45 Separation of Variables

which is zero because each term in the last row is zero, since {fj } is a fundamental set of
solutions of the homogeneous equation L[y] = 0. Hence

L[yp ] = −f (2.81)

Consider the case where {fj } is the fundamental set of solutions of L[y] = 0: they satisfy the
ICs
The Kronecker delta δij is 1 (n−1)
fj (x0 ) = δij
for i = j and 0 for i 6= j.
Then from Theorems 2.7 and 2.12 it follows that
n Z x ∼
X −f (t) W (x, t)
y(x) = αj fj (x) + dt (2.82)
j=1 x0 a0 (t) W (t)

is the unique solution of the ODE L[y] = −f with ICs y (j−1) (x0 ) = αj , j = 1, 2, . . . , n, where
αj are constants.

Example 2.16. Using the one-sided Green’s function find a particular solution of
d3 y dy
L[y] = 3
− = y (3) − y (1) = x (2.83)
dx dx
From eq. 2.72, a particular solution of L[y] = −f is given by
∼
x
−f (t) W (x, t)
Z
yp = dt
x0 a0 (t) W (t)
where W (x) is the Wronskian determinant corresponding to a fundamental solution set {fj }
of the homogeneous solution L[y] = 0. First we therefore find the fundamental set solving
y (3) − y (1) = 0. Using y = emx , the equation becomes m3 − m = 0 which gives m = 0, 1, −1.
Hence f1 = 1, f2 = ex , and f3 = e−x . The Wronskian determinant is nonzero and confirms
that this is a fundamental set:
   
 f1 (x)
 f2 (x) f3 (x)  1 ex e−x 
W (x) = f1(1) (x) f2(1) (x) (x) 
f3  = 0 ex −e−x  = 2
  
 (2)
f (x) f (2) (x) f (2) (x) 0 ex e−x 
  
1 2 3

Also,
   
∼
 f1 (t)
 f2 (t) f3 (t)  1 et e−t 
W (x, t) = f1(1) (t) f2(1) (t) f3(1) (t) = 0 et −e−t  = −2 + et−x + e−t+x
   
ex e−x 
   
 f1 (x) f2 (x) f3 (x)  0

The particular solution then is

Z x
t 1
yp (x) = (−2+et−x +e−t+x )dt = [−2+(x20 −x2 )+(1−x0 )ex0 −x +(1+x0 )ex−x0 ] (2.84)
x0 2 2

2.3 Separation of Variables

Introduction

Consider the transient conduction equation for T (x, t). The PDE is
 
∂ ∂T ∂
k = (Cp ρT )
∂x ∂x ∂t
 46 Differential Equations

with t > 0 and x ∈ (0, L). If the properties are constant this becomes

∂2T 1 ∂T
2
= (2.85)
∂x α ∂t
Let the BCs be
T (0, t) = Ta = T (L, t) (2.86)

An IC could be
T (x, 0) = T0 (x), x ∈ (0, L) (2.87)

The BCs are nonhomogeneous; we change variables and make them homogeneous.

y = T − Ta (2.88)

∂2y 1 ∂y
= (2.89)
∂x2 α ∂t

y(0, t) = 0 = y(L, t) (2.90)

y(x, 0) = T0 (x) − Ta = y0 (x) (2.91)

We can use the method of separation of variables:

y(x, t) = X(x)Θ(t) (2.92)

After some manipulation, we get

∂2y 1 ∂y d2 X 1 1 dΘ
= ⇒ = = −λ
∂x2 α ∂t X dx2 α Θ dt

λ is a constant. d2 X
= −λX, x ∈ (0, L) (2.93)
dx2

dΘ
= −αλΘ, t>0 (2.94)
dt
The BCs become
X(0)Θ(t) = 0 = X(L)Θ(t), t>0

Θ(t) 6= 0 for all t > 0 . Hence

Else y = 0 for any arbitrary
X(0) = 0 = X(L) (2.95)
y0 (x).
The IC implies
X(x)Θ(0) = y0 (x) (2.96)

The BVP for X us

Such eigenvalue problems d2 X
which are easy to solve will L[X] = = −λX (2.97)
dx2
turn out to involve a self-
adjoint differential operator X(0) = 0 = X(L) (2.98)
L[ ] along with self-adjoint
The trivial solution is X(x) = 0, but this causes y(x, t) = 0 which is unacceptable. Nontrivial
homogeneous BCs.
solutions exist for specific values of the constant λ. This is similar to the matrix eigenvalue
problem: Ax = λx.
 47 Separation of Variables

Example 2.17. Consider the eigenvalue problem

d2 y
L[y] = = −λy (2.99)
dx2
y(0) = 0 = y(L) (2.100)

The general solution is

√ √
y = c1 sin λx + c2 cos λx (2.101)

At x = 0, c2 = 0 and hence
√
y = c1 sin λx (2.102)
√ √
At y(L) = 0, c1 sin λL = 0 which implies either c1 = 0 or sin λL = 0. Therefore
c1 = 0 results in y = 0 which √
sin λL = 0 (2.103)
is a trivial solution.
√ n2 π 2
⇒ λL = ±nπ, n = 0, 1, 2, . . . ⇒ λ= , n = 1, 2, . . .
L2
n = 0 sets up a trivial solution again:
n2 π 2
λ = λn = (2.104)
L2
There are infinite eigenvalues and the corresponding solutions (eigenfunctions) are
p  nπx 
yn (x) = sin λn x = sin (2.105)
L
Can λ be complex instead of real as assumed above in Eq. 2.104? If sin z is to be zero for
The eigenvalues are real, as complex z, let z = a + ib with b 6= 0. Then sin(a + ib) = 0 which on expansion gives
proven by contradiction.
sin a cos(ib) + cos a sin(ib) = 0

But cos(ib) = cosh b and sin(ib) = i sinh b and hence

sin a cosh b + i cos a sinh b = 0

Both real and complex parts need to be zero and hence

sin a cosh b = 0 cos a sinh b = 0

For b 6= 0, sinh(b) and cosh(b) are both > 0. Hence a is a real number which satisfies sin a = 0
and cos a = 0 which is not simultaneously possible. Hence the eigenvalues are real.

Finding separable solutions

For the heat equation above, the process involved assuming a product form for the solution
(y(x, t) = X(x)Θ(t)) and then using some relevant algebra to arrive at the form
If there are more than two
Function of X only = Function of Θ only = constant
variables, it may be better
to separate out one variable Usually you have to divide by the product (X(x)Θ(t)) to achieve this. If there are constants,
at a a time. it is more convenient to take it to the Θ (first order ODE) side. In this case the constant is
written as −λ where λ is real.

1. For λ < 0, (Aλ and Bλ are arbitrary constants)

h √ √ i
yλ (x, t) = Aλ e λx + Bλ e− λx e−αλt (2.106)

2. For λ = 0, (A0 and B0 are arbitrary constants)

yλ (x, t) = A0 x + B0 (2.107)
 48 Differential Equations

3. For λ > 0,
h √ √ i
yλ (x, t) = Aλ cos( λx) + Bλ sin(− λx) e−αλt (2.108)

Superposition and the role of BCs and ICs

The heat PDE has the set of separable solutions listed in Eqs. 2.106 - 2.108. We need to next
use BCs to identify the set of separable functions {yi (x, t)} that satisfy the PDE and the
BCs.
X
y(x, t) = cj yj (x, t) (2.109)
j

If all the yj (x, t) satisfy the PDE and the given BCs, it may seem that the superposition
principle generates y(x, t) as a solution as well. This need not be so, for example when the
cj s have been identified from the ICs: as an example, if all yj (x, t) satisfy y(0, t) = 10 (for
t > 0), then
X X
y(0, t) = cj yj (0, t) = 10 cj
j j
P
and then y(0, t) itself can be a solution only if cj = 1 which may not be permitted by the
ICs.
Note that if the BC had been y(0, t) = 0 for t > 0, then there would have been no contradiction,
P
with y(0, t) = j cj yj (0, t) = 0. Such BCs, which facilitate the superposition of solutions are
called ‘suitable’ BCs. The suitable BCs that we are likely to see are
If yj (x, t) satisfy the PDE
and a suitable BC, so 1. Homogeneous/regular BCs: At x = a
will a linear combination, 
∂y 
P αy|x=a + β  =0 (2.110)
j cj yj (x, t), of them. ∂x x=a

where α and β are constants, with at at least one of them being nonzero. Typical cases of
this are:

∂y 
y(a, t) = 0 and =0
∂x x=a

2. Boundedness BCs. A solution should not ”blow up” at x = a:

| y|x=a | < ∞ (2.111)

For example conditions at r = 0 in cylindrical and spherical coordinates.

3. Periodic BCs: Behavior at two different points a and b must match. Examples are (typically
in polar/spherical coordinates)
 
∂y  ∂y 
αy|x=a = αy|x=b and = (2.112)
∂x x=a ∂x x=b

For the problem stated in 2.99 and 2.100, if you revisit Eqs. 2.106 - 2.108, you should be able
to show that for λ < 0 and λ = 0, we end up with the trivial solution yλ = 0. For λ > 0, the
solution is 2.105.

2.4 Inner products and Orthogonality

The set of all functions f (x), g(x), . . . on the interval α ≤ x ≤ β forms a linear vector space.