MATH 219

Fall 2017
Lecture 25
Lecture notes by Özgür Kişisel

Content: Heat Equation: Separation of Variables

Suggested Problems:
§10.1 : 2, 5, 6, 11, 13, 14, 15, 16, 20
In these three lectures, we will use some of the previous techniques together with
some new ones in order to solve a partial differential equation. This will only serve
as a first introduction to the vast subject of partial differential equations.

1 Heat Conduction on a Rod

Suppose that we have a uniform rod of length L, with an ignorable thickness. We
will assume that there is an initial temperature distribution on the rod which evolves
in time. We would like to understand how exactly it evolves. Let us place an axis
labeled by the variable x, parallel to the rod and assume that it stays between x = 0
and x = L. Let u(x, t) denote the temperature of point x at time t. We will assume
that u(0, t) = u(L, t) = 0 for all t. In other words, the two ends are kept at the
constant temperature 0. Heat can escape from the ends, but not through the lateral
surface.
In order to proceed, we need a differential equation which models how the temper-
ature at each point changes with time. This is governed by the heat equation:

ut = α2 uxx ,

where α is a constant that depends on the material that the rod is made up of.
Here is a partial justification of why heat equation holds: Ignoring the boundary
conditions for now, at an equilibrium, the temperature of each point would be equal
to the average temperature of the points neighboring it. In 1 dimension, the only
functions that satisfy this property at each point are linear functions. If uxx < 0

1
at some point p, then the graph of u with respect to x is concave up at p and the
temperature at p is less than the average of the temperatures of its neighboring
points. Therefore, the temperature of the point p will increase (ut > 0) which is in
accordance with the equation. Likewise, if uxx < 0 at p, then the temperature at p is
greater than the average temperature in its neighborhood. In this case u(p) should
decrease, namely ut (p) < 0. This partially justifies the heat equation by showing
that ut and uxx have the same sign. The heat equation says something stronger: ut
is directly proportional to uxx .
Our mathematical problem then is to solve the PDE
ut = α2 uxx
subject to the boundary conditions
u(0, t) = u(L, t) = 0
and an initial condition
u(x, 0) = f (x).
The domain of u(x, t) is the subset 0 ≤ x ≤ L, t ≥ 0 of R2 which has the shape of
an infinite rectangular strip. On the other hand, if we wish to graph the solution
u(x, t), we need one more dimension to place it. So the graph of u(x, t) will be a
surface in R3 over the rectangular strip; the boundary conditions and the initial
condition tell us what happens on the boundary of this surface.

2 Strategy for Solving the Problem

We break the problem into two major steps:

1. Find as many solutions of the problem

ut = α2 uxx , u(0, t) = 0, u(L, t) = 0
as possible.
2. Among these solutions, find one that satisfies u(x, 0) = f (x).

The first part is about separation of variables and solving the resulting boundary
problem. The second part is about finding a Fourier series.

2
3 Separation of Variables
Let us look for solutions of the problem

ut = α2 uxx , u(0, t) = 0, u(L, t) = 0

of the form
u(x, t) = X(x)T (t),
namely as a function of x times a function of t. It might be unclear whether or not
the problem has any interesting solutions of this form. There is the trivial solution
0, but it is unclear whether there are any nontrivial solutions or not. At the end of
our analysis, we will see that there are many. Rewriting the PDE in terms of these
functions, we have

X(x)T 0 (t) = α2 X 00 (x)T (t)

T 0 (t) X 00 (x)
= .
α2 T (t) X(x)
The left hand side of the last equation depends only on t. The right hand side
depends only on x. But they are equal, so this quantity must be a constant.
T 0 (t) X 00 (x)
= = −λ.
α2 T (t) X(x)
This gives us two ordinary differential equations

T 0 (t) + λT (t) = 0
X 00 (x) + λX(x) = 0.

In addition, we have

u(0, t) = X(0)T (t) = 0

u(L, t) = X(L)T (t) = 0.

Then, either T (t) = 0 for all t, or X(0) = X(L) = 0. The first of these would imply
that u(x, t) = X(x)T (t) = 0, so we would have a trivial solution. Therefore we
assume from now on that X(0) = X(L) = 0. Together with the ODE for X(x), we
have the following boundary value problem:

X 00 (x) + λX(x) = 0, X(0) = X(L) = 0.

3
The characteristic equation for the ODE is r2 + λ = 0. Depending on the sign of λ,
we have three possibilities:
(i) λ < 0: Set λ = −k 2 for convenience. Then, r1,2 = ±k are two distinct, real
roots.
X(x) = c1 ekx + c2 e−kx .
The two boundary conditions give us

X(0) = c1 + c2 = 0
X(L) = c1 ekL + c2 e−kL = 0.

In matrix form, these equations can be written as

    
1 1 c1 0
kL −kL = .
e e c2 0

This linear system has a nontrivial solution if and only if the determinant of the
coefficient matrix A is zero. However,

det(A) = e−kL − ekL 6= 0.

Therefore, no nontrivial solutions arise from this case.

(ii)λ = 0: In this case, r1 = r2 = 0. Therefore,

X(x) = c1 + c2 x.

Plugging in the boundary conditions, we get

X(0) = c1 = 0, X(L) = c1 + c2 L = 0

from which we easily get c1 = c2 = 0. Again, no nontrivial solutions arise from this
case.
(iii) λ > 0 Set λ = k 2 . Then, r1,2 = ±ik are two complex conjugate roots. Therefore,

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

The boundary conditions give

X(0) = c1 = 0
X(L) = c1 cos(kL) + c2 sin(kL) = 0.

4
Write these equations in matrix form:
    
1 0 c1 0
=
cos(kL) sin(kL) c2 0.

Again, let A be the coefficient matrix. The system has nontrivial solutions if and
only if the following equivalent conditions hold:

det(A) = 0 ⇔ sin(kL) = 0
⇔ kL = nπ, n ∈ Z
nπ
⇔ k= , n∈Z
L
n2 π 2
⇔ λ = 2 n ∈ Z.
L
For each such value of λ, the nontrivial solutions that we get are constant multiples
of  nπx 
Xn (x) = sin .
L
It is enough to take n ∈ {1, 2, 3, . . .} since n = 0 gives a trivial solution and −n
gives the negative of the solution for n.
Next, let us solve the ODE for T (t) for the special values of λ above.

n2 π 2 α 2
T 0 (t) + T (t) = 0.
L2
Select a solution,
n 2 π 2 α2
Tn (t) = e− L2
t
.
Set un (x, t) = Xn (x)Tn (t). We have one solution for each n = 1, 2, 3, . . .
n 2 π 2 α2
 nπx 
un (x, t) = e− L2 t sin .
L

4 Superposition
In the previous section, we produced infinitely many solutions to the problem

ut = α2 uxx , u(0, t) = u(L, t) = 0.

5
We can still produce more solutions by using the principle of superposition: If u and
v satisfy all three conditions above, then it is easy to check that c1 u + c2 v satisfies
the same conditions for any choice of constants c1 , c2 . We can also try to use the
same principle for an infinite linear combination of the form
∞
X
u(x, t) = cn un (x, t)
n=1
∞  nπx 
X n 2 π 2 α2
= cn e− L2
t
sin .
n=1
L

Everything works out fine, except that there is a convergence issue. An infinite
series may converge or diverge, depending on how fast the coefficients cn grow with
n. In this case, the negative exponentials in the sum decay vary rapidly to 0. So
if the coefficients cn do not grow very quickly with n, then the sum converges and
principle of superposition holds. This gives us a very large family of solutions.