Notes On Guenter Lee PDEs PDF
Notes On Guenter Lee PDEs PDF
Notes On Guenter Lee PDEs PDF
Differential Equations
Bernard Deconinck
Department of Applied Mathematics
University of Washington
Campus Box 352420
Seattle, WA, 98195, USA
April 23, 2010
i
Prolegomenon
These are the lecture notes for Amath 569: Methods of applied mathematics III: partial
differential equations. 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 Partial differential equations of mathematical physics and integral
equations, by Ron Guenther and John Lee (Dover, 1996). Other recommended sources are
Whithams Linear and nonlinear waves (Wiley, 1999), and Evans Partial differential
equations (AMS, 1998).
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 Motivation and introduction
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2 Well-posedness of partial differential equations . . . . . . . . . . . . . . . . .
2 The method of characteristics
2.1 First-order equations . . . . . . . . . . . .
1.
Homogeneous case . . . . . . . . .
2.
Nonhomogeneous case . . . . . . .
2.2 What happens after shock formation? . . .
1.
Introducing dissipation . . . . . . .
2.
Rankine-Hugoniot conditions . . .
2.3 Classifying higher-order equations . . . . .
1.
Hyperbolic equations: z1 6= z2 , real
2.
Parabolic equations: z1 = z2 , real .
3.
Elliptic equations: z2 = z1 , complex
4.
A general classification . . . . . . .
2.4 Exercises . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
1
1
3
.
.
.
.
.
.
.
.
.
.
.
.
5
5
5
12
15
15
18
21
23
25
25
29
29
.
.
.
.
.
.
.
.
.
.
.
.
35
35
36
42
44
47
50
51
55
56
56
58
61
iv
CONTENTS
6.7
.
.
.
.
.
.
. . . . . . .
. . . . . . .
. . . . . . .
dependence
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
heat equation
A homogeneous initial-boundary-value problem . . . .
A nonhomogeneous initial-boundary-value problem . .
Nonconstant boundary conditions . . . . . . . . . . . .
The maximum principle in one dimension . . . . . . . .
The maximum principle in a multi-dimensional setting
The Greens function . . . . . . . . . . . . . . . . . . .
1.
The heat equation on a finite domain, revisited
2.
The heat equation on an infinite domain . . . .
3.
The heat equation on a semi-infinite domain . .
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
7 Sturm-Liouville problems
7.1 Greens functions for Sturm-Liouville problems . . . . . . . .
1.
Definition and derivation . . . . . . . . . . . . . . . .
2.
Properties of the Greens function . . . . . . . . . . .
3.
The Sturm-Liouville problem as an integral equation
7.2 Inner product spaces . . . . . . . . . . . . . . . . . . . . . .
7.3 The Fredholm alternative . . . . . . . . . . . . . . . . . . .
1.
Proof for degenerate kernels . . . . . . . . . . . . . .
2.
Continuous kernels . . . . . . . . . . . . . . . . . . .
3.
The strong form of the Fredholm alternative . . . . .
7.4 Eigenvalues and eigenfunctions . . . . . . . . . . . . . . . .
7.5 Eigenfunction expansions . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
63
63
63
64
65
66
66
70
72
77
.
.
.
.
.
.
79
80
80
81
83
85
86
.
.
.
.
.
.
.
.
.
.
89
89
91
92
93
98
100
101
104
107
109
.
.
.
.
.
.
.
.
.
.
.
111
111
112
115
118
119
124
125
126
128
130
132
CONTENTS
7.6
7.7
1.
Preliminaries . . . . . . . . . . .
2.
The Hilbert-Schmidt theorem . .
Sturm-Liouville problems in applications
1.
Some more lemmas and theorems
2.
Example: a nonhomogeneous heat
Exercises . . . . . . . . . . . . . . . . . .
v
. . . . . .
. . . . . .
. . . . . .
. . . . . .
equation
. . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
8 Laplaces equation
8.1 Laplaces equation on separable domains . . . . . . . .
1.
Separation of variables . . . . . . . . . . . . . .
2.
The Greens function . . . . . . . . . . . . . . .
8.2 The Hankel transform . . . . . . . . . . . . . . . . . .
8.3 The maximum principle . . . . . . . . . . . . . . . . .
8.4 Solving Poissons equation . . . . . . . . . . . . . . . .
1.
The fundamental solution of Laplaces equation
2.
Poissons equation on Rn . . . . . . . . . . . . .
8.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
132
135
136
136
139
141
.
.
.
.
.
.
.
.
.
143
144
144
146
147
150
153
154
155
159
vi
CONTENTS
Chapter 1
Motivation and important problems
related to partial differential
equations
1.1
Motivation
Most courses on partial differential equations start off by deriving some of the basic equations
from the physical models from which they come. You have probably already seen such
derivations before, so we will spend little time on this. Lets look at some of the basic partial
differential equations well consider in this course.
a) utt = c2 (uxx + uyy + uzz ). This is the wave equation. It arises in all kinds of
settings where signals propagate without undergoing any damping. Typically, youll
see a derivation of this equation in textbooks using a vibrating string setting.
b) ut = (uxx + uyy + uzz ). This is the heat equation. It often arises in settings where
dissipation is an important mechanism. A typical example is the redistribution of heat
in some domain.
c) uxx + uyy + uzz = 0. This is Laplaces equation. It can be derived in its own right,
but it is also seen to govern the stationary solutions of the wave and heat equations.
I will not take you through these derivations. Rather, Id like to discuss some of the
important systems of partial differential equations, and what some outstanding problems
are.
The Navier-Stokes equations
There are many incarnations of the Navier-Stokes equations. One of the most popular is
u
+ u u = u p
t
u = 0,
1
where the three-dimensional vector u represents the velocity field of the flow, is the viscosity, and p is the pressure. These equations describe incompressible viscous flow. Proving
that these equations have solutions, and establishing these solutions, is the topic of one of
the Clay Institute million dollar prizes.
Maxwells equations
These equations describe the dynamical response of the electric and magnetic fields to different media. They are a popular T-shirt topic.
D
D
H
t
B
E+
t
B
= ,
= J,
= 0,
= 0,
with the additional two equations D = E and B = H. Here D is the electric displacement, is the charge density, H is the magnetic field, J is the current density, E is the
electric field, and B is the magnetic induction. Lastly, and represent the dielectric tensor
and the magnetic permeability tensor respectively. Additional constitutive relationships give
connections between and J and the other field quantities.
Note that Maxwells equations are linear. Any nonlinearity enters through the response
of the fields to the medium.
The Schr
odinger equation
The single-particle Schrodinger equation is given by
~2
+ V (x) = i~ .
2m
t
This equation is also linear. It has to be, due to the essence of the superposition principle
in quantum mechanics.
The nonlinear Schr
odinger equation
it = xx + ||2 .
This equation describes the slow modulation of a wave packet in a nonlinear medium. Higherdimensional versions of this equation exist as well.
1.2
Given that we typically use partial differential equations to obtain information about physical
or other application problem, it is important that the partial differential equation is wellposed.
What do we mean when we say a problem is well posed?
There are a sufficient number of initial conditions: if there are time derivatives
in the problem, an equal amount of initial conditions should be specified, just like for
ordinary differential equations.
Since were dealing with a physical problem, we would like these problems to (a) have a
solution, (b) not to be too sensitive with respect to measured quantities such as initial and
boundary conditions.
Example: Consider the heat equation
ut = u
u(x, 0) = 1,
for x = (x, y, z) R3 and > 0, R Suppose we have a final condition u(x, 0), but
we want to know where it came from. In other words, we want to solve the heat equation
backwards. Thus, wed have to solve
ut = u
u(x, 0) = 1,
Note that
un (x, t) = 1 +
1 3n2 t
e
sin(nx) sin(ny) sin(nz)
n
Chapter 2
Partial differential equations of the
first order: the method of
characteristics
2.1
First-order equations
(2.1)
Such an equation is called quasi-linear, because it is linear in the derivatives of u(x, t).
We could have written down a slightly more general equation, by considering c(u) to be an
explicit function of x and t as well.
We will use the method of characteristics to solve equations like this. This method is
more important than just for solving equations like (2.1), as it also allows us to classify
partial differential equations of higher order. Well get back to that later.
Let u(x, 0) = u0 (x) be the initial condition for (2.1), given on the whole line x R, see
Fig. 2.1. We want to know what happens to u(x, t) as t > 0. Lets look at the homogeneous
case first.
1.
Homogeneous case
du
= 0,
dt
du
u dx u
=
+
,
dt
t
dt x
5
(2.2)
u0 (x)
x
Figure 2.1: The initial condition u = u0 (x) for equation (2.1)
u = constant along C
u = u0 (),
along the characteristic curve that starts at x = in the (x, t)-plane. What is the equation
of this curve? Well, we have along this curve
dx
= c(u(x, t))
dt
dx
= c(u0 ())
dt
x = + tc(u0 ()).
Thus, the solution of the equation (2.2) with initial condition u(x, 0) = u0 (x) is given by
u = u0 (),
where is determined by the implicit equation
x = + tc(u0 ()).
x
Figure 2.2: The initial condition u(x, 0) = u0 (x) being transported away from the x axis, in
the directions of the arrows. (2.1)
x
Figure 2.3: The characteristics for ut + cux = 0, demonstrating the uniform translation of
the initial condition over time. (2.1)
These characteristic lines are fanning out, thus the solution spreads out. It is defined for all
t 0. It is given by
u(x, t) = u0 () = arctan,
where = (x, t) is determined by the implicit relationship
x = + t arctan.
In this case, we cannot write down an explicit solution, as we cannot solve x = + t arctan
explicitly for as a function of x and t. However, it is still straightforward to analyze the
behavior of the solution. The solution at different times and the characteristics are shown
in Fig. 2.4.
Example: Consider the system
ut + uux = 0,
u(x, 0) = arctanx.
This is almost the same example, but now the sign of the initial condition is switched.
Following what we did in the previous example, the solution is given by
u(x, t) = u0 () = arctan,
where = (x, t) is determined by the implicit relationship
x = t arctan.
As before, we cannot solve this explicitly for , but we can analyze the behavior of the
solution. The plot of the characteristics is shown in Fig. 2.5. We see there is a time t
after which the solution obtained through the methods of characteristics is no longer valid:
multiple characteristics cross, and in this wedge of overlapping characteristics, it is not
possible to assign a unique value to u(x, t).
We will return to this example soon, to resolve the issues we just encountered. When a
situation with crossing characteristics is encountered, what happens for t > t , the time at
10
t
1
Figure 2.4: The characteristics for ut + uux = 0, demonstrating the spreading of the initial
condition u0 =arctanx over time.
which they first cross? This is a question well look into later. For now, lets find a condition
to determine the time of first crossing t . This time is known as the breaking time, as it
is the time at which the wave first breaks: the solution develops an infinite derivative in
space x or time t (if one happens, the other one happens as well, because of the partial
differential equation). We have
ut = u0 t ,
ux = u0 x ,
using the chain rule. These will become infinite when both of x and t become infinite,
since u0 is a well-behaved function. Taking x and t derivatives of the equation for the
characteristics gives
x = + tc(u0 ())
x =
1
tc0 (u
1+
0 ())u0
c(u0 ())
.
t =
1 + tc0 (u0 ())u0
These have the same denominators, so that the breaking time is determined by when these
denominators are zero. Thus
1
t = 0
,
c (u0 ( ))u0 ( )
11
1
0
Figure 2.5: The crossing of the characteristics for ut +uux = 0, demonstrating the steepening
of the initial condition u0 = arctanx over time.
c(u) = u
u0 = arctan
Thus is determined by
c0 u0 < 0
1
< 0,
1 + 2
c0 = 1,
1
u0 =
.
1 + 2
12
1
c0 (u
( ))u
( )
1
1+(1 )2
= 1 + ( )2 = 1,
2.
Nonhomogeneous case
du
dt = F (u, x, t),
dx
= c(u, x, t).
dt
This system of equations should be solved as seen fit, along the characteristic for which x =
and u = u0 at t = 0. After doing so, is eliminated from this procedure, resulting in the
final solution u(x, t). Note that since now were allowing c to depend explicitly on x and/or
t, the characteristics are not necessarily straight lines.
Example: Consider the equation
ut +
u+1
x
ux =
,
x+t
x+t
with
u(x, 0) = u0 (x) = x2 .
The characteristics are given by
dx
x
=
,
dt
x+t
and the solution along characteristics satisfies
du
u+1
=
.
dt
x+t
From the first equation we get
dt
x+t
t
=
=1+
dx
x
x
dt
1
t=1
dx x
0
t
1
=
x
x
t
= ln |x| + c.
x
13
We want this to be the equation of the characteristic curve that passes through x = when
t = 0, thus
0 = ln || + c c = ln ||,
from which
x
t = x ln .
This is the equation for the characteristics. From Fig. 2.6, one sees that there will be
problems with the solution straight away, as the characteristics have a singularity at = 0,
and multiple characteristics become tangent immediately, for any t > 0. We may still hope
to construct a solution valid in other regions of the (x, t)-plane, if we dont necessarily think
of this as an initial-value problem.
Figure 2.6: Some characteristics for the problem ut + ux x/(x + t) = (u + 1)/(x + t).
The solution along these characteristics is determined by
du
u+1
u+1
u+1
=
=
=
dt
x+t
x + x ln |x/|
x(1 + ln |x/|)
du dx
u+1
=
dx dt
x(1 + ln |x/|)
14
du 1
u+1
=
dx dt/dx
x(1 + ln |x/|)
1
u+1
du
=
dx (1 + t/x)
x(1 + ln |x/|)
1
u+1
du
=
dx 1 + ln |x/|
x(1 + ln |x/|)
du
u+1
=
dx
x
du 1
1
u=
dx x
x
u 0
1
= 2
x
x
1
u
= +c
x
x
u = 1 + cx.
At x = , we want u = u0 , so that
u0 = 1 + c c =
u0 + 1
.
This gives
u0 + 1
x
2 + 1
u = 1 +
x
1
u = 1 + x +
.
u = 1 +
Note that we dont need absolute value signs anymore. The last equation implies that x and
have the same sign. Finally,
1 t/x
t/x
u = 1 + x xe
+ e
x
= 1 + x2 et/x + et/x .
Let me emphasize again that in general an explicit solution cannot be obtained using
this procedure. We got lucky in the previous example, since we could solve the equation for
the characteristics explicitly for , which we could then use to eliminate for the solution
15
formula for u. Even so, we can still obtain many important properties of the solution from
this procedure1 . In this particular case breaking starts immediately at t = 0, and a shock is
formed.
2.2
1.
Introducing dissipation
The first method well see modifies the governing equation to prevent the formation of
shocks. There are many ways of doing this, and typically one looks at the application being
considered to figure out what the right way is to achieve this. Our equation is supposed to
model some process associated with an application. When we get close to shock formation
and derivatives get large, it is possible that other terms that we have previously omitted
become important.
Example: We have found that the dissipationless Burgers equation
ut + uux = 0
develops a shock if the initial condition has a profile as shown in Fig. 2.7. In general, any
profile with negative derivative will lead to steepening. Now suppose that the problem we
are modeling actually corresponds to the differential equation
ut + uux = uxx ,
where > 0 is very small. Because of the small coefficient we may have thought it was
fine to ignore this second derivative term. This approximation will be valid, as long as
derivatives in the solution are small (meaning not large). However, as we approach the
shock, derivatives become very large and the term on the right does become significant.
Lets verify that this is indeed the case. Close to the shock time t , we have
ux = u00 ()x =
1
u00 ()
,
1 + tu00 ()
and what else but this do we want an explicit solution formula for anyways?
16
u(x,t)
Figure 2.7: Part of an initial profile that will lead to shock formation in the dissipationless
Burgers equation.
x = + tu0 ()
1 = x + tu00 ()x
0 = xx + tu00 ()xx + tu000 ()x2
xx (1 + tu00 ()) = tu000 ()x2
tu000 ()x2
.
xx =
1 + tu00 ()
17
and we see that uxx much faster than ux or ut . Thus, the right-hand side term cannot
be ignored close to the shock formation.
Next, well show that taking this second-derivative term into account does indeed have
the desired effect of arresting the shock formation. Near the shock we have
ut = uxx uux uxx ,
Since uxx is much larger than the product of uux just before the shock formation. Just before
a shock is formed, the solution has a singular inflection point (i.e., the second derivative
changes sign, but by passing through infinity instead of zero). Above this inflection point
ut uxx < 0, whereas below the inflection point ut uxx > 0. The combination of these
is to counteract the steepening, as desired, see Fig. 2.8.
Figure 2.8: The effect of second-derivative dissipation on the near-formation of a shock near
an inflection point P .
For other examples, and depending on the applications, different mechanisms may prevent
shock formation. For instance for dispersive processes, the introduction of a third-derivative
term may be more appropriate:
ut + uux = uxxx .
The resulting partial differential equation is known as the Korteweg-deVries equation. It is
one of the seminal equations of nonlinear wave theory and integrable systems, as treated in
Amath573. Investigating the behavior of solutions as 0 is nowhere near as straightforward as for the Burgers equation with dissipation. It is the topic of what is now known as
Lax-Levermore theory, where one examines the zero-dispersion limit of wave equations with
increasingly small dispersion. We will have more to say about dispersive equations in later
chapters.
For yet different applications, one may have to resort to higher-order dissipation or dispersion, or different types of nonlinear terms.
18
2.
Rankine-Hugoniot conditions
Another approach to the problem of shocks forming is to embrace them: instead of trying
to prevent them, lets just figure out what to do with them once they form. Lets figure out
where the shocks go, and how fast they move. The key to doing this is to work with an integral
form of the partial differential equation. This integral form typically corresponds to the
conservation law form of the application being modeled. Most partial differential equations
arising in applications are originally derived in integral form, expressing the conservation of
various quantities, such as mass, charge, momentum, energy, etc. This is true for Maxwells
equations in electro-magnetism, for the equations of continuum mechanics, including the
Navier-Stokes equations.
The idea behind using an integral form of the partial differential equation is that integrals
do not require continuous integrands to act on, whereas the arguments of derivatives should
be smooth. Thus, using the integral form, we can deal with discontinuous solutions. Such
solutions are also known as weak solutions of the differential equation.
Example: Again we consider the dissipationless Burgers equation
ut + uux = 0.
We can rewrite this equation as
(u) +
t
x
1 2
u
2
= 0.
This form of the equation is called the conservation law form: integrating over the entire
spatial domain, we find a conserved quantity for the partial differential equation if suitable
boundary conditions are used. For instance
Z
Z
1 2
d
udx = 0,
(u) +
u
dx = 0
t
x 2
dt
R
where weve assumed that u 0 as x . Thus udx is a constant of the motion.
Lets integrate the conservation law form over only part of space:
x=x
Z x2
1 2 2
(u)dx + u
=0
2 x=x1
x1 t
Z
d x2
1 2
udx =
u1 u22 ,
dt x1
2
where u1 = u(x1 , t) and u2 = u(x2 , t). As long as our solution has continuous derivatives,
this integral equation, which is valid for all x1 and x2 , is completely equivalent with the
original partial differential equations form of the dissipationless Burgers equation.
Now suppose that the solution exhibits a shock at x = s(t). Also, assume that x2 > s(t)
and x1 < s(t). Lastly, let u = limx<s(t) u(x, t), and u+ = limx>s(t) u(x, t). Thus u and u+
19
are the values of u just past and in front of the shock. Then we get
!
Z s(t)
Z x2
d
1 2
udx =
u1 u22
udx +
dt
2
s(t)
x1
Z x2
Z s(t)
1 2
0
+ 0
ut dx =
ut dx +
u s u s +
u1 u22 .
2
s(t)
x1
Now we let x1 s and x2 s. Then both of
Z
Z s(t)
ut dx and
x2
ut dx
s(t)
x1
+ 0
+2
(u u )s =
u u
2
1 (u u+ )(u + u+ )
s0 =
2
u u+
1
s0 = (u + u+ ).
2
We find that the velocity of the shock is equal to the average of the solution amplitude in
front of and past the shock.
How can we use this? For typical initial conditions the values of u and u+ are not
constant, and the shock velocity condition represents an ordinary differential equation for
the shock position. However, we know at which time t and on which characteristic the
shock first forms. We may use our condition to move the shock approximately, using an
Euler-like numerical method for instance.
Lets see how this works: at time t , a shock forms on characteristic , from which we
find the initial position s(t ) of the shock. Then we know the values of u = u(s(t ), t ) = u+ .
Note that u (t ) = u+ (t ) for smooth initial conditions. This is not the case if the initial
conditions are discontinuous. Using this, we may propagate the shock with a small time step
t:
1
s(t + t) = s(t ) + (u + u+ )t + O((t)2 ).
2
+
Next, we update the values of u and u , using the values transported to the shock line
along the new characteristics that now contribute to the shock. The introduction of the
shock, splits the region where characteristics used to cross in two regions, one to the left of
the shock line, one to the right. The characteristics can now safely be followed all the way
to the shock line, without risking any multivaluedness. A schematic of this idea is presented
in Fig. 2.9.
In general, for an equation of the form
+
j = 0,
t
x
20
Figure 2.9: A shock region before and after the introduction of the shock. In the original
shock region (left, filled in), the characteristics cross. After the shock is introduced (right,
bold line), the characteristics are followed all the way to the shock.
j
,
s0 =
2 u u+
2 u + u u+ + u+
=
.
3 u 2 u+ 2
3
u + u+
This is different from the result we have obtained before! So, when going from the differential
form of the partial differential equation to the integral form (which is of course what were
21
doing when we invoke the Rankine-Hugoniot condition), we might lose uniqueness. This
is unsatisfying, but not unexpected. As previously announced, the application in question
(physics, engineering, biology, etc.) will have to decide which integral formulation to use.
The above calculation merely illustrates that the mathematics alone cannot tell us the right
answer.
Lets summarize: typically an integral formulation is obtained first for many applications.
Then one often switches to a differential formulation, because differential equations are easier
to solve than integral equation. Or maybe we just have more experience with them. In
any case, the limit process of going from an integral equation to a differential equation is
straightforward and it results in a unique differential formulation. Then, when shocks are
found in the differential equation, it is important to know which integral formulation the
differential equation originated from, so as to be able to resolve the shock in a consistent
manner.
2.3
In this section, well use our understanding of characteristics for first-order equations to
classify equations of the second order. Specifically, well look at equations of the form
a(x, y)uxx + 2b(x, y)uxy + c(x, y)uyy = f (x, y, u, ux , uy ).
Weve denoted the independent variables by x and y. One of them might be interpreted as
a time-like variable t, depending on the results well find.
Using an invertible (at least locally) transformation, we will transform the above equation
to one of the normal forms
1. uxx uyy = F , or
2. uy = uxx + F , or
3. uxx + uyy = F ,
where in all cases F = F (x, y, u, ux , uy ). In the first case, well call the equation hyperbolic.
The independent variable x will get the interpretation of time t. The second case is called
parabolic, and y plays the role of time t. The last case is that of an elliptic equation, and
neither variable can be interpreted as time.
The classification well carry out is very similar to that of classifying conic sections in
the plane. If we cared to, and had enough time and energy, we could deal with equations
with a third independent variable z, which would be similar to classifying quadrics in threedimensional space.
In order to achieve the classification above, we use a transformation on the independent
variables
= (x, y)
,
= (x, y)
22
which we want to be invertible, at least locally. This implies that the Jacobian
x
J(x, y) = det
x
6= 0.
y
Using the chain rule, we see that the first derivatives are transformed according to
ux = u x + u x
.
uy = u y + u y
More chain ruling
uxx
uyy
uxy
(Are you excited yet? Were just getting started!) We now use these expressions to transform
the original second-order equation. We get
a(x, y)uxx + 2b(x, y)uxy + c(x, y)uyy =
a(x, y)(u xx + u xx + u x2 + 2u x x + u x2 )
+2b(x, y)(u xy + u xy + u x y + u x y + u (x y + y x ))
+c(x, y)(u yy + u yy + u y2 + 2u y y + u y2 ) =
f (x, y, u, u x + u x , u y + u y ),
So that
A(, )u + 2B(, )u + C(, )u = F (, , u, u , u ),
where
f (x, y, u, ux , uy )
.
2b(x, y)
We may assume that b 6= 0, since otherwise our equation is not of second order. Using the
transformation
x = +
,
y =
23
u =
(using an invertible transformation, it is not possible to make B zero as well) The example above shows how this form is easily transformed to the standard form of a hyperbolic
equation.
For starters, we divide by 2y to obtain
a
x
y
2
+ 2b
x
y
+ c = 0,
a(z z1 )(z z2 ) = 0,
1.
In this case
x = z1 y
.
x = z2 y
We can solve these equations using the method of characteristics. For we have
d
= 0,
dx
dy
= z1 ,
dx
where the second equation defines the characteristic y = y(x) passing through y = y0 at
x = x0 , and
= (x0 , y0 )
24
along this characteristic. Note that z1 typically depends on x and y, since a, b and c do. We
have similar equations for . These equations for and are valid as long as x0 and y0 are
in regions of the (x, y)-plane where z1 and z2 are real. Then in that region of the (x, y)-plane
the equation is reduced to the standard form
2B(, )u = F (, , u, u , u )
if the Jacobian condition holds:
x y
x y
det
6= 0
z1 y y
det
6 0
=
z2 y y
y y (z1 z2 ) 6= 0
y y 6= 0,
F (, , u, u , u )
,
2B(, )
valid in the region where b2 ac > 0. In this region the partial differential equation is called
hyperbolic. We can think of the partial differential equation as transporting information
along the characteristics
x = z1 y
.
x = z2 y
One consequence of this is that information contained in initial data for a hyperbolic problem
travels at finite speed, dictated by the inverse slope of the characteristics. Well have more
to say about this when we study the wave equation
utt c2 uxx = 0,
which is the prototypical example of a second-order hyperbolic equation. Note that we have
identified y with t.
2.
25
If b2 ac 0, then z1 = z2 R. In this case the quadratic equation has only one distinct
solution, and we can only define one characteristic. Clearly, were not allowed to pick and
both this way, as the resulting transformation (x, y) (, ) would not be invertible.
Choosing to satisfy
b
x = z1 y x = y ,
a
we may choose to be anything that is linearly independent of our choice for . Now we
have
B = ax x + b(x y + y x ) + cy y
b
= by x + b y y + y x + cy y
a
2
b
= y y + cy y
a
ac b2
= y y
a
= 0,
and our normal form is
u =
F (, , u, u , u )
.
C(, )
Equations that reduce to this form are called parabolic. The prototypical example of a
parabolic equation is the heat equation
ut = uxx ,
with parameter > 0. Note that we have identified y with t. For the heat equation, the xaxis is a characteristic. This is often interpreted in the sense that initial data information for
the heat equation propagates infinitely fast: any disturbance in u(x, t0 ) immediately impacts
u(x, t) for any t > t0 .
3.
Lastly, we turn to the case where b2 ac < 0, and z1 and z2 are complex conjugate functions of
x and y. Now, assuming that a, b and c are suitable analytic, the whole preceding calculation
for the hyperbolic case may be repeated here. This leads to the normal form
u = G(, , u, u , u ),
as before, valid whenever b2 ac < 0. However, now
x = z1 y
,
x = z2 y
26
and z1 and z2 are complex, leading to complex characteristics. To eliminate these complex
quantities, we set
= + i
= ( + )/2
.
= i
= ( )/2i
Note that these substitutions are consistent with our equations, since z1 = z2 . We get
u = (u + u )
1
= (u iu )
2
1
= (u + u iu iu )
2
1
= (u + u ) ,
4
which leads us to the standard form of an elliptic equation
u + u = 4G(, , u, u , u ),
valid whenever b2 ac < 0. Since the characteristics are complex, we no longer have the
notion of information being transported along them. A disturbance anywhere in the (, )domain has an instantaneous impact on the entire domain. This is even worse than in the
parabolic case! The prototypical equation of elliptic type is Laplaces equation
uxx + uyy = 0,
which well say more about later as well.
Example: Consider the equation
yuxx xuyy = 0,
with (x, y) R2 . Here we have
a = y, c = x, b = 0.
Then
b2 ac = xy,
and the classification of the equation will be different in the four main quadrants. The
boundaries of the different domains are determined by b2 ac = 0, which give the x and y
axes, as shown in Fig. 2.10.
The hyperbolic region is the only region with (real) characteristics. They are determined
by
x = z1 y
.
x = z2 y
2
b ac<0
Elliptic
2
b ac>0
Hyperbolic
27
y
b2ac>0
Hyperbolic
x
2
b ac<0
Elliptic
Figure 2.10: The different regions for the equation yuxx xuyy = 0. It may be tempting to
say that the equation is parabolic on the coordinate axis, but this is not correct. In order
for partial derivatives to be defined at a point, we need an open region around that point.
Thus the classification weve been going through only is valid in such open regions.
az 2 + 2bz + c = 0
yz 2 x = 0
x
z2 =
y
r
r
x
x
z1 =
, z2 =
,
y
y
x
y
y
r
d
dy
x
= 0, along
=
dx
dx
y
x =
y 1/2 dy = x1/2 dx
2 3/2
2
y = x3/2 + c.
3
3
28
3/2
x0
3/2
3/2
y0
2/3
This specifies a family of curves along which is constant. On the other hand will
change along these curves. These curves play the role of coordinate lines for our new
coordinates (, ).
Second characteristic: (in the first quadrant) A transverse (but not necessarily
orthogonal) set of coordinate curves is given
p by the second set of characteristics. Repeating the same calculation with z2 = x/y for gives
y= x
3/2
3/2
x0
3/2
y0
2/3
Both families of characteristics are plotted in Fig. 2.11. A similar picture is valid for
the third quadrant. Notice that the two families become tangent to each other on the
original coordinate axes, reflecting the breakdown of the hyperbolic character of the partial
differential equation.
2.5
1.5
0.5
0.5
1.5
2.5
Figure 2.11: The two families of characteristic curves for yuxx xuyy = 0. Curves of constant
are solid. Curves of constant are displayed using fat dots.
2.4. EXERCISES
4.
29
A general classification
For completeness, we give some standard terminology here. Consider the second-order partial
differential equation
Lu = F (x, u, u),
with
Lu =
n
X
i,j=1
aij
2u
,
xi xj
and x = (x1 , . . . , xn ).
This partial differential equation is called elliptic in a certain region of space if its
coefficient matrix (aij )ni,j=1 is positive definite. Then the linear differential operator L
is called elliptic as well.
The partial differential equation
ut = Lu + F
is called parabolic if the operator L is elliptic.
The partial differential equation
utt = Lu + F
is called hyperbolic if the operator L is elliptic.
It should be stated that sometimes other conventions are used, for instance in such areas
as relativity or field theory. In all cases, the classification is valid locally, in some open
neighborhood of a point.
2.4
Exercises
ut + tuux = 0,
u(x, 0) = sin(x), x R.
Solve this initial-value problem for t 0. Draw the characteristics. For how long is
the solution valid?
2. Consider the initial-value problem
ut + tuux = 0,
u(x, 0) = cos(x), x R.
Solve this initial-value problem for t 0. Draw the characteristics. For how long is
the solution valid?
30
ut + 2xtux = 1,
u(x, 0) = ex , x R.
Solve this initial-value problem for t 0. Draw the characteristics. For how long is
the solution valid?
4. Consider the initial-value problem
ut + xux = 1,
u(x, 0) = ex , x R.
Solve this initial-value problem for t 0. Draw the characteristics. For how long is
the solution valid?
5. Consider the initial-value problem
ut + uux = 0,
u(x, 0) = u0 (x), x R.
Solve this initial-value problem for t 0 with (a) u0 (x) = x, (b) u0 (x) = 1 x2 , (c)
u0 (x) = sin x. For each case, draw the characteristics. For how long is the solution
valid?
6. The method of characteristics is not restricted to initial-value problems. (a) Consider
the problem
xux + yuy = u + 1,
u(x, x2 ) = x2 , (x, y) R2 .
This is Example 2 from Section 2.1 of Guenther and Lee. Use our version of the method
of characteristics to solve this problem. You can use either x or y as your time-like
variable. Where is your solution defined? (b) Now consider the same problem, but
with u(x, x) = x2 , instead of u(x, x2 ) = x2 . Show that this problem cannot be solved
using the method of characteristics.
Explanation: This is because the data from which you are to recover the solution
is given on a characteristic curve (check this). Since the solution is transported along
characteristic curves, giving data along a certain characteristic curve does not allow
one to construct the solution along the other characteristics. In general, the method
of characteristics will work when the data is given along a curve that crosses the
characteristics.
7. We have seen an example where the breaking time t depends on the initial data. Now
consider the initial-value problem
(t )ut + xux = u,
u(x, 0) = f (x), x R, > 0.
2.4. EXERCISES
31
Solve this initial-value problem for t 0, i.e., give an explicit solution for u(x, t) in
terms of the initial data f (x). Draw the characteristics. For how long is the solution
valid? As you see, in this problem the breaking time t is independent of the initial
data f (x).
8. The method of characteristics is easily extended to multidimensional problems. Consider the initial-value problem
ut + c1 (u)ux + c2 (u)uy = 0,
u(x, y, 0) = u0 (x, y), (x, y) R2 .
Generalize the method of characteristics to such two-dimensional equations: write
down the equations for the characteristics and solve them if possible. Write down
the exact solution for u(x, y, t), in explicit form if possible, in implicit form otherwise.
Apply the above method to solve the two-dimensional transport equation ut +
c1 ux + c2 uy = 0, with initial condition u(x, y, 0) = u0 (x, y). Find the explicit
solution for u(x, y, t).
Use a slight variation of your method to solve the damped transport initial-value
problem
ut + c1 ux + c2 uy + u = 0, > 0
u(x, y, 0) = u0 (x, y), (x, y) R2 .
9. Consider the IVP
ut + uux = 0,
u(x, 0) = (x),
x0
1,
1 x, 0 < x < 1
(x) =
0,
x 1.
(a) Show that a shock forms at t = 1. (b) You know that different integral representations of the differential equation can lead to different shock dynamics. By multiplying
the differential equation by u , (1, +), find different shock positions for any
t > 1, depending on . (c) Show that by a suitable choice of , the shock can be
continued along any of the characteristics starting in x0 (0, 1). (d) For any choice of
, draw the solution u(x, t) at t = 0, t = 1/2, t = 1, t = 2, and t = 5.
10. Consider the IVP
ut + uux = 0,
u(x, 0) = (x),
32
x < 1
1,
x2 , 1 < x < 0
(x) =
0,
0<x
Determine the solution of this problem in all regions of x R and t > 0, using the
integral representation.
b
Z
d b
1 2
udx + u = 0.
dt a
2 a
Are there any shocks? If so, where and when do they occur? Also draw all relevant
phases the solution goes through, and provide a picture of the characteristics. If there
are shocks, include a picture of the characteristics without and with the shock lines.
x R,
and
u(x, 0) =
0 for x 0,
1 for x > 0.
Graph the characteristics for this problem in the (x, t)-plane (t 0) and show that
there are no characteristics in a whole sector of the plane.
(b) As with shocks, there are several ways to deal with this problem. Well examine
one here: consider the same differential equation as above, but with initial condition
u (x, 0) = 1/2 + tanh(x/)/2, with > 0. Show that in this case, there is exactly
one characteristic in every area of the (x, t)-plane (t 0) (a graphical proof is
fine). What happens to the initial condition u (x, 0) as 0? What happens to the
characteristics as 0?
(c) We now use the limit characteristics found above to solve the initial-value problem.
It is clear what to do on the original characteristics (i.e., the ones we had before we
talked about this limit process), as we can propagate the initial condition along them.
What do we do along the characteristics in the sector? The solution should be constant
along them, but we dont know what initial condition to propate along them. In order
to be constant along these characteristics, the solution u(x, t) should be of the form
u(x, t) = g(x/t). Substitute this ansatz in the partial differential equation to find an
equation for g(z), z = x/t (ignore the possibility g 0 = 0, which cannot result in a
continuous solution. It can be shown that this choice can indeed be ignored).
(d) Finally, write down the solution to the original initial-value problem obtained this
way and plot it for various instances of time. This solution satisfies the original partial
2.4. EXERCISES
33
differential equation at all time, except at the corner points emanating from the original
discontinuity.
12. Rarefaction waves: another example. Use the same process to solve the initialvalue problem
0 for x 0,
2
ut + u ux = 0, x R, and u(x, 0) =
1 for x > 0.
13. Shock waves. It is not often that shock conditions like the ones we derived in class
can be solved as a differential equation for the shock position, hence determining the
shock position s(t) at any time after its formation. But sometimes, this is possible.
Consider the following initial-value problem:
2
ut + u ux = 0,
x R,
and
u(x, 0) =
1 for x 2,
0 for x > 2.
Solve this initial-value problem in the region of the (x, t)-plane (t 0) where no shock
develops. Next, assuming that
Z x2
Z
d x2
udx +
u2 ux dx = 0
dt x1
x1
is the correct integral formulation to use, derive the shock condition governing the
motion of the shock position x = s(t). In this case, this condition can be solved for the
shock position. Proceed to solve it, and hence find the position of the shock for any
time after its formation. Give the solution to the initial-value problem, for any time
t 0. Illustrate this with some appropriately chosen snapshots of the solution u(x, t).
14. Rarefaction waves and shock waves combined. Consider the following initialvalue problem:
0 for x 0,
1 for 0 < x < 1,
ut + uux = 0, x R, and u(x, 0) =
0 for x 1.
(a) Find the solution of this problem in all areas of the (x, t) (t 0)-plane where there
is exactly one characteristic through each point.
(b) Using the same process as before, fill in the rarefaction region with characteristic
lines, and write down the solution to the initial-value problem in this region.
(c) Assuming that the integral formulation
Z
Z x2
d x2
1 2
udx +
u
dx = 0
dt x1
2
x1
x
34
0 for
3 for
ut + uux = 0, x R, and u(x, 0) =
2 for
1 for
Using rarefaction waves and the integral formulation you used in the previous problem,
solve this problem for all t > 0.
16. Consider the nonlinear partial differential equation
tt 2x 2xt x t + 2t xx = 0.
This equation is a special case of the so-called Monge-Amp`ere equation. Show that in
terms of the auxiliary function u = t /x this equation may be reduced to a first-order
equation. Assuming initial conditions (x, 0) = 1 + 2e3x and t (x, 0) = 4e3x , find
(x, t) for x R and t 0.
Chapter 3
Fourier series and Fourier transforms
I am sure you are all familiar with aspects of Fourier analysis from previous courses. The
main result I want to achieve here are a review of some of the main theorems, this time
around with rigorous statements and proofs. Then later, I want to illustrate why this level
of rigor can matter in an application.
3.1
Fourier series
Lets start with Fourier series. Suppose f (x) can be expanded in a Fourier series:
f (x) = A +
X
an cos
n=1
nx
nx
+ bn sin
,
L
L
f (x)dx.
L
The other coefficients are computed in a similar manner: lets multiply the equation by
cos(mx/L), or by sin(mx/L). Then only one of the terms on the right-hand side has
non-zero average. This is easily seen from the following orthogonality conditions (check
35
36
these!):
Z
mx
nx
cos
dx = 0,
L
L
L
Z L
nx
mx
sin
sin
dx = Lnm ,
L
L
L
Z L
mx
nx
cos
dx = Lnm ,
cos
L
L
L
sin
nx
nx
a0 X
an cos
+
+ bn sin
,
S(x) =
2
L
L
n=1
where a0 , a1 , . . ., b1 , b2 , . . . are given by the preceding formulas, our main question for now
will be to figure out under what conditions S(x) = f (x). When these are not equal, what
can be said?
1.
The formulas for the Fourier coefficients involve definite integrals of a function multiplied
by a trigonometric function. The frequency of this trigonometric function as n .
What can we say about the behavior of the Fourier coefficients for large n? This question
is answered (at least to leading order) by the Riemann-Lebesgue lemma. This lemma is one
of the truly outstanding results of advanced calculus, and well treat it here in its own right.
Well use it later on many occasions.
Lemma 1 (Riemann-Lebesgue) Let g(x) be continuous (except at at most a finite number
of points) in [a, b], and let
Z b
|g(x)|dx < .
a
37
Then
Z
lim
g(x)
cos x
dx = 0.
sin x
and similar for the sine. The lemma states that this bounded integral approaches zero as
.
Ill give you three proofs, only one of which really deserves to be called a proof. But the
first two provide some intuition of what is going on, and why the Riemann-Lebesgue lemma
has to be true.
First proof: graphical. Since the integrand is a continuous function (otherwise we
break the problem up in smaller integrands), we have a graph as shown in Fig. 3.1: the graph
of a continuous function g(x) is multiplied by a trigonometric function of high frequency. This
trig function, for large enough , oscillates on a scale that is much faster than the scale on
which g(x) changes. As a consequence, the figure has the appearance of a slowly modulated
trig function, where g(x) plays the role of the modulation. The definite integral that we are
interested in is the area between this rapidly oscillating function and the horizontal axis. For
sufficiently large, any one of the positive areas is almost exactly canceled out by one of
the neighboring negative areas. This heuristic statement becomes more accurate as n ,
and it seems reasonable to believe the Riemann-Lebesgue lemma.
A numerical evaluation of
Z 2
g(x) cos xdx
I() =
0
a
a a
Z
1 b 0
= lim
g (x) sin xdx.
a
Note that weve assumed that limxa,b g(x) are finite.
We see that integration by parts buys us one factor ofR in the denominator, with an
b
integral of the same type. If in addition we assume that a |g 0 (x)|dx < , then we may
Rb
conclude that a g(x) cos xdx 0, at a rate of 1 .
38
y
5
0.5
1.5
Figure 3.1: The function y = g(x) cos x, with g(x) = x3 x + 2 and = 100. Also shown
are g(x) and g(x), which play the role of modulating amplitudes.
Rb
If g(x) is twice differentiable, limxa,b g 0 (x) are finite, and a |g 0 (x)|dx < , then we can
Rb
repeat this game, to find that a g(x) cos xdx 0, at a rate of 2 .
In general, if g(x) is n times differentiable, suitable limits of various derivatives at
the endpoints are finite, and all derivatives are integrable over [a, b], then we conclude
Rb
that a g(x) cos xdx 0, at a rate of n . Specifically, if g(x) is analytic on [a, b],
Rb
g(x) cos xdx 0 at a rate faster than any power of . This hints at the convergence
a
being exponentially fast!
Third proof: and now for real! Lets look at
Z b
I=
g(x) sin xdx.
a
Lets assume that g(x) is continuous. If its not, we break the integral up in smaller parts,
on which g(x) is continuous. That way, the only discontinuities we have to worry about are
at the endpoints. Well investigate these later.
Let x = t + /. Then t = x /, and
Z b/
I=
g t+
sin(t + )dt
a/
39
y
0.5
20
40
60
80
100
0.5
R2
0
b/
=
a/
g t+
sin tdt.
We now add this result to our definition of I, to obtain 2I. Also, we rename t inside the
integral to be x. This is allowed, since the integration variable is only a dummy variable.
We get
b/
sin xdx
g(x) sin xdx
g x+
a
a/
Z a
=
g x+
sin xdx+
a/
!
Z b/
Z b/
+
g(x) sin xdx
g x+
sin xdx +
a
a
Z b
+
g(x) sin xdx
2I =
b/
=I1 + I2 + I3
|2I| |I1 | + |I2 | + |I3 |,
40
where I1 , I2 and I3 are the first, second (the stuff inside the parentheses) and third term of
the previous line.
Well now proceed to show that each of I1 , I2 and I3 can be made arbitrarily small. For
I1 and I3 this is due to our integration being over a tiny interval. For I2 , however, this is
due to the cancelation that we discussed in our graphical proof. Indeed, the transformation
x = t + / is exactly the linear shift that moves us over half a period, where the integrand
has the opposite sign, and cancellations can be expected. This is illustrated in Fig. 3.3.
y
2
0.02
0.04
0.06
0.08
0.1
Figure 3.3: Shifting by / moves you one half-period of the trig function. Shown are the
function y = g(x) sin x, with g(x) = x3 x + 2 and = 100. Also shown are g(x) and
g(x), which play the role of modulating amplitudes.
Lets look at I3 .
Z b
|I3 | =
g(x) sin xdx
b/
Z b
|g(x)|dx
b/
M ,
where
M = max |g(x)|.
x[a,b]
41
|I1 | M .
Lastly, we look at I2 . In doing so, well assume that is sufficiently large, so that
b / > a. Then
Z
b/
(g(x) g(x + /)) sin xdx
|I2 | =
a
b a
max |g(x) g(x + /)| .
x[a,b/]
2M
|2I|
Since a continuous function on a bounded closed interval is uniformly continuous, this last
maximum goes to zero as . Thus
lim I = 0,
which is what we had to prove. The cases with or with cos x are all done similarly,
and require no more work.
Now well look at what needs to be done if there are discontinuities at the endpoints of
the interval. Notice that the preceding part did not require us to use that
Z
|g(x)|dx < .
a
Well use this now: the existence of this integral allows us to control how bad the singularities
in g(x) at the endpoints a and b are allowed to be. Since the integral converges, we have
that for any > 0, we can find a > 0 such that
Z
a
a+
|g(x)|dx < ,
2
Z
and
|g(x)|dx < .
2
b
42
Then (using our newly established result for the continuous case)
Z a+ Z b Z b
Z b
g(x) sin xdx =
+
+
g(x) sin xdx
a
lim
a+
a+
Z
lim
= lim
lim 2
2
= .
Since can be chosen arbitrarily small, we get
Z b
lim
g(x) sin xdx = 0,
under the conditions stated in the theorem. This completes the proof of the RiemannLebesgue lemma.
2.
Lets look at Fourier series with L = , which simplifies some of the expressions a bit. We
now have
a0 X
+
(an cos nx + bn sin nx) ,
S(x) =
2
n=1
so that the N -th partial sum SN is
N
SN (x) =
a0 X
+
(an cos nx + bn sin nx) .
2
n=1
Lemma 2 (Partial sum lemma) Let SN (x) be the N -th partial sum of the 2-periodic
Fourier series obtained from the function f (x), i.e.,
Z
Z
1
1
an =
f (x) cos nxdx, bn =
f (x) sin nxdx,
43
for n = 0, 1, . . .. Then
1
SN (x) =
2
f (x + u)
sin(2N + 1) u2
du.
sin u2
a0 X
+
(an cos nx + bn sin nx)
2
n=1
Z
Z
N Z
X
1
1
1
=
f (t)dt +
f (t) cos ntdt cos nx +
f (t) sin ntdt sin nx
2
n=1
"
#
Z
N
X
1
1
(cos nt cos nx + sin nt sin nx) dt
=
f (t)
+
2 n=1
#
"
Z
N
1
1 X
=
cos n(t x) dt.
f (t)
+
2 n=1
SN (x) =
1
1
= +<
2
ei(tx)/2 ei(tx)/2
"
#
1
1
ei(tx)(N + 2 ) ei(tx)/2
= +<
2
2i sin tx
2
h
i
1
1
i(tx)(N + 12 )
i(tx)/2
<
i
e
=
e
2 2 sin tx
2
1
1
tx
tx
= +
sin(2N + 1)
sin
2 2 sin tx
2
2
2
tx
1 sin(2N + 1) 2
=
.
2
sin tx
2
44
Ah, the wonders of complex numbers! In the above < denotes the real part of the expression.
Substituting this back in to the expression for SN (x), we get
Z
sin(2N + 1) tx
1
2
f (t)
dt.
SN (x) =
2
sin tx
2
Notice that the integrand is periodic with period 2. As the integral is over an entire period,
it is not relevant what its upper and lower integration bounds are. Letting u = t x, we get
Z
sin(2N + 1) u2
1
SN (x) =
f (x + u)
du,
2
sin u2
where we did not bother with the bounds of integration, because of our observation above.
.
This is what we had to prove.
The partial sum lemma provides an elegant, compact form for a partial sum of a Fourier
series. The form even appears in a format that tempts us to use the Riemann-Lebesgue
lemma as N . Thats not allowed though: the limit as u 0 of the integrand is
(2N+1)f(x), which is not of the form suitable for the Riemann-Lebesgue lemma. Nevertheless, the partial sum lemma will be useful later on (otherwise it wouldnt be much of a
lemma).
3.
You may have seen the Weierstrass approximation theorem before, probably in a numerical
analysis class. The theorem is usually phrased in its polynomial form, so that it states that
every continuous function on an interval [a, b] can be approximated arbitrarily close by a
polynomial (the Weierstrass polynomial) so that the absolute error of the approximation
does not exceed > 0 anywhere in [a, b]. The degree of the Weierstrass polynomial depends
on , of course.
In this section, we state and prove the theorem in its trigonometric form.
Theorem 1 (Weierstrass approximation) Let > 0 be arbitrary, and f (x) is a 2periodic function, continuous on [, ]. There exists a trigonometric polynomial T (x) such
that |f (x) T (x)| < , for all x [, ].
Note that a trigonometric polynomial is any expression of the form
n
0 X
+
(k cos kx + k sin kx).
2
k=1
In other words, a trigonometric polynomial is an expression that has a truncated Fourier
series.
Proof. Define
T (x) = cn
cos2n
u
2
f (x + u)du.
45
Using the addition and subtraction formulas for cosine, this may be reduced to an expression
with powers of cos x and sin x. These powers can come out of the integral, which only cares
about t. Thus T (x) is a combination of powers of sines and cosines, with constant coefficients.
These powers of sine and cosine are now reduced to sines and cosines of integer multiples of
x, using the multi-angle formulas. Thus, indeed, T (x) is a trigonometric polynomial.
Now, we choose cn so that
Z
cn
u
cos2n du = 1.
2
With a little effort and a lot of calculus, the integral may be done. You find
2 4 6 2n
2 1 3 (2n 1)
2 4
2n 2 2n
=
3 5
2n 1 2
n
.
cn =
Then
Z
2n u
cos
|T (x) f (x)| = cn
f (x + u)du f (x)
2
Z
Z
u
u
2n
2n
f (x + u)du cn
cos
f (x)du
= cn
cos
2
2
Z
u
= cn
(f (x + u) f (x)) du
cos2n
2
Z
u
cn
cos2n
|f (x + u) f (x)| du
2
As is seen from Fig. 3.4, the kernel function cos2n (u/2) is sharply peaked at zero, but goes
to zero rapidly elsewhere, for sufficiently high n. Because of this, we split the integral up
in three parts: a central part over a small interval where the kernel is not small, and two
outlying parts where the kernel is small. Well use different techniques to control these
different parts.
Z
|T (x) f (x)| cn
Z
+
cn
2n
cos
u
2
cos2n
u
2
|f (x + u) f (x)|du+
|f (x + u) f (x)|du,
46
0.8
0.6
0.4
0.2
Z
u
cn
cos2n
du
2
2
= .
2
Next we look at the the other, noncentral parts. Because f (x) is uniformly continuous,
there exists a constant M such that
|f (x)| M.
Then
Z
cn
Z
+
2n
cos
u
2
Z
|f (x + u) f (x)|du cn
Z
+
cos2n
u
Z
2M cn 2
cos2n
u
2
du
2M du
47
cos
du
2
2n
,
4M n cos
2
4M n
2n
where the next-to-last step is based on cos(u/2) being a decreasing function on [0, ]. Notice
that
2n
n cos
0,
2
as n . Thus we can choose n so large that
2n
4M n cos
.
2
2
Putting everything together gives
+ = .
2 2
This finishes the proof of the Weierstrass approximation theorem.
|T (x) f (x)|
4.
2
f
(x)dx
is finite. Then the n-th partial sum of the Fourier series provides the best least
square approximation to f (x) of all trigonometric polynomials of order n.
Proof. Let
0 X
Tn =
+
(k cos kx + k sin kx) ,
2
k=1
a trigonometric polynomial of order n. Then
Z
Z
Z
2
2
[f (x) Tn (x)] dx =
f (x)dx 2
f (x)Tn (x)dx +
Tn2 (x)dx.
k=1
Z
Z
Z
n
n
X
X
0
=
f (x)dx +
k
f (x) cos kxdx +
k
f (x) sin kxdx
2
k=1
k=1
n
X
= 0 a0 +
(k ak + k bk ) ,
2
k=1
48
!
!
Z
n
n
0 X
0 X
+
+
(m cos mx + m sin mx) dx
=
(k cos kx + k sin kx)
2
2
m=1
k=1
n
n
X
X
02
2
= 2 +
k +
k2
4
k=1
k=1
n
X
02
+
k2 + k2 ,
=
2
k=1
and weve used the orthogonality of the cosines and sines, as seen at the beginning of this
chapter.
Putting things together, we get
Z
1
[f (x) Tn (x)]2 dx
Z
n
n
X
1 2
a20 X 2
=
(k ak + k bk ) +
f (x)dx 0 a0 2
+
k + k2
2
k=1
k=1
Z
n
n
2
2
X
(a0 0 )
a0 X
1
2
2
2
f (x)dx +
+
(k ak ) + (k bk )
(a2k + b2k ).
=
2
2
k=1
k=1
Some of the terms in this expression depend on f (x) only (the first, third, and last term)
and are not affected by changes in Tn (x). The other terms are all positive, since they are
sums of squares. These terms do change as Tn (x) is changed. It follows that the expression
is minimized when these terms are. But these terms can all be made to vanish, exactly by
choosing k = ak and k = bk , for k = 0, 1, . . .. Doing this makes
Tn (x) = Sn (x),
the n-th partial sum of the Fourier series of f (x), proving the theorem.
Thus the optimal least-square approximation to any square-integrable periodic function
on a finite interval in the family of trigonometric polynomials of order n is given by the n-th
partial sum of the Fourier series of f (x). Note that we did not need to impose continuity of
f (x).
The least-square approximation theorem has a few important consequences. Letting
Tn (x) = Sn (x), we get
Z
Z
n
1
1 2
1 2 X 2
2
(ak + b2k ).
[f (x) Sn (x)] dx =
f (x)dx a0
2
k=1
49
1 2 X 2
ak + b2k .
f (x)dx a0 +
2
k=1
This inequality is known as Bessels inequality. The reason that we cant write an equality
(without introducing more conditions on f (x), see below) is that we cant conclude without
extra work that the least-square error term will vanish in the limit.
Using the Weierstrass approximation theorem, we can do better than Bessels inequality.
Of course, this requires us to impose that f (x) is continuous on [, ]. We get
Z
Z
2
[f (x) Sn (x)] dx
[f (x) T (x)]2 dx
where T (x) is the Weierstrass polynomial of order n. The inequality holds because the leastsquare error among all trigonometric polynomials of order n is minized by Sn (x). But we
know that for a sufficiently large n we have
|f (x) T (x)| < ,
uniformly in x, by the Weierstrass approximation theorem. This results in
Z
[f (x) Sn (x)]2 dx 22 .
lim
[f (x) Sn (x)]2 dx = 0,
provided f (x) is continuous and square integrable. This allows us to strengthen Bessels
inequality to obtain Parsevals relation: if f (x) is continuous and 2 periodic, then
1
1 2 X 2
f (x)dx = a0 +
ak + b2k .
2
k=1
50
5.
Before we move on to one of the main results of Fourier analysis (Dirichlets theorem; next
section), we need to recall some results from real analysis.
If the functions {fn (x), n = 1, 2, . . .} are all continuous, and fn (x) f (x) uniformly
in x as n , then f (x) is continuous.
Another way to phrase this result is: the uniform limit of a sequence of continuous
functions is continuous.
If gn (x) g(x), uniformly on x [a, b] as n , then
Z
lim
Z
gn (x)dx =
g(x)dx.
Or, rephrased: the limit of the integral is the integral of the limit.
These results can be restated in terms of series, by identifying fn (x) with partial sums
(check this!):
If the functions {g
n (x), n = 1, 2, . . .} are all continuous, and
P
convergent, then
k=1 gk (x) is continuous.
P
Let
k=1 gk (x) be uniformly convergent on x [a, b], then
lim
or
Z bX
n
Z
gk (x)dx =
a k=1
Z
X
k=1
gk (x)dx =
lim
n
X
a n k=1
Z bX
Pn
k=1
gk (x) is uniformly
gk (x)dx,
gk (x)dx.
a k=1
In other words, under these conditions, we can integrate a series term by term.
An important consequence of these results is the following theorem, which gives us conditions under which we can differentiate a series term by term.
Theorem 3 (Term-by-term differentiation)
If {gn (x), n = 1, 2, . . .} are continuously
P
P
0
differentiable on an interval I, and (i)
g
(a)
converges at some a I, (ii)
k=1 k
k=1 gk (x)
is uniformly convergent on I, then
!0
X
X
gk (x) =
gk0 (x)
k=1
on I.
k=1
51
Proof. Consider
gk0 (x).
k=1
Z x
X
X
X
X
0
gk (t)dt =
gk0 (t)dt =
(gk (x) gk (a)) =
gk (x)
gk (a),
a
k=1
k=1
k=1
k=1
k=1
xX
Z
gk (x) =
a
k=1
gk0 (t)dt
k=1
gk (a).
k=1
!0
gk (x)
k=1
gk0 (x),
k=1
If the series k=1 Mk is convergent, then the series k=1 gk (x) is absolutely and uniformly
convergent on S.
6.
Dirichlets theorem
Now were armed and ready to go. Lets give a first answer to the question of convergence
of a Fourier series.
Theorem 4 (Fourier series convergence) If the Fourier series S(x) of a 2-periodic
continuous function f (x) converges uniformly for all x [, ], then S(x) = f (x).
Proof. Let
a0 X
S(x) =
+
(ak cos kx + bk sin kx) ,
2
k=1
the Fourier series obtained from f (x). Then clearly S(x) is 2-periodic, and S(x) is continuous, as the uniform limit of a sequence of continuous functions (the partial sums).
Now consider S(x) cos nx. This series is also uniformly convergent, so it can be integrated
term by term. Similarly with S(x) sin nx. This way we find that the Fourier coefficients of
S(x) are identical to those of f (x). Since both of f (x) and S(x) are continuous and periodic,
and their Fourier coefficients are equal, we have that f (x) = S(x), as a consequence of
Parsevals relation.
As immediate consequence of the Weierstrass M -test is the following.
52
and uniformly convergent to f (x) if k=1 |ak | and k=1 |bk | are convergent.
Next we come to Dirichlets theorem, which is one of the most important results in Fourier
analysis.
Theorem 5 (Dirichlet) Let f (x) be 2-periodic and piecewise smooth. Then its Fourier
series converges for each x [, ]: for each fixed x
lim Sn (x) = S(x) =
f (x ) + f (x+ )
,
2
where
f (x ) = lim f (x h),
h>0
Further, if f (x) is continuous, then S(x) converges absolutely and uniformly on any closed
subinterval.
Proof. According to the partial sum lemma
Z
sin(2n + 1) u2
1
f (x + u)du.
Sn (x) =
2
sin u2
Using f (x) = 1 in this identity gives
Z
1 sin(2N + 1) u2
1=
du
2 sin u2
Z
1 sin(2N + 1) u2
1
=
du,
2
0
2 sin u2
Z
0
sin(2N + 1) u2
+
f (x
0 ) + f (x0 ) du.
u
2 sin 2
Also,
Sn (x0 ) =
=
=
=
Z
sin(2n + 1) u2
1
f (x0 + u)du
2
sin u2
Z
Z
1 sin(2n + 1) u2
1 0 sin(2n + 1) u2
f (x0 + u)du +
f (x0 + u)du
0
2 sin u2
2 sin u2
Z
Z
1 sin(2n + 1) u2
1 0 sin(2n + 1) u2
f (x0 + u)du
f (x0 u)du
0
2 sin u2
2 sin u2
Z
1 sin(2n + 1) u2
[f (x0 + u) + f (x0 u)] du.
0
2 sin u2
53
+
=
f
(x
+
u)
+
f
(x
u)
f
(x
)
f
(x
)
du
0
0
0
0
0
2 sin u2
Z
f (x0 u) f (x
1 sin(2n + 1) u2
f (x0 + u) f (x+
0)
0)
+
du.
=
u
0
2 sin u2
u
u
54
Thus the Riemann-Lebesgue lemma applies, proving the first part of Dirichlets theorem.
To prove the second part of the theorem, assume that f (x) is 2-periodic, continuous
and piecewise smooth. It suffices to show that
k=1
|bk | < .
k=1
f (x)
dx
=
n
n
Z
1
=
f 0 (x) sin nxdx
n
1
= n ,
n
where n is a Fourier coefficient of f 0 (x). Using
1 2
a + b2
2
ab
we have
1
1
|an | = |n |
n
2
1
2
+ |n | ,
n2
or
1X
1X 1
+
|n |2
|an |
2
2
n
2
n=1
n=1
n=1
2 1 X
=
|n |2
+
12 2 n=1
< ,
by Bessels inequality, since
1
f 02 (x)dx
is finite.
P
0
The proof for
n=1 |bn | < is similar. If there are points of discontinuity in f (x) at
isolated points, the interval has to be split up in parts on which the function is continuous,
as in the proof of the Riemann-Lebesgue lemma.
.
3.2
55
In this section well introduce Fourier transforms in much the same way that we introduced
Fourier series. Since many of the results are quite similar, we wont present the results here
with the same level of rigor. Few proofs will be given.
Given a function f (x), the Fourier transform of f (x), for x R is defined as
Z
eix f (x)dx,
F[f (x)]() = f() =
A few notes:
The above definition of the Fourier transform is one of many youll find in the literature.
Theyre all equivalent, but minus signs, factors of 2, etc. may be in different places.
Im using the above definition, which may be the most commonly used.
From the Riemann-Lebesgue lemma (or an extension of it to the whole real line), it
follows that f() 0 as , provided f (x) is continuous, or has at most a finite
number of jump discontinuities.
Theorem 6 (Inverse of the Fourier transform) Let f (x) be continuous, piecewise smooth,
and absolutely integrable. Then
Z
1
f (x) =
eix f()d.
2
We wont prove this theorem, but a few notes are in order.
If f (x) is only piecewise continuous, then
f (x+ ) + f (x )
1
=
2
2
eix f()d.
56
Theorem 7 Let f (x) be piecewise continuous and absolutely integrable. Then f() is continuous on R.
Theorem 8 Let f (x), f 0 (x), f 00 (x), . . . be continuous and absolutely integrable. Then
F f (k) (x) () = (i)k F [f (x)] (),
for all k for which these conditions are satisfied. Here f (k) (x) denotes the k-th derivative of
f (x).
1.
In Rn , we have
f() =
ei x f (x)dx,
Rn
where x Rn , and = (1 , 2 , . . . , n ). This form is valid whenever the integral exists. Its
inversion is given by
Z
1
ei x f()d,
f (x) =
(2)n Rn
for f (x) continuous, piecewise smooth, and absolutely integrable.
We wont go into the details of Fourier transforms to the same extent as we did for Fourier
series. As stated before, the proofs are very similar, with the role of uniform convergence
of infinite series taken over by uniform convergence of improper integrals. Lets state the
theorems from real analysis that are required to do these proofs, as theyll be useful later.
2.
is uniformly convergent for all x [a, b], then F (x) is continuous for x [a, b].
57
Theorem 11 (Switching the order of integration) Under the circumstances of the previous theorem (uniform convergence of F (x), piecewise continuity in y) we have that
Z Z b
Z b Z
g(x, y)dx dy.
g(x, y)dy dx =
a
Using a similar proof to that for switching infinite sums and derivatives we can prove a
similar theorem for improper integrals and differentiation.
Theorem 12 (Switching derivatives and improper integration) Suppose that g(x, y)
and gx (x, y) are continuous for < x < , y R, then
Z
Z
g
d
(x, y)dy,
g(x, y)dy =
dx
x
for x (a, b), provided both integrals exist, and provided the one on the right is uniformly
convergent on bounded, closed subintervals of (, ).
Example: As an example of what these theorems can do for us, well calculate
Z
sin y
I=
dy.
y
0
Well use the different theorems many times over when we start solving partial differential
equations in the following chapters, but this example may keep the dogs down for now.
Usually, this integral is done using methods of complex analysis, especially the residue
theorem. Well do it a different way here.
Consider
Z
sin y
dy,
L(x) =
exy
y
0
for x 0. We wish to compute L(0). In terms of the preceding theorems we have
g(x, y) = exy
sin y
,
y
and
gx (x, y) = exy sin y.
Both of these are continuous on the intervals considered. Thus, for x > 0 (the above theorem
does not allow us to include x = 0) we have
Z
1
0
L (x) =
exy sin ydy =
.
1 + x2
0
This last integral is easily done using standard calculus methods, such as good ol integration
by parts. Another way is to use Laplace transforms.
58
arctan x.
2
3.3
Lets look at an example where the standard methods involving blindfolded term-by-term
differentiation etc fail.
Consider the following conducting wave slab problem:
Eyy + Ezz + K 2 E = i0 J.
Here E is the x-component of the electric field and J is a volume source, such as a current,
prescribed inside the area of the slab, see Fig. 3.5. This problem describes the linear dynamics
of the electric field in a planar wave slab, used for communication purposes such as antennas.
Assume that the volume source J is given as
J=
Bk (z) sin
k=1
ky
L
.
As indicated in Fig. 3.5, the electric field E vanishes on the edges of the wave conductor,
except at perhaps a few places where there is a hole in the surrounding (insulating) apperture.
At those places we may assume the field is known. For now well consider the case of point
sources on the boundary, for instance in the form of a Dirac delta function.
Because of the boundary conditions and the prescribed form of J, it is reasonable to
expand E in a Fourier sine series:
E=
X
k=1
Ek (z) sin
ky
L
,
59
z
E=0
E=0
E=0
E=0
0
E=0
for 0 < x < L. A typical way to proceed would be to take as many derivatives of this
expression as are required for the partial differential equation. In our case, wed compute
Ezz
Eyy
ky
=
,
L
k=1
2
X
k
ky
=
Ek (z)
sin
.
L
L
k=1
Ek00 (z) sin
60
Fk (z) sin
k=1
ky
L
.
Our goal is now to find a relationship between the Fourier coefficients {Fk (z)} of Eyy and
the Fourier coefficients {Ek (z)} of E. By definition of the Fourier coefficients we have
Z
ky
2 L
Eyy sin
Fk (z) =
dy
L 0
L
y=L
!
Z
2
ky
ky
k L
=
Ey cos
Ey sin
dy
L
L y=0
L 0
L
Z
ky
2k L
Ey cos
dy
= 2
L 0
L
y=L Z L
!
ky
k
ky
2k
+
E
sin
= 2 E cos
dy
L
L y=0
L
L
0
=
k22
2k
k
E(L,
z)(1)
E(0,
z)
2 Ek (z).
L2
L
We see that this relationship is different from the one we find using term-by-term differentiation. In that case we find
k22
Fk (z) = 2 Ek (z).
L
The extra terms are all due to the boundary effects at the holes in the apperture, as expected.
Since the electric field there is supposed to be known, the relation
Fk (z) =
k22
2k
k
E(L,
z)(1)
E(0,
z)
2 Ek (z)
L2
L
k22
2k
00
2
k
sin
Ek + K Ek 2 Ek 2 E(L, z)(1) E(0, z)
L
L
k=1
X
ky
=
Bk sin
i0 .
L
k=1
ky
L
3.4. EXERCISES
61
Here the prime denotes differentiation with respect to z. Using the orthogonality of the
different sines, we equate the coefficients of like terms. This gives
k22
2k
2
00
Ek + K 2 Ek = i0 Bk + 2 E(L, z)(1)k E(0, z)
L
L
The first term on the right-hand side is due to the volume source in the problem, whereas
the other terms on the right are a consequence of the boundary sources in the holes of the
insulating apperture. This is the first instance of a few that well encounter in this course of
the boundary conditions entering the governing equations as source terms.
The above ordinary differential equation is now solved as a superposition of a homogeneous and a particular solution. In this example the term-by-term differentiation caused
wrong results, but the difficulties were relatively easy to circumvent. The important part is
to be aware of a potential problem like this one.
3.4
Exercises
1. We have seen that the Fourier series of f (x) can be termwise differentiated if f (x)
is continuous and piecewise smooth. Apart from functions whose Fourier series has
a finite number of terms, can you think of any other functions in the mathematical
compendium whose Fourier series can be differentiated ad exhaustiam? Essentially this
question is asking to write down an f (x) whose Fourier coefficients decay faster than
any power of n, as taking derivatives multiplies these Fourier coefficients by powers of n.
Fourier series are quite different from Taylor series in this way: we know many Taylor
series whose derivatives have Taylor series that are equally well converging. Examples
are exponentials, hyperbolics, trig, and others in their domain of convergence. Usually
the worst thing that happens when we take a termwise derivative of a Taylor series is
that the domain of convergence becomes a bit smaller. In the case of a Fourier series it
is most common that the domain of convergence completely dissapears if we take too
many termwise derivatives. So: do we know any other nice periodic functions other
than the trig functions?
2. Invent a function defined by a Fourier series with an infinite number of terms that is
infinitely differentiable.
3. About the Riemann-Lebesgue lemma The Riemann-Lebesgue lemma can easily
be extended to improper integrals, using essentially the same assumptions as in Lemma
1, with a = and/or b = .
R
2
(a) Consider I() = ex /2 cos(x)dx. The R-L lemma states that lim I() =
0. (a) Whats the best estimate you can get for how fast I() approaches 0? (b)
Calculate I(), by obtaining a first-order differential equation for I(), and solving it. Having found I(), how fast does it approach 0? (c) Calculate the Fourier
2
transform of ex /2 .
62
2
2
sin(x )dx =
.
cos(x )dx =
2
One might use a similar argument as that for the cancelations in the R-L lemma
to argue the convergence of these integrals. Assuming the above result, show that
2
Z
+
2
sin(x ) cos(x)dx = cos
,
4
Chapter 4
Wave propagation in one spatial
dimension
In this chapter we investigate the one-dimensional wave equation
utt = c2 uxx .
As weve seen before, this is the prototype of a second-order hyperbolic equation. For
definiteness, well assume that c > 0.
4.1
1.
= u + u
= u + 2u + u
= c (u u )
= c2 (u 2u + u ) .
64
Here A and B are arbitrary functions of their arguments. Just like the general solution of a
second-order ordinary differential equation depends on two arbitrary constants, we see that
the general solution of this second-order partial differential equation with two independent
variables depends on two arbitrary functions of one variable.
Note that A(x + ct) represents a wave traveling to the left with velocity c, whereas
B(x ct) represents a wave going to the right with velocity c. We see that the wave
equation splits given initial data up in a left-going and a right-going part. This is illustrated
in Fig. 4.1.
t
c
x
Figure 4.1: The wave equation splits given initial data up in a left-going and a right-going
part.
2.
65
V (z)dz + c1 ,
0
where c1 is constant. Adding to this the equation for A(x) + B(x), we get
Z
1 x
V (z)dz + c1
2A(x) = U (x) +
c 0
Z x
1
1
c1
A(x) = U (x) +
V (z)dz + .
2
2c 0
2
Similarly, by subtracting the two equations we get B(x). Alternatively,
B(x) = U (x) A(x)
Z
1
1 x
c1
= U (x)
V (z)dz
2
2c 0
2
Z 0
1
c1
1
V (z)dz .
= U (x) +
2
2c x
2
Putting these results together results in
u(x, t) = A(x + ct) + B(x ct)
Z
Z
1
1 x+ct
c1 1
1 0
c1
= U (x + ct) +
V (z)dz + + U (x ct) +
V (z)dz
2
2c 0
2
2
2c xct
2
Z x+ct
U (x + ct) + U (x ct)
1
=
+
V (z)dz,
2
2c xct
which is dAlemberts solution of the wave equation with initial data specified on the real
line. It generalizes to any number of spatial dimensions without much trouble. If follows
from DAlemberts solution that the wave equation with initial data defined on the entire
real line is well-posed.
3.
From the characteristic picture discussed in Chapter 2, but also from the preceding calculation, it is clear that the influence of any point (x0 , y0 ) is restricted to a cone emanating from
this point with slopes c, as shown in Fig. 4.2. This cone is called the cone of influence of
(x0 , t0 ): no matter the size of the disturbance at (x0 , t0 ), its effect will never be felt outside
of its cone of influence.
Similarly, the point (x0 , t0 ) does not depend on any disturbances outside of its cone of
dependence, given by the same slopes c, but for t < t0 . The cone of dependence of
(x0 , t0 ) is the collection of all points that have (x0 , y0 ) in their cone of influence. This is also
illustrated in Fig. 4.2.
66
II
t0
I
x
x0
Figure 4.2: The cone of dependence (region I) and the cone of influence (region II) of the
point (x0 , t0 ).
4.2
utt = c2 uxx
u(x, 0) = f (x)
ut (x, 0) = g(x)
u(0, t) = 0 = u(L, t)
This problem is an initial-boundary-value problem, since it involves both initial and boundary
conditions. It is also a homogeneous problem because there is no forcing in the partial
differential equation (making u = 0 a solution, albeit not a very interesting one), and the
boundary conditions are zero as well.
1.
Separation of variables
Well solve this problem using the standard separation of variables method. The idea is
to look for simple solutions that we will superimpose later.
Lets try to find solutions of the form
u(x, t) = X(x)T (t).
67
X 00
T 00
=
.
c2 T
X
The left-hand side of this equation is a function of t alone, whereas the right-hand side
depends only on x. These two sides can only be equal if both sides are constant. Thus
T 00
X 00
=
= K,
c2 T
X
where K is a constant. This implies the two ordinary differential equations
X 00 KX = 0,
T 00 c2 KT = 0.
and
Lets look at the first equation first, since the boundary conditions look simpler than the
initial conditions. Using the boundary conditions, we impose
X(0) = 0, and X(L) = 0.
So, we have to solve the ordinary differential equation boundary-value problem
00
X KX = 0
X(0) = 0,
X(L) = 0.
This is an eigenvalue problem for the function X(x), with eigenvalue K.
There are several ways to establish that K must be negative to have nontrivial solutions to
this problem. The most obvious one is to go through all the possibilities separately (K > 0,
K = 0, and K < 0) and see that only the last option gives nontrivial (i.e., nonzero) solutions
of the eigenvalue problem. However, that approach requires that we are able to solve the
ordinary differential equation for its general solution. Below we use an approach that does
not require that we know a single solution. In that sense it is more general. It is the first hint
of what well refer to as an energy method (due to its similarities with physical systems,
where the computed quantity corresponds to the total energy of the system).
Multiply the equation by X to get
XX 00 KX 2 = 0
XX 00 KX 2 dx = 0
X 02 + KX 2 dx = 0
X 02 + KX 2 dx = 0,
where weve used integration by parts and the boundary conditions for X(x).
68
X(0) = 0
X(L) = 0
c1 = 0,
c2 sin L = 0.
n
, n N.
L
Note that n = 0 results in the trivial solution with K = 0, not surprisingly. Negative values
of n lead to the same value of K and can be ignored.
The eigenfunctions corresponding to these eigenvalues
n2 2
, n = 1, 2, . . .
L2
are
nx
Xn (x) = sin
, n = 1, 2, . . . ,
L
where we have chosen c2 = 1, since multiples of eigenfunctions remain eigenfunctions.
Kn = 2n =
69
Weve given Tn (t) a subindex n, to indicate that a different function will be obtained depending on which eigenvalue we use. We have
cnt
cnt
+ n sin
, n = 1, 2, . . . ,
Tn (t) = n cos
L
L
where n and n are arbitrary constants.
In summary, we find the special class of solutions, one for each eigenvalue,
nx
cnt
cnt
n cos
+ n sin
, n = 1, 2, . . . .
un (x, t) = Xn (x)Tn (t) = sin
L
L
L
Since the wave equation is linear, we get a more general solution by superimposing all the
special solutions we just found. You saw this coming a mile away, right?
u(x, t) =
un (x, t)
n=1
sin
n=1
nx
L
n cos
cnt
L
+ n sin
cnt
L
.
In order to determine the constants n and n , we get to (finally) use the initial conditions.
At t = 0, we get for u(x, 0)
u(x, 0) =
n sin
n=1
nx
L
= f (x),
by the first initial conditions. Next we need to take a derivative of u(x, t) with respect to t,
and evaluate this at t = 0. Lets not worry for now about taking the derivative inside the
sum. Thats the topic of the next section. We get
ut (x, 0) =
X
n=1
nx
cn
sin
= g(x).
L
L
Both series are Fourier sine series, requiring us to use odd extensions of the given initial
conditions. The constants n and n are nothing but Fourier coefficients, and we have
Z
nx
2 L
n =
f (x) sin
dx,
L 0
L
and
Z
nx
cn
2 L
n
=
g(x) sin
dx
L
L 0
L
Z L
nx
2
n =
g(x) sin
dx,
nc 0
L
70
2.
Justification
Theres a number of things we need to justify: can we switch infinite sums and derivatives?
Can we use Fourier expansions for f (x) and g(x)? Well now justify the above solution
process by taking the solution we obtained formally, and showing under what conditions it
is actually a solution.
We have
nx
X
cnt
cnt
n cos
+ n sin
u(x, t) =
sin
L
L
L
n=1
In order to substitute this expression into the differential equation, we need to take partial
derivatives of it, both with respect to x and with respect to t. This requires us to switch
infinite sums and derivatives. If allowed, wed get
X
n
nx
cnt
cnt
ux =
cos
n cos
+ n sin
,
L
L
L
L
n=1
nx
X
n 2
cnt
cnt
uxx =
sin
n cos
+ n sin
.
L
L
L
L
n=1
These differentiations are valid provided the obtained series are uniformly convergent. By
the Weierstrass M -test, it suffices to find a constant C such that
2 2
n n C ,
L
n2
()
n2 2
,
2
L
n2
71
for n = 1, 2, . . .. Note that if these conditions are satisfied that the series of ux is automatically uniformly convergent, as its terms would approach zero as C/n3 . The calculation of
ut and utt leads to similar series. You should also verify that a similar requirement leads to
the validity of the the Fourier expansions of f (x) and g(x).
Next, we wish to translate the conditions () into conditions on the input data of the
problem, namely f (x) and g(x). If we manage to do this, we will have obtained a criterion
for the solvability of our original problem that will only involve easily verifiable conditions
on f (x) and g(x)!
Lets look at n . We have (n = 1, 2, . . . in what follows)
2
n =
L
f (x) sin
0
nx
L
dx.
If we impose the compatibility conditions (requiring the initial conditions to match the
boundary conditions at the edges) f (0) = 0, f (L) = 0, then
Z L
nx
2 L
f 0 (x) cos
dx
n =
L n 0
L
Z L
nx
2
=
f 0 (x) cos
dx
n 0
L
Assuming that f 0 (x) is integrable has already bought us one factor of 1/n. Were getting
closer! If we impose the additional conditions f 00 (0) = 0, f 00 (L) = 0, we get
Z L
nx
2L
f 00 (x) sin
dx
n = 2 2
n 0
L
Z L
nx
2L2
000
= 3 3
f (x) cos
dx
n 0
L
Z L
nx
2L3
= 4 4
f (iv) (x) sin
dx.
n 0
L
This immediately gives rise to
Z L
2L3
|n |
|f (iv) (x)|dx
(n)4 0
2 2
Z L
n n
C
2L
|f (iv) (x)|dx,
L2 n 2 = 2 n 2
0
|f (iv) (x)|dx.
A similar calculation places restrictions on g(x). All this leads to the following theorem.
72
Theorem 13 Let f (x) have a fourth derivative which is absolutely integrable on [0, L], and
let g(x) have a third derivative which is absolutely integrable on [0, L], such that
f (0) = f (L) = f 00 (0) = f 00 (L) = 0
g(0) = g(L) = g 00 (0) = g 00 (L) = 0.
Then the formal solution obtained using separation of variables solves the initial-boundaryvalue problem
utt = c2 uxx
0 < x < L, t > 0
u(x, 0) = f (x)
0xL
.
ut (x, 0) = g(x)
0xL
u(0, t) = 0 = u(L, t)
t0
As announced, the theorem list conditions on the input functions that are easy to check.
If these conditions are satisfied, we may proceed with our regular methods to find a formal
solution. The theorem then guarantees the validity of this formal solution.
3.
Well-posedness
We could study well-posedness for the same initial-boundary-value problem as in the previous
section, but what would be the fun of that? Lets look at a slightly more general problem
instead. Consider the problem
u(x, 0) = f (x)
0xL
.
u
(x,
0)
=
g(x)
0
xL
u(0, t) = 0 = u(L, t)
t0
Here k > 0, representing the damping in a string. This problem can be solved in an almost
identical way as the previous one, using separation of variables. Lets consider its wellposedness instead.
Existence of a solution
As stated, using separation of variables, a solution to this problem can be constructed.
Perhaps we have to impose some regularity conditions on f (x) and g(x) to make this rigorous.
So be it. Thus a solution exists.
Uniqueness of a solution
Having established that a solution exists, lets prove that it is unique. Proving uniquess of a
solution is almost always done the same way: by contradiction. Well assume theres another
solution, which is of course assumed to be different. If u1 (x, t) and u2 (x, t) represent our two
distinct solutions, what is the problem that is satisfied by v = u1 u2 ?
73
v(x, 0) = 0
0xL
.
vt (x, 0) = 0
0xL
v(0, t) = 0 = v(L, t)
t0
This is a somewhat easier problem than the original one, since the initial conditions are
homogeneous. Well now show that the only solution of this problem is the zero solution
v(x, t) = 0, so that u1 (x, t) = u2 (x, t), proving uniqueness of the solution of the original
problem. We cant use separation of variables to show that v = 0. The whole point is to
argue that the solution produced by separation of variables is the only solution. We need to
rule out that another method might produce another solution.
74
Here
1 2 c2 2
E(t) =
v + vx dx
2 t
2
0
can be interpreted as the total energy of the system, comprised of both kinetic (first term)
and potential (second term) energy. We have just shown that the energy is always decreasing
as a consequence of the damping, implying that the energy at any time t > 0 is less than or
equal to its initial amount.
At t = 0, we have v(x, 0) = 0, from which vx (x, 0) = 0. The other initial condition states
that v(x, 0) = 0. Thus,
Z L
c2 2
1 2
v (x, 0) + vx (x, 0) dx = 0.
E(0) =
2 t
2
0
Z
Thus
E(t) 0,
for all t 0. Since the integrand of the energy integral is always non-negative, this is possible
only if
1 2 c2 2
v + vx 0 vx 0, vt 0 v(x, t) = constant = 0,
2 t
2
where the last equality follows from the boundary or initial conditions. Thus
u1 (x, t) = u2 (x, t),
for all x [0, L] and all t 0. Thus the solution of the original problem is unique.
1
No! Integration by parts and the boundary conditions are our main weapons!
75
cnt
cnt
1n cos
+ 1n sin
sin
u 1 u2 =
L
L
L
n=1
nx
X
cnt
cnt
sin
2n cos
+ 2n sin
L
L
L
n=1
nx
X
cnt
cnt
=
sin
(1n 2n ) cos
+ (1n 2n ) sin
L
L
L
n=1
|u1 u2 |
nx
(|1n 2n | + |1n 2n |) ,
n=1
where
1n 2n
2
=
L
(f1 f2 ) sin
nx
dx,
and
1n 2n
2
=
nc
(g1 g2 ) sin
0
nx
L
dx.
Lets assume that solutions exist corresponding to the input data f1 (x) and g1 (x), as
well as to the f2 (x) and g2 (x). Otherwise the above examinations dont make much sense.
Assuming the existence of solutions, we get to impose the assumptions of our theorem,
namely that f1 (x), f2 (x) (g1 (x), g2 (x)) has four (three) absolutely integrable derivatives.
Furthermore, these functions and their second derivatives vanish as the endpoints x = 0 and
76
Similarly,
1n 2n =
=
=
|1n 2n |
Z L
nx
2
dx
(g1 g2 ) sin
nc 0
L
Z L
nx
2
L
0
0
(g1 g2 ) cos
dx
nc n
L
0
Z L
nx
2L
0
0
(g
g
)
cos
dx
2
n2 2 c2 0 1
L
Z L
2L
|g10 g20 |dx.
2
2
2
n c 0
(|1n 2n | + |1n 2n |)
n=1
Z L
Z L
X
2L
2L
00
00
00
00
|f1 f2 |dx + 2 2
|f1 f2 |dx
n2 2 0
n 0
n=1
Z
Z L
X
X
2L L 00
1
1
2L
00
0
0
= 2
|f1 f2 |dx
+ 2 2
|g1 g2 |dx
2
0
n
c 0
n2
n=1
n=1
Z
Z L
L L 00
L
00
=
|f1 f2 |dx + 2
|g10 g20 |dx.
3 0
3c 0
We see that u1 u2 will stay small, as long as
Z L
Z
00
00
|f1 f2 |dx and
0
and
G1 = max |g10 g20 |
x[0,L]
4.3. EXERCISES
77
1
F2 + 2 G1
c
x[0,L]
and
G0 = max |g1 g2 |
x[0,L]
|f10 (0)
is small. Indeed, if
dx <
(f1 f2 )dx <
0
f20 (0)|
dx
which is small. Similar calculations show that |f1 f2 | and |g1 g2 | are small as well.
4.3
Exercises
1. Find a formal solution for the following IVP with boundary conditions:
u(x, 0, t) = u(x, b, t) = 0, 0 x a.
78
Chapter 5
General remarks about linear
evolution equations
An evolution equation is any partial differential equation (vector or scalar) that may be
solved for the time derivative of the dependent variables. Well only look at autonomous
evolution equations, which do not depend on the time or spatial variables explicitly. Thus,
in one spatial dimension such equations may be written as
ut = F (u, ux , . . . , uN x ),
where N is the highest order of the x-derivatives. Generalizing this to multiple space dimensions is straightforward. In this chapter, well look at linear evolution equations in some
detail. Such equations are of the form
ut = L(u, ux , . . . , uN x ),
where L is linear in its arguments. In other words
ut =
N
X
k ukx ,
k=0
and the k (k=0,1, . . . , N) are constants. Here u could be a vector function, in which case F
and L are vectors with the same dimension as u. In what follows, well consider the partial
differential equation with vanishing boundary conditions as x :
lim u = 0.
Well impose similar conditions on whatever number of derivatives we need, in order for
the solution formulas below to make sense. A similar treatment could be used for periodic
boundary conditions. In that case, wed find solutions expressed as Fourier series, whereas
now well need Fourier transforms.
79
80
5.1
In what follows, we work with the scalar case for simplicity, unless otherwise stated. Since
our linear evolution equation has constant coefficients, we may look for fundamental solutions
of the form
u = Aeikxit ,
where k and are constants, to be determined. Then
ut = iu,
ux = iku,
uxx = (ik)2 u,
5.2
Superposition
81
The factor 2 is chosen entirely for convenience, but could be omitted. In order to determine
A(k), we evaluate the above at t = 0. We get
Z
1
u(x, 0) =
A(k)eikx dk = F 1 [A(k)](x)
2
u(x,0)
u(x,t)
Fourier
Inverse Fourier
A(k)
A(k)ei(k ) t
Figure 5.1: The schematic of using Fourier transforms for solving initial-value evolution
partial differential equations. (2.1)
Thus, if our initial condition is absolutely integrable (so that its Fourier transform is
defined), we have found a unique solution of the initial-value problem for our given evolution
equation. What can be concluded from this? Specifically, what conclusions follow from the
knowledge of the dispersion relationship? The answer is quite a few. Well show in the
next section that nearly all important information about the equation is contained in the
linear dispersion relationship.
5.3
82
1
|A(k)|2 e2I (k)t dk,
=
2
since
u(k) = A(k)eik0 xi(k)t .
R
Lets call |u2 |dx the energy. Then the energy is conserved only if (k) is real
for real k. In this case there is no growth or decay as a consequence of loss of energy.
Well show shortly that decay is still possible, but a different process will determine it.
e) If there is a term
xn u
in the differential equation, then its contribution to the dispersion relationship is
i(ik)n = k n in+1 .
If n > 1 is odd, then this is real. Such a term is called dispersive. If n > 1 is even,
such a term results in an imaginary contribution to (k), and the term is dissipative
or amplifying, depending on its sign.
There are more consequences, but they require a bit more work. Lets give them their
own section.
5.4
83
If there is growth or decay due to dissipation, the asymptotics of the solution is dominated
by this behavior. If these two processes are absent, how can we determine the behavior of
the solution as t if it is either purely propagating, or dispersive?
Assume that = (k) = R (k) is real for real k. Then the solution of the problem is
given by (see above)
R
1
A(k)eikxi(k)t dk,
u(x, t) = R2
(k)
= c = constant,
k
84
which is large for large t. This leads to destructive interference of such neighboring waves,
unless x 0 (k)t, in which case phases remain approximately constant. In such directions no
destructive phase interference occurs. This is the direction in which information propagates.
More explicitly, consider a case where A(k) is sharply peaked at k = k0 , and decays
rapidly to zero away from k = k0 . We have
u(x, t) =
=
=
=
Z
1
A(k)eikxi(k)t dk
2
Z
eik0 xi(k0 )t
A(k)ei(kk0 )xi((k)(k0 ))t dk
2
Z
eik0 xi(k0 )t
0
2
A(k)eixki( (k0 )k+O((k) ))t dk
2
ik0 xi(k0 )t Z
e
0
A(k)eitk(x/t (k0 )+O(k)) dk.
2
k+k
|A(k)|2 dk.
kk
As Fig. 5.2 demonstrates this energy is spreading over a region that is growing linearly in t.
Since this energy is conserved, with density |u(x, t)|2 , we expect
|u(x, t)|2
1
t
u(x, t)
1
t1/2
Thus the amplitude of the solution should decay as a the inverse of a square root. This
estimate can be made rigorous, using the method of stationary phase or the method of
steepest decent. Weve barely hinted at the method of stationary phase here.
5.5. EXAMPLES
85
x=cg (k k)t
Figure 5.2: A simple cartoon illustrating the idea of different wavenumbers dominating along
different directions in the (x, t)-plane. (2.1)
5.5
Examples
as t .
Example: Consider the equation
ut = uxxx ,
known as the linear Korteweg-deVries equation. Its dispersion relationship is
i = (ik)3 = ik 3 = k 3 .
86
= k2,
k
5.6
cg = 2k.
Exercises
= x/tq ,
5.6. EXERCISES
87
(a) What are the conditions on p and q for this equation to have solutions of the form
u = tp f (), with = x/tq that satisfy an ordinary differential equation? (b) Solve
this differential equation with p = 1/2. (c) Show that the solution you have obtained
of the linear Schrodinger equation cannot be captured by Fourier transform methods.
88
Chapter 6
The heat equation
Lets turn our attention to parabolic problems, more specifically the heat equation
ut = auxx ,
with one spatial dimension, or
ut = au,
in an arbitrary number of dimensions. Here = x2 + y2 in two dimensions. This definition
of the Laplacian generalizes to three and more dimensions in an obvious manner.
6.1
ut
u(0, t)
u(L, t)
u(x, 0)
=
=
=
=
auxx ,
0,
0,
f (x),
X 00 cX = 0,
89
90
nx
= sin n x.
L
n
, n = 1, 2, . . . .
L
Turning to the equation for T , we find that
n =
Tn0
2
= 2n Tn = n ean t ,
aTn
again for n = 1, 2, . . .. Further, n denotes an arbitrary constant, possibly different for every
n.
A superposition of the fundamental solutions
2
n ean t sin n x.
n=1
X
n=1
n sin n x
2
n =
L
We have constructed a formal solution to the heat equation initial-boundary-value problem. This will be an actual solution, provided we can justify our Fourier series and the
termwise differentiation. This proceeds along similar lines to what we did for the wave equation, so it will not be repeated here. We should just remark that for any t > 0, it suffices to
have n < M (for some positive M ), without needing any decay with increasing n. This is
because the exponential decay in time provides the necessary decay of the Fourier coefficients
to guarantee absolute and uniform convergence of the Fourier series. In summary, we have
the following theorem.
Theorem 14 If f (x) is bounded and integrable, the formal Fourier series solution for u(x, t)
gives an actual solution of the heat equation initial-boundary-value problem. If f (x) has a
continuous second derivative with f (0) = 0 = f (L), then u(x, t) thus found is uniformly
convergent and continuous for t 0, 0 x L, and satisfies the initial condition.
6.2
91
ut
u(0, t)
u(L, t)
u(x, 0)
=
=
=
=
The function F (x, t) represents an external time- and space-dependent forcing. Suppose it
is given by
X
F (x, t) =
Fn (t) sin n x.
n=1
This form of F (x, t) implies that weve used an odd extension of F (x, t) for its Fourier series
representation. This is consistent with the boundary conditions: an external forcing at the
boundaries will likely result in a function u(x, t) which is not able to satisfy the boundary
conditions.
Since we know how to solve the homogeneous solution, it suffices to find one particular
solution. Once we have a particular solution, the general solution is given by the superposition of the homogeneous and particular solutions. Lets use variation of parameters to
construct a particular solution. Thus, well look for a particular solution of the form
up (x, t) =
cn (t)ean t sin n x.
n=1
The temporal functions cn (t) replace the previously constant n . They are our varying
parameters. Since they are to be determined, we might as well save ourselves some writing
and bundle them with the exponentials. This means we can look for
up (x, t) =
dn (t) sin n x.
n=1
X
n=1
d0n
sin n x = a
2n dn
sin n x +
n=1
Fn (t) sin n x.
n=1
n = 1, 2, . . . ,
92
so that
ean t dn
0
= ean t Fn (t)
Z t
2
a2n t
ean Fn ( )d
e
dn =
0
Z t
2
a2n t
ean Fn ( )d
dn = e
0
Z t
2
=
ean (t ) Fn ( )d.
0
Note that we only care about finding one particular solution, so that we may ignore any
constants during the integration process, as weve done here.
Thus the general solution is
u(x, t) =
X
n=1
a2n t
n e
sin n x +
X
n=1
Z
sin n x
ean (t ) Fn ( )d.
At this point, the n s may be determined from the initial conditions. Weve been smart
enough to choose a particular solution that vanishes at t = 0, so that the formulas for the
n in terms of the initial data are identical to those for the homogeneous case. As before,
we can find conditions on F (x, t) and f (x) (our input functions) that ensure that the formal
solution is indeed a solution.
6.3
Suppose that the previous problem is modified so that we not have homogeneous boundary
conditions, but instead some time-dependent behavior is fixed at the edges x = 0 and x = L.
Thus, we have a problem as given below:
u(0, t) = g(t),
t 0,
u(L,
t)
=
h(t),
t 0,
u(x, 0) = f (x),
0 x L.
Physically, it would appear we are putting in time-dependent heat sources or sinks. This is
indeed the case. Let
Lx
x
u(x, t) = v(x, t) +
g(t) + h(t).
L
L
Then
u(0, t) = v(0, t) + g(t) v(0, t) = 0
u(L, t) = v(L, t) + h(t) v(L, t) = 0.
93
Thus the new function v(x, t) satisfies homogeneous boundary conditions. Furthermore,
ut = vt +
x
Lx 0
g (t) + h0 (t),
L
L
uxx = vxx ,
since the added terms are linear in x. Thus
vt = avxx + F (x, t) + xL
g 0 (t) Lx h0 (t), t > 0, 0 < x < L,
v(0, t) = 0,
t 0,
v(L, t) = 0,
t 0,
x
v(x, 0) = f (x) Lx
g(0)
h(0),
0 x L.
L
L
This is a similar problem to the one we solved in the previous section, but with a slightly
different forcing term. We see that the effect of the time-dependent boundary conditions is
indeed that of a forcing. In good ol mathematical tradition, weve reduced the problem to
one weve solved before. Therefore were done.
6.4
So far, weve focused on providing exact solutions of partial differential equations. You should
realize that in many situations, this is not possible, even for linear partial differential equations. Maybe the boundary conditions are too hard, maybe the domain is too complicated.
Then for nonlinear equations all bets are off. We will most likely have to resort to numerical
methods, or to other approximation methods such as asymptotics and perturbation theory.
In situations like that it is crucial that we are able to determine some a priori bounds on
the solution of the problem, which can act as a check on the approximation method. The
maximum principle is a method that allows us to determine such bounds.
Lets get things going with a lemma. Why not indeed?
Lemma 3 Let a > 0, T > 0 and L > 0. Suppose that w satisfies wt awxx < 0 in 0 < x < L
and 0 < t T . Then w(x, t) cannot assume a local maximum at any point in 0 < x < L,
0 < t T.
Proof. Well proceed by contradiction. Suppose there does exist a local maximum in
our domain x (0, L) and t (0, T ]. At a local maximum
wxx (
x, t) 0.
Furthermore,
wt (
x, t) 0.
Note that usually wed have wt (
x, t) = 0 at a maximum, but in this case we have to allow
for the possibility of wt (
x, t) > 0, if t = T . Thus we have
wt (
x, t) awxx (
x, t) 0,
94
t
t=T
wt a wxx <0
x=L
Figure 6.1: The domain under discussion in the lemma. The domain includes the top thick
black line. (2.1)
which is a contradiction. This shows that such a local maximum cannot exist.
The situation discussed in this lemma is illustrated in Fig. 6.1.
After every lemma, theres a theorem:
for
on
on
on
95
for
on
on
on
Introducing the small positive parameter allows us to write the first expression with a
strict inequality. Thats what we need in order to invoke the previous lemma. Crafty, no?
It follows from the lemma that w(x, t) does not have a local maximum on the inside of
our domain. This means that its largest value is obtained on the boundary. But all these
boundary values are either zero or negative. However, for sufficiently small we have
w(
x, t) = u(
x, t) t > 0,
since u(
x, t) > 0. This provides a contradiction, finishing the proof.
The following theorem is entirely similar to the previous one.
Theorem 16 Suppose that u(x, t) is continuous on 0 x L, 0 t T , and that
ut auxx 0
u(0, t) 0
u(L, t) 0
u(x, 0) 0
for
on
on
on
t[0,T ]
t[0,T ]
and
m = min{ min u(x, 0), min u(0, t), min u(L, t)}.
x[0,L]
t[0,T ]
t[0,T ]
Then
m u(x, t) M for all (x, t) [0, L] [0, T ].
96
Proof. The proof is straightforward: apply the previous two theorems to M u(x, t)
and u(x, t) m respectively.
The Minimum-Maximum Principle states that a solution of the heat equation on a
bounded domain is always in between the largest and smallest of its values on the boundary
of its domain of definition. Well see in the next setting that this conclusion holds in an
arbitrary number of dimensions. Now well show that in addition to providing stringent
a-priori bounds, the Minimum-Maximum Principle may also be used to prove some more
fundamental results.
Corollary 2 Let f (x), g(t) and h(t) be continuous on their respective domains of definition.
Then the solution of
ut = auxx + F (x, t)
u(0, t) = g(t)
u(L, t) = h(t)
u(x, 0) = f (x)
for
for
for
for
is unique.
Proof. As before, we consider w = u1 u2 , where u1 and u2 both solve the initialboundary-value problem posed above. Then w satisfies
wt = awxx
w(0, t) = 0
w(L, t) = 0
w(x, 0) = 0
for
for
for
for
97
Well now look at an example of how the Minimum-Maximum Principle is used to construct a-priori bounds on a solution, far improving on those mentioned in the theorem itself.
Example: Consider the initial-boundary-value problem
ut = uxx
u(0, t) = 0
u(1, t) = 0
u(x, 0) = x(1 x)ex
for
for
for
for
x(1 x) 2
2
x(1 x)
8.
We also have
w(0, t) = 0 = w(1, t),
for t 0; and
w(x, 0) = x(1 x)ex x(1 x)
= x(1 x) (ex ) .
98
Making these choices for and , we can apply the Minimum-Maximum Principle to the
function w(x, t), to find that
w(x, t) 0 u(x, t) v(x, t),
so that we now know that
0 u(x, t) x(1 x)e18t .
Since this upper and lower bound approach each other as t , these inequalities provide
quite a good test to any approximation method!
The real power of the Minimum-Maximum Principle becomes apparent in higher-dimensional
settings, on irregular domains, as well see next.
6.5
ut = auxx
ut = a(uxx + uyy )
ut = a(uxx + uyy + uzz ).
At this point we note that there was almost nothing in our proofs of the one-dimensional
form of the Minimum-Maximum Principle that restricted it to one dimension. In this section,
well extend it to the multidimensional setting.
Well repeat some of those results here again, for completeness. This will also give us a
preview on how to deal with the higher-dimensional setting, which will be useful later.
Lets introduce some notation:
{t [0, T ]}.
ST = D
denotes a closed domain. This a domain that includes its own boundary:
Here the notation D
if D is an open domain then
= D D,
D
where D denotes the boundary of the domain D. For our purposes, D is a domain in which
T is an open domain, not including
the d-dimensional spatial variable lives. In what follows, S
its spatial boundary, but including its temporal boundary. As before, we have a lemma.
99
Lemma 4 Let the function u be twice differentiable in the spatial variables, and once in the
time variable. Let a > 0. If
ut au < 0
T , then u has no local maximum in S
T .
on S
Proof. Assume u has a local maximum, occurring at
(
x, t) ST .
Here x is a spatial coordinate of dimension d. Then, at this point,
u(
x, t) 0.
Usually we have a strict inequality here, but equality could occur in degenerate cases (e.g.,
the function x4 ). Also
ut (
x, t) 0.
A similar comment is in order: usually we have equality here, but inequality could occur if
the maximum is attained at t = T . Thus
ut au 0,
at (
x, t), which provides a contradiction. This ends the proof.
100
Since t < T is finite, we can let the term t become infinitely small, resulting in
u 0, on ST .
This is what we had to prove.
A similar theorem holds for positive values on the boundary, and a differential inequality
of different sign. This is proven by letting u u in the preceding lemma and theorem.
We now get the Minimum-Maximum Principle for the heat equation in any number of
dimensions. This result, much more than its one-dimensional counterpart, provides strong
a-priori bounds on the solution u of the heat equation, in complicated domains where the
problem cannot be solved analytically.
T , twice differentiable in the spatial variTheorem 19 Suppose that u is continuous on S
ables, and once in time. Lastly, let u be continuous on ST . If
ut = au,
T ,
on S
then
min u u max u
ST
ST
Proof. m u satisfies the condition of our previous theorem, and so does u M , from
which the desired result follows.
As in the one-dimensional case, the Minimum-Maximum Principle allows us to prove
uniqueness of the solution of the heat equation, even with forcing or nonhomogeneous boundary conditions. Try this as a good exercise.
6.6
Lets recall the notion of a fundamental matrix for a system of ordinary differential equations.
For simplicity, well restrict ourselves to the case of constant coefficients. Thus we consider
the system
x0 = Ax,
where x Rd , and A Rdd , a constant matrix. We can write the solution of this system as
x = (t)x0 ,
where (t) is the fundamental matrix of the system, and x0 is the initial condition for x:
x(0) = x0 .
For a constant coefficient system we can be more explicit: the fundamental matrix is given
by
(t) = eAt ,
101
defined using the standard matrix exponential (for instance, using a Taylor expansion). One
of the good things about a fundamental matrix is that it allows us to solve the system
many times for different initial conditions, merely by multiplying by a different vector. In
other words, the fundamental matrix encodes everything about the system, except the initial
condition.
We seek to generalize this concept to partial differential equations. Of course, for partial
differential equations there are even more types of input: theres the equation itself, there are
the initial conditions, and there are the boundary conditions. Depending on the context, a
Greens function will encode all information contained in the partial differential equation
(or ordinary differential equation, as well see later), and one of the initial data or boundary
conditions. The idea is to get an operation that can easily be repeated many times, as one
piece of the input data is changed. The context will dictate which input function we wish
to be varied. Well see that Greens functions have many nice properties, many of which are
similar to those of the fundamental matrix.
Remark. This concept of a Greens function can be generalized even further to nonlinear
partial differential equations, using the idea of a semi-group. We wont get further into this.
Lets make the concept of a Greens function a bit more concrete, before we construct a
few for specific examples. Consider the linear partial differential equation
ut = L(u),
where L is a linear differential operator in x acting on u. Suppose that were given initial
conditions
u(x, 0) = f (x),
defined for x R. Were looking for solutions such that u 0 as x . In this context,
were looking for an operator such that
u(x, t) = (x, t)f (x).
We will see that in many examples this operator (x, t) is an integral operator. Its kernel
will be the Greens function.
1.
ut
u(0, t)
u(L,
t)
u(x, 0)
=
=
=
=
auxx ,
0,
0,
f (x),
X
n=1
n ean t sin n x,
102
for n = 1, 2, . . ..
Substituting this in the general solution formula, we find
u(x, t) =
Z
X
2
n=1
Z
=
0
L
2
L
!
a2n t
n=1
We call
G(x, y, t) =
2 X a2n t
e
sin n x sin n y
L n=1
the Greens function for the heat equation on a finite domain with homogeneous boundary
conditions. Then
Z L
u(x, t) =
G(x, y, t)f (y)dy,
0
as desired.
Some of the steps taken above were formal (switching integrals and infinite sums, for
instance), so some conditions on f (x) are required for all steps to make sense. At this point,
we can take the definition of the Greens function and the formula as a starting point for
solving the given problem. It turns out that by doing this, we can dispense with many of
our previous conditions. The reason for this is that integrals are smoothing operators.
Properties of the Greens function
1. The Greens function satisfies the partial differential equation, in both spatial variables
for t > 0 and 0 < x < L: thus
Gt = aGxx , Gt = aGyy .
This follows immediately from the uniform convergence of the series representation of
the Greens function and its derivatives, and from the functional form of the summands.
2. The Greens function satisfies the boundary conditions:
G(0, y, t) = 0, G(L, y, t) = 0,
for 0 y L, and for t > 0. This is immediately verified from the form of the Greens
function.
103
(Thus the Greens function behaves like a delta function as t 0). This one requires
a bit more work. Lets prove it.
Proof. Choose a sequence of twice differentiable functions fn (x) (n=1,2,. . . ), uniformly convergent to f (x), with fn (0) = 0 = fn (L), for all n. Then, for all > 0, there
exists an N () such that n, m > N () implies that
|fn (x) fm (x)| < ,
for x [0, L]. Let un (x, t) be the solution of the heat equation with homogeneous
boundary conditions corresponding to initial condition fn (x). We know from our theorem using the separation of variables solution that all the functions un are continuous.
By the Minimum-Maximum Principle we have
min (fn (x) fm (x)) un (x, t) um (x, t) max (fn (x) fm (x)),
x[0,L]
x[0,L]
or
|un um | .
Hence the sequence {un , n = 1, 2, . . .} is a uniform Cauchy sequence in the space of
continuous functions. Since the space of continuous functions is complete, as long as
consider uniform convergence (the uniform limit of a sequence of continuous functions
is continuous), this Cauchy sequence is a convergent sequence, say un u, where the
convergence is uniform, and u is continuous. Then
u(x, t) = lim un (x, t)
n
Z L
= lim
G(x, y, t)fn (y)dy
n 0
Z L
=
lim G(x, y, t)fn (y)dy (by uniform convergence)
0 n
Z L
=
G(x, y, t) lim fn (y)dy
n
0
Z L
G(x, y, t)f (y)dy.
=
0
104
= lim fn (x)
n
= f (x).
We are allowed to switch the limits since the convergence in n is uniform. Taking the
limit as t 0+ of the previous result proves our delta function property.
Putting all these properties together, we have shown that the following theorem holds.
Theorem 20 Let f (x) be continuous on x [0, L] with f (0) = 0 = f (L). Then the problem
u(0, t) = 0,
t 0,
u(L,
t)
=
0,
t 0,
u(x, 0) = f (x), 0 x L.
has a unique, continuous solution for t > 0, given by
Z L
G(x, y, t)f (y)dy.
u(x, t) =
0
This is quite a surprising result: we have found that using the Greens function formalism,
we can obtain a solution of the heat equation requiring less regularity than we could before,
using separation of variables. Well take it!
2.
Consider the heat equation with initial data defined on the entire real line:
x R,
x R,
ut = auxx
u(x, 0) = f (x)
where f (x) is sufficiently fast decaying, to allow for the formal analysis that follows.
As in our discussion of evolution equations, we look for solutions of the form
u = Aeikxit ,
so that
ux = iku,
uxx = (ik)2 u,
ut = iu.
105
Thus i = a(ik)2 = aik 2 , which is the dispersion relationship. Since its imaginary
part is negative, we see that the solutions of the heat equation will decay as t . We get
2
uk (x, t) = A(k)eikxak t .
Superimposing these solutions gives
1
u(x, t) =
2
A(k)eikxak t dk.
where weve introduced the dummy variable y. This allows us to rewrite the solution as
Z
Z
1
ikxak2 t
iky
u(x, t) =
e
e
f (y)dy dk
2
Z Z
1
ikxak2 tiky
=
e
dk f (y)dy
2
Z
=
G(x, y, t)f (y)dy.
106
=
e
sin kz
cos kze
dk
2 2at
2at
z
G(z)
=
2at
d
z
ln G =
dz
2at
z2
G
=
ln
G0
4at
G = G0 e
(xy)2
4at
G0 = G(x, x, t)
Z
1
2
eak t dk
=
2
Z
Z
1
2
2
ak2 t
G0 = 2
e
dk
eal t dl
4
Z Z
1
2
2
= 2
eat(k +l ) dkdl
4
Z
1
2
= 2 2
eatr rdr
4
0
r=
atr2
1 e
=
2 2at
r=0
1 1
=
2 2at
1
G0 =
,
4at
so that
(xy)2
1
e 4at .
4at
This allows us to write the solution of the heat equation on an infinite domain as
Z
(xy)2
1
e 4at f (y)dy.
u(x, t) =
4at
G(x, y, t) =
Note that G(x, y, t) has an essential singularity as t 0 (t > 0), in which limit
G(x, y, t) (x y). As before, G(x, y, t) satisfies the heat equations
Gt = aGxx = aGyy .
107
With a little bit more effort, we can establish the following theorem.
Theorem 21 Let f (x) be bounded and piecewise continuous on < x < . Then the
heat equation problem on the whole line is solved by
( R
(xy)2
1 e 4at f (y)dy
for t > 0, .
4at
u(x, t) =
f (x)
for t = 0.
This solution is continuous for all t > 0. It is continuous at t = 0, except at the discontinuities of f (x).
The proof is technical, but entirely similar to some of the previous ones. A surprising
side note is that the above theorem does not require f (x) to be integrable on the whole line!
This is a testament to the smoothing and damping effect of the solution operator of the heat
equation. Lets do an example.
Example: Solve the initial-value problem
ut = auxx
u(x, 0) = f (x) = 1.
for x R. Note that in this example, f (x) = 1 is not integrable. It is however piecewise
continuous and it therefore satisfies the conditions of the theorem. Lets see what we get.
For t > 0,
Z
(xy)2
1
e 4at dy.
u(x, t) =
4at
Let
yx
=
4at
and
d =
Z
1
dy,
4at
1
2
e d = 1.
Direct substitution confirms that this is the correct solution. In this case there is no decay of
energy: weve started with an infinite amount, and we keep an infinite amount throughout.
As a consequence, theres no decay in the solution.
u(x, t) =
3.
We can use the result of the previous section to solve the heat equation on a semi-infinite
domain, such as the one shown in Fig. 6.2.
Thus we are interested in solving the problem
ut = auxx
u(x, 0) = f (x)
u(0, t) = 0
108
u(x,t)
f(x)
x
Figure 6.2: The domain under discussion for the heat equation with initial data on a half
line, with homogeneous boundary conditions. (2.1)
Wed like to solve this problem using as much as possible from the previous section1 . In
order to do this, we need to extend the initial condition to the whole line in such a way
that the u(0, t) = 0 boundary condition is automatically satisfied. It seems reasonable well
accomplish this by using an odd extension of f (x). In other words, well consider the problem
ut = auxx
u(x, 0) = g(x)
where g(x) is the odd extension of f (x):
f (x)
0
g(x) =
f (x)
for x R, t > 0,
for x R,
for x > 0,
for x = 0,
for x < 0.
e 4at g(y)dy
u(x, t) =
4at
Z 0
Z
(xy)2
(xy)2
1
1
4at
=
e
f (y)dy +
e 4at f (y)dy
4at
4at
0
1
6.7. EXERCISES
109
Z
(x+y)2
(xy)2
1
1
4at
=
e
f (y)dy +
e 4at f (y)dy
4at
4at
0
Z
2
2
(xy)
(x+y)
1
e 4at e 4at
f (y)dy.
=
4at
0
Z
If this result is correct (wait for it... Well check it in a minute) then the Greens function
for the heat equation problem on the half-line is
(x+y)2
(xy)2
1
4at
4at
e
.
G(x, y, t) =
e
4at
Lets check that the solution weve constructed is indeed a solution of the original initial-value
problem with homogeneous boundary conditions.
The solution satisfies the partial differential equation. We could calculate some
fancy derivatives and do some tedious calculus. Lets not. Instead, we can notice that
the solution constructed is a superposition of two solutions to the heat equation. Since
the heat equation is linear, this superposition still satisfies the equation.
The solution satisfies the initial condition. Lets use the alternate form
Z
(xy)2
1
e 4at g(y)dy.
u(x, t) =
4at
u(0, t) =
e
e
f (y)dy = 0.
4at
0
We have verified directly that we have constructed a solution of the heat equation on the
positive half line.
6.7
Exercises
110
2. Find a formal solution to the following initial-value problem. Here x [0, L].
u(x, 0) = 0,
u(0, t) = 0,
ux (L, t) = 0.
Chapter 7
Integral eqations, Sturm-Liouville
problems, and eigenfunction
expansions
So far, weve looked at expansions of functions in terms of a simpler set of basis functions in
the simplest case: rectangular domains, and problems with constant coefficients.
In general, more complicated scenarios arise if either of these conditions is relaxed. In
many cases, separation of variables is still successful, but the ordinary differential equations
that need to be solved are more complex. One often has to solve a Sturm-Liouville problem
of the form
0
(p(x)y 0 ) + q(x)y = r(x)y,
with boundary conditions
a1 y(0) + a2 y 0 (0) = 0,
b1 y(1) + b2 y 0 (1) = 0.
Note that we can always scale x so that its domain is x [0, 1]. Lets define the differential
operator L so that
0
Ly = (k(x)y 0 ) + l(x)y.
This allows us to rewrite the Sturm-Liouville problem as
Ly = r(x)y.
7.1
Well review many properties of these systems as we proceed in this chapter. Well even
prove them. We start by discussing the nonhomogeneous problem
(p(x)y 0 )0 + q(x)y = f (x),
for 0 < x < 1,
y(0) = h0 y 0 (0), y(1) = h1 y 0 (1),
where p(x) 6= 0, p0 (x), q(x) and f (x) are continuous on x [0, 1].
111
112
1.
It is natural to look for a solution of this problem using a Greens function: since we regard
f (x) as our input function, wed like to write
1
y(x) =
0
where g(x, s) is the Greens function were looking for. You might want to think of this
formula as kind of a variation of parameters idea. In that case the solution is also given as
an integral over the forcing function.
A function g(x, s) is called a Greens function for our Sturm-Liouville problem if it is
continuous on x [0, 1] and
Z 1
g(x, s)f (s)ds
y(x) =
0
d
(p(x)W (u, v)) ,
dx
Lv = 0,
v(1) = 0.
yLu uLy =
d
pW (v, y)
dx
d
vf =
pW (v, y).
dx
yLv vLy =
Z0 x
and
v(x)
Z
u(s)f (s)ds + u(x)
= pW (u, v)y.
We can get the Wronskian using Abels formula:
LY = 0
pY 00 p0 Y 0 + qY = 0
p0 0 q
00
Y + Y Y =0
p
p
W = Ce
p0
dx
p
= Ce ln p
C
= ,
p
113
114
where C is a constant. It can be chosen arbitrarily, as long as its not chosen to be zero.
Well choose C = 1 here. Thus
Z x
Z 1
y = v(x)
u(s)f (s)ds + u(x)
v(s)f (s)ds
0
x
Z 1
g(x, s)f (s)ds,
y=
0
with
g(x, s) =
v(x)u(s)
u(x)v(s)
0 s x 1,
0 x s 1.
115
or
W (u, v) = 1/p.
If this is possible weve constructed our Greens function. Suppose that it is not possible,
which happens if W (u, v) = 0 and C = 0. In that case u and v have to be proportional,
implying that
u(1) = 0, v(0) = 0.
Then u (or v) is a nontrivial solution of the homogeneous problem, violating our assumption.
Thus the functions u and v constructed are not proportional, and therefore not linearly
dependent. Hence their Wronskian does not vanish and C can be chosen to be any nonzero
constant, especially -1. Thus if the homogeneous problem only has the trivial solution, we
can contruct the Greens function.
This completes the proof.
2.
116
x=s
Lg = 0 there. This proves (3). It remains to prove (5). This is also a straightforward
calculation:
0
v (x)u(s),
0 s x 1 gold triangle,
gx (x, s) =
0
u (x)v(s),
0 x s 1 purple rectangle.
Thus
1
gx (s+ , s) gx (s , s) = u0 (s)v(s) + v 0 (s)u(s) = W (u, v) = ,
p
which finishes the proof.
Example: Consider the constant coefficient problem
ky 00 = f (x)
y(0) = 0,
y(1) = 0.
Here k is a positive constant. Lets follow the steps outlined above and construct the Greens
function.
1. find u, which satisfies the initial-value problem
ku00 = 0
u = Ax,
u(0) = 0
117
where A is a constant.
2. find v, which satisfies the initial-value problem
kv 00 = 0
v = B(1 x),
v(1) = 0
where B is a constant.
3. Next we need to ensure that the Wronskian of these two solutions is 1/p = 1/k.
We have
W = uv 0 u0 v
= Ax(B) AB(1 x)
= AB.
Thus we need to impose
1
1
AB = .
k
k
Theres some freedom to solve this. We can choose A = 1, B = 1/k. Then u = x,
v = (1 x)/k. Note that this choice of constants does not influence the Greens
function which only depends on the product of u and v.
AB =
(py 0 )0 + qy = 0
(py 0 )0 y + qy 2 = 0
Z 1
Z 1
0 0
(py ) ydx +
qy 2 dx = 0
0
0
Z 1
1
py 0 y|0 +
py 02 + qy 2 dx = 0
Z0 1
2
2
p(1)y 0 (1)h(1) + p(0)y 0 (0)h(0) +
py 02 + qy 2 dx = 0.
0
118
Figure 7.2: The Greens function for the constant coefficient example.
All terms in this last expression are positive, so all terms need to vanish individually. In
particular we need y 0 2 (0) = 0, which implies that y(x) is constant. This constant is forced
to be zero by the boundary conditions. This concludes the proof.
3.
Next we recast the general Sturm-Liouville eigenvalue problem in integral form. It is this
form that we will use later on.
As a starting point we have
Ly r(x)y = f (x)
0<x<1
.
y(0) = h0 y 0 (0), y(1) = h1 y 0 (1)
Any solution of
Ly = f (x) + r(x)y
satisfies
Z
y=
119
or
y(x) =
0
where
Z
F (x) =
0
This problem is completely equivalent with the original one. The integral equation we have
obtained is known as a Fredholm integral equation of the second kind with kernel
g(x, s)r(s).
Note that the kernel of this integral equation can be symmetrized. Assuming that r(x)
(which plays the role of a weight function) is positive, we have that
Z 1
p
p
p
p
p
r(x)y =
g(x, s) r(x) r(s) r(s)y(s)ds + F (x) r(x).
0
Lets introduce
z(x) =
r(x)y(x).
Then
Z
z(x) =
with
p
p
k(x, s) = g(x, s) r(x) r(s),
p
G(x) = F (x) r(x).
In the immediate future, well see more about Fredholm integral equations, setting us up for
using all of it in the context of partial differential equations.
7.2
Lets introduce a somewhat more abstract setting. This will allow us to do things in more
generality.
Definition (Inner product space). A linear space V is an inner product space if
for each u, v V , there is a scalar hu, vi (their inner product) such that for all u, v, w V
(1) hu, vi = hv, ui ( denotes complex conjugate),
(2) hu + v, wi = hu, wi + hv, wi,
(3) hu, vi = hu, vi, for all C,
(4) hv, vi 0, and hv, vi = 0 implies v = 0.
With the inner product, we define the norm of v V to be
p
||v|| = hv, vi.
120
n
X
xi yi ,
i=1
where x = (x1 , . . . , xn ) and y = (y1 , . . . , yn ). The norm induced by this inner product is the
regular Euclidean norm:
v
u n
p
uX
x2i .
||x|| = hx, xi = t
i=1
Example: The space of all continuous functions on an interval [a, b] is denoted by C([a, b]).
It is an inner product space with
Z
f (x)g (x)dx,
hf, gi =
a
where f and g are continuous functions on [a, b]. The induced norm is
s
Z b
p
||f || = hf, f i =
|f |2 (x)dx.
a
f (x)g (x)dx,
121
122
| hv, vn i |2 ||v||2 ,
n=1
for any v V .
Proof. The proof is straightforward calculation.
*
+
N
N
N
X
X
X
2
||v
hv, vk i vk || = v
hv, vk i vk , v
hv, vn i vn
k=1
k=1
N
X
= ||v||
n=1
hv, vk i hvk , vi
hv, vn i hv, vn i +
n=1
k=1
N
N X
X
N
X
k=1 n=1
= ||v||2 2
= ||v||2
N
X
| hv, vk i |2 +
N
X
| hv, vk i |2
k=1
k=1
N
X
| hv, vk i |2 .
k=1
Letting N we obtain
lim ||v
N
X
hv, vk i vk || = ||v||
k=1
| hv, vk i |2
k=1
| hv, vk i |2 ||v||2 ,
k=1
| hv, vk i |2 = ||v||2 .
k=1
The following result does not require that we have an inner product space. However, the
generality of the result mandates that we include in this section, rather than include it with
more specific sections later.
123
n R
m || = ||Rm+1 I + R + . . . + Rnm1 ||
||R
||Rm+1 || ||I + R + . . . + Rnm1 ||
||R||m+1 (1 + ||R|| + ||R||2 + . . . + ||R||nm1 )
0, as m, n
124
7.3
Lets return to our study of integral equations. A Fredholm integral equation has the form
1
y(x) =
0
which is the generic form of a linear nonhomogeneous problem. The operator K is given by
Z
k(x, s)y(s)ds.
Ky(x) =
0
In this section we will establish the Fredholm alternative using various kernels k(x, s).
Lets preview the result well get (well be more precise later):
The Fredholm alternative: either the equation
(I K)y = f
has a unique solution for each f , or else3
(I K)y = 0
has at least one nontrivial solution.
You have already encountered the Fredholm alternative in many settings:
Linear algebra: For the nonhomogeneous equation Ax = b to have a unique solution,
it is necessary that the homogeneous problem Ax = 0 only has the trivial solution
x = 0.
Linear ordinary differential equations: The solution of an inhomogenous linear
ordinary differential equation are given by the solution of the homogeneous problem,
plus a particular solution.
Well put = 1 from now on, for convenience. We can always absorb it in the definition
of K.
3
1.
125
First we prove the Fredholm alternative for degenerate kernels. A degenerate kernel is a
kernel of the form
n
X
k(x, s) =
aj (x)bj (s).
j=1
y(x) = f (x) +
n
X
= f (x) +
n
X
aj (x)bj (s)y(s)ds
j=1
aj (x) hbj , yi .
i=1
Heres a trivial observation from the above equation: if we can find hbj , yi for all functions
bj in the degenerate kernel k(x, s), then we can find y from the above. This observation is
actually useful: lets multiply the above equation by bi (x) and integrate from 0 to 1. Then
hbi , yi = hbi , f i +
n
X
hbi , aj i hbj , yi .
j=1
Y = F + CY
(I C)Y = F.
126
2.
Continuous kernels
In this section we prove the Fredholm alternative for kernels that are continuous functions
of x and s.
Let k(x, s) be continuous on [0, 1] [0, 1]. Let K be the operator defined by
Z
k(x, s)y(s)ds.
Ky =
0
We see that K maps functions that are continuous on [0, 1] to functions that remain continuous on [0, 1]:
K : y C([0, 1]) Ky C([0, 1]).
As a norm we use
||y|| = max |y(x)|,
x[0,1]
implying that we are dealing with functions that are uniformly continuous.
The first step of our proof consists in approximating k(x, s) by a polynomial p(x, s) such
that
1
|k(x, s) p(x, s)| < ,
with > 1. This is easily done, using the polynomial form of the Weierstrass approximation
theorem. Let P denote the integral operator corresponding to p(x, s).
Let
r(x, s) = k(x, s) p(x, s),
with R being the integral operator with r(x, s) as its kernel:
Z
Ry =
r(x, s)y(s)ds.
0
127
Then
Z 1
r(x, s)y(s)
|Ry| =
0
Z 1
|r(x, s)||y(s)|ds
0
Z 1
|k(x, s) p(x, s)|ds
||y||
0
||y||
1
||R|| < 1.
(I K)y
(I R P )y
(I R)y P y
y (I R)1 P y
=f
=f
=f
= (I R)1 f
(1 L)y = f,
where f = (I R)1 f and L = (I R)1 P . Weve used that I R has an inverse since our
space is complete and > 1 implies that ||R|| < 1.
If we can establish the Fredholm alternative for this last equation, it holds for the original
problem: these two problems are completely equivalent as the above shows. We now show
that our last equation has a degenerate kernel. We have
1
Ly = (I R)
p(x, s)y(s)ds.
0
with
Bi (x) = xi .
128
= (I R)1
Z
=
0
i=0
1
n
X
Bi (x)Ai (s)y(s)ds
i=0
n
X
(I R)1 Bi (x) Ai (s)y(s)ds,
i=0
which has a degenerate kernel! This concludes the proof of the Fredholm alternative for
continuous kernels.
Remark: completely analogous to the previous two proofs, we can prove the Fredholm
alternative for kernels that are square integrable (this requires Lebesgue integration, instead of Riemann integration), or that are weakly singular (i.e., have singularities that are
integrable).
3.
The Fredholm alternative may be rephrased in a somewhat more comprehensive form than
the one weve been using to far. Lets illustrate this for the linear algebra setting first, where
youve probably seen it all before.
Strong form of the Fredholm alternative: matrix formulation
Consider the matrix problem
Ax = b
and its adjoint problem
A y = c,
where A is the adjoint matrix of A, determined by
x, A y = hAx, yi .
If A is real, then A is the transposed matrix: A = AT . The following theorem holds:
Theorem 28 (Fredholm alternative (strong form; matrix setting)) If the homogeneous equations A y = 0 and Ax = 0 only have the trivial solution, then there is a unique
solution to the nonhomogeneous equations.
If there is a nontrivial solution to the homogeneous problem, then
a) Ax = 0 and A y = 0 have the same number of linearly independent solutions (dim ker A=
dim ker A ),
129
hKy1 , y2 i = y1 , K y2 .
In terms of integral kernels, this implies that the kernel of the integral operator K is
k (x, s) = k(s, x) . We have the following theorem.
Theorem 29 (Fredholm alternative (strong form; integral operator setting)) If the
homogeneous equations (I K)y = 0 and (I K )y = 0 only have the trivial solution, then
there is a unique solution to the nonhomogeneous equations.
If there is a nontrivial solution to the homogeneous problem, then
a) (I K)y = 0 and (I K )y = 0 have the same number of linearly independent solutions
(dim ker (I K)= dim ker (I K )),
b) (IK)y = f has a solution if and only if f is orthogonal to all solutions of (IK )y = 0
(f ker (I K )),
c) The general solution of (I K)y = f (if it exists) is given by a particular solution
of (I K)y = f added to the most general solution of the homogeneous problem
(I K)y = 0.
These additions to the form of the Fredholm alternative follow almost immediately from
our previous considerations.
130
7.4
hy, yi = hKy, yi
= hy, Kyi
D yE
= y,
= hy, yi
( ) hy, yi = 0
( )||y||2 = 0
=
R.
This proves that all eigenvalues of self-adjoint integral operators are real.
131
Next, let
y = Ky,
and
z = Kz,
with 6= . Then
hy, zi = hKy, zi
= hy, Kzi .
On the other hand, we also have
hy, zi = hy, Kzi
= hy, Kzi
= hy, Kzi .
We conclude that
hy, Kzi = hy, Kzi
132
We may refer to this equation as the eigenvalue equation. As stated above, the left-hand
side is an analytic function of as long as < . Since an analytic function has at most
a finite number of zeroes in a bounded region of the complex plane, it follows that the
eigenvalue equation has a finite number of solutions in < . Equivalently, it follows that
the operator K only has a finite number of eigenvalues in < . Dividing the complex plane
in a countable number of concentric annuli, and letting , we reach our conclusion.
This concludes the proof of the theorem.
In principle, we still have to prove that eigenvalues actually exist. We will not prove
that in this course. It is a rather obvious thing to believe, given what we know about linear
equations in various settings, especially matrix problems. As were not proving theorems
right now, lets also not prove the following theorem.
Theorem 34 (Rayleigh-Ritz quotient) Assume that K is self-adjoint, and not identically equal to zero. Then K has an eigenvalue of smallest modulus 1 , and
1
hKy, Kyi
= max
.
2
y6=0
1
hy, yi
7.5
Eigenfunction expansions
In this section we prove the Hilbert-Schmidt theorem, which is a statement about the completeness of the set of eigenfunctions of a self-adjoint kernel.
1.
Preliminaries
(n)
(x, s) = k(x, s)
n
X
j (x)j (s)
j=1
is an eigenfunction of k(x, s) with the same eigenvalue. Here j (x) denotes an eigenfunction
of k(x, s) corresponding to eigenvalue j . These eigenvalues are considered to be ordered by
modulus, with multiplicity taken into account.
Proof. The operator K (n) corresponding to the kernel k (n) (x, s) is defined by
K
(n)
z = Kz
= Kz
n Z
X
j=1
n
X
j=1
j (x)j (s)
z(s)ds
j
hz, j i
j .
j
133
h, r i = K (n) , r
n
X
h, j i hj , r i
= hK, r i
j
j=1
h, r i
r
= h, Kr i
h, r i
r
=
(h, r i h, r i)
r
= 0.
= hK, r i
We conclude that
K
(n)
= K
n
X
h, j i
j=1
j = K,
n
X
j=1
aj j .
134
The calculation in the proof of the previous lemma shows that for all j = 1, 2, . . . , n we have
h, j i = 0.
This implies that aj = 0 for all j, from which 0. This is a contradiction, from which the
theorem follows.
The next theorem is the last one we need before we can prove our completeness result.
Theorem 36 Let k be a self-adjoint, non-degenerate kernel with eigenvalues {1 , 2 , . . .}
(including multiplicities) and orthonormal eigenfunctions {1 , 2 , . . .}. Then k (n) (x, s) has
eigenvalues {n+1 , n+2 , . . .} and for any g
lim K (n) g, K (n) g = 0.
n
n
X
hr , j i
j=1
= Kr ,
since hr , j i = 0, for j = 1, 2, . . . , n. This implies that
r = r Kr = r K (n) r .
Furthermore, all eigenvalues and eigenfunctions of K (n) are obtained this way. Indeed,
suppose there is an eigenfunction of K (n) that is not spanned by {n+1 , n+2 , . . .}. We
know that is also an eigenfunction of K:
= K (n) = K.
Since is not spanned by {n+1 , n+2 , . . .}, it must be spanned by {1 , 2 , . . . , n }:
=
n
X
aj j .
j=1
But for the eigenfunctions of K (n) we know that aj = 0, which provides the required contradiction.
Lastly, we show that
lim K (n) g, K (n) g = 0,
n
hg, gi
K (n) g, K (n) g 2 ,
n+1
135
1
2n+1
K (n) g, K (n) g
= max
.
g6=0
hg, gi
lim K (n) g, K (n) g = 0,
2.
hf, n i n f,
n=1
n
X
j=1
n
X
j=1
n
X
j=1
n
X
j=1
hf, j i j
hKg, j i j
hg, Kj i j
hg, j i
j .
j
Then
f Sn = Kg
n
X
hg, j i
j=1
=K
(n)
g.
We know that
hf Sn , f Sn i = K (n) g, K (n) g 0,
136
lim
2
n
X
hf, j i j dx = 0.
f (x)
j=1
We should note that this proof can be strengthened to give absolute and uniform convergence
of the eigenfunction expansion series of f in the case when f is continuous on 0 x 1.
An immediate consequence of the above is that Parsevals identity holds.
7.6
Here p(x), p0 (x), q(x) and r(x) are continuous and p(x), r(x) > 0, q(x) 0, and h0 , h1 0.
Under these conditions, weve shown that this problem is equivalent to that of studying
the integral equation
Z 1
k(x, s)z(s)ds,
z(x) =
0
p
with z(x) = r(x)y(x), and
p
p
k(x, s) = g(x, s) r(x) r(s),
where g(x, s) is the Greens function for the Sturm-Liouville problem:
g(x, s) =
u(s)v(x)
u(x)v(s)
0sx1
.
0xs1
The two fundamental solutions u(x) and v(x) satisfy the boundary conditions at the left and
right sides respectively. They are connected by
pW (u, v) = 1.
1.
Having set up all our integral equation machinery, we can now state the following theorem.
137
ym yn r(x)dx = nm .
0
(c) If f , f 0 , f 00 are continuous on [0, 1], and satisfy the boundary conditions, then f (x) has
an absolutely and uniformly convergent eigenfunction expansion
f (x) =
cn yn (x),
n=1
with
f (x)yn (x)r(x)dx.
cn =
0
Proof.
(a-b) The kernel k(x, s) is real and symmetric, thus self-adjoint. Thus k(x, s) has a (possibly
finite) set of eigenvalues (real) with corresponding eigenfunctions, such that
1
zm (x)zn (x)dx = nm ,
0
from which
Z
ym (x)yn (x)r(x)dx = nm .
0
This proves (a-b), apart from the statement that there are an infinite number of eigenvalues and eigenvectors. Well come back to this.
(c) Since k(x, s) is symmetric and continuous, the Hilbert-Schmidt theorem states that any
function in its range has an absolutely and uniformly convergent expansion in terms
of z1 (x), z2 (x), . . . . Define
h(x) = (pf 0 )0 + qf.
Then y(x) = f (x) solves the boundary-value problem
Ly = h,
y(0) = h0 y 0 (0),
0<x<1
y(1) = h1 y 0 (1).
138
y=
g(x, s)h(s)ds
0
f (x) =
Thus
get
k(x, s)
p
h(s)ds
r(x) r(s)
0
Z 1
p
h(s)
r(x)f (x) =
k(x, s) p
ds.
r(s)
0
r(x)f (x) is in the range of the k(x, s). Using the Hilbert-Schmidt theorem we
E
X Dp
p
r(x)f (x) =
r(x)f (x), zj (x) zj (x),
j
X
f (x) =
rf, zj yj (x) =
cj yj (x).
j
p
r(x)f (x)zj (x)dx
cj =
Z0 1 p
p
r(x)f (x)yj (x) r(x)dx
=
Z0 1
=
f (x)yj (x)r(x)dx,
0
139
Theorem 39 Under the assumptions of the previous theorem, all eigenvalues are positive,
and each eigenvalue is simple, i.e., has multiplicity one.
Proof. We have
2.
Let us consider an example to illustrate the use of all of the previous material on SturmLiouville problems, as applied to a nonhomogeneous heat equation.
Consider the following problem:
u
u
(x)c(x)
=
k(x)
l(x)u,
t
x
x
with (x) denoting the density, c(x) the specific heat, u(x) the temperature, k(x) the thermal
conductivity and l(x) a radiation coefficient.
We have boundary conditions
u(0, t) = h0 ux (0, t), u(1, t) = h1 ux (1, t).
This form for the boundary conditions is an immediate consequence of Newtons law of
cooling. Further we have an initial temperature distribution given by
u(x, 0) = f (x).
140
In the above the functions (x), c(x) are strictly positive, and k(x) 0. Lastly, the constants
h0 and h1 are positive or zero as well.
Lets try using separation of variables on this problem. Set
u(x, t) = X(x)T (t),
then
X
u(x, t) =
an en t Xn (x).
n=1
an Xn (x),
n=1
so that
Z
cn =
f (x)Xn (x)r(x)dx.
0
This concludes the solution of the problem of heat conduction in a nonhomogeneous rod,
at least at a formal level. Going beyond this formal level, we know that the series for u(x, t)
is absolutely and uniformly convergent (this follows from the Hilbert-Schmidt theorem),
but some more work is required to establish this for its derivatives. The positivity of the
eigenvalues makes this an easy exercise.
7.7. EXERCISES
7.7
141
Exercises
||Lu||
= sup ||Lu||.
||u||
||u||=1
142
Chapter 8
Laplaces equation and potential
theory
Potential theory deals with the solutions of the two partial differential equations
u = 0,
and
u = f (x).
The first equation is called Laplaces equation, the second one Poissons equation. Solution
to Laplaces equation are called harmonic functions. We will consider these two equations
in two or three dimensions, in an open domain D, whose boundary is denoted by D.
The Dirichlet problem for either of these equations imposes given values of u on D:
for x D,
for x D.
u = f (x)
u = h(x)
On the other hand, the Neumann problem imposes conditions on the normal derivative
of u on the boundary:
(
u = f (x)
for x D,
u
= g(x)
for x D.
n
Here n denotes the notmal vector to D. Lastly, the Robin problem mixes these two types
of boundary conditions:
(
u = f (x)
u
= k(x)
(x)u + (x)
n
143
for x D,
for x D.
144
8.1
The domain D for our equations can be arbitrarily complicated since we are no longer in one
dimension. The separation of variables method works essentially in rectangular domains, or
domains that become rectangular after a transformation to a different coordinate system.
Such domains include spherical domains, cylindrical domains, or any other domains where
one of the families of orthogonal coordinates transforms our boundaries to line segments
parallel to the new coordinate axes. We refer to such domains as separable.
As an example, we consider Laplaces equation in two dimensions, defined on the inside
of a circle of radius a:
u = 0
for x {|x| < a},
u(a, ) = f () = f ( + 2)
for |x| = a.
Here is the angular variable along the circle. The requirement that
f () = f ( + 2)
is imposed to ensure continuity of the boundary condition.
We start by expressing Laplaces equation in polar coordinates:
u =
1 2u
2 u 1 u
+
+
= 0,
r2
r r r2 2
1.
Separation of variables
1
1
R00 + R0 + 2 R00
r
r
R00
R0
1 00
+
+ 2
R
rR r
00
R
R0 00
r2
+r +
R
R
00
0
R
R
r2
+r
R
R
=0
=0
=0
=
00
= c,
Imposing not only that this function is periodic, but that its period is 2 requires
c = n2 , for n = 0, 1, 2, . . . .
Thus
n = c1 cos n + c2 sin n.
Next we consider the problem for R(r). We have
r2 R00 + rR0 n2 R = 0.
This is an Euler equation with indicial equation
s(s 1) + s n2 = 0
s2 = n 2
s = n.
R0 = 0 + 0 ln r.
This solution has a singularity at r = 0, thus we need to choose 0 = 0. Thus
R0 = 0 ,
a constant.
2. n > 0: now s = n, so that
Rn = n rn + n rn .
The second term is singular as r 0, so we impose n = 0. Thus
Rn = n rn .
Superimposing all the solutions weve just found results in
u(r, ) = 0 +
rn (n cos n + n sin n) .
n=1
145
146
an (n cos n + n sin n) .
n=1
2.
Using the above solution, we can proceed to find the Greens function for our problem. We
have
Z 2
Z
1
1 X r n 2
u=
f ()d +
f (s)(cos ns cos n + sin ns sin n)ds
2 0
n=1 a
0
Z 2
Z
X
1
1 2
r n
=
f ()d +
f (s)ds
cos n(s )
2 0
0
a
n=1
!
Z 2
1
1 X r n
=
f (s)ds
+
cos n(s ) .
0
2 n=1 a
Well come back to this result. As an aside, we now compute for |v| < 1 and v, w R
!
1 X n
1 X
n
+
+
v cos nw = <
veiw
2 n=1
2 n=1
1
1
iw
=<
+ ve
2
1 veiw
1
1
1 veiw
iw
=<
+ ve
2
1 veiw 1 veiw
1
veiw (1 veiw )
=<
+
2 1 + v 2 2v cos w
147
1 + v 2 2v cos w + 2veiw 2v 2
=<
2(1 + v 2 2v cos w)
1 v2
=
2(1 + v 2 2v cos w)
8.2
In this section we introduce an integral transform other than the Fourier transform. Well
see that in essence it is a Fourier transform suitable for domains with cylindrical symmetry.
The Bessel function of the first kind of order 0 is defined as
Z 2
1
J0 (z) =
cos(z cos )d.
2 0
We can use this definition to show that the Bessel function satisfies the following second-order
ordinary differential equation:
zJ000 (z) + J00 (z) + zJ0 (z)
Z 2
1
z cos(z cos ) cos2 sin(z cos ) cos + z cos(z cos ) d
=
2 0
Z 2
1
=
z cos(z cos )(1 cos2 ) sin(z cos ) cos d
2 0
Z 2
1
=
z cos(z cos ) sin2 sin(z cos ) cos d
2 0
Z 2
1
=
( sin(z cos ) sin ) d
2 0
= 0.
Thus
zJ000 (z) + J00 (z) + zJ0 (z) = 0.
The Hankel transform of f (x) is
f() =
f (r)J0 (r)rdr.
0
148
Theorem 40 (Inverse Hankel transform) The inverse of the Hankel transform is given
by
Z
f()J0 (r)d.
f (r) =
0
Thus, the functional recipes for the Hankel transform and its inverse are identical.
Proof. Let F (x, y) = f (r) be a radially symmetric function. Its two-dimensional Fourier
transform is
Z Z
1
F (k1 , k2 ) =
F (x, y)ei(xk1 +yk2 ) dxdy.
2
We introduce polar coordinates in both (x, y)-space and in (k1 , k2 )-space:
x = r cos
,
y = r sin
k1 = cos
.
k2 = sin
R 2
We see that f(, ) depends only on . Note that 0 sin(r cos )rdr = 0. This is easily
seen by replacing by + . Since were integrating over an entire period of a periodic
function, this shift of the independent variable has no effect on the overall result.
Given that were using two-dimensional Fourier transforms, we know what the inverse
transform is:
Z Z
1
f (r) =
f()ei(xk1 +yk2 ) dk1 dk2
2
Z Z 2
1
=
f()eir cos() dd
2 0
0
Z
=
f()J0 (r)d.
0
149
1
u
2u
for z > 0, r 0,
r
+ 2 =0
r r
r
z
u(r, 0) = f (r),
for (x, y) R2 .
We multiply this partial differential equation by rJ0 (r) and integrate over all r values. In
what follows, Hankel transforms are denoted by a hat.
Z
u
2u
J0 (r)
r
+ rJ0 (r) 2 dr = 0
r
r
z
0
Z
2
u
u
J0 (r)
+
r
dr = 0
z 2
r
r
0
Z
2 u
u
u 0
+
J0 (r)r
r J0 (r) dr = 0
z 2
r
r
r
0
r Z
2
u
u
u
+ J0 (r)r
r J00 (r)dr = 0
2
z
r r=0
r
0
Z
2
u
u
r J00 (r)dr = 0.
2
z
r
0
The integrated term vanishes because we want our solution and its derivatives to decay as
r . Thus
Z
2 u
u
=
r J00 (r)dr
2
z
r
Z0
u
=
rJ00 (r)dr
r
0
Z
=
u (rJ00 (r)) dr
r
Z0
=
u J00 (r) + 2 rJ000 (r) dr
0
150
u
rJ00 (r) + 2 r2 J000 (r) dr
=
r
Z 0
u 2 2
=
r J0 (r)dr
r
0
Z
2
=
uJ0 (r)rdr
0
2
= u.
This differential equation is easily solved:
u = A()ez + B()ez .
It follows immediately that we need to eliminate the first term, as it as z . Thus
u = B()ez .
It follows that
Z
u(r, z) =
uJ0 (r)d
0
Z
=
B()ez J0 (r)d.
B() = f =
f (r)J0 (r)rdr.
0
This concludes the solution of our Dirichlet problem on the half-space. We could put these
results together to obtain a Greens function, for instance. Provided with a radially symmetric boundary condition at z = 0, the above formula determines B(), which in turn
determines the final solution by
Z
u(r, z) =
B()ez J0 (r)d.
0
8.3
We have studied the maximum principle for the heat equation in reasonable detail. Such
principles can be used for the equations of potential theory with equal success.
where D is connected and bounded1 . If u 0 in D
Lemma 7 Let u C 2 (D) C 1 (D),
and u 0 on D, then u 0 on D. Similarly, if u 0 in D and u 0 on D, then
u 0 on D.
1
= D D.
This means that u is twice continuously differentiable on D, and once on D
151
Proof. Well prove the first statement. The second statement follows from it by replacing
u by u. We proceed by contradiction. Suppose that there exists an x0 D such that
u(x0 ) < 0.
Then u achieves a negative minimum value somewhere in D, say at x = c:
u(c) < 0.
For > 0 construct
v(x) = u(x) |x c|2 .
Note that v(c) = u(c). We choose so that for all x D
v(x) = u(x) |x c|2 |x c|2 d2 > u(c) = v(c).
Here d is the diameter of D:
d = max |x1 x2 |.
x1 ,x2 D
It follows that v(x) achieves its minimum at some point interior to D, say at p D. Then
v(p) 0.
Let D denote the dimension of D. Then
u(p) = v(p) + 2D > 0,
which is a contradiction. Thus u(x) 0, for all x D. This is what we had to show.
With this lemma at hand, it is easy to prove the Maximum-Minimum Principle.
Theorem 41 (Maximum-Minimum Principle) Let D be connected and bounded, and
Then
let u C 2 (D) C 1 (D).
(i) if u 0 in D, then u(x) maxyD u(y), for all x D,
(ii) if u 0 in D, then u(x) minyD u(y), for all x D,
(iii) if u = 0 in D, then minyD u(y) u(x) maxyD u(y), for all x D.
Proof (i) Let M = maxyD u(y). Then
(u M ) 0
for x D, and
uM 0
for x D. By the previous lemma
uM 0
on D. This proves (i). Items (ii) and (iii) follow similarly.
We can put this result to good use to prove uniqueness for the Dirichlet problem.
152
Theorem 42 Let D be connected and bounded. Then there is at most one solution to
u = h(x)
for x D
,
u = f (x)
for x D
u = 0 for x D.
for x D,
which implies
w 0,
for x D,
and
w F,
for x D.
Next we consider
W = w u.
153
x1 =c1
x1 =c1 +d
Then
W = w u
H h
0, for x D.
On D we have
W = w u F u 0.
Now we can use the Maximum-Minimum Principle to conclude that
W 0 w u 0,
for x D.
8.4
So far, weve mainly talked about solving Laplaces equation. Well continue this for a bit
longer. It turns out that the current discussion of singular solutions of Laplaces equation
will lead to solutions of Poissons equation.
154
1.
x2j
j=1
= 0.
We will look for special solutions to this equation that have spherical symmetry: solutions
that depend only on the radial variable, and that have no angular dependence. Thus
u(x) = v(r),
where
2
r =
n
X
x2j .
j=1
We immediately get
2r
r
= 2xj
xj
r
xj
= ,
xj
r
v
(r)
.
=
v
(r)
x2j
r2
r
r3
1
v 00 +
v0 = 0
r
v 00
1n
=
0
v
r
1
n
(ln v 0 )0 =
r
v0
ln = (1 n) ln r
c
c
v 0 = cr1n = n1 .
r
v 00 + n
c
v = c ln r + d,
r
where d is a second arbitrary constant.
v0 =
155
c
+ d.
(2 n)rn2
With an appropriate choice of the constants we call these solutions the fundamental solution
(x) of Laplaces equation:
ln |x|
for n = 2
2
(x) =
for n = 3,
4|x|
and so on for other dimensions. Note that these solutions are singular at x = 0. This should
make you ponder. Are these solutions useful? Or are they mere mathematical artifacts only
to be appreciated by the true aesthete? The answer as always is all of the above. Lets
demonstrate this.
Clearly, (x y) is a solution of Laplaces equation as well, for any fixed y. So is
(x y)f (y),
for any function f (y). Could we argue that the superposition
Z
(x y)f (y)dy
u(x) =
Rn
is the general solution? This turns out to be wrong: due to the singularity at the origin, we
cannot bring the Laplacian inside the integral sign. The infinite extent of the integration
domain does not help either. So, what can be said about this superposition?
2.
Poissons equation on Rn
u(x) =
Rn
It follows that
u
=
xj
Z
(z)
Rn
f
(x z)dz,
xj
156
and
Z
(z)x f (x z)dz
u =
n
ZR
Z
(z)x f (x z)dz +
(z)x f (x z)dz
Rn B(0,)
B(0,)
= I + J .
Here x denotes the Laplacian with partial derivatives with respect to the components of
x. Also, B(0, ) denotes the ball of radius centered at 0.
Lets look at I first. This is the part containing the singularity of the fundamental
solution, but its a small part. In contrast J does not contain any singularities, but its
huge. We have
Z
|I | maxn |x f (x z)|
zR
|(z)|dz.
B(0,)
1
|(z)|dz = 2
| ln r| rdr
B(0,)
0 2
Z
=
r| ln r|dr
0
2 | ln |,
since r| ln r| is monotone increasing for small . Thus
|I | max2 |x f (x z)|2 | ln | 0,
zR
as 0.
2. If n = 3 then
1 1 2
|(z)|dz = 4
r dr
B(0,)
0 4 r
Z
=
rdr
1
= 2 .
2
Thus
|I | max3 |x f (x z)|
zR
as 0.
2
0,
2
157
w
d,
n
where is a surface element on D. Here surface denotes anything thats on the boundary
D. It does not necessarily entail a two-dimensional space. Using this identity we get
Z
Z
f
z (z) z f (x z)dz
J =
(z) (x z)d
n
Rn B(0,)
B(0,)
= K + L .
Again, weve split things in two parts. for the first part, we get
Z
f
K =
(z) (x z)d
n
B(0,)
Z
|K | maxn |f |
|(z)|d.
zR
B(0,)
Z
|(z)|dz
B(0,)
B(0,)
2
1
| ln r| d
2
1
| ln |d
2
0
= | ln |.
Thus
|K | maxn |f || ln | 0,
zR
as 0.
2. If n = 3 then
Z
1
d
4r
|(z)|dz =
B(0,)
B(0,)
2 Z
Z
=
= .
1 2
sin dd
4
158
as 0.
We see that also K does not contribute to the integral, as 0. All our hope lies with L .
We have
Z
L =
z (z) z f (x z)dz.
Rn B(0,)
We use the same Greens identity again, but with the roles of v and w switched. This gives
us
Z
Z
f (x z)z (z)dz
f (x z) (z)d
L =
n
Rn B(0,)
B(0,)
Z
f (x z) (z)d,
=
n
B(0,)
p
since z (z) = 0. Lets compute the normal derivative of : let r denotes z12 + . . . + zn2 .
Then
zj
= 0 (r)
zj
r
zj
and
ni =
r
n
X
=
nj
n
z
j
j=1
= 0 (r)
2r
=
4r2
for n = 2
.
for n = 3
We see that
0
L = ()
B(0,)
1
|B(0, )|
f (x z)d
Z
f (x z)d
B(0,)
= hf (x z)iB(0,)
= hf (x)iB(0,) .
Here hf (x z)i denotes the average of f (x z) over the indicated region. Clearly, as 0,
we have
lim L = lim hf (x z)iB(0,) = f (x).
8.5. EXERCISES
159
= f (x).
We see that the fundamental solutions of Laplaces equation may be used as Greens functions
for Poissons equation. Thus singular solutions can be of great interest indeed!
8.5
Exercises
1. We have derived Poissons formula for the solution of Laplaces equation inside of a
circle of radius a with Dirichlet boundary conditions given on r = a. Show that this
formula, when properly interpreted, is also valid for the solution of Laplaces equation
outside of a circle of radius a with Dirichlet boundary conditions given on r = a.
Note: you cant merely use a transformation r 1/r for this, as this transformation
is singular. Youre supposed to repeat what we did in this chapter, but taking into
account different boundary conditions, because of the different type of domain.
2. Find a formal solution to the following initial-value problem for the heat equation with
two spatial dimensions with diffusion coefficient :
ut (uxx + uyy ) = f (x, y),
u(x, y, 0) = 0,
p
where f (x, y) only depends on the radial variable x2 + y 2 with (x, y) R2 and t 0.
3. Find a formal solution to the following initial-value problem for the wave equation with
speed c:
utt = c2 (uxx + uyy ) + f (x, y),
u(x, y, 0) = 0, ut (x, y, 0) = 0,
p
where f (x, y) only depends on the radial variable x2 + y 2 with (x, y) R2 and t 0.
4. Let D be a bounded, connected region in Rn . Suppose that u satisfies the Neumann
problem
u = F (x), x D,
u
= h(x), x D.
n
160