AMSI 2up

AMSI
Jan. 14 Feb. 8, 2008
Partial Differential Equations
Jerry L. Kazdan
c 2008 by Jerry L. Kazdan

Copyright
[Last revised: March 28, 2011]
iv
CONTENTS
7. Dirichlets principle and existence of a solution

Chapter 6. The Rest
Contents
Chapter 1. Introduction
1. Functions of Several Variables
2. Classical Partial Differential Equations
3. Ordinary Differential Equations, a Review
1
2
3
5
Chapter 2. First Order Linear Equations

1. Introduction
2. The Equation uy = f (x, y)
3. A More General Example
4. A Global Problem
5. Appendix: Fourier series
11
11
11
13
18
22
Chapter 3. The Wave Equation

1. Introduction
2. One space dimension
3. Two and three space dimensions
4. Energy and Causality
5. Variational Characterization of the Lowest Eigenvalue
6. Smoothness of solutions
7. The inhomogeneous equation. Duhamels principle.
29
29
29
33
36
41
43
44
Chapter 4. The Heat Equation

47
1. Introduction
47
2. Solution for Rn
47
3. Initial-boundary value problems for a bounded region, part
1
50
4. Maximum Principle
51
5. Initial-boundary value problems for a bounded region, part
2
54
6. Appendix: The Fourier transform
56
Chapter 5. The Laplace Equation
1. Introduction
2. Poisson Equation in Rn
3. Mean value property
4. Poisson formula for a ball
5. Existence and regularity for u + u = f on Tn
6. Harmonic polynomials and spherical harmonics
iii
59
59
60
60
64
65
67
69
75
1. INTRODUCTION
Kazdan, Jerry, Lecture Notes on Applications of Partial Differential

Equations to Some Problems in Differential Geometry, available at
http://www.math.upenn.edu/ kazdan/japan/japan.pdf
CHAPTER 1
Gilbarg, D., and Trudinger, N. S., Elliptic Partial Differential Equations of Second Order, 2 nd Edition, Springer-Verlag, 1983.
Introduction
Partial Differential Equations (PDEs) arise in many applications to
physics, geometry, and more recently the world of finance. This will be
a basic course.
In real life one can find explicit solutions of very few PDEs and many
of these are infinite series whose secrets are complicated to extract. For
more than a century the goal is to understand the solutions even
though there may not be a formula for the solution.
The historic heart of the subject (and of this course) are the three fundamental linear equations: wave equation, heat equation, and Laplace
equation along with a few nonlinear equations such as the minimal surface equation and others that arise from problems in the calculus of
variations.
We seek insight and understanding rather than complicated formulas.
Prerequisites: Linear algebra, calculus of several variables, and basic
ordinary differential equations. In particular Ill assume some experience with the Stokes and divergence theorems and a bit of Fourier
analysis. Previous acquaintantance with normed linear spaces will also
be assumed. Some of these topics will be reviewed a bit as needed.
References: For this course, the most important among the following
are the texts by Strauss and Evans.
Strauss, Walter A., Partial Differential Equations: An Introduction,
New York, NY: Wiley, 1992.
John, Fritz. Partial Differential Equations, 4th ed., Series: Applied
Mathematical Sciences, New York, NY: Springer-Verlag.
Axler, S., Bourdin, P., and Ramey, W., Harmonic Function Theory,
accessible at
http://www.axler.net/HFT.pdf.
Courant, Richard, and Hilbert, David, Methods of Mathematical Physics,
vol II. Wiley-Interscience, New York, 1962.
Evans, L.C., Partial Differential Equations, American Mathematical
Society, Providence, 1998.
Jost, J., Partial Differential Equations, Series: Graduate Texts in Mathematics, Vol. 214 . 2nd ed., 2007, XIII, 356 p.
1
1. Functions of Several Variables

Partial differential equations work with functions of several variables,
such as u(x, y). Acquiring intuition about these can be considerably
more complicated than functions of one variable. To test your intuition,
here are a few questions concerning a smooth function u(x, y) of the
two variables x, y defined on all of R2 .
Exercises:
1. Say u(x, y) is a smooth function of two variables that has an isolated critical point at the origin (a critical point is where the gradient is zero). Say as you approach the origin along any straight
line u has a local minimum. Must u have a local minimum if you
approach the origin along any (smooth) curve? Proof or counter
example.
2. There is no smooth function u(x, y) that has exactly two isolated
critical points, both of which are local local minima. Proof or
counter example.
3. Construct a function u(x, y) that has exactly three isolated critical
points: one local max, one local min, and one saddle point.
4. A function u(x, y), (x, y) R2 has exactly one critical point, say
at the origin. Assume this critical point is a strict local minimum,
so the second derivative matrix (or Hessian matrix ).
u (x, y) =

uxx uxy
uxy uyy
is positive definite at the origin. Must this function have its global
minimum at the origin, that is, can one conclude that u(x, y) >
u(0, 0)
for
all
(x, y)
6
=
(0, 0)?
Proof or counter example.
2. CLASSICAL PARTIAL DIFFERENTIAL EQUATIONS
2. Classical Partial Differential Equations

Three models from classical physics are the source of most of our knowledge of partial differential equations:
1. INTRODUCTION
flow away from that point and contradict the assumed equilibrium.
This is the maximum principle: if u satisfies the Laplace equation then
min u u(x, y) max u
utt = uxx + uyy

ut = uxx + uyy
wave equation
heat equation
uxx + uyy = f (x, y)
Laplace equation
The homogeneous Laplace equation, uxx + uyy = 0, can be thought

of as a special case of the wave and heat equation where the function
u(x, y, t) is independent of t. This course will focus on these equations.
For all of these equations one tries to find explicit solutions, but this
can be done only in the simplest situations. An important goal is to
seek qualitative understanding, even if there are no useful formulas.
Wave Equation: Think of a solution u(x, y, t) of the wave equation
as describing the motion of a drum head at the point (x, y) at time
t. Typically one specifies
initial position: u(x, y, 0),
initial velocity: ut (x, y, 0)
boundary conditions: u(x, y, t) for (x, y) , t 0
and seek the solution u(x, y, t).
Heat Equation: For the heat equation, u(x, y, t) represents the temperature at (x, y) at time t. Here a typical problem is to specify
initial temperature: u(x, y, 0)
boundary temperature: u(x, y, t) for (x, y) , t 0
and seek u(x, y, t) for (x, y) , t > 0. Note that if one investigates
heat flow on the surface of a sphere or torus (or compact manifolds without boundary), then there are no boundary conditions for the simple
reason that there is no boundary.
Laplace Equation: It is clear that if a solution u(x, y, t) is independent of t, so one is in equilibrium, then u is a solution of the Laplace
equation (these are called harmonic functions). Using the heat equation model, a typical problem is the Dirichlet problem, where one is
given
boundary temperature u(x, y, t)
for (x, y)
and one seeks the (equilibrium) temperature distribution u(x, y) for

(x, y) . From this physical model, it is intuitively plausible that in
equilibrium, the maximum (and minimum) temperatures can not occur
at an interior point of unless u const., for if there were a local
maximum temperature at an interior point of , then the heat would
for (x, y) .
Of course, one must give a genuine mathematical proof as a check that

the differential equation really does embody the qualitative properties
predicted by physical reasoning such as this.
For many mathematicians, a more familiar occurrence of harmonic
functions is as the real or imaginary parts of analytic functions. Indeed,
one should expect that harmonic functions have all of the properties of
analytic functions with the important exception that the product or
composition of two harmonic functions is almost never harmonic (that
the set of analytic functions is also closed under products, inverse (that
is 1/f (z)) and composition is a significant aspect of their special nature
and importance).
Some Other Equations: It is easy to give examples of partial differential equations where little of interest is known. One example is the
so-called ultrahyperbolic equation
uww + uxx = uyy + uzz .
As far as I know, this does not arise in any applications, so it is difficult
to guess any interesting phenomena; as a consequence it is of not much
interest.
We also know little about the local solvability of the Monge-Amp`ere
equation
uxx uyy u2xy = f (x, y)
near the origin in the particularly nasty case f (0, 0) = 0, although at
first glance it is not obvious that this case is difficult. This equation
arises in both differential geometry and elasticity and any results
would be interesting to many people.
In partial differential equations, developing techniques are frequently
more important than general theorems.
Partial differential equations, a nonlinear heat equation, played a central role in the recent proof of the Poincare conjecture which concerns
characterizing the sphere, S 3 , topologically.
They also are key in the Black-Scholes model of how to value options
in the stock market.
Our understanding of partial differential equations is rather primitive.
There are fairly good results for equations that are similar to the wave,
heat, and Laplace equations, but there is a vast wilderness, particularly
for nonlinear equations.
3. ORDINARY DIFFERENTIAL EQUATIONS, A REVIEW
3. Ordinary Differential Equations, a Review

Since some of the ideas in partial differential equations also appear in
the simpler case of ordinary differential equations, it is important to
grasp the essential ideas in this case.
We briefly discuss the main ODEs one can solve.
du
a). Separation of Variables. The equation
= f (t)g(u) is
dt
solved using separation of variables:
du
= f (t)dt.
g(u)
Now integrate both sides and solve for u. While one can rarely explicitly compute the integrals, the view is that this is a victory and is as
much as one can expect.
du
A special case is
+ a(t)u = 0, the homogeneous first order linear
dt
equation. Separation of variables gives
u(t) = e
Rt
a(x) dx
b). First Order Linear Inhomogeneous Equations. These have

the form
du
+ a(t)u = f (t).
dt
When I first saw the complicated explicit formula for the solution of
this, I thought it was particularly ugly:
Z t
Rt
Rx
u(t) = e a(x) dx
f (x) e a(s) ds dx
but this really is an illustration of a beautiful simple, important, and

really useful general idea: try to transform a complicated problem into
one that is much simpler. Find a function p(t) so that the change of
variable
u(t) = p(t)v(t)
reduces our equation to the much simpler
dv
(1.1)
= g(t),
dt
which we solve by integrating both sides. Here are the details. Since
Lu = L(pv) = (pv) + apv = pv + (p + ap)v,
if we pick p so that p + ap = 0 then solving Lu = f becomes pv = f

which is just Dv = (1/p)f , where Dv := v , as desired in (1.1). More
abstractly, with this p define the operator Sv := pv which multiplies
v by p. The inverse operator is S 1 w = (1/p)w . The computation we
just did says that for any function v
LSv = SDv,
that is S 1 LS = D,
1. INTRODUCTION
so using the change of variables defined by the operator S , the differential operator L is similar to the basic operator D . Consequently
we can reduce problems concerning L to those for D .
Exercise: With Lu := Du + au as above, we seek a solution u(t), periodic with period 1 of Lu = f , assuming a(t) and f are also periodic,
a(t + 1) = a(t) etc. It will help to introduce the inner product
Z 1
g(t)h(t) dt.
hg, hi =
0
We say that g is orthogonal to h if hg, hi = 0. Define the operator L

by the rule L w = Dw + aw .
a) Show that for all periodic u and w we have hLu, wi = hu, L wi.
b) Show that for a given function f there is a periodic solution of
Lu = f if and only if f is orthogonal to all the (periodic) solutions
z of the homogeneous equation L z = 0.
d2 u 2
+c u = 0, with c 6= 0 a constant. Before doing anything
dt2
else, we can rescale the variable t, replacing t by t/c to reduce to the
special case c = 1. Using scaling techniques can lead to deep results.
The operator Lu := u + c2 u = 0 has two types of invariance: i).
linearity in u and translation invariance in t.
Linearity in u means that
c).
L(u + v) = Lu + Lv,
and L(au) = aLu
for any constant a.

To define translation invariance, introduce the simple translation operator T by
(T u)(t) = u(t + )
Then L being translation invariant means that
(1.2)
L(T u) = T L(u)
for any function u. There is an obvious group theoretic property:

T T = T+ .
Lemma [Uniqueness] If Lu = 0 and Lv = 0 with both u(0) = v(0)
and u(0) = v (0), then u(t) = v(t) for all t.
Proof: Let w = u v . Introduce the energy
E(t) = 12 (w 2 + w 2 ).
By linearity w + w = 0 so E (t) = w (w + w) = 0. This proves that

E(t) is a constant, that is, energy is conserved. But w(0) = w (0) = 0
also implies that E(0) = 0, so E(t) 0. Consequently w(t) = 0 for
all t, and hence u(t) = v(t).
We now use this. Since cos t and sin t are both solutions of Lu = 0, by
linearity for any constants a and b the function (t) := a cos t+b sin t =
0 is a solution of L(t) = 0. By translation invariance, for any constant

, the function z = cos(t + ) satisfies L(z) = 0. Claim: we can find
constants a and b so that i) z(0) = (0) and ii) z (0) = (0). These
two conditions just mean
cos() = a and
sin() = b.
Consequently, by the uniqueness lemma, we deduce the standard trigonometry formula

cos(t + ) = cos cos t sin sin t.
Moral: one can write the general solution of u + u = 0 as either
u(t) = C cos(t + )
for any constants C and , or as
u(t) = a cos t + b sin t.
Physicists often prefer the first version which emphasises the time invariance, while mathematicians prefer the second that emphasizes the
linearity of L.
Exercise: Consider solutions of the equation
Lu := u + b(t)u + c(t)u = f (t),
where for some constant M we have | b(t) | < M and | c(t) | < M .
Generalize the uniqueness lemma. [Suggestion. Use the same E(t)
(which is an artificial substitute for energy) but this time show that
E (t) kE(t)
for some constant k .
This means [ekt E(t)] 0. Use this to deduce that E(t) ekt E(0)
for all t 0, so the energy can grow at most exponentially].
Exercise: If a map L is translation invariant [see (1.2)], and q(t; ) :=
Let , show that q(t; ) = g(0; )et . Thus, writing Q() = q(0; ),
conclude that
Let = Q()et ,
that is, et is an eigenfunction of L with eigenvalue Q(). You find
special solutions of the homogeneous equation by finding the values of
where Q() = 0.
Exercise: Use the previous exercise to discuss the second order linear
difference equation u(x + 2) = u(x + 1) + u(x). Then apply this to find
the solution of
u(n + 2) = u(n + 1) + u(n),
n = 0, 1, 2, . . .
with the initial conditions u(0) = 1, and u(1) = 1.
1. INTRODUCTION
d). Group Invariance. One can use group invariance as the key
to solving many problems. Here are some examples:
a) au + bu + cu = 0, where a, b, and c are constants. This linear equation is also invariant under translation t 7 t + , as the
example above. One seeks special solutions that incorporate the
translation invariance and then use the linearity to build the general solution.
b) at2 u + btu + cu = 0, where a, b, and c are constants. This is
invariant under the similarity t 7 t. One seeks special solutions
that incorporate the similarity invariance and then use the linearity
to build the general solution.
du
at2 + bu2
c)
= 2
, where a, b, c, and d are constants. This is
dt
ct + du2
invariant under the stretching
t 7 t,
u 7 u,
for > 0.
In each case the idea is to seek a special solution that incorporates the
u
invariance. For instance, in the last example, try v(t) = .
t
Lie began his investigation of what we now call Lie Groups by trying to
use Galois group theoretic ideas to understand differential equations.
e). Local vs Global: nonlinear. . Most of the focus above was
on local issues, say solving a differential equation du/dt = f (t, u) for
small t. A huge problem remains to understand the solutions for large
t. This leads to the qualitative theory, and requires wonderful new
ideas from topology. Note, however, that for nonlinear equations (or
linear equations with singularities), a solution might only exist for finite
t. The simplest example is
du
= u2 with initial conditions u(0) = c.
dt
The solution, obtained by separation of variables,
c
u(t) =
1 ct
blows up at t = 1/c.
f). Local vs Global: boundary value problems. Global issues
also arise if instead of solving an initial value problem one is solving a
boundary value problem such as
(1.3)
d2 u
+ a2 u = f (x) with boundary conditions u(0) = 0, u() = 0.
dx2
Here one only cares about the interval 0 x . As the following
exercise illustrates, even the case when a is a constant gives non-obvious
results.
Exercise:
a) In the special case of (1.3) where a = 0, show that a solution exists
for any f .
R
b) If a = 1, show that a solution exists if and only if 0 f (x) sin x dx =
0.
c) If 0 a < 1 is a constant, show that a solution exists for any f .
Exercise: [Maximum Principle]
a) Let u(x) be a solution of u + u = 0 for 0 < x < 1. Show that
at a point x = x0 where u has a local maximum, u cannot be
positive. If u(x0 ) = 0, what can you conclude?
b) Generalize to solutions of u + b(x)u + c(x)u = 0, assuming
c(x) > 0.
c) Say u and v both satisfy u + u = f (x) for 0 < x < 1 with
u(0) = v(0) and u(1) = v(1). Show that u(x) = v(x) for all
0 x 1.
d) Say u is a periodic solution, so u(1) = u(0) and u (1) = u(0), of
u = 1 h(x)eu
for 0 x 1,
where h is also periodic and satisfies 0 < a h(x) b. Find

upper and lower bounds for u in terms of the constants a and b.
10
12
CHAPTER 2
First Order Linear Equations

1. Introduction
The local theory of a single first order partial differential equation, such
as
2
u
u
3
= f (x, y),
x
y
is very special since everything reduces to solving ordinary differential

equations. However the theory gets more interesting if one seeks a
solution in some open set or if one looks at a global problem.
Well see some of the standard ideas here. Because the main basic ideas
in studying partial differential equations arise more naturally when one
investigates the wave, Laplace, and heat equations, we will not linger
long on this chapter.
The story for a nonlinear equation, such as Inviscid Burgers Equation,
ut + uux = 0, is much more interesting. We may discuss it later.
2. The Equation uy = f (x, y)
The simplest partial differential equation is surely
The differential equation asserts that u(x, y) is constant on the vertical

lines. The vertical lines are called the characteristics of this differential equation. If is a more complicated region (see
figure), then the above result is not the most general solution since to the right of the y -axis one can
y
use two different functions 1 (x) and 2 (x), one in
each region. Thus, for simplicity we will restrict our
attention to vertically convex domains , that is,

x
ones in which every vertical line intersects in a
single line segment.
Figure 1-1
By analogy with ordinary differential equations, if one prescribes the
initial value
(2.4)
uy (x, y) = f (x, y),
so given f (x, y) one wants u(x, y). This problem is not quite as trivial
as one might think.
u(x, 0) = h(x)
on the line y = 0, then in a convex there will be a unique solution of

the initial value problem (2.2) (2.4), namely, the solution is u(x, y) =
h(x) for all (x, y) R2 . Again, one must be more careful for more
complicated regions.
Exercise: Solve uy + u = 0 with initial condition u(x, 0) = 2x 3.
Instead of specifying the initial values on the line y = 0, one can
prescribe them on a more general curve (t) = (x(t), y(t)), say
(2.5)
u(x(t), y(t)) = h(t).
In this case, using (2.3) one finds that

(2.6)
(2.1)
2. FIRST ORDER LINEAR EQUATIONS
(x(t)) = h(t).
However, one cannot use an arbitrary curve . For an extreme example, if is vertical, that is, x(t) = const., then one cannot solve
the initial value problem (2.2) (2.5) unless h(t) const. Thus one cannot prescribe arbitrary initial data on an arbitrary curve. Even more
seriously, if one differentiates (2.6), then one finds
a). The homogeneous equation. If R2 is a disk, the most

general solution of the homogeneous equation
(2.7)
(2.2)
so if is vertical for some value t0 , then h (t0 ) = 0.
uy (x, y) = 0
in is
(2.3)
u(x, y) = (x),
for any function depending only on x.

11
(x(t))
dx
dh
=
,
dt
dt
Moral: if is tangent to a characteristic curve at some point, then

one cannot solve (2.2) with initial condition (2.5) unless h satisfies
some additional conditions. However, if is nowhere nowhere tangent
to a characteristic, then one can solve the problem at least locally
given any h.
3. A MORE GENERAL EXAMPLE
13
b). The inhomogeneous equation. One can readily extend this

discussion to the inhomogeneous equation uy = f (x, y). The only new
issue is finding one particular solution of the inhomogeneous equation
vy = f . Using this the initial value problem (2.1), (2.5) is reduced to
the homogeneous case (2.2) (2.5) by letting w = uv . Then w satisfies
the homogeneous equation wy = 0.
14
alternate method Another approach to (2.9) is to change to new

coordinates so that (2.9) is the simpler ur = 0. A linear change of
variables is clearly appropriate.
r = ax + by
s = cx + dy
By the chain rule
Exercise: Solve uy = 1 2xy with u(x, 0) = 0.
u
u r u s
u
u
=
+
=a
+c
x
r x s x
r
s
u r u s
u
u
u
=
+
=b
+d
y
r y
s y
r
s
If one attempts to find a particular solution of the inhomogeneous equation uy = f in a domain , where f C (), then vertical convexity
is again needed. In fact
Proposition 2.1. One can solve uy = f for all f C ()
is vertically convex.
Proof: Just integrate.
A proof can be found in [Hormander-1, Theorems
3.5.4 and 3.7.2]. However the following argument (I
learned it from G. Schwarz) is adequate for many
domains such as the region
p in the figure. Let
f (x, y) = 1/r , where r = x2 + y 2 . Assume there
is a solution u of uy = f . Then for any > 0,
(2.8)
u(, 1) u(, 1) =
uy (, t) dt =
sin x = u(x, 0) = (2x) so (x) = sin(x/2).
Consequently, just as above,
1
dt.
2
+ t2
Now as 0, the left side is finite but the right side becomes infinite.
This contradiction completes the proof.
3. A More General Example
a). Constant coefficient. Other first order equations can be treated
similarly. For example, the equation
(2.9)
0 = ux + 2uy = (a + 2b)ur + (c + 2d)us .

Since we want the form ur = 0, let c + 2d = 0 and a + 2b = 1. Then
s = d(2x + y). The choice of d is unimportant, so we just pick d = 1.
The solution of ur = 0 is u(r, s) = (s) = (2x + y) for any function
. Using the initial condition we find
1
1
Thus
u(x, y) = sin(x 12 y).

The transport equation is
(2.10)
ut + cux = 0.
It is a simple model for the following situation. Say one has water
flowing at a constant velocity c in a horizontal cylindrical pipe along
the x-axis. Initially , near x = 0 a colored dye is inserted in the water.
Ignoring possible dispersion of the dye, it will simply flow along the
pipe. The concentration u(x, t) of the dye, then is reasonably described
by the transport equation. If the initial concentration is u(x, 0) = f (x),
then by our discussion in the previous paragraph, the solution is
u(x, t) = f (x ct).
ux + 2uy = 0,
can be interpreted as the directional derivative in the direction of the

constant vector field V (x, y) = (1, 2) is zero: V u = 0. Thus u
is constant along lines of slope 2, that is, on lines of the form 2x
y =const. Thus u(x, y) depends only on which straight line one is on,
that is, on the value of 2x y . Hence u(x, y) = h(2x y) for some
function h(s). If we also ask, for instance, that u also satisfy the initial
condition u(x, 0) = sin x, then h(2x) = u(x, 0) = sin x so the solution
is u(x, y) = sin(x 12 y)
The lines 2x y =const. are the characteristics of (2.9). There is an
obvious analog of the vertically convex domains for this equation.
This solution f (x ct) represents a density wave traveling to the

right with velocity c. To see this we sketch u(x, t) = f (x ct) for a
specific choice of f .
t=0
t = t2=1
c c 2cc 2c 2c
u(x,0)
u(x,1)
f(x-c)
u(x,2)===f(x)
f(x2c)
zz
15
16
t=0
Exercise: Solve xux + uy = 0 with u(x, 0) = g(x). Consider solving

the same equation but with the initial condition u(0, y) = h(y).
Exercise: Solve yux xuy = xy with u(x, 0) = 0 for x > 0. Is your

solution valid for all (x, y) other than the origin?
2c
u(x,0) = f(x)
Exercise: Solve ux + uy = 0 with the initial value u(0, y) = 3 sin y .
Exercise: Solve ux + uy + 2u = 0 with the initial value u(0, y) = 3y .
b). Variable coefficient. First order linear equations with variable coefficients
(2.11)
a(x, y)ux + b(x, y)uy = f (x, y)
are also easy to understand. Let V (x, y) be the vector field

V (x, y) = (a(x, y), b(x, y)).
Then aux + buy is just the directional derivative of u along V .
Let (x(t), y(t)) be the integral curves of this vector field
dx
dy
(2.12)
= a(x, y),
= b(x, y).
dt
dt
In the homogeneous case f = 0, (2.11) means
that u(x, y)is a solution if and only if it is constant along these curves. These are the characteristics of (2.11). To
solve (2.11) one introduces these characteristic curves as new coordinates. This will enable us to reduce the equation to the simple form
(2.1).
Example 2.2. x
u
u
+y
= f (x, y).
x
y
The integral curves of the vector field (x, y) are x(t) = x0 et ,

y(t) = y0 et . These are half-rays through the origin (the origin is a singular point of the vector field so for the present we
delete it from consideration). These curves are the characteristics of our example. These radial lines tell us to introduce polar
coordinates. Then the equation becomes simply
rur = f (r cos , r sin ),
that is
ur =
f (r cos , r sin )
,
r
that is exactly of the form (2.1).

Exercise: Use this procedure to obtain solutions of ux uy = 0. Find a
solution satisfying the initial condition u(x, 0) = xe2 x. Is this solution
unique?
For two independent variables, the following small modification is sometimes convenient. Consider for the homogeneous equation
(2.13)
a(x, y)ux + b(x, y)uy = 0
We find a solution using our previous observation that the solutions

are precisely those functions that are constant along the characteristic
curves. For convenience, say a(x0 , y0) 6= 0 so near (x0 , y0 ) (2.12) can
be written as
dy
b(x, y)
=
.
dx
a(x, y)
(2.14)
Write the solution of this that passes through the point x = x0 , y = c

as y = g(x, c). Then of course c = g(x0 , c). This curve is the characteristic that passes through (x0 , c). By the implicit function theorem the
equation y = g(x, c) can be solved for c to rewrite the equation of the
characteristics in the form (x, y) = c. But u = (x, y) is constant
along these characteristic curves, as is u = F ((x, y)) for any function F . Thus u(x, y) = F ((x, y)) is a solution of the homogeneous
equation (2.13) for any F .
Exercise: Use this procedure to find the general solution of xux +buy =
1. Here b is a constant. [Suggestion: To apply the method, first find
(by inspection) a particular solution of the homogeneous equation.]
We next extend these ideas to n independent variables x := (x1 , . . . , xn ).
Let a1 (x), . . . , an (x) be real functions. Then locally one can solve
(2.15)
P u :=
n
X
j=1
aj (x)
u
= f (x)
xj
by observing that P u is the directional derivative of u in the direction

of the vector field V (x) := (a1 (x), . . . , an (x)), which is assumed nonsingular. Thus P u = f specifies the directional derivative of u along
the integral curves to this vector field. These integral curves are the
solutions of the ordinary differential equation
(2.16)
dxj
= aj (x(t)),
dt
j = 1, . . . , n;
that is
dx
= V (x).
dt
17
If this vector field is differentiable, then (locally) through every point

there is a a unique solution., this solution being the integral curve of V through the point.
Let S be an n 1 dimensional surface that is
transversal to these integral curves (transversal means that the integral curves are not
tangent to S ). Let := (1, . . . , n1 ) be
local coordinates on S . Pick the parameter t so that at t = 0 the integral curves
are on S and let x = x(t, ) be the integral curve passing through when t = 0, so
x(0, ) = S . Introduce the new coordinates (t, ) in place of (x1 , . . . , xn ) and notice
P u dxj
P
u
that ut =
j xj dt =
j aj (x) xj . Thus
(2.15) assumes the simple canonical form
u
= f (t, ).
t
This is exactly the special case (2.1) we have already solved. After one
has solved this, then one reverts to the original x coordinates.
Example: Solve 2ux + yuy uz = 0 with u(x, y, 0) = (x y)2 .
To solve this, write V = (2, y, 1). Then the differential equation

states that V u=0, that is, u is constant along the integral curves
of this vector field v . These integral curves are the characteristics of
the differential equation. To find the characteristics we integrate
dy
dz
dx
= 2,
= y,
= 1.
dt
dt
dt
The solution is
x = 2t + ,
y = et
z = t + ,
where , , and are constants. Since the parameter t is arbitrary,

we pick t = 0 when z = 0 This gives = 0, so z = t. Then we can
replace the parameter t by z .
x = 2z + ,
y = ez .
The solution u(x, y, z) depends only on the integral curve passing through
the point (x, y, z), so it depends only on = x + 2z and = yez :
u(x, y, z) = h(x + 2z, yez )
for some function h which we now determine from the initial condition
2
Consequently,
(x y) = u(x, y, 0) = h(x, y).

z 2
u(x, y, z) = (x + 2z ye ) .
18
Exercise: Use this approach to solve 2ux + uy xuz = 2x with the

initial condition u(x, 0, z) = 0.
Exercise: (Transport equation). Consider functions u(x, t) in the n+1
variables (x, t) := (x1 , . . . , xn , t) and let c := (c1 , . . . , cn ) Rn . Solve
the transport equation in n space variable: ut + c1 ux1 + + cn uxn = 0,
that is, ut + c u = 0, with initial condition u(x, 0) = F (x). You
should be led to the solution u(x, t) = F (x ct).
Exercise: Discuss how to solve a(x, y)ux +b(x, y)uy +c(x, y)u = f (x, y)
by introducing the characteristics as coordinates and reducing to an
equation of the form ut + p(t, )u = h(t, ), which can be solved locally
by ODE techniques.
Solve 2ux + uy xuz + u = 2x with the initial condition u(x, 0, z) = 0.
4. A Global Problem
a). Statement. So far we have limited our discussion to local and
semi-local problems. Let T2 = { (x, y) R2 : 0 x 2, 0 y
2 } be the torus, where we identify x = 0 with x = 2 , and y = 0
with y = 2 . Then C (T2 ) is just the set of smooth functions that
are 2 periodic in both x and y . Let 6= 0 and c be a real constants.
The problem is: given f C (T2 ), find u C (T2 ) so that
(2.17)
Lu := ux uy + cu = f (x, y).
Without loss of generality we may assume that > 0, since if it is not,

we can replace y by y . For c there are two cases, c > 0 and c = 0
(if c < 0, replace x by x and y by y ).
If needed, refer to the Section 5 of this chapter for a speedy tour of

Fourier series. They are essential here.
b). Application of Fourier Series to ux uy + cu = f (x, y).
Case 1: c > 0 We seek a solution u of (2.17) as a Fourier series
X
uk ei(kx+y) .
u(x, y) =
k,
Differentiating term-by-term we obtain

X
Lu =
i(k )uk ei(kx+y) .
Thus, if u is to satisfy (2.17) Lu = f , matching the above Fourier series

for Lu with that (2.33) for f and using the orthogonality of ei(kx+y) ,
we find the equation
(2.18)
Thus,
uk =
i(k + c)uk = fk ,
fk
,
i(k ) + c
so
k, = 0, 1, 2, . . .
u(x, y) =
fk
ei(kx+y) .
i(k ) + c
4. A GLOBAL PROBLEM
19
Since c > 0 then | i(k ) + c | c > 0 so | uk | | fk |/c. Thus, by

(2.39) if f C (T2 ) then u C (T2 ).
Case 2: c = 0 If we integrate both sides of the equation over T2 ,
by the periodicity of u we immediately find the necessary condition
Z
f (x, y) dx dy = 0.
T2
We seek a solution u of (2.17) as

X
u(x, y) =
uk ei(kx+y) .
k,
Formally, after differentiating term-by-term we obtain

X
Lu =
i(k )uk ei(kx+y) .
Thus, if u is to satisfy (2.17) Lu = f , matching the above Fourier series

for Lu with that (2.33) for f and using the orthogonality of ei(kx+y) ,
we find the equation
(2.19)
i(k )uk = fk ,
k, = 0, 1, 2, . . .
R
If k = = 0 this implies f00 = 0, which is just T2 f = 0 dx (again).
Moreover, if = p/q is rational, then fk = 0 whenever k/ = p/q .
This gives infinitely many conditions on f . We will not pursue this case
further and consider only the case when is irrational. Then solving
(2.19) for uk and using them in the Fourier series for u we obtain
X ifk
(2.20)
u(x, y) =
ei(kx+y) .
k
It remains to consider the convergence of this series. Well use Lemma
2.39 to determine if u is smooth. It is clear that there will be trouble if
can be too-well approximated by rational numbers, since then the denominator (k/) will be small. This is the classical problem of small
divisors. Of course every real number can be closely approximated by
a rational number p/q . The issue is how large the denominator q must
be to get a good approximation.
Definition 2.3. An irrational number is a Liouville
number if for

k
p

every positive integer and any k > 0 then < for infinitely
q
q
many pairs of integers (p, q). Thus, is not a Liouville number if there
p

k

exist numbers and k so that > for all but a finite number
q
q
of integers p, q .
The results in the remainder of this section will not be used elsewhere in these
notes.
20
Remark 2.4. If is a real algebraic number of degree m 2 over the

rational numbers, then it is not a Liouville number and one may pick
= n. Here is the proof. Say the real irrational number is a root of
h(x) := an xn + + a0 = 0
with integer coefficients and an 6= 0 and that p/q is so close to that

h(p/q) 6= 0. Then
p | a pn + a pn1 q + |
1

n
n1
n
h
=
q
qn
q
since the numerator is a non-zero integer. Thus by the mean value
theorem
p p

1
p

h
h() = | h (c) |
= h
n
q
q
q
q
for some c between and p/q . Thus,

p M

(2.21)
n ,
q
q
where M = 1/| h (c) |.

Liouville used this approach to exhibit the first transcendental number
around 1850; only later were e and proved to be transcendental.
The inequality (2.21) was subsequently improved successively by Thue,
Siegel, Dyson, Gelfond, and Roth. Roths final result is that the exponent n on the right side of (2.21) can be replaced by 2. He was
awarded a Fields Medal for this.
P n!
P n!
Exercise: Show that :=
2
and :=
10
are Liouville
numbers and hence transcendental.
Exercise: Show that the set of Liouville numbers 0 < < 1 has
measure zero.
We are now in a position to prove the following striking result on the
global solvability of (2.17) on the torus. It will be convenient to use
the following equivalent definition of a Liouville number:
Lemma 2.5. is a Liouville
number if for every positive integer and

p

k

any k > 0 then <
for infinitely many pairs of
2
q
(1 + p + q 2 )/2
integers (p, q).
Proof: Since 1/(1 + p2 + q 2 )/2 < 1/q , a Liouville number in this
sense also satisfies the previous definition. Conversely, if is a Liouville
number in the previous sense, then for any integer > 0 and any k > 0,
we know | p/q | < 1/q for infinitely many p, q . This, with = 0
and k = 1, gives the crude estimate | p | | p q | + | q | 1 + q
4. A GLOBAL PROBLEM
21
so 1 + p2 + q 2 c(1 + q 2 ) (we can let c = 3 + 2 2 , but this value is

unimportant). Since q 1,
1
2
2c
.
q2
1 + q2
1 + p2 + q 2
Thus, if is a Liouville number in the previous sense then it is also a

Liouville by this alternate definition.
Theorem 2.6. Let be an irrational number. Then the equation (2.17)
has a solution u C (T2 ) for all f C (T2 ) if and only if is not
a Liouville number.
Proof: Say is irrational but not a Liouville number. Our goal is to
estimate the growth of the Fourier coefficients in (2.20) and show the
series converges to some smooth u. Since is not a Liouville number,
there is an integer > 0 and a number k such that for all but a finite
number of integers p, q we have

p
k

.
>
q
(1 + p2 + q 2 )/2
This will allow us to estimate the denominators in (2.20). To estimate

the numerators we use the above Lemma 2.5. Consequently, for any s
there is come constant c(s) so that
| uk | =
| fk |
c(s)(1 + k 2 + 2 )/2
c(s)
=
.
| k |
(1 + k 2 + 2 )s
(1 + k 2 + 2 )s(/2)
Since s is arbitrary, by the Lemma again we find that u is smooth.

Therefore we can differentiate (2.20) term-by-term and verify that it
satisfies the differential equation (2.17).
Conversely, if is a Liouville number, we will exhibit a smooth f so
that with this f the equation (2.17) has no smooth solution. Since
is Liouville, for any k > 0, > 0 there are infinitely many pairs of
integers (p, q) that satisfy
| p q | <
(1 +
p2
k
+ q 2 )/2
Using this with = 2j and k = 1, for each j pick one point (pj , qj ).
(2.22)
| pj qj | <
1
(1 + p2j + qj2 )j
2
We may assume that p2j + qj2 < p2j+1 + qj+1
to insure that each of these
lattice points selected is associated with only one index j . Define f by
setting
1
fpj qj =
2
(1 + pj + qj2 )j
22
for these lattice points (pj , qj ) while for all other lattice points (k, )
we set fk = 0. Then by Lemma 2.5 f is smooth. However from (2.19)
and (2.22)
| fpj qj |
| upj qj | =
>1
| pj qj |
so u is not smooth. In fact, u is not even in L2 (T2 ). Consequently if
is a Liouville number, then there is no smooth solution.
5. Appendix: Fourier series
Many problems in science and technology lead naturally lead one to
Fourier series. They are a critical tool in these notes.
a). Fourier series on S 1 . Say a function f (x) is periodic with period 2 . It is useful to think of these as functions on the unit circle, S 1 .
The simplest functions with this periodicity are eikx , k = 0, 1, 2 . . .
(or, equivalently, cos kx and sin kx). One tries to write f as a linear
combination of these functions
X
(2.23)
f (x)
a eix .
=
But how can you find the coefficients a ? What saves the day (and was
implicitly realized by Euler as well as Fourier) is to introduce the inner
product
Z
h, i =
(x)(x) dx,
and say that is orthogonal to when h, i = 0. Note that if

and are orthogonal, then the Pythagorean formula is valid:
k + k2 = kk2 + kk2 .
In this inner product eikx and eix are orthogonal for integers k 6= .
As in Rn we also write the norm
Z
1/2
1/2
2
(2.24)
kk = h, i =
| (x) | dx
[to keep history in perspective, the inner product in Rn was introduced

only in the late nineteenth century]. Since keikx k2 = heikx , eikx i = 2 ,
it is convenient to use the orthonormal functions eikx / 2 and write
X
eix
(2.25)
f (s)
c .
2
=
Formally taking the inner product of both sides of this with eikx / 2
we obtain the classical formula for the Fourier coefficients
Z
eikx
eikx
(2.26)
ck = hf, i =
f (x) dx.
2
2
5. APPENDIX: FOURIER SERIES
23
Understanding the convergence of the Fourier series (2.25) is fundamental. This convergence clearly depends on the decay of the Fourier
coefficients c . First we discuss convergence in the norm (2.24).
Let TN be the (finite dimensional) space of trigonometric polynomials
whose degree is at most N , that is, these functions have the form
P
P
eikx
ikx
tN (x) =
. Also let PN (f ) :=
| k |N ak e
| k |N ck 2 TN be
the terms in (2.25) with | k | N . By (2.26), note that f PN (f )
is orthogonal to TN because if | | N then hf PN (f ), eix i = 0.
Thus, we have written
f = PN (f ) + [f PN (f )]
as the sum of a function in TN and a function orthogonal to TN so we

call PN (f ) the orthogonal projection of f in the subspace TN . Since
both PN (f ) and tN are in TN , then by the Pythagorean theorem
(2.27) kf tN k2 = kf PN (f )k2 + kPN (f ) tN k2 kf PN (f )k2.
In other words, in this norm the function PN (f ) is closer to f than

any other function tN in TN .
The useful Bessels inequality is a special case of the computation (2.27)
when we pick tN 0. It says
X
| ck | 2 .
(2.28)
kf k2 kPN (f )k2 =
| k |N
In particular, if f is piecewise continuous, so kf k < , then

converges.
k | ck
|2
We will use the observation (2.27) to prove that if f C(S 1 ), that

is, if f is continuous and 2 periodic, then PN (f ) converges to f in
our norm (2.24). For this we use the Weierstrass approximation theorem to P
uniformly approximate f by some trigonometric polynomial
tN (x) = | k |N ak eikx . Thus, given > 0 there is some trigonometric polynomial tN (x) so that maxxS 1 | f (x) tN
(x) | < (here N is
determined by ). This implies that kf tn k 2 . Consequently
(2.27) gives the desired convergence in this norm:
(2.29)
kf PN (f )k kf tN k 2 ,
that is, limN kf PN (f )k 0. Since kf k2 = kPN (f )k2 + kf

PN (f )k2 , this also implies the Parseval identity
Z
X
| f |2 =
(2.30)
kf k2 = lim kPN (f )k2 , that is
| ck | 2 .
N
Remark 2.7. While this reasoning used that f C(S ), it is straightforward to see that the results hold only assuming f is piecewise continuous (or even square integrable). For this we use that in the norm
24
(2.24) one can approximate a piecewise continuous function on [, ]

by a continuous 2 periodic function.
Exercise: Let f (x) := x for x . Compute its Fourier series

and the consequent formula that Parsevals identity (2.30) gives.
To obtain the uniform convergence
P of (2.25) we will prove that if f is
smooth enough then the series
| ck | converges. By the Weierstrass
M-Test this will give the uniform convergence. Thus we need to discuss
the decay of the Fourier coefficients c .
To understand this decay, without worrying about convergence formally
take the derivative of both sides of (2.25) to find that
(2.31)
f (s)
eix
ic .
2
=
Thus, we suspect that the Fourier coefficients of f are ikck . This is

easy to prove directly if f is periodic and has a continuous derivative;
just use integration by parts in (2.26) to obtain
Z
eikx
f (x) dx.
ck =
ik 2
Consequently,
| ck |
2
max| f (x) |.
|k|
Repeating this procedure, we find that if f C j (S 1 ) and, with its

derivatives, is periodic, then
2
| ck |
max| D j f (x) |.
| k |j
Thus, the smoother f is, the faster its Fourier coefficients decay. In
particular, if f C 2 (S 1 ) (so f , f , and f are periodic), then | ck |
const /k 2 so the series | ck | converges and hence the Fourier series (2.25)
converges uniformly to f .
By being more careful, we can prove that the Fourier series converges
uniformly if f C 1 (S 1 ); in fact, all we will really require is that f is
square integrable. For this we use Bessels inequality (2.28) applied to
f:
X
(2.32)
| kck |2 kf k2 .
k
25
Therefore, by the Schwarz inequality,

X
X
p
1
p
1 + | k | 2 | ck |
| ck | =
1 + | k |2
| k |N
| k |N
X
1/2
1/2 X
1
2
2
.
(1 + | k | )| ck |
1 + | k |2
| k |N
| k |N

The second series converges by (2.32) (in fact, it converges to kf k2 +

P
1/2
kf k2
), while the first by comparison to
1/| k |2 .
Exercises:
1. Let ck be the Fourier coefficients of f C(S 1 ). Show that if f and
all of its derivatives exist and are continuous, then for any integer
s 0 there is a constant M(s) so that | ck | M(s)/(1 + | k |2 )s/2 .
2. Conversely, if for any integer s 0 there is a constant M(s) so
that | ck | M(s)/(1 + | k |2 )s/2 , show that f C (S 1 ).
1
3. If f C (S , use Fourier series to solve u + u = f on S .

4. If f C 1 (S 1 , use Fourier series to discuss when one can solve
u = f on S 1 .
b). Fourier series on Tn . With the above theory for Fourier series in one variable as motivation, we now investigate Fourier series in
n variables, that is, on the n-dimensional torus Tn . As a warm-up, in
two variables we write
X
(2.33)
f (x, y) =
fk ei(kx+y)
k,
where the Fourier coefficients are given by

Z
1
fk = (2)
f (x, y)ei(kx+y) dx dy.
2
T2
Note that we have switched normalization from (2.25) to (2.23).

For n variables, to avoid a mess we introduce some notation. Write
x = (x1 , . . . , xn ) Tn , let k = (k1 , . . . , kn ) be a multi-index (vector)
with integer coefficients, | k | = [k12 + , kn2 ]1/2 and let k x = k1 x1 +
, kn xn . Then the (formal) Fourier series for f is
Z
X
1
ikx
(2.34)
f (x) =
fk e
where fk =
f (x)eikx dx.
n
(2)
n
T
k
Parsevals theorem
states that if f is square integrable, then in the L2
R
2
norm kf k := Tn | f (x) |2 dx we have
X
kf k2 = (2)n
| fk |2 .
26
We next generalize the insight we found with Fourier series in one

variable that the smoothness of a function is encoded in the decay
of its Fourier coefficients. If u has a Fourier series
X
(2.35)
u(x) =
uk eikx
k
then, formally,
(2.36)
( + 1)u(x) =
X
k
(1 + | k |2 )uk eikx .
Using this and the divergence theorem we observe that for a real function u
Z
Z
X
2

2

2
u uu] dx =
| u | + | u | dx =
(1 + | k |2 )| uk |2 .
Tn
Tn
and
Z
Tn
2

| u | + 2| u |2 + | u |2 dx =
Tn
X

u(1 )2 u] dx =
(1+| k |2 )2 | uk |2
k
(for complex functions u one just adds a few complex conjugate signs).
Using this as motivation, define the Sobolev spaces H s (Tn ) to be the
space of functions with finite norm
X
kk2H 1 (Tn ) :=
(1 + | k |2 )s | k |2 < .
k
Of course H (T ) = L (T ).
One thinks of H s (Tn ) as the space of functions on Tn whose derivatives
up to order s are square integrable.
P To see this, letr r = (r1 , . . .r,1 rn ) be
any multi-index of integers with
rj = r and let D = (/x1 ) (/xn )rn
be a partial derivative of order r s. Then, using
X
k eikx
(2.37)
(x) =
k
we have
D r (x) =
X
k
But | k1r1 knrn | | k |r so
(i)r (k1r1 knrn )k eikx .
kD r k2L2
X
k
| k |2r | k |2 .
Because r s we have | k |2r (1 + | k |2 )s so the above sum is finite.
It should be clear that if f C s (Tn ), that is, if the derivatives of f up

to order s are continuous and periodic, then f H s (Tn ) but H s (Tn )
is a much larger space. However, we will show that if H s (T n )
for sufficiently large s, then is continuous. Using the Weierstrass
27
M-test, from (2.37) it is enough to show that | k | < . But by the

Schwarz inequality
X
X
1
| k | =
(1 + | k |2 )s/2 | k |
2 )s/2
(1
+
|
k
|
k
X
1/2 X
1/2
1
2 s
2
(1
+
|
k
|
)
|
|
k
(1 + | k |2 )s
k
k
1/2
X
1
kkH 2 (Tn ) .
=
(1 + | k |2 )s
k
P
The series
1/(1 + | k |2 )s converges for all s > n/2. One way to see
this is by comparison with an integral using polar coordinates
Z
Z n1
dx
r
dr
n1
=
Area
(S
)
.
2 s
(1 + r 2 )s
Rn (1 + | x | )
0
This integral converges if 2s (n 1) > 1, that is, if s > n/2. Thus,
if s > n/2, there is a constant c so that if H s (Tn ) then
kkC 0 (Tn ) ckkH s (Tn ) .
One consequence is that if we have a Cauchy sequence in H s (Tn ) and

if s > n/2, then it is Cauchy in C 0 (Tn ) and hence converges uniformly
to a continuous function. This is expressed as the Sobolev embedding
theorem: if s > n/2, then C 0 H s .
If we apply this to a j th derivative of , we find the following basic

result.
Theorem 2.8. Sobolev inequality
(2.38)
If
s > j + n/2
then
kkC j (Tn ) ckkH s (Tn )
and corresponding embedding theorem:

Theorem 2.9. Sobolev embedding theorem If s > j + n/2 then
Cj Hs .
This shows that
Corollary 2.10. If is in H s (T2 ) for all positive integers s, then
is smooth: C (Tn ). That is, C (Tn ) = s H s (Tn ).
Equivalently, C (Tn ) if and only if its Fourier coefficients decay
faster than any polynomial: for any integer s 0 there is a constant
c(s) so that
c(s)
(2.39)
| k |
.
(1 + | k |2 )s/2
28
30
3. THE WAVE EQUATION
The formula (3.2) implies an interesting identity we will need t

R x+ct=q
x-ct=b
soon. Let P , Q, R, and S be the successive vertices of a
Q
S
parallelogram whose sides consist of the four characteristic lines
x-ct=a
x ct = a, x ct = b, x + ct = p, and x + ct = q . If u(x, t) is x+ct=p P
a solution of the wave equation, then
x
CHAPTER 3
(3.3)
u(P ) + u(R) = u(Q) + u(S).
This is clear since u(P ) = F (a) + G(p), u(Q) = F (a) + G(q), u(R) =
F (b) + G(q), and u(S) = F (b) + G(p).
The Wave Equation

1. Introduction
a). Infinite string, < x < . On physical grounds based on

experiments with the motion of particles, we anticipate that we should
specify the following initial conditions:
Light and sound are but two of the phenomena for which the classical
wave equation is a reasonable model. This study is one of the real success stories in mathematics and physics. It has led to the development
of many valuable techniques.
(3.4)
Upon studying the motion of a vibrating string one is led to the simple
differential equation
utt = c2 uxx ,
Exercise: Consider the equation
where u(x, t) denotes the displacement of the string at the point x at

time t and c > 0 is a constant that involves the density and tension of
the string. Well shortly show how to interpret c as the velocity of the
propagation of the wave.
By making the change of variables = x ct and = x + ct in (3.1),
we find
(3.6)
u = u .
b) Use this to solve (3.6) with the initial conditions
Integrating this twice reveals the general solution u(, ) = f () +

g() for any twice differentiable functions f and g . Untangling the
change of variables give us the general solution of (3.1):
u(x, t) = F (x ct) + G(x + ct).
The term F (x ct) represents a wave traveling to the right with velocity c. We saw this in the previous Section a) when we discussed the
transport equation. The sketches there substantiate the statement that
c is the velocity of propagation of the wave. Similarly, G(x + ct) represents a wave traveling to the left with velocity c, so the general solution
is composed of waves traveling in both directions. The two families of
straight lines x ct =const, and x + ct =const are the characteristics
of the wave equation (3.1).
29
uxx 3uxt 4utt = 0.
a) Find a change of variable = ax + bt, = cx + dt so that in the

new coordinates the equation is the standard wave equation
u = 0.
(3.2)
u(x, 0) = f (x)
ut (x, 0) = g(x).
Using these conditions we can uniquely determine F and G in (3.2).

This gives dAlemberts solution of the initial value problem (3.1), (3.4):
Z
f (x + ct) + f (x ct)
1 x+ct
(3.5)
u(x, t) =
g(s) ds.
+
2
2c xct
2. One space dimension
(3.1)
initial position
initial velocity
u(x, 0) = x2 , ut (x, 0) = 2ex .
(x,t)
x
x-ct
It is instructive to note that the solution at (x, t)

depends only on the initial data in the interval
between the points xct and x+ct . This interval
is called the domain of dependence of the point
(x, t).
x+ct
Similarly, the initial data at a point (x0 , 0) can
only affect the solution u(x, t) for points in the
triangular region | x x0 | ct. This region is
called the domain of influence of the point (x0 , 0)
(x0 ,0)
2. ONE SPACE DIMENSION
31
32
b). Semi-infinite string, 0 < x < . Semi-infinite strings can

also be treated.
General Case. For a semi-infinite string, the general problem with

the initial and boundary conditions
Special Case 1. As an example, we specify zero initial position and

velocity but allow motion of the left end point:
(3.7)
u(x, 0) = 0, ut (x, 0) = 0 for x > 0,
while u(0, t) = h(t) for t > 0.
u(x, 0) = f (x),
Well assume that h(0) = 0 to insure continuity at the origin.

The critical characteristic x = ct is important here. The domain of
dependence of any point to the right of this line does not include the
positive t-axis. Thus, if x ct, then u(x, t) = 0. Next we consider
a point (, ) above this characteristic. The simplest approach is to
use the identity (3.3) with a characteristic parallelogram having its
base on the critical characteristic x = ct. The
characteristic of the form xct = const. through
(, ) intersects the t-axis at t = /c. Since
u(x, t) = 0 on the base of this parallelogram,
(,)
then by (3.3) we conclude that u(, ) = h(
t
/c). To summarize, we see that
(0,/c )
(
x
0
for 0 t x
u(x, t) =
h(t x/c)
for 0 x t.
Exercise: For the semi-infinite string 0 < x, solve the initial-boundary

value problem where the end at x = 0 is free (Neumann boundary
condition):
Special Case 2. A clever observation helps to solve the related problem for a semi-infinite string:
(3.8)
u(x, 0) = f (x), ut (x, 0) = g(x) for x > 0,
while u(0, t) = 0 for t > 0.
The observation is that for the infinite string < x < , if the
initial position u(x, 0) = f (x) and velocity ut (x, 0) = g(x) are odd
functions, then so is the solution u(x, t) (proof?). Thus, to solve (3.8)
we simply extend f (x) and g(x) to all of R as odd functions fodd (x)
and godd (x) and then use the dAlembert formula (3.5).
Exercise: Carry this out explicitly for the special case where (3.8)
holds with g(x) = 0. In particular, show that for x > 0 and t > 0
(
1
[f (x + ct) + f (x ct)] for x > ct
u(x, t) = 12
[f (ct + x) f (ct x)] for x < ct.
2
The boundary condition at x = 0 serves as a reflection. One can
see this clearly from a sketch, say with the specific function f (x) =
(x 2)(3 x)) for 2 x 3 and f (x) = 0 for both 0 x 2 and
x > 3.
ut (x, 0) = g(x), for x > 0,
while u(0, t) = h(t) for t > 0
can now be solved by simply adding the solutions from the two special
cases (3.7) (3.8) just treated.
u(x, 0) = f (x),
ut (x, 0) = g(x) for x > 0,
while ux (0, t) = 0 for t > 0.
c). Finite string: 0 < x < L. In the case of a finite string, such
as a violin string, one must evidently also say something about the
motion of the end points x = 0 and x = L. One typical situation is
where we specify the position of these boundary points:
(3.9)
left end: u(0, t) = (t),
right end: u(L, t) = (t).
Thus, if the ends are tied down we would let f (t) = g(t) = 0. The
equations (3.9) are called boundary conditions. As an alternate, one
can impose other similar boundary conditions. Thus, if the right end
is allowed to move freely and the left end is fixed ((t) = (t) = 0),
then the above boundary conditions become
u
(3.10)
u(0, t) = 0
(L, t) = 0,
x
The condition at x = L asserts the slope is zero there (that the slope at
a free end is zero follows from physical considerations not given here).
There is no simple closed form solution of the mixed initial-boundary
value problem (3.1),(3.4), (3.9), even in the case f (t) = g(t) = 0. The
standard procedure one uses is separation of variables (see section c)
below). The solution is found as a Fourier series.
d). Conservation of Energy. For both physical and mathematical reasons, it is important to consider the energy in a vibrating string.
Here we work with an infinite string.
Z
(3.11)
E(t) = 12
(u2t + c2 u2x ) dx
u2t
The term
is for the kinetic energy and c2 u2x the potential energy.
(Here we have assumed the mass density is 1; otherwise E(t) should be
multiplied by that constant.) For this integral to converge, we need to
assume that ut and ux decay fast enough at . From the dAlembert
formula (3.5), this follows if the initial conditions decay at infinity.
We prove energy is conserved by showing that dE/dt = 0. This is a
straightforward computation involving one integration by parts in
3. TWO AND THREE SPACE DIMENSIONS
33
which the boundary terms dont appear because of the decay of the
solution at infinity.
Z
Z
dE
2
(3.12)
ut (utt c2 uxx ) dx = 0,
(ut utt + c ux uxt ) dx =
=
dt
where in the last step we used the fact that u is a solution of the wave
equation.
Exercises
1. For a finite string 0 < x < L with zero boundary conditions:
u(0, t) = u(L, t) = 0, define the energy as
Z L
(3.13)
E(t) = 21
(u2t + c2 u2x ) dx.
0
Show that energy is conserved. Show that energy is also conserved

if one uses the free boundary condition u/x = 0 at either or
both endpoints.
2. For a finite string 0 < x < L let u be a solution of the modified

wave equation
(3.14)
utt + b(x, t)ut = uxx + a(x, t)ux
with zero Dirichlet boundary conditions: u(0, t) = u(L, t) = 0,

where we assume that | a(x, t) |, | b(x, t) | < M for some constant
M . Define the energy by (3.13).
a) Show that E(t) et E(0) for some constant depending only
on M . [Suggestion: Use the inequality 2ab a2 + b2 .]
b) What happens if you replace the Dirichlet boundary conditions
by the Neumann boundary condition uN = 0 on the boundary (ends) of the string?
c) Generalize part a) to a bounded region in Rn .
3. Two and three space dimensions
In higher space dimensions, the wave equation is utt = c2 u. Thus, in
two and three space dimensions
(3.15)
utt = c2 (uxx + uyy )
and
utt = c2 (uxx + uyy + uzz ).
Two dimensional waves on a drum head and waves on the surface of a

lake are described by the first equation while sound and light waves are
described by the second. Just as in the one dimensional case we can
prescribe the initial position and initial velocity of the solution. For
instance, in two space variables
(3.16)
initial position
u(x, y, 0) = f (x, y)
(3.17)
initial velocity
ut (x, y, 0) = g(x, y).
34
a). Formulas for the solution in R2 and R3 . There are standard formulas for the solution of the initial value problem (the term
Cauchy problem is often called).
Technical Observation Let x = (x1 , . . . , xn ) Rn . Say we want
to solve
(3.18)
utt = u,
with u(x, 0) = f (x) and ut (x, 0) = g(x).
Let v(x, t) and w(x, t), respectively, be the solutions of

(3.19)
vtt = v,
with v(x, 0) = 0 and vt (x, 0) = f (x).
wtt = w,
with w(x, 0) = 0 and wt (x, 0) = g(x).
and
(3.20)
Then vt also satisfies the wave equation but with initial conditions
vt (x, 0) = f (x) and vtt = 0. Thus the solution of (3.18) is u(x, t) =
vt (x, t)+w(x, t). Since both (3.19) and (3.20) have zero initial position,
one can find u(x, t) after solving only problems like (3.20). This is
utilized to obtain the following two formulas.
For the two (space) dimensional wave equation it is
(3.21)
ZZ
ZZ
1
f (, )
g(, )
1
u(x, y, t) =
d d +
d d,
2
2
2
2c t
2c
c t r
c2 t2 r 2
rct
rct
where r = (x ) + (y ) .
In three (space) dimensions one has

(3.22)
ZZ

ZZ
1
1
f
(,
,
)
dA
+
g(, , ) dA,
u(x, y, z, t) =
4c2 t
4c2 t
r=ct
r=ct
where r = (x ) + (y ) + (z ) and dA is the element of

surface area on the sphere centered at (x, y, z) with radius r = ct.
These are called Kirchoff s formulas. It is simplest first to obtain the
formula in the three space dimensional case (3.22), and then obtain
the two dimensional case (3.21) from the special three dimensional case
where the initial data f (x, y, z) and g(x, y, z) are independent of z .
This observation is called Hadamards method of descent.
Exercises
1. Maxwells equations for an electromagnetic field E(x, t) = (E1 , E2 , E3 ),
B(x, t) = (B1 , B2 , B3 ) in a vacuum are
Et = curl B,
Bt = curl E,
div B = 0,
div E = 0.
Show that each of the components Ej and Bj satisfy the wave

equation utt = uxx . Also, show that if initially div B(x, 0) = 0 and
3. TWO AND THREE SPACE DIMENSIONS
35
div E(x, 0) = 0, then div B(x, t) = 0 and div E(x, t) = 0 for all
t > 0.
n
2. Let x = (x1 , . . . , xn ) R and consider the equation

n
X
2u
2u
ajk
,
=
2
t
xj xk
j,k=1
where the coefficients ajk are constants and (without loss of generality why?) akj = ajk . If the matrix A = (ajk ) is positive
definite, show there is a change of variable x = Sy , where S is
an n n invertible matrix, so that in these new coordinates the
equation becomes the standard wave equation
n
2u X 2u
.
=
t2
2 y
=1
b). Domain of dependence and finite signal speed. As before, it is instructive to examine intersection of the domain of dependence with the plane t = 0, in other words, to
determine the points x for which the initial data
can influence the signal at a later time. In the two
dimensional case (3.21), the intersection of the domain of dependence of the solution at (x0 , y0 , t0 )
t
with the plane t = 0 is the entire disc r ct0 ,
(x,y,t)
while in the three dimensional case (3.22), the domain of dependence is only the sphere r = ct0 ,
not the solid ball r ct0 . Physically, this is intery
ct
preted to mean that two dimensional waves travel
with a maximum speed c, but may move slower, x
while three dimensional waves always propagate
with the exact speed c.
This difference in observed in daily life. If one drops a pebble into
a calm pond, the waves (ripples) move outward from the center but
ripples persist even after the initial wave has passed. On the other
hand, an analogous light wave, such as a flash of light, moves outward
as a sharply defined signal and does not persist after the initial wave
has passed. Consequently, it is quite easy to transmit high fidelity
waves in three dimensions but not in two. Imagine the problems
in attempting to communicate using something like Morse code with
waves on the surface of a pond.
For the two space variable wave equation, the characteristics are the
surfaces of all light cones (x )2 + (y )2 = c2 t2 . In three space
dimensions, the characteristics are the three dimensional light cones.
They are the hypersurfaces in space-time with (x )2 + (y )2 +
(z )2 = c2 t2 .
36
4. Energy and Causality

One can also give a different prove of results concerning the domain of
dependence using an energy method. This technique is especially useful
in more general situations where explicit formulas such as (3.21)(3.22)
are not available.
Let x = (x1 , . . . , xn ) and let u(x, t) be a smooth solution of the ndimensional wave equation
(3.23)
utt = c2 u
where
u = ux1 x1 + + uxn xn ,
with initial data

(3.24)
u(x, 0) = f (x),
ut (x, 0) = g(x)
( Physicists often write the Laplacian, , as 2 . Some mathematicians

define with a minus sign, so for them, in R1 , u = u . Thus, one
must be vigilant about the sign convention.)
a). Conservation of energy. Just as in the one dimensional case,
we use the energy
Z
E(t) = 12
(u2t + c2 | u |2) dx,
Rn
where we assume the solution is so small at infinity that this integral

(and those below) converges. To prove conservation of energy, we show
that dE/dt = 0, The computation is essentially identical to the one
dimensional case we did above, only here we replace the integration by
parts by the divergence theorem.
Z
Z
dE
ut (utt c2 u) dx = 0.
(ut utt + c2 u ut ) dx =
=
dt
Rn
Rn
An immediate consequence of this is the uniqueness result: the wave
equation (3.23) with initial conditions (3.24) has at most one solution.
For if there were two solutions, v and w , then u := v w would be
a solution of the wave equation with zero initial data, and hence zero
initial energy. Since energy, E(t), is conserved, E(t) = 0 for all time
t 0. Because the integrand in E is a sum of squares, then ut = 0
and u = 0 for all t 0. Thus u(x, t) const.. However u(x, 0) = 0
so this constant can only be zero.
In two and three space dimensions this uniqueness also follows from the
explicit formulas (3.21)(3.22). However, the approach using energy
also works when there are no explicit formulas.
4. ENERGY AND CAUSALITY
37
b). Causality using energy. Energy gives another approach

to determine the domains of dependence and influence of the wave equation. Let P = (X, T ) be a point in
space-time and let
KP = { (x, t) : kx Xk c| t T | }
be the light cone with apex P . This

cone has two parts, that with t > T
is the future light cone while that
with t < T is the past light cone.
In the two and three (space) dimenfuture
sional case, from the explicit formulas for the solution we have seen
that the value of the solution at P
only depends on points in the past
P=(X, T)
light cone, and can only influence
the solution at points in the future t = t
1
light cone. Here we give another
past
demonstration of this that does not
rely on the earlier explicit formulas.
First, say t1 < T and let D(t1 ) be the intersection of KP with the
plane t = t1 . Define the energy function as
E(t) =
1
2
D(t)
D(t2 )
(u2t + c2 | u |2) dx
1
2
D(t1 )
Proof: We will show that dE(t)/dt 0 for 0 t T . In R we

use spherical coordinates centered at X , we have dx = dr dr , where
dr is the element of area on the n 1 sphere of radius r . Since the
radius of the ball D(t) is c(T t) we find that
Z
c(T t)
Z
S(r)

u2t + c2 | u |2 dr
Note that the integral on the right is just

Z

u2t + c2 | u |2 dA,
D(t)
where dA is the element of area on D(t). Since

u ut = (ut u) ut u,
then by the divergence theorem we have

Z
Z
Z
ut (u ) dA
u ut =
D(t)
D(t)
ut u dx,
D(t)
where is the unit outer normal vector to D(t). Upon substituting

into the formula for dE/dt, we find
Z
Z

2cut u u2t + c2 | u |2 dA.
ut (utt c2 u) dx+ 2c
E (t) =
D(t)
D(t)
Next, we note that utt = c u so the first integral is zero. For the
second term we use the standard inequality 2ab a2 + b2 for any real
a, b to obtain the estimate
dr,
There are two immediate consequences of the energy inequality of this

theorem.
Corollary 3.2 (uniqueness). . There is at most one solution of the
inhomogeneous wave equation utt c2 u = f (x, t) with initial data
(3.24).
(u2t + c2 | u |2) dx,
1
2
where S(r) is the n1 sphere (in the plane) with radius r and centered
at (X, t). Hence
Z
Z

dE
ut utt + c2 u ut dx 2c
=
u2t + c2 | u |2 dc(T t) .
dt
D(t)
S(c(T t)))
Consequently, E (t) 0. This completes the proof.
that is, energy E(t) is non-increasing for t T .

Consequently, if for some t1 < T we have u(x, t1 ) = 0 and ut (x, t1 ) = 0
for all x D(t1 ), then u(x, t) = 0 for all points in the cone with
t1 t T .
E(t) =
| 2cutu | +2c| utu | u2t + c2 | u |2.
(u2t + c2 | u |2) dx.
Theorem 3.1. If u(x, t) is a solution of the wave equation, and if

t1 < t2 < T , then
1
2
38
Corollary 3.3 (domain of influence). . Let u be a solution of the

initial value problem (3.23)(3.24). If u(x, 0) and ut (x, 0) are zero
outside the ball { kx Xk < }, then for t > 0, the solution u(x, t) is
zero outside the forward light cone { kx Xk < + ct }, t > 0.
Thus, for t > 0, the domain of influence of the ball { kx Xk < }

is contained in the cone { kx Xk < + ct } in the sense that if one
changes in the initial data only in this ball, then the solution can change
only in the cone.
Exercise: Let u(x, t) be a solution of the wave equation (3.14) for

x R. Use an energy argument to show that the solution u has the
same domain of dependence and range of influence as in the special
case where a(x, t) = b(x, t) = 0.
4. ENERGY AND CAUSALITY
39
c). Mixed Initial-Boundary Value Problems. The above formulas (3.21), (3.22) were for waves in all of space. In the case of a
vibrating membrane , we must also impose boundary values on ,
the boundary of . Similarly, in the case of light or sound waves outside of , we put boundary conditions on both and at infinity
(this is sometimes referred to as an exterior problem, while a vibrating
membrane is an interior problem. Just as for the vibrating string (...
), two typical boundary conditions are
u(x, t) = f (x, t) for x
(Dirichlet conditions)
(3.25) u
(x, t) = g(x, t) for x
(Neumann conditions),
where / means the directional derivative in the outer normal direction to . Of course this presumes that the boundary is smooth
enough to have an outer normal direction. One also has situations
where one of these conditions holds on part of the boundary and the
other on another part. The vibrating string (3.10) is an example.
We now restrict our attention to waves in a bounded region , such
as a vibrating membrane, and use the method of separation of variables to solve the wave equation with homogeneous Dirichlet boundary
conditions:
(3.26)
utt = u
(3.26a)
u(x, t) = 0
(3.26b)
u(x, 0) = f (x),
for
x ,
ut (x, 0) = g(x),
where we have let c = 1. We seek special standing wave solutions in

the form of a product
(3.27)
u(x, t) = W (x)T (t).
If one takes a sequence of photographs of such a solution at various

times t1 , t2 , . . ., then you see the graph of W (x) multiplied by the
factor T (t). The wave does not move horizontally, only up and down.
In order to satisfy the boundary condition (3.26a) we need W (x) = 0
for x . Substituting (3.27) into the wave equation we find that
W (x)
T (t)
=
.
W (x)
T (t)
Since the left side depends only on x while the right depends only on
t, they must both be equal to a constant, say . Thus we obtain the
two equations
(3.28)
T T = 0
W = W.
We next observe that for a non-trivial solution (3.27) we must have

< 0. To see this, multiply W = W by W and integrate by parts
40
over :
W 2 (x) dx =
W (x)W (x) dx =
| W |2 dx,
where we use the boundary condition W (x) = 0 on to eliminate

the boundary term. Solving this for we clearly see that < 0. In
view of this, it will be convenient to write = , so > 0. Then we
have
(3.29)
with
(3.30)
W = W
in
w = 0 on
| W |2 dx
| W |2 dx
Thus is an eigenvalue of the operator, with corresponding eigenfunction W (x). For a membrane , these eigenvalues are essentially
the squares of the various frequencies with which the membrane can
vibrate and the eigenfunctions are the normal modes. It turns out that
only a discrete sequence of eigenvalues 0 < 1 2 are possible
with j as j . Write the corresponding
eigenfunctions
as
p
p
j . From (3.28) the functions Tj (t) = aj cos j t + bj sin j t so the
standing wave solutions (3.27) are
p
p
(3.31)
uj (x, t) = aj cos j t + bj sin j t j (x),
where the aj and bj are arbitrary constants.

Next we seek the solution of (3.26), (3.26a), (3.26b) as a linear combination of these special standing wave solutions:
X
p
p
(3.32)
u(x, t) =
aj cos j t + bj sin j t j (x).
To satisfy the initial conditions (3.26b) we choose the constants aj and

bj so that
X
X p
f (x) =
aj j (x)
g(x) =
bj j j (x).
It is always possible to find these constants because the eigenfunctions

j are a complete orthonormal set on L2 (). This series formally
satisfies the differential equation, boundary conditions, and initial conditions. If f and g are sufficiently differentiable, then one can legitimately differentiate the above infinite series term-by-term to rigorously
verify that u(x, t) is an honest solution.
We can carry out these computations in only the simplest situations.
The most basic is for a vibrating string of length , so we let = { 0 <
x < }. Then (3.29) is
W + W = 0,
W (0) = 0,
W () = 0.
5. VARIATIONAL CHARACTERIZATION OF THE LOWEST EIGENVALUE 41
Thus W (x) = A cos t + B sin t. Now W (0) = 0 implies that

A = 0. Then to obtain a non-trivial solution, W () = 0 implies
that k = k 2 , k = 1, 2, . . . so the eigenfunctions are k (x) = sin kx,
k = 1, 2, . . . and the series (3.31) is a classical Fourier series. The lowest
eigenvalue, 1 , is the fundamental tone of the string while the higher
eigenvalues give the possible overtones or harmonics.
Exercises:
1. In the above, investigate what happens if you replace the Dirichlet
boundary condition u(x, t) = 0 for x by homogeneous Neumann boundary condition u/N = 0 for x . Note here that
= 0 is now an eigenvalue. What is the corresponding eigenfunction? Carry out the details for a vibrating string on the interval
0 < x < .
2. Find the motion u(x, t) of a string 0 x whose motion is
damped:
utt + 3ut = uxx ,
with
u(x, 0) = sin 3x 2 sin 5x,
ut x, 0 = 0,
u(0, t) = u(, t) = 0.
3. Prove the uniqueness of this solution of the problem (3.26) by an

energy argument using (3.13).
5. Variational Characterization of the Lowest Eigenvalue
The formula (3.30) is essentially identical to the formula = hx, Axi/kxk2
for the eigenvalues of a self-adjoint matrix A. A standard fact in linear
algebra is that the lowest eigenvalue is given by 1 = minx6=0 hx, Axi/kxk2
(proof?). It is thus natural to surmise that the lowest eigenvalue of the
Laplacian satisfies
Z
| |2 dx
(3.33)
1 = min Z
,
| |2 dx
where the minimum is taken over all C 1 functions that satisfy the
Dirichlet boundary condition = 0 on . Assuming there is a function C 2 ()) C 1 () that minimizes (3.30), we will show that it is
an eigenfunction with lowest eigenvalue 1 . To see this say such a
minimizes the functional
Z
| v |2 dx
J(v) = Z
,
| v |2 dx
42
so J() J(v) for all v C 1 () with v = 0 on . Let F (t) :=

J( + th) for any h C 1 () with h = 0 on and all real t. Then
F (t) has its minimum at t = 0 so by elementary calculus, F (0) = 0.
By a straightforward computation, just as in the case of matrices,
Z
( h 1 h) dx
Z
F (0) = 2
.
| |2 dx
We integrate the first term in the numerator by parts. There are no

boundary terms since h = 0 on . Thus
Z
[( 1 )h] dx
F (0) = 2 Z
.
| |2 dx
Since F (0) = 0 for all of our h, we conclude the desired result:

= 1 .
Equation (3.33) is called the variational characterization of the lowest

eigenvalue. There are analogous formulas for higher eigenvalues. Such
formulas useful for computing numerical approximations to eigenvalues,
and also to prove the existence of eigenvalues and eigenfunctions. The
fraction in (3.33) is called the Raleigh (or Raleigh-Ritz ) quotient.
Equation (3.33) implies the Poincare inequality
Z
Z
| |2 c() | |2 dx
(3.34)
for all C 1 () that vanish on (these are our admissible ).

Moreover, it asserts that 1/1 () is the best value for the constant c.
It is instructive to give a direct proof of the Poincare inequality since
it will give an estimate for the eigenvalue 1 (). Let V be a vector
field on Rn (to be chosen later). For any of our admissible , by the
divergence theorem
Z
Z
Z
div (2 V ) dx = [2 div V + V ] dx,
2 V N dA =
0=
where N is the unit outer normal vector field on . Now pick V so

that div = 1, say V = (x1 , 0, . . . , 0). Then picking the constant
appropriately, | V | w/2, where w is the width of in the x1
direction. Therefore, by the Schwarz inequality,
Z
1/2 Z
1/2
Z
w
2
2
2
[ dx
dx
| | dx
.
2
Squaring both sides and canceling gives (3.34) with c = (w/2)2 , so

1 () w 2 /4.
6. SMOOTHNESS OF SOLUTIONS
43
Using the variational characterization(3.33), it is easy to prove a physically intuitive fact about vibrating membranes: larger membranes have
a lower fundamental frequency. To prove this, say + are bounded
domains with corresponding lowest eigenvalues 1 () and 1 (+ ).
Both of these eigenvalues are minima of the functional (3.33), the only
difference being the class of functions for which the minimum is taken.
Now every admissible function for the smaller domain is zero on
and hence can be extended to the larger domain by setting it to be zero
outside . It is now also an admissible function for the larger domain
+ . Therefore, for the larger domain the class of admissible functions
for J(v) is larger than for the smaller domain . Hence its minimum
1 (+ ) is no larger than ().
Using similar reasoning, one can prove a number of related facts, and
also get explicit estimates for eigenvalues. For instance, if we place
R2 in a rectangle + , since using Fourier series we can compute
the eigenvalues for a rectangle, we get a lower bound for ().
6. Smoothness of solutions
From the formula u(x, t) = F (x ct) + G(x + ct) for the solution of the
one dimensional wave equation, since formally F and G can be any
functions, it is clear that a solution of the wave equation need not be
smooth (this is in contrast to the solutions of the Laplace equation, as
we shall see later). In fact, in higher dimensions, even if the initial data
(3.16) are smooth, the solution need not even be continuous. This can
be seen intuitively for three space variables by choosing initial conditions on a sphere so that light rays are focuses at the origin at a later
time. This is commonly done with a lens. To see this with formulas,
notice that for any smooth f C (R) the function
u(x, y, z, t) =
f (r + ct)
,
r
where r 2 = x2 + y 2 + z 2 ,
2
formally satisfies the wave equation utt = c u. For small t it is a

classical (that is, C 2 ,) solution even at r = 0 if we pick a smooth function f so that f (s) = 0 for | s | < 1. The solution represents spherical
waves coming to a focus at the origin. For such f both u(x, y, z, 0) and
ut (x, y, z, 0) are smooth everywhere., however, if say f (2) 6= 0, then
at time t = 2/c the solution u(x, y, z, t) will blow-up at the origin.
Nonetheless, one can make both physical and mathematical sense of
this physically common situation. Since energy is conserved, the solution and its first derivatives are square-integrable. This can be used to
define the concept of a weak solution of the wave equation. We take
this up later (see ??).
44
7. The inhomogeneous equation. Duhamels principle.

There are also formulas for the solution of the inhomogeneous wave
equation
(3.35)
Lu := utt c2 u = F (x, t).
The approach is analogous to Lagranges method of variation of parameters, which gives a formula for the solution of an inhomogeneous
equation such as u +u = F (t) in terms of solutions of the homogeneous
equation. The method is called Duhamels principle.
We illustrate it for the wave equation, seeking a solution of (3.35) with
initial conditions
u(x, 0) = 0 ut (x, 0) = 0.
Since we are solving a differential equation, it is plausible to find a
solution as an integral in the form
Z t
v(x, t; s) ds
(3.36)
u(x, t) =
0
where the function v(x, t; s), which depends on a parameter s, is to be

found. This clearly already satisfies the initial condition u(x, 0) = 0.
Working formally, we have
Z t
vt (x, t; s) ds + v(x, t; t),
ut (x, t) =
0
so ut (x, 0) = 0 implies v(x, 0; 0) = 0. In fact, we will further restrict v

by requiring that v(x, t; t) = 0 for all t 0. Then the formula for ut
simplifies and
Z t

vtt (x, t; s) ds + vt (x, t; s)s=t .
utt (x, t) =
0
The similar formula for u is obvious. Substituting these into the wave
equation (3.35) we want
Z t

Lv(x, t; s) ds + vt (x, t; s)s=t .
F (x, t) = Lu(x, t) =
0

This is evidently satisfied if Lv = 0 and vt (x, t; s)s=t = F (x, t) along
with v(x, t; t) = 0 for all t 0.
Because the coefficients in the wave equation do not depend on t, our
results can be simplified a bit by writing v(x, t; s) = w(x, t s; s) so
for each fixed s, the function w(x, t; s) satisfies
(3.37) wtt = c2 w
with w(x, 0, ; s) = 0 and wt (x, 0; s) = F (x, s).
7. THE INHOMOGENEOUS EQUATION. DUHAMELS PRINCIPLE.
45
We can now find w by using our earlier formulas. For instance, in three
space variables, from (3.22)
ZZ
1
F (, s) dA ,
w(x, t; s) =
4c2 t
kxk=ct
where dA is the element of surface area on the sphere centered at x

with radius ct, that is, k xk = ct. Therefore from (3.36)
ZZ
Z t
1
1
u(x, t) =
F (, s) dA ds
4c2 0 t s
=
1
4c
kxk=c(ts)
ZZ
kxk=c(ts)
F (, t k xk/c)
dA ds.
k xk
But in spherical coordinates, the element of volume d = cdA ds so we

finally obtain
ZZZ
F (, t k xk/c)
1
d.
(3.38)
u(x, t) =
4c2
k xk
kxkct
Thus, to solve the inhomogeneous equation we integrate over backward

cone k xk ct, which is exactly the domain of dependence of the
point (x, t).
Exercises
1. Use Duhamels principle to obtain a formula for the solution of
u + k 2 u = f (x),
x R,
with u(0) = 0, u (0) = 0.
Similarly, do this for u k 2 u = f (x).

2. Use (3.21) to derive the analog of (3.38) for one and two space
variables.
3. Let x Rn .
a) If function w(x) depends only on the distance to the origin,
r = kxk, show that
2 u n 1 u
+
.
2r
r r
b) Investigate solutions u(x, t), x R3 of the wave equation
utt = u where u(x, t) = v(r, t) depends only on r and t.
For instance, let v(r, t) := rw(r, t) and note that v satisfies a
simpler equation. Use this to solve the wave equation in R3
where the initial data are radial functions:
u =
v(r, 0) = (r),
vt (r, 0) = (r).
[Suggestion: Extend both and as even functions of r .]
46
Are there solutions of wt t(r, t) = (r, t) with the form w(r, t) =

h(r)g(r t)?
48
4. THE HEAT EQUATION
for all x. Thus, the solution of the heat equation is a smoothing

operator.
CHAPTER 4
The Heat Equation

1. Introduction
If we have a body in Rn , then under reasonable assumptions the
differential equation
(4.1)
ut = ku,
governs the temperature u(x, t) at a point x at time t. Here k > 0,

assumed constant in this example, describes the thermal conductivity
of the body. It is large for copper and small for wood. By scaling x
we can let k = 1. From experience in daily life, everyone has already
done many experiments with heat flow. As we will see, to a surprising
extent, the simple model of equation (4.1) embodies this intuition. This
equation also describes diffusion.
2. Solution for Rn
.
a). Homogeneous equation. There are many approaches to get
the formula for the solution of (4.1) in the special case where is all of
Rn . Perhaps the most straightforward but not the most elementary
is to use the Fourier transform.
R1 We first treat the one dimensional case of an infinite rod <
x < , so the problem is to solve the standard initial value problem
(4.2)
ut = uxx
with
u(x, 0) = f (x).
Assuming a mild growth condition on f , say it is bounded and continuous, the solution is
Z
(xs)2
1
(4.3)
u(x, t) =
f (s)e 4t ds.
4t
Before going further it is useful to make some observations based on this
formula. First, it implies that if the initial temperature is non-negative
but not identically zero, then the solution is positive everywhere, even
for very small t. Thus, in contrast to the solution of the wave equation,
heat conduction has an infinite signal speed. We also observe that even
if f is, say, only piecewise continuous, the solution is smooth in both
x and t for all t > 0. In fact, it has a power series in x that converges
47
To derive (4.3) for the moment we work formally and assume all integrals make sense. First take the Fourier transform of ut = uxx with
respect to the space variable x,
Z
u(x, t)eix dx.
u(, t) :=
Then from (4.27) in the Appendix to this chapter u(, t) satisfies the
ordinary differential equation
ut = | |2 u
with
u(, 0) = f()
in which appears only as a parameter. Its solution is
2
u(, t) = e| | t f(t).
Thus, by Fourier inversion (4.24) and the computation (4.23) we get

the desired formula
Z
1
2
u(x, t) =
et| | +ix f() d
2
Z

Z
1
2
f (y)
et| | +i(xy) d dy
=
2
Z
| xy |2
1
=
f (y)e 4t dy.
4t
This derivation was purely formal. Since the resulting formula may well
hold under more general conditions than this derivation admits, instead
of checking each step we verify directly that it solves the heat equation
and satisfies the initial condition. By differentiating under the integral
we immediately verify that it satisfies the heat equation ut = uxx for
all t > 0. Moreover u is a smooth function of x and t for all t > 0.
It remains to show that limt0 u(x, t) = f (x). This is a special case of
the next lemma.
Let C(R) have the properties
(1)
R (x) 0,
(2) R (x) dx = 1,
R
(3) For any > 0, lim0 | y | (y) dy = 0.
Let
(4.4)
f (x) :=
f (y)(x y) dy.
In applying this to the heat equation we will let

| x |2
1
e 4 .
(x) =
4
Comparing with (4.3) we see that f (x) = u(x, ).
2. SOLUTION FOR Rn
49
Lemma 4.1. If f C(R) is bounded and as above, then lim0 f (x) =

f (x), where the limit is uniform on compact subsets. Moreover, if is
smooth, then so is f .
Proof. To prove the uniform convergence in a compact interval
K R, given > 0, use the uniform continuity of the continuous
function f on a slightly larger interval K1 to find > 0 so that if
x K and | w | < with (x w) K1 , then | f (w) f (x) | < . Also,
say | f (x) | M . After the change of variable x y = z we get
Z
Z
f (x) f (x) = [f (y) f (x)] (x y) dy = [f (x z) f (x)] (z) dz
R
Z
ZR
[f (x z) f (x)] (z) dz
[f (x z) f (x)] (z) dz +
=
| z |
| z |<
Thus,
| f(x) f (x) | < + 2M
| z |
Remark In R the convolution f g of f and g is defined as

Z
f (y)g(x y) dy.
(f g)(x) =
R
The definition (4.4) defines f as a convolution. Note that f is only

continuous but g is smooth, then f g is smooth assuming the integral exists. If the are smooth, the above proof shows that on compact subsets we can uniformly approximated f by the smooth function
f . This technique of smoothing (or mollifying) a function is valuable.
Weierstrass used (4.3) in his original proof of what we now call the
Weierstrass Approximation Theorem.
Exercise:
a) Solve ut = uxx + au for x R, where a = const, with u(x, 0) =
f (x). [Suggestion: Let u(x, t) = (t)v(x, t), picking cleverly.
b) Solve ut = uxx bux for x R, where b = const, with u(x, 0) =
f (x). The term bux introduces convection. [Suggestion: Introduce a moving frame of reference by letting y = x bt.]
Rn There is a similar formula for the solution of the heat equation
for x = (x1 , . . . , xn ) in Rn . In this case we seek a solution of
ut = u
This solution is given by the formula

Z
| xy |2
1
(4.6)
u(x, t) =
f (y)e 4t dy.
n/2
(4t)
Rn
To verify this, one uses the routine generalization of the above lemma.
Exercise: Use Fourier transforms to obtain (4.6).
Below we will use the maximum principle to show that with an essential
boundedness assumption, the solution of (4.5) is unique.
b). Inhomogeneous equation. Using Duhamels principle it is
straightforward to obtain a formula for the solution of the inhomogeneous equation
(4.7)
ut u = F (x, t)
Seek u in the form

(z) dz
By Property 3) the last integral can be made arbitrarily small by choosing sufficiently small. Since the right hand side is independent of x
(as long as x K ), the convergence is uniform.
(4.5)
50
with initial temperature u(x, 0) = f (x).
u(x, t) =
with u(x, 0) = 0.
v(x, t; s) ds.
Clearly u(x, 0) = 0. Also, working formally,

Z t
vt (x, t; s) ds + v(x, t; t).
ut =
0
with a similar formula for u. Consequently

Z t
[vt v] ds + v(x, t; t).
ut u =
0
Since we want to solve ut u = F , it is natural to specify

vt v = 0
with
v(x, t; s)|s=t = F (x, t).
The function v is given by (4.6) except that we specify the initial

temperature at t = s. Thus the desired solution u(x, t) of (4.7) for
x Rn , t > 0, is
Z t
Z
| xy |2
1
u(x, t) =
f (y, s)e 4(ts) dy ds.
n/2
0 [4(t s)]
Rn
3. Initial-boundary value problems for a bounded region,
part 1
To determine the temperature in a bounded region , it is clear that
we will need to know the initial temperature u(x, 0) and also something
about the boundary. Two typical situations are that we might specify
the temperature u(x, t) at some boundary points x while ask that the
boundary be insulated at other boundary points. As mentioned above,
at a point where the boundary is insulated, the appropriate boundary
condition is that the directional derivative in the outer normal direction
is zero there: u/N = 0.
4. MAXIMUM PRINCIPLE
51
Thus, if we specify the temperature at all boundary points, we are

asking to solve the heat equation with
(4.8)
initial temperature
u(x, 0) = f (x)
(4.9)
boundary temperature
u(x, t) = g(x, t)
for x
for x .
We call (4.9) a Dirichlet boundary condition.

The special case of the boundary condition u(x, t) = g(x) means that
the temperature at all boundary points does not depend on the time.
Assuming this, here are two assertions that are intuitively clear.
Say the initial and boundary temperatures are at most M .
Then at any time in the future, the maximum temperature is
at most M . This is called the maximum principle. We discuss
it in the next section.
Eventually, the temperature throughout the body tends to
some equilibrium temperature, u(x, t) v(x), where v(x)
depends only on the boundary temperature, not on the initial
temperature. This will be treated later in this chapter.
One test of the mathematical model is to prove these assertions from
the data specified.
Above we specified Dirichlet boundary conditions. As an alternate, on
some or all of the boundary of one can prescribe the outer normal
u
derivative, N
:= N u. This is the directional derivative in the
direction of the outer normal:
u
= g(x, t), for x .
(4.10)
N
This is called a Neumann boundary condition. The special case of an
u
insulated boundary, so N
= 0, arises frequently.
Mixed boundary conditions
u(x, t) + c(x, t)
u(x, t)
= g(x, t),
N
for x
also arise occasionally.
52
and
(4.12)
w(x, 0) 0 for
while
w(x, t) 0 for x ,
0tT
then either w(x, t) > 0 for all x , 0 < t T or else w(x, t) 0

for all x , 0 t T .
For simplicity we prove only the weaker statement that w(x, t) 0.
First, to make the proof more transparent first assume that wt w >
0. Reasoning by contradiction, say w(x, t) < 0 somewhere in S :=
{ [0, T ] }. Then it is negative at its absolute minimum at some
interior point (x0 , t0 ) with x0 T . But at this point, if 0 < t0 < T ,
we know that wt = 0, while if t0 = T then wt 0. Moreover, by the
second derivative test for a minimum we know that w 0 at x0 .
These facts contradict our assumption that wt w > 0.
Next, assume only that wt w 0. We will use a limiting argument
to prove that w(x, t) 0. Again by contradiction, say w(x0 , T ) = m <
0 at some interior point x0 . Let z(x, t) := w(x, t) | x x0 |2 .
Pick > 0 so small that z(x, t) > m on PT . Then z has its minimum
at a point (x1 , t1 ) where x1 T . Since zt z > 0, we can apply
the reasoning of the above paragraph to obtain a contradiction.
Corollary 4.3 (Strong Maximum Principle). In T assume the solution u(x, t)of the heat equation is in C 2 for in x , C 1 for t in
(0, T ] in t. Also assume that u C(T ). Then
max u(x, t) = max u(x, t).
PT
Moreover, if u(x, t) attains its maximum at some point (x0 , t0 ) T ,

then u is constant throughout the cylinder t0 .
a). Applications of the maximum principle. Here are several
typical consequences of the maximum principle.
Say
(4.13)
ut ku = F (x, t) vt kv = G(x, t)
(4.14)
u(x, 0) = f (x)
v(x, 0) = g(x)
To state the maximum principle we introduce some notation. If Rn

is a bounded connected open set, for a fixed T > 0 let T := (0, T ]
so is a cylinder in space-time. Its parabolic boundary is PT = T
T . This consists of the sides and bottom of the closed cylinder T .
The maximum principle will be a consequence of the assertion
(4.15)
u(x, t) = (x, t)
for x
v(x, t) = (x, t)
for x , t > 0.
Theorem 4.2. In a bounded open set , if the function w(x, t) satisfies
then u(x, T ) v(x, T ), with strict inequality holding unless F (x, t)

G(x, t), f (x) g(x), and (x, t) (x, t) for all 0 t T .
4. Maximum Principle
(4.11)
wt w 0
for x T ,
Corollary 4.4 (Comparison of solutions). If

F (x, t) G(x, t), f (x) g(x),
and (x, t) (x, t)
for all 0 t T,
4. MAXIMUM PRINCIPLE
53
Corollary 4.5 (Growth estimate). Say | F (x, t) | M , | f (x) | c

and | (x, t) | c. Let w(x, t) be a solution of
(4.16)
wt = w + 1 with w(x, 0) = 0 in and w(x, t) = 0, x .
Then
(4.17)
| u(x, t) | c + M| w(x, t) | for
t 0.
Corollary 4.6 (Uniqueness). There is at most one solution of ut

ku = F (x, t) with
u(x, 0) = f (x) (x ),
and
u(x, t) = (x, t)
(x , t > 0).
Corollary 4.7 (Stability). If the functions F , f , and are perturbed

slightly, then the solution is perturbed only slightly. To be specific, say
| F (x, t)G(x, t) | < , | f (x)g(x) | < ,
and
| | <
for
x , t 0.
Then | u(x, t) v(x, t) | < for x , t o, where is small if ,

, and are small.
This is essentially just a restatement of (4.17) applied to u v .

Exercises
1. Prove Corollary 4.
3. Find an explicit estimate for the solution w(x, t) in (4.16). The
estimate will involve some property of , such as its diameter.
54
theorem for the initial value problem (4.5) for the heat equation in all of
Rn . We prove uniqueness for this in the class of bounded solutions (one
2
can weaken this to allow u(x, t) consteconst | x | , see [PW], p. 181,
but there are examples of non-uniqueness if one allows faster growth).
Say u(x, t) satisfies the heat equation ut = u in all of Rn with
u(x, 0) = 0 and | u(x, t) | M . Inside the disk {| x | < a} consider
the comparison function v(x, t; a) := M(| x |2 + 2nt)/a2 . Then v also
satisfies the heat equation with
v(x, 0; a) 0 while v(x, t; a) M u(x, t) for | x | = a, t > 0.
Thus by the maximum principle u(x, t) v(x, t; a) for | x | a. Fixing
(x, t) but letting a we conclude that u(x, t) 0. Replacing u
by u we then get u(x, t) = 0 for all t 0.
Exercises
1. Let u(x, t) be a bounded solution of the heat equation ut = uxx
with initial temperature u(x, 0) = f (x). If f (x) is an odd function
of x R, show that the solution u(x, t) is also an odd function of
x.
2. Semi-infinite interval Solve the heat equation on a half-line:
0 < x < with u(x, 0) = f (x) for x 0 and the following
conditions:
a) u(x, 0) = f (x) for x 0 and u(0, t) = 0 for t 0. [Suggestion: Extend f (x) cleverly to x < 0.]
b) u(x, 0) = 0 and u(0, t) = g(t). [Suggestion: Let v(x, t) =
u(x, t) g(t).]
c) u(x, 0) = f (x) for x 0 and u(0, t) = g(t) for t 0.
b). Symmetry of solutions. Uniqueness is often the easiest approach to show that a solution possesses some symmetry. One example makes the ideas transparent. Let R2 be the rectangle
{ | x | < 1, 0 < y < 1 } and let : be the reflection across
the y -axis. Assume the initial and boundary temperatures are invariant under , so they are even functions of x. We claim the solution is also invariant under . This is obvious since both u(x, t) and
v(x, t) := u((x), t) are solutions of the heat equation with the same
initial and boundary values.
c). Uniqueness in Rn . If is unbounded, such as an infinite rod
{ < x < }, then the simple example u(x, t) := 2t + x2 which
satisfies the heat equation but whose maximum does not occur at t = 0
shows that the maximum principle fails unless it is modified. However,
we can still use it as a tool. To illustrate, well prove a uniqueness
5. Initial-boundary value problems for a bounded region,

part 2
a). Using separation of variables. We seek special solutions of
the heat equation in a bounded region with zero Dirichlet boundary
conditions:
(4.18)
ut = u for x , with u(x, 0) = f (x) and u(x, t) = 0 x .
Because regions are rarely simple, one can almost never fill-in many
details, yet even working crudely one can get useful information. Just as
for the wave equation, one can use separation of variables to seek special
solutions u(x, t) = v(x)T (t) with v = 0 on the boundary of . As
before, v must be an eigenfunction vk of the Laplacian, with eigenvalue
k , that is, vk = v vk . We may assume the eigenfunctions are
orthonormal. Then Tk (t) = ek t so the special solutions are uk (x, t) =
5. INITIAL-BOUNDARY VALUE PROBLEMS FOR A BOUNDED REGION, PART55

2
vk (x)ek t . We build the general solution as a linear combination:

X
(4.19)
u(x, t) =
ak vk (x)ek t ,
where the ak are found using the initial condition

X
f (x) = u(x, 0) =
ak vk (x)
so ak = hf, vk i.
56
3. Let u(x, t) be a solution of the heat equation ut = u in with

homogeneous Neumann boundary conditions, u/N
R = 0 on ,
so the boundary is insulated. Show that Q(t) := u(x, t) dx =
constant.
4. Find a modified version of Exercises 1-2 above for the case of homogeneous Neumann boundary conditions.
Consequently, working formally,
6. Appendix: The Fourier transform
(4.20)
u(x, t) =
(4.21)
=
hf, vk ivk (x)ek t =
f (y)G(x, y)dy,
XZ
f (y)vk (y)vk (x)ek t dy
where G(x, y, t) =
vk (y)vk (x)ek t .
Here G(x, y, t) is called Greens function for the problem. Because the
eigenvalues, k are all positive, it is clear from (4.19) that u(x, t) 0
as t tends to infinity. This should agree with your physical intuition.
The lowest eigenvalue, 1 , determines the decay rate.
Exercise: Repeat this using homogeneous Neumann boundary conditions u/N = 0 on the boundary. What can you say about limt u(x, t)?
b). Another approach. Using techniques similar to the energy
methods we used for the wave equation, we can also obtain information
about solutions of the heat equation. These are reasonable exercises.
Exercises
1. Let u(x, t) be a solution of the heat equation ut = u in with
u = 0 on . Define
Z
H(t) := 12
u2 (x, t) dx.
Show that dH/dt 0. Then use this to prove a uniqueness theorem.

2. [Improvement of the previous Exercise] Use the variational characterization of 1 (see our discussion of the wave equation) to show
that
dH
1 H(t).
dt
Then use this to show that H(t) ce1 t for some constant c. This
proof of decay is independent of the previous version that used
separation of variables.
To derive the standard formula for the solution of the heat equation of
an infinite rod, we used the Fourier transform. Here is a brief summary
of basic facts about the Fourier transform. If u L1 (Rn ), its Fourier
transform u() is defined as
Z
u(x)eix dx,
(4.22)
u() :=
Rn
It is evident that | u
() | kukL1 (Rn ) .
a). A special integral.
Lemma 4.8. Let A be a real n n positive definite symmetric matrix

and b a complex vector. Then
n 21
Z
1
xAx+bx
e
dx =
(4.23)
I :=
ebA b/4 ,
det
A
n
R
where b x is the usual inner product in Rn .
Proof. Since A is positive definite, it has a positive definite square

root P , P 2 = A; this is obvious in a basis in which A is diagonalized.
Make the (real) change of variables y = P x in the above integral.
Then we have dy = (det P ) dx = (det A)1/2 dx and with := 12 P 1 b,
by completing the square
x Ax b x = | y |2 2 y = | y |2 14 b A1 b.
Let z := y and c := Im{}. Then the above integral, I becomes

!
Z icj
1
n
ebA b/4 Y
zj2
I=
e dzj .
(det A)1/2 j=1
icj
To complete the computation we need to evaluate the complex integrals

on the right. In the complex plane, integrate around the rectangle
with vertices at (R, Riq), where R and q are real, and let R
to conclude that
Z
Z iq
2
2
e d = .
e d =
iq
Combined with the above formula for I this gives the desired formula.
6. APPENDIX: THE FOURIER TRANSFORM
57
| x |2 /2
We use this to compute the Fourier transform of (x) = e

Z
2
2
=
e| x | /2 ix dx = (2)n/2 e| | /2 .
(4.24)
()
Rn
| x |2 /2
Thus e
is an eigenfunction of the Fourier transform operator.
b). Inversion of the Fourier transform. The formula

Z
1
(4.25)
u(x) =
u()eix d
(2)n
shows how to recover a function from its Fourier transform. To prove
this, say u L1 (Rn ) is a bounded function and pick some so that
both and are bounded and in L1 (Rn ) (below we make the specific
2
choice (x) = e| x | /2 ). Use the notation () := (). Then by an
Now
easy computation its Fourier transform is (y) = n (y/).
(4.26)

Z Z
Z
u(y)eiy dy ()eix d
u() ()eix d =
Rn
Rn
Z

Z
u(y)
()ei(yx) d dy
=
n
Rn
Z
ZR
dt.
u(x + t)(t)
u(y) (y x) dy =
=
Rn
Rn
In this computation we were permitted to interchange the orders of

integration since the integrals all converge absolutely. Because u is
bounded, by the dominated convergence theorem we can let 0 to
obtain the identity
Z
Z
ix
dt.
(t)
u()e d = u(x)
(0)
Rn
Rn
| x |2 /2
Choosing (x) = e
and using (4.24) gives the desired Fourier
inversion formula, at least for bounded functions u L1 (R).
c). Fourier transform of the derivative. One reason the Fourier
transform is so useful when discussing linear differential equations with
constant coefficients is that the Fourier transform changes differentiation into multiplication by a polynomial. This is easily seen by integrating by parts
Z
j u(x)eix dx = ij u().
(4.27)
c
j u() =
[
In particular, (u)()
= | |2 u(). so for any integer k 0
(1 + | |2 )k u() = [(1 \
)k u]().
58
Therefore, if u C (Rn ), then for any integer k 0 there is a constant

c depending on u and k so that
c
(4.28)
|u
() |
.
(1 + | |2 )k
It is useful to compare this to the result in Chapter 2 concerning the
decay of Fourier coefficients of smooth functions. They are essentially
identical.
60
5. THE LAPLACE EQUATION
2. Poisson Equation in Rn
CHAPTER 5
The Laplace Equation

1. Introduction
As we saw in the previous chapter, if v(x, t) is a solution of the heat
equation and if that solution converges to an equilibrium state u(x),
then u is a solution of the Laplace equation:
(5.1)
u = 0.
These are called harmonic functions.

Harmonic functions are invariant under both translations: x x a,
orthogonal transformations: x Rx, and scalings: x x.
Although we will not exploit it here, less obvious is that is the behavior
under inversions in the unit sphere: x x = x/| x |2 . Note that x
is on the same ray from the origin as x and | x || x | = 1. Given a
domain , let be its image under this inversion. For instance, the
inversion of the unit ball | x | < 1 is the exterior of this same ball. In
dimension two, harmonic functions are invariant under inversions. For
higher dimensions, define the Kelvin transform by
K(u)(x) :=
u(x )
.
| x |n2
If u(x) is harmonic in , then K(u) is harmonic in . This follows

from the identity (K(u)) = K(| x |4 u) which is most easily proved
first for homogeneous polynomials and then use that one can approximate any u C 2 (in the C 2 norm) by a polynomial.
In this chapter we will also briefly discuss both harmonic functions and
solutions of the inhomogeneous equation
u = f (x),
which is called the Poisson Equation.
A useful reference for this chapter is the first part of the book
Axler, S., Bourdin, P., and Ramey, Harmonic Function Theory, accessible
at
http://www.axler.net/HFT.pdf.
59
We first seek a particular solution of the Poisson Equation in Rn . For

this, we look for a solution of the very special equation
(5.2)
= 0 ,
where 0 (x) is the Dirac delta measure concentrated at the origin.

Since 0 (x) = 0 except at x = 0 and since the Laplacian is invariant
under orthogonal transformations, it is plausible to seek a solution (x)
of (5.2) as a function depending only on the radial direction r = | x |,
so (x) = v(r) is harmonic away from the origin. By the chain rule,
d2 v n 1 dv
.
+
dr 2
r dr
Thus we seek solutions of the ordinary differential equation
n1
v +
v = 0.
r
This is straightforward and gives
(
a log r + b for n = 2,
v(r) =
a
+b
for n 3.
r n2
v(r) =
To get a solution of (5.2), one lets b = 0 and picks a appropriately to

define
(
1
log| x |,
for n = 2
2
(5.3)
(x) =
1
for n 3,
n(n2)n | x |n2
where n is the volume of the unit ball B(0, 1) = {x Rn || x | < 1},
so nn is the area of the unit sphere S n1 = B(0, 1). This function
(x) is called the fundamental solution of the Laplacian.
Since = 0 , we guess that a solution of u = f is given by
Z
f (y)(x y) dy.
(5.4)
u(x) =
Rn
To verify this, it is tempting to take the Laplacian of both sides, but

since we believe (x y) = x , which is highly singular at y = x, we
must proceed more carefully and need to assume some smoothness
for f (assuming f C1 is more than enough). The details of this
verification are in many books.
3. Mean value property
The mean value property for a harmonic function u states that u(x) is
the average of its values on any sphere centered at x
Z
1
(5.5)
u(x) =
u(y) dAy .
Area (B(x, r)) B(x,r)
3. MEAN VALUE PROPERTY
61
62
To prove this, let d be the element of area on the unit sphere; then
on B(x0 , r) we have dA = r n1d ,
somewhere inside and thus attains its maximum or minimum at an

interior point. This contradicts the maximum principle.
(5.6)
One can use an energy approach to give an alternative proof. By the

divergence theorem, if u is harmonic in and zero on the boundary,
then
Z
Z
uu dx = | u |2 dx
0=
0=
Z
u
u(x + r)
dA = r n1
d
r
B(x,r) N
| |=1

Z
d
u(x + r) d .
= r n1
dr
| |=1
u(y) dy =
B(x,r)
(5.7)
Thus the last integral on the right is independent of r . Letting r 0

we obtain
Z
u(x + r) d = Area (B(0, 1))u(x),
| |=1
which is (5.5).
The solid mean value property is
Z
1
u(x) =
u(y) dy.
Vol (B(x, r)) B(x,r)
It follows from the mean value property for spheres by simply multiplying both sides by Area (B(x, r)) and integrating with respect to
r.
The maximum principle is an easy consequence. It asserts that if u
is harmonic in a connected bounded open set and continuous in ,

then
max u(x) max u(x).
Moreover, if u attains its maximum at an interior point, then u

constant in .
and hence attains
Since u is a continuous function on the compact set
its maximum somewhere, we need only prove the second assertion. Say
u attains its maximum at an interior point x0 and say u(x0 ) = M .
Let Q = {x |u(x) = M}. Since u is continuous then Q is closed.
By the mean value property, Q is open. Since Q is not empty and
is connected, it follows that Q = .
There is an obvious minimum principle which follows by replacing u
by u.
Uniqueness for the Dirichlet problem in a bounded connected region
u = f in ,
with u = on
is easy. We need only prove that if u = 0 in and u = 0 on , then

u 0. But if u is not identically zero, it is either positive or negative
so u constant . But since u = 0 on the boundary, u 0.
Exercises:
1. Show that this second proof also works with Neumann boundary
conditions u/N = 0, except that with these boundary conditions we can only conclude that u constant. Indeed, if u is any
solution, then so is u + const.
2. If in a bounded domain say u = 0 with u = f on the boundary
while v = 0 with v = g on the boundary. If f < g what can you
conclude? Proof?
3. If u satisfies u 0, show that the average of u on any sphere
is at least its value at the center of the sphere. Use this to conclude
that if u 0 on the boundary of a bounded domain , then u 0
throughout .
4. In a domain Rn let u(x) be a solution of u + a(x)u = 0,
where a(x) > 0.
a) Show there is no point where u has a positive local maxima (or
negative minima).
b) In a bounded domain, show that there is at most one solution
of the Dirichlet problem
+ a(x)u = F (x) in with u = on .
[Give two different proofs, one using part a), the other using
Energy.]
5. In a domain Rn let the vector u(x) be a solution of the system
of equations u + A(x)u = 0 with u = 0 on the boundary.
Here A(x) is a symmetric matrix and u means apply to each
component of u; a useful special case is the system of ordinary
differential equations u + A(x)u = 0. Assume A(x) is a positive
definite matrix, show that u 0. Also, give an example showing
that if one drops the assumption that A(x) is positive definite, then
there may be non-trivial solutions. Suggestion: As just above,
there are two distinct approaches, both useful:
i). Use energy methods directly.
3. MEAN VALUE PROPERTY
63
ii). Let (x) = | u(x) |2 and apply the scalar maximum principle to .
2
6. On the torus, T , let u be a solution of

u = 1 h(x)eu ,
where h C(T2 ), so in particular, h is a periodic function of its

variables. If 0 < a h(x) b, find upper and lower bounds for u
in terms of a and b.
7. Let (aij (x)) be a positive definite nn matrix for x Rn . Assume
u(x) C 2 satisfies
n
X
i,j=1
aij
2u
+ c(x)u = 0,
xi xj
where c(x) > 0.

a) Show that u cannot have a local positive maximum. Also show
that u cannot have a local negative minimum.
b) If a function u satisfies the above equation in a bounded region
D Rn and is zero on the boundary of the region, show that
u(x) is zero throughout the region.
Using only the solid mean value property one can prove a weak Harnack
inequality. Assume the harmonic function u 0 in the ball | x x0 |
R. Then for any point x1 in this ball
(5.8)
0 u(x1 )
Rn
u(x0 ).
(R | x1 |)n
To prove this, since the ball B(x1 , R| x1 |) B(x0 , R) we use the solid
mean value property in B(x1 , R | x1 |), the assumption that u 0,
and the mean value property a second time to find
Z
1
u(x) dx
u(x1 ) =
Vol (B(x1 , R | x1 |) B(x1 , R| x1 |)
Z
1
u(x) dx
Vol (B(x1 , R | x1 |) B(x0 ,R)
=
Vol (B(x0 , R))

Rn
u(x0 ) =
u(x0 ).
Vol (B(x1 , R | x1 |)
(R | x1 |)n
Inequality (5.8) quickly implies a Liouville theorem: If u is harmonic

on all of Rn and u 0, then u(x) const. Indeed, by letting R
in (5.8) we find that u(x1 ) u(x0 ) for any two points x0 , x1 .
Interchanging the roles of these points we see that u(x0 ) u(x1 ).
Consequently u(x1 ) = u(x0 ).
64
4. Poisson formula for a ball

There are very few domains for which one has an explicit formula
for the solution u(x)of the Dirichlet problem
(5.9)
u = 0 in with u = f on .
A valuable special case is if is a ball B(0, R) Rn . Then the solution

is given by the Poisson formula
Z
R 2 | x |2
f (y)
(5.10)
u(x) =
dA(y).
nn R
|
x
y |n
B(0,R)
The function
P (x, y) =
is called the Poisson kernel.
R 2 | x |2
nn R| x y |n
There are several ways to derive (5.10). If n = 2 one can use separation
of variables in polar coordinates. Other techniques are needed in higher
dimensions. The details are carried out in all standard texts.
The mean value property is the special case of (5.10) where x = 0.
One easy, yet important, consequence of the Poisson formula is that if
a function u is harmonic inside a domain , then it is smooth (C )
there. To prove this near a point x consider a small ball B(x0 , R)
containing x and use (5.10) to obtain a formula for u in terms of its
values on the boundary of the ball:
Z
R 2 | x |2
u(y)
u(x) =
dA(y).
nn R
|
x
y |n
B(x0 ,R)
Since y is on the boundary of the ball and x is an interior point,

one can repeatedly differentiate under the integral sign as often as one
wishes. By a more careful examination, one can even see that u(x) is
real analytic, that is, locally it has a convergent power series expansion.
Another consequence of the Poisson formula is a removable singularity

assertion. Say u(x) is harmonic in the punctured disk 0 < | xx0 | R
and bounded in the disk | x x0 | R. Then u and be extended
uniquely to an harmonic function in the whole disk | x x0 | R.
The proof goes as follows. Without loss of generality we may assume
that x0 = 0. Using the Poisson formula, in {| x | R} we can find a
harmonic function v with v(x) = u(x) for | x | = R. Given any > 0
let
w(x) = u(x) v(x) [(x) (R)],
where (x) is the fundamental solution (5.3) of the Laplacian. Clearly
w(x) = 0 on | x | = R while, since u(x) is bounded, then w(x) < 0
on | x | = for > 0 sufficiently small. Consequently w(x) 0 in
5. EXISTENCE AND REGULARITY FOR u + u = f ON Tn
65
66
< | x | R; equivalently, u(x) v(x) + [(x) (R)]. Since is

arbitrary, u(x) v(x) in the annular region < | x | R.
Similarly, by considering
a). u + u = f on Tn . If f has a Fourier series (2.34) and we

seek a solution u of
w(x) = u(x) v(x) + [(x) (R)],
having a Fourier series (2.35), then from (2.36), matching the coefficients we find that
fk
uk =
1 + | k |2
so
X
fk
(5.12)
u(x) =
eikx .
2
1
+
|
k
|
k
we deduce that u(x) v(x) in this same annular region. Consequently,

u(x) = v(x) in this region. Since we can make arbitrarily small and
since v(x) is continuous at the origin, if we define u(0) = v(0), the
function u(x) is harmonic throughout the disk | x | R.
Remark: In this proof, we could even have allowed u(x) to blow-up
near the origin, as long as it blows-up slower than the fundamental
solution (x). To state it we use little o notation:
g(s) = o(h(s)) as s s0
means
lim g(s)/h(s) = 0.
ss0
For example x2 = o(x) as x 0. In this notation, the precise assumption needed on u(x) for a removable singularity is
u(x) = o((x)) as x 0.
Exercises:
1. Use separation of variables in polar coordinates to obtain the Poisson formula for the unit disk in R2 .
2. Use separation of variables in polar coordinates to solve the Dirichlet problem for the annulus 0 < a2 < x2 + y 2 < 1 in R2 .
3. Let uk be a sequence of harmonic functions that converge uniformly
to some function u(x) in a domain . Show that u is also harmonic.
4. [Harnack inequality] Let u(x) 0 be harmonic in the ball
B(0, R). Use the Poisson formula to show that
Rn2
R |x|
R + |x|
u(0) u(x) Rn2
u(0).
(R + | x |)n1
(R | x |)n1
5. Existence and regularity for u + u = f on Tn
We will use Fourier series to solve u + u = f on the torus T n . [See

the last section of Chapter 2 for basics on Fourier series.] This equation
is a bit simpler than u = f since the homogeneous equation u = 0
has the non-trivial solution u =const. Despite that the solution of
u + u = g will be an infinite series, the results and insight gained
are valuable.
(5.11)
u + u = f
on Tn
Moreover, if f H s (Tn ) then

X
X
| fk |2
(1+| k |2 )s | fk |2 = kf k2H s (Tn ) .
(1+| k |2 )s+2
=
kk2H s+2 (Tn ) =
2
2
(1 + | k | )
k
We summarize this.
Theorem 5.1. Given any f H s (Tn ), there is a unique solution u

of u + u = f . Moreover, u is in H s+2(Tn ), that is, it has two
more derivatives than f in L2 . If s > j + n/2, then, by the Sobolev
embedding theorem 2.9, u C j (T2 ).
So far we only considered the case where f H s (Tn ). This is a global
assumption on the smoothness of f . What can one say if f happens
to be smoother only near a point x0 ? We suspect that the smoothness
of u near x0 will depend only on the smoothness of f near x0 . This
is easy.
Say we know that F is smoother in the ball B(x0 , R). Pick an r < R
and a non-negative smooth function (x) so that (x) > 0 in the ball
B(x0 , R) with
(
1 for x in B(x0 , r),
.
(x) =
0 for x outside B(x0 , R).
Extend f (x) to Tn by f (x) = 0 outside B(x0 , R). Our vague smoothness assumption on f near x0 is now made precise by formally assuming
that (f ) H s (Tn ) for all r < R (one can use this to define the spaces
s
Hloc
of functions that are locally in H s ).
If u H 2 (Tn ) satisfies u + u = f , consider v(x) := (x)u(x) and
use the obvious extensions to Tn . Then,
where
(5.13)
v + v = (u + 2 u + u) + u = G,
G = f 2 u u.
6. HARMONIC POLYNOMIALS AND SPHERICAL HARMONICS
67
Say f H s (Tn ) for some s 1. Since u H 2 (Tn ), then G is in

H 1 (Tn ). Thus by Theorem 5.1, v H 3 (Tn ). If s 2 we can repeat
this to find that G H 2 (Tn ) so v H 3 (Tn ). Continuing, we conclude
that f H s (Tn ) implies v H s+2(Tn ). But u = v in B(x0 , r) so u is
in H s in this ball. This proves
2
Corollary 5.2. [Local Regularity] If u H (T ) satisfies u+

u = f and f H s in a neighborhood of x0 , then u H s+2 in this
neighborhood.
Exercise: If f H s (Tn ), discuss the existence, uniqueness, and regularity of solutions to u = f on Tn .
Exercise: Use Duhamels Principle to find a simple formula for the
solution of u + u = f (x) for 0 < x < , with u(0) = u() = 0.
Compare this with the solution obtained using Fourier series.
6. Harmonic polynomials and spherical harmonics
Consider the linear space P of polynomials of degree at most in
the n variables x1 , . . . , xn and let P be the sub-space of polynomials
homogeneous of degree . A polynomial u(x) is called a harmonic
polynomial if u = 0. We wish to compute the dimension of the
subspace H of P consisting of harmonic polynomials, homogeneous
of degree . If n = 2, and 1 the dimension is 2, since for 1
one basis for the space of harmonic polynomials of degree exactly is
the real and imaginary parts of the analytic function (x + iy) .
For the general case, observe that : P+2 P and define the linear
map L : P P by the formula

(5.14)
Lp(x) := (| x |2 1)p(x) ,
where | x | is the Euclidean norm. Now Lp = 0 means the polynomial

u(x) := (| x |2 1)p(x) P+2 is harmonic. But clearly u(x) = 0 on
the sphere | x | = 1, so u 0. Thus ker L = 0 so L is invertible.
In particular, given
a homogeneous q P there is a p P with
(| x |2 1)p(x) = q . Let v P denote the homogeneous part of
p that has highest degree . Since reduces the degree by two, we
deduce that in fact (| x |2 v) = q . Therefore this map v 7 q from
P P is onto and hence an isomorphism.1 Here are two consequences.
1) Since the map : P P2 is onto, again by linear algebra,
we can compute the dimension of the space of homogeneous harmonic
1One can also give a purely algebraic proof that if p P satisfies (| x |2 p) = 0,
then p 0 hence the map M : P 7 P defined by M p := (| x |2 p) is an

isomorphism of P .
68
polynomials:
dim H = dim P dim P2 =

(n + 2 2)(n + 3)!
n+1
n+3
=
.
!(n 2)!
For instance if n = 3 then dim H = 2 + 1.

2) Any homogeneous polynomial q P can be written (uniquely) in
the form q = h + | x |2 v, where h H and v P2 . To prove this,
first compute q and then use the above to find a unique v P2 so
that (| x |2 v) = q P2 . The function h := q | x |2v is clearly
harmonic. Applying this again to v and so on recursively we conclude
that q = h + | x |2 h2 + | x |4h4 + , where hj Hj . This yields
the direct sum decomposition P = H | x |2H2 . Since both
the Laplacian and the operation of multiplying by | x |2 commute with
rotations, the summands in this decomposition are SO(n)-invariant, a
fact that is useful in discussing spherical harmonics and the symmetry
group SO(n).
The idea behind the definition of L in (5.14) was that to solve u =
q P , we seek u in the special form u = (| x |2 1)p(x) to obtain a
new problem, Lp = q, whose solution is unique. Frequently it is easier
to solve a problem if you restrict the form of the solution to obtain
uniqueness.
Homogeneous harmonic polynomials arise since, when restricted to the

unit sphere these are exactly the eigenfunctions of the Laplacian on the
sphere. These are called the spherical harmonics. The dimensions of
the eigenspaces are then the numbers just computed. for instance, when
n = 3 this number is 2 + 1. We carry our part of this computation.
In spherical coordinates on Rn , the Laplacian is
2 u n 1 u
1
+ 2 S n1 u,
+
2
r
r r r
where S n1 is the Laplacian on the standard sphere S n1 . If p(x) is
a polynomial, homogeneous of degree k , then p(x) = r k v(), where
is a point on the unit sphere. Thus, if p is also harmonic, then using
(5.15)
Rn u =
(5.15)
0 = Rn r k v() = r k2[(k(k 1) + (n 1)k)v + S n1 ]v.

Consequently,
S n1 v = k(n + k 2)v.
In other words, v is an eigenfunction of the Laplacian on S n1 and

the corresponding eigenvalue is k(n + k 2). The missing piece is to
show that every eigenfunction of the Laplacian has this form. This can
be done, for instance, by proving that the dimension of the eigenspace
is the same as the dimension of the space of homogeneous harmonic
polynomials.
7. DIRICHLETS PRINCIPLE AND EXISTENCE OF A SOLUTION
69
Application. Atoms are roughly spherically symmetric. The maximum number of electrons in the k th atomic subshell is related to the
dimension of the eigenspace corresponding to the k th eigenvalue. The
Pauli exclusion principle asserts that no two electrons can be in the
same state. But electrons can have spins 1/2, There are 2k + 1 electrons with spin 12 , so 2(2k + 1) in all. Thus the subshells contain at
most 2, 6, 10, 14, . . . electrons.
7. Dirichlets principle and existence of a solution
a). History. To solve the Dirichlet problem (5.9), Dirichlet proposed to find the function u that minimizes the Dirichlet integral
Z
(5.16)
J() := | |2 dx
among all functions , say piecewise smooth, with = f on .

To see this, let h be any function that is piecewise smooth in and
zero on the boundary. If u minimizes J(), then for any t the function
= u+th has the correct boundary values so J(u+th) has a minimum
at t = 0. Taking the first derivative gives
Z
dJ(u + th)
(5.17)
0=
= 2 u h dx.
dt
t=0
If u has two continuous derivatives, we can now integrate by parts and

use that h = 0 on the boundary to find
Z
0 = (u) h dx.
Since h can be any piecewise smooth function that is zero on the boundary, this implies that u = 0, as desired. [Proof: if not, say u > 0
somewhere, then u > 0 on a small ball. Pick a function h that
is
R positive on this ball and zero elsewhere, giving the contradiction
u h dx > 0.]
Riemann adopted this reasoning in his proof of what we now call the
Riemann mapping theorem. Weierstrass pointed out that although
J(u) is bounded below and hence has an infimum, it is not evident that
that there is some function u satisfying the boundary conditions for
which J(u) has that minimal value. To make his argument convincing,
he gave the example
Z 1
x2 (x)2 dx
with (1) = 1.
J() =
1
For this consider the sequence
1 for 1 x 1/k,
k (x) = kx for 1/k x 1/k,
1 for 1/k x 1.
70
2
Then J(k ) = 3k
0, so inf J() = 0. But if J(u) = 0, then
u = const and cant satisfy the boundary conditions.
Since Riemanns application of Dirichlets principle was important,
many people worked on understanding the issues. Using other methods
Poincare gave a rather general proof that one could solve the Dirichlet
problem (5.9) while around 1900 Hilbert showed that under reasonable
conditions, Dirichlets principle is indeed valid.
,
b). A modified problem. In subsequent years the tools developed to understand the issues have led to a considerable simplification.
First, instead of solving (5.9) solve the related inhomogeneous equation
(5.18)
u = F in
with
u = 0 on .
To reduce (5.9) to this form, let fe (x) be a smooth extension of f from

to all of . We assume this can be done since if there is a solution
of (5.9), then the solution itself gives a very special extension. Then
let w := u fe . This satisfies w = fe , which has the form (5.18)
with F = fe .
For (5.18) the analogue of (5.16) is the functional
Z

| |2 2F dx.
Q() :=
Imitating the procedure Dirichlet followed, we seek to minimize Q from

an appropriate class of functions that vanish on the boundary. If u
minimizes Q, then, Q(u + th) has its minimum at t = 0. Computing
dQ/dt|t=0 gives
Z
(u h F h) dx = 0
(5.19)
for all h that vanish on the boundary. As before, assuming this function u has two continuous derivatives, an integration by parts shows
that u = F , as desired. It is not difficult to show that Q is bounded
below, but even knowing this we still dont know that Q achieves its
minimum. Instead of perusing this, we take a slightly different approach.
For a bounded open set , use the space Cc1 () of functions with
compact support in (the support of a function is the closure of the
set where the function is not zero). Thus, the functions in Cc1 () are
zero near the boundary of . For Cc1 () define the norm
Z
kk2H 1 () = | |2 dx.
0
Because of the Poincare inequality (3.34), this is a norm, not a seminorm. Define the Sobolev space H01 () as the completion of Cc1 () in
71
this norm. This is a Hilbert space with inner product

Z
dx.
h, iH01 () =
72
dimensions. In an orthonormal basis e1 ,. . . en , say x = x1 e1 + xn en .

Then by linearity
(x) = x1 (e1 ) + + xn (en ).
Motivated by (5.19) If F L2 (), we say u H01 () is a weak solution

of (5.18) if
(5.20)
Z
Z
Z
f v dx
F v dx,
that is
hu, viH01 () =
u v dx =
H01 ().
If u C () satisfies (5.18), then it is clearly a

for all v
weak solution: just integrate by parts. Conversely, if u C2 () is a
weak solution, then after an integration by parts,
Z
[u F ]v dx = 0
for all v H01 (),
so, arguing as above, u = F . By a separate argument that we do

not give, u = 0 on .
Note that a weak solution, if one exists, is unique, since if there were
two, u and w let = u w H01 (). Then
Z
w v dx = 0
for all v H01 ().
R
Letting v = w we conclude that | w |2 =0 so, using the Poincare
inequality (3.34), w = 0. Consequently, if we have a weak solution and
if we believe there is a classical solution, then the only possibility is
that the weak solution is also the desired classical solution.
Our strategy is to break the proof of the existence of a solution into
two parts:
Existence: Prove there is a weak solution.
Regularity: Prove that this weak solution is a classical solution
if f is smooth enough.
Consequently, if we let z = (x1 )e1 + + (en )en , then (x) = hx, zi.
Geometrically one can interpret z as a vector orthogonal to the kernel
of .
Essentially the same proof works in any separable Hilbert space. Pick
a (countable) orthonormal basis and write x H in this basis. Then,
as above, we are led to let
z = (x1 )e1 + + (en )en + .
However, it is not yet evident that this series converges in H. Thus,

for any N let
zN = (e1 )e1 + + (eN )eN .
Then,
| (zN ) | = (e1 )(e1 ) + + (eN )(eN ) = kzN k2 .
But by hypothesis, | (x) | ckxk, so kzN k2 ckzN k. Thus kzN k
c. Because this estimate is independent of N , the series defining zN
converges in H to an element of H and we have (x) = hx, zi for all
x H.
To prove the existence of a weak solution, motivated by (5.20), for any
v H01 (), define the linear functional
Z
F v dx.
(v) :=
Then by the Schwarz and Poincare inequalities
| (v) | = kF kL2 () kvkL2 () ckvkL2() ckvkH01 () ,
By the Riesz representation theorem, there is a u H01 () such that

Z
F v dx,
(v) = hu, viH01 () , that is, hu, viH01 () =
c). Existence of a weak solution. The key ingredient in the

following proof of the existence of a weak solution is a standard result
in elementary functional analysis: the Riesz representation theorem for
a Hilbert space H. To state it, recall that a bounded linear functional
(x) is a linear map from elements x H to the complex numbers
with the property that | (x) | ckxk, where the real number c is
independent of x. A simple example is (x) = hx, zi for some z
H. The Riesz representation theorem states that every bounded linear
functional has this form.
For those new to this result, here is a primitive proof (using coordinates)
that works for separable Hilbert spaces. As a warm-up, first in finite
just as desired.
Note that this proof works for any bounded open set , no matter
how wild its boundary. For instance, if is the punctured sphere
0 < kxk < 1 in R3 and try to solve u = 4 there with u = 0 on the
boundary, the unique solution in polar coordinates is u = (1 | r |2 )
which does not satisfy the boundary condition we attempted to impose,
u(0) = 0. That jump discontinuity is a removable singularity. The
existence theorem is smart enough to ignore bad points we may have
on the boundary of .
Exercises:
73
1. In the one dimensional case, so R1 is a bounded interval, if

and u = 0 on .
u H01 (), show that u C()
satisfy
2. Let R1 be a bounded open interval, a(x) C 1 (
0 < a(x) and c(x) 0 a bounded continuous function.
Consider solving
Lu := (a(x)u ) + c(x)u = f L2 ()
Define u
H01 ()
Z
with u = 0 for x .
to be a weak solution of Lu = f if
Z
f v dx
[a(x)u v + c(x)uv] dx =
for all v H01 (). Prove that there exists exactly one weak soDefine and use a Hilbert space that uses
Rlution. [Suggestion:

[a(x)
+
c]
dx
as
its inner product. Show that the norm on
this space is equivalent to the H01 () norm.]

and F L2 (), prove
3. If c(x) > 0 is a continuous function in
1
there is a unique weak solution u H0 () of u + c(x)u = F .
[The first step is to define a weak solution].
4. Let A := (aij (x)) be a positive definite nn matrix of continuously
where Rn is a bounded open
differentiable functions for x ,
set. In particular, there are constants m, M so that for any vector
v Rn we have mkvk2 hv, Avi Mkvk2 . Consider

n
X
u
Lu :=
aij (x)
+ c(x)u = F (x)
xi
xj
i,j=1
where 0 c(x) is continuous and F L2 (). Show there is
a unique weak solution u H01 () of Lu = F L2 (). [The first
step is to define a weak solution].
d). Regularity of the weak solution. If needed, dilate our bounded

domain , so it is inside the box | xj | , j = 1, . . . , n, which
we view as the torus Tn . Let u H01 () be our weak solution of
u = F L2 (Tn ). Rewrite this as u + u = f (x), where now
f (x) = F (x) + u(x) L2 (Tn ) is considered to be a known function. By
Theorem 5.1, u H 2 (T2 ).
Although F might be smoother in , our extension of F to Tn likely
looses this additional smoothness across . However, the local regularity Corollary 5.2 implies that if F is in H s near a point x0 , then u
is in H s+2 near x0 .
74
CHAPTER 6
The Rest
In the last part of the course I outlined the several topics, mainly following various parts from my old notes
Lecture Notes on Applications of Partial Differential Equations to Some
Problems in Differential Geometry, available at
http://www.math.upenn.edu/kazdan/japan/japan.pdf
In addition, there is a bit of overlap with my expository article Solving
Equations available at
http://www.math.upenn.edu/kazdan/solving/solvingL11pt.pdf
Topics
Defined both the Holder spaces C k+ , 0 1 and Sobolev
spaces H p,k and illustrated how to use them in various regularity assertions for solutions of some linear and nonlinear elliptic
partial differential equations.
Defined ellipticity for nonlinear equations, giving several examples including a Monge-Amp`ere equation.
Discussed issues concerning qualitative properties and existence for the minimal surface equation, equations of prescribed
mean and Gauss curvature (for surfaces) and some equations
for steady inviscid fluid flow.
Discussed techniques for proving that a partial differential equation has a solution. The techniques included:
a) iteration using contracting mappings,
b) direct methods in the calculus of variations,
c) continuity method
d) fixed point theorems (Schauder and Leray),
e) heat equation (R. Hamilton).
75

AMSI 2up

Uploaded by

Copyright:

Available Formats

AMSI 2up

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

AMSI 2up

Uploaded by

Copyright:

Available Formats

AMSI

Jan. 14 Feb. 8, 2008

Partial Differential Equations

c 2008 by Jerry L. Kazdan

[Last revised: March 28, 2011]

7. Dirichlets principle and existence of a solution

Chapter 2. First Order Linear Equations

Chapter 3. The Wave Equation

Chapter 4. The Heat Equation

Kazdan, Jerry, Lecture Notes on Applications of Partial Differential

1. Functions of Several Variables

2. CLASSICAL PARTIAL DIFFERENTIAL EQUATIONS

2. Classical Partial Differential Equations

utt = uxx + uyy

uxx + uyy = f (x, y)

The homogeneous Laplace equation, uxx + uyy = 0, can be thought

and one seeks the (equilibrium) temperature distribution u(x, y) for

Of course, one must give a genuine mathematical proof as a check that

3. ORDINARY DIFFERENTIAL EQUATIONS, A REVIEW

3. Ordinary Differential Equations, a Review

b). First Order Linear Inhomogeneous Equations. These have

but this really is an illustration of a beautiful simple, important, and

if we pick p so that p + ap = 0 then solving Lu = f becomes pv = f

We say that g is orthogonal to h if hg, hi = 0. Define the operator L

and L(au) = aLu

for any constant a.

for any function u. There is an obvious group theoretic property:

By linearity w + w = 0 so E (t) = w (w + w) = 0. This proves that

3. ORDINARY DIFFERENTIAL EQUATIONS, A REVIEW

0 is a solution of L(t) = 0. By translation invariance, for any constant

Consequently, by the uniqueness lemma, we deduce the standard trigonometry formula

for some constant k .

with the initial conditions u(0) = 1, and u(1) = 1.

3. ORDINARY DIFFERENTIAL EQUATIONS, A REVIEW

where h is also periodic and satisfies 0 < a h(x) b. Find

First Order Linear Equations

is very special since everything reduces to solving ordinary differential

The differential equation asserts that u(x, y) is constant on the vertical

attention to vertically convex domains , that is,

uy (x, y) = f (x, y),

on the line y = 0, then in a convex there will be a unique solution of

u(x(t), y(t)) = h(t).

In this case, using (2.3) one finds that

2. FIRST ORDER LINEAR EQUATIONS

a). The homogeneous equation. If R2 is a disk, the most

so if is vertical for some value t0 , then h (t0 ) = 0.

for any function depending only on x.

Moral: if is tangent to a characteristic curve at some point, then

3. A MORE GENERAL EXAMPLE

b). The inhomogeneous equation. One can readily extend this

2. FIRST ORDER LINEAR EQUATIONS

alternate method Another approach to (2.9) is to change to new

Exercise: Solve uy = 1 2xy with u(x, 0) = 0.

sin x = u(x, 0) = (2x) so (x) = sin(x/2).

Consequently, just as above,

0 = ux + 2uy = (a + 2b)ur + (c + 2d)us .

u(x, y) = sin(x 12 y).

can be interpreted as the directional derivative in the direction of the

This solution f (x ct) represents a density wave traveling to the