Macquarie University

Department of Mathematics
F
F
o
o
u
u
r
r
i
i
e
e
r
r
T
T
h
h
e
e
o
o
r
r
y
y
B
B
.
.
M
M
.
.
N
N
.
.
C
C
l
l
a
a
r
r
k
k
e
e
Table of Contents
1. Introduction................................................................................................ ................................. 1
2. Linear differential operators ................................................................................................ ...... 3
3. Separation of variables ................................................................................................ ............... 5
4. Fourier Series................................................................................................ .............................. 9
5. Bessel's inequality................................................................................................ ..................... 14
6. Convergence results for Fourier series .................................................................................... 16
7. Differentiation and Integration of Fourier Series ................................................................... 20
8. Half-range Fourier series ................................................................................................ ......... 23
9. General Intervals ................................................................................................ ..................... 25
10. Application to Laplace's equation ............................................................................................ 27
11. Sturm-Liouville problems and orthogonal functions .............................................................. 33
12. Bessel Functions ................................................................................................ ....................... 37
13. Fourier Transforms ................................................................................................ .................. 41
14. Inverse Fourier Transforms ................................................................................................ ..... 47
15. Applications to Differential Equations .................................................................................... 50
16. Plancherel's and Parseval's Identities ..................................................................................... 54
17. Band Limited Functions and Shannon's Sampling Theorem................................................. 56
18. Heisenbergs Inequality ................................................................................................ ............ 59
1. Introduction. 1
1. Introduction.
Fourier theory is a branch of mathematics first invented to solve certain problems in
partial differential equations. The most well-known of these equations are:
Laplace's equation,

2
u
x
2
+

2
u
y
2
= 0, for u(x, y) a function of two variables,
the wave equation,

2
u
t
2
c
2

2
u
x
2
= 0, for u(x, t) a function of two variables,
the heat equation,
u
t

2
u
x
2
= 0, for u(x, t) a function of two variables.
In the heat equation, x represents the position along the bar measured from some origin, t
represents time, u(x, t) the temperature at position x, time t. Fourier was initially
concerned with the heat equation. Incidentally, the same equation describes the
concentration of a dye diffusing in a liquid such as water. For this reason the equation is
sometimes called the diffusion equation.
In the wave equation, x, represents the position along an elastic string under tension,
measured from some origin, t represents time, u(x, t) the displacement of the string from
equilibrium at position x, time t.
In Laplace's equation, u(x, y) represents the steady temperature of a flat conducting plate
at the position (x, y) in the plane.
Since both the heat equation and the wave equation involve a single space variable x, we
sometime refer to them as the one dimensional heat equation and the one dimensional
wave equation respectively.
Laplace's equation involves two spatial variables and is therefore sometimes called the
two-dimensional laplace equation. Laplace's equation is connected to the theory of analytic
functions of a complex variable. If f(z) = u(x, y) + iv(x, y), the real and imaginary parts
u(x, y), v(x, y) satisfy the Cauchy-Riemann equations,
u
x
=
v
y
,
u
y
=
v
x
,
Then

2
u
x
2
=

2
v
xy
=

2
u
y
2
or

2
u
x
2
+

2
u
y
2
= 0.
Similarly,

2
v
x
2
+

2
v
y
2
= 0.
Heat conduction and wave propagation usually occur in 3 space dimensions and are
described by the following versions of Laplace's equation, the heat equation and the wave
equation;

2
u
x
2
+

2
u
y
2
+

2
u
z
2
= 0,
1. Introduction. 2
u
t

,
_

2
u
x
2
+

2
u
y
2
+

2
u
z
2
= 0,

2
u
t
2
c
2

,
_

2
u
x
2
+

2
u
y
2
+

2
u
z
2
= 0.
2. Linear differential operators. 3
2. Linear differential operators.
All of the above mentioned partial differential equations can be written in the form
L[u] = F
where
L[u]
2
u =

2
u
x
2
+

2
u
y
2
+

2
u
z
2
in Laplace's equation,
L[u]
u
t

,
_

2
u
x
2
+

2
u
y
2
+

2
u
z
2
=
u
t

2
u in the heat equation,
and
L[u]

2
u
t
2
c
2

,
_

2
u
x
2
+

2
u
y
2
+

2
u
z
2
=

2
u
t
2
c
2

2
u in the heat equation.
L [u ] is in each case, a linear partial differential operator. Linearity means that for any
two functions u
1
, u
2
, and any two constants c
1
, c
2
,
L[c
1
u
1
+ c
2
u
2
] = c
1
L[u
1
] + c
2
L[u
2
].
In other words, L is linear if it preserves linear combinations of u
1
, u
2
. This definition
generalises to
L[c
1
u
1
+
....
+ c
n
u
n
] = c
1
L[u
1
] +
...
+ c
n
L[u
n
]
for any functions u
1
,..., u
n
and constants c
1
,..., c
n
.
Let u(x
1
,x
2
,...,x
n
) be a function of n variables x = (x
1
,x
2
,...,x
n
). Then the most general linear
partial differential operator is of the form
L[u] =

i=1
n

j=1
n
a
ij
(x)

2
u
x
i
x
j
+

i=1
n
b
i
(x)
u
x
i
+ c(x)u
where a
ij
(x), b
i
(x), c(x) are given coefficients.
The highest order partial derivative appearing is the order of the partial differential
operator. Henceforth we will consider only second order partial differential operators of the
form
L[u] =

i=1
n

j=1
n
a
ij
(x)

2
u
x
i
x
j
+

i=1
n
b
i
(x)
u
x
i
+ c(x)u.
The general linear second order partial differential equation is of the form
L[u] = F(x)
where F(x) is a given function. When F(x) 0, the equation L[u] = 0 is called homogeneous.
If F(x) 0, the equation L[u] = F(x) is called non-homogeneous.
Linearity of L is essential to the success of the Fourier method. There are usually
(infinitely) many solutions of a linear partial differential equation. The number of
2. Linear differential operators. 4
solutions may be restricted by imposing extra conditions. Often these extra conditions are
given as linear equations involving u and its derivatives on the boundary of some region
R
n
. These equations are written as boundary conditions
B[u(x)] = (x)
where B is a partial differential operator defined on the boundary of the region .
As an example let u(x, t) be the temperature at x , at time t, of a conducting body. Then
u satisfies the heat equation

u
t

2
u = 0, x , t > 0.
Suppose the temperature at the boundary is maintained at a given temperature , then
u(x, t) = (x, t), x , t > 0.
Also the initial temperature of the body is given by
u(x, 0) = f(x), x .
Suppose u
1
,..., u
k
satisfy the linear partial differential equations
L[u
j
] = F
j
, j = 1, ... , k
and boundary conditions
B[u
j
] =
j
, j = 1, ... , k,
then the linear combination u = c
1
u
1
+ .... + c
k
u
k
satisfies the linear partial differential
equation
L[u] = c
1
F
1
+ .... + c
k
F
k
and the boundary condition
B[u] = c
1

1
+ .... + c
k

k
.
This result is commonly referred to as the principle of superposition and it is of paramount
importance for the Fourier method. It shows that by taking linear combinations of
solutions of related linear partial differential equations, other solutions can be constructed
for source and boundary terms F, which are linear combinations of simpler terms.
3. Separation of variables. 5
3. Separation of variables.
Example 1. To motivate Fourier series, consider a heat conducting bar of length l,
insulated along its length, so that heat can flow only along the bar. The temperature u(x,t)
along the bar, satisfies the heat equation
u
t
u
xx
= 0, 0 < x < l, t > 0
and boundary conditions
u(0, t) = 0, u(l, t) = 0, t > 0.
Let the initial temperature along the bar at t = 0 be given by
u(x, 0) = f(x), 0 < x < l.
Assume a solution of the form
u(x, t) = X(x)T(t).
Such a solution is called a separation-of-variables solution. Substituting into the heat
equation,
X(x)T'(t) = X"(x)T(t).
Dividing by X(x)T(t), leads to
T'(t)
T(t)
=
X"(x)
X(x)
.
Since the variables x, t appear on separate sides of this equation, each side of this equation
can only be equal to a constant, say . Then
T' = T, and X" = X.
These are constant coefficient ordinary differential equations. is real but could be
positive, negative or zero. Assume for the moment that > 0. The solutions are
T(t) = Ce
t
, X(x) = Ae
x

+ Be
x

for A, B, C constants. Then u(x, t ) = X(x)T(t) satisfies u(0, t) = 0, u(l, t) = 0, t > 0, if and
only if X(0) = 0, X(l) = 0. That is
0 = X(0) = A + B
0 = X(l) = Ae
l

+ Be
l

.
Eliminating A, B leads to the condition
e
l

e
l

= 0
or
3. Separation of variables. 6
sinh ( ) l

= 0.
There are no values of > 0 which satisfy this condition. So cannot be positive.
Suppose that < 0 and let =
2
for some real . Then

= i and hence satisfies

e
il
e
i l
= 0
or
sin( ) l = 0.
This has solutions l = n, n = t1, t2,..., and hence
n
=
n
l
, n = t1, t2,... ,
T
n
(t) = e

n
2

2
t
l
2
, X
n
(x) = e

inx
l
e

inx
l
= 2i sin

_
nx
l
, n = 1, 2,...
For = 0,
T' = 0, and X" = 0
or
T(t) = C, and X(x) = A + Bx.
The boundary conditions u(0, t) = 0, u(l, t) = 0, t > 0, are satisfied if and only if X(0) = 0, X(l)
= 0. That is
0 = X(0) = A , 0 = X(l) = Bl
or
A = B = 0.
Therefore the non-trivial solutions of the heat equation
u
t
u
xx
= 0, 0 < x < l, t > 0
satisfying boundary conditions
u(0, t) = 0, u(l, t) = 0, t > 0
are
u
n
(x, t) = e

n
2

2
t
l
2
sin

_
nx
l
, n = 1, 2,... .
By superposition
u(x, t) =

n=1

b
n
u
n
(x, t) =

n=1

b
n
e

n
2

2
t
l
2
sin

_
nx
l

also satisfies the heat equation and the boundary conditions. For the initial condition to be
satisfied
u(x, 0) = f(x) =

n=1

b
n
sin

_
nx
l
, 0 < x < l.
That is we must show that f(x) can be represented as a series expansion of sines. We are
mainly concerned wuth those functions f(x) which have this property.
A series expansion such as
3. Separation of variables. 7
f(x) =

n=1

b
n
sin

_
nx
l

is called a Fourier series expansion.

Example 2. Suppose the conducting bar is insulated at each end, the temperature u(x,t)
satisfies the same heat equation and initial condition but different boundary conditions
u
x
(0, t) = 0, u
x
(l, t) = 0, t > 0.
Separation of variables u(x, t ) = X(x)T(t) in the heat equation leads to the same ordinary
differential equations
T' = T, and X" = X
and assuming for the moment that > 0, solutions
T(t) = Ce
t
, X(x) = Ae
x

+ Be
x

.
The boundary conditions are satisfied if and only if
0 = X'(0), 0 = X'(l).
That is
0 =

(A B), 0 =

( )
Ae
l

Be
l

.
Eliminating A, B again leads to the condition, 0,
e
l

e
l

= 0
or
sinh( ) l

= 0.
There are no values of > 0 which satisfy this condition. So cannot be positive.
Suppose that < 0 and let =
2
for some real . Then

= i and hence satisfies

e
il
e
i l
= 0
or
sin( ) l = 0.
This has solutions
n
l = n, n = t1, t2,..., and hence
n
=
n
l
, n = t1, t2,... ,

n
=
n

2
=

_
n
l

2
, n = 1, 2,... ,
T
n
(t) = e

n
2

2
t
l
2
, X
n
(x) = e

inx
l
+ e

inx
l
= 2 cos

_
nx
l
, n = 1, 2,...
If = 0, T(t) = C, and X(x) = A + Bx. The boundary conditions u(0, t) = 0, u(l, t) = 0, t >
0, if and only if X'(0) = 0, X'(l) = 0. That is
0 = X'(0) = B , 0 = X'(l).
3. Separation of variables. 8
or
X(x) = A
0
, T(t) = C
0
Therefore the non-trivial solutions of the heat equation
u
t
u
xx
= 0, 0 < x < l, t > 0
satisfying boundary conditions
u
x
(0, t) = 0, u
x
(l, t) = 0, t > 0
are
u
n
(x,t) = e

n
2

2
t
l
2
cos

_
nx
l
, n = 0, 1, 2,... .
By superposition
u(x,t) =

n=0

a
n
u
n
(x,t) =

n=0

a
n
e

n
2

2
t
l
2
cos

_
nx
l

also satisfies the heat equation and the boundary conditions. For the initial condition to be
satisfied
u(x, 0) = f(x) =

n=0

a
n
cos

_
nx
l
, 0 < x < l.
This is also called a Fourier series expansion.
4. Fourier Series. 9
4. Fourier Series.
A function f(x), f: RR is called periodic if f(x + P) = f(x) for all x R. P > 0 is called the
period of f.
Suppose f(x ) is periodic with period 2, then an important question is whether f(x ) has a
Fourier series expansion of the form
f(x) =
a
0
2
+

n=1

( )
a
n
cos nx+b
n
sin nx , 0 < x < 2.
The constant term is taken as
a
0
2
as a matter of convenience.
The formulas
e
ix
= cos x + i sin x, e
ix
= cos x i sin x
can be used to write the Fourier series expansion as
f(x) =

c
n
e
inx
where the coefficients are
c
n
=
a
n
ib
n
2
, c
n
=
a
n
+ib
n
2
, n = 1, 2, ... , c
0
=
a
0

2
.
Assuming for the moment that the 2-periodic function f has a Fourier series expansion,
the Fourier coefficients c
n
will be determined using the following orthogonality property of
the complex exponentials e
inx
, n = 0, t1, t2, ... ,.
Lemma.

e
inx
e
imx
dx =

'

0; nm

2; n=m
.
Proof. For n m,

e
inx
e
imx
dx =

e
i (nm)x
dx
=
1
i(nm)
[ ] e
i (nm)
e
i (nm)

=
2
(nm)
sin(n m) = 0.
If n = m,

e
inx
e
inx
dx =

1 dx = 2.

4. Fourier Series. 10
Since f(x) =

c
n
e
inx
,

f(x) e
imx
dx =

c
n
e
inx
e
imx
dx
=

c
n

e
i (nm)x
dx = 2 c
m
.
Hence
c
m
=
1
2

f(x) e
imx
dx , m = 0, t1, t2, ... ,
which are the Fourier coefficients of f(x).
Since
c
n
=
a
n
ib
n
2
, c
n
=
a
n
+ib
n
2
, n = 1, 2, ... , c
0
=
a
0

2
,
the Fourier cosine and sine coefficients are given by
a
0
= 2c
0
=
1

f(x) dx,
a
n
= c
n
+ c
n
=
1
2

f(x) ( ) e
inx
+e
inx
dx =
1

f(x) cos nx dx
b
n
= i(c
n
c
n
) =
i
2

f(x) ( ) e
inx
e
inx
dx =
1

f(x) sin nx dx.

Either of the expansions

c
n
e
inx
,
a
0
2
+

n=1

( )
a
n
cos nx+b
n
sin nx ,
is called a Fourier series expansion of f(x). The first, the exponential form and the second
the trigonometric form.
Notice that if f(x) is an even periodic function, i.e. f(x) = f(x) for all x, then
b
n
=
1

f(x) sin nx dx =
1

0
f(x) sin nx dx +
1

f(x) sin nx dx
=
1

0
f(x) sin( nx) (dx) +
1

f(x) sin nx dx
=
1

(f(x) + f(x)) sin nx dx = 0, n = 1, 2, ... .

4. Fourier Series. 11
Thus the Fourier series expansion of an even function f(x) has only cosine terms

a
0
2
+

n=1

a
n
cos nx
Furthermore, the coefficients a
n
a
n
=
1

f(x) cos nx dx =
1

0
f(x) cos nx dx +
1

f(x)cos nx dx
=
1

0
f(x) cos( nx) (dx) +
1

f(x) cos nx dx
=
1

(f(x) + f(x)) sin nx dx

=
2

f(x) cos nx dx, n = 0, 1, 2, ... .

Similarly if f(x) is an odd periodic function, f(x) = f(x), then
a
n
=
1

f(x) cos nx dx =
1

0
f(x) cos nx dx +
1

f(x)cos nx dx
=
1

0
f(x) cos( nx) (dx) +
1

f(x) cos nx dx
=
1

(f(x) + f(x)) sin nx dx = 0, n = 0, 1, 2, ... .

Thus the Fourier series expansion of an odd function f(x) has only sine terms

n=1

b
n
sin nx,
and the coefficients are
b
n
=
1

f(x) sin nx dx =
1

0
f(x) sin nx dx +
1

f(x) sin nx dx
=
1

0
f(x) sin( nx) (dx) +
1

f(x) sin nx dx
=
1

(f(x) + f(x)) sin nx dx

=
2

f(x) sin nx dx , n = 1, 2, ... .

It is worth noting that for a periodic function f(x), the constant term in the Fourier series

a
0
2
=
1
2

f(x) dx
4. Fourier Series. 12
is the average value of f(x) over a period, < x < .
Example (a). Let f(x) be periodic with period 2,
f(x) =

'

1; <x<0

1; 0< x<
.
-5 -3 -1 1 2 3 4 5 6
-1
0.5
-6.28 6.28
-1.5
1.5
This is called a squarewave function. It is obviously an odd function, hence its Fourier
cosine coefficients a
n
are all zero and
b
n
=
2

f(x) sin nx dx =
2

1 sin nx dx
=
2
n
[ ] cosn+cos0 =
2
n

[ ]
(1)
n
+1
=

'

0; neven

4
n
; nodd
.
Therefore b
2n
= 0, b
2n1
=
4
(2n1)
, n = 1, 2, ... . The Fourier series of f(x) is

n=1

b
n
sin nx, =

n=1

b
2n1
sin (2n 1)x
=

n=1

4
(2n1)
sin (2n 1)x
=
4

n=1

1
2n1
sin (2n 1)x.
(b). Let f(x) be periodic with period 2, f(x) = |sin x|, < x < ,
f(x) =

'

sinx; <x<0

sin x; 0< x<

.
4. Fourier Series. 13
-5 -3 -1 1 2 3 4 5 6
0.25
0.75
1.25
-6.28 6.28
-0.5
1.5
f(x) is an even periodic function, hence b
n
= 0 for all n.
a
n
=
2

f(x) cos nx dx =
2

sin x cos nx dx
=
2

1
2

[ ]
sin(n+1)x+sin((n+1)x) dx
=
1

[ ]
sin(n+1)xsin(n1) x dx
=
1

]
1
1

cos(n+1)
n+1
+
cos(n1)
n1
+
1
n+1

1
n1

=
1

]
1
1
n+1
(1cos( n+1))
1
n1
(1cos( n1) )
=

'

0; nodd

]
1
2
n+1

2
n1
;neven
=

'

0; nodd

4
(n
2
1)
;neven
.
Therefore a
2n
=
4
((2n)
2
1)
, a
2n1
= 0 and the Fourier series of f(x) is

a
0
2
+

n=1

a
n
cos nx =
2

n=1

1
4n
2
1
cos 2nx.
5. Bessel's Inequality. 14
5. Bessel's inequality.
Theorem. Let f be 2 -periodic and Riemann integrable on [, ]. Then

n=

|c
n
|
2

1
2

|f(x)|
2
dx.
Proof. Using z z = |z|
2
for complex z,
0

f(x)

n=N
N
c
n
e
inx

2
=

_
f(x)

n=N
N
c
n
e
inx

_
f(x)

m=N
N
c
m
e
imx

= |f(x)|
2

n=N
N
c
n
e
inx
f(x)

m=N
N
c
m
e
imx
f(x) +

n=N
N

m=N
N
c
n
c
m
e
i (nm)x
Dividing by 2 and integrating over [, ],
0
1
2

|f(x)|
2
dx

n=N
N
c
n

1
2

e
inx
f(x)dx

m=N
N
c
m

1
2

e
imx
f(x) dx
+

n=N
N

m=N
N
c
n
c
m

1
2

e
i (nm)x
dx
=
1
2

|f(x)|
2
dx

n=N
N
c
n
c
n

m=N
N
c
m
c
m
+

n=N
N
c
n
c
n
=
1
2

|f(x)|
2
dx

n=N
N
|c
n
|
2
for any N.
Letting N leads to the result.

Bessel's inequality will later be shown to be actually an equality but for now it implies that
the series

n=

|c
n
|
2
converges where c
n
are the Fourier coefficients of the Riemann
integrable function f.
Using the equations
c
n
=
a
n
ib
n
2
, c
n
=
a
n
+ib
n
2
, n = 1, 2, ... , c
0
=
a
0

2
,
Bessel's inequality can be written as
5. Bessel's Inequality. 15

n=

|c
n
|
2
=
a
0
2
4
+

n=1

a
n
ib
n
2
2
+

n=1

a
n
+ib
n
2
2
=
a
0
2
4
+
1
2

n=1

[ ]
|a
n
|
2
+|b
n
|
2

1
2

|f(x)|
2
dx.
This implies that the series

n=1

|a
n
|
2
,

n=1

|b
n
|
2
also converge, where a
n
, b
n
are the Fourier cosine and
sine coefficients of f.
6. Convergence results for Fourier series. 16
6. Convergence results for Fourier series.
We will consider the question: For what functions f do the Fourier series

c
n
e
inx
,
a
0
2
+

n=1

( )
a
n
cos nx+b
n
sin nx ,
converge?
Because we are dealing with functions, the concept of convergence must be made precise.
Do we mean the numerical series converging at every x [ , ]? Can convergence be
different at different points x, indeed can we have convergence at some points and not at
others and if so, which are the points of convergence? Can we have some sort of average
convergence on [ , ]?
A function f is called piecewise continuous on an interval [a, b] if it is continuous
everywhere except at finitely many points x
1
, x
2
, ... , x
k
[a, b] and the left-hand and
right-hand limits of f exist at each of the points x
1
, x
2
, ... , x
k
. The set of all piecewise
continuous functions on [a, b] is denoted by PC[a, b].
f is called piecewise smooth if f and its derivative f' are piecewise continuous on [a, b].
Consider the N
th
partial sum of the complex Fourier series of f,
S
N
f
(x) =

n=N
N
c
n
e
inx
=

n=N
N

_
1
2

e
iny
f(y)dy e
inx
=

_
1
2

n=N
N
e
in(xy)
f(y) dy.
We define D
N
(x) =
1
2

n=N
N
e
inx
. Then
S
N
f
(x) =

D
N
(x y)f(y) dy =

D
N
(y)f(x y) dy .
[The last equality follows from a change of variable and the periodicity of the integrand
D
N
(x y)f(y)]. The function D
N
(x) is called the Dirichlet kernel and
D
N
(x ) =
1
2
e
iNx

n=0
2N
e
inx
=
1
2
e
iNx

_
e
i (2N+1) x
1
e
ix
1

6. Convergence results for Fourier series. 17

=
1
2

_
e
i (N+1) x
e
iNx

e
ix
1

=
1
2

_
e
i (N+
1
2
)x
e
i (N+
1
2
)x
e

ix
2
e

ix
2

=
1
2

sin
( )
N+
1
2
x
sin( )
x
2
.
The question of convergence of the Fourier series reduces to question:
Does S
N
f
(x) converge as N for x [ , ]? If so, to what does it converge?
We prove a preliminary result.
Lemma.

0
D
N
(x) dx =

D
N
(x) dx =
1
2
.
Proof. D
N
(x) =
1
2
+
1
2

n 0

n=N
N
e
inx
=
1
2
+
1

n=1
N
cos nx
From the even-ness of D
N
(x) and integrating,

0
D
N
(x) dx =

D
N
(x) dx =

]
1
1
1
2
+
1

n=1
N
cosnx dx
=

]
1
1 x
2
+
1

n=1
N

sin nx
n

0
=
1
2
.

Theorem. Let f be a piecewise smooth 2-periodic function on R. Then

lim
N
S
N
f
(x) =
1
2
[ ]
f(x

)+f(x
+
)
for every x. Hence lim
N
S
N
f
(x) = f(x) for each point x of continuity of f.
Proof. S
N
f
(x)
1
2
[ ]
f(x

)+f(x
+
) =

D
N
(y)f(x y) dy
1
2
[ ]
f(x

)+f(x
+
)
=

D
N
(y)[f(x y) (f(x

)]dy +

0
D
N
(y)[f(x y) f(x
+
)]dy
(from lemma above)
6. Convergence results for Fourier series. 18
=

D
N
(y)[f(x y) f(x

)]dy +

0
D
N
(y)[f(x + y) f(x
+
)](dy)
=

D
N
(y)[f(x y) f(x

) + f(x + y) f(x
+
)]dy
=

1
2

sin
( )
N+
1
2
y
sin( )
y
2
[f(x y) f(x

) + f(x + y) f(x
+
)]dy
=
1

sin
( )
N+
1
2
y

_
y
2
sin( )
y
2

]
1
f(xy)f(x

)
y
+
f(x+y) f(x
+
)
y
dy
For fixed x define
g(y) =

_
y
2
sin( )
y
2

]
1
f(xy)f(x

)
y
+
f(x+y) f(x
+
)
y
,
which is an odd piecewise continuous function for [, ] by the condition of piecewise
smoothness on f.
Then
S
N
f
(x)
1
2

[ ]
f(x

)+f(x
+
) =
1

sin
( )
N+
1
2
y g(y)dy
=
1

sin (Ny){ } cos( )

y
2
g(y) dy +
1

cos (Ny){ } sin( )

y
2
g(y) dy
The last two terms are the Fourier sine and cosine coefficients B
N
, A
N
of

1
2
cos( )
y
2
g(y),
1
2
sin( )
y
2
g(y) respectively. By Bessel's inequality, B
N
, A
N
0 as N
.
Therefore
S
N
f
(x)
1
2
[ ]
f(x

)+f(x
+
) = B
N
+ A
N
0, as N .

This result, that the partial sums of the Fourier series of a piecewise smooth 2periodic
function converges pointwise to the mean of the left and right hand limits at x. If x is a
point of continuity of f, then the partial sums converge pointwise to f(x).
6. Convergence results for Fourier series. 19
Fourier series provide a useful method for summing certain numerical series.
Example. The Fourier series of the continuous periodic function f(x) =|sin x|, x [ , ],
is

2

n=1

1
4n
2
1
cos 2nx
-5 -3 -1 1 2 3 4 5 6
0.25
0.75
1.25
-6.28 6.28
-0.5
1.5
Because x = 0 is a point of continuity of f, the Fourier series converges at x = 0 to f(0) = 0,
0 =
2

n=1

1
4n
2
1
or

n=1

1
4n
2
1
=
1
2
.
At x =

2
, the Fourier series converges to
f
( )

2
= 1 =
2

n=1

1
4n
2
1
cos(n)
=
2

n=1

(1)
n
4n
2
1

n=1

(1)
n
4n
2
1
=

_
2
4
.
7. Differentiation and Integration of Fourier Series. 20
7. Differentiation and Integration of Fourier Series.
Fourier series can be differentiated term-by-term but the question is does the resulting
series converge and if so, to what does it converge? Similar concerns apply to the series
resulting from the term-by-term integration of a Fourier series.
Theorem. Let a
n
, b
n
, c
n
be the Fourier coefficients of a 2-periodic piecewise smooth
function f and a'
n
, b'
n
, c'
n
be the Fourier coefficients of f'. Then
a'
n
= nb
n
, b'
n
= na
n
, c'
n
= inc
n
.
Proof. By integration by parts,
c'
n
=
1
2

f'(x) e
inx
dx
=
1
2
f(x)e
inx

1
2

f(x) (ine
inx
)dx
= in
1
2

f(x) e
inx
dx = inc
n
.
A similar proof works for a'
n
= nb
n
, b'
n
= na
n
.

Theorem. Let f be 2-periodic, piecewise smooth, with piecewise smooth derivative f'.
Then the Fourier series of f' is

inc
n
e
inx
=

n=1

( )
nb
n
cos nx na
n
sin nx
and converge for each x where f'(x) exists. If f' is not continuous at x, then the series above
converge to
1
2
[f'(x

) + f'(x
+
)].
Proof. This result follows by combining the previous two theorems.

Integration of Fourier series is not so straightforward since the anti-derivative of a

periodic function need not be periodic. For example, f(x) = 1 is periodic but its anti-
derivative F(x) = x is not. However, since all but the constant term of a Fourier series has a
periodic anti-derivative, the following result is true.
Theorem. Let f be 2-periodic, piecewise continuous with Fourier coefficients a
n
, b
n
, c
n
and let F(x) =

0
x
f(t) dt. If c
0
=
1
2
a
0
= 0, then for all x,
F(x) = C
0
+
n 0

n=

c
n
in
e
inx
=
A
0
2
+

n=1

,
_

,
_
b
n
n
cos nx+

,
_
a
n
n
sin nx
7. Differentiation and Integration of Fourier Series. 21
where C
0
=
A
0
2
=
1
2

F(x) dx the average value of F on [, ].

If c
0
0, then the series above converges to F(x) c
0
x.
Proof. Since f(x) is piecewise continuous, F(x) =

0
x
f(t) dt is continuous.
If c
0
= 0, F(x) is 2-periodic since
F(x + 2) F(x) =

0
x+2
f(t) dt

0
x
f(t) dt =

x
x+2
f(t) dt =

f(t) dt = 2c
0
= 0.
Therefore the Fourier series of F converges pointwise at each x [, ] to F(x). Since f(x) =
F'(x), by the two previous theorems,
a
n
= nB
n
, b
n
= nA
n
, c
n
= inC
n
,
where a
n
, b
n
, c
n
are the Fourier coefficients of f and A
n
, B
n
, C
n
are the Fourier coefficients
of F. If n 0, this implies that
A
n
=
b
n
n
, B
n
=
a
n
n
, C
n
=
c
n
in
.
The constant C
0
=
A
0
2
=
1
2

F(x) dx is the constant term in the Fourier series of

F(x) =

0
x
f(t) dt.
If c
0
0, f(x) c
0
has zero mean value on [, ] and therefore its zeroth Fourier coefficient
is

1
2

(f(x) c
0
)dx =
1
2

f(x) dx c
0
= c
0
c
0
= 0.
Applying the result just obtained to f(x) c
0
and its anti-derivative F(x) c
0
x completes
the theorem.

Integrating and differentiating known Fourier series using the above results is a useful
way of obtaining new Fourier series.
Example. The 2-periodic function f :RR, f(x) = x( |x|), x [, ], is continuous with
piecewise smooth derivative. Therefore its Fourier series converges at each x to f(x). The
Fourier series is by the table #17
f(x) = x( |x|) =
8

n=1

1
(2n1)
3
sin (2n 1)x.
By the above results,
f'(x) = 2|x| =
8

n=1

1
(2n1)
2
cos (2n 1)x
7. Differentiation and Integration of Fourier Series. 22
or
|x| =

2

4

n=1

1
(2n1)
2
cos (2n 1)x
which agrees with table #2.
Example. f(x) =

'

0; <x<0

1; 0<x<
-5 -3 -1 1 2 3 4 5 6
0.25
0.75
1.25
-6.28319 6.28319
-0.5
1.5
is piecewise continuous with Fourier series
f(x) =
1
2
+
2

n=1

1
2n1
sin (2n 1)x
(table #7). Since c
0
=
1
2
0, F(x) =

0
x
f(t) dt. =

'

0; <x<0

x; 0<x<
.
Therefore
F(x)
x
2
=
A
0
2

2

n=1

1
(2n1)
2
cos (2n 1)x
Since
A
0
2
=
1
2

F(x) dx =

4
, F(x)
x
2
=

4

2

n=1

1
(2n1)
2
cos (2n 1)x.
But F(x)
x
2
=

'

x
2
; <x<0

x
2
; 0<x<
=
|x|
2
Therefore

|x|
2
=

4

2

n=1

1
(2n1)
2
cos (2n 1)x
or
|x| =

2

4

n=1

1
(2n1)
2
cos (2n 1)x.
This result agrees with the previous example
8. Half-range Fourier series. 23
8. Half-range Fourier series.
It is often convenient to represent a given function as a Fourier series which contains only
cosine terms or only sine terms, as in the initial examples.
If f(x) is piecewise continuous on [0, ], it can be extended to [, 0] as either an even
function or as an odd function.
Then f is periodically extended to the whole real line as a 2-periodic function by
f(x + 2) = f(x), x R.
Let f(x) be a piecewise smooth function given on the interval [0, ] and extended to [, 0]
as an even function. That is, f(x) = f(x), x [, 0]. Then the periodic extension to R, is
piecewise smooth and has Fourier series

a
0
2
+

n=1

a
n
cos nx
where
a
n
=
2

f(x) cos nx dx.

Similarly, extending f(x) to [, 0] as an odd function and periodically to R, the extension
has Fourier series

n=1

b
n
sin nx
where
b
n
=
2

f(x) sin nx dx.

The Fourier series
a
0
2
+

n=1

a
n
cos nx,

n=1

b
n
sin nx are known as half-range series for
f(x), x [0, ].
Example. Let f(x) = sin x, x [0, ]. Extend f to [, 0] as an even function.
That is f(x) =

'

sin(x)=sin x; <x<0

sin x; 0<x<
= |sin x|
8. Half-range Fourier series. 24
-2 -1 1 2 3
0.25
0.75
1.25
-3.14159 3.14159
-0.5
1.5
The half-range Fourier series of f is given by
f(x) =
2

n=1

1
2n
2
1
cos 2nx (table #8).
9. General Intervals. 25
9. General Intervals.
The theory of Fourier series can be expanded to include functions which have arbitrary
period. Let f :RR have period 2l > 0. That is f(x + 2l) = f(x), x R. Also define : RR by
(x) = f

_
lx

. Then
(x + 2) = f

_
l(x+2)

= f

_
lx

+2l = f

_
lx

= (x)
So has period 2. Applying the earlier results to , results in the Fourier series
(x) =

c
n
e
inx
=
a
0
2
+

n=1

( )
a
n
cos nx+b
n
sin nx ,
where
a
n
=
1

(x) cos nx dx, b

n
=
1

(x) sin nx dx, c

n
=
1
2

(x) e
inx
dx.
The change of variable y =
lx

leads to
a
n
=
1
l

l
l

_
y
l
cos

_
ny
l
dy =
1
l

l
l
f(x) cos

_
ny
l
dy,
b
n
=
1
l

l
l

_
y
l
sin

_
ny
l
dy =
1
l

l
l
f(y) sin

_
ny
l
dy,
c
n
=
1
2l

l
l

_
y
l
e

iny
l
dy =
1
2l

l
l
f(y) e

iny
l
dy.
Therefore
f(x) =

_
x
l
=

c
n
e

inx
l
=
a
0
2
+

n=1

_
a
n
cos

_
nx
l
+b
n
sin

_
nx
l
is the Fourier series of a 2l-periodic function f.
It follows from the above calculations that if f(x) is an even function on [l, l], its Fourier
series contains no sine terms and is of the form

a
0
2
+

n=1

a
n
cos

_
nx
l
,
with
a
n
=
2
l

0
l
f() cos

_
n
l
d
9. General Intervals. 26
Likewise if f(x) is an odd function on [l, l], its Fourier series contains no cosine terms and
is of the form

n=1

b
n
sin

_
nx
l
,
with
b
n
=
2
l

0
l
f() sin

_
n
l
d.
A function f(x) defined on an interval [0, l] can be extended to [l, l] either as an even or an
odd function and the extended periodically with period 2l to the real line. The Fourier
series of such functions are half-range expansions on [0, l] and contain no sine terms in the
case of an even extension or no cosine terms for an odd extension.
10. Application to Laplace's equation. 27
10. Application to Laplace's equation.
Example 1. Let be a rectangle 0 x a, 0 y b, and consider the boundary-value
problem

2
u
x
2
+

2
u
y
2
= 0, 0 x a, 0 y b
u(x, 0) = 0, u(x, b) = f(x), 0 x a
u(0, y) = u(a, y) = 0, 0 y b.
u = 0
u = f
u = 0
u = 0
x = a
y = b
Assuming a separation of variables solution of the form u(x, y) = X(x)Y(y) and substituting
into Laplaces equation,
X
"
Y + X Y
"
= 0
or

X
"
X
=
Y
"
Y
= constant =
For > 0, X
"
+ X = 0, Y
"
Y = 0, hence
X(x) = A cos

x + B sin

x , Y (y) = C cosh

y + D sinh

y .
To satisfy the boundary conditions u(0, y) = u(a, y) = 0, 0 y b, requires that
X(0) = A = 0, X(a) = B sin( ) a

= 0.
Therefore a

= n, =

_
n
a
2
, n = 1, 2, ... , and X(x) = B sin

_
nx
a
.
To also satisfy the boundary condition u(x, 0) = 0, 0 x a, requires that
10. Application to Laplace's equation. 28
Y(0) = C = 0 and therefore Y (y) = D sinh

_
ny
a
. The separation of variables solutions for
> 0 are
u(x, y) = X(x)Y(y) = BD sin

_
nx
a
sinh

_
ny
a
, n = 1, 2, ... ,
for an arbitrary constant BD.
For < 0, let = , > 0. Then X
"
X = 0, Y
"
+ Y = 0, hence
X(x) = A cosh

x + B sinh

x , Y (y) = C cos

y + D sin

y .
To satisfy the boundary conditions u(0, y) = u(a, y) = 0, 0 y b, requires that
X(0) = A = 0, X(a) = B sinh( ) a

= 0.
But B = 0 gives a trivial solution and sinh( ) a

0 for < 0. So there is only the trivial

separation of variables solution u = 0, for < 0.
Finally suppose that = 0, then X
"
= 0, Y
"
= 0, hence
X(x) = A + Bx, Y(y) = C + Dy.
To satisfy the boundary conditions u(0, y) = u(a, y) = 0, 0 y b, requires that
X(0) = A = 0, X(a) = Ba = 0.
That is, A = B = 0 and the only separation of variables solution for = 0 is the trivial
solution.
Summarising thus far, there is a sequence of non-trivial solutions of the form
u
n
(x, y) = B
n
sin

_
nx
a
sinh

_
ny
a
, n = 1, 2, ... ,
where B
n
are the constants BD for each n = 1, 2, ... . By superposition, a solution is also
given by
u(x, y) =

n =1

B
n
u
n
(x, y)
=

n =1

B
n
sin

_
nx
a
sinh

_
ny
a
.
The boundary condition u(x, b) =f(x), 0 x a, implies that
10. Application to Laplace's equation. 29
u(x, b) = f(x) =

n =1

B
n
sin

_
nx
a
sinh

_
nb
a
.
Since (x), 0 x a, has a Fourier half-range series f(x) =

n =1

b
n
sin

_
nx
a
, where
b
n
=
2
a

0
a
f() sin

_
n
a
d = B
n
sinh

_
nb
a
.
Therefore
B
n
=
b
n
sinh

_
nb
a
=
2
asinh

_
nb
a

0
a
f() sin

_
n
a
d
u(x, y) =

n =1

B
n
sin

_
nx
a
sinh

_
ny
a
.
Suppose all four sides of the rectangle have non-zero boundary conditions, then we can
break up the problem into four sub-problems each similar to the one we have just solved.
Each of these subproblems will have zero boundary conditions on three of the four sides
and can be solved as above. Then the solution to the boundary-value problem is the sum of
the four sub-problems by superposition.
Example 2. We seek a solution to Laplaces' equation on a circle, satisfying Dirichlet
boundary conditions. It is natural to choose a circle centre the origin, radius a, and
Laplaces' equation in polar coordinates,
u(r, ) =

2
u
r
2
+
1
r

u
r
+
1
r
2

2
u

2
= 0, r < a, 0 < 2,
subject to boundary conditions
u(a, ) = f(), 0 < 2,
where is a continuous function on 0 < 2.

u =
a
10. Application to Laplace's equation. 30
Using separation of variables, assume that
u(r, ) = R(r)()
and substitute into the differential equation to obtain
R
"
+
1
r
R
'
+
1
r
2
R
"
= 0,
or separating variables,
r
2
R
"
+rR
'
R
=

"

= constant =
2
(say),
which leads to the two families of ordinary differential equations
r
2
R
"
+ rR
'

2
R = 0,
"
+
2
= 0.
Consider the case = 0; then the solutions of the above equations are
= A
0
+ B
0
,
R = C
0
+ D
0
log r
Since solutions u must be periodic in , and bounded for r a, we conclude that B
0
= 0, D
0
= 0.
In the case =/ 0 the solutions of the above equations are
= A

cos + B

sin
R = C

+ D

where the coefficients depend on . Again because solutions are periodic in with period
2, we conclude that = n = 1,2,3,.., and because solutions are bounded for r a, D
n
= 0, n
= 1,2,3,... By superposition we can combine these solutions to obtain
u(r,) =

n = 0

C
n
r
n
( )
A
n
cos n+B
n
sin n
We may as well incorporate the constant C
n
into A
n
and B
n
, hence
u(r, ) =

n = 0

r
n
( )
A
n
cos n+B
n
sin n
At the boundary r = a,
f() = u(a, ) =

n = 0

a
n
( )
A
n
cos n+B
n
sin n
10. Application to Laplace's equation. 31
=
a
0
2
+

n = 1

( )
a
n
cos n+b
n
sin n
where
a
n
=
1

0
2
f(t)cosntdt , b
n
=
1

0
2
f(t)sinntdt
are the Fourier coefficients of the function f. Equating coefficients of cos n, sin n, in
these two expressions for f(), we obtain that
A
n
=
a
n
a
n
, B
n
=
b
n
a
n
and hence
u(r, ) =
a
0
2
+

n = 1

,
_
r
a
n
( )
a
n
cos n+b
n
sin n
This series converges uniformly for 0 < 2, r < a.
We can get a closed form expression for u by substituting the formulae for the Fourier
coefficients as follows.
u(r, ) =
1
2

0
2
f(t)dt +
1

n = 1

,
_
r
a
n

0
2
f(t)( ) cos nt cosn+sinnt sinn dt
=
1
2

0
2
f(t)dt +
1

n = 1

,
_
r
a
n

0
2
f(t)cos n(t)dt
=
1
2

0
2
f(t)dt +
1

0
2
f(t)

n = 1

,
_
r
a
n
cos n(t ) dt
where the interchange of the order of summation and integration is allowed by the uniform
convergence of the series.
Therefore u(r, ) =
1
2

0
2
f(t)

]
1
1
1+2

n = 1

,
_
r
a
n
cos n(t) dt.
Using the exponential form, 2 cos x = e
ix
+ e
i x
the term in the square brackets reduces to
two infinite geometric series whose sum is
1 + 2

n = 1

,
_
r
a
n
cos n(t ) =
a
2
r
2
a
2
2arcos(t)+ r
2
(exercise). Therefore
u(r, ) =
1
2

0
2
f(t)
a
2
r
2
a
2
2ar cos(t)+r
2
dt,
10. Application to Laplace's equation. 32
which is Poissons' integral formula for the circle.
11. Sturm-Liouville problems and orthogonal functions. 33
11. Sturm-Liouville problems and orthogonal functions.
The functions sin nx , cos nx, e
inx
, e
inx
, n = 0, 1, 2, ... , are examples of orthogonal
functions on [, ] Their orthogonality properties follow from the fact that they solutions
of linear second order ordinary differential equations.
For example sin nx , cos nx, e
inx
satisfy
u"(x) = u(x),
u() = u(), u'() = u'()
for = n
2
. The ordinary differential equation together with the boundary conditions is
called a boundary-value problem.
The set {e
inx
, n = 0, t1, t2, ... } form a basis for the vector space L
2
[, ] consisting of
functions f :[, ] R for which

|f(x)|
2
dx < . A set of functions {
n
(x); n = 1, 2, ... } is
said to be a basis for L
2
[, ] if for any function f L
2
[, ] , there is a unique set of
scalars c
n
, n = 1, 2, ... such that
lim
N

f(x)

n=1
N
c
n

n
(x)
2
dx = 0.
Then we say that f has the expansion f =

n=1

c
n

n
.
Furthermore, if the set {
n
(x); n = 1, 2,... } is orthonormal, that is

n
(x)
m
(x) dx =

'

0; nm

1; n=m
then the coefficients c
n
are given by
c
n
=

f(x)
n
(x) dx, n = 1, 2, ... .
The interval [, ] can be replaced by [a, b] and the theory of Fourier series generalised
to the expansion of arbitrary functions f L
2
[a, b] in terms of an orthonormal basis {
n
(x);
n = 1, 2,... } consisting of functions which are solutions of a boundary-value problem for
certain second order linear ordinary differential equations.
Let p(x), q(x), w(x) be real continuous functions on the interval [a, b]. Let p(x) be
continuous and p(x) > 0, q(x) 0, w(x) 0 on [a, b]. Let
0
,
1
,
0
,
1
, be real constants.
A Sturm-Liouville problem is a boundary value problem

( )
p(x)u' ' + q(x)u = w(x)u, x [a, b]
11. Sturm-Liouville problems and orthogonal functions. 34

0
u(a) +
1
u'(a) = 0,

0
u(b) +
1
u'(b) = 0.
The value of for which this boundary-value problem has non-trivial solutions are called
eigenvalues of the boundary-value problem and the corresponding solutions u(x) are called
eigenfunctions. It can be shown that the eigenvalues form a countable set {
n
; n = 1, 2,... }
and the corresponding eigenfunctions {
n
(x); n = 1, 2,... } are orthonormal with respect to
the weighted inner-product
< >
f; g
w
=

a
b
w(x) f(x) g(x) dx.
Let the vector space L
2
w
[a, b] consist of functions f :[a, b] R for which
||f||
2
=

a
b
w(x)|f(x)|
2
dx < . A set of functions {
n
(x); n = 1, 2, ... } is said to be a basis for
L
2
w
[a, b] if for any function f L
2
w
[a, b] , there is a unique set of scalars c
n
, n = 1, 2, ... such
that
lim
N

a
b
w(x)

f(x)

n=1
N
c
n

n
(x)
2
dx = 0.
Let the linear Sturm-Liouville differential operator L be defined as
L[f] =
1
w(x)
[ ]

( )
p(x)f' '+q(x)f .
Let f(x), g(x) be C
2
functions on [a, b]. Then
< >
L[f]; g
w
=

a
b

[ ] ( )
p(x)f' '+q(x)f g(x) dx
= p(x)f' g

b
a
+

a
b

( )
p(x)f' g' + q(x)f g dx
Also
< >
f; L[g]
w
=

a
b
f(x)
[ ] ( )
p(x)g' '+q(x)g dx
= p(x)f g'

b
a
+

a
b

( )
p(x)f' g' + q(x)f g dx.
Subtracting these two expressions,
< >
L[f]; g
w

< >
f; L[g]
w
= p(x)
( )
f' g f g'

b
a
.
This is Lagrange's identity.
11. Sturm-Liouville problems and orthogonal functions. 35
Lemma. If f, g are C
2
functions on [a, b] which satisfy

0
f(a) +
1
f'(a) = 0,
0
g(a) +
1
g'(a) = 0,

0
f(b) +
1
f'(b) = 0,
0
g(b) +
1
g'(b) = 0,
then
< >
L[f]; g
w
=
< >
f; L[g]
w
.
Proof. Suppose
0
0,
0
0, then by Lagrange's identity,
< >
L[f]; g
w

< >
f; L[g]
w
= p(x)
( )
f' g f g'

b
a
= p(b)

]
1
1
f'(b)

0
g'(b)

0
f' (b) g'(b)
+ p(a)

]
1
1
f'(a)

0
g'(a)

0
f' (a) g'(a)
= 0.
Similarly if
1
0,
1
0 or
1
0,
0
0 or
0
0,
1
0, the result follows with minor
changes to the argument.

Lemma. The eigenvalues of the Sturm-Liouville operator L are real.

Proof. Let be an eigenvalue with corresponding eigenfunction . Then satisfies
L[] =
1
w(x)
[ ] ( ) p(x)' '+q(x) = ,

0
(a) +
1
'(a) = 0,

0
(b) +
1
'(b) = 0.
Then
||||
w
2
=
< >
;
w
=
< >
;
w
=
< >
L[];
w
=
< >
;L[]
w
(above lemma)
=
< >
;
w
=
< >
;
w
= ||||
w
2
.
Therefore
( )||||
w
2
= 0.
Since 0, = and is real.

11. Sturm-Liouville problems and orthogonal functions. 36

Lemma. Let (x), (x) be eigenfunctions of the Sturm-Liouville boundary-value problem
with corresponding eigenvalues , respectively. Then If , (x), (x) are orthogonal on
[a, b] with respect to the weighted inner product
< >
f; g
w
.
Proof.

'

L[]=

0
(a)+
1
'(a)=0

0
(b)+
1
'(b)=0

'

L[]=

0
(a)+
1
'(a)=0

0
(b)+
1
'(b)=0
From a previous lemma,
< >
L[];
w
=
< >
;L[]
w
,
< >
;
w
=
< >
;
w
,
( )
< >
;
w
= 0.
Since ,
< >
;
w
= 0.

These results can be used to prove the following result:

Theorem. The eigenfunctions of a regular Sturm-Liouville problem are a countable set
{
n
(x); n = 1, 2, ... } and form an orthonormal basis for L
2
w
[a, b]. That is, each function f
L
2
w
[a, b] has a series expansion f =

n=1

c
n

n
where the c
n
=
< >
f;
n w
, n = 1, 2, ... , and
convergence is in the sense that
lim
N

a
b
w(x)

f(x)

n=1
N
c
n

n
(x)
2
dx = 0,
or
lim
N
||f

n=1
N
c
n

n
||
w
2
= 0.
If f(x) C
2
[a, b] and satisfies the boundary conditions
0
f(a) +
1
f'(a) = 0,

0
f(b) +
1
f'(b) = 0, the series f(x) =

n=1

c
n

n
(x) converges uniformly on [a, b]. That is
lim
N
max
x [a; b]

f(x)

n=1
N
c
n

n
(x) = 0.
12. Bessel Functions. 37
12. Bessel Functions.
Consider the heat equation on a disc 0 r < a, 0 < 2,,
u
t
u(r, ) =
u
t

2
u
r
2
+
1
r

u
r
+
1
r
2

2
u

2
= 0,
0 r < a, 0 < 2, t > 0, where u(r, , t) is the temperature, > 0 a constant.
Let the initial temperature at t = 0, be
u(r, , 0) = f(r, ),
0 r < a, 0 < 2, and boundary condition at r = a,
u(a, , t) = 0,
0 < 2, t > 0.
Assuming a solution of the form u(r, , t) = R(r) () T(t), substituting in the heat equation,
R T'

,
_
R"T+
1
r
R' T+
1
r
2
R"T = 0,
T'
T
=
R"
R
+
R'
rR
+
"
r
2

= constant =
2
(say),
T' +
2
T = 0,
r
2
R"
R
+
rR'
R
+
2
r
2
=
"

= constant =
2
(say),
" +
2
= 0, r
2
R" + rR' + (
2
r
2

2
)R = 0.
The equations for T, are familiar but the equation for R is Bessel's equation and has
non-constant coefficients. We can write it in the Sturm-Liouville form,

( )
rR' ' +

2
r
R =
2
rR,
where p(r) = r 0, q(r) =

2
r
> 0, w(r) = r 0.
A change of variable = r in Bessel's equation results in

2
d
2
S
d
2
+
dS
d
+ (
2

2
)S = 0,
where S() = R(r) = R

. We denote a solution of this Bessel's equation as S() = J

()
and then a solution of the original form of Bessel's equation is R(r) = S(r) = J

(r). A
second solution of Bessel's equation is Y

(r), a Bessel function of the second kind, and

therefore the general solution is
R(r) = AJ

(r) + BY

(r) .
12. Bessel Functions. 38
If 0, 1, 2, ..., then J

(r) is also a solution and

{ }
J

(r);J

(r) is a linearly
independent set of solutions to Bessel's equation. The general solution is then
R(r) = AJ

(r) + BJ

(r)
for constants A, B. The solution J

(r) is called a Bessel's function of the first kind and a

series representation can be found
J

(r) = (r)

]
1
1
1

2
r
2
2(2+2)
+

4
r
4
2.4.(2+2)(4+2)
+
....
.
For > 0, J

(0) = 0, J
0
(0) = 1. Y

(r), J

(r) are unbounded as r , hence the only

bounded solutions R(r) occur when B = 0.
The Bessel functions J

(r) are oscillatory for r > 0. For > 0, let the zeros of J

() be
{
m
; m = 0, 1, ... }, where
0
= 0 for all > 0.
Then R(r) = J

(r) solves the Sturm-Liouville problem

(rR'(r))' +

2
r
R(r) =
2
rR(r),
R(0) = J

(0) = 0, R(a) = J

(a) = 0,
if a =
m
, m = 1, 2, ... , for each > 0.
The Bessel functions

'

m
r
a

m=1
, are orthogonal on the interval (0, a), with respect to
the weighted inner product
< >
f; g
w
=

0
a
r f(r) g(r) dr. That is,

0
a
r J

m
r
a
J

l
r
a
dr = 0, if m l.
Returning to the heat equation on the disc 0 r < a, 0 < 2, the solutions of the form
u(r, , t) = R(r) () T(t) are given by
R(r) = AJ

(r) + BY

(r),
() = C cos () + D sin (),
T(t) = E e

2
t
.
For () to have period 2, = n = 0, 1, 2, ... .
For R(r) to be bounded at r = 0, B = 0.
12. Bessel Functions. 39
For u(a, , t) = 0, R(a) = J
n
(a) = 0 or a =
nm
, n = 0, 1, 2, ... , m = 1, 2, ... , the zeros of
J
n
(). That is
nm
=

nm
a
, n = 0, 1, 2, ... , m = 1, 2, ... .
By superposition, a solution of the heat equation on the disc satisfying the boundary
condition u(a, , t) = 0, is given by
u(r, , t) =

n=o

m=1

exp

nm
2
t
a
2
J
n

nm
r
a

( )
C
nm
cos(n)+D
nm
sin(n) .
The orthogonality relations

0
a
r J
n

nm
r
a
J
n

nl
r
a
dr = 0, if m l, and

cos(n) sin (k) d, if n k,

imply the orthogonality of

'

J
n

nm
r
a
cos(n);J
n

nm
r
a
sin(n)

n=0; m=1
on the rectangle
(r, ) (0, a)(, ) with respect to the inner product
< >
f; g =

0
a

r f(r, ) g(r, ) d dr.

That is,

0
a

_
J
n

nm
r
a
cos(n)

_
J
k

kl
r
a
cos(k) d dr
=

'

0; n k
0; n=kandml

0
a
rJ
k

kl
r
a
2
dr;n=kandm=l

0
a

_
J
n

nm
r
a
sin(n)

_
J
k

kl
r
a
sin(k) d dr
=

'

0; n k
0; n=kandml

0
a
rJ
k

kl
r
a
2
dr;n=kandm=l

0
a

_
J
n

nm
r
a
cos(n)

_
J
k

kl
r
a
sin(k) d dr = 0 for all n, m, k and l.
The coefficients C
nm
, D
nm
are determined using these orthogonality properties. For
example,
12. Bessel Functions. 40

0
a

r J
k

kl
r
a
cos(k) f(r, ) d dr = C
kl

0
a
r J
k

kl
r
a
2
dr,
hence
C
kl
=
1

0
a
rJ
k

kl
r
a
2
dr

0
a

r J
k

kl
r
a
cos(k) f(r, ) d dr.
Similarly,
D
kl
=
1

0
a
rJ
k

kl
r
a
2
dr

0
a

r J
k

kl
r
a
sin(k) f(r, ) d dr,
k = 0, 1, 2, ... , l = 1, 2, ... .
13. Fourier Transforms. 41
13. Fourier Transforms.
A function f : RR is said to be integrable on R if

|f(x)|dx < . We call the class of all

such functions L
1
(R). Similarly, f : RR is said to be square integrable on R if

|f(x)|
2
dx
< . and we call the class of all such functions L
2
(R).
Examples. Let f : RR and g : RR be defined by
f(x) =

'

2
3
;0<x<1

0;otherwise
,
g(x) =

'

2
3
; x>1

0;otherwise
.
Then

|f(x)|dx =

0
1
x

2
3
dx = 3x
1
3
|
1
0
= 3 <
and

|f(x)|
2
dx =

0
1
x

4
3
dx = 3x

1
3
|
1
0
= .
Therefore f L
1
(R) but f L
2
(R). On the other hand,

|g(x)|dx =

2
3
dx = 3x
1
3
|

1
=
and

|g(x)|
2
dx =

0
1
x

4
3
dx = 3x

1
3
|

1
= 3 < .
Therefore g L
1
(R) but g L
2
(R).

Given two function f L

1
(R) and g L
1
(R) , the product fg is not necessarily in L
1
(R).
Counter-examples are given by
f(x) = g(x) =

'

2
3
;0<x<1

0;otherwise
.
There is a product
*
for which f
*
g L
1
(R) whenever f L
1
(R) and g L
1
(R). We define
the convolution product
13. Fourier Transforms. 42
(f
*
g)(x) =

f(x y) g(y) dy.

Lemma. If f L
1
(R) and g L
1
(R), then f
*
g L
1
(R).
Proof.

|(f
*
g)(x)|dx =

f(x y)g(y) dy dx

|f(x y) g(y)| dydx

|f(x y)|

|g(y)| dydx

|f(x)| dx

|g(y)| dx < .

The Fourier transform of a function f L

1
(R) is defined to be
f
^
() =

f(x)e
ix
dx,
where R. It is easy to see that the Fourier transform of a function f L
1
(R) is a
bounded continuous function on R. The boundedness follows easily from
|f
^
()| =

f(x)e
ix
dx

|f(x)| dx < .
The continuity follows from the following.
Theorem. (Riemann-Lebesgue). If f L
1
(R), then f
^
C(R) and f
^
() 0 as || .
Proof. Let R, 0. Then
f
^
() =

f(x)e
ix
dx = e
i

f(x)e
ix
dx
=

f(x)e
i

,
_
x+

dx =

_
x

e
ix
dx.
Therefore
2|f
^
()| =

f(x)e
ix
dx

_
x

e
ix
dx

f(x) f

_
x

|e
ix
| dx
13. Fourier Transforms. 43

f(x) f

_
x

dx 0 as || .
To prove continuity of f
^
, let > 0 be given and a > 0 chosen such that

|x|>a
|f(x)|dx <

4
and
> 0 chosen such that 2a

|x|<a
|f(x)|dx < . Then for || < ,
| | f
^
(+) f
^
() =

( ) f(x)e
i(+ )x
f(x)e
ix
dx

|f(x)|
.
|e
ix
(e
ix
1)| dx = 2

|f(x)|
| |
sin
( )
x
2
dx
2

|x|>a
|f(x)|
| |
sin
( )
x
2
dx + 2

|x|<a
|f(x)|
| |
sin
( )
x
2
dx
2

|x|>a
|f(x)| dx + 2

|x|<a
|f(x)|
| |
x
2
dx

2
+ a

|x|<a
|f(x)|dx

2
+

2
= .
Therefore f
^
() is uniformly continuous on R.

We define the linear transformation F: f f

^
defined by
(Ff)() = f
^
() =

f(x)e
ix
dx.
F is called the Fourier integral transformation and f
^
the Fourier transform of f L
1
(R).
13. Fourier Transforms. 44
Properties of Fourier transforms.
1. Translation (a). Let f : R R and h R. The translate of f by h is the function
h
f
defined by
(
h
f)(x) = f(x h), x R.
The Fourier transform of
h
f is
( )

h
f
^
() =

(
h
)(x)e
ix
dx
=

f(x h)e
ix
dx
=

f(x )e
i(x+ h)
dx
= e
ih

f(x )e
ix
dx
= e
ih
f
^
().
(b). For real c,
( ) e
icx
f(x)
^
=

e
ix
( ) e
icx
f(x) dx
=

f(x )e
i(xc)
dx
= f
^
( c) =
( )

c
f
^
().
2. Dilation. Let R, > 0, the dilation of f by is defined as

f where
(

f)(x) =

1
2
f(
1
x), x R.
The Fourier transform of

f is
( )

f
^
() =

f)(x)e
ix
dx
=

1
2

f(
1
x) e
ix
dx
=

1
2

f(y) e
iy
dy
=

1
2
f
^
() =
( )

1
f
^
()
13. Fourier Transforms. 45
3. Differentiation. Let f and f' L
1
(R) and denote by D the differential operator D =

x
.
The Fourier transform of Df is
( )
Df
^
() =

(Df)(x)e
ix
dx
= f(x)e
ix
|

f(x)( ) De
ix
dx
= (i)f
^
().
By induction
( )
D
k
f
^
= (i)
k
f
^
, k = 1, 2, ... .
4. Multiplication. We denote by the differential operator =

. If f and xf L
1
(R)
then
( )
f
^
() =

f(x)e
ix
dx
=

(ix)f(x)e
ix
dx.
Therefore
( )
ixf
^
= f
^
. By induction it follows that
k
f
^
=
( )
(ix)
k
f
^
, k = 1, 2, ... .
5. Convolution. Let f, g L
1
(R). Then f
*
g L
1
(R) and has Fourier transform
(f
*
g)
^
() =

(f
*
g)(x)e
ix
dx
=

f(x y)g(y) dy e
ix
dx
=

f(x) g(y) e
i(x+ y)
dx dy
=

f(x)e
ix
dx

g(y)e
iy
dy
= f
^
()g
^
().
Example 1. Let f(x) = e

x
2
2
. Then f L
1
(R) and has Fourier transform
f
^
() =

e

x
2
2
e
ix
dx
13. Fourier Transforms. 46

f
^

() =

e

x
2
2
e
ix
dx
=

e

x
2
2
(ix)e
ix
dx
= i

,
_
e

x
2
2
e
ix
dx
= ie
ix
e

x
2
2
|

e

x
2
2

x
( ) e
ix
dx
=

e

x
2
2
e
ix
dx
= f
^
().
Therefore f
^
satisfies a first order ordinary differential equation with solution f
^
() =
ce

2
2
.
Since f
^
(0) = c =

e

x
2
2
dx =

2,

,
_
e

x
2
2

^
=

2 e

2
2
.

Example 2. Let f(x) =

'

1;|x| a

0;otherwise

a
(x) L
1
(R) and has Fourier transform
( )

a
^
() =

a
(x) e
ix
dx =

a
a
e
ix
dx
=
1
i
e
ix
|
a
a
=
1
i

( )
e
ia e
ia
=
2sina

.
14. Inverse Fourier Transforms. 47
14. Inverse Fourier Transforms.
The Fourier transform of f L
1
(R) is continuous, bounded and |f
^
()| 0 as || .
However f
^
is not neccesarily in L
1
(R). Assuming f L
1
(R), f piecewise smooth, and
f
^
L
1
(R), define
f
a
(x) =
1
2

a
a
f
^
() e
ix
d
=
1
2

a
a

f(y)e
iy
dy e
ix
d
=
1
2

a
a
e
i(xy)
d f(y) dy
=
1
2

,
_
e
ia(xy)
e
ia(xy)
i(x y)
f(y) dy
=

D
a
(x y) f(y) dy =

D
a
(y) f(x y) dy
=
( )
D
a
*
f (x)
where D
a
(x ) =
sinax
x
.
Now
f
a
(x)
1
2
[ ]
f(x

)+f(x
+
) =
1

f(x y)
sinay
y
dy
1
2
[ ]
f(x

)+f(x
+
)
=
1

[ ]
f(x y)f(x

)
sinay
y
dy
+
1

[ ]
f(x y)f(x
+
)
sinay
y
dy

_
since

sinay
y
dy=

2
=
1

[ ]
f(x y)f(x

)
sinay
y
dy
+
1

[ ]
f(x+y) f(x
+
)
sinay
y
dy
14. Inverse Fourier Transforms. 48
In the second integral,

1

[ ]
f(x+y) f(x
+
)
sinay
y
dy
=
1

0
K

[ ]
f(x+y) f(x
+
)
sinay
y
dy +
1

[ ]
f(x+y) f(x
+
)
sinay
y
dy
If K 1,

f(x+y)
sinay
y
dy

| |
f(x+y) dy
and

f(x
+
)
sinay
y
dy = f(x
+
)

sinay
y
dy.
Since

|f(x)|dx ,

sinay
y
dy are both convergent integrals,

sinay
y
dy 0,

| |
f(x+y) dy 0 as K .
For the integrals over [0, K],

0
K

[ ]
f(x+y) f(x
+
)
sinay
y
dy =

,
_
e
iay
e
iay
2i
g(y) dy
=
1
2i

[ ]
g
^
(a)g
^
(a)
where g(y) =

'

f(x+y) f(x
+
)
y
;0<y<K

0;otherwise
.
Since f is piecewise smooth, f'(x
+
) exists for all x R and lim
y0
+
g(y) = f'(x
+
). Therefore
g is bounded on [0, K] and hence g L
1
(R). By the Riemann-Lebesgue lemma, g
^
exists, is
continuous, g
^
(ta) 0 as a and therefore

0
K

[ ]
f(x+y) f(x
+
)
sinay
y
dy 0 as a
for K 1. A virtually identical argument works for the first integral.
Therefore
| |
f
a
(x)
1
2
[ ]
f(x

)+f(x
+
) 0 as a
or
lim
a
f
a
(x) = lim
a

1
2

a
a
f
^
() e
ix
d
=
1
2

f
^
() e
ix
d =
1
2
[ ]
f(x

)+f(x
+
) .
Summarising
14. Inverse Fourier Transforms. 49
Theorem. Let f L
1
(R) and let f be piecewise smooth on R. Then for every x R,

1
2

f
^
() e
ix
d =
1
2
[ ]
f(x

)+f(x
+
) .
If x is a point of continuity of f, then
1
2

f
^
() e
ix
d = f(x).

For L
1
(R), we call

(x) =
1
2

() e
ix
d the inverse Fourier transform of and
F
1
defined by ( ) F
1
(x) =
1
2

() e
ix
d in called the inverse Fourier transformation.
If is the Fourier transform of a piecewise smooth function f L
1
(R), that is = f
^
, then
we define f(x) =
1
2
[ ]
f(x

)+f(x
+
) at a point of discontinuity of f.
Then f(x) =

(x) =
1
2

f
^
() e
ix
d. That is F
1
Ff = f.
Example. The inverse Fourier transform is useful in computing Fourier transforms.
Since F
( )
e
a|x|
=
2a
a
2
+
2
, it follows that F
1
F
( )
e
a|x|
= F
1

_
2a
a
2
+
2
or
e
a|x|
= F
1

_
2a
a
2
+
2
. Interchanging the roles of x and , multiplying by 2, leads to
F

,
_
2a
a
2
+x
2
= 2 e
a||
or F

,
_
1
a
2
+x
2
=

a
e
a||
. In a similar fashion, every Fourier
transform pair defines a dual pair using the inverse Fourier transform.
15. Applications to Differential Equations. 50
15. Applications to Differential Equations.
1. The wave equation. The equation

2
u
t
2
c
2

2
u
x
2
= 0, < x < , t > 0,
describes the vertical vibrations of an infinite stretched elastic string, where u(x, t) is the
vertical displacement of the string from its rest position at position x, time t. Let the initial
displacement and velocity be given as
u(x, 0) = f(x),
u
t
(x, t) = g(x), < x < .
We take Fourier transforms of the wave equation and the initial conditions with respect to
the x variable and denote by u
^
(, t ) the Fourier transform F
( )
u(x, t) . The using the
derivative properties of the Fourier Transform,

2
u
^
t
2
(ci)
2
u
^
= 0, < < , t > 0,
or

2
u
^
t
2
+ (c)
2
u
^
= 0.
Then
u
^
(, t) = A() cos(ct) + B() sin(ct).
When t = 0, u
^
(, t) = A() = f
^
(),
u
^
t
(, t) = g
^
() = cB(). Therefore
u
^
(, t) = f
^
() cos(ct) +
g
^
()
2c

2sin(ct)

= f
^
()

_ e
i(ct)
+e
i(ct)

2
+
1
2c

( )

ct
*
g
^
().
Using the translation property and the convolution theorem,
u(x, t) =
1
2

[ ]
f(x+ct)+f(xct) +
1
2c

ct
(x ct) g(y) dy
=
1
2

[ ]
f(x+ct)+f(xct) +
1
2c

|xct|<ct
g(y) dy
=
1
2

[ ]
f(x+ct)+f(xct) +
1
2c

x ct
x + ct
g(y) dy.
This is d'Alemberts solution to the one-dimensional wave equation.
15. Applications to Differential Equations. 51
2. Laplace's Equation.
Consider Laplace's equation in two variables on the upper half-plane y > 0. Then

2
u
x
2
+

2
u
y
2
= 0, < x < , y > 0.
Let the boundary condition
u(x, 0) = f(x), < x <
be given for a function f L
1
(R). Then taking the Fourier transform in the variable x,
(i)
2
u
^
(, y) +

2
u
^
y
2
(, y) = 0, < x < , y > 0.
or

2
u
^
y
2
(, y)
2
u
^
(, y) = 0,
which has solutions
u
^
(, y) = A()e
y
+ B()e
y
.
We need two conditions to determne the functions A(), B(). In addition to u(x, 0) = f(x),
< x < , let
u
y
(x, 0) = g(x), < x < for some function g L
1
(R). We will find g
such that the solution u(x, y) is bounded for y > 0, in fact such that u(x, y) 0 as y .
Taking transforms,
u
^
(, 0) = f
^
(),
u
^
y
(, 0) = g
^
(),
or
A() + B() = f
^
(), A() B() = g
^
().
Solving for A(), B(),
A() =
1
2

_
f
^
()+
g
^
()

, B() =
1
2

_
f
^
()
g
^
()

,
and
u
^
(, y) =
1
2

_
f
^
()+
g
^
()

e
y
+
1
2

_
f
^
()
g
^
()

e
y
.
For > 0, u
^
(, y) 0 as y 0 if and only if f
^
() +
g
^
()

= 0, and for < 0, u

^
(, y) 0 as
y 0 if and only if f
^
()
g
^
()

= 0. Therefore
g
^
() =

'

f
^
();>0

f
^
();<0
= || f
^
().
Hence
15. Applications to Differential Equations. 52
u
^
(, y) =

'

f
^
()e
y
;>0

f
^
()e
y
;<0
= f
^
()e
|y|
.
Since e
|y|
=

_
y
(x
2
+y
2
)
^
, by the convolution theorem,
u(x, y) =

y
(x
2
+y
2
)
*
f =

y
((x s)
2
+y
2
)
f(s) ds.
3. The Heat Equation.
The heat equation,
u
t

2
u
x
2
= 0, < x < , t > 0
for u(x, t) a function of two variables, with initial condition
u(x, t) = f(x), < x <
for f L
1
(R), f continuous and bounded, can be solved using Fourier transforms. Taking
transforms in the variable x,
u
^
t
(, t) (i)
2
u
^
(, t) = 0,
or
u
^
t
(, t) +
2
u
^
(, t) = 0.
Solving this first order ordinary differential equation,
u
^
(, t) = u
^
(, 0)e

2
t
= f
^
()e

2
t
.
Now

,
_
e

x
2
2

^
=

2 e

2
2
and using the dilation property of Fourier transforms

1
2
e

(
1
x)
2
2
^
=

1
2
e

()
2
2
e

()
2
2
=
1

_
e

(
1
x)
2
2
^
.
Let

2
2
= t or =

2t , then
15. Applications to Differential Equations. 53
e

2
t
=
1

2

1

2t

,
_
e

1
2

,
_
x

2t
2
^
=
1

4t

,
_
e

1
2

,
_
x

2t
2
^
.
Therefore from the convolution theorem, since
u
^
(, t) = f
^
()
1

4t

,
_
e

1
2

,
_
x

2t
2
^
u(x, t) =

_
f
*
1

4t
e

1
2

,
_
x

2t
2
(x)
=
1

4t

(xy)
2
4t
f(y) dy.
The function h(x, t) =
e

x
2

4t

4t
, is called the heat kernel. We can then write
u(x, t) =

h(x y, t) f(y) dy.

16. Plancherels' and Parsevals' Identities.
54
16. Plancherels' and Parsevals' Identities.
By the convolution theorem, for f, g L
1
(R)L
2
(R),
( )
f
*
g
^
= f
^
g
^
.
Therefore
f
*
g =
( )
f
^
g
^

or

f(x y) g(y) dy =
1
2

e
i x
f
^
() g
^
() d.
Set x = 0,

f( y) g(y) dy =
1
2

f
^
() g
^
() d
Replacing g(x) by g(x) , the Fourier transform g
^
() is replaced by g
^
() , hence

f( y) g(y) dy =
1
2

f
^
() g
^
() d
or

f( y) g(y) dy =
1
2

f
^
() g
^
() d.
This is Plancherels' identity. When f = g we obtain Parsevals' identity,

|f( y)|
2
dy =
1
2

|f
^
()|
2
d.
Examples.
1. Let f(x) =
a
(x) =

'

1;|x|< a

0;otherwise
.
Then f
^
() =
2sina

, and by Parsevals' identity,

|
a
(x)|
2
dx =
1
2

sina

2
d

_
2sina

2
d = 4a
or

_
sina

2
d = a.
16. Plancherels' and Parsevals' Identities.
55
2. Let f(x) = e
a|x|
, f
^
() =
2a
a
2
+
2
and by Parsevals' identity

( )
e
a|x| 2
dx =
1
2

2a
a
2
+
2
2
d
1
a
=
2a
2

1
(a
2
+
2
)
2
d
or

1
(a
2
+
2
)
2
d =

2a
3
.
17. Band Limited Functions and Shannon's Sampling Theorem.
56
17. Band Limited Functions and Shannon's Sampling Theorem.
The Fourier transform variable has the role of frequency and f
^
() is referred to as the
frequency representation of f(x).
If f
^
() = 0 for || >
c
> 0, then f(x) is called a band-limited function and
c
is called the
cut-off frequency. Many functions from science and technology, are band-limited. For
example, human hearing is assumed to be limited to frequencies below about 20 kHz.
Therefore the acoustic signals recorded on compact discs are limited to a bandwidth of 22
kHz.
A first step in the processing of signals, is sampling. A signal represented by a continuous
function f(x), is replaced by its samples at regular intervals, {f(nL), n = 0, t1, t2, ... }.
Shannons' Theorem shows that it is possible to exactly recover the band-limited
continuous function f(x) from knowledge of its samples, provided the sampling interval L,
is sufficiently short.
Consider the functions
^
n
() given by

^
n
() =

'

0;| |>
c

2
c

e

in

c
; ||
c
The inverse Fourier transforms of
^
n
() are given by

n
(x) =
1
2

e
i x

^
n
() d
2
n
(x) =

c
e
i x

1

2
c

e

in

c
d
=
1

2
c

c
e
i

,
_
x
n

c
d
=
1

2
c

]
1
1
1
e
i

,
_
x
n

c
i

,
_
x
n

c
=

2
c

_
e
i (
c
xn)
e
i (
c
xn)
2i(
c
xn)

2
c

sin(
c
xn)

c
xn
17. Band Limited Functions and Shannon's Sampling Theorem.
57
=

2
L

sin
c
(xnL)

c
(xnL)
where L =

c
.
Consider the inner products
< >

^
n
;
^
m
=

^
n
()
^
m
() d
=
1
2
c

c
e

in

_
e

im

c
d
=
1
2
c

c
e

i(mn)

c
d
=

'

0; mn

1; m=n
.
So the functions {
^
n
() ; n = 0, t1, t2, ... }, form an orthogonal set in L
2
(
c
,
c
). Since f
^
()
is band limited to ||
c
, it has a Fourier series
f
^
() =

n=

c
n

^
n
()
=

n=

c
n

1

2
c

in

c
d
where the Fourier coefficients c
n
are given by
c
n
=
< >
f
^
;
^
n
=

c
f
^
()
1

2
c

in

c
d.
By Plancherel's theorem
c
n
=
< >
f
^
;
^
n
= 2
< >
f;
n
so
f
^
() =

n=

c
n

^
n
() =

n=

2
< >
f;
n

^
n
().
Taking inverse Fourier transforms,
f(x) =

n=

< >
f;
n
2
n
(x)
=

n=

< >
f;
n

2
L

sin
c
(xnL)

c
(xnL)
.
17. Band Limited Functions and Shannon's Sampling Theorem.
58
The samples at intervals of length L =

c
are
f(kL) =

n=

< >
f;
n

2
L

sin
c
(kLnL)

c
(kLnL)
=

n=

< >
f;
n

2
L

sin(kn)
(kn)
=
< >
f;
k

2
L
.
So
< >
f;
k
= f(kL)

L
2
, k = 0, t1, t2, ... .
Finally therefore
f(x) =

n=

< >
f;
n

2
L

sin
c
(xnL)

c
(xnL)
=

n=

f(nL)
sin
c
(xnL)

c
(xnL)
.
Summarising we have Shannon's theorem,
Theorem. Let f L
1
(R)L
2
(R) be continuous and band limited to ||
c
. Then
f(x) =

n=

f(nL)
sin
c
(xnL)

c
(xnL)
where L =

c
.

The relationship L = 2 , where is the frequency of sampling in cycles per unit length,
shows that from Shannon's theorem, to reconstruct a band-limited function, it suffices to
sample at a frequency =
2
L
= 2
c
, twice the cut-off frequency.
18. Heisenbergs Inequality.
59
18. Heisenbergs Inequality.
Let f L
2
(R), xf L
2
(R). Then the quantity

a
f

(xa)
2
|f(x)|
2
dx

|f(x)|
2
dx
is called the dispersion about the point x = a of f. The reasoning behind the definition is
that if f(x) is concentrated near x = a , then
a
f is smaller than when f is not close to zero
far from x = a.
Example. Consider the characteristic function

b
(x) =

'

1;|x| b

0;otherwise
which has Fourier transform
(
b
)
^
() =
2sinb

.
Notice that
b
is concentrated near x = 0 for small b. The dispersion about the origin is

b
=

b
b
x
2
dx

b
b
dx
=
b
2
3
and it clear that the dispersion increases as b increases.
Notice that the Fourier transform
^
b
does not have a finite dispersion about the origin,
since

2
|
^
b
()|
2
d =

_
2sin(b)

2
d
= 4

sin
2
(b) d = .
This indicates that
^
b
is spread out away from x = 0.
The following result shows that there is a type of inverse relationship between the
dispersion of a function and that of its Fourier transform.
Theorem. Let f L
1
(R)L
2
(R). Then for all a, R,
18. Heisenbergs Inequality.
60
( )

a
f
( )

f
^
=

x
2
|f(x)|
2
dx

|f(x)|
2
dx

2
|f
^
()|
2
d

|f
^
()|
2
d

1
4
and equality holds if and only if f(x) = c e
kx
2
for constants c R and k > 0.
Proof. We firstly prove the result for a = = 0.

x
2
|f(x)|
2
dx and

2
|f
^
()|
2
d are
both assumed finite since otherwise the result is trivial.
Let f
*
(x) ((if
^
())

or (f
*
)
^
() = if
^
(). Then f
*
L
2
(R), and

x
2
|f(x)|
2
dx

2
|f
^
()|
2
d =

x
2
|f(x)|
2
dx

|(i)f
^
()|
2
d
=

x
2
|f(x)|
2
dx

|(f
*
)
^
()|
2
d
= 2

x
2
|f(x)|
2
dx

|f
*
(x)|
2
dx (by Parsevals identity).
Since

]
1
1

1
2
x
( )
f
*
(x) f(x) + f
*
(x) f(x) dx
2
=

]
1
1

xRe
( )
f
*
(x) f(x) dx
2
=

]
1
1
Re

xf (x) f
*
(x) dx
2

xf (x) f
*
(x) dx
2

x
2
|f(x)|
2
dx

|f
*
(x)|
2
dx
(by the Cauchy-Schwartz inequality).
We will show that

x
( )
f
*
(x) f(x) + f
*
(x) f(x) dx =

|f(x)|
2
dx
from which the result follows, for then

x
2
|f(x)|
2
dx

2
|f
^
()|
2
d = 2

x
2
|f(x)|
2
dx

|f
*
(x)|
2
dx
2

]
1
1

1
2
x
( )
f
*
(x) f(x) + f
*
(x) f(x) dx
2
18. Heisenbergs Inequality.
61
=

]
1
1

|f(x)|
2
dx
2
=

|f(x)|
2
dx

_ 1
2

|f
^
()|
2
d
=
1
4

|f(x)|
2
dx

|f
^
()|
2
d .
To complete the proof, we assume that f is continuous and piecewise smooth, This
assumption can be removed since functions in L
1
(R) are the uniform limit of such
functions.
Then from the property of Fourier transforms,
f
*
(x) = f'(x)
wherever the derivative exists. Then for any interval [a, b] ,
b |f(b)|
2
a |f(a)|
2
=

a
b

d
dx
( ) x|f(x)|
2
dx
=

a
b

( )
xf' (x) f(x) +x f(x) f' (x) + |f(x)|
2
dx
=

a
b
x
( )
f
*
(x) f(x) + f
*
(x) f(x) dx +

a
b
|f(x)|
2
dx
The assumption f L
2
(R) implies that b |f(b)|
2
0 as b and a |f(a)|
2
0 as a
since otherwise |f(x)| > c |x|
1/2
as |x| , which is not integrable. Taking the limit as
b and a ,
0 =

x
( )
f
*
(x) f(x) + f
*
(x) f(x) dx +

|f(x)|
2
dx
as required.
As for the case of equality in Heisenbergs inequality, this holds if and only if f(x) f
*
(x) is
real and f
*
(x) = K x f(x) for some complex constant K. That is,
f(x) f
*
(x) = f(x) K xf(x) = x |f(x)|
2
K is real.
Therefore K is real.
The differential equation
f'(x) = f (x) = K x f(x)
has solutions of the form
f(x) = c e

Kx
2
2
, c any real constant,
18. Heisenbergs Inequality.
62
and f(x) = c e

Kx
2
2
L
2
(R) if and only if K > 0. Therefore equality holds in Heisenbergs
inequality only if f(x) = c e
kx
2
for constants c R and k > 0.
Conversely, let f(x) = e

Kx
2
2
for constant K > 0. Then

x
2
|f(x)|
2
dx

2
|f
^
()|
2
d = 2

x
2
|f(x)|
2
dx

|f
*
(x)|
2
dx
= 2

x
2
|f(x)|
2
dx

|K xf(x)|
2
dx
= 2 K
2

x
2
|f(x)|
2
dx
2
= 2 K
2

x
2
e
Kx
2
dx
2
= 2 K
2

]
1
1

,
_
x
2K

( )
2Kxe
Kx
2
dx
2
= 2 K
2

]
1
1

,
_
x
2K

( )
2Kxe
Kx
2
dx
2
= 2 K
2

]
1
1

,
_
x
2K

d
dx

( )
e
Kx
2
dx
2
= 2 K
2

]
1
1

,
_
1
2K
e
Kx
2
dx
2
(integration by parts)
=

]
1
1

e
Kx
2
dx
2
=

2K

]
1
1

e
t
2
dt
2
=

2
2K
.
Whereas,

1
4

|f(x)|
2
dx

|f
^
()|
2
d =
1
4

|f(x)|
2
dx 2

|f(x)|
2
dx
=

|f(x)|
2
dx
2
18. Heisenbergs Inequality.
63
=

]
1
1

e
Kx
2
dx
2
and equality holds.
The case of a 0, 0, follows by observing that F(x) = e
ix
f(x + a) satisfies the same
hypotheses as f(x) and
a
f =
0
F and

f
^
=
0
F
^
for any a 0, 0.
As a consequence of the inequality
( )

a
f
( )

f
^

1
4
, we see that it is impossible for
both
a
f and

f
^
to be simultaneously small. That is, if one of
a
f or

f
^
is very small
then the other must be large.