2.

Fourier Transforms
How do we decompose a signal into its frequency components? Ba-
sic properties. The time domain and the frequency domain. Fre-
quency transfer functions and frequency response functions. What is
the fft?
2.1 Introduction
We have seen how to decompose a signal in the time domain into its
harmonic components when the signal is a periodic function that has a
well-dened fundamental period T (and frequency = 2/T). This gives
a discrete spectrum (set of frequencies).
But many signals when represented in the frequency domain have a
continuous spectrum. These represent signals in the time domain that
do not have a single fundamental frequency. That is they are not periodic
functions.
Of course, in engineering, most signals are not periodic; e.g. the ac-
celerations in earthquake, random vibrations, chaotic outputs of simple
nonlinear circuits, freak waves such as tsunamis the Severn bore, etc.. . .
So, we need a technique to nd the frequency content of an arbitrary
function f(t). The only stipulation is that the signal contain a nite
amount of energy. A very conservative way of achieving this is to assume
that
f(t) 0 as t .
In particular, we need the approach to 0 not to be too slow:
16
Such functions could have a nite duration, e.g.
f(t) =
_
1 for 1 < t < 1
0 otherwise
or have decaying tails
f(t) = e
|t|
.
We will motivate the idea of the Fourier Transform as the analogue of
the Fourier series but with continuous frequencies rather than discrete
frequencies
n
= 2n/T.
f(t) =
1
2
_

F()e
jt
d
F() =
_

f(t)e
jt
dt
Denition: We say that F() is the Fourier transform of f(t)
and f(t) is the inverse Fourier transform of F().
Sometimes f(t) and F() are called a Fourier transform pair
and written:
f(t) F()
17
2.2. When does this integral even exist?
(reference: James p 365)
Require
_

0
|f(t)|dt <
(for example, f(t) = 1/t doesnt satisfy this, even though it ap-
proaches 0 as t approaches ).
For technical reasons we also require that f has at most a nite number
of maxima and minima and a nite number of discontinuities in any
nite interval. For example f(t) = sin(
1
t
) doesnt work.
2.2. Why does this work?
(reference: James p 330 for the complex version of fourier series; 365++
for the Fourier transform)
Now, the Fourier series for a periodic function with xed frequency
0
can be expressed conveniently as a complex function if we remember that
cos(
0
t) =
1
2
(e
j
0
t
+ e
j
0
t
)
and
sin(
0
t) =
1
2j
(e
j
0
t
e
j
0
t
) =
j
2
(e
j
0
t
e
j
0
t
).
18
Hence the Fourier series
f(t) =
a
0
2
+

n=1
(a
n
cos(
0
nt) + b
n
sin(
0
nt))
=
a
0
2
+
_
a
n
2
[e
j
0
nt
+ e
j
0
nt
] +
jb
n
2
[e
j
0
nt
e
j
0
nt
]
_
=
a
0
2
+
1
2

n=1
(a
n
jb
n
)e
j
0
nt
+ (a
n
+ jb
n
)e
j
0
nt
=
a
0
2
+
1
2

n=1
(c
n
e
j
0
nt
+ c
n
e
j
0
(n)t
)
=

c
n
e
j
0
nt
, (2.1)
where c
n
= a
n
jb
n
, c
n
= a
n
+ jb
n
and, by denition of the Fourier
series, we have
c
n
=
1
T
_ T
2

T
2
f(t)e
jn
0
t
dt (2.2)
This is the complex representation of the same old Fourier series that we
have been computing for the past week.
How are Fourier transforms and Fourier series related?
Suppose we let T and
0
0.
We let the discrete variable n
0
be replaced by the continuous vari-
able . Thus is a continuous analogue of the integer variable n
0
.
Then we have
1
T
=

0
2
=

2
and
c
n

2
_

f(t)e
jt
dt
19
Convention is to let 2c
n
represent the amount of frequency in
the signal
Setting
2c
n
= F()d
gives
F() =
_

f(t)e
jt
dt

F()d = amount of
the total frequency that's
between and + d
The set of coecients c
n
has become the function F() and the sum
over n has become the integral over .
20
2.3 Notation, notation, notation
Commonly used notations for the representation of data in the time
domain and frequency domain are:
F denoting Fourier transform of . . .
F [f(t)] F() =
_

f(t)e
jt
dt
and F
1
denoting inverse Fourier transform of . . .
F
1
[F()] = f(t) =
1
2
_

F()e
jt
d
Sometimes in place of f(t) F() we use other notations such as
x(t) X() and f(t) F().
The Fourier Transform F[f(t)] = F() is in general complex i.e.
F() = R() + jX() = |F()|e
j()
|F()| is the magnitude spectrum (or amplitude spectrum)
() is the phase spectrum (it is equal to tan
1
(X/R))
Warning: Some text books dene the Fourier transform with an extra
factor

f(t) =

2f(t),

F(t) = F(t)/

f(t) =
1

2
_

F()e
jt
d
and

F() =
1

2
_

f(t)e
jt
dt
21
Worked example 2.1 A unit impulse is applied to an electronic
circuit and is found to give an impulse response corresponding to the
following function:
f(t) =
_
0 for t < 0
e
t
for t > 0
( > 0).
Find the Fourier transform of f(t) and plot the magnitude and phase
spectra.
Note, such a function f(t) in the above example is often written H(t)e
t
,
where H is the so-called Heaviside step function
H(t) =
_
0 for t < 0
1 for t 0
2.4 Properties of the Fourier transform
Linearity
F[ax(t) + by(t)] =
_

[ax(t) + by(t)]e
jt
dt
= a
_

x(t)e
jt
dt + b
_

y(t)e
jt
dt
If x(t) X() and y(t) Y ()
then ax(t) + by(t) aX() + bY ()
Symmetry: by denition :-
x(t) =
1
2
_

X()e
jt
d
22
Replacing t by t :-
2x(t) =
_

X()e
jt
d
Interchanging t and :-
2x() =
_

X(t)e
jt
dt = F[X(t)]
If x(t) X() then X(t) 2x()
Time-delay
F[x(t t
0
)] =
_

x(t t
0
)e
jt
dt
Put t

= t t
0
so dt

= dt and
F [x(t

)] =
_

x(t

)e
j(t

+t
0
)
dt

= e
jt
0
_

x(t

)e
jt

dt

= e
jt
0
X()
If x(t) X() then x(t t
0
) X()e
jt
0
The Convolution Theorem
The convolution of two functions x(t) and y(t) is dened by the function:
f(u) =
_

x(v)y(u v)dv
and is generally written :
f(u) = x(u) y(u)
23
Convolution theorem :
F[x(t) y(t)] = X()Y ()
or
F
1
[X()Y ()] = x(t) y(t)
Proof
F
__

x(u)y(t u)du
_
=
_

e
jt
(
_

x(u)y(t u)du)dt
by denition of the fourier transform
=
_

x(u)
__

e
jt
y(t u)dt
_
du
where we have now switched order of integration
Now we use the result for delays to write the integral in brackets in terms
of the fourier transform of y:
__

e
jt
y(t u)dt
_
= e
ju
Y ()
so that
F
__

x(u)y(t u)du
_
=
_

x(u)
_
e
ju
Y ()

du
= Y ()
_

x(u)e
ju
du
= Y ()X()
24
and we can invert this as well:
F
1
[X()Y ()] =
_

x(u)y(t u)du = x(t) y(t)

Dierentiation
F
_
df(t)
dt
_
= (j)F()
F
_
d
n
f(t)
dt
n
_
= (j)
n
F()
Worked example 2.4 Prove this!
2.5 Frequency response function
Worked example 2.3 Show that if the time signals y(t) and u(t)
have Fourier transforms Y () and U() respectively, and if y(t) and
u(t) satisfy the dierential equation
d
2
y(t)
dt
2
+ 3
dy(t)
dt
+ 7y(t) = 3
du(t)
dt
+ 2u(t), (2.3)
then Y () = G()U() for some function G().
In this example, Y () is called the frequency response function
of the linear ordinary dierential equation (2.3) corresponding to input
u(t).
We say that Y () is the representation in the frequency domain of
the output y(t) and U() is the representation of the input to the linear
system in the frequency domain.
In contrast, y(t) and u(t) represent the inputs and outputs in the time
domain.
25
The function G() which is the multiplicative factor between the input
and the output in the frequency domain is called the frequency transfer
function.
These notations are very important in vibration analysis and in AC
circuit theory. Often we represent a system by its frequency transfer
function, which enables us to represent the output of that system as a
simple mapping between the inputs and the outputs in the frequency
domain. This leads to problems in system identication
2.6 The fast Fourier transform (t)
In 1965 Cooley and Tukey came up with a really clever algorithm for
computing the Fourier transform exceptionally quickly directly from data.
This is known as the fast Fourier transform and has revolutionised engi-
neering data processing. For example there is a simple MATLAB function
called fft that can calculate approximate Fourier transforms of data.
Thus if we have the output of any experiment, data capture in the eld,
or numerical model in the time domain. In the blink of an eye we can
transform that data into the frequency domain.
There are many variants of this algorithm available in almost all data
analysis software e.g. the fastest Fourier transform in the West) fftw.
26
2.7 Summary
Fourier transform:
F() =
_

f(t)e
jt
dt
inverse transform:
f(t) =
1
2
_

F()e
jt
d
The frequency transfer function for a linear system with input
u(t) and output y(t) is G() such that Y () = G()U(). Y () is
the frequency response function.
Useful formulae The following are often useful for Fourier series
and Fourier transforms (Chapters 1 and 2)
cos(n) = (1)
n
, sin(n) = 0, e
in
= (1)
n
_
b
a
f(t)dt =
_
a
b
f(t)dt,
_
b
a
f(t)dt =
_
0
a
f(t)dt +
_
b
0
f(t)dt
See James pp 365 - 381 for details, worked examples, questions and so
on.
27