Discrete Time Fourier Methods

Discrete-Time Fourier Series Concept
A signal can be represented as a linear combination of sinusoids.
Discrete-Time Fourier Methods
M. J. Roberts - All Rights Reserved. Edited by Dr. Robert Akl

1

2


The relationship between complex and real sinusoids















10

The Discrete-Time Fourier Series

The discrete-time Fourier series (DTFS) is similar to the CTFS. A periodic discrete-time signal can be expressed as x[n] =
k= N
The Discrete Fourier Transform

The discrete Fourier transform (DFT) is almost identical to the DTFS. A periodic discrete-time signal can be expressed as x[n] = 1 N
k= N
c x [ k ] e j 2 kn / N
cx [ k ] =
1 N
n0 + N 1 n = n0
x [ n ] e j 2 kn / N
X[ k ]e
j 2 kn / N
X[ k ] =
n0 + N 1 n = n0
x[n]e
j 2 kn / N
where c x [ k ] is the harmonic function, N is any period of x [ n ] and the notation,

k= N
where X [ k ] is the DFT harmonic function, N is any period of x [ n ] and the notation, means a summation over any range of
k= N
means a summation over any range of
consecutive ks exactly N in length.
consecutive ks exactly N in length. The main difference between the DTFS and the DFT is the location of the 1/N term. So X [ k ] = N c x [ k ].

11


12



Because the DTFS and DFT are so similar, and because the DFT is so widely used in digital signal processing (DSP), we will concentrate on the DFT realizing we can always form the DTFS from cx [ k ] = X[ k ] / N .

13


14

DFT Example

Find the DFT harmonic function for x n = u n u n 3 5 n using its fundamental period as the representation time. X k =
n= N 4
The DFT Harmonic Function

We know that x [ n ] = 1 N
x n e
j 2 kn / N
k= N
X[ k ]e
j 2 kn / N
so we can find x [ n ]
from its harmonic function. But how do we find the harmonic
= u n u n 3 5 n e
n= 0 2
j 2 kn /5
function from x [ n ]? We use the principle of orthogonality like we did with the CTFS except that now the orthogonality is in discrete time instead of continuous time.
= e
n= 0 2
j 2 kn /5
1 e j 6 k /5 e j 3 k /5 e j 3 k /5 e j 3 k /5 = = j k /5 j k /5 1 e j 2 k /5 e j k /5 e e sin 3 k / 5 sin k / 5
= e j 2 kn /5 = e j 2 k /5
n= 0
( (
) = 3e )
j 2 k /5
drcl k / 5,3
)
15
16





17


18

Below is a set of complex sinusoids for N = 8. They form a set of basis vectors. Notice that the k = 7 complex sinusoid rotates counterclockwise through 7 cycles but appears to rotate clockwise through one cycle. The k = 7 complex sinusoid is exactly the same as the k = 1 complex sinusoid. This must be true because the DFT is periodic with period N .


The projection of a real vector x in the direction of another real vector y is xT y y yT y If p = 0, x and y are orthogonal. If the vectors are complexvalued p= p= xH y y yH y
where the x H is the complex-conjugate transpose of x. xT y and x H y are the dot product of x and y.
19
20




21


22




23


24


The most common definition of the DFT is X [ k ] = x [ n ] e j 2 kn / N ,
n=0 N 1
Convergence of the DFT

The DFT converges exactly with a nite number of terms. It does not have a Gibbs phenomenon in the same sense that the CTFS does

x[n] =
1 N
k= N
X[ k ]e
j 2 kn / N
Here the beginning point for x [ n ] is taken as n0 = 0 . This is the form of the DFT that is implemented in practically all computer languages.

25


26


harmonic number just as we have seen in the CTFS. x [ n ] and X [ k ] form a DFT pair based on N points.
D FT x [ n ] X[ k ] N j 2 kn / N
DFT Properties

X [ k ] is called the DFT harmonic function of x [ n ] and k is the
1 From x [ n ] = N
k= N
X[ k ]e
we see that x [ n ] is formed
by a linear combination of functions of the form e j 2 kn / N each of which has a period N . Therefore x [ n ] must also be periodic with period (but not necessarily fundamental period) N .

27


28

DFT Properties

DFT Properties

It can be shown (and is in the text) that if x [ n ] is an even X [ k ] is purely imaginary. function, X [ k ] is purely real and if x [ n ] is an odd function

29


30

The Dirichlet Function

The functional form sin ( Nt ) N sin ( t )
DFT Pairs

D FT cos ( 2 qn / N ) ( mN / 2 ) ( mN [ k mq ] + mN [ k + mq ]) mN D FT e j 2 n / N mN mN [ k m ] mN D FT sin ( 2 qn / N ) ( jmN / 2 ) ( mN [ k + mq ] mN [ k mq ]) mN D FT N [ n ] m mN [ k ] mN D FT 1 N N [ k ] N D FT N
appears often in discrete-time signal analysis and is given the special name Dirichlet function. That is drcl ( t , N ) = sin ( Nt ) N sin ( t )
( u [ n n ] u [ n n ]) [ n ] e
0 1 N D FT N
D FT sinc ( n / w ) N [ n ] wrect ( wk / N ) N [ k ] N
( n1 n0 ) drcl ( k / N , n1 n0 ) e j k / N 2 tri ( n / N w ) N [ n ] N w drcl ( k / N , N w ) , N w an integer
j k ( n1 + n0 ) / N

31


32

The Fast Fourier Transform

One could write a MATLAB program to implement the DFT.
. % . % Initialize the DFT array to a column vector of zeros. (Acquire the input data in an array x with N elements.)
The Fast Fourier Transform

There is a function in MATLAB fft which accomplishes the same goal and is typically much faster. This table compares the speeds of the two methods. M stands for computer multiplies and A stands for computer additions.
X = zeros(N,1) ; % Compute the Xns in a nested, double for loop.
for k = 0 :N - 1 for n = 0 :N - 1 X(k + 1) = X(k + 1) + x(n + 1) * exp(-j * 2 * pi * n *k / N) ; end end

33

N = 2 1 2 2 4 3 8 4 16 5 32 6 64 7 128 8 256 9 512 10 1024
ADFT 2 12 56 240 992 4032 16256 65280 261632
M DFT 4 16 64 256 1024 4096 16384 65536 262144
AFFT 2 8 24 64 160 384 896 2048 4608
M FFT 1 4 12 32 80 192 448 1024 2304 5120
ADFT / AFFT 1 1.5 2.33 3.75 6.2 10.5 18.1 31.9 56.8 102.3
M DFT / M FFT 4 4 5.33 8 12.8 21.3 36.6 64 113.8 204.8

34

1047552 1048576 10240

Generalizing the DFT for Aperiodic Signals

Pulse Train


DFT of
Pulse Train

As the period of the rectangular wave increases, the period of the DFT increases

This periodic rectangular-wave signal is analogous to the

continuous-time periodic rectangular-wave signal used to
illustrate the transition from the CTFS to the CTFT.


35


36


Normalized
DFT of
Pulse Train

By plotting versus k / N 0 instead of k , the period of the normalized DFT stays at one.

The normalized DFT approaches this limit as the
period approaches innity.


37


38

Denition of the Discrete-Time Fourier Transform (DTFT)

Forward x[n] = X( F ) e
1 j 2 Fn
The Discrete-Time Fourier Transform

The function e j appears in the forward DTFT raised to the nth power. It is periodic in with fundamental period 2 . n is an integer. Therefore e- jn is periodic with fundamental period 2 / n and 2 is also a period of e jn . The forward DTFT is X e j =
F Form
F
Inverse
n =
dF X(F ) = Form
x[n]e
j 2 Fn
Forward x[n] = 1 2
Inverse
( ) x[n]e
n =
jn
F X e j e jn d X e j =
( )
( ) x[n]e
n =
jn
a weighted summation of functions of the form e jn , all of which repeat with every 2 change in . Therefore X e j is always periodic in with period 2 . This also implies that X ( F ) is always periodic in F with period 1.
( )

39


40

DTFT Pairs

We can begin a table of DTFT pairs directly from the definition. (There is a more extensive table in the text.)
F [ n ] 1
The Generalized DTFT

By generalizing the CTFT to include transform that have impulses we were able to find CTFT's of some important practical functions. The same is true of the DTFT. The DTFT of a constant X(F ) =
n =
e j 1 e j 1 F = , <1 , n u [ n 1] j = , >1 e 1 e j e 1 e j e j sin ( 0 ) e j sin ( 0 ) F F n n sin ( 0 n ) u [ n ] j 2 , < 1 , sin ( 0 n ) u [ n 1] j 2 , >1 j 2 j e 2 e cos ( 0 ) + e 2 e cos ( 0 ) + 2
F n u [ n ] j F n cos ( 0 n ) u [ n ]
Ae
j 2 Fn
= A e j 2 Fn
n =
j 2
j e j e cos ( 0 ) j
2 e cos ( 0 ) +
F , < 1 , n cos ( 0 n ) u [ n 1]
j 2
j e j e cos ( 0 ) j
2 e cos ( 0 ) +
, >1
does not converge. The CTFT of a constant turned out to be an impulse. Since the DTFT must be periodic that cannot the the transform of a constant in discrete time. Instead the transform must be a periodic impulse.

41


42


To shorten the process, find the inverse DTFT of a periodic impulse of the form A1 ( F ) . Using the formula x [ n ] = A1 ( F ) e j 2 Fn dF = A
1 1/2 1/2

Now consider the function X ( F ) = A1 ( F F0 ) , 1 / 2 < F0 < 1 / 2. Its inverse DTFT is x [ n ] = A1 ( F F0 ) e j 2 Fn dF = A
1 1/2
(F)e
j 2 Fn
dF = A
Now change x [ n ] to x [ n ] = A cos ( 2 F0 n ) = ( A / 2 ) e j 2 F0 n + e j 2 F0 n . Then

F A cos ( 2 F0 n ) ( A / 2) 1 ( F F0 ) + 1 ( F + F0 )
1/2
(F F )e
0
j 2 Fn
dF = Ae j 2 F0 n
proving that the DTFT of a constant A is A1 ( F ) or, in radianF frequency form A 2 A 2 ( ) .
or
F A cos ( 0 n ) A 2 ( 0 ) + 2 ( + 0 )

43


44

Forward DTFT Example

Find the forward DTFT of x [ n ] = u [ n n0 ] u [ n n1 ].
F u [ n n0 ] u [ n n1 ] n =

n1 1 n = n0
u [ n n0 ] u [ n n1 ] e j 2 Fn =
n1 n0 1 m=0
e j 2 Fn
F u [ n n0 ] u [ n n1 ] e j 2 Fn0
Let m = n n0 . Then u [ n n0 ] u [ n n1 ]
F n1 n0 1 m=0
Factor out e j F ( n1 n0 ) from the numerator and e j F from the denominator e j F ( n1 n0 ) e j F ( n1 n0 ) e j F ( n1 n0 ) e j F e j F e j F By the definition of the sine function in terms of complex exponentials
F u [ n n0 ] u [ n n1 ] e j 2 Fn0 F u [ n n0 ] u [ n n1 ]
1 e j 2 F ( n1 n0 ) 1 e j 2 F
j 2 F ( m + n0 )
=e
j 2 Fn0
j 2 Fm
Summing this geometric series 1 e j 2 F ( n1 n0 ) u [ n n0 ] u [ n n1 ] e j 2 Fn0 1 e j 2 F

F
e j F ( n0 + n1 ) sin F ( n1 n0 ) e j F sin ( F )

45


46


Consider the special case of n0 + n1 = 1 n0 = 1 n1 (making the function a discrete-time rectangular pulse of width 2 n0 +1 centered at n = 0. u [ n n0 ] u [ n n1 ]
F
More DTFT Pairs

We can now extend the table of DTFT pairs.
F [ n ] 1
sin F ( n1 n0 ) sin ( F )
, n0 + n1 = 1
1 + (1 / 2 )1 ( F ) 1 e j 2 F F sinc ( n / w ) w rect ( wF ) 1 ( F )
F u [ n ] F tri ( n / w ) w drcl2 ( F, w ) F 1 1 ( F )
, , , , , ,
1 F u [ n ] + 1 ( ) 1 e j F sinc ( n / w ) w rect ( w / 2 ) 2 ( )
F tri ( n / w ) w drcl2 ( / 2 , w ) F 1 2 2 ( )
Compare this to the CTFT of a rectangular pulse of width w centered at t = 0. sin ( wf ) f The DTFT is a periodically-repeated sinc function.
F rect ( t / w ) w sinc ( wf ) =
F cos ( 2 F0 n ) (1 / 2 ) 1 ( F F0 ) + 1 ( F + F0 )
F N0 [ n ] (1 / N 0 )1/ N0 ( F )
F sin ( 2 F0 n ) ( j / 2) 1 ( F + F0 ) 1 ( F F0 ) , F u [ n n0 ] u [ n n1 ] F u [ n n0 ] u [ n n1 ]
F cos ( 0 n ) 2 ( 0 ) + 2 ( + 0 )
F N0 [ n ] ( 2 / N 0 ) 2 / N0 ( )
F sin ( 0 n ) j 2 ( + 0 ) 2 ( 0 )
e j 2 F e j F ( n0 + n1 ) e j 2 n0 F e j 2 n1 F = ( n1 n0 ) drcl ( F, n1 n0 ) e 1 e j F
j 2 F
e j e j( n0 + n1 ) /2 e jn0 e jn1 = ( n1 n0 ) drcl ( / 2 , n1 n0 ) e j 1 e j /2

47


48

DTFT Properties

F F x [ n ] + y [ n ] X ( F ) + Y ( F ) , x [ n ] + y [ n ] X ( e j ) + Y ( e j ) F x [ n n0 ] e j 2 Fn0 X ( F ) , F x [ n n0 ] e jn0 X e j F e j0 n x [ n ] X e j (0 )
DTFT Properties

( )
m =
x [ m ] 1 e
F
X(F )
j 2 F
1 X ( 0 )1 ( F ) , 2
F x [ n ] X(F ) ,
F e j 2 F0 n x [ n ] X ( F F0 ) ,
x [ n / m ] , n / m an integer F F If y [ n ] = X ( mF ) or y [ n ] X e jm then y [ n ] , otherwise 0
F x [ n ] y [ n ] X ( F ) Y( F ) ,
F x [ n ] y [ n ] X ( F ) Y( F )
see note
( ) + X ( e ) ( ) x [ n ] X(e ) x [ n ] y [ n ] X (e ) Y(e ) x [ n ] y [ n ] (1 / 2 ) X ( e ) Y ( e )
m =
x [ m ] 1 e
F F F j
X e j
j
j0
j see note
F x* [ n ] X* ( F ) ,
F x [ n ] x [ n 1] 1 e j 2 F X ( F ) ,
) x [ n ] x [ n 1] (1 e ) X ( e )
F x* [ n ] X* e j F j j
n =
j 2 Fn
= 1 ( F ) ,
2
n =
e
2
jn
= 2 2 ( ) X e j
(Note:
n =
x[n]
= X ( F ) dF ,
1 T
n =
x[n]
= (1 / 2 )
( )
x ( t ) y ( t ) = x ( ) y ( t ) d where T is a period of both x and y

49


50

DTFT Properties

Find and graph the inverse DTFT of X(F ) = rect ( 50 ( F 1 / 4 )) + rect ( 50 ( F + 1 / 4 )) 1 ( F ) Start with
F sinc ( n / w ) wrect ( wF ) 1 ( F )
DTFT Properties

Frequency shifting the other direction
F e j n /2 (1 / 50 ) sinc ( n / 50 ) rect ( 50 ( F + 1 / 4 )) 1 ( F )
In this case w = 50.

F rect ( 50 F ) 1 ( F ) (1 / 50 ) sinc ( n / 50 )
Combining the last two results and using cos ( x ) =
Then, using the frequency-shifting property e

j n /2 F rect ( 50 ( F 1 / 4 )) 1 ( F ) (1 / 50 ) sinc ( n / 50 ) F e j 2 F0 n x [ n ] X ( F F0 )
F (1 / 25 ) sinc ( n / 50 ) cos ( n / 2 ) rect ( 50 ( F 1 / 4 )) + rect ( 50 ( F + 1 / 4 )) 1 ( F )
e jx + e jx 2

51


52

DTFT Properties

Time scaling in discrete time is quite different from time scaling in continuous-time. Let z [ n ] = x [ an ]. If a is not an be found for it. If a is an integer greater than one, some values of x [ n ] will not appear in z [ n ] because of decimation and there cannot be a unique relationship between their DTFTs
53

DTFT Properties

Time scaling does not work for time compression because of decimation. But it does work for a special type of time expansion. x [ n / m ] , n / m an integer Let z [ n ] = . Then Z ( F ) = X ( mF ) . , otherwise 0 So the time-scaling property of the DTFT is x [ n / m ] , n / m an integer z[n] = , , otherwise 0
F z [ n ] X ( mF ) F or z [ n ] X e jm
integer some values of z [ n ] are undefined and a DTFT cannot

54

DTFT Properties

In the time domain, the response of a system is the convolution of the excitation with the impulse response y[n] = x[n] h[n] In the frequency domain the response of a system is the product of the excitation and the frequency response of the system Y e j = X e j H e j
DTFT Properties

Find the signal energy of x [ n ] = (1 / 5 ) sinc ( n /100 ) . The straightforward way of finding signal energy is directly from the definition E x = Ex =
n =
x[n]
.
2
( )
( ) ( )
n =
(1 / 5 ) sinc ( n /100 )
= (1 / 25 ) sinc 2 ( n /100 )
n =
In this case we run into difficulty because we don't know how to sum this series.

55


56

DTFT Properties

We can use Parseval's theorem to find the signal energy from the DTFT of the signal.
F 20rect (100 F ) 1 ( F ) (1 / 5 ) sinc ( n /100 )
Transform Method Comparisons

Parseval's theorem is
n =
x[n]
= X ( F ) dF
2 1 1/2
For this case E x = 20rect (100 F ) 1 ( F ) dF =

2 1 1/200
1/2
20rect (100 F ) dF
2
E x = 400
1/200
57
58

dF = 4



59


60

10


Using the equivalence property of the impulse and fact that both e j 2 F and 1 ( F ) have a fundamental period of one,

The DFT can often be used to find the DTFT of a signal. The DTFT is defined by X ( F ) =
N 1 n=0 n =
( F 1 /12 ) j /6 1 ( F + 1 /12 ) Y ( F ) = (1 / 2 ) e j /6 1 j /6 +e e 0.9 e j /6 0.9 Finding a common denominator and simplifying,

Y ( F ) = (1 / 2 ) 1.81 1.8 cos ( / 6 )
x[n]e
j 2 Fn
and the DFT
is defined by X [ k ] = x [ n ] e j 2 kn / N . If the signal x [ n ] is causal and time limited, the summation in the DTFT is over a finite range of n values beginning with 0 and we can set the value of N by letting N 1 be the last value of n needed to cover that finite range. Then X ( F ) = x [ n ] e j 2 Fn . Now let F k / N yielding
n=0 N 1
1 ( F 1 /12 ) (1 0.9 e j /6 ) + 1 ( F + 1 /12 ) (1 0.9 e j /6 )
Y ( F ) = 0.4391 1 ( F 1 /12 ) + 1 ( F + 1 /12 )
y [ n ] = 0.8782 cos ( 2 n /12 ) + 1.7914 sin ( 2 n /12 ) y [ n ] = 1.995 cos ( 2 n /12 1.115 )
+ j 0.8957 1 ( F + 1 /12 ) 1 ( F 1 /12 )
X ( k / N ) = x [ n ] e j 2 kn / N = X [ k ]
n=0
61
62

N 1


The result X( k / N ) = x[n]e
n=0 N 1 j 2 kn / N

We can also use the DFT to approximate the inverse DTFT. The inverse DTFT is defined by x [ n ] = X ( F ) e j 2 Fn dF
1
= X[ k ]
is the DTFT of x [ n ] at a discrete set of frequencies F = k / N or = 2 k / N . If that resolution in frequency is not sufficient, N can be made larger by augmenting the previous set of x [ n ] values with zeros. That reduces the space between frequency points thereby increasing the resolution. This technique is called zero padding.
and the inverse DFT is defined by x [ n ] = We can approximate the inverse DTFT by x[n]
k=0 N 1 ( k +1) / N
1 N 1 X [ k ] e j 2 kn / N . N k=0
( k +1) / N
k/N
k/N
X ( k / N ) e j 2 Fn dF = X ( k / N )
k=0
N 1
e j 2 Fn dF
x[n] X( k / N )
k=0
N 1
e j 2 ( k +1)n / N e j 2 k / N e j 2 n / N 1 N 1 = X ( k / N ) e j 2 kn / N j 2 n j 2 n k = 0 1 N 1 X ( k / N ) e j 2 kn / N N k=0
64

x [ n ] e j n / N sinc ( n / N )
63



For n << N , x[n] 1 N 1 X ( k / N ) e j 2 kn / N N k=0

N = 512 ; k = [0 : N - 1]' ; % % % % Number of pts to approximate X(F) Harmonic numbers Compute samples from X(F) between 0 and 1 assuming periodic repetition with period 1
This is the inverse DFT with X [ k ] = X ( k / N ) . Use this result to find the inverse DTFT of X(F ) = rect ( 50 ( F 1 / 4 )) + rect ( 50 ( F + 1 / 4 )) 1 ( F ) with the inverse DFT.
X = rect(50 * (k / N - 1 / 4)) + rect(50 * (k / N - 3 / 4)) ; % % Compute the approximate inverse DTFT and center the function on n = 0
xa = real(fftshift(ifft(X))) ;
n = [-N / 2 : N / 2 - 1]' ; % % Vector of discrete times for plotting
Compute exact x[n] from exact inverse DTFT
xe = sinc(n / 50).* cos(pi * n / 2) / 25 ;

65


66

11


The Four Fourier Methods


67


68

Relations Among Fourier Methods

Multiplication-Convolution Duality

Discrete Frequency Continuous Time Discrete Time
FS x ( t ) y ( t ) X [ k ] Y[ k ]

Parseval s Theorem

Discrete Frequency 1 Continuous Time T0 Discrete Time Continuous Frequency
2
Continuous Frequency
F x ( t ) y ( t ) X ( f ) Y( f )
FS F x [ n ] y [ n ] Y[ k ] X [ k ] x [ n ] y [ n ] X ( F ) Y( F )
x (t )
T0 2
dt =
k =
X[ k ]
x (t )
dt =
2
X( f )
1
df
Discrete Frequency Continuous Time Discrete Time x ( t ) y ( t ) T0 X [ k ] Y[ k ]

FS
Continuous Frequency x ( t ) y ( t ) X ( f ) Y( f )
F
n= N
x[n]
1 = N
k= N
X[ k ]
n =
x[n]
= X ( F ) dF
2
FS F x [ n ] y [ n ] N 0 Y[ k ] X [ k ] x [ n ] y [ n ] X ( F ) Y( F )

69


70


Time and Frequency Shifting

FS x [ n n0 ] X [ k ] e j 2 kn0 / N FS x ( t t 0 ) c x [ k ] e j 2 kt0 / N

F x [ n n0 ] X e j e jn0 F x ( t t 0 ) X ( j ) e j t0
( )

FS x ( t ) e+ j 2 kt /T c x [ k k0 ] T FS x [ n ] e+ j 2 kn / N X [ k k0 ] N
F x ( t ) e+ j 0t X j ( 0 ) F x [ n ] e+ j0 n X e j (0 )

71


72

12

Discrete Time Fourier Methods

Încărcat de

Informații document

Titlu original

Drepturi de autor

Formate disponibile

Partajați acest document

Partajați sau inserați document

Opțiuni de partajare

Vi se pare util acest document?

Este necorespunzător acest conținut?

Drepturi de autor:

Formate disponibile

Discrete Time Fourier Methods

Încărcat de

Drepturi de autor:

Formate disponibile

Discrete-Time Fourier Series Concept

A signal can be represented as a linear combination of sinusoids.