Partial Differential Equations and Waves

Bernard Deconinck
Department of Applied Mathematics
University of Washington
Campus Box 352420
Seattle, WA, 98195, USA

May 22, 2017

 i

Prolegomenon

These are the lecture notes for Amath 353: Partial Differential Equations and Waves. This
is the first year these notes are typed up, thus it is guaranteed that these notes are full of
mistakes of all kinds, both innocent and unforgivable. Please point out these mistakes to
me so they may be corrected for the benefit of your successors. If you think that a different
phrasing of something would result in better understanding, please let me know.

These lecture notes are not meant to supplant the textbook used with this course. The
main textbook is Roger Knobel’s “An introduction to the mathematical theory of waves”,
American Mathematical Society 1999, Student Mathematical Library Vol 3.

These notes are not copywrited by the author and any distribution of them is highly
encouraged, especially without express written consent of the author.
ii
Contents

1 Introduction 1
1.1 An introduction to waves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 A mathematical representation of waves . . . . . . . . . . . . . . . . . . . . 3
1.3 Partial differential equations . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

2 Traveling and standing waves 11

2.1 Traveling wave solutions of PDEs . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2 The sine-Gordon equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.3 The Korteweg-de Vries equation . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.4 Wave fronts and pulses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.5 Wave trains and dispersion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.6 Dispersion relations for systems of PDEs . . . . . . . . . . . . . . . . . . . . 21
2.7 Information from the dispersion relation . . . . . . . . . . . . . . . . . . . . 22
2.8 Pattern formation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.9 A derivation of the wave equation . . . . . . . . . . . . . . . . . . . . . . . . 27
2.10 d’ Alembert’s solution of the wave equation . . . . . . . . . . . . . . . . . . 30
2.11 Characteristics for the wave equation . . . . . . . . . . . . . . . . . . . . . . 33
2.12 The wave equation on the semi-infinite domain . . . . . . . . . . . . . . . . . 36
2.13 Standing wave solutions of the wave equation . . . . . . . . . . . . . . . . . 42

3 Fourier series 49
3.1 Superposition of standing waves . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.2 Fourier series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.3 Fourier series solutions of the wave equation . . . . . . . . . . . . . . . . . . 56
3.4 The heat equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
3.5 Laplace’s equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
3.6 Laplace’s equation on the disc . . . . . . . . . . . . . . . . . . . . . . . . . . 64

4 The method of characteristics 69

4.1 Conservation laws . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.2 Examples of conservation laws . . . . . . . . . . . . . . . . . . . . . . . . . . 71
4.3 The method of characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . 73
4.4 Breaking and gradient catastrophes . . . . . . . . . . . . . . . . . . . . . . . 80

iii
iv CONTENTS

4.5 Shock waves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89

4.6 Shock waves and the viscosity method . . . . . . . . . . . . . . . . . . . . . 95
4.7 Rarefaction waves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
4.8 Rarefaction and shock waves combined . . . . . . . . . . . . . . . . . . . . . 103
4.9 Weak solution of PDEs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
Chapter 1

Introduction

1.1 An introduction to waves

In this course we will learn different techniques for solving partial differential equations
(PDEs), specifically with an emphasis on wave phenomena. We begin by defining, in a very
loose sense, what we mean by a wave.
Let’s start with some examples. We have students majoring in very many different fields
enrolled in this class. Waves are relevant for all of you. I will not always pick the most obvious
examples below, as you’ll have to come up with some examples on the first homework set.

ˆ ACMS and Mathematics: The study of wave phenomena has produced some of
the biggest mathematical breakthroughs of the last decades. Waves have helped our
understanding of geometry, algebraic geometry, analysis, and just about any other area
of mathematics. And, obviously, as an ACMS major, all applications listed here are
relevant to you.

ˆ Aeronautical and Aerospace Engineering: Radar is an important example of the

use of sound waves: waves are sent out and their reflection is observed. This is the
same process used by bats to determine where they are in the middle of the night or
in a dark cave. Sonic booms are another example we will talk about.

ˆ Atmospheric Science: Patterns in clouds are an obvious example of waves. Spe-

cific examples are the Morning Glory phenomenon off the coast of Australia (https:
//en.wikipedia.org/wiki/Morning_Glory_cloud) and the Kevin-Helmholtz insta-
bility http://en.es-static.us/upl/2014/05/kelvin-helmhotz-clouds1.jpg.

ˆ Biology: Dispersal of seeds by wind waves is one of the most important means of
plant regeneration.

ˆ Computer Science and Computer Engineering: Electromagnetic waves are re-

sponsible for how we see, how we communicate (using our smart phones, for instance,
or even through old-fashioned wired communications). These signals are put in on one

1
2 CHAPTER 1. INTRODUCTION

end, they are transmitted, and finally they are received on the other end. If we’re
dealing with digital communication, the waves we’re talking about are sequences of
zeros and ones.

ˆ Bioresource Engineering: the movement of plant habitats as a function of a chang-

ing climate is described by a wave that moves a LOT slower than most of the ones we
describe here: a noticable change can take decades to be observable.

ˆ (Bio) Chemistry and BioEngineering: The Belousov-Zhabotinsky reactions dis-

plays both temporal and spatial oscillations that are easily observed with the naked
eyes, see https://www.youtube.com/watch?v=IBa4kgXI4Cg. This is not your typical
chemistry 101 reaction, where the reactants react, and the final product appears.

ˆ Mechanical Engineering: The understanding of tidal and other water waves is im-
portant for harbor and ship design, see https://www.dropbox.com/s/mlittwqaxk6j3we/
hexagons-bw.jpg?dl=0.

ˆ Oceanography: Do I need to say tsunami? Or rogue wave? In fact, a lot of the

terminology we will introduce originates from the study of water waves. That is most
likely because water waves are so obvious to observe: at the very least we can see the
patterns we’re talking about.

ˆ Physics: Waves in Quantum Mechanics: the electron microscope allows us to “see”

matter at an atomic scale by sending in an electron wave (a beam of electrons) which
interacts with the surface of the material we are “looking at”. By analyzing the reflected
and the transmitted wave, we can determine the nature of the surface.

ˆ Political Science and Economics, International Studies: The propagation of

political or economical ideas can be described by a wave. As we are leading up to
an election, different candidates generate different waves of enthusiasm based on a
variety of external forces (media, advertizing, etc). These waves are also affected by
the medium in which they propagate: Iowa first, New Hampshire second, and so on.

ˆ Psychology: A delta wave is a brain wave that occurs during what is known as
deep sleep. It is known that delta wave activity is vastly reduced in people with
schizophrenia.

What do all the phenomena mentioned above have in common? It turns out that it is
not so easy to give a precise definition of a wave that captures everything we want. We can
agree on the following.

1. A wave is the result of a disturbance propagating through a medium, with finite ve-
locity, and

2. Associated with waves are signals: the result of any kind of measurement of a wave.
The outcome could be called Amplitude, Frequency, etc.
1.2. A MATHEMATICAL REPRESENTATION OF WAVES 3

u(x, tn )

u(x, 0) u(x, 3)
u(x, 2)

u(x, 1)

Figure 1.1: A wave of decreasing height (i.e., amplitude) traveling to the right.

Example. Ripples in a pond. Waves travel horizontally across the surface of the
pond. A signal might be the vertical displacement of the crests.
Example. Waves in traffic disturbances can be caused by an accident, a police car,
a traffic light, a car merging, etc. We will discuss this example near the end of the course.
And I apologize: this knowledge will not let you race through Seattle traffic. But it will tell
you why you are stopped. That should make you happier, in a zen-like way, no?
Example. The wave in a sports stadium.

1.2 A mathematical representation of waves

Let’s start with one-dimensional waves. These are waves that propagate in one direction
only, say the x direction. For the sake of argument, think of a signal along a string (like a
guitar or violin string) or a channel, or a queue of people.
Clearly, the wave signal will depend on where we measure it, and when we measure it.
In other words, the signal is a function of both x (space) and t (time). Thus, we are looking
for a function of two variables, u(x, t).
At a fixed time t0 , we can take a snapshot of the wave signal u(x, t0 ). Similarly, we can
put a probe in a specific location x0 to get the time signal u(x0 , t). To visualize wave signals,
we often take a series of snapshots at different fixed times t1 , t2 , t3 , . . . . This is illustrated
in Fig. 1.1 and Fig. 1.2. Of course, we can always make a three-dimensional plot too (with
x and t as the horizontal axes, u(x, t) the vertical), or even a movie. Both those last two
options don’t work as well on a two-dimensional sheet of paper.
Example. The function

u(x, t) = f (x − vt), v > 0,

4 CHAPTER 1. INTRODUCTION

u(x, tn )

u(x, 0)
u(x, 2.6)

u(x, 3.9) u(x, 1.3)

Figure 1.2: A shock wave traveling to the left.

gives a profile f (x) that is moving to the right (because v is positive) with speed v. If v < 0,
the profile would be moving to the left, with speed |v|. Indeed, at t = 0, we have
u(x, 0) = f (x),
so this is our initial profile. At a later time t > 0, we get the same profile, but it has moved.
Suppose we wanted to track the value f (0). At any time t, this value can be found at the
position given by
x − vt = 0 ⇒ x = vt,
which is positive, and increasing linearly with time t. Specifically, the speed at which the
value f (0) moves to the right is given by
dx
= v,
dt
which verifies our claim. The same reasoning works for tracking any value of f (x)1 .
Example. As an example of our example (Really? Yup, really.), we consider
u(x, t) = sech2 (x − 5t).
I know you all love hyperbolic functions, so I won’t waste much time recalling what they
are. We have that
1
sech(x) = ,
cosh(x)
and
ex + e−x
cosh(x) = .
2
1
When I make such gratuitous statements, you should check them. I do not make them to sound smart.
If I did, I’d make them in Latin.
1.2. A MATHEMATICAL REPRESENTATION OF WAVES 5

cosh(x) sech(x) sech2 (x)

Figure 1.3: The graphs for cosh(x) (left), sech(x) (middle), and sech2 (x) (right).

Figure 1.3 shows what the plot of this function looks like at time t = 0. Once we know that,
we know that u(x, t) has the same plot at all time, but translated to the right by an amound
5t.
Example. The wave u(x, t) = sin(x + t) represents a sine function at t = 0, which moves
to the left with speed 1.
Example. Consider
u(x, t) = H(x − 7t),
where H(x) is the Heaviside step function:

1 if x ≥ 0,
H(x) =
0 if x < 0.

Its graph is drawn in Fig. 1.4. Thus u(x, t) is a step profile, moving to the right with
velocity 7.

Visualizing functions of two variables

We have different options for visualizing functions u(x, t) of two continues variables x and
t. Which one we prefer often depends on how we are communicating our results. Check out
the visualization.mw file in the software folder.

ˆ Animation. We show a movie of u(x, t) as t changes from an initial to a final value.

Not so great on paper (unless you have a flip book, which people sometimes do!), but
works really well in a presentation.

ˆ Slice plots. We create a bunch of slices and we plot them together, either in 2-D or
in 3-D. Great for paper communication.
6 CHAPTER 1. INTRODUCTION

H(x)

0
x

Figure 1.4: The graphs for cosh(x) (left), sech(x) (middle), and sech2 (x) (right).

ˆ Surface plots. We plot u(x, t) as a surface depending on the two variables x and
t. Also good for paper communication, provided that the three dimensions come out
well. This depends heavily on u(x, t).

ˆ (x, t)-plots, contour plots. We plot u(x, t) as a function in the (x, t) plane, but
looked at from above. We use different colors or grey scale to indicate the height
u(x, t). Great on paper.

1.3 Partial differential equations

A partial differential equation (PDE) for a function u(x, t) is a differential equation that
relates different derivatives of u(x, t) to each other and to u(x, t). That’s the same definition
as for ordinary differential equations, but now, because u(x, t) depends on more than one
variable, some of the derivatives will be with respect to x, others will be with respect to t.
We often use shorthand:
∂u ∂u ∂ 2u
ut := , ux := , uxt := ,
∂u ∂x ∂x∂t
and so on.
Example. The advection equation is given by

ut + cux = 0,

where c is a parameter. This equation is linear, since it contains no products of u with itself
or any of its derivatives, and first order in both x and t, since no derivatives of higher than
1.3. PARTIAL DIFFERENTIAL EQUATIONS 7

first order appear. The equation is also homogeneous, because u = 0 is a solution, albeit2
a not very interesting one.
Example. The diffusion equation is given by

ut = σuxx ,

where σ > 0 is a parameter. This equation is also linear, but is is of second order in x, first
order in t. The equation is homogeneous. Since it arises in the study of heat transport, it is
also known as the heat equation.
Example. The three-dimensional diffusion or heat equation is

ut = σ(uxx + uyy + uzz ), σ > 0.

As before, the equation is linear and homogeneous. It is first order in t, and second order in
the spatial variables x, y and z.
Example. The Burgers equation is given by

ut + uux = 0.

It is first order in x and t. It is our first example of a nonlinear PDE, because of the second
term. It is still homogeneous, but for nonlinear equations that won’t buy us much.
Example. The equation
ut + uux + uxxx = g(t)
is nonlinear (evil second term), first order in t, third order in x, and nonhomogeneous if
g(t) 6≡ 0.
Example. The sine-Gordon equation3 is given by

uxt = sin u.

It is nonlinear (why?), first order in x and t. It is homogeneous.

Example. In contrast, the sadly unnamed equation

uxt = cos u,

is not homogeneous. It is nonlinear (again, why?), first order in x and t.

Let’s bring some intuition in. The quantity ut (x0 , t) (where x0 is fixed) denotes the rate
of change of u as t changes, at a fixed location. In other words, we put a probe at a certain
spot, and we look at how fast the time signal changes. Similarly, ux (x, t0 ) (where t0 is fixed)
2
That’s a funny word!
3
Yes, that’a a pun, which you will recognize if you are a quantum physicist. If you are a quantum
physicist, this class may be a bit too basic for you.
8 CHAPTER 1. INTRODUCTION

uxx > 0
uxx > 0

uxx < 0

Figure 1.5: An initial profile for the heat equation and how it will evolve.

is the rate of change of u as x changes, at a fixed time. This corresponds to taking a snapshot
of u, and looking at the slope of the curve in the picture. It follows that both ux and ut are
interpreted as velocities. Similarly, both uxx and utt can be seen as accelerations.
Example. Let u represent the temperature in a metal rod at position x, at time t. Then
u satisfies the so-called heat equation
ut = Duxx ,
where D > 0 is the heat conductivity of the rod, assumed to be constant here. Suppose we
start with an initial temperature profile, as in Fig. 1.5. In those regions where uxx > 0 (in
other words, the function is concave up), the heat equation states that ut will be positive,
so u will increase in time. For those regions where the profile is concave down, the heat
equation gives that ut will be negative, i.e., u will decrease in time, as indicated in the
figure. It follows that the heat equation likes to smear out profiles to a constant value.
Example. Next, we consider u to be a solution of the transport equation
ut = ux .
We choose an initial profile u(x, 0) as in Fig. 1.6. Where u(x, 0) is increasing as a function
of x, ux > 0 and therefore ut = ux > 0, thus u is increasing in time. Similarly, if u(x, 0)
decreases as a function of x, ux < 0 and it follows that u decreases in time, as indicated by
the arrows in Fig. 1.6. It follows that the overall shape of the profile will move to the left.
1.3. PARTIAL DIFFERENTIAL EQUATIONS 9

ux < 0

ux > 0

Figure 1.6: An initial profile for the transport equation and how it will evolve.
10 CHAPTER 1. INTRODUCTION
Chapter 2

Traveling and standing waves

2.1 Traveling wave solutions of PDEs

We have already seen that functions of the form

u(x, t) = f (x − vt)

represent profiles f (x) that move to the right with velocity v is v > 0, and to the left if
v < 0.
In general, solutions of PDEs are functions of both x and t independently. Sometimes
they have solutions where x and t always show up in the special combination x − vt. We call
such solutions traveling waves. We are especially interested in the case when f (x) is not
constant, or in the case where f (x) is bounded for all values of x. Unbounded signals (i.e.,
u → ±∞) are usually1 not relevant for applications. Then f (x − vt) represents a disturbance
moving through a medium with velocity v.
Example. Consider
u(x, t) = sin(3x − t).
This is a traveling wave moving to the right with velocity v = 1/3. Indeed,

u(x, t) = sin(3x − t)
= sin 3(x − t/3)
= f (x − vt),

so that v = 1/3 and

f (z) = sin 3z.

Example. Let’s find traveling wave solutions of the wave equation

utt = a2 uxx ,
1
In the “always” sense.

11
12 CHAPTER 2. TRAVELING AND STANDING WAVES

where we may assume that the constant a > 0. We let

u = f (x − vt).

Our task is to find the function f and the constant v. Let z = x − vt. We get

u(x, t) = f (z)
⇒ ux = f 0 ,
⇒ uxx = f 00 ,
⇒ ut = −vf 0 ,
⇒ utt = (−v)2 f 00 = v 2 f 00 ,

where we have used the chain rule and the fact that
∂z ∂z
= 1, = −v.
∂x ∂t
Thus, traveling wave solutions of the PDE

utt = a2 uxx

satisfy the ODE

v 2 f 00 = a2 f 00
⇒ (v 2 − a2 )f 00 = 0.

Either f 00 = 0 or v 2 = a2 . Using the first possibility, we get

f 00 = 0 ⇒ f (z) = Az + B,

independent of what v is. Here A and B are constants. The second possibility results in

v 2 = a2 ⇒ v = ±a,

independent of what f (z) is. In summary, we have obtained the following solutions:

ˆ A(x − vt) + B, for any value of A, B, v. These solutions are not very interesting:
in order for them to be bounded, we need A to be zero. But that leaves us with a
constant solution, which is unexciting.

ˆ f1 (x − at), for any function f1 (z), and

ˆ f2 (x + at), for any function f2 (z).

2.2. THE SINE-GORDON EQUATION 13

Since the equation is linear, we can superimpose the solutions to get a more general
solution:
u(x, t) = A(x − vt) + B + f1 (x − at) + f2 (x + at).
Some remarks are in order.

ˆ Usually, we would include multiplicative constants c1 , c2 and c3 to get a solution of the

form
u(x, t) = c1 (A(x − vt) + B) + c2 f1 (x − at) + c3 f2 (x + at).
Since these constants can be absorbed in the forms of f1 , f2 , and the values of A and
B, we may omit them.

ˆ Note that the superposition of different traveling waves is not necessarily a traveling
wave! Our superposition consists of three different parts. Two of these parts (f1 (x−at)
and f2 (x + at)) even move in opposite directions.

2.2 The sine-Gordon equation

We consider a more complicated example. The sine-Gordon equation is

utt = uxx − sin u.

First we substitute u = f (z), z = x − vt into the equation. We want to find an ODE for
f (z), which we want to use to determine f (z) and perhaps also v. In this case, we get

v 2 f 00 = f 00 − sin f
⇒ (1 − c2 )f 00 = sin f.

This is a messy second-order ODE. We can reduce it to a first order ODE by multiplying by
f 0 , which results in an equation we may integrate once:

(1 − c2 )f 0 f 00 = f 0 sin f
02
 
d 2 f d
⇒ (1 − c ) = [− cos f ]
dz 2 dz
⇒ (1 − c2 )f 02 = A − 2 cos f,

where we have used that f 0 f 00 is the derivative of f 02 /2 and that f 0 sin f is the derivative of
− cos f . Here A is an arbitrary constant. Our new equation is a first-order ODE for f (z).
This is progress! It’s still a messy ODE, but we can actually solve this ODE using separation
of variables. Here I just give one class of solutions. You should check that these are, in fact,
solutions2 .  √ 
z/ 1−c2
f (z) = 4 arctan e ,
2
You should never, ever, trust me.
14 CHAPTER 2. TRAVELING AND STANDING WAVES

u(x, t)

Figure 2.1: The profile of a traveling wave solution of the sine-Gordon equation. For this
specific profile, c = 1/2 and the whole graph moves to the right with velocity 1/2.

which corresponds to A = 2, and is valid for c ∈ (0, 1).

It follows that  √ 
2
u(x, t) = 4 arctan e(x−ct)/ 1−c
is a traveling wave solution of utt = uxx − sin u. Since this equation is nonlinear, we can’t
simply superimpose a bunch of these solutions to get new solutions. A plot of a traveling
wave solution is shown in Fig. 2.1. Note the horizontal asymptotes at 0 and 2π.

2.3 The Korteweg-de Vries equation

In 1834, John Scott Russell, a Scottish engineer3 was on horseback, following a ship as it was
pulled by horses along one of the canals in Scotland. At some point, the ship hit something
in the water. What happened next is best told in his own words, and preferably with a
Scottish accent. More information can be found at http://www.macs.hw.ac.uk/~chris/
scott_russell.html.

“I was observing the motion of a boat which was rapidly drawn along a narrow channel
by a pair of horses, when the boat suddenly stopped - not so the mass of water in the
channel which it had put in motion; it accumulated round the prow of the vessel in a
state of violent agitation, then suddenly leaving it behind, rolled forward with great
velocity, assuming the form of a large solitary elevation, a rounded, smooth and well-
defined heap of water, which continued its course along the channel apparently without
change of form or diminution of speed. I followed it on horseback, and overtook it still
rolling on at a rate of some eight or nine miles an hour, preserving its original figure
some thirty feet long and a foot to a foot and a half in height. Its height gradually
3
I am not making this up!
2.3. THE KORTEWEG-DE VRIES EQUATION 15

diminished, and after a chase of one or two miles I lost it in the windings of the channel.
Such, in the month of August 1834, was my first chance interview with that singular
and beautiful phenomenon which I have called the Wave of Translation”
It took 10 years for J. S. Russell to publish his results. It took another 50 for them
to become appreciated (no Twitter yet). Two Dutchmen Korteweg (PhD advisor) and de
Vries (his student) derived the equation that now bears their name, although we usually
abbreviate it the KdV equation:
ut + uux + uxxx = 0.
They derived the equation to describe long waves in shallow water, like tsunamis. But it
describes much more than this: in general it describes the propagation of long waves in a
dispersive medium4 . This covers water waves, but also waves in plasmas, like the northern
lights, or like the light flickering we sometimes see in (large) LED lights.
Let’s look for traveling wave solutions of the KdV equation. Thus we let
u(x, t) = f (z), z = x − vt,
where v is the velocity of the wave. For simplicity we will look for waves traveling to the
right, thus v > 0. Further, we will limit our investigations to waves like the ones that Russell
saw, namely waves that decay to zero as x → ∞ or x → −∞. As before, we have

ux = f 0 , uxxx = f 000 , ut = −vf 0 .

The KdV equation becomes
−vf 0 + f f 0 + f 000 = 0
 
d 1 2 00
⇒ −vf + f + f =0
dz 2
1
⇒ −vf + f 2 + f 00 = A,
2
where A is an integration constant. Since A is constant, we may use any value of x to
evaluate the left-hand side. The most convenient values of x are ±∞, because we are
looking for solutions for which f , f 0 , etc, all decay to zero there. Evaluating our expression
as ±∞ gives A = 0, so that we are left with
1
−vf + f 2 + f 00 = 0.
2
0
Multiplying this by f , we have
1
−vf f 0 + f 2 f 0 + f 0 f 00 = 0
 2 
d v 2 1 3 1 02
⇒ − f + f + f =0
dz 2 6 2
v 1 1
⇒ − f 2 + f 3 + f 02 = B,
2 6 2
4
We’ll learn what “dispersive” means soon.
16 CHAPTER 2. TRAVELING AND STANDING WAVES

where B is a second constant of integration. evaluating the left-hand side once again at ±∞,
we find that B = 0. Thus, we find the first-order ordinary differential equation
v 1 1
− f 2 + f 3 + f 02 = 0.
2 6 2
It follows that
3f 02 = (3v − f )f 2 .
Since v > 0 (by choice), we find that f ≤ 3v. That’s good: we are looking for a bounded
solution, and it looks like we’ll find one. It remains to solve the first-order equation using
separation of variables. Ready to have a good time with integration? We have
√ 0
3f
√ =1
f 3v − f
√ Z df
⇒ 3 √ = z + α,
f 3v − f
where α is a constant of integration. We use substitution to simplify the integral. The
hardest part is the square root in the denominator. Let
p
g = 3v − f ⇒ f = 3v − g 2 .

It follows that df = −2gdg. We get

√ Z −2gdg
3 =z+α
(3v − g 2 )g
√ Z dg
⇒ −2 3 = z + α.
3c − g 2
This integral can be done using partial fractions. This results in
√ !
3v + g √
ln √ = − v(z + α)
3v − g
√
√ e− v(z+α) − 1
⇒ g = 3v −√v(z+α)
e √+1
√ v
= − 3v tanh (z + α).
2
Returning to f = 3v − g 2 , we get
√
2 v
f (z) = 3vsech (z + α).
2
It follows that √
2v
u(x, t) = 3v sech (x − vt + α)
2
2.4. WAVE FRONTS AND PULSES 17

v
3v

Figure 2.2: The profile of a traveling wave solution of the KdV equation. The profile moves
to the right with velocity v and has height 3v.

is a traveling wave solution of the KdV equation, for any v > 0. It is illustrated in Fig. 2.2. Its
amplitude (i.e., height) is 3v, while its velocity is v. Thus higher waves of the KdV equation
travel
√ faster, as we observe at the beach. Further, the width of the profile is proportional to
1/ v, thus taller, faster waves are more narrow. Lastly, the maximum of the wave profile
occurs when x − vt + α = 0, or at x = vt − α.

2.4 Wave fronts and pulses

A traveling wave
u(x, t) = f (x − ct)
is called a front if for a fixed t, we have

lim u = k1 , lim u = k2 ,
x→−∞ x→∞

and
k1 6= k2 .
In other words, our profile approaches different limit values at −∞ and +∞. The traveling
wave solution of the sine-Gordon equation is a good example. Note, however, that the
transition between the two limiting values does not have to be monotone.
If on the other hand,
k1 = k2 ,
then we call the solution a pulse. The profile drawn in Fig. 1.6 is an example. Thus, for
pulses, the beginning and end state are the same: the medium returns back to its original
state after a pulse passes through. On the other hand, the passing of a front forever alters
the state of the medium.
Examples of fronts.
18 CHAPTER 2. TRAVELING AND STANDING WAVES

ˆ Weather fronts: different meteorological signals are altered by the passing of a weather
front, such as the air pressure.
ˆ Sonic booms: the density in the air is altered by the passing of the sonic boom. Of
course, dissipation effects eventually return it to its original value, but for quite a while,
the air density is changed. That is different from a subsonic plane passing through,
which creates a localized disturbance in the density.
ˆ Bob Dylan, Beethoven: music was forever altered by both of them.

Examples of pulses.

ˆ Optical flashes

ˆ Bits

ˆ One-hit wonders: they have no lasting effect on the musical landscape.

2.5 Wave trains and dispersion

Different types of traveling waves, other than fronts or pulses, exist.
Example. Consider
u(x, t) = 3 cos(7x − 5t) = 3 cos 7(x − 5t/7).
This is a traveling wave train of amplitude 3 and velocity 5/7. It has wave number 7 and
frequency 5.
In general, (linear) traveling wave trains are expressions of the form
u = A cos(kx − ωt + φ),
where instead of cos we may have sin or exp(i . . .). Using one of these functional forms, the
following quantities are defined.

ˆ A is the amplitude. This is the largest value the wave train attains. Similarly, −A is
its smallest value.
ˆ The wave number k denotes how many oscillations occur during an interval of length
2π. Indeed, the period of the wave train is 2π/k. Since one oscillation occurs per
period, k oscillations occur over an interval of length 2π.
ˆ The frequency ω denotes how many oscillations happen during a time interval of
length 2π. The same argument as above shows this, but we consider the temporal
dependence instead of the spatial dependence. It follows that signals with high wave
number or frequency are very oscillatory. That means that if we want to plot them
accurately, we’ll need to use many points at which to sample the function.
2.5. WAVE TRAINS AND DISPERSION 19

ˆ The phase shift φ simply shift the origin of space or time. That sounds very philo-
sophical5 , but it really isn’t. We can rewrite the wave train signal as (using the cosine
form, for instance)
u(x, t) = A cos(k(x − x0 ) − ωt) or u(x, t) = A cos(kx − ω(t − t0 )),
where −kx0 = φ and ωt0 = φ. Thus φ simply changes when we start measuring where
or when we start.
ˆ The velocity of the traveling wave is v = ω/k, since we can rewrite the signal as
u(x, t) = A cos(k(x − vt) + φ),
where v = ω/k.
Often, the frequency ω and the wave number k are related. This relation is called the dis-
persion relation. Since sin and cos can be written as linear combinations of exponentials,
it suffices to consider expressions of the form
u = Aeikx−iωt .
The idea is to substitute this into the PDE and find the relationship between ω and k, if
there is one. This is easy: note that whenever we take an x derivative, we get
ux = ikAeikx−iωt = iku,
and so on for higher-order derivatives. Thus taking an x derivative simply multiplies our
solution by a factor ik. Similarly, taking a time derivative,
ut = −iωAeikx−iωt = −iωu,
and taking a time derivative results in multiplying u by −iω. Taking more time derivatives
results in more powers of −iω.
Example. Consider the wave equation
utt = a2 uxx , a > 0.
We already know that this equation has traveling wave solutions that travel to the left with
velocity −a and to the right with velocity a. Plugging in
u = Aeikx−iωt ,
we get
(−iω)2 u = a2 (ik)2 u
⇒ −ω 2 = −k 2
⇒ ω 2 = k 2 a2 ,
5
That’s OK: I have a doctorate in philosophy. Really.
20 CHAPTER 2. TRAVELING AND STANDING WAVES

from which it follows that

ω1 = ka, ω2 = −ka.

Thus there are two possible branches of the dispersion relation. They give rise to the solutions

u1 = A1 eik(x−at) ,

and
u2 = A2 eik(x+at) .

Example. Consider the PDE

utt = auxx − bu, a, b > 0.

Looking for the dispersion relation, we find

(−iω)2 u = a(ik)2 u − bu
⇒ −ω 2 = −ak 2 − b
⇒ ω 2 = ak 2 + b
√
⇒ ω1,2 = ± ak 2 + b,

leading to the two classes of solutions

u1 = A1 eikx−iω1 t ,

and
u2 = A2 eikx−iω2 t .

Example. Consider the linear free6 Schrödinger equation:

iϕt = −ϕxx .

We look for solutions of the form

ϕ = Aeikx−iωt ,

to get

i(−iω)ϕ = −(ik)2 ϕ
⇒ ω = k2.
6
because there is no potential
2.6. DISPERSION RELATIONS FOR SYSTEMS OF PDES 21

2.6 Dispersion relations for systems of PDEs

We will examine how to do this by example. Consider the PDE system

ut = αux + vxxx
vt = βvx − uxxx .

We wish to look for wave train solutions, whose x and t dependence is of the form

eikx−iωt .

For the scalar case, the amplitude A never came into play, since for linear equations, we
could simply divide it out. That is no longer true here, since there is no reason why u and
v should have the same amplitude.
This situation is similar to what we do with systems of ODEs: if we wish to solve

ay 00 + by 0 + cy = 0,

we guess
y = eλt .
On the other hand, if we want to solve the system

y 0 = Ay,

where A is a matrix and y is a vector, then we guess

y = eλt v,

where v is a vector.
We do the same for our system of PDEs. We guess
   
u U
= eikx−iωt .
v V

Substituting this into the PDE, we get

−iωU = αikU + (ik)3 V

−iωV = βikV − (ik)3 U
  
−iω − αik ik 3 U
⇒ = 0.
ik 3 −iω − βik V

Since we are not interested in the zero solution

 
U
= 0,
V
22 CHAPTER 2. TRAVELING AND STANDING WAVES

we need that  
−iω − αik ik 3
3
ik −iω − βik
is a singular matrix. Thus
 
−iω − αik ik 3
det 3 = 0.
ik −iω − βik

This the dispersion relation, determining the frequency ω as a function of the wave number
k. The rest is algebra. We get

(−iω − αik)(−iω − βik) − k 6 =0

⇒ −ω 2 + iω(αik + βik) − αβk 2 − k 6 =0
⇒ −ω 2 − ω(α + β)k − αβk 2 − k 6 =0
⇒ ω 2 + ω(α + β)k + αβk 2 + k 6 =0
1 p 
⇒ ω= −(α + β)k ± (α + β)2 k 2 − 4(αβk 2 + k 6 ) .
2
Thus, there are two possibilities for ω: ω1 and ω2 , corresponding to the + and − signs above.

2.7 Information from the dispersion relation

Suppose we have found and solved the dispersion relation, so that we have ω(k). What do
we get out of this? Well, of course we can conclude that the PDE has wave train solutions
whose x and t dependence if of the form

u = eikx−iωt .

Since the equation is linear, we may superimpose such solutions. But what can we say about
these solutions? What are their properties?

ˆ Phase velocity. We may rewrite the solution as

u = eik(x−cp (k)t) ,

where
ω(k)
cp (k) =
k
is called the phase velocity. It is the velocity with which a single wave train travels.

ˆ Group velocity. Another velocity matters for wave trains:

dω
cg (k) = ,
dk
2.7. INFORMATION FROM THE DISPERSION RELATION 23

cg

cp

Figure 2.3: A wave packet consisting of two traveling wave trains. The phase speed is the
speed of waves inside the packets, whereas the envelope of the packets moves with the group
speed.

which we call the group velocity. The full importance of the group velocity is hard to
explain (see Amath569), but let’s give it a shot.
Suppose that in a given signal, the most important wave number is k0 , with corre-
sponding frequency ω0 = ω(k0 ). We rewrite exp(ikx − iω(k)t) as
eikx−ω(k)t = eik0 −iω0 t ei(k−k0 )x−i(ω(k)−ω0 )t
0 2 )−ω )t
= eik0 −iω0 t ei(k−k0 )x−i(ω(k0 )−(k−k0 )ω (k0 )+O((δk) 0

0 2 ))t
= eik0 −iω0 t ei(k−k0 )x−i((k−k0 )ω (k0 )+O((∆k)
0
= eik0 −iω0 t ei∆k(x−ω (k0 )t)+... .
The second exponential factor acts as a slowly varying amplitude to the first one.
Indeed, the wave number ∆k and frequency ω 0 (k0 )∆k are both small, assuming that
k is close to k0 . The second, slowly-varying factor moves with the group velocity! A
wave packet is used to illustrate this in Fig. 2.3.
Thus individual waves move with the phase velocity, while wave packets move with the
group velocity. Note that it is perfectly possible for the group and phase velocity to
have opposite signs.
Although we will not show this here, the group velocity is also the velocity with which
the energy associated with a wave moves. Often we care about this much more than
we care about individual waves.
ˆ Dispersive vs. non-dispersive waves. An equation or a system is called dispersive
if (a) the phase speed cp is real for real k, and (b) the phase speed is not constant. This
implies that wave trains with different wave numbers will move with different speeds.
24 CHAPTER 2. TRAVELING AND STANDING WAVES

Figure 2.4: A decaying wave train (ωI < 0, left) and a growing wave train (ωI > 0, right).

But there is more! In general, the dispersion relation may be complex, even for real k.
If that happens, then for those values of k the equation is not dispersive.
Example. Consider the PDE
ut = uxx + cux .

This equation gives rise to the dispersion relation

−iω = (ik 2 ) + cik

⇒ ω = −ik 2 − ck.

In general, we can split ω(k) into its real and imaginary parts: ω = ωR + iωI , where ωR
and ωI are the real and imaginary parts of ω, respectively. It follows that our wave train
solutions are of the form

eikx−iωt = eikx−it(ωR +iωI )

= eikx−iωR t+ωI t
= eωI t eikx−iωR t .

Thus, if ωI > 0, the signal will grow in time. If ωI < 0, the signal will decay in time. Both
situations are illustrated in Fig. 2.4.
If ωI < 0 and amplitudes decay, we say the PDE is dissipative. if ωI > 0 and amplitudes
grow, we call the system unstable.
2.8. PATTERN FORMATION 25

2.8 Pattern formation

Often, linear systems are obtained by ignoring nonlinear term, which are assumed to be less
important, provided the solution is small. Such small solutions can be considered distur-
bances to the zero solution. When we see instabilities in a linear PDE or a system of linear
PDES, they result in the growth of disturbances to the zero solution. In actual application
settings, this growth cannot go on forever, as there is only a finite amount of energy in the
system, for instance.
Recall that the linear system is only valid provided the solution is small. If the distur-
bances grow exponentially, the linear system will not be valid for long, and nonlinear terms
will have to be considered. Often they have the effect of arresting the growth due to linear
instabilities. The nonlinear effects that we ignored to get to the linear system may be small
initially, but they will start to matter as the solution grows.
Even so, the study of the dispersion relation often allows us to predict many aspects of the
solution of the full, nonlinear problem, even if we cannot solve that problem completely. Let’s
see how we can predict what kinds of patterns form in the so-called Kuramoto-Sivasinski
(KS) equation

ut + uux + uxx + auxxxx = 0, a > 0.

This equation arises a lot in applications. Among other things, it describes the dynamics of
flame fronts.

ˆ We see immediately that u = 0 is a solution.

ˆ Let’s investigate the dynamics of solutions close to this solution. Such solutions are
small, so that we may ignore the term uux . Indeed, if

u ∼ Aeikx−iωt ,

with A small, then

ux ∼ ikAeikx−iωt ,
and
uux ∼ ikA2 e2ikx−2iωt ,
which is a lot smaller than any of the linear terms, since A2  A, for small A. Thus,
we are justified in studying the linearized KS equation:

ut + uxx + auxxxx = 0.

ˆ We find its dispersion relation:

−iω + (ik)2 + a(ik)4 = 0

⇒ −iω − k 2 + ak 4 = 0
⇒ −iω = k 2 − ak 4 ,
26 CHAPTER 2. TRAVELING AND STANDING WAVES

k 2 − ak 4

Figure 2.5: The graph of k 2 − ak 4 with a = 4.

so that
2 −ak 4 )t
eikx−iωt = eikx+(k .
It follows that the solution grows in t if k 2 − ak 4 > 0 and it decays if k 2 − ak 4 < 0. We
plot k 2 − ak 4 in Fig. 2.5.
√ √
If k ∈ (−1/ a, 1/ a), the growth rate is positive. Otherwise it is negative. Since the
equation is linear, the general solution is a superposition of a bunch of solutions of the
form
2 4
eikx+(k −ak )t .
We have √ just concluded that any part of this superposition that has k > 1/sqrta or
k < −1/ a will decay. Thus after some time, these parts will not come into play
anymore. Since for all of these parts, |k| is large, the period is small. In other words,
these are highly oscillatory signals. The equation appears to get rid of them quickly.
All the contributions from the other wave numbers k grow (except from k = 0). Which
one grows the most? Let

f (k) = k 2 − ak 4 ⇒ f 0 (k) = 2k − 4ak 3 = 2k(1 − 2ak 2 ).

Thus the growth rate is maximal for k ∗ = ±1/sqrt2a. All other solutions grow slower
than this one. Thus, in comparison, they decay. Indeed, if we have

y = c1 eax + c2 ebx ,
2.9. A DERIVATION OF THE WAVE EQUATION 27

u(x, t)

Figure 2.6: The set-up for the derivation of the wave equation modeling a plucked string.

with a > b > 0 (so both exponentials grow as x → ∞), then

y = eax c1 + c2 e(b−a)x ,

and the exponential in the parentheses decays! Thus, so sufficiently large x, we see
y ∼ c1 eax .
The same conclusion holds for our case: for large enough t, we see only
∗ x+(k ∗2 )−ak ∗4 t ∗ x+(k ∗2 )−ak ∗4 t
eik and eik .
√
Thus, the equation naturally creates periodic patterns with period 2π/k ∗ = 2π 2a,
independent of the initial conditions!

ˆ After some t, these solutions become too big and we can no longer ignore the nonlinear
terms. But, the stage is set and the linear problem has already selected the period of
the solution!

2.9 A derivation of the wave equation

The wave equation
utt = c2 uxx ,
shows up in many applications. Let’s actually derive it in one setting, namely that of a
plucked string, as illustrated in Fig. 2.6
We begin with the following assumptions.

ˆ The equilibrium position of the string is at u(x, t) = 0.

ˆ The string has constant density ρ.

ˆ The vibration of the string stays in the plane: there is no dependence on the transverse
variable y.
28 CHAPTER 2. TRAVELING AND STANDING WAVES

T x

x x + ∆x

Figure 2.7: We apply Newton’s law to a little piece of string.

ˆ Tension is uniform: a string extends a force only in the direction parallel to the string.
In other words, the force a piece of string exerts on neighboring pieces , keeping the
string together, is tangential to the string.
ˆ We assume that tension is constant anywhere along the string.
ˆ There are no other forces.
ˆ All vibrations are small. This is a physical way of saying that, mathematically, we will
be ignoring nonlinear effects.
Next, we apply Newton’s law of motion to an itty bitty7 piece of string, lying between x
and x + ∆x, where ∆x is very small. We have
Mass of S × Acceleration of S = Net force on S.
This is Newton’s law, perpendicular to the x axis. We could also write it parallel to the x
axis, but that would result in a perfect force balance. This would offer no information about
u(x, t), which is the vertical displacement.
First, we get an expression for the mass of S.
Mass of S = ρ × arclength
Z x+∆x p
=ρ 1 + u2x (s, t)ds
x
Z x+∆x  
1 2
≈ρ 1 + ux (s, t) + . . . ds
x 2
≈ ρ∆x,
7
Technical term.
2.9. A DERIVATION OF THE WAVE EQUATION 29

where we have used that the vibrations are small, thus all nonlinear terms are ignored.
Next, the acceleration of S is simply utt (x, t), by definition. Last, we turn to the net
force. The net force is pulling on the left and right ends of S by the string parts to the
immediate left and right of S. On the left end, the tension pulls with magnitude T in the
direction of the tangent vector. This normalized tangent vector is given by
−(1, ux (x, t))
p .
1 + u2x (x, t)
Thus the left force is
 
−T (1, ux (x, t)) 1 2
p ≈ −T (1, ux ) 1 − ux + . . .
1 + u2x (x, t) 2
≈ −T (1, ux ).
It follows that the left force in the vertical direction is −T ux (x, t). We repeat these consid-
erations on the right end. The tangent vector is
(1, u (x + ∆x, t))
p x ,
1 + u2x (x + ∆x, t)
so that the force is
 
T (1, ux (x + ∆x, t)) 1 2
p ≈ −T (1, ux (x + ∆x)) 1 − ux (x + ∆x, t) + . . .
1 + u2x (x + ∆x, t) 2
≈ −T (1, ux (x + ∆x, t)).
Thus, after ignoring the nonlinear terms, the right force in the vertical direction is given by
T ux (x + ∆x, t).
Combining all of this, Newton’s law becomes
ρ∆xutt (x, t) = T (ux (x + ∆x, t) − ux (x, t))
ux (x + ∆x, t) − ux (x, t)
⇒ ρutt (x, t) = T .
∆x
Taking the limit ∆x → 0, we obtain
ρutt = T uxx ,
or
utt = c2 uxx ,
where
T
c2 = ,
ρ
and we have derived the wave equation! We have already seen that f (x − ct) and f (x + ct)
are traveling wave solutions of the wave equation. It follows that their velocity is given by
s
T
±c = ± .
ρ
30 CHAPTER 2. TRAVELING AND STANDING WAVES

2.10 d’ Alembert’s solution of the wave equation

We want to find the general solution of the wave equation

utt = c2 uxx .

We already know that this equation has traveling wave solutions f (x − ct) and g(x + ct),
for arbitrary profiles f and g. We show that any solution of the wave equation is a linear
combination of such traveling waves.
Inspired by the form of the traveling waves, we use a coordinate transformation

ξ = x − ct,
η = x + ct.

Let us find out how the derivatives transform. First, we consider ux .

∂u ∂u ∂ξ ∂u ∂η
= +
∂x ∂ξ ∂x ∂η ∂x
= uξ · 1 + uη · 1
= uξ + uη .

Next up, the second derivative uxx :

∂ 2u ∂
2
= (uξ + uη )
∂x ∂x
∂ ∂ξ ∂ ∂η
= (uξ + uη ) + (uξ + uη )
∂ξ ∂x ∂η ∂x
∂ ∂
= (uξ + uη ) · 1 + (uξ + uη ) · 1
∂ξ ∂η
∂ ∂
= (uξ + uη ) + (uξ + uη )
∂ξ ∂η
= (uξξ + uηξ ) + (uξη + uηη )
= uξξ + 2uξη + uηη ,

where we have assumed that u is smooth, so that uξη = uηξ . Next, we do the same for the t
derivatives.

∂u ∂u ∂ξ ∂u ∂η
= +
∂t ∂ξ ∂t ∂η ∂t
= uξ · (−c) + uη · c
= −cuξ + cuη .
2.10. D’ ALEMBERT’S SOLUTION OF THE WAVE EQUATION 31

Next up, the second derivative utt :

∂ 2u ∂
= (−cuξ + cuη )
∂x2 ∂x
∂ ∂ξ ∂ ∂η
= (−cuξ + cuη ) + (−cuξ + cuη )
∂ξ ∂t ∂η ∂t
∂ ∂
= (−cuξ + cuη ) · (−c) + (−cuξ + cuη ) · c
∂ξ ∂η
∂ ∂
= −c (−cuξ + cuη ) + c (−cuξ + cuη )
∂ξ ∂η
= −c(−cuξξ + cuηξ ) + c(−cuξη + cuηη )
= c2 uξξ − 2c2 uξη + c2 uηη .
Substituting the expressions for uxx and utt into the wave equation we get
utt = c2 uxx
⇒ c2 uξξ − 2c2 uξη + c2 uηη = c2 (uξξ + 2uξη + uηη )
⇒ uξη = 0.
This is fantastic progress! Now we proceed to solve this equation. From uξη = 0, it
follows that (uξ )η = 0, thus uξ is independent of η: it is a function of ξ only. Thus
uξ = F (ξ),
where F is any function. We integrate once more to find
Z ξ
u= F (ξ)dξ + g(η),

where g is any function of η. The first term is the anti-derivative of any function of ξ, so it’s
another arbitrary function of ξ. Let’s denote it by f . Thus we have that any solution of the
wave equation can be written in the form
u = f (ξ) + g(η),
or, returning to the original variables:
u = f (x − ct) + g(x + ct),
which proves our claim.
Example. You can easily show that u = cos t sin x solves the wave equation with c = 1.
Strange. . . . It doesn’t appear to be of the form given above. But it is, as we show now. Fun
with trig identities! From the trig addition formulas, we have
cos t sin x + sin t cos x = sin(x + t)
cos t sin x − sin t cos x = sin(x − t),
32 CHAPTER 2. TRAVELING AND STANDING WAVES

from which it follows that

1
u(x, t) = cos t sin x = (sin(x + t) + sin(x − t)),
2
which shows that u can indeed be written as the sum of a function which depends on x + t
only, and another one which depends on x − t only.
Next, we will solve our first initial-value problem in this course. Consider the problem

utt = c2 uxx
u(x, 0) = f (x)
ut (x, 0) = g(x),

where f and g are given functions. Here f (x) represents the initial position of the string and
g(x) represents its initial velocity.
We know that u(x, t) can be written as

u(x, t) = F (x − ct) + G(x + ct).

Our task, should we choose to accept it8 , is to find F and G in terms of f and g. At t = 0,
we have u(x, t) = f (x), thus
f (x) = F (x) + G(x).
Taking a time derivative of our solution formula, we have

ut (x, t) = −cF 0 (x − ct) + cG(x + ct).

Evaluating this at t = 0, we get

g(x) = −cF 0 (x) + cG0 (x).

Thus we need to solve the equations


F (x) + G(x) = f (x)
,
−cF (x) + cG0 (x) = g(x)
0

for the unknown functions F and G.

From the second equation,

1
−F 0 (x) + G0 (x) = g(x)
cZ
1 x
⇒ −F (x) + G(x) = g(s)ds + α,
c 0
8
We will. We’re awesome. We’re fearless.
2.11. CHARACTERISTICS FOR THE WAVE EQUATION 33

where α is an integration constant. Our system becomes


F (x) + G(x) = f (x)
Rx ,
−F (x) + G(x) = 1c 0 g(s)ds + α.

Adding these two equations, we find that

Z x
1 1 α
G(x) = f (x) + g(s)ds + .
2 2c 0 2

Subtracting the two equations gives F :

Z x
1 1 α
F (x) = f (x) − g(s)ds − .
2 2c 0 2

Putting all of this together, we find that

1 x−ct 1 x+ct
Z Z
1 α 1 α
u(x, t) = f (x − ct) − g(s)ds − + f (x + ct) + g(s)ds +
2 2c 0 2 2 2c 2
Z 0 Z x+ct 0
1 1 1 1
= f (x − ct) + g(s)ds + f (x + ct) + g(s)ds
2 2c x−ct 2 2c 0
1 x+ct
Z
1
= (f (x − ct) + f (x + ct)) + g(s)ds.
2 2c x−ct

This is the d’ Alembert solution for the initial-value problem for the wave equation.
Don’t take this for granted. The wave equation is one of very few PDEs for which we
can write down the solution of the initial-value problem so explicitly!
Example. Consider the initial-value problem

utt = c2 uxx
2
u(x, 0) = e−x
ut (x, 0) = 0,

then d’Alembert gives

1  −(x−ct)2 −(x+ct)2

u(x, t) = e +e .
2
Thus the solution splits the initial condition in two parts: one goes to the left, one goes to
the right. Both propagate with speed c.

2.11 Characteristics for the wave equation

Consider the transport equation
ut + cux = 0.
34 CHAPTER 2. TRAVELING AND STANDING WAVES

x0

Figure 2.8: The initial condition is transported along the characteristics.

Using the transformation

ξ = x − ct,
τ = t,

we see that the general solution of this problem is

u = f (x − ct),

where f (x) = u(x, 0), the initial profile. Thus, the transport equation simply moves the
initial profile to the right with speed c, where we have assumed that c > 0.
This is illustrated in Fig. 2.8. The straight lines x − ct = x0 , originating at (x0 , 0) are
called the characteristics. It follows that the value of u(x, t) along these characteristics is
given by
u(x, t) = f (x − ct) = f (x0 ).
Thus, u is constant along the characteristics: if we know u anywhere along a characteristic,
when we know it anywhere on that characteristic.
We aim to get a similar understanding for the dynamics of the wave equation

utt = c2 uxx .

Recall that the d’Alembert solution gives

Z x+ct
1 1
u(x, t) = (u(x − ct, 0) + u(x + ct, 0)) + ut (s, 0)ds.
2 2c x−ct
2.11. CHARACTERISTICS FOR THE WAVE EQUATION 35

t0

x0 − ct0 x0 x0 + ct0

Figure 2.9: The domain of dependence (grey) for the solution of the wave equation.

It is clear from this formula that the value of u at (x = x0 , t = t0 ) is determined only by the
initial values for x ∈ [x0 − ct0 , x0 + ct0 ]. This is illustrated in Fig. 2.9.
The cone given by the grey region in Fig. 2.9 is called the domain of dependence. It
indicates that the solution at place x0 and time t0 depends on all values in that cone, but
on no values outside of it.
If we consider the special situation where
ut (0, t) = 0,
in other words, the string starts with an initial profile, but is released without any initial
velocity, then d’Alembert giveth
1
u(x, t) = (u(x − ct, 0) + u(x + ct, 0)),
2
which shows that u depends only on the initial condition on the boundary of the domain
of dependence. These two boundary lines are called the characteristics for the wave
equation. Thus, the characteristics determine the domains of dependence.
By turning all of the above around, we can answer the question of which (x, t) points
have a given point (x0 , t0 ) in their domain of dependence. This is known as the domain of
influence. It is illustrated in Fig. 2.10. The domain of influence is also bordered by the
characteristics, but now on the lower side.
The domain of influence shows that information in the wave equation travels at a finite
speed c: it takes a definite time for the effect from x0 to be felt at any other x.
Example. Sometimes the characteristics may be used to examine the solution of an
36 CHAPTER 2. TRAVELING AND STANDING WAVES

(x0 , t0 )

Figure 2.10: The domain of influence (grey) of the point (x0 , t0 ) for the wave equation.

initial-value problem. Consider the problem

utt = 4uxx

1 if x ∈ [0, 1]
u(x, 0) =
0 if x ∈
6 [0, 1]
ut (x, 0) = 0.

Here the speed of propagation is 2:

c2 = 4 ⇒ c = 2.

We have the set-up illustrated in Fig. 2.11.

From this characteristic plot we can immediately read off the values of u(x, t), for any
x and t. For instance, we can plot the solution u(x, t) at different instances of t, by taking
horizontal slices of the characteristic plot above. This is illustrated in Fig. 2.12.

2.12 The wave equation on the semi-infinite domain

Fixed end
Let’s consider a string of semi-infinite extent, fixed at the left end, as drawn in Fig. 2.13.
The condition u(0, t) = 0 for all t is our first example of a boundary condition: a
2.12. THE WAVE EQUATION ON THE SEMI-INFINITE DOMAIN 37

1/2
1/2 1/2

0 (1/2, 1/4) 0 x
1
0 0 1 1 0

Figure 2.11: The different solution regions for the example problem. The values in blue are
values of u(x, t) in that region. The characteristics drawn have slope 1/2.

condition on the function we are looking for, given at the end of our physical domain.
Thus, we want to solve the boundary-value problem


 utt = c2 uxx , x ∈ (0, ∞)
u(0, t) = 0, for all t > 0

,

 u(x, 0) = f (x), x ∈ (0, ∞)
ut (x, 0) = g(x), x ∈ (0, ∞)


where f (x) and g(x) are given functions. We will solve this using d’Alembert’s solution and
the insight we gained from using characteristics. From d’Alembert,
Z x+ct
1 1
u(x, y) = (f (x − ct) + f (x + ct)) + g(s)ds.
2 2c x−ct

This formula can only be valid if x − ct > 0. Otherwise, we’d be asking to evaluate f (x) at
arguments that are negative, but we have been given f (x) only for positive arguments. A
conundrum! Thus the solution ceases to be valid when

x − ct < 0 ⇒ t > x/c.

This region is illustrated in Fig. 2.14.

This is not surprising: the shaded region in Fig. 2.14 would be a part of the domain
of influence of negative x values, but we have no information there. So, what happens?
Boundary condition to the rescue: we have not used the condition u(0, t) = 0 for all t. We
will do so now.
In the shaded region, we still have that

u(x, t) = F1 (x − ct) + G(x + ct),

38 CHAPTER 2. TRAVELING AND STANDING WAVES

t=0
x
0 0
0 1
1

0 < t < 1/4 1/2 1/2 x

0 0
1/2

t = 1/4 1/2 x
0 0
1/2

1/2 1/2 x
t > 1/4 0 0
0
1/2

Figure 2.12: The different stages of the solution. As before, values of u are in blue.

since all solutions of the wave equation are of this form. Using the boundary condition, we
have

0 = F1 (−ct) + G(ct)
⇒ F1 (z) = −G(−z),

so that
u(x, t) = −G(−x + ct) + G(x + ct).

Using the initial conditions, we find that

Z z
1 1
G(z) = f (z) + g(s)ds,
2 2c 0

as before. Note that to determine G(x + ct), it is fine to use the initial conditions, since the
information in G(x + ct) comes only from positive values of x. Alternatively, we can impose
the continuity of our solution at x = ct, finding the same result for G(x). Either way, it
2.12. THE WAVE EQUATION ON THE SEMI-INFINITE DOMAIN 39

u(0, t) = 0, for all t.

Figure 2.13: The wave equation string, on the half line x > 0 with fixed end u(0, t) = 0, for
all t.

follows that
1 x+ct 1 −x+ct
Z Z
1 1
u(x, t) = f (x + ct) + g(s)ds − f (−x + ct) − g(s)ds
2 2c 0 2 2c 0
1 ct+x
Z
1
= (f (x + ct) − f (ct − x)) + g(s)ds,
2 2c ct−x
which is valid for t > x/c, while we still have that
Z x+ct
1 1
u(x, t) = (f (x − ct) + f (x + ct)) + g(s)ds,
2 2c x−ct

for t ≤ x/c.
Let’s examine the characteristics to make sense out of this, see Fig. 2.15. At x = 0,
the region of dependence would usually come partially from x < 0, but that is not allowed
now. Rather, all the information comes from x > 0, and from the boundary itself. We can
pretend there is an x < 0 region, as long as we always satisfy u(0, t) = 0, for all t. This is
easily done: let us extend u for x < 0 to be the opposite of u for the corresponding positive
value −x. In other words, the overall u(x, t) is odd, as a function of x: u(−x, t) = −u(x, t):

u(x, t), x≥0
û(x, t) = ,
−u(−x, t) x≤0
then clearly û(0, t) = 0. Also, the function behaves as if it switches sign every time it hits the
boundary, as the derivative will be even9 . Indeed, when the value from point (a) in Fig. 2.15
hits the boundary, it meets the value from (b), which has the opposite value. This opposite
value continues on for x > 0. Thus, in effect, the value flips at the boundary.
9
Recall that the derivative of an even function is odd and the derivative of an odd function is even. Whoa!
40 CHAPTER 2. TRAVELING AND STANDING WAVES

t = x/c

Figure 2.14: Our original d’Alembert solution is not valid in the shaded region, which
corresponds to t > x/c.

Free end
We consider the boundary-value problem


 utt = c2 uxx , x ∈ (0, ∞)
ux (0, t) = 0, for all t > 0

,

 u(x, 0) = f (x), x ∈ (0, ∞)
ut (x, 0) = g(x), x ∈ (0, ∞)


Thus, according to the boundary condition, the string is horizontal at the boundary, for all
time. We proceed as before. By d’Alembert,
Z x+ct
1 1
u(x, t) = (f (x − ct) + f (x + ct)) + g(s)ds.
2 2c x−ct

By the same reasoning as for the fixed-end case, this is only valid when x − ct > 0, and
another solution form has to be found in the shaded region of Fig. 2.14. Again this is
expected: in the shaded region, part of the solution would come from x < 0, but we have no
information there.
As always, we have that in the shaded region10

u(x, t) = F2 (x − ct) + G(x + ct),

10
Because we’re solving the wave equation.
2.12. THE WAVE EQUATION ON THE SEMI-INFINITE DOMAIN 41

u=0

(a) (b) x

x<0 0 x>0

Figure 2.15: The characteristics for the wave equation on the half line, including the
“phantom” characteristics for x < 0.

where G(z) is determined as before (once again: the information in G(z) comes only from
positive x values):
1 z
Z
1
G(z) = f (z) + g(s)ds.
2 2c 0
Thus
1 x+ct
Z
1
u(x, t) = F2 (x − ct) + f (x + ct) + g(s)ds.
2 2c 0
In order to impose the boundary condition, we calculate
1 1
ux (x, t) = F20 (x − ct) + f 0 (x + ct) + g(x + ct).
2 2c
Evaluating this at x = 0, we get
1 1
0 = F20 (−ct) + f 0 (ct) + g(ct)
2 2c
0 1 0 1
⇒ F2 (−ct) = − f (ct) − g(ct).
2 2c
Equating −ct = z, we get
1 1
F20 (z) = − f 0 (−z) − g(−z)
2 2c
1 −z
Z
1
⇒ F2 (z) = f (−z) + g(s)ds,
2 2c 0
by the fundamental theorem of calculus. Thus
Z 0
1 1
F2 (x − ct) = f (ct − x) − g(s)ds,
2 2c ct−x
42 CHAPTER 2. TRAVELING AND STANDING WAVES

so that for t > x/c,

Z ct−x Z ct+x
1 1 1
u(x, t) = (f (x + ct) + f (ct − x)) + g(s)ds + g(s)ds.
2 2c 0 2c 0

If we let g(x) ≡ 0, we can understand this result, using the characteristics, see Fig. 2.15.
In order to satisfy the boundary condition, we extend u(x, 0) to negative values to be an
even function. Then its derivative will be odd, and its value at 0 will be 0. We see that at
(x, t), there are two contributions:

ˆ One from (ct + x, 0) and one from (x − ct, 0).

ˆ The first one is f (x + ct)/2, the second is given by f (ct − x)/2, by our result above.

2.13 Standing wave solutions of the wave equation

Standing waves are different from traveling waves in that they stay where they are, but their
profile may change over time.
Example. The wave profile

u(x, t) = 2 cos t sin x

represents a standing wave. It looks like 2 sin x at all time t, but multiplied by a time-
dependent amplitude cos t. It is illustrated in Fig. 2.16. It is clear11 that standing waves can
often be written as linear combinations of traveling waves. Here we have

u(x, t) = sin(x − t) + sin(x + t),

but this is not always the case.

To look for traveling wave solutions of the wave equation

utt = c2 uxx ,

we let
u(x, t) = T (t)X(x).
Thus X(x) represents the spatial profile, while T (t) gives the temporal part of the profile,
its time-dependent amplitude. We get

T 00 (t)X(x) = c2 T (t)X 00 (x)

T 00 X 00
⇒ = c2 .
T X
11
As clear as some trig identities.
2.13. STANDING WAVE SOLUTIONS OF THE WAVE EQUATION 43

t=0
t = π/4

t = π/2 x

t=π

Figure 2.16: The standing wave profile u = 2 cos t sin x, for different values of t.

Note that the left-hand side depends on t only, while the right-hand side is a function of
only x. It follows that both sides have to be constant. Indeed, if we take an x derivative of
the equation, we get  00 0
2 X
0=c ,
X
since the left-hand side does not depend on x. Thus c2 X 00 /X is constant. The same argument,
but by taking a derivative with respect to t, gives that T 00 /T is constant too. Thus we have
T 00 X 00
= c2 = λ,
T X
where λ is a constant. We get two ordinary differential equations:
 00
T = λT,
X 00 = cλ2 X.
Here the first equation is an ODE in t, while the second one depends on x.
As you know, we get different kinds of solutions, depending on the sign of λ. Let us
investigate all possibilities.
ˆ λ = 0. we get T 00 = 0 and X 00 = 0. It follows that T = A + Bt and X = C + Dx. Thus
u = (C + Dx)(A + Bt).
In particular, we want D = 0 and B = 0, if x ∈ R, since we want bounded solutions.
On the other hand, if x is restricted to a smaller domain, the solution with D 6= 0 6= B
may be perfectly acceptable.
44 CHAPTER 2. TRAVELING AND STANDING WAVES

ˆ λ > 0. We write λ = r2 , where r > 0. We have to solve

 00
T = r2 T,
2
X 00 = rc2 X,

from which it follows that12

T = Aert + Be−rt ,
and
X = Cerx/c + De−rx/c ,
resulting in
u = (Aert + Be−rt )(Cerx/c + Be−rx/c ).

ˆ λ < 0. For the last case, we write λ = −r2 , with r > 0. We have to solve
 00
T = −r2 T,
2
X 00 = − rc2 X,

from which it follows that

T = A cos(rt + B sin(rt),

and
X = C cos(rx/c) + D sin(rx/c),
resulting in

u = (A cos(rt + B sin(rt))(C cos(rx/c) + D sin(rx/c)).

All of these result in standing wave solutions of the wave equation. Next, we impose
some boundary conditions.

Standing waves on a finite string

All the profiles above solve the wave equation. If we impose boundary conditions, many of
them are rules out and only some remain. As an important example, let’s consider the wave
equation on a finite interval with fixed ends, like a guitar string of length L. We have, for
x ∈ (0, L), t > 0,

utt = c2 uxx ,
u(0, t) = 0,
u(L, t) = 0.
12
Or, you could use hyperbolic functions. Always fun!
2.13. STANDING WAVE SOLUTIONS OF THE WAVE EQUATION 45

We have that
u(x, t) = X(x)T (t),
for standing waves. We want u(0, t) = 0, which implies that
X(0)T (t) = 0.
Since we don’t want that T (t) = 013 , we need that X(0) = 0. Similarly, we need that
X(L) = 0. We already know the allowed forms for X(x):
X(x) = C + Dx,
or X(x) = Cerx/c + De−rx/c ,
or X(x) = C cos(rx/c) + D sin(rx/c).
We now impose the conditions
X(0) = 0, X(L = 0),
on these possibilities.
ˆ With X = C + Dx, we get
X(0) = 0 = C,
X(L) = 0 = C + DL,
from which it follows that both C and D are zero, so that X = 0. Not interesting.
ˆ Next, we consider X = Cerx/c + De−rx/c . Imposing X(0) = 0, we get
C + D = 0 ⇒ D = −C.
This allows us to rewrite X(x) as
X(x) = C(erx/c − e−rx/c ) = 2C sinh(rx/c).
Next, we impose X(L) = 0. We get
0 = 2C sinh(rL/c).
Clearly14 , we don’t want C = 0, since this would result in u(x, t) = 0. That is not
exciting. If we want excitement, we have to impose
sinh(rL/c) = 0.
Unfortunately, the sinh function is zero only when it’s argument is zero, which would
imply r = 0, since L 6= 0 (otherwise we’d have no string. But r = 0 is not allowed for
this case, since r = 0 implies λ = 0, which was the previous case. Bummer. All that
work and no solutions. What is this? A course on how not to find solutions to PDEs?
Patience, my young apprentices. . .
13
Resulting in only the extremely boring u = 0 solution. . .
14
The most hated word in a math text, with the possible exception of “obviously”.
46 CHAPTER 2. TRAVELING AND STANDING WAVES

ˆ Inevitably, we end up with all our money on the last case15 . Imposing X(0) = 0, with

X(x) = C cos(rx/c) + D sin(rx/c),

we get
C = 0.
Not a good start! We’ve just thrown out half of our last remaining non-zero solutions.
This leaves us with
X(x) = D sin(rx/c).
Next we impose X(L) = 0. We obtain

D sin(rL/c) = 0.

Since we don’t want D = 0, we need sin(rL/c) = 0, which is satisfied if

rL
= nπ,
c
where n is any integer. Since r is not allowed to be zero, we have to exclude n = 0.
Further, since we may assume that r > 0, we need only consider n ∈ Z+ 0 , the set of
strictly positive integers: n ∈ {1, 2, 3, 4, . . .}. Thus we have found
nπx
Xn (x) = sin , n = 1, 2, . . .
L
We can ignore the constant multiplication factor D, since it can be absorbed into the
multiplying function T (t), which we still have to determine. We have endowed X(x)
with an index n, to distinguish the different solutions we have found.

In summary, we find that the only standing wave solutions of the wave equation that
satisfies the boundary conditions for a fixed finite-length string are given by
   
nπct nπct  nπx 
un (x, t) = (A cos + B sin ) sin ,
L L L

for all n = 1, 2, . . ..

Modes of vibration
We call un the n-th mode of vibration of the string. The wave number of the n-th mode is
nπ/L. Over the domain x ∈ (0, L), the solution has exactly n − 1 zeros16 Note that all of
the Xn are sin functions, with increasing wave number as n increased. In other words, for
15
This statement is in no way an endorsement of any gambling activity. The University of Washington
and its employees do not encourage gambling in any way or form. Except for solving PDEs.
16
You should convince yourself of this statement. If you have extra time, convince your neighbor too.
2.13. STANDING WAVE SOLUTIONS OF THE WAVE EQUATION 47

larger n, Xn has a smaller period. For increasing n we are simply cramming more periods
in the interval [0, L].
We briefly revisit the x-problem we solved:

c2 X 00 = λX,
X(0) = 0,
X(L) = 0.

This is called a Sturm-Liouville problem for X(x). The constant λ is called the eigen-
value, X(x) the eigenfunction corresponding to λ.
The method we have used to solve for the modes of the wave equation is called Sepa-
ration of Variables, because we looked for that solutions that are multiples of functions
that depend on x and t separately.
In the next chapter, we look at how we can construct very general solutions from linear
superpositions of these standing wave solutions.
48 CHAPTER 2. TRAVELING AND STANDING WAVES
Chapter 3

Fourier series and solutions of partial

differential equations on a finite
interval

3.1 Superposition of standing waves

Since the wave equation is linear and homogeneous, we should be able to superimpose dif-
ferent solutions, in order to get new solutions.
Consider the problem
utt = c2 uxx , x ∈ (0, L), t > 0,
u(0, t) = 0, t > 0,
u(L, t) = 0, t > 0.
We show explicitly that
v = α1 u1 (u, t) + α2 u2 (x, t)
solves this problem, provided that u1 and u2 do. Here α1 and α2 are constants.
First, we check that v satisfies the PDE:
(α1 u1 + α2 u2 )tt = α1 u1tt + α2 u2tt
= α1 c2 u1xx + α2 c2 u2xx
= c2 (α1 u1xx + α2 u2xx )
= c2 (α1 u1 + α2 u2 )xx .
It follows that vtt = c2 vxx .
Next we verify that v satisfies the boundary conditions. We have
v(0, t) = α1 u1 (0, t) + α2 u2 (0, t) = α1 · 0 + α2 · 0 = 0,
and
v(L, t) = α1 u1 (L, t) + α2 u2 (L, t) = α1 · 0 + α2 · 0 = 0.

49
50 CHAPTER 3. FOURIER SERIES

This proves what we had to show.

Example. But, be careful: this does not work for nonhomogeneous problems, even if
the nonhomogeneous part is in the boundary conditions. For instance, consider

utt = c2 uxx , x ∈ (0, L), t > 0,

u(0, t) = 0, t > 0,
u(L, t) = 1, t > 0.

As you repeat the calculation above for this problem, you encounter a problem with the
second boundary condition. Bummer!
The above implies we can add our standing wave solutionsto get new, more general
solutions of the wave equation. Let

N
X
u(x, t) = un (x, t)
n=1
N     
X nπct nπct  nπx 
= An cos + Bn sin sin .
n=1
L L L

This is a superposition of a finite-number of standing waves. Any constant choice of An and

Bn gives a new solution. Thus u(x, t) depends on 2N parameters. The idea would be to pick
these parameters An and Bn so we can satisfy the initial conditions, if any are given.
Example. Consider the initial-value problem

utt = uxx , 0 < x < 1, t > 0,

u(0, t) = 0, t > 0,
u(1, t) = 0 t > 0,
u(x, 0) = 0, 0 < x < 1,
ut (x, 0) = 2 sin(πx) − 3 sin(4πx), 0 < x < 1.

Let
N
X
u(x, t) = [An cos (nπt) + Bn sin (nπt)] sin (nπx) .
n=1

At this point, we have satisfied the PDE and the boundary conditions. From the initial
conditions, we get
N
X
u(x, 0) = 0 = An sin(nπx),
n=1
3.1. SUPERPOSITION OF STANDING WAVES 51

and
N
X
ut (x, 0) = [(−An )nπ sin(nπt) + Bn nπ cos(nπt) sin(nπx)]|t=0
n=1
XN
= Bn nπ sin(nπx)
n=1
= 2 sin πx − 3 sin 4πx.
The first equation is easily satisfied by choosing
An = 0,
for all n. Next, we can choose N = 4, so that the second equation becomes
B1 π sin πx + B2 2π sin 2πx + B3 3π sin 3πx + B4 4π sin 4πx = 2 sin πx − 3 sin 4πx,
which is solved by choosing
2 −3
B1 =
, B2 = 0, B3 = 0, B4 = .
π 4π
This solves the given initial-value problem.
This whole thing feels like a cheat, right? Those were very special initial conditions! Can
we do something in general? For starters, can we show that the above is the unique solution
to the problem? The answers to these questions will be “Yes!” and “Yes!”1 .
Let’s kick this up a notch© . How about if we include all standing wave solutions we
know? Can we let N → ∞? We would have
∞     
X nπct nπct  nπx 
u(x, t) = An cos + Bn sin sin .
n=1
L L L
Does this make sense? Perhaps. If An , Bn → 0 sufficiently fast, we might get a convergent
series. Such a convergent series is called a Fourier series. The big trick is to see whether
we can find An and Bn such that we can satisfy general initial conditions, not just the really
special ones we used above. Perhaps this is possible.
Example. Suppose the initial condition is given as
1 1
u(x, 0) = sin πx − sin 3πx + sin 5x − . . .
9 25
∞
X (−1)n+1
= sin(2n − 1)πx,
n=1
(2n − 1)2
and ut (x, 0) = 0. Clearly, we would need
1 1
A1 = 1, A2 = 0, A3 = − , A4 = 0, A5 = , . . .
9 25
Can we do this for more general initial conditions?
1
Yes, with exclamation points.
52 CHAPTER 3. FOURIER SERIES

3.2 Fourier series

Here is the concrete question we will answer: given f (x), c ∈ [−L, L], we wish to write f (x)
as a linear combination of sines and cosines, where all of these have period 2L. Thus, we’d
want the terms in our linear combination to contain
 nπx   nπx 
sin and cos ,
L L
for n = 0, 1, 2, . . .2 . Before we proceed, here are some identities we need that will be very
helpful:
Z L  nπx   mπx 
sin sin dx = Lδnm ,
−L L L
Z L  nπx   mπx 
cos cos dx = Lδnm ,
−L L L
Z L  nπx   mπx 
sin cos dx = 0,
−L L L
where δnm is the Kronecker delta:

1, n = m,
δnm =
0, n 6= m.
These are easy to prove. Let’s look at the first one3 . Recall the following trig identities:
sin α sin β − cos α cos β = − cos(α + β)
sin α sin β + cos α cos β = cos(α − β)
−cos(α + β) + cos(α − β)
⇒ sin α sin β = .
2
This allows for an easy evaluation of the integrals. Assuming that n 6= m, we get
Z L
1 L
Z     
 nπx   mπx  (n + m)πx (n − m)πx
sin sin dx = − cos + cos dx
−L L L 2 −L L L
   L   (n−m)πx  L
(n+m)πx
1 sin sin
= −
L
 +1 L

2 (n + m)π/L 2 (n − m)π/L
−L −L
 
1 sin(n + m)π + sin(n + m)π
= − +
2 (n + m)π/L
 
1 sin(n − m)π + sin(n − m)π
2 (n − m)π/L
= 0.
2
Fourier was declared crazy by his contemporaries for wanting to answer this question, even though he
was successful!
3
You know the drill: you should look at the other two.
3.2. FOURIER SERIES 53

On the other hand, if n = m, then

Z L  nπx   mπx  Z L  nπx 
sin sin dx = sin2 dx
−L L L −L L
1 L
Z   
2nπx
= 1 − cos dx
2 −L L
= L,

since the integral of the second term is zero. Combining these two results, we get the desired
identity: Z L  nπx   mπx 
sin sin dx = Lδnm .
−L L L
The others are proven in a similar way.
Back to our main business: we wish to find an and bn such that
∞ 
X  nπx   nπx 
f (x) = A + an cos + bn sin .
n=1
L L

Let’s start by taking the average of this expression:

Z L Z L ∞ 
1 1 X  nπx   nπx 
f (x)dx = A + dx an cos + bn sin .
2L −L 2L −L n=1
L L

Since the average of all the sines and cosines is zero, we get
Z L
1
A= f (x)dx.
2L −L

Next, let m = 1, 2, . . .. Then

 mπx   nπx  X ∞  nπx   mπx  X ∞  nπx   mπx 
f (x) sin = A sin + an cos sin + bn sin sin .
L L n=1
L L n=1
L L

Once again, we integrate:

Z L  mπx  Z L  mπx  ∞
X Z L  nπx   mπx 
f (x) sin dx = A sin dx + + an cos sin dx
−L L −L L n=1 −L L L
∞
X Z L  nπx   mπx 
+ bn sin sin dx
n=1 −L L L
∞
X Z L  nπx   mπx 
= bn sin sin dx.
n=1 −L L L
54 CHAPTER 3. FOURIER SERIES

The only nonzero terms occur for n = m, so that

1 L
Z  mπx 
bm = f (x) sin dx, m = 1, 2, . . . .
L −L L

Similarly, we get
Z L
1  mπx 
am = f (x) cos dx, m = 1, 2, . . . .
L −L L

Note that A = a0 /2, so that we may write

∞
a0 X   nπx   nπx 
f (x) = + an cos + bn sin ,
2 n=1
L L

with
1 L
Z  nπx 
an = f (x) cos dx, n = 0, 1, 2, . . . ,
L −L L
1 L
Z  nπx 
bn = f (x) sin dx, n = 1, 2, . . . .
L −L L

This series for f (x) using trig functions is called its Fourier series. It arises not only in
wave and PDE problems, but also in image processing, data analysis, etc.
Example. Consider
f (x) = x2 , x ∈ [−1, 1],
as plotted in Fig. 3.1. For this example, L = 1, and we get

1 L
Z  nπx 
an = f (x) cos dx
L −L L
Z 1
= x2 cos(nπx)dx
−1
4nπ(−1)n
=
n3 π 3
4(−1)n
= 2 2 ,
nπ
where we have used integration by parts a few times4 . Next,

1 L
Z  nπx  Z 1
bn = f (x) sin dx = x2 sin(nπx)dx = 0,
L −L L −1

4
Per the Constitution, you should check this.
3.2. FOURIER SERIES 55

f (x)

Figure 3.1: The function f (x) = x2 over its domain of definition x ∈ [−L, L].

since the integrand is an odd function. It follows that for x ∈ [−1, 1],
∞
a0 X
x2 = + an cos nπx
2 n=1
∞
1 4 X (−1)n
= + 2 cos nπx.
3 π n=1 n2

This example immediately leads to some special cases and remarks.

Remarks.
ˆ When we construct the Fourier series for f (x) = x2 , for x ∈ [−1, 1], we are in fact
constructing the Fourier series for the periodic extension of f (x), consisting of the
periodic repetition of the function between [−1, 1], as plotted in Fig. 3.2.
ˆ If f (x) is even then bn = 0, and
1 L
Z  nπx 
an = f (x) cos dx
L −L L
2 L
Z  nπx 
= f (x) cos dx,
L 0 L
and ∞
a0 X  nπx 
f (x) = + an cos .
2 n=1
L
56 CHAPTER 3. FOURIER SERIES

f (x)

Figure 3.2: The periodic extension function of the function f (x) = x2 .

This is called the Fourier cosine series.

ˆ If f (x) is odd then an = 0, and
1 L
Z  nπx 
bn = f (x) sin dx
L −L L
2 L
Z  nπx 
= f (x) sin dx,
L 0 L
and ∞
X  nπx 
f (x) = bn sin .
n=1
L
This is called the Fourier sine series.
Next, we will start to use Fourier series to solve PDEs.

3.3 Fourier series solutions of the wave equation

Let’s return to the problem of a finite string with fixed ends:

utt = c2 uxx , x ∈ (0, L), t > 0,

u(x, 0) = f (x), x ∈ (0, L),
ut (x, 0) = g(x), x ∈ (0, L),
u(0, t) = 0, t > 0,
u(L, t) = 0, t > 0.
3.3. FOURIER SERIES SOLUTIONS OF THE WAVE EQUATION 57

We already know that

∞  nπx      
X nπct nπct
u(x, t) = A + sin an cos + bn sin ,
n=1
L L L

is the most general solution of the wave equation that satisfies the boundary conditions,
found using separation of variables. It remains to impose the initial conditions. First, by
letting t = 0, we get
∞
X  nπx 
f (x) = A + an sin .
n=1
L
Next, we take a derivative of u(x, t) with respect to t, and we let t = 0. This results in
∞
X nπc  nπx 
g(x) = bn sin .
n=1
L L

Because of the boundary conditions (fixed end), we use an odd extension of f (x). Using the
Fourier sine series, we get
A = 0,
and Z L
2  nπx 
an = f (x) sin dx,
L 0 L
for n = 1, 2, . . .. Similarly, we use an odd extension for g(x), since the above indicates we
wish to use a Fourier sine series. We get

2 L
Z
nπc  nπx 
bn = g(x) sin dx,
L L 0 L
Z L
2  nπx 
⇒ bn = g(x) sin dx.
nπc 0 L
also for n = 1, 2, . . ..
This completely determines the solution of the wave equation problem with the given
initial data.
Let’s consider a different problem, namely that with two free ends.

utt = c2 uxx , x ∈ (0, L), t > 0,

u(x, 0) = f (x), x ∈ (0, L),
ut (x, 0) = g(x), x ∈ (0, L),
ux (0, t) = 0, t > 0,
ux (L, t) = 0, t > 0.
58 CHAPTER 3. FOURIER SERIES

We begin by using separation of variables to find the standing wave solutions. Let

u(x, t) = X(x)T (t).

Then
T 00 X 00
= = λ,
c2 T X
so that λ has to be constant. It follows that

X 00 = λX.

Note that the boundary conditions imply that X 0 (0) = 0 = X 0 (L). We consider three
different cases.

ˆ λ = 0. We have
X 00 = 0 ⇒ X = D,
where we have already used the boundary conditions. This case results in a constant
(as a function of x) standing wave, given by

u0 (x, t) = A0 + B0 t,

where we have equated D = 1, since we can absorb the constant in the values of A0
and B0 .

ˆ λ > 0. We set λ = r2 , where r > 0. Then

X(x) = c1 cosh(rx) + c2 sinh(rx).

To impose the boundary conditions, we need

X 0 (x) = c1 r sinh(rx) + c2 cosh(rx).

From X 0 (0) = 0, it follows that

0 = c2 .
Next, from X 0 (L) = 0, we are left with

0 = c1 r sinh(rL),

which requires c1 = 0, so that the exponential case does not result in any solutions.

ˆ λ < 0. Now we let λ = −r2 , again with r > 0. We know we get solutions of the form

X(x) = c1 cos(rx) + c2 sin(rx).

For the boundary conditions, we need

X 0 (x) = −c1 r sin(rx) + c2 r cos(rx).

3.3. FOURIER SERIES SOLUTIONS OF THE WAVE EQUATION 59

From X 0 (0) = 0 we get

c2 r = 0 ⇒ c2 = 0.
Next, using X 0 (L) = 0, we get

−c1 r sin(rL) = 0.

Since we wish to avoid c1 = 0, we impose that

nπ
r= ,
L
so that  nπx 
Xn (x) = cos ,
L
for n = 1, 2, 3, . . .. We get the standing wave solutions
 nπx   
nπct
 
nπct

un (x, t) = cos An cos + Bn sin
L L L

Using superposition, we get that the most general solution satisfying the boundary
conditions is given by
∞  nπx      
X nπct nπct
u(x, t) = A0 + B0 t + cos An cos + Bn sin .
n=1
L L L

It remains to impose the initial conditions, so as to determine the constants An and

Bn . First, we plug in x = 0. We get
∞
X  nπx 
f (x) = A0 + An cos .
n=1
L

Using an even extension of f (x), we get

Z L
2  nπx 
An = f (x) cos dx, n = 1, 2, . . . ,
L 0 L

and Z L
1
A0 = f (x)dx,
L 0
using the formulae for the Fourier cosine series. Lastly, to impose the second initial
condition, we take a derivative with respect to t, and we let t = 0. We get
∞
nπcX  nπx 
g(x) = B0 + Bn cos .
n=1
L L
60 CHAPTER 3. FOURIER SERIES

Using an even extension for g(x), we get that

1 L
Z
B0 = g(x)dx,
L 0
and
2 L
Z
nπc  nπx 
Bn = g(x) cos dx
L L 0 L
Z L
2  nπx 
⇒ Bn = g(x) cos dx.
nπc 0 L
Note that in a practical problem, you might have to impose that the average of the
initial velocity g(x) is zero. If this is not the case, your solution to the string-with-
free-ends problem will have a component that is linearly growing!

3.4 The heat equation

Consider the problem

ut = σuxx , x ∈ (0, L), t > 0,

u(x, 0) = f (x), x ∈ (0, L),
u(0, t) = 0, t > 0,
u(L, t) = 0, t > 0.

Here σ > 0 is the heat conductivity coefficient. The heat equation describes heat flow in a
medium with heat conductivity σ. We will consider only the one-dimensional heat equation.
As for the wave equation, we begin by looking for solutions of the form

u(x, t) = X(x)T (t).

We get

XT 0 = σX 00 T
T0 X 00
⇒ = = λ.
σT X
Here λ is a separation constant, using the same argument we used for the wave equation:
since the left-hand side depends only on t, and the right-hand side depends only on x, they
must both be constant. It follows that X(x) satisfies the following problem:

X 00 − λX = 0,
X(0) = 0,
X(L) = 0.
3.5. LAPLACE’S EQUATION 61

This is the same exact problem for X as we had for the wave equation with fixed ends
(Dirichlet boundary conditions). As a consequence, we know we get solutions only for
λ = −r2 < 0, and  nπx 
Xn (x) = sin , n = 1, 2, . . . ,
L
and
n2 π 2
λn = − 2 .
L
It follows that Tn satisfies the ordinary differential equation
Tn0 n2 π 2 2 π 2 t/L2
= 2 ⇒ Tn = e−σn .
σTn L
We could have included a multiplicative constant, but this is not necessary, as the next step
is to take a linear superposition of the solutions un (x, t) = Xn (x)Tn (t) we have just found:
2 2 2
 nπx 
un (x, t) = e−σn π t/L sin .
L
The superposition results in the general solution
∞  nπx 
−σn2 π 2 t/L2
X
u= cn e sin .
n=1
L

It remains to impose the initial condition. At t = 0, we get

∞
X  nπx 
f (x) = cn sin ,
n=1
L

since all the exponentials become 1, when evaluated at t = 0. Using an odd extension of
f (x), we get that
2 L
Z  nπx 
cn = f (x) sin dx, n = 1, 2, . . . .
L 0 L
Thus, we have solved the initial-value problem for the heat equation with prescribed zero
temperature at the ends5 .

3.5 Laplace’s equation

Consider the problem

uxx + uyy = 0, x ∈ (0, L), y ∈ (0, M ).

This is Laplace’s equation, posed on a rectangle. You could imagine getting to Laplace’s
equation by wanting to find time-independent (or stationary) solutions of the multi-dimensional
5
That was quick! We’re getting to be good at this.
62 CHAPTER 3. FOURIER SERIES

M
u(x, M ) = 0

ux (0, y) = 0 ux (L, y) = 0

0 u(x, 0) = f (x) L

Figure 3.3: The domain for Laplace’s equation on the rectangle.

wave equation utt = c2 (uxx +uyy ) or of the multi-dimensional heat equation ut = σ(uxx +uyy ).
We impose boundary conditions as follows, see Fig. 3.3:

u(x, 0) = f (x),
ux (0, y) = 0,
ux (L, y) = 0,
u(x, M ) = 0.

First, we look for separated solutions:

u(x, y) = X(x)Y (y),

and we get
X 00 Y 00
X 00 Y + XY 00 = 0 ⇒ =− = λ.
X Y
As before, λ is a constant, since X 00 /X and −Y 00 /Y are dependent on x and y separately.
We have that

X 00 − λX = 0,
X 0 (0) = 0,
X 0 (L) = 0.

As usual, we consider different cases for λ.

ˆ λ > 0. You should check that this does not result in any solutions for X(x).
3.5. LAPLACE’S EQUATION 63

ˆ λ = 0. This results in the solution X0 (x) = 1. Similarly, Y0 (y) = A + By. Imposing

Y (M ) = 0, we find 0 = A + BM , so that A = −BM . Thus Y0 (y) = B(y − M ).
This gives rise to the constant solution u0 (x, y) = y − M , where we have omitted the
multiplicative constant6 .

ˆ λ < 0. As before, we get

X = c1 cos(rx) + c2 sin(rx).

To impose the boundary conditions, we need

X 0 = −c1 r sin(rx) + c2 r cos(rx).

From the first boundary condition, we get

0 = c2 r ⇒ c2 = 0.

The second boundary condition gives

nπ
0 = −c1 r sin(rL) ⇒ r = .
L
This gives rise to  nπx 
Xn (x) = cos , n = 1, 2, . . . .
L
Next, we have to solve for Yn (y). We have

n2 π 2
Yn00 − Yn = 0
L2  nπ   nπ 
⇒ Yn = An cosh (y − M ) + Bn sinh (y − M ) ,
L L
where we have opted to write the solution using hyperbolic functions of a shifted
argument. Imposing the boundary condition yn (M ) = 0, we find that

An = 0.

It follows that  nπ 
Yn (y) = Bn sinh (y − M ) .
L
The linear superposition of all solutions is7
∞
X  nπ   nπx 
u(x, y) = B0 (y − M ) + Bn sinh (y − M ) cos .
n=1
L L
6
For now. Don’t worry. It’ll come back.
7
It’s back!
64 CHAPTER 3. FOURIER SERIES

Imposing the one remaining boundary condition, we get

∞  
X nπM  nπx 
u(x, 0) = f (x) = −M B0 − Bn sinh cos .
n=1
L L

the rest is some simple Fourier series stuff, using an even extension8 of f (x)!

ˆ For n = 0,
Z L
1
−M B0 = f (x)dx
L 0
Z L
1
⇒ B0 = − f (x)dx.
ML 0

ˆ For n = 1, 2, . . .,
  Z L
nπM 2  nπx 
−Bn sinh = f (x) cos dx
L L 0 L
Z L
2  nπx 
⇒ Bn = − nπM

f (x) cos dx.
L sinh L 0 L

This completely determines the solution of the Laplace problem on the rectangle.

3.6 Laplace’s equation on the disc

Consider the problem (see Fig. 3.4)

uxx + uyy = 0, x2 + y 2 < R2 ,

u(x, y) = f (x, y), x2 + y 2 = R 2 .

It is clear that we should reformulate this problem in polar coordinates. Let

v(r, θ) = u(x, y).

Then9
∂u ∂v ∂r ∂v ∂θ
ux = = + .
∂x ∂r ∂x ∂θ ∂x
From
r 2 = x2 + y 2 ,
8
Because the right-hand side is a cosine series.
9
You can see the chain rules coming a mile away!
3.6. LAPLACE’S EQUATION ON THE DISC 65

x2 + y 2 = R 2

Figure 3.4: The domain for Laplace’s equation on the disc.

it follows that
∂r ∂r x
2r = 2x ⇒ = .
∂x ∂x r
similarly, from
y
tan θ = ,
x
we get
∂θ y r2 ∂θ y ∂θ y
sec2 θ =− 2 ⇒ = − ⇒ = − 2.
∂x x x2 ∂x x2 ∂x r
Thus
x y
ux = vr − vθ 2
r r
sin θ
= vr cos θ − vθ .
r
Next, we apply this same process to get uxx = (ux )x : in other words, we repeat the above,
but with v replaced by the expression we found for ux . This gives
   
sin θ sin θ sin θ
uxx = vr cos θ − vθ cos θ − vr cos θ − vθ
r r r θ r
   
sin θ sin θ sin θ cos θ sin θ
= vrr cos θ − vrθ + vθ 2 cos θ− vrθ cos θ − vr sin θ − vθθ − vθ
r r r r θ r
2 2
sin θ cos θ sin θ cos θ sin θ sin θ
= vrr cos2 θ − 2vrθ + 2vθ + v r + vθθ .
r r2 r r2
66 CHAPTER 3. FOURIER SERIES

Similarly, we find

2 sin θ cos θ sin θ cos θ sin2 θ cos2 θ

uyy = vrr sin θ + 2vrθ − 2vθ + vr + vθθ 2 .
r r2 r r
Adding these expressions, we get amazing simplifications10 :
vr vθθ
uxx + uyy = 0 ⇒ vrr + + 2 = 0.
r r
We can reformulate our original problem in polar coordinates:
vr vθθ
vrr + + 2 = 0,
r r
v(R, θ) = f (θ), θ ∈ [0, 2π),

since the boundary condition is given on the circle of radius R.

As usual, we apply separation of variables to look for solutions of the form

v(r, θ) = S(r)T (θ).

We get

00 S 0T T 00 S
S T+ + 2 =0
r r
00 0
S rS T 00
⇒ r2 + + =0
S S T
S 00 rS 0 T 00
⇒ r2 + =− = λ,
S S T
where λ is a separation constant. Indeed, in this last line, the left-hand side is a function of
r only, while −T 00 /T depends only on θ. Thus both are constant.
Since the equation for T is the simplest, we solve it first.

ˆ If λ = 0, then
T 00 = 0 ⇒ T = a + bθ.
We want v(r, θ) to be single-valued as a function of both r and θ. This implies that T
should be a periodic function of θ, with period 2π. Thus, b = 0. This leaves us with

T0 = 1,

ignoring the multiplicative constant a.

10
Always a good sign.
3.6. LAPLACE’S EQUATION ON THE DISC 67

ˆ If λ = α2 > 0, with r > 0, we get

T = a cos αθ + b sin αθ.

Requiring that T is periodic with period 2π, we find that

α = n,

an integer. Then λn = n2 . This gives

Tn = an cos nθ + bn sin nθ.

ˆ If λ = −α2 < 0, we find no 2π-periodic solutions11 .

Next, we solve for S(r).

ˆ With λ0 = 0, then

r2 S000 + rS00 =0
⇒ rS000 + S00 =0
⇒ (rS00 )0 =0
⇒ rS00 = c1
⇒ S0 = c1 ln r + c2 .

This is infinite as r → 0, which we do not want to allow12 . Thus we require c1 = 0, so

that
S0 = 1,
where we, once again, have ignored the multiplicative constant. Thus,

v0 (r, θ) = S0 (r)T0 (θ) = 1.

ˆ Next, we examine λn = n2 , n = 1, 2, . . .. We have

r2 Sn00 + rSn0 − n2 Sn = 0.

This is an Euler or equivariant ordinary differential equation. Thus we look for

solutions of the form
S = rp .
We get

p(p − 1) + p − n2 = 0
⇒ p2 = n2
⇒ p = ±n.
11
Don’t make me say it. OK, fine: you should check this.
12
Nothing goes to infinity on our watch!
68 CHAPTER 3. FOURIER SERIES

This gives

Sn = cn rn + dn r−n .

Since n > 0, the second term → ∞ as r → 0. Thus we require dn = 0. Thus

Sn = r n ,

up to a multiplicative constant.

The general solution is a linear superposition of all solutions we have found:

∞
a0 X n
v(r, θ) = + r (an cos nθ + bn sin nθ).
2 n=1

Having found the general solution, we impose the boundary condition

v(R, θ) = f (θ).

This results in ∞
a0 X n
f (θ) = + R (an cos nθ + bn sin nθ).
2 n=1

Using the Fourier series formulae, we find that

Z 2π
1
an = f (θ) cos nθdθ,
πRn 0
for n = 0, 1, . . ., and Z 2π
1
bn = f (θ) sin nθdθ,
πRn 0
for n = 1, 2, . . .. This completely determines the solution to the Laplace equation posed on
the inside of a circle.
Chapter 4

The method of characteristics

4.1 Conservation laws

Let’s go back to wave behavior. We’ll go beyond the wave equation.

Derivation of a general scalar conservation law

A conservation law tells us the way in which a particular quantity can change. The simplest
examples you know are that in conservative systems, energy is conserved.
Suppose we have a one-dimensional setting, for convenience we call it the x-axis. Suppose
the quantity Q changes with our dynamics.
Example. Q could represent
ˆ the number of cars in a traffic flow problem,

ˆ the number of particles in molecular chemistry problem,

ˆ the energy of a physical system,

ˆ the number of people in a social dynamics problem.

Let u(x, t) denote the density of Q. In other words, Q in S is given by

Z b
Q= u(x, t)dx.
a

S x

x=a x=b

Figure 4.1: A one-dimensional domain, setting up for the derivation of a conservation law.

69
70 CHAPTER 4. THE METHOD OF CHARACTERISTICS

The quantity Q in S can change in two ways:

1. Q could enter or leave S through a or b, and

2. Q is created or destroyed in S.

It follows that the rate of change of Q, dQ/dt, is given by the rate at which Q enters or
leaves at x = a, plus the rate at which Q enters or leaves at x = b, plus the rate at which Q
is created or destroyed in S. In equations,
Z b Z b
d
u(x, t)dx = φ(a, t) − φ(b, t) + f (x, t)dx,
dt a a

where φ(x, t) is the rate at which Q moves past x at time t. If φ(x, t) > 0, then the flow is
in the positive x direction, otherwise it is in the negative x direction. Thus, the net rate at
which Q enters through the ends of S is

φ(a, t) − φ(b, t).

The − sign for the second term is a consequence of our flow convention: if the flow of Q
through x = b is to the right, it is leaving S, thus it results in a decrease.
Lastly, if Q is created or destroyed in S, this happens with a source or sink function
f (x, t), resulting in an amount of Q that is added equal to
Z b
f (x, t)dx.
a

Our rate-of-change equation becomes

Z b Z b
d
u(x, t)dx = φ(a, t) − φ(b, t) + f (x, t)dx.
dt a a

Suppose that u and φ have continuous derivatives, then

Z b Z b Z b
ut (x, t)dx = − φx (x, t)dx + f (x, t)dx
a a a
Z b
⇒ (ut + φx − f )dx = 0.
a

Since this is true for all a and b, we get that

ut + φx = f.

This is called the differential form of the conservation law.

4.2. EXAMPLES OF CONSERVATION LAWS 71

Constitutive relations
Even if we consider f (x, t) as given, we still have one partial differential equation for two
quantities u and φ. A constitutitve relation relates u and φ. In many cases such a relation
gives φ as a function of u. Then φ = φ(u), and we get

ut + φ0 (u)ux = f.

Example. The inviscid Burgers equation

ut + uux = 0,

is a conservation law with f = 0, φ = u2 /2. However, there are more possibilities. The same
equation can be written as
   
2 1 2 1 3
uut + u ux = 0 ⇒ u + u = 0.
2 t 3 x

If we let
1
v = u2 ,
2
then
u = (2v)1/2 ,
and
u3 1
φ= = (2v)3/2 ,
3 3
and the Burgers equation can be rewritten as
 
1 3/2
vt + (2v) = 0.
3 x

Which form of the equation we choose will matter in the following lectures. In practice, it
is of course dictated by the application we are working on.

4.2 Examples of conservation laws

Diffusion
Consider the undesirable scenario of a pollutant spreading in stagnant water in a horizontal
pipe, see Fig. 4.2. Let u(x, t) denote the concentration of pollutant (in mass/length). Then

ut + φx = f
is our conservation law. Assume that we have no magical pollutant eating piranhas in the
pipe, and no pollutant is destroyed or created, then f = 0. Next, we need to relate the flux
function φ to the concentration u.
72 CHAPTER 4. THE METHOD OF CHARACTERISTICS

Figure 4.2: Diffusion of a pollutant in stagnant water.

(a) (b) Less

Less
Lots of pollutant

Figure 4.3: Initial concentration of a pollutant in stagnant water.

Suppose we have an initial concentration u0 (x) shown below in Fig. 4.3. We expect that
pollutant will flow from areas where there is a lot to areas where there is less. If ux is
positive, as in (a), then φ should be negative so that pollutant flows to the left. If, at (b),
ux is negative, then φ should be positive and u will flow to the right. The simplest way to
make this happen is Fick’s Law:

φ(x, t) = −Dux (x, t),

where D > 0 is the diffusion constant.

Putting all of this together, we get

ut + φx = 0
⇒ ut + (−Dux )x = 0
⇒ ut = Duxx ,

the heat or diffusion equation!

Traffic flow
This goes back to studies by Whitham and Lighthill in the 50s. We approximate the number
of cars per unit length by a continuous function u(x, t). Assuming there are no exits or
entrances, we get
f = 0,
4.3. THE METHOD OF CHARACTERISTICS 73

so that our conservation law becomes

ut + φx = 0.

As before, we have to determine a relation between φ and u.

φ = cars/time unit
= Rate at which cars are passing x at t
= u × v,

where u is the car density and v is the velocity. Thus, we need to relate the velocity v to u,
and we will be done. Clearly1 , if u is high then v will be low, and if u is near zero, v should
be maximal, the speed limit. The simplest way to do this is

v = v1 − au,

where v1 is the speed limit and a is a positive constant. As u → v1 /a, v → 0, thus v1 /a = u1

is the maximal density possible. We can rewrite the velocity as
 
u
v = v1 1 − ,
u1

for u ∈ [0, u1 ]. Our constitutive relation becomes

u2
   
u
φ = uv = v1 u 1 − = v1 u − .
u1 u1

This last form is less instructional, but it is easier to take a derivative. The conservation law
becomes

ut + φx = 0
 
2uux
⇒ ut + v1 ux − =0
u1
 
2u
⇒ ut + v1 1 − ux = 0.
u1

4.3 The method of characteristics

We will use the method of characteristics to solve conservation laws of the form

ut + φx = f,
u(x, 0) = u0 (x),
1
That word again!
74 CHAPTER 4. THE METHOD OF CHARACTERISTICS

x(t)

x0

Figure 4.4: A characteristic curve in the (x, t) plane.

where −∞ < x < ∞, t > 0, unless otherwise stated.

Let’s start with the simplest equation of this form, the advection equation. Suppose
that φ = cu, then we have, with f = 0,

ut + cux = 0,
u(x, 0) = u0 (x).

The method of characteristics looks for special curves in the (x, t) plane along which our
PDE becomes an ODE. In other words, we wish to find x(t) in the (x, t) plane such that we
only have to solve ODEs along this curve, see Fig. 4.4.
Along these curves u(x, t) becomes a function of t only. Then

d dx
u(x(t), t) = ux + ut .
dt dt
We compare this expression with the PDE we wish to solve

ut + cux = 0.

We see that if we pick the curves (called the characteristic curves or characteristics)
such that
dx
= c,
dt
then our PDE simply becomes
du
(x(t), t) = 0.
dt
Thus u is constant along characteristic curves in this case. The value of the constant is of
course determined by the initial conditions.
4.3. THE METHOD OF CHARACTERISTICS 75

Thus, we have to solve

du
= 0, along curves for which
dt
dx
= c, so that
dt
u(x, 0) = u0 (x),

the given initial condition. We get

dx
= 0 ⇒ x = ct + x0 ,
dt
where x0 is the starting point of the characteristic curve at t = 0. Next, from
du
= 0 ⇒ u = A(x0 ),
dt
a constant, which could be different, depending on which characteristic we are solving for u.
We know that at t = 0

u(x(t), t)|t=0 = u(x(0), 0) = u(x0 , 0) = u0 (x0 ).

We also have that

x0 = x − ct,
and thus
u = u0 (x − xt),
which is the general solution to our advection equation. It shows that the advection equation
simply moves the initial condition to the right with velocity c, see Fig. 4.5.
Example. Consider

ut + 4ux = 0,
u(x, 0) = arctan(x).

We have to solve
du
= 0,
dt
along curves for which
dx
= 4 ⇒ x = 4t + x0 .
dt
It follows that
u = arctan(x0 ) = arctan(x − 4t).

Next, we consider the case of a nonhomogeneous advection equation. We have

ut + cux = f (x, t).

76 CHAPTER 4. THE METHOD OF CHARACTERISTICS

slope 1/c
x−x0
t= c

x
x0

Figure 4.5: The characteristics for the advection equation.

Proceeding as before, we look for curves x(t) so that we get ODEs. Along these curves,

d dx
u(x(t), t) = ut + ux .
dt dt
Thus we solve
dx
= c,
dt
x(0) = x0 .

The characteristics are the solutions of this system. Along the characteristics, we have to
solve
du
= f (x(t), t),
dt
which is easily solved by simply integrating both sides.
Example. Consider the nonhomogeneous problem

ut + 4ux = 1,
u(x, 0) = arctan(x).

We have to solve
dx
= 4 ⇒ x = 4t + x0 .
dt
Along these straight-line characteristics, we solve

du
= 1 ⇒ u = t + A,
dt
4.3. THE METHOD OF CHARACTERISTICS 77

where A is a constant of integration. Evaluating this at t = 0, we get

u(x0 , 0) = A = arctan(x0 ).

It follows that
u = t + arctan(x − 4t).

General linear conservation laws

A general linear conservation law is of the form

ut + c(x, t)ux = f (x, t),

where we assume we have initial conditions

u(x, 0) = u0 (x).

As before, we have to solve the ODEs

du
= f (x, t), along curves determined by
dt
dx
= c(x, t), x(0) = x0 .
dt
The second equation is an ODE for x(t), but in general it may be hard to solve. Once it is
solved, the first equation gives us u as a function of t, along these characteristics. Note that
in this case, the characteristics are typically not straight lines.
Example. Consider the initial-value problem

ut + xux = 0,
1
u(x, 0) = .
1 + x4
We solve for the characteristics first. We have
dx
=x
dt
d
⇒ ln |x| = 1
dt
x
⇒ ln =t
x0
⇒ x = x0 et .

Next, since
du
= 0,
dt
78 CHAPTER 4. THE METHOD OF CHARACTERISTICS

Figure 4.6: The characteristics for different values of x0 with c(x, t) = x.

along the characteristics, with initial condition

1
u(x, 0) = .
1 + x4
We get
u = A,
a constant. Imposing the initial condition at t = 0,
1
u(x0 , 0) = = A.
1 + x40

It follows that
1
u= ,
1 + x4 e−4t
where we have used that x0 = x exp(−t). The characteristics are not straight lines in this
case. They are shown in Fig. 4.6.

Nonlinear conservation laws

Suppose that φ = φ(u), then our conservation law ut + φx = f becomes

ut + φ0 (u)ux = f.

Define
c(u) := φ0 (u).
4.3. THE METHOD OF CHARACTERISTICS 79

Our PDE is rewritten as

ut + c(u)ux = f.
It follows that we have to solve
du
= f, u(x0 , 0) = u0 (x0 ),
dt
along curves determined by
dx
= c(u), x(0) = x0 .
dt
In general this is a coupled, messy system of ODEs. We can make some more progress if
f ≡ 0. Then the first ODE becomes

du
= 0 ⇒ u = u0 (x0 ).
dt
Thus this equation tells us that u is constant along characteristics, even though at this point
we do not yet know what these characteristics are. We can substitute this result into our
ODE for the characteristics:
dx
= c(u0 (x0 )).
dt
Since the right-hand side is constant, we get

x = x0 + tc(u0 (x0 )).

It follows that the characteristics are all straight lines, but with varying slopes depending
on x0 , as illustrated in Fig. 4.7
The slope depends not only on where (x0 ) we start, but also on what the initial condition
u0 is there. In order to get the full solution to the problem, we need to solve the characteristic
equation x = x0 + tc(u0 (x0 )) for x0 as a function of x and t. More often than not, this is
not possible. In most cases, we have to be satisfied with an implicit representation of the
solution.
Example. Let φ = u2 /2. Then c(u) = u, and we consider the initial-value problem

ut + uux = 0,

0, x ≤ 0,
u(x, 0) = −1/x
e , x > 0.

The characteristics are given by

x = x0 , x0 ≤ 0,
−1/x0
x = x0 + te , x0 > 0.
80 CHAPTER 4. THE METHOD OF CHARACTERISTICS

x
x0
slope= c(u01(x0 ))

Figure 4.7: The characteristics for different values of x0 with different slopes 1/c(u0 (x0 )).

There is no way for us2 to solve this3 for x0 as a function of x and t. However, we can obtain
a perfectly fine implicit solution:

0, x ≤ 0,
u(x, t) =
e−1/x0 , x > 0,

where x0 is defined by the equation

x = x0 + te−1/x0 .

A plot of the characteristics is shown in Fig. 4.8.

4.4 Breaking and gradient catastrophes

We have seen that the characteristics for the equation

ut + c(u)ux = 0, u(x, 0) = u0 (x),

are all straight lines, with slope 1/c(u0 (x0 )), where x0 is the starting point of the character-
istic at t = 0.
Even though the characteristics are just straight lines, things4 can get very interesting.
Let’s see what can happen.
2
Or anyone!
3
Other than numerically.
4
Technical term.
4.4. BREAKING AND GRADIENT CATASTROPHES 81

Figure 4.8: The characteristics for the problem with the piecewise defined initial condition.

u0 (x)

Figure 4.9: The initial profile u0 (x).

1. The characteristics are parallel. Our solution is implicitly given by

x = x0 + tc(u0 (x0 )),
u = u0 (x0 ).
Suppose that we start with a profile as shown in Fig. 4.9.
Following the initial condition along the characteristics, we see that the initial values
simply translate as we go forward in time, see Fig. 4.10. Indeed, if c(u0 (x0 )) = c, a
constant, then x0 = x − ct, and
u = u0 (x − ct).

2. The characteristics are spreading out. Starting from an arctan-like profile, we get
the situation depicted in Fig. 4.11. Now the values of the solution get spread out, as
the values follow the characteristics. Thus, spreading characteristics lead to smoother
solutions.
82 CHAPTER 4. THE METHOD OF CHARACTERISTICS

u0 (x − ct)

u0 (x)

Figure 4.10: Moving the initial condition along the parallel characteristics.

3. The characteristics are crossing, as in Fig. 4.12. If the characteristics cross, then
the values of u0 (x) between two crossing characteristics starting at a and b get squeezed
together as t increases. Thus the solution becomes locally steeper, since rise/run→ 0,
since run→ 0 as t approaches the time at which the characteristics cross. At that time,
the solution becomes infinitely steep (i.e., it has a vertical tangent), which implies
that the differential form of the conservation law is no longer valid. Indeed,
to derive the differential form, we assumed that all derivatives of u existed and were
continuous. If one or all of these derivatives → ∞, we have to revisit the integral form
of the conservation law. Up to the time where we first get the vertical tangent, the
differential form works well. Past that time, we have a problem. As is seen in Fig. 4.12,
the crossing characteristics give rise to a wedge-like region where any point has multiple
characteristics passing through it. Outside of this wedge, we can immediately see what
the value of u is by tracing back the unique characteristic going through the point. The
value of the initial condition on that characteristic is the value of the solution. For
points in the wedge, this does not work, as it is unclear what characteristic to follow
back.
The formation of the solution with a vertical tangent is called a gradient catastro-
phe5 .
Let’s do a few examples.
Example. Consider the problem
ut + uux = 0,
5
Wow. Sounds serious. It is.
4.4. BREAKING AND GRADIENT CATASTROPHES 83

Figure 4.11: Moving the initial condition along the spreading characteristics.

u(x, 0) = arctan(x).

The implicit solution to this problem is given by

u(x, t) = arctan(x0 ),
x = x0 + tc(u0 (x0 ))
= x0 + tu0 (x0 )
= x0 + t arctan(x0 ).

We cannot solve this last equation for x0 as a function of x and t, and an implicit solution
is the best we can do. However, we can plot the characteristics. We have
x − x0
t= .
arctan(x0 )

These characteristics, for varying x0 are plotted in Fig. 4.11. We see that the characteristics
are spreading out, and the initial profile of a front-like arctan becomes less steep as time
progresses, since the characteristics in the region of the front are fanning out.
Example. Consider the problem

ut + uux = 0,
u(x, 0) = − arctan(x).
84 CHAPTER 4. THE METHOD OF CHARACTERISTICS

u(x, t = 0.8)

u0 (x)
x

Figure 4.12: Moving the initial condition along the crossing characteristics.

This is almost the same problem as above, but the sign of the initial condition is flipped.
The implicit solution to this problem is given by

u(x, t) = − arctan(x0 ),
x = x0 + tc(u0 (x0 ))
= x0 + tu0 (x0 )
= x0 − t arctan(x0 ).

As before, we cannot solve this last equation for x0 as a function of x and t, and an implicit
solution is the best we can do. However, we can plot the characteristics. We have

−x + x0
t= .
arctan(x0 )

These characteristics, for varying x0 are plotted in Fig. 4.13 (left). We see that the charac-
teristics are crossing, with an apparent gradient catastrophe occurring at t = 16 . A wedge
region where the characteristics cross is formed. A few time slices of solution profiles are
shown in the right panel of Fig. 4.13. We note that the front profile becomes steeper as we
approach the gradient catastrophe time, referred to as the breaking time.
As discussed, if the characteristics cross, we will get ux , ut → ∞, as t → tb , the so-called
breaking time. Now what? The PDE is no longer and we need to rethink what we are doing.
6
We’ll check whether this is correct soon.
4.4. BREAKING AND GRADIENT CATASTROPHES 85

t u(x, 1)

u(x, 0.8)

u(x, 0.4) x

u(x, 0)

Figure 4.13: Crossing characteristics and the solution profiles that go with them.

We know that the slope of the characteristics is the inverse of the velocity:
1
slope = .
c(u0 (x0 ))

Let’s look at the effect this has on different profiles. Let’s assume, as in our examples that
c(u) is an increasing function of u. In other words, higher values of u will move with higher
velocities. Suppose we start with an initial profile u0 (x) that resembles an increasing front,
as in Fig. 4.14. Since higher values of u travel faster, the top part of the profile will move
more ahead of the bottom part, and the overall effect is that of the spreading out of the
solution, i.e., the solution becomes less steep. As shown in the figure, this corresponds to
two characteristics where the right one has a lesser slope 1/c2 than the left one 1/c1 , leading
to the characteristics fanning out. Thus the values of u in between u1 and u2 are being
spread out over a larger x interval.
On the other hand, if we start with a downward front, we obtain the situation depicted
in Fig. 4.15. Now the higher velocities of the higher values of u lead to the steepening of the
profile. The characteristics cross, and the values of u between u1 and u2 are condensed in an
x interval that shrinks to a point at the crossing of the characteristics, leading to a profile
that is infinitely steep: the solution becomes steeper as the interval length decreases, until
it becomes vertical.
It might be tempting to guess that the solution behaves as illustrated in Fig. 4.16. This
is incorrect! The evolution from the first panel to the second one is correct. But at the
second panel, the solution has a vertical tangent, and we cannot rely on the characteristics
or anything else coming from the PDE anymore. Since the PDE is no longer valid, we cannot
86 CHAPTER 4. THE METHOD OF CHARACTERISTICS

u0 (x)

u2
c2 > c1
u1
c1
x

t
1
c1

1 1
c2
< c1

Figure 4.14: The effect of spreading characteristics

use it to move from the second to the last panel. Of course, the solution u(x, t) is supposed
to be a single-valued function of x and t. In the third panel, there exists an entire interval
of x values for which the solution is tripple valued. Woe!
Before we figure out what happens after a shock (a profile with vertical tangent) forms,
we should first determine when a shock forms. Thus, we want to determine the so-called
breaking time tb ≥ 0.

Example. We revisit the example we plotted the characteristics and the solution profiles
for in Fig. 4.12. The problem is given by

ut + uux = 0,
2
u(x, 0) = e−x .

The characteristics plotted in Fig. 4.12 are given by

2 2
x = x0 + te−x0 ⇒ t = (x − x0 )ex0 .

The plot seems to indicate that tb ≈ 1.2. How can we find this value?
4.4. BREAKING AND GRADIENT CATASTROPHES 87

u0 (x)

u1 c2 > c1

c1
u2

1
c2

1
c1

Figure 4.15: The effect of crossing characteristics

The breaking time tb

The breaking time tb is the first time for which ux or ut become infinite. Let us calculate ux
and ut . We have

u = u0 (x0 ),
∂x0
⇒ ut = u00 (x0 ) , and
∂t
∂x0
⇒ ux = u00 (x0 ) ,
∂x
where we have used the chain rule, since x0 depends implicitly on x and t. Assuming that
u0 (x0 ) is a nice profile (i.e., no vertical tangents), we see that we need to determine when
∂x0 /∂x and/or ∂x0 /∂t are infinite. We have

x = x0 + tc(u0 (x0 )).

Taking an x derivative, we get

∂x0 ∂x0
1= + tc0 (u0 (x0 ))u00 (x0 )
∂x ∂x
88 CHAPTER 4. THE METHOD OF CHARACTERISTICS

x x

Figure 4.16: The incorrect evolution of the spatial steepening profile

∂x0 1
⇒ = .
∂x 1 + tc (u0 (x0 ))u00 (x0 )
0

On the other hand, if we take a t derivative:

∂x0 ∂x0
0= + c(u0 (x0 )) + tc0 (u0 (x0 ))u00 (x0 )
∂t ∂t
∂x0 −c(u0 (x0 ))
⇒ = .
∂x 1 + tc0 (u0 (x0 ))u00 (x0 )
In order for one of these expressions to be infinite, their denominator needs to be zero. They
have the same denominator, as you might expect. Thus we need

1 + tc0 (u0 (x0 ))u00 (x0 ) = 0

−1
⇒ t= .
c0 (u0 (x0 ))u00 (x0 )
We want to find the smallest such time, as long as it is positive. Thus
−1
tb = min ≥ 0.
x0 c0 (u0 (x0 ))u00 (x0 )
The x0 for which the minimum value tb is attained gives the characteristic along which the
gradient catastrophe will happen.
Example. Revisiting the example above, we have

c(u0 (x0 )) = u0 (x0 ) ⇒ c0 = 1,

4.5. SHOCK WAVES 89

and
2
u0 (x) = e−x .
The expression for the breaking time becomes
−1
tb = min 2
x0 −2x0 e−x0
1
= min 2.
x0 2x0 e−x0

Thus, we can find this minimum by maximizing the function

2
F (x) = 2xe−x .

We have
2 2
F 0 = 2e−x − 4x2 e−x
2
= 2e−x (1 − 2x2 ),

which is zero for

±
x= √ .
2
Further, √ √ √ √
F (1/ 2) = 2e−1/2 > 0, F (−1/ 2) = − 2e−1/2 < 0.
We discard the second possibility, since it leads to a negative breaking time. Thus
r
e
tb = ≈ 1.16,
2
which occurs along the characteristic that starts at
1
x0 = √ ≈ 0.707,
2
which is in excellent agreement with Fig. 4.12.
You should use this method to find the breaking time in the example with u0 (x) =
− arctan(x).

4.5 Shock waves

When we derived the differential form of the conservation law, we assumed that our functions
had continuous derivatives. If they do not, we have to work with the integral form. In other
words, once a gradient catastrophe happens and the derivatives of the solution → ∞, the
differential form of the equation ceases to be valid and we turn to the integral form. In this
section, having determined the breaking time, we figure out how to move beyond it.
90 CHAPTER 4. THE METHOD OF CHARACTERISTICS

Again, let us consider

ut + c(u)ux = 0,
u(x, 0) = u0 (x),
with implicit solution given by
u(x, t) = u0 (x0 ),
x = x0 + tc(u0 (x0 )).
As discussed, this solution is valid as long as the derivatives ux and ut are finite, i.e., up
until t = tb . For t > tb , we have to return to the integral form, which is
d b
Z
u(x, t)dx = φ(a, t) − φ(b, t),
dt a
where we have equated f ≡ 0, and
Z
φx = c(u)ux ⇒ φ = c(u)du.

Once the derivative becomes infinite, we should expect the solution to develop a discontinuity,
known as a shock. In other words, u(x, t) will have different values u− (before) and u+ (after)
the shock, where the derivative is vertical. At either side of the shock, the solution satisfies
the PDE, but the PDE cannot capture the shock itself.
Suppose we have a wedge-like region of the (x, t) plane where the characteristics cross,
as in Fig. 4.17. The idea is to insert a path in the (x, t) plane along which the shock will
propagate. Up until the shock path, our previous solution is valid, and we continue to follow
the characteristics, as before, until they hit the shock. Such a shock path x = xs (t) is
inserted in red in Fig. 4.17. Once we have inserted the shock path, we can merrily continue
the characteristics until the hit the shock, even into the wedge region!
So, how do we find how the shock moves? The integral form is given by
d b
Z
u(x, t)dx = φ(a, t) − φ(b, t).
dt a
Suppose a shock exists at x = xs (t), in between x = a and x = b. Then
Z xs Z b 
d
u(x, t)dx + u(x, t)dx = φ(a, t) − φ(b, t).
dt a xs

Now we use the Leibniz rule7 , which is really just a big ol’ chainrule. Since xs depends on t,
we get
Z xs Z b
− dxs dxs
ut (x, t)dx + u(xs , t) + ut (x, t)dx − u(x+
s , t) = φ(a, t) − φ(b, t),
a dt xs dt
R b(t) R b(t)
7
Recall, d
dt a(t)
f (x, t)dx = a(t)
ft (x, t)dx + f (b(t), t)b0 (t) − f (a(t), t)a0 (t)
4.5. SHOCK WAVES 91

t
x = xs (t)

tb

x0b xb

Figure 4.17: The wedge region where characteristics cross, with a shock curve inserted.

where

u− := u(x−
s , t) = lim− u(x, t),
x→xs
+
u := u(x+
s , t) = lim+ u(x, t),
x→xs

are the limits as x approaches xs from the left and right, respectively.
Next, since a and b are completely arbitrary in this process, we now let a → x− s and
+
b → xs . This eliminates the integrals above, since they are integrals of bounded functions
over an interval that shrinks to zero. We are left with

u− x0s − u+ x0s = φ− − φ+
dxs φ− − φ+ ∆φ
⇒ = − +
:= .
dt u −u ∆u
This is known as the Rankine-Hugoniot condition. It dictates the speed at which the
shock moves. Note that for the PDE

ut + uux = 0,

we have that φ = u2 /2, so that

1 − 2 2
φ− − φ+ u − u+ 1
x0s = − +
= 2
− +
= (u− + u+ ),
u −u u −u 2
92 CHAPTER 4. THE METHOD OF CHARACTERISTICS

Figure 4.18: The characteristics before a shock path is inserted.

and the shock speed is simply the average value of the solution to the left and right of it.
Let’s see how the Rankine-Hugoniot condition works.
Example. Consider the problem

ut + uux = 0,

1, x ≤ 0,
u(x, 0) = u0 (x) =
0, x > 0.
The implicit solution is given by

u(x, t) = u0 (x0 ),
x = x0 + tu0 (x0 ).

For x0 > 0, this becomes u = 0 and x = x0 , while for x0 ≤ 0, we get u = 1 and x = x0 + t.

These characteristics are drawn in Fig. 4.18.
We see there is a triangular region where the characteristics cross. We wish to insert a
shock path there, starting at (x, t) = (0, 0). According to Rankine-Hugoniot,
dxs ∆φ
= .
dt ∆u
Here φ = u2 /2. Further, on the left-side of the shock, we will have u− = 1, thus φ− = 1/2.
On the right side of the shock, u+ = 0, so that φ+ = 0. Thus
dxs 1 1
= ⇒ xs = t + α,
dt 2 2
4.5. SHOCK WAVES 93

u=1 u=0

Figure 4.19: The characteristics with the shock path is inserted.

where α is an integration constant. Since xs = 0 at t = 0, we see that α = 0. Thus the

shock path is given by
t
xs = .
2
We insert this path in the (x, t) plane, which results in Fig. 4.19, where we have continued
the characteristics up to the shock line, but not beyond. The solution for our problem is
given by 
1, x < t/2,
u(x, t) =
0, x > t/2.

Example. We consider a second example, from traffic flow. Recall that we the traffic
flow model is governed by
ut + φx = 0,
with  
u
φ = v1 u 1 − .
u1
Let’s assume that the speed limit v1 = 45 miles/hour. Similarly, we work with a maximal
density of cars of u1 = 300 cars/mile. Thus

u2
 
φ = 45 u − .
300

As an initial condition, we have a traffic moving to the right with a velocity of v =

30miles/hour, running into a traffic jam at location x = 0. This is plotted in Fig. 4.20.
94 CHAPTER 4. THE METHOD OF CHARACTERISTICS

u0 (x)

300 (stopped)

v = 30miles/hour, density= 100 cars/mile

x
0

Figure 4.20: The initial condition for the traffic jam problem.

Indeed, if v = 30miles/hour, then

 
u
v = 30 = v1 1− ⇒ u = 100.
u1
Thus, 
100, x < 0,
u0 (x) =
300, x ≥ 0.
The implicit solution is given by
x = x0 + tc(u0 (x0 )),
u = u0 (x0 ).
Here
0 u 

c(u) = φ (u) = 45 1 − .
150
The characteristics and the solution are different, depending on whether x0 < 0 or x0 > 0.
ˆ For x0 < 0, u0 (x0 ) = 100, so that c(u0 (x0 )) = c(100) = 15. Thus the characteristics
are given by
x = x0 + 15t,
i.e., lines of slope 1/15 in the (x, t) plane. Along these lines, the value of the solution
is u = 100.
ˆ For x0 > 0, u0 (x0 ) = 300, so that c(u0 (x0 )) = −45. Thus the characteristics are
x = x0 − 45t,
lines of slope −1/45 in the (x, t) plane. Along these lines, the value of the solution is
u = 300.
4.6. SHOCK WAVES AND THE VISCOSITY METHOD 95

Figure 4.21: The characteristics for the traffic jam problem, without the insertion of a shock
path.

Once again, there is a (very large) triangular region where characteristics cross. We use
the Rankine-Hugoniot condition, because it’s what the cool kids do. We have
dxs ∆φ
= .
dt ∆u
On the left-side of the shock, we will have u− = 100, thus φ− = 3000. On the right side of
the shock, u+ = 300, so that φ+ = 0. Thus
3000
x0s = = −15 ⇒ xs = −15t,
−200
where we have used that the shock starts at xs = 0 at t = 0. The characteristics with this
shock line inserted are shown in Fig. 4.22.
Finally, our solution is given by

300, x > −15t,
u(x, t) =
100, x < −15t.

Furthermore, we learn that the traffic jam backs up at a rate of 15 miles/hour.

4.6 Shock waves and the viscosity method

So far, we have calculated where and when shocks form, and when they form, we have
examined how they move, using the Rankine-Hugoniot condition.
96 CHAPTER 4. THE METHOD OF CHARACTERISTICS

u = 300

u = 100

Figure 4.22: The characteristics for the traffic jam problem, with the insertion of a shock
path.

In real applications, a shock never forms. Instead, we get a solution with a very steep,
but not vertical profile. The presence of extra physical effects precludes the formation of
actual shocks.
As an example, we consider once again traffic flow. How shall we update our model to
take into account extra effects which presumably will prevent shock formation?
We have
ut + φx = 0.
So far, we have used φ = uv, with
 
u
v = v1 1 −
u1

as our velocity profile. We have assumed that drivers adjuct their speed based on the car
density they observe where they are. Let’s give drivers a bit more credit8 . Let’s assume that
drivers can adjust their speed based on what they observe ahead of them.
For instance, suppose that drivers observe that the traffic is becoming more dense. In
other words, ux > 0. It would seem natural that they would decrease their speed. Further,
if u is very small, any chance appears huge and might have a big effect. Similarly, if u
is large, and change is not very important, this it is the relative change that matters: we
wish to modify our equation for v as a function of u with a term −rux /u. Here ux /u is the
relative density change, and r is a positive proportionality constant. The − sign ensures that
the velocity increases as the density increases. Notice that the opposite happens if we see
8
You can let me know in 30 years whether we should or not.
4.6. SHOCK WAVES AND THE VISCOSITY METHOD 97

the density is lighter ahead: the drivers will start to speed up. Thus, our new velocity law
becomes  
u ux
v = v1 1 − −r .
u1 u
It follows that
u2
 
φ = uv = v1 u−− rux
u1
 
2u
⇒ φx v1 1 − ux − ruxx .
u1

Our PDE becomes

 
2u
ut + φx = 0 ⇒ ut + v1 1 − ux = ruxx .
u1

We impose the boundary condition

lim u(x, t) = u1 ,
x→∞
lim u(x, t) = u0 < u1 .
x→−∞

In other words, we have gridlock on the right, and some lower-density moving traffic moving
into the gridlock. In addition, we have

lim ux (x, t) = 0,
x→∞
lim u(x, t) = 0.
x→−∞

Traveling wave solutions

If we look for solutions of the form

u(x, t) = f (x − ct),

we get
 
0 2f
−cf + v1 1 − f 0 = rf 00
u1
v1
⇒ −cf + v1 f − f 2 = rf 0 + k.
u1
At this point, we can find c and k, using the boundary conditions. Evaluating the above at
+∞, we get
−cu1 + v1 u1 − v1 u1 − r · 0 + k ⇒ k = −cu1 .
98 CHAPTER 4. THE METHOD OF CHARACTERISTICS

u(x, t)

u = u1 = 300

c = −v1 u0 /u1

x
u = u0 = 100

Figure 4.23: The traveling-wave solution for the traffic flow problem with smart drivers.
Here u1 = 300, u0 = 100, v1 = 45. Different values of r are used, ranging from 20 to 5.

From an evaluation at −∞, we get

v1 u20 u0
−cu0 + v1 u0 − = r · 0 + k ⇒ c = −v1 .
u1 u1

Solving the ODE, we get9

u0 − u1
u(x, t) = u1 +  .
v1 (u1 −u0 )
1 + exp ru1
(x + v1 u0 t/u1 )

This solution is drawn in Fig. 4.23, for different values of r. Note that as r → 0,

u1 , x − ct > 0,
u(x, t) =
u0 , x − ct < 0.

Thus, the solution becomes steeper and steeper as r → 0, ultimately resulting in the shock
solution as r → 0.
Unfortunately, we cannot devote more time to the viscosity method at this point. This
example gives a flavor, but it also indicates that the way to introduce extra effects which
might arrest shock formation is very problem dependent.

4.7 Rarefaction waves

We have now seen how to follow the solution along characteristics as long as these don’t
cross. Next, we have learned what to do when they do cross: a shock is formed, and we
9
Come on, you know the drill. Check this. I’ll wait. Done already? OK. Let’s move on.
4.7. RAREFACTION WAVES 99

u(x, 0) = u0 (x)

x
0
0

Figure 4.24: The initial condition u0 (x) for the rarefaction example.

know how to propagate it. We know everything!

Example. Consider the problem

ut + uux = 0,

0, x < 0,
u(x, 0) = u0 (x) =
1, x > 0.

The initial condition is shown in Fig. 4.24.

Figure 4.25: The characteristics with the empty region.

100 CHAPTER 4. THE METHOD OF CHARACTERISTICS

u(x, 0) = u0 (x)

− x
0


Figure 4.26: A new, improved, smooth initial condition.

The characteristics are given by

x = x0 + u0 (x0 )t,

which give two different cases:

x0 < 0 : x = x0 ,
x0 > 0 : x = x0 + t.

These characteristics are shown in Fig. 4.25. We face the immediate question what the value
of the solution is in the red region without characteristics.
To answer this question, we revisit the same problem, but now with an initial condition
that is smooth as opposed to discontinuous. Such an initial condition is drawn in Fig. 4.26.
It is zero outside of |x| > , but it changes smoothly from zero to one in |x| < . The new
initial condition is constructed so that it limits to the discontinuous one, u0 (x), as  → 0.
This gives rise to a new characteristics picture, shown in Fig. 4.27. The characteristics
are given by
x − x0
t= ,
u0 (x0 )
and it is clear that for x0 between − and , the characteristics will start at x0 and have a
slope that varies smoothly from 0 for x0 = − to 1 for x0 = .
In the limit as  → 0, we get a fan of characteristics, all starting at (0, 0), but with slope
ranging from ∞ (on the left) to 1 on the right. This seems like a reasonable way to fill in the
empty region: we assume an initial profile with high steepness that is smooth, and we let
the steepness → ∞. The limiting characteristic plane is shown in Fig. 4.28. We call the red
region in Fig. 4.25 the rarefaction region and the characteristics that fill it the rarefaction
fan.
4.7. RAREFACTION WAVES 101

−  x

Figure 4.27: The characteristics corresponding to the smooth initial condition.

Now that we have determined the rarefaction characteristics, what is the solution u(x, t)
there? We know that u(x, t) should be constant along characteristics. Since these are of the
form
x = ωt,
it follows that
u(x, t) = u(ωt, t).
Since u(x, t) is supposed to be constant along these characteristics, it follows that this last
quantity has to be independent of t. Thus
u(x, t) = g(ω) = g(x/t).
Now we have to find the function g(ω). Using the PDE ut + uux = 0, we get
∂ω x
ut = g 0 (ω) = − 2 g 0 (ω),
∂t t
∂ω 1
ux = g 0 (ω) = g 0 (ω).
∂x t
Thus
x 1
− 2 g 0 (ω) + g(ω) g 0 (ω) = 0
t t
x 0
⇒ − g (ω) + g(ω)g 0 (ω) = 0
t
⇒ −ωg 0 (ω) + g(ω)g 0 (ω) = 0
⇒ g 0 (ω)(g(ω) − ω) = 0.
102 CHAPTER 4. THE METHOD OF CHARACTERISTICS

Figure 4.28: The characteristics with the rarefaction fan in place.

There are two possibilities:

1. g 0 (ω) = 0 ⇒ g(ω) =constant. This would mean that u(x, t) is still discontinuous,
leading to a shock. But the characteristics do not cross for t > 0. Thus this doesn’t
work. We may ignore this possibility.

2. g(x) = ω, so that
x
u(x, t) = ,
t

in the rarefaction region.

In summary, our rarefaction problem has the solution


 0, x < 0,
u(x, t) = x/t, 0 < x < t,
1, x > t.


A few time slices of the solution are shown in Fig. 4.29. The time slices make sense: due
to the fanning out of the characteristics in the rarefaction region, the values of u get spread
out over a larger region, as t increases.
4.8. RAREFACTION AND SHOCK WAVES COMBINED 103

u(x, 0)

x
0
u(x, 1)

x
0
u(x, 3)

x
0

Figure 4.29: Different time slices of the solution.

4.8 Rarefaction and shock waves combined

Consider the problem

ut + uux = 0,

1, x ∈ (0, 1),
u(x, 0) = u0 (x) =
0, x 6∈ (0, 1).

The initial condition u0 (x) for this problem is shown in Fig. 4.30. The characteristics are
given by
x = x0 + tu0 (x0 ),
or,

x0 ∈ (0, 1) : x = x0 ,
x0 6∈ (0, 1) : x = x0 + t.
104 CHAPTER 4. THE METHOD OF CHARACTERISTICS

u0 (x)

0 0 x
0

Figure 4.30: Different time slices of the solution.

The characteristics, as given above, are drawn in Fig. 4.31.

Regions A, B, and C present no problems. However, we see there are two immediate
problems, already at t = 0, in regions D and E. In D, no characteristics are present, and
a rarefaction fan is needed, while in E, a shock needs to be inserted. And we have no clue
what’s going to happen in region F! But, that’s a concern for another page. Right now, one
thing at a time: let’s make sure we can move beyond the initial condition, by solving our
problems at t = 0.

1. In region D, we insert characteristics x = ωt, ω ∈ (0, 1), as in the previous section.

Along these characteristics,
u(x, t) = g(ω).
Substitution in the PDE gives, as before, g(ω) = ω, so that
x
u(x, t) = ω = ,
t
in region D.

2. In region E, we use the Rankine-Hugoniot condition to insert a shock path. The shock
will receive the value u− = 1 from the left (the characteristics from region B) and the
value u+ = 0 from the right (the characteristics from region C). Thus
1 1 1
x0s = (u− + u+ ) = ⇒ xs = t + 1,
2 2 2
where the integration constant has been chosen to ensure that the shock line starts at
x0 = 1.

The characteristic plane with shock and fan inserted is shown in Fig. 4.32. We have
resolved all problems at t = 0 and we can move the initial condition forward in time. The
4.8. RAREFACTION AND SHOCK WAVES COMBINED 105

D B C
E
x

Figure 4.31: The characteristics, pre rarefaction, pre shock.

solution is perfectly well defined in regions A, B, c, D, and we have a beautiful shock moving
along shock path E. But, we’re not quite done yet! At (x, t) = (2, 2), the shock line xs
(in red) hits the first characteristic (in bold black) of the rarefaction fan in region B. This
implies that the left and right values we used in the Rankine-Hugoniot condition will have
to be altered. Past t = 2, the value of u+ (coming in from the right) remains at 0, following
the characteristics in region C. However, the value u− will no longer be 1, as now the values
come from the rarefaction fan. Thus
xs
u− = ,
t
where we have imposed that x = xs on the shock line. The problem determining our new
shock path is

dxs 1 xs
= (u− + u+ ) = ,
dt 2 2t
xs (2) = 2,

where the last equation simply state that the new shock path is a continuation of the old
one. We get

dxs 1 dt
=
xs 2 t
1
⇒ ln xs = ln t + c
2√
⇒ ln xs = ln Ct,
106 CHAPTER 4. THE METHOD OF CHARACTERISTICS

A
E
D
B C
x

Figure 4.32: The new and improved characteristics, now with shock and fan.

where c and C are constants, related by c = ln(C)/2. Applying the initial condition, we get
√
ln 2 = ln 2C ⇒ C = 2,

and our new shock path becomes

√ 1
xs = 2t ⇒ t = x2s .
2
This is an upward parabola starting at (x, t) = (2, 2). Note that at this point
dt
= xs = 2,
dxs
so that the shock line starts of tangent to our original shock path.
The complete characteristic plane is drawn in Fig. 4.33, where all problems have been
resolved. Region D contains a rarefaction fan, and we have a two-stage shock, following a
straight line along E, and a parabolic path along F.
There are six different solution regions.

1. In x < 0, t > 0 the solution is u = 0. This is region A in Fig. 4.33.

2. In t < x < t/2 + 1, t > 0, the solution is u = 1. This corresponds to region B in

Fig. 4.33.

3. If x > t/2 + 1, t ∈ (0, 2), then u = 0. This corresponds to the part of region C below
the red straight-line shock path in Fig. 4.33.
4.8. RAREFACTION AND SHOCK WAVES COMBINED 107

A E
D
B C x

Figure 4.33: The newer and more improved characteristics, now with complete shock and
fan.

4. If x ∈ (0, t), t ∈ (0, 2), then u = x/t. This corresponds to the lower part of region D
in Fig. 4.33, to the left of the black line.
√
5. Similarly, if x ∈ (0, 2t), t > 2, then u = x/t, corresponding to the upper part of
region D in Fig. 4.33, to the left of the blue path.
√
6. Lastly, if x > 2t and t > 2, then u = 2, corresponding to the part of region C, below
the blue path.

The solution profiles at t = 0, t =√ √ t = 3 are shown in Fig. 4.34. As t → ∞,

1, t = 2 and
we have that the shock speed x0s = ( 2t)0 = 1/ 2t → 0, and the shock slows down as we
proceed.
Note that

ut + uux = 0
Z ∞ Z ∞
d
⇒ udx + uux dx = 0
dt −∞ −∞
Z ∞ ∞
d 1
⇒ udx + u2 =0
dt −∞ 2 −∞
d ∞
Z
⇒ udx = 0
dt −∞
Z ∞
⇒ u(x, t)dx = 1.
−∞

The last line comes from the fact that the area underneath the solution is 1 initially, while
the line before it says that this area is conserved. We have also used that limx→±∞ u = 0. We
easily verify this: at t = 1, the area is the area of a trapezoid. It is equal to 1×(1/2+3/2)/2 =
1. Next, at t = 2, we need the √ p a triangle, equal to 2 × 1/2 = 1. Lastly, at t = 3, we
area of
still have a triangle, with area 6 × 2/3/2 = 1. It also follows from this argument that as
108 CHAPTER 4. THE METHOD OF CHARACTERISTICS

u=1

x0s = 1/2
t=0
u=0 u=0
x=0 x=1
u=1

u=x x02 = 1/2

t=1
u=0 u=0
x=0 x = 1 x = 3/2
u=1

u = x/2 x0s = 1/2

t=2
u=0 u=0
x=0 x=2
p
u= 2/3
√
t=3 u = x/3 x0s = 1/ 6
u=0 √u=0
x=0 x= 6

Figure 4.34: Three different solution profiles, at t = 0, t = 1, t = 2 and t = 3.

we march on, the top of the triangle is lowered, while its base grows so that its area remains
unchanged.

A warning
Let’s reconsider the example on page 92. We have

ut + uux = 0,

1, x < 0,
u(x, 0) =
0, x > 0.

Previously, we used φ = u2 /2. With u− = 1, u+ = 0, φ− = 1/2, φ+ = 0, the Rankine-

Hugoniot condition gives
1
x0s = ,
2
4.9. WEAK SOLUTION OF PDES 109

which we used to determine a perfectly fine shock. This condition is based on the conservation
of u.
However, consider the same problem, but written as

uut + u2 ux = 0,

1, x < 0,
u(x, 0) =
0, x > 0.
Now, the PDE is written
ρt + Φx = 0,
with ρ = u2 /2 and Φ = u3 /3. The Rankine-Hugoniot condition becomes
1 − 3 3
∆Φ u − 1 u+ 2
x0s = = 31 − 2 31 + 2 = ,
∆ρ 2
u − 2u 3

and we find a different, but still equally fine shock. Thus, the same PDE can give rise to
two (or many more) conclusions about the shock velocity. How do we choose? Which one is
correct?
It turns out the answer is we don’t choose. We are lowly mathematicians, who solve
problems given to us by our experimentalist friends10 . The application we’re dealing with
dictates which quantity is conserved, which implies how we should write the PDE. One
we know this, we can propagate the shock. But simply knowing the PDE is not enough
information to work with the shock, as the PDE breaks down once we have a shock.

4.9 Weak solution of PDEs

Classical solutions of conservation laws
When we are faced with solving a PDE with initial and boundary conditions, we usually
want to find a solution that satisfies the equation everywhere in our domain of interest.
This implies that everywhere in this domain of interest, the solution should have as many
continuous derivatives as appear in the equation. Such a solution is referred to as a classical
solution.
What we have been doing recently does not result in classical solutions: we have often had
solutions that are piecewise defined, and that have corners (i.e., discontinuous derivatives)
or even shocks (discontinuous function). So, apart from at a few points, such solutions
satisfy the equation, just not everywhere. And we know this. After all, we returned to the
integral form (the conservation law form) of the PDE to get such solutions. We refer to such
piecewise defined solutions as weak solutions.
The notion of a weak solution allows us to get away with just a bit more: weak solutions
(at least so far) satisfy the equation almost everywhere. In this lecture, we look at a different
10
Everyone should have at least one experimentalist friend. Get yours today!
110 CHAPTER 4. THE METHOD OF CHARACTERISTICS

Figure 4.35: The test function T (x).

way to characterize weak solutions of the PDE, through the so-called weak formulation of
the PDE.

The weak form of the conservation law. Test functions

Consider the function  2 /(1−x2 )
e−x , |x| < 1,
T (x) =
0, |x| ≥ 0.
This function is plotted in Fig. 4.35.
Let’s investigate this function at x = 1. Clearly, as x → 1 from the right, T (x) and all
of its derivatives are zero at x = 1. How about the left limits?
2
lim T (x) = e lim e−1/(1−x ) .
x<1 x<1

The denominator in the exponent is always positive and approaches zero. Thus −1/(1 −
x2 ) → −∞ and
lim T (x) = 0.
x<1

Next, we calculate the derivative of T (x) for |x| < 111 :

−2x −x2 /(1−x2 )
T 0 (x) = e .
(x2 − 1)
Taking the limit as x → 1 from the left, it is clear that the exponential will dominate and
once again,
lim T 0 (x) = 0.
x<1

The same argument holds for all derivatives: at all orders, we get the same exponential12
and some rational function of x. Thus
lim T (n) (x) = 0, for n ≥ 0.
x<1
11
No, this is not 111 . Rather, it’s the obligatory reminder that you should check this.
12
Exponentials do that, you know.
4.9. WEAK SOLUTION OF PDES 111

Remark. As a side note, this implies that the Taylor series of T (x) around x = 1 is
given by
∞
X T (n) (1)
T (x) = (x − 1)2 ≡ 0.
n=0
n!
Quite a weird result!
We call T (x) a test function or, sometimes, a Schwarz function. It is a function that is
infinitely differentiable but has compact support. This means that it is nonzero only in
a finite region. By shifting x and rescaling, we can control where the support of the test
function is located and how large it is.
Similarly, consider
 −(x2 +t2 )/(1−(x2 +t2 ))
e , x2 + t2 < 1,
T (x, t) =
0, x2 + y 2 ≥ 0.

This is a test function in the (x, t) plane. Basically, it looks like a hat with compact support.
Suppose we wish to solve

ut + φx = 0,
u(x, 0) = u0 (x),

for x ∈ R, t¿0. Let T (x, t) be a test function. Then T (x, t)u(x, t) is zero at every point
outside the circle x2 + y 2 = 1, so we have isolated a small part of u(x, t). Then

T (x, t)ut + T (x, t)φx = 0

Z ∞ Z ∞
⇒ (T (x, t)ut + T (x, t)φx ) dx dt = 0
0 −∞
Z ∞ Z ∞
⇒ dt dx (T (x, t)ut + T (x, t)φx ) = 0.
0 −∞

Here the last line is just a less confusing version of the previous line.
Now we integrate by parts13 .
Z ∞  Z ∞ 
t=∞
dx u(x, t)T (x, t)|t=0 − dtu(x, t)Tt (x, t) +
−∞ 0
Z ∞  Z ∞ 
x=∞
dt φ(x, t)T (x, t)|x=−∞ − dxφ(x, t)Tx (x, t) = 0
0 −∞
Z ∞  Z ∞ 
⇒ dx 0 − u0 (x)T (x, 0) − u(x, t)Tt (x, t) +
−∞ 0

13
Secret: integration by parts is the main weapon of a good applied mathematician.
112 CHAPTER 4. THE METHOD OF CHARACTERISTICS

Z ∞  Z ∞ 
dt 0 − dxφ(x, t)Tx (x, t) = 0
0 −∞
Z ∞ Z ∞ Z ∞
⇒ − u0 (x)T (x, 0) − dt dx (u(x, t)Tt (x, t) + φ(x, t)Tx (x, t)) = 0.
−∞ 0 −∞

This is known as the weak form of the PDE. It involves derivatives only of the test function,
and just the value of u and φ, which are defined everywhere, even for weak solutions. We
should remark that the weak form of the equation gives us back the PDE, provided that
the solutions are classical solutions. In that case, we can start from the weak form, use
integration by parts and get the PDE.
There are many reasons for wanting to work with the weak form of an equation. Here
are two.

1. Larger function spaces. The modern theory of PDEs likes to talk about solutions
is specific function spaces. Examples are C(R), the space of continuous functions,
defined for all x ∈ R, C 1 (R), the space of differentiable functions on R, etc. Clearly,
C 1 (R) ⊂ C(R), and so on. In general, we like to solve PDEs in the largest function
space possible: a larger space implies fewer conditions on the solution, so the solution
is more general. Since the weak form of the PDE has fewer conditions on the solution
u(x, t), by only needing its value, and not its derivatives, it allows us to work in larger
function spaces.

2. Numerical solutions. An important example here is the finite element method:

we approximate functions using basis elements, like piecewise constant functions, or
piecewise linear functions, etc. Thus, we’re giving up on classical solutions when we
do this. This is not a problem at all for the weak formulation. Thus finite element
methods are applied to the weak formulation, not to the PDE itself.
CHAPTER 5 TREFETHEN 1994 191

Chapter 5.
Dissipation, Dispersion, and Group Velocity

5.1. Dispersion relations

5.2. Dissipation
5.3. Dispersion and group velocity
5.4. Modied equations
5.5. Stability in `p norms
5.6. Notes and references

Things fall apart the center cannot hold

Mere anarchy is loosed upon the world.
| W. B. Y EATS, The Second Coming (190{)
CHAPTER 5 TREFETHEN 1994 192

It would be a ne thing if discrete models calculated solutions to partial

di erential equations exactly, but of course they do not. In fact in general they
could not, even in principle, since the solution depends on an innite amount
of initial data. Instead, the best we can hope for is that the errors introduced
by discretization will be small when those initial data are reasonably well-
behaved.
This chapter is devoted to understanding the behavior of numerical errors.
From truncation analysis we may have a bound on the magnitude of discretiza-
tion errors, depending on the step sizes h and k, but much more can be said,
for the behavior of discretization errors exhibits great regularity, which can be
quantied by the notions of numerical dissipation and dispersion. Rounding
errors too, though introduced essentially at random, propagate in the same
predictable ways.
So long as we can estimate the magnitude of the discretization and round-
ing errors, what is the point in trying to investigate their behavior in more
detail? There are several answers to this question. One is that it is a good idea
to train the eye: a practitioner familiar with articial numerical e ects is less
likely to mistake spurious features of a numerical solution for mathematical or
physical reality. Another is that in certain situations it may be advantageous
to design schemes with special properties|low dissipation, for example, or
low dispersion. A third is that in more complicated circumstances, the mag-
nitude of global errors may depend on the behavior of local errors in ways
that ordinary analysis of discretization and rounding errors cannot predict. In
particular, we shall see in the next chapter that the stability of boundary con-
ditions for hyperbolic partial di erential equations depends upon phenomena
of numerical dispersion.
One might say that this chapter is built around an irony: nite di erence
approximations have a more complicated \physics" than the equations they are
designed to simulate. The irony is no paradox, however, for nite di erences
are used not because the numbers they generate have simple properties, but
because those numbers are simple to compute.
5.1. DISPERSION RELATIONS TREFETHEN 1994 193

5.1. Dispersion relations

Any time-dependent scalar, linear partial di erential equation with con-
stant coecients on an unbounded space domain admits plane wave solutions
u(xt) = ei(x+!t)  2 R (5:1:1)
where  is the wave number and ! is the frequency. (Vector di erential
equations admit similar modes multiplied by a constant vector the extension
to multiple space dimensions is described at the end of this section.) For each
 , not all values of ! can be taken in (5.1.1). Instead, the PDE imposes a
relationship between  and !,
! = !( ) (5:1:2)
which is known as the dispersion relation, mentioned already in x3.1. In
general each wave number  corresponds to m frequencies !, where m is the
order of the di erential equation with respect to t, and that is why (5.1.2) is
called a relation rather than a function. For most purposes it is appropriate
to restrict attention to values of  that are real, in which case ! may be real or
complex, depending on the PDE. The wave (5.1.1) decays as t ! 1 if Im ! > 0,
and grows if Im ! < 0.
For example, here again are the dispersion relations for the model equa-
tions of x3.1, and also for the second-order wave equation:
ut = ux : ! =  (5:1:3)
utt = uxx : !2 =  2 i.e., ! =  (5:1:4)
ut = uxx : i! = ; 2 (5:1:5)
ut = iuxx : ! = ; 2: (5:1:6)
These relations are plotted in Figure 5.1.1. Notice the double-valuedness of
the dispersion relation for utt = uxx, and the dashed curve indicating complex
values for ut = uxx.
More general solutions to these partial di erential equations can be ob-
tained by superimposing plane waves (5.1.1), so long as each component sat-
ises the dispersion relation the mathematics behind such Fourier synthesis
5.1. DISPERSION RELATIONS TREFETHEN 1994 194

! !

(a) ut = ux (b) utt = uxx

i! !

(c) ut = uxx (d) ut = iuxx

Figure 5.1.1. Dispersion relations for four model partial di erential
equations. The dashed curve in (c) is a reminder that ! is complex.
5.1. DISPERSION RELATIONS TREFETHEN 1994 195

was described in Chapter 2, and examples were given in x3.1. For a PDE of
rst order in t, the result is
Z 1
u(xt) = 2 1 i(x+! ( )t)
;1 e u^0( ) d: (5:1:7)
Since most partial di erential equations of practical importance have variable
coecients, nonlinearity, or boundary conditions, it is rare that this integral
representation is exactly applicable, but it may still provide insight into local
behavior.
Discrete approximations to di erential equations also admit plane wave
solutions (5.1.1), at least if the grid is uniform, and so they too have dispersion
relations. To begin with, let us discretize in x only so as to obtain a semidis-
crete formula. Here are the dispersion relations for the standard centered
semidiscretizations of (5.1.3){(5.1.6):
ut = 0u : ! = h1 sin h (5:1:8)
utt = u : !2 = h42 sin2 h 2 (5:1:9)
ut = u : i! = ; h42 sin2 h 2 (5:1:10)
ut = i u : ! = ; h42 sin2 h 2
: (5:1:11)
These formulas are obtained by substituting (5.1.1) into the nite di erence
formulas with x = xj . In keeping with the results of x2.2, each dispersion
relation is 2=h-periodic in  , and it is natural to take  2 ;=h=h] as
a fundamental domain. The dispersion relations are plotted in Figure 5.1.2,
superimposed upon dotted curves from Figure 5.1.1 for comparison.
Stop for a moment to compare the continuous and semidiscrete curves in
Figure 5.1.2. In each case the semidiscrete dispersion relation is an accurate
approximation when  is small, which corresponds to many grid points per
wavelength. (The number of points per spatial wavelength for the wave (5.1.1)
is 2=h.) In general, the dispersion relation for a partial di erential equation
is a polynomial relation between  and !, while a discrete model amounts to
a trigonometric approximation. Although other design principles are possible,
the standard discrete approximations are chosen so that the trigonometric
function matches the polynomial to as high a degree as possible at the origin
 = ! = 0. To illustrate this idea, Figure 5.1.3 plots dispersion relations for the
standard semidiscrete nite di erence approximations to ut = ux and ut = iuxx
of orders 2, 4, and 6. The formulas were given in x3.3.
5.1. DISPERSION RELATIONS TREFETHEN 1994 196

! !

(a) ut = 0 u (b) utt = u

i! !

(c)ut = u (d) ut = i  u

Figure 5.1.2. Dispersion relations for centered semidiscrete approx-

imations to the four model partial di erential equations. Each func-
tion is 2=h-periodic in   the plots show the fundamental domain
 2 ;=h=h].
5.1. DISPERSION RELATIONS TREFETHEN 1994 197

6
4
2

ut = ux
(a) (b) ut = iuxx

Figure 5.1.3. Dispersion relations for semidiscrete centered di er-

ence approximations to ut = ux and ut = iuxx of orders 2, 4, 6.

Now let us turn to fully discrete nite di erence formulas: discrete in time
as well as space. The possibilities become numerous. For example, substituting
the plane wave
vjn = ei(xj +!tn) = ei(jh+!nk)
into the leap frog approximation of ut = ux yields the dispersion relation
ei!k ; e;i!k = (eix ; e;ix )
where = k=h, that is,
sin !k = sin h: (5:1:12)
Similarly, the Crank-Nicolson approximation of ut = uxx has the dispersion
relation
ei!k ; 1 = (e 2 +1) eih ; 2+ eih 
i!k h i

which reduces to
i tan !k
2 = ;2 sin2 :
h
2 (5:1:13)
 5.1. DISPERSION RELATIONS TREFETHEN 1994 198

These dispersion relations are 2=h -periodic in  and 2=k -periodic in !.

With the use of inverse trigonometric functions it is possible to solve such
equations for !, so as to exhibit the functional dependence explicitly, but the
resulting formulas are less appealing and often harder to work with. Equations
like (5.1.12) and (5.1.13) have a certain elegance|one sees at a glance that
the time and space derivatives have been replaced by trigonometric analogs.
Tables 5.1.1 and 5.1.2* consider once again the various nite di erence
approximations to ut = ux and ut = uxx that appeared in Tables 3.2.1/3.2.2
and 4.4.1/4.4.2. In each case the dispersion relation is both listed and plotted.
Since h and k are independent parameters, there is now a range of possible
plots we have arbitrarily taken = k=h = 0:5 in the rst table and = k=h2 =
0:5 in the second. That is why each plot occupies a rectangle of aspect ratio 2
in wave number space.y Notice that the multistep (leap frog) formulas contain
two branches of ! values.
For partial di erential equations in several space dimensions, the notion of
dispersion relation generalizes just as x2.6 would suggest: a plane wave takes
the form
u(xt) = ei(x+!t)  2 R  (5:1:14)
where  and x are vectors, and (5.1.2) becomes a scalar function (or relation)
of a vector argument. For example, the wave equation
utt = uxx + uyy
has the dispersion relation
!2 = 2 + 2 (5:1:15)
if the vector  is written ( ), so that the lines of constant ! in the ( )
plane are concentric circles. On the other hand the leap frog approximation
vijn+1 ; 2vijn + vijn;1 = 2(vin+1j + vin;1j + vij
n
+1 + vij ;1 ; 4vij )
n n

appropriate to a uniform grid with h = !x = !y, has the dispersion relation

sin2 !k = 2 sin2 h +sin2 h 
h i
(5:1:16)
2 2 2
which is plotted in Figure 5.1.4 for  0. Once again the dispersion relation
is accurate for small wave numbers but diverges dramatically elsewhere.
EXERCISES
. 5.1.1. What are the coecients as trigonometric functions of the dispersion relations plotted
in Figure 5.1.3?
. 5.1.2. Sketch the dispersion relation for the leap frog model of ut = ux with  > 1|say,
 = 2. How is the instability of this nite dierence formula reected in your sketch?
*Not yet written.
y analogous to a Brillouin zone in crystallography (C. Kittel, Introduction to Solid State Physics,

Wiley, 1976).
5.2. DISSIPATION TREFETHEN 1994 199

BEx = Backward Euler ;i(1 ; e;i!k ) = sin h

CNx = Crank-Nicolson 2 tan !k2 = sin h

LF = Leap Frog sin !k = sin h

BOXx = Box tan !k2 = tan h2

LF4 = 4th-order Leap Frog sin !k = 34 sin h ; 61 sin 2h

LXF = Lax-Friedrichs ei!k = cos h + i sin h

UW = Upwind ei!k ; 1 = (eih ; 1)

LW = Lax-Wendro ;i(ei!k ; 1) = sin h +2i 2 sin2 h2

Table 5.1.1. Dispersion relations for various nite di erence approx-

imations to ut = ux with = k=h = 0:5. See Tables 3.2.1 and 4.4.1.
The dashed lines indicate the slope d!=d at isolated points where !
is real.
 5.2. DISSIPATION TREFETHEN 1994 200

Figure 5.1.4. Dispersion relation for the leap frog model of utt =
uxx + uyy in the limit ! 0. The region shown is the domain ;=h
=h]2 of the ( ) plane. The concentric curves are lines of constant
! for !h = 41  21 ::: 114 .

5.2. Dissipation
Even though a partial dierential equation may conserve energy in the L2 norm, its
nite dierence models will often lose energy as t increases, especially in the wave numbers
comparable to the grid size. This property is numerical dissipation, and it is often advan-
tageous, since it tends to combat instability and unwanted oscillations. In fact, articial
dissipation* is often added to otherwise nondissipative formulas to achieve those ends. An
example of this kind appeared in Exercise 3.2.1.
To make the matter quantitative, suppose we have a linear partial dierential equation
or nite dierence approximation that admits waves (5.1.1) with ! given by a dispersion
relation (5.1.2). Since  is assumed to be real, it follows that the wave has absolute value
jei(x+!t) j = e;t Im ! (5:2:1)
*In computational uid dynamics one encounters the more dramatic term articial viscosity.
 5.2. DISSIPATION TREFETHEN 1994 201

as a function of t, and thus decays exponentially if Im ! > 0. By Parseval's equality, the L2

norm of a superposition of waves (5.1.1) is determined by a superposition of such factors:
1 
u( t) = e;t Im !() u^0 ( ): (5:2:2)
k
2
k

As an extreme case the heat equation ut = uxx, with dispersion relation ! = i 2 , dissipates
nonzero wave numbers strongly|and indeed it does nothing else at time t, only wave
p
numbers  = O( t ) remain with close to their initial amplitudes. But it is principally the
dissipation introduced by nite dierence formulas that we are concerned with here.
The following denitions are standard:
A nite dierence formula is nondissipative if Im ! = 0 for all  . It is dissipative if
Im ! > 0 for all  = 0. It is dissipative of order 2r if ! satises
6

Im !k 2r ei !k j  1 ; 2 (h)2r
 1 (h)  i.e., j (5:2:3)
for some constants j > 0. In each of these statements,  varies over the interval
 =h=h], and ! represents all possible values ! corresponding to a given  . For
;

problems in multiple space dimensions, (h) is replaced by h in any norm.

k k

For example, the leap frog and Crank-Nicolson models of ut = ux are nondissipative,
while the upwind and Lax-Wendro formulas are dissipative.
Dissipative and nondissipative are mutually exclusive, but not exhaustive: a nite
dierence formula can be neither dissipative nor nondissipative. See Exercise 5.2.1.
According to the denition, no consistent nite dierence approximation of ut = ux + u
could be dissipative, but customary usage would probably use the term dissipative sometimes
for such a problem anyway. One could modify the denition to account for this by including
a term O(k) in (5.2.3).
A more serious problem with these standard denitions arises in the case of multistep
formulas, for which each  corresponds to several values of !. In such cases the denition
of dissipative, for example, should be interpreted as requiring Im ! > 0 for every value of !
that corresponds to some  = 0. The diculty arises because for multistep formulas, that
6

condition ensures only that the formula dissipates oscillations in space, not in time. For
example, the leap frog model becomes dissipative if a small term such as k  vn is added
to it, according to our denitions, yet the resulting formula still admits the wave (5.1.1)
with  = 0, ! = =h, which is sawtoothed in time. To exclude possibilities of that kind it is
sometimes desirable to use a stronger denition:
A nite dierence formula is totally dissipative if it is dissipative and in addition,
Im ! = 0 implies ! = 0.

EXERCISES
. 5.2.1. Determine whether each of the following models of ut = ux is nondissipative, dissi-
pative, or neither. If it is dissipative, determine the order of dissipativity.
(a) Lax-Wendro, (b) Backward Euler, (c) Fourth-order leap frog, (d) Box, (e) Upwind.
5.3. DISPERSION AND GROUP VELOCITY TREFETHEN 1994 202

5.3. Dispersion and group velocity

This section is not yet properly written. The next few pages contain a few remarks, followed by
an extract from my paper \Dispersion, dissipation, and stability."]
The general idea. Whereas dissipation leads to decay of a wave form, dispersion leads
to its gradual separation into a train of oscillations. This phenomenon is a good deal less
obvious intuitively, for it depends upon constructive and destructive interference of Fourier
components. It is of central importance in nite dierence modeling, because although many
partial dierential equations are nondispersive, their discrete approximations are almost
invariably dispersive. (Spectral methods are an exception.)
Caution. Phase and group velocity analysis depend upon the problem being linear and
nondissipative|i.e., ! must be real when  is real. (However, similar predictions hold if
there is a suciently slight amount of dissipation.)
Phase velocity. Suppose that a PDE or nite dierence formula admits a solution
ei(!t+x) . It's then a triviality to see that any individual \wave crest" of this wave, i.e., a
point moving in such a way that the quantity inside the parentheses has a constant value
(the phase), moves at the velocity
!
Phase velocity: c(! ) = ; :

(5:3:1)
Group velocity. However, early in the twentieth century it was realized that wave
energy propagates at a dierent velocity,
d!
Group velocity: cg (! ) = ;
d
: (5:3:2)
The algebraic meaning of this expression is that we dierentiate the dispersion relation with
respect to  . (In a plot in  {! space, c is minus the slope of the line through (!) and
the origin, and cg is minus the slope of the line tangent to the dispersion relation at (!).)
The physical meaning is that, for example, a wave packet|a smooth envelope times an
oscillatory carrier wave with parameters (!)|will move approximately at the velocity cg .
The same goes for a wave front and for any other signal that can carry information.
Dispersive vs. nondispersive systems. If the dispersion relation is linear, i.e., ! =
const  , then (5.3.1) and (5.3.2) are equal and the system is nondispersive. If the dispersion
relation is nonlinear, the system is dispersive. Finite dierence formulas are almost always
dispersive, since their dispersion relations are periodic and therefore nonlinear. (However,
see Exercise 5.3.2.)
Precise meaning of group velocity. The meaning of group velocity can be made precise
in various asymptotic fashions, for example by considering the limit t . The mathemat-
!1

ics behind this is usually a stationary phase or steepest descent argument.

Simple explanations of group velocity. There are a number of intuitive ways to under-
stand where the derivative (5.3.2) comes from. One is to superimpose two waves with nearby
5.3. DISPERSION AND GROUP VELOCITY TREFETHEN 1994 203

parameters (1 !1 ) and (2 !2 ). It is then readily seen that the superposition consists of a
smooth envelope times a carrier wave at frequency 21 (!1 + !2) and wave number 21 (1 + 2 ),
and the envelope moves at velocity (!2 !1)=(2 1 ), which approaches (5.3.2) in the
; ; ;

limit 2 1 , !2 !1 . Another approach is to take a pure exponential ei(!t+x) , with 

! !

and ! real, and change  slightly to a complex value  + i  \in order to visualize which
way the envelope is moving." If the eect on ! is to make it change to ! + i !, it is readily
calculated that the resulting evanescent wave has an envelope that moves laterally at the
velocity !=  , and this again approaches (5.3.2) in the limit  0, ! 0. A third,
; ! !

more \PDE-style" explanation is based upon advection of local wave number according to
a simple hyperbolic equation with coecient cg  see Lighthill.
Group velocity in multiple dimensions. If there are several space dimensions, the group
velocity becomes the gradient of ! with respect to the vector  , i.e., cg =  !.
;r

Phase and group velocities on a grid. There is another sense in which group velocity
has more physical meaning than phase velocity on a nite dierence grid: the former is
well-dened, but the latter is not. On a periodic grid, any Fourier mode can be represented
in terms of innitely many possible choices of  and ! that are indistinguishable physically,
and according to (5.3.1), each choice gives a dierent phase velocity. What's going on here
is that naturally one can't tell how fast a pure complex exponential wave is moving if one
sees it at only intermittent points in space or time, for one wave crest is indistinguishable
from another. By contrast, the group velocity is well-dened, since it depends only on the
slope, which is a local property formula (5.3.2) has the same periodicity as the dispersion
relation itself.
Computation of a group velocity on a grid. To compute the group velocity for a
nite dierence formula, dierentiate the dispersion relation implicitly and then solve for
cg = d!=d . For example, the wave equation ut = ux has c = cg = 1 for all  . For the leap
; ;

frog approximation the dispersion relation is sin !k =  sin ! , which implies k cos !k d! =
h cos ! d , and since k = h, cg (! ) = cos h= cos !k .
;

Parasitic waves. Many nite dierence formulas admit parasitic waves as solutions, i.e.,
waves that are sawtoothed with respect to space or time. These correspond to  = =h, 

! = =k , or both. It is common for such waves to have group velocities opposite in sign


to what is correct physically. In the example of the leap frog formula, all four parasitic
modes 1, ( 1)j , ( 1)n , and ( 1)j+n are possible, with group velocities 1, 1, 1, and 1,
; ; ; ; ;

respectively.
Spurious wiggles near interfaces and boundaries. It is common to observe spurious
wiggles in a nite dierence calculation, and they appear most often near boundaries, in-
terfaces, or discontinuities in the solution itself. The explanation of where they appear is
usually a matter of group velocity. Typically a smooth wave has passed through the discon-
tinuity and generated a small reected wave of parasitic form, which propagates backwards
into the domain because its group velocity has the wrong sign. More on this in the next
chapter.
Waves in crystals. Dispersion relations for vibrations in crystals are also periodic with
respect to  . As a result, sound waves in crystals exhibit dispersive eects much like those
associated with nite dierence formulas, including the existence of positive and negative
group velocities.
5.3. DISPERSION AND GROUP VELOCITY TREFETHEN 1994 204

Figure 5.3.1. Dispersion under the leap frog model of ut = ux with  = 0:5. The
lower mesh is twice as ne as the upper.
 5.3. DISPERSION AND GROUP VELOCITY TREFETHEN 1994 205

EXERCISES
. 5.3.1. The Box formula.
(a) Write out the BOXx formula of Table 3.2.1 in terms of vjn , vjn+1 , etc.
(b) Determine the dispersion relation (expressed in as simple a form as possible).
(c) Sketch the dispersion relation.
(d) Determine the group velocity as a function of  and !.
. 5.3.2. Schrodinger equation.
(a) Calculate and plot the dispersion relation for the Crank-Nicolson model of ut = iuxx of
Exercise 3.2.1(f).
(b) Calculate the group velocity. How does it compare to the group velocity for the equation
ut = iuxx itself?
. 5.3.3. A paradox. Find the resolution of the following apparent paradox, and be precise in
stating where the mistake is. Draw a sketch of an appropriate dispersion relation to explain
your answer.
One the one hand, if we solve ut = ux by the leap frog formula with  = 1, the results will
be exact, and in particular, no dispersion will take place.
On the other hand, as discussed above, the dispersion relation on any discrete grid must be
periodic, hence nonlinear|and so dispersion must take place after all.
Chapter 2

Waves, dispersion and dissipation

The main reference for this chapter is §1.1 of the book [Drazin and Johnson, 1989].

2.1 Dispersion
Usually, localised waves spread out (“disperse”) as they travel. This prevents them from
being solitons. Let’s understand this phenomenon first.

EXAMPLES:

1. ADVECTION EQUATION (linear, 1st order):

1
ut ` ux “ 0 (2.1)
v

ÝÑ Solution
upx, tq “ f px ´ v tq for any function f ,
i.e. a wave moving with velocity v (right-moving if v ą 0, left-moving if v ă 0). The
wave keeps a fixed profile f pξ q and moves rigidly at velocity v (indeed ξ “ x ´ v t):

12
CHAPTER 2. WAVES, DISPERSION AND DISSIPATION 13

So in this case there is no dispersion, but nothing else happens either.

2. “THE” WAVE EQUATION or D’ALEMBERT EQUATION (linear, 2nd order):

1
utt ´ uxx “ 0 pv ą 0 wlogq (2.2)
v2
ÝÑ Solution

upx, tq “ f px ´ v tq ` g px ` v tq for any functions f , g ,

i.e. the superposition of a right-moving and a left-moving wave with velocities ˘v :

All waves move at the same speed, so there is no dispersion, but there is no interaction
either, so this is also not very interesting for our purposes.

3. KLEIN-GORDON EQUATION1 (linear, 2nd order):

1
utt ´ uxx ` m2 u “ 0 , (2.3)
v2
where we take v ą 0 wlog.
This is a more interesting equation. Let us try a complex “plane wave” solution2

upx, tq “ eipkx´ωtq . (2.4)

Substituting the plane wave (2.4) in the Klein-Gordon equation (2.3), we find:

ω 2 ipkx´ωtq
´ e ` k 2 eipkx´ωtq ` m2 eipkx´ωtq “ 0
v2
ω2
ùñ ´ 2 ` k 2 ` m2 “ 0 .
v
1
This is the first relativistic wave equation (with v the speed of light). It was introduced independently by
Oskar Klein [Klein, 1926] and Walter Gordon [Gordon, 1926], who hoped that their equation would describe
electrons. It doesn’t, but it describes massive elementary particles without spin, like the pion or the Higgs boson.
2
This is called a “plane wave” because its three-dimensional analogue up⃗x, tq “ exprip⃗k ¨ ⃗x ´ ω tqs has constant
u along a plane ⃗k ¨ ⃗x “ const at fixed t. Unless specified, in this course we are interested in real fields u. It is
nevertheless convenient to use complex plane waves (2.4) and eventually take the real or imaginary part to find
a real solution, rather than working with the real plane waves cospk x ´ ω tq and sinpk x ´ ω tq from the outset.
CHAPTER 2. WAVES, DISPERSION AND DISSIPATION 14

So the plane wave (2.4) is a solution of the Klein-Gordon equation (2.3) provided that ω
satisfies ?
ω “ ω pk q “ ˘ v k 2 ` m2 . (2.5)
We will usually ignore the sign ambiguity and only consider the ` sign in (2.5) and
similar equations.3

VOCABULARY:
k wavenumber λ“ 2π
k
wavelength (periodicity in x)
ω angular frequency τ“ 2π
ω
period (periodicity in t)
A formula like (2.5) relating ω to k : dispersion relation.

The maxima of a real plane wave, like for instance Re eipkx´ωpkqtq or Im eipkx´ωpkqtq , are
called “wave crests”. By a slight abuse of terminology, we will refer to the wave crests
of the real or imaginary part of a complex plane wave like (2.4) simply as the wave crests
of the complex plane wave.

By rewriting the complex plane wave solution (2.4) of the Klein-Gordon equation as
eikpx´cpkqtq , we see that its wave crests move at the velocity
c
ω pk q m2
cpk q “ “ v 1 ` 2 signpk q .
k k

Plane waves with different wavenumbers move at different velocities, so if we try

to make a lump of real Klein-Gordon field by superimposing different plane waves
ż `8
upx, tq “ Re dk f pk q eipkx´ωpkqtq , (2.6)
´8

it will disperse.

In fact, there are two different notions of velocity for a wave:

- PHASE VELOCITY
ω pk q
cpk q “ , (2.7)
k
which is the velocity of wave crests.
3
We do not lose generality here, since we can obtain the plane wave solution with opposite ω by taking the
complex conjugate plane wave solution and sending k Ñ ´k .
CHAPTER 2. WAVES, DISPERSION AND DISSIPATION 15

- GROUP VELOCITY
dω pk q
cg pk q “ , (2.8)
dk
which is the velocity of the lump of field while it disperses.

We will understand better the relevance of the group velocity in the next section.

REMARK:
The energy (and information) carried by a wave travels at the group velocity, not at the
phase velocity. For a relativistic wave equation with speed of light v , no signals can be
transmitted faster than the speed of light. So it should be the case that |cg pk q| ď v for all
wavenumbers k , but there is no analogous bound on the phase velocity. For example, for the
Klein-Gordon equation (2.3), we can calculate

- |Group velocity|: ˇ ˇ
ˇ dω pk q ˇ
|cg pk q| “ ˇ
ˇ ˇ“ b v ďv
dk ˇ 1` m2
k2

consistently with the principles of relativity.

- |Phase velocity|: c
ˇ ˇ 2
ˇ ω pk q ˇ
|cpk q| “ ˇˇ ˇ“v 1` m ěv,
k ˇ k2
which is faster than the speed of light v for all k , but this is not a problem.

2.2 Example: the Gaussian wave packet

The simplest example of a localised field configuration obtained by superposition of plane
waves is the “GAUSSIAN WAVE PACKET”, which is obtained by choosing a Gaussian
2 pk ´k̄ q2
f pk q “ e´a pa ą 0, k̄ P Rq

in the general superposition (2.6). This represents a lump of field with

average wavenumber k̄
spread of wavenumber „ 1{a ,

see fig. 2.1.

CHAPTER 2. WAVES, DISPERSION AND DISSIPATION 16

Figure 2.1: Gaussian wavepacket in Fourier space.

Then upx, tq “ Re z px, tq is a real solution of the Klein-Gordon equation, where

ż `8
2 2
z px, tq “ dk e´a pk´k̄q eipkx´ωpkqtq , (2.9)
´8

?
provided that ω pk q “ v k 2 ` m2 .4

Since most of the integral (2.9) comes from the region k « k̄ , we can obtain a good approxi-
mation to (2.9) by Taylor expanding ω pk q about k “ k̄ . Expanding to first order in pk ´ k̄ q we
obtain
ω pk q “ ω pk̄ q ` ω 1 pk̄ q ¨ pk ´ k̄ q ` Oppk ´ k̄ q2 q
“ ω pk̄ q ` cg pk̄ q ¨ pk ´ k̄ q ` Oppk ´ k̄ q2 q
« ω pk̄ q ` cg pk̄ q ¨ pk ´ k̄ q ,

where in the second line we used (2.8) and in the third line we introduced a short-hand « to
4
z px, tq is a complex solution of the Klein-Gordon equation. Since the Klein-Gordon equation is a linear
equation with real coefficients, the complex conjugate z px, tq˚ is also a solution of the Klein-Gordon equation,
as are Re z px, tq and Im z px, tq.
CHAPTER 2. WAVES, DISPERSION AND DISSIPATION 17

avoid writing Oppk ´ k̄ q2 q every time. Substituting in (2.9), we find

ż `8
2 2
z px, tq « dk e´a pk´k̄q eitkx´rωpk̄q`cg pk̄q¨pk´k̄qstu
´8
ż `8
irk̄x´ω pk̄qts 2 2
“e dk e´a pk´k̄q eipk´k̄qrx´cg pk̄qts
´8
ż `8
irk̄x´ω pk̄qts 2 2
“ e dk e´a k `ikrx´cg pk̄qts
kÑk`k̄ ´8
ż `8 2
irk̄x´ω pk̄qts ´ 4a12 rx´cg pk̄qts2 2 i
“ e e dk e´a tk´ 2a2 rx´cg pk̄qtsu
complete ´8
the square
?
π ´ 12 rx´cg pk̄qts2
“ eirk̄x´ωpk̄qts ¨ e 4a ,
a
loooomoooon
Gaussian
CARRIER WAVE
integral
loooooooooomoooooooooon
ENVELOPE

where in the second line we factored out a plane wave with k “ k̄ , in the third line we
changed integration variable replacing k by k ` k̄ , in the fourth line we completed the square
2
Ak 2 ` B k “ Apk ` 2BA q2 ´ B
4A
, and in the last line we used the Gaussian integral formula
ż `8`ic c
2 π
e´Ak “ ,
´8`ic A

which holds for all A ą 0 and c P R. The final result is the product of a:

1. “CARRIER WAVE”:
a plane wave moving at the phase velocity

ω pk̄ q
cpk̄ q “
k̄

2. “ENVELOPE”:
a localised profile (or “wave packet”) mov-
ing at the group velocity

cg pk̄ q “ ω 1 pk̄ q .

Click here to see an animation of a Gaussian wavepacket with a (Gaussian) envelope and a
carrier wave moving at different velocities. In the animation the phase velocity is much larger
than the group velocity.

To this order of approximation, the spatial width of the lump has the parametric dependence
CHAPTER 2. WAVES, DISPERSION AND DISSIPATION 18

WIDTH „ a,

meaning that the width doubles if a is doubled, and is constant in time. (Indeed, a simultaneous
rescaling of x ´ cg pk̄ qt and a by the same constant λ leaves the envelope invariant.)

˚ EXERCISE: Improve on the previous approximation by including the 2nd order in k ´ k̄ .

Show that [Ex 10]
ω 2 pk̄q 2
WIDTH2 „ a2 ` 4a2
t

and that the amplitude of the wave packet also decreases as time increases.

This leads to the phenomenon of DISPERSION, whereby the profile of the wave packet
changes as it propagates. In particular, starting from a localised wave packet, dispersion makes
the wave packet spread out: the width of the initial wave packet grows and the amplitude de-
creases as time increases. See this animation of the time evolution of the Gaussian wave-packet
up to second order in pk ´ k̄ q.

2.3 Dissipation
So far we have considered wave equations which lead to a real dispersion relation, so ω pk q P R.
If instead ω pk q P C, then a new phenomenon occurs: DISSIPATION, where the amplitude
of the wave decays (or grows) exponentially in time. For a plane wave
upx, tq “ eipkx´ωpkqtq “ eipkx´ Re ωpkq¨tqq e Im ωpkq¨t (2.10)
and we have two cases:

• Im ω pk q ă 0: “PHYSICAL DISSIPATION”
The amplitude decays exponentially with time.

• Im ω pk q ą 0: “UNPHYSICAL DISSIPATION”
The amplitude grows exponentially with time (physically unacceptable).

EXAMPLES:

1.
1
ut ` ux ` αu “ 0 pα ą 0, v ą 0q (2.11)
v
Sub in a plane wave u “ eipkx´ωtq :
ω
´i ` ik ` α “ 0 ùñ ω pk q “ v pk ´ iαq ,
v
CHAPTER 2. WAVES, DISPERSION AND DISSIPATION 19

leading to a complex dispersion relation. The plane wave solution is therefore

upx, tq “ eikpx´vtq e´αvt

and the wave decays exponentially, or “dissipates”, to zero as t Ñ `8. This is an

example of physical dissipation. (αv ă 0 would have led to unphysical dissipation.)

2. HEAT EQUATION:
ut ´ αuxx “ 0 pα ą 0q (2.12)

˚ EXERCISE: Sub in a plane wave and derive the dispersion relation ωpkq “ ´iαk . 2

So the plane wave solution of the heat equation is

2t
upx, tq “ eikx e´αk

and the waves dissipates as time passes.

2.4 Summary
• Linear wave equation ÝÑ (Complex) plane wave solutions u “ eipkx´ωtq .
Sub in to get ω “ ω pk q dispersion relation.

• Wave crests move at cpk q “ ω pk q{k phase velocity.

(If ω pk q P C, then we define the phase velocity as cpk q “ Re ω pk q{k .)

• Lumps of field move at cg pk q “ ω 1 pk q group velocity.

/wave packets
(If ω pk q P C, then we define the group velocity as cg pk q “ Re ω 1 pk q.)

• Dispersion (real ω , width increases and amplitude decreases) and dissipation (complex
ω , amplitude decreases exponentially) smooth out and destroy localised lumps of energy
in linear wave (or field) equations.

• Non-linearity can have an opposite effect (steepening and breaking, see chapter 1).

• For solitons the competing effects counterbalance one another precisely, leading to
stable lumps of energy, unlike for ordinary waves.