Advanced Classical Mechanics

Prof. Iain W. Stewart

Classical Mechanics III

MIT 8.09 & 8.309

Editors: Francisco Machado, Thanawuth Thanathibodee,

Prashanth S. Venkataram

c 2016 Iain W. Stewart

Contents

1 A Review of Analytical Mechanics 1

1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Lagrangian & Hamiltonian Mechanics . . . . . . . . . . . . . . . . . . . . . . 1
1.3 Symmetry and Conservation Laws . . . . . . . . . . . . . . . . . . . . . . . . 12
1.4 Constraints and Friction Forces . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.5 Calculus of Variations & Lagrange Multipliers . . . . . . . . . . . . . . . . . 22

2 Rigid Body Dynamics 29

2.1 Coordinates of a Rigid Body . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.2 Time Evolution with Rotating Coordinates . . . . . . . . . . . . . . . . . . . 33
2.3 Kinetic Energy, Angular Momentum, and
the Moment of Inertia Tensor for Rigid Bodies . . . . . . . . . . . . . . . . . 37
2.4 Euler Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
2.5 Symmetric Top with One Point Fixed . . . . . . . . . . . . . . . . . . . . . . 45

3 Vibrations & Oscillations 52

3.1 Simultaneous Diagonalization of T and V . . . . . . . . . . . . . . . . . . . . 53
3.2 Vibrations and Oscillations with Normal Coordinates . . . . . . . . . . . . . 55

4 Canonical Transformations, Hamilton-Jacobi Equations, and Action-Angle

Variables 60
4.1 Generating Functions for Canonical Transformations . . . . . . . . . . . . . 60
4.2 Poisson Brackets and the Symplectic Condition . . . . . . . . . . . . . . . . 66
4.3 Equations of Motion & Conservation Theorems . . . . . . . . . . . . . . . . 69
4.4 Hamilton-Jacobi Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
4.5 Kepler Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
4.6 Action-Angle Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83

5 Perturbation Theory 92
5.1 Time Dependent Perturbation Theory for the
Hamilton-Jacobi Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
5.2 Periodic and Secular Perturbations to Finite Angle Pendulum . . . . . . . . 95

ii
 CONTENTS

5.3 Perihelion Precession from Perturbing a Kepler Orbit . . . . . . . . . . . . . 98

6 Fluid Mechanics 104

6.1 Transitioning from Discrete Particles to the Continuum . . . . . . . . . . . . 104
6.2 Fluid Equations of Motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
6.2.1 Continuity Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
6.2.2 Ideal Fluid: Eulers Equation and Entropy Conservation . . . . . . . 112
6.2.3 Conservation of Momentum and Energy . . . . . . . . . . . . . . . . 113
6.3 Static Fluids & Steady Flows . . . . . . . . . . . . . . . . . . . . . . . . . . 115
6.4 Potential Flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
6.5 Sound Waves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
6.6 Viscous Fluid Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
6.7 Viscous Flows in Pipes and Reynolds Number . . . . . . . . . . . . . . . . . 130
6.8 Viscous Flow Past a Sphere (Stokes Flow) . . . . . . . . . . . . . . . . . . . 135

7 Chaos and Non-Linear Dynamics 142

7.1 Introduction to Chaos . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
7.1.1 Evolution of the system by first order differential equations . . . . . . 143
7.1.2 Evolution of Phase Space . . . . . . . . . . . . . . . . . . . . . . . . . 145
7.1.3 Fixed Points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
7.1.4 Picturing Trajectories in Phase Space . . . . . . . . . . . . . . . . . . 149
7.2 Bifurcations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
7.2.1 Saddle-Node Bifurcation . . . . . . . . . . . . . . . . . . . . . . . . . 154
7.2.2 Transcritical Bifurcation . . . . . . . . . . . . . . . . . . . . . . . . . 157
7.2.3 Supercritical Pitchfork Bifurcation . . . . . . . . . . . . . . . . . . . 159
7.2.4 Subcritical pitchfork bifurcation . . . . . . . . . . . . . . . . . . . . . 160
7.3 Fixed Points in Two-Dimensional Systems . . . . . . . . . . . . . . . . . . . 164
7.3.1 Motion Near a Fixed Point . . . . . . . . . . . . . . . . . . . . . . . . 164
7.3.2 Systems with a conserved E(~x) . . . . . . . . . . . . . . . . . . . . . 171
7.4 Limit Cycles and Bifurcations . . . . . . . . . . . . . . . . . . . . . . . . . . 173
7.4.1 Poincare-Bendixson Theorem . . . . . . . . . . . . . . . . . . . . . . 175
7.4.2 Fixed Point Bifurcations Revisited and Hopf Bifurcations . . . . . . . 176
7.5 Chaos in Maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
7.6 Chaos in Differential Equations, Strange Attractors, and Fractals . . . . . . 189
7.6.1 The Lorenz Equations . . . . . . . . . . . . . . . . . . . . . . . . . . 189
7.6.2 Fractals and the Connection to Lyapunov Exponents . . . . . . . . . 191
7.6.3 Chaos in Fluids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194

iii
Chapter 1

A Review of Analytical Mechanics

1.1 Introduction
These lecture notes cover the third course in Classical Mechanics, taught at MIT since
the Fall of 2012 by Professor Stewart to advanced undergraduates (course 8.09) as well as
to graduate students (course 8.309). In the prerequisite classical mechanics II course the
students are taught both Lagrangian and Hamiltonian dynamics, including Kepler bound
motion and central force scattering, and the basic ideas of canonical transformations. This
course briefly reviews the needed concepts, but assumes some familiarity with these ideas.
References used for this course include

Goldstein, Poole & Safko, Classical Mechanics, 3rd edition.

Landau and Lifshitz vol.6, Fluid Mechanics. Symon, Mechanics for reading material
on non-viscous fluids.

Strogatz, Nonlinear Dynamics and Chaos.

Review: Landau & Lifshitz vol.1, Mechanics. (Typically used for the prerequisite
Classical Mechanics II course and hence useful here for review)

1.2 Lagrangian & Hamiltonian Mechanics

Newtonian Mechanics
In Newtonian mechanics, the dynamics of a system of N particles are determined by solving
for their coordinate trajectories as a function of time. This can be done through the usual
vector spatial coordinates ri (t) for i {1, . . . , N }, or with generalized coordinates qi (t) for
i {1, . . . , 3N } in 3-dimensional space; generalized coordinates could be angles, et cetera.

1
 CHAPTER 1. A REVIEW OF ANALYTICAL MECHANICS

Velocities are represented through vi ri for spatial coordinates, or through qi for

generalized coordinates. Note that dots above a symbol will always denote the total time
derivative dtd . Momenta are likewise either Newtonian pi = mi vi or generalized pi .
For a fixed set of masses mi Newtons 2nd law can be expressed in 2 equivalent ways:
1. It can be expressed as N second-order equations Fi = dtd (mi ri ) with 2N boundary
conditions given in ri (0) and ri (0). The problem then becomes one of determining the
N vector variables ri (t).
2. It can also be expressed as an equivalent set of 2N 1st order equations Fi = pi &
pi /mi = ri with 2N boundary conditions given in ri (0) and pi (0). The problem then
becomes one of determining the 2N vector variables ri (t) and pi (t).
Note that F = ma holds in inertial frames. These are frames where the motion of a
particle not subject to forces is in a straight line with constant velocity. The converse does not
hold. Inertial frames describe time and space homogeneously (invariant to displacements),
isotropically (invariant to rotations), and in a time independent manner. Noninertial frames
also generically have fictitious forces, such as the centrifugal and Coriolis effects. (Inertial
frames also play a key role in special relativity. In general relativity the concept of inertial
frames is replaced by that of geodesic motion.)
The existence of an inertial frame is a useful approximation for working out the dynam-
ics of particles, and non-inertial terms can often be included as perturbative corrections.
Examples of approximate inertial frames are that of a fixed Earth, or better yet, of fixed
stars. We can still test for how noninertial we are by looking for fictitious forces that (a) may
point back to an origin with no source for the force or (b) behave in a non-standard fashion
in different frames (i.e. they transform in a strange manner when going between different
frames).
We will use primes will denote coordinate transformations. If r is measured in an inertial
frame S, and r0 is measured in frame S 0 with relation to S by a transformation r0 = f (r, t),
then S 0 is inertial iff r = 0 r0 = 0. This is solved by the Galilean transformations,
r0 = r + v0 t
t0 = t,
which preserves the inertiality of frames, with F = mr and F0 = mr0 implying each other.
Galilean transformations are the non-relativistic limit, v  c, of Lorentz transformations
which preserve inertial frames in special relativity. A few examples related to the concepts
of inertial frames are:
1. In a rotating frame, the transformation is given by
 0   
x cos() sin() x
=
y0 sin() cos() y
If = t for some constant , then r = 0 still gives r0 6= 0, so the primed frame is
noninertial.

2
 CHAPTER 1. A REVIEW OF ANALYTICAL MECHANICS

Figure 1.1: Frame rotated by an angle

2. In polar coordinates, r = rr, gives

dr d
= , = r (1.1)
dt dt
and thus  
+ r
r = rr + 2r 2 r . (1.2)

Even if r = 0 we can still have r =6 0 and 6= 0, and we can not in general form
a simple Newtonian force law equation mq = Fq for each of these coordinates. This
is different than the first example, since here we are picking coordinates rather than
changing the reference frame, so to remind ourselves about their behavior we will call
these non-inertial coordinates (which we may for example decide to use in an inertial
frame). In general, curvilinear coordinates are non-inertial.

Lagrangian Mechanics
In Lagrangian mechanics, the key function is the Lagrangian
L = L(q, q, t). (1.3)
Here, q = (q1 , . . . , qN ) and likewise q = (q1 , . . . , qN ). We are now letting N denote the
number of scalar (rather than vector) variables, and will often use the short form to denote
dependence on these variables, as in Eq. (1.3). Typically we can write L = T V where
T is the kinetic energy and V is the potential energy. In the simplest cases, T = T (q)
and V = V (q), but we also allow the more general possibility that T = T (q, q, t) and
V = V (q, q, t). It turns out, as we will discuss later, that even this generalization does not
describe all possible classical mechanics problems.
The solution to a given mechanical problem is obtained by solving a set of N second-order
differential equations known as Euler-Lagrange equations of motion,
 
d L L
= 0. (1.4)
dt qi qi

3
 CHAPTER 1. A REVIEW OF ANALYTICAL MECHANICS

These equations involve qi , and reproduce the Newtonian equations F = ma. The principle
of stationary action (Hamiltons principle),
Z t2
S = L(q, q, t) dt = 0, (1.5)
t1

is the starting point for deriving the Euler-Lagrange equations. Although you have covered
the Calculus of Variations in an earlier course on Classical Mechanics, we will review the
main ideas in Section 1.5.
There are several advantages to working with the Lagrangian formulation, including
1. It is easier to work with the scalars T and V rather than vectors like F.

2. The same formula in equation (1.4) holds true regardless of the choice of coordinates.
To demonstrate this, let us consider new coordinates

Qi = Qi (q1 , . . . , qN , t). (1.6)

This particular sort of transformation is called a point transformation. Defining the

new Lagrangian by
t) = L(q, q, t),
L0 = L0 (Q, Q, (1.7)
we claim that the equations of motion are simply
d L0 L0
 
= 0. (1.8)
dt Q i Qi

Proof: (for N = 1, since the generalization is straightforward)

t) = L(q, q, t) with Q = Q(q, t) then
Given L0 (Q, Q,
d Q Q
Q = Q(q, t) = q + . (1.9)
dt q t

Therefore

Q Q
= , (1.10)
q q

a result that we will use again in the future. Then

L L0 L0 Q L0 Q
= = + , (1.11)
q q Q q Q q
L L0 L0 Q L0 Q
= = = .
q q Q q Q q

4
 CHAPTER 1. A REVIEW OF ANALYTICAL MECHANICS

Q L0 Q
Since q
= 0 there is no termQ q 
in the last line.
Plugging these results into 0 = dtd qL L
q
gives

d L0 Q L0 d Q L Q L0 Q
     0 
0= + +
dt Q q Q dt q Q q Q q
  0
L0 Q

d L
= , (1.12)
dt Q Q q

since dtd Q
q

= (q q Q
+ t
) q = q
(q q
+ t )Q = Q
q
so that the second and fourth terms
cancel. Finally for non-trivial transformation where Q q
6= 0 we have, as expected,

L0 L0
 
d
0= . (1.13)
dt Q Q

Note two things:

This implies we can freely use the Euler-Lagrange equations for noninertial coor-
dinates.
We can formulate L in whatever coordinates are easiest, and then change to
convenient variables that better describe the symmetry of a system (for example,
Cartesian to spherical).

3. Continuing our list of advantages for using L, we note that it is also easy to incorporate
constraints. Examples include a mass constrained to a surface or a disk rolling without
slipping. Often when using L we can avoid discussing forces of constraint (for example,
the force normal to the surface).

Lets discuss the last point in more detail (we will also continue to discuss it in the next
section). The method for many problems with constraints is to simply make a good choice for
the generalized coordinates to use for the Lagrangian, picking N k independent variables
qi for a system with k constraints.
Example: For a bead on a helix as in Fig. 1.2 we only need one variable, q1 = z.
Example: A mass m2 attached by a massless pendulum to a horizontally sliding mass m1
as in Fig. 1.3, can be described with two variables q1 = x and q2 = .
Example: As an example using non-inertial coordinates consider a potential V = V (r, )
in polar coordinates for
a fixed mass m at position r = rr. Since r = rr + r we have
T = m2 r2 = m2 r2 + r2 2 , giving

m 2 
L= r + r2 2 V (r, ). (1.14)
2
5
 CHAPTER 1. A REVIEW OF ANALYTICAL MECHANICS

Figure 1.2: Bead on a helix

Figure 1.3: Pendulum of mass m2 hanging on a rigid bar of length ` whose support m1 is a
frictionless horizontally sliding bead

For r the Euler-Lagrange equation is

 
d L L d V
0= = (mr) mr2 + . (1.15)
dt r r dt r

This gives
V
mr mr2 = = Fr , (1.16)
r
6 mr. For the Euler-Lagrange equation is
from which we see that Fr =
 
d L L d  2  V
0= = mr + . (1.17)
dt dt

This gives
d  2  V
mr = = F , (1.18)
dt
which is equivalent to the relation between angular momentum and torque perpendicular to
the plane, L z = F = z . (Recall L = r p and = r F.)

6
 CHAPTER 1. A REVIEW OF ANALYTICAL MECHANICS

Figure 1.4: Particle on the inside of a cone

Example: Let us consider a particle rolling due to gravity in a frictionless cone, shown
in
p Fig. 1.4, whose opening angle defines an equation for points on the cone tan() =
x2 + y 2 /z. There are 4 steps which we can take to solve this problem (which are more
general than this example):
1. Formulate T and V by N = 3 generalized coordinates. Here it is  most convenient
 to
m 2 2 2 2
choose cylindrical coordinates denoted (r, , z), so that T = 2 r + r + z and
V = mgz.

2. Reduce the problem to N k = 2 independent coordinates and determine the new

Lagrangian L = T V . In this case we eliminate z = r cot() and z = r cot(), so
mh i
1 + cot2 r2 + r2 2 mgr cot .


L= (1.19)
2

3. Find the Euler-Lagrange equations. For r, 0 = dtd rL rL , which here is

d 
m 1 + cot2 r mr2 + mg cot


0= (1.20)
dt
giving
1 + cot2 r r2 + g cot = 0.


(1.21)
 
d L L
For we have 0 = dt

, so

d  2 
0= mr 0, (1.22)
dt
giving
(2r + r)r
= 0. (1.23)

4. Solve the system analytically or numerically, for example using Mathematica. Or we

might be only interested in determining certain properties or characteristics of the
motion without a full solution.

7
 CHAPTER 1. A REVIEW OF ANALYTICAL MECHANICS

Hamiltonian Mechanics
L
In Hamiltonian mechanics, the canonical momenta pi q i
are promoted to coordinates
on equal footing with the generalized coordinates qi . The coordinates (q, p) are canonical
variables, and the space of canonical variables is known as phase space.
L
The Euler-Lagrange equations say pi = q i
. These need not equal the kinematic momenta
mi qi if V = V (q, q). Performing the Legendre transformation

H(q, p, t) = qi pi L(q, q, t) (1.24)

(where for this equation, and henceforth, repeated indices will imply a sum unless otherwise
specified) yields the Hamilton equations of motion

H
qi = (1.25)
pi
H
pi =
qi

which are 2N 1st order equations. We also have the result that
H L
= . (1.26)
t t
Proof: (for N = 1) Consider

H H H
dH = dq + dp + dt (1.27)
q p t
L L L
= pdq + qdp dq dq dt . (1.28)
q q t
L L
Since we are free to independently vary dq, dp, and dt this implies q
= p, q
= p, and
H
t
= L
t
.
We can interpret the two Hamilton equations as follows:
H L
qi = pi
is an inversion of pi = qi
= pi (q, q, t).

pi = H
qi
provides the Newtonian dynamics.

However, these two equation have an have equal footing in Hamiltonian mechanics, since
L
the coordinates and momenta are treated on a common ground. We can use pi = q i
to
construct H from L and then forget about L.

8
 CHAPTER 1. A REVIEW OF ANALYTICAL MECHANICS

As an example of the manner in which we will usually consider transformations between

Lagrangians and Hamiltonians, consider again the variables relevant for the particle on a
cone from Fig. 1.4:

z) z=r cot
L(r, , z, r, , new L(r, , r, ) Euler-Lagrange Eqtns. (1.29)
l l l
not here
H(r, , z, pr , p , pz ) = H(r, , pr , p ) Hamilton Eqtns.

Here we consider transforming between L and H either before or after removing the redun-
dant coordinate z, but in this course we will only consider constraints imposed on Lagrangians
and not in the Hamiltonian formalism (the step indicated by =). For the curious, the topic
of imposing constraints on Hamiltonians, including even more general constraints than those
we will consider, is covered well in Diracs little book Lectures on Quantum Mechanics.
Although Hamiltonian and Lagrangian mechanics provide equivalent formalisms, there is of-
ten an advantage to using one or the other. In the case of Hamiltonian mechanics potential
advantages include the language of phase space with Liouvilles Theorem, Poisson Brackets
and the connection to quantum mechanics, as well as the Hamilton-Jacobi transformation
theory (all to be covered later on).

Special case: Let us consider a special case that is sufficient to imply that the Hamiltonian
is equal to the energy, H = E T + V . If we only have quadratic dependence on velocities
in the kinetic energy, T = 21 Tjk (q)qj qk , and V = V (q) with L = T V , then

L 1 1
qi pi = qi = qi T ikqk + qj Tji qi = 2T. (1.30)
qi 2 2
Hence,
H = qi pi L = T + V = E (1.31)
which is just the energy.

Another Special case: Consider a class of Lagrangians given as

1
L(q, q, t) = L0 + aj qj + qj Tjk qk (1.32)
2
where L0 = L0 (q, t), aj = aj (q, t), and Tjk = Tkj = Tjk (q, t). We can write this in shorthand
as
1
L = L0 + ~a ~q + ~q T ~q . (1.33)
2
Here the generalized coordinates, momenta, and coefficients have been collapsed into vectors,
like ~q (rather than the boldface that we reserve for Cartesian vectors), and dot products of

9
 CHAPTER 1. A REVIEW OF ANALYTICAL MECHANICS

vectors from the left imply transposition of that vector. Note that ~q is an unusual vector,
since its components can have different dimensions, eg. ~q = (x, ), but nevertheless this
notation is useful. To find H,
L
pj = = aj + Tjk qk , (1.34)
qj
meaning p~ = ~a + T ~q. Inverting this gives ~q = T1 (p~ ~a), where T1 will exist because
of the positive-definite nature of kinetic energy, which implies that T is a postive definite
matrix. Thus, H = ~q p~ L yields
1
H = (p~ ~a) T1 (p~ ~a) L0 (q, t) (1.35)
2
as the Hamiltonian. So for any Lagrangian in the form of Eq. (1.32), we can find T1 and
write down the Hamiltonian as in Eq. (1.35) immediately.

Example: let us consider L = 21 mv2 e + eA v, where e is the electric charge and SI

units are used. In Eq. (1.32), because the coordinates are Cartesian, a = eA, T = m1, and
L0 = e, so
1
H= (p eA)2 + e . (1.36)
2m
As you have presumably seen in an earlier course, this Hamiltonian does indeed reproduce
the Lorentz force equation e(E + v B) = mv.

A more detailed Example. Find L and H for the frictionless pendulum shown in Fig. 1.3.
This system has two constraints, that m1 is restricted to lie on the x-axis sliding without
friction, and that the rod between m1 and m2 is rigid, giving
y1 = 0 , (y1 y2 )2 + (x1 x2 )2 = `2 . (1.37)
Prior to imposing any constraints the Lagrangian is
m1 2 m2 2
L=T V = x + (x + y22 ) m2 gy2 m1 gy1 . (1.38)
2 1 2 2
Lets choose to use x x1 and the angle as the independent coordinates after imposing
the constraints in Eq. (1.37). This allows us to eliminate y1 = 0, x2 = x + ` sin and
y2 = ` cos , together with x2 = x + ` cos , y2 = ` sin ,
x1 = x. The Lagrangian with
constraints imposed is
m1 2 m2 2
x + 2` cos x + `2 cos2 2 + `2 sin2 2 + m2 g` cos .


L= x + (1.39)
2 2
Next we determine the Hamiltonian. First we find
L = (m1 + m2 )x + m2 ` cos ,
px = = m1 x + m2 (x + ` cos ) (1.40)
x
L
p = = m2 ` cos x + m2 `2 .

10
 CHAPTER 1. A REVIEW OF ANALYTICAL MECHANICS

Note that px is notsimply pr oportional

 to x here (actually px is the center-of-mass momen-
px x
tum). Writing = T gives
p
 
m1 + m2 m2 ` cos
T = , (1.41)
m2 ` cos m2 `2
 
1 x
with L = (x ) T + L0 where L0 = m2 g` cos . Computing
2
 
1 1 m2 `2 m2 ` cos
T = , (1.42)
m1 m2 `2 + m2 `2 sin2 m2 ` cos m1 + m2
we can simply apply Eq. (1.35) to find the corresponding Hamiltonian
 
1 1 px
H = (px p ) T m2 g` cos (1.43)
2 p
1 h
2 2 2
i
= m2 ` px + (m1 + m2 )p 2m2 ` cos px p m2 g` cos .
2m2 `2 (m1 + m2 sin2 )
Lets compute the Hamilton equations of motion for this system. First for (x, px ) we find
H px cos p
x = = , (1.44)
px m1 + m2 sin `(m1 + m2 sin2 )
2

H
px = = 0.
x
As we might expect, the CM momentum is time independent. Next for (, p ):
H 1 h i
= = (m 1 + m 2 )p m 2 ` cos p x , (1.45)
p m2 `2 (m1 + m2 sin2 )
H sin cos h
2 2 2
i
p = = 2 m2 ` p x + (m 1 + m 2 )p 2m 2 ` cos p x p
` (m1 + m2 sin2 )2
sin px p
m2 g` sin .
`(m1 + m2 sin )
These non-linear coupled equations are quite complicated, but could be solved in math-
ematica or another numerical package. To test our results for these equations of motion
analytically, we can take the small angle limit, approximating sin ' , cos ' 1 to obtain
px p 1 h i
x = , px = 0 , = ( m1 + m )p
2 m2 x ,
`p
m1 `m1 m1 m2 `2
h 2 2 2
i p p
x
p = 2 2 m2 ` px + (m1 + m2 )p 2m2 ` cos px p m2 g` . (1.46)
` m1 `m1

11
 CHAPTER 1. A REVIEW OF ANALYTICAL MECHANICS

To simplify it further we can work in the CM frame, thus setting px = 0, and linearize the
equations by noting that p should be small for to remain small, and hence p2 is a
higher order term. For the non-trivial equations this leaves
p p
x = , = 2 , p = m2 g` , (1.47)
`m1 `

where = m1 m2 /(m1 + m2 ) is the reduced mass for the two-body system. Thus =
p /(`2 ) = m2`g as expected for simple harmonic motion.

1.3 Symmetry and Conservation Laws

A cyclic coordinate is one which does not appear in the Lagrangian, or equivalently in the
Hamiltonian. Because H(q, p, t) = qi pi L(q, q, t), if qj is absent in L for some particular j,
it will be absent in H as well. The absence of that qj corresponds with a symmetry in the
dynamics.
In this context, Noethers theorem means that a symmetry implies a cyclic coordinate,
which in turn produces a conservation law. If qj is a cyclic coordinate for some j, then
we can change that coordinate without changing the dynamics given by the Lagrangian
or Hamiltonian, and hence there is a symmetry. Furthermore the corresponding canonical
momentum pj is conserved, meaning it is a constant through time.
L
The proof is simple. If q j
= 0 then pj = dtd q
L
j
L
= qj
H
= 0, or even more simply, q j
=0
is equivalent to pj = 0, so pj is a constant in time.
Special cases and examples of this abound. Lets consider a few important ones:
1. Consider a system of N particles where no external or internal force acts
P on the center
1
of mass (CM) coordinate R = M mi ri , where the total mass M = i mi . Then the
CM momentum P is conserved. This is because

FR = R V = 0 (1.48)

so V is independent of R. Meanwhile, T = 21 i mi ri2 , which when using coordinates

P
relative to the center of mass, r0i ri R, becomes
! !
1 X +R d X
0 1X 1 2 1X
T = mi R 2
mi ri + mi r02
i = MR + mi r02
i . (1.49)
2 i
dt i
2 i 2 2 i

Note that i mi r0i = 0 from the definitions of M , R, and r0i , so T splits into two
P
terms, one for the CM motion and one for relative motion. We also observe that T is
independent of R. This means that R is cyclic for the full Lagrangian L, so P = M R
is a conserved quantity. In our study of rigid bodies we will also need the forms of M
and R for a conRtinuous body with massR distribution (r), which for a three dimensional
body are M = d3 r (r) and R = M1 d3 r (r) r.

12
 CHAPTER 1. A REVIEW OF ANALYTICAL MECHANICS

Note that P = 0 is satisfied by having no total external force, so Fext = i Fext

P
i =
rd
0, and by the internal forces obeying Newtons 3 law Fij = Fji . Hence,
X X
MR = Fext
i + Fij = 0. (1.50)
i i,j

2. Let us consider a system that is invariant with respect to rotations of angle about a
symmetry axis. This has a conserved angular momentum. If we pick as a generalized
coordinate, then L = T V is independent of , so p = L

= 0 meaning p is constant.
In particular, for a system where V is independent of the angular velocity we have
T X ri X ri
p = = mi ri = mi vi . (1.51)
i
i

Simplifying further using the results in Fig. 2.2 yields

X X
p = mi vi (n ri ) = n ri mi vi = n Ltotal . (1.52)
i i

Figure 1.5: Rotation about a symmetry axis

Note
P that L about the CM is conserved for systems with no external torque,
ext ext
= i ri Fi = 0 and internal forces that are all central. Defining rij ri rj and
its magnitude appropriately, this means Vij = Vij (rij ). This implies that Fji = i Vij
(no sum on the repeated index) is parallel to rij . Hence,
dL X X X
= ri pi = ri Fext
i + ri Fji . (1.53)
dt i i i,j

13
 CHAPTER 1. A REVIEW OF ANALYTICAL MECHANICS

ri Fext
P
However, i i = 0, so
dL X
= rij Fji = 0. (1.54)
dt i<j
3. One can also consider a scaling transformation. Suppose that under the transformation
ri ri the potential is homogeneous and transforms as V k V for some constant
k. Letting T be quadratic in ri and taking time to transform as t 1k/2 t then
gives ri k/2 ri . So by construction T k T also, and thus the full Lagrangian
L k L. This overall factor does not change the Euler-Lagrange equations, and hence
the transformation is a symmetry of the dynamics, only changing the overall scale or
units of the coordinate and time variables, but not their dynamical relationship. This
can be applied for several well known potentials:

a) k = 2 for a harmonic oscillator. Here the scaling for time is given by 1 k/2 = 0,
so it does not change with . Thus, the frequency of the oscillator, which is a
time variable, is independent of the amplitude.
b) k = 1 for the Coulomb potential. Here 1k/2 = 3/2 so there is a more intricate
relation between coordinates and time. This power is consistent with the behavior
of bound state orbits, where the period of the orbit T obeys T 2 a3 , for a the
semi-major axis distance (Keplers 3rd law).
Here 1 k/2 = 1/2 so for a freely falling
c) k = 1 for a uniform gravitational field.
object, the time of free fall goes as h where h is the distance fallen.

4. Consider the Lagrangian for a charge in electromagnetic fields, L = 21 mr2 e + eA r.

As a concrete example, let us take and A to be independent of the Cartesian coor-
dinate x. The canonical momentum is p = L r
= mr + eA, which is notably different
from the kinetic momentum. Then x being cyclic means the canonical momentum px
is conserved.

5. Let us consider the conservation of energy and the relationship between energy and
the Hamiltonian. Applying the time derivative gives H = H
q
q + H
p
p + H
t
. However,
q = H
p
and p = H
q
. Thus

H L
H = = . (1.55)
t t
There are two things to consider.

If H (or L) has no explicit time dependence, then H = qi pi L is conserved.

Energy is conserved if E = 0, where energy is defined by E = T + V .

If H = E then the two points are equivalent, but otherwise either of the two could be
true while the other is false.

14
 CHAPTER 1. A REVIEW OF ANALYTICAL MECHANICS

Example: Let us consider a system which provides an example where H = E but

energy is not conserved, and where H = 6 E but H is conserved. The two situations
will be obtained from the same example by exploiting a coordinate choice. Consider a
system consisting of a mass m attached by a spring of constant k to a cart moving at a
constant speed v0 in one dimension, as shown in Fig. 1.6. Let us call x the displacement

Figure 1.6: Mass attached by a spring to a moving cart

of m from the fixed wall and x0 is its displacement from the center of the moving cart.
Using x,
m k
L(x, x) = T V = x2 (x v0 t)2 , (1.56)
2 2
where the kinetic term is quadratic in x and the potential term is independent of x.
This means that H falls in the special case considered in Eq. (1.31) so
p2 k
H =E =T +V = + (x v0 t)2 , (1.57)
2m 2
However Ht
6= 0 so the energy is not conserved. (Of course the full energy would
be conserved, but we have not accounted for the energy needed to pull the cart at a
constant velocity, treating that instead as external to our system. That is what led to
the time dependent H.)
If we instead choose to use the coordinate x0 = x v0 t, then
m 02 m k
L0 (x0 , x0 ) = x + mv0 x0 + v02 x02 . (1.58)
2 2 2
Note that p0 = mx0 +mv0 = mx = p. This Lagrangian fits the general form in equation
(1.32) with a = mv0 and L0 = mv02 /2 kx02 /2. So
1 2 k m
H 0 (x0 , p0 ) = x0 p0 L0 = (p0 mv0 ) + x02 v02 , (1.59)
2m 2 2
15
 CHAPTER 1. A REVIEW OF ANALYTICAL MECHANICS

Here the last terms is a constant shift. The first and second terms in this expression for
H 0 look kind of like the energy that we would calculate if we were sitting on the cart
and did not know it was moving, which is not the same as the energy above. Hence,
0
H 0 6= E, but H 0 = 0 because H
t
= 0, so H 0 is conserved.

1.4 Constraints and Friction Forces

So far, weve considered constraints to a surface or curve that are relationships between
coordinates. These fall in the category of holonomic constraints. Such constraints take the
form
f (q1 , . . . , qN , t) = 0 (1.60)
where explicit time dependence is allowed as a possibility. An example of holonomic constrain
is mass in a cone (Figure 1.4), where the constrain is z r cot = 0. Constraints that violate
the form in Eq. (1.60) are non-holonomic constraints.
An example of a non-holonomic constraint is a mass on the surface of a sphere. The

Figure 1.7: Mass on a sphere

constraint here is an inequality r2 a2 0 where r is the radial coordinate and a is

the radius of the sphere.

Another example of a non-holonomic constraint is an object rolling on a surface with-

out slipping. The point of contact is stationary, so the constraint is actually on the
velocities.
A simple example is a disk of radius a rolling down an inclined plane without slipping,
as shown in Fig. 1.8. Here the condition on velocities, a = x is simple enough that it can
be integrated into a holonomic constraint.
As a more sophisticated example, consider a vertical disk of radius a rolling on a horizontal
plane, as shown in Fig. 1.9. The coordinates are (x, y, , ), where (x, y) is the point of
contact, is the rotation angle about its own axis, and is the angle of orientation along the
xy-plane. We will assume that the flat edge of the disk always remain parallel to z, so the

16
 CHAPTER 1. A REVIEW OF ANALYTICAL MECHANICS

Figure 1.8: Disk rolling down an incline without slipping

Figure 1.9: Vertical rolling disk on a two dimensional plane

disk never tips over. The no-slip condition is v = a where v is the velocity of the center of
the disk, and v = |v|. This means x = v sin() = a sin() and y = v cos() = a cos(),
or in differential notation, dx a sin()d = 0 and dy + a cos()d = 0.
In general, constraints of the form
X
aj (q)dqj + at (q)dt = 0 (1.61)
j

are not holonomic. We will call this a semi-holonomic constraint, following the terminology
of Goldstein.
Let us consider the special case of a holonomic constraint in differential form, f (q1 , ..., q3N , t) =
0. This means
X f f
df = dqj + dt = 0, (1.62)
j
qj t

17
 CHAPTER 1. A REVIEW OF ANALYTICAL MECHANICS

f f
so aj = qj
and at = t
. The symmetry of mixed partial derivatives means

aj ai at ai
= , = . (1.63)
qi qj qi t

These conditions imply that a seemingly semi-holonomic constraint is in fact holonomic. (In
math we would say that we have an exact differential form df for the holonomic case, but
the differential form in Eq.(1.61) need not always be exact.)

Example: To demonstrate that not all semiholonomic constrants are secretly holo-
nomic, consider the constraint in the example of the vertical disk. Here there is no func-
tion h(x, y, , ) that we can multiply the constraint df = 0 by to make it holonomic.
For the vertical disk from before, we could try (dx a sin() d)h = 0 with ax = h,
a
a = a sin()h, a = 0, and ay = 0 all for some function h. As we must have = a

,
h k ax a h
then 0 = a cos() a sin() , so h = sin() . That said, = x gives = 0 which is a
contradiction for a non-trivial h with k 6= 0.
If the rolling is instead constrained to a line rather than a plane, then the constraint is
holonomic. Take as an example = 2 for rolling along x, then x = a and y = 0. Integrating
we have x = a + x0 , y = y0 , and = 2 , which together form a set of holonomic constraints.

A useful concept for discussing constraints is that of the virtual displacement ri of

particle i. There are a few properties to be noted of ri .

It is infinitesimal.

It is consistent with the constraints.

It is carried out at a fixed time (so time dependent constraints do not change its form).

Example: let us consider a bead constrained to a moving wire. The wire is oriented along

Figure 1.10: Bead on a moving wire

the x-axis and is moving with coordinate y = v0 t. Here the virtual displacement of the

18
 CHAPTER 1. A REVIEW OF ANALYTICAL MECHANICS

bead r is always parallel to x (since it is determined at a fixed time), whereas the real
displacement dr has a component along y in a time interval dt.

For a large number of constraints, the constraint force Zi is Rperpendicular to ri , meaning

Zi ri = 0, so the virtual work (in analogy to work W = F dr) Pof a constraint force
vanishes. More generally, there is no net work from constraints, so i Zi ri = 0 (which
holds for the actions of surfaces, rolling constraints, and similar things). The Newtonian
equation of motion is pi = Fi + Zi , where Fi encapsulates other forces. Vanishing virtual
work gives X
(pi Fi ) ri = 0 (1.64)
i
which is the DAlembert principle. This could be taken as the starting principal for classical
mechanics instead of the Hamilton principle of stationary action.
Of course Eq.(1.64) is not fully satisfactory since we are now used to the idea of working
with generalized coordinates rather than the cartesian vector coordinates used there. So lets
ri
transform to generalized coordinates through ri = ri (q, t), so ri = q j
qj , where again we
sum over repeated indicies (like j here). This means
ri
Fi ri = Fi qj Qj qj (1.65)
qj
where we have defined generalized forces
ri
Qj Fi . (1.66)
qj
We can also transform  the pi ri term using our earlier point transformation results as well
2 ri
d ri
P 2 ri vi
as the fact that dt qj = q j t
+ k qj qk qk = qj . Writing out the index sums explicitly,
this gives
X X ri
pi ri = mi ri qj
i i,j
qj
X d  ri
 
d ri

= mi ri mi ri qj
i,j
dt q j dt q j

X d  vi

vi

= mi vi mi vi qj
i,j
dt qj qj
X  d  T  T 
= qj (1.67)
j
dt q j q j

for T = 12 i mi vi2 . Together with the DAlembert principle, we obtain the final result
P

X  d  T  T 
Qj qj = 0. (1.68)
j
dt qj qj

19
 CHAPTER 1. A REVIEW OF ANALYTICAL MECHANICS

We will see momentarily that this result is somewhat more general than the Euler-Lagrange
equations, containing them as a special case.
We will start by considering systems with only holonomic constraints, postponing other
types of constraints to the next section. Here we can find the independent coordinates qj
with j = 1, . . . , N k that satisfy the k constraints. This implies that the generalized virtual
displacements qj are independent, so that their coefficients in Eq. (1.68) must vanish,
 
d T T
Qj = 0 . (1.69)
dt qj qj

There are several special cases of this result, which we derived from the dAlembert principle.

1. For a conservative force Fi = i V , then

ri V
Qj = (i V ) = (1.70)
qj qj

where we assume that the potential can be expressed in the generalized coordinates
as V = V (q, t). Then using L = T V , we see that Eq. (1.69) simply reproduces the
d L L
Euler-Lagrange equations dt qj q j
= 0.
 
V d V
2. If Qj = qj
+ dt
for V = V (q, q, t), which is the case for velocity dependent
qj
forces derivable from a potential
  (like the electromagnetic Lorentz force), then the
d L L
Euler-Lagrange equations dt qj
qj
= 0 are again reproduced.

3. If Qj has forces obtainable from a potential as in case 2, as well as generalized forces

Rj that cannot, then  
d L L
= Rj (1.71)
dt qj qj
is the generalization of the Euler-Lagrange equations with non-conservative generalized
forces.

An important example of a nonconservative forces Rj is given by friction.

Static friction is Fs Fsmax = s FN for a normal force FN .

Sliding friction is F = FN vv , so this is a constant force that is always opposite the

direction of motion (but vanishes when there is no motion).

Rolling friction is F = R FN vv .

Fluid friction at a low velocity is F = bv vv = bv.

20
 CHAPTER 1. A REVIEW OF ANALYTICAL MECHANICS

A general form for a friction force is Fi = hi (vi ) vvii (where as a reminder there is no
implicit sum on i here since we specified i on the right-hand-side). For this form
X vi rj X vi vi
Rj = hi = hi . (1.72)
i
vi qj i
vi qj

Simplifying further gives

X hi X vi X vi Z vi X vi 0
Z
0 0
2
dvi hi (vi0 )


Rj = vi = hi = dvi hi (vi ) =
i
2vi q j i
q j i
q j v i 0 q j i 0

F
= (1.73)
qj

where vi
XZ
F= dvi0 hi (vi0 ) (1.74)
i 0

is the dissipation function. This is a scalar function like L so it is relatively easy to work
with.

Example: Consider a sphere of radius a and mass m falling in a viscous fluid. Then
T = 12 m0 y 2 where m0 < m accounts for the mass of displaced fluid (recall Archimedes princi-
ple that the buoyant force on a body is equal to the weight of fluid the body displaces). Also
V = m0 gy, and L = T V . Here h y, so F = 3ay 2 , where by the constant
 of proportion-
ality is determined by the constant , which is the viscosity. From this, dt yL yL = F
d
y
gives the equation of motion m0 y + m0 g = 6ay. The friction force 6ay is known as
Stokes Law. (We will derive this equation for the friction force from first principles later
on, in our discussion of fluids.) This differential equation can be solved by adding a par-
ticular solution yp (t) to a solution of the homogeneous equation m0 yH + 6ayH = 0. For
the time derivatives the results are yp = m0 g/(6a) and yH = A exp(6at/m0 ), where
the constant A must be determined by an initial condition. The result y = yH + yp can be
integrated in time once more to obtain the full solution y(t) for the motion.

Example: if we add sliding friction to the case of two masses on a plane connected by
a spring (considered on problem set #1), then hi = f mi g for some friction coefficient f ,
and  q q 
F = f g(m1 v1 + m2 v2 ) = f g m1 x21 + y12 + m2 x22 + y22 . (1.75)

If we switch to a suitable set of generalized coordinates qj that simplify the equations of

F
motion without friction, and then compute the generalized friction forces Rj = q j
, we can
get the equations of motion including friction. Further details of how this friction complicates
the equations of motion were provided in lecture.

21
 CHAPTER 1. A REVIEW OF ANALYTICAL MECHANICS

1.5 Calculus of Variations & Lagrange Multipliers

Calculus of Variations
In the calculus of variations, we wish to find a set of functions yi (s) between s1 and s2 that
extremize the following functional (a function of functions),
Z s2
J[yi ] = ds f (y1 (s), . . . , yn (s), y1 (s), . . . , yn (s), s) , (1.76)
s1

where for this general discussion only we let yi dy

ds
i
rather than dtd . To consider the action of
the functional under a variation we consider yi0 (s) = yi (s) + i (s) where i (s1 ) = i (s2 ) = 0,
meaning that while the two endpoints are fixed during the variation yi = i , the path in
between is varied. Expanding the variation of the functional integral J = J[yi0 ] J[yi ] = 0
to 1st order in yi we have
Z s2 X  
f f
0 = J = ds yi + yi . (1.77)
s1 i
y i y i

Using integration by parts on the second term, and the vanishing of the variation at the
R s2 P h f d f
i
endpoints to remove the surface term, J vanishes when s1 i yi ds yi yi (s) ds =
0. For independent variations yi (for example, after imposing holonomic constraints), this
can only occur if  
f d f
= 0. (1.78)
yi ds yi
The scope of this calculus of variation result for extremizing the integral over f is more
general than its application to classical mechanics.

Example: Hamiltons principle states that motion qi (t) extremizes the action, so in this
case s = t, yi = qi , f = L, and J = S. Demanding S = 0 then yields the Euler-Lagrange
equations of motion from Eq. (1.78).

Example: As an example outside of classical mechanics, consider showing that the shortest
distance between points on aR sphere of radius a are great circles. This can be seen by
s
minimizing the distance J = s12 ds where for a spherical surface,
p q
ds = (dx) + (dy) + (dz) = a2 (d)2 + a2 sin2 ()(d)2
2 2 2 (1.79)

since dr = 0. Taking s = and y = , then

s  2
2 d
ds = a 1 + sin () d, (1.80)
d

22
 CHAPTER 1. A REVIEW OF ANALYTICAL MECHANICS

q  
so f = 1 + sin2 () 2 . The solution for the minimal path is given by solving d
d f

f
=
0. After some algebra these are indeed found to be great circles, described by sin( ) =
cot() where , are constants.

Example: Hamiltons principle can also be used to yield the Hamilton equations of motion,
by considering the variation of a path in phase space. In this case
Z t2 h i
J[q, p] = dt pi qi H(q, p, t) = 0 (1.81)
t1

must be solved with fixed endpoints: qi (t1 ) = qi (t2 ) = 0 and pi (t1 ) = pi (t2 ) = 0. Here,
the role of yi , of is played by the 2N variables (q1 , . . . , qN , p1 , . . . , pN ). As f = pi qi H, then
 
d f f H
=0 = pi = , (1.82)
dt qi qi qi
 
d f f H
=0 = qi = ,
dt pi pi pi
giving Hamiltons equations as expected. Note that because f is independent of pi , the
term (f /pi )pi = 0, and it would seem that we do not really need the condition that
pi (t1 ) = pi (t2 ) = 0 to remove the surface term. However, these conditions on the variations
pi are actually required in order to put qi and pi on the same footing (which we will exploit
later in detail when discussing canonical transformations).
It is interesting and useful to note that DAlemberts principle
   
d L L
Rj qj = 0 (1.83)
dt qj qj
is a differential version of the equations that encode the classical dynamics, while Hamil-
tons principle Z t2   
L d L
J = dt qj = 0 (1.84)
t1 qj dt qj
(for Rj = 0 where all forces come from a potential) is an integrated version.

Method of Lagrange Multipliers

Next we will consider the method of Lagrange multipliers. For simplicity we will assume
there are no generalized forces outside the potential, Rj = 0, until further notice. The
method of Lagrange multipliers will be useful for two situations that we will encounter:
1. When we actually want to study the forces of constraint that are holonomic.

2. When we have semi-holonomic constraints.

23
 CHAPTER 1. A REVIEW OF ANALYTICAL MECHANICS

Let us consider k constraints for n coordinates, with {1, . . . , k} being the index
running over the constraints. These holonomic or semi-holonomic constraints take the form
g (q, q, t) = aj (q, t)qj + at (q, t) = 0 (1.85)
where again repeated indices are summed. Thus, g dt = aj dqj + at dt = 0. For a virtual
displacement qj we have dt = 0, so
n
X
aj qj = 0 , (1.86)
j=1

which gives us k equations constraining the virtual displacements. For each equation we
can multiply by a function (t) known as Lagrange multipliers, and sum over , and the
combination will still be zero. Adding this zero to DAlemberts principle yields
   
d L L
aj qj = 0 (1.87)
dt qj qj
where the sums implicitly run over both and j. Its clear that the Lagrange multiplier
term is zero if we sum over j first, but now we want to consider summing first over for
a fixed j. Our goal is make the term in square brackets zero. Only n k of the virtual
displacements qj are independent, so for these values of j the square brackets must vanish.
For the remaining k values of j we can simply choose the k Lagrange multipliers to force
the k square bracketed equations to be satisfied. This is known as the method of Lagrange
multipliers. Thus all square bracketed terms are zero, and we have the generalization of the
Euler-Lagrange equations which includes terms for the constraints:
 
d L L
= aj . (1.88)
dt qj qj
This is n equations, for the n possible values of j, and on the right-hand-side we sum over
for each one of these equations. The sum aj can be interpreted as a generalized constraint
force Qj . The Lagrange multipliers and generalized coordinates qj together form n + k
parameters, and equation (1.88) in conjunction with g = 0 for each from (1.85) together
form n + k equations to be solved.
There are two important cases to be considered.
1. In the holonomic case, f (q, t) = 0. Here, g = f = f
q + ft , so aj = f
qj j qj

. This
gives
  k
d L L X f
= (1.89)
dt qj qj =1
qj
for holonomic constraints. The same result can be derived from a generalized Hamil-
tons principle Z t2
J[qj , ] = (L + f ) dt (1.90)
t1

24
 CHAPTER 1. A REVIEW OF ANALYTICAL MECHANICS

by demanding that J = 0. It is convenient to think of f as an extra potential

energy that we add into L so that a particle does work if it leaves the surface defined
by f = 0. Recall that given this potential, the Forceq = q ( f ) = q f ,
where the derivative q f gives a vector that is normal to the constraint surface of
constant f = 0. This agrees with the form of our generalized force above.
2. In the semi-holonomic case, we just have g = aj (q, t)qj +at (q, t) = 0, with aj = g
qj
.
This gives
  k
d L L X g
= (1.91)
dt qj qj =1
q j

for semi-holonomic constraints. This result cannot be derived from Hamiltons principle
in general, justifying the time we spent discussing dAlemberts principle, which we
have used to obtain (1.91). Recall that static friction imposes a no-slip constraint in
the form of our equation g = 0. For g q, the form , g
q
, is consistent with the form
of generalized force we derived from our dissipation function, F q
from our discussion
of friction.
We end this chapter with several examples of the use of Lagrange multipliers.

Example: Consider a particle of mass m at rest on the top of a sphere of radius a, as shown
above in Fig. 1.7. The particle is given an infinitesimal displacement = 0 so that it slides
down. At what angle does it leave the sphere?
We use the coordinates (r, , ) but set = 0 by symmetry as it is not important. The
constraint r a is non-holonomic, but while the particle is in contact with the sphere the
constraint f = r a = 0 is holonomic. To answer  this question
 we will look for the point
m 2 2 2
where the constraint force vanishes. Here T = 2 r + r and V = mgz = mgr cos() so
that L = T V , then dtd rL rL = f


r
gives
mr mr2 + mg cos() = , (1.92)
 
while d
dt
L

L

= f

= 0 gives

d  2 
mr mgr sin() = 0. (1.93)
dt
This in conjunction with r = a gives 3 equations for the 3 variables (r, , ). Putting them
together gives r = 0 so r = 0. This means
ma2 = mga sin(), ma2 + mg cos() = .
Multiply the first of these by and integrate over time, knowing that = 0 when = 0,
gives 2 = 2g
a
(1 cos()). Thus,
= mg(3 cos() 2) (1.94)

25
 CHAPTER 1. A REVIEW OF ANALYTICAL MECHANICS

is the radial constraint force. The mass leaves the sphere when = 0 which is when
cos() = 23 (so 48o ).
What if we instead imposed the constraint f 0 = r2 a2 = 0? If we call its Lagrange
0
multiplier 0 we would get 0 f
r
= 2a0 when r = a, so 2a0 = is the constraint force from
0
before. The meaning of is different, and it has different units, but we still have the same
constraint force.
What are the equations of motion for > arccos 32 ? Now we no longer have the

constraint so

2 d  2 
mr mr + mg cos() = 0 and mr mgr sin() = 0.
dt
The initial conditions are r1 = a, 1 = arccos 23 , r1 = 0, and 12 = 3a
2g


from before. Simpler
coordinates are x = r sin() and z = r cos(), giving
m 2
x + z 2 mgz,


L= (1.95)
2

so x = 0 and z = g with initial conditions z1 = 2a 3
, x1 = 35a , and the initial velocities
simply left as z1 and x1 for simplicity in writing (though the actual values follow from
z1 = a sin 1 1 and x1 = a cos 1 1 ). This means

x(t) = x1 (t t1 ) + x1 , (1.96)
g
z(t) = (t t1 )2 + z1 (t t1 ) + z1 , (1.97)
2
where t1 is the time when the mass leaves the sphere. That can be found from
 
2 2g 4g 2
= (1 cos()) = sin , (1.98)
a a 2
q
a
R arccos( 23 ) d
so t1 = 4g 0 sin( 2 )
where 0 is the small initial angular displacement from the top
of the sphere.

Example: Consider a hoop of radius a and mass m rolling down an inclined plane of angle
without slipping as shown in Fig. 1.11, where we define the x direction as being parallel to
the ramp as shown. What is the friction force of constraint, and how does the acceleration
compare to the case where the hoop is sliding rather than rolling?
The no-slip constraint means a = x, so h = a x = a, which can be made holonomic
but which we will treat as semi-holonomic. Then T = TCM + Trotation = 12 mx2 + 12 ma2 2 as
Ihoop = ma2 . Meanwhile, V = mg(l x) sin() so that V (x = l) = 0. This means

m 2 ma2 2
L=T V = x + + mg(x l) sin(). (1.99)
2 2
26
 CHAPTER 1. A REVIEW OF ANALYTICAL MECHANICS

Figure 1.11: Hoop rolling on inclined plane

 
d L L
h d L L
= h


The equations of motion from dt x
x
= x
and dt

are

mx mg sin() = and ma2 = a, (1.100)

along with x = a. Taking a time derivative of the constraint gives x = a, so mx = , and

g
x = 2 sin(). This is one-half of the acceleration of a sliding mass. Plugging this back in we
find that
1
= mg sin() (1.101)
2
is the friction force in the x direction for the no-sliding constraint, and also = g sin(). 2a

Example: Consider a wedge of mass m2 and angle resting on ice and moving without
friction. Let us also consider a mass m1 sliding without friction on the wedge and try to
find the equations of motion and constraint forces. The constraints are that y2 = 0 so the

Figure 1.12: Wedge sliding on ice

wedge is always sitting on ice, and xy11 y2

x2
= tan() so the point mass is always sitting on the
wedge. (We will ignore the constraint force for no rotation of the wedge, and only ask about
these two.) The kinetic energy is simply T = m21 (x21 + y12 ) + m22 (x22 + y22 ), while the potential

27
 CHAPTER 1. A REVIEW OF ANALYTICAL MECHANICS

energy is V = m1 gy1 + m2 g(y2 + y0 ), where y0 is the CM of the wedge taken from above its
bottom. Then L = T V , with the constraints f1 = (y1 y2 ) (x1 x2 ) tan() = 0 and
f2 = y2 = 0. The equations of motion from the Euler-Lagrange equations with holonomic
constraints are
d L L f1 f2
= 1 + 2 = m1 x1 = 1 tan(), (1.102)
dt x1 x1 x1 x1
d L L f1 f2
= 1 + 2 = m1 y1 + m1 g = 1 ,
dt y1 y1 y1 y1
d L L f1 f2
= 1 + 2 = m2 x2 = 1 tan(),
dt x2 x2 x2 x2
d L L f1 f2
= 1 + 2 = m2 y2 + m2 g = 1 + 2 ,
dt y2 y2 y2 y2

which in conjunction with y1 y2 = (x1 x2 ) tan() and y2 = 0 is six equations. We number

them (1) to (6). Equation (6) gives y2 = 0 so (4) gives m2 g = 2 1 where 2 is the force
of the ice on the wedge and 1 is the vertical force (component) of the wedge on the point
mass. Adding (1) and (3) gives m1 x1 + m2 x2 = 0 meaning that the CM of m1 and m2 has
no overall force acting on it.
Additionally, as (5) implies y1 = (x1 x2 ) tan(), then using (1), (2), and (3) we find
the constant force
g
1 = 2 . (1.103)
1
m1 cos2 ()
+ tanm2()
With this result in hand we can use it in (1), (2), and (3) to solve for the trajectories. Since

tan()
x2 = 1 , (1.104)
m2
tan()
x1 = 1 ,
m1
1
y1 = g,
m1
the accelerations are constant. As a check on our results, if m2 , then x2 = 0 so indeed
the wedge is fixed; and for this case, x1 = g sin() cos() and y1 = g sin2 () which both
vanish as 0 as expected (since in that limit the wedge disappears, flattening onto the
icy floor below it).

28
Chapter 2

Rigid Body Dynamics

2.1 Coordinates of a Rigid Body

A set of N particles forms a rigid body if the distance between any 2 particles is fixed:

rij |ri rj | = cij = constant. (2.1)

Given these constraints, how many generalized coordinates are there?
If we know 3 non-collinear points in the body, the remaing points are fully determined
by triangulation. The first point has 3 coordinates for translation in 3 dimensions. The
second point has 2 coordinates for spherical rotation about the first point, as r12 is fixed.
The third point has one coordinate for circular rotation about the axis of r12 , as r13 and r23
are fixed. Hence, there are 6 independent coordinates, as represented in Fig. 2.1. This result
is independent of N , so this also applies to a continuous body (in the limit of N ).

Figure 2.1: 3 non-collinear points can be fully determined by using only 6 coordinates. Since
the distances between any two other points are fixed in the rigid body, any other point of
the body is fully determined by the distance to these 3 points.

29
 CHAPTER 2. RIGID BODY DYNAMICS

The translations of the body require three spatial coordinates. These translations can
be taken from any fixed point in the body. Typically the fixed point is the center of mass
(CM), defined as:
1 X
R= mi ri , (2.2)
M i

where mi is the mass of the i-th particle and ri the position of that particle with respect to
a fixed origin and set of axes (which will notationally be unprimed) as in Fig. 2.2. In the
case of a continuous body, this definition generalizes as:
Z
1
R= r(r) dV, (2.3)
M V

where (r) is the mass density at position r and we integrate over the volume V.

Figure 2.2: The three translational coordinates correspond to the position of the Center of
Mass, and the three rotational coordinates correspond to the three angles necessary to define
the orientation of the axis fixed with the body.

Rotations of the body are taken by fixing axes with respect to the body (we will denote
these body fixed axes with primes) and describing their orientation with respect to the
unprimed axes by 3 angles (, , ).
A particularly useful choice of angles are called Euler angles. The angle is taken as a
rotation about the z-axis, forming new x- and y-axes while leaving the z-axis unchanged, as
shown in Fig. 2.3. The angle is then taken as a rotation about the x-axis, forming new
y 0 - and z 0 -axes while leaving the x-axis unchanged, as shown in Fig. 2.4. Finally, the angle
is taken as a rotation about the z 0 -axis, forming new x0 - and y 0 -axes while leaving the
z 0 -axis unchanged, as shown in Fig. 2.5. (The x-axis is called the line of nodes, as it is the
intersection of the xy- and xy-planes.)

30
 CHAPTER 2. RIGID BODY DYNAMICS

Figure 2.4: Second rotation Figure 2.5: Final rotation

Figure 2.3: First rotation is is by around the final z 0
by around the original z is by around the interme-
diate x axis. axis.
axis.

Rotations can be described by 3 3 matrices U . This means each rotation step can be
described as a matrix multiplication. Where r = (x, y, z), then

cos() sin() 0 x
r = U r = sin() cos() 0 y . (2.4)
0 0 1 z
Similar transformations can be written for the other terms:
r0 = U r , r0 = U r0 = U U r = U U U r.
Defining the total transformation as U , it can be written as:
U U U U r0 = U r. (2.5)
Care is required with the order of the terms since the matrices dont commute. Writing U
out explicitly:

cos() sin() 0 1 0 0 cos() sin() 0
U = sin() cos() 0 0 cos() sin() sin() cos() 0 . (2.6)
0 0 1 0 sin() cos() 0 0 1
All rotation matrices, including U , U , U , and U are orthogonal. Orthogonal matrices W
satisfy
W > W = W W > = 1 W > = W 1 , (2.7)
where 1 refers to the identity matrix and > to the transpose. This ensures that the length
of a vector is invariant under rotations:
r02 = r> (W > W )r = r2 . (2.8)

31
 CHAPTER 2. RIGID BODY DYNAMICS

Orthogonal matrices W have 9 entries but need to fulfill 6 conditions from orthogonality,
leaving only 3 free parameters, corresponding to the 3 angles necessary to determine the
rotation.
We can also view r0 = U r as a transformation from the vector r to the vector r0 in
the same coordinate system. This is an active transformation, as opposed to the previous
perspective which was a passive transformation.
Finally, note that inversions like

1 0 0 1 0 0
U = 0 1 0 or U = 0 1 0 (2.9)
0 0 1 0 0 1

are not rotations. These have det(U ) = 1, so they can be forbidden by demanding that
det(U ) = 1. All orthogonal matrices have det(W ) = 1 because det(W > W ) = (det(W ))2 =
1. In the language of group theory, the restriction to det(W ) = 1 gives the special or-
thogonal group SO(3) as opposed to simply O(3), the orthogonal group. We disregard the
det(U ) = 1 subset of transformations because it is impossible for the system to undergo
these transformations continuously without the distance between the particles changing in
the process, so it would no longer be a rigid body.

Intuitively, we could rotate the coordinates (x, y, z) directly into the coordinates (x0 , y 0 , z 0 )
by picking the right axis of rotation. In fact, the Euler theorem states that a general dis-
placement of a rigid body with one point fixed is a rotation about some axis. This theorem
will be true if a general rotation U leaves some axis fixed, which is satisfied by

Ur = r (2.10)

for any point r on this axis. This is an eigenvalue equation for U with eigenvalue 1. To
better understand this, we need to develop a little linear algebra.
Although the notion of an eigenvalue equation generally holds for linear operators, for
now the discussion will be restricted to orthogonal rotation matrices U . The eigenvalue
equation is
U = , (2.11)
where is an eigenvector and is the associated eigenvalue. Rewriting this as

(U 1) = 0 (2.12)

requires that det(U 1) = 0, so that U 1 is not invertible and the solution can be
6 0. det(U 1) = 0 is a cubic equation in , which has 3 solutions, which
non-trivial, =
are the eigenvalues for {1, 2, 3}. The associated eigenvectors are () and satisfy

U () = () , (2.13)

32
 CHAPTER 2. RIGID BODY DYNAMICS

where no implicit sum over repeated indices is taken. Forming a matrix from the resulting
eigenvectors as columns:


X = (1) (2) (3) (2.14)

then we can rewrite Eq.(2.13) as

U X = X diag(1 , 2 , 3 ) X 1 U X = diag(1 , 2 , 3 ) (2.15)

This means X diagonalizes U . Since U is orthogonal, the matrix X is unitary (X X =

XX = 1). Note that > indicates transposition whereas indicates Hermitian conjugation
(i.e. complex conjugation ? combined with transposition > ).
Next we note that since det(U ) = 1, then 1 2 3 = 1. Additionally, | |2 = 1 for each
because:
U = U > = ? = U > U = ||2 . (2.16)
Finally, if is an eigenvalue, then so is ? :

U = U ? = ? ? (2.17)

where ? is still a column vector but with its elements undergoing complex conjugation with
respect to . Without loss of generality, let us say for a rotation matrix U that 2 = ?3 .
Then 1 = 1 |2 |2 = 1 , so one of the eigenvalues is 1, giving Eq.(2.10), and thus proving
Eulers Theorem. The associated eigenvector (1) to the eigenvalue 1 = 1 is the rotation
axis, and if 2 = ?3 = ei then is the rotation angle about that axis.

In fact, we can make good use of our analysis of Eulers theorem. Together the rotation
axis and rotation angle can be used to define the instantaneous angular velocity (t) such
that:
|| = and k (1) . (2.18)
The angular velocity will play an important role in our discussion of time dependence with
rotating coordinates in the next section. If we consider several consecutive displacements of
the rigid body, then each can have its own axis (1) and its own , so changes at each
instance of time, and hence = (t) (for the entire rigid body).

2.2 Time Evolution with Rotating Coordinates

Lets use unprimed axes (x, y, z) for the fixed (inertial) axes, with fixed basis vectors ei . We
will also use primed axes (x0 , y 0 , z 0 ) for the body axes with basis vectors e0i .

33
 CHAPTER 2. RIGID BODY DYNAMICS

If we consider any vector then it can be decomposed

with either set of basis vectors:
X X
b= bi e i = b0i e0i . (2.19)
i i

For fixed axes basis vectors by definition ei = 0, while

for those in the body frame,

e0i = (t) e0i (2.20)

meaning vectors of fixed length undergo a rotation at

a time t. The derivation of this result is shown in the
figure on the right, by considering the change to the
vector after an infinitesimal time interval dt.
Summing over repeated indices, this means:

b = b i ei = b 0i e0i + (t) (b0i e0i ) = b 0i e0i + (t) b

d dR
Defining as the time evolution in the fixed (F) frame and the time evolution in
dt dt
the rotating/body (R) frame, then vectors evolve in time according to
db dR b
= + b. (2.21)
dt dt
As a mnemonic we have the operation (d/dt) = dR /dt + which can act on any vector.
Let us apply this to the position r of a particle of mass m, which gives
dr dR r
= +r vF = vR + r. (2.22)
dt dt
Taking another time derivative gives us the analog for acceleration:
F dvF dR v F
= = + vF (2.23)
m dt dt
dR v R dR dR r
= + r+ + vR + ( r) .
dt dt dt
As ddt
Rr
= vR is the velocity within the rotating body frame and dR vR
dt
= aR is the acceleration
within the body frame, then
dR
maR = F m ( r) 2m vR m r (2.24)
dt
gives the acceleration in the body frame with respect to the forces that seem to be present
in that frame. The terms m ( r) and 2m vR are, respectively, the centrifugal

34
 CHAPTER 2. RIGID BODY DYNAMICS

and Coriolis ficticious forces respectively, while the last term m dRdt r is a ficticious force
that arises from non-uniform rotational motion, so that there is angular acceleration within
the body frame. The same result could also have been obtained with the Euler-Lagrange
equations for L in the rotating coordinates:

m
L= (r + r)2 V , (2.25)
2

and you will explore this on a problem set.

Note that the centrifugal term is radially outward and perpendicular to the rotation axis.
To see this, decompose r into components parallel and perpendicular to , r = rk + r , then
r = r , so ( r ) = w2 r . This term is present for any rotating body.
On the other hand, the Coriolis force is nonzero when there is a nonzero velocity in the
rotating/body frame: vR = 6 0.

Example: Consider the impact of the Coriolis force on projectile motion on the rotating
Earth, where the angular velocity is Earth = 2436002
s
7.3 105 s1 . We work out the
cross-product vr as shown in Fig. 2.6 for a particle in the northern hemisphere, where
points to the north pole. Thus a projectile in the northern/southern hemisphere would be
perturbed to the right/left relative to its velocity direction vr .

Figure 2.6: For a projectile, in the Northern Hemisphere, the Coriolis pushes it to its right,
relative to its direction of motion.

Example: Consider a Foucault pendulum which hangs from a rigid rod, but is free to os-
cillate in two angular directions, as shown in Fig. 2.2. For  1 and working to first order
in the small , the result derived from the Coriolis force gives Earth sin(). Here is
the latitude angle measured from equator. The precession is clockwise in the northern hemi-
sphere, and is maximized at the north pole where = 90 . (This is proven as a homework
problem.)

35
 CHAPTER 2. RIGID BODY DYNAMICS

Example: Consider the Coriolis deflection of a freely falling body on Earth in the northern
hemisphere. We use the coordinate system shown below, where z is perpendicular to the
surface of the earth and y is parallel to the earths surface and points towards the north
pole.

Working to first order in the small gives us

maR = mvR = mgz 2m v, (2.26)
where the centrifugal terms of order O( 2 ) are dropped. As an initial condition we take
v(t = 0) = v0 z. The term v points along x, so:
z = g + O( 2 ) vz = v0 gt (2.27)
Moreover implementing the boundary condition that x(t = 0) = 0:
 g 
x = 2( v)x = 2 sin()vz (t) x = 2 sin() v0 t t2 . (2.28)
2
Taking also x(t = 0) = 0, and integrating one further time, the motion in the x direction is:
v g 
0 2
x(t) = 2 sin() t t3 . (2.29)
2 6
36
 CHAPTER 2. RIGID BODY DYNAMICS

Lets consider this effect for a couple simple cases. If the mass m is dropped from a height
z(t = 0) = hmax with zero velocity, v0 = 0, then:
g
z = hmax t2 (2.30)
2
and the mass reaches the floor at time
s
2hmax
t1 = . (2.31)
g

From Eq.(2.28) we see that x > 0 for all t, and that:

8 sin()h3max
x(t = t1 ) = > 0.
3g 2
However, if the mass m is thrown up with an initial z(t = 0) = v0 > 0 from the ground
(z = 0), then :
g
z(t) = v0 t t2 > 0. (2.32)
2
Here the particle rises to a maximum height z = v02 /(2g) at time t = v0 /g, and then falls
back to earth. Using Eq.(2.28) we see that x < 0 for all t. If t1 is the time it reaches the
ground again (t1 = 2vg0 ), then:

4 sin()v03
x(t = t1 ) = < 0. (2.33)
3g 2

2.3 Kinetic Energy, Angular Momentum, and

the Moment of Inertia Tensor for Rigid Bodies
Returning to rigid bodies, consider one built out of N fixed particles The kinetic energy
is best expressed using CM coordinates, where R is the CM and we here take ri to be
the displacement of particle i relative to the CM. Once again making sums over repeated
subscripts as implicit, the kinetic energy (T ) of the system is given by:
1 2 1
T = MR + mi r2i . (2.34)
2 2
As the body is rigid, then points cannot translate relative to the body but can only rotate
so that ri = ri . The rotational kinetic energy is then
1 1 1 
TR = mi r2i = mi ( ri )2 = mi 2 r2i ( ri )2 .

(2.35)
2 2 2
37
 CHAPTER 2. RIGID BODY DYNAMICS

Labeling Cartesian indices with a and b to reserve i and j for particle indices, then we
can write out this result making the indicies all explict as
1X
TR = mi (ab r2i ria rib )a b . (2.36)
2 i,a,b

It is convenient to separate out the parts in this formula that depend on the shape and
distributions of masses in the body by defining the moment of inertia tensor I for the discrete
body as X
Iab mi (ab ri2 ria rib ) . (2.37)
i

The analog for a continuous body of mass density (r) is:

Z

Iab (r2 ab ra rb )(r) dV. (2.38)
V

In terms of the moment of inertia tensor, the kinetic energy from rotation can now be
written as:
1X 1
TR = Iab a b = I , (2.39)
2 a,b 2
where in the last step we adopt a convenient matrix multiplication notation.
The moment of inertia tensor can be written with its components as a matrix in the form
2
X yi + zi2 xi yi xi zi
I = mi xi yi x2i + zi2 yi zi , (2.40)
2 2
i xi zi yi zi xi + yi
where the diagonal terms are the moments of inertia and the off-diagonal terms are the
products of inertia. Note also that I is symmetric in any basis, so Iab = Iba .

Special case: if the rotation happens about only one axis which can be defined as the z-axis
for convenience so that = (0, 0, ), then TR = 12 Izz 2 which reproduces the simpler and
more familiar scalar form of the moment of inertia.

Lets now let ri be measured from a stationary point in the rigid body, which need not
necessarily be the CM. The angular momentum can be calculated about this fixed point.
Since vi = ri , we can write the angular momentum as:

L = mi ri vi = mi ri ( ri ) = mi ri2 ( ri )ri .
 
(2.41)

Writing out the components

X X X
La = mi (r2i a ( ri )ria ) = b mi (ab r2i ria rib ) = Iab b , (2.42)
i i,b b

38
 CHAPTER 2. RIGID BODY DYNAMICS

which translates to the matrix equation:

L = I . (2.43)
This allows us to write the corresponding rotational kinetic energy as:
1
TR = L. (2.44)
2
Note that in general, L is not parallel to . We will see an explicit example of this below.
Also note that the formula that we used for I in this case is the same as we gave above. We
use these formulas whether or not ri is taken with respect to the CM.

It is useful to pause to see what precisely the calculation of I depends on. Since it in-
volves components of the vectors ri it depends on the choice of the origin for the rotation.
Furthermore the entries of the matrix Iab obviously depend on the orientation of the axes
used to define the components labeled by a and b. Given this, it is natural to ask whether
given the result for Iab with one choice of axes and orientation, whether we can determine
an I0 a0 b0 for a different origin and axes orientation. This is always possible with the help of
a couple of theorems.

The parallel axis theorem: Given ICM about the CM, it is simple to find IQ about a
different point Q with the same orientation for the axes. Referring to the figure below,

we define r0i as the coordinate of a particle i in the rigid body with respect to point Q and
ri to be the coordinate of that particle with respect to the CM, so that:
r0i = R + ri . (2.45)
By definition of the CM:
X X
mi ri = 0 and we let M= mi . (2.46)
i i

39
 CHAPTER 2. RIGID BODY DYNAMICS

The tensor of inertia around the new axis is then:

Iab
Q
= mi (ab r02 0 0
i ria rib ) (2.47)
= mi (ab (r2i + 2ri R + R2 ria rib ria Rb Ra rib Ra Rb ) , (2.48)

where the cross terms involving a single ri or single component ria sum up to zero by
Eq.(2.46). The terms quadratic in r are recognized as giving the moment of inertia tensor
about the CM. This gives the parallel axis theorem for translating the origin:

Iab
Q
= M (ab R2 Ra Rb ) + Iab
CM
, (2.49)

If we wish to carry out a translation between P and Q, neither of which is the CM, then
we can simply use this formula twice. Another formula can be obtained by projecting the
parallel axis onto a specific axis n where n2 = 1 (giving a result that may be familiar from
an earlier classical mechanics course):

n IQ n = M (R2 (n R)2 ) + n ICM n = M R2 [1 cos2 ()] + n ICM n

= M R2 sin2 () + n ICM n (2.50)

where n R R cos().

Example: Lets consider an example of the calculation of I for a situation where L is not
parallel to . Consider a dumbbell made of 2 identical point passes m attached by a massless
rigid rod (but with different separations r1 and r2 from the axis of rotation), spinning so
that = z and so that the rod makes an angle with the axis of rotation, as shown

We define body axes where the masses lie in the yz-plane. Here,

r1 = (0, r1 sin , r1 cos ) and r2 = (0, r2 sin , r2 cos ). (2.51)

40
 CHAPTER 2. RIGID BODY DYNAMICS

Then using the definition of the moment inertia tensor:

Izz = m(x12 + y12 ) + m(x22 + y22 ) = m(r12 + r22 ) sin2

Ixx = m(y12 + z12 ) + m(y22 + z22 ) = m(r12 + r22 )
Iyy = m(x21 + z12 ) + m(x22 + z22 ) = m(r12 + r22 ) cos2 (2.52)
Ixy = Iyx = m(x1 y1 x2 y2 ) = 0
Ixz = Izx = m(x1 z1 x2 z2 ) = 0
Iyz = Izy = m(y1 z1 y2 z2 ) = m(r12 + r22 ) sin cos

Plugging these into L = I , recalling that only z is non-zero, this gives

L = (0, Iyz , Izz ). (2.53)

Thus in this example L is not parallel to .

Next, instead of translating the axes in a parallel manner, let us keep the origin fixed and
rotate the axes according to an orthogonal rotation matrix U satisfying U > U = U U > = 1.
Vectors are rotated as

L0 = U L , 0 = U and therefore = U > 0 . (2.54)

Putting these together

L0 = U I = (U IU
>) 0 I0 = U IU
>, (2.55)

where I0 is the new moment of inertia tensor. (The fact that it transforms this way defines
it as a tensor.) This allows us to calculate the new moment of inertia tensor after a rotation.

For a real symmetric tensor I, there always exists a rotation from an orthogonal matrix
U that diagonalizes I giving a diagonal matrix I0 :

I1 0 0
ID = 0 I2 0 . (2.56)
0 0 I3

The entries of the diagonal moment of inertia tensor, I , are real and positive. This is
just a special case of saying a Hermitian matrix can always be diagonalized by a unitary
transformation (which is often derived in a Quantum Mechanics course as part of showing
that a Hermitian matrix has real eigenvalues and orthogonal eigenvectors). The positivity
of diagonal matrix follows immediately from the definition of the moment of inertia tensor
for the situation with zero off-diagonal terms.

41
 CHAPTER 2. RIGID BODY DYNAMICS

The axes that make I diagonal are called the principal axes and the components I are
the principal moments of inertia. We find them by solving the eigenvalue problem
I = , (2.57)
where the 3 eigenvalues give the principal moments of inertia I , and are obtained from
solving det(I 1) = 0. The corresponding  3 real orthogonal eigenvectors
()
are the
>
 (1) (2) (3)
principal axes. Here U = , where the eigenvectors vectors fill out the
columns. Then, without summing over repeated indices:
1X
L = I and T = I 2 , (2.58)
2
where L and are the components of L and , respectively, evaluated along the principal
axes.

To summarize, for any choice of origin for any rigid body, there is a choice of axes that
For T to separate into translational and rotational parts, we must pick the
diagonalizes I.
origin to be the CM. Often, the principal axes can be identified by a symmetry of the body.

Example: for a thin rectangle lying in the yz-plane with one edge coinciding with the z-axis,
and the origin chosen as shown below, then Iyz = 0 as the body is symmetric under z z,
while Ixz = Ixy = 0 as the body lies entirely within x = 0. Hence these are principal axes.

Sometimes, symmetry allows multiple choices for the principal axes.

Example: for a sphere, any orthogonal axes through the origin

are principal axes.

Example: for a cylinder whose central axis is

aligned along the z-axis, because of rotational
symmetry any choice of the x- and y-axes gives
principal axes.

42
 CHAPTER 2. RIGID BODY DYNAMICS

Example: Lets consider an example where the principal axes may not be apparent, which
we can solve through the eigenvalue problem. Consider a uniform cube with sides of length
a, mass m, and having the origin at one corner, as shown below.

By symmetry we have
m a a a 2
Z Z Z
2
Ixx = Iyy = Izz = 3 (x + y 2 ) dx dy dz = ma2 , (2.59)
a 0 0 0 3
Z aZ aZ a
m 1
Ixy = Iyz = Ixz = 3 xz dx dy dz = ma2 .
a 0 0 0 4
Thus the matrix is
+ 23 14 14

I = ma2 14 + 23 14 . (2.60)

14 14 + 23
The principal moments of inertia are found from
 2  
11 1
det(I 1) = 2
ma 2
ma = 0 . (2.61)
12 6
This gives I1 = 1 = 16 ma2 . Solving

(I 1 1) (1) = 0 we find (1) = (1, 1, 1). (2.62)

The remaining eigenvalues are degenerate:
1
I2 = I3 = 2 = 3 = ma2 (2.63)
12
2 1) (2,3) =
so there is some freedom in determining the corresponding principal axes from (I
(1)
0, though they still should be orthogonal to each other (and ). One example of a solution
is:
(2) = (1, 1, 0) and (3) = (1, 1, 2) (2.64)

43
 CHAPTER 2. RIGID BODY DYNAMICS

Using these principal axes and the same origin, the moment of inertia tensor becomes

1 0 0
ma2 1
ID = 0 2 0 . (2.65)
6 1
0 0 2

In contrast, if we had chosen the origin as the center of the cube, then one choice for the
principal axes would have the same orientation, but with ICM = 16 ma2 1. This result could
be obtained from Eq. (2.65) using the parallel axis theorem.

2.4 Euler Equations

Consider the rotational motion of a rigid body about a fixed point (which could be the CM
but could also be another point). We aim to describe the motion of this rigid body by
exploiting properties of the body frame. To simplify things as much as possible, for this
fixed point, we choose the principal axes fixed in the body frame indexed by {1, 2, 3}.
Using the relation between time derivatives in the inertial and rotating frames, the torque
is then given by:
dL dR L
= = +L (2.66)
dt dt
where = (t). For example:

dR L1
1 = + 2 L3 3 L2 . (2.67)
dt
Not summing over repeated indices and using the formula for angular momentum along
the principal axes gives L = I . Since we have fixed moments of inertia within the
body we have dR I /dt = 0. Note that d/dt = dR /dt + = dR /dt, so its rotating
and inertial time derivatives are the same, and we can write without possible cause of
confusion. Thus dR L /dt = I . This yields the Euler equations:

I1 1 (I2 I3 )2 3 = 1
I2 2 (I3 I1 )3 1 = 2 (2.68)
I3 3 (I1 I2 )1 2 = 3

where in all of these and are calculated in the rotating/body frame. This can also be
written as

= I + lk l k Ik , (2.69)

with fixed but a sum implied over the repeated l and k indicies. Here abc is the fully
antisymmetric Levi-Civita symbol.

44
 CHAPTER 2. RIGID BODY DYNAMICS

Solving these equations gives (t). Since the result is expressed in the body frame,
rather than the inertial frame of the observer, this solution for (t) may not always make
the physical motion transparent. To fix this we can connect our solution to the Euler angles
using the relations

1 = x0 = sin sin + cos ,

2 = y0 = sin cos sin , (2.70)
3 = z0 = cos + .

These results should be derived as exercise for the student.

Example: let us consider the stability of rigid-body free rotations ( = 0). Is a rotation
= 1 e1 about the principal axis e1 stable?
Perturbations can be expressed by taking = 1 e1 + 2 e2 + 3 e3 , where 2 and 3 are small
and can be treated to 1st order. The Euler equations are:

(I2 I3 )
1 = 2 3 = O(2 ) 0, (2.71)
I1

so 1 is constant at this order, and

(I3 I1 ) (I1 I2 )
2 = 1 3 and 3 = 1 2 . (2.72)
I2 I3

Combining these two equations yields

(I3 I1 )(I1 I2 )12

 
2 = 2 . (2.73)
I2 I3

The terms in the square bracket are all constant, and is either negative = w2 with an
oscillating solution 2 cos(wt + ), or is positive = 2 with exponential solutions 2
aet + bet . If I1 < I2,3 or I2,3 < I1 then the constant prefactor is negative, yielding stable
oscillatory solutions. If instead I2 < I1 < I3 or I3 < I1 < I2 then the constant prefactor
is positive, yielding an unstable exponentially growing component to their solution! This
behavior can be demonstrated by spinning almost any object that has three distinct principal
moments of inertia.

2.5 Symmetric Top with One Point Fixed

This section is devoted to a detailed analysis of a particular example that appears in many
situations, the symmetric top with one point fixed, acted upon by a constant force.

45
 CHAPTER 2. RIGID BODY DYNAMICS

Labeling the body axes as (x, y, z) and the fixed axes

as (xI , yI , zI ), as depicted in the right, symmetry implies
that I1 = I2 , and we will assume that I1,2 = 6 I3 . The
Euler angles are as usual (, , ). From the figure we
see that is the rotation rate of the top about the
(body) z-axis, is the precession rate about the zI fixed
inertial axis, and is the nutation rate by which the
top may move away or towards the zI axis. The Euler
equations in this case are

I1 1 (I2 I3 )2 3 = 1 ,
I1 2 (I3 I1 )3 1 = 2 , (2.74)
I3 3 = 0 = 3 .
Since the CM coordinate R is aligned along the z-axis there is no torque along z, 3 = 0,
leading to a constant 3 .

There are two main cases that we will consider.

Case: = 0 and = 0

The first case we will consider is when = 0 (so there is no gravity) and = 0 (so there
is no nutation). Then
dL
= = 0 L = constant (2.75)
dt
Let us define the constant:
I3 I1
3 . (2.76)
I1
Then the Euler equations for this situation reduce to:

1 + 2 = 0 and 2 1 = 0 . (2.77)

The simplest solution correspond to 1 (t) = 2 (t) = 0, where we just have a rotation about
the z-axis. Here:

L = L3 e3 where L3 = I3 3
1 = 2 = 0 = = 0 and = 3 . (2.78)

In this case L k . A more general situation is when L and are not necessarily parallel,
and 1 and 2 do not vanish. In this case Eq. (2.77) is solved by:

1 = C sin(t + D) and 2 = C cos(t + D). (2.79)

46
 CHAPTER 2. RIGID BODY DYNAMICS

The simple case corresponds to C = 0, so now we take

C > 0 (since a sign can be accounted for by the constant
phase D). This solution means precesses about the
body z-axis at the rate , as pictured on the right.
Since 12 + 22 is constant, the full = || is constant,
and is given by 2 = C 2 + 32 .
The energy here is just rotational kinetic energy TR =
1
2
L which is constant too, since both and L are
constant. Thus also precesses about L.
We can picture this motion by thinking about a body cone that rolls around a cone in the
fixed coordinate system, where in the case pictured with a larger cone about L we have
I1 = I2 > I3 .

To obtain more explicit results for the motion we can relate Eq.(2.79) to Euler angles. Since
= 0, we take = 0 to be constant. The other Euler angles come from:

sin(0 ) sin()

C sin(t + D)
= C cos(t + D) = sin(0 ) cos() . (2.80)
3
cos(0 ) +

Adding the squares of the 1st and 2nd components gives

C 2 = sin2 (0 ) 2 . (2.81)

To be definite, take the positive square root of this equation to give

C C
= = t + 0 . (2.82)
sin(0 ) sin(0 )
The first two equations in Eq. (2.80) are then fully solved by taking = t D, so we
find that both and have linear dependence on time. Finally the third equation gives a

47
 CHAPTER 2. RIGID BODY DYNAMICS

relation between various constants

3 = C cot(0 ) . (2.83)
Thus, we see that the solution has and are constants with = 0. If we had picked the
opposite sign when solving Eq. (2.81) then we would have found similar results:
C
= = t D and 3 = C cot(0 ) . (2.84)
sin(0 )

6 0 and =
Case: = 6 0

Now we consider the general case where 6= 0 and 6= 0. It is now more convenient to
use the Lagrangian than the Euler equations directly. Since I1 = I2 , using
sin() sin() + cos()

1
I1 (12 + 22 ) + I3 32 and = sin() cos() sin() ,


T = (2.85)

2
cos() +
gives us the kinetic energy
I1  2  I 
3
2
T = + sin2 2 + + cos . (2.86)
2 2
Moreover, V = mgR cos(), so in the Lagrangian L = T V the variables and are cyclic.
This means that the momenta
L 
= I1 sin2 () + I3 cos2 () + I3 cos()

p = (2.87)

L
p = = I3 ( + cos())
= I3 3 (2.88)

are conserved (constant). Here p is same as the angular momentum L3 discussed in the case
above. The torque is along the line of nodes, and p and p correspond to two projections
of L that are perpendicular to this torque (i.e. along zI and z). Additionally, the energy is
given by
I1  2 2 2
 I 
3
2
E =T +V = + sin () + + cos() + mgR cos() (2.89)
2 2
and is also conserved. Solving the momentum equations, Eq. (2.87), for and gives
p p cos()
= (2.90)
I1 sin2 ()
p (p p cos()) cos()
= .
I3 I1 sin2 ()

48
 CHAPTER 2. RIGID BODY DYNAMICS

Note that once we have a solution for (t) that these two equations then allow us to imme-
diately obtain solutions for (t) and (t) by integration. Eq. (2.90) can be plugged into the
energy formula to give

I1 2 (p p cos())2 p2
E = + + + mgR cos(), (2.91)
2 2I1 sin2 () 2I3

which is a (nonlinear) differential equation for , since all other quantities that appear are
simply constants. To simplify this result take u = cos() so that:

u2
1 u2 = sin2 () , u = sin() , 2 = . (2.92)
1 u2
Putting all this together gives:

u2 2EI3 p2 mgR
1 p p u 2
   
2
= u 1u Veff (u) , (2.93)
2 2I1 I3 I1 2 I1

which is a cubic polynomial that weve defined to be the effective potential Veff (u). The
solution to this from
du
dt = p (2.94)
2Veff (u)
yields a complicated elliptic function, from which it is hard to get intuition for the motion.

49
 CHAPTER 2. RIGID BODY DYNAMICS

Veff

u1 u2 u
1
0 1
=180 =0
Allowed region

Figure 2.7: Allowed region for solutions for the tops nutation angle that solve Eq. (2.95).

Instead, we can look at the form of Veff (u), because

1 2
u + Veff (u) = 0 (2.95)
2
is the equation for the energy of a particle of unit mass m = 1, kinetic energy u2 /2, a
potential Veff (u), and with vanishing total energy. The cubic equation will have in general
three roots where Veff (u) = 0. Since the kinetic energy is always positive or zero, the potential
energy must be negative or zero in the physical region, and hence the particle can not pass
through any of the roots. The roots therefore serve as turning points. Furthermore, physical
solutions are bounded by 1 (u = cos ) 1. While the precise values for the roots will
depend on the initial conditions or values of E, p , and p , we can still describe the solutions
in a fairly generic manner.
Consider two roots u1 and u2 (corresponding respectively to some angles 1 and 2 as
u = cos()) satisfying Veff (u1 ) = Veff (u2 ) = 0, where Veff (u) < 0 for u1 < u < u2 ; as shown
in Fig. 2.7. We see that u1 and u2 correspond to the turning points of the motion. The
region u1 < u < u2 corresponds to the region where the motion of our top lives and gives
rise to a periodic nutation, where the solution bounces between the two turning points.
Depending on the precise value of the various constants that appear in this Vef f this gives
rise to different qualitative motions, with examples shown in Figs. 2.82.11. Recalling that
= (p p u)/[I1 (1 u2 )], we see that the possible signs for will depend on p and p .
In Fig. 2.8 the top nutates between 1 and 2 while always precessing in the same direction

50
 CHAPTER 2. RIGID BODY DYNAMICS

Figure 2.9: has both Figure 2.10: at 2 we Figure 2.11: No nu-

Figure 2.8: > 0
signs have = 0, = 0 tation

with > 0, whereas in Fig. 2.9 the precession is also in the backward direction, < 0, for
part of the range of motion. In Fig. 2.10 the top has = 0 at 2 , before falling back down
in the potential and gaining > 0 again. This figure also captures the case where we let go
of a top at = 2 0 that initially has > 0 but = 0. Finally in Fig. 2.11 we have the
situation where there is no nutation oscillation because the two angles coincide, 1 = 2 .

51
Chapter 3

Vibrations & Oscillations

The topic of vibrations and oscillations is typically discussed in some detail in a course
on waves (at MIT this is 8.03). Our goal for this chapter is to revisit aspects of oscillation
phenomena using generalized coordinates. Many equations of motion we have encountered
have been nonlinear. Here, we will expand about a minimum of the potential V (q1 , . . . , qn ),
yielding linear equations.
Let us take qi = q0i + i , where ~q0 minimizes V (q), and expand in the i . Henceforth and
until further notice, repeated indices will implicitly be summed over. Then
1 2 V 
 
V 
V (q1 , . . . , qn ) = V (q01 , . . . , q0n ) + i + i j + . . . , (3.1)
qi 0 2 qi qj 0

V 
where |0 means evaluate the quantity at ~q0 . We already know that q  = 0 as by
i 0

definition ~q0 minimizes V (q). As a matter of convention, we choose V (q0 ) = 0, since this
just corresponds to picking  the convention for the zero of the Energy. Finally, we define the
2V 
constants Vij qi qj  . There is no time dependence in the definition of our generalized
0
coordinates, so the kinetic energy is
1 1
T = mij (q1 , . . . , qn )qi qj = mij (q01 , . . . , q0n )i j + O( 2 ), (3.2)
2 2
where mij (q01 , . . . , q0n ) Tij are constants, and terms of O( 2 ) and beyond are neglected.
Thus, the Lagrangian to quadratic order in the i s is
1
(Tij i j Vij i j ) .
L= (3.3)
2
From this, the equations of motion are
Tij j + Vij j = 0 (3.4)
to the same order. These are coupled linear equations of motion.

52
 CHAPTER 3. VIBRATIONS & OSCILLATIONS

3.1 Simultaneous Diagonalization of T and V

To solve Eq. (3.4) lets try
i = ai eit (3.5)
where ai C for all i {1, . . . , n}, and the frequency is the same along all directions in
the generalized
coordinate space. Notationally, i and j will denote coordinate indices, while
i = + 1 is the imaginary unit. This gives

Vij aj = 2 Tij aj (3.6)

which can be rewritten in matrix form as

V ~a = T ~a (3.7)

with = 2 . This looks like an eigenvalue equation except that when we act with the linear
operator V on ~a we get back T ~a instead of just the eigenvector ~a. This can be rewritten as

(V T) ~a = 0 (3.8)

where V and T are real and symmetric nn matrices. In order to have a non-trivial solution
of this equation we need
det(V T) = 0 (3.9)
which is an nth order polynomial equation with n solutions eigenvalues with {1, . . . , n}.
The solutions of (V T) ~a() = 0 are the eigenvectors ~a() . This means

V ~a() = T ~a() , (3.10)

and the solutions are much like a standard eigenvalue problem. Here and henceforth, there
will be no implicit sum over repeated eigenvalue indices (so any sums that are needed will
be made explicit), but we will retain implicit sums over repeated coordinate indices i & j.

There are two cases we will consider.

1) Let us start by considering the case when T is diagonal. In particular, let us consider the
even easier case proportional to the unit matrix, where Tij = mij . This means

mi + Vij j = 0 . (3.11)

Here we have the standard eigenvalue equation

V ~a() = m~a() . (3.12)

53
 CHAPTER 3. VIBRATIONS & OSCILLATIONS

The eigenvalues are real and nonnegative as = 2 ; the quantities are the normal
mode frequencies. The eigenvectors ~a() are orthogonal, and we can choose their normaliza-
tion so that
m ~a() ~a() = (or ~a() ~a() = ) . (3.13)
This implies that
= ~a() V ~a() (or m = ~a() V ~a() ) . (3.14)
The time-dependent eigenvectors are then
() ()
~ () = ~a() ei t , or i = ai ei t . (3.15)
These are the normal mode solutions for the n coordinates labeled by i, and there are n such
solutions labeled by . The general solution of a linear equation is a superposition of the
independent normal mode solutions:
X
~ = C ~ () (3.16)

where C C are fixed by initial conditions. To find real coordinate solutions, we take the
real parts of these equations.
Lets prove the statements made above. Again, there will be no implicit sum over the
eigenvalue index . Dotting in ~a() into Eq. (3.12) gives
~a() V ~a() = m~a() ~a() , (3.17)

taking the Hermitian conjugate of both sides, noting that V = V , and then swapping
gives ~a() V ~a() = m?~a() ~a() . Taking the difference of these results gives

( ? )~a() ~a() = 0, (3.18)

and if = then ( ? )~a() ~a() = 0 implies the eigenvalues are real R. For
6= , Eq. (3.18) then implies ~a() ~a() = 0 so the eigenvectors are orthogonal. If by
chance = for some = 6 then we can always simply choose the corresponding eigen-
vectors to be orthogonal. By convention, we then normalize the eigenvectors so that they
satisfy Eq. (3.13). Finally, if = then Eq. (3.17) now gives = ~a() V ~a() . The
statement that we are at a local minimum of the multivariable potential and not a saddle
point or a maximum implies then that 0 (we have positive second derivatives in each
eigenvector direction).

2) Let us now consider when T is not diagonal and summarize which parts of the result
are the same and where there are differences. Here we have (V T) ~a = 0. Again, the
eigenvalues are real and nonnegative, with = 2 . Now, however,
~a() T ~a() = 0 , (3.19)

54
 CHAPTER 3. VIBRATIONS & OSCILLATIONS

for 6= , and we can replace the old normalization condition by a new one stating that

~a() T ~a() = , (3.20)

which up to an overall prefactor reduces to the old orthonormality condition when T = m1.
Here again,

= ~a() V ~a() , (3.21)

and the th normal mode solution is

~ () = ~a() ei t . (3.22)

The general solution is again the superposition

X
~ = C ~ () , (3.23)

with the complex coefficients C fixed by the initial conditions (and a real part taken to get
real coordinates).
Lets repeat the steps of our proof for this case. Dotting ~a() into Eq. (3.10) gives

~a() V ~a() = ~a() T ~a() . (3.24)

Taking the Hermitian conjugate of both sides yields ~a() V ~a() = ?~a() T ~a() .
Subtracting the two results this gives

( ? ) ~a() T ~a() = 0, (3.25)

and if = then ( ? )~a() T ~a() = 0 implies R since T = T? = T> and

physically we know that the kinetic energy T = ~ T ~ > 0 for any ~ 6= 0. For 6= ,
then the condition instead implies ~a() T ~a() = 0 so the eigenvectors are orthogonal; if
by chance = for some 6= then we can choose the corresponding eigenvectors to be
orthogonal. By convention, we normalize the eigenvectors so that they will be orthonormal
as in Eq. (3.20). Finally, if = then = ~a() V ~a() , which is positive, so > 0 also.
The statement that we are at a local minimum of the potential and not a saddle point or a
maximum implies then that 0.

3.2 Vibrations and Oscillations with Normal Coordi-

nates
Given these results, it is natural to ask whether a different set of generalized coordinates
might be better for studying motion about the minimum?

55
 CHAPTER 3. VIBRATIONS & OSCILLATIONS

We form the matrix A by placing the eigenvectors in columns

 
A = ~a(1) ~a(2) . . . ~a(n) (3.26)

and construct a diagonal eigenvalue matrix = diag(1 , 2 , . . . , n ). The matrix A can be

ensured to be real because each ~a only has at most an overall phase1 , and these can be
()

removed by putting them into the coefficients C . The matrix A simultaneously diagonalizes
T and V since
A> TA = 1 and A> V A = . (3.27)
We choose new normal coordinates ~ by letting

~ = A ~ and ~ = A ~ (3.28)

so that the Lagrangian

1 1
L = ~ T ~ ~ V ~ (3.29)
2 2
1 ~ 1
= (A> T A) ~ ~ (A> V A) ~
2 2
1 X  2 
= 2 2 .
2

This gives the simple equations of motion for each :

+ 2 = 0 . (3.30)

Thus, each normal coordinate describes the oscillations of the system with normal mode
frequency .

Example: Let us consider the triatomic molecule CO2 shown in Figure 3.1. We can picture it
as a carbon atom of mass M in the middle of two oxygen atoms each of mass m. For the three
particles there are 9 coordinates given by r1 , r2 , and r3 . Six of these coordinates correspond
to translations and rotations of the mass system treated as a rigid body. This leaves 3
coordinates that correspond to internal motions of the system. To model the potential we
connect each oxygen atom to the carbon atom with a spring of constant k and relaxed
length b. This does not add any cost to relative motion of the atoms with fixed spring
length, which we will address below by adding another potential term in order to favor the
linear configuration.
It is straightforward to guess what the normal modes could be:
1
Why is it just an overall phase? The equation V ~a() = T ~a() alone does not fix the normalization
() () () ()
of ~a . Let us say we pick ai R for some i. Then Vkj aj = Kkj aj is a set of equations with all real
() ()
coefficients and one real term in the sums. Hence the solutions aj /ai R for all j {1, . . . , n}, implying
that at most there is an overall phase in ~a() .

56
 CHAPTER 3. VIBRATIONS & OSCILLATIONS

Figure 3.1: The CO2 molecule.

The oxygen atoms moving in the same direction along the line and the carbon atom
moving in the opposite direction. This is a longitudinal oscillation.

The oxygen atoms opposing each other along the line while the carbon atom remains
at rest. This is a longitudinal oscillation.

The oxygen atoms move in the same direction perpendicular to the line and the carbon
atom moving in the opposite direction. This is a transverse oscillation.

These three normal modes are shown in Figure 3.2.

Figure 3.2: The three Normal Modes of the CO2 molecule

We pick the body frame axes as follows:

the three particles are in the xy-plane fixing 3 coordinates zi = 0 for i {1, 2, 3},

the origin is the CM so m(x1 + x3 ) + M x2 = m(y1 + y3 ) + M y2 = 0, which fixes two

more coordinates,

the axes are oriented so that y1 = y3 , which fixes one coordinate.

m
Defining the mass ratio as M , then x2 = (x1 + x3 ) and y2 = 2y1 can be eliminated.
Altogether this fixes 6 coordinates, leaving the coordinates (x1 , x3 , y1 ). This setup is shown
in Figure 3.3.
For the potential we take

k k b2 2
V = (s1 b)2 + (s2 b)2 + (1 + 22 ) . (3.31)
2 2 2
57
 CHAPTER 3. VIBRATIONS & OSCILLATIONS

Figure 3.3: The orientation of the CO2 molecule on xy plane

The first two terms are the springs discussed previously, and the last two provide a quadratic
energy cost to the springs rotating away from the linear configuration, with strength given
by . The spring lengths are
p q
s1 = (x1 x2 ) + (y1 y2 ) = [x1 + (x1 + x3 )]2 + (1 + 2)2 y12
2 2 (3.32)
p q
s2 = (x2 x3 )2 + (y2 y3 )2 = [x3 + (x1 + x3 )]2 + (1 + 2)2 y12 ,

and the two angles are

y y   
1 3 2 1 (1 + 2)y1
1 = tan = tan , (3.33)
x3 x 2 (1 + )x1 + x3
 
1 y2 y1 (1 + 2)y1
 
1
2 = tan = tan .
x2 x 1 (1 + )x3 + x1

These results give V = V (x1 , x3 , y1 ). [to be continued]

58
 CHAPTER 3. VIBRATIONS & OSCILLATIONS

For the kinetic energy we have

m 2
M m 2
M
T = x1 + x23 + x22 + y1 + y32 + y22 , (3.34)
2 2 2 2
which after eliminating coordinates becomes
m
(1 + ) x21 + x32 + mx1 x3 + m (1 + 2) y12 .


T = (3.35)
2
Equilibrium comes from taking y1 = 0, x3 = x1 = b, which implies 1 = 2 = 0,
s1 = s2 = b, and V = 0. We define coordinates for expanding about this potential minimum 
V 
as 1 = x1 + b, 3 = x3 b, and 2 = y1 . Then as V (b, b, 0) = 0 in equilibrium and i
 =0
0
then V = 12 Vij i j + . . . where
2

2
 k(1 + 2 + 2 ) 0 2k(1 + )
V 
Vij = = 0 2(1 + 2)2 0 (3.36)
i j 0 2
2k(1 + ) 0 k(1 + 2 + 2 )
for this system. Additionally,

m(1 + ) 0 m
Tij = 0 2m(1 + 2) 0 (3.37)
m 0 m(1 + )
for this system. Since there are no off-diagonal terms in the 2nd row or 2nd column in either
V or T, the transverse and the longitudinal modes decouple. For the transverse mode, we
are left with
2(1 + 2)2
y1 + y1 = 0 , (3.38)
2m(1 + 2)
which is a simple harmonic oscillator. For the longitudinal modes, we have ~ = (1 , 3 ). The
frequencies come from
 
k(1 + 2 + 22 ) m(1 + ) 2k(1 + ) m
det =0 (3.39)
2k(1 + ) m k(1 + 2 + 22 ) m(1 + )
The solutions give the normal mode frequencies
k k
1 = 12 = , 2 = 22 = (1 + 2) , (3.40)
m m
with associated eigenvectors
   
(1) 1 1 (2) 1 1
~a = , ~a =p , (3.41)
2m 1 2m(1 + 2) 1

which were chosen to satisfy ~a() T ~a() = . Thus, the normal coordinates for the
longitudinal modes are 1 x1 x3 and 2 x1 + x3 . Oscillations in these coordinates
correspond to the normal mode motions in Fig. 3.2(b) and Fig. 3.2(a) respectively.

59
Chapter 4

Canonical Transformations,
Hamilton-Jacobi Equations, and
Action-Angle Variables

Weve made good use of the Lagrangian formalism. Here well study dynamics with the
Hamiltonian formalism. Problems can be greatly simplified by a good choice of generalized
coordinates. How far can we push this?

Example: Let us imagine that we find coordinates qi that are all cyclic. Then pi = 0, so
pi = i are all constant. If H is conserved, then:

H = H(1 , . . . , n ) (4.1)

is also constant in time. In such a case the remaining equations of motion:

H
qi = = i () qi = i t + i (4.2)
i

All coordinates are linear in time and the motion becomes very simple.

We might imagine searching for a variable transformation to make as many coordinates as

possible cyclic. Before proceeding along this path, we must see what transformations are
allowed.

4.1 Generating Functions for Canonical Transforma-

tions
Recall the the Euler-Lagrange equations are invariant when:

60
 CHAPTER 4. CANONICAL TRANSFORMATIONS, HAMILTON-JACOBI
EQUATIONS, AND ACTION-ANGLE VARIABLES
a point transformation occurs Q = Q(q, t) with L[q, t] = L0 [Q, t];
dF [q, t]
a total derivative is summed to the Lagrangian L0 = L + .
dt
For H we consider point transformations in phase space:

Qi = Qi (q, p, t) and Pi = Pi (q, p, t), (4.3)

where the Hamiltons equations for the evolution of the canonical variables (q, p) are satisfied:
H H
qi = and pi = . (4.4)
pi qi
Generally, not all transformations preserve the equations of motion. However, the trans-
formation is canonical if there exists a new Hamiltonian:

K = K(Q, P, t), (4.5)

where
K K
Qi = and Pi = . (4.6)
Pi Qi
For notational purposes let repeated indices be summed over implicitly.
Hamiltons principle can be written as:
Z t2
(pi qi H(q, p, t)) dt = 0, (4.7)
t1

or in the new Hamiltonian as:

Z t2  
Pi Q i K(Q, P, t) dt = 0. (4.8)
t1

For the Eq.(4.7) to imply Eq.(4.8), then we need:

(pi qi H) = Pi Q i K + F . (4.9)

Since F is a total derivative and the ends of the path are fixed:

q|tt21 = 0 and p|tt21 = 0 F |tt21 = 0 (4.10)

There are a few things to be said about transformations and .

If = 1 then the transformation is canonical, which is what we will study.

If =
6 1 then the transformation is extended canonical, and the results from = 1 can
be recovered by rescaling q and p appropriately.

61
 CHAPTER 4. CANONICAL TRANSFORMATIONS, HAMILTON-JACOBI
EQUATIONS, AND ACTION-ANGLE VARIABLES

If Qi = Qi (q, p) and Pi = Pi (q, p) without explicit dependence on time, then the

transformation is restricted canonical.

We will always take transformations Qi = Qi (q, p, t) and Pi = Pi (q, p, t) to be invertible

in any of the canonical variables. If F depends on a mix of old and new phase space variables,
it is called a generating function of the canonical transformation. There are four important
cases of this.

1. Let us take
F = F1 (q, Q, t) (4.11)

where the old coordinates qi and the new coordinates Qi are independent. Then:

F1 F1 F1
pi qi H = Pi Q i K + F 1 = Pi Q i K + + qi + Qi (4.12)
t qi Qi

from this we see that Pi Q i cancels and equating the terms with a qi , a Q i and the
remaining terms gives:

F1 F1 F1
pi = , Pi = and K=H+ , (4.13)
Qi Qi t

which gives us formula for a transformation:

pi = pi (q, Q, t) and Pi = Pi (q, Q, t) (4.14)

and connects K to an initial H.

Example: if
Q
F1 = , (4.15)
q

then:
F1 Q F1 1
p= = 2 and P = = . (4.16)
q q Q q
Writing the new coordinates as function of the old ones yields

1
Q = pq 2 and P = (4.17)
q

Example: Given the transformations

q2
   
p p
Q = ln and P = +1 , (4.18)
q 2 q

62
 CHAPTER 4. CANONICAL TRANSFORMATIONS, HAMILTON-JACOBI
EQUATIONS, AND ACTION-ANGLE VARIABLES

we can prove they are canonical by finding a corresponding generating function. We

know:
F1
= p = qeQ , (4.19)
q
which gives us
q2 Q
Z
F1 = qeQ dq + g(Q) = e + g(Q), (4.20)
2
and

q2 q2 q2
   
F1 dg p
P = = eQ = +1 = + 1 eQ
Q 2 dQ 2 q 2
g(Q) = eQ . (4.21)

Thus F1 is given by:

q2
 
F1 = + 1 eQ . (4.22)
2

2. Let:
F = F2 (q, P, t) Qi Pi (4.23)
where we wish to treat the old coordinates qi and new momenta Pi as independent
variables. Then:
F2 F2 F2
qi pi H = Q i Pi K + F 2 Q i Pi Qi P i = Qi P i K + + qi + Pi (4.24)
t qi Pi

This corresponds to

F2 F2 F2
pi = ; Qi = and K=H+ . (4.25)
qi Pi t

3. We could also take

F = F3 (p, Q, t) + qi pi (4.26)
with the new coordinates Qi and the old momenta pi as independent variables.

4. Finally we could take

F = F4 (p, P, t) + qi pi Qi Pi (4.27)
with the old momenta pi and new momenta Pi as independent variables.

This can be summarized in the table below.

63
 CHAPTER 4. CANONICAL TRANSFORMATIONS, HAMILTON-JACOBI
EQUATIONS, AND ACTION-ANGLE VARIABLES

Function Transformations Simplest case

F1 F1
F1 (q, Q, t) pi = , Pi = F1 = qi Qi Qi = pi , Pi = qi
qi Qi
F2 F2
F2 (q, P, t) pi = , Qi = F 2 = q i Pi Qi = qi , Pi = pi
qi Pi
F3 F3
F3 (p, Q, t) qi = , Pi = F3 = pi Qi Qi = qi , Pi = pi
pi Qi
F4 F4
F4 (p, P, t) qi = , Qi = F 4 = p i Pi Qi = pi , Pi = qi
pi Pi

The simplest case of the 2nd (F2 ) transformation is just an identity transformation. For any
of these Fi cases we also have:
Fi
K=H+ . (4.28)
t
If Fi is independent of time then this implies

K=H (4.29)

Mixed cases may also occur when more than two old canonical coordinates are present. (In
this chapter we will be using Einsteins repeated index notation for implicit summation,
unless otherwise stated.)

Example: consider
F2 = fi (q, t)Pi (4.30)
for some functions fi where i {1, . . . , n}. Then

F2
Qi = = fi (q, t) (4.31)
Pi
is a coordinate point transformation. It is canonical with

fi
pi = Pj , (4.32)
qj

which can be inverted to get Pj = Pj (q, p, t).

Example: Consider the harmonic oscillator:

p2 kq 2
H= + where k = m 2 (4.33)
2m 2

64
 CHAPTER 4. CANONICAL TRANSFORMATIONS, HAMILTON-JACOBI
EQUATIONS, AND ACTION-ANGLE VARIABLES

Let us try the transformation:

p = 2mP cos(Q) (4.34)

q= 2mP sin(Q)
m
for constant. Then:
K = H = P 2 cos2 (Q) + sin2 (Q) = P 2 ,


(4.35)
so the new momentum
E
P = (4.36)
2
is just proportional to the energy, while Q is a cyclic variable.
Is this transformation canonical? We can find a generating function F = F1 (q, Q) by
dividing the old variables:
p
= m cot(Q). (4.37)
q
This gives us:
Z
F1 1
p= F1 = p(q, Q) dq + g(Q) = mq 2 cot(Q) + g(Q) (4.38)
q 2
F1 mq 2 dg
P = = 2
Q 2 sin (Q) dQ
Setting:
dg 2P
= 0 q2 = sin2 (Q), (4.39)
dQ m

which tells us the transformation is canonical if = . This means:
E
P = (4.40)

By Hamiltons equations Eq.(4.4):
K
Q = = Q = t + . (4.41)
P
Putting this altogether, this gives the familiar results:
r
2E
q= 2
sin(t + ) (4.42)
m
p = 2mE cos(t + ).
Lets record for future use our final canonical transformation here:

r
2P
q= sin(Q) , p = 2mP cos(Q) .
m
So far, a transformation Q = Q(q, p, t) and P = P (q, p, t) is canonical if we can find a
generating function F . This involves integration, which could be complicated, so it would
be nice to have a test that only involves differentiation. There is one!

65
 CHAPTER 4. CANONICAL TRANSFORMATIONS, HAMILTON-JACOBI
EQUATIONS, AND ACTION-ANGLE VARIABLES

4.2 Poisson Brackets and the Symplectic Condition

In Classical Mechanics II (8.223) the Poisson bracket of the quantities u and v was defined
as
X  u v v u

{u, v}q,p = (4.43)
i
qi pi qi pi
It is easy to check that the following fundamental Poisson bracket relations are satisfied:
{qi , qj }q,p = {pi , pj }q,p = 0 and {qi , pj }q,p = ij . (4.44)
There are a few other properties of note. These include:
{u, u} = 0, (4.45)
{u, v} = {v, u}, (4.46)
{au + bv, w} = a{u, w} + b{v, w}, (4.47)
{uv, w} = u{v, w} + {u, w}v, (4.48)
{u, {v, w}} + {v, {w, u}} + {w, {u, v}} = 0, (4.49)
for a, b constants. Eq.(4.49) is the Jacobi identity.
The above looks a lot like the commutators of operators in quantum mechanics, such as:
[x,
p] = i~ (4.50)
Indeed, quantizing a classical theory by replacing Poisson brackets with commutators through:
[u, v] = i~{u, v} (4.51)
is a popular approach (first studied by Dirac). It is also the root of the name canonical
quantization. (Note that Eq.(4.48) was written in a manner to match the analogous formula
in quantum mechanics where the operator ordering is important, just in case its familiar.
Here we can multiply functions in either order.)
Now we can state the desired criteria that only involves derivatives.
Theorem: A transformation Qj = Qj (q, p, t) and Pj = Pj (q, p, t) is canonical if and only if:
{Qi , Qj }q,p = {Pi , Pj }q,p = 0 and {Qi , Pj }q,p = ij . (4.52)
To prove it, well need some more notation. Lets get serious about treating qi and pi on an
equal footing together, defining the following two quantities:

q1
..
.
0nn 1nn

qn
~ =

and J =

(4.53)
p1

1nn 0nn
..


.
pn

66
 CHAPTER 4. CANONICAL TRANSFORMATIONS, HAMILTON-JACOBI
EQUATIONS, AND ACTION-ANGLE VARIABLES

where 0nn is the n n zero matrix, 1nn is the n n identity matrix. The following
properties of J will be useful:

J > = J , J 2 = 12n2n and J > J = JJ > = 12n2n . (4.54)

We also note that det(J) = 1.

With this notation Hamiltons equations, Eq.(4.4), can be rewritten as:

H
~ = J or ~ = J ~ H . (4.55)
~
The notation ~ H better emphasizes that this quantity is a vector, but we will stick to using
the first notation for this vector, H/~ , below.
Although the Theorem is true for time dependent transformations, lets carry out the proof
for the simpler case of time independent transformations Qi = Qi (q, p) and Pi = Pi (q, p).
This implies K = H. Let us define:

Q1
..

.
Qn
~ =



(4.56)
P1
..

.

Pn

which is a function of the original coordinates, so we can write:

~ = (~
~ ) (4.57)

~
Now consider the time derivative of :
i i
i = j ~ = M~ where Mij = . (4.58)
j j

Here M corresponds to the Jacobian of the transformation.

From the Hamiltons equations, we know that
H
~ = J . (4.59)
~
We want to show that :
H
~ = J for ~ = (~
~ ) a canonical transformation. (4.60)
~

67
 CHAPTER 4. CANONICAL TRANSFORMATIONS, HAMILTON-JACOBI
EQUATIONS, AND ACTION-ANGLE VARIABLES

Let us now consider:

i i H i l H H
i = j = Jjk = Jjk ~ = M JM > (4.61)
j j k j k l ~

for any H. Then

~ = (~
~ ) is a canonical transformation iff M JM > = J (4.62)

is satisfied. This is known as the symplectic condition. Moreover, since

M J = J(M > )1 and J 2 = 1, (4.63)

we can write:

J(M J)J = JM = J(JM >1 )J = M >1 J JM = M >1 J . (4.64)

Thus we see that M JM > = J is equivalent to:

M > JM = J . (4.65)

Now consider Poisson brackets in this matrix notation:

 >
u v
{u, v}q,p = {u, v}~ = J (4.66)
~ ~

and the fundamental Poisson brackets are:

{i , j }~ = Jij (4.67)

Then we can calculate the Poisson brackets that appeared in the theorem we are aiming to
prove as
 >
i j
{i , j }~ = J = (M > JM )ij (4.68)
~ ~
This last equation is the same as Eq.(4.65). The new variables satisfy the Poisson bracket
relationships Eq.(4.67):
{i , j }~ = Jij (4.69)
if and only if
M > JM = J (4.70)

which itself is true if, and only if, ~ = (~

~ ) is canonical, Eq.(4.65), completing the proof.
There are two facts that arise from this.

68
 CHAPTER 4. CANONICAL TRANSFORMATIONS, HAMILTON-JACOBI
EQUATIONS, AND ACTION-ANGLE VARIABLES

Poisson brackets are canonical invariants

{u, v}~ = {u, v}~ = {u, v}. (4.71)

This is true because:

  > >  
u > u v > v
{u, v}~ = M J =
J M (4.72)
~ ~~ ~
 >  >
u v u v
= M JM > = J = {u, v }~ (4.73)
~
~
~ ~

Phase space volume elements are preserved by canonical transformations, as discussed

in 8.223. Phase space volume is given by:
Y Y
V~ = dQi dPi = |det(M )| dqj dpj = |det(M )|V~ . (4.74)
i j

However, we also have:

det(M > JM ) = det(J) = (det(M ))2 det(J) |det(M )| = 1. (4.75)

4.3 Equations of Motion & Conservation Theorems

Let us consider a function:
u = u(q, p, t) (4.76)
Then:
u u u u H u H u
u = qi + pi + = + , (4.77)
qi pi t qi pi pi qi t
which can be written more concisely as

u
u = {u, H} + (4.78)
t
for any canonical variables (q, p) and corresponding Hamiltonian H. Performing canonical
quantization on this yields the Heisenberg equation of time evolution in quantum mechanics.
There are a few easy cases to check.

If u = qi then:
qi H
qi = {qi , H} + = (4.79)
t pi

69
 CHAPTER 4. CANONICAL TRANSFORMATIONS, HAMILTON-JACOBI
EQUATIONS, AND ACTION-ANGLE VARIABLES

If u = pi then:
pi H
pi = {pi , H} + = (4.80)
t qi
Together the above two cases yield Hamiltons equations of motion.

Also, if u = H then:
H H
H = {H, H} + = (4.81)
t t

Next consider what we can say if a quantity u is conserved. Here:

u
u = 0 = {u, H} + . (4.82)
t
As a corollary, if
u
= 0, (4.83)
t
then
{u, H} = 0 u is conserved. (4.84)
(In quantum mechanics this the analog of saying that u is conserved if u commutes with H.)
Another fact, is that if u and v are conserved then so is {u, v}. This could potentially
provide a way to compute a new constant of motion. To prove it, first consider the special
case where:
u v
= =0 (4.85)
t t
then using the Jacobi identity we have:

{H, {u, v}} = {u, {v, H}} {v, {H, u}} = {u, 0} {v, 0} = 0 (4.86)

For the most general case we proceed in a similar manner:

   
v u
{{u, v}, H} = {u, {v, H}} + {v, {H, u}} = u, + v,
t t
d
= {u, v} {u, v } = 0 (4.87)
t dt
Infinitesimal Canonical Transformations

Let us now consider the generating function:

F2 (q, P, t) = qi Pi + G(q, P, t), (4.88)

70
 CHAPTER 4. CANONICAL TRANSFORMATIONS, HAMILTON-JACOBI
EQUATIONS, AND ACTION-ANGLE VARIABLES

where F2 = qi Pi is an identity transformation, and ||  1 is infinitesimal. The function

G(q, P, t) is known as the generating function of an infinitesimal canonical transformation.
Using the properties of an F2 generating function we have:
F2 G G
pj = = Pj +  pj = Pj pj =  (4.89)
qj qj qj

giving the infinitesimal transformation in the momentum. Likewise:

F2 G
Qj = = qj +  (4.90)
Pj Pj

Since Pj = pj + O() and  is infinitesimal we can replace G(q, P, t)/Pj = G(q, p, t)/pj +
O(). Therefore we have:
G G
Qj = qj +  + O(2 ) qj = Qj qj =  (4.91)
pj pj

where now we note that we can consider G = G(q, p, t), a function of q and p, to this order.
Returning to the combined notation of ~ > = (q1 , . . . , qn , p1 , . . . , pn ), Eq.(4.89) and Eq.(4.90)
can be consisely written as the following Poisson bracket:

~ = {~ , G} (4.92)

Example: if G = pi then pi = 0 and qj = ij , which is why momentum is the generator

of spatial translations.

Example: if G is the z component of the angular momentum:

X
G = Lz = (xi piy yi pix ) and  = (4.93)
i

then the infinitesimal change correponds to a rotation

xi = yi , yi = xi , zi = 0 (4.94)
pix = piy , piy = pix , piz = 0 (4.95)

which is why angular momentum is the generator of rotations.

Important Example: if G = H and  = dt then

{~ , G} = {~ , H }dt = ~ dt = d~

On the left hand side we have the change to the phase space coordinates due to our transfor-
mation. On the right hand side we have the physical increment to the phase space variables

71
 CHAPTER 4. CANONICAL TRANSFORMATIONS, HAMILTON-JACOBI
EQUATIONS, AND ACTION-ANGLE VARIABLES

that takes place in time dt. The fact that these are equivalent tells us that the Hamilto-
nian is the generator of time evolution. The infinitesimal transformation generated by the
Hamiltonian is corresponds with the physical motion.
Rather than trying to think of this as a passive coordinate change (q, p) (Q, P ), it is
useful if we can take an active view of the infinitesimal canonical transformation generated
by H. Let the time t be a parameter for the family of transformations with  = dt: the
initial conditions are:
~0 (t0 ) = ~0 (4.96)
The result is a series of transformations of ~ that move us in a fixed set of phase space
coordinates from one point to another:

(t3 )
(t2 ) (tn )

(t1 ) ~0 (t0 ) ~1 (~0 , t1 ) . . . ~n (~n1 , tn )

where tn = t is the final time (4.97)
(t0 )

Phase Space 
All together, combining an infinite number of infinitesimal transformations allows us to make
a finite transformation, resulting in:
~ = ~ (~0 , t) or ~0 = ~0 (~ , t) (4.98)
This is a canonical transformation that yields a solution for the motion!
How could we directly find this transformation, without resorting to stringing together
infinitesimal transformations? We can simply look for a canonical transformation with new
coordinates Qi and new momenta Pi that are all constants, implying an equation of the type:
~0 = ~0 (~ , t) (4.99)
Inverting this then gives the solution for the motion.
This logic can be used to extend our proof of the Theorem in Section 4.2 to fully account
for time dependent transformations. (see Goldstein). Using K = H +G/t, Goldstein also
describes in some detail how the change to the Hamiltonian H under an active infinitesimal
canonical transformation satisfies:
G
H = {G, H}  = G (4.100)
t
This says the constants of motion are generating functions G of the infinitesimal canonical
transformation that leave H invariant; that is, G = 0 if and only if H = 0 under the
transformation. Thus a conservation law exists if and only if there is a symmetry present.

72
 CHAPTER 4. CANONICAL TRANSFORMATIONS, HAMILTON-JACOBI
EQUATIONS, AND ACTION-ANGLE VARIABLES

4.4 Hamilton-Jacobi Equation

Let us take the suggestion from the end of the previous section seriously and look for new
canonical variables that are all cyclic, such that:
Q i = P i = 0 (Q, P ) are all constants . (4.101)
If the new Hamiltonian K is independent of (Q, P ) then:
K K
Qi = =0 and Pi = = 0. (4.102)
Pi Qi
We could look for a constant K, but it it is simplest to simply look for K = 0.
Using a generating function F = F2 (q, P, t), then we need
F2
K = H(q, p, t) + = 0. (4.103)
t
F2
Because pi = , then we can rewrite this condition as
qi
 
F2 F2 F2
H q1 , . . . , q n , ,..., ,t + = 0, (4.104)
q1 qn t
which is the time dependent Hamilton-Jacobi equation (henceforth abbreviated as the H-J
equation). This is a 1st order partial differential equation in n + 1 variables (q1 , . . . , qn , t)
for F2 . The solution for F2 has n + 1 independent constants of integration, One of these
constants is trivial (F2 F2 + C for a pure constant C), so well ignore this one. Hence,
suppose the solution is:
F2 S = S(q1 , . . . , qn , 1 , . . . n , t), (4.105)
where S is called Hamiltons principal function and each i is an independent constant. We
can pick our new momenta to be the constants of integration Pi = i for i {1, . . . , n} (so
that Pi = 0), thus specifying F2 = F2 (q, P, t) as desired. Then, using again the property of
an F2 generating function (and K = 0), we have that the new constant variables are:
S(q, , t)
Pi i and Qi i = . (4.106)
i
We introduce the notation i to emphasize that these are constants.
From these results we can obtain a solution for the motion as follows. From the invert-
ibility of our transformations we have:
S
i (q, , t) = qi = qi (, , t) , (4.107)
i
S
pi (q, , t) = pi = pi (q, , t) = pi (q(, , t), , t) = pi (, , t) .
qi

73
 CHAPTER 4. CANONICAL TRANSFORMATIONS, HAMILTON-JACOBI
EQUATIONS, AND ACTION-ANGLE VARIABLES

(Note that function argument notation has been abused slightly here since pi (q, , t) and
pi (, , t) are technically different functions of their three arguments. Since we are always
sticking explicit variables into the slots this should not cause confusion.) If desired, we can
also swap our 2n constants i and i for 2n initial conditions qi0 and pi0 , to obtain a solution
for the initial value problem. We obtain one set of constants in terms of the other set by
solving the 2n equations obtained from the above results at t = t0 :

qi0 = qi (, , t0 ) , pi0 = pi (, , t0 ) . (4.108)

Thus we see that Hamiltons principal function S is the generator of canonical transforma-
tions of constant (Q, P ), and provides a method of obtaining solutions to classical mechanics
problems by way of finding a transformation.
There are a few comments to be made about this.

1. The choice of constants i is somewhat arbitrary, as any other independent choice

i = i () is equally good. Thus, when solving the H-J equation, we introduce the
constants i in whatever way is most convenient.

2. What is S? We know that:

S S S
S = qi + Pi + , (4.109)
qi Pi t

but we also know that:

S S
= pi , Pi = 0 and = H (4.110)
qi t

Putting Eq.(4.109) and Eq.(4.110) together we have:

Z
S = pi qi H = L S= L dt (4.111)

Thus S is the classical action which is an indefinite integral over time of the Lagrangian
(so it is no coincidence that the same symbol is used).

3. The H-J equation is also the semiclassical limit of the quantum mechanical Schrodinger
equation (0th order term in the WKB approximation). To see this consider the
Schrodinger equation

 ~2 2 
i~ = H = + V (q) , (4.112)
t 2m q 2

74
 CHAPTER 4. CANONICAL TRANSFORMATIONS, HAMILTON-JACOBI
EQUATIONS, AND ACTION-ANGLE VARIABLES

with the wavefunction = exp(iS/~). At this point we are just making a change of
variable, without loss of generality, and S(q, t) is complex. Plugging it in, and canceling
an exponential after taking the derivative, we find
S i~ 2 S 1  S 2
= + + V (q) . (4.113)
t 2m q 2 2m q
This equation is just another way of writing the Schrodinger equation, to solve for a
complex S instead of . If we now take ~ 0 then we find that the imaginary term
goes away leaving
S 1  S 2 S  S 
0= + + V (q) = + H q, , (4.114)
t 2m q t q
which is the Hamilton-Jacobi equation for S with a standard p2 /2m kinetic term in H.
Having set things up, it is always good for us to test a new formalism on an example
where we know the solution.

Example: let us consider the harmonic oscillator Eq.(4.33):

1
p2 + (mq)2 = E


H= (4.115)
2m
Here we will look for one constant P = and one constant Q = . The H-J equation says
 2 !
1 S 2 S
+ (mq) + = 0. (4.116)
2m q t

In solving this, we note that the dependence of S on q and t is separable

S(q, , t) = W (q, ) + g(, t), (4.117)
which gives: !
 2
1 W g
+ (mq)2 = = . (4.118)
2m q t
Since the left side is independent of t and the right hand side is independent of q, then the
result must be equal to a separation constant that is independent of q and t. We will
choose our new P = . Now we have
g
= g = t (4.119)
t
where we have avoided the addition of a further additive constant (since our convention was
to always drop an additive constant when determining S). To identify what is note that
S g
H= = = , (4.120)
t t

75
 CHAPTER 4. CANONICAL TRANSFORMATIONS, HAMILTON-JACOBI
EQUATIONS, AND ACTION-ANGLE VARIABLES

which corresponds to the constant energy,

= E. (4.121)

The other equation we have to solve is

 2 !
1 W
+ (mq)2 = , (4.122)
2m q

so rearranging and integrating this gives the indefinite integral

Z p
W = 2m (mq)2 dq , (4.123)

which we will leave unintegrated until we must do so. The full solution is then given by:
Z p
S = t 2m (mq)2 dq. (4.124)

With this result for Hamiltons Principal function in hand we can now solve for the equations
of motion. The equations of motion come from (we now do the integral, after taking the
partial derivative):
r !
m 2
Z
S dq 1
= = t m p t + = arcsin q . (4.125)
2m (mq)2 2

Inverting gives: r
2
q= sin((t + )), (4.126)
m 2
so is related to the phase. Next we consider p and use this result to obtain:
S p
p= = 2m (mq)2 = 2m cos((t + )) (4.127)
q
These results are as expected. We can trade (, ) for the initial conditions (q0 , p0 ) at t = 0.
The choice of phase (from shifting so that + ) allows taking the positive sign
of each square root in the solutions above.
Separation of variables is the main technique to solve the H-J equation. In particular,
for a time independent H where
H
H = =0 (4.128)
t
we can always separate time by taking:

S(q, , t) = W (q, ) 1 t, (4.129)

76
 CHAPTER 4. CANONICAL TRANSFORMATIONS, HAMILTON-JACOBI
EQUATIONS, AND ACTION-ANGLE VARIABLES

where 1 has been chosen as the separation constant, then plugging this into the time
dependent H-J equation yields (just as in our Harmonic Oscillator example):
 
W W
H q1 , . . . , q n , ,..., = 1 . (4.130)
q1 qn
This result is referred to as the time independent Hamilton-Jacobi equation. Since H = 0,
H is conserved, and equal to a constant 1 . If H = E then this constant is energy, 1 = E.
The function W is called Hamiltons characteristic function.
The idea is now to solve the time independent H-J equation for W = W (q, ) where P =
still. If we follow the setup from our time dependent solution above then the equations of
motion are obtained from the following prescription for identifying variables:
W
pi = for i {1, . . . , n} , (4.131)
qi
S W
Q 1 = 1 = = t,
1 1
W
Qj = j = for j {2, . . . , n} for n > 1 .
j
Here all the Qi are constants.
There is an alternative to the above setup, which allows us to not refer to the time
dependent solution. The alternative is to consider W = F2 (q, P ) as the generating function,
instead of S and only demand that all the new momenta Pi are constants with P1 = 1 = H
for a time independent Hamiltonian H. At the start of chapter 4 we saw that this less
restrictive scenario would lead to Qs that could have a linear time dependence, which is still
pretty simple.
This is almost identical to the above setup but we rename and reshuffle a few things.
The following three equations are the same as before:
 
W W
pi = , Pi = i and H q, = 1 (4.132)
qi q
However, now we have a non-zero K and different equation for Q1 :
W
K = H = 1 and Qi = for all i {1, . . . , n}. (4.133)
i
This means:
K W
Q1 = = 1 Q1 = t + 1 = (4.134)
1 1
which is Eq. (4.131) but rearranged from the perspective of Q1 . For j > 1, the equations are
the same as before Eq.(4.131):
K W
Qj = =0 Qj = j = (4.135)
j j

77
 CHAPTER 4. CANONICAL TRANSFORMATIONS, HAMILTON-JACOBI
EQUATIONS, AND ACTION-ANGLE VARIABLES

In this language we do not need to mention S only W . There are a few comments to be
made:

1. Again, the choice of is arbitrary, and i = i () is fine. If we do replace 1 = 1 ()

K
then Q i = = vi is a constant so that (potentially) all of the Qi become linear in
i
time:
Qi = vi t + i for all i {1, . . . , n} (4.136)

2. What is W ? We know that:

Z Z
W
W = qi = pi qi W = pi qi dt = pi dqi , (4.137)
qi

which is a different sort of action.

3. The time independent H-J equation has some similarity to the time-independent Schrodinger
energy eigenvalue equation (both involve H and and constant E, but the former is a
non-linear equation for W , while the latter is a linear equation for the wavefunction
).

The H-J method is most useful when there is a separation of variables in H.

Example: if
H = h1 (q1 , q2 , p1 , p2 ) + h2 (q1 , q2 , p1 , p2 )f (q3 , p3 ) = 1 , (4.138)
so that q3 is separable, then
1 h 1
f (q3 , p3 ) = (4.139)
h2
is a constant because the right hand side is independent of q3 and p3 . Thus we assign

f (q3 , p3 ) = 2 (4.140)

for convenience. We can then write:

 
0 W3
W = W (q1 , q2 , ) + W3 (q3 , ) f q3 , = 2 and (4.141)
q3
W 0 W 0 W 0 W 0
   
h1 q1 , q2 , , + 2 h2 q1 , q2 , , = 1 (4.142)
q1 q2 q1 q2

Here, q1 and q2 may or may not be separable.

If all variables are separable then we use the solution:
X
W = Wi (qi , ) (4.143)
i

78
 CHAPTER 4. CANONICAL TRANSFORMATIONS, HAMILTON-JACOBI
EQUATIONS, AND ACTION-ANGLE VARIABLES

We can simply try a solution of this form to test for separability.

Note that cyclic coordinates are always separable.
Proof: let us say that q1 is cyclic. Then
 
W W W
p1 and H q2 , . . . , qn , , ,..., = 1 , (4.144)
q1 q2 qn
where is constant. Let us now write

W (q, ) = W1 (q1 , ) + W 0 (q2 , . . . , qn , ). (4.145)

This gives us:

W1
p1 = = W1 = q1 . (4.146)
q1
Which gives us:
W (q, ) = q1 + W 0 (q2 , . . . , qn , ) (4.147)
This procedure can be repeated for any remaining cyclic variables.
Note that the choice of variables is often important in finding a result that separates. A
problem with spherical symmetry may separate in spherical coordinates but not Cartesian
coordinates.

4.5 Kepler Problem

As an extended example, let us consider the Kepler problem of two masses m1 and m2 in a
central potential (with the CM coordinate R = 0). The Lagrangian is:
1 1 1 1
L = mr2 V (r) where + , (4.148)
2 m m1 m2
and here m is the reduced mass. Any V (r) conserves L = r p, so the motion of r and p
is in a plane perpendicular to L. The coordinates in the plane can be taken as (r, ), so:
m 2 2 2

L= r + r V (r), (4.149)
2
with being cyclic, which implies:

p = mr2 is a constant. (4.150)

In fact p = |L| `. Notationally, we use ` for the magnitude of the angular momentum L
to distinguish it from the Lagrangian L.
The energy is then:
m `2
E = r2 + + V (r), (4.151)
2 2mr2

79
 CHAPTER 4. CANONICAL TRANSFORMATIONS, HAMILTON-JACOBI
EQUATIONS, AND ACTION-ANGLE VARIABLES

which is constant, and this can be rewritten as:

m 2 `2
E= r + Veff (r) where Veff (r) = V (r) + , (4.152)
2 2mr2
where Veff is the effective potential, as plotted below for the gravitational potential.
Veff(r)

Figure 4.1: Plot of the effective potential Vef f along with the different qualitative orbits
allowed in a gravity-like potential. The points correspond to turning points of the orbit.

Writing the E-L equation for r = dr/dt = . . . and then solving for it as dt = dr/(. . .),
and integrating yields
Z r
dr0
t = t(r) = q (4.153)
2 `2


r0 0
E V (r ) 2mr02
m

as an implicit solution to the radial motion.

The orbital motion comes as r = r() or = (r) by using Eq.(4.150) and substituting,
in Eq.(4.153). We have = d/dt = `/(mr2 ), so we can use this to replace dt by d in
dt = dr/(. . .) to get an equation of the form d = dr/(. . .). The result is given by
Z r
dr0
0 = ` q (4.154)
`2


r0 r 02 2m E V (r0 ) 2mr 02

80
 CHAPTER 4. CANONICAL TRANSFORMATIONS, HAMILTON-JACOBI
EQUATIONS, AND ACTION-ANGLE VARIABLES

k
In the particular case of V (r) = , the solution of the orbital equation is:
r
1 mk
= 2 (1 + cos( 0 )) (4.155)
r() `

where the eccentricity is given by:

r
2E`2
1+ (4.156)
mk 2
Below are plotted the different qualitative orbits for the gravitic potential, with different
or E (circular, elliptical, parabolic, and hyperbolic respectively).

=0 , E = Emin

0 < <1 , E <0

=1 , E =0

>1 , E >0

Figure 4.2: Different Orbits for the gravity-like potential. The orbits colors match those of
Fig.(4.1). The unbounded orbits occur for E 0. The different curves correspond to the
different possible conic sections.

Consider solving this problem instead by the H-J method. Lets start by considering as
the variables (r, ) so that we assume that the motion of the orbit is in a plane. Here

p2
 
1 2
H= p + + V (r) = 1 = E . (4.157)
2m r r2

81
 CHAPTER 4. CANONICAL TRANSFORMATIONS, HAMILTON-JACOBI
EQUATIONS, AND ACTION-ANGLE VARIABLES

As is cyclic, then p is constant. Using:

W = W1 (r) + , (4.158)

then the time independent H-J equation is:

2 !
2

1 W1
+ 2 + V (r) = 1 . (4.159)
2m r r

This is simplified to s
W1 2
= 2m(1 V (r)) 2 (4.160)
r r
and solved by s
2
Z
W = + 2m(1 V (r)) dr. (4.161)
r2
The transformation equations are:
Z
W dr
t + 1 = =m q ,
1 2

2m(1 V (r)) r2
Z
W dr
2 = = q . (4.162)
2

r2 2m(1 V (r)) r2

Thus we immediately get the radial equation t = t(r) and orbital equation = (r) from
this, with = ` and 1 = E, showing that the constants are physically relevant parameters.
Lets solve this problem again, but suppose the motion is in 3 dimensions (as if we did
not know the plane of the orbit). Using spherical coordinates (r, , ) this corresponds to

p2 p2
 
1 2
H= p + + + V (r) = 1 . (4.163)
2m r r2 r2 sin2
Lets try a separable solution

W = Wr (r) + W () + W () . (4.164)

Since is cyclic we know it is separable and that:

W () = . (4.165)

Together, this leaves;

2 2 !
2
 
Wr 1 W
+ 2 + + 2mV (r) = 2m1 . (4.166)
r r sin2 ()

82
 CHAPTER 4. CANONICAL TRANSFORMATIONS, HAMILTON-JACOBI
EQUATIONS, AND ACTION-ANGLE VARIABLES

Because the term: 2

2

W
+ (4.167)
sin2 ()
only depends on while the rest of the equation depends on r, it must be a constant so we
can say: 2
2

W
+ 2 (4.168)
sin2 ()
and the separation works. This then gives:
2
2

Wr
+ 2 = 2m(1 V (r)), (4.169)
r r
which is the same equation we considered before when assuming the motion was in a plane,
with 1 = E and = `. Eq. (4.168) says that
p2
p2 + 2 = `2 . (4.170)
sin ()
Here p is the constant angular momentum about the z axis.

4.6 Action-Angle Variables

For many problems, we may not be able to solve analytically for the exact motion or for
orbital equations, but we can still characterize the motion. For periodic systems we can find
the frequency by exploiting action-angle variables.
The simplest case is for a single dimension of canonical coordinates (q, p). If H(q, p) = 1
then p = p(q, 1 ). There are two types of periodic motion to consider.
1. Libration (oscillation) is characterized by a closed phase space orbit, so that q and p
evolve periodically with the same frequency.

Figure 4.3: Phase space orbit of a libration (oscilation). The trajectory closes on itself, the
state returns to the same position after some time .

83
 CHAPTER 4. CANONICAL TRANSFORMATIONS, HAMILTON-JACOBI
EQUATIONS, AND ACTION-ANGLE VARIABLES

2. Rotation is characterized by an open phase space path, so p is periodic while q evolves

without bound.

Figure 4.4: Phase space orbit of a rotation. Although the orbit is not closed, each period the
evolution of the system is the same, leading to a orbit that repeats itself with a translation.

Example: a pendulum of length a may be characterized by canonical coordinates (, p ),

where:
p2
E=H= mg a cos (4.171)
2ma2
This means: p
p = 2ma2 (E + mga cos ) (4.172)
must be real. A rotation occurs when E > mga, and oscillations occur when E < mga. The
critical point in between (when the pendulum just makes it to the top) is when E = mga
exactly, and is depicted by a dashed line in the figure below.

Rotation

Libration/Oscilation

3.142 3.142

Figure 4.5: The pendulum exhibits both librations and rotations depending on the initial
conditions.

84
 CHAPTER 4. CANONICAL TRANSFORMATIONS, HAMILTON-JACOBI
EQUATIONS, AND ACTION-ANGLE VARIABLES

For either type of periodic motion, it is useful to replace P = 1 by a different (constant)

choice called the action variable I
J = p dq, (4.173)
H
where refers to a definite integral over one period in phase space. To see that J is
constant, recall that p = p(q, 1 ), so plugging this into the definite integral we are left
with J = J(1 ). Also, we have the inverse, 1 = H = H(J), and can rewrite Hamiltons
characteristic function in terms of J by W = W (q, 1 ) = W (q, H(J)) = W (q, J) (where
again the argument notation is abused slightly).
The coordinate conjugate to J is the angle variable
W
= (4.174)
J
(where is not meant to imply an angular velocity). This means
H(J)
= = (J) is a constant. (4.175)
J
As a result the angle variable has linear time dependence,

= t + , (4.176)

for some initial condition . Dimensionally, J has units of angular momentum, while has
no dimensions (like an angle or a phase).
To see why it is useful to use the canonical variables (, J), let us consider the change in
when q goes through a complete cycle.
I I 2 I I
W W
= dq = dq = dq = p dq = 1 (4.177)
q qJ J q J
where in the last equality we used the definition of J in Eq.(4.173). Also, we have =
where is the period. Thus
1
= (4.178)

is the frequency of periodic motion. If we find H = H(J) then
H(J)
= (4.179)
J
immediately gives the frequency = (J) for the system. Often, we then J = J(E) to
get = (E) the frequency at a given energy. This is a very efficient way of finding the
frequency of the motion without solving for extraneous information.
Example: let us consider a pendulum with action-angle variables. We define:
E
E = (4.180)
mga

85
 CHAPTER 4. CANONICAL TRANSFORMATIONS, HAMILTON-JACOBI
EQUATIONS, AND ACTION-ANGLE VARIABLES

so that E > 1 corresponds to rotation and E < 1 corresponds to oscillation. This means
p p
p = 2m2 ga3 E + cos . (4.181)

For E > 1: Z p
p
2
J = 2m ga3 d E + cos , (4.182)

which is an elliptic integral. For E < 1:

p Z +0 p p Z 0 h p i
J = 2m ga 2 3 2
d E + cos + 2m ga3 d E + cos
0 0
p Z 0 p
= 4 2m2 ga3 d E + cos , (4.183)
0

as the contributions from the four intervals that the pendulum swings through in one period
are all equivalent. Here 0 is the turning point of the oscillation, and E = cos(0 ).
From this:  1
E J
= = (4.184)
J E
which we can solve graphically by making a plot of J vs E, then dJ/dE versus E, and finally
the inverse = dE/dJ versus E.

Figure 4.8: Plot of the fre-

Figure 4.6: Plot of J(E) dJ
The discontinuity Figure 4.7: Plot of . quency of oscillation (E)
versus E. dE
versus E. As E 1
corresponds to the transi- The discontinuity is loga-
rithmic divergent so it is in- we approach the small am-
tion from Oscilation to Ro- plitudeplimit, where =
tation. tegrable.
(2)1 g/a.

Example: let us consider the limit ||  1 of a pendulum, so:

p2 mga 2
H= 2
+ mga (4.185)
2ma 2

86
 CHAPTER 4. CANONICAL TRANSFORMATIONS, HAMILTON-JACOBI
EQUATIONS, AND ACTION-ANGLE VARIABLES

We can actually consider this in the context of a general harmonic oscillator

p2 m2 2
H= + x, (4.186)
2m 2
where: r
g p
= , x = a and p = (4.187)
a a
Notationally, is used for the harmonic oscillator frequency to distinguish from the trans-
formed angle variable . We then have:
I I
J = p dq = 2mE m2 2 x2 dx (4.188)

Note that the coordinate does not need to be an angle, as may be the case for general x.
This gives:
Z x0 s
x2 2mE
J = 4 2mE 1 2 dx where x0 . (4.189)
0 x0 m
Solving the integral yields
2mE 2E
J = 2mEx0 = = , (4.190)
m
which gives us
E
= , (4.191)
J 2
the expected cyclic frequency for the harmonic oscillator.

Multiple Variables: We can treat multiple variables (q1 , . . . , qn , p1 , . . . , pn ) with the action-
angle formalism if each pair (qi , pi ) has an oscillatory or rotating orbit. Lets also assume
that the H-J equation is completely separable into:
X
W = Wj (qj , ). (4.192)
j

Here we have
I
Wi
pi = = pi (qi , 1 , . . . , n ) Ji = pi dqi = Ji (1 , . . . , n ) (4.193)
qi
where repeated indices do not correspond to implicit sums here. This implies that the inverse
will be j = j (J1 , . . . , Jn ) and thus 1 = H = H(J1 , . . . , Jn ). Likewise:
W X Wj
i = = = i (q1 , . . . , qn , J1 , . . . , Jn ) . (4.194)
Ji j
Ji

87
 CHAPTER 4. CANONICAL TRANSFORMATIONS, HAMILTON-JACOBI
EQUATIONS, AND ACTION-ANGLE VARIABLES

Just as in the one dimensional case the time derivative of the angle variables is a constant
H
i = = i (J1 , . . . , Jn ) (4.195)
Ji
which are the frequencies describing motion in this multi-periodic system. Due to the
presence of multiple frequencies, the motion through the whole 2n-dimensional phase space
need not be periodic in time.
Example: in the 2-dimensional harmonic oscillator:
x = A cos(21 t) and y = B cos(22 t) (4.196)
px = mx and py = my (4.197)
1
The overall motion is not periodic in time unless is a rational number.
2
Kepler Problem Example:
Let us do a more extended and detailed example. Returning to the Kepler problem:
k
V (r) = (4.198)
r
with its separable W :
W = Wr (r, ) + W (, ) + W (, ) . (4.199)
If we take E < 0, we have oscillation in r and , along with a rotation in . In particular
from solving our earlier differential equations for W and Wr , we have
W =
Z s
2
W = 2 d
sin2 ()
r
2
Wr = 2m(1 V (r)) dr
r2
Here we have I I I
W
J = p d = d = d (4.200)

For the cyclic variable , we still call the constant p periodic and will take the period to
be 2 (arbitrarily since any period would work), which corresponds to particle returning to
the original point in space. Thus
J = 2 , (4.201)
where is the angular momentum about z.
Continuing, in a similar manner we have
s
2
I I I
W
J = p d = d = 2 d (4.202)
sin2 ()

88
 CHAPTER 4. CANONICAL TRANSFORMATIONS, HAMILTON-JACOBI
EQUATIONS, AND ACTION-ANGLE VARIABLES

Let us call:

cos() (4.203)

the angular momentum fraction. Then:
s
cos2 ()
I
J = 1 d (4.204)
sin2 ()

If we let sin(0 ) = cos(), then p = 0 at the turning points, {0 , 0 }, as expected.

Here one oscillator goes from 0 0 when p > 0, and in reverse for p < 0.
Moreover, sin()2 is even about = 2 . This gives
Z 0 s
cos2 ()
J = 4 1 2 d. (4.205)

2
sin ()

Making two more substitutions

cos() sin() sin() , and then u tan() , (4.206)

after some work the expression becomes

Z 
cos2 ()

1
J = 4 du = 2 (1 cos())
0 1 + u2 1 + u2 cos2 ()
= 2( ), (4.207)

which gives
J + J = 2 . (4.208)
Finally we can consider
I r
(J + J )2
Jr = 2mE 2mV (r) dr (4.209)
4 2 r2
We can immediately make some observations. We observe that Jr = Jr (E, J + J ) is a
function of two variables for any V = V (r), and thus if we invert E = E(Jr , J + J ). This
implies:
E E
= = (4.210)
J J
The two frequencies are degenerate for any V = V (r).
For the V (r) = kr1 potential, the integration can be performed (for example, by
contour integration) to give (for E < 0)
r r
2m 2m
Jr = (J + J ) + k Jr + J + J = k . (4.211)
E E

89
 CHAPTER 4. CANONICAL TRANSFORMATIONS, HAMILTON-JACOBI
EQUATIONS, AND ACTION-ANGLE VARIABLES

This means:
2 2 k 2 m
E= = = r (4.212)
(Jr + J + J )2
In particular:
r
E 1 2E 3
r = = 4 2 k 2 (Jr + J + J )3 = (4.213)
Jr k m
which is the correct orbital frequency in a bound Kepler orbit.
Using the relations between {1 = E, , } and {Jr , J , J }, we can also get Hamiltons
characteristic function for this system as

W = W + W + Wr
Z s 2 Z s
J J d (2mk)2 2mk (J + J )2
= (J + J )2 2 + dr.
2 sin () 2 (Jr + J + J )2 r (2r)2

This then gives the angle variables:

W
r = = r (r, Jr + J + J , J + J )
Jr
W
= = (r, , Jr + J + J , J + J , J ) (4.214)
J
W
= = (r, , , Jr + J + J , J + J , J )
J

where r = r , = , and = . Of course, in this case, r = = .

At this point we can identify five constants for the Kepler problem

J1 = J
J2 = J + J
J3 = Jr + J + J (4.215)
1 =
2 = r .

(These 5 constants could also be identified from the angular momentum L, ~ energy E, and
~
Laplace-Runge-Lenz vector A.) What are they? There are two constants specifying the
plane of the orbit (the x0 y 0 -plane), which are the inclination angle i and the longitude of the
ascending node . There are three constants specifying the form of the ellipse, which are
the semi-major axis a (giving the size), the eccentricity (giving the shape), and the angle
(giving the orientation within the plane). These are all shown in Fig. 4.9.

90
 CHAPTER 4. CANONICAL TRANSFORMATIONS, HAMILTON-JACOBI
EQUATIONS, AND ACTION-ANGLE VARIABLES

Figure 4.9: Picture of an orbit in 3d and the five parameters necessary to fully specify it.
The angles i, and provide the orientation in space while a and provide the size and
shape of the conic section.

It can be shown that the relations between these constants and the ones above are
s  2
2
J1 k J3 J2
cos(i) = a= = 2 = 1
J2 2E 4 mk J3
= 21 = 22

providing a fairly simple physical interpretations to (J1 , J2 , J3 , 1 , 2 ). Also recall that J2 =

2 = 2`.
When the orbit is perturbed by additional forces (like a moon, another planet, general
relativistic corrections, and so on), these action-angle variables provide a natural way to
describe the modified orbit. We will soon see that they become functions of time. For
example, from the above list of constants, the perturbed = (t) is the precession of the
perihelion of an orbit. We will learn how to derive this time dependence next, in our chapter
on Perturbation Theory.

91
Chapter 5

Perturbation Theory

In this chapter we will discuss time dependent perturbation theory in classical mechanics.
Many problems we have encountered yield equations of motion that cannot be solved ana-
lytically. Here, we will consider cases where the problem we want to solve with Hamiltonian
H(q, p, t) is close to a problem with Hamiltonian H0 (q, p, t) for which we know the exact
solution. We say
H(q, p, t) = H0 (q, p, t) + H(q, p, t) , (5.1)
where H is small. The general idea is to expand variables
z(t) = z0 (t) + z1 (t) + 2 z2 (t) + . . . , (5.2)
for z {q, p} and use the expanded equations of motion to determine the series
k
X
z0 (t) z (1) (t) z (2) (t) . . . , where z (k) (t) = j zj (t) . (5.3)
j=0

We can do this with any of our methods for solving problems in classical mechanics, including
the Euler-Lagrange equations, Hamilton equations, Poisson bracket equations, or Hamilton-
Jacobi equations. Since there are some practical benefits, our focus will be on doing this for
the Hamilton-Jacobi equations, but lets first start with an example where we carry out an
expansion for the Hamilton equations.
p 2 2
Example Consider H0 = 2m the free Hamiltonian, and H = m2
x2 . Here is an oscillator
frequency. The full Hamiltonian H = H0 + H in this case is just a Harmonic oscillator
where we already know the solution, so we have the opportunity to see how this solution is
built up perturbatively. Without any approximation, the Hamilton equations are
p
x = , p = m 2 x . (5.4)
m
To carry out perturbation theory we are going to count w2 as O() and then at each order
we balance the number of s on each side of the equations of motion. For H0 , we have

92
 CHAPTER 5. PERTURBATION THEORY

p0 = 0 so the momentum p0 is a constant which we fix as the initial condition value. We

also have x0 = pm0 (since w2 does not appear we have not dropped anything in this equation).
Integrating we get x0 = pm0 t, where we have taken the initial condition x(t = 0) = 0 for
simplicity.
Having setup the 0th order solution, lets now consider determining the solution at 1st
order. At first order the RHS of the equations of motion should be O(). Therefore
p0
p(1) = m 2 x(0) = m 2 t = 2 t p0 , (5.5)
m
1
p(1) (t) = p0 p0 2 t2 .
2
For the other equation of motion at this order we then have

p(1) p0 p0 2 t2
x(1) = = , (5.6)
m m 2m
p0 p 0 2 t3
x(1) (t) = t .
m 6m
These are precisely the 1st order terms in the full solution
p0
p(t) = p0 cos(t), x(t) = sin(t). (5.7)
m

5.1 Time Dependent Perturbation Theory for the

Hamilton-Jacobi Equations
From here on we will focus on using H-J methods. If H = H0 + H, then the solution for
H0 has a principal function S(q, , t) that is the generating function that makes a canonical
transformation (q, p) (, ), so that
 S  S
H0 q, ,t + = 0. (5.8)
q t

For the dynamics generated by H0 the variables (, ) are constants. However, the result-
ing canonical transformation provides a new set of variables that is valid for use with any
Hamiltonian, they are just particularly simple variables for H0 . Therefore, for H, we can
still use the canonical transformation generated by S, but now the new variables

Pi = i = i (p, q) , Qi = i = i (p, q) , (5.9)

93
 CHAPTER 5. PERTURBATION THEORY

will no longer be constant in time. The new Hamiltonian is

S
K = H0 + H + = H = H(, , t) (5.10)
t

The new Hamilton equations Q i = K

Pi
and Pi = Q
K
i
now yield exact equations of motion
for these variables
H H
i = , i = . (5.11)
i i

The idea of perturbation theory is to solve these equations with an expansion. Since here
the small H  appears on the RHS of both equations, we will always use lower order
solutions on the RHS to obtain the higher order results on the LHS. Thus we use (0) and
(0) to get the first order (1) and (1) :
 
(1) H  H 
i = = , (5.12)
i i =(0)
i i 0
(0)
=
 i i 
(1) H H
i =
 
(0) =
  ,
i i =i
 i 0
(0)
i =i

where the |0 is a shorthand notation. We then use (1) and (1) to get (2) and (2) , and so
on. At nth order we have
 
(n) H  H 
i = 1) = , (5.13)
i i =(ni i n1
(n1)
=
 i i 
(n) H  H 
i = 1) = .
i i =(n
i i n1
(n1)
i =i

p 2 2
Example Lets once again consider H0 = 2m and H = m 2
x2 . For H0 , the H-J equation is
1 S 2 2
+ S


2m x t
= 0. As x is cyclic, the solution is S = x 2m t. Here,

S
P = , Q== = x t, (5.14)
m
giving the exact transformation equations

S
x= t+, p= = .
m x

94
 CHAPTER 5. PERTURBATION THEORY

For simplicity, we can take the initial constants as (0) = 0 and (0)

2 = 0 = 0. In terms of
m 2
the new variables our perturbing Hamiltonian is H = 2 m t + , so prior to expanding
the full equations of motion are

H 2
 
= = m t+ , (5.15)
m
H  
= = 2t t+ .
m
Plugging in 0th order solutions on the RHS, to 1st order we have
1
(1) = 2 0 t (1) (t) = 0 2 0 t2 , (5.16)
2
(1) 2 2
0 t3
= 0 t2 (1)
(t) = .
m 3m
If we change back to our original variables with the inverse transformation (which we may
wish to do at any point) this gives

1
p(1) = (1) = 0 2 0 t2 , (5.17)
2
and
(1) (t) 0 2 0 t 3 2 0 t 3 0 2 0 t3
x(1) (t) = t + (1) (t) = t + = t , (5.18)
m m m 2 m 3 m m 3!
which are the same results we previously obtained by solving Hamiltons equations pertur-
batively.

5.2 Periodic and Secular Perturbations to Finite Angle

Pendulum

Example Let us consider a case where we do not have a simple solution. Consider a
pendulum, with

p2
H= mga cos() (5.19)
2ma2
with  1. Expanding the cosine term we have

p2
 2
4

mga 2 mga 2
H = mga + + + + + ... . (5.20)
2ma2 2 2 12 360

95
 CHAPTER 5. PERTURBATION THEORY

In this case, the first term is a constant that will not play a role in our equations of motion,
p2 mga 2
so we can identify H0 = 2ma 2 + 2
. If we are only interested in applying first order
perturbation theory we can simply take H = mga 24
4 and drop terms of O(6 ) and higher.
The Hamiltonian H0 is just a harmonic oscillator with moment of inertia I = ma2 and
frequency 2 = ag . Again we use here for angular frequency of the H0 harmonic oscillator,
to avoid confusion with the angle variable .
The action-angle variables for H0 are

= H0 = J, = t + (5.21)
2

where = 2
, J is the action variable, and is the angle variable. This gives
r r
2 J  
= sin(t + ) = sin 2(t + ) , (5.22)
I 2 I

r
IJ  
p = 2I cos(t + ) = cos 2(t + ) .

Since and are linearly related, we are free to take (J, ) as our new canonical variables
when using the transformation in Eq. (5.22).
If we use (J, ) as the new variables, with J (0) = J0 and (0) = 0 as given constants
fixed by the initial conditions, then in terms of the new variables

mga 4 J2
H = = 2
sin4 (2(t + )). (5.23)
24 24 I

Expanding by using the 0th order solution gives


(1) H  J0
= = sin4 (2(t + 0 )) , (5.24)
J 0  12 2 I
J02

(1) H 
J =  = sin3 (2(t + 0 )) cos(2(t + 0 )) .
0 3I


These results can be integrated to give (1) = (1) (J0 , 0 , , t) and J (1) = J (1) (J0 , 0 , , t).
Before we consider computing these functions, lets pause to characterize two types of solu-
tion that occur in a more general context than simply this example.

Often we can characterize the nature of the perturbative solution without requiring a full
study of the analytic form of a solution. A common situation where this is the case is when
H0 exhibits periodic orbits (as in the harmonic oscillator) with some frequency . In this
case a relevant question is the following: what cumulative effect does the small perturbation
have after going through one or more periods T = 1 ? There are two possibilities:

96
 CHAPTER 5. PERTURBATION THEORY

The perturbation itself could be periodic, where the parameter returns to its initial
value. Here the perturbed trajectory looks much like the unperturbed one.

Alternatively, we could have a net increment in the parameter after each orbit, called
a secular change. After many periods, the parameter will be quite different from its
value in H0 .

Example Returning to our pendulum from before, the interesting quantity to study is the
average over one period of the time rate of change of the variable,

1 T (1) J (1) (T ) J (1) (0)

Z

J =(1) J (t) dt = , (5.25)
T 0 T

since thisR tells us how much the variable changes over one period. For our example J(1) = 0
2
because 0 sin3 () cos() d = 0, and therefore the perturbation to J is periodic. Actually,
from integrating Eq. (5.24) we have

J02
J (1) (t) = J0 + sin4 (2(t + 0 )) . (5.26)
24 2 I
Note from Eq. (5.22) that J determines the amplitude for (t) and p(t). A comparison
between the trajectory with J0 and with J (1) (t) is made in Fig. 5.1, where for this figure we
set 0 = 0.

with
J (1)
with J 0

Figure 5.1: Comparison of the pendulums periodic phase space trajectory using J0 and
J (1) (t).
R 2
In contrast, using 0
sin4 () 2
d
= 38 , we find
T
(1) (T ) (1) (0)
Z
1 J0
(1) = (1) (t) dt = = , (5.27)
T 0 T 32 2 I

97
 CHAPTER 5. PERTURBATION THEORY

which means experiences a secular change. After many periods (t  T ) the change
continues to build up, and we have on average that (1) (t) (1) t + 0 . (If we look at
the exact solution for (1) (t) then it has precisely this linear term in t, plus terms that are
periodic over the period , and that is what we mean by the here.) Looking back at how
the (t) dependence appears in = (J, , t) and p = p(J, , t) from Eq. (5.22), we see that
on average the 1st order perturbation simply shifts the frequency to 0 = + (1) .
Recall that we determined the full frequency full (E) numerically as an example in our
study of action-angle variables, which is shown below in Figure 5.2. Recalling that J0 =

full

= E/mga
E
-1 1

Figure 5.2: The full frequency full vs. E

H0 E+mga

=
, we can write our perturbative shift to the frequency as a function of energy

(E + mga)
0 = (1) = . (5.28)
32 2 ma2

This is the first order correction to full (E) when it is expanded about the simple harmonic
oscillator minimum at E = mga, which in Fig. 5.2 gives the negative linear correction to
the frequency that occurs just above E = E/(mga) = 1.

5.3 Perihelion Precession from Perturbing a Kepler

Orbit
Kepler Example: Consider a central force perturbation

p2
 
1 2 k h
H= pr + 2 n (5.29)
2m r r |{z
r}
| {z }
H0 H

98
 CHAPTER 5. PERTURBATION THEORY

where the coupling parameter h is small and n 2. The action-angle variables for H0 are

J1 = J 1 =
J2 = J + J 2 = r
J3 = Jr + J + J 3 = r

where only 3 = r 6= 0, and all the others are constant in time. The pairs (i , Ji ) for
i {1, 2, 3} are all canonically conjugate. One way to see this is to note that we can
implement a change of variables from the canonical pairs {(r , Jr ), ( , J ), ( , J )} that we
considered earlier, to these variables by using the generating function

F2 = ( )J1 + ( r )J2 + r J3 . (5.30)

Let us study the perihelion precession, with the parameter = 22 determining the
perihelion angle. Some examples of precession are shown in Fig. 5.3, where in the case of a
planet, the precession is like that of Fig. 5.3(b) with the sun at the focus of the ellipse. We

Figure 5.3: Precession of the perihelion from the point of view of coordinates centered on (a)
the center of the ellipse, and (b) the focus of the ellipse. The latter is relevant for a planet
like mercury orbiting the sun (shown with exaggerated eccentricity here).

know that
H
2 = , J2 = J + J = 2 = 2` (5.31)
J2

where ` = |L| is the magnitude of the angular momentum. From the equation of motion
w2 = H/J2 we therefore have

H H
= 2 = , (5.32)
J2 `

99
 CHAPTER 5. PERTURBATION THEORY


H 
and perturbatively, (1) = ` 0
. We can average over the orbit to find the secular change:

1 T H 
Z 
(1)
= dt. (5.33)
T 0 ` 0
Recall for the Kepler problem that E (0) = (2 2 k 2 m)/J32 , and that at 0th order the period
is r
1 m
T = = k = T (J3 ). (5.34)
r 2E (0)3

Therefore ` = 2 J 2 does not act on T = T (J3 ), so we can pull the partial derivative outside
the integral,
1 T
 Z 
(1)

= H |0 dt = H|0 . (5.35)
` T 0 `
Thus we must calculate the average of H over one period,
h T dt
Z
H|0 = . (5.36)
T 0 rn (t)
2
Using ` = mr2 so that dt = mr` d to switch variables from t to , and then using the
orbital equation for r() we have
n2 Z 2
hm 2 d
Z 
hm mk  0 n2

H|0 = = 1 +  cos( ) d , (5.37)
`T 0 rn2 () `T `2 0
q
2
where the eccentricity  = 1 + 2E`mk2
also depends on `. There are two simple cases where
can perform this integral:
2hm 2hm
If n = 2, then H = , so (1) = 2 .
`T `T
hkm2 R 2
If n = 3, then H = 3 d(1 +  cos( 0 )), where the cos term vanishes
`T 0
H|0 6m2 hk
upon integration, so (1) = = .
` `4 T
The latter type of potential (n = 3) is induced by corrections from general relativity to the
Newtonian potential. The Schwarzschild metric is
 rS  2 (dr)2
ds2 = c2 (d )2 = 1 c (dt)2 r2 (d)2 r2 sin2 ()(d)2 (5.38)
r 1 rrS
for rs = 2GM
c2
where M is the central mass (say of the sun), G is Newtons gravitational
constant, and c is the speed of light. The geodesic equation for radial motion is given by
 2
m dr
E= + Veff (r) (5.39)
2 d

100
 CHAPTER 5. PERTURBATION THEORY

2 2 2
where Veff (r) = mc2rrS + 2mr
` rS `
2 2mr 3 . (More background details on the Schwarzchild metric

and the derivation of this geodesic equation are discussed below in the Side Note on page
102.) Defining k = GM m, then the effective potential can be rewritten as
k `2 k`2
Veff (r) = + , (5.40)
r 2mr2 c2 m2 r3
2
from which we can identify h = ck` 2 m2 . Note that h must b e treated as a constant independent

of the canonical variable ` for the purpose of the above perturbative analysis (we simply
substitute this value for h at the end).
For Mercury, T = 0.2409TEarth ,  = 0.2056, and a = 5.79 107 km, while GMc2sun =
1.4766 km, so we get a precession rate of (1) = 42.98 arcseconds/century from general rel-
ativity. (An arcsecond is 1/3600th of a second.) After removing other contributions, such
as a shift of 531.54 arcseconds/century from perturbations by other planets, the data on
mercurys orbit shows a shift of 43.1 arcseconds/century (excellent agreement!). This was
historically one of the first tests of general relativity, and still remains an important one.

We could also consider perturbations involving momentum variables rather than coordinates.

Example Consider the relativistic correction to harmonic oscillator where the relativistic
energy
p p2 p4
E = c4 m2 + c2 p2 = mc2 + + ... (5.41)
2m 8m3 c2
In this case to analyze the first order perturbative correction we take
1 p4
p2 + m2 2 q 2 2 3 .


H= (5.42)
|2m {z } | 8c{zm }
H0 H

From H0 , the variables have a canonical transformation from the H-J analysis that gives
r r
J Jm
q= sin(2(t + )) , p= cos(2(t + )) . (5.43)
m
This gives
J 2 2
H = 2 2 cos4 2(t + ) .
 
(5.44)
8 c m
Since J(1) is odd over one period, it turns out that J is periodic once again

H 
(1)
J = J(1) = 0. (5.45)
0
Meanwhile, the change for (1) is secular,
J0 2 3J0 2

(1) H 
cos4 2(t + )
 
= = 2 2
(1) = . (5.46)
J 0 4 mc 32 2 mc2

101
 CHAPTER 5. PERTURBATION THEORY

Thus, (1) from the relativistic correction H is again a negative shift to the frequency of
the oscillator.

This ends our discussion of perturbation theory in classical mechanics.

Side Note: The Schwarzchild Geodesic from Action Minimization

The Schwarzchild metric is given by
 rs  2 2 dr2
ds2 = g dx dx = 1 ct rs

r2 d2 r2 sin2 d2 , (5.47)
r 1 r

where rs = 2GM
c2
is the Schwartzchild radius. The geodesic orbit for a test particle is a curve
which minimizes proper distance with this metric. In this case, we have
1
dx dx 2
Z Z 
0 = s = ds = g d (5.48)
d d

where is the proper time and ds2 = c2 d 2 . (One method of determining the geodesic
2 dx
equations is to use 0 = ddx2 + dx
d d
with the Christoffel symbols determined from
the metric, but we will follow a different approach.)
The minimization in Eq. (5.48) is equivalent to applying the minimal action principal for
the Lagrangian

m dx dx
L= g (5.49)
2 " d d #
 2  2  2  2
m  rs  2 dt 1 dr 2 d 2 2 d
= 1 c r r sin
1 rrs


2 r d d d d

1
since the presence of the extra square root (. . .) 2 does not matter for this minimization. Here
we have the generalized coordinates x = (t, r, , ) which are to be considered as functions
of the proper time variable . Also, the mass m is a test mass (which also gives us the proper
units).
Because t and are cyclic variables in L, we have

L  rs  2 dt
pt = =m 1 c = E tot energy (5.50)
t r d
L d
p = = mr2 sin2 =` angular momentum
d

102
 CHAPTER 5. PERTURBATION THEORY

Using the E-L equation on we obtain

d L L
0= (5.51)
d
 
d 2 d
= mr mr2 sin cos 2
d d

where planar motion with = 2
is a solution that suffices for our purposes. Now
2
ds2 2
`2

2
 rs  2 Etot 1 dr
= c = 1 c 2 r2 (5.52)
1 rrs
2


d r m2 1 rrs c4 d m2 r4

gives a radial equation. Separating out the rest mass and expanding for E  mc2 we have:

2
E tot = mc2 + E E tot m2 c4 + 2mc2 E (5.53)

Finally,
 2
l2
 
m 2 rs  m 2 m dr
c + 2 2 1 c + =E (5.54)
2 mr r 2 2 d
| {z }
Veff (r)

Note that the rest mass mc2 /2 terms cancel. Therefore the effective potential is

rs mc2 l2 rs l2
Veff (r) = +
2r 2mr2 2mr3
2
k l kl2
= +
r 2mr2 m2 c2 r3
2GM
where rs = c2
and k = GM m, so mrs c2 = 2k. This is the result that was quoted above in
Eq. (5.40).

103
Chapter 6

Fluid Mechanics

So far, our examples of mechanical systems have all been discrete, with some number of
masses acted upon by forces. Even our study of rigid bodies (with a continuous mass distri-
bution) treated those bodies as single objects. In this chapter, we will treat an important
continuous system, which is that of fluids.
Fluids include both liquids and gases. (It also includes plasmas, but for them a proper
treatment requires the inclusion of electromagnetic effects, which will not be discussed here.)
For our purposes, a fluid is a material which can be treated as continuous, which has the
ability to flow, and which has very little resistance to deformation (that is, it has only a
small support for shear stress, which refers to forces parallel to an applied area). Applica-
tions include meteorology, oceanography, astrophysics, biophysics, condensed matter physics,
geophysics, medicine, aerodynamics, plumbing, cosmology, heavy-ion collisions, and so on.
The treatment of fluids is an example of classical field theory, with continuous field
variables as the generalized coordinates, as opposed to the discrete set of variables qi that we
have treated so far. Therefore the first step we have to take is understanding how to make
the transition from discrete to continuum.

6.1 Transitioning from Discrete Particles to the Con-

tinuum
Rather than starting with fluids, lets instead consider a somewhat simpler system, that of
an infinite one dimensional chain of masses m connected by springs with spring constant k,
which we take as an approximation for an infinite continuous elastic one-dimensional rod.

104
 CHAPTER 6. FLUID MECHANICS

If the equilibrium separation of masses is a and the distance the ith mass is translated
from equilibrium is i , then
kX mX 2
V = (i+1 i )2 T = , (6.1)
2 i 2 i i
where V is the potential energy from the springs, and T is the kinetic energy. It is convenient
to write the Lagrangian as
 2 !
1X m 2 i+1 i
L=T V = a ka , (6.2)
2 i a i a
and the corresponding equations of motion obtained from the Euler-Lagrange equations as
m    
i+1 i i i1
i ka 2
+ ka 2
= 0. (6.3)
a a a
Technically both the Lagrangian and the equations of motion are independent of a, but we
have introduced factors of a to facilitate taking the continuous limit a 0. In this limit
the masses become progressively more densely distributed along the line. The important
question when taking this limit is which quantities do we hold fixed.
Lets define ma
as the mass density and Y = ka as the Youngs modulus. Here Y is
equivalent to the spring constant for a continuous rod. (For a rod, the force F = Y where
is the longitudinal extension per unit length, or in other words, the strain.) We intend to
hold and Y fixed when taking the continuous limit.
The key change in the continuous limit is that the discrete position index i becomes a con-
tinuous position label x, so instead of i = i (t), now x = x (t), or with more conventional
notation, = (x, t). This also means that
(x + a, t) (x, t)
, (6.4)
a x
2
 
1 (x + a, t) (x, t) (x, t) (x a, t)
, (6.5)
a a a x2
X Z
a dx . (6.6)
i
R
Using these results in Eq. (6.2) gives L = L dx where
 2  2 !
1 (x, t) (x, t)
L= Y (6.7)
2 t x
is the Lagrangian density. Likewise, using them in Eq. (6.3) gives the equations of motion
2 2
Y = 0, (6.8)
t2 x2
which we recognize as the wave equation.

105
 CHAPTER 6. FLUID MECHANICS

The results for the Lagrange density and Euler Lagrange equations are of course not
indepedent. We can also use
 
L = L , , , x, t (6.9)
x t
with Hamiltons principle, Z Z t2
S = L dx dt = 0. (6.10)
t1
to formulate the dynamics, and thus derive the Euler-Lagrange equations. Because =
(x, t) has two parameters, if we follow the standard procedure of varying the path takes
between the two endpoints, we get variations from the dependence of the Lagrange density
on its first three arguments. Integrating by parts in each of t and x, and setting the surface
terms to zero, then yields
! !
L L L
+ =0 (6.11)
t



t
x x
as the continuum Euler-Lagrange equation. Recall that for N particles we expect N E-L
equations for the time dependence, but here we have just one equation. However actually
by that counting, this result corresponds to an infinite number of equations, one for each
value of x. From this point of view, the derivatives with respect to x are what couples these
equations together.

Example For the Lagrangian density of the elastic rod,

L L L


= ,

= Y , = 0. (6.12)
t
t x
x
2 2
Putting these results together in Eq. (6.11) gives the wave equation t2 Y x
2 = 0 as

anticipated.
In our above analysis, = (x, t) is a continuum generalized coordinate called a classical
field. Here t is a parameter and x is a continuous label as well.
Although we have been talking about one dimension so far, it is easy to generalize the
above discussion to a higher number of dimensions. For example, in three dimensions we
simply have dependence on three continuous label parameters, = (x, y, z, t) or = (r, t).
The field (r, t) is called a scalar field because the output is a single number. With multiple
dimensions we also have vector fields (r, t), where the output is a vectors. An example of
vector fields that you are familiar with are the electromagnetic fields E(r, t) and B(r, t). In
fact, classical fields of exactly this sort are also the starting point for formulating quantum
field theory. One formulates a classical Lagrangian density L (that is most often Lorentz
invariant) which depends on fields like the electromagnetic scalar and vector potentials (r, t)
and A(r, t). Then one quantizes these fields.
Our description of fluids will make use of classical field variables in 3-dimensions without
considering quantization.

106
 CHAPTER 6. FLUID MECHANICS

Eulerian & Lagrangian Variables

Let us consider fluid flowing in a tube. A natural set

of variables would be to take some initial time t = 0
and label all fluid elements by their coordinates at this
time: r0 . We will often refer to fluid element, which
is a point in the fluid, or sometimes an infinitesimally
small region of the fluid. The motion of the fluid could
then be described by r = r(r0 , t), which determines the
subsequent coordinates of the fluid element labeled by
r0 at the later time t.
If we continue in this direction it leads to the Lagrangian formulation of fluid dynamics.
The advantage is that the usual laws of classical mechanics apply to fluid particles. The
disadvantage is that it is more work to make the connection to measurable quantities that
characterize the fluid.
Instead, we will explore an alternate formulation.
Pretend we sit at a fixed point r in the fluid and ask
what happens there as a function of the time t. We can
think of it like placing a measuring device at r and mea-
suring, at any time t, quantities like the fluid velocity
v, density , or pressure P. In fact, these 5 variables
(as density and pressure are scalars, while velocity is a
vector) are enough to describe the state of the moving
fluid. This is the Eulerian formulation of fluid dynam-
ics. Additionally, and P are enough to determine
all thermodynamic quantities (assuming we know the
equation of state).

Example for an ideal gas at temperature T , we have PV = nR T , where V is volume, n

is the number of moles, R is contant, and T is temperature. Dividing by V we have that
pressure is simply proportional to density, P = R0 T for a rescaled gas constant R0 .
Pressure is isotropic in a fluid. This means it is the same from all directions, so there is
only 1 number for each (r, t); thus, pressure is a scalar field, P(r, t).
To prove this consider the infinitesimal wedge-shaped fluid element below, which we take
to be at rest in a gravitational field. Recall that pressure is force per unit area, P = n F/A,
where n is a unit vector perpendicular to the area A that the force F is acting on. In the figure
we label three pressures P1 , P1 , P1 acting on three of the sides, that have corresponding force
components in the x and z directions. (There is also pressure on the other two faces with
forces in the y direction, but the above will suffice to argue why the pressure is isotropic.)

107
 CHAPTER 6. FLUID MECHANICS

P
P

Next we balance forces for the wedge at rest. Gravity pulls down, and the volume of the
wedge is dxdydz/2 so the force of gravity is

dx dy dz
Fg = g . (6.13)
2
Also by simple trigonometry the area dA of the slanted face can be written in two different
ways

dz dx
dA = dy = dy .
sin() cos()

Balancing forces in the x and z directions then means

0 = dFx = P2 dy dz P1 sin() dA = (P2 P1 ) dy dz, (6.14)

1
0 = dFz = P3 dx dy P1 cos() dA g dx dy dz.
2
The first equation implies P1 = P2 . In the second equation we can pull out a common dxdy
to give
1
P3 = P1 + g dz, (6.15)
2
then as the infinitesimal distance dz 0 we have

P1 = P3 . (6.16)

Thus, pressure is the same in all directions. Even if the fluid is moving or even accelerating
we would come to the same conclusion. For example, if we had to balance the force against
acceleration this would lead to adding term

a dx dy dz (6.17)

108
 CHAPTER 6. FLUID MECHANICS

for acceleration a, which again drops out for an infinitesimal fluid element just like the
gravitational force did.
Time Derivatives:
The total time derivative dtd tells us the rate at which a quantity changes as we move

with a fluid element. The partial time derivative t tells us the rate of change of a quantity
at a fixed position r. We can work out a relation between them.
Example For P
dP P P P P
= + x + y + z (6.18)
dt t x y z
P
= + v P.
t
In general, the time derivative acts as
d
= +v (6.19)
dt t
on any fluid field (scalar, vector, or tensor) that is a function of (x, y, z, t).

6.2 Fluid Equations of Motion

6.2.1 Continuity Equations
Let us consider a volume V of fluid. Then from Gauss
divergence theorem (which you may recall from elec-
tomagnetism, but which is a general result of vector
calculus for any vector field):
Z Z Z
dV v = dS n v = dS v, (6.20)
V V V

where V is the closed area that bounds the volume V ,

n is unit vector orthogonal to the surface element dS,
and dS dSn.

Lets ask: As the fluid moves, how does V change?

The quantity n v is the outward velocity of the surface

dS, so n v dS dt is the volume added in a time interval
dt, as illustrated on the right.

109
 CHAPTER 6. FLUID MECHANICS

This means that the change in volume can be determined by adding up all the changes
from integrating over the entire surface
Z Z
dV
= v dS = v dV . (6.21)
dt V V

This result becomes even simpler if we consider an infinitesimal volume V over which v
does not vary, then we can pull v outside the integral to obtain simply

dV
= V v. (6.22)
dt

Thus the divergence of the velocity, v, controls how the fluid volume expands with time.
If v = 0 everywhere then we say the fluid is incompressible because for every volume
element dV
dt
= 0.
Even if the volume changes, the mass of the fluid element will not,

dm d
= (V ) = 0. (6.23)
dt dt

This implies that

 
d dV d
0 = V + = V + v , (6.24)
dt dt dt
so
d
+ v = 0. (6.25)
dt
Expanding out the time derivative into partial derivatives this yields

+ v + v = 0, (6.26)
t

and simplifying the result gives

+ (v) = 0 . (6.27)
t
This is an important result the continuity equation for mass of the fluid, which is a partial
differential equation in the fluid variables.
Here v is the mass density flux (the flow of mass density), and this equation says that
nowhere is the matter making up the fluid created or destroyed. If the density changes, then
there must be an incoming or outgoing flux. This is easier to see if we use the divergence
theorem for the vector (v) to write this result in integral form,

110
 CHAPTER 6. FLUID MECHANICS

Z Z

dV = dS n (v) . (6.28)
t V V

R R
Here t V
dV is the increase of mass in the volume V , while V dSn (v) is the outflow of
mass through the surface surrounding this volume (which becomes an inflow with the minus
sign).
Weve talked about v thus far, so its natural to ask: is there a physical interpretation
to v? There is indeed.
The quantity = v is the vorticity. If the velocity
is constant or uniform in a single direction ~v = vx (x)x,
then its obvious that = 0. To consider what it mea-
sures we can use Stokes theorem for the velocity vector
field, Z I
( v) n dS = v dl, (6.29)
S S
where S is now an open surface and S is its closed
boundary curve. We can use this result to determine
when = v is non-zero.

Consider a rotating fluid and the circular curve shown.

When we integrate the velocity along this circular curve
the terms always add up, so 6= 0. It thus provides a
measures of the rate of rotation of the fluid.

If we consider flow in one direction with a velocity gra-

dient, again 6= 0. Here a larger positive contribution
to the integral is obtained from the right side of the cir-
cle, relative to the negative contribution from the left
side.

If v = 0 everywhere in a moving fluid, we say the flow is irrotational.

We can determine more precisely what is as follows. Consider a rotating coordinate

111
 CHAPTER 6. FLUID MECHANICS

system with constant angular velocity , so v = v0 + r. Then

v = v0 + ( r)
= v0 + r r
= v0 + 2 .
Now was constant, but we have not yet specified its value. If we pick = 12 v at
a point r, then v0 = 0 and the fluid is irrotational at r in the rotating frame. This
argument can be repeated for other points r. Considering this from the point of view of the
origianl frame we thus see that /2 is the angular velocity of the fluid at position r.

6.2.2 Ideal Fluid: Eulers Equation and Entropy Conservation

Let us consider an ideal fluid which has no energy dissipation due to internal friction (mean-
ing no viscosity) and no heat exchange between different parts of the fluid (meaning no
thermal conductivity).
The force on a fluid element V from pressure in the x
direction is

Fx = Fx (x) Fx (x + x) =
P(x) P(x + x) P
= y z x = V . (6.30)
x x
More generally accounting for all directions we have

F = P V. (6.31)
From external forces F it is useful to define the force density f by F = f V . This means
that Newtons law, ma = F becomes
dv
V = (P + f )V. (6.32)
dt
Writing out the total time derivatives this becomes
v P f
+ v v + = (6.33)
t
which is the Euler equation for fluid dynamics. There are two special cases which are useful
to consider.
1. Often f can be derived from a potential: f = . Note that here is a potential
energy per unit mass. For example, with gravity, = gz, so f = gz. In general,
then
v P
+ v v + + = 0 (6.34)
t
is a rewriting of the Euler equation.

112
 CHAPTER 6. FLUID MECHANICS

 2
v
2. We can use v = v ( v) = 2
v v. This gives

v  v2  P
+ + + v=0 (6.35)
t 2
as another rewriting of the Euler equation. For constant , we can take the curl and
use the fact that (h) = 0 for any h to obtain


= v . (6.36)
t
We will have occasion to use this result later on.
To solve the Euler and continuity partial differential equations we need boundary condi-
tions. The boundary conditions for ideal fluids are simply that they cannot penetrate solid
surfaces, so if a surface with normal vector n is stationary then

v n  = 0.
surface
(6.37)
If the surface does move then 
v n surface = vsurface . (6.38)
So far we have four equations (continuity for the scalar density , and the Euler equation
for the vector velocity v) for five unknowns. For an ideal fluid the 5th equation,
dS
= 0, (6.39)
dt
is the statement that the entropy S is conserved, so there is no heat exchange. Effectively,
this provides a relationship between pressure and density through P = P(, S). A simple
example is an ideal gas at constant temperature, where P = R0 T .

6.2.3 Conservation of Momentum and Energy

Due to the term (v )v which has a vi vj , the Euler equation is nonlinear. For simpler
situations it is therefore very useful to consider conservation laws.
Let us start by considering Momentum Conservation. The quantity v is the flux of mass
density, which is also the density of momentum (in direct analogy to p = mv). Consider

t
(vi ) = v + v
t i t
i
. Using the continuity and Euler equations to replace these two partial
derivatives, and once again implicitly summing over repeated indices until further notice,
this becomes  
vi 1 P
(vi ) = vi (vj ) + vj + fi . (6.40)
t xj xj xi
This is rearranged to give
P
(vi ) + + (vi vj ) = fi . (6.41)
t xi xj

113
 CHAPTER 6. FLUID MECHANICS

We define the stress tensor for an ideal fluid as

Tij = Pij + vi vj (6.42)

which gives the momentum flux density (which is to say, the density of momentum in the
direction ei flowing in the direction of ej ). Note that Tij is symmetric. Then, in vector form,
the equation above becomes

(v) + T = f . (6.43)
t
This describes the conservation of linear momentum density with sources given by the exter-
nal force densities f . Comparing to the continuity equation where the density is a scalar
so its flux v is a vector, here the momentum density v is a vector so its flux T is a tensor.
Next consider Conservation of Energy. Conservation of energy can be said to arise from
the Euler equation dv/dt + P/ + = 0. Note that here we are switching back to the
total time derivative since this is more useful for our discussion of energy. For a volume
element V , we take the inner product of the Euler equation with vV to obtain
dv
V v + V v P + V v = 0 . (6.44)
dt
Given that dtd (V ) = 0 by the conservation of mass we can move this combination inside of
total time derivatives. Also recall that v = d/dt /t. Using these two facts we obtain
 
d 1 2 d
v V + (v P)V + (V ) V = 0. (6.45)
dt 2 dt t

Next we try to convert the second term to time derivatives. Consider using
d P
(PV ) = V + (v P)V + P( v)V, (6.46)
dt t
where we recalled that dV /dt = ( v)V . Using this to eliminate (v P)V gives
   
d 1 2 P
v V + V + PV = + V + P vV, (6.47)
dt 2 t t

where 21 v2 V is the kinetic energy, V is the external potential energy, and PV is the
internal potential energy due to pressure. The terms with partial time derivatives act like
sources. Unfortunately there is still a term without a total or partial time derivative, however
this term is easier to interpret. It is related to the work Wu done by V when it expands
and exerts pressure on the surrounding fluid. In particular

dWu d(V ) d
=P = P v V (U m) (6.48)
dt dt dt

114
 CHAPTER 6. FLUID MECHANICS

where in the last step we have defined the work as a negative potential energy U (per unit
mass). If the equation of state is given, this U can be calculated either as an integral in
density or pressure,
Z 0 Z P 
P 0 P( )
Z
0 1 d

U = d(V ) = d 02
= P 2 dP 0
(P 0 ) dP 0 (6.49)
m 0 P0

where we used the fixed m = V to switch variables between V and . Using Eq. (6.48)
means we can now write everything in terms of time derivatives,
    
d 1 2 P
v + + P + U V = + V (6.50)
dt 2 t t
This is the equation for energy conservation in an ideal fluid. If P and are not explicitly
dependent on time at any point in space (which is often the case), then any fluid element
has a constant total energy as it moves along (recall that this is the meaning of d/dt). For
applications to fluids it is more convenient to divide this result by m = V to give
d v2
 
P 1
++ +U = + . (6.51)
dt 2 t t
This is Bernoullis equation.

6.3 Static Fluids & Steady Flows

Having derived the equations of motion and conservation laws for ideal fluids, let us now
consider some important special cases.
Static Fluids
Static fluids have v = 0 everywhere, so the fluid is at rest (implying mechanical equi-
librium). Continuity then says t
= 0, so and P are independent of time. If = gz for
gravity, the Euler equation says
P
= gz, (6.52)

1 P
so P and can each only be functions of z, while in fact z
is independent of z.

Example if the density is constant, then if = gz,

then
P(z) = P0 gz, (6.53)
so the pressure decreases with height.

115
 CHAPTER 6. FLUID MECHANICS

Example let us pretend the atmosphere is an ideal gas at a uniform constant temperature
T . As P = R0 T , then
P g
= 0 P, (6.54)
z RT
so gz
P(z) = P0 e R0 T , (6.55)
and the pressure falls exponentially.

Example the Archimedes principle says that the pres-

sure balances the weight of any displaced fluid. This
follows from our first example above. The pressure on
the bottom of the object displacing the fluid is gz,
where z is the distance from the surface. The force
is pressure times area, so that is the volume displaced
times g, or the mass displaced times g.
Note that if the temperature is not uniform, then the mechanical equilibrium is not stable,
as the temperature gradients result in convection currents which mix the fluid. Therefore
we have used the fact that we are discussing an ideal fluid.
Steady Flows
Steady flows are ones in which t
= 0, tP = 0, and
v
t
= 0 at every position r.
In this case, the continuity equation becomes
(v) = 0, so what flows in must flow out. This is most
easily implemented by using the integral form, where
for any closed surface S we have
Z
dS (v) = 0 . (6.56)
S
For a steady flow the Bernoulli equation becomes the statement that
v2 P
++ +U =B (6.57)
2
where B is a constant along the paths of fluid elements. Most often we will consider gravity
where = gz.
A steady incompressible fluid is one where v = 0 and U is constant. Continuity now
says = 0 so is constant as well. Moreover, the Bernoulli equation now says
v2 P
+ + = B0 (6.58)
2
is also constant.

116
 CHAPTER 6. FLUID MECHANICS

Example for a horizontal pipe filled with an incompressible fluid (which is approximately
true of water at room temperature) of constant density .

Lets consider the two ends to be at approximately at the same height so we can drop the
term = gz. At the hatched areas shown we know that the flow must be tangential to the
edge of the pipe, so a valid solution is to simply consider the velocities to be uniform and
tangential to the enclosing pipe across each of these areas. Using Bernoulli this implies that

v12 v 2
+ P1 = 2 + P2 . (6.59)
2 2
Furthermore, the continuity equation for the enclosed areas shown (some sides being those of
the pipe) implies that the flux in at one end must equal the flux out at the other A1 v1 = A2 v2 .
Together this gives
 2 !
1 2 A2
P1 = P2 + v2 1 . (6.60)
2 A1
Since A1 > A2 we have v2 > v1 , and this implies P1 > P2 .

Example let us consider a water tank filled to a height z1 with a hole at height z2 < z1 that
produces a jet of water.
Let us assume v1 0 (so the tank is much larger than
the hole). Then
1 2
v + P2 + gz2 = P1 + gz1 . (6.61)
2 2
Additionally, P2 = P1 = Patmosphere , so the pressure
terms cancel out, and we can solve for the v2 velocity
to give p
v2 = 2g(z1 z2 ) . (6.62)
This is the same velocity as that for any mass falling
from rest through a height z1 z2 . Of course a key
difference for the jet of water is that this velocity is
horizontal rather than vertical.

117
 CHAPTER 6. FLUID MECHANICS

Lets discuss two common ways to picture flows. One is through stream lines, which are
lines that are everywhere tangent to the instantaneous velocity, meaning

dx(s)
v =0 (6.63)
ds
for some parameter s that determines the distance along the streamline. These lines are
drawn at some fixed time and never cross since there is a unique velocity at every point.

Writing out the cross product we find 0 = dx/ds v = z vy dx vx dy

ds ds
+ . . . implying that
we have
dy vy dz vz  dz vz 
= , = , or = . (6.64)
dx vx dx vx dy vy

When we use the equations in this form we would need to be able to switch from s to the
variable x to uniquely parameterize the curve.
Another method of picturing the flow is through flow lines, which are paths that are
followed by fluid elements, meaning
dx
= v(x, t). (6.65)
dt

Since flow lines are time dependent, they can in general cross since the path a fluid element
takes through a point may differ at a later time.
For a steady flow, the lines are time independent, and stream lines and flow lines are
identical. To prove this we consider the flow line equations dx/dt = vx , dy/dt = vy , and note
that since the velocities are time independent that we can eliminate time through the ratio

118
 CHAPTER 6. FLUID MECHANICS

dy/dx = vy /vx (and similar for other directions), which is the equation for the stream lines.
To go the opposite direction we simply pick s = t to parameterize the streamline, and note
that v v = 0. Furthermore, for a steady fluid we have Bernoullis law

v2 P
++ +U =B (6.66)
2

which we now understand is constant along stream lines.

Stagnation Points and a Pitot Tube

Consider an ideal incompressible fluid in a steady flow which flows in an almost uniform
manner, and then hits a wall.
Since the velocity must be tangential at the sur-
face, the flow can either curve left or right, and
it is often the case that parts of the flow will go
in both directions. In this case there is a stag-
nation point o, where vo = 0. If we consider the
streamline that hits the stagnation point then
Po = P + 21 v 2
, where v and P are, re-
spectively, the velocity and pressure infinitely
far away. Thus the largest pressure in the en-
tire flow occurs at the stagnation point where
there is only pressure and no kinetic energy.

Now let us consider a Pitot tube, which is a device used to measure velocity (for example
on airplanes).

119
 CHAPTER 6. FLUID MECHANICS

O
a

h
l

If the density is a in the air and l in the liquid, we can write down the Bernoulli equations
obtained by comparing the air and liquid flows at , the stagnation point o, at the point
s near the surface (where the air velocity is the same as at and liquid is at rest), and at
the point m at the top of the liquid inside the column. This gives:
1 2
Po = P + a v
2
1
Ps = P + a g(zo + zs )
2
Pm = Po + a g(z0 + zs + h)
= Ps + l gh
Subtracting the 2nd equation from the 1st , subtracting the 4th equation from the 3rd , and
then adding these two results cancels all the pressure terms, and leaves
1 2
a v = g(l a )h. (6.67)
2
This can be rearranged to write
 
2 l l
v = 2g 1 h 2g h (6.68)
a a
allowing us to determine the velocity of the air v in terms of the known ratio of densities
l
a
 1 and simply the measured height between the liquid on each side, h.

6.4 Potential Flow

When a flow is everywhere both irrotational and incompressible it is known as potential
flow. Such flows may be steady or not steady. Since = v = 0 the velocity field is
conservative. This means there exists a velocity potential
Z r
(r) = v(r0 , t) dr0 (6.69)
r0

120
 CHAPTER 6. FLUID MECHANICS

which depends only on the endpoints of the integration, and not the path taken between
them, such that
v = (6.70)
Since v = 0 for an incompressible fluid, then must solve Laplaces equation

2 = 0 . (6.71)

Solving for the scalar (with suitable boundary conditions), then immediately gives v.
We can then use the Euler equation to immediately get the pressure. For = 0 one
form of the Euler equation was
 2 
v v P
+ + + = 0. (6.72)
t 2
If is constant, then
v2
 
P
+ ++ = 0, (6.73)
t 2
so
v2 P
+ + + = b(t) (6.74)
t 2
for some function b. For each tR we can pick the zero of so that b(t) is constant. (This is
t
equivalent to shifting + b(t0 ) dt0 , where adding this constant that is independent of
x gives a solution that is equally valid.) The remaining constant b be fixed by a boundary
condition on the pressure. Thus the full pressure as a function of x and t is determined by
P v2
= + b, (6.75)
t 2
where in principal the first three terms on the right hand side carry both spatial and time
dependence. Often we are interested in a steady flow, in which case the term /t = 0.

Example if = v0 x, then v = v0 x, which is a specific case of a uniform flow.

Example if = A ln(r) in 2 dimensions, then for all r > 0

1 A
v= r + = r, (6.76)
r r r
which is a point source. An analogous point source in 3 dimensions would arise from = Ar ,
which gives v = rA2 r for all r > 0. (At r = 0 there would be a delta function source, so we
do not satisfy Laplaces equation at this point.)

Example Consider = 2 in polar coordinates in the 2 dimensional plane. Then vr =
1
r
= 0 and v = r = 2r for all r > 0. This corresponds to a potential vortex about the
point r = 0.

121
 CHAPTER 6. FLUID MECHANICS

Example Consider a sphere of radius R moving with constant velocity u through an incom-
pressible ideal fluid and find its velocity by solving for its velocity potential. Equivalently we
can consider the problem of finding the velocity when a sphere is held at rest and the fluid
flows in from far away with uniform velocity u at infinity. The first situation is shown in
figure a) and the second as figure b).

a) b)

We use coordinates centered on the sphere, and define the axes so that u = ux. The
problem is spherically symmetric other than the directionality from u, so its natural to expect
v u. Since = v this means that we expect u. (Effectively the boundary condition
is linear in u and the equation 2 = 0 is linear. We could also explicitly demonstrate the
proportionality v u using dimensional analysis, as we will discuss in more detail later in
this chapter.)
It is actually easier to consider the sphere being at rest with the fluid moving past it as
in b), so lets start with this case. Since 2 = 0 is linear, we can solve using superposition.
The velocity potential
(x) = ux + 0 (x) (6.77)
has a term ux giving the correct uniform flow far away from the sphere. Therefore, with
r the distance from the center of the sphere, we have

lim 0 (x) = 0. (6.78)

r

Another way to see this is that taking the gradient of Eq. (6.77) gives

v = ux + v0 , (6.79)

which is simply the translation between the velocity field v for b) and the velocity field v0 for
a). For the situation a) we would anticipate Eq. (6.78) as the correct boundary condition,
since the fluid is at rest at infinity when it is infinitely far away from the disturbance caused
by dragging the sphere. We can look for a solution for 0 .
As 2 0 = 0 with limr 0 = 0, one option could be 0 = 1r in 3 dimensions, but this
would give a point source solution with velocity moving radially outward from our sphere
and hence make it impossible to satisfy the appropriate boundary condition on the sphere
(its also not u). Instead, let us consider a dipole source
 
0 1
= Au (6.80)
r

122
 CHAPTER 6. FLUID MECHANICS

for some constant A. This proposal is linear in u, and satisfies 2 0 = 0 for r > 0 since
the derivatives commute: 2 0 = A(u )2 (1/r) = 0. It remains to compute 0 and v
explicitly and check that we can satisfy the boundary conditions (and proper dimensions)
with this solution. Since  
1 r
= 3, (6.81)
r r
we have
Au r
0 = . (6.82)
r3
Moreover,
(u r) = u, rn = nr(n+2) r, (6.83)
so the solution for v0 is  
0 A0 3(u r)r
v = = 3 u + (6.84)
r r2
and we then also have obtained v = u + v0 . The boundary condition on the surface of the
sphere (which is r = R for case b) where the sphere is at rest) is v r = 0. This means
Au r 3Au r
0 = u r + , (6.85)
R3 R3
R3
which has the solution A = 2
. Thus,

R3
 
3(u r)r
v = u + 3 u . (6.86)
2r r2
This solution has the right dimensions and satisfies the boundary conditions on the sphere
and at infinity.
For steady flow, we can then use Bernoullis equation to get the pressure on the sphere,
constant = (P + 21 v 2 )r=R = (P + 12 v 2 )r= . Squaring our result for the velocity on the
sphere, setting u r = ur cos , and simplifying we find
u2
9 cos2 5 .


P = P + (6.87)
8
This result for the pressure says that it is the same on the front and back of the sphere, since
its unchanged by taking /2 . This is quite counterintuitive, since we expect a force
on the sphere in b) that would try to push it downstream. This actually results from our
approximation that the fluid is ideal (viscosity can not be neglected when trying to answer
questions near surfaces).
Another possibility is that our approximation of potential flow is suspect. To explore
this, lets ask how common is potential flow? Consider
 
d dv v
( v) = = + v v , (6.88)
dt dt t

123
 CHAPTER 6. FLUID MECHANICS

which expands into

  2 
v
( v) + v ( v) = (v ) = 0 (6.89)
t 2 t
where in the last step we used our result in Eq. (6.36) that we derived from the Euler
equation. Thus, v is conserved along flow lines. If we consider a steady flow which
starts out as uniform infinitely far away, then there is no vorticity at infinity, and
v =0 (6.90)
on every stream line, and remains that way for the entire flow. Thus we have a potential
flow with no vorticity.
If we wanted to get around the counterintuitive behavior we have found in our ideal
fluid solution, but stick with the ideal fluid framework, then we would have to allow for
the existence of discontinuous solutions. For an ideal fluid flowing past a sphere, we could
propose that stream lines exist that start tangential to the spherical surface, and hence
satisfy the boundary conditions, and can have v = 6 0 since they are not connected to
infinity. Behind the sphere we could then imagine having fluid at rest.

The correct treatment of boundary layers near objects and of wakes, does not require
discontinuous solutions of this sort, but instead simply requires the inclusion of viscosity,
which we will turn to shortly, after treating one final important example from ideal fluids.

6.5 Sound Waves

In this section we will explore an example where v 6= 0 plays an important role. To set
things up, consider a compressible fluid at rest with pressure P0 and density 0 in equilibrium
with an external force density f0 . If P0 and 0 are constant and uniform, then
P0 f0
= (6.91)
0 0

124
 CHAPTER 6. FLUID MECHANICS

from the Euler equation. Now lets add disturbances P 0 and 0 to this system
P = P0 + P 0 , (6.92)
= 0 + 0 , (6.93)
with P 0  P0 and 0  0 . These disturbances will induce a velocity field as well, v(r, t),
which we will also assume is small, so that perturbation theory applies. We will therefore
drop terms that are second order or higher in any of {P 0 , 0 , v}. Using perturbation theory
on the Euler equation, that is
v v
0= + (v )v + p f0 = P0 f0 + 0 + P 0 + . . . (6.94)
t t
where we have dropped terms v v = O(0 v2 ) and O(0 v). Using f0 = P0 then gives
v P 0
= (6.95)
t 0
as the Euler equation to 1st order in perturbations. To the same order, continuity says
0
= (v) = 0 v. (6.96)
t
where we have again dropped second order terms. Finally the appropriate thermodynamic
equation of state is
0
0 = P 0 (6.97)
B
where B is a constant known as the bulk modulus. The bulk modulus describes a substances
0
resistance to compression, and this formula arises from B = P
0 P0 . Using this result
we can eliminate density 0 to get a second equation involving only the pressure and velocity
disturbances,
P 0 B 0
= = B v (6.98)
t 0 t
Combining Eq. (6.95) and Eq. (6.98) we can derive a differential equation for the pressure
disturbance
2P 0 v B
2
= B = 2 P 0 , (6.99)
t t 0
which can be written more simply as
2P 0
c2S 2 P 0 = 0 (6.100)
t2
q
which is a wave equation for P 0 , whose solutions move at a velocity cS = B0 which is known
as the speed of sound. Due to the simple proportionality from the equation of state we also
immediately know that
2 0
c2S 2 0 = 0 (6.101)
t2

125
 CHAPTER 6. FLUID MECHANICS

so the same wave equation also holds for the density.

It remains to derive a differential equation for the velocity. Taking the curl of Eq. (6.95)
yields
1
( v) = P 0 = 0. (6.102)
t 0

This means v does not explicitly depend on time, so we can conveniently take v = 0
initially everywhere, and hence for all times. Using Eq. (6.95) and Eq. (6.98) now gives
 0
2v 1 P B
2
= = ( v) (6.103)
t 0 t 0
B  
= 2 v ( v)
0
B
= 2 v ,
0

since ( v) = 0. This yields

2v
c2S 2 v = 0 (6.104)
t2
which means that the velocity of the fluid disturbance also satisfies the same wave equation.
The solutions are thus sound waves in pressure, density, and velocity, with speed cS . An
example of a solution is a plane wave (here written for pressure) which looks like

P 0 = P 0 (r n cS t) (6.105)

when traveling in the direction of n.

Note that v 6= 0 was important for this derivation. It is reasonable to ask if there is
a way that we can determine when the approximation v = 0 may be justified. For a flow
with characteristic velocity v0 this can be done by defining the Mach number

v0
M= , (6.106)
cS

since the scaling of terms involving v will be determined by this ratio. If M  1 then
we can treat flow as approximately incompressible, with v = 0.
Considering flows with large values of M leads to the concept of shock waves. Consider a
flow with initial velocity v0 in which there is a disturbance. If M < 1 then the flow is said to
be subsonic, and the perturbation spreads everywhere, because the speed of the perturbation
is larger than that of the flow.

126
 CHAPTER 6. FLUID MECHANICS

On the other hand if M > 1, then the disturbance is swept downstream to the right by
the flow, and actually propagates downstream within a cone of angle defined by sin() =
cs /v0 = M1 , as shown below.

If we consider a supersonic plane, then we should view this picture the other way around,
where the fluid is static and the disturbance (plane) moves through it, traveling to the left
at faster than the speed of sound. This causes a sonic boom, which is the air pressure wave
given by the dashed lines trailing the plane, which moves at speed cS . (Another example is
thunder, where the rapid increase in temperature of plasma of ions causes rapid air expansion,
making a shockwave.)

6.6 Viscous Fluid Equations

Internal friction occurs in a fluid when neighboring fluid particles move with different veloc-
vi
ities, which means that x j
6= 0. Viscous stresses oppose such relative motion. Adding these

127
 CHAPTER 6. FLUID MECHANICS

friction terms changes some of our fluid equations. In particular, the continuity equation re-
mains unchanged, while the Euler equations along with the conservation laws for momentum,
energy, and entropy must be modified.
vi
To consider this friction we will work to first order in the x j
partial derivatives, treating
these as the most important terms. These derivatives can be arranged into
 
vi vj 2
ij = + ij v + ij v (6.107)
xj xi 3

to define the viscous stress tensor with elements ij . The constant coefficients of the two
terms are the shear viscosity and the bulk viscosity , where > 0 & > 0. The form of
the viscous stress tensor ij is dictated by the fact that it must vanish for constant v and for
uniform rotation v = r where there is no friction. Writing out v = x (yz zy) + . . . we
see that v = 0 and vy
z
+ v
z
y
= 0, etc., for the uniform rotation, dictating the symmetric
form of the terms in ij . The remaining organizational choice is to let multiply a traceless
tensor.
Momentum conservation still comes from

(v) + T = f (6.108)
t
where now we include a friction term in the stress tensor to account for the viscous transfer
of momentum. Thus
Tij = Pij + vi vj ij (6.109)
is the new total stress tensor.
A simple rule for incorporating ij is simply to replace Pij Pij ij . With this we
can add friction to the Euler equation. In particular we have


(P)i = ki P ki P ki , (6.110)
xk xk
where we can compute that

 2v 2 
i
ki = + v v + v
xk xk xk xi 3 xi xi
 
= 2 vi + + v. (6.111)
3 xi
Plugging this into the Euler equation yields
v P 2 1  f
+ (v )v + v + ( v) = (6.112)
t 3
which is the Navier-Stokes equation.

128
 CHAPTER 6. FLUID MECHANICS

A common case we will study is when f = 0 & v = 0, which reduces the Navier-Stokes
equation to
v P
+ v v = + 2 v (6.113)
t
where

(6.114)

is the kinematic viscosity, and the bulk viscosity term has dropped out. The dimensions of
the kinematic viscosity are [] = m2 /s, which is simpler than [] = kg/(ms).
Other useful equations can be derived for the situation where v = 0 & is constant.
Taking the divergence of the Navier-Stokes equation removes both the v/t term, and the
2 v term, leaving [(v )v] = 2 P. Writing this out in components we find
vi vj
2 P = vi vj = , (6.115)
xj xi xj xi
since vj /xj = v = 0. This equation can be used to compute the pressure if the velocity
is determined, since it simply acts like a source term. Taking the curl of the Navier-Stokes
equation, and recalling that = v we find

(v ) = 2 , (6.116)
t
where the algebra to arrive at the terms on the LHS was consider earlier in our discussion
of the Euler equation, and the new pieces is the term on the RHS.
In the presence of viscosity the boundary conditions change from what we had previously.
Molecular forces between the viscous fluid & any surface mean that the fluid adheres to the
surface, and hence that the velocity of the fluid and surface must be the same for both the
tangential and longitudinal components. Therefore the boundary condition for a moving
wall is v = vwall , which also covers the case v = 0 for a wall at rest.
Another important concept is the force exerted by the fluid on a surface. This has a
contribution both from the pressure as well as from the friction. The force per unit area Fi
is given by
Fi = nj (Pji ji ) = Pni + ij nj , (6.117)
where n is the normal vector pointing out of the surface, and the first term is the pres-
sure acting along this normal vector, while the second is the friction that has tangential
components.
Starting with the Navier-Stokes equation we can also derive a modified form for energy
conservation. Rather than carrying out this derivation explicitly, we will just examine the
final result in integral form, which is a bit more intuitive:
Z   I   2  
P
Z
1 2 v vi
v dV = + vi vj ji dSi ij dV . (6.118)
t V 2 V 2 V xj

129
 CHAPTER 6. FLUID MECHANICS

Here the term on the LHS is the change of the kinetic energy in the volume V with time.
The first integral on the RHS is the energy flux through the closed bounding surface V ,
and the second integral is the decrease in energy that is caused by dissipation. To see this
even more clearly we can consider integrating over the whole fluid with v = 0 at (or on
V ). This removes the flux term and leaves
Z   Z  2
E vi vj vi vi vj
= + dV = + dV < 0 . (6.119)
t V xj xi xj 2 V xj xi
where we can check the second equality by squaring and manipulating the summed over
dummy indices i and j. Thus we see that friction causes energy to dissipate just as we would
expect (and this also justifies our sign choice of > 0).
Entropy conservation is modified at temperature T to T S = ij x
vi
j
, where the left-hand
side of the equation is the heat gain per unit volume, and the right-hand side of the equation
is the energy dissipated by viscosity. (If we allow thermal conduction (from temperature
gradients) in the fluid, then there is another term on the right-hand side that appears as
(T ) for conductivity .)

6.7 Viscous Flows in Pipes and Reynolds Number

We start our investigation of fluid flow with viscosity, by studying steady flows in pipes,
typically caused by a pressure gradient. This type of steady viscous flow is also called
Poiseuille flow, after Jean Poiseuille who first studied them experimentally. We take the flow
to be incompressible, v = 0, and hence the continuity equation implies that is constant
(just as it did in the ideal fluid case). Thus the Navier-Stokes equation for such flows reduces
to
P
v v = + 2 v (6.120)

Example Lets start by considering flow in a long rectangular duct, aligned with the x-axis,
with height h and width w. We also take it be a thin duct with h  w, and with a pressure
gradient along the pipe in the x-direction.

130
 CHAPTER 6. FLUID MECHANICS

This can be approximated by infinite parallel plates, taking w and holding h fixed.
Here

v = vx (y) x,
(6.121)

since there is no dependence on x or z by the translational symmetry in the plane, and we

equally well can not develop velocity components vy or vz due to this symmetry. This implies
that the term v v = vx v
x
= 0 Taking the inner product of Eq.(6.120) with y and z then
removes all the terms that depends on the velocity (which is only in the x direction), giving

P P
= = 0. (6.122)
y z

Thus the pressure P = P(x) and can have a gradient only in the x direction. Taking the
inner product of Eq.(6.120) with x gives

P 2 vx
= 2 vx = 2 = k , (6.123)
x y
2
where we have introduced a constant k. Since P x
only depends on x, while yv2x only
depends on y, they must both be equal to a constant. Let us say P x
= k < 0, so that the
pressure drops as we move in the x direction. (In the Navier-Stokes equation, this pressure
drop balances the viscous stress term.) Then integrating the equation for vx (y) gives

k 2
vx (y) = y + ay + b , (6.124)
2

where we have introduced two integration constants a and b. To solve for a and b we impose
the boundary conditions that the velocity must vanish at y = 0 and y = h, giving

(k )
vx (y) = y(h y) . (6.125)
2

Recalling that k < 0 we see that vx (y) > 0, flowing down the pipe. (If we had reversed k
the flow would be in the opposite direction.) The velocity field we have derived flows in the
pipe with a parabolic profile with its maximum in the center:

131
 CHAPTER 6. FLUID MECHANICS

Lets also calculate the friction force per unit area that the fluid exerts on the pipe wall.
The bottom plate at y = 0 has a unit vector n = y, so from Eq. (6.117) the force along x is

vx hk
Fx = yx = |y=0 = > 0. (6.126)
y 2
Intuitively this is the direction we expect, the fluid tries to drag the pipe along with it.

Example Lets now consider the same example of fluid between infinite parallel plates, but
now with no pressure gradient. Instead we take the top plate to move with velocity u = ux.
Here P = P0 is constant and uniform with k = 0, so the second derivative of vx (y) is zero,
2 vx = 0, and the solution for vx (y) can at most be linear. The solution for this case is
vx (y) = uyh
, which satisfies the boundary conditions vx (0) = 0 and vx (h) = u. Thus the
fluids velocity field is linear for this case:

Example Next consider a long cylindrical pipe of radius L, oriented along x, again with a
pressure gradient along the pipe. We will approximate the pipe as being infinitely long so
there is a translational symmetry along x.

Due to the translational symmetry we know that v = v(y, z). To fully exploit the conse-
quences of the symmetry it is useful to use cylindrical coordinates (x, r, ) so we can also
easily impose the rotational symmetry about x to conclude v = v(r). The fact that there is

132
 CHAPTER 6. FLUID MECHANICS

r dependence makes sense since we know that v must vanish at the edge of the pipe, r = L,
but we do not want it to vanish everywhere. Continuity and symmetry also imply that the
velocity is only in the x direction, so in fact

v = vx (r)x . (6.127)

For example, consider an annulus shaped closed surface formed by the region between two
cylinders cocentric with the pipe. The flow into and out of this surface must be balanced
by continuity. The flow on the ends of the surface automatically balance each other since v
is independent of x. But since v is r dependent, the only way the flow through the circular
sides can balance each other is if there is no flow in the r direction.

With this setup we can again confirm that (v )v = vx (r) x vx (r)x = 0, leaving
2
P = v from the Navier-Stokes equation. Taking the inner product with y and z
we see that P
y
= zP = 0, so P = P(x). Taking the inner product with x gives

P(x)
= 2 vx (r) = k (6.128)
x
where since P
x
depends only on x while 2 vx (r) is independent of x, the two must be equal
to a constant k. Again we choose k < 0 to have a pressure gradient that pushes the fluid
down the pipe in the x direction (k > 0 would simply reverse the flow). For the velocity this
gives  
1 vx k
r = , (6.129)
r r r
and integrating this gives  
k r
vx (r) = r2 + a ln +b (6.130)
4 r0
for some constants a and b. (The constant r0 is introduced to make the equation dimension-
fully correct, but is not independent, since any change to r0 can be compensated by a change
to b). Since vx (r) has to be finite at r = 0 we must have a = 0 (if the geometry excluded
the region at the middle ). The condition vx = 0 at r = L fixes b so that

(k ) 2
L r2 .


vx (r) = (6.131)
4
Lets calculate the discharge rate of fluid in such a pipe, as mass per unit time. This is given
by
Z L
kL4
discharge rate = 2rvx (r) dr = > 0. (6.132)
0 8

Note that this rate is quite sensitive to the radius, being proportional to L4 . This is why
you dont want narrow blood vessels.

133
 CHAPTER 6. FLUID MECHANICS

Reynolds Number and Similarity

How do we quantify the importance of viscosity? Let us consider a flow and identify a
characteristic length L (examples of which might be the radius of a pipe or the size of a
sphere inserted in a flow) and a characteristic speed u (examples of which might be the fluid
velocity in the middle of a pipe or the fluid velocity upstream from flow around a sphere).
The quantity
uL uL
R= = (6.133)

is the dimensionless Reynolds number. A large R means viscosity is not as important com-
pared to the characteristic u and L of the system, while small R means that viscosity plays
an important role. It is important to note that viscosity is always important near a surface
where it modifies the boundary condition on v. (The above examples of Poiseuille flow have
R . 2000.)
The introduction of R leads nicely into a discussion of the utility of exploiting dimensional
analysis in fluid dynamics problems. This is best explained by way of examples.

Example Consider a steady incompressible flow in a system that is characterized by a single

length scale L and initial velocity u. What could the possible functional form be for the
velocity and pressure in such a system? Considering the coordinates and fluid field variables
we can only have
v r  P r 
=h ,R , =g ,R . (6.134)
u L u2 L

Here h and g are a dimensionless vector and scalar function respectively. The ratio r/L
is dimensionless, as is R, and the dimensions of u and P are compensated by u and u2
respectively. Note that if we consider flows that have the same R, then the properties of
those flows are related by changes to the scales associated with v, r, or P. Such flows are
called similar.

Example Consider a viscous flow past a sphere of radius a with initial velocity given by
limx v = ux. Here the Reynolds number is R = ua . If we double and double u then R
is unchanged. Due to the relations in Eq. (6.134) we thus can predict that we will have the
exact same form for the solutions with v twice as large as before, and P being four times as
large as before.
Note that in general other dimensionless ratios, like the ratio of two length scales, or the
Mach number M = cuS could also appear. (For M  1 we treat the fluid as incompressible
and neglect M for the dimensional analysis.) To determine how many independent dimen-
sionless ratios can be formed for the analysis of a general dimensional analysis problem, one
counts the number of variables and subtracts the number of unrelated types of dimensions
that appear in these variables. For most fluid problems this will mean subtracting three for
kg, meters, and seconds.

134
 CHAPTER 6. FLUID MECHANICS

6.8 Viscous Flow Past a Sphere (Stokes Flow)

Lets consider a steady, incompressible, viscous fluid with small R flowing past a sphere of
radius a. The fluid has velocity u when it is far away from the sphere. This is the same
problem we treated for an ideal fluid with potential flow, but now we want to consider the
impact of viscosity, and resolve the puzzle we found in our solution for an ideal fluid. To
make the problem solvable we work in the limit R  1.
Here the Navier-Stokes equation is
P 2
v v = + v. (6.135)

Lets determine the relative importance of the two velocity terms in the R  1 limit. Using
a to characterize spatial variations, and u to characterize velocity we find
 u2  2  u   u2 
v v = O , v=O =O . (6.136)
a a2 aR
Therefore the viscosity term, which is enhaced by a factor of 1/R, dominates. Neglecting
the v v term, the Navier-Stokes equation reduces to
P = 2 v . (6.137)
Note that when written in terms of the shear viscosity , that the density has dropped out
of this equation, and hence the constant will not play a part in the solution. Taking the
divergence of this equation we find 2 P = 2 v = 0, so the pressure satisfies Laplaces
equation. Using dimensional analysis for the pressure we expect a solution of the form
u r  r 
P= g , (6.138)
a2 a
where the dimensional analysis requirement of having only a single factor of the velocity u
and a scalar result, leads to including the factor of u r. Note that we have not included R
as a possible argument since we are expanding with R  1. Due to the fact that P must
satisfy Laplaces equation and is proportional to u r, we can immediately recognize the
dipole solution (which we met earlier for the potential in the ideal fluid case):
au r
P = P0 + k . (6.139)
r3
Here we have included an overall constant pressure P0 to satisfy any boundary condition on
the pressure at r , and a dimensionless constant k which is still to be determined.
Next we note that the vorticity of the fluid satisfies
1
= ( v) = ( v) 2 v = 2 v = P

 
ka 3(u r)r
= 3 u . (6.140)
r r2

135
 CHAPTER 6. FLUID MECHANICS

Enforcing that vanishes as r , the solution for this equation is

ur
= ka . (6.141)
r3
Thus we see that unlike the ideal fluid case, there is now a non-zero vorticity in the fluid
here.
Next we turn to determining the velocity, which can only depend on v = v(r, u, a). Again
we do not include R as an argument since we are expanding for R  1. By dimensional
analysis the velocity must be linear in u so the most general possible solution takes the form
 r  r(u r)  r 
v = uf + g , (6.142)
a a2 a
where the functions f and g are dimensionless. The gradient of one of these dimensionless
functions gives f (r/a) = (r/(ar))f 0 (r/a). Computing the divergence of the velocity with
this form we have
u r 0  r  (u r)r 0  r  4(u r)  r 
0=v = f + g + g . (6.143)
ar a a3 a a2 a
Therefore we find that the scalar functions must satisfy
r r2  r  4r  r 
f0 = 2 g0 g . (6.144)
a a a a a
Next we equate the Laplacian of the velocity and gradient of the pressure, which can be
simplified to give
 2g   4g 0  (u r)r
2 v = 2 f + 2 u + 2 g +
a ar a2
3
1 ka  3ka (u r)r
= P = 3 u + 5 . (6.145)
r r a2
Note that here 2 g(r/a) = (1/r2 )(d/dr)r2 (d/dr)g(r/a). Equating the coefficients of the two
structures we find
4g 0 g 00 6g 0 3ka3
2 g + = 2+ = 5 ,
ar a ar r
00 0
f 2f 2g ka
2
+ + 2 = 3 . (6.146)
a ar a r
To solve the equation for g we try a polynomial solution of the form g(x) = Cn xn giving

Cn n(n 1) + 6n xn2 = 3kx5 .

 
(6.147)

Here n = 3 is a particular solution to the full inhomogeneous equation with C3 = k/2.

Also n = 0 and n = 5 are homogeneous solutions where the LHS vanishes, and the

136
 CHAPTER 6. FLUID MECHANICS

corresponding coefficients C0 and C5 must be fixed by boundary conditions. Looking back

at our starting point in Eq. (6.142) we see that the full set of boundary conditions are

v = u at r = : lim r2 g(r/a) 0 , lim f (r/a) 1 ,

r r
v = 0 at r = a : g(1) = 0 , f (1) = 0 , (6.148)

This fixes C0 = 0 and C5 = k/2 so that

ka3  a2 
g(r/a) = 1 . (6.149)
2r3 r2
Using Eq. (6.144) and integrating once we find that

ka ka3
f (r/a) = + 3, (6.150)
2r 6r
where we have set the integration constant to zero to satisfy the boundary condition at
r = . The final boundary condition, f (1) = 1 then requires us to take the constant
k = 3/2. Note that this fixes the constant k that appeared in the vorticity and in the
pressure P. All together we have that the final solution for the velocity is
 3a a3  3a3 r(u r)  a2 
v =u 1 3 3 1 . (6.151)
4r 4r 4r a2 r2
The flow looks like:

Next we turn to determining the drag force on the sphere. In general the drag force on
an object in the direction j is given by an integral of the force per unit area over the surface,
Z
FDj = dSi (Pij ij ) . (6.152)
V

Lets take the inflowing velocity to be in the x direction, v(r ) = ux. Then FD x
will be the drag force on our sphere in the direction of the bulk fluid flow. With spherical
coordinates (r, , ) where is the polar angle, we have dS = ra2 d cos d with r = a, as
well as r x = cos and x = sin . Thus
Z  
2
x FD = a dcos d P cos + rr cos r sin . (6.153)

137
 CHAPTER 6. FLUID MECHANICS

Computing the needed components on the sphere r = a we find


vr   1 v
r v v   3 
rr = 2 = 0, r = + = u sin ,
r r=a r r r r=a 2a
 3 
P = P0 u cos . (6.154)
2a

Thus the drag force on the sphere is

 3u  Z
x FD = dcos d a2 = (6a)u , (6.155)
2a

which is the famous Stokes formula for the viscous (friction) drag force on a sphere. (Note
that we could have obtained the factor of au by dimensional analysis.)
In addition to drag forces like this, that point in the direction of the fluid flow, objects
may also experience lift forces that are tangential to direction of the fluid flow. Such forces
occur for wing-shaped objects and are important for many physical phenomena, including
lift on airplanes.
Dynamic Vortices and Turbulence
For our flow about the fixed sphere, lets consider what happens as we increase R. From
our analysis above it is clear that at some point the non-linear v v term we dropped
will become important. The v/t will also become important, with flows that are more
dynamical, changing with time. Lets consider how the flow appears for various values of R:

For R  1, the flow is symmetric and is (somewhat counter intuitively) qualitatively

like the case of 0. This is also called Stokes flow or a laminar flow.

For R 1, the flow is still like Stokes flow, but the stream lines are no longer as
symmetric, with a more clear wake developing behind the sphere.

138
 CHAPTER 6. FLUID MECHANICS

For 10 . R . 100, detached vortices called eddies form behind the sphere, though
the flow is still steady. Note that directly behind the sphere between the vortices
that the fluid is now flowing in the opposite direction to the asymptotic inflow u.
As R increases, the flow becomes looses its steady nature, with the time dependence
emerging by having through oscillations of the vortices.

For 150 . R . 1000, vortices start to be cyclically shed and drift downstream in a
wake behind the sphere. This time dependent solution appears like it has interaction
between the eddies, where one pair pushes the next downstream.

For 103 . R . 2 105 , the wake becomes highly irregular, exhibiting a phenom-
ena known as turbulence which we will discuss in more detail below. Here there are
unsteady, interacting vortices at all length scales.

139
 CHAPTER 6. FLUID MECHANICS

For R & 2105 , the turbulent wake narrows and the boundary layer around the sphere
is no longer laminar, also becoming turbulent.

Turbulence is characterized by a flow that: i) is highly irregular (technically chaotic) in

x and/or t, ii) has nonzero vorticity, iii) has much greater dissipation of energy compared
to more uniform laminar viscous flows, and iv) has eddies and vortices forming over many
length scales with energy that in a three dimensional flow cascades from the largest eddies
down to the smallest eddies where it dissipates into heat due to viscous friction. Turbulent
mixing is a very effect mechanism of transport for energy, momentum, temperature, and so
on. Examples of turbulence include many familiar phenomena: the circulation of air in the
atmosphere, the water flow from a faucet which becomes white at a certain point when the
flow rate is increased, water in rapids, dust kicked up by wind, water beside a ship moving
in an otherwise smooth lake, clear air turbulence causing a drop in lift for airplanes, and so
on.
In the last value of R discussed for flow around our shere, a turbulent boundary layer
appeared. This causes an abrupt drop in drag in the flow over objects, and is a very
useful phenomena. In particular, by introducing imperfections we can cause this turbulent
boundary layer to form at smaller values R, meaning smaller velocities. This is why golf
balls have dimples and baseballs and tennis balls have visible seams.
We also get turbulence in flow through pipes at large R. The viscous flow in pipes we
previously considered were laminar flow at smaller R values and had velocity distributions

140
 CHAPTER 6. FLUID MECHANICS

that were parabolic, while in contrast a turbulent flow will be non-uniform at small scales,
but when averaged causes a more uniform flow down the pipe at larger length scales.

In general the chaotic and irregular nature of turbulence makes it difficult to treat with
analytic methods, and a complete description of turbulence remains an unsolved problem.

141
Chapter 7

Chaos and Non-Linear Dynamics

By a deterministic systems of equations, we mean equations that given some initial conditions
have a unique solution, like those of classical mechanics. In a deterministic system we will
define chaos as aperiodic long-term behavior that exhibits sensitive dependence on initial
conditions.

Here aperiodic behavior means that phase space trajectories do not converge to a
point or a periodic orbit, they are irregular and undergo topological mixing (discussed
below).

By sensitive to initial conditions we mean that trajectories that start nearby initially,
separate exponentially fast. Defining (t) as the difference between points on two such
trajectories at time t, then this means that |(t)| 0 et for some > 0, as depicted
in Fig. 7.1.

Figure 7.1: The difference in initial condition leads to different orbits. Their difference is
given by (t), which grows exponentially with time.

This means that even though they are deterministic, chaotic systems are most often not
predictable. In particular, there will always be a small difference 0 between the true and
measured initial conditions for the system (from statistical or systematic measurement error),
which grows exponentially to yield inaccurate predictions for predictions far enough in the
future.

142
 CHAPTER 7. CHAOS AND NON-LINEAR DYNAMICS

The sensitivity to initial conditions is important to chaos but does not itself differentiate
from simple exponential growth, so the aperiodic behavior is also important. In the definition
of this somewhat undescriptive phrase we include that the system should undergo Topological
Mixing. This means that any points starting in a region (open set) of the phase space will
evolve to overlap any other region of the phase space, so chaotic systems tend to explore a
larger variety of regions of the phase space.1

7.1 Introduction to Chaos

We will now explore some properties of non-linear dynamical systems, including methods to
characterize solutions, and the study of solutions with chaotic behavior.

7.1.1 Evolution of the system by first order differential equations

The dynamical system can be defined by a system of first order differential equations:
x1 = f1 (x1 , . . . , xn )
x2 = f2 (x1 , . . . , xn ) (7.1)
..
.
xn = fn (x1 , . . . , xn )
where the quantities xi , for i {1, . . . , n}, are any variables that evolve in time, which
could be coordinates, velocities, momenta, or other quantities. For our applications in this
chapter we will often assume that the xi equations are also chosen to be dimensionless, and
the procedure for this type of conversion will be discussed further below.
Example: the Hamilton equations of motion are 1st order equations in the canonical vari-
ables, so they are an example of equation of the form in Eq. (7.1) with an even number of
xi variables.
Deterministic evolution from the existence and uniqueness theorem
Assume that we have a set of differential equations in the form in Eq. (7.1), which we
can write in a shorthand as
~x = f~(~x) , (7.2)
f
and that fj and xjj (for i, j {1, . . . , n}) are continuous in a connected region D Rn .
Then if we have an initial condition ~x(t = 0) = ~x0 D, then the theorem states that there
exists a unique solution ~x = ~x(t) on some interval (, ) about t = 0. Time evolution in
such a system is therefore deterministic from this existence and uniqueness theorem.
1
For a dissipative chaotic system there are further restrictions on the choice of the open sets in this
definition of topological mixing since it is otherwise obvious that we could pick a region that the system will
not return to.

143
 CHAPTER 7. CHAOS AND NON-LINEAR DYNAMICS

For this chapter the damped nonlinear oscillator will be a good to base our discussion.
In the case of a pendulum with damping and a periodic driving force, its evolution is given
by the equation of motion:

ml2 + ml2 + mgl sin() = A cos(D t) , (7.3)

where ` is the length of the pendulum, is the oscillator angle, is the damping coefficient,
and A cos(D t) is the driving force. It is useful to turn this into a dimensionless equation.
First we divide by mgl to make the third term dimensionless, defining
A
a , (7.4)
mgl
to give a dimensionless amplitude for the forcing term. This leaves
l l
+ + sin = a cos(D t) . (7.5)
g g
Next to make the first term dimensionless we rescale the time derivatives so that they involve
a dimensionless time t0 , and change to a dimensionless frequency D
0
for the forcing term via
r s
g l du du
t0 t, 0
D D , u u 0 . (7.6)
l g dt dt

As indicated we also now let dots indicate derivatives with respect to the dimensionless time.
Finally we define
s
1 l
, (7.7)
q g

where q is the dimensionless quality factor for the damping term.

Dropping the newly added primes, our final differential equation is now fully dimension-
less:
1
+ + sin() = a cos(D t) (7.8)
q
Here a, q, and D are all dimensionless constants. We can convert this into 1st order form by
defining D t to get rid of the explicit time dependence in the forcing term, and
to eliminate the double time derivatives. This gives the system of three equations that are
in the form in Eq. (7.1) with ~x = (, , ):

= ,
1
= sin() + a cos() , (7.9)
q
= D .

144
 CHAPTER 7. CHAOS AND NON-LINEAR DYNAMICS

7.1.2 Evolution of Phase Space

Phase space trajectories never cross
From the uniqueness theorem, phase space trajectories never cross. To prove this, note
that any point ~x(t) on a trajectory could be used as an initial condition for a new trajectory.
Since a point can only be part of one single trajectory, no crossings can occur.

Figure 7.2: By the uniqueness theorem, no two trajectories can cross, only come arbirtrarily
close.

Evolution of phase space volume

The phase space volume is given by:
n
Z Y
V= dxj (7.10)
V j=1

Recall that for Hamiltonian systems, canonical transformations do not change volume ele-
ments. If we view this transformation as a solution for motion (via the H-J equation), then
it is clear that the motion generated by a Hamiltonian preserves the volume, so V = 0.
What happens with damping/friction (which is not in our Hamiltonian formalism)? To
determine the answer we can exploit an analogy with our results for changes in volume for
fluids:

x = v(x) ~x = f~(~x) , (7.11)

Z Z
V = dV v V = dV f~ .

where in the context of a general nonlinear system, refers to derivatives with respect to
~x. Thus we see that f~ determines the change to a volume of our phase space variables.
For this reason we define f~ = 0 as a conservative system (whether or not a general
Hamiltonian exists), while f~ < 0 is a dissipative system where the phase space volume
shrinks.

145
 CHAPTER 7. CHAOS AND NON-LINEAR DYNAMICS

For our damped nonlinear driven oscillator example we have:

 
~ 1 D 1
f = + sin() + a cos() + = < 0, (7.12)
q q
as expected for a dissipative system.
For the special case of q and a = 0 (undamped and undriven system), then:

= , and = sin() . (7.13)

The corresponding trajectories in phase space are illustrated below:

3
5 0 5

Figure 7.3: Phase space picture of the undamped, unforced oscillator. Filled circles are the
stable fixed points and empty circles are the saddle points which are fixed points that are
unstable in one direction and stable in another.

7.1.3 Fixed Points

Of particular interest in a system are its fixed points, ~x? , defined as the locations where

f~(~x? ) = 0 . (7.14)

At these points the state of the system is constant throughout time. Depending on the
behavior of the trajectories nearby the fixed point they can be characterized as:
Stable - nearby trajectories approach the stable point

146
 CHAPTER 7. CHAOS AND NON-LINEAR DYNAMICS

Unstable - nearby trajectories move away from the fixed point

Saddle Point - in different directions trajectories can either approach or move away

For the undriven, undamped oscillator (Eq.(7.13)), the system has fixed points for = 0
and = n for any integer n. For this pendulum, the fixed point at the bottom = 2n is
stable, while the fixed point at the top is unstable = (2n + 1), as shown in Fig. 7.3. Note
that this fixed point at the top is not a crossing trajectory because we can only get to this
point if E = 0 exactly, and in that case the trajectory would stop at this fixed point. Any
small perturbation knocks it off the unstable point at the top and determines which way it
goes.
If there is dissipation, then all trajectories in the neighborhood of a stable fixed point
converge upon it, so this region is called the basin of attraction and the fixed point is an
attractor ; energy dissipates as motion decays to the attractor. In our example it occurs if
q is finite, and the basins of attraction in this case are diagonal strips in phase space. The
result for two trajectories in phase space are shown below.

2.0

1.5

1.0

0.5

0.0

0.5

1.0

1.5

2.0
4 2 0 2 4

Figure 7.4: With finite damping (q = 5) in our oscillator example the trajectories converge
to the stable fixed points of the system with spiraling motion.

Conditions for chaotic behavior

In general, two necessary conditions for chaos are:

The equations of motion must be nonlinear. (For linear systems we already know the
solutions, which are exponential or oscillating and hence not chaotic.)

147
 CHAPTER 7. CHAOS AND NON-LINEAR DYNAMICS

There must be at least 3 variables, so if i {1, . . . , n}, then n 3. (We will see why
this is necessary later.)

In our non-linear damped oscillator example, now including a non-zero forcing term gives
rise to a wider range of qualitative behaviors. In particular for certain values of (a, q, D )
the system can be chaotic.

If we start instead with the linearized version of the forced damped oscillator then we
have:

1
= + a cos() (7.15)
q

For this case the solution, which is non-chaotic, are well known and often studied in ele-
mentary courses in classical mechanics or waves. The general solutions come in three cases,
underdamped (q > 1/2), critically damped (q = 1/2), or overdamped (q < 1/2). For example
the general underdamped solution is given by:

r 
t
2q 1 a
(t) = Be cos t 1 2 + 0 + q cos(D t ) , (7.16)
4q 2 1
D q + (D D )2

2
where tan() = D /(q qD ), and B and 0 are constants that are determined by the initial
conditions. The first term in Eq. (7.16) is the transient that decays away exponentially,
whereas the second term describes the steady state forced motion (whose amplitude exhibits
resonant behavior at D = 1).

A projection of the trajectories into the 2-dimensional plane, as shown in Fig. 7.5
shows that they converge onto ellipses after many cycles. This does not break the uniqueness
theorem since = D t is increasing, so the trajectory never crosses itself when all three
variables are plotted. If restrict [0, 2] then the trajectory converges to a closed orbit.
Note that the nonlinear forcing term cos() is important to ensure that this closed orbit is
an isolated stable endpoint for the motion.

148
 CHAPTER 7. CHAOS AND NON-LINEAR DYNAMICS

2.0
2.0

1.5
1.5

1.0 1.0

0.5
0.5

0.0

0.0
0.5

1.0 0.5

1.5
1.0
2.0

2.0 1.5
1.5
0 1.0 2.0
5
10
15
20
0.5
0.5 0.0 0.5 1.0 1.5 2.0

25

30 0.0
35
40
450.5

Figure 7.5: Although the projection of the system shows crossings, which would seem
to violate the uniqueness theorem, plotting the 3 variables we see that no crossing occurs,
and uniqueness is perserved. In the projection plot we also clearly see the system evolving
to a closed orbit.

An attractor that is a closed orbit rather than a single point is called a limit cycle.

7.1.4 Picturing Trajectories in Phase Space

2-dim projections
To solve the full nonlinear damped forced oscillator, described by the solution to Eq. (7.9),
we use a computer. Note that we can examine chaos and sensitivity to initial conditions on
a computer since the various phenomena, including the exponential growth of differences
do to chosen initial conditions, occur much before the differences due to truncation errors
associated with machine precision take over. In order to give a taste of what chaos looks
like, we will first simply examine some of the results especially as applied to the nonlinear
oscillator.
One way to see chaos is to do a projection of trajectories in the full n-dimensional space
of variables to a lower dimension (usually down to 2 dimensions so we can plot results in a
plane). For the nonlinear oscillator, this is typically the plane where we project away
(as in the right most images of Fig. 7.5). For chaotic motion this projection yields a two
dimensional picture which in general gets quite messy, with the trajectory filling out a large
area of the plane.
Poincare Section (Poincare Map)
To simplify things further we can can use a Poincare section (also called a Poincare map).
Here we sample the trajectory periodically in say when = 2n which is the periodicity
of cos for our example, and plot only these values (n , n ) in the plane. The results we

149
 CHAPTER 7. CHAOS AND NON-LINEAR DYNAMICS

track are then much like what we would observe by looking at the system with a stroboscope.
For creating the Poincare section of any such system, we wait until the transients die out.

For the nonlinear oscillator, this might be at = 2n for integer n, yielding a sampling
frequency of exactly D , so the map is a plot of only these values (n , n ). For example, we
could take D = 32 and q = 2 while varying a as in Fig. 7.6; where we have waited for 30
cycles to ensure that the transients have died out.

In figure Fig. 7.6 we show both 2-dimensional phase portraits and Poincare maps for
various values of a. As a increases the plots show singly periodic long term behavior (a = 0.9),
to doubly periodic (a = 1.07), to chaotic (a = 1.19), and finally to periodic again occurring
amidst neighboring chaos (a = 1.35).

150
 CHAPTER 7. CHAOS AND NON-LINEAR DYNAMICS

/ /

2.00
3

2
1.95

= 1

1.90

- -/2 /2
-1
1.85
-2

-3

-1.0 -0.8 -0.6 -0.4 -0.2 0.0

/ /

2.00
3

2 1.98

= 1
1.96

- -/2 /2 1.94
-1

-2 1.92

-3

-1.0 -0.8 -0.6 -0.4 -0.2 0.0

/ /

2.5
3
2.0
2

= 1 1.5

1.0

- -/2 /2
-1 0.5

-2 0.0

-3

-2 -1 0 1 2 3

/ /

0.90
3

2
0.85

= 1

0.80

- -/2 /2
-1
0.75
-2

-3

-1.0 -0.8 -0.6 -0.4 -0.2 0.0

Figure 7.6: Phase portraits and Poincar/e sections for the nonlinear driven damped oscil-
lator with D = 2/3, q = 2, and various values of a. The plots show singly periodic, doubly
periodic, chaotic, and singly periodic behavior respectively. (Plots generated with the Math-
ematica demonstration package, Chaotic Motion of a Damped Driven Pendulum, by Nasser
Abbasi.)
151
 CHAPTER 7. CHAOS AND NON-LINEAR DYNAMICS

Bifurcation Map
Yet another way is through a bifurcation diagram, which takes the Poincare map results
but plots one of the variables against one of the parameters of the system. This plot allows us
to see the transitions between different behaviors, in particular a change in the fixed points
of the system. For the nonlinear oscillator, this could be a plot of against a, as shown in
Fig. 7.7.

2.5

2.0

1.5

1.0

0.5

0.0
1.0 1.2 1.4 1.6 1.8
-0.5

Figure 7.7: For the Driven damped nonlinear oscillator, plot of = values obtained from
the Poincare map as a function of a with Q = 2 and D = 2/3 fixed. This bifurcation plot
show the qualitative transitions of the system, such as where period doubling/bifurcation oc-
curs, and where chaos starts. (Plot generated with the Mathematica demonstration package,
Chaotic Motion of a Damped Driven Pendulum, by Nasser Abbasi.)

There are a few notable features in this bifurcation plot which we summarize in the following
table:

a Features
1.0 only a single
1.07 two values of from the same initial conditions (period doubling)
1.15-1.28 mostly chaos (some periodic windows)
1.35 periodic again

Other parameter choices also lead to qualitatively similar bifurcation plots, with quantita-
tively different windows of periodic behavior and chaos. We can also obtain bifurcation plots
which exhibit both periodic and chaotic windows by plotting against other parameters of
the system, such as D .

152
 CHAPTER 7. CHAOS AND NON-LINEAR DYNAMICS

7.2 Bifurcations

In our analysis of the nonlinear damped forced oscillator, we tooks snapshots (n , n ) at

= 2n for integer n to form the Poincare map. When we changed the driving amplitude a,
there were qualitative changes to the (, ) projected trajectories (which are also generally
called phase portraits) captured by the Poincare map results. In particular, we observed
period doubling at certain values of a; period doubling is a particular example of a bifurcation
(Fig.(7.7)).

A simple example of an abrupt change is when the existence/type of fixed points changes
with the systems parameters (or limit cycles, attractors, or so on) abruptly changes. These
changes are generally known as bifurcations. Since bifurcations already occur in 1-dimensional
systems, so we will start by studying these systems. We will later on find out that many
examples of bifurcations in higher-dimensions are simple generalizations of the 1D case.

For a 1-dimensional system we study the equation:

x = f (x) (7.17)

Trajectories in 1 dimension are pretty simple, we either have flow to a finite fixed point
x x or a divergence to x .

Example: The system x = x2 1, pictured in Fig. 7.8, has a stable fixed point at x? = 1
and an unstable fixed point at x? = 1. For one dimension the motion is simple enough
that we can determine whether fixed points are stable or unstable simply from this picture.
Imagine a particle moving on the x-axis. For x < 1 the red curve of x2 1 is above the
x-axis, so x > 0 and the particle moves to the right, as indicated by the blue arrow. For
1 < x < 1 the red curve is below, x < 0, and the particle moves to the left. For x > 1
the curve is again above, x > 0 and the particle moves to the right. The left point is stable
since the particle always moves towards it, while the right point is unstable and the particle
moves away from it.

153
 CHAPTER 7. CHAOS AND NON-LINEAR DYNAMICS

x 1

2.0 1.5 1.0 0.5 0.0 0.5 1.0 1.5 2.0

Figure 7.8: In this system there are two fixed points, one stable (represented by a full circle)
and one unstable (represented by the empty circle)

Stability can also be determined by linearizing about a fixed point. Although this is
overkill for one variable, the general method remains applicable for analyzing situations
with more variables, so its useful to discuss it here. Using x = x? + and expanding to
O(), then = x = x2 1 2x? , so for x? = 1, then = 2 which decays according to
e2t making the fixed point stable, while for x? = 1, then = 2 which grows according
to e2t and the fixed point is unstable.
To find the stability of fixed points in multiple dimensions, we would similarly set ~x =
~x + ~ and expand, giving a linearized system of equations after dropping O( 2 ) terms:
?

~ = M~ (7.18)
Here M is a n n matrix, whose eigenvalues and eigenvectors give us the solutions near the
fixed point, of the form ~ = ~aet . We will come back later on to discuss higher dimensional
fixed points in much more detail.
First we will categorize several types of bifurcations in one dimension, by considering the
equation
x = f (x, r) , (7.19)
where r is a parameter that we vary. The fixed points x of f (x, r) are functions of r, and
drawing them in the r x-plane gives a bifurcation diagram.

7.2.1 Saddle-Node Bifurcation

A saddle-node bifurcation is the basic mechanism by which fixed points are created and
destroyed. As we vary r two fixed points can either appear or disappear (one stable and one
unstable).

154
 CHAPTER 7. CHAOS AND NON-LINEAR DYNAMICS

Example: Consider the equation

x = r + x2 , (7.20)

which exhibits a saddle-node bifurcation at r = 0. The two fixed points disappear as we

increase r from negative to positive values, as shown in the images below.

4 4

3 3

2
x

2
x

x
1

1 1

0 0

2.0 1.5 1.0 0.5 0.0 0.5 1.0 1.5 2.0 1 1

x 2.0 1.5 1.0 0.5 0.0 0.5 1.0 1.5 2.0 2.0 1.5 1.0 0.5 0.0 0.5 1.0 1.5 2.0

x x

Figure 7.9: Two fixed

Figure 7.10: A single semi- Figure 7.11: No fixed points
points, one stable and one
stable point exists for r = 0 occur for r > 0
unstable exist for r < 0

This saddle-node bifurcation transition can be best pictured by the bifurcation diagram in
Fig. 7.12 below, where the full lines correspond to the stable fixed points and the dashed
lines the unstable ones.

2.0 2.0

1.5 1.5

1.0 1.0

0.5 0.5
x

0.0 0.0

0.5 0.5

1.0 1.0

1.5 1.5

2.0 2.0
2.0 1.5 1.0 0.5 0.0 0.5 1.0 1.5 2.0 2.0 1.5 1.0 0.5 0.0 0.5 1.0 1.5 2.0

r r

Figure 7.12: Bifurcation diagram for the Figure 7.13: Bifurcation diagram for the
system x = r + x2 system x = r x2

For the analogous equation x = r x2 we can obtain the results by interchanging x x

and r r. This gives the bifurcation diagram shown in Fig. 7.13.

155
 CHAPTER 7. CHAOS AND NON-LINEAR DYNAMICS

Example: Some flow equations like

x = r x ex = f (x, r) (7.21)

are hard to solve analytically for the fixed points, which are given by the transcendental
equation
?
r x? = ex (7.22)

Here a graphical approach suffices, where we separately plot r x and ex and look for
intersections of the curves to provide the position of the fixed points, as displayed in Fig. 7.14.

5 5 5

4 4 4

3 3 3

2 2 2
x

x
1 1 1

0 0 0

1 1 1
3 2 1 0 1 2 3 3 2 1 0 1 2 3 3 2 1 0 1 2 3
x x x

Figure 7.14: Fixed points of the system correspond to the intersections of the curves ex
and r x for r = 1.5, r = 1.0, r = 0.5 respectively. As r is varied the position of the fixed
points varies and a Saddle-Node Bifurcation occurs.

Examining which curve is larger also determines the direction of the one-dimensional flow,
and hence the stability of the fixed points.
Here the bifurcation occurs at r = rC , when the two curves are tangential and hence only
touch once: 
f 
=0 (7.23)
x x=x? ,r=rC
This gives 1 = exp(x? ) so x? (rC ) = 0. Plugging x? = 0 into Eq. (7.22) we find that
rC = 1.
By a simple generalization, we can argue that the quadratic examples x = r x2 are
representative of all saddle-node bifurcations. Taylor expanding f (x, r) near the bifurcation
point and fixed point we have
? 2 2 
  
f  f  (x x ) f
x = f (x, r) = f (x? , rC ) + (x x? )  + (r rC )  +  + ...
x x? ,rC r x? ,rC 2 x2 x? ,rC
= a(r rC ) + b(x x? )2 + . . . , (7.24)

where we have kept the first non-trivial dependence on r and x (noting that the partial
derivatives are simply some constants a and b), and two terms have vanished due to the

156
 CHAPTER 7. CHAOS AND NON-LINEAR DYNAMICS

fixed point and by the tangential bifurcation conditions:


? f 
f (x ; rC ) = 0 and = 0. (7.25)
x x=x? ,r=rC

Thus x = r x2 is the normal form of a saddle-node bifurcation. This can be determined

p
explicitly from Eq. (7.24) by making the change of variable r0 = a(r rC ) and x0 = |b|(x
x? ) to obtain x0 = r0 x0 2 where the sign is determined by the sign of b.

7.2.2 Transcritical Bifurcation

In a transcritical bifurcation a fixed point exists for all values of the parameter, but changes
its stability as the parameter is varied.

Example: Consider the equation

x = x(r x) . (7.26)

Here there are fixed points at x? = 0 and x? = r. These fixed points change their stability
at r = 0 but never disappears as illustrated graphically in Fig. 7.15.

1.0 1.0 1.0

0.5 0.5 0.5

0.0 0.0 0.0

x

0.5 0.5 0.5

1.0 1.0 1.0

1.5 1.5 1.5

1.5 1.0 0.5 0.0 0.5 1.0 1.5 1.5 1.0 0.5 0.0 0.5 1.0 1.5 1.5 1.0 0.5 0.0 0.5 1.0 1.5

x x x

Figure 7.15: Analysis of x = x(r x) for r = 1, r = 0 and r = 1 respectively. As r changes

the same type of fixed points remain, but the stability of the fixed points is swapped.

This gives us the following bifurcation diagram:

157
 CHAPTER 7. CHAOS AND NON-LINEAR DYNAMICS

2.0

1.5

1.0

0.5
x
0.0

0.5

1.0

1.5

2.0
2.0 1.5 1.0 0.5 0.0 0.5 1.0 1.5 2.0

Figure 7.16: Bifurcation diagram for the system x = x(r x) which plots the position of
the fixed points, with a full and dashed line for the stable and unstable points respectively.
Here the transcritical bifurcation at r = 0 becomes clear.

In fact, the equation x = x(r x) is the normal form of a transcritical bifurcation obtained
by expanding in a Taylor series near x = x? and r = rC .
Example: Lets consider an example with physical content, namely a model for the threshold
behavior of a laser. This can be modeled as:

n = GnN Kn ,
N = GnN f N + p . (7.27)

where the variables are N the number of excited atoms and n the number of laser photons.
The constant parameters include, f for the term governing the spontaneous emission decay
rate, G for the stimulated emission gain coefficient, K as the photon loss rate, and p as the
pump strength. Since there are two equations this is in general a two dimensional system
(which we will discuss how to analyze shortly). Here to make the equation one dimensional
we will assume rapid relaxation so that N 0, this allows us to solve for N (t) from the
second equation in Eq. (7.27) to give
p
N (t) = . (7.28)
Gn(t) + f

Plugging this back into the first equation in Eq. (7.27) then gives
n h i
n = pG K(Gn + f ) n(r x) + O(n3 ) (7.29)
Gn + f

158
 CHAPTER 7. CHAOS AND NON-LINEAR DYNAMICS

Expanding this result near n = 0 we find

n = n(r bn) + O(n3 ) (7.30)
where the constant coefficients are
pG G2 p
r K, b . (7.31)
f f2
Only n > 0 makes sense, so the critical parameter value is when r = 0, or when the pump
strength p = Kf
G
. For larger values of p this lamp turns into a laser, with a fixed point at a
non-zero n = r as illustrated in Figs. 7.17 and 7.18. The fixed point for p > Kf G
indicates
the coherent laser action.

1.0
1.0

0.5
0.5

0.0
0.0
n
0.5
n

0.5

1.0
1.0

1.5
1.5

2.0
2.0 0.0 0.5 1.0 1.5 2.0
0.0 0.5 1.0 1.5 2.0
n
n

Figure 7.18: When p > Kf /G, the stable

Figure 7.17: When p < Kf /G, the only
point of the system is with a non-zero num-
stable point is when there are no photons.
ber of laser photons.

7.2.3 Supercritical Pitchfork Bifurcation

A supercritical pitchfork bifurcation is a type of bifurcation common in problems with sym-
metries such that fixed points appear or disappear in pairs. In particular, as the parameter
is varied, one fixed point is always present but changes from being stable to being unstable
at the same place where two stable fixed points appear.
Example: The normal form for this type of bifurcation is
x = rx x3 . (7.32)
This equation is invariant under x x, so the fixed point x? = 0 is always present. On
the other hand, the fixed points at x? = r only appear when r crosses from negative to
positive values. The different cases that occur as we change r are plotted in Fig. 7.19.

159
 CHAPTER 7. CHAOS AND NON-LINEAR DYNAMICS

10 10 4

5 5 2

1
x

x
0 0 0

5 5 2

10 10 4
2.0 1.5 1.0 0.5 0.0 0.5 1.0 1.5 2.0 2.0 1.5 1.0 0.5 0.0 0.5 1.0 1.5 2.0 2.0 1.5 1.0 0.5 0.0 0.5 1.0 1.5 2.0

x x x

Figure 7.19: Plots of x = rx x3 for r = 2, 0, +2 respectively. When r becomes positive

the fixed point at x = 0 looses its stability, and two new stable fixed points appear in the
system.

This gives rise to the following bifurcation diagram:

2.0

1.5

1.0

0.5
x

0.0

0.5

1.0

1.5

2.0
2.0 1.5 1.0 0.5 0.0 0.5 1.0 1.5 2.0

Figure 7.20: Diagram for the supercritical pitchfork bifurcation. The stability of the fixed
point at x = 0 changes while two new stable points appear.

7.2.4 Subcritical pitchfork bifurcation

A subcritical pitchfork bifurcation essentially the opposite of a supercritical pitchfork bifur-
cation in that if the parameter is varied, one fixed point that is always present changes from
unstable to stable, while two unstable fixed points appear.
Example As an example consider the normal form

x = rx + x3 , (7.33)

160
 CHAPTER 7. CHAOS AND NON-LINEAR DYNAMICS

which again has a x x symmetry. Here the cubic term is destabilizing, so this exhibits
a subcritical pitchfork bifurcation as depicted in Fig. 7.21.

2.0

1.5

1.0

0.5
x

0.0

0.5

1.0

1.5

2.0
2.0 1.5 1.0 0.5 0.0 0.5 1.0 1.5 2.0

Figure 7.21: Bifurcation Diagram for the Subcritical Bifurcation x = rx + x3 . Here there is
a change in the stability of the fixed point at x = 0 and the appearance of two new fixed
points at x = r as r becomes negative.

It is interesting to consider what happens if we add a higher order stabilizing term, such as
in the equation

x = rx + x3 x5 . (7.34)

This equation supports five real solutions for a finite range of r values. This system supports
hysterisis as we increase and decrease r as illustrated in Fig. 7.22. We can imagine a path
where we start with a particle at x = 0 and r = 0.2 and then slowly increase r. When we
get to r = 0 the x = 0 fixed point becomes unstable and a small perturbation will push the
particle to another branch, such as that at x > 0. Increasing r further the particle travels
up this branch. If we then start to decrease r, the particle will travel back down this same
branch, and continue on it even below r = 0, and thus not following the same path. Then
suddenly at the critical rC < 0 where there is a saddle-node bifurcation, the particle will
again loose its stability and will jump back down to x = 0, after which the process can be
repeated.

161
 CHAPTER 7. CHAOS AND NON-LINEAR DYNAMICS

Figure 7.22: Bifurcation Diagram of the system x = rx + x3 x5 . The arrows show the
motion of the sytem as we increase and decrease r; it undergoes hysterisis.

Example: Lets consider a bead on a rotating hoop with friction, described by the equation
of motion:

ma + b = m sin (a 2 cos g) . (7.35)

Here is the angle of the bead of mass m from the bottom of the hoop, a is the radius of the
hoop, is the constant angular velocity for the rotation of the hoop (about an axis through
the center of the hoop and two points on the hoop), and g is the coefficient of gravity. Once
again to turn this into a one-dimensional problem we consider the overdamped solution.
Overdamping means we can take ma 0. The fixed points are then ? = 0 which changes
from being stable (when a 2 < g) to being unstable (when a 2 > g), while ?
= is always
present and unstable. Additionally, the stable fixed points ? = arccos ag 2 appear when
a 2 > g. This corresponds to a supercritical pitchfork bifurcation. The systems bifurcation
diagram is shown in Fig. 7.23.

162
 CHAPTER 7. CHAOS AND NON-LINEAR DYNAMICS

3
2
1
0

1
2
3
0 1 2 3 4 5 6
2
a
g

Figure 7.23: When 2 > g/a the bottom of the loop becomes and unstable fixed point and
two new stable fixed points arise that move away from the bottom as the rotation speed is
increased.

Example: As another example, consider an overdamped pendulum driven by a constant

torque described by:

= b sin() , (7.36)

where > 0, b > 0, and [, ]. For b > , the gravity beats the torque and there is
one stable and one unstable fixed points as shown in Fig. 7.24. For b < , there are no fixed
points as shown in Fig. 7.25, and here the torque wins resulting in a rotating solution. Even
when b < , there is a remnant of the influence of the fixed point in the slowing down of the
pendulum as it goes through the bottleneck to overcome gravity. Combined this is thus a
saddle-node bifurcation at = b as shown in Fig. 7.26.

163
 CHAPTER 7. CHAOS AND NON-LINEAR DYNAMICS

3.0

2.5

2.0

1.5

1.0

0.5

0.0
0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4

/b

Figure 7.24: as a function Figure 7.25: as a function Figure 7.26: Bifurcation

of when < b. Gravity of when > b. Torque plot of the system as a func-
dominates torque and there dominates gravity so there tion of the applied torque
is a stable fixed point. are no fixed points. ( /b).

7.3 Fixed Points in Two-Dimensional Systems

7.3.1 Motion Near a Fixed Point
General Categorization
In 2-dimensions, to analyze the trajectories near a fixed point ~x? = (x? , y ? ), we can
again linearize the equations of the system. Therefore, well start by analyzing a general
2-dimensional linear system with ~x? = 0. This can be written as

x = ax + by ,
y = cx + dy , (7.37)

or  
a b
~x = M~x where M= (7.38)
c d
and the matrix of coefficients here has no restrictions.
Example: Let us consider a system of equations that consists of two independent 1-
dimensional flows,

x = ax , y = y . (7.39)

We have the two independent solutions:

x(t) = x0 eat and y(t) = y0 et (7.40)

The parameter regions a < 1, a = 1, and 1 < a < 0 all produce a stable and
attracting fixed point ~x? = 0 in qualitatively different ways, because the decay rate of x(t) is

164
 CHAPTER 7. CHAOS AND NON-LINEAR DYNAMICS

either larger, equal, or smaller than that of y(t). This is shown in Figs. 7.27, 7.28, and 7.29.
If a = 0, then ~x? = 0 is no longer isolated as there is a line of fixed points at y = 0 and for
all values of x, see Fig. 7.30. If a > 0, then ~x? = 0 is a saddle point (with the y-axis being
the stable manifold and the x-axis being the unstable manifold), see Fig. 7.31.

Figure 7.27: Stable Node Figure 7.28: Stable Node Figure 7.29: Stable Node
a < 1 a = 1 1 < a < 0

Figure 7.30: Non-isolated Figure 7.31: Saddle Point

fixed points a = 0 a>0

In general in two dimensions there are more possibilities for the motion than in one-
dimension and we should be more careful about our definition for when a fixed point is
stable. For a fixed point ~x? we will say that

it is attracting if all trajectories starting in its neighborhood approach it as t ,

it is Lyapunov stable if all trajectories starting in its neighborhood remain in that

neighborhood for all time,

it is stable if it is both attracting and Lyapunov stable.

165
 CHAPTER 7. CHAOS AND NON-LINEAR DYNAMICS

Picking one fixed point from Fig. 7.30, most trajectories will be Lyapunov stable but not
attracting. If a fixed point allows a trajectory to wander away but eventually return, then
it may also be attracting but not Lyapunov stable.
In the general case defined in Eq.(7.38), we need to solve the eigenvalue problem for the
linear system, and thus find the eigenvalues and eigenvectors of M . Here we have

~a = M~a = ~a ~a(t) = a~0 et (7.41)

Therefore, as usual, we set det(M 1) = 0 where 1 is the identity matrix of the same
dimension as M . From this, defining

det(M ) = ac bd , tr(M ) = a + d, (7.42)

then the eigenvalues are given by

2 4
= . (7.43)
2

The corresponding eigenvectors are then ~a , and for a generic M they will not be orthogonal.
Assuming that two different eigenvectors exist a general solution is by linearity given by

h i
~x(t) = Re C+~a+ e+ t + C~a e t (7.44)

assuming for the moment that + 6= and taking the real part at the end if needed. There
are three main cases to consider.

1. Real eigenvalues + , R with + 6= . This is like the system in Eq.(7.40), but

with the x and y axes replaced by the directions defined by ~a+ and ~a .

Example: Consider for example a solution where ~a+ = (1, 1) and ~a = (1, 4),
ignoring normalization. If < 0 < + , then growth occurs along ~a+ and decay
occurs along ~a , so ~x? = 0 is a saddle point, as drawn in Fig. 7.32

If instead < + < 0, then decay occurs slower with + so it occurs first onto ~a+ ,
making ~x? = 0 a stable node, as drawn in Fig. 7.33

166
 CHAPTER 7. CHAOS AND NON-LINEAR DYNAMICS

Figure 7.32: Saddle Point with ~a+ = Figure 7.33: Stable Node with ~a+ = (1, 1)
(1, 1) and ~a = (1, 4) and ~a = (1, 4)

2. Let us now consider when + = = R. In this situation there can either be two
independent eigenvectors or only one. Two independent eigenvectors can only occur if

M = 1, (7.45)

in which case the fixed point is called a star, and is shown in Fig. 7.34.

If instead there is only one independent eigenvector, then the fixed point is called a
degenerate node. An example of this is


b
M= ,
0

where the eigenvalue is and which has ~a = (1, 0) as its only independent eigenvector.
Here the phase space portrait is as given in Fig. 7.35, where the trajectory decays first
onto the eigenvalue direction and then down onto the fixed point.

167
 CHAPTER 7. CHAOS AND NON-LINEAR DYNAMICS

Figure 7.35: Degenerate Node with = 1

Figure 7.34: Unstable star with > 0.
and b = 2.

3. The final case to consider is when the eigenvalues have complex parts, = i
for 6= 0. If < 0, the fixed point is a stable spiral where the trajectories spiral into
it, as in Fig. 7.36. If = 0, the fixed point is a center, with neighboring trajectories
being closed orbits around it, as in Fig. 7.37. If > 0, the fixed point is an unstable
spiral where trajectories spiral out from it, as in Fig. 7.38.

Figure 7.36: Stable spiral Figure 7.37: Trajectories Figure 7.38: Unstable Spi-
with > 0 about a center fixed point ral with < 0

As a summary if < 0 then the fixed points are saddle points, while if = 0 then the
fixed points are not isolated but form a continuous line of fixed points. If > 0, then there
are a number of possibilities:

< 2 produces stable nodes;

> 2 produces unstable nodes;

168
 CHAPTER 7. CHAOS AND NON-LINEAR DYNAMICS

2 < < 0 produces stable spirals;

= 0 produces centers;

0 < < 2 produces unstable spirals;

= 2 produces stars or degenerate nodes.

Note that all unstable fixed points have > 0, while all stable fixed points have < 0; this
is true even for stars and degenerate nodes. This information can be summarized by the
following diagram:
Non-isolated Points

Figure 7.39: Diagram determining the type of fixed point given the determinant and trace
of the linearized system.

This linearized analysis yields the correct classification for saddle points, stable/unstable
nodes, and stable/unstable spirals, but not necessarily for the borderline cases that occur
on a line rather than in an area of the plane (centers, stars, degenerate nodes, or non-
isolated fixed points). Nonlinear terms can tip a borderline case to a nearby case in the
plane. This implies nonlinear terms may only affect the stability of centers.
Analysis of a General 2-Dimensional System
Consider a general 2-dimensional system:

~x = f~(~x) = fx (x, y), fy (x, y) ,

(7.46)

169
 CHAPTER 7. CHAOS AND NON-LINEAR DYNAMICS

which may have several fixed points (x? , y ? ). We can analyze their types by linearizing about
each one, defining u = x x? and v = y y ? and expanding about (u, v) = (0, 0). Defining
~u = (u, v), then this expansion yields

fi 
~u = M~u where Mij = (7.47)
xj ~x=~x?

This is the same as a Taylor series about ~x = ~x? , where f~(~x? ) = 0.

Example: Lets consider a population growth model where rabbits (x) compete with sheep
(y). With one species, the model might look like x = x(1 x), where for small x there
is population growth, but above x > 1 food resources become scarce and the population
shrinks. For two species there can be coupling between the equations, so we could consider

x = x(3 x 2y) ,
y = y(2 y x) . (7.48)

which is called the Lotka-Volterra model. Here the parameters have been chosen to model
the fact that rabbits produce faster (3 > 2 in the linear terms) and sheep compete better
for resources (2 > 1 in the quadratic cross terms). To determine how solutions to these
equations behave we can analyze the structure of the fixed points.
The fixed points for this system are:

~x {(0, 0), (0, 2), (3, 0), (1, 1)} (7.49)

For each one we carry out a linear analysis:

~x? = (0, 0) simply gives x = 3x and y = 2y, so it is an unstable node.

~x? = (0, 2). Here we define u = x and v = y 2 and the linear equations become
u = u and v = 2u 2v. Taking the trace and determinant we find = 3 and
= 2 giving + = 1 and = 2. This is a stable node.

~x? = (2, 0) gives + = 1 and = 3, making it too a stable node.

~x? = (1, 1) gives = 1 2, making it a saddle point.

From knowing the behavior of trajectories near these fixed points we can complete the picture
for an approximate behavior of the entire system, as shown in Fig. 7.40. A diagonal line
passing through the unstable node and saddle point divides the basins of attraction for the
fixed points where the sheep win (0, 2) or where the rabbits win (3, 0).

170
 CHAPTER 7. CHAOS AND NON-LINEAR DYNAMICS

3.0

2.5

2.0

1.5
Sheep

1.0

0.5

0.0

0.5
0.5 0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0
Rabbits
Figure 7.40: Picture of the behavior of trajectories for the population growth model in
Eq. (7.48).

7.3.2 Systems with a conserved E(~x)

The mechanical systems of equations that we are most familiar with are those where the
energy is constant along trajectories. We will generalize this slightly and say systems with
any function E = E(~x) that is conserved (so E = 0) are conservative systems with E. To
rule out taking a trivial constant value of E (which would work for any system), we demand
that E(~x) must not be constant on any closed region in the space of ~x. Note that E = 0 is
generally not equivalent to f~ = 0, and hence we do not simply call these conservative
systems.
Several results follow from considering conservative systems with an E:
Conservative systems with E do not have at-
tracting fixed points. We can prove this by con-
tradiction, by imagining that such a point did
exist. Since all the points in the basin of attrac-
tion of that point must go to this single fixed
point, they must all share the same value of E,
which contradicts E not being constant within
a closed region.

171
 CHAPTER 7. CHAOS AND NON-LINEAR DYNAMICS

From our experience in expanding about the minima of potentials, we also expect to
find stable centers for conservative systems with E. This result is achieved by the
following theorem which we have essentially proven in our analysis in the chapter on
vibrations:
For ~x = f~(~x), where fi is continuous for all i, j, if E = E(~x) is conserved with an
xj
isolated fixed point ~x? at the minimum of E, then all trajectories sufficiently close are
centers.
In 2 dimensions the f~ = 0 definition of conservative is equivalent to having a
conserved E = E(~x) along the systems trajectories.
Knowing that ~x = f~(~x) and f~ = 0, then let us define:
Z y Z x
0 0
H(~x) = fx (x, y ) dy fy (x0 , y) dx0 (7.50)
Z x Z x
H fy (x0 , y) 0 fx (x0 , y) 0
= fx (x, y) dx = fx + dx
y y x0
Z y Z y
H fx (x, y 0 ) 0 fy (x, y 0 ) 0
= fy (x, y) + dy = fy dy
x x y 0

Then Hy
{fx , 2fx } and H
x
{fy , 2fy }. The first case of each occurs if fx = fx (y)
and fy = fy (x), respectively. Thus x = H y
and y = Hx
for {1, 2}. After a
trivial rescaling, these are the Hamilton equations for a conserved Hamiltonian H(~x)
(independent of t) which serves here as our function E(~x). Additionally, from the
relations the critical points ~x? of H where H|~x=~x? = 0 are identical to the fixed
points where f~(~x? ) = 0.

Example: Consider the one-dimensional classical mechanics motion given by:

x = ax x2 U 0 (x) (7.51)
with a > 0. We first turn this into one-dimensional form by writing
x = y = fx , y = ax x2 = fy . (7.52)

Since here fx is independent of x, and fy is independent of y, we obviously have f~ = 0.

We can define a conserved scalar quantity from fx and fy using Eq. (7.50) to give
y 2 ax2 x3
H= + = K(y) + U (x) (7.53)
2 2 3
where we have Hamiltons equations
H K H U
x = = and y = = . (7.54)
y y x x

172
 CHAPTER 7. CHAOS AND NON-LINEAR DYNAMICS

For this system the fixed points occur at ~x? = (0, 0) and (a, 0), which are also the extremal
points of H. For ~x? = (0, 0) we have

2H 2H 2H
= a < 0 , = 1 > 0 , and = 0, (7.55)
a2 y 2 xy
so the fixed point is a saddle point. For ~x? = (a, 0) we have

2H 2H 2H
=a>0 , = 1 > 0 , and = 0, (7.56)
x2 y 2 xy
so the fixed point is a center. These fixed points and some representative trajectories are
illustrated in Fig. 7.41. Here the bound trajectories have H < 0, while the unbound trajec-
tories have H > 0. The dividing case with energy H = 0 is the trajectory is that would stop
at the saddle point (0, 0).

0
y

3
3 2 1 0 1 2 3 4
x

Figure 7.41: Phase space picture of the system x = ax x2 with a = 2.

7.4 Limit Cycles and Bifurcations

In two dimensions, we can have a new type of 2-dimensional attractor called a limit cycle,
which is an isolated closed trajectory. For stable limit cycles trajectories nearby converge to
it as in Fig. 7.42, while for unstable limit cycles the nearby trajectories diverge from it as in

173
 CHAPTER 7. CHAOS AND NON-LINEAR DYNAMICS

Fig. 7.43. We could also imagine a semi-stable limit cycles, where the trajectories converge
or diverge on opposite sides of the cycle (an example is shown in Fig. 7.44).

Figure 7.42: Stable Limit Figure 7.43: Unstable limit Figure 7.44: Semi-stable
Cycle Cycle Limit Cycle

Note that a limit cycle is not like a center trajectory about a fixed point, because a limit
cycle is isolated from other closed trajectories, whereas around centers nearby tranjectories
are also closed.
Example: Lets consider a system of equations written with polar coordinates, x = r cos()
and y = r sin() so that
r = r(1 r2 ) , = 1 , (7.57)
with r 0. Here the circle r? = 1 corresponds to a stable limit cycle, as in Fig. 7.42. Since
only the radial coordinate matters for the stability of the limit cycle we can picture this in
one dimension, as in Fig. 7.45.

1.0

0.5

0.0
r

0.5

1.0
0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4

Figure 7.45: Behavior of the radial component of the system. The stable point is at r = 1,
meaning the system has a stable limit cycle of radius r = 1.

174
 CHAPTER 7. CHAOS AND NON-LINEAR DYNAMICS

Example: Lets consider the van der Pol oscillator (first studied in 1927 in electric circuits
and found to exhibit chaotic noise when driven)
(
x =
x + (x2 1)x + x = 0 (7.58)
= (1 x2 ) x

If x2 > 1 then the term involving gives (nonlinear) positive damping, while if x2 < 1 then
the term involving gives (nonlinear) negative damping, which is growth. For different
values the phase portrait is depicted in the figures below.

Figure 7.46: Van Der Pol Oscillator with Figure 7.47: Van Der Pol Oscillator with
= 0.1 =2

There are several known methods for ruling out limit cycles, but we will instead focus on
a method for showing they exist.

7.4.1 Poincare-Bendixson Theorem

Take a 2-dimensional system x = fx (x, y) and y = fy (x, y) with continuous and differentiable
f~. Let D be a closed, bounded region. Suppose there exists a trajectory C confined inside
D for all times t 0, then C either goes to a fixed point or a limit cycle as t .
The proof requires the use of some topology, so we wont study it. To understand how
we can use this theorem, let us suppose we have determined that there are no fixed points in
a closed, bounded region D, and at the boundarys surface the ~x points inward to trap the
trajectory in D. An example of this situation is shown in Fig.(7.48). Then due to the theorem
we must have a limit cycle in this region. Intuitively, the trajectory C wanders around D,
but it cannot self intersect and it eventually runs out of room to wander. Therefore, it must
converge to a fixed point or a limit cycle. This implies that there is no chaos in 2 dimensions.

175
 CHAPTER 7. CHAOS AND NON-LINEAR DYNAMICS

In 3 or more dimensions, trajectories have more room to wander and can do so forever,
allowing for chaos!

Figure 7.48: If at the boundary, the flow of a two-dimensional system pushes it into a region
where there are no fixed points, then the system has a stable limit cycle in that region.

Example: Lets consider whether a limit cycle exists for

(
x = x y x(x2 + 5y 2 )
(7.59)
y = x + y y(x2 + y 2 )

Using polar coordinates

rr = xx + yy rr = r2 (1 r2 r2 sin2 (2)) (7.60)

In particular, 1 r2 r2 sin2 (2) > 0for r < 1/ 2, while 1 r2 r2 sin2 (2) < 0 for r > 1.
Since there are no fixed points for 1/ 2 < r < 1 there must be a limit cycle.

7.4.2 Fixed Point Bifurcations Revisited and Hopf Bifurcations

We can revisit bifurcations by adding a varying parameter to the discussion of fixed points
and limit cycles. In particular, we now include limit cycles popping in or out of existence in
the range of things that can occur if we change a parameter.
Saddle-node, transcritical, and pitchfork bifurcations for fixed points can still occur here.
Example: As a simple example consider a system of uncoupled equations

x = x2 , y = y . (7.61)

which has a saddle-node bifurcation at = 0, as shown in the phase portraits below.

176
 CHAPTER 7. CHAOS AND NON-LINEAR DYNAMICS

2.0 2.0 2.0

1.5 1.5 1.5

1.0 1.0 1.0

0.5 0.5 0.5

0.0 0.0 0.0

0.5 0.5 0.5

1.0 1.0 1.0

1.5 1.5 1.5

2.0 2.0 2.0

3 2 1 0 1 2 3 3 2 1 0 1 2 3 3 2 1 0 1 2 3

Figure 7.49: System with Figure 7.50: System with Figure 7.51: System with
= 1 with two fixed points = 0 and one fixed point = 1 and no fixed points

More generally, we can think about determining the fixed points by drawing the curves
x = fx (x, y) = 0 and y = fy (x, y) = 0 separately. Fixed points require both equalities to be
satisfied, so we look for crossings of these two curves. Varying a parameter of the system
then leads the fixed points to slide into one another, which corresponds to a 1-dimensional
motion. This is why our study of the various types of bifurcation of fixed points in one-
dimension (saddle-node, transcritical, supercritical and subcritical pitchforks) immediately
carry over to bifurcation of fixed points in higher dimensional equations.
Example: consider the system of equations

x = x + y + sin(x) , y = x y . (7.62)

Note that these equations have a symmetry under x x and y y. This always has
~x? = (0, 0) as a fixed point. Linearizing for this fixed point yields = and = ( + 2).
So the fixed point is stable if < 2 or a saddle point if > 2.
Do to the symmetry we might expect a pitchfork bifurcation. If so, then near = 2,
there should be two more fixed points. We would need x = y, so expanding and solving we
write
x3
x = ( + 1)x + x + ... = 0. (7.63)
6
Since we are studying points near x ' 0, but with ' 2 the term with x3 can be equally
? ?
imp
portant, whereas the higher terms are subleading. This yields a solution where x = y =
6( + 2) for > 2, implying that there is a supercritical pitchfork bifurcation. This
occurs when = + = 0, which actually means + = 0 first. As we vary here the
eigenvalue crosses from negative to positive values and the stability changes.
Hopf Bifurcations
A Hopf bifurcation occurs when a spiral trajectory changes stability when a parameter
is varied, and this stability change is accompanied by the creation or destruction of limit

177
 CHAPTER 7. CHAOS AND NON-LINEAR DYNAMICS

cycles. A Hopf bifurcation is like a pitchfork bifurcation except that the limit cycle replaces
the fork in the pitchfork. Both supercritical and subcritical Hopf bifurcations exist in
analogy to pitchfork bifurcations. Here the transition of the eigenvalues of the linearized
system is different, with the real part of both eigenvalues switching sign simultaneously, as
pictured below:

Re
Re

Re

Im Im
Im

Figure 7.52: Re() < 0 Figure 7.54: Re() > 0

which gives us a stable spi- Figure 7.53: Re() = 0 which gives us an unstable
ral spiral

Example: Consider in polar coordinates the system

r = r r3 , = + br2 . (7.64)

It has a stable spiral into r? = 0 for < 0 and no limit cycles. For > 0, then r? =
is a stable limit cycle, while the spiral from r? = 0 becomes unstable. Thus, = 0 is a
supercritical Hopf bifurcation.

If we look at the eigenvalues of the linearized system at r = 0 by setting x = r cos()

and y = r sin(), then x x y and y x + y, so = i which indeed hits
Re() = 0 when = 0 as expected. The flows for this Hopf bifurcation are depicted below.

178
 CHAPTER 7. CHAOS AND NON-LINEAR DYNAMICS

Figure 7.55: System for = 0.5, b = 2 Figure 7.56: System for = 1, b = 2 and
and = 1 =1

Example: Consider the following system of equations in polar coordinates:

r = r + r3 r5 , = + br2 , (7.65)

which has a subcritical Hopf bifurcation at = 0.

Figure 7.57: System for = 0.2, b =

Figure 7.58: System for = 1, b = 2 and
2 and = 1. One of the inner orbits
= 1. There is no longer an unstable
converges to the center while the other
limit cycle in the inner region of the phase
converges to the outer limit cycle, there is
space.
an unstable limit cycle between the two.

Example: As a physics example with a limit cycle, lets consider a damped pendulum driven

179
 CHAPTER 7. CHAOS AND NON-LINEAR DYNAMICS

by a constant torque whose equation of motion is

1 =
+ + sin() = 1 (7.66)
q = sin() +
q

For 1, the fixed points are ? = 0 and sin(? ) = , for which there are two solutions
given by the solutions to ? = arcsin( ). The graphical solution for the fixed points is shown
below where we compare sin to the constant and observe where they cross. One fixed
point is stable and the other is a saddle point.2

1.0

0.5

0.0

0.5

1.0

3 2 1 0 1 2 3

Figure 7.59: Graphical determination of the value of the fixed points. We see that they
cannot occur if > 1.

What if > 1? It turns out that there is a unique stable limit cycle attractor. Consider
1h i
= q( sin ) (7.67)
q

For > 1 there are no fixed points, however for very negative , then > 0 and for very
positive , < 0. There is thus a trapping region where the system has no fixed points,
which by the Poincare-Bendixson theorem implies the existence of a limit cycle. This limit
2
See also our earlier analysis of the overdamped oscillator in Eq. (7.35), which used a slightly different
definition for the constants (q and q b).

180
 CHAPTER 7. CHAOS AND NON-LINEAR DYNAMICS

cycle corresponding to rotations of the pendulum over the top. The motion of two trajectories
with the same initial conditions, but with < 1 and > 1, are shown in Fig. 7.60.

2.5

2.0

1.5

1.0

0.5

0.0

0.5

1.0
0 2 4 6 8

Figure 7.60: Two trajectories shown from the same initial condition, one with < 1 (where
the trajectory converges to stable point), and one with > 1 (where the trajectory continues
indefinitely).

1.4
stable limit cycle
1.2
1.0
0.8 both exist

0.6 stable fixed point

0.4
0.2
0.0 q
0.0 0.5 1.0 1.5 2.0 2.5

Figure 7.61: Stable attractors and bifurcation transitions for a pendulum with a constant
applied torque.

In fact for q > 1 the limit cycle also exists for a range of values c < < 1. Since both
the fixed points and limit cycle exist for these parameter values the endpoint of the motion

181
 CHAPTER 7. CHAOS AND NON-LINEAR DYNAMICS

depends on the choice of initial conditions. Here c is a constant that depends on q, and
as q 1 then c 1. For q 1 the limit cycle only exists for 1. The boundaries
between these regions are sketched in Fig. 7.61. For the transition between the cycle and
fixed points shown as a solid (red) line, the saddle and stable node fixed points are born on
the cycle which then disappears (called an -period bifurcation). The transition across the
dashed (black) line is a saddle node bifurcation where the two fixed points are born, but the
saddle persists. Finally, for the transition across the dot-dashed (blue) line the saddle point
collides with and destabilizes the cycle, so that it seeks to exist in the region to the left (this
is called a homoclinic bifurcation). Although we have not tried to classify the full range of
possible bifurcations for systems involving a limit cycle, this example has illustrated a few
of the possibilities.

7.5 Chaos in Maps

In nonlinear systems with 2 variables, we have obtained a qualitative analytic understanding

of the motion by analyzing fixed points and limit cycles. The analysis of 2 variables includes
the possible motion for a 1-dimensional particle with two phase space variables. There is no
chaos with 2 variables. We could study chaos with 3 variables, but is there a simpler way?
Recall that chhaos in the Poincare map of the damped driven nonlinear oscillator could
be found from
(
N +1 = f1 (N , N )
(7.68)
N +1 = f2 (N , N )

which are 2 discrete variables. Here we set = 2N to be discrete with N an integer.

Uniqueness for the 3 continuous variable solution implies the existence of f1 and f2 . In fact,
for general systems, we can go a step further. Chaos already exists in 1-dimensional maps
xN +1 = f (xN ) for a nonlinear function f .
The example we will be using to illustrate chaos in maps is the logistic map:

xn+1 = f (xn ) = rxn (1 xn ) (7.69)

which has a parameter r. If we take 0 < r 4, then the {xn } are bounded by 0 x 1,
since the maximum is f (1/2) = r/4. We can visualize this solution by a plot in the xn xn+1
plane:

182
 CHAPTER 7. CHAOS AND NON-LINEAR DYNAMICS

1.0

r =1
0.8 r =2
r =4
0.6

xn +1
0.4

0.2

0.0
0.0 0.2 0.4 0.6 0.8 1.0

xn
The fixed points of a general map satisfy:

x? = f (x? ) (7.70)

which is slightly different from nonlinear differential equations, as these are now iterated
difference equations. For our logistic map example this gives

x? = 0 for all r
? ? ?
x = rx (1 x ) 1 (7.71)
x? = 1 for r > 1
r

The stability of a fixed point can be found by checking a small perturbation


? ? ? ? df 
xN = x + N xN +1 = x + N +1 = f (x + N ) = f (x ) + N + O( 2 ) (7.72)
dx x=x?

to obtain 
df 
N +1 = N (7.73)
dx x=x?
Therefore if
 
 df 
  < 1 lim N = 0: x? is stable.
dx x=x? N

 
 df 
  > 1 lim N : x? is unstable.
dx x=x? N

 
 df 
  = 1, then x? is marginal (requiring an expansion beyond linear analysis).
dx x=x?

183
 CHAPTER 7. CHAOS AND NON-LINEAR DYNAMICS

For the logistic map, Eq. (7.69) we have:


df 

=r x? = 0


df dx x=0
= r 2rx  (7.74)
dx df 
? 1
=2r x =1 r
dx x=x?


The first case is stable if r < 1, and the second is stable if 1 < r < 3 and unstable otherwise,
which we show graphically in Fig. 7.62. Thus we find that

0 r<1
lim xn = 1 . (7.75)
n 1 1<r<3
r
For r > 3, the limN xN is not well-defined as a single number given by a fixed point. So
what happens?

1.0

r >3
0.8

1 <r <3
0.6
xn +1

0.4

0.2
r <1

0.0
0.0 0.2 0.4 0.6 0.8 1.0

xn

Figure 7.62: Stability of fixed points of the Logistic Map.

To find out, consider two iterations of the map:

xn+2 = f (f (xn )) (7.76)

which makes xn+2 a 4th order polynomial of xn as shown in Fig. 7.63 on the right. Here there
are three fixed points of the double iterated map, two are stable (which we call p and q) and
one is unstable. Furthermore we find that p = f (q) and q = f (p), so the n state of the

184
 CHAPTER 7. CHAOS AND NON-LINEAR DYNAMICS

Logistic map is an oscillating 2-cycle as shown in Fig. 7.63 on the left. Thus the (discrete)
period has doubled and we call this a pitchfork bifurcation of the map at r = 3.

1.0 1.0

0.8 0.8

0.6 0.6
xn +1

xn +2
0.4 0.4

0.2 0.2

0.0 0.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

xn xn

Figure 7.63: The logistic map for r > 3 has fixed points for the double iterated mapping,
which are a two-cycle for the original map.

If we analyze the stability of p and q, we find that more bifurcations occur for higher
values of r. Since

d df (p) df (q)
(f (f (x))) |x=x? = (7.77)
dx dp dq

? ?
for
 x =  p or x = q, this implies that p and q lose their stability simultaneously when
 df (p) df (q) 
 dp dq  > 1. At this point the 2-cycle bifurcates into a 4-cycle. This pattern of period
doubling continues, 2 4 8 16 32 . . ., until r = 3.5699456 . . .. Beyond that
point the map becomes chaotic. This behavior is shown in Fig. 7.64.

185
 CHAPTER 7. CHAOS AND NON-LINEAR DYNAMICS

x*
1.0

0.8

0.6

0.4

0.2

r
1.0 1.5 2.0 2.5 3.0 3.5 4.0
Figure 7.64: Bifurcation plot for the Logistic Map, showing points that are part of the non-
transient motion as a function of r. Below
r = 3 there is a single fixed point. The first two
bifurcations occur for r = 3 and r = 1 + 6. Each new bifurcation is closer to the previous,
until we reach the chaotic regime. In the middle of the chaotic region there are non-chaotic
regions, such as the one near r = 0.384 visible as a white stripe.

This is called a period doubling road to chaos and is one common mechanism by which
chaos emerges. Indeed, this phenomena also occurs in the nonlinear damped driven oscillator.
But how do we know that it is chaos?
If chaos occurs in a map, then we should have sensitivity to initial conditions. Examine
x0 x1 x2 . . .
x 0 + 0 x 1 + 1 x 2 + 2 . . .
where n is the separation between two initially neighboring trajectories after n iterations.
As such we expect limn1 |n | |0 |en ; there should be exponential separation with > 0
for chaos to occur, where is called the Lyapunov exponent.
For maps we can derive a formula for as follows. We know that:
1  f n (x0 + 0 ) f n (x0 ) 
   
1  n 
= lim ln   = lim ln  (7.78)
n n 0 n n 0 

where f n (x) = f (f (f (. . . x))). Assuming that 0 is very small this implies

| {z }
n times
 

1  d n

 1 n1
Y df (xj ) 
' lim ln  f (x0 ) = lim ln  , (7.79)

n n dx0 n n  dxj 
j=0

186
 CHAPTER 7. CHAOS AND NON-LINEAR DYNAMICS

where the xj are the points along the map trajectory so far. This gives

n1  
1 X  df (xj ) 
= lim ln  (7.80)
n n dxj 
j=0

as a formula we can use to compute the Lyapunov exponent by keeping track of this sum as
we increase n. The result is shown in Fig. 7.65. In period doubling regions < 0, while in
chaotic regions > 0. There may also be periodic windows with chaos on either side. For
the logistic map, the largest such window is the 3-cycle near r 3.83. These windows are
also clearly visible in the bifurcation diagram.

1.0

0.5

r
3.2 3.4 3.6 3.8 4.0
-0.5

-1.0

-1.5

Figure 7.65: Value of the Lyapunov Exponent as a function of r. The chaotic regimes
correspond to > 0. (Finite sampling leads to the discrete points.)

You may have noticed that period doubling occurs after progressively shorter intervals as
the parameter r is increased in the case of the logistic map. In fact, for a wide class of maps
(and nonlinear differential equations), this speed-up is characterized by a universal number.
For a parameter r, denoting r as the value where the th period doubling occurs, then

r r1
F = lim 4.669201 (7.81)
r+1 r

is the Feigenbaum number. For the logistic map, it is easy to check that we are already
pretty close to this number for small . Given this, we can estimate where chaos starts as
follows:
n1 1
1 = r2 r1 = 62 and n = rn+1 rn = = . . . = n1 , (7.82)
F F

187
 CHAPTER 7. CHAOS AND NON-LINEAR DYNAMICS

so our estimate for where chaos starts is

X X (n1) 1
r = lim rn = 3 + n ' 3 + 1 F =3+ = 3.572 (7.83)
n
n=1 n=1
1 1F

which is fairly close to the real value of r = 3.5699456.

You may be wondering how trajectories can diverge exponentially (initially) while still
remaining bounded. The mechanism is by stretching and folding of trajectories. Think of
a drop of food coloring on cookie dough which you then fold and knead. To see this more
explicitly, lets once again returning to the logistic map in Eq.(7.69), but now we take r ' 4.
Then:
 
1
xn 0, is sent to xn+1 (0, 1) (7.84)
2
 
1
xn , 1 is sent to xn+1 (1, 0) (7.85)
2

so the two resulting intervals are the same however in opposite directions. Together the
original (0, 1) interval is both stretched and bent. We then repeat this, with the phase space
structure getting progressively more complicated as depicted below:

Finally, there is a self-similarity property of the bifurcation diagram for the logistic map.
When we zoom in on regions of smaller scales of r, we see the same picture again, including
the periodic windows, chaotic regions, and period doubling. This is a property of fractals
that well see shortly.

188
 CHAPTER 7. CHAOS AND NON-LINEAR DYNAMICS

7.6 Chaos in Differential Equations, Strange Attrac-

tors, and Fractals
7.6.1 The Lorenz Equations
The prototypical example of chaos in differential equations consists of the Lorenz equations

x = (y x)

y = rx y xz (7.86)

z = bz + xy

where the 3 parameters , r, and b are all positive. Note the symmetry under x x
and y y. Lorenz discovered chaotic behavior in his study of atmospheric modeling,
which he showed also appeared in these simpler three equations. This serves as a simplified
model of a fluid in a convection roll, with x being the average velocity in the loop, y being
the temperature difference between the flow on the two halves of the roll, and z being
the temperature difference between the inside and outside of the roll. One can think of
these equations as an approximation arising from the full Navier-Stokes and heat transfer
equations.
The fixed points arepx? = y ? = z ? = 0 which is stable for r < 1 or a saddle point for
r > 1, and x? = y ? = b(r 1) and z ? = r 1 which only exist for r > 1. At r = 1 their
is a supercritical pitchfork bifurcation of the fixed point. With some work, we can show that
the r > 1 stable fixed points only remain stable up to r = rH , and are unstable beyond that.
At this point r = rH the stable fixed point prongs each become unstable under subcritical
Hopf bifurcations, which involve a collision with an unstable limit cycle that shrinks onto
each of the fixed points. This is shown in Fig. 7.66.

Figure 7.66: Bifucation Diagram of the x coordinate of the Lorentz system

189
 CHAPTER 7. CHAOS AND NON-LINEAR DYNAMICS

We also know that in this system the phase space volumes contract f~ = (+b+1) < 0
(so the system is dissipative). It can also be shown that trajectories are (eventually) bounded
by a sphere x2 + y 2 + (z r )2 = constant.
In the Lorenz system, for r < rH the trajectories converge on a stable fixed point. What
happens for r > rH ? The trajectories are bounded and the phase space volume shrinks, but
there are no stable fixed points or stable limit cycles to serve as attractors. Instead, we have
chaos with a strange attractor, which is depicted in Fig. 7.67.

z
45

40

35

30

25

20

15

x
-10 -5 5 10

Figure 7.67: Strange attractor in the chaotic regime of the Lorenz equations, shown in the
3-dimensional space as well as for the xz projection.

In a strange attractor, the trajectories still never cross (in the 3 dimensions), and the
attractor trajectory exhibitsRexponential sensitivity to initial conditions. It also has zero
volume consistent with V = f~ dV , but interestingly, it has infinite surface area! There
are infinitely many surfaces traced out by cycles near the fixed points, so the attractor is a
fractal. For the Lorenz system, surfaces are different after each pass from x > 0 to x < 0
and vice versa; this attractor is a fractal with dimension 2 < D < 3. In fact D ' 2.05 in
this case.
How can we have exponential divergence of trajectories while the phase space volume
shrinks? For the Lorenz system, we have 3 variables, so there are 3 directions in which
trajectories can converge or diverge. (In general, these directions are more complicated than
simply fixed Cartesian axes. We must find the principal axes at each time.) Thus there are
in principal 3 exponents governing the trajectories:

j = j0 ej t for j {1, 2, 3} V (t) V0 e(1 +2 +3 )t (7.87)

190
 CHAPTER 7. CHAOS AND NON-LINEAR DYNAMICS

For a case where f~ is constant this means that:

Z
V = (1 + 2 + 3 )V (t) = f~ dV = f~ V (t) < 0 .
(7.88)

Note that f~ = ( + b + 1) is constant for the Lorenz system. This means 1 + 2 +

3 = f~ < 0, and the system is dissipative if the sum of the exponents is negative,
indicating that the volume shrinks overall. But exponential sensitivity to initial conditions
only requires j > 0 for at least one value of j. Here, the Lyapunov exponent is defined as
max(1 , 2 , 3 ).
For the nonlinear damped driven oscillator, we also have 1 + 2 + 3 = f~ = 1q < 0.
Here, things are even simpler because = D t has 3 = 0, so 1 + 2 = 1q . For the
undamped case (q ), we can still have chaos with 1 = 2 > 0. Thus we note that
chaos can occur in both conservative and dissipative systems.
We can think of an area in phase space as it gets stretched and contracted as pictured
below. Here it is stretched by the exponent 1 > 0 and contracted by 2 < 0. If trajectories
are bounded then it must also get folded.

Figure 7.68: The action of the system leads to stretching and rotation of phase space volume.

7.6.2 Fractals and the Connection to Lyapunov Exponents

Fractals are characterized by a nontrivial structure at arbitrarily small length scales. In
particular, they are self-similar (in that they contain copies of themselves at arbitrarily
small scales).
Example: The Cantor set fractal is created by iteratively removing the middle 13 of a line
segment. If S0 is the line segment 0 x 1, then S1 is formed by removing the middle 13
of S0 , S2 is formed by removing the middle 13 of each piece of S1 , and so on until the true
Cantor set emerges as S . This is pictured in Fig. 7.69. Here the number of separate pieces
grows infinitely large (and is in fact non-denumerable), while the total length of the pieces

191
 CHAPTER 7. CHAOS AND NON-LINEAR DYNAMICS

tends to zero. (This is the one dimensional analog of area with volume 0 for the
Lorentz equations strange attractor.) The Cantor set also has the self-similar property.

Figure 7.69: Illustration of the iterative procedure that produces the Cantor Set.

How do we define a non-integer dimension for fractals? Let us consider covering a line of
length a0 with segments of length a. We would need N (a) = aa0 segments. For a square of side

2
length a0 covered by smaller squares of side length a, we would need N (a) = aa0 squares.
In general, for a D-dimensional hypercube of side length a0 covered by D-dimensional hy-

D
percubes of side length a, we would need N (a) = aa0 such hypercubes for integer D. This
can be generalized beyond integers to
ln(N (a))
dF = lim (7.89)
a0 ln a0


a

which is the Hausdorff dimension (also called the capacity dimension or the fractal dimen-
sion).
Example: in the Cantor set, after n steps, the number of segments is:
N (a) = 2n (7.90)
while the length of each segment goes as:
a0
an = (7.91)
3n
Thus the fractal dimension is given by:
ln(2n ) ln(2)
dF = lim = ' 0.6309 (7.92)
n ln(3n ) ln(3)
indicating that it is less than a line with dF = 1 but more than a point with dF = 0.
In general, fractal dimensions are not integers and are usually irrational.
Example: The Koch curve is like the Cantor set, except that instead of deleting the middle
1
3
of every segment, we replace it by an equilateral triangle on the other two sides, so segments
are overall added rather than removed. The Koch curve corresponds to one of the sides of
the Koch Snowflake depicted below in Fig. 7.70. In this case:
a0 ln(4)
N (a) = 4n and an = dF = ' 1.262 (7.93)
3n ln(3)

192
 CHAPTER 7. CHAOS AND NON-LINEAR DYNAMICS

which satisfies 1 < dF < 2. This means the Koch curve has infinite length (1-dimensional
volume) but zero area (2-dimensional volume).

Figure 7.70: The Koch Curve corresponds to starting with just one of the 3 sides of the
triangle used to generate the Koch Snowflake shown here.

We can connect the notion of a fractal dimension to Lyapunov exponents which govern
the behavior of physical trajectories. For simplicity, let us consider an example with 1 > 0,
2 < 0, and 3 = 0. The area of a square of phase space points evolves as:

A(t = 0) = a20 A0 (t) = a20 e(1 +2 )t (7.94)

while the squares covering it have area A(t) = a20 e22 t , see Fig. 7.71. Therefore

A0 (t)
N (t) = = e(1 2 )t (7.95)
A(t)
This gives rise to a fractal dimension of:
1
dF = 1 + (7.96)
|2 |
which is the Kaplan-Yorke relation. A fixed point attractor has dF = 0, and a limit cycle
attractor has dF = 1. By contrast, a strange attractor generally has a non-integer dF , and
this dimension is related to the sensitivity to initial conditions (given by 1 ) as well as to
the contraction of phase space (given by 2 ).

193
 CHAPTER 7. CHAOS AND NON-LINEAR DYNAMICS

Figure 7.71: As the system evolves the phase space volume changes, so our tiling volume
changes as well.

7.6.3 Chaos in Fluids

Chaos can occur in fluids as well. If we take v = 0 and to be constant and uniform,
the Navier-Stokes equation says:

v P
= v v + 2 v (7.97)
t

and this should be used in conjunction with the heat transfer equation. In the language
we have been using in this chapter, the velocity field v(x, t) corresponds to a continuum of
variables (each labeled by x). One can also think of the terms involving v as couplings
between these variables, like finite differences, for example:

vx vx (x + ) vx (x )
(7.98)
x 2
In some cases (as in convection rolls per the Lorenz equations), we can have aperiodic
time dependence but spatial regularity in x. Here, many of the ideas that we have studied
(like, for example, the period doubling road to chaos) apply. In other cases, the spatial
structure in x also becomes irregular. The regularity (or lack thereof) can also depend
on initial conditions. This happens, for example, in fat convection rolls in shallow fluids.
Essentially there could be multiple attractors present. For the case with irregularity in x,
the dimensionality of the attractor is proportional to the size of the system, which is very
large! Here it makes more sense to speak of a dimension density.
Strong turbulence in a fluid falls in the category of being irregular in x with no charac-
teristic size for features. This is certainly more advanced than our examples, and indeed a
full formalism for turbulence remains to be invented. One thing we can do to characterize
strong turbulence is apply dimensional analysis.

194
 CHAPTER 7. CHAOS AND NON-LINEAR DYNAMICS

There are several scaling laws for turbulence in 3 dimensions. Recall that vortices (eddies)
appear at all length scales and are efficient at transferring energy. Let us define L as the
size of the fluid container, 0 as the scale where dissipation is important (for Reynolds
number R 1),  as the mean energy transfer per unit time per unit mass, and v as
the velocity variation at length scale . Note that the dimensions [] = m2 /s and [] =
(kgm2 /s2 )(1/skg) = m2 /s3 . There are three scales to consider.
vL3
1. At L, there can be no dependence on , so  . (This is the scale with the
L
most kinetic energy and the largest energy.)
v3
2. At 0   L, there can still be no , so here  . Note that this is independent

of the properties , and the scale L of the fluid!

3. At 0 , because R = v00 1, then v0 0 . This is where the energy dissipation

3
occurs. Here we only have and 0 present, so  4 .
0

Rather than using and v , the universal result for the case 0   L is often written
1
in terms of the wavenumber k and kinetic energy per unit mass per unit wave number,

E(k). The kinetic energy per unit mass can be written as E(k) dk. Here E(k) behaves as a
rescaled version of the energy with slightly different dimensions, [E(k)] = m3 /s2 . Analyzing
its dimensions in relation to  and k we note that m3 /s2 = (m2 /s3 )2/3 (1/m)5/3 which yields
2 5
E(k)  3 k 3 (7.99)

This is the famous Kolmogorov scaling law for strong turbulence. It provides a mechanism
by which we can make measurements and probe a universal property of turbulence in many
systems.
The End.

195
MIT OpenCourseWare
https://ocw.mit.edu

8.09 Classical Mechanics III

Fall 2014

For information about citing these materials or our Terms of Use, visit: https://ocw.mit.edu/terms.