Chapter 1 Introduction

1.1 Definitions

The simplest differential equation is perhaps the Malthusian equation modelling population growth:

du
= ku, (1.1.1)
dt
where k is a constant, u is the unknown function (representing population size) and t is the independent
variable. A differential equation involving only one independent variable is called ordinary differential
equation(ODE). So (1.1.1) is such. An example of a different type of differential equations is the heat
equation
∂u ∂2u ∂2u ∂2u
= k( 2 + 2 + 2 ) (1.1.2)
∂t ∂x ∂y ∂z
where k is the heat conductivity constant, u is the unknown function (representing temperature) and x,
y, z and t are the independent variables. This is an example of a partial differential equation.

Definition 1.1.1 (Partial differential equations)

An equation involving partial derivatives of an unknown function is called partial differential equation
(PDE). The general form of PDEs is

∂u ∂u ∂u ∂ 2 u ∂ 2 u ∂2u
F (x, u, , ,···, , 2 , 2 , · · · , 2 , · · ·) = 0, (1.1.3)
∂x1 ∂x2 ∂xn ∂x1 ∂x2 ∂xn
or in the operator form:
L[u](x) = f (x), (1.1.4)
where x = (x1 , x2 , · · · , xn ) are the independent variables, u = u(x1 , x2 , · · · , xn ) the unknown function,
f = f (x) = f (x1 , x2 , · · · , xn ) the given function. It is agreed that L[u] includes all the terms in F that
involve u.

Definition 1.1.2 (order, linearity, homogeneity, superposition principle)

The highest order of the partial derivative of the unknown function u(x) that appears in a PDE is called
the order of the PDE. The PDE is said to be linear, if the differential operator L satisfies the following
condition:

L[α1 u1 + α2 u2 ] = α1 L[u1 ] + α2 L[u2 ], (1.1.5)

where α1 , α2 are arbitrary constants. If the PDE (1.1.4) is linear and if f ≡ 0, then the PDE is said to be
homogeneous. For nonlinear PDEs, we do not talk about their homogeneity (because doing so yields no
benefit).
The heat equation (1.1.2) is second order, linear and homogeneous. In this course, we will focus on linear
PDES and, unfortunately, will not spend enough time on nonlinear ones.
For a linear homogeneous PDE, u(x) ≡ 0 is always a solution, which is called the trivial solution.
Suppose that u1 and u2 are two solutions of the homogeneous linear PDE, i.e.

L[u1 ](x) = 0, L[u2 ](x) = 0.

It follows from (1.1.5) that

L[α1 u1 + α2 u2 ] = 0,
2

i.e. the linear combination of two (hence any number of ) solutions is also a solution of the PDE. This is
called the superposition principle for linear PDEs, and does not hold for non-linear ones.

More examples

ut + V ux = 0 (transport equation) (1st order, linear, homogeneous),

utt − c2 uxx = A sin t (wave equation) (2nd order, linear, inhomogeneous),
uxx + uyy = 0 (Laplace equation) (2nd order, linear, homogeneous),
ut + u ux = kuxx (dif f usion − convection equation) (2nd order, nonlinear),

∂u
where ux denotes the partial derivative
, etc.
∂x
General solutions, initial and boundary conditions
For ODEs, we talked about their general solutions; for example, the general solution, i.e. the set of all
solutions of (1.1.1) is
u(t) = const. Cekt (1.1.6).

For PDEs, we can do the same. For example, consider the transport equation

ut + V ux = 0, (1.1.7)

where u represents the density of a substance (e.g. a pollutant) in a one dimensional medium (e.g. a river)
that moves with constant velocity V . Let

u(x, t) = φ(x − V t), (1.1.8)

where φ is an arbitrary smooth function. We check that this is always a solution of (1.1.7):

ut + V ux = −V φ0 (x − V t) + V φ0 (x − V t) = 0.

On the other hand, we will show in the next chapter that given any solution u of (1.1.7), there exists a
function φ such that (1.1.7) holds. Thus the general solution of (1.1.7) is given by (1.1.8). Observe that
in the ODE general solution (1.1.6) there is an arbitrary constant C, while in the PDE general solution
(1.1.8) there is an arbitrary function. This is typical for PDEs.
As in the case of ODEs, to pick out a unique solution from a family of solutions, we need initial and
boundary conditions that are mostly physically motivated/demanded. For (1.1.7), if the initial condition

1
u(x, 0) = , x ∈ (−∞, ∞) (1.1.9)
1 + x2
is given (physically the initial density is given), then

1
φ(x) = , x ∈ (−∞, ∞) (1.1.10)
1 + x2
and so u can be uniquely determined
1
u(x, t) = , (x, t) ∈ (−∞, ∞) × [0, ∞).
1 + (x − V t)2

If the “river”(medium) runs from x = 0 to x = ∞, then to determine a unique solution of (1.1.7) we need
not only the initial condition (1.1.9) for x ∈ [0, ∞), but also a boundary condition at x = 0 (the origin
of the river) such as

u(0, t) = 1 + sin t, t ≥ 0 (so the pollutant is discharged at x = 0 periodically) : (1.1.11)

 3

by (1.1.9) with x ∈ [0, ∞), we have (1.1.10) for x ∈ [0, ∞) and so

1
u(x, t) = , if x − V t ≥ 0;
1 + (x − V t)2
by the boundary condition,
φ(−V t) = 1 + sin t, t ≥ 0.
Labelling −V t (≤ 0) by z, we have
z
φ(z) = 1 − sin( ), if z ≤ 0,
V
hence
x−Vt
u(x, t) = 1 − sin( ), if x − V t ≤ 0.
V
Note u(x, t) is defined in two pieces, one for x − V t ≥ 0, the other for x − V t ≤ 0.
This is wonderful, but, unfortunately, rare: for most PDEs, we cannot find formulas for their general
solutions; moreover, even if we can, often times it is still hard to pick out the particular solution that
satisfies the initial and boundary conditions. Thus, the methodology of starting with general solutions (
as in ODEs) is often abandoned when solving PDEs.

1.2 Classification of second order PDEs

Let us recall the Laplace equation

uxx + uyy = f. (1.2.1)
If we replace uxx by ξ 2 , uyy by η 2 , and f by 1, then we get

ξ 2 + η 2 = 1, (1.2.2)

which is the equation of an ellipse. ((1.2.2) can be obtained by taking Fourier transform of both sides
with f being a δ−function, and then dropping the Fourier transform of u.) In this connection, we say
that the Laplace equation (1.2.1) is elliptic. In the same fashion, the wave equation

utt − uxx = f

corresponds to the equation of a hyperbola

ξ2 − η2 = 1

hence we say the wave equation is hyperbolic. For the heat equation

ut − uxx = f,

if we take Laplace transform in the t-variable and then Fourier transform in the x-variable, we then obtain
the equation of a parabola
ξ = −η 2 .
Thus the heat equation is said to be parabolic.
It turns out that all 2nd-order linear PDEs

a11 ux1 x1 + 2a12 ux1 x2 + a22 ux2 x2 + b1 ux1 + b2 ux2 + c u = f, (1.2.3)

can be classified into three types: elliptic, parabolic and hyperbolic. Here the coefficients a11 , a12 , a22 ,
b1 , b2 , c and the righthand side f are functions of the variables x1 and x2 in general. (Note ux1 x2 = ux2 x1
for smooth u and so there is no need to include a ux2 x1 -term in (1.2.3).) Before discussing how to do the
4

classification, we mention the significance of classifying PDEs: it can be proved by using advanced PDE
theory that solutions of an elliptic equation are as smooth as allowed by the coefficients and the righthand
side of the equation; solutions of a parabolic equation are smoothed right after the initial time, while the
smoothness of solutions of a hyperbolic equation is neither improved nor destroyed after the initial time.
Other differences are: second order elliptic and parabolic equations satisfy the maximum principle; the
speed of propagation of disturbance is finite for hyperbolic equations, infinite for parabolic equations.
Now let us come back to the question of classifying (1.2.3). The type of this equation is a local property
so we freeze the coefficients at a fixed point on the x1 x2 −plane (this is related to the “freezing coefficient
method”in the regularity theory for elliptic and parabolic equations). So we now assume the coefficients in
(1.2.3) are constants. The type of this PDE does not depend on the lower order terms and the righthand
side. Thus we focus on the second order terms.
We determine the type of (1.2.3) by checking to see if there exists a linear change of variables
2
X
yi = bij xj , i = 1, 2,
j=1

so that in y1 , y2 variables, the highest order terms of (1.2.3) are the same as that of the Laplace, or the
wave or the heat equation. By chain rule, we have
2
X
uxk xl = uyi yj bik bjl .
i,j=1

Since ux1 x2 = ux2 x1 , we can let a21 = a12 and rewrite the leading terms of (1.2.3) as
 
X 2 2
X X2
uxk xl akl =  akl bik bjl  uyi yj . (1.2.4)
k,l=1 i,j=1 k,l=1

The coefficient of uyi yj on the righthand side is just the ij−element of matrix BAB T , where A = (aij )
and B = (bij ). Since A is symmetric, there exists an orthogonal matrix B such that BAB T is diagonal:

BAB T = diag(λ1 , λ2 )

where the λ’s are the eigenvalues of the matrix A. Then in y1 , y2 variables, the leading terms of (1.2.3)
are written as
λ1 uy1 y1 + λ2 uy2 y2 .

Case 1. det A > 0, i.e, a11 a22 > a212 . In this case the λ’s have the same sign (recall det A = λ1 λ2 ); without
√ √
loss of the generality, assume that both are positive. If we let z1 = y1 / λ1 and z2 = y2 / λ2 , then the
leading terms of (1.2.3) can be further reduced to

uz1 z1 + uz2 z2 .

Thus in Case 1, we say (1.2.3) is elliptic, the same type as the Laplace equation.
Case 2. det A < 0, i.e, a11 a22 < a212 . In this case the λ’s have the opposite signs. We say that (1.2.3) is
hyperbolic, the same type as the wave equation.
Case 3. det A = 0, i.e, a11 a22 = a212 . In this case one of the λ’s is 0, the other is not (if both are 0, then
A = 0 and (1.2.3) is not a second order PDE). We say that (1.2.3) is parabolic, the same type as the
heat equation.
If the coefficients of the leading order terms in (1.2.3) are not constant, then we define the type of (1.2.3)
at a point P on x1 x2 −plane, as above.
 5

Example 1.2.1 Consider

xuxx + yuyy − uxy + 2011ux + 210uy + u = x2 + y.

In what region on xy−plane is the PDE of elliptic, hyperbolic or parabolic type?

In this example, a11 = x, a22 = y and a12 = a21 = −1/2;
1
a11 a22 − a212 = xy − .
4
Thus on the hyperbola xy = 1/4, the PDE is parabolic; in the region xy < 1/4 (which is between the two
branches of the hyperbola) the PDE is hyperbolic; in the rest of the xy−plane, it is elliptic.

Assignment 1

1. For each of the PDEs below, find its order, linearity and homogeneity.
(1) ut + uux = 0 (Burger’s equation).
(2) xut − uxx + 2x + sin t = 0 (Degenerate heat equation).
(3) utt − (uxx + uyy + uzz ) = −u + u3 (Klein-Gordon equation).
(4) (1 + u2y )uxx − 2ux uy uxy + (1 + u2x )uyy = 0 (Minimal surface equation).

2. Classify the following equations as hyperbolic, parabolic or elliptic

(1) uxx + 4 uxy + 5 uyy + ux + 2 uy = 0,
(2) uxx − 4 uxy + 4 uyy + 3 ux + 4 u = 0,
(3) uxx + 2 uxy − 3 uyy + 2 ux + 6 uy = 0.

3. Consider the PDE with constant coefficients

a11 uxx + 2a12 uxy + a22 uyy + b1 ux + b2 uy + c u = 0.

Show that the only ones that are unchanged under all axis-rotations (rotation invariant) have the form

a · (uxx + uyy ) + b u = 0,

where a and b are constants. Hint: Use the discussion in the text, especially (1.2.4)

4. Classify the following equations as hyperbolic, parabolic or elliptic. If the type changes in the xy−plane,
find the region for each type.
(1) (1 + x2 ) uxx + (1 + y 2 ) uyy + x ux + y uy = 0
(2) uxx + (1 + y)2 uyy = 0
(3) e2x uxx + 2ex+y uxy + e2y uyy = 0
(4) uxx + y uyy = 0
(5) uxx + xy uyy = 0
(6) y uxx − x uyy + ux + y uy = 0
6

5. Show by direct substitution that u(x, t) = f (x + 2t) + g(x − 2t) is a solution of the PDE

utt − 4uxx = 0

for arbitrary smooth functions f and g.

6∗ . Given the partial differential equation

A uxx + B uxy + C uyy = 0, (∗)

where A, B and C are constants. Find the general solution of the above equation when
(1) equation (∗) is hyperbolic;
(2) equation (∗) is parabolic.

Chapter 2 First-order Partial Differential Equations

2.1 Transport equation: derivation

Perhaps the most important first order PDE is the transport equation, also called continuity equation.
It can be used to describe the evolution of distribution of mass, energy, electric charges and biological
population, etc. Thus this equation is fundamental to applied sciences.
Here is the general idea in the derivation of the transport equation. Let x represent location in Rn
(n = 1, 2, 3 are the most relevant cases) and let t represent time. Consider a substance that moves in
Rn ; this substance may be a pollutant in a river, population of a biological species, even thermal energy.
Let u(x, t) be the density function of the substance, measured in mass/volume (we emphasize that here
“mass”may mean energy); suppose the velocity vector of the particle of the substance which is at location
x at time t is V(x, t). In multi-variate Calculus we learned that given a surface S with unit normal vector
field n, the rate at which mass crosses the surface in the direction of normal n is given by the “flux integral”
Z
u(x, t)V · n(x, t)dS. (2.1.1)
S

Note that the mass may flow across one part of the surface in the direction of n, and across the other
part in the opposite direction. Thus (2.1.1) is really the net rate at which mass crosses the surface in
the direction of normal n. The substance may be created or degraded so we assume that the creation-
degradation rate is f (x, t). f (x, t) is calculated by taking a small neighborhood of point x with volume
∆V , then dividing the rate of change of mass in the neighborhood by ∆V and sending ∆V to zero. Thus
the unit of f is mass/time/volume.
Now take an arbitrary bounded region Ω in Rn with smooth boundary ∂Ω. By conservation of mass, the
net rate of change of mass in Ω = net rate at which mass crosses ∂Ω in the direction of the inner normal
+ total creation-degradation rate in Ω. Thus we have
Z Z Z
d
u(x, t)dx = u(x, t)V · (−n)(x, t)dS + f (x, t)dx, (2.1.2)
dt Ω ∂Ω Ω

where n is the unit outer normal vector field of ∂Ω. By Divergence Theorem, the surface integral in the
above equation is equal to Z
− ∇ · (uV)dx,
Ω
and so (2.1.2) can be rewritten as
Z
(ut + ∇ · (uV) − f )(x, t)dx = 0.
Ω
 7

If we assume the integrand is continuous, then it must be identically equal to 0: otherwise, there exists a
small ball B on which the integrand is positive or negative, so if we take Ω to be B then the integral is
non-zero; a contradiction!
Now we have derived the transport equation

ut + ∇ · (uV) = f. (2.1.3)

We introduce

J = uV
which is called flux. Then (2.1.3) is rewritten as

ut + ∇ · J = f. (2.1.4)

Traffic equation
This is a specific example of the transport equation. Of concern is the traffic problem on one straight lane:

       
       

- x
direction of traffic

Figure 2.1.1 One straight lane traffic

We use the continuum hypothesis to describe the traffic flow. Let V (x, t) be the velocity of traffic at
position x and time t ( unit : length/time ). The traffic density at time t is defined by

N o. of cars at time t in (x, x + ∆x)

u(x, t) = lim
∆x→0 ∆x
Even though the limit ∆x → 0 is not achievable in the practice, the above definition is commonly accepted
as a continuum model for the traffic flow. The traffic flux at position x is defined by

J(x, t) = u(x, t)V (x, t).

(In (2.1.1), we take the “surface”to be the point x and the normal n to be 1 (one dimensional vector), so
the flux integral is simply the flux J(x, t).) The flux J(x, t) is the rate at which cars cross point x (in the
positive direction) at time t. Now by the general transport equation (2.1.4), we have

∂u ∂J
+ =0
∂t ∂x

or,

∂u ∂
+ (u · V ) = 0,
∂t ∂x

which is called the conservation law or conservation equation for the traffic problem.

2.2 First-order linear PDEs: method of characteristics, general solutions and

break-down of smoothness
8

Example 2.2.1 Let us start with the simplest transport/traffic equation, already discussed in Section 1.1:

∂u ∂u

 +V =0
∂t ∂x (2.2.1)

u(x, 0) = φ(x),


where V is a constant. Recall that in Section 1.1, we announced that the solution is

u(x, t) = φ(x − V t). (2.2.2)

The idea used to solve this problem is to reduce the PDE to an ODE, by restricting the solution u(x, t)
to a curve on the xt−plane so that u now is a function of one variable and satisfies an ODE. The curve
is called characteristic curve of the PDE. This is called the method of characteristics. Let us carry
out the plan now. Consider a curve x = x(t); after restricting u on the curve it becomes a function of t
only (the function u(x(t), t)). Then by chain rule, we have
du dx
= ut + ux . (2.2.3)
dt dt
We wish to relate the righthand side to the PDE, so we demand that
dx
= V.
dt
Thus
x = x(t) = V t + C, (2.2.4)
where C is an arbitrary constant. (Note that all these characteristics are parallel lines and fill the entire
x-t plane.) Now by our choice of the curve, (2.2.3) and the PDE, we have
du
= ut + V ux = 0,
dt
hence u = M where the constant M depends on the characteristic curve and hence on C, i.e, u is a
function C, which is just x − V t according to (2.2.4). Thus we write

u(x, t) = f (x − V t) (2.2.5)

for some function f . On the other hand, as has already been verified in Section 1.1, (2.2.5) is always
a solution of the PDE (2.2.1) for any smooth function f . Thus the set of all solutions, or the general
solution of the PDE is given by (2.2.5).
We invoke the initial condition in (2.2.1) as has been done in Section 1.1 and so the solution of (2.2.1) is

u(x, t) = φ(x − V t).

Before we discuss the method of characteristics further, let us get some feel of this solution. If we take the
snapshot of the graph of u as a function of x at time t, it is just the graph of the initial value φ(x) shifted
to the right by V (to the left if V < 0). The graph is travelling with velocity V without changing its
shape. For this reason, u is called a travelling wave solution. This solution can be interpreted as modelling
a lump of (non-diffusive) pollutant in a river with water moving with constant speed V .
Example 2.2.2

∂u ∂u

 +y =u+y
∂x ∂y (2.2.6)

u(x, 1) = φ(x),


This is a linear inhomogeneous equation. The characteristics are give by

dy
= y,
dx
 9

hence by
y = Cex .
So u, when restricted on a fixed characteristic curve (so constant C is fixed), satisfies
du
= ux + yuy = u + Cex .
dx
Solving this ODE, we obtain
u = Cxex + M ex ,
where M is a constant that depends on the characteristic curve and hence on C, and thus M = f (C) for
a function f . But on the curve, C = ye−x . We then have the general solution of the PDE:

u(x, y) = xy + f (ye−x )ex .

Because of the initial condition u(x, 1) = φ(x), we have

φ(x) = x + f (e−x )ex ,

(φ(x) − x)e−x = f (e−x ),

f (z) = (φ(− ln z) + ln z)z.
Therefore,
u(x, y) = y ln y + yφ(x − ln y).
In this example, the initial value is not given on the x−axis, because the x−axis is a characteristic curve
on which the only choice for u is M ex , and even such an initial value is prescribed, it does not affect the
value of u off the x−axis, that is, such an initial value does not determine a unique solution. Let’s examine
the case u(x, 0) = ex . Then by the general solution formula,

ex = u(x, 0) = xy + f (ye−x )ex |y=0 = f (0)ex .

So f (0) = 1 and nothing else we can say about f . Then we cannot uniquely determine u(x, y) when y 6= 0.
The moral of this story is that if an initial condition is imposed on a characteristic curve, then there may
not exist a solution; if there exists a solution, the solution is not unique.

Characteristic curves for general first-order PDEs

After these two examples, it is now obvious that the characteristic curves of the general “quasilinear ”PDE

a(x, t, u)ut + b(x, t, u)ux = f (x, t, u)

are defined by the differential equation

a(x, t, u)dx − b(x, t, u)dt = 0. (2.2.7)

(We use the differentials because we have the advantage of re-writing the equation in the form of either
dx dt
dt = · · · or dx = · · ·.) In the linear case (a and b independent of u), (2.2.7) is an ODE that may be solved
as in the previous examples. But in the nonlinear case, (2.2.7) appears to be useless because it involves
the unknown solution u. This is not so as we can see from the following.
Example 2.2.3 Consider the following quasilinear traffic/transport equation

∂u ∂u

 + c(u) =0
∂t ∂x (2.2.8)

u(x, 0) = φ(x).


The ODE for characteristic curves is

dx
= c(u(x, t)). (2.2.9)
dt
10

On a fixed characteristic curve, we have

du dx
= ut + ux = ut + c(u)ux = 0.
dt dt
Thus on each characteristic curve, u = Const. M . Now going back to (2.2.9), we obtain the equation for
the characteristic curves
x = c(M )t + Const. x0 ,
which is a straight line that intersects the x−axis at x0 . So

u(x, t) = u(x0 , 0) = φ(x0 ) = φ(x − t · c(u(x, t))). (2.2.10)

This formula for u is not given in explicit form; this is a nonlinear phenomenon.
There is a more striking phenomenon due to nonlinearity: the initial value can be smooth, yet the solution
develops a singularity in finite time. This occurs when c0 (u) > 0 for all u and when there exist two points
x1 < x2 such that φ(x1 ) > φ(x2 ). In this case, the two characteristic lines issued from x1 and x2 on the
x−axis intersect at time
x2 − x1
t=− , (2.2.11)
c(φ(x2 )) − c(φ(x1 ))
which is positive. But along each characteristic line, the solution u must be constant. This shows that at
and after the time given by (2.2.11), a smooth solution is impossible to exist; something that determines
the smoothness of the solution must have broken down before or at this time. The question is: exactly
what has broken down? To answer this question, (2.2.10) is useful: differentiating both sides of it with
respect to x, we have
ux = φ0 (x − c(u)t)(1 − tc0 (u)ux ),
φ0 (x − c(u)t)
ux = .
1 + tc0 (u)φ0 (x − c(u)t)
Suppose there exists x0 such that
φ0 (x0 ) < 0.
Then along the characteristic line issued from the point x0 on the x−axis,
φ0 (x0 )
ux = .
1 + tc0 (φ(x0 ))φ0 (x0 )
Thus along the characteristic line as t tends to time
−1
> 0,
c0 (φ(x 0
0 ))φ (x0 )

the slope of u becomes unbounded. This is called the steepening effect. At or before the time given
above, the classical solution ceases to exist; the breakdown time or life-span of the classical solution is given
by
−1
ts = min .
x0 ∈R: φ (x0 )<0 c (φ(x0 ))φ0 (x0 )
0 0

When a classical solution stops to exist, we say a shock occurs. The subscript s in ts refers to the word
“shock”. Now the natural question is: how do we define a solution after the shock occurs? The resolution
is to introduce the notion of “weak solution”. In this course, we will not cover this topic.
We remark that if c0 (u) > 0 for all u and if the initial value φ is non-decreasing, then no two characteristic
lines intersect and the smooth solution exist for all time t ≥ 0.
It is helpful to interpret physically the appearance of shock: consider the case where c(u) = u and think
of (2.2.8) as a traffic/transport equation modelling the density of a substance flowing in a pipe. Then
the velocity function V is u/2 and so the more dense the substance is distributed, the faster it moves. If
the initial value φ is non-decreasing, particles move no slower than the ones behind (to the left), so we
 11

do not expect the steepening effect. On the other hand, if the initial value is not non-decreasing, then
some particles will catch up some others that are initially ahead of them, causing “collision”, and thus the
steepening effect.
Example 2.2.4 (Discontinuous initial value.) Find the solution of the following Burger’s equation with
discontinuous initial condition:

 ut + uux = 0 
−1, x < 0,
 u(x, 0) = φ(x) =
1, x > 0.

Solution
We study the following problem:

 ut + uux = 0
 
 −1, x < −,


 u(x, 0) = φ(x) = x/, − ≤ x ≤ ,
 
1, x > .

By (2.2.10), we have

 −1, x − ut < −,
u(x, t) = φ(x − u(x, t)t) = (x − ut)/, − ≤ x − ut ≤ ,
x − ut > ,

1,

i.e. 
 −1, x < −(t + ),
u(x, t) = x/(t + ), −(t + ) ≤ x ≤ (t + ),

1, x > t + .
Sending  → 0, we get the solution of the original problem

 −1, x < −t,
u(x, t) = x/t, −t ≤ x ≤ t,

1, x > t.
2
It is interesting to observe that the initial discontinuity is smoothed as illustrated in following figure. We
can say Burger’s equation favors non-decreasing initial values and dislikes other ones.
u(x, 0) 6 u(x, t0 ) 6

1 1

-x -x
x = t0

−1 −1

More general definition of characteristic curves

Consider a first-order partial differential equation or system on the xt−plane. A curve Γ on the plane is
said to be non-characteristic at a point P = (x0 , t0 ) ∈ Γ if the solution u of the PDE can be uniquely
determined near P via Taylor series by the PDE, the (prescribed) values of u on Γ and Γ itself. Here
“uniquely determined via Taylor series”means that all the partial derivatives of u at P can be uniquely
12

determined. If a curve is not non-characteristic at every point on it, then it is said to be a characteristic
curve of the PDE.
Now let us show that in the case of

a(x, t, u)ut + b(x, t, u)ux = f (x, t, u), (2.2.12)

this new definition of characteristic curves is equivalent to (2.2.7). Consider the following equations
(
aut + bux = f (x, t, u) (PDE)
(2.2.13)
ut dt + ux dx = du|Γ (du is known on Γ)

i.e.
" #( ) ( )
a b ut f (x, t, u)
=
dt dx ux du Γ

Note that (2.2.7) is equivalent to

" #
a b
det = 0,
dt dx

which in turn is equivalent to the fact that ut and ux cannot be uniquely determined. Thus (2.2.7) implies
that the curve Γ is characteristic in the sense of the new definition. Conversely, if (2.2.7) does not hold
at a point P (and for prescribed values of u on Γ), then both ut and ux at P = (x0 , t0 ) can be uniquely
determined. Moreover all higher order partial derivatives of u can be determined uniquely: differentiating
(2.2.13) with respect to t and then plugging in (x, t) = (x0 , t0 ), we have
(
autt + buxt = known quantity
utt + uxt dx/dt = known quantity,

thus (2.2.7) implies that utt and uxt at point P can be uniquely determined. Other higher-order derivatives
can be determined uniquely in the same fashion. Therefore if (2.2.7) fails at P , then Γ is non-characteristic
there. We have completed the proof of the equivalence of (2.2.7) and the new definition of characteristic
curves in the case of (2.2.12).
The advantage of the new definition of characteristic curves is that it applies to systems of PDEs where
we cannot foresee the likes of (2.2.7) easily. To make this point, we supply the following examples
Example 2.2.5 Consider the following first-order system of partial differential equations:
(
a1 ux + b1 uy + c1 vx + d1 vy = f1 (x, y)
a2 ux + b2 uy + c2 vx + d2 vy = f2 (x, y)

Find the characteristic equation.

Solution
Suppose Γ is a characteristic curve. Consider the following equations


 a1 ux + b1 uy + c1 vx + d1 vy = f1 (x, y)


 a2 ux + b2 uy +

c2 vx + d2 vy = f2 (x, y)


 ux dx + uy dy = du|Γ (du is known on Γ)



vx dx + vy dy = dv|Γ (dv is known on Γ).
 13

i.e.  


ux

 
 f1 (x, y) 

a1 b1 c1 d1 
   
   
 f2 (x, y)
   
 a2  uy
   
b2 c2 d2 
 

 dx
 =
dy 0 0  du
 vx
 
 
 

  
0 0 dx dy   
 
 


vy
 
 dv 

Γ

Notice that {ux , uy , vx , vy } cannot be uniquely determined if and only if

a1 b1 c1 d 1
 
 a2 b2 c2 d 2 
det 
 dx dy
 = 0.
0 0 
0 0 dx dy

As in the previous discussion regarding (2.2.12), if the ODE system fails, then Γ is non-characteristic.
Thus this ODE system defines the characteristic equation.
2
Unlike in the case of single equation (2.2.12), the system of PDEs in the above example can no longer be
reduced to a systems of ODEs along a characteristic curve. Then what is the point of finding these curves?
The answer lies partially in the fact that if the “initial value”is prescribed on a characteristic curve, then
generally either the existence or the uniqueness of solution fails, as we have seen in Example 2.2.2. To have
the existence and uniqueness, we should prescribe the initial value on a non-characteristic curve: indeed, a
general local existence theorem called Cauchy-Kovalevski Theorem says that if everything (coefficients
and given functions) in the system is analytic, if the non-characteristic curve and the initial value are also
analytic, then in a neighborhood of the curve, the initial value problem has one and only one analytic
solution, which is obtained from the Taylor series mentioned in the definition of non-characteristic curves.
We close this section by one more example.
Example 2.2.6∗ (Advanced problem)
The rotation-free 2-D gas equations are given by

uux + vuy + ρ−1 px = 0



 (1)

 uvx + vvy + ρ−1 py = 0

(2)

(2.2.14)


 (ρu)x + (ρv)y = 0 (3)


vx − uy =

0 (4),
where u and v are the velocity along the x and y directions respectively, p is the pressure, and ρ is the
density of the gas. In addition, p and ρ satisfy the relation
dp
= c2 (c is the speed of sound).
dρ

From the first two equations of (2.2.14), we have

u2 ux + vuuy + ρ−1 upx = 0 (5)

2 −1
uvvx + v vy + ρ vpy = 0 (6)

From dp = c2 dρ, we have

px = c2 ρx , py = c2 ρy .
Thus the third equation in (2.2.14), ρx u + ρux + ρy v + ρvy = 0 becomes

1
−ρ(ux + vy ) = ρx u + ρy v = (upx + vpy ).
c2
14

Substituting this into (5)+(6), we have

( 2
(u − c2 )ux + uv(uy + vx ) + (v 2 − c2 )vy = 0 (7)
vx − uy = 0 (4).

Therefore the characteristic equation is

 2
u − c2 uv uv v 2 − c2

 0 −1 1 0 
det 
 dx
 = 0,
dy 0 0 
0 0 dx dy

which is √
dy uv ± c u2 + v 2 − c2
= .
dx u2 − c2
2

Assignment 2

1. Solve the following initial value problem: 3ut + 5ux = 0, u(x, 0) = exp(−x2 ).

2. Find the general solution of ux + xuy = u.

3. Solve the initial value problem: ut + ux = x, u(x, 0) = 1/(1 + x2 ).

4. This exercise makes the point that the boundary condition for transport equations has to be given
carefully: show that the PDE ut + ux = 0, x ∈ [0, 1], t ∈ R, has no smooth solutions satisfying the
boundary condition u(0, t) = 1, u(1, t) = 2. Explain this physically. Hint: Draw several characteristic
curves.

5. Consider the following initial value problem for Burger’s equation


 ut + uux = 0
 
1, x ≤ 0,



 u(x, 0) = φ(x) = 1 − x, 0 < x ≤ 1,
 
0, x > 1.

(a) Find the time ts when shock first occurs; (b) Solve the initial value problem before time ts .

6. Let u be a positive C 1 -smooth solution of Burger’s equation

ut + uux = 0, x ∈ (−∞, ∞), t ≥ 0.

Prove that (a) for each fixed t ≥ 0, u is non-decreasing in x; (b) for each fixed x, u is non-increasing in
t ≥ 0. Hint: Argue by contradiction to prove (a).

Chapter 3 Parabolic Equations

3.1 Heat equation and reaction-diffusion equation: derivation

 15

The heat equation

We use the transport equation to derive the PDE satisfied by the temperature function. Let u(x, t) be
the temperature at location x ∈ Rn and at time t (n = 1, 2, 3), let E(x, t) be the thermal energy density
function. E(x, t) is calculated by first taking a small neighborhood of point x with volume ∆x, dividing
the thermal energy inside the neighborhood by ∆x, then taking the limit of the quotient as ∆x → 0. Let
F (x, t) be the creation-degradation rate of thermal energy (unit: energy/time/volume). Let J(x, t) be the
thermal energy flux which satisfies the same mass flux property mentioned in (2.1.2) in 3D: the net rate
at which thermal energy crosses a surface S in the direction of normal n is given by
Z
J · n(x, t) dS.
S

The unit for J is energy/time/length (length2 should be replaced in by lengthn−1 if n = 1, 2). Now by
2

the transport equation (2.1.4), we have

Et + ∇ · J = F. (3.1.1)
Common sense tells us that the thermal flux J should point in the direction of decreasing temperature.
In fact, for the case of an isotropic medium Fourier’s law states that

J(x, t) = −k∇u(x, t), (3.1.2)

where the gradient is taken in the x−variable, and k > 0 (so J points in the direction in which the
temperature function decreases most rapidly). Moreover an empirical law says that thermal energy density
is proportional to the temperature; more precisely,

E = cρu, (3.1.3)

where ρ is the density function of the medium, c the specific heat which is positive. In principle, k, c, ρ
depend on location and time, however, for simplicity we assume in this course that they all are constant.
k is called the thermal conductivity of the medium. Combining (3.1.1-3), we have the heat equation

ut − a2 ∆u = f (x, t), (3.1.4)

where a2 = k/(cρ) which is called the heat diffusion coefficient of the medium, and f = F/(cρ). The
unit of heat diffusion coefficient is length2 /time.

Reaction-diffusion equation
Let u(x, t) be the density function of a diffusive substance. Diffusion refers to the tendency of particles of
the substance moving from the region of higher concentration to the region of lower concentration. Then
Fick’s law, an analog of Fourier’s law, states that the flux of the substance is given by

J(x, t) = −k∇u(x, t),

where k is positive. Again, for simplicity in this course we assume that k is constant; it is called diffusion
coefficient. Now by the transport equation (2.1.4), we have the reaction-diffusion equation

ut − k∆u = f (x, t), (3.1.5)

where f is the creation-degradation rate of the substance.

3.2 Boundary conditions for heat and diffusion equations

If heat conduction occurs in a region Ω with boundary ∂Ω, obviously what happens on the boundary to
the temperature function affects the temperature function inside the region. Thus it is necessary to give
a boundary condition before solving the heat equation. (This also holds for reaction-diffusion equations.)
16

Typically there are three kinds of boundary conditions.

First kind B.C. (Dirichlet B.C.)

u(x, t) = µ(x, t), x ∈ ∂Ω.

The physical meaning of the first kind B.C. is that the temperature at boundary is specified. For example,
if a rod of length l with the left end being put in ice-water and the right end being put in boiling water,
then the boundary condition is u(0, t) = 0 (◦ C) and u(l, t) = 100 (◦ C).

Second kind B.C. (Neumann B.C.)

If the boundary is well-insulated, then there is no penitration of heat or substance through ∂Ω. It means
the normal component of the flux (thermal or otherwise) is zero:

∂u
J · n = −k = 0,
∂n
where n is the unit normal vector field on the boundary. In general, we can specify the normal component
of the flux on the boundary:
∂u
−k (x, t) = µ(x, t), x ∈ ∂Ω.
∂n
This is the general Neumann B.C..

Third kind B.C. (Robin B.C.)

Let the temperature of the medium surrounding Ω be µ(x, t). If for x ∈ ∂Ω, u(x, t) denotes the inner limit
limy∈Ω→x u(y, t), and if the surrounding medium has an intimate contact with ∂Ω, then we can assume that
the inner limit is just the temperature of the surrounding medium, ı.e. the Dirichlet B.C. u(x, t) = µ(x, t).
If the contact is not so intimate, e.g, optically flat surfaces lightly pressed, then by Newton’s cooling law
we have that the normal component of the thermal flux on the boundary is proportional to u(x, t)−µ(x, t):

∂u
−k (x, t) = H(u(x, t) − µ(x, t)),
∂n
where H is a positive constant, called coefficient of surface heat transfer. It can be re-written as

∂u
(x, t) + h(u(x, t) − µ(x, t)) = 0, x ∈ ∂Ω
∂n
where h = H/k is positive. This is the general Robin B.C. It is also called the radiation B.C.. Observe
in the limit h → 0, Robin B.C. becomes the homogeneous Neumann B.C.; in the limit h → ∞, Robin B.C.
becomes the Dirichlet B.C..
All these boundary conditions make sense for the reaction-diffusion equation.

3.3 Uniqueness of solution of heat equation via energy method

We now establish the uniqueness of the solution of the heat equation with initial condition and one of
the three boundary conditions discussed in the previous section. Let Ω be a bounded region in Rn with
piecewise C 1 -smooth boundary ∂Ω (so the divergence theorem applies).

ut = a2 ∆u + f (x, t)


 x ∈ Ω, t > 0

u(x, 0) = φ(x), x∈Ω (3.3.1)



Boundary conditions on ∂Ω,
 17

where φ(x) is the initial temperature distribution. Since the PDE and the boundary condition are lin-
ear, the uniqueness of the above problem is satisfied if and only if the homogeneous heat equation with
homogeneous boundary condition has only the trivial solution u = 0.

ut = a2 ∆u


 x ∈ Ω, t > 0

u(x, 0) = 0, x∈Ω (3.3.2)



Homogeneous boundary conditions on ∂Ω,

Here by homogeneity in the boundary condition, we mean that µ is zero. Multiplying both sides of the
homogeneous heat equation by u and integrating on Ω, we have
1 d
R 2 R
2 dt Ω u (x, t)dx = uu (x, t)dx
ΩR t
2
= a Ω u∆u(x, t)dx
= a2 Ω (∇ · (u∇u) − |∇u|2 )(x, t)dx
R

= a2 [ ∂Ω u∇u · n(x, t)dS − Ω |∇u|2 (x, t)dx]

R R

≤ a2 ∂Ω u∇u · n(x, t)dS,

R

where we have used the divergence theorem. The boundary integral is zero in the cases of Dirichlet
and Neumann boundary conditions; it is non-positive in the Robin case because h is positive. Thus the
following function of t is nonincreasing: Z
u2 (x, t)dx. (3.3.3)
Ω
This together with the facts that this function is non-negative and is equal to zero initially implies that
Z
u2 (x, t)dx ≡ 0, t > 0.
Ω

Thus u(x, t) ≡ 0, x ∈ Ω, t > 0.

The method we used above is called the energy method. Here the term “energy”refers to the function
(3.3.3) (in physics, a quantity decreasing in time is often called energy), though (3.3.3) is not a physical
energy at all.

3.4 Method of separation of variables

Model problem
Consider the following initial-boundary value problem for the heat equation:

ut = a2 uxx


 (0 < x < l, t > 0)

u(0, t) = 0, u(l, t) = 0 (3.4.1)



u(x, 0) = φ(x).

In the above model problem, the PDE and boundary condition are homogeneous.

Separation of variables
The strategy to solve (3.4.1) is to first find solutions that satisfy the PDE and the boundary condition,
and then take care of the initial condition. The solutions to start with have the following form

u(x, t) = X(x) · T (t). (3.4.2)

We say that the variables x and t are separated. Substituting this into the PDE, we have

X(x) · T 0 (t) = a2 X 00 (x) · T (t),

18

i.e.
X 00 (x) T 0 (t)
= 2 .
X(x) a T (t)

6 6
function of t
function of x

Notice that the left hand side of the above equation is a function only of x, while the right hand side is a
pure function of t. The only possibility is that both sides of the above equation are constants, i.e.

X 00 (x) T 0 (t)
= 2 ≡ −λ (constant),
X(x) a T (t)

where the negative sign is selected because we will show that λ > 0. Thus we have two separated equations:
( 00
X (x) + λ X(x) = 0 (0 < x < l),
(3.4.3)
T 0 (t) + a2 λ T (t) = 0 (t > 0).

Now plugging (3.4.2) into the boundary condition, we have

X(0) · T (t) = 0 and X(l) · T (t) = 0 (t > 0). (3.4.4)

Eigenvalue problem
The boundary conditions in (3.4.4) leads to that X(0) = 0 and X(l) = 0. Thus we get

X 00 (x) + λ X(x) = 0
(
(0 < x < l),
(3.4.5)
X(0) = 0, X(l) = 0

X(x) = 0 is obviously a solution of the above problem. But it gives only a trivial solution. Thus we demand
that X 6= 0. (3.4.5) is analogous to the eigenvalue problem of a linear transformation and so (3.4.5) is said
to be an eigenvalue problem, the unknown constant λ is called eigenvalue, any nontrivial function X
is called eigenfunction, the set of all eigenfunctions corresponding to a fixed eigenvalue λ, plus the zero
function, is called the eigenspace. Any eigenspace is linear in the sense that if u1 and u2 belong to the
eigenspace, then so does c1 u1 + c2 u2 for any constants c1 and c2 .
It is easy to check that λ = 0 leads to a trivial solution and thus 0 is not an eigenvalue. For λ 6= 0, the
general solution of the ODE in (3.4.5) is clearly given by
√ √
−λ x
X(x) = C1 e + C2 e− −λ x
.

By the boundary condition, we have

(
C1 + C2 = 0, ( by B.C. at x=0)
√ √
−λ l
C1 e + C 2 e− −λ l
= 0. ( by B.C. at x=l)

For nonzero solution {C1 , C2 }, one must have

" #
1 1 √ √ √
det √ √ = e− −λ l
−e −λ l
= 0, i.e. e2 −λ l
= 1.
−λ l
e e− −λ l

This leads to
√
2 −λ l = i 2nπ (n = 1, 2, · · ·).
 19

Then the eigenvalues are given by

 nπ 2
λn = (n = 1, 2, · · ·). (3.4.6)
l
and the corresponding eigenfunctions are
nπx nπx nπx
Xn (x) = C1 (e l i − e− l i ) = Const. C sin (n = 1, 2, · · ·). (3.4.7)
l
(So each eigenspace is one-dimensional, meaning that any eigenfunction corresponding to a fixed λn is a
constant multiple of a fixed eigenfunction.)

Principle of superposition
For each n, the second equation in (3.4.3) becomes
 nπa 2
Tn0 (t) + Tn (t) = 0 ∀ n.
l
It is easy to show that   
nπa 2
Tn (t) = φn exp − t , (n = 1, 2, · · ·) (3.4.8)
l
where φn is an arbitrary constant. Thus we have infinitely many functions
  
nπa 2 nπx
un (x, t) = Xn (x) · Tn (t) = φn exp − t sin , (n = 1, 2, · · ·), (3.4.9)
l l

that satisfy the PDE and the B.C. (3.4.1). According to the principle of superposition for homogeneous
linear equations, a finite sum of solutions is still a solution; this still holds for an infinite sum if the
convergence of the sum (series) is guaranteed. To take care of the initial condition, we boldly form the
sum of all solutions un to obtain
∞   
X nπa 2 nπx
u(x, t) = φn exp − t sin . (3.4.10)
n=1
l l

(The first person to do so was Fourier.) Formally, that is, non-rigorously, u satisfies the PDE and the B.C.
Indeed, because of the exponential decay of un and its partial derivatives as n → ∞ if t > 0, it is possible
to prove that if all the φn are bounded, then u is well-defined (meaning: the series converges) and all its
partial derivatives exist and are continuous in the upper plane t > 0; moreover u satisfies the PDE and
the B.C. for t > 0. The rigorous proofs of these statements are not required for this course, unless your
instructor insists otherwise.
It is our hope that by choosing the coefficients φn suitably, u satisfies the initial condition. We wish
∞
X nπx
u(x, 0) = φn sin = φ(x). (3.4.11)
n=1
l

That means the coefficients φn in (3.4.10) are determined by the Fourier expansion of initial temperature
φ(x). Recall the formula
(
2 l
Z
nπx mπx 0 m 6= n
sin sin dx = (3.4.12)
l 0 l l 1 m = n.
nπx
Then multiplying (3.4.11) by sin and integrating term-by-term on [0, l], we have
l
2 l
Z
nπx
φn = φ(x) sin dx (3.4.13)
l 0 l
20

The problem is completely solved. 2

Remark 1
Notice that
∞   
X nπa 2 nπx
u(x, t) = φn exp − t sin −→ 0 as t → + ∞.
n=1
l l
This result coincides with our physical intuition. The temperature will be zero eventually, since the
terminals of the rod are put in ice-water mixture and there is neither creation nor degradation of thermal
energy inside of the rod (f (x, t) = 0).
Remark 2
Notice that the convergence rate of the term
  
nπa 2
Tn (t) = exp − t −→ 0 as t → + ∞
l

is faster, if the integer n is larger. That means the higher frequency components, which is related to
the larger n in sin nπxl , will vanish faster than the lower frequency components. Thus we expect the
temperature will be smoothed as time increases. Indeed, as we mentioned before u(x, t) as a function of
x becomes infinitely smooth right after t = 0, even if the initial φ is very rough (e.g. nowhere continuous
but is square-integrable in the sense of Lesbegue). This phenomenon is called the smoothing effect of
heat equation, which also holds for general parabolic equations.
Remark 3
According to Remark 2, a solution u of the heat equation (with a B.C.) at time t = 1 must be infinitely
smooth in x. So if we want to solve the heat equation backward from time t = 1, the initial value u(x, 1)
must be at least this smooth. So in general, solving the heat equation backward in time is an ill-posed
problem. It is ill-posed because of another reason: the solutions
  
1 nπa 2 nπx
un (x, t) = exp − t sin , (n = 1, 2, · · ·),
n l l

at time t = 0 are all bounded between −1/n and 1/n and hence are small if n is large, but at a negative
time, say, t = −1, un becomes unbounded as n → ∞. Thus small input leads to unbounded output; this
makes no physical sense and is highly unstable. This discussion of ill-posedness also applies to general
parabolic equations. The most famous backward parabolic equation is the Black-Scholes PDE that arises
in Finance
∂V 1 ∂2V ∂V
+ σ 2 S 2 2 + rS − rV = 0. (3.4.14)
∂t 2 ∂S ∂S
2
Note that the sign of the coefficient of ∂∂SV2 is positive, unlike the heat equation. The coefficient becomes
negative if we change variable by replacing t by T − t (T is an arbitrary constant). Black-Scholes equation
in its original form (3.4.14) can only be solved if a terminal condition V (S, T ) = φ(S) is given, that is,
V (S, t) can be determined by φ for t ≤ T . Fortunately, solving Black-Scholes equation with a terminal
condition at the “expiration day”(of the call option) is helpful for stock investors. See Exercise 9.
In Chapter 5, we will see that the backward problem is well-posed for wave propagation problems (hyper-
bolic equation), e.g. sonar, electro-magnetic wave.
Remark 4
The methodology of separation of variables applies to general linear parabolic equations in higher spatial
dimensions. The technical difficulty in handling the higher dimensional case is that X 00 in the eigenvalue
problem (3.4.5) now becomes the Laplacian ∆X, so we can no longer use the ODE method to solve the
eigenvalue problem easily. In this course, we will not go into this in detail.
 21

Summary The following is a summary of the separation of variables:

Step 1: Let u(x, t) = X(x) · T (t). Derive the eigenvalue problem for X(x) from PDE and BCs.
Step 2: Find the eigenvalues λn and eigenfunctions Xn (x) from the eigenvalue problem.
Step 3: Find Tn (t) for each λn from the equation derived in Step 1.
Step 4: Set
∞
X
u(x, t) = φn Tn (t) Xn (x).
n=1

Find the φn from initial condition.

3.5 Eigenvalue problems: Sturm-Liouville theory and eigen-expansion

(3.4.12) makes it possible to find Fourier coefficients in the fashion of (3.4.13). Here is its generalization.

Theorem 3.5.1
Suppose that λm and λn (λm 6= λn ) are two eigenvalues of the following general eigenvalue problem:
 00

 X (x) + λ X(x) = 0 (a < x < b),

α1 X 0 (a) + α2 X(a) = 0, (3.5.1)


β1 X 0 (b) + β2 X(b) = 0,


where {α1 , α2 } are not all zeroes, {β1 , β2 } are not all zeroes, and Xm (x), Xn (x) are eigenfunctions
corresponding to λm , λn respectively. Then the eigenfunctions Xn (x) and Xm (x) are orthogonal in the
following sense:
Z b
Xn (x)Xm (x) dx = 0. (3.5.2)
a

Proof
Since Xm (x), Xn (x) are eigenfunctions w.r.t. eigenvalues λm , λn , we have
( 00
Xm (x) + λm Xm (x) = 0, (1)
Xn00 (x) + λn Xn (x) = 0. (2)

Xn (x)×(1)-Xm (x)×(2) leads to

Rb 00
0 = [Xn (Xm
a
+ λm Xm ) − Xm (Xn00 + λn Xn )]dx
Rb 00
Rb
= a (Xn Xm − Xm Xn00 )dx + (λm − λn ) a Xn Xm dx (3.5.3)
0
Rb
= (Xn Xm − Xm Xn0 )|ba + (λm − λn ) a Xn Xm dx.

Notice that
0
(
α1 Xm (a) + α2 Xm (a) = 0,
α1 Xn0 (a) + α2 Xn (a) = 0.
{α1 , α2 } is a non-trivial solution of the above system, we must have
0
 
Xm (a) Xm (a)
det 0 = 0,
Xn (a) Xn (a)

i.e.
0
Xm (a)Xn (a) − Xm (a)Xn0 (a) = 0.
22

Similarly we have
0
Xm (b)Xn (b) − Xm (b)Xn0 (b) = 0.
Then (3.5.3) becomes
Z b
(λm − λn ) Xn Xm dx = 0.
a
Since λm 6= λn , one must have
Z b
Xn Xm dx = 0.
a
2
In the previous section, we have seen that the eigenvalue problem (3.4.5) with Dirichlet B.C. has infinitely
many eigenvalues (3.4.6) that diverge to infinity, each of which has a one dimensional eigenspace. We also
note that all the eigenfunctions corresponding to λn change sign exactly n − 1 times. All these can be
generalized for the more general eigenvalue problem (3.5.1) as follows.

Theorem 3.5.2
(i) The eigenvalues of (3.5.1) are real and form an increasing and diverging sequence

λ1 < λ2 < · · · < λn < · · · → ∞.

(ii) For each n, the eigenspace Vn corresponding to the eigenvalue λn is one dimensional.
(iii) Any eigenfunction corresponding to λn changes sign exactly n − 1 times; in particular, every eigen-
function corresponding to λ1 is either positive or negative on (a, b).
We point out that if the B.Cs are not given as in (3.5.1), then eigenspaces may not be one-dimensional
(e.g. in the case of the periodic B.C, all except the first eigenspaces are two dimensional - see Exercise
3, part (2)). Theorems 3.5.1 and 3.5.2 form the Sturm-Liouville theory that actually works when X 00 in
(3.5.1) is replaced by the more general operator (a(x)X 0 )0 + c(x)X. We will not go into details in this
course.
Suppose we have obtained all eigenvalues {λn , n = 1, 2, · · ·}, and their corresponding eigenfunctions
{Xn (x), n = 1, 2, · · ·}. To solve the heat equation ut = a2 uxx with the B.C. in (3.5.1), we form

un (x, t) = φn exp(−λn a2 t)Xn (x).

To satisfy the initial condition u(x, 0) = φ(x), we construct

∞
X ∞
X
u(x, t) = un (x, t) = φn exp(−λn a2 t) Xn (x). (3.5.4)
n=1 n=1

Formally, to satisfy the initial condition, we choose φn such that

∞
X
φ(x) = φn Xn (x). (3.5.5)
n=1

Then by Theorem 3.5.1, we deduce

Z b ∞
X Z b Z b
φ(x)Xm (x) dx = φn Xn (x) Xm (x) dx = φm (Xm (x))2 dx ∀ m.
a n=1 a a

Thus the coefficients φn in (3.5.4) are given by

Rb
a
φ(x)Xn (x) dx
φn = Rb . (3.5.6)
a
(Xn (x))2 dx
 23

The right hand side of (3.5.5) with φn given by (3.5.6) is called the generalized Fourier series or the
eigenexpansion of φ. There is a subtle point in the above discussion: what we have actually shown is
that if the right hand side of (3.5.5) converges to φ, then φn is given by (3.5.6), namely, only the generalized
Fourier series can converge to φ. But then logically, we must answer the following question: what’s the
requirement on φ so that the the generalized Fourier series/eigenexpansion of φ converges to φ in some
sense? The following theorem answers the question.

Theorem 3.5.3
(i) If φ is square integrable, i.e, φ2 has a finite integral on [a,b], then the eigenexpansion of φ converges to
φ in the L2 sense:
Z b Xk
|φ(x) − φn Xn (x)|2 dx → 0, as k → ∞.
a n=1

(ii) If φ, φ0 and φ00 all are continuous on [a, b] and φ satisfies the B.C. in (3.5.1), then the eigenexpansion
of φ converges to φ uniformly on [a, b].

Remark In some special cases, such as the case when the B.C. at both x = 0 and x = l is Dirichlet or
when the B.C. at both boundary points is Neumann (so the eigen-expansion is either the classical sine
series or the cosine series), to have the uniform convergence in (ii), we actually do not even need the
existence of φ00 , and in the Neumann case we do not need φ to satisfy the B.C.. In these two special
cases, weaker assumptions yield weaker conclusions: the Fourier series converges to φ(x) at every x in the
open interval (a, b) if φ is continuous and φ0 is piecewise continuous on [a, b] (piecewise continuity means
continuity everywhere except at finitely many points where the function still has finite one-sided limits);
In the more general case when φ and φ0 are piecewise continuous on [a, b], the Fourier series converges to
the average of the right and left limits of φ

1
(φ(x − 0) + φ(x + 0))
2
at every x in (a, b); the sine series converges to 0 at x = 0, l (because all the sine terms in the series are
equal to 0 at these two end points); the cosine series at x = 0 converges to φ(0 + 0), at x = l to φ(l − 0).

Special cases of the eigenvalue problem

We summarize four special cases of (3.5.1) below and then supply more of them with details.

1◦
X 00 (x) + λ X(x) = 0
(
(0 < x < l),
X(0) = 0, X(l) = 0
nπ 2


Eigenvalues: λn = l , n = 1, 2, · · ·
Eigenfunctions: Xn (x) = sin nπx
l , n = 1, 2, · · ·.
Rl l
Integral formula: 0
(Xn (x))2 dx = , n = 1, 2, · · · .
2
2◦
X 00 (x) + λ X(x) = 0
(
(0 < x < l),
X 0 (0) = 0, X 0 (l) = 0
nπ 2


Eigenvalues: λn = l , n = 0, 1, 2, · · ·
Eigenfunctions: Xn (x) = cos nπx
l , n = 0, 1, 2, · · ·.
24
Rl Rl l
Integral formula: 0
(X0 (x))2 dx = l; 0
(Xn (x))2 dx = , n = 1, 2, · · · .
2
3◦
X 00 (x) + λ X(x) = 0
(
(0 < x < l),
X(0) = 0, X 0 (l) = 0
2
(n+ 21 )π

Eigenvalues: λn = l , n = 0, 1, 2, · · ·
(n+ 12 )πx
Eigenfunctions: Xn (x) = sin l , n = 0, 1, 2, · · ·.
Rl l
Integral formula: 0
(Xn (x))2 dx = , n = 0, 1, 2, · · · .
2

4◦
X 00 (x) + λ X(x) = 0
(
(0 < x < l),
X 0 (0) = 0, X(l) = 0
2
(n+ 12 )π

Eigenvalues: λn = l , n = 0, 1, 2, · · ·
(n+ 12 )πx
Eigenfunctions: Xn (x) = cos l , n = 0, 1, 2, · · ·.
Rl l
Integral formula: 0
(Xn (x))2 dx = , n = 0, 1, 2, · · · .
2

Example 3.5.1
X 00 (x) + λ X(x) = 0
(
(−π < x < π),
X(−π) = 0, X 0 (π) = 0

Solution
Let ξ = x + π and X(x) = X(ξ − π) ≡ y(ξ). Then the above eigenvalue problem becomes
( 00
y (ξ) + λ y(ξ) = 0 (0 < ξ < 2π),
y(0) = 0, y 0 (2π) = 0
According to the result 3◦ , the eigenvalues and eigenfunctions are given by
2  2
(n + 21 )π

n 1
λn = = + , n = 0, 1, 2, · · ·
2π 2 4
and  
n 1
yn (ξ) = sin + ξ, n = 0, 1, 2, · · · .
2 4
Therefore the solution of the original problem is given by
 2
n 1
λn = + , n = 0, 1, 2, · · ·
2 4
and  
n 1
Xn (x) = sin + (x + π), n = 0, 1, 2, · · · .
2 4
2

Example 3.5.2
X 00 (x) + λ X(x) = 0
(
(0 < x < l),
X(0) = 0, X 0 (l) + hX(l) = 0
 25

where h is a positive constant.

Solution
Let’s first prove that λ ≤ 0 cannot be an eigenvalue. Multiplying the ODE by X(x) and integrating on
[0, l], we have
Rl Rl
0 = 0 XX 00 (x)dx + λ 0 X 2 (x)dx
Rl Rl
= XX 0 (l) − XX 0 (0) − 0 (X 0 )2 (x)dx + λ 0 X 2 (x)dx
Rl Rl
= −hX 2 (l) − 0 (X 0 )2 (x)dx + λ 0 X 2 (x)dx,
Z l Z l
λ 2 2
X (x)dx = hX (l) + (X 0 )2 (x)dx.
0 0

Hence λ ≥ 0. If λ = 0, then the right hand of the last equation is zero, implying X 0 (x) ≡ 0 on [0, l]. So
X(x) ≡ Const. on [0, l]. But X(0) = 0 and so X(x) ≡ 0, which contradicts the early agreement that an
eigenfunction is not identically equal to zero. We have shown that any eigenvalue must be positive.
Now we can write the general solution of the ODE in the eigenvalue problem as :
√ √
X(x) = C1 sin( λx) + C2 cos( λx).

The boundary condition X(0) = 0 yields that C2 = 0. Thus we have

√
X(x) = sin( λx).

The boundary condition X 0 (l) + hX(l) = 0 leads to

√ √ √
λ cos( λ l) + h sin( λ l) = 0,

i.e.
µ √
= − tan µ (µ = λ l).
hl
This equation has infinitely many nonzero roots as illustrated below.

µ
(
(((
((( (( h l
(((
( (((((
(((
((((
((( • π • 2π • 3π -µ
µ1 µ2 µ3
− tan µ

Therefore we get infinitely many eigenvalues and eigenfunctions:

  2
 λn = µn
 n = 1, 2, 3, · · ·
l
µ x
 Xn (x) = sin n
 n = 1, 2, 3, · · · .
l
2
26

3.6 Non-homogeneous problem

In section 3.4, the method of separation of variables is applied to solve the homogeneous heat conduction
problem with homogeneous boundary conditions. For non-homogeneous boundary conditions, the method-
ology is to transfer the non-homogeneous BCs to homogeneous ones. For non-homogeneous equations, the
methodology is to first expand the inhomogeneous term by the eigenfunctions and then use the technique
for solving non-homogeneous ODE.

Model problem
ut = a2 uxx + f (x, t)


 (0 < x < l, t > 0),

u(0, t) = µ1 (t), u(l, t) = µ2 (t) (t > 0), (3.6.1)



u(x, 0) = φ(x) (0 < x < l).

Transforming non-homogeneous boundary conditions

Let
u(x, t) = U (x, t) + w(x, t), (3.6.2)
where U (x, t) is supposed to satisfy the homogeneous boundary condition, i.e.

U (0, t) = 0, U (l, t) = 0 (t > 0).

Thus the function w(x, t) must satisfy the non-homogeneous boundary condition, i.e.

w(0, t) = µ1 (t), w(l, t) = µ2 (t) (t > 0). (3.6.3)

It is easy to select w(x, t) to satisfy the boundary condition (3.6.3)

x
w(x, t) = µ1 (t) + (µ2 (t) − µ1 (t)). (3.6.4)
l
Then the non-homogeneous boundary condition in (3.6.3) is satisfied. Substituting (3.6.2) with (3.6.4)
into (3.6.1), we get that

Ut = a2 Uxx + F (x, t)


 (0 < x < l, t > 0),

U (0, t) = 0, U (l, t) = 0 (t > 0), (3.6.5)



U (x, 0) = Φ(x) (0 < x < l),

where 
∂w
 F (x, t) = f (x, t) −

∂t (3.6.6)

Φ(x) = φ(x) − w(x, 0).


The original problem (3.6.1) for u(x, t) with non-homogeneous BC is transformed to the problem (3.6.5)
for U (x, t) with homogeneous BC.

Solutions of non-homogeneous equations with homogeneous BCs

For the inhomogeneous problem (3.6.5) with homogeneous BC, we still follow Step 1 and Step 2 for
solving homogeneous PDE with homogeneous BCs described in section 3.4. That means we find the
eigenvalues and eigenfunctions (3.4.1)
  nπ 2
 λn = (n = 1, 2, · · ·),


l
 nπx
 Xn (x) = sin
 (n = 1, 2, · · ·).
l
 27

Now we expand F (x, t) and initial condition Φ(x) in {Xn (x), n = 1, 2, · · ·}, i.e.
∞
X
F (x, t) = Fn (t) Xn (x) (3.6.7)
n=1

and
∞
X
Φ(x) = Φn Xn (x), (3.6.8)
n=1

where Rl
0
F (x, t)Xn (x) dx
Fn (t) = Rl
0
(Xn (x))2 dx
and Rl
0
Φ(x)Xn (x) dx
Φn = Rl .
0
(Xn (x))2 dx
We also expand the solution
∞
X
U (x, t) = Tn (t) Xn (x). (3.6.9)
n=1

Substituting (3.6.7)-(3.6.9) into (3.6.5), we get

P∞ P∞
n=1 Tn0 (t) Xn (x) = n (a2 Tn (t) Xn00 (x) + Fn (t) Xn (x)),
P∞
= n (−a2 λn Tn (t) Xn (x) + Fn (t) Xn (x)),
∞
X ∞
X
Tn (0)Xn (x) = Φn Xn (x).
n n

Multiplying both sides of each of the above equation by Xk (x) and integrating on [0, l], then using the
orthogonality proved in Theorem 3.5.1, we have

Tk0 (t) + a2 λk Tk (t) = Fk (t)

(
∀k
(3.6.10)
Tk (0) = Φk .

Solving this initial value problem for the linear ODE, we obtain
Z t
2 2
Tn (t) = e−a λn (t−τ )
Fn (τ ) dτ + Φn e−a λn t
. (3.6.11)
0

In this fashion, the solution U (x, t) of (3.6.5) and the solution u(x, t) = U (x, t) + w(x, t) of (3.6.1) are
determined.

Summary
The following is a step-by-step description for the method of separation of variables:
Step 1: (Nonhomogeneous BCs)
Let u(x, t) = U (x, t) + w(x, t). Select w(x, t) to transform the non-homogeneous BC to a homogeneous
one.
Step 2: (Eigenvalue problem)
Let U (x, t) = X(x) · T (t). Derive the eigenvalue problem for X(x) from the homogeneous PDE and BC.
Step 3: (Eigenvalues and eigenfunctions)
Find the eigenvalues λn and eigenfunctions Xn (x) from the eigenvalue problem.
28

Step 4: (Eigenexpansions)
Find Fourier/eigenexpansions of the non-homogeneous term in the equation, F (x, t), and of the initial
value Φ(x).
Step 5: (Solution of non-homogeneous equation)
P
Set U (x, t) = Tn (t) Xn (x). Find Tn (t) from the ODE derived from the nonhomogeneous PDE and the
initial condition.

Example 3.6.1
ut = a2 uxx + x − π


 (0 < x < π, t > 0),

u(0, t) = π, u(π, t) = 0 (t > 0), (3.6.12)


u(x, 0) = π − x

(0 < x < π).

Solution
Step 1: (Nonhomogeneous BC)
Let u(x, t) = U (x, t) + w(x, t) and w(x, t) = π − x. The problem (3.6.12) becomes

Ut = a2 Uxx + x − π


 (0 < x < π, t > 0),

U (0, t) = 0, U (π, t) = 0 (t > 0), (3.6.13)



U (x, 0) = 0 (0 < x < π).

Step 2: (Eigenvalue problem)

Let U (x, t) = X(x) · T (t). The eigenvalue problem is found to be
( 00
X (x) + λ X(x) = 0 (0 < x < π),
X(0) = 0, X(π) = 0.

Step 3: (Eigenvalues and eigenfunctions)

According to 1◦ in Section 3.5, the eigenvalues and their corresponding eigenfunctions are given by

λn = n2 ; Xn (x) = sin(nx), n = 1, 2, · · · .

Step 4: ( Eigen-expansions)
Let
∞
X
F (x, t) = x − π = Fn sin(nx).
n=1

The Fourier coefficients {Fn , n = 1, 2, · · ·} are found by

2 π
Z
2
Fn = (x − π) sin(nx) dx = − n = 1, 2, · · · .
π 0 n

Step 5: (Solution of non-homogeneous equation)

Set
∞
X
U (x, t) = Tn (t) Xn (x).
n=1

Substituting this into (3.6.13), we get that

 T 0 (t) + (na)2 T (t) = − 2


n n ∀n
n
Tn (0) = 0.

 29

The general solution of the homogeneous ODE is C exp(−(na)2 t), and the particular solution of the above
non-homogeneous ODE is (by inspection) −2/(n3 a2 ). Thus the general solution is given by
2 2 2
Tn (t) = C e−n a t
− .
n3 a 2
The initial condition Tn (0) = 0 leads to

2  2 2

Tn (t) = e−n a t
−1 .
n3 a2
Therefore we get that
∞ ∞
X X 2  −n2 a2 t 
U (x, t) = Tn (t) Xn (x) = e − 1 sin(nx),
n=1 n=1
n3 a2

and
∞
X 2  −n2 a2 t 
u(x, t) = U (x, t) + w(x, t) = π − x + e − 1 sin(nx).
n=1
n3 a2
2

3.7 Fundamental solution of heat equation

Define function G(x, t; ξ) by

(x − ξ)2
 
1
G(x, t; ξ) = √ exp − (t > 0), (3.7.1)
2a πt 4a2 t

where x and t are variables, ξ ∈ R a parameter. It can be directly verified that G ∈ C ∞ w.r.t. x or t, and
satisfies the homogeneous heat conduction equation for any ξ, i.e.

Gt = a2 Gxx (t > 0). (3.7.2)

G(x, t; ξ) is called the fundamental solution of heat conduction equation.

Remark
There are several approaches to the derivation of the fundamental solution G(x, t; ξ) . In Appendix 3.1,
G(x, t; ξ) is derived by the method of separation of variables. Other approaches, e.g. similarity analysis,
Fourier transformation, may be found in other textbooks.

Property 1◦

The graph of G(x, t; ξ) is illustrated as below

G(x, t; ξ)
1 6
√
2a πt

-x
ξ
30

Figure 3.7.1 Fundamental solution G(x, t; ξ) at time t and position ξ

Property 2◦

∂m
G(x, t; ξ) −→ 0 as x → ∞, f or t > 0, ξ ∈ (−∞, + ∞). (3.7.3)
∂xm

Property 3◦
Z +∞
G(x, t; ξ)dx = 1, f or t > 0, ξ ∈ (−∞, + ∞). (3.7.4)
−∞

Property 4◦

0 x 6= ξ
lim+ G(x, t; ξ) = (3.7.5)
t→0 +∞ x = ξ.

Property 5◦
Z +∞
lim f (ξ)G(x, t; ξ)dξ = f (x), (3.7.6)
t→0+ −∞

where f (x) is any bounded continuous function.

The proofs of properties 2◦ -4◦ are simple. The proof of property 5◦ is given in Appendix 3.2.

Theorem 3.7.1
Consider Cauchy problem for the heat equation:

 ut = a2 uxx −∞ < x < +∞, t>0
(3.7.7)
 u(x, 0) = φ(x) −∞ < x < +∞,

where φ(x) is bounded and continuous on R. The following function gives a bounded solution
Z +∞
u(x, t) = φ(ξ)G(x, t; ξ)dξ. (3.7.8)
−∞

Proof
Z +∞
ut − a2 uxx = φ(ξ)(Gt − a2 Gxx )dξ = 0.
−∞
Z +∞
lim+ u(x, t) = lim+ φ(ξ)G(x, t; ξ)dξ = φ(x).
t→0 t→0 −∞

We shall prove in the future the uniqueness of Cauchy problem, i.e. it has at most one bounded solution.
Thus (3.7.8) is the only bounded solution of the Cauchy problem. We emphasize that the Cauchy problem
has no boundary condition. In applications, the initial values are often piecewise continuous, then at every
continuous point x of φ, we still have limt→0+ u(x, t) = φ(x); at a discontinuous point x, we have
1
lim u(x, t) = (φ(x − 0) + φ(x + 0)),
t→0+ 2
that is, the limit is the average of the left and right limits of the initial value. Since there are just discretely
many discontinuous points of φ where the initial condition is not satisfied exactly, we still say (3.7.8) is
the solution of the Cauchy problem.
 31

Physical interpretation of fundamental solution and a critique of the heat/diffusion equation

Consider the Cauchy problem (3.7.7) with the following initial condition:
∆Q


 |x − x0 | < δ
φ(x) = 2cρδ (3.7.9)

|x − x0 | > δ,

0
where c is the specific heat and ρ is the mass density of the rod. The physical meaning of the above initial
condition is that the heat energy ∆Q is concentrated on the interval (x0 − δ, x0 + δ).
According to Theorem 3.7.1 and the comment following it, the solution of (3.7.7) with initial condition
(3.7.9) is given by
Z +∞
u(x, t) = φ(ξ)G(x, t; ξ)dξ
−∞
Z x0 +δ
∆Q
= G(x, t; ξ)dξ
x0 −δ 2cρδ
∆Q ¯
= G(x, t; ξ),
cρ
where ξ¯ ∈ (x0 − δ, x0 + δ) whose existence is guaranteed by the integral mean value theorem. Now send
δ → 0+ . Then ξ¯ → x0 , i.e. the heat energy source is put right at the position x0 . The temperature
distribution becomes
∆Q
u(x, t) = G(x, t; x0 ).
cρ
∆Q
The physical quantity is called thermal density. Therefore the fundamental solution G(x, t; x0 )
cρ
is the temperature distribution when one unit of thermal density is concentrated at position x0
when t = 0.
The physical interpretation of the fundamental solution G in the context of diffusion is similar. Recall
that in that context, u represents the density function of a substance. So it is natural to replace in (3.7.9)
∆Q by ∆m, and eliminate cρ, where ∆m is the total mass in the interval (x0 − δ, x0 + δ). Then we have

u(x, t) = ∆m G(x, t; x0 ).

Thus the fundamental solution G(x, t; x0 ) is the mass distribution function when a unit point mass is put
at x0 at time t = 0.
Critique. The above interpretation of the fundamental solution actually reveals a flaw of the heat/diffusion
equation: if we put one unit of thermal density or point mass at 0 at time t = 0 and if there is no cre-
ation of thermal energy or mass, then even if the thermal energy/mass travels at the speed of light - the
supposedly highest speed at which any object can travel according to Einstein - we still need to wait for a
while to feel a fragment of the thermal energy/mass 10 miles away; but the temperature or mass density
10 miles away from the origin at any time t is raised from 0 to G(10, t; 0) > 0 ! This is certainly absurd,
and is enough for the purist to immediately abandon the heat equation. But wait: if a2 = 1 and t = 0.1,
then G(10, t; 0) = 2.3811 × 10−109 , so tiny that the flaw of the heat/diffusion equation is practically harm-
less! Moreover, the heat/diffusion equation has been proved to match experimental data so well that it is
accepted as the standard model in the science and engineering community.

Art is a lie that tells the truth. − −P ablo P icasso

Example 3.7.1
32

Solve the following Cauchy problem on the half-line:

ut = a2 uxx


 0 < x < +∞, t > 0,

u(x, 0) = 0 0 < x < +∞, (∗)



u(0, t) = N0 , t > 0.

Solution
Consider the standard Cauchy problem on the whole line:

ut = a2 uxx
(
−∞ < x < +∞, t > 0,
(∗)0
u(x, 0) = φ(x) −∞ < x < +∞,

where the initial condition φ(x) is set to be

(
0 x>0
φ(x) =
Φ(x) x < 0.

The solution of (∗)0 satisfies problem (*), except the boundary condition u(0, t) = N0 . We will select a
suitable Φ(x) to satisfy this boundary condition. According to Theorem 3.7.1, the solution of (∗)0 is given
by
Z +∞ Z 0
u(x, t) = φ(ξ)G(x, t; ξ)dξ = Φ(ξ)G(x, t; ξ)dξ.
−∞ −∞

The condition u(0, t) = N0 leads to

Z 0
u(0, t) = Φ(ξ)G(0, t; ξ)dξ = N0 .
−∞

Notice that G(0, t; ξ) is an even function of ξ. Thus we have

Z +∞
Φ(ξ)G(0, t; ξ)dξ = 2N0 .
−∞

The above condition is satisfied, if we take Φ(x) as

Φ(x) = 2N0 .

Therefore the solution of (*) is given by

Z 0 Z 0
u(x, t) = Φ(ξ)G(x, t; ξ)dξ = 2N0 G(x, t; ξ)dξ
−∞ −∞
Z +∞
= 2N0 G(x, t; −ξ)dξ
0
+∞
(x + ξ)2
Z  
1
= 2N0 √ exp − dξ.
0 2a πt 4a2 t
2
Remark
x+ξ
The solution of the above example is usually expressed in another form. Letting y = √ , we have
2a t
 33

Z +∞
2 2
u(x, t) = √ N0 e−y dy
π x√
2a t
Z +∞ Z x√ !
2 −y 2 2a t
−y 2
= N0 √ e dy − e dy
π 0 0
 
x√
= N0 1 − erf( ) ,
2a t
where the so-called error function is a standard special function in most of mathematical software and
function tables, defined by
Z x
2 2
erf(x) = √ e−u du
π 0

Higher dimensional case

For the heat equation in higher spatial dimensions

ut = a2 ∆u, x ∈ Rn , t > 0,

the fundamental solution is given by

1 −|x − ξ|2
G(x, t; ξ) = exp( ).
(4πa2 t)n/2 4a2 t
Properties 1◦ -5◦ , and Theorem 3.7.1 hold with the obvious modifications.

3.8 The Maximum principles

This section is devoted to qualitative study of the heat equation. Consider a refrigerator occupying region
Ω; and think about the maximum of the temperature function u(x, t) inside the refrigerator during time
interval [0, T ]. The maximum of u(x, t) must be achieved either at a boundary point at some time between
t = 0 and t = T , or inside Ω at time t = 0 (as in the case of a refrigerator which is turned on at time
t = 0). Thus
max u = max u, (3.8.1)
DT ΓT

where DT = Ω × (0, T ] and ΓT = (∂Ω × [0, T ]) ∪ (Ω × {0}). DT is called the parabolic interior , and ΓT
the parabolic boundary of the cylinder Ω × [0, T ].
This is the physical interpretation of the weak maximum principle for the heat equation that we
will formulate mathematically now. First let us think about how to use the PDE language to model a
refrigerator. We have a refrigerator if and if the creation-degradation rate of thermal energy is non-positive,
i.e. f in (3.1.4) is non-positive in DT . Now we are ready to state

Weak maximum principle

Let Ω be a bounded region in Rn and T > 0. Suppose on DT u(x, t) satisfies

ut − a2 ∆u ≤ 0. (3.8.2)

Then (3.8.1) holds.

Proof.
We first claim that if the strict inequality in (3.8.2) holds, then the maximum of u on DT cannot be
achieved at a point (x0 , t0 ) in the parabolic interior DT . Suppose otherwise. Then u(x0 , t) as a function
34

of t, achieves its maximum on [0, T ] at t0 . Depending on whether t0 < T or t0 = T , we have

ut (x0 , t0 ) = 0, or ≥ 0. (3.8.3)

On the other hand, consider the function u(x, t0 ), which assumes its maximum value at x0 . By the Second
Derivative Test, the Hessian matrix

∂2u
( (x0 , t0 ))1≤i, j≤n
∂xi ∂xj

is non-positive-definite. So the trace of this matrix is non-positive. But the trace is just the Laplace of u
and hence
∆u(x0 , t0 ) ≤ 0.
Combining this with (3.8.3) we reach a contradiction to (3.8.2) at point (x0 , t0 ). We have proved (3.8.1)
when the strict inequality in (3.8.2) holds.
Now suppose we do not have the strict inequality. Define a new function

v(x, t) = u(x, t) + e−t ,

where  > 0 is a constant. The new function satisfies the strict inequality in (3.8.2) and so it satisfies

max v = max v.
DT ΓT

Sending  → 0, we conclude the proof of the weak maximum principle.

Remark. There is a subtle point about (3.8.1): it does say that the maximum of u on DT is achieved on
the parabolic boundary ΓT ; but it does not rule out the possibility that the maximum is also achieved at
a parabolic interior point. Indeed, if u is a constant function on DT , then of course (3.8.2) is fulfilled but
the maximum is also assumed at parabolic interior points. It turns out that being a constant function is
the only way for u to satisfy (3.8.2) and to achieve its maximum at a parabolic interior point. This result
is called

Strong maximum principle

Suppose all the assumptions in the weak maximum principle hold; assume that the maximum of u on DT
is taken at a point (x0 , t0 ) ∈ Ω × (0, T ]. Then u is a constant function on Ω × [0, t0 ].
The proof of this is too technical to be taught at the undergraduate level. We emphasize that after time
t0 , u is not guaranteed to be a constant - think about a well-insulated refrigerator which is turned on at
time t0 !
If the inequality sign in (3.8.2) is reversed, then physically we have the scenario of an oven, and mathe-
matically we have the weak and strong minimum principles: i.e. (3.8.1) with all “max”replaced by
“min”, etc.

Comparison principle
Let u and v satisfy

ut − a2 ∆u ≥ vt − a2 ∆v


 x ∈ Ω, T ≥ t > 0

u(x, t) ≥ v(x, t), x ∈ ∂Ω, T ≥ t > 0 (3.8.4)


u(x, 0) ≥ v(x, 0), x ∈ Ω.


Then
u(x, t) ≥ v(x, t), x ∈ Ω × [0, T ].
 35

Moreover, if there exists (x0 , t0 ) ∈ Ω × (0, T ] where u and v touch each other, then u and v are identical,
at least before time t0 .
Proof
Let w = v − u. Then w satisfies (3.8.2) and hence by the weak maximum principle, we have

max w = max w.
DT ΓT

But on the vertical component of ΓT , w ≤ 0 by the B.C.; the same is true at time t = 0 by the initial
condition. Thus w ≤ 0 and u ≤ v on DT .
If u and v touch at (x0 , t0 ), then w achieves its maximum in the parabolic interior DT . It follows from
the strong maximum principle that w is identically equal to zero before time t0 . This completes the proof.

Corollary 1. Uniqueness of heat equation

The initial-Dirichlet boundary value problem for heat equation

ut − a2 ∆u = f (x, t)


 x ∈ Ω, T ≥ t > 0

u(x, t) = g(x, t), x ∈ ∂Ω, T ≥ t > 0


u(x, 0) = φ(x), x ∈ Ω.


has at most one solution.

Proof
Suppose v is another solution. Then we can apply the comparison principle twice to obtain u ≡ v. Recall
this uniqueness was proved in section 3.3 via the energy method.

Corollary 2. Structural stability

For each of i = 1, 2, let ui be the solution of

uit − a2 ∆ui = fi (x, t)



 x ∈ Ω, T ≥ t > 0

ui (x, t) = gi (x, t), x ∈ ∂Ω, T ≥ t > 0


ui (x, 0) = φi (x), x ∈ Ω.


Then
max |u1 − u2 | ≤ max |φ1 − φ2 | + max |g1 − g2 | + T max |f1 − f2 |. (3.8.5)
DT Ω ∂Ω×[0,T ] DT

Proof
Let v = u1 − u2 and

w(x, t) = max |φ1 − φ2 | + max |g1 − g2 | + t max |f1 − f2 |.

Ω ∂Ω×[0,T ] DT

Then
wt − a2 ∆w ≥ vt − a2 ∆v


 x ∈ Ω, T ≥ t > 0

w(x, t) ≥ v(x, t), x ∈ ∂Ω, T ≥ t > 0


w(x, 0) ≥ v(x, 0), x ∈ Ω.


By the comparison principle, we have v ≤ w on DT . Similarly, −v ≤ w on DT . These and the fact that
w is less than or equal to the right hand side of (3.8.5) imply (3.8.5).
36

In applications, the structure components, i.e. the source term f , the boundary value g and the initial
value φ are measured experimentally and hence are not given precisely. Equation (3.8.5) says that small
errors in measuring these data result in a small error in the solution. Thus we have structural stability.

Weak maximum principle for Cauchy problem

Let u(x, t) be a bounded function and satisfy

ut − a2 ∆u ≤ 0, x ∈ Rn , 0 < t ≤ T,

(3.8.6)
u(x, 0) = φ(x), x ∈ Rn .

Then
sup u = sup φ. (3.8.7)
Rn ×[0,T ] Rn

Proof
The idea is the same as in the case of bounded region. The new difficulty is that the sup of u may not be
assumed at a finite point. To overcome this, we use a function Γ(x, t) that satisfies the heat equation, but
has (exponential) growth as |x| → ∞, so when we subtract u by Γ, the sup is achieved at a finite point.
This Γ is a modification of the fundamental solution:
1 2 2
Γ(x, t) = √ e|x| /(4a (T +1−t)) .
T +1−t
We check by direct computation that
Γt − a2 ∆Γ = 0.
Take a small constant  and define

v(x, t) = u(x, t) − Γ(x, t) + e−t .

Then v satisfies the strict inequality in (3.8.6). The sup of v on Rn × [0, T ] is assumed somewhere at
(x0 , t0 ). If t0 > 0, then we reach a contradiction as in the case of bounded region. So t0 = 0 and hence

sup v = sup(φ(x) − Γ(x, 0) + ).

Rn ×[0,T ] Rn

Sending  → 0, we obtain (3.8.7).

From this maximum principle, it follows immediately the uniqueness of bounded solutions of the Cauchy
problem
ut − a2 ∆u = f (x, t), x ∈ Rn , 0 < t ≤ T,


u(x, 0) = φ(x), x ∈ Rn .
We mention that without the boundedness condition, there is no uniqueness: the Cauchy problem has
infinitely many unbounded solutions that have the 0 initial value.

Appendix 3.1: Fundamental Solution of Heat Equation

Consider the model problem in section 3.4:

2

 ut = a uxx


(0 < x < l, t > 0),
u(0, t) = 0, u(l, t) = 0 (t > 0), (1)



u(x, 0) = φ(x) (0 < x < l).

The solution of (1) is given by (3.4.10) and (3.4.13), i.e.

 37

∞   
X nπa 2 nπx
u(x, t) = φn exp − t sin
n=1
l l
∞
!
2 l
Z   
X nπξ nπa 2 nπx
= φ(ξ) sin dξ exp − t sin
n=1
l 0 l l l
∞
Z l X   !
2 nπξ nπx nπa 2
= sin sin exp − t φ(ξ) dξ
0 n=1
l l l l
Z lX
≡ ·φ(ξ) dξ
0
nπ P
Let ωn = . The in the above formula becomes
l
∞
X X 2 2 2
= sin(ωn ξ) sin(ωn x)e−ωn a t

n=1
l

Notice that
nπ π π
ωn = = n · ≡ n · ∆ω (∆ω = )
l l l

-ω
ω0 ω1 ω2 ωn

Let l → +∞ (∆ω → 0). We have

∞
X X 2 2 2
= sin(ωn ξ) sin(ωn x)e−ωn a t · ∆ω
n=1
π
Z +∞
2 2 2
−→ sin(ωξ) sin(ωx)e−ω a t · dω
0 π
Z +∞
1 2 2
= [cos ω(ξ − x) − cos ω(ξ + x)] e−ω a t · dω
0 π
To compute this, let
Z +∞
2
I(c) = cos(ωc)e−ω b dω,
0

where b is a positive constant. Integrating by parts leads us to

R +∞ 2
I(c) = 0 cos(ωc)e−ω b dω,
R +∞ 2
= 1c 0 e−ω b d sin(ωc),
+∞ 2
= − 1c 0 sin(ωc)de−ω b ,
R
+∞ 2
= 2b sin(ωc)ωe−ω b dω,
R
c 0
dI(c)
= − 2b
c dc .

Thus
dI(c) c
= − I(c).
dc 2b
Solving this (linear and also separable) ODE, we have
c2
I(c) = I(0)e− 4b ,
38

where by changing variable, R +∞ −ω2 b

I(0) = 0 R
e dω,
1 +∞ −η 2
= √
b 0
e dη,
√
= √π .
2 b
Now define Z +∞
1 2 2
G(x, t; ξ) = cos ω(ξ − x)e−ω a t
· dω.
π 0
Then

(x − ξ)2
 
1
G(x, t; ξ) = √ exp −
2a πt 4a2 t

Notice that
Z lX Z +∞
u(x, t) = ·φ(ξ) dξ −→ [G(x, t; ξ) − G(−x, t; ξ)]φ(ξ) dξ as l → +∞.
0 0

Then the function

Z +∞
U (x, t) = [G(x, t; ξ) − G(−x, t; ξ)]φ(ξ) dξ (2)
0

might be the solution of the following problem

2

 Ut = a Uxx


0 < x < +∞, t>0
U (x, 0) = φ(x) 0 < x < +∞ (3)



U (0, t) = 0, t>0
If φ is odd on R, then (2) becomes (3.7.8).

Appendix 3.2: Weak Convergence of Fundamental Solution

We are going to prove Property 5◦ in Section 3.7:

Z +∞
lim+ f (ξ)G(x, t; ξ)dξ = f (x), (∗)
t→0 −∞

where f (·) is bounded on R and is continuous at x. By (3.7.4) and the the formula for G, we have
Z +∞ Z +∞
f (x) = f (x) G(x, t; ξ)dξ = f (x)G(x, t; ξ)dξ
−∞ −∞

so we can rewrite (*) as

Z +∞
lim (f (ξ) − f (x))G(x, t; ξ)dξ = 0. (∗∗)
t→0+ −∞
Introducing a new variable
ξ−x
η= √ ,
4ta
we rewrite the integral on the left hand side of (**) as

1
Z +∞
2 √
√ e−η (f (x + 4taη) − f (x))dη. (∗ ∗ ∗)
π −∞
 39

Now (**) follows from Lesbegue dominated convergence theorem.

If the reader has not learned the Lesbegue theorem, then we offer a more elementary proof. Since f is
2
bounded, we can find a constant M such that |f (y)| ≤ M for any y. Since e−η is integrable on R, for
any  > 0, there exists a large L such that
Z
1 2 
√ e−η dη ≤ .
π |η|≥L 2M

Then
1
Z
2 √ 
√ | e−η (f (x + 4taη) − f (x))dη| ≤ 2M · = .
π |η|≥L 2M
Since f is continuous at x, there exists a small τ > 0 such that
√
|f (x + 4taη) − f (x)| ≤  for any 0 < t < τ, |η| ≤ L,

and hence 2 √ 2
√1 | e−η (f (x √1 e−η  dη
R R
π |η|≤L
+ 4taη) − f (x))dη| ≤ π |η|≤L
R +∞ −η2
≤ √1 e dη · 
π −∞
= .
Then the absolute value of (***) is less than or equal to
Z Z
| |+| | ≤  +  for any 0 < t < τ.
|η|≥L |η|≤L

This completes the proof of (**).

According to the definition of weak convergence, the above identity means G(x, t; ξ) weakly converges
to Dirac-δ function δ(x − ξ), i.e.
G(x, t; ξ) * δ(x − ξ). (4)
The identity (*), therefore, is often referred as the weak convergence property of the fundamental solution
of the heat conduction equation.

Assignment 3

1. (Transmission conditions)
Consider a surface S that separates two media with different thermal conductivities k1 and k2 . Let u1
and u2 be the temperature in the media. Suppose the media are in intimate contact along the surface S
so we have
u1 = u2 on S. (1)
Prove that on S,
∂u1 ∂u2
−k1 = −k2 , (2)
∂n ∂n
where n is the unit normal vector field of the surface S. ( (1) and (2) are called transmission conditions.)
Hint: Take an arbitrary patch ∆S of S, and think about the rate at which thermal energy crosses the
patch in the direction of the normal.

2. (Effective boundary condition on a coated body)

Let a body Ω1 (space shuttle or turbine engine) be thermally insulated by a thin coating Ω2 of thickness
δ; assume the outer boundary of the coating is subject to a high exterior temperature H. Let u1 be the
temperature function in Ω1 and u2 be that in Ω2 that satisfies (1) on ∂Ω1 . Let the thermal conductivities
40

of the body and the coating be k1 and k2 , respectively. Prove that on the boundary ∂Ω1 of the body, we
have approximately Robin boundary condition
∂u1 k2
k1 + (u1 − H) = 0, (3)
∂n δ
where n is the unit outer normal vector field of ∂Ω1 . (Equation (3) is called the effective boundary
condition; its significance is that with it we do not need to solve, analytically or numerically, the heat
equation inside the coating–we just need to solve it inside the body with (3) as the B.C.) To insulate the
body well, what should be the scaling relationship of k2 and δ? Hint: start with (2); fix a point x on ∂Ω1 ,
and define f (τ ) = u2 (x + τ n). Then perform a Taylor expansion of f at 0.

3. Solve the following eigenvalue problems.

(1)
X 00 (x) + λX(x) = 0
(
−l < x < l
X 0 (−l) = 0, X(l) = 0

(2)
X 00 (x) + λX(x) = 0,
(
0<x<l
X(x) is a periodic function with period l

4. Find the eigenvalues of the following problem graphically

00

 X + λX = 0, x ∈ (0, l),
X(0) = 0,
X 0 (l) + hX(l) = 0


where h is a nonzero constant that may not be positive. Note that negative eigenvalues may appear. In
this case, what is the behavior of the solution for the corresponding initial-boundary value problem for
the homogeneous heat equation?

5. Solve the following boundary-initial value problems.

(1)
ut = a2 uxx


 0 < x < l, t>0

u(0, t) = u1 , u(l, t) = u2 , t>0



u(x, 0) = u0 0<x<l
where u0 , u1 and u2 are constants. After solving it, find the limit of u(x, t) as t → ∞. Show that the limit
is a steady-state, ı.e. a time-independent solution of the PDE and B.C.
(2)
ut = a2 uxx − h u + g


 0 < x < l, t>0

u(0, t) = 0, u(l, t) = 0, t>0



u(x, 0) = 0 0<x<l
where g and h are constants.
(3)
ut = k 2 (uxx + uyy )


 0 < x < a, 0 < y < b, t>0


 u(0, y, t) = u(a, y, t) = 0




 u(x, 0, t) = u(x, b, t) = 0



u(x, y, 0) = xy
 41

Reference answer:
∞ X
∞   2
m2
 
X
m+n 4ab n 2 2 nπx mπy
u(x, y, t) = (−1) 2
exp − 2
+ 2 π k t · sin sin .
m=1 n=1
mnπ a b a b

6. Consider the initial-Neumann boundary value problem

2

 ut = a uxx


0 < x < l, t > 0
ux (0, t) = 0 = ux (l, t), t>0



u(x, 0) = x 0<x<l

Find the limit of u(x, t) as t → ∞ by inspecting the general solution formula obtained by separation of
variables. (You do not need to compute all the Fourier coefficients.) Interpret your result physically;
generalize it, without proof, to the case of general initial value and higher spatial dimensions.

7. Solve the Cauchy problem for heat equation

ut − a2 uxx = 0

x ∈ (−∞, +∞), t > 0,
(4)
u(x, 0) = φ(x), x ∈ (−∞, +∞),
where 
1, if |x| ≤ 1,
φ(x) =
0, if |x| > 1.
Show that u decays as t → ∞ and find the decay rate. Explain, physically, why u decays as t → ∞.

8. (Symmetry of heat equation)

Let u be the bounded solution of the Cauchy problem (4) with a general initial value. Show that if the
initial value φ is even, then so is u in x; likewise, if φ is odd, then so is u in x. Hint: either use the explicit
solution formula or use the maximum principle for the Cauchy problem.

9. (Black-Scholes equation)
Consider the terminal value problem for the Black-Scholes equation
(
∂V 1 2 2 ∂2V ∂V
∂t + 2 σ S ∂S 2 + rS ∂S − rV = 0, S > 0, 0 < t < T,
(5)
V (S, T ) = φ(S), S > 0,

where S is the price of a stock (as independent variable), V the call option value (as the dependent
variable), σ the volatility of the stock, r the risk-free interest rate, T the expiration day of the option.
This homework is designed to show that this terminal value problem can be transformed to the Cauchy
problem for the heat equation and therefore (5) can be solved explicitly.
(a) Introduce new variables
S = Kex , t = T − τ /(σ 2 /2),
where the constant K is the striking price. Let v(x, τ ) = V (S, t). Show that
∂v ∂2v 2r ∂v 2r
= + ( 2 − 1) − v.
∂τ ∂x2 σ ∂x σ 2

(b) Choose constants α and β such that

u(x, τ ) = exp(αx + βτ )v(x, τ )

satisfies the heat equation

∂u ∂2u
= .
∂τ ∂x2
42

(c) Now solve (5) with

φ(S) = max(S − K, 0) (European call).
Express your answer in terms of the distribution function of the normal distribution
Z y
1 2
Φ(y) = √ e−ξ /2 dξ.
2π −∞

10. (Application of maximum principles)

Let u be a smooth solution of the initial-boundary value problem
2

 ut − a uxx = 0, x ∈ (0, l), t > 0
u(0, t) = 0 = u(l, t), t>0 (6)
x ∈ [0, l],

u(x, 0) = φ(x),

where φ ≥ 0 but is not identically equal to zero on [0, l], satisfying φ00 < 0 on (0, l).
(a) Prove that that u(x, t) > 0 for (x, t) ∈ (0, l) × (0, ∞). Hint: first use the weak minimum principle and
then the strong minimum principle.
(b) Prove that ut (x, t) < 0 for (x, t) ∈ (0, l) × (0, ∞). Hint: Let w = ut ; first find the initial-boundary
problem that w solves, then apply the maximum principles to w.
(c) Draw the graph of u vs x and put arrows on the graph to indicate the behavior of the graph as t
increases. Can you predict the behavior without proving (b) rigorously? What if the initial value changes
its concavity?

11. Consider the solution (3.7.8) of the Cauchy problem (3.7.7). If φ is bounded on R and has a jump
discontinuity at point x, prove that
1
lim u(x, t) = (φ(x − 0) + φ(x + 0)).
t→0+ 2

12. (Backward uniqueness)

We have already discussed the ill-posedness of solving the heat equation backwards in time. But, perhaps
surprisingly, the backward heat equation has uniqueness, as we will prove in this exercise.
Let u be a smooth solution of
ut = a2 uxx


 (0 < x < l, t < 0),

u(0, t) = 0, u(l, t) = 0 (t < 0),



u(x, 0) = 0 (0 < x < l).

Prove that u is identically equal to zero on [0,l] for all t ≤ 0. To this end, recall the energy that we have
defined before
Z l
E(t) = u2 (x, t)dx.
0
We just need to prove that E(t) is identically equal to zero for all t < 0. We argue by contradiction by
assuming that there exists t0 < 0 such that E(t0 ) > 0. By continuity, there exists a t1 ∈ (t0 , 0] such that
E is positive on [t0 , t1 ) and is equal to zero at t1 . Without loss of generality, assume t1 = 0. Now proceed
as follows
(a) Prove
Z l
E 00 (t) = 4a4 u2xx (x, t)dx.
0
 43

(b) Prove Cauchy-Schwarz inequality

Z l Z l Z l
1 1
| f (x)g(x)dx| ≤ ( f 2 (x)dx) 2 ( g 2 (x)dx) 2 .
0 0 0

Hint: the quadratic polynomial in r defined by

Z l
(f (x) + rg(x))2 dx
0

is always non-negative for all r; think about its discriminant.

(c) Use the formula for E 0 and E 00 to prove

(E 0 )2 ≤ E E 00 , t ∈ [t0 , 0).

(d) Prove that

(ln E(t))00 ≥ 0, t ∈ [t0 , 0).

(e) Prove that (d) contradicts the assumption that u(x, 0) ≡ 0 (E(0) = 0).

Chapter 4 Elliptic Equations

4.1 Laplace and Poisson equations

The Laplace equation is defined by

∆u = 0 ,

or its component form:

∂2u ∂2u
+ 2 =0 ( in 2 − D ),
∂x2 ∂y
(4.1.1)
∂2u ∂2u ∂2u
+ 2 + 2 =0 ( in 3 − D ).
∂x2 ∂y ∂z

A solution of the Laplace equation is called a harmonic function. The inhomogeneous version of Laplace
equation

−∆u = f , (4.1.2)

where f is a given function, is called Poisson’s equation.

Motivations
Steady states of heat conduction equation
Consider the heat equation
ut − a2 ∆u = f (x), x ∈ Rn .
If we have a solution u = u(x) which is independent of time t, then we call u steady state of the heat
equation. Steady states therefore satisfy the Poisson equation.
Analytic function of a complex variable
44

Write the complex variable z = x + i y and the complex function

f (z) = u(x, y) + i v(x, y),

where u(x, y) and v(x, y) are real-valued functions. An analytic function is one that is expressible as a
power series in z, i.e.
+∞
X
f (z) = an z n ,
n=0

with complex coefficients an . That is

+∞
X
u(x, y) + i v(x, y) = an (x + iy)n .
n=0

Formal differentiations w.r.t. x and y of this series show that

∂u ∂v ∂u ∂v
= and =− ,
∂x ∂y ∂y ∂x

which are called the Cauchy-Riemann equations. If we differentiate them, we find that

uxx = (vy )x = (vx )y = −uyy .

So that ∆u = 0. Similarly we have ∆v = 0. Thus the real and imaginary parts of an analytic function are
harmonic.
Irrotational and incompressible fluid
Consider a fluid that is irrotational, that is, its velocity vector field v(x) (x ∈ R3 ) satisfies

curl v = 0.

By multivariate calculus, if the region occupied by the fluid has no holes, then the velocity vector field has
a potential φ: ∇φ = v. Suppose the fluid is also incompressible, that is,

∇ · v = 0.

Then we see the potential φ satisfies the Laplace equation

∆φ = 0.

Electrostatics
Suppose at point x0 there is a point charge qx0 . By Coulomb’s law, the electric force exerted by qx0 on a
positive unit point charge located at point x is

qx0 (x − x0 )
Ex0 (x) = ,
4π|x − x0 |3

where we define the unit of charges so that the Coulomb constant is just 1/(4π). The electric potential
induced by the charge qx0 is a scalar function Ux0 of x such that

−∇Ux0 (x) = Ex0 (x).

It can be checked directly that

qx0
Ux0 (x) = .
4π|x − x0 |
 45

It can also be verified that the “electric field ”Ex0 is divergence-free, hence the electric potential function
satisfies the Laplace equation

∆Ux0 (x) = 0, x 6= x0 .
Now consider the scenario of a continuous distribution of charges inside a dielectric material Ω. Sup-
pose the density function of charges is f (x) (unit: charges/volume). Then the amount of charge in an
infinitesimal volume element dVx0 at x0 is qx0 = f (x0 )dVx0 . Treating these volume elements as point
charges, by the superposition principle we have that the induced electric potential is
Z
f (x0 )
V (x) = dVx0
Ω 4π|x − x0 |

where the integral is with respect to x0 . If f is smooth, then one can prove (very hard!) that V satisfies
Poisson equation
−∆V (x) = f (x), x ∈ Ω. (4.1.3)
This is also true outside Ω if we interpret that the density function f is zero outside Ω.

4.2 Separation of variables

4.2.1 Rectangular coordinates

Example 4.2.1 Solve


 ∆u = 0 in Ω = {M = (x, y); 0 < x < 1, 0 < y < 1}

ux (0, y) = 0 = ux (1, y),



u(x, 0) = 0, u(x, 1) = x

Solution
We use a trial solution in the form of
u(x, y) = X(x)Y (y),
and take care of the PDE, and the B.C. at the lateral boundary and at the bottom of the square first
(because there, the B.C. is homogeneous). By the PDE, we have

X 00 (x)Y (y) + X(x)Y 00 (y) = 0

and so
X 00 (x) Y 00 (y)
=− = Const. = −λ.
X(x) Y (y)
X satisfies the Neumann B.C. X 0 (0) = 0 = X 0 (1). Thus

λn = n2 π 2 , Xn (x) = cos(nπx), n = 0, 1, · · · .

On the other hand, Y satisfies B.C. Y (0) = 0 (again, do not worry about the B.C. at y = 1 at this
moment). The general solution for the Y −ODE is given by Yn (y) = aenπy + be−nπy for n = 1, 2, · · · ; and
Y0 (y) = ay + b for n = 0. By B.C. Y (0) = 0, we have

Y0 (y) = a0 y, Yn (y) = an (enπy − e−nπy ), n = 1, 2, · · · .

Now we form
∞
X ∞
X
u(x, y) = un (x, y) = a0 y + an (enπy − e−nπy ) cos(nπx),
n=0 n=1
46

which satisfies the PDE, and the B.C. at the lateral boundary and the bottom of the square. To satisfy
the B.C. at the top of the square, we demand

∞
X
a0 + an (enπ − e−nπ ) cos(nπx) = x,
n=1

from which we have

Z 1
1
a0 = x dx = ,
0 2

Z 1
2
an (enπ − e−nπ ) = 2 x cos(nπx) dx = (cos(nπ) − 1), n = 1, 2, · · · .
0 n2 π 2

Thus
∞
1 X 2((−1)n − 1) enπy − e−nπy
u(x, y) = y+ cos(nπx).
2 n=1
n2 π 2 enπ − e−nπ

4.2.2 Polar coordinates

In the previous example, we use the rectangular coordinates to solve the boundary value problem for the
Laplace equation because of the shape of the region. If the region is circular, then it is convenient to use
polar coordinates.
Laplace operator in polar coordinates
y
6
y * (x, y)

r
 

θ -x
x


x = r cos θ
y = r sin θ

The Laplace operator has the form (See Appendix 3.1 for proof.) :

∂2 1 ∂ 1 ∂2
∆= + + (4.2.1)
∂r2 r ∂r r2 ∂θ2

Laplace operator in spherical coordinates

z
6

* (x, y, z)

r
  z
θ 

P -y

ψ P

P
 47


 x = r sin θ cos ψ
y = r sin θ sin ψ

z = r cos θ.

The Laplace operator has the form (See Appendix 3.1 for proof.) :

∂2 ∂2
 
2 ∂ 1 ∂ ∂ 1
∆= 2 + + 2 sin θ + 2 2
∂r r ∂r r sin θ ∂θ ∂θ r sin θ ∂ψ 2

Separation of variables
We assume that

u(r, θ) = R(r) · Θ(θ).

Substituting this into (4.2.1), we have

r2 R00 + rR0 Θ00

= ≡ −λ.
−R Θ

This leads to equations

Θ00 (θ) + λΘ(θ) = 0, (4.2.2)

and
r2 R00 (r) + rR0 (r) − λR(r) = 0. (4.2.3)

Eigenvalue problem with periodic condition

The function Θ(θ) must be a 2π-periodic function because u is so in θ. Then the equation (4.2.2) becomes
an eigenvalue problem with periodic condition:
( 00
Θ (θ) + λΘ(θ) = 0
Θ(θ + 2π) = Θ(θ).

The solutions of the ODE (4.2.2) are exponential functions (for λ < 0), linear functions (for λ = 0) and
trigonometric functions (for λ > 0). Because of the periodicity, one must have

eigenvalues : λ n = n2 n = 0, 1, 2, · · ·
eigenfunctions : Θn = An cos(nθ) + Bn sin(nθ) n = 0, 1, 2, · · · .

General solution of Laplace equation in polar coordinates

Now we solve the Euler’s equation (4.2.3) with λ = λn . When λ = λ0 = 0, equation (4.2.3) is reduced to
r2 (R0 (r))0 + r(R0 (r)) = 0. The solution is clearly given by

R0 (r) = C0 + D0 ln r.

When λ = λn = n2 > 0, we assume that R(r) = rk . Substituting this into equation (4.2.3), we have

k(k − 1) + k − n2 = 0, or k = ±n.
That means
R(r) = Cn rn + Dn r−n .
48

Now we see that the general solution of the Laplace equation in polar coordinates has the following series
form:
+∞
X
Cn rn + Dn r−n (An cos(nθ) + Bn sin(nθ)) .


u(r, θ) = C0 + D0 ln r + (4.2.4)
n=1

Remark
From formula (4.2.4) it follows that the following functions are special harmonic functions:

1, ln r
rn cos(nθ), rn sin(nθ) n = 1, 2, · · ·
r−n cos(nθ), r−n sin(nθ) n = 1, 2, · · · .

The first two harmonic functions, which are functions of r only, may be derived directly from the Laplace
equation. Let

u(r, θ) = R(r).
Then ∆u = 0 leads to ODE
1
R00 + R0 = 0.
r
The solution is clearly

u(r, θ) = R(r) = a + b ln r.
1
The constant function is a trivial solution of the Laplace equation. The singular function − ln r, in fact,
2π
plays an important role in solving the Poisson equation, which will be studied in Sections 4.3 and 4.4.

Harmonic function in unit circle

We consider the following Laplace equation in the unit circle with arbitrary boundary condition:
(
∆u = 0 0 ≤ r < 1, 0 ≤ θ < 2π
(4.2.5)
u(1, θ) = φ(θ) 0 ≤ θ < 2π.

The solution u(r, θ) must be continuous, and then bounded, at r = 0. Therefore the coefficients Dn ,
including D0 , in (4.2.4) must vanish. The solution of (4.2.5) may be written as
+∞
a0 X n
u(r, θ) = + r (an cos(nθ) + bn sin(nθ)) , (4.2.6)
2 n=1

where the coefficients {an , bn } are determined by the boundary condition

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

That means the coefficients {an , bn } must be the Fourier coefficients of given function φ(θ) i.e.

1 2π
Z
an = φ(ξ) cos(nξ) dξ n = 0, 1, 2, · · · (4.2.7a)
π 0
1 2π
Z
bn = φ(ξ) sin(nξ) dξ n = 1, 2, 3, · · · . (4.2.7b)
π 0
 49

Example 4.2.2
(
∆u = 0 a < r < b, 0 ≤ θ < 2π
u(a, θ) = 0, u(b, θ) = f (θ).

Solution
This is a problem in the interior of an annulus. The general solution (4.2.4) and boundary condition at
r=a:
+∞
X
Cn an + Dn a−n (An cos(nθ) + Bn sin(nθ)) = 0


u(a, θ) = C0 + D0 ln a +
n=1

lead to

C0 + D0 ln a = 0
Cn an + Dn a−n = 0.
Let C0 = −D0 ln a and Cn an = −Dn a−n . The solution is written as
+∞  
r X  r n  r −n
u(r, θ) = D0 ln + − (an cos(nθ) + bn sin(nθ)) .
a n=1 a a

The boundary condition at r = b leads to

"
+∞  n  −n #
b X b b
u(b, θ) = D0 ln + − (an cos(nθ) + bn sin(nθ)) = f (θ).
a n=1 a a

Thus we have
Z 2π
1 b
D0 = f (ξ) dξ (µ = )
2π ln µ 0 a
Z 2π
1
an = f (ξ) cos(nξ) dξ
π(µn − µ−n ) 0
Z 2π
1
bn = f (ξ) sin(nξ) dξ.
π(µ − µ−n )
n
0
2

Example 4.2.3

 ∆u = 0
 r > a, 0 ≤ θ < 2π
∂u

 = 0, lim (u(r, θ) − V0 r cos θ) = 0.
∂r r=a r→+∞

Solution
Notice that the boundary condition at infinity is inhomogeneous. Let

u(r, θ) = V0 r cos θ + w(r, θ).

Thus the boundary condition of the new unknown function w(r, θ) at infinity is homogeneous. Notice that
r cos θ (= x) is harmonic, which leads to ∆u = ∆w. Then the original problem is transformed to

 ∆w = 0
 r > a, 0 ≤ θ < 2π
∂w

 = −V0 cos θ, lim w(r, θ) = 0.
∂r r=a r→+∞
50

According to the homogeneous boundary condition at infinity, the coefficients D0 and Cn , including C0 ,
in (4.2.4) are zero. Then the solution of this problem should have the form
+∞
X
w(r, θ) = C0 + r−n (An cos(nθ) + Bn sin(nθ)) ,
n=1

and
+∞
∂w X
= − n r−(n+1) (An cos(nθ) + Bn sin(nθ)) .
∂r n=1

According to the boundary condition at r = a, we have

+∞
X
− n a−(n+1) (An cos(nθ) + Bn sin(nθ)) = −V0 cos θ.
n=1

Comparing both sides of this equation we see that all of the coefficients An and Bn are zero except A1 .
The coefficient A1 is determined by −a−2 A1 = −V0 , i.e. A1 = a2 V0 . Thus we have

w(r, θ) = C0 + a2 V0 r−1 cos θ

Finally the solution of the original problem is given by

u(r, θ) = V0 r cos θ + w(r, θ) = V0 r cos θ + a2 V0 r−1 cos θ,

i.e.
a2
 
u(r, θ) = V0 cos θ r + .
r
2

Example 4.2.4
Let us solve the following boundary value problem for a Poisson equation
(
∆u = cos θ 1 ≤ r ≤ 2, 0 ≤ θ < 2π
u|r=1 = 0, u|r=2 = 2.
Solution
Unlike in the previous two examples, u is no longer harmonic on the region and hence the formula (4.2.4)
for the generation solution does not work for this problem. What we can do now is to use the idea used
before for solving inhomogeneous heat equations: for each fixed r ∈ [1, 2], think of u(r, θ) as an 2π-periodic
function of θ, and expand it into a Fourier series involving An cos nθ + Bn sin nθ (n = 0, 1, · · ·):
∞
X
u(r, θ) = A0 (r) + (An (r) cos nθ + Bn (r) sin nθ).
n=1

(The coefficients An and Bn may vary when r varies and hence in general, they are functions of r. ) Now
the formula (4.2.1), the PDE and the above expansion for u imply
∞ 
A00 (r) X A0n (r) n2 An (r) Bn0 (r) n2 Bn (r)

A000 (r) + + 00
(An (r) + − 00
) cos nθ + (Bn (r) + − ) sin nθ = cos θ.
r n=1
r r2 r r2

Comparing both sides, we have that all the An = 0 = Bn , except A0 and A1 which must satisfy
A00 (r) A0 (r) A1 (r)
A000 (r) + = 0; A001 (r) + 1 − = 1.
r r r2
Multiplying the A0 -equation by r, we have

(rA00 )0 = 0 ⇒ rA0 (r) = const. C ⇒ A0 (r) = C ln r + D.

 51

For the A1 -equation, by guessing, we have a particular solution r2 /3; the general solution for the corre-
sponding homogeneous equation
A0 (r) A1 (r)
A001 (r) + 1 − =0
r r2
is given by C1 r + C2 r−1 . Thus for the original inhomogeneous equation

A01 (r) A1 (r)

A001 (r) + − =1
r r2

the general solution is given by C1 r + C2 r−1 + r2 /3. Now we see that u must be in the form of

u(r, θ) = C ln r + D + (C1 r + C2 r−1 + r2 /3) cos θ.

Using the boundary conditions, we are led to


C ln 1 + D + (C1 + C2 + 1/3) cos θ = 1
C ln 2 + D + (C1 + C2 /2 + 4/3) cos θ = 2.

Comparing both sides of each equation, we see that


D=1
C ln 2 + D = 2,

and

C1 + C2 + 1/3 = 0
C1 + C2 /2 + 4/3 = 0.

Solving these simultaneous equations, we have D = 1, C = 1/ ln 2, C1 = −7/9 and C2 = 4/9. Finally, we

obtain
ln r −7r 4 r2
u(r, θ) = 1 + +( + + ) cos θ.
ln 2 9 9r 3
2

Poisson’s formula
Substituting (4.2.7) into (4.2.6), we have

+∞
a0 X n
u(r, θ) = + r (an cos(nθ) + bn sin(nθ))
2 n=1
+∞
1 2π
Z  Z 2π 
X 1
= φ(ξ)dξ + rn φ(ξ) cos(nξ)dξ cos(nθ)
2π 0 n=1
π 0
 Z 2π  
1
+ φ(ξ) sin(nξ)dξ sin(nθ)
π 0
+∞
!
1 2π 1 2π
Z Z X
= φ(ξ)dξ + φ(ξ) rn cos n(ξ − θ) dξ
2π 0 π 0 n=1
Z 2π +∞
!
1 X
= φ(ξ) 1 + 2 rn cos n(ξ − θ) dξ
2π 0 n=1

Notice that for r < 1 one has

52

+∞
X
1+2 rn cos n(ξ − θ)
n=1
+∞
X +∞
X
= 1+ rn ein(ξ−θ) + rn e−in(ξ−θ) ( i2 = −1)
n=1 n=1
+∞ +∞
X
n X
n
= 1+ rei(ξ−θ) + re−i(ξ−θ)
n=1 n=1

e in(ξ−θ)
e−in(ξ−θ)
= 1+ +
1 − rein(ξ−θ) 1 − re−in(ξ−θ)
1 − r2
= .
1 − 2r cos(ξ − θ) + r2
Finally we get the so-called Poisson’s formula or Poisson’s integral for the solution of Laplace equation
in the unit circle:
Z 2π
1 1 − r2
u(r, θ) = φ(ξ) dξ. (4.2.8)
2π 0 1 − 2r cos(ξ − θ) + r2

Let ρ = Rr (a > 0). The above formula leads to the Poisson’s formula for the solution of Laplace equation
in the circle with radius R:
Z 2π
1 R 2 − ρ2
u(ρ, θ) = φ(ξ) 2 dξ, (4.2.9)
2π 0 R − 2Rρ cos(ξ − θ) + ρ2
where 0 ≤ ρ < R, 0 ≤ θ < 2π. This formula has theoretical value and is beautiful; but you will have a
very hard time computing it when, say, φ(θ) = cos θ, while by the separation of variables method, you can
easily get the solution u = r cos θ = x of (4.2.5) with B.C u(1, θ) = cos θ. Thus when solving boundary
value problem (4.2.5), the first thing to try is still the separation of variables method!

4.3 Fundamental solution of Laplace equation

Preparation: Dirac’s δ-function

Let M0 be a fixed point and M an arbitrary point in Rn . Dirac’s δ-function, or simply the δ-function,
centered at a fixed point M0 is defined by the following three properties:

1◦ δ(M ; M0 ) = 0 ∀ M ∈ Rn and M 6= Mo

2◦ δ(M0 ; M0 ) = +∞
Z
3◦ φ(M )δ(M ; M0 )dM = φ(M0 ), ∀ bounded and continuous function φ defined on Rn .
Rn

Observe that δ(M ; M0 ) = δ(M − M0 ; 0). So it makes sense to write δ(M ; M0 ) = δ(M − M0 ).
If we take φ ≡ 1, then Z
δ(M − M0 )dM = 1.
Rn

But since δ is everywhere equal to zero except at M0 , the integral must be zero (recall the value of a
function at one point does not influence the value of its integral). We have a contradiction! Thus, strictly
speaking the δ-function is not a function. In fact, it is a “functional”: it acts on a function φ (as the input)
and the output is a number given by φ(M0 ). This is the point of view taken in advanced mathematics
 53

courses. Here in this course, for the sake of simplicity, we still treat δ-function as a function. This practice,
as was intended by its inventor the physicist Dirac, produces correct results as long as we restrain ourselves
from doing wild things such as squaring the δ-function.
The mysterious δ-function can be approximated by “earthly”, i.e. ordinary functions, if they are concen-
trated at a point to form spikes:

Theorem 4.3.1
Let {fm (x)} be a sequence of functions satisfying that
(i) there exists a constant K such that for every m,
Z
|fm (x)|dx ≤ K;
Rn

(ii) the functions concentrate at a fixed point x0 in the following sense:

Z
lim |fm (x)|dx = 0
m→∞ |x−x0 |≥r

for any fixed r > 0;

(iii) the total “mass”of the functions have a limit in the following sense:
Z
lim fm (x)dx exists and = A.
m→∞ Rn

Then the sequence {fm (x)} converges to Aδ(x − x0 ) in the following weak sense:
Z Z
lim fm (x)φ(x)dx = Aφ(x0 ) = Aφ(x)δ(x − x0 )dx,
m→∞ Rn Rn

for any φ which is bounded and continuous on Rn . In fact, we only need the boundedness φ on Rn and
its continuity at point x0 .
Proof
We just need to show Z
lim fm (x)(φ(x) − φ(x0 ))dx = 0. (4.3.1)
m→∞ Rn

Since φ is continuous at x0 , for any  > 0, there exists r > 0 such that |φ(x) − φ(x0 )| <  if |x − x0 | ≤ r.
Then R R
| |x−x0 |≤r fm (x)(φ(x) − φ(x0 ))dx| ≤ |x−x0 |≤r |fm (x)||φ(x) − φ(x0 )|dx
R
≤ |x−x0 |≤r |fm (x)|dx
R (4.3.2)
≤  Rn |fm (x)|dx
= K.
On the other hand, let L be an upper bound of |φ|. Because of the concentration assumption, there exists
M such that if m ≥ M , we have Z
|fm (x)|dx < .
|x−x0 |>r

Then for such m,

R R
| |x−x0 |>r fm (x)(φ(x) − φ(x0 ))dx| ≤ |x−x0 |>r |fm (x)||φ(x) − φ(x0 )|dx
R
≤ 2L |x−x0 |>r |fm (x)|dx
≤ 2L.
54

Combining this with (4.3.2) we have that for m ≥ M

Z Z Z
fm (x)(φ(x) − φ(x0 ))dx ≤ + ≤ (K + 2L),
Rn |x−x0 |≤r |x−x0 |>r

which implies (4.3.1). This completes the proof of the theorem.

Example 4.3.1
In Section 3.7, the fundamental solution of heat equation

(ξ − ξ0 )2
 
1
G(ξ, t; ξ0 ) = √ exp − (t > 0)
2a πt 4a2 t

is introduced. Property 3◦ there implies conditions (i) and (iii) with x0 = ξ0 . Now for any fixed r > 0,
Z Z
1 −|η|2
G(ξ, t; ξ0 )dξ = √ √ e dη,
|ξ−ξ0 |>r π |η|>r/(2a t)

which converges to zero as t → 0+ . Thus condition (ii) is also satisfied and Theorem 4.33.1 implies that
G(ξ, t; ξ0 ) converges to δ(ξ − ξ0 ) weakly as t → 0+ . Recall this is proved in Appendix 3.2 where the
arguments are similar to the ones in the proof of Theorem 4.3.1.

G(ξ, t; ξ0 )
1 6
√
2a πt

-ξ
ξ0

Fundamental solution
The solution G0 (x) of the following equation:

−∆G0 (x) = δ(x), x ∈ Rn (4.3.3)

is called the fundamental solution, or free Green’s function of Laplace equation. We set out to find a
formula for the fundamental solution. We first present a formal method, which is simple and transparent;
we then present a rigorous treatment which relies on regularization G ( > 0) of singular solutions of
the Laplace equation. In the rigorous approach, (4.3.3) is interpreted so that the fundamental
solution is the limit G0 of a family of smooth functions G as  → 0 satisfying

−∆G (x) → δ(x) weakly.

In fact, our choice of G will satisfy the stronger property

Z
lim+ (−∆G )(x)φ(x)dx = φ(0) for any φ that is bounded on Rn and continuous at 0. (4.3.4)
→0 Rn

The rigorous treatment is a nice application of Theorem 4.3.1 and will make the future proofs (such as
Property 2◦ in Section 4.3) involving the fundamental solution rigorous. Let us consider the 2D case first.

Formal treatment-2D case

 55

Notice that δ(x) is radially symmetric about the origin. Then it is reasonable to seek radially symmetric
solution:
G0 (x) = F (r), r = |x|.
Then we have

d2 F
 
1 dF 1 d dF
∆G0 (x) = + = r .
dr2 r dr r dr dr
(4.3.3) leads to
 
d dF
r = −r δ(x),
dr dr
Z r
dF
r = − r δ(x)dr
dr 0 Z
1 2π r
Z
=− δ(x)r · drdθ
2π Z0 Z 0
1
=− δ(x) dx
2π Z Z|x|<r
1
=− δ(x) dx
2π R2

1
=− .
2π
1
Therefore we have the ODE dF/dr = −1/2πr and its solution F (r) = − ln r +C, where C is a constant.
2π
Since constant C is a trivial solution of Laplace equation ∆u = 0, we simply take C = 0. Finally we get
that
1
G0 (x) = − ln |x|. (4.3.5).
2π

Rigorous treatment-2D case

By finding radially symmetric solutions of Laplace equation, we find that F (r) = C ln r for any constant
C is a solution of the Laplace equation except at the origin. Our aim is to find the value of C so that it
is the fundamental solution interpreted as (4.3.4).
We smooth out the singularity of F at 0 by defining
p C
G (r) = C ln r 2 + 2 = ln(r2 + 2 ) (4.3.6)
2
where  is a positive constant. We compute

∂ 2 G 1 ∂G 22 C
−∆G = −( 2
+ )=− 2 .
∂r r ∂r (r + 2 )2

We want to choose C so that the righthand side converges to δ(x) weakly and (4.3.4) holds. The integral
of the righthand side on R2 is equal to
Z 2π Z ∞
22 C
− 2 2 2
rdrdθ = −2πC,
0 0 (r +  )

so conditions (i) and (iii) in Theorem 4.3.1 are satisfied with A = −2πC. To verify condition (ii), take a
fixed r0 > 0, we estimate Z Z ∞
|22 C| 4π2 |C|
(r 2 +2 )2 dx = (r 2 +2 )2 rdr
|x|>r0 r0 Z ∞
2 1
≤ 4π |C| r 4 rdr
r0
56

which converges to zero as  → 0+ . Now Theorem 4.3.1 implies that

−∆G → −2πCδ(x) weakly as  → 0+ .

Thus C = −1/(2π). By the last sentence in the statement of Theorem 4.3.1, we actually have (4.3.4).

Higher dimensional case

In the 3D case, the fundamental solution is given by
1
G0 (x) = . (4.3.7)
4π|x|

In the nD case with n ≥ 3, it is given by

1
G0 (x) = , (4.3.8)
(n − 2)ωn |x|n−2

where ωn is the surface area of the unit sphere in Rn . We can use either of the two methods presented
above to obtain these formulas; (4.3.4) holds if the regularization is defined by

1
G (x) = . (4.3.8)
(n − 2)ωn (|x|2 + 2 )(n−2)/2

Fundamental solution with singularity at x0

Let x0 be a fixed point in Rn . The fundamental solution with singularity at x0 is

G0 (x − x0 ).
Formally, by the chain rule we have

−∆(G0 (x − x0 )) = −(∆G0 )(x − x0 ) = δ(x − x0 ). (4.3.9).

Rigorously, in (4.3.4) we replace φ(x) by φ(x + x0 ) and then change variable y = x + x0 , we are led to
Z
lim+ (−∆G )(y − x0 )φ(y)dy = φ(x0 ) for any φ that is bounded on Rn and continuous at x0 .
→0 Rn
(4.3.10)
Notice that in the 3D case, the fundamental solution with singularity at x0 is exactly the electric potential
induced by a unit point charge located at x0 (see Section 4.1).

4.4 Green’s identities and applications

We only derive the Green’s identities in 2-D space. But the results are still true for 3-D or higher
dimensional spaces.
Notice that
(vux )x = vx ux + vuxx
and
(vuy )y = vy uy + vuyy .
The addition of the above two identities leads to the identity in the vector form

5 · (v 5 u) = 5v · 5u + v∆u.

Integrating this equation on a bounded region Ω with piecewise smooth boundary ∂Ω, we have
 57

Z Z Z Z Z Z
5 ·(v 5 u)dA = 5 v · 5udA + v∆udA (4.4.1)
Ω Ω Ω

Recall the divergence theorem

Z Z I
∇ · F dA = F · n dS,
Ω ∂Ω

where F is a smooth vector field in Ω and n is the unit outward normal vector field on ∂Ω. Using this in
(4.4.1) we obtain
Z Z Z Z I I
∂u
5 v · 5udA + v∆udA = (v 5 u) · ndS = v dS.
Ω Ω ∂Ω ∂Ω ∂n

The above result is often written in the following form

Z Z Z Z I
∂u
v∆udA = − 5 v · 5udA + v dS , (4.4.2)
Ω Ω ∂Ω ∂n

which is called the Green’s first identity or first Green’s formula. What we are doing here is to shift
the “burden”(the derivative) on u to v; in doing so, the boundary integral is a necessary evil. The reader
may benefit from thinking about the simple case where Ω is an interval [a, b]: by integration by parts, we
have Z b Z Z b b
vu00 dx = vdu0 = − v 0 u0 dx + vu0 (b) − vu0 (a),
a a a

where the last two terms can be regarded as an integral on two boundary points a, b. Because of this, in
PDE literature, (4.4.2) is called integration by parts.
From the Green’s first identity, one can easily get that

Z Z I
∂v ∂u
(u∆v − v∆u)dA = (u −v )dS , (4.4.3)
Ω ∂Ω ∂n ∂n

which is called the Green’s second identity/formula.

We now discuss several applications of Green’s identities.

Properties of harmonic functions

In the sequel, until we say otherwise we assume that u satisfies the Laplace equation

∆u = 0 in a bounded region Ω.

Property 1◦
I
∂u
dS = 0. (4.4.4)
∂Ω ∂n

It can be easily proved by taking v(x, y) = 1 in either Green’s first identity or Green’s second identity.
This property is a necessary condition for a harmonic function and often used to validate the boundary
conditions for Laplace equation.

Property 2◦
58

I
∂u ∂G0
u(M0 ) = (G0 (M ) − u(M ) )dSM , (4.4.5)
∂Ω ∂n ∂n

where M0 is an interior point in Ω and G0 = G0 (M − M0 ) the fundamental solution with singularity at

M0 .
Formal proof

I
∂u ∂G0
(G0 (M ) − u(M ) )dSM
∂Ω ∂n ∂n
Z Z
= (G0 ∆u(M ) − u(M )∆G0 )dAM ( take v = G0 )
Ω
Z Z
= u δ(M − M0 )dAM ( − ∆G0 = δ(M − M0 ) and ∆u = 0 )
Ω

= u(M0 ). ( third property of δ−f unction )

2
This proof is simple and nice. But there is a problem here: Green’s second identity is used, requiring that
u and v be smooth enough (precisely, u and v and their partial derivatives up to order 2 be continuous on
Ω); however G0 (M − M0 ) is singular at M0 . To overcome this difficulty, in the following rigorous proof
we start with replacing G0 by its regularization G (see (4.3.6) and (4.3.8)), then sending  → 0+ .
Rigorous proof
By Green’s second identity, we have
I Z Z
∂u ∂G
(G (M ) − u(M ) )dSM = (G ∆u(M ) − u(M )∆G )dAM
∂Ω ∂n ∂n Z ZΩ (4.4.6)
= (−u(M )∆G )dAM .
Ω

Sending  to zero, the left hand side becomes the one with  replaced by 0 because the the singularity is
away from the boundary; the right hand side can be re-written as the same integral on the entire space
Rn by defining u ≡ 0 outside Ω, then by (4.3.10), the right hand side of (4.4.6) converges to u(M0 ). 2
Remark 1. (4.4.5) is called the boundary representation of harmonic functions. Since for M ∈ ∂Ω,
G0 (M − M0 ) and ∂G∂n (M − M0 ) are infinite smooth functions of M0 ∈ interior of Ω, by the boundary
0

representation any harmonic function is infinitely smooth in the interior of Ω.

Remark 2. If u is not harmonic but satisfies the Poisson equation

−∆u(M ) = f (M ), M ∈ Ω,

then the proofs above also lead to a representation formula

I Z
∂u ∂G0
u(M0 ) = (G0 (M ) − u(M ) )dSM + G0 (M − M0 )f (M ) dAM . (4.4.50 )
∂Ω ∂n ∂n Ω

Property 3◦
I
1
u(M0 ) = u dS (2 − D),
2πR ∂B(M0 ,R)
I (4.4.7)
1
u(M0 ) = u dS (3 − D),
4πR2 ∂B(M0 ,R)
 59

where B(M0 , R) is the ball centered at M0 with radius R that is contained in Ω.

Proof

n
M b 
*


ρ
b
 
R
M0

We start with (4.4.5) with Ω replaced by B(M0 , R). In the 2D case, the fundamental solution is given by
(4.3.5). Let ρ = |M − M0 |. We compute
∂ ∂ 1 1
(ln ρ) = (ln ρ) = = .
∂n ∂B(M0 ,R) ∂ρ ∂B(M0 ,R) ρ ∂B(M0 ,R) R

Then the formula (4.4.5) becomes

I   I I
1 1 ∂u 1 ln R ∂u
u(M0 ) = u − (ln R) dS = u dS − dS
2π ∂B(M0 ,R) R ∂n 2πR ∂B(M0 ,R) 2π ∂B(M0 ,R) ∂n

The second integral is equal to zero according to Property 1◦ . So we have

I
1
u(M0 ) = u dS.
2πR ∂B(M0 ,R)

The proof for 3-D problem is similar. 2

Remark 1. Notice that 2πR is the circumference of the circle ∂B(M0 , R) (2-D) and 4πR2 is the surface
area of the sphere ∂B(M0 , R) (3-D), and so the righthand sides of (4.4.7) are the average of u on the
circle/sphere. Thus (4.4.7) is called the mean value property of harmonic functions. In the 1D case,
this is trivial: any harmonic function is linear and hence its value at the midpoint of an interval is equal
to the average of the function at the end points of the interval.
Remark 2. The mean value property of harmonic functions has another version:
Z
1
u(M0 ) = u dA (2 − D),
πR2 B(M0 ,R)
Z (4.4.8)
1
u(M0 ) = u dV (3 − D),
4πR3 /3 B(M0 ,R)

where the right hand sides are the averages of u on the disk/ball. This can be proved, in the 2D case for
example, by replacing every R by r < R, by multiplying (4.4.7) by 2πr, and then integrating in r:
Z R Z RI
u(M0 )2πr dr = u dS dr.
0 0 ∂B(M0 ,r)

The left hand side is equal to πR2 u(M0 ); the right hand side is equal to
RR
B(M0 ,R)
u dA because the
length element dS on ∂B(M0 , r) times dr is just the area element dA on B(M0 , R).
Remark 3. If u is not harmonic but satisfies the Poisson equation with a source term

−∆u(M ) = f (M ) ≥ 0, M ∈ Ω,
60

then by modifying the proof of (4.4.7) and using (4.4.5’) (the details are supplied at the end of this remark),
we have I
1
u(M0 ) ≥ u dS (2 − D),
2πR ∂B(M0 ,R)
I (4.4.70 )
1
u(M0 ) ≥ u dS (3 − D),
4πR2 ∂B(M0 ,R)
and Z
1
u(M0 ) ≥ u dA (2 − D),
πR2 B(M0 ,R)
Z (4.4.80 )
1
u(M0 ) ≥ u dV (3 − D).
4πR3 /3 B(M0 ,R)

Of course, if f ≤ 0, then we just reverse the direction of the inequality sign “≥”.
Now we give the details of the proof of (4.4.70 ) in the 2D case. By (4.4.50 ) and the proof of Property 3,
we see
I I Z
1 ln R ∂u 1
u(M0 ) = u dS − dS − ln |M − M0 |f (M ) dAM .
2πR ∂B(M0 ,R) 2π ∂B(M0 ,R) ∂n 2π B(M0 ,R)

To rewrite the middle integral on the right hand side, we take in Green’s first identity v = 1 and then use
the PDE for u. Then
I Z Z Z Z
∂u
dS = ∆u(M ) dAM = − f (M ) dAM .
∂B(M0 ,R) ∂n B(M0 ,R) B(M0 ,R)

Thus I Z
1 1
u(M0 ) = u dS + (ln R − ln |M − M0 |)f (M ) dAM .
2πR ∂B(M0 ,R) 2π B(M0 ,R)

Since the integrand in the last integral is nonnegative, we complete the proof of (4.4.70 ) in the 2D case.

Uniqueness of boundary problems for Poisson equation

Let Ω be a bounded region in Rn with piecewise smooth boundary ∂Ω. Consider the Poisson equation

−∆u(x) = f (x), x ∈ Ω. (4.4.9)

As in the case of the heat equation, three kinds of boundary conditions can be prescribed:

u(x) = φ(x), ∀ x ∈ ∂Ω − − − − Dirichlet boundary condition,

∂u
(x) = ψ(x), ∀ x ∈ ∂Ω − − − − Neumann boundary condition,
∂n
∂u
(x) + hu(x) = µ(x), ∀ x ∈ ∂Ω − − − − Robin boundary condition, (h>0)
∂n

Theorem 4.4.1 The solution of Poisson equation (4.4.9) with Dirichlet or Robin boundary condition
is unique; with Neumann boundary condition, the solution is unique up to an additive constant.
Proof.
Since the Poisson equation and the three boundary conditions are all linear, it is only required to prove
that the following problem

∆u = 0 in Ω
homogenuous B.C. on ∂Ω
 61

has unique solution u(x) = 0 for Dirichlet or Robin boundary condition, and u(x) = constant for Neumann
boundary condition.
The PDE ∆u = 0 and Green’s first identity lead to
Z I
∂u
5 u · 5u dV = u dS.
Ω ∂Ω ∂n

∂u
For the homogeneous Dirichlet B.C. (u = 0) or Neumann B.C. ( = 0), we have
∂n
Z I
∂u
5 u · 5u dV = u dS = 0.
Ω ∂Ω ∂n

∂u
For the homogeneous Robin boundary condition ( + hu = 0), we have
∂n
Z I I
∂u
5 u · 5u dΩ = u dS = − hu2 dS ≤ 0.
Ω ∂Ω ∂n ∂Ω

Since h > 0, one must have hu2 ≥ 0. So for all three cases, we have
Z
5 u · 5u dV = 0.
Ω

The integrand is non-negative and continuous, and so

5u ≡ 0, i.e. u = constant.

In the Dirichlet case, u is zero on the boundary and so the constant function must be identically equal
to zero; in the Robin case, since u is constant, the B.C. also implies that u is zero on the boundary and
hence in the interior. 2

Dirichlet Principle
Again, let Ω be bounded region in Rn with piecewise smooth boundary. For any smooth function v defined
on Ω, motivated by electrostatic considerations, we define the energy of v as
Z
1
E(v) = |∇v|2 dV.
2 Ω

Consider the Dirichlet boundary value problem for the Laplace equation
(
∆u = 0 in Ω
(4.4.10)
u=φ on ∂Ω.

Theorem 4.4.2 (Dirichlet Principle) Let u be the unique smooth solution of (4.4.10). Let Γφ be the
set of all smooth functions v defined on Ω satisfying the boundary condition i.e.

v=φ on ∂Ω.

Then u is a minimizer of the energy functional E:

E(v) ≥ E(u) ∀ v ∈ Γφ .

Conversely, any minimizer u of E in Γφ is a solution of the Dirichlet boundary value problem (4.4.10).
Proof
62

Suppose u is a solution of (4.4.10). Let w = v − u. Then w vanishes on the boundary, i.e. w|∂Ω = 0. We
compute
Z Z
1
E(v) = E(u + w) = | 5 (u + w)|2 dV = E(u) + 5 u · 5w dV + E(w).
2 Ω Ω

From Green’s first identity, it follows that

Z I Z
∂u
5 w · 5u dV = w dS − w∆udV = 0.
Ω ∂Ω ∂n Ω

This result leads to

E(v) = E(u) + E(w) ≥ E(u).
This completes the proof of the first part of the theorem.
Now suppose u is a minimizer mentioned in the second part of the theorem. Then automatically u satisfies
the B.C. in (4.4.10) because u belongs to Γφ . It remains to show that u is harmonic. To this end, pick an
arbitrary smooth function w with w ≡ 0 on ∂Ω. Define

f (t) = E(u + tw).

Since u + tw ∈ Γφ , f achieves its minimum at t = 0. Thus

Z
0
0 = f (0) = ∇u · ∇w dV.
Ω

Now we use Green’s first identity to shift the gradient on w to u:

Z I
∂u
0=− w∆u dV + w dS.
Ω ∂Ω ∂n

The boundary integral is zero because w = 0 on the boundary. Hence

Z
w∆u dV = 0, ∀ smooth w with w = 0 on ∂Ω. (4.4.11)
Ω

If there exists a point x0 ∈ Ω where ∆u 6= 0, say, ∆u(x0 ) > 0, then by continuity, ∆u > 0 in a small ball
B(x0 , ). Pick a smooth function w that is positive inside the ball, equal to zero outside it. Then
Z Z
w∆u dV = w∆u dV > 0,
Ω B(x0 ,)

which contradicts (4.4.11). Thus u is harmonic in Ω. 2

Remark
The proof of Theorem 4.4.2 is simple, but this is an important mathematical theorem based on the
physical idea of energy. It is a general principle in physics that any system prefers to approach the state
of lowest energy, called the ground state. Theorem 4.4.2 is a mathematical manifestation of this physical
principle.
Equation (4.4.11) characterizes harmonic functions, that is, it is a necessary and sufficient condition for
u to be harmonic. In advanced PDE courses, it is called the weak formulation of harmonic functions;
any function that satisfies (4.4.11), even those that we do not know their smoothness, is called a weak
solution of the Laplace equation. It can be proved a weak solution is necessarily a smooth solution.

Using fundamental solution to solve Poisson equation in whole space

 63

Consider the Poisson equation in the whole of Rn :

−∆u(x) = f (x), x ∈ Rn , (4.4.12)

where f is a smooth function with bounded support (meaning that f ≡ 0 outside a large ball). Motivated
by the electrostatic consideration in Section 4.1, we conjecture that
Z
u(x) = G0 (x − x0 )f (x0 )dVx0 (4.4.13)
Rn

is a solution.

Formal proof
Recall from (4.3.9)
−∆G0 (x − x0 ) = δ(x − x0 ).
We compute R
−∆u(x) = Rn (−∆G0 )(x − x0 )f (x0 )dVx0
R
= Rn δ(x − x0 )f (x0 )dVx0
R
= Rn δ(x0 − x)f (x0 )dVx0
= f (x).

Rigorous proof
Choose a large L such that outside the ball B(0, L), f = 0. Change variable in (4.4.13) by letting
y = x − x0 . Then Z
u(x) = G0 (y)f (x − y)dVy ,
Rn
R
−∆u(x) = Rn G0 (y)(−∆x ) (f (x − y)) dVy
R
= Rn G0 (y)(−∆y ) (f (x − y)) dVy (by chain rule)
R
= B(x,L+1) G0 (y)(−∆y ) (f (x − y)) dVy (by assumption on f )
R
= lim→0+ B(x,L+1) G (y)(−∆y ) (f (x − y)) dVy
R
= lim→0+ B(x,L+1) (−∆y )G (y)f (x − y)dVy +
 
+ lim→0+ ∂B(x,L+1) ∂G∂n  (y)
f (x − y) − G (y) ∂f (x−y)
H
∂n dS ( Green second formula)
R
= lim→0+ B(x,L+1) (−∆y )G (y)f (x − y)dVy (f terms in boundary integral = 0)
R
= lim→0+ Rn (−∆y )G (y)f (x − y)dVy
= f (x) (by (4.3.10)).

The most physically meaningful case of (4.4.13) is the 3D case:

Z
1
u(x) = f (x0 )dVx0 ,
R3 4π|x − x0 |

where f is the density function of charges continuously distributed in the space, and u is the electric
potential thus induced.

4.5 Maximum-minimum principle

We start with the simplest case: suppose −u00 ≤ 0 on an interval I. Then u is concave up. If u achieves
its maximum value at an interior point of I, then u must be a constant function. In general, we have the
following

Theorem 4.5.1 (Strong maximum principle) Suppose in a connected region Ω function u satisfies

−∆u ≤ 0, (4.5.1)
64

then the maximum function value is taken only on the boundary unless u is a constant function.
Proof.
We just need to show that if u(M0 ) reaches the maximum function value umax and M0 is an interior point
of Ω, one must have

u(M ∗ ) = umax = constant ∀ M ∗ ∈ Ω.

#  Γ -+ 
# c c c c
c
c   M∗
"! M M3 Mk−1
"! M1 2
M0

Since Ω is a connected region, one can find a continuous curve Γ ⊂ Ω connecting M0 and M ∗ . Since M0
is an interior point in Ω, one can find a disk B(M0 , R0 ) ⊂ Ω. According to (4.4.8’), in the 2D case (the
higher dimensional case can be handled in the same way) one has
Z
1
u(M0 ) ≤ u dA.
πR02 B(M0 ,R0 )

But u(M0 ) is the maximum, we then have

u(M ) = umax ∀ M ∈ B(M0 , R0 ).

Denote the intersection of curve Γ and circle ∂B(M0 , R0 ) along the positive direction by M1 . Then we
have M1 ∈ Ω and u(M1 ) = umax . The above argument can be iteratively repeated, leading to a series of
interior points {M0 , M1 , M2 , · · ·} such that

u(M ) = umax ∀M ∈ B(Mi , Ri ) i = 1, 2, 3, · · · .

Eventually we reach a finite number k such that M ∗ falls into the disk B(Mk , Rk ). This completes the
proof. 2

Remark 1. There is another proof of this result: observe that u also satisfies (3.8.2) so the parabolic
strong maximum principle applies, which yields the elliptic strong maximum principle above.
Remark 2. If the inequality in (4.5.1) is reversed, then we have the strong minimum principle. If u
is harmonic, then we have both principles.
Remark 3. The strong maximum principle implies the weak maximum principle:

max u = max u.
Ω ∂Ω

4.6 Method of Green’s function

 65

The goal of this section is to solve the Dirichlet boundary value problem

∆u = 0, x ∈ Ω,
(4.6.1)
u(x) = φ(x), x ∈ ∂Ω,

where for the time being, we assume that Ω is a bounded region in Rn with piecewise smooth boundary.
Recall we have obtained (4.4.5) from which we have:
I
∂u ∂G0
u(x0 ) = (G0 (x − x0 ) (x) − φ(x) (x − x0 ))dSx .
∂Ω ∂n ∂n
∂u
This almost achieves the goal of solving (4.6.1), except that we do not know ∂n on the boundary. To get
rid of this term on the righthand side, for each fixed x0 in the interior of Ω, we find a function ψ(x; x0 )
which is smooth in x on the closure of Ω, satisfying

∆x ψ(x; x0 ) = 0, x ∈ Ω,
ψ(x; x0 ) = −G0 (x − x0 ), x ∈ ∂Ω.

The Green’s function for Ω with Dirichlet boundary condition is defined as

G(x; x0 ) = G0 (x − x0 ) + ψ(x; x0 ). (4.6.2)

It satisfies 
−∆x G(x; x0 ) = δ(x − x0 ), x ∈ Ω,
(4.6.3)
G(x; x0 ) = 0, x ∈ ∂Ω.
The physical interpretation of Green’s function in the 3−dimensional case is that it is the electric
potential function induced by a unit point-charge located at x0 with ∂Ω grounded (so the potential on
the boundary is zero). ψ in (4.6.2) is called the regular part of Green’s function.
Now repeating the proof of (4.4.5) with G0 replaced by G, we have the solution formula for (4.6.1)
I
∂G
u(x0 ) = − φ(x) (x; x0 )dSx . (4.6.4)
∂Ω ∂n

Thus solving the Dirichlet problem (4.6.1) boils down to merely finding the Green’s function, which
sounds very promising. But finding Green’s function is equivalent to finding its regular part; and finding
the regular part ψ is still a Dirichlet boundary problem, though with a special boundary value. This is
the reason why for general regions, it is impossible to find explicit formulas for the Green’s functions; in
fact, only for balls and half-spaces, will we be able to find these in this course. The Green’s function for a
rectangle is, surprisingly, not easy to find and has to be given by a series (thus in this case the preferred
method is still the method of separation of variables).

Green’s function for half-space

The above discussion was carried out for the case of bounded regions Ω; on unbounded regions Green’s
second identity (based on which (4.6.4) is derived) does not hold without additional assumptions on the
decay of the functions at infinity. The right philosophy is not to try to justify rigorously every step that
leads to (4.6.4), but to first find a formula for the Green’s function and then verify rigorously that (4.6.4)
is indeed a solution of (4.6.1).
We now do so for the case of the upper half-space of Rn (n = 2, 3).
(
∆u = 0 in Ω = {x = (x1 , · · · , xn ); xn > 0 }
(4.6.5)
u(x1 , · · · , xn−1 , 0) = φ(x1 , · · · , xn−1 ).
To obtain the Green’s function, we first consider the 3D case so we can use our knowledge of physics.
As mentioned above, Green’s function G(x; x0 ) is the electric potential induced by a unit point charge
66

at x0 in the upper space with the plane x3 = 0 grounded (so the potential is 0 on the plane). To
ground the plane, we simply put a negative unit point charge at the mirror image x∗0 of x0 (so x∗0 =
(x01 , x02 , · · · , x0n−1 , −x0n )). In this fashion, we get the Green’s function of the upper half-space

G(x; x0 ) = G0 (x − x0 ) − G0 (x − x∗0 ). (4.6.6)

Indeed, G(x; x0 ) is identically equal to zero on the plane xn = 0 and is harmonic on the upper half-space
(note the singularity of G(x; x∗0 ) occurs in the lower half-space). And these hold in case of any spatial
dimensions (recall that G0 is a radial function and note that x0 and x∗0 are equidistance from any point
x on the plane xn = 0); the 2D case is illustrated Figure 4.6.1.
x2
6
xb0 = (x01 , x02 )
@
@
@
@
@
@ x
@b - x1

b
x∗0 = (x01 , − x02 )

Figure 4.6.1 Mirror reflected point x∗0 of x0 for the upper half-plane

In the 2D case, the Green’s function is given by

G(x; x0 ) 1
= − 2π ln |x − x0 | + 1
2π ln |x − x∗0 |
1 (x1 −x01 )2 +(x2 +x02 )2
= 4π ln (x1 −x01 )2 +(x2 −x02 )2 ;

in the 3D case, it is given by

1 1
G(x; x0 ) = − .
4π|x − x0 | 4π|x − x∗0 |

∂G
Now let us come back to (4.6.4) by first computing on the plane xn = 0. In the 2D case,
∂n
∂G(x; x0 ) ∂G −x02 1
=− = · . (4.6.7)
∂n x2 =0 ∂x2 x2 =0 π (x1 − x01 )2 + x202
In the 3D case,
∂G(x; x0 ) ∂G x03 1
=− =− ·p 3. (4.6.8)
∂n x3 =0 ∂x3 x3 =0 2π (x1 − x01 )2 + (x2 − x02 )2 + x203

Combining (4.6.4) and (4.6.7) and (4.6.8), for the 2D case we have
y0 +∞
Z
φ(x)
u(x0 , y0 ) = dx, (4.6.9)
π −∞ (x − x0 )2 + y02
for 3D case we have
Z +∞ Z +∞
z0 φ(x, y)
u(x0 , y0 , z0 ) = 3 dxdy. (4.6.10)
2π
p
−∞ −∞ (x − x0 )2 + (y − y0 )2 + z02
 67

These two formulas for the solution of the boundary problem (4.6.5) are called Poisson integrals for the
half-spaces.
We now check rigorously that if φ is bounded and continuous, then (4.6.9) and (4.6.10) indeed are solutions
of the boundary value problem (4.6.5) in 2D and 3D cases, respectively. Since the 3D case is similar to
the 2D case, we shall focus on the latter case. By inspecting the explicit formula for the Green’s function,
we note the following symmetry

G(x; x0 ) = G(x0 ; x), x 6= x0 .

Because of this and the fact that G(x; x0 ) is harmonic in x-variable, it is also harmonic in x0 -variable,
except at x0 = x. Thus for any fixed x = (x, y) on the x-axis,

∂G(x; x0 ) y0
− =
∂y π((x − x0 )2 + y02 )

is harmonic in x0 = (x0 , y0 )-variable in the upper half-plane {y0 > 0}. Now it follows that

Z +∞
y0
∆x0 u(x0 ) = ∆x 0 [ ]φ(x)dx = 0.
−∞ π((x − x0 )2 + y02 )

Now we check that that u given by (4.6.9) satisfies the boundary condition in (4.6.5) in the sense of

lim u(x0 , y0 ) = φ(x0 ), for x0 ∈ R. (4.6.11)

y0 &0+

To this end, we apply Theorem 4.3.1: denote

y0
Γ(x; (x0 , y0 )) = ,
π((x − x0 )2 + y02 )

which is called the Poisson kernel. For every y0 > 0,

R∞ 1 ∞
R 1 x−x0
−∞
Γ(x; (x0 , y0 )) dx = π −∞ z 2 +1 dz, (z = y0 )
1
= π (arctan ∞ − arctan(−∞)) = 1;

and for any fixed δ > 0,

Z Z
1 1
Γ(x; (x0 , y0 )) dx = dz → 0
|x−x0 |≥δ π |z|≥ yδ z2 + 1
0

as y0 & 0. Now (4.6.11) follows from Theorem 4.3.1.

Example 4.6.1


 ∆u = 0
 in Ω = {M = (x, y); − ∞ < x < +∞, y > 0 }

u0 x>0
 u(x, 0) =
 .
0 x<0

Solution
Using the Poisson’s integral, we find
68

Z +∞
y f (ξ)
u(x, y) = dξ
π −∞ (ξ − x)2 + y 2
Z +∞
y u0
= dξ
π 0 (ξ − x)2 + y 2
+∞  
ξ−x
Z
u0 dζ
= set ζ =
π −x/y 1 + ζ2 y
u0
+∞
= tan−1 ζ −x/y
π
  
u0 2 −1 x
= 1 + tan .
2 π y
2

Green’s function for a ball with Dirichlet boundary condition

Let BR (0) be the open ball in Rn with center at 0 and radius R. Recall that given any fixed x0 ∈ BR (0),
the Green’s function G(x; x0 ) for BR (0) that is singular at x0 is given by (4.6.2) and satisfies (4.6.3). The
regular part ψ(x; x0 ) is harmonic in x−variable and on the boundary ∂BR (0), G ≡ 0. Let us start again
with the 3-dimensional case where we can use our knowledge of physics. Imagine a unit positive charge is
put at location x0 , resulting in electric potential G0 (x − x0 ). To ground the boundary of the ball BR (0),
we recall a fact that we learn from high school science: the electric potential induced by a unit positive
charge located at x0 equals that induced by a positive charge of |xR0 | units located at the point x∗0 so that
x0 and x∗0 are symmetric about ∂BR (0). Recall that x∗0 is on the ray starting at the origin and passing
through x0 , satisfying
|x0 ||x∗0 | = R2 , (4.6.12)
and so
R2
x∗0 = x0 . (4.6.13)
|x0 |2
The electric potential induced by a positive charge of R
|x0 | units located at the point x∗0 is given by

R R 1 |x0 |
G0 (x − x∗0 ) = = G0 ( (x − x∗0 )). (4.6.14)
|x0 | |x0 | 4π|x − x∗0 | R

So we take the regular part of the Green’s function to be ψ(x; x0 ) = −G0 ( |xR0 | (x − x∗0 )). It is harmonic in
the ball and it is equal to − |xR0 | G0 (x − x∗0 ) = −G0 (x − x0 ) for x ∈ ∂BR (0). Thus the Green’s function is
given by
|x0 |
G(x; x0 ) = G0 (x − x0 ) − G0 ( (x − x∗0 )). (4.6.15)
R

x
bH
 H
 HH
H
 HH
ψ b Hb
H -
O x0 x∗0
BR (0)
 69

Figure 4.6.2 Reflected point x∗0 of x0 for Dirichlet problem in the ball BR (0)
It turns out that (4.6.15) is also the formula for the Green’s function for the ball in all spatial dimensions
n ≥ 2: By the chain rule, it is easy to check that the regular part in (4.6.15) is harmonic in the ball; it
remains to check that the right-hand side is 0 for x on the boundary of the ball. By (4.6.12) and R = |x|,
we have
|x0 | |x|
= ∗ .
|x| |x0 |
Thus we have two similar triangles ∆Oxx0 ' ∆Ox∗0 x. This in turn implies that

|x − x0 | |x0 |
= , (4.6.16)
|x − x∗0 | |x|

|x0 |
|x − x0 | = |x − x∗0 |.
R
Then because G0 is a radial function, we obtain

|x0 |
G0 (x − x0 ) = G0 ( (x − x∗0 )).
R
This completes the checking that (4.6.15) indeed is the Green’s function for the ball in any spatial dimen-
sions bigger than 1.
Now we use the Green’s function to solve the Dirichlet boundary value problem:

∆u = 0, x ∈ BR (0),
(4.6.17)
u(x) = φ(x), x ∈ ∂BR (0),

According to (4.6.4), we need to compute − ∂G(x;x

∂n
0)
for x ∈ ∂BR (0). Let r = |x − x0 | and ρ = |xR0 | |x − x∗0 |.
Then G(x; x0 ) = G0 (r) − G0 (ρ) (recall G0 is a radial function and so the notation G0 (r) and G0 (ρ) make
sense). So

∂G(x; x0 )
− = −∇G(x; x0 ) · n = (−G00 (r)∇r + G00 (ρ)∇ρ) · n (by chain rule)
∂n  1−n
ρ1−n

r x
= ∇r − ∇ρ · (by (4.3.8))
ωn ωn R
r1−n x
= (∇r − ∇ρ) · (for x ∈ ∂BR (0), r = ρ)
ωn  R
1−n
|x0 | x − x∗0

r x − x0 x
= − ·
ωn  |x − x0 | R |x − x∗0 |  R
r1−n x − x0 |x0 |2 x − x∗0 x
= − · (we use (4.6.16))
ωn |x − x0 | R2 |x − x0 | R
r1−n R2 (x − x0 ) · x − |x0 |2 (x − x∗0 ) · x
=
ωn |x − x0 |R3
1−n 2 2
r R − |x0 |
= (use (4.6.13))
ωn |x − x0 |R
2 2
R − |x0 | 1
= .
ωn R |x − x0 |n

Now we use (4.6.4) to obtain the solution of (4.6.17)

R2 − |x0 |2 φ(x)
Z
u(x0 ) = dS (4.6.18)
∂BR (0) ωn R |x − x0 |n

This is called the Poisson formula for the ball BR (0). In the 2D case (n = 2), this formula is just
(4.2.9) if we express every item in polar coordinates.
70

Appendix 3.1: Laplace Operator

in Polar and Spherical Coordinates
For polar coordinates, we have 
x = r cos θ
(1)
y = r sin θ

Then 
∂ ∂ ∂

 = cos θ + sin θ
∂r ∂x ∂y


 1 ∂ ∂ ∂
= − sin θ + cos θ



r ∂θ ∂x ∂y
i.e.

∂ ∂ 1 ∂

 = cos θ − sin θ
∂x ∂r r ∂θ


(2)
 ∂ ∂ 1 ∂
= sin θ + cos θ



∂y ∂r r ∂θ

Notice that
∂2
  
∂ 1 ∂ ∂ 1 ∂
= cos θ − sin θ cos θ − sin θ
∂x2 ∂r r ∂θ ∂r r ∂θ
∂2 1 ∂2
 
1 ∂
= cos2 θ − cos θ sin θ − +
∂r2 r2 ∂θ r ∂r∂θ
2
∂2
   
1 ∂ ∂ 1 ∂
− sin θ − sin θ + cos θ + 2 sin θ cos θ + sin θ 2 .
r ∂r ∂r∂θ r ∂θ ∂θ

∂2
  
∂ 1 ∂ ∂ 1 ∂
= sin θ + cos θ sin θ + cos θ
∂y 2 ∂r r ∂θ ∂r r ∂θ
∂2 1 ∂2
 
1 ∂
= sin2 θ + sin θ cos θ − +
∂r2 r2 ∂θ r ∂r∂θ
∂2 ∂2
   
1 ∂ 1 ∂
+ cos θ cos θ + sin θ + 2 cos θ − sin θ + cos θ 2 .
r ∂r ∂r∂θ r ∂θ ∂θ

Finally

∂2 ∂2 ∂2 1 ∂ 1 ∂2
∆= + = + + . (3)
∂x2 ∂y 2 ∂r2 r ∂r r2 ∂θ2

For spherical coordinates, let s = r sin θ. Then the 3-D coordinate transformation becomes a pair of 2-D
transformations:
 
x = s cos ψ z = r cos θ
and (4)
y = s sin ψ s = r sin θ .

Thus we have
∂2 ∂2 ∂2 1 ∂ 1 ∂2
+ = + + . (5)
∂x2 ∂y 2 ∂s2 s ∂s s2 ∂ψ 2

∂2 ∂2 ∂2 1 ∂ 1 ∂2
+ = + + . (6)
∂z 2 ∂s2 ∂r2 r ∂r r2 ∂θ2
 71

The formulas (5) and (6) lead to

∂2 ∂2 ∂2 ∂2 1 ∂ 1 ∂2 1 ∂ 1 ∂2
2
+ 2+ 2 = 2+ + 2 2+ + 2 . (7)
∂x ∂y ∂z ∂r r ∂r r ∂θ s ∂s s ∂ψ 2

From the definition


z = r cos θ
s = r sin θ
in (4), instead of (1), and the second relation in (2), one has
∂ ∂ 1 ∂
= sin θ + cos θ
∂s ∂r r ∂θ

Thus we get that

1 ∂ 1 ∂ 1 ∂
= + 2 cos θ .
s ∂s r ∂r r sin θ ∂θ

Substituting into (6), we have

∂2 ∂2 ∂2 ∂2 ∂2 ∂2
 
2 ∂ 1 ∂ 1
+ + = + + 2 sin θ 2 + cos θ + .
∂x2 ∂y 2 ∂z 2 ∂r 2 r ∂r r sin θ ∂θ ∂θ r2 sin2 θ ∂ψ 2

or

∂2 ∂2 ∂2 ∂2 ∂2
 
2 ∂ 1 ∂ ∂ 1
∆= + + = + + 2 sin θ + . (8)
∂x2 ∂y 2 ∂z 2 ∂r 2 r ∂r r sin θ ∂θ ∂θ r2 sin2 θ ∂ψ 2

Assignment 4

1. Derive the free Green’s function G0 (x) in 3-D, i.e. (4.3.7).

2. A spherical shell with inner radius 1 and outer radius 2 has s steady-state temperature distribution,
∂u
u. Its inner boundary is held at 100◦ C. Its outer boundary satisfies = −γ, where γ is a constant.
∂n
(a) Find the temperature u. Hint: everything is radial and hence so is u.
(b) What are the hottest and coldest temperatures?
(c) Can you choose γ so that the temperature on the outer boundary is 20◦ C?

3. Suppose that u is a harmonic function in disk D = {r < 2} and that u = 3 sin(2θ) + 1 for r = 2.
Without finding the solution, answer the following questions:
(a) Find the maximum value of u in D;
(b) Calculate the value of u at the origin.

4. Find the Green’s function G(M ; M0 ) for Dirichlet problem in the first quadrant of plane:
(
−∆G = δ(M − M0 ) in Ω = {M = (x, y); x > 0, y > 0}
G|x=0 = 0, G|y=0 = 0
72

5. Select a suitable method to solve the following boundary value problems:

(1)
(
∆u = 0 in Ω = {M = (r, θ); 0 ≤ r < R, 0 ≤ θ < 2π}
u(R, θ) = A cos θ

(2)
(
∆u = 1 in Ω = {M = (r, θ); 0 ≤ r < R, 0 ≤ θ < 2π}
u(R, θ) = 0
Hint: the region and B.C. are radially symmetric and hence the solution should be radially symmetric.
(3) 
 ∆u = A r2 sin(2θ) in Ω = {M = (r, θ); 0 ≤ r < R, 0 ≤ θ < 2π}
2
u(R, θ) = 0


Hint: For each fixed r, u(r, θ) is 2π−periodic function of θ which can be expanded by the eigenfunctions
with 2π-period B.C.. Thus u takes the form of
∞
X
u(r, θ) = A0 (r) + (An (r) cos(nθ) + Bn (r) sin(nθ)).
n=1

(4)


 ∆u = 0 in Ω = {M = (x, y); 0 < x < π, 0 < y < π}

u(0, y) = 0 u(π, y) = cos2 y



uy (x, 0) = 0 uy (x, π) = 0

(5)


 ∆u = 0 in Ω = {M = (x, y); 0 < x < a, 0 < y < b}


u(0, y) = 0 u(a, y) = 0

 

 ∂u
+u = 0 u(x, b) = g(x)


∂y

y=0

6. Find the solutions that depend only on r of the Helmholtz equation −∆u = λ2 u in 3-D, where
λ > 0 is a constant.

7. Show that there is no solution of

in Ω ⊂ R3

 ∆u = f
 ∂u = g on ∂Ω
∂n
unless
Z I
f dx = g dS.
Ω ∂Ω

8. Let Ω be a bounded domain in Rn with smooth boundary. Consider Poisson equation with Neumann
boundary condition 
−∆GN (x; x0 ) = δ(x − x0 ), x ∈ Ω,
∂GN
∂n (x; x0 ) = const. C, x ∈ ∂Ω,
 73

where n is the unit outer normal of ∂Ω, x0 is a fixed point in Ω. Do this problem formally.
(i) Find the value of const. C such that the above BVP has a solution.
(ii) By using GN , find a formula for u(x0 ), where u is a solution of

−∆u = f (x), x ∈ Ω,
∂u
∂n = g(x), x ∈ ∂Ω.

9. Consider Poisson equation

−∆u = f (x), x ∈ R3 ,
lim u = 0,
|x|→∞

where 
1, if |x| ≤ 1,
f (x) =
0, if |x| > 1.

(i) Solve this equation (leave your answer as an integral).

(ii) Find lim|x|→∞ |x|u(x).
(iii) Let c be the limit found in (ii). Then
c
u(x) ≈ for |x| large .
|x|

Interpret this physically.

10. (Harnack inequality) Let u be a nonnegative harmonic function in Rn . Prove that

sup u ≤ 2n infn u.
Rn R

Hint: take an arbitrary pair of points P and Q. Let R = |P −Q|. Use the mean value property of harmonic
functions on the balls BR (P ) and B2R (Q) (balls centered at P and Q with radius R and 2R, respectively).

11. (Liouville Theorem) Prove that any harmonic function u in the whole Rn that is either bounded
from below or above must be a constant function. Hint: Consider either u − inf u or sup u − u.

12. (Decay rate of harmonic functions) Suppose u is harmonic in the exterior of the ball BR (0) in
R3 such that it decays at infinity:
lim u(x) = 0.
|x|→∞

(i) Define
v(x) = M G0 (x) − u(x),

where G0 is the fundamental solution of Laplace equation, and the constant M is taken large enough such
that v > 0 on ∂BR (0). Prove that v is positive in the exterior of BR (0). Hint: argue by contradiction and
use the strong minimum principle.
(ii) Prove that u decays at infinity at least as fast as the fundamental solution.

Chapter 5 Hyperbolic Equations

74

5.1 Wave equation: string vibration

Physical description of a vibrating string

We consider a soft string of length l, of which the two end points are tied at two pegs located at x = 0
and x = l, respectively. We assume that the motion of the string is transverse: the particles on the
string move only in the direction perpendicular to x−axis. Then it makes sense to use x to represent the
particle that moves on the vertical line passing through x on the x−axis. We use u(x, t) to represent the
position of particle x at time t. (So for a fixed t the graph of u(x, t) is the shape of the string at time t.)
We also assume that the motion of the string is mild, which means the slope of the string is small, i.e ∂u ∂x
is small. This is an idealization that will simplify the PDE for u.
u
6

b b - x
0 l

Derivation of wave equation

Let ρ be the density of the string (unit: mass/length). Since the total length of string

Z lp
1 + u2x (x, t) dx
0

may change as t changes (while the total mass does not change), strictly speaking ρ is a function of t.
But because of the idealization that ∂u
∂x is small, the total length of the string is approximately constant
in t. On the other hand, the string may be inhomogeneous and thus we assume ρ depends on x only. We
introduce the tension T (x) of the string at x, which is the magnitude of the tensile force exerted on the
part of the string on one side of the particle x by the part on the other side. For a string that is “aging
”quickly, T should vary with respect to t appreciably. Thus T may depend on both x and t (however,
as we will see, the transversality of the motion of the string implies that T is independent of x). We
assume that an external force (such as damping force) is exerted on the string in the transverse direction;
to describe this force, we let

external force exerted on piece between x and x + ∆x

F (x, t) = lim
∆x→0 ∆x

so the dimension for F is force/length.

: T (x2 , t)


b 

u(x, t)

T (x1 , t)

x1 x2
Now we take an arbitrary piece of the string between particles x1 and x2 with x1 < x2 . We apply Newton’s
Second Law to this piece. The tensile force exerted to the piece at x2 (by the part of string on the right-
hand side) is tangent to the string and has magnitude T (x2 , t), and so the force is T (x2 , t) times the unit
tangent vector. To get the tangent vector, think of the string at time t being parameterized by the function
 75

x 7→ (x, u(x, t)). Then the unit tangent vector that we are seeking for is
(1, u (x , t))
p x 2 ,
1 + u2x (x2 , t)
and so the tensile force at particle x2 at time t is given by
(1, u (x , t))
p x 2 T (x2 , t).
1 + u2x (x2 , t)
Similarly, the tensile force at particle x1 at time t is given by
(1, ux (x1 , t))
−p T (x1 , t).
1 + u2x (x1 , t)
Notice that the external (non-tensile) force exerted on the piece of the string is given by
Z x2
F (x, t) dx(0, 1).
x1

Since the string moves only transversely, the horizontal components of the tensile forces cancel:
T (x2 , t) T (x1 , t)
p =p .
2
1 + ux (x2 , t) 1 + u2x (x1 , t)
Because of our assumption that ux is small, this leads to, approximately,

T (x2 , t) = T (x1 , t),

i.e, T is independent of x: T = T (t).

On the other hand, applying Newton’s Second Law (ma = F ) in the vertical direction, we have
Z x2 Z x2
T (t)ux (x2 , t) T (t)ux (x1 , t)
utt (x, t)ρ(x) dx = p −p + F (x, t) dx.
x1 1 + u2x (x2 , t) 1 + u2x (x1 , t) x1
√
By the assumption that ux is small, and the Taylor expansion v/ 1 + v 2 = v(1 − 12 v 2 + O(v 4 )) for v
small, we can drop the ux in the denominators on the right hand side. Then we can use the Fundamental
Theorem of Calculus to re-write the resulting equation as
Z x2 Z x2 Z x2
∂T (t)ux (x, t)
utt (x, t)ρ(x) dx = dx + F (x, t) dx.
x1 x1 ∂x x1

Combining all the integrals into one, we have

Z x2  
∂ux (x, t)
utt (x, t)ρ(x) dx − T (t) − F (x, t) dx = 0.
x1 ∂x
If the integrand is piecewise continuous, then because of the arbitrariness of x1 and x2 , the integrand must
be identically equal to 0, namely,

utt (x, t)ρ(x) = T (t)uxx (x, t) + F (x, t), ∀x ∈ (0, l).

Dividing both sides by ρ, we have the wave equation in the standard form

utt = a2 uxx + f (x, t) x ∈ (0, l), t ∈ (−∞, ∞) (5.1.1)

where a2 = T /ρ, f (x, t) = F (x, t)/ρ(x). In the rest of this chapter, for simplicity we assume that T and ρ
are constants and hence a is also a constant. The dimension of a is
s
p mass × length/time2
f orce/density = = length/time.
mass/length
76

So the unit of a is that of speed! Indeed, we will see in the future that a is the speed of wave propagation
along the string.

Initial and boundary conditions

From the view point of physics, the solution u(x, t) of wave equation (5.1.1) cannot be uniquely determined
unless the boundary conditions at terminal points x = 0 and x = l and the following two initial conditions
are provided:

u(x, 0) = f (x) and ut (x, 0) = g(x), (5.1.2)

where f (x) is the initial displacement (i.e. the initial shape of string) and g(x) the initial velocity (i.e.
the initial motion of the string).

The boundary conditions for wave equation may be classified into three types:
First kind B.C. (Dirichlet B.C.)
u(0, t) = µ1 (t); u(l, t) = µ2 (t). (5.1.3)
The physical meaning of the first kind B.C. is that the terminal points are forced to move vertically
according to certain formulas µ1 (t) and µ2 (t), respectively. For the simplest case i.e. u(0, t) = 0, the
terminal point x = 0 is fixed.

Second kind B.C. (Neumann B.C.)

ux (0, t) = µ1 (t); ux (l, t) = µ2 (t). (5.1.4)
Interpretation: a vertical force f1 (t) acts on the left end of the string, creating a tensile force there whose
vertical component has equal magnitude but opposite direction, and so
f1 (t) = −ux (0, t)T (t)
p
where again we use the approximation ux / 1 + u2x ≈ ux as in the derivation of the wave equation.
Dividing both sides of the equation by −T (t) we are led to the Neumann B.C. at x = 0. The B.C at x = l
can be interpreted similarly. In the special case when there are no external vertical forces at the ends of
the string, we have the homogeneous Neumann B.C
ux (0, t) = 0 = ux (l, t).
For this reason, the homogeneous Neumann boundary condition is also called a free boundary condition.

Third kind B.C. (Robin B.C.)

(ux − h u)|x=0 = 0; (ux + h u)|x=l = 0 (h > 0). (5.1.5)

?
u(0, t)
- x
6

The physical meaning of the Robin B.C. is that the left-end point of the string is put on an elastic
foundation, which is modelled by an elastic spring, and so the external force f1 (t) in the discussion of
 77

Neumann B.C. is now given by Hooke’s law: f1 (t) = −ku(0, t), where k > 0 is the spring constant, and
we assume that u = 0 is the equilibrium position of the spring. Now

−ku(0, t) = −ux (0, t)T (t)

k
ux (0, t) − u(0, t) = 0,
T (t)
which is the Robin B.C at x = 0. The Robin B.C at x = l can be interpreted in the same fashion.

Wave equation in higher dimensions

Consider an elastic, flexible membrane stretched over a rigid frame C (a closed curve) that lies in the x-y
plane. Let the equilibrium position of the membrane be on the x-y plane. We assume again that the
motion of the membrane is transverse and mild. We use a function u(x, y, t) to represent the position of
the membrane at time t. The tensile force in the membrane is now defined such that given an infinitesimal
piece of a curve on the membrane, the tensile force exerted on the piece from the part of the membrane
on one side of the curve is tangent to the membrane, perpendicular to the piece and has magnitude T ds0 ,
where T is the tension of the membrane and ds0 is the arc-length of the piece. Thus tension T now has
the unit force/length.
Take an arbitrary region D inside the frame C so that the boundary ∂D is smooth. Let Γt be the curve
on the membrane at time t that corresponds to ∂D: Γt consists of all points (x, y, u(x, y, t)), (x, y) ∈ ∂D.
Parameterize the closed curve ∂D by (x(s), y(s)) where s is the arclength variable so that as s increases,
(x(s), y(s)) traces out ∂D in the counter-clockwise direction. Then ~t = (x0 (s), y 0 (s)) is the unit tangent
vector of ∂D pointing in the counter-clockwise direction, and n = (y 0 (s), −x0 (s)) is the unit outer normal
on ∂D. Recall that the direction of the tensile force on Γt from the membrane outside Γ is tangent to
the membrane and perpendicular to Γt . The direction vector D ~ can be obtained by forming the cross
product of the unit tangent vector of Γt and the upward-pointing unit normal vector of the membrane. Γt
is parameterized by (x(s), y(s), u(x(s), y(s), t)) and so its unit tangent vector is given by

(x0 (s), y 0 (s), ux (x(s), y(s), t)x0 (s) + uy (x(s), y(s), t)y 0 (s)) ~t + (~t · ∇u)~k
T~ = p =q
0 2 0 2 0
(x ) (s) + (y ) (s) + (ux (x(s), y(s), t)x (s) + uy (x(s), y(s), t)y (s)) 0 2
1 + (∇u · ~t)2

The unit normal vector of the membrane at a point of Γt is given by

~ = (−∇u(x(s),
N p
y(s), t), 1)
.
1 + |∇u|2

Now we have
0 ~ ~ − (x0 (s) − ux ∇u · ~t)~j + (y 0 (s)ux − x0 (s)uy )~k
~
D = T~ × N ~ = (y (s) − uy ∇u · t)iq
p
1 + (∇u · ~t)2 1 + |∇u|2
≈ (y 0 (s), −x0 (s), y 0 (s)ux − x0 (s)uy ),

where in the last step we drop all the quadratic terms of ∇u (which are much smaller than the linear
terms of ∇u, by our assumption that the motion of the membrane is mild). So the horizontal component
of the tensile force is
Z Z q Z
0 0 0 ~ 2
(y (s), −x (s))T ds = nT 1 + (∇u · t) ds ≈ nT ds
Γ Z ∂D ∂D
= (Tx , Ty )dxdy (Divergence Theorem) = 0
D

because there is no horizontal motion. Since region D is arbitrary, we have (Tx , Ty ) ≡ 0 and so T depends
78

only on t: T = T (t). On the other hand, the vertical component of the tensile force on Γt is
Z Z q
~ · ~kT ds0 =
D (y 0 (s)ux − x0 (s)uy ))T 1 + (∇u · ~t)2 ds
Γt Z ∂D
≈ ∇u · nT (t)ds
Z ∂D
= T (t)∆u dxdy.
D

Suppose the density function of the membrane is ρ(x, y) and the external (non-tensile) force on the mem-
brane is modelled by F (x, y, t) with unit force/area. Then by Newton Second Law, we obtain
Z Z Z
utt ρ dxdy = T (t)∆u dxdy + F dxdy.
D D D

From the arbitrariness of D, it follows that

utt = a2 ∆u + f (x, y, t) (x, y) ∈ Ω, t > 0 (5.1.6)

where a2 = T /ρ, f = F/ρ.

Wave equations in 3D have the same form. They are satisfied by sound pressure, electric and magnetic
fields, etc.

5.2 Energy and uniqueness

Total energy of vibrating string

Energy is one of the most important concepts in physics. The law of conservation of energy states
that in the absence of external force the total energy E(t) of a dynamic system remains constant in time
t, where
E(t) = K(t) + P (t), (5.2.1)

with K(t) and P (t) being kinetic energy and potential energy of the dynamic system, respectively.
For a vibrating string, at the beginning we do not know how to define its potential energy; but we can
easily define its kinetic energy: for a moving particle with mass m, the kinetic energy of the particle is
1
defined as mv 2 , where v is the velocity of the particle. Thus naturally the kinetic energy of a vibrating
2
string should be given by
Z l Z l
1 1
K(t) = (ρdx)(ut )2 = ρu2t dx. (5.2.2)
0 2 2 0

The conservation of energy E(t) means that

dE dP dK
= + = 0.
dt dt dt
 
dK
Then the potential energy P (t) may be derived from − mathematically. Suppose that there is
dt
no external force applied to the string, that tension T is constant, and that we have the free boundary
condition at the both ends of the string. Observe
Z l Z l Z l Z l
dK d 1 2
= ρut utt dx = ρut (a2 uxx ) dx = − T utx ux dx = − T u dx.
dt 0 0 0 dt 0 2 x
 79

Therefore the potential energy of a vibrating string should be

1 l
Z
P (t) = T u2x dx. (5.2.3)
2 0
The total energy is given by

1 l
Z
E(t) = (ρu2t + T u2x ) dx. (5.2.4)
2 0
Since the multiplication of the energy E by a constant will not change the conservation property, the
mathematical version of the total energy is written in the following form

1 l 2
Z
E(t) = (ut + a2 u2x ) dx. (5.2.5)
2 0

Uniqueness of wave equation

Theorem 5.2.1 Consider the wave equation (5.1.1) with initial condition (5.1.2) and one of the bound-
ary conditions (5.1.3-5). The initial-boundary value problem has at most one solution.
Proof.
Because of the linearity of the wave equation and boundary conditions, it suffices to prove that the following
problem

utt = a2 uxx


 x ∈ (0, l), t ∈ (−∞, ∞)

u(x, 0) = 0, ut (x, 0) = 0 x ∈ (0, l)


t ∈ (−∞, ∞)

homogenuous B.C. x = 0, l,

has unique solution u(x, t) = 0.

Now we consider the following identity
Z l
0= ut · (utt − a2 uxx ) dx
0
Z l
= (ut · utt − a2 ut · uxx ) dx
0
Z l
l
= (ut · utt + a2 uxt · ux ) dx − a2 ut · ux 0
0
Z l
1 d

l
= u2t + a2 u2x dx − a2 ut · ux 0
,
2 dt 0
i.e.

dE l
− a2 ut · ux


0
= 0. (5.2.6)
dt

dE
In the case of either homogeneous Dirichlet or Neumann B.C., the boundary term is 0. Thus = 0,
dt
i.e. the total energy E(t) is a constant (conservation of energy). Now the zero initial conditions lead to
E(t) = E(0) = 0. Hence ux ≡ 0 ≡ ut and u ≡ constant. Now by the initial condition u(x, 0) = 0, we have
u ≡ 0.
In the case of homogeneous Robin boundary condition (5.1.5), one has
l ha2 d 2
a2 ut · ux = −ha2 [ut (l, t) · u(l, t) + ut (0, t) · u(0, t)] = − [u (l, t) + u2 (0, t)]


0 2 dt
80

Substituting this into (5.2.6), one has

!
Z l
1 d k 2 k
u2t a2 u2x 2
(ha2 =


+ dx + [u (l, t) + u (0, t)] = 0. )
2 dt 0 ρ ρ

which leads to Z l
e =1 u2t + a2 u2x
1k 2
[u (l, t) + u2 (0, t)] = constant.


E(t) dx +
2 0 2ρ

The integral above is the total energy of spring and the boundary part is the energy stored in the elastic
foundation. E(t),
e in fact, is the total energy of the whole system: the vibrating string and the elastic
foundation. By the initial conditions, E(0)
e = 0; hence E(t) ≡ 0, from which it follows that u ≡ constant.
Finally the initial condition u(x, 0) = 0 implies u ≡ 0.
2

5.3 Method of separation of variables

The method of separation of variables introduced in Chapters 3 and 4 can be also applied to the hyperbolic
equations. We will not repeat the description for the method, since it is almost the same as that discussed in
Section 3.4-3.6. We select several typical examples and pay more attentions to the physical understanding
of the Fourier series form of the solutions.

Example 5.3.1
A string with length l is fixed at two ends, x = 0 and x = l. The middle point of the string is lifted to
height h and released at time t = 0. Find the subsequent vibration of the string.
Solution
The problem is to solve the following wave equation with boundary and initial conditions:

utt = a2 uxx


 0 < x < l, t ∈ (−∞, ∞)

u(0, t) = 0, u(l, t) = 0 t ∈ (−∞, ∞) (5.3.1)



u(x, 0) = φ(x), ut (x, 0) = 0 0 < x < l,

where
2h l

 l x

 0≤x≤
2
φ(x) =
 2h (l − x) l

≤ x ≤ l.

l 2
We start with a trial solution of the form

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

Then the PDE leads to

X 00 (x) T 00 (t)
= 2 = −λ ( constant ).
X(x) a T (t)
This and the homogeneous Dirichlet boundary condition leads to the eigenvalue problem
( 00
X (x) + λX(x) = 0
X(0) = 0, X(l) = 0.

Based on the result in Section 3.5, the eigenvalues and eigenfunctions are given by
 81

 nπ 2
λn = n = 1, 2, 3, · · ·
l
nπx
Xn = sin n = 1, 2, 3, · · · .
l
Now the ODE for the T −part is
T 00 (t) + a2 λn T (t) = 0.

The general solution of the above ODE is given by

nπa nπa
Tn (t) = an cos t + bn sin t. (5.3.2)
l l

Now let
+∞
X
u(x, t) = Tn (t)Xn (x).
n=1

One should notice that the PDE and the boundary condition are already satisfied. So we only need to
take care of the initial conditions.
By the first initial condition, we have
+∞
X
φ(x) = an Xn (x),
n=1

and so Z l
2 nπx
an = φ(x) sin dx
l 0 l
l
!
Z Z l
2 2 2h nπx 2h nπx
= x sin dx + (l − x) sin dx
l 0 l l l
2
l l

8h nπ
= sin
n2 π 2 2

8h(−1)k

n = 2k − 1, k = 1, 2, · · ·

= (2k − 1)2 π 2
0 n = 2k, k = 1, 2, · · ·


By the second initial condition ut (x, 0) = 0, we have

∞
X nπa
0= bn Xn (x),
n=1
l

from which we have that all the bn are zero. Finally the solution u(x, t) is given by
+∞
X  nπa   nπ 
u(x, t) = an cos t sin x .
n=1
l l
i.e.
+∞
8h(−1)k
   
X (2k − 1)πa (2k − 1)π
u(x, t) = cos t sin x .
(2k − 1)2 π 2 l l
k=1
2
Harmonic vibrations
If the initial conditions in (5.3.1) are changed to general ones, the solution of Tn (t) is still in the form of
(5.3.2). Then the string vibration can be expressed as
82

+∞ 
X nπa nπa   nπ 
u(x, t) = an cos t + bn sin t sin x .
n=1
l l l

The above formula can be written as

+∞
X  nπa   nπ 
u(x, t) = αn cos t + θn sin x . (5.3.3)
n=1
l l

Now let us consider a fixed point, x = x0 . The motion at this position is clearly given by
+∞
X  nπa 
u(x0 , t) = An cos t + θn ,
n=1
l
 nπ 
where An = αn sin x0 . That means the vibration at every particle of the string is a combination of
l
a series of harmonic vibrations with frequency:
nπa
ωn = n = 1, 2, 3, · · · ,
l
with a phase shift θn . The fundamental tone of the string vibration is determined by the lowest frequency
,e.g. ω1 = πa/l in Example 5.3.1. If the length of the string is cut to be one-half, e.g. the mid-point of
the string is fixed by the finger of a violin player, the frequency of fundamental tone is doubled.

Comparison of heat and wave equations, part 1.

• The heat equation, as we have noticed, cannot be solved backward in time; however, for the wave
equation the solution formula (5.3.3) makes as much sense for t > 0 as for t < 0. So the wave
equation can be solved both forward and backward.

• The heat equation enjoys the maximum principle, while the wave equation does not: consider the
initial-boundary value problem again, but with φ(x) = sin(πx/l); then the unique solution is u(x, t) =
cos(πat/l) sin(πx/l) which obviously changes sign, while u(x, 0) ≥ 0 on the interval [0, l].

• The solution of the heat equation with homogeneous Dirichlet boundary condition decays exponen-
tially as t → ∞, while the solution of the wave equation with the same boundary condition oscillates
with its total energy remaining constant.

To close this section, we supply an example which has Neumann boundary condition:

Example 5.3.2

 utt − 4uxx = 0, x ∈ (0, 1), t ∈ R,
ux (0, t) = 0 = ux (1, t), t ∈ R,
u(x, 0) = 0, ut (x, 0) = x, x ∈ (0, 1).


Solution
This initial-boundary value problem models the situation that the two ends of string move freely in the
vertical direction and the string initially lies on the x−axis and is given an upward velocity x at point x.
Our intuition tells us that the string should “float”to infinity and at the meantime vibrate as t ↑ ∞. Let’s
compute and then see if this is the case from the formula for the solution.
 83

Similarly as in Example 5.3.1, we start with the trial solution u(x, t) = X(x)T (t). Then again the PDE
leads to
X 00 (x) T 00 (t)
= = −λ ( constant ).
X(x) 4T (t)
The corresponding eigenvalue problem is

X 00 (x) + λX(x) = 0
(

X 0 (0) = 0, X 0 (1) = 0.

The eigenvalues and eigenfunctions are given by

λ n = n2 π 2 n = 0, 1, 2, · · ·

Xn = cos nπx n = 0, 1, 2, · · · .

The ODE for the T −part is

T 00 (t) + 4λn T (t) = 0,

whose general solution is given by

Tn (t) = an cos 2nπt + bn sin 2nπt, for n = 1, 2, · · · ;

and
T0 (t) = a0 + b0 t.

Now we form
+∞
X
u(x, t) = Tn (t)Xn (x),
n=0
which satisfies the PDE and the boundary condition.
By the initial condition u(x, 0) = 0, we have
+∞
X
0= an Xn (x),
n=0

and so all an , including a0 , are zero. By the second initial condition ut (x, 0) = x, we have
+∞
X
x = b0 + bn 2nπ cos nπx.
n=1

Thus Z 1
1
b0 = xdx = ,
0 2
1
(−1)n − 1
Z
bn 2nπ = 2 x cos nπxdx = .
0 n2 π 2

Now
+∞
t X (−1)n − 1
u(x, t) = + cos 2nπt cos nπx.
2 n=1 2n3 π 3

Now we check to see if this match our physical intuition about the string floating to infinity and oscillating
as t ↑ ∞: observe that the absolute value of the general term in the series is dominated by 1/(n3 π 3 ), and
so the absolute value of the series is dominated by
+∞
X 1
n 3 π3
n=1
84

which, by the “p-test”, is finite. Thus u(x, t) grows in the order of t/2 (the first term in the formula for
u) as t ↑ ∞. On the other hand, the cosine functions in t−variable in the formula make u oscillate when
it ascends to infinity.

5.4 d’Alembert formula and wave propagation

Cauchy problem of wave equation

In this section we study the wave propagation in 1-D infinite domain without boundary. The problem is
stated as following:

utt = a2 uxx
(
x ∈ (−∞, +∞), t ∈ (−∞, ∞)
(5.4.1)
u(x, 0) = φ(x), ut (x, 0) = ψ(x).
Such kind of problems, in which only initial conditions are imposed, are called Cauchy problems. In
Section 3.7, we use the fundamental solution to study the Cauchy problem for the heat equation. Now we
use the so-called d’Alembert formula to study this problem.

d’Alembert’s formula
We re-write the wave equation in the operator form
 
∂ 2 ∂ 2
( ) − (a ) u(x, t) = 0.
∂t ∂x

The operator reminds us the algebraic expression A2 − B 2 which can be factored as (A + B)(A − B).
Factoring the operator in the same fashion we have
∂ ∂ ∂ ∂
( + a )( − a )u(x, t) = 0.
∂t ∂x ∂t ∂x
We are tempted to introduce new independent variables η and ξ such that
∂ ∂ ∂ ∂ ∂ ∂
−a = , +a = .
∂t ∂x ∂η ∂t ∂x ∂ξ
This and the chain rule imply that
∂x ∂t ∂x ∂t
= −a, = 1, = a, = 1.
∂η ∂η ∂ξ ∂ξ
Then it makes sense to choose
x = −aη + aξ, t = η + ξ,
or equivalently
x − at x + at
η=− , ξ= .
2a 2a
Now in the new independent variables, the wave equation becomes

uηξ = 0.

Then the general solution is given by

u = F (η) + G(ξ),
where F and G are arbitrary smooth functions. F depends on η and hence on x − at; similarly, G on
x + at. Thus the general solution of the wave equation is given by

u(x, t) = f (x − at) + g(x + at). (5.4.3)

 85

The functions f (·) and g(·) are determined by the initial conditions. Using the initial conditions, we have
(
f (x) + g(x) = φ(x)
−af 0 (x) + ag 0 (x) = ψ(x)
i.e. 

 f (x) + g(x)
= φ(x)
1 x
Z
 −f (x) + g(x) =
 ψ(s)ds + C,
a 0

where C is a constant. Thus

1 x
 Z
1
f (x) = φ(x) − ψ(s)ds − C2


2 2a 0


1 x
Z
1
ψ(s)ds + C2

 g(x) = φ(x) +


2 2a x0
From this and (5.4.3) it follows that
Z x+at
1 1
u(x, t) = [φ(x − at) + φ(x + at)] + ψ(s)ds, (5.4.4)
2 2a x−at

which is called d’Alembert’s formula. Since the derivation of d’Alembert’s formula is constructive, the
uniqueness of the solution of Cauchy problem (5.4.1) is also proved.
This formula does not apply to the initial-boundary value problem such as (5.3.1) when the interval
(x − at, x + at) is not included in the interval (0, l).

Travelling waves
We recall that the solution u(x, t) of the wave equation is decomposed into two parts

u(x, t) = f (x − at) + g(x + at).

We consider f (·) first. The function f (x−at) is a shift of f (x) to the right with a distance at as illustrated
below.
u
6
 at -
a a

T ravelling-
-x
x0 x = x0 + at

Figure 5.4.1 Travelling wave

Obviously the travelling speed is a. Thus f (x − at) is called a right-travellingpwave with speed a.
Similarly g(x + at) is called a left-travelling wave with speed a. Notice that a = T /ρ. Therefore the
wave propagation is faster, if the tension T is larger or the density ρ is smaller. One can get a better
understanding of the travelling waves from the following example.

Example 5.4.1 (Plucked string)

The initial conditions of Cauchy problem (5.4.1) is given by
  
 h 1 − |x| ∀ |x| ≤ l
ψ(x) = 0, φ(x) = l
0 elsewhere.

86

l 3l l 2l
Sketch the solution of problem (5.4.1) at times: t = 0, , , , and .
2a 4a a a
(This is a “three-finger” pluck, with all three fingers removed from the string at once.)
Solution
u
6
h
@
t = 0 @
@ h/2
@
@ - x
−2l −l 0 l 2l

6
t = l/2a

 H
HH h/2
  HH
 - x
−2l −l 0 l 2l

6
t = 3l/4a
HH h/2

 H H
 H
HH
 H - x
−2l −l 0 l 2l

6
t = l/a
H
 H H
 HH h/2
  HH  HH
 H - x
−2l −l 0 l 2l

6
t = 2l/a
H
 H H H h/2
  HH  HHH -
 H  - x
−2l −l 0 l 2l

Figure 5.4.2 Separation, propagation and interaction of travelling waves in plucked string

At time t = 0, the original wave with amplitude h separates to a pair of waves which start to travel in the
two directions at speed a and at half the original amplitude. These two travelling waves interact within
the period t ∈ (0, l/a). They are completely separated after time t = l/a, since the base length of two
travelling waves is 2l and the relative travelling speed is 2a.

Reflection of waves
d’Alembert’s formula cannot be applied directly to the situation when there is a boundary point. Indeed,
when the infinitely long string is clamped at a point, and when an incident wave reaches this boundary
point, then the string exerts an, say, upward tensile force to the fixed boundary point, which in turn
exerts a downward force to the string with equal magnitude (Newton’s third law), resulting in a reflected
wave. This effect is not included in (5.4.4). However, as we will see in the following example, after some
preparation d’Alembert’s formula can still be used to solve the problem.
 87

utt = a2 uxx


 x ∈ (0, +∞), t ∈ (−∞, ∞)

u(x, 0) = φ(x), ut (x, 0) = ψ(x), x ∈ (0, +∞) (5.4.5)


t ∈ (−∞, ∞).

u(0, t) = 0,

We use the method of reflection, which was used in Section 3.7 to solve the heat conduction problem
on the half-line. Consider the standard Cauchy problem (5.4.1) with the following initial conditions:

utt = a2 uxx


 x ∈ (−∞, +∞), t>0


  
 φ(x) x>0  ψ(x) x>0 (5.4.6)
u(x, 0) = φ̃(x) = 0 x=0 , ut (x, 0) = ψ̃(x) = 0 x=0





−φ(−x) −ψ(−x)
 
x<0 x<0

where φ̃(x) and ψ̃(x) defined in (−∞, +∞) are the odd extension of functions φ(x) and ψ(x) which were
originally defined on the half-line (0, +∞). By Exercise 6, the solution of (5.4.6) is odd in x and hence it
is also the solution of (5.4.5). Then the solution of (5.4.5) is given by

1 x+at
Z
1
u(x, t) = [φ̃(x − at) + φ̃(x + at)] + ψ̃(s)ds.
2 2a x−at
We will compute the solution only for t > 0.
Case I x − at > 0, i.e. x > at
We have φ̃(x − at) = φ(x − at) and ψ̃(s) = ψ(s) for s ≥ x − at. Then the solution keeps its original form
in (5.4.4), i.e.

1 x+at
Z
1
u(x, t) = [φ(x − at) + φ(x + at)] + ψ(s)ds (x > at).
2 2a x−at

Case II x − at < 0, i.e. x < at

Notice that x + at is always positive. Then the solution may be written as
Z 0 Z x+at
1 1 1
u(x, t) = [φ̃(x − at) + φ(x + at)] + ψ̃(s)ds + ψ(s)ds
2 2a x−at 2a 0

Substituting φ̃(ξ) = −φ(−ξ) and ψ̃(ξ) = −ψ(−ξ), where ξ = x − at < 0, into the above formula, we get
that
Z 0 Z x+at
1 1 1
u(x, t) = [−φ(−x + at) + φ(x + at)] + − ψ(−s)ds + ψ(s)ds.
2 2a x−at 2a 0

Notice that
Z 0 Z at−x
− ψ(−s)ds = − ψ(w)dw (set w = −s).
x−at 0

So
at−x x+at
−1
Z Z
1 1 1
u(x, t) = φ(−x + at) − ψ(s)ds + φ(x + at) + ψ(s)ds, (5.4.7)
2 2a 0 2 2a 0

that is,
Z at+x
1 1
u(x, t) = [−φ(at − x) + φ(at + x)] + ψ(s)ds (x < at). (5.4.8)
2 2a at−x
88

If both φ and ψ are positive (on the interval (0, ∞)), then the third and the fourth terms in (5.4.7) form
a positive, left-travelling wave, and the first and the second terms form a negative, right-travelling wave
which is the reflected wave. 2

Domain of dependence
Now let us investigate the solution of (5.4.1) at (x0 , t0 ), which is given by

1 x0 +at0
Z
1
u(x0 , t0 ) = [φ(x0 − at0 ) + φ(x0 + at0 )] + ψ(s)ds,
2 2a x0 −at0
It is clear that the solution at (x0 , t0 ) is determined by the initial information in interval [x0 −at0 , x0 +at0 ].

t
6
(x0 , t0 )
t0 b

• • • - x
x0 − at0 6
x0 x0 + at0 x∗

In daily life, the phenomenon of sound propagation is the simplest example. A person who is standing at
position x0 can only hear at time t0 > 0 the sound produced at the distance at0 at time t = 0, where a
is the speed of sound propagation (a = 344 M/Sec.); the solution value at (x0 , t0 ) has no relation to the
initial “information”at position x∗ (x∗ > x0 + at0 or x∗ < x0 − at0 ), since the initial information is too
far away to reach the position x0 .
We further investigate the line segment M1 M2 as illustrated below.

t
6
(x0 , t0 )
t0 b
M1 = (x1 , t∗ )
t∗ b b • M2 = (x2 , t∗ )
M1 M2 x∗
• • - x
x0 − at0 6
x0 x0 + at0

The solution at any point on M1 M2 , u(x, t∗ ) (x1 < x < x2 ) can be determined by the initial conditions
in (x0 − at0 , x0 + at0 ) and is well-defined. We denote

φ∗ (x) = u(x, t∗ ) and ψ ∗ (x) = ut (x, t∗ ), ∀ x1 < x < x 2 .

With respect to t0 > t∗ , t∗ may also be considered as the “initial time”; then the elapsed time from t∗ to
t0 is t0 − t∗ ≡ τ0 . So by d’Alembert’s formula the solution value u(x0 , t0 ) is given by

1 x0 +aτ0 ∗
Z
1
u(x0 , t0 ) = [φ∗ (x0 − aτ0 ) + φ∗ (x0 + aτ0 )] + ψ (s)ds,
2 2a x0 −aτ0
Since t∗ < t0 is arbitrary, the solution u(x0 , t0 ) can be determined by the information on any line segment
 89

in the triangle. Meanwhile the information in the exterior of the triangle has no effect to the solution
value u(x0 , t0 ). Thus the triangle is called the domain of dependence of point (x0 , t0 ).
t
6
(x0 , t0 )

- x
x0 − at0 x0 + at0

Figure 5.4.2 Domain of dependence at (x0 , t0 ) in phase plane

Now let us consider the domain of dependence for the problem (5.4.5) which takes into account the
reflection of waves from boundary x = 0. For Case I, since x0 − at0 > 0, i.e. t0 < x0 /a, the influence
of boundary reflection has not yet reached x0 . Thus the domain of dependence at (x0 , t0 ) is the same as
illustrated in Figure 5.4.2. For Case II, x0 − at0 < 0. By (5.4.8) the domain of dependence of (x0 , t0 ) is
as illustrated below.

6
(x0 , t0 )

- x
x0 − at0 at0 − x0 x0 + at0

Figure 5.4.3 Domain of dependence at (x0 , t0 ) due to the reflection of waves at boundary x = 0

Domain of influence
Now let us consider the influence of the initial conditions at point x0 . Pick a point (x, t) in upper-half of
the x-t plane and draw its domain of dependence as illustrated below.
t
6
(x, t)

• - x
6 x0 6
x − at x + at

It is clear that the initial conditions at x0 will affect the solution at (x, t), if and only if (x0 , 0) is an interior
point of the bottom of the domain of dependence, i.e. x0 > x − at and x0 < x + at. In other words, the
initial conditions at x0 have influence on the domain

Domain of influence ≡ {(x, t), x < x0 + at and x > x0 − at, t > 0} ,

which is illustrated as below

90

t
6
x + at = x0 x − at = x0
@
@
@
@
@
@• - x
x0
Figure 5.4.4a Domain of influence for initial conditions at x0

For the initial conditions in the interval (x1 , x2 ), the domain of influence is shown as below.
t
6
x + at = x1 x − at = x2
@
@
@
@
@
@• • - x
x1 x2
Figure 5.4.4b Domain of influence for initial conditions in interval (x1 , x2 )

Domain of influence ≡ {(x, t), x < x2 + at and x > x1 − at, t > 0} ,

Thus if at t = 0 there is a localized perturbation to the string near point x = x0 or in the entire interval
(x1 , x2 ), the perturbation will propagate in both directions with speed a.

Characteristic lines
We have seen that the two sets of parallel straight lines in x-t plane
t
x − at = constant
6

- x

Figure 5.4.5a Characteristic lines with positive slope

and
t
6
x + at = constant
@ @ @ @ @ @ @ @ @ @
@ @ @ @ @ @ @ @ @ @
@ @ @ @ @ @ @ @ @ @
@ @ @ @ @ @ @ @ @ @
@ @ @ @ @ @ @ @ @ @
@ @ @ @ @ @ @ @ @ @ - x
 91

Figure 5.4.5b Characteristic lines with negative slope

are extremely important. The domain of dependence and domain of influence are bounded by these lines.
These lines are called characteristic lines of the wave equation.

Comparison of heat and wave equations, part 2. (This is a continuation of the discussion at the
end of Section 5.3.)

• As we have seen in Section 3.7, the heat equation predicts infinite speed of propagation of heat; on
the other hand, we have just found that the speed of propagation vibration waves modelled by the
wave equation is finite.

• Heat equations have a smoothing effect, but wave equations do not enhance the smoothness of initial
conditions (see Example 5.4.1).

Assignment 5

1. Derive the wave equation for a string that moves in a medium in which the resistance force between
the string and the medium is proportional to the velocity of the string.

2. Solve the following initial-boundary value problems for the wave equation
(i)

 utt − 4uxx = 0, x ∈ (0, 1), t ∈ R,
u(0, t) = 0 = u(1, t), t ∈ R,
u(x, 0) = sin(πx), ut (x, 0) = sin(4πx), x ∈ (0, 1).


(ii)
utt − a2 uxx = 0, x ∈ (0, 1), t ∈ R,


ux (0, t) = 0 = ux (1, t), t ∈ R,
u(x, 0) = x, ut (x, 0) = 0, x ∈ (0, 1).


(iii)
utt − a2 uxx = 0, x ∈ (0, 1), t ∈ R,


u(0, t) = 0, ux (1, t) = 1, t ∈ R,
u(x, 0) = 0, ut (x, 0) = cos(πx), x ∈ (0, 1).


3. Solve the Cauchy problem

(
utt − a2 uxx = 0, x ∈ R, t ∈ R,
2
u(x, 0) = e−x , ut (x, 0) = sin(x), x ∈ R.

4. Solve the Cauchy problem

utt − a2 uxx = 0, x ∈ R, t ∈ R,


u(x, 0) = 0, ut (x, 0) = ψ(x), x ∈ R,

where 
1, for |x| < a,
ψ(x) =
0, for |x| ≥ a.
92

Sketch the graph of u vs x at times t = 1/2, 1, 3/2, 2.

5. Let u be a solution of the wave equation

utt = uxx , x ∈ R, t ∈ R.

Is it possible that u(x, 1) is smoother than u(x, 0)? Is it possible to have a maximum principle for the
wave equation?

6. Let u be the solution of

utt − a2 uxx = 0, x ∈ R, t ∈ R,


u(x, 0) = φ(x), ut (x, 0) = ψ(x), x ∈ R.

Show that if both φ and ψ are even functions, then so is u in x. Formulate and prove the analog when
both φ and ψ are odd.

7. Let u be a smooth solution of the wave equation

utt = uxx , x ∈ R, t ≥ 0.

Prove that for any (x0 , t0 ) ∈ (−∞, +∞) × (0, +∞),

Z x0 +t0 −t Z x0 +t0
1 2 1 2
(ut + u2x )(x, t) dx ≤ (ut + u2x )(x, 0) dx, ∀ 0 < t < t0 .
x0 −t0 +t 2 x0 −t0 2
Hint: multiply the wave equation by ut and then trying to re-write the equation in the form

Ft − Gx = 0.

Integrate this equation and apply Green’s Theorem to the trapezoid bounded by the x-axis, the charac-
teristic lines passing through (x0 , t0 ), and the horizontal line t = t.