Chapter10: Fourier Transform Solutions of PDEs

In this chapter we show how the method of separation of variables may be extended to solve PDEs defined on
an infinite or semi-infinite spatial domain. Several new concepts such as the Fourier integral representation
and Fourier transform of a function are introduced as an extension of the Fourier series representation to
an infinite domain.

We consider the heat equation

u 2u
= k 2 , < x < (1)
t x
with the initial condition
u(x, 0) = f (x), < x < (2)
As physical conditions at we impose

u(, t) = 0, u(, t) = 0, t > 0 (3)

Using separation of variables,

u(x, t) = (x)h(t) (4)
we obtain the differential equations
dh
= kh h(t) = cekt (5)
dt
d2
= (6)
dx2
First problem we face is to impose boundary conditions at for (x).

Q: Show that equation (6) with the boundary conditions () = () = 0 has no solutions.

Therefore, the correct boundary condition for (x) at x = is different from the boundary condition for
u(x, t) at x = . We require that (x) is bounded at x =

|()| < (7)

Q: Show that problem (6)-(7) has no solutions if < 0.

If > 0 the solution of the problem (6),(7) is

(x) = c1 cos x + c2 sin x (8)

which remains bounded at x = for all > 0. The major difference from the finite domain problem is
that in this case all > 0 are eigenvalues!

Q: Show that = 0 is an eigenvalue of the problem (6),(7) and the corresponding eigenfunction is (x) c.

The set of eigenvalues for a problem is usually referred to as the spectrum. For the heat equation on a finite
domain we have a discrete spectrum n = (n/L)2 , whereas for the heat equation defined on < x <
we have a continuous spectrum 0.

Superposition principle. From (5) and (8) we obtain the product solutions u(x, t) = sin xekt and
u(x, t) = cos xekt for all 0. The general solution is obtained using the principle of superposition:
Z
u(x, t) = [c1 () cos xekt + c2 () sin xekt ]d (9)
0

1
where c1 () and c2 () are arbitrary functions of .

Q: Show that (9) is a solution of the equation (1) for any c1 () and c2 ().

If we let = 2 then (9) becomes

Z
2 2
u(x, t) = [A() cos x ek t + B() sin x ek t ]d (10)
0

where A() = 2c1 ( 2 ), B() = 2c2 ( 2 ) are arbitrary functions.

To satisfy the initial condition (2) we must have

Z
f (x) = [A() cos x + B() sin x]d (11)
0

The problem is now to determine the functions A() and B() such that (11) is satisfied.

Complex exponentials

Complex exponentials may be used to express the sin and cos functions (Eulers formulas):

ei + ei ei ei
cos = , sin = (12)
2 2i
With these formulas, we may write (10) as
Z
eix + eix eix eix k2 t
u(x, t) = [A() + B() ]e d
0 2 2i
Z
eix + eix eix eix k2 t
= [A() iB() ]e d
2 2
Z0 Z
A() iB() ix k2 t A() + iB() ix k2 t
= e e d + e e d
0 2 0 2
Z 0 Z
A() iB() ix k2 t A() + iB() ix k2 t
= e e d + e e d
2 0 2

such that we obtain Z

2
u(x, t) = c()eix ek t d (13)

where the function c() : R C (takes a real argument and returns a complex value) is defined as

A()+iB()
2 , >0
c() = (14)
A()iB()
2 , <0

Remark: Notice that from (14) it follows that c() = c().

At t = 0 we obtain from (13) Z
f (x) = c()eix d (15)

Q: Show that (15) with c() given by (14) is equivalent to (11).

To determine the coefficients c() from (15) we need to introduce a couple of new concepts: Fourier trans-
form and Fourier integral representation of a function.

2
Fourier transform pair

Given a piecewise smooth function f (x) defined on L x L, the Fourier series representation if f is

f (x+) + f (x) X nx X nx
= a0 + an cos + bn sin (16)
2 n=1
L n=1
L

where Z L
1
a0 = f (x)dx
2L L
Z L
1 nx
an = f (x) cos dx
L L L
Z L
1 nx
bn = f (x) sin dx
L L L
The complex form of (16) is (see Section 3.6 for a proof):

f (x+) + f (x) X
= cn einx/L (17)
2 n=

where the the coefficients cn are complex numbers given by

Z L
1
cn = f (x)einx/L dx (18)
2L L

By replacing (18) in (17) we obtain the Fourier series identity

" Z L #
f (x+) + f (x) X 1
= f (x)einx/L
dx einx/L (19)
2 n=
2L L

For functions defined for < x < the following Fourier integral identity is valid
Z Z 
f (x+) + f (x) 1
= f (x)e dx eix d
ix
(20)
2 2
You may think about (20) as a limit of (19) as L .

Definition: Given a function f (x), < x < , we define the Fourier transform of f(x) as the function
Z

F () = f (x)eix dx (21)
2

where is an arbitrary constant (6= 0).

From (20) and (21) it follows that
Z
f (x+) + f (x) 1
= F ()eix d (22)
2

Relation (22) is called the Fourier integral representation of f (x) and f (x) determined by (22) is called
the inverse Fourier transform of F (). If f is continuous we simply have

1
Z
f (x) = F ()eix d (23)

3
 R
The relations (21) and (22) are valid for any function f (x) that satisfies |f (x)|dx < and are also
known as the Fourier transform pair. In our applications we will let = 1.

Next we mention several properties of the Fourier transform.

1. The Fourier transform is a linear operator:

F[c1 f (x) + c2 g(x)] = c1 F[f (x)] + c2 F[g(x)] (24)

where F[f (x)] = F () denotes the Fourier transform of f (x).

2. Given a real valued function f (x) we have

F () = F () (25)

where F () denotes the complex conjugate of F ().

3. If f (x) is an odd function then F () is an odd function. In addition, in this case we have

1 2i
Z Z
f (x) = F ()(cos x i sin x)dx = F () sin x d
0

Z Z

F () = f (x)eix dx = f (x)(cos x + i sin x) dx
2 2
i i
Z Z
= f (x) sin x dx = f (x) sin x dx
2 0

If we choose, for convenience, = 2i, then we obtain

Z
f (x) = F () sin x d (26)
0
Z
2
F () = f (x) sin x dx (27)
0

Relations (26) and (27) are called the Fourier sine transform pair and are valid if f (x) is an odd function.
F () S[f (x)] is called the Fourier sine transform of f (x) and f (x) S 1 [F ()] is called the inverse
Fourier sine transform of F ().

4. Similarly, if f (x) is an even function then F () is an even function and we obtain the Fourier cosine
transform pair Z
f (x) = F () cos xd (28)
0
Z
2
F () = f (x) cos x dx (29)
0

In this case F () C[f (x)] is called the Fourier cosine transform of f (x) and f (x) C 1 [F ()] is called
the inverse Fourier cosine transform of F ().

5. The Fourier transform of a Gaussian is a Gaussian and the inverse Fourier transform of a Gaussian is a
Gaussian
2 1 2
f (x) = ex F () = e 4 (30)
4

4
 r
x2 2
f (x) = e 4 F () = e (31)

6. Convolution Theorem: If F () and G() are the Fourier transforms of f (x) and g(x), respectively, then
Z
1
h(x) = g(x)f (x x)dx (32)
2

is the inverse Fourier transform of the product F ()G(). The function h(x) defined in (32) is called the
convolution of the functions f and g and is denoted h = f g. Notice that f g = g f .

Fourier transform and the heat equation

We return now to the solution of the heat equation on an infinite interval and show how to use Fourier
transforms to obtain u(x, t). From (15) it follows that c() is the Fourier transform of the initial temperature
distribution f (x):
Z
1
c() = f (x)eix dx (33)
2
such that by replacing (33) in (13) we may express the solution u(x, t) as
Z  Z 
1 2
u(x, t) = f (x)e dx eix ek t d
ix
(34)
2

which may be written in the equivalent form

Z Z 
1 2
u(x, t) = f (x) ek t ei(xx) d dx (35)
2

Notice that Z
2
g(x) = ek t eix d (36)

2
is the inverse Fourier transform of ek t . With this notation, the solution (35) becomes
Z
1
u(x, t) = f (x)g(x x)dx (37)
2

The only problem now is to obtain an explicit formula for g(x) defined by (36).

Inverse Fourier Transform of a Gaussian

Functions of the form 2

G() = e
where > 0 is a constant are usually referred to as Gaussian functions. The function g(x) whose Fourier
transform is G() is given by the inverse Fourier transform formula
Z Z
2
ix
g(x) = G()e d = e eix d (38)

The last integral in (38) may be evaluated (the proof is not trivial, see the Appendix to 10.3 for details) to
obtain
r
x2
g(x) = e 4 (39)

5
showing that the inverse Fourier transform of a Gaussian is itself a Gaussian.

If we let = kt in (38), (39) we obtain

Z r
2 x2
ek t eix d = e 4kt
kt

such that replacing in (36) it follows that

r
x2
g(x) = e 4kt (40)
kt

The solution u(x, t) is obtained by replacing (40) in (37)

Z r
1 (xx)2
u(x, t) = f (x) e 4kt dx (41)
2 kt

which may be written as Z

1 (xx)2
u(x, t) = f (x) e 4kt dx (42)
4kt
Remark: Notice that if the initial condition is specified as a Dirac impulse concentrated at the origin,
f (x) = (x), then from (42) we get
1 x2
u(x, t) = e 4kt (43)
4kt
which is called the fundamental solution of the heat equation.

Transform of the derivatives: the following properties hold:

1. The Fourier transform of a time derivative equals the time derivative of the Fourier transform
 
u
F = U (, t) (44)
t t

2. In general, the Fourier transform of the nth derivative of a function u(x, t) with respect to x equals
(i)n times the Fourier transform of u(x, t), if u(x, t) 0, sufficiently fast as x . In particular,
 2 
u
F = (i)2 U (, t) (45)
x2
Q: Prove (44) and (45).

Fourier transforming the heat equation

Next we show how the Fourier transform may be used to solve directly the heat equation on an infinite
interval

u 2u
= k 2 , < x < (46)
t x
with the initial condition
u(x, 0) = f (x), < x < (47)
u(, t) = 0, u(, t) = 0, t > 0 (48)

6
Since the Fourier transform is a linear operator, by applying the Fourier transform to the equation (46) we
obtain    2 
u u
F = kF (49)
t x2
The Fourier transform of u(x, t) is
Z
1
F[u] U (, t) = u(x, t)eix dx (50)
2

If we are able to find F[u] then the solution u(x, t) is given by the inverse Fourier transform formula
Z
u(x, t) = U (, t)eix d (51)

After replacing (44) and (45) in (49) we obtain an ordinary differential equation for the Fourier transform
U (, t)

U (, t) = k 2 U (, t) (52)
t
Therefore, the Fourier transform operation converts a linear PDE with constant coefficients into an ODE!

The general solution of (52) is

2
U (, t) = c()ek t
(53)
To determine c() we use the initial condition (47) and require that U (, 0) is the Fourier transform of f (x).
Then Z
1
c() = f (x)eix dx (54)
2
After replacing (53) in (51) we obtain the solution u(x, t)
Z
2
u(x, t) = c()eix ek t d (55)

which may be written after replacing (54) and rearranging the terms
Z Z 
1 k 2 t i(xx)
u(x, t) = f (x) e e d dx (56)
2

This is the same formula we obtained by using the separation of variables.

Heat equation on semi-infinite intervals

We consider the heat equation on a semi-infinite interval x > 0

u 2u
= k 2, 0 < x < (57)
t x
with the initial condition
u(x, 0) = f (x), 0 < x < (58)
u(0, t) = 0, u(, t) = 0, t > 0 (59)
Using separation of variables,

u(x, t) = (x)h(t)
we obtain
dh
= kh
dt

7
and the boundary value problem
d2
=
dx2
(0) = 0
|()| <
whose eigenvalues are all positive and the eigenfunctions are

(x) = c sin x = c sin x

where > 0. Therefore, the product solutions are
2
u(x, t) = A sin xek t

and using the principle of superposition we obtain

Z
2
u(x, t) = A() sin xek t d (60)
0

To satisfy the initial condition we require

Z
f (x) = A() sin x d
0

If we consider the odd extension f (x) of f (x) to < x < then

Z
f (x) = f (x) = F () sin x d, x > 0
0

where F () is the Fourier sine transform of f (x)

2 2
Z Z
F () = f (x) sin x dx = f (x) sin x dx
0 0

Therefore, A() is given by the Fourier sine transform of f (x)

2
Z
A() = f (x) sin x dx (61)
0

8
Fourier sine and cosine transforms of derivatives

We defined the Fourier sine and cosine transforms of a function f (x) as

2
Z
S[f (x)] = f (x) sin x dx (62)
0
2
Z
C[f (x)] = f (x) cos x dx (63)
0
For the Fourier sine and cosine transforms of derivatives, assuming that limx f (x) = 0, we have the
following formulas:
 
df 2
C = f (0) + S[f ] (64)
dx
 
df
S = C[f ] (65)
dx
 2   
d f 2 df df 2 df
C 2
= (0) + S = (0) 2 C[f ] (66)
dx dx dx dx
 2   
d f df 2
S = C = f (0) 2 S[f ] (67)
dx2 dx
Q: Prove formulas (64-67).

Formulas (66) and (67) suggest the appropriate choice of the Fourier sine/cosine transform when solving
PDEs involving second order derivatives on a semi-infinite interval x 0. Replace f (x) by u(x, t) in the
formulas above: if u(0, t) is given we will use a Fourier sine transform; if u/x(0, t) is given we will use a
Fourier cosine transform.

Example: Consider the heat equation on 0 < x < with nonhomogeneous boundary condition at x = 0
u 2u
= k 2, 0 < x < (68)
t x
u(x, 0) = f (x), 0 < x < (69)
u(0, t) = g(t), u(, t) = 0, t > 0 (70)
We introduce the Fourier sine transform of u(x, t)
2
Z
U (, t) = u(x, t) sin x dx (71)
0
By applying the Fourier sine transform to equation (68) and using the property (67) we obtain a first order
linear ODE  
U 2 2
=k g(t) U (72)
t
We require that U (, 0) must be the Fourier sine transform of u(x, 0) = f (x) such that
2
Z
U (, 0) = f (x) sin x dx (73)
0
By solving (72-73) we obtain U (, t). The solution u(x, t) is given then by the inverse Fourier sine transform
Z
u(x, t) = U (, t) sin x d (74)
0

9
Worked examples using transforms

One-dimensional wave equation on an infinite interval

Consider the one-dimensional wave equation

2u 2u
2
= c2 2 , < x < (75)
t x
with the initial conditions
u(x, 0) = f (x) (76)
u
(x, 0) = 0 (77)
t
To solve this problem we consider the Fourier transform
Z
1
U (, t) = u(x, t)eix dx (78)
2
Z
u(x, t) = U (, t)eix d (79)

By applying the Fourier transform to equation (75) and using the properties of the transforms of the deriva-
tives we obtain the ODE
2U
= c2 2 U (80)
t2
with the initial conditions Z
1
U (, 0) = F[u(x, 0)] = f (x)eix dx (81)
2

U (, 0) = 0 (82)
t
The general solution of equation (80) is

U (, t) = A() cos ct + B() sin ct (83)

and from the initial conditions (81-82) we obtain the coefficients

B() = 0 (84)
Z
1
A() = U (, 0) = f (x)eix dx (85)
2

Therefore,
U (, t) = U (, 0) cos ct
and from (79) we obtain the solution
Z
u(x, t) = U (, 0) cos cteix d (86)

To simplify this expression we use Eulers formula

eict + eict
cos ct =
2
such that (86) becomes
Z
1 h i
u(x, t) = U (, 0) ei(xct) + ei(x+ct) d (87)
2

10
From (81) we have (inverse Fourier transform)
Z
f (x) = U (, 0)eix d

such that (87) becomes

1
[f (x ct) + f (x + ct)]
u(x, t) = (88)
2
which is the same expression we obtained using the method of characteristics.

Laplaces equation in a semi-infinite strip

Consider the Laplaces equation in a semi-infinite strip

2u 2u
2 u = + 2 = 0, 0 < x < L, y > 0 (89)
x2 y
with the boundary conditions
u(0, y) = g1 (y), y > 0 (90)
u(L, y) = g2 (y), y > 0 (91)
u(x, 0) = f (x), 0 < x < L (92)
In addition we assume that
lim u(x, y) = 0, 0 < x < L (93)
y

We split this problem into two subproblems that are easier to solve:

u(x, y) = u1 (x, y) + u2 (x, y) (94)

where u1 and u2 satisfy the Laplaces equation (89) and respectively, the boundary conditions

u1 (0, y) = g1 (y), u1 (L, y) = g2 (y), u1 (x, 0) = 0, lim u1 (x, y) = 0 (95)

y

u2 (0, y) = 0, u2 (L, y) = 0, u2 (x, 0) = f (x), lim u2 (x, y) = 0 (96)

y

Zero-temperature sides

Using separation of variables, we search for a solution

u2 (x, y) = (x)h(y) (97)

After replacing (97) in (89) and using the boundary conditions (95) we obtain

d2
= , (0) = 0, (L) = 0 (98)
dx2
d2 h
= h (99)
dy 2
From (98) we obtain the eigenvalues n = (nx/L)2 and the corresponding eigenfunctions n (x) = sin nx
L .
The general solution of (99) is
h(y) = an e n y + bn e n y

11
such that using the principle of superposition

X nx ny/L X nx ny/L
u2 (x, y) = an sin e + bn sin e (100)
n=1
L n=1
L

The boundary condition

lim u2 (x, y) = 0
y

implies bn = 0, n = 1, 2, . . .. The nonhomogeneous boundary condition u2 (x, 0) = f (x) implies then

X nx
f (x) = an sin
n=1
L

such that an are the Fourier sine coefficients of f (x)

Z L
2 nx
an = f (x) sin dx (101)
L 0 L

Therefore, the solution u2 (x, y) is

X nx ny/L
u2 (x, y) = an sin e (102)
n=1
L

where the coefficients an are evaluated according to (101).

Zero-temperature bottom

To find the solution u1 (x, y) we introduce the Fourier sine transform in y

Z
u1 (x, y) = U1 (x, ) sin y d (103)
0
Z
2
U1 (x, ) = u1 (x, y) sin y dy (104)
0
Taking the Fourier sine transform with respect to y of the Laplaces equation and using the properties of
the transforms of the derivatives, we obtain

2 2
U1 (x, ) + U1 (x, 0) 2 U1 (x, ) = 0 (105)
x2
Now, U1 (x, 0) = S[u1 (x, 0)] = S[0] = 0 such that (105) becomes

2
U1 (x, ) 2 U1 (x, ) = 0 (106)
x2
The solution of (106) may be expressed in terms of sinh as

U1 (x, ) = a() sinh x + b() sinh (L x) (107)

The coefficients a() and b() are obtained from the boundary conditions

2
Z Z
2 1
U1 (0, ) = S[g1 (y)] = g1 (y) sin y dy = b() sinh L b() = g1 (y) sin y dy
0 sinh L 0

2
Z Z
2 1
U1 (L, ) = S[g2 (y)] = g2 (y) sin y dy = a() sinh L a() = g2 (y) sin y dy
0 sinh L 0

12
This completes the evaluation of the Fourier sine transform U1 (x, ) and the solution u1 (x, y) is given by
(103).

Remark: The solution u(x, y) = u1 (x, y) + u2 (x, y) of the Laplaces equation in a semi-infinite strip is the
sum of a solution u1 (x, y) obtained using the Fourier sine transform and a solution u2 (x, y) obtained using
a Fourier sine series.

An alternative approach to obtain the solution u(x, y) is to directly apply the Fourier sine transform in y to
the equation (89) with nonhomogeneous boundary conditions
Z
u(x, y) = U (x, ) sin y d
0

The differential equation for U (x, ) is then nonhomogeneous

2U 2
2 U = f (x)
x2
and must be solved with two nonhomogeneous boundary conditions at x = 0 and x = L.

Laplaces equation in a half-plane

We consider the Laplaces equation in a half-plane

2u 2u
2 u = + 2 = 0, < x < , y > 0 (108)
x2 y
with the boundary conditions
u(x, 0) = f (x) (109)
lim u(x, y) = 0 (110)
x

lim u(x, y) = 0 (111)

x

lim u(x, y) = 0 (112)

y

To solve this problem we consider the Fourier transform in x

Z
u(x, y) = U (, y)eix d (113)

Z
1
U (, y) = u(x, y)eix dx (114)
2

Taking the Fourier transform in x of the Laplaces equation (108) we obtain the ODE

2U
2 U = 0 (115)
y 2
whose general solution is
U (, y) = a()ey + b()ey (116)
Next we use the boundary conditions to determine a() and b().

lim u(x, y) = 0 lim U (, y) = 0

y y

13
such that a() = 0 if > 0 and b() = 0 if < 0. Therefore, (116) becomes
(
a()ey , < 0
U (, y) = (117)
b()ey , > 0

which may be written in a compact form

U (, y) = c()e||y (118)

where c() = a(), < 0 and c() = b(), > 0.

From the boundary condition (109) we have
Z
1
U (, 0) = F[f (x)] = f (x)eix dx
2

such that c() in (118) must be the Fourier transform of f (x):

Z
1
c() = f (x)eix dx (119)
2
The solution u(x, y) is then given by (113) where U (, y) is evaluated according to (118)-(119).

Q: Using the convolution theorem, show that the solution u(x, y) may be expressed as
Z
1 2y
u(x, y) = f (x) dx (120)
2 (x x)2 + y 2

Laplaces equation in a quarter-plane

To solve the Laplaces equation in a quarter-plane x > 0, y > 0

2u 2u
2 u = + 2 = 0, x > 0, y > 0 (121)
x2 y
with the boundary conditions
u(0, y) = g(y), y > 0 (122)
u
(x, 0) = f (x), x > 0 (123)
y
lim u(x, y) = lim u(x, y) = 0 (124)
x y

we decompose the problem into two subproblems that are easier to solve:

u(x, y) = u1 (x, y) + u2 (x, y)

where

2 u1 = 0 2 u2 = 0

u1 (0, y) = g(y) u2 (0, y) = 0

u1 (x, 0) = 0 y u2 (x, 0) = f (x)
y
lim u1 (x, y) = 0 limx u2 (x, y) = 0
x

lim u1 (x, y) = 0 limy u2 (x, y) = 0

y

14
The techniques for solving these problems are very similar, and we only indicate how to solve the problem
for u1 (x, y). We consider the Fourier cosine transform in y since we have a homogeneous boundary condition
u1 /y(x, 0) = 0. Z
u1 (x, y) = U1 (x, ) cos y d (125)
0
Z
2
U1 (x, ) = u1 (x, y) cos y dy (126)
0
Taking the Fourier cosine transform in y of the Laplaces equation for u1 and using the homogeneous boundary
condition u1 /y(x, 0) = 0 we obtain the ordinary differential equation

2 U1
2 U1 = 0
x2
whose general solution is
U1 (x, ) = a()ex + b()ex , x > 0, > 0
To determine a() and b() we use the boundary conditions.

lim u1 (x, y) = 0 lim U1 (x, ) = 0 b() = 0

x x

Therefore,
U1 (x, ) = a()ex (127)
Z
2
U1 (0, ) = C[u1 (0, y)] = C[g(y)] a() = g(y) cos y dy (128)
0

The solution u1 (x, y) is obtained from (125) with U1 (x, ) given by (127)-(128).

15