The Discrete Fourier Transform (DFT)

Introduction
Here we will start with the Fourier series of a band-limited, periodic function and see what happens when it
is sampled using Ideal sampling. The result will be the Discrete Fourier Transform (DFT). We can then
understand the characteristics of periodic, discrete-time signals in the frequency domain.
The Periodic Signal
Assume x(t) is a band-limited, periodic function with maximum frequency W and period T. Its complex
exponential series can then be written as:
( )

+

2
1
2
2
*
N
N
i
t
T
i
j
i
X t x

Where N/2 WT is sufficient to cover all the non-zero coefficients (those with low enough energy to be
approximated away) and:
( )

T t
T
i
j
i
dt t x
T
X
0
2
*
1

The Fourier Transform, X(j), of x(t) is then a finite sum of -functions and can be sketched as a function
of frequency.
X(f)
0
0.2
0.4
0.6
0.8
1
1.2
-6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6
Frequency index (i)
J. N. Denenberg Copyright DTS June 13, 2005
DFT Tutorial Page 2 of 4
Sampling x(t)
If we now sample (we shall assume Ideal sampling to simplify the derivation) x(t) at the Nyquist rate of 2W,
the spectrum becomes periodic as was determined in the discussion on sampling. The spectrum also gets a
gain factor (not discussed earlier) which is introduced in the sampling process.
Assume that we sample f(t) ideally over the time interval (0, a) with n samples.
( )

,
_

,
_

N
a
k t
N
a
k f T t f
N k
k
* ,
1
0
Integrating (using the sifting property of -functions)
( )

,
_

1
0
0
,
N k
k
a
N
a
k f dt T t f
But
( )

,
_

,
_

1
0
1
0
0
*
N k
k
N k
k
a
N
a
k f
N
a
N
a
N
a
k f dt t f
So the sampled function is N/a time stronger than the original. We can adjust for this by
scaling the impulse train by T/M in our development.
And
( )

,
_

,
_

N
a
k t
N
a
k x
M
T
T t x
N k
k
* ,
1
0
Where M is the largest integer that corresponds to a frequency that is less than 2TW at the selected sampling
rate.
The Fourier series for the discrete-time signal is then

,
_

2
1
2
2
*
N
N
i
ik
M
j
i k
X x
M
kT
t x

Or, if we sample at exactly the Nyquist rate, M=N and

2
1
2
2
*
N
N
i
ik
N
j
i k
X x

Or Since the X
i
are periodic with period N

1
0
2
*
N
i
ik
N
j
i k
X x

Which is one half of the two DFT equations

J. N. Denenberg Copyright DTS June 13, 2005
DFT Tutorial Page 3 of 4
The second half of the DFT is obtained by starting with
( )

T t
T
i
j
i
dt T t x
T
X
0
2
* ,
1

We can now substitute the sampling sum of -functions for x(t,T)

1
]
1

,
_

T t
T
i
j
N k
k
k i
dt
N
T
k t x
N
T
T
X
0
2
1
0
* *
1


Interchanging the order of integration and summation and again using the sifting property of -functions
yields the second half of the DFT pair
ik
N
j
N k
k
k i
x
N
X

2
1
0
*
1

Often these two equations are simplified by assuming that

N
j
N
W

Resulting in the DFT transform pair

Direct DFT
ik
N
k
k i
W x
N
X

*
1
1
0
Inverse DFT

1
0
*
N
i
ik
i k
W X x
It should be noted that the 1/N factor is sometimes placed in the Inverse DFT equation. This can be a source
of confusion.
DFT: The Matrix Formulation
The equations above can be restated as a pair of matrix equations. Let
( ) ( )
( ) ( )
( ) ( )
1
1
1
1
1
1
]
1

2
2
1 1 0
2 2 0
1 2 1 0
0 0 0 0
...
... ...
...
...
n n
n n
n n
W W W
W W W
W W W W
W W W W
W

then, its inverse is

( ) ( )
( ) ( )
( ) ( )
1
1
1
1
1
1
]
1

2
2
1 1 0
2 2 0
1 2 1 0
0 0 0 0
1
...
... ...
...
...
n n
n n
n n
W W W
W W W
W W W W
W W W W
W

J. N. Denenberg Copyright DTS June 13, 2005

DFT Tutorial Page 4 of 4
Now define
( )
( )
1
1
1
1
1
1
]
1

1
2
1
0
...
n
n
X
X
X
X
X

and
( )
( )
1
1
1
1
1
1
]
1

1
2
1
0
...
n
n
x
x
x
x
x

The DFT transform pair is then given by the matrix equations

x W
N
X


* *
1
and
X W x

*
1

J. N. Denenberg Copyright DTS June 13, 2005