2.

1
ECE 410 DIGITAL SIGNAL PROCESSING D. Munson
University of Illinois Chapter 2 A. Singer


Fourier Representation of Discrete-Time Signals

As with continuous-time signals, it is often convenient to represent discrete-time signals as a
linear combination of simpler signals, or basis signals. For discrete-time systems, we have
seen the eigenfunctions of discrete-time signals include all signals that can be written in the form
of a complex exponential, or z
n
for all n and for any, possibly complex, z. By restricting the
class of such signals to have unity magnitude, we arrive at the class of complex exponentials of
the form e
jn
for all n and for real-valued . These signals play a particularly important role in
the analysis of discrete-time systems due to this eigenfunction property, which implies that the
response of a linear shift-invariant system to a complex exponential input will be a complex
exponential output of the same frequency with amplitude and phase determined by the system.
For real-valued systems, i.e. systems with real-valued impulse responses, when the input is
sinusoidal of a given frequency, the output will remain sinusoidal of the same frequency, again
with amplitude and phase determined by the system. This important property of linear shift
invariant systems makes the representation of signals in terms of complex exponentials
extremely useful for studying linear system theory.

The Fourier transform enables the construction of a wide class of signals from a superposition of
complex exponentials. Through the eigenfunction property, the response of a linear shift
invariant system to any signal in this class, i.e. any signal with a Fourier transform, can be
constructed by adding up the responses to each of the eigenfunctions that make up the original
signal. By linearity of the system, the response of the system to a linear combination of complex
exponentials will be given by the same linear combination of the responses to the complex
exponentials. The eigenfunction property of LSI systems enables us to express very simply the
response of the system to each of these complex exponentials.

The Fourier transform for a continuous-time signal x(t), defined over the infinite interval t is
given by

= dt e t x X
t j
) ( ) (
where is the real-valued frequency in radians indicating the amount of the
eigenfunction
t j
e

in the signal ). (t x The inverse Fourier transform is given by

= d e X t x
t j
) ( ) (
2
1

The reader may recall the relationship between the CT Fourier transform and the Laplace
transform,

= dt e t x s X
st
L
) ( ) (
namely, that
=
=
j s
L
s X X ) ( ) ( ,
2.2


such that the Fourier transform can be viewed as the Laplace transform evaluated along the j -
axis, when it is in the region of convergence of the Laplace transform.

The Discrete-Time Fourier Transform (DTFT)

The discrete-time Fourier transform, or DTFT, enables the representation of discrete-time
sequences by a superposition of complex exponentials. Many sequences of interest can be
represented by the following Fourier integral,

d e X n x
n j
d
) (
2
1
] [
where,
( ) [ ]
j n
d
n
X x n e

=
=

is the discrete-time Fourier transform of the sequence. These two expressions comprise the
Fourier representation of the sequence x[n]. Note that the DTFT, X
d
(), is a complex-valued
function of the real-valued variable , when the sum exists. The integral corresponds to the
inverse DTFT and represents the synthesis of the signal x[n] from a superposition of signals of
the form

d e X
n j
d
) (
2
1
,
where we interpret the integral as the limit of a Riemann sum. The value of the DTFT, X
d
(),
determines the relative amount of each of the complex exponentials that is required to construct
x[n]. The DTFT is referred to as Fourier analysis, as we analyze the composition of the signal in
terms of the complex exponentials that make it up. The inverse DTFT is referred to as Fourier
synthesis, as it can be viewed as synthesizing the signal from these basic components that make
it up.

There is a strong similarity between the Fourier transform and the z-transform for discrete-time
signals. This relationship is similar to that between the Fourier transform and the Laplace
transform for continuous-time signals. Just as the Fourier transform for continuous-time signals
can be viewed as evaluating the Laplace transform along a specific curve, namely the imaginary
axis in the s-plane, the DTFT can be viewed as evaluating the more general z-transform along a
specific curve in the z-plane. The z-transform of a discrete-time sequence, given by,

=
n
n
z n x z X ] [ ) ( ,
is the same as the DTFT for values of z evaluated for a particular slice of the complex z-plane.
Specifically, the DTFT can be seen to be the same as the z-transform evaluated along a curve in
the z-plane corresponding to the unit-circle, i.e.,
2.3




X
d
() = X(e
j
)
angle =
e
j
1


The DTFT exists as a regular function if and only if the region of convergence of the z-
transform, ROC
X
, includes the unit circle, i.e. |z| = 1. For the case of sinusoidal sequences,
where X( z) contains poles on the unit circle, X
d
() can be defined in terms of impulse
distributions.

While for continuous-time signals, the notion of angular frequency is relatively well-defined, for
discrete-time signals, we also refer to the variable in X
d
() as the digital frequency. In some
textbooks this variable is used to represent analog frequency in the continuous-time Fourier
transform. Here, we will use the variable to denote continuous-time frequency. Of course, we
could use any variable for the DTFT and the continuous-time Fourier transform. By selecting
to represent digital frequency and to represent analog frequency, we can maintain consistency
in our discussion as well as with a number of texts on the topic. It is important to recall that
while continuous-time sinusoids have a fixed relationship between their frequency of oscillation
and the period of the periodic time-domain waveform, discrete-time sinusoids may not be
periodic at all. Recall that a signal of the form
0
[ ]
j n
x n e

=
is only periodic if the following relation holds
[ ] [ ] x n x n P = + .
Spefically, we must have that
0 0
( 2 ) ( [ ]) j n k j n P
e e
+ +
=
which corresponds to requiring that
0
2 P
k

= ,
i.e., the digital frequency must be a rational multiple of . This relationship will certainly only
hold for a small set of digital frequencies. Since the rational numbers are countable, and the real
numbers are uncountable, this relationship does not hold almost everywhere. That is, for
practical purposes, almost any digital frequency you come up with will correspond to a complex
exponential sequence that is not periodic. As a result, we rarely discuss the period of complex
exponentials and refer only to their digital frequency.

To demonstrate that the Fourier transform synthesis equation, or inverse DTFT, in fact inverts
the DTFT, we can simply plug the definition of the DTFT into the synthesis equation as follows.
2.4


( )

( )

1
[ ]
2
1
[ ]
2
1
[ ]
2
[ ] [ ]
[ ]
j n
d
j m j n
m
j n m
m
m
x n X e d
x m e e d
x m e d
x m n m
x n

=
=
| |
=
|
\ .
=
=
=

where we used that
( )
( ) ( )

1,
1
, 2
2 ( )
1,
1 1
,
2 ( )
1,
0,
[ ].
j n m
j n m j n m
n m
e d
e e
n m
j m n
n m
n m
j m n
n m
n m
n m

= +

=
=

Properties of DTFT

A number of important properties of the DTFT follow directly from those of the z-transform.

Linearity:
The DTFT is a linear transform, in that super position holds as follows. If
[ ] ( )
DTFT
d
x n X
and
[ ] ( )
DTFT
d
y n Y ,
then,
[ ] [ ] ( ) ( )
DTFT
d d
ax n by n aX bY + + .
The proof follows by noting that the DTFT is a special case of the z-transform which was shown
to be linear. We can also show this directly from the definition of the DTFT,
[ ] [ ] ( [ ] [ ]) [ ] [ ] ( ) ( )
DTFT j n j n j n
d d
n n n
ax n by n ax n by n e a x n e b y n e aX bY




= = =
+ + = + = +

.

Periodicity:
2.5
The DTFT of every sequence is always periodic in that the following relation holds
( ) ( 2 )
d d
X X = + .
The proof follows from the definition of the DTFT,
( 2 )

( 2 ) [ ]
[ ]
[ ] (1)
[ ]
( ).
j n
d
n
j n j n
n
j n n
n
j n
n
d
X x n e
x n e e
x n e
x n e
X

+
=

=
+ =
=
=
=
=

Real-part and Imaginary-part Symmetry Properties:
For real-valued sequences x[n], we have that the real-part of the DTFT is even, and the
imaginary part is odd, i.e.,
| | | |
| | | |
( ) ( ) ,
( ) ( ) .
d d
d d
X X
X X


=
=

The proof follows from trigonometric properties of the real and imaginary parts. Specifically, If
x[n] is real, then
| |
| |
( ) [ ]cos( )
[ ]cos( ) ( ) .
d
n
d
n
X x n n
x n n X

=
=
= =

For the imaginary part, we have that
| |
| |
( ) [ ]sin( )
[ ]sin( ) ( ) .
d
n
d
n
X x n n
x n n X

=
=
= =

Magnitude and Phase Symmetries:
For real-valued sequences x[n], we have that the magnitude of the DTFT is an even function and
the phase of the DTFT is an odd function, i.e.,
( ) ( ) ,
( ) ( ).
d d
d d
X X
X X


=
= X X


For example, ( ) and ( )
d d
X X X might look like
2.6


0
0.2
0.4
0.6
0.8
1
|X
d
()|
frequency,
0 -2 - 2
-4
-2
0
2
4
X
d
()
frequency,
2 - -2
0


The proof of the magnitude and phase symmetries follows from the definition of the DTFT as
below. From the definition of the DTFT we have
1/ 2
2 2
1/ 2
2 2
1/ 2
2 2
( ) [ ]cos [ ]sin
[ ]cos ( ) [ ]sin ( )
[ ]cos ( ) [ ]sin ( )
( )
d
n n
n n
n n
d
X x n n x n n
x n n x n n
x n n x n n
X

= =

= =

= =
(
| | | |
= + (
| |
\ . \ . (

(
| | | |
= + (
| |
\ . \ . (

(
| | | |
= + (
| |
\ . \ . (

=




and for the phase of the DTFT we have that
2.7
sin
( ) arctan
cos
sin( )
arctan
cos( )
sin( )
arctan ( )
cos( )
n
n
d
n
n
n
n
n
n
n
n
d
n
n
x n
X
x n
x n
x n
x n
X
x n

(
(
=
(
(

(
(
=
(

(

(
(
= =
(

X
X

as desired. The last line above follows since arctan is an odd function of its argument.

Shifting Property:
The shifting properties of the z-transform translate directly into properties of the DTFT by
placing z = e
j
. We have that
[ ] ( ),
DTFT j k
d
x n k e X

+
where k is an integer (positive or negative). This can also be shown directly, as
( )
[ ] [ ] [ ] [ ] ( )
DTFT j n j m k j k j m j k
d
n m m
x n k x n k e x m e e x m e e X

= = =
+ + = = =



Modulation:
By modulating a sequence by a complex exponential, the Fourier transform is translated in
frequency just as with the continuous-time Fourier transform. The modulation property for the
DTFT is
0
0
[ ] ( ),
j n DTFT
d
x n e X

which can be used to find the effect of modulation by a sinusoidal sequence,
| |
0 0 0
1
[ ]cos( ) ( ) ( )
2
DTFT
d d
x n n X X + + .
To prove the modulation property, we begin with the definition of the DTFT,
( )
0 0
( )

0
[ ] [ ]
( )
j n j n j n
n n
d
x n e e x n e
X




= =
=
=


which is the desired result. The result for a sinusoidal modulation follows directly from the
construction of a sinusoid using complex exponentials, i.e.
( )

0
1
cos( ) [ ] [ ].
2
o o
j n j n
n x n e e x n

= +
Then by applying the modulation property, along with linearity of the DTFT, we have the
desired result.

Example

If X
d
() has the form shown in the following figure,
2.8


X
d
()
2 2

1


then Y
d
(), the DTFT of the sequence [ ] [ ]
o
j n
y n e x n

= has the form in the following figure.

Y
d
()
2 2

0
1


Example

If X
d
() has the form shown in the following figure,

X
d
()
2 2

1


then Y
d
(), the DTFT of ( ) [ ] cos [ ]
o
y n n x n = has the form shown in the following figure.

Y
d
()
2 2
1
2

0

0




Convolution Property:
As we have seen, the z-transform maps convolution of two sequences into the multiplication of
their z-transforms. Since the DTFT is a special case of the z-transform, then we have that
2.9
[ ]* [ ] ( ) ( )
DTFT
d d
x n h n X H ,
assuming that the discrete-time Fourier transform of each of the two sequences are well-defined,
that is, assuming that the unit circle lies within the region of convergence of the z-transform of
each sequence. This result follows directly from the convolution property for z-transforms. This
can also be shown directly as
[ ]* [ ] [ ] [ ] [ ] [ ]
[ ] ( ) ( ) ( )
DTFT j n j n
n m m n
j m
d d d
m
x n h n x m h n m e x m h n m e
x m e H X H

= = = =

=
| | | |
=
| |
\ . \ .
= =

Where the second line follows from the delay property.

Parsevals Relation:
A particularly useful relationship between the energy of a sequence in the time-domain and the
energy contained in its Fourier transform is captured by Parsevals relation. Since sequences
with convergent DTFTs are square summable, they have finite energy and we have that
2
2

1
| [ ] | ( ) .
2
d
n
x n X d

=
=

This result can be computed by substituting in the inverse DTFT into the left hand side of the
above relation to obtain
( )
( )
( )
*

[ ]( [ ]) [ ] ( )
[ ] ( )
( ) [ ]
( ) ( )
( ) .
j n
d
n n
j n
d
n
j n
d
n
d d
d
x n x n x n X e d
x n X e d
X x n e d
X X d
X d

(

(
(
(
= = (

1
=
2
1
=
2
1
| |
=
|
2
\ .
1
=
2
1
=
2

Example:
Parsevals relation can be used to compute the energy of a signal in the time domain or the
frequency domain. As a result, one of these is often simpler than the other. For example,
consider the sequence x[n] = u[n]-u[n-10]. A rather complicated integral can be reduced to a
simple sum by noting the transform pair
sin(5 / 2)
[ ] [ 10]
sin( / 2)
DTFT
u n u n

and using Parsevals relation as follows:
2.10


2

2
9
0

( ) [ ]
sin(5 / 2)
(1)
sin( / 2)
sin(5 / 2)
10.
sin( / 2)
d
n
n
X d x n
d
d

=
=
1
=
2
1
=
2
1
=
2

2.11
Examples of DTFT


We continue our discussion of the discrete-time Fourier transform by considering a few simple
examples.

Example
Consider the following sequence containing two non-zero samples,
[ ] [ ] [ 1]. x n n n = +
The discrete-time Fourier transform of this sequence can be computed directly from the
definition of the DTFT as
( ) 1 .
j
d
X e

= +
The magnitude of the discrete-time Fourier transform can be easily computed as
2
2
2
2 2
2 2
( ) 1
1 cos sin
1 cos sin
1 2cos cos sin
2 2cos .
j
d
X e
j
j

= +
= +
= + +
= + + +
= +

This result could also have been obtained by noting that when two exponential terms (or one
exponential term and one constant term) have the same magnitude, by factoring out a common
phase factor, a sinusoid can be constructed as follows,
2
2
2 2
/ 2 / 2 / 2
2 2
2
( ) 1
1 2cos( / 2)
4cos ( / 2)
2 2cos( ),
j
d
j j j
X e
e e e

= +
= +
=
=
= +

where, in the last line, the trigonometric identity that cos
2
(x) = (1+cos(2x)) / 2. The magnitude
squared of the discrete-time Fourier transform is shown below:

0

4
2
X
d
()
2

2
3
2

2.12


Since
2
( )
d
X is both periodic with period 2 and symmetric around the origin,
2
( )
d
X is
completely determined by its values on the interval 0 . This is similarly true for X
d
().
Because of this, when [ ] x n is real (so that ( )
d
X and X
d
() have even and odd symmetry,
respectively) we will often plot ( )
d
X and X
d
() on just the interval 0 .

To find X
d
() in this example, write
( )
( )
/ 2 / 2 / 2
/ 2
( ) 1
2cos / 2 ( )
j j j j
d
j
X e e e e
e

= + = +
=

Now, since cos( / 2) 0 > for < < , equation (
*
) expresses X
d
() in polar form, so that
( ) / 2
d
X = , < < . The phase is plotted below as

2
2

X
d
()

2


Notice that the phase of X
d
() is an odd function. Also note that for > , the expression
( ) / 2
d
X = is not valid, however we simply use our knowledge that the DTFT is periodic
with period 2. While it is clear from this example that the discrete-time Fourier transform is
periodic in the variable with period 2, the DTFT is, in fact, always periodic with period 2.
This can be seen easily from the definition of the DTFT as follows,
1
2
( 2 )
( ) [ ]
[ ] ( )
[ ]
( 2 ).
j n
d
n
j n j
n
j n
n
d
X x n e
x n e e
x n e
X

+
=
=
=
=
= +

We see that since
2
1
j n
e

= for integer n, the DTFT is always periodic. This is different from
the continuous-time Fourier transform, where we were interested in frequencies spanning an
2.13
infinite range of real values. In contrast, in discrete-time, all digital frequencies can be captured
in a single interval of length 2. The reason for this periodicity stems directly from the
observation that complex exponentials of the for
j n
e

are only unique over an interval of this

range. That is the sequence
j n
e

is identical to the sequence

( 2 ) j n
e
+
. Since these two
sequences have identical values, for all n, then the composition of x[n] in terms of these
sequences, i.e. the DTFT, must only require a single interval containing them. Since the DTFT
is periodic with period 2, the DTFT only needs to be specified over an interval of length 2. It
is often convenient to use the interval - < < , so that the low frequencies are centered around
= 0. Note that since all frequencies that are multiples of 2 are indistinguishable, the low
frequencies are also those centered around any multiple of 2. Similarly, the highest digital
frequency corresponds to = , as well as all odd multiples of . We will return to this issue
again when we discuss the discrete-time frequency response of linear shift-invariant systems.

Example
Consider the sequence x[n] = [n 1] |n + 1]. For this sequence, we will plot the magnitude
( )
d
X and phase X
d
().
( )
2 sin
( ) 2sin
j j
d
d
X e e
j
X

=
=
=

which can be plotted as follows

X
d
()
2
2 2



Note that ( )
d
X is again an even function of , and its appearance is completely specified by
( )
d
X on just the interval 0 . The phase of X
d
() is found by noting
{ }
{ }

2
2
2 sin : sin 0
( )
2 sin : sin 0
2sin 0
2 sin 0
d
j
j
j
X
j
e
e

>

<

< <

< <

Both the top and bottom lines within the bracket are written in polar form, since sin > 0 for
0 < < and sin > 0. Thus,
2.14

0
2
( )
0
2
d
X

< <

< <

which can be plotted as
2 2

X
d
()

2

Notice that X
d
() is an odd function of the variable . This is again due to the symmetry
properties of the DTFT for real-valued x[n]. It is important to recall that these symmetry
properties of ( )
d
X and X
d
() hold only for real-valued sequences x[n]. If x[n] is not real-
valued, then this symmetry will not be present, as shown in the next example.

Example
Consider the sequence x[n] = |n] + j [n 1]. For this sequence, we have the following
discrete-time Fourier transform,
2
( ) 1
( ) (1 )(1 )
(1 1)
2 2sin( )
j
d
j j
d
j j
X je
X je je
je je

= +
= +
= + +
= +

which is plotted below and clearly does not display the same symmetries of a real-valued
sequence.
4
2

2
3
2
2

X
d
()
2

Here, we see that ( )
d
X = ( )
d
X does not hold.
2.15

Example
Consider the following sequence x[n] = a
n
u[n], with a real-valued and |a| < 1. We can use our
knowledge of the z-transform for this sequence, and that
( ) ( )
j
d
z e
X X z

=
=
X
d
() is well defined here since the region of convergence of the z-transform, ROC
X
, includes
the unit circle. We have that

2
2 2
2 2 2 2
2
( ) ,| |

1
( )
1
1
( )
(1 cos ) ( sin )
1
1 2 cos cos sin
1
1 2 cos
j
d
j j
d
z
X z z a
z a
e
X
e a ae
X
a a
a a a
a a

= >
= =

=
+
=
+ +
=
+

For 0 < a < 1, the magnitude squared would appear as depicted in the following figure.

2

X
d
()
1
1 a
1
1+ a


Example
We next consider a sinusoidal input, x[n] = cos
o
n, for all n. The z-transform for this sequence
is undefined for all z, including z on the unit circle, where it coincides with the discrete-time
Fourier transform. However, we are willing to extend the notion of existence of the discrete-
time Fourier transform to include such signals with the aid of impulse distributions. Similar to
their continuous-time counterparts, we can define the DTFT of discrete-time sinusoidal signals
in terms of impulses. As such, we can define the DTFT of such a signal to satisfy
( )
0

0 0
( ) cos( )
( ) ( ) ,| |
j n
d
n
X n e

=
=
= + + <

in the sense that the inverse discrete-time Fourier transform of this distribution would yield the
original signal x[n]. Here, X
d
() is a distribution, not a function, so the DTFT does not really
2.16


exists in the normal sense, and the summation defining the DTFT does not converge in any
meaningful sense to any function. However, if we use this distribution as the operational DTFT
of the sequence, then taking its inverse DTFT would yield
( ) ( )
0 0

0
1
( )
2 2
1
2
cos( ).
j n j n
d o o
j n j n
X e d e d
e e
n

= + + (

( = +

=



If we graphically depict the distribution X
d
(), using our notation for representing impulse
distributions, we would obtain the following sketch below.

X
d
()

o

o

While an impulse distribution cannot be considered a function, nor can it be considered a proper
limit of a sequence of functions, we could approximate the above figure using tall, narrow
rectangles around the frequencies of interest, i.e. in place of (
o
) and ( +
o
). This
would correspond to the discrete-time Fourier transform of an approximation to cos(
o
n).
Additionally, the quality of the approximation improves as the rectangles get narrower and taller.
This is explored further in the next Example.

Example

Consider the finite-length sequence x[n] described below,
cos , 0 1
[ ]
0, else
o
n n N
x n

=

where the sequence is finite-length in that it has a finite number, N, of non-zero samples. This
sequence can be constructed as a windowed version of the original infinite-length sequence
cos(
o
n). The DTFT of the sequence can be written,
2.17
0 0
0 0
0 0
0 0 0
0
1
0
0
1

0
1 1
( ) ( )
0 0
( ) ( )
( ) ( )
( ) / 2 ( ) / 2 ( ) / 2
( ) / 2
( ) cos( )
1
2
1 1
2 2
1 1 1 1
2 1 2 1
1
2
o o
N
j n
d
n
N
j n j n j n
n
N N
j n j n
n n
j N j N
j j
j N j N j N
j
X n e
e e e
e e
e e
e e
e e e
e e

=
+
= =
+
+


=
( = +

= +

= +

0 0 0
0 0 0 0 0
0 0
0
( ) / 2 ( ) / 2 ( ) / 2
( ) / 2 ( ) / 2 ( ) / 2 ( ) / 2 ( ) / 2
( 1) ( 0
( ) ( )
2
0
periodic sinc centered at
1
2
sin ( )
1 1 2
1 2 2
sin ( )
2
j N j N j N
j j j j j
N N
j j
e e e
e e e e
N
e e






+ + +
+ + +

+
=

+

| |

|
\ .
= +
| |

|
\ .
_
1) 0
2
0
sin ( )
2
.
1
sin ( )
2
N

| |
+
|
\ .
| |
+
|
\ .


Note that this expression contains two terms; one term corresponding to a ratio of sin
expressions, multiplied by a linear phase term, and another corresponding to a similar ratio of sin
expressions and a similar linear phase term. Each of these terms, corresponds to a periodic sinc
function, centered at the corresponding positive and negative frequencies of the original cosine
expression. The periodic sinc function is simply the DTFT of a length-N sequence of ones, and
is depicted in the following figure.
-1 -0.5 0 0.5 1 1.5 2 2.5 3
-10
-5
0
5
10
15
20
25
/
sin(N/2)/sin(/2), for N=25

Periodic Sinc. Note that the first zero-crossing occurs at = 2/.
2.18


For large N, the main lobe in the periodic sinc becomes narrow and large in amplitude so that the
two terms in the DTFT of the windowed cosine sequence do not overlap much and we have that
( )
( )
( )
( )
sin sin
1 1 2 2
( )
1 1 2 2
sin sin
2 2
o o
d
o o
N N
X

| | | |
+
| |
\ . \ .
+
| | | |
+
| |
\ . \ .
.
This relation is expressed in the following figure, again for the case of N = 25.
0
2
4
6
8
10
12
14
-

0
-
0
|X
d
()| for N=25

0


Now, as N becomes large, this figure begins to resemble, in some sense, the figure containing
two impulses. Similarly, as N becomes large, the windowed (truncated) cosine sequence
becomes a better approximation of the infinite-length cosine sequence.



Discrete Fourier Transform (DFT)


The DFT of x
n
{ }
n=0
N1
is a set of evenly spaced samples of X
d
() on [0, 2), and is denoted
by X
m
{ }
m=0
N1
. Specifically,


2.19

X
m
= X
d
2
N
m

= x
n
e
j
2
N
mn
n=0
N1

m=0, 1, , N1


(1)


Picture:

2

X
d
()
X
2
X
0
X
1
X
N1


Note that the last DFT sample, X
N1
is not at = 2. It is to the left of = 2.

Inverse DFT:


x
n
=
1
N

m=0
N-1
X
m
e
j
2
N
nm
n=0, 1, , N-1


Proof of DFT
-1
:



1
N
X
m
e
j
2
N
nm
m=0
N1

=
1
N
x
/
e
j
2
N
/m
/=0
N1

m=0
N1

e
j
2
N
nm


=

1
N
x
/
/=0
N1

e
j
2
N
m(n/)
m=0
N1

(
*
)


Second sum =

N / = n
1 e
j
2
N
(n/)N
1 e
j
2
N
(n/)
= 0 / n







So, the second sum can be written as N
/n
. Substituting into (
*
) gives

2.20



1
N
X
m
e
j
2
N
nm
m=0
N1

=

1
N
x
/
N
/n
/=0
N1

= x
n




Properties of the DFT


1) If x
n
{ }
n=0
N1
is real, then

a) X
m
= X
<Nm>
N
0 m N 1 (magnitude of X
m
{ } is even)

b) X
m
= X
<N-m>
N
0 m N 1 (angle of X
m
{ } is odd)

where <k>
N
= k mod N, i.e.,

<k>
N
= r if k = /N + r with 0 r N 1


e.g.,

<7>
4
= 3, <4>
4
= 0, <5>
4
= 3, <2>
4
= 2

2(4) + 3 1(4) + 2

Note: It is sometimes helpful to use < k>
N
= N <k>
N
.

In the above DFT property, computing the index modulo N makes a difference only for m = 0
since <Nm>
N
= N m for 1 m N1, and <N 0>
N
= 0.

2.21
Proof: Follows from X
m
= X
d

2 m
N
|
\
|
.
and

X
d
() = X
d
() = X
d
(2 )

even periodic

X
d
() = X
d
() = X
d
(2)

odd periodic

Example

Suppose have real-valued x
n
{ }
n=0
7
. Then X
m
, 0 m 7 might look like the DTFT samples
below

X
0
2

X
d
()
X
1
X
2
X
6
X
7


Obviously:

X
1
= X
7


X
2
= X
6

etc.

Similarly:

X
1
= X
7

X
2
= X
6

etc.

2) Time Shift

DFT
1
X
m
e
j
2
N
mk



`
)
N1
m=0



(

(

= x
<nk>
N
{ }
N1
n=0


2.22


This property states that multiplication of a DFT X
m
by a complex exponential e
j
2
N
mk

corresponds to a circular shift by an amount k.

Proof:


1
N
X
m
e
j
2
N
mk
m=0
N1

e
j
2
N
nm
=
1
N
X
m
e
j
2
N
m(n k)
m=0
N1

=
1
N
X
m
e
j
2
N
m<nk>
N
m=0
N1

(
*
)

= x
<nk>
N
0 n N 1

since if (nk) = q N + r, 0 r N 1

e
j
2
N
m(nk)
= e
j
2
N
m(qN+r)
= e
j
2
N
mr
= e
j
2
N
m<nk>
N



Note: The step in Eq. (
*
) is necessary because x
nk
is not defined outside the range
0 n - k N - 1.

Example

Let x
n
{ }
n=0
4
= {1, 3, 5, 7, 9}

Then
(k= 1) DFT
1
X
m
e
j
2
5
m



`
)



(

(
= {9, 1, 3, 5, 7}


(k = 2) DFT
1
X
m
e
j
2
5
2m



`
)



(

(
= {7, 9, 1, 3, 5}

(k = 3) DFT
1
X
m
e
j
2
5
3m



`
)



(

(
= {5, 7, 9, 1, 3}


etc.

3) DFT X
m
{ }
m=0
N1
| |
= N x
<Nn>
N
{ }
n=0
N1


2.23

This DFT property states that applying a forward DFT twice gives {x
n
} back but scaled by N
and flipped around (except the n = 0 term x
0
).


Example

Let X
m
{ }
m=0
5
be the DFT of

x
n
{ }
n=0
5
= {2, 4, 6, 8, 10, 12}.

Then DFT X
m
{ }
m=0
5
| |
= 6 {2, 12, 10, 8, 6, 4} = {12, 72, 60, 48, 36, 24}.

This property gives a way to compute an inverse DFT using a forward DFT algorithm, which is
handy on some rare occasions.


Proof of Property 3):

Let y
n
{ }
n=0
N1
= DFT X
m
{ }
m=0
N1
| |

y
n
= X
m
m=0
N1

e
j
2
N
nm


=

m=0
N1

x
/
/=0
N1

e
j
2
N
m/
e
j
2
N
nm


=

x
/
/=0
N1

e
j
2
N
m(/+n)
m=0
N1

=

x
/
/=0
N1

e
j
2
N
m</+n>
N
m=0
N1

But, the second sum above can be written as



e
j
2
N
m</+n>
N
m=0
N1

=
N < / + n >
N
= 0
0 < / + n >
N
0





Since / {0, 1, , N1} we have </+n>
N
= 0 if / = <n>
N
= <Nn>
N
.

2.24



Thus, the second sum can be written as N
/<Nn>
N
so that

y
n
=

/=0
N1

x
/
N
/<Nn>
N


= N x
<Nn>
N
0 n N 1

4) Parseval


m=0
N1

X
m
2
= N
n=0
N1

x
n
2


This property shows how to compute the energy of a sequence in terms of its DFT.

Proof:


m=0
N1

X
m
2
=
m=0
N1

X
m
X
m

=
m=0
N1

X
m
x
n
e
j
2
N
nm
n=0
N1

(

(
(

=
n=0
N1

x
n

m=0
N1

X
m
e
j
2
N
nm


=
n=0
N1

x
n

N x
n


= N
n=0
N1

x
n
2




5) Cyclic Convolution

C
m
= A
m
B
m
0 m N 1

if and only if

2.25
c
n
=

/=0
N1

a
/
b
<n/>
N


This summation is called a cyclic convolution. It will be described later in great detail. It differs
from the regular linear convolution because taking the index n/ modulo N causes the b
/

sequence to wrap back around the a
/
sequence in a circular fashion. Our interest in this
property arises because it suggests that a cyclic convolution can be computed as

c
n
{ } = DFT
1
DFT a
n
{ }
| |
DFT b
n
{ }
| |
{ }
. (*)

Later we will study extremely fast algorithms for computing DFTs and inverse DFTs, and we
will also show that by zero-padding {a
n
} and {b
n
}, a cyclic convolution can produce a linear
convolution. Together, these results will provide a fast method for performing regular, linear
convolution.


Proof of Cyclic Convolution Property

Show that (*) results in a cyclic convolution:

c
n
{ } = DFT
1
A
m
B
m
{ }
m=0
N1
| |

=
1
N
m=0
N1

a
k
e
j
2
N
mk
k=0
N1

|
\

|
.
|
B
m
e
j
2
N
mn


= a
k
k=0
N1

1
N
B
m
m=0
N1

e
j
2
N
m(nk)


= a
k
k=0
N1

1
N
B
m
m=0
N1

e
j
2
N
m<nk>
N

_

b
<nk>
N

= a
k
k=0
N1

b
<nk>
N
0 n N 1

Note: We needed the next-to-last step, where we inserted the modulo notation, because b
/
is
defined only for 0 / N 1.