Professional Documents
Culture Documents
The DFT and FFT
The DFT and FFT
X ( e j ) = x[n ]e j n
n =0
(1)
1
X ( e j ) = X ( e j ) = X a j
T
T
4-1
1
X e j X a j
T
T
( )
= 2 k / N ;
k = 0,1,..., N 1
j
use X ( e ) instead.
j
associated with X ( e ) and simply
X [k ] = X ( e
N 1
N 1
= 2 k / N
= x[n ]W Nkn ;
= x[ n]e
j ( 2 k / N )n
n =0
k = 0,1,..., N 1
n =0
4-2
where
WN = e j 2 / N
1
x[ n] =
N
X [k ]W
kn
N
n = 0,1,..., N 1
k =0
Proof:
1
N
N 1
X [k ]W
kn
N
k =0
1
=
N
N 1
km
kn
x
[
m
]
W
N WN
k =0 m =0
1
=
N
N 1 k ( m n )
x[ m] WN
m= 0
k =0
N 1
N 1
But
N 1
W
k =0
k ( m n )
N
1 WN( mn ) N 1 e j (2 / N )(m n )N
=
=
(m n )
1 WN
1 e j (2 / N )(m n )
N
=
0
m=n
mn
So
4-3
1
N
N 1
X [ k ]W
kn
N
= x[ n]
k =0
The DFT and the IDFT pair can also be represented in matrix form as
1
1
X [0] 1
X [1] 1 W 1
2
W
N
N
X [2] = 1 WN2
WN4
M
M
M
M
X [ N 1] 1 WN( N 1) WN( N 1)2
or
L
L
L
L
L
x[0]
WNN 1 x[1]
WN2( N 1) x[2]
M
M
( N 1)( N 1)
WN
x[ N 1]
1
X = Wx
where
x[0]
x[1]
,
x=
M
x[ N 1]
X [0]
X [1]
,
X =
X [ N 1]
1
1
1
1 W 1
2
W
N
N
W = 1 WN2
WN4
M
M
M
1 WN( N 1) WN( N 1)2
L
1
N 1
L
WN
2( N 1)
L WN
L
M
( N 1)( N 1)
L WN
Note that
W 1 =
1
W
N
4-4
We can deduce from the matrix representation of the DFT that its
computational complexity is in the order of O ( N 2 ) .
X (e
N 1
) = x[n]e
j n
n= 0
1 N 1
= X [ k ]WN kn e j n
n= 0 N k = 0
1 N 1
N 1 j ( 2 k / N ) n
= X [k ] e
N k =0
n =0
N 1
1 N 1
1 e j ( 2 k / N ) N
= X [k ]
N k =0
1 e j ( 2 k / N )
1 e j N
=
N
N 1
X [k ]
j j 2 k / N
e
k =0 1 e
Since the FT is the ZT evaluated on the unit circle, this means we can also
obtain the ZT from the DFT coefficients by replacing the term e j in the
above equation by z . In other word
4-5
1 z N
X ( z) =
N
N 1
1 z
k =0
X [k ]
1 j 2 k / N
e
y[ n] = x1[ n] x2[n]
j
j
j
Let X 1 ( e ) , X 2 ( e ) , and Y ( e ) be the FTs of x1[n ] , x2 [n] , and y[ n]
respectively. Then
( )
( ) ( )
Y e j = X 1 e j X 2 e j
4-6
X 1[k ] = x1[n]WNkn = X 1 ( e j )
n= 0
= 2 k / N
N 1
X 2 [ k ] = x2 [ n]WNkn = X 2 ( e j )
n =0
= 2 k / N
= 2 k / N
= X 1[k ] X 2[ k ]
1
y [n ] =
N
1
=
N
N 1
Y[k ]W
kn
N
k =0
j 2 km / N j 2 kn / N
y
[
m
]
e
e
k =0 m =
N 1
N 1 j 2 k ( n m )/ N
y[m ] e
m =
k =0
1
1 e j 2 k ( n m )
=
y[m ]
N m=
1 e j 2 k ( n m )/ N
1
=
N
Since
4-7
m = n + rN
1 e j 2 k ( n m)
N
=
1 e j 2 k ( n m ) / N 0
otherwise ,
y [n ] =
y[n + rN ]
r =
Example: Let
1
x1[ n] = x2 [n] =
0
0 n3
otherwise .
2
1
3 4 3
1
n
0 1 2 3 4 5 6
and y[n] = 4 , n = 0,1,2,3.
4-8
The cause for aliasing in y[n ] is the lack of sufficient DFT coefficients in
its construction. Note that y[ n] is of duration 2N 1 samples so we need as
many DFT coefficients to capture all the information about the signal.
M = 2N
samples, where x1[ n] = x2 [n] = 0 when N n 2 N 1 . Note that in
general, M can be any number as long as it is not smaller than the duration
of y[ n] .
We then take the M-point DFTs of x1[ n] and x2[n ] to obtain
M 1
X 1[k ] = x1[n]WMnk = X 1 ( e j )
n= 0
X 2[k ] =
M 1
x [ n]W
2
n =0
nk
M
= X 2 ( e j )
= k / N
= k / N
Y [ k ] = X1[ k ] X 2 [k ] = Y ( e j )
= k / N
k = 0,1,..., M 1
4-9
1
N
M 1
Y[k ]W
k =0
kn
M
r =
r =
y[ n + rM ] = y[ n + 2 rN ]
1
N
M 1
Y [k ]W
kn
M
k =0
= y [ n]
y[ n] =
y[n + rN ] ,
r =
(2)
4-10
Let
x%1[n] = x1 ( n ) N
and
x%2[ n] = x2 ( n ) N ,
and
x%1[m + rN ] = x1[m],
m = 0,1,..., N 1
x%2 [m + rN ] = x2 [m],
m = 0,1,..., N 1
Conversely, the signals x1[ n] and x2[n ] can be obtained from x%1[n ] and
x%2 [n ] according to
x2 [ n] = x%2 [ n]RN [ n] ,
where
1
RN [n ] =
0
0 n N 1
otherwise
4-11
4-12
We want to show that the signal y[ n] in Eqn (2) is simply one period of
x%3 [ n] , i.e.
y[ n] = x%3 [ n]RN [n ]
Proof:
- Since the signal x%3 [ n ] is periodic, we resctrict our attention to time
indices n between 0 and N-1.
- Since 0 m N 1 , the term x%1[ m] in the circular convolution is
simply x1[ m] .
- The index n m of the term x%2 [n m] can be as small as 1 N or as
large as N 1 . When n m 0 ,
x%2 [n m] = x2 [ n m];
nm 0
x%2 [n m] = x2 [ N + n m];
nm<0
= x%1[ m]x%2[ n m] +
N 1
m=0
m=n +1
N 1
= x1[ m]x2[ n m] +
m=0
m=n +1
x%1[ m]x%2[ n m]
x1[ m]x2[ N + n m]
4-13
1
=
m= 0 N
n
[ r ]W r ( n m)
X
2
N
k =0
r =0
N 1
N 1
1 N 1
km 1
+ X 1[k ]WN X 2 [r ]WN r ( N + n m )
m = n +1 N k = 0
N r = 0
1
= 2
N
1
= 2
N
N 1
X 1[k ]W
km
N
1
N
N 1
n
nr
( rk ) m
X
[
k
]
X
[
r
]
W
W
1
2
N
N
k =0 r =0
m =0
N 1
1 N 1 N 1
(n +N )r
( r k )m
+ 2 X 1[ k ]X 2 [ r ]WN
W
N
N k =0 r =0
m= n +1
N 1 N 1
X 1[ k ]X 2 [ r ]WN nr WN( r k ) m
k =0 r =0
m =0
N 1 N 1
N 1
But
N 1
W
m= 0
( r k ) m
N
N
=
0
r=k
otherwise
So
1
x% 3[ n ] =
N
N 1
X [k ]X
1
k =0
[k ]WN nk
= IDFT X 1[ k ] X 2[k ]
= y [n ]
for 0 n N 1
4-14
0 n N 1
1
w[n] =
0
otherwise
Let x[n ] and y[n ] = x[n ]w[ n ] be the signals before and after truncation. As
shown in Chapter 2, the FT of y[ n] is
( )
Y e
1
=
2
X (e )W ( e
j
j ( )
) d ,
where X ( e j ) and W ( e j ) are respectively the FTs of the signal x[n ] and
the windowing function w[n ] .
Y [ k ] = y[n ]WNkn
n= 0
N 1
= x[n ]WNkn
n= 0
= Y ( e j )
= 2 k / N
k = 0,1,..., N 1
4-15
Since
( )
( )
( )
X e j at = 2 k / N .
Hamming,
Hanning,
Kaiser, and
Blackman
We will talk more about them in the Chapter on FIR filter design.
x[n ] = a nu[n] .
where the parameter a is real, positive, and less than 1. The FT of this
signal is
( )
X e j =
1
1 ae j
x[n ]
y[ n] =
0
0 n N 1
otherwise
4-16
Y [ k ] = y[n ]W
kn
N
n= 0
N 1
= a W
n
N 1
= ( ae j 2 k / N )
kn
N
n= 0
1 ( ae j 2 k / N )
n= 0
1 ae j 2 k / N
1 aN
=
1 ae j 2 k / N
= (1 a N ) X (e j )
= 2 k / N
X [k ] = x[n ]WNkn = X ( e j )
n =0
=2 k / N
k = 0,1,..., N 1
What happens when the signal is not casual, like the discrete-time version
of the SQRC pulse used in our project?
N N
Let x[n ] be a non-casual signal time-limited to the range 2 , 2 1,
where the duration N is assumed to be an even number.
4-17
N 1
X [k ] = x n N2 WNkn
n= 0
N 1
= x n N2 e j 2 kn / N
n= 0
N 1
= x n N2 e j 2 k ( n N / 2 ) /N e j k
n= 0
N / 2 1
x [ m]e
j 2 km / N
j k
m = N / 2
= e j k X ( e j )
j
However, this is not exactly X ( e )
= 2 k / N
=2 k / N
j
In order that the k-th DFT coefficient equals X ( e ) at = 2 k / N when
the signal is 2-sided, we should take instead the DFT of the signal
0 n N2 1
x[ n]
y[ n] =
x[ n N ]
N
2
n N 1
Proof:
X (e
N / 2 1
)=
x[n]e j n
n = N / 2
N / 2 1
x[n ]e
j n
n =0
n =0
x[n ]e j n
n = N / 2
N / 2 1
x[n ]e
j n
N 1
x[ n N ]e j ( n N )
n= N / 2
4-18
( )
X e
=2 k / N
N / 21
x[n]e
j 2 kn / N
n =0
N / 21
n =0
N 1
x[ n N ]e j 2 k ( n N )/ N
n =N /2
kn
N
x[ n]W
N 1
n=N / 2
x[ n N ]WNkN
N 1
= y[n ]W Nkn
n =0
= Y [k ]
x[ n]
4 3 2 1 0 1
n
2 3
y[ n]
n
0 1
2 3 4
5 6
4-19
As discussed in the last section, the direct method of computing the DFT
coefficients has a complexity of N 2 complex multiplications and N ( N 1)
complex additions.
Each complex multiplication itself requires 4 real multiplications and 2 real
additions.
Each complex addition involves two real additions.
So the total number of real multiplications is 4N 2 and the total number of
real addition is 4 N 2 2 N .
WNk ( N n ) = WN kn = (WNkn )
N = 2v
where v is an integer. The idea, however, can be generalized to any other
composite (i.e. non-prime) value of N .
X [ k ] = x[n ]WNkn ;
k = 0,1,..., N 1
n =0
X [k ] =
x[n ]WNkn +
n even
2 kr
N
x[2r ]W
r =0
N / 2 1
r =0
N / 2 1
kn
N
n odd
N / 2 1
x[n ]W
2 kr
N
x[2r ]W
r =0
+W
N / 2 1
k
N
x[2r + 1]W N2 kr
r =0
WN2 kr = e j 2 (2 kr )/ N = e j 2 kr /( N / 2 ) = WNkr/ 2
4-21
X [k ] =
N / 2 1
2 kr
N
x[2r ]W
N / 2 1
+W
k
N
r =0
N / 2 1
x[2r + 1]WN2 kr
r =0
r=0
N / 2 1
r=0
= G[ k ] + H [ k ]
where
G[ k ] =
N / 2 1
r =0
x[2r ]WNkr/ 2 ;
k = 0,1,..., N 1
and
H [k ] =
N / 2 1
r=0
x[2r + 1]WNkr/ 2 ;
k = 0,1,..., N 1
Since
WN( m/ 2+ N / 2 ) r = WNmr/ 2 ,
this means
G[ m + N /2] = G[ m];
and
H [ m + N /2] = H [ m];
m = 0,1,...N / 2 1
m = 0,1,...N / 2 1 .
In the above diagram, the 8 DFT coefficients X [0] to X [7] are computed
from G[0], G[1], G[2], G[3] and H[0], H [1], H[2], H[3] via N / 2 = 4 butterfly
structures. Each butterfly is of the form
G [k ]
1
G [k ]
X [k ]
WNk
1
H [k]
WNk + N / 2
X [ k + N / 2]
4-23
sk = s jt jk
j =1
WNk + N / 2 = WNk ,
the number of multiplications can be reduced to 1 if the following structure
is adopted instead
1
G [k ]
G [k ]
X [k ]
1
1G[k ]
H [k]
WNk
X [ k + N / 2]
4-24
The overall computational structure now takes on the form shown in the
diagram in the next page. The two N / 2 -point DFTs are now replaced by 4
butterflies.
Since the number of butterflies in the second stage (from the right) is the
same as that in the first stage, the computational complexities of both
stages are identical and equal to 4 CMs and 8 CAs.
In general, (when N = 2v ,) there are always N / 2 butterflies at each stage.
4-25
Each of the four N / 4 -point DFTs in the above example can be computed
using the following butterfly
1
G [k ]
W20 = 1
1
W21 = WNN / 2 = 1
4-26
4-27
nM = v
N
1
1 = N log 2 ( N )
2
2
n A = 2n M = N log 2 ( N ) .
In conclusion, the complexity of this decimation-in-time algorithm is in the
order of N log N , which is much smaller than N 2 for large N.
be the array containing the shuffled data. In the N = 8 example, this vector
is A = ( x[0], x[4], x[2], x [6],x [1], x[5], x[3], x[7] ) . Furthermore, let
B = ( bv 1, bv 2 ,..., b0 );
b j {0,1}
n = bj 2 j
j =0
Then
a[n] = x[m]
4-28
where
v1
m = b j 2v 1 j
j= 0
In-place computations
The FFT algorithm computes the DFT coefficients in stages.
The input data at the (m-1)-th stage ( m = 0 to v from left to right) are
transformed into another set of data that will be used in the computations in
the m-th stage.
Clearly, we can store the data to be computed in one array and those used
in the computation in another array. This requires a total of 2N storage.
However, the FFT algorithm we just described can be used to implement
in-place computations. This will essentially reduce the storage to about
50%.
Let us first index the data at the m-th stage in the signal flow graph as
X m [l] , where l = 0,1,..., N 1 (from top to bottom), then it can be easily
shown that the pair
( X m [ p], Xm [ q])
4-29
( X m 1[ p], X m 1[ q])
using the butterfly
1
G [k ]
X m 1[ p ]
X m[ p ]
1
1G[k ]
X m 1 [q ]
r
N
X m [q ]
This means as far as data storage is concerned, we need only one array of
N samples (and a temporary buffer of 2 samples). At the beginning of the
m-th stage, this array stores all the X m 1[l] , l = 0,1,..., N 1 . As soon as
X m [ p ] and X m [q ] are computed, the corresponding entries in the array are
then updated.
There are other versions of FFT that do not require re-ordering of the input;
see Fig. 9.14-9.16 of Text. Some of these algorithms, however, have the
disadvantage that computations can not be implemented in place. Others,
on the other hand, requires re-ordering of the output (see for example the
decimation-in-frequency FFT algorithm described in Section 9-4).
4-30
QAM Modulator
z% (t )
Bit duration
fc
Conventional serial modem
4-31
New bit
duration
QAM
Modulator
f N / 2
QAM
Modulator
s% (t )
f1 N / 2
Bit duration
M
QAM
Modulator
f N / 21
OFDM Modulator
1/ Ts
f
fc
Spectral characteristics of conventional serial modem. Ts is the
symbol duration of the signal z% (t ) .
1/ Ta
f
1/ T
fc
4-33
Why OFDM?
In many applications, the communication channel is not ideal, i.e. its
frequency response is not flat (nor linear phase) over the transmission band.
A nonideal channel causes severe intersymbol interference (ISI) in serial
modems.
By dividing the entire transmission band into subchannels, then as long as
the frequency response over a subchannel is flat (and linear phase), that
subchannel is ideal.
Without loss in generality, we can focus on the frame in the interval [0, Ta ] .
The OFDM signal in this interval can be written as
1
s% (t ) =
N
N / 2 1
k = N / 2
s%k (t );
0 t Ta
where
4-34
and
Y [ k ] {1, 3, K , (Q 1)}
are respectively the in-phase (I) and quadrature (Q) channel data of the kth modulator in the interval [0, Ta ] ,
M = Q2
is the number of points in the signal constellation, and
fk = fc +
k
Ta
is the k-th subcarrier frequency. The term fc is the center frequency of the
OFDM signal.
4-35
4-36
Exercise: Show that the overall bit rate of the above OFDM signal is
Rb = N log 2 M
1
Ta
Exercise: Convince yourself that when Q=2, the signal in each subchannel
is a QPSK signal with rectangular pulse shaping.
Ta
s% (t ) s%
k
Ta
(t ) dt = { X [ k ]cos ( 2 fk t ) Y[ k ]sin ( 2 fk t )}
0
{ X [ m]cos ( 2 f t ) Y [m]sin ( 2 f t )} dt
m
Ta
= X [ k ]X [m]cos ( 2 f k t ) cos ( 2 f m t ) dt
0
Ta
Ta
X [k ]Y [m]cos ( 2 fk t ) sin ( 2 fm t ) dt
0
Ta
Y[ k ]X [m]sin ( 2 fk t ) cos ( 2 fm t ) dt
0
Since
4-37
Ta
1
0 cos ( 2 f kt ) cos ( 2 f mt )dt = 2
Ta
{cos ( 2[ f
+ f m ]t ) + cos ( 2 [ f k f m ] t )} dt
{ (
)}
1 a
= cos 4 fc t + 2 kT+am t + cos 2 kTam t dt
20
T / 2
= a
0
Ta
1
2
km
Ta
{cos ( 2 [ f
fm ]t ) cos ( 2 [ fk + fm ]t)} dt
Ta
k =m
{ (
)}
1
cos 2 kTam t cos 4 f ct + 2 kT+am t dt
2 0
T / 2
= a
0
k =m
km
and
Ta
1 a
sin
2
f
t
cos
2
f
t
dt
=
(
)
(
)
{sin ( 2 [ fk fm ]t ) + sin ( 2 [ fk + fm ]t)} dt
k
m
0
2 0
T
{ (
)}
1 a
= sin 2 kTam t + sin 4 f ct + 2 kT+am t dt
20
=0
X 2 [ k ] + Y 2 [k ] Ta / 2
0 s%k (t) s%m ( t) dt = 0
Ta
k =m
k m
4-38
To ensure orthogonality, the different QAM modulators must be in phasesynchronsim. This is because
Ta
{ X [k ]cos ( 2 f t + ) Y [k ]sin ( 2 f t + )}
k
{ X [m]cos ( 2 f
t + m ) Y[ m]sin ( 2 f mt + m )} dt 0
Consider the k-th subchannel signal again. This signal can be written as
k
sk (t ) = S[ k ]exp j 2 t
Ta
4-39
1 N / 21
s% (t ) =
s%k (t )
N k = N / 2
1 N / 21
=
Re {sk (t )exp ( j 2 fc t )}
N k = N / 2
1 N / 21
= Re
s
(
t
)
exp
j
2
f
t
N / 2 k (
c )
N
k
=
= Re{ s (t )exp ( j 2 f c t )}
where
1
s( t ) =
N
N / 2 1
k
S
[
k
]exp
j
2
T
k = N / 2
a
= u (t ) + jv(t )
1
s
(
t
)
=
k
N
k = N / 2
N / 2 1
4-40
S [ N / 2 ]
s[0]
S [ N / 2 + 1]
s [1]
S [ N / 2]
IFFT
Re [ ]
Im [ ]
s [ N 1]
u [ n]
v [n ]
DAC
u (t )
DAC
v( t )
cos(2 f c t )
u ( t)
s% ( t)
v (t )
sin(2 f c t )
IF/RF segment of OFDM modulator
Let
s[ n ] = s(t )
t = nTa / N
4-41
1
s[ n] =
N
1
=
N
1
=
N
1
=
N
k
S [k ]exp j2 T t
k = N / 2
a t = nT / N
a
N / 2 1
k nTa
S
[
k
]exp
j
2
Ta N
k = N / 2
N / 21
N / 21
S [k ]W N kn
k = N / 2
N / 2 1
S [k ]W
k =0
kn
N
1
+
N
1 N / 21
=
S [k ]W N kn +
N k =0
1 N / 21
=
S [k ]W N kn +
N k =0
1 N 1
= D[k ]W N kn ;
N k =0
144244
3
1
N
1
N
k = N / 2
N 1
k= N / 2
N 1
S [k ]WN kn
S [ k N ]WN ( k N )n
S [ k N ]WN kn
k= N / 2
n = 0,1,2,..., N 1
IDFT of D [ k ]
where
S [k ]
D[ k ] =
S [k N ]
k = 0,1,..., N / 2 1
k = N / 2,..., N 1
At this point, it is obvious that we can exploit the efficiency of the FFT to
compute the s[n ] s.
The signal s( t) is obtained by passing the s[n ] s to a digital to analog
converter (DAC).
4-42
4-43
r (t ) = s (t ) + n (t )
be the received signal. Here n(t) represents channel noise. If the channel is
noiseless, then r (t ) = s(t ) . It is obvious in this ideal case that if the receiver
samples r (t) at t = nTa / N , n = 0,1,..., N 1 , and then takes a FFT of the
resultant DT signal, it will obtain the permutated data D[k ] . In the
presence of noise, the result simply becomes D[k ] plus a noise term.
1
a (t ) =
N
k
S
[
k
]exp
j
2
T
(
)
G ,
T
=
k N /2
a
N / 2 1
0 t T
where
T = Ta + TG
4-44
is the new OFDM symbol period, with Ta being the fundamental symbol
duration, and TG being the duration of the guard interval.
The frequency-spacing between adjacent sub-channels is still 1/ Ta .
Exercise: Determine the bit rate of the above modified OFDM signal,
assuming all the other system parameters remain the same .
The modified OFDM signal a (t ) is divided into two segments, one in the
interval [0, TG ] , and the other in the interval [TG , T ] . The latter segment is
called the effective data while the former is called the cyclic prefix or
guard interval. It should be clear that the effective data segment is
a(t ) = s (t TG ) ;
TG t T ,
where s (t ) is the basic OFDM signal in Section 4.3.1. The cyclic prefix is
simply a repeat of the last portion of the effective data; see figure below.
Effective
Data
Cyclic
prefix
t
0
TG
T
4-45
Assuming
TG = L ( Ta / N )
(i.e. the Guard interval is made up of L samples) and let
a[n] = a (t )
t = nTa / N
n = 0,1,..., N + L 1
a[ n] = s [ n L] ;
n = L, L + 1,..., L + N 1 .
a[n ] = s [ N L + n] ;
n = 0,1,..., L 1 .
g0 , g1 ,..., g L 1
and delays
4-46
0, t , 2 t ,K , ( L 1)t .
Here
t = Ta / N
O
O
O
O
O
g L 1 s[ N L],|| g L 1 s[ N L + 1], K ,
O
g L 1 s[ N L],|| g L 1 s[ N L + 1],K , g L 1 s[ N 1]
where the sign || || are the delimiters for the time interval [TG ,T]. Let the
received samples in this interval be denoted by r [n ] , n = 0,1,...., N 1 , and
define the vector r as
r[0]
r [1]
r=
M .
r [N 1]
It is observed that
L 1
r[ n] = gm s ( n m ) N ;
n = 0,1,..., N 1
m= 0
4-47
L 1
r = g m P s = g m P m W 1 D
m =0
m =0
L 1
where
0
1
P = 0
M
0
0 0 L
0 0 L
1 0 L
M M L
0 0 L
0 1
0 0
0 0
M M
1 0
W = 1 WN2
WN4
M
M
M
(
N
1)
(
N
1 WN
WN 1)2
L
L
L
L
M
( N 1)( N 1)
L WN
N 1
N
2( N 1)
N
1
W ,
N
and
D[0]
D[1]
,
D=
M
D[ N 1]
4-48
R = Wr
L1
= W g m Pm W1 D
m =0
L1
= g m WP m W 1 D
m =0
L1
= g mQ m D
m =0
WN0
m
W
N
1
2
m
m
1
m
WN
Qm = WP W = WP W =
N
O
WN( N 1)m
L 1
R = g m Qm D = HD ,
m =0
where
4-49
H0
L 1
H = g mQ m =
m =0
H1
H N 1
H k = g mWNkm ,
m= 0
fk = fc +
k
Ta
R[ k ] = H kD [k ] .
The result implies that there is no ISI, even though the channel is a nonideal channel.
4-50