version of 21 Nov 1997

IX. PDEs in space †

Until now we have studied partial differential equations in one space dimension, x. Now
we are ready to consider some problems in two or three space dimensions. Fortunately,
the technique of separation of variables in more dimensions presents only a few new
conceptual issues, and the technical complications are quite manageable..

In three space dimensions the wave equation has the form:

∂ 2u = c 2 ∇⋅∇u := c 2 ∂2 u + ∂2 u + ∂2 u .
∂t2 ∂x2 ∂y2 ∂z 2

It describes many sorts of waves that move through space, such as acoustic waves in air or
another fluid medium, where c is the speed of sound in that medium. It also describes any
component of an electromagnetic field in free space if c is the speed of light (see the
derivation from Clerk Maxwell's equations in the appendix). The expression on the right is
c2 times the Laplacean of u, which is abbreviated ∇ 2u (some texts prefer the notation ∆u).
It is sometimes convenient to think of ∇ 2u as the Laplace operator ∇ 2 applied to u.

The heat equation in more than one space dimension likewise involves the Laplace operator:

∂u = k ∇ 2u.
∂t

Both these equations describe how a quantity will change in time - but it might be static. If
an electromagnetic field or a temperature is at equilibrium, the time derivatives will be 0,
and we are led to the potential equation of Laplace,

† Copyright © 1994, 1997 by Evans M. Harrell II and James V. Herod.

126 Linear Methods of Applied Mathematics

∇2 u := ∂ u2 + ∂ u2 =0,
2 2
(9.1)
∂x ∂y
or, in three dimensions,

∇2 u := ∂ u2 + ∂ u2 + ∂ u2 =0.
2 2 2
(9.2)
∂x ∂y ∂z

Laplace's equation arises in many other applications as well. It is the equation which
describes the
• displacement of an elastic membrane
• electrostatic potential function
• equilibrium concentration of a suspensate in a still fluid
and still further physical quantities.

Solutions of Laplace's equation are called harmonic functions.

Let's begin with the two-dimensional Laplace equation on a rectangle

0<x<a
0 < y < b.

This could arise from the wave equation for a long, rectangular wave-guide, assuming that
the electric field is independent of z and static, that is, independent of t; or it might also
arise from heat flow in a rectangular solid at equilibrium. In either case we are left with an
equation in only two independent variables. It should thus be no more complicated than the
earlier equations in the variables t and x.

Laplace's equation is the prototype of an elliptic equation, with different qualitative

properties from either hyperbolic equations (like the wave equation) or parabolic equations
(like the heat equation). The auxiliary conditions that are imposed are quite different from
those for the wave equation. Specifically, we are normally given boundary conditions on
the entire boundary, and are not specially concerned with an "initial" value of either variable
x or y. Dirichlet boundary conditions for this problem are typically nonhomogeneous, of
the form:

u(0,y) = f1(y) (9.3)

 PDEs in space 127

u(a,y) = f 2(y) (9.4)

u(x,0) = f3(x) (9.5)
u(x,b) = f4(x) (9.6)

where the four boundary functions are supposed given. In electromagnetic theory, for
example, they are determined if the charge is measured on the boundary. In
thermodynamics, they are the temperature determined by the temperature at the edge of the
rectangle.

Separation of variables works much as it did for the heat equation and the wave equation.
As with our earlier equations, we begin with the ansatz that u is a product solution
u(x,y) = X(x) Y(y)
and substitute into Laplace's equation; it is again convenient to divide the result by u. We
use first information from the partial differential equation, then the information from
homogeneous boundary conditions, and lastly information from inhomogeneous boundary
conditions, which are treated much like the initial conditions for the earlier PDE's.

Rather than tackling four non-homogeneous boundary conditions (9.3)-(9.6) all at once,
we begin by setting three of the four boundary functions to 0. For definiteness, we also
make a specific choice of the fourth while developing the ideas:

Model Problem IX.1. Let us solve Laplace's equation with boundary conditions which
are homogeneous on three of the four sides of a rectangle:

u(0,y) = 0 (9.7)
u(1,y) = 1 (9.8)
u(x,0) = 0 (9.9)
u(x,2) = 0. (9.10)
128 Linear Methods of Applied Mathematics

Solution.

Separation begins much as before, but note that boundary condition (9.8) is not
homogeneous, so it is not consistent with the superposition principle. If we add two
functions satisfying (9.8), for example, we will get a function with the boundary value 2,
not 1, when x = a. As in our earlier solutions by separating variables, let us guess that the
solution is a product, u(x,y) = X(x) Y(y) and subsitute into Laplace's equation (9.1).
Dividing through by X(x) Y(y), we find that

X''(x) Y''(x)
X(x) + Y(x) = 0.

Evidently, if the y-independent quantity X''/X equals the x-independent quantity -Y''/Y,
both must be a constant, and we have the familiar ordinary differential equations,
X'' = µ X
-Y'' = µ Y
(notice the difference of sign).

A good rule of thumb in this subject is to deal with the most homogeneous boundaries first.
It is the variable y which has two homogeneous boundary conditions in this case. It
satisfies essentially the same eigenvalue problem as we have seen in previous chapters:

-Y'' = µ Y
Y(0) = 0, Y(2) = 0

At this stage we can probably recognize the eigenfunctions and eigenvalues as

Yn(y) = sin nπ y , µn = nπ 2.
2 2

This same constant µn enters in to the X equation, but with the other sign. The general
solution for X is therefore a linear combination of exp(n π x/2) and exp(- n π x/2), rather
than sines and cosines. There is a better choice for the basis of this solution space, though,
namely
Xn(x) = C1 cosh(n π x/2) + C2 sinh(n π x/2).

One homogeneous boundary condition applies to this function, forcing

 PDEs in space 129

X n(0) = 0.

The Dirichlet boundary condition thus eliminates the hyperbolic cosine, allowing us to
normalize the function Xn as

Xn(x) = sinh(n π x/2).

For the general solution we get:

∞ ∞
u(x, y) = Σ c X (x) Yk(y) = n=Σ1 c n sin(n π y / 2)sinh(n π
k= 1 k k
x / 2).

There is only one boundary condition left, the nonhomogeneous (9.8) at x = 1. Before
incorporating it, we make a general linear combination of all the product solutions which
are consistent with what we know so far:
∞
u(x, y) = Σ
n=1
cn sin(n π y / 2) sinh(n π x / 2).

This is a Fourier sine series in the variable y with the complication that the coefficients have
an extra factor,

bn = c n sinh(π n x/2).
The series must be matched to the series for the function f2(y) = 1, the coefficients in which
can be calculated with the familiar formula. If
∞
Σ b sin(n π y / 2) = 1,
n =1 n

then the Fourier coefficients are easily calculated as:

4 if n is odd, and otherwise 0.

b n = nπ

To find cn, now divide by sinh(n π/2). The solution becomes

∞
4 sin(n π y / 2) sinh(n π x / 2) .
u(x, y) = Σ
n= 1 nπ sinh(n π/ 2)
nodd
130 Linear Methods of Applied Mathematics

---------------------------------------------------
It wasn't critical that the function at x=1 was just the constant function 1; given any other
function of y at x=1 as a Fourier sine series,
∞
u(1, y)= f 2(y) = Σ
n=1
b 2,n sin(n π y / 2),
the solution we get would still be of the form:
∞ sinh(n π x / 2)
u(x, y) = Σ
n= 1
b 2,n sin(n π y / 2) sinh(n π/ 2) .

Now let us return to the full problem with boundary conditions (9.3) -(9.6). Don't start
over and rederive everything! Use the solution we already obtained as a springboard, as
follows.

Let us rename the solution we just obtained for the simplified boundary conditions

u(0,y) = 0
u(a,y) = f 2(y)
u(x,0) = 0
u(x,b) = 0,

with only the function f2 different from 0; for future bookkeeping u2 will be better than u:

∞
sinh(n π x / 2)
u2(x, y) = Σ b sin(n π
n = 1 2,n
y/ 2)
sinh(n a π/ 2)
.

(Notice the a in the denominator, which was fixed as 1 in the model problem.)

Next, suppose that we had the boundary conditions:

u(0,y) = 0
u(a,y) = 0
u(x,0) = 0
u(x,b) = f4(x).

The only differences here are that f2 becomes f4, a becomes b, and the variables x and y are
switched. Thus:
 PDEs in space 131

∞ sinh(n π y / 2)
u4(x, y) = Σ b sin(n π
n = 1 4,n
x/ 2)
sinh(n b π/ 2)
. (9.11)

As the next piece of the puzzle, suppose that

u(0,y) = f1(y)
u(a,y) = 0
u(x,0) = 0
u(x,b) = 0.

We can get these boundary conditions from our original ones with f2 by interchanging x
and a-x. Notice that this does not affect the potential equation, because under this change of
variable there are two compensating changes of sign in ∂2 u/ ∂ x2; each time you use the
chain rule there is a factor of -1. The answer has to be:

∞
sinh(n π (a–x) /2)
u1(x, y) = Σ b sin(n π
n = 1 1,n
y/ 2)
sinh(n a π/ 2)
. (9.12)

Similarly, if
u(0,y) = 0
u(a,y) = 0
u(x,0) = f3(x)
u(x,b) = 0,
then
∞ sinh(n π (b–y) / 2)
u3(x, y) = Σ b sin(n π
n = 1 3,n
x/ 2)
sinh(n b π/ 2)
. (9.13)

Because of the principle of superposition, the entire solution of the boundary-value

problem with boundary conditions (9.3)- (9.6) is the sum:

u(x,y) = u1(x,y) + u2(x,y) + u3(x,y) + u4(x,y).

------------------------------------------------------------

Let us turn our attention now to a fully multidimensional problem. There will be few new
concepts, though as we shall see the additional dimensions require some extra book-
keeping with several indices to label the pieces of the solution correlating with the various
dimensions. We illustrate the topic in a specific example.
132 Linear Methods of Applied Mathematics

Model Problem IX.2. How a mathematician cooks a cube steak. Consider how to cook
a steak which is one meter on a side (it is a whale steak) under the following conditions.
BC: The steak will be put into a preheated oven at temperature 200 C at time t=0.
IC: At time t=0, the steak is pulled directly from the freezer (u(0, x,y,z) = 0) and put into
the oven for four hours.

We wish to find the temperature throughout the interior of the steak at that time. Since the
mathematician shops at a terrible meat market (but doesn't really notice), we will assume
that the steak has the thermal properties of wood, so that in the heat equation,
k = 2500 cm2/hr.

Solution.

Instead of directly solving for the temperature, let u(t,x,y,z) be the temperature minus 200
C, in order to have homogeneous Dirichlet boundary conditions. With this change, for all
t > 0 the conditions on the six faces of the cube are

u(t,0,y,z) = 0
u(t,x,0,z) = 0
u(t,x,y,0) = 0
u(t,100,y,z) = 0
u(t,x,100,z) = 0
u(t,x,y,100) = 0, (9.14)

while the PDE is the usual heat equation (HE),

ut := ∂u
∂t = k ∇ u,
2
(HE)

and the initial condition becomes

u(0,x,y,z) = -200

Solution.

Step 1. Construct a general solution by separation of variables.

 PDEs in space 133

Suppose as usual that a solution of the heat equation is a product of the form

u(t,x) = T(t) Q(x);

a new feature here is the vector variable x. Plugging into the heat equation and dividing by
T Q, we find:
T'(t) ∇ 2Q(x)
= k
T(t) Q(x) .
We have a multidimensional eigenvalue problem to solve, viz.,
–∇2 Q(x) = µ Q(x),
subject to the Dirichlet boundary conditions (9.14). I inserted the minus sign, as before,
because experience teaches me to expect µ to be positive this way. (It is a matter of book-
keeping and not essential.)

How do we solve this equation? Why, by separating variables again, of course! Let

Q(x) = X(x) Y(y) Z(z)

and evaluate
∇ 2Q(x)
µ = Q(x) .

The result, after cancelling common factors is

X″(x) Y″(y) Z″(z)
µ= + + . (9.15)
X(x) Y(y) Z(z)
which is a very curious equation, since any one of the three terms on the right could be
isolated on one side of the equation. For instance, if we solve for X''/X, we find

X''(x)/X(x) = (something independent of x).

Similarly for Y''/Y and Z''/Z. The conclusion is that we have three one-dimensional
eigenvalueproblems, all precisely like the familiar eigenvalue problem from Chapter 6:

-X'' = µ1 X
-Y'' = µ2 Y
-Z'' = µ3 Z
134 Linear Methods of Applied Mathematics

If we compare with (9.15), we see that

µ = µ1 + µ2 + µ3,

and by this stage we can immediately see that µ1 and X are of the form

µ1 = (m1 π/100)2,
X(x) = sin(m1 π x /100), m1 = 1, 2, ....

Likewise,

µ2 = (m2 π/100)2, Y(y) = sin(m2 π y /100), m2 = 1, 2, ..., and

µ3 = (m3 π/100)2, Z(z) = sin(m3 π z /100), m3 = 1, 2, ....

The product solutions thus look like:

π 2 m 2 +m 2+m 2 kt sin m 1πx sin m 2πy sin m 3πz .

um 1m2m3 t,x,y,z = exp – 10000 1 2 3 100 100 100

The general solution is a triple sum:

∞ ∞ ∞
u t,x,y,z = Σ Σ
m1 = 1 m2 = 1 m3 = 1
Σ cm 1m2m3u m1m2 m3 t,x,y,z ,

where the multi-indexed constants cm1,m2,m3 can have arbitrary values, subject only to
convergence of the series.

Step 2.

The constants will be determined by the initial conditions when we set t=0, obtaining a
triple Fourier sine series. In this model problem the coefficients need to satisfy

∞ ∞ ∞
u 0,x,y,z =–200 = Σ Σ
m1 = 1 m2 = 1 m3 = 1
Σ c m1m2m3 um1m 2m3 0,x,y,z ,
m1πx m 2πy m 3πz
= Σ1
Σ mΣ c m m m sin
m m 2 3 1 2 3 100 sin 100 sin 100 .
 PDEs in space 135

There are two ways to proceed here. It turns out that the products of three sines compose a
complete orthogonal set for the cube, so we could directly evaluate the coefficients by the
use of projections in function space. In this case the inner product uses a thre-dimensional
integral over the cube.

You may find it more congenial, however, to rely on the usual, one-variable Fourier sine
series, as follows. Imagine for the moment that y and z are fixed. What remains is a
function of x, and we could use the orthogonality of the functions sin(m1 π x/L) to remove
the sum over m1:

100
n 1πx 100
n1πx m 1πx
u(0,x,y,z) sin 100 dx= Σ Σ2 mΣ3 c m1m2m3 0 sin 100 sin 100 dx ×
0 m1 m
m πy m πz
sin 2 sin 3
100 100
m2πy m 3πz
= Σ Σ2 mΣ3 c m1 m2m350 δ n1msin 100 sin 100
m1 m
m2 πy m3πz
= Σ Σ c 50 sin sin
m 2 m3 n m m
1 2 3 100 100 .

The point here is that sin(m1 π x /100) and sin(n1 π x /100) are orthogonal unless m1=n1,
so only one term survives from the sum over m1.

If we now multiply by sin(n2 π y/100) and integrate from 0 to 100, which will eliminate all
but one term in the sum over m2:

100 100
n1πx n 2πy m3 πz
0 0
u(0,x,y,z) sin 100 sin 100 Σ3 c n1n 2m3sin 100
dxdy = 502 m .

Finally, let us multiply the result by sin(n3 π z /100) and integrate over z:

100 100 100

n 1πx n 2πy n3πz
u(0,x,y,z) sin 100 sin 100 sin 100 dx dy dz = 503c n1 n2n 3.
0 0 0

The result of the calculation is

136 Linear Methods of Applied Mathematics

n1 n n3
8 1 –(–1) 1 – (–1) 2 1 – (–1)
c n1n 2n3 = 3 .
n1n2 n3π

Notice that this is 0 unless all three integers nk are odd, in which case it is 64/(n1 n2 n3) π 3.
------------------------------------------------------------

Exercises IX

IX.1. Derive solutions (9.11)-(9.13) carefully.

IX.2. Find the solution of Laplace's equation with mixed Dirichlet and Neumann
boundary conditions :

u(0,y) = f1(y)
u(a,y) = f 2(y)
∂ u(x,0) /∂ y = f3(x)
∂ u(x,b) /∂ y = f4(x)

Hint: Solve four simpler problems, each of which has three of these functions equal to 0.
Then sum the result.

IX.3. Find the solution of Laplace's equation with the more specific mixed Dirichlet and
Neumann boundary conditions :
u(0,y) = y
u(1,y) = 0
∂ u(x,0) / ∂ y = 0
∂ u(x,2) /∂ y = 1

IX.4. Find the specific solution of Model Problem IX.2 by regarding the functions
sin(n1 π x /100) sin(n2 π y /100) sin(n3 π z /100)
directly as an orthonormal set on the cube.

IX.5. Change the boundary conditions of Model Problem IX.2 so that two of the faces of
the cube are insulated, i.e., Neumann boundary conditions. Discuss the differences
depending on whether the two insulated faces are adjacent or opposite.
 PDEs in space 137

IX.6. Consider a rubber cube of side 1 firmly encased in a metal box of the same
dimensions. The vertical displacement of the cube satisfies the wave equation with c=1000,
and the box enforces zero Dirichlet boundary conditions analogout to (9.14) on all six
faces (but at coordinates 0 and 1, not 0 and 100).

a) Find the normal modes of vibration.

b) Solve the initial-boundary-value problem with initial conditions:

u(0,x,y,z) = xyz sin(π x)sin(π y)sin(2 π z)

ut(0,x,y,z) = sin(2 π x).

Appendix

Clerk Maxwell codified the laws of electricity and magnetism in four partial differential
equations, which in in rationalized mks units read:

∇×H (≡Curl H)= ∂D

∂t + J (extension of Ampère's law)

∇×E = – ∂B (Faraday's law of induction)

∂t
∇⋅ B (≡Div B)=0 (Gauß's law for magnetism, in the absence of monopoles)
∇⋅ D = ρ (Gauß-Coulomb law)

Here, H is the magnetic force and B = µ H is the magnetic induction; E is the electric
intensity and D = ε E is the electric induction; J is the current density and ρ is the charge
density. In free space, ρ = 0 and J = 0, while the value of µ ε turns out to equal 1/c2,
where c is the speed of light.

Let us calculate the second time derivative of B:

∂ 2B = –∇× ∂Ε = – 1 ∇× ∂D = – 1 ∇× ∇× B .
∂t2 ∂t ε ∂t εµ

There happens to be a vector identity for the curl of a curl:

∇× ∇× v =–∇2 v + ∇ ∇⋅ v ,
where the Laplacian of a vector v = (v1, v 2,v 3) is defined by its action
on each component:
138 Linear Methods of Applied Mathematics

∇2 v := ∇2v1, ∇ 2v2 , ∇2v3 .

Because of the absence of magnetic monopoles, the vector identity as applied
to B becomes
∇× ∇× B =– ∇2 B,

which means that

∂ 2B = 1 ∇2 B.
∂t2 εµ

Clerk Maxwell recognized that this equation was the wave equation in three dimensions,
for each component of the magnetic field B. A very similar analysis shows that each
component of E satisfies the same wave equation. Of course, Clerk Maxwell's equations
imply that the electric and magnetic fields in a wave are coupled; if we assume a traveling
wave, the fields E and B turn out to oscillate exactly out of phase.

After deducing the possibility of these electromagnetic waves, Clerk Maxwell was
interested in studying the properties of these waves to see whether they could be
experimentally verified. Imagine his excitement in 1854(?) when he calculated the
characteristic speed of the waves with the best known values of µ and ε for a vacuum and
discovered that it was quite close to the observed value of the speed of light. This was
surely one of the crowning moments in the history of physics. In Clerk Maxwell's own
words,

...we can scarcely avoid the inference that light consists in the transverse
undulations of the same medium which is the cause of electric and
magnetic phenomena.

(W.D. Niven, ed., The Scientific Papers of J. Clerk Maxwell, Cambridge, 1890, vol. I, p.
500).