plane.
Problem 5.5 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 rst pendulum is stationary and all the energy has gone to the second.
Problem 5.6 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
Most people have 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 xed point, and is uniquely characterized by its angle of rotation
. Its action on the coordinates is
_
x
_
=
_
cos sin
sin cos
__
x
y
_
. (6.1)
What does this mean? Actually it can mean two quite dierent 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 dierent values of the coordinates. These are two very dierent
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 g
1
, g
2
, etc, together with a rule for combining them, so that g
1
g
2
= g
3
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 nally every element g possesses an inverse g
1
such that
gg
1
= g
1
g = e. In general it may or may not be true that g
1
g
2
= g
2
g
1
.
For the two dimensional rotation group the elements can be written g(), so
the number of elements are continuously innite. 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 is given by an analytic function of the
parameters of its factors. Thus
g(
1
) g(
2
) = g(
3
(
1
,
2
)) . (6.2)
6.1 Rotations 59
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 proper-
ties 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 rst
place because rotations do not commute in general. Rotations can be repre-
sented by matrices acting on vectors,
x
i
= R
ij
x
j
. (6.4)
By denition this is a rotation if the lengths of all vectors are preserved,
x
i
x
i
= x
i
x
i
. (6.5)
Hence the matrix R
ij
must be subject to some restrictions. On the other hand
it is immediate from the denition 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:
x
i
y
i
= R
ki
x
i
R
kj
y
j
= x
i
R
ki
R
kj
y
j
= x
i
ij
y
j
= x
i
y
i
. (6.6)
Here
ij
is the Kronecker delta. Since the vectors are arbitrary this is equivalent
to
kl
R
ki
R
lj
= R
ki
R
kj
=
ij
. (6.7)
In matrix notation this is
R
T
R = RR
T
= 1 . (6.8)
The group properties can be now be checked on the level of matrices. Matrix
multiplication is associative, and moreover
R
1
R
2
= R
3
R
T
3
R
3
= (R
1
R
2
)
T
R
1
R
2
= R
T
2
R
T
1
R
1
R
2
= 1 . (6.9)
This conrms that R
3
is a group element. We observe that
1 = det RR
T
= det Rdet R
T
= (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:
R
im
R
jn
R
kp
mnp
=
ijk
det R =
ijk
, (6.11)
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.
where the denition (sic!) of the determinant was used. In N dimensions the re-
stricted rotation group is called the special orthogonal group, denoted SO(N).
Special refers to the restriction that the determinant equals one. If matri-
ces with determinant 1 are admitted the group is called O(N), and includes
reections.
Eulers theorem states that any rotation in 3-space is a rotation around a
xed axis. To prove it, note that an orthogonal matrix is unitary, and the
eigenvalues of a unitary matrix always take the form
= e
i
(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 xed axis
of rotation. Note that rotations in even dimensions (like two dimensions) work
dierently. Note also that rotations can be hard to grasp: if the rotation axes
of R
1
and R
2
are known, what is the rotation axis of R
1
R
2
?
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 identied 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 identied because of the periodicity of . It sounds simple, but there
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
6.1 Rotations 61
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
dicult to make them appear natural. We introduce them by brute force, as
follows: Dene
R
=
_
_
cos sin 0
sin cos 0
0 0 1
_
_
R
=
_
_
1 0 0
0 cos sin
0 sin cos
_
_
R
=
_
_
cos sin 0
sin cos 0
0 0 1
_
_
.
Compute
R(, , ) = R
= (6.14)
=
_
_
cos cos cos sin sin cos sin + cos cos sin sin sin
sin cos cos sin cos sin sin + cos cos cos cos sin
sin sin sin cos 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 or-
thogonal 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 ll out a right
handed orthonormal triad. All this is true provided that
0 < 2 , 0 < 2 , 0 . (6.15)
We accept this, and now we have a coordinate system on the group manifold
SO(3) available whenever we need one.
In writing eq. (6.14) we proved that an arbitrary rotation can be eected by
rst rotating around the z-axis, then around the x-axis, and nally around the
z-axis again. There are 322 = 12 dierent ways of choosing the axes here,
leading to 12 dierent 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 denitions that was used.
A nal 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 = e
A
1 +A +
1
2!
A
2
+. . . . (6.16)
This is an orthogonal matrix if and only if the matrix A is anti-symmetric,
A
T
= A . (6.17)
We can dene a special class of curves in the group by
R(t) = e
tA
R(0) = 1 . (6.18)
We would like to claim that every rotation group element can be written as
e
tA
, for some choice of anti-symmetric matrix A. This is actually so for any
SO(3) rotation, but not for the additional reections present in O(3). To see
how it works, we begin by using the denition (6.16) to conclude that
exp
_
t
_
0 1
1 0
__
=
_
cos t sin t
sin t cos t
_
. (6.19)
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 non-
commutativity means in terms of what goes on close to the identity element:
R
1
(t
1
)R
2
(t
2
)R
1
1
(t
1
)R
1
2
(t
2
)
(1 +t
1
A
1
+
t
2
1
2
A
2
1
)(1 +t
2
A
2
+
t
2
2
2
A
2
2
)(1 t
1
A
1
+
t
2
1
2
A
2
1
)(1 t
2
A
2
+
t
2
2
2
A
2
2
)
(6.20)
1 +t
1
t
2
(A
1
A
2
A
2
A
1
) .
The group elements commute only if the commutator [A
1
, A
2
] 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
[A
1
, A
2
] 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
A
1
=
_
_
0 0 0
0 0 1
0 1 0
_
_
A
2
=
_
_
0 0 1
0 0 0
1 0 0
_
_
A
3
=
_
_
0 1 0
1 0 0
0 0 0
_
_
. (6.21)
6.2 Rotating coordinate systems 63
They obey
[A
1
, A
2
] = A
3
[A
2
, A
3
] = A
1
[A
3
, A
1
] = A
2
. (6.22)
They form a basis for the Lie algebraand would perhaps look more familiar
had we multiplied the Lie algebra elements A
i
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
A
ij
=
ikj
k
, (6.23)
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
, (6.24)
then
i
is referred to as the angular velocity vector at time t = 0. In space it is
pointing along Eulers xed axis of rotation. If we return to the picture of the
group manifold as a solid ball with antipodal points on its surface identied,
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.
6.2 Rotating coordinate systems
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
gure out the laws if all experiments had been carried out on roundabouts.
Let X
i
(t) denote the trajectory of a particle relative to an inertial system, and
let x
i
(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
X
i
(t) = R
ij
(t)x
j
(t) =
_
ij
+
ikj
k
t +o(t
2
)
_
x
j
(t) . (6.25)
At time t = 0 the inertial and rotating coordinate systems agree, and we
assume that the angular velocity vector
i
is constant. The time derivatives
are related by
X
i
= R
ij
x
j
+

R
ij
x
j
. (6.26)
At t = 0 (a point in time picked by convention) our assumption gives
X
i
= x
i
+
ikj
k
x
j
. (6.27)
The dot-notation can get confusing at this point, so it may be helpful to write
this as an operator relation
X
i
= D
tij
x
j
=
_
ij
d
dt
+
ikj
k
_
x
j
, (6.28)
If we dierentiate twice we obtain Newtons second law in the form
F
i
= m

X
i
= mD
tij
D
tjk
x
k
= m x
i
+ 2m
ijk
j
x
k
+m
ijk
kmn
m
x
n
. (6.29)
By the --identity, or equivalently by means of the formula for repeated cross
products, this is
m x
i
= F
i
2m
ijk
j
x
k
+m
_
ij

i
j
_
x
j
. (6.30)
This is what Newtons second law looks like on the roundabout. The extra
terms on the right hand side are known as inertial or ctitious forcesbut
they feel real enough.
The third term on the right hand side of eq. (6.30) is known as the cen-
trifugal force. It is responsible for the repulsive part of the eective potential
(4.8) in the two body problem. The second term is the Coriolis force. It is per-
pendicular 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 = k
2
x + 2
z
y y = k
2
y 2
z
x . (6.31)
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 = e
t
:
w +k
2
w + 2i
z
w = 0
2
+k
2
+ 2i
z
= 0 . (6.32)
The solutions are
= i
z
i
_
k
2
+
2
z
i
z
i
_
k +

2
z
2k
_
i
z
ik . (6.33)
6.2 Rotating coordinate systems 65
Figure 6.2. This picture illustrates Einsteins article on Beers Law, so presum-
ably this is what the great mans tea cup looked like.
Therefore
w = e
izt
_
c
1
e
ikt
+c
2
e
ikt
_
. (6.34)
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 ctitious 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 ow 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 prole will be set
up in the cup. Therefore the tea leavesstaying close to the bottom due to
their specic 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 prole will carry fast moving surface water to the outer river
bank, and will cause erosion there. The latter will therefore suer 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
law does not apply to small and bending rivers, where the centrifugal force
dominates.
Even more dramatically, x 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. The Coriolis force steps
in to the rescue and prevents the sun from disappearing into outer space.
6.3 The inertia tensor
A rigid body is dened as a number of particles at xed distances from each
other, or equivalently as a continuous mass distribution whose shape is xed.
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 conguration space of a rigid
body is R
3
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 xed at some
point O within the body. These coordinates are called x
i
. We also introduce
an inertial coordinate system so that O has coordinates X
i
. The velocity of
any given point x
i
within the body with respect to the inertial coordinate
system is denoted v
i
, and the velocity of the point O is denoted V
i
. 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 v
i
is not
the time derivative of x
i
, and that we will not be bothered by ctitious forces.
The various quantities we introduced are related by
v
i
= V
i
+ x
i
= V
i
+
ikj
k
x
j
, (6.35)
To what extent does it matter where the origin O is placed? If we translate it
so that
x
i
x
i
= x
i
+a
i
, (6.36)
we see that the instantaneous velocities transform according to
v
i
= V
i
+
ikj
k
x
j
= V
i
ikj
k
a
j
+
ikj
k
x
j
= V
i
+
ikj
k
x
j
. (6.37)
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
T =
m
2
(V
i
+
ikj
k
x
j
)
2
=
(6.38)
=
_
m
2
V
2
+m
ikj
V
i
k
x
j
+
m
2
(
ikj
k
x
j
)
2
_
.
6.3 The inertia tensor 67
(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
T =
M
2
V
2
+
1
2
m(x
2
ij
x
i
x
j
)
i
j
+
ikj
V
i
mx
j
. (6.39)
If O, the origin within the body, is placed at the center of mass the last term
vanishes. Of course it also vanishes if the origin coincides with a point that is
xed in space. Now dene the inertia tensor with respect to O,
I
ij
=
m(x
2
ij
x
i
x
j
) . (6.40)
We observe that the angular momentum with respect to O is
L
i
=
ijk
x
j
m x
k
=
ijk
kmn
mx
j
m
x
n
= I
ij
j
. (6.41)
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
T =
1
2
MV
2
+
1
2
I
ij
j
=
1
2
M
1
P
i
P
i
+
1
2
I
1
ij
L
i
L
j
. (6.42)
The inertia tensor describes the resistance of the body to changes of its rota-
tion, 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
I =
_
_
_
_
_
_
m(x
2
2
+x
2
3
)
mx
1
x
2

mx
1
x
3
mx
2
x
1
m(x
2
1
+x
2
3
)
mx
2
x
3
mx
3
x
1

mx
3
x
2
m(x
2
1
+x
2
2
)
_
_
_
_
_
_
. (6.43)
This is a symmetric matrix, hence it can be diagonalized by rotations. This
means that we can rotate the internal coordinate system so that
I =
_
_
I
1
0 0
0 I
2
0
0 0 I
3
_
_
. (6.44)
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 n
i
,
n
i
I
ij
n
j
=
m
_
n
2
x
2
(n
i
x
i
)
2
_
0 . (6.45)
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
(n x)
2
= n
2
x
2
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
I
1
+I
2
=
m(x
2
1
+x
2
2
+ 2x
2
3
)
m(x
2
1
+x
2
2
) = I
3
. (6.46)
Hence no moment of inertia can exceed the sum of the other two. Equality
happens if and only if x
3
= 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 reection in a plane, two of the principal
axes lie in it.
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 xed 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 more than three higher order symmetry axesthis is true for a cube, say
it follows that all the eigenvalues are equal, so in fact every axis is a principal
axis.
The shape of the body is reected 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 an ellipsoid, and nothing else matters as far as its 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 I
ij
be its inertia tensor
relative to the center of mass, and I
ij
its inertia tensor relative to a point
translated from the center of mass by the vector a
i
. Then
I
ij
= I
ij
+M(a
2
ij
a
i
a
j
) . (6.47)
Roughly speaking it is easiest to rotate the body around its center of mass.
The proof is easy, once we recall that
ma
i
x
i
= a
i
mx
i
= 0 , (6.48)
where x
i
is position relative to the center of mass.
It is instructive to prove that any inertia tensor can be reproduced by placing
6.4 Eulers equations 69
four particles of equal mass at appropriate distances from each other. To do
so, let the coordinates of the particles be
(x
i
, y
i
, z
i
) , 1 i 4 . (6.49)
This permits us to dene three vectors x, y, z in R
4
. We dene a fourth aux-
iliary vector whose components are equal,
v
T
= (1, 1, 1, 1) . (6.50)
That is to say that we are looking at the matrix
[x y z v] =
_
_
x
1
y
1
z
1
1
x
2
y
2
z
2
1
x
3
y
3
z
3
1
x
4
y
4
z
4
1
_
_
, (6.51)
where the row index labels the four particles. Now we insist that the inertia
tensor be diagonal and that the center of mass is at the origin. This translates
into the six conditions
x y = x z = y z = v x = v y = v z = 0 . (6.52)
In other words the four vectors must be mutually orthogonal, which is easily
arranged in R
4
. It remains to check that we can reproduce the most general
diagonal inertia tensor; the solution is given in exercise 3.
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,
that are associated with its shape andwith a little practicecan 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.
When its center of mass is xed, the equation of motion for a rigid body
subject to a torque
i
is
L
i
=
m
ijk
x
j
F
k
=
i
, (6.53)
where both the angular momentum and the torque are computed with respect
to the center of mass. The equation is complicated by the fact that the inertia
tensor is time dependent,
L
i
= I
ij

j
+

I
ij
j
. (6.54)
This is awkward, whether there is a torque or no.
For the rest of this section we conne ourselves to torque free motion, with
the center of mass kept xed. (Even if we are in the gravitational eld of the
earth it is possible to manufacture tops that are suspended at their centers
of mass, and which therefore realize torque free motion.) Theoretically, the
guiding idea will be to formulate the equations of motion with respect to a
coordinate system xed within the body, because then the inertia tensor is
time independent. The direction of the angular momentum vector will change
with time though, because we are in a rotating coordinate system. We denote
it by J
i
. We proceed in analogy with eq. (6.27), that is to say we assume that
the intertial and body xed coordinate systems agree at t = 0, and get for
torque free motion
J
i
+
ikj
k
J
j
= I
ij

j
+
ikj
k
J
j
= 0 . (6.55)
Being moderately intelligent we adapt the rotating coordinate system so that
its axes coincide with the principal axes, and arrive at Eulers equations
I
1

1
+ (I
3
I
2
)
2
3
= 0
I
2

2
+ (I
1
I
3
)
3
1
= 0
I
3

3
+ (I
2
I
1
)
1
2
= 0 .
(6.56)
These equations can be solved exactly in terms of elliptic functions, which
means that the spinning top in the absence of torque 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
, (6.57)
where the
i
time dependent and assumed to be small. Then expand the
equations to rst order in
i
:
I
1

1
= 0
I
2

2
+ (I
1
I
3
)
0
3
= 0
I
3

3
+ (I
2
I
1
)
0
2
= 0
I
2

2
+
(I1I3)(I1I2)
I3
2
0
2
= 0
I
3

3
+
(I2I1)(I3I1)
I2
2
0
3
= 0 .
(6.58)
The perturbation will grow, and the solution will be unstable, unless
I
1
> I
2
, I
3
or I
1
< I
2
, I
3
. (6.59)
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
6.4 Eulers equations 71
Figure 6.3. The inertia ellipsoid, and its intersections (as curves) with spheres
of three dierent sizes.
J
2
1
I
1
+
J
2
2
I
2
+
J
2
3
I
3
= 2E . (6.60)
The angular momentum vector itself is changing (in this coordinate system),
but its magnitude remains constant:
J
2
1
+J
2
2
+J
2
3
= constant . (6.61)
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 xed points
at the major and minor axes, and hyperbolic xed 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 dene a dy-
namical system whose trajectories are conned to a three dimensional space,
for which we can use the coordinates J
1
, J
2
, J
3
. A solution of the Euler equa-
tions is a curve in this three dimensional space, and curves in a three dimen-
sional 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 denes a set of easy-to-describe surfaces that ll 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 Eu-
ler equations. The Lorenz equations (1.44) behave quite dierently: for them
there are no well behaved constants of the motion, the solutions are unruly,
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. Using inertial coordinates
we dene the ellipsoid of inertia
X
i
I
ij
X
j
= 2E . (6.62)
Think of it as a massless shell surrounding the body. If we compute the inertia
tensor with respect to Absolute Space we will nd 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
L
i
i
= 2E . (6.63)
Since the angular momentum vector is xed in absolute space we can assume
that it points along the X
3
-axis, and then
3
is a constant, which means that
the herpolhode is a plane curve. The plane to which it is conned is known
as the invariable plane, and is orthogonal to L
i
. Now the normal vector of the
ellipsoid of inertia is
n
i
(X) = 2I
ij
X
j
. (6.64)
Evaluated at the point
i
this is
n
i
() = 2I
ij
j
= 2L
i
. (6.65)
This coincides with the normal of the invariable plane. Because of eq. (6.63)
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 X
3
-axis. Hence there are two
frequencies involved, and if they are not commensurable in the sense of eq.
(1.36) the herpolhode never closes. This is quite reminiscent of the Lissajous
gures.
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 L
i
. This motion is called regular
precession, and is not to be confused with the precession of a top placed in
a gravitational eld, or with the precession of the equinoxes caused by the
spinning earth. Note that Eulers equations (6.56) for a symmetrical top are
easily integrated in terms of trigonometric functions.
The Earth is a symmetrical top, with I
1
,= I
2
= I
3
and
6.5 The Lagrangian description 73
I
1
I
3
I
1
1
305
. (6.66)
This is the ellipticity of the Earth. Moreover its rotation axis diers slightly
from its geometrical symmetry axis. Judging from Eulers equations, 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 eect 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 uid core inside a rigid
shell need not invalidate the argument.
Incidentally, note that it would be dicult 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 eects.
6.5 The Lagrangian description
Evidently the Lagrangian of the rigid body must be L = T V , but going back
to our expression (6.42) 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 conguration 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
A
ij
=
ikj
k
A =
1
A
1
+
2
A
2
+
3
A
3
, (6.67)
where we use the basis introduced in eqs. (6.21) and
i
is the angular velocity
vector. Furthermore
R = e
tA
A =

RR
1
. (6.68)
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 o the angular velocity vector by comparing to eq. (6.67).
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:
A =

RR
1
=
d
dt
(R
)(R
)
1
=
=

R
R
1
+R
R
1
R
1
+R
R
1
R
1
R
1
(6.69)
=

A
3

(cos A
1
+ sin A
2
)

(sin (sinA
1
cos A
2
) + cos A
3
) .
Comparing to eq. (6.67) we read o that
1
=

cos +

sin sin
2
=

sin

sin cos
3
=

+

cos .
(6.70)
This we can insert in a Lagrangian.
We will choose a Lagrangian to describe a symmetrical top with one point
xed, subject to a gravitational force. The restriction to a symmetrical top
implies that two of the eigenvalues of the inertia tensor are equal, say I
1
= I
2
;
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 xed 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 the 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 () = Mgl cos . (6.71)
Now we can write the Lagrangian. We obtain
L =
I
1
2
_
2
1
+
2
2
_
+
I
3
2

2
3
V () =
(6.72)
=
I
1
2
_
2
+

2
sin
2
_
+
I
3
2
(

+

cos )
2
Mgl cos .
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 nd the constants of motion
L
z
=
L
= I
1

sin
2
+I
3
cos (

+

cos ) (6.73)
L
3
=
L
= I
3
(

+

cos ) . (6.74)
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:
+

cos =
L
3
I
3
(6.75)
=
L
z
L
3
cos
I
1
sin
2
. (6.76)
Next we turn to the energy, which is
E =
I
1
2
(
2
+

2
sin
2
) +
I
3
2
(

+

cos )
2
+Mgl cos =
(6.77)
=
I
1
2
2
+
(L
z
L
3
cos )
2
2I
1
sin
2
+Mgl cos +
L
2
3
2I
3
.
This leads to a rst order dierential equation which can be explicitly solved
in terms of elliptic functions. Inserting the result successively in eqs. (6.76)
and (6.75) will lead us to the complete solution for the heavy symmetrical
top.
As usual one can go a long way with qualitative arguments. Eq. (6.77),
which governs the nutation of the top, can be described as one dimensional
motion in an eective potential. It can be written on the form
=

2
+
(b a cos )
2
sin
2
+ cos , (6.78)
where
=
2EI
3
L
2
3
I
1
I
3
, b =
L
z
I
1
, a =
L
3
I
3
, (6.79)
=
2Mgl
I
1
> 0 . (6.80)
If we introduce the variable u = cos the eective potential becomes a cubic
polynomial,
u
2
= u
3
( +a
2
)u
2
+ (2ab )u + b
2
. (6.81)
The velocity

can change sign only when the right hand side vanishes. The
polynomial on the right hand side goes from to , and
Figure 6.4. Precession might possibly turn retrograde.
u = 1 u
2
= (b a)
2
. (6.82)
Except for the special case that b = a it follows that the cubic polynomial has
two roots in the physically relevant interval 1 u 1. This observation is
enough to reveal the key qualitative features of the nutation of the top.
For the precession, described by eq. (6.76), an interesting qualitative ques-
tion is whether the derivative

changes its sign as evolves. The answer
depends on the relative size of the initial data L
z
and L
3
; see Fig. 6.4 for the
possibilities.
To see why there is precession in the rst place, take a top whose point
of contact with the ground is xed 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 L
z
must remain
constant. The top manages to keep L
z
constant because it begins to precess,
thus adding an extra precessional contribution to L
z
.
These considerations are applicable to the Earth, which is an oblate sym-
metrical top subject to tidal forces, primarily from the Moon. 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. The nutation of the Earth was reported by Bradley
in the 18th century, after rst observing it for a complete period, 18.7 years. A
lesson, perhaps, for modern astronomers. It is a small eect, but signicantly
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 Place four equal masses at the corners of a regular tetrahedron
and compute the inertia tensor with respect to their center of mass.
Problem 6.3 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 6.4 Compute the inertia tensor with respect to the center of mass
for a sphere, a cube, a circular cylinder and a circular cone, all of them having constant
density.
Problem 6.5 The mass of the Sun is 2 10
30
kg, its equatorial radius is 7 10
8
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 210
27
kg, its semi-
major axis is 8 10
11
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 6 10
24
kg and its equatorial radius is 6 10
6
m, the mass of the Moon is 7 10
22
kg and its
semi-major axis is 4 10
8
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 Repeat the derivation leading to eqs. (6.70), but use the equally
valid denition
A = R
1

R .
What dierence does it make?
Problem 6.8 For the symmetrical top in a gravitational eld, show that
rotation around the vertical axis is stable if and only if
L
2
z
> 4MglI
1
. (6.83)
Problem 6.9 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 formu-
lation 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 eect 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 dierentiable,
and obeys
d
2
f
dx
2
> 0 . (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 dened as
g(p) = max
x
_
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 monoto-
neously 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) denes the function g(p) uniquely.
A function and its Legendre transform are related by
Interlude: Legendre transformations 79
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 denes 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) denes 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
d
2
g
dp
2
=
dx
dp
=
_
dp
dx
_
1
=
_
d
2
f
dx
2
_
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. Generalisation to an arbitrary number of variables is quite
straightforward.
80 Interlude: Legendre transformations
There is an interesting geometric interpretation of the Legendre transfor-
mation, and I hope this can be deciphered from Fig. 7.1: to nd out the value
of the function g at p
0
, start by drawing the straight line p
0
x, and then use the
denition of the Legendre transform to nd g(p
0
) 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
f(x) =
_
_
_
1 x , x < 1
0 , 1 x 2
x 2 , 2 < x
. (7.7)
Compute the Legendre transform using the denition (7.2).
8 The Hamiltonian formulation
Any set of ordinary dierential equations can be written in rst 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 ows
For a Lagrangian of the form L(q, q), which is general enough for our purposes,
the Euler-Lagrange equations are
d
dt
L
q
i
=
L
q
i
. (8.1)
These are second order ODEs. One obvious way to turn them into rst order
equations is to dene the canonical momenta
p
i

L
q
i
. (8.2)
The expression on the right hand side will be equal to m q
i
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 rst order form. The use of q for
generalized coordinates and p for their momenta goes back to Jacobi, and
was solidied by Whittaker in a famous textbook.
Eqs. (8.1) now take the rst order form
p
i
=
L
q
i
. (8.3)
But we need equations for q too. To this end we assume that eqs. (8.2) can be
inverted, that is
p
i
= p
i
(q, q) q
i
= q
i
(q, p) . (8.4)
Now we do have a rst 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) = q
i
p
i
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
q
i
=
H
p
i
p
i
=
H
q
i
. (8.6)
These are Hamiltons equations, and the function H is known as the Hamilto-
nian. The variables p
i
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
L =
m
2
q
2
H =
1
2m
p
2
. (8.7)
An apparently more complicated case, involving many degrees of freedom, is
L =
1
2
M
ij
(q) q
i
q
j
V (q) . (8.8)
Here the matrix M depends on the conguration 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
H =
1
2
p
j
M
1
ij
p
j
+V (q) . (8.9)
We only have to check the matrix M is invertible, everywhere in congura-
tion space. This in fact amounts to checking that the Lagrangian is a convex
function of the variables q
i
.
Our phase space is spanned by the 2n variables q
i
, p
i
. 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 dimen-
sional. But there is another and more dramatic dierence. In section 1.4 time
8.1 Hamiltons equations and Hamiltonian ows 83
evolution was described by the equations z
i
= f
i
(z), so that the general case
is obtained by choosing 2n independent functions f
i
. 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 = q
i
H
q
i
+ p
i
H
p
i
+
H
t
=
H
p
i
H
q
i
H
q
i
H
p
i
+
H
t
=
H
t
= 0 . (8.10)
Hence the Hamiltonian is always a conserved quantity, unless it depends ex-
plicitly 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 ow of a uid. The ow lines
are dened by the little arrows in phase space. The uid is incompressible
if, for any xed nite volume in phase space, the amount of uid going out
through its surface equals the amount that is going in. By Gauss theorem the
dierence between them is equal to the integral of the divergence of the ow
over the volume. Since this must vanish for every volume we conclude that the
ow is that of an incompressible uid if and only if its divergence vanishes.
For Hamiltonian time evolution this is indeed so:
div z = v =
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 ow 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 xed points that occur are either elliptic or hyperbolic.
On reection 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 uid passing close to a hyperbolic xed 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.
Figure 8.1. Liouvilles theorem in action: a volume element in phase space pre-
serves its volume, but may become very distortedhere because it approaches
a hyperbolic xed 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 [). They form a vector space because
one can take linear combinations of kets: if [
1
) and [
2
) are kets and a
1
, a
2
are numbers then [
3
) is a ket too, where
[
3
) = a
1
[
1
) +a
2
[
2
) . (8.13)
In quantum mechanics the numbers a
1
, a
2
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 [. By denition a bra is a
linear function from the vector space of kets to the real numbers. Thus, given
a bra [,
[) = a (8.14)
is dened for all kets [). 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 innite 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,
(a
1
1
[ +a
2
2
[)[) = a
1
1
[) +a
2
2
[) . (8.15)
Hence
3
[ = a
1
1
[ +a
2
2
[ is a bra.
8.2 Kets and bras and all that 85
At this point we have two vector spaces, and a priori they are dierent. The
equation [ = [) makes no sense at all. Apples and pears are never equal.
However, there is a nal twist to the story, because there does exist a one-to-
one 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
[ [). It means that we can associate a unique number to any ket, namely
the real number [). This number is the length squared of the ket [).
How can we make the distinction between bras and kets in the index no-
tation? 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 impor-
tant 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 dene this map as
u(x) = u
= a , (8.16)
where Einsteins summation convention is understood. It is important to note
that in the upstairs-downstairs notation, the expressions x
and u
are
simply not allowed, so that we cannot dene the length squared of a vector
x
as 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
[) = [e
)x
, (8.17)
where the vectors [e
) form a basis (and must not be confused with the com-

ponents 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
and v
= u
(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 dierent kinds of sense. The vector
[
) = [e
)y
= [e
)A
= A[) (8.19)
is a new vector, obtained by transforming the old vector [). This is an active
transformation of the vector. On the other hand the vector
[) = [e
)x
= [e
)(A
1
)
= [e
)y
(8.20)
is just the original vector [) 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
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
1
)
. (8.21)
This is part of the origin of the names contravariant and covariant vectors
they transform in the opposite direction, to ensure that u
remains un-
changed.
We can go on to dene tensors with more than one index in the same way.
A tensor with k indices running from 1 to n is dened as a collection of kn
components transforming in specic way under changes of basis. Examples
include
T
, S
(A
1
)
(A
1
)
(A
1
)
. (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
denition 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 dene the analogue of [).

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 denition 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
. (8.24)
Here the left hand side denes an operation which acts as the identity on both
our vector spaces,
8.2 Kets and bras and all that 87
Figure 8.2. Bras and kets: kets are arrows, bras collections of parallel planes
(or measuring tapes). Their lengths are undened but one can multiply them
with real numbers, and the quantity u
has a clear meaning.
= x
and u
= u
. (8.25)
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
= g
= g
. (8.26)
We dene the length squared of a vector x
as x
. If g
we recover
the naive denition, used implicitly in section 6.1. Finally, let us consider eq.
(6.7), which dened 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
= 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 precisely as a set of parallel
hyperplanes with constant spacing. If the bra u
is multiplied with 2, the

spacing of the level surfaces is multiplied with 1/2. The real number u
will be represented by the number of parallel planes pierced by the arrow.

However, the true length of x
is not dened yet. You might think that it

could be dened by choosing some special measuring tape, rotating this tape
until it is aligned with the arrow, and afterwards compute u
. But this
assumes that the measuring tape is unchanged by the rotation, and the catch
is that rotation is dened as a transformation that preserves the lengths of
all arrows, or the spacing of all measuring tapes. We are caught in a circular
denition, and forced to introduce some further structurea metric!before
we can dene 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 identied them from the outset, without
comment. But the more careful treatment here is useful for various generali-
sations. In particular it is useful to understand the Hamiltonian formulation
of the equations of motion. There may even be a moral here. Forbidding cer-
tain 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 sin-
gle vector x
, how can we assign a number to it? We know the answer. We

introduce a metric tensor, and dene
[[x[[
2
= 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
= [[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
A =
x
1
y
1
x
2
y
2
. (8.30)
We can write this as
8.3 The symplectic form 89
Figure 8.3. Two ways of associating a number to a pair of vectors: the angle
they subtend, and the area they span.
A = y
, (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 dening an innitesimal rotation in 3-space
necessarily has a zero eigenvalue, and hence the rotation it generates has a
xed axis.)
Just as we may wish to dene metrics on curved spaces, not only on vector
spaces, so we may wish to dene 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 nal requirement on
a symplectic form is
= 0 . (8.35)
We will mostly be interested in the case when the components of
are con-
stant, 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 8.7.
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 . . . 0 0
1 0 0 0 . . . 0 0
0 0 0 1 . . . 0 0
0 0 1 0 . . . 0 0
.
.
.
.
.
.
.
.
.
.
.
.
0 0 0 0 . . . 0 1
0 0 0 0 . . . 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 R
2n
, 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 that metric tensors behave in a completely dierent
way.
Because the symplectic form has an inverse, it can be used to relate con-
travariant 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 91
z
=
_
_
_
_
_
_
_
_
_
_
_
z
1
z
2
z
3
z
4
.
.
.
z
2n1
z
2n
_
_
_
_
_
_
_
_
_
_
_
=
_
_
_
_
_
_
_
_
_
_
_
q
1
p
1
q
2
p
2
.
.
.
q
n
p
n
_
_
_
_
_
_
_
_
_
_
_
. (8.38)
Remembering that we agreed to be careless about the position of Latin indices
we then nd that
z
H
_
_
_
q
i
=
H
pi
p
i
=
H
qi
.
(8.39)
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 specied by
choosing a denite Hermitian operator

H. In the Heisenberg picture the time
evolution of an arbitrary operator is determined by the equation
d

A
dt
= i[

A,

H] . (8.40)
Here the bracket represents a commutator. It obeys the algebraic relations
[a
1

A
1
+a
2

A
2
,

B] = a
1
[

A
1
,

B] +a
2
[

A
2
,

B] , (8.41)
[

A,

B] = [

B,

A] , (8.42)
[

A, [

B,

C]] + [

C, [

A,

B]] + [

B, [

C,

A]] = 0 , (8.43)
where a
1
, a
2
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 rst sight there does not
seem to be a classical analogue of the commutator. But there is one! Using the
symplectic form we dene the Poisson bracket of two arbitrary phase space
functions as
A, B
B . (8.44)
By inspection we see that the Poisson bracket obeys
a
1
A
1
+a
2
A
2
, B = a
1
A
1
, B +a
2
A
2
, 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 (com-
plex) linear space, and by replacing the Poisson brackets with commutators
according to the rule
A, B
1
i
[

A,

B] . (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 (q
i
, p
i
). Then the Poisson bracket
reads
A, B =
A
q
i
B
p
i
A
p
i
B
q
i
. (8.51)
(Yes, as always a sum over repeated indices is understood.) Moreover eq. (8.48)
becomes
8.5 Innitesimal canonical transformations 93
q = q, H p = p, H . (8.52)
This form of Hamiltons equations is frequently seen.
For the explicit evaluation of Poisson brackets we look carefully at the def-
inition, and observe that
A, F =
F = (
q
A
p
p
A
q
)F . (8.53)
Hence we can think of A, as a dierential operator acting on phase space
functions. It obeys Leibniz rule
A, BC = BA, C +A, BC . (8.54)
In this sensewhich is a very important oneit acts like a kind of deriva-
tive. Using this the evaluation of any Poisson bracket can ultimately be made
starting from the fundamental Poisson brackets
q
i
, p
j
=
ij
, q
i
, q
j
= p
i
, p
j
= 0 . (8.55)
We can now conrm that the Poisson bracket acts like a derivative in the
sense that
q
i
, A =
pi
A p
i
, 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 Innitesimal canonical transformations
We have yet to discuss the classical analogue of unitary transformations. Uni-
tary transformations preserve commutators, so we dene canonical transfor-
mations as transformations that preserve Poisson brackets. This is the same
thing as saying that canonical transformations leave the symplectic form in-
variant, and is analogous to the denition of rotations as those transformations
that leave a specic 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 an-
other. 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) , (8.57)
then phase space functions are transformed according to
A
(q
, p
) = A(q, p) . (8.58)
(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
. (8.59)
How do we nd such transformations?
We begin by looking for innitesimal canonical transformations. Choose a
function F = F(q, p) on phase space, and set
A = A, F . (8.60)
Computing to rst order in , and using the Jacobi identity, we nd that
A
, B
= A, B + ( A, F, B +A, B, F ) =
(8.61)
= C F, A, B = C +C, F = C +C = C
(to that order). This works for any phase space function F. In principle it is
possible to integrate eq. (8.60) to obtain a nite canonical transformation
the equation has the same form as Hamiltons equations of motion, and the
only catch may be that it will be dicult to do the integration explicitly.
Again, note the similarity to quantum mechanics: in quantum mechanics
any Hermitian operator generates a unitary transformation, in classical me-
chanics any phase space function generates a canonical transformation. At the
same time there is an interesting dierence 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 nite dimension, there are indeed rather few rotations, but the set of
canonical transformations is always innite dimensional because the space of
all functions has innite 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 denition 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
8.6 The symplectic one-form 95
L
i
=
ijk
x
j
p
k
x
i
,
j
L
j
=
ijk
j
x
k
= x
i
. (8.62)
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 . (8.63)
By means of this identity it is easy to show that
L
i
, L
j
=
ijk
L
k
. (8.64)
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.
It should come as no surprise to hear that the results concerning rotations
and their Noether charges can be generalized to 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
nd that
x
i
= x
i
= x
i
, H . (8.65)
Again the Noether charge generates the transformation to which it owes its
existence.
8.6 The symplectic one-form
To proceed further with canonical transformations we need to develop some
more machinery. First a little discussion of vector elds. A contravariant vector
eld (with indices upstairs) is a 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 owlines of the vector eld. Hence the general denition:
A dynamical system is a vector eld 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 eld 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 eld
is the gradient of some function, u
f. And it can be shown that this is

so if and only if
= 0 . (8.66)
If this condition fails the level planes that exist at each point do not t together
to form level surfaces that extend over all space. Actually this phenomenon
is well known from the theory of curve integrals. A covariant vector eld is
just what we need in the integrand of such an integral. Suppose that a curve
is dened explicitly by the functions z
= z
(t), where t is some parameter

along the curve. Then
_
dz
(x) =
_
t2
t1
dt
dz
dt
u
(x(t)) . (8.67)
It is known that this integral is independent of the path and denes a function
f(x(t)), if and only if condition (8.66) holds. But the curve integral is well
dened regardless of whether this is true or not. This 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 eld is a map from the set of all owlines to
the real numbers.
This leads to a useful piece of notation. If the coordinates on our space
are denoted by z
and if u
is a covariant vector eld, then we dene the

dierential one-form u as
u = dz
. (8.68)
That is, the coordinate diererentials dz
are chosen as the basis in which we

expand the covariant vector. The gradient of a function f denes a one-form
of a special kind, namely
df = dx
f . (8.69)
The notation is useful since it gives the correct behaviour of u
(x) under
coordinate transformations. Let x
= x
(x). Then
u = dx
= dx
= dx
(x
) =
x
(x) . (8.70)
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 denition 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
. (8.71)
Here
is some covariant vector eld. In symplectic geometry it is known as

the symplectic one-form. (If you like, one-form is simply another name for
covariant vector eld.)
8.7 General transformation theory 97
When electrodynamics is written in tensor notation one nds that the elec-
tric and magnetic elds can be written as a two-form
F
=
_
0 E
i
E
i

ijk
B
k
_
=
_
_
_
_
0 E
1
E
2
E
3
E
1
0 B
3
B
2
E
2
B
3
0 B
1
E
3
B
2
B
1
0
_
_
_
_
. (8.72)
Eq. (8.71) are part of Maxwells equations; they guarantee the existence of
the vector potential, which is a one-form There is also a second tensor equa-
tion 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 coordi-
nates (q
i
, p
i
) on phase space, so that the symplectic two-form takes the form
(8.36)? There is no unique answer to this question since the previous discussion
implies that if
(8.73)
then
and
give rise to the same symplectic two-form. However, a possible

choice of symplectic one-form is
= p
i
dq
i
. (8.74)
This will be our standard choice. To make sure that you see how it works,
consider a two dimensional phase space with
z
=
_
q
p
_
,
=
_
p , 0
_
. (8.75)
It is simple to verify that
=
_
0
q
2
0
_
=
_
0 1
1 0
_
q, p = 1 . (8.76)
More generally, using eq. (8.38) we recover our standard symplectic two-form
(8.36) in any dimension.
8.7 General transformation theory
So far we have discussed innitesimal canonical transformations only, and we
must now turn to nite 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) . (8.77)
This is a canonical transformation if the Poisson brackets obey
Q, P = q, p . (8.78)
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
p
i
dq
i
= P
i
dQ
i
+dF , (8.79)
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.
Denote the function F by F
1
= F
1
(q, Q). Then we derive that
(p
q
F
1
) dq (P +
Q
F
1
) dQ = 0 . (8.80)
Since q and Q serve as coordinatessince dq and dQ are linearly independent
one-formswe can conclude from this that
p =
q
F
1
(q, Q) P =
Q
F
1
(q, Q) . (8.81)
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)
. (8.82)
This is the canonical transformation generated by the generating function
F
1
(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.
(8.81) in order to obtain the explicit canonical transformation. If you think
about it, it is clear that there cannot exist a generating function F
1
(q, Q)
generating the simplest canonical transformation we can think of, namely the
identity transformation
Q = q P = p . (8.83)
Evidently the pair q and Q does not coordinatize phase space. To circumvent
this diculty we dene a new generating function by means of a Legendre
transformation from the rst:
F
2
= F
1
(q, Q) +QP . (8.84)
The sign conventions here are a little odd. Still the relation between the func-
tions F
2
and F
1
is recognisably a Legendre transformation, and it followsby
virtue of eqs. (8.81)that F
2
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 F
2
to generate canonical transformations
by repeating the previous logic:
pdq = d(PQ)QdP +F
1
= QdP +dF
2
= QdP +
q
F
2
dq+
P
F
2
dP . (8.85)
Comparing coecients in front of the linearly independent one-forms, in this
case dq and dP, we conclude that
p =
q
F
2
Q =
P
F
2
. (8.86)
Again we assume that these equations can be inverted to yield eqs. (8.82) in
explicit form. The identity transformation is generated by
F
2
(q, P) = qP . (8.87)
We can go on to dene the generating functions F
3
= F
3
(p, Q) and F
4
=
F
4
(p, P) by means of further Legendre transformations, in the obvious way.
We have recovered the conclusion from section 8.5, namely that every func-
tion on phase space generates a canonical transformation, but now we have a
recipe for how to nd such transformations in nitenot only innitesimal
form. Hamiltons equations retain their form under all canonical transforma-
tions, with a new Hamiltonian dened as the function
K(Q, P) = H(q, p) = H(q(Q, P), p(Q, P)) . (8.88)
In chapter 9 we will regard canonical transformations as passive coordinate
transformations, and we will try to design a canonical transformation to ensure
that 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 re-
semble those of thermodynamics. In fact there is a close relation between
symplectic geometry and thermodynamics, although in the latter case the sit-
uation is a little dierentin 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 trans-
formations 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 eld, eq. (2.10).
Problem 8.2 The Lagrangian for a relativistic particle is
L = mc
_
c
2
x
2
. (8.89)
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
. Prove that this implies that the gradient of a function transforms like a co-
variant vector,
f(A
1
)
.
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
, in such a way that

x
= X
, (8.90)
and such that the new metric tensor G
becomes equal to a Kronecker delta. Show

that if we instead require
y
= Y
(8.91)
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 arbi-
trary 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 . (8.92)
Is there an analogue of Liouvilles theorem? If you choose H to be the harmonic
oscillator Hamiltonian, what would the phase space ow look like?
Problem 8.7 Consider the Lagrangian L =
m
2
( r
2
+ sinh
2
r

2
). In section
2.2 three conserved charges J
X
, J
Y
, and J
Z
are given, see eqs. (2.32). Evaluate the
Poisson brackets that these three charges enjoy.
Problem 8.8 Consider two Lagrangians related by a total derivative,
L
(q
i
, q
i
) = L(q
i
, q
i
) +
d
dt
(q
i
) =
1
2
q
i
q
i
V (q
i
) +
d
dt
(q
i
) , (8.93)
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?)
Problem 8.9 Show that the transformation
Q = ln
_
sin p
q
_
P = q cot p (8.94)
is canonical, and determine the generating functions F
1
and F
2
.
Problem 8.10 Find the canonical transformation generated by
F
1
= kq
2
cot Q , (8.95)
where k is a constant.
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 rst order coupled ordinary dierential equationsexists and is unique,
given the initial data, it is not clear that we can expect to actually nd these
solutions with a reasonable expense of eort. 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. (1.38) does not depend explicitly on timethe key
observation is that the ow lines in phase space never cross themselves. If we
ll a two dimensional plane (or some other two dimensional surface) with non-
intersecting ow 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 =
p
2
2
+V (q) , (9.1)
we obtain a dierential equation that determines the shape of the owlines:
_
_
_
q = p
p =
q
H
dp
dq
=
q
V
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? 103
where the energy E is an integration constant that labels the individual ow
lines. And this is almost the end of the story.
If the dimension of phase space is three things can become much more
complex. Once the ow 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 (6.56) the phase space coordinates
can be taken to be J
1
, J
2
, J
3
. We can solve the equations because any solution
J
i
(t) is a curve that lies on two well behaved surfaces
J
2
1
+J
2
2
+J
2
3
= constant ,
J
2
1
I
1
+
J
2
2
I
2
+
J
2
3
I
3
= 2E . (9.4)
Hence any solution curve is the intersection of a sphere and an ellipsoid, and
such a curve is not dicult 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
(1.44) things are very dierent: 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.
Note that for non-autonomous systems the evolution equations depend ex-
plicitly on t, as in eq. (1.37). This includes systems that are driven by some
time-dependent external force. For them it is no longer true that the phase
space ow lines are non-intersecting. In this case we expect that trouble will
arise already in two dimensional phase spaces.
Still, the Lorenz equations are not Hamiltonian. They cannot be, because
their phase space has odd dimension. Now Hamiltonian phase space ows have
some rather special properties, such as that uncovered by Liouvilles theorem
they are like the ow of an incompressible uid, and certain special kinds of
xed points cannot occur. Hence there is still some hope that the long time
behaviour of Hamiltonian systems may be signicantly simpler than that of
a typical dynamical system. This is what we will look into now. (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 ows that can occur we must be able to
picture them eectively. If the space to which the ow is conned is three
dimensionaleither because the phase space has three dimensions or because
the ow is conned to a three dimensional energy surface in a four dimensional
phase spacethe Poincare section solves the problem for us. The idea is to
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 denite 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 nite. 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 ow takes place on circles surrounding
a single xed point, and the idea is to introduce new canonical coordinates
(Q, P) such that P labels the ow line we are on, while Q tells us where we are
on the given ow 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
H =
1
2
p
2
+

2
2
q
2
, (9.5)
and the explicit solution of Hamiltons equations is given by
q = Asin (t +) p = Acos (t +) , (9.6)
9.2 Integrable systems 105
where the amplitude A is given in terms of the energy E by
E =

2
2
A
2
A =
2E
. (9.7)
We dene 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 ow lines in
phase space. Moreover, using the tricks introduced in section 8.4, a calculation
veries that
q, p = 2
P sin Q,
P cos Q = = Q, P . (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 equa-
tions of motion for n degrees of freedom. In the rst step we must nd n
constants of the motion I
i
, such that
I
i
, I
j
= 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
q
i
= q
i
(, I) p
i
= p
i
(, 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)

I
i
=
i
H = 0 . (9.15)
The new canonical variables I
i
and
i
are called action-angle variables. If such
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 conned to an n dimensional submanifold, dened by the n conditions I
i
=
constant. We also have available n vector elds
I
i
=
I
i
. (9.16)
The normal vector of any hypersurface dened by the equation I
i
= constant
is
n
i
I
i
, (9.17)
and it is clear that
I
i
, I
j
= 0 n
i
j
= 0 . (9.18)
That is to say that these vector elds are tangential to all the n hypersur-
faces, or in other words they represent n everywhere non-vanishing vector
elds pointing along the surface of the n dimensional submanifold dened by
the n equations I
i
= 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 eld. (The circle does, but the sphere does notyou cannot
comb a sphere.) Closer inspection shows that eq. (9.12) means that an ad-
ditional technical condition on these vector elds is obeyed, namely that the
vector elds 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 elds 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 conned 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 gure that never repeats, see section 1.3).
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
9.3 Canonical perturbation theory 107
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) L
3
and L
2
. We know that
L
i
, L
j
=
ijk
L
k
L
3
, L
2
= 0 . (9.19)
Therefore L
3
and L
2
can serve as action variables. Clearly this is reminiscent
of quantum mechanics, with its emphasis on commuting operators.
9.3 Canonical perturbation theory
The canonical transformation to action-angle variables will be dicult to nd
in general, and we will have to resort to canonical perturbation theory to con-
struct it order by order in some parameter. There is a well developed technol-
ogy for doing this. In 1969 it took us to the Moon.
The rst step then is to split the Hamiltonian function into two parts,
H(, I) = H
0
(I) +H
int
(, I) , (9.20)
where the split is dened in such a way that the equations of motion coming
from H
0
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. Con-
sider the case of a pendulum, with the Hamiltonian
H =
1
2
p
2
+

2
2
cos q . (9.21)
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 re-
gions separated by separatrices. The two separatrices, and the hyperbolic xed
point, all have the same energy E, so the conserved energy will not do as a co-
ordinate 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 innite amount of time to traverse it.
When the energy is large the eect of gravity is small (the pendulum is
rotating almost freely). Thus we can choose
H
0
=
1
2
p
2
, (9.22)
and treat as a small parameter. For small oscillations on the other hand we
can set
H
0
=
1
2
p
2
+

2
2
q
2
. (9.23)
The unperturbed system behaves as a harmonic oscillator. Supposing that we
obtain the solution as two dierent power series based on these two dierent
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
n
1
1
+n
2
2
= 0 (9.24)
for some integers n
1
, n
2
. Moreover the trouble is most severe if these integers
are small.
So are almost all Hamiltonian systems integrable? No. But if the Hamilto-
nian is of the form (9.20), with an that is not too large, then large regions
of phase space will still be lled 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 rst are those tori for
which the unperturbed motion obeys eq. (9.24) for small integers n
1
, n
2
. 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 dis-
tribution 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 func-
tion. Then, as time goes on, an asteroid whose trajectory was on a torus in
phase space will remain there, while an asteroid on a chaotic orbit sooner or
later changes its angular frequency signicantly. 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 aster-
oids with an
1
obeying eq. (9.24) for small values of n
1
and n
2
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.
Another beautiful illustration is provided by the Henon-Heiles Hamiltonian
H =
1
2
(p
2
x
+p
2
y
+x
2
+y
2
) +x
2
y
y
3
3
. (9.25)
It originally arose in the study of stars moving in the gravitational eld of
a galaxy, but here we treat it as a toy model. The inuence of the cubic
terms becomes more pronounced as the energy goes up, and in fact the energy
surface becomes unbounded above E = 1/6. 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 p
y
.
9.3 Canonical perturbation theory 109
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 conned 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 ac-
cording to an eight order canonical perturbation theory calculation performed
by Gustavson. Above E = 1/9, when the importance of the cubic terms is
larger, one nds orbits that denitely do not lie on tori since they ll 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.
Problem 9.1 Consider the Hamiltonian (9.1), and let V (q) be a fourth order
polynomial. Draw the phase space ow for all qualitatively dierent choices of this
polynomial. Pay special attention to the transition cases, when a maximum and a
minimum merge to produce an inection point on the graph of V (q).
Problem 9.2 Solve the evolution equation
x = x +x
2
(9.26)
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 conve-
nient to write the Henon-Heiles Hamiltonian as
H =
1
2
(p
2
x
+p
2
y
+x
2
+y
2
) +(x
2
y
y
3
3
) . (9.27)
Explain why small coupling constants correspond to small values of the total energy
in eq. (9.25). What is the nature of the energy surface when is small? When it is
large?
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 beau-
tiful 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.
You may also enjoy:
Julian Barbour: The End of Time, Weidenfeld and Nicolson 1999. A stim-
ulating account of philosophical issues raised by Newtons laws.
Ernst Mach: Die Mechanik in ihrer Entwickling, many editions (avoid the
ninth!). For the history of mechanics.
A. Dahan Dalmedico et al.: Chaos et determinisme,

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

Anmek Analytic Mech

Uploaded by

Copyright:

Available Formats

Anmek Analytic Mech

Uploaded by

Document Information

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 Analytic Mech

Uploaded by

Copyright:

Available Formats

Notes on Analytical Mechanics

Ingemar Bengtsson (2012)

= gml sin . (1.25)

). To see if the phase space ow has any

= gml sin = 0 . (1.42)

) = (0, 0) and (, 0). Linearizing around

(t) = q(t) + q(t), where the variation is of the

= t. Given a solution q(t) there is another

the period of the oscillations are unaected by the scaling,

LC. There are numerous useful analogies between electrical,

, that is to say with its index

. It is forbidden. There is no notion of length squared until we have

simply stands for the components of a ket relative to

) form a basis (and must not be confused with the com-

of a bra vector!). A matrix operating on the ket vector space will

. Then the equations

should remain unchanged. We can arrange this by insisting that whenever we

and the set

, and then dene the analogue of [).

. In fact we can be a bit more general. We introduce

, in the sense that

has a clear meaning.

is a special example of a metric tensor, while

is multiplied with 2, the

will be represented by the number of parallel planes pierced by the arrow.

is not dened yet. You might think that it

, how can we assign a number to it? We know the answer. We

is especially interesting, but in principle any

is allowed, as long as it is invertible and symmetric. Once we

= [[x[[ [[y[[ cos . (8.29)

, but in principle its components may

H. Let us relate them:

f. And it can be shown that this is

(t), where t is some parameter

is a covariant vector eld, then we dene the

are chosen as the basis in which we

. To see why we come back to eq. (8.35), which

is some covariant vector eld. In symplectic geometry it is known as

give rise to the same symplectic two-form. However, a possible

(with positive eigenvalues)

, in such a way that

becomes equal to a Kronecker delta. Show

You might also like