Discrete Time Signal Processing Solution Manual
Discrete Time Signal Processing Solution Manual
Discrete Time Signal Processing Solution Manual
Solutions - Chapter 2
Discrete-Time Signals and Systems
3
= L a :z:,[n] + b L z 2[n]
A:=no .t=no
= aT(z,[n]) + bT(z2[n])
•
The system is linear.
• Not Tl:
·-L...
y[n- no] =
"' ....... :z:[l:]
The system is not TI.
• Not Memoryless: Values of y[n] depend on past values for n > no, so tbis is not memoryless.
(c) T(:z:[n]) L:::~ ... :z:[k]
...
• Stable: IT(z[n])l :;; I:::.~ ... lz[k]l $ L::.!..~ :z:[k]M $ l2no +liM for lz[n]l $ M, so it is
stable.
• Not Causal: T(:z:[n]) depends on future values of :z:[n], so it is not causal.
4
• Linear:
n+no
T(az1 [n] + bz2[n]) = L azt[k] + bz2[k]
t=n-no
n+no n+no
= a L :tt[k] +b L :t2[k] = aT(z,[n]) + bT(:t 2[n])
This is linear.
• TI:
a+no
T(z[n- no] = L:
.t=n-ne
z[k- nol
= L:
t=n-no
:t[k]
= 11!n- no]
This is TI.
• Not memoryless: The values of 11[n] depend on 2no other values of :t, not memoryless.
(d) T(:t[n]) = z[n- no]
• Stable: IT(z[n])l = [z[n- no]!~ M if [z[n] ~ M, so stable.
• Causality: If no ~ 0, this is causal, otherwise it is not causal.
• Linear:
T(az,[n] + bz•[n]) = az,[n- no]+ bx.[n- no]
= aT(:t,[n]) + bT(x.[n])
This is linear.
• TI: T(:t[n- nd] = :t[n- no- nd] = 11[n- n•l· This is TI.
• Not memoryless: Unless no = 0, this is not memoryless.
=
(e) T(:t[n]) e•l•l
• Stable: jz[n]l ~ M, ]T(x[n])l = [e•l•lj ~ el•l•ll ~eM, this is stable.
• Causal: It doesn't use future values of z[n], so it causal.
• Not linear:
T(az 1 [n] + b:t 2[n]) = eu•I•J+i>z•l•l
= eAZt(n)eb,[n]
2.2. For an LTI system, the output is obtained from the convolution of the input with the impulse response
of the system:
00
y(n] = L h(k]z(n - k]
b-oo
Note that the minimum value of (n - l:) is N2. Thus, the lower bound on n, which occurs for
k =No is
(b) H :r[n] # 0, for some n 0 :,; n:,; (no+ N- 1), and h[n] # 0, for some n, :,; n:,; (n, + M -1), the
results of part (a) imply that the output is nonzero for:
(n0 + n,) :,; n :,; (n0 + n 1 + M + N- 2)
So the output sequence is M + N- 1 samples long. This is an important quality of the convolution
for finite length sequences as we shall see in Chapter 8.
y[n] = L h[k]:r[n - k]
.k=-~
The step response results when the input is the unit step:
forn2:0
:r[n] = u[n] = { 0,1• for n < 0
Substitution into the convolution sum yields
00
y[n] = L 4-•u[-k]u[n- k]
b-oo
For n :,; 0:
00
y[n] = I:
1=-CX)
4-·
00
= I: 4•
b-n
4-n
= 1-4
Forn>O:
0
y[n] = I: 4-·
=
.. •=-em
r:4•
"="
1
= 1-4
7
The impulse response (for :r[n] = 6[n]) is the inverse Fourier transform of H(eiw).
H(eiw)- -8 + 8
- 1 + le-Jw 1- le-Jw
• 2
Thus,
h[n] = -8( 41)"u[n] + 8( 21)"u[n]. :::: y [>1)
2.5. (a) The homogeneous difference equation:
= 1 Sz- 1 + 6z- 2
-2 2
= 1 - 2z 1 +::---;:--,.
1 - 3z-l '
where the region of convergence is outside the outermost pole, ~ the system is causal. Hence
the ROC is jzj > 3. Taking the uiverse z-transform, the impulse response is
and
1 + 2e-;... +.-;:a...
=
1- !e-iw
H(eiw)
cross multiplying,
2.8. We take the Fourier transform of both h[n] and :t[n], and then use the fact that convolution in the time
domain is the same as multiplication in the frequency domain.
5
H(e'w) = 1 +le-i""'
2
Y(e"w) = H(e"w)X(e"w)
5 1
= 1 +le-i~>~ 1- fe
2
jw
3 2
= 1 + ~e jw + 1 - ie J·w
1 1
y[n] = 2(3)nu[n] + 3(-2)nu[n]
Y(eiw}
H(.,JW) = X(eiw)
le-2jw
= 3
1 - !e-iw + le 2jw
• •
Now we take the ;,;...,rse Fourier transform to find the impulse response:
H(eJ'W) =
h(n]
10
•
= L: h[kJ
1:=-oo
1- (1/3)•+• 1 - (1/2)n+'
= -2 1- 1/3 u[n] + 2 1- 1/2 u[n]
= (1 + (~)"- 2(~)")u[n]
(b) The homogeneous solution ll>[n] solves the difference equation when :[n] = 0. It is in the form
· ll>[n] = 2: A(c)", where the c's solve the quadrMic equation
5 1
c"--c+-=0
6 6
So for c = 1/2 and c = 1/3, the general form for the homogeneous solution is:
=
{ t_, ..
lo=-oo
n::; -1
L a•,
.......
a•
n> -1
= { 1-1/a'
1/a
1- 1/a'
n$ -1
n > -1
11
(b) First, let us define u[n] = 2nu(-n- 1]. Tben, from part (a), we know that
2'* 1 n < -1
w(n] = u[n] ov(n] ={ 1, ' n; _1
Now,
n>3
(c) Given the same definitions for u(n] and w(n] from part(b), we use the fact that h(n] = 2n- 1 u(-(n-
1) - 1] = u(n- 1] to reduce our work:
y[n] = r(n] • h(n]
= r[n]• u(n- 1]
= w{n -1]
{ 2n, n$0
= 1, n>O
(d) Again, we use u(n] and w[n] to help us.
H(eif) =
We get:
2../2£-i•l•ei•n/4 _ 2,.T2.J•I<e-i•n/4
y(n] =
2j
= 2J2sin(7:n/4- •/4).
12
:[n] = 6[n- 1]
the recursion yields
y[n] = 0, for n <0
y[O] = 0
y[1] = 1
y[2] =2
11[3] =6
11[4] = 24
Using h[n] from part (a),
2.13. Eigenfunctions of LTI systems are of the form a", so functions (a), (b), and (e) are eigenfunctions.
Notice that part (d), cos(Won) = .S(ei"'•" + .-;"''") is a sum of two a" functions, and is therefore not
an eigenfunction itself.
2.14. (a) The ioformation given shows that the system satisfies the eigenfunction property of exponential
sequences for LTI systems for one particular eigenfunction input. However, we do not know the
system response for any other eigenfunction. Hence, -can say that the system may be LTI, but
we cannot uniquely determine it. ==> (iv).
(b) If the system were LTI, the output should he in the form of A(1/2)", since (1/2)" would have been
an eigenfunction of the system. Since this is not true, the system cannot be LTI. ==> (i).
(c) Given the ioformation, the system may be LTI, but does not have to be. For example, for any
input other than the given one, the system may output 0, making this system non-LTI. ==> (iii).
If it were LTI, its system function can be found by using the DTIT:
Y(ei"')
H(ei"') = X(e;"')
1
= 1- le-;~o~
2
1
h(n] = (2')"u(n]
2.15. (a) No. Consider the following input/outputs:
The causal impulse response corresponds to assuming that the region of convergence extends outside
the outermost pole, making
hc[n] =((-1/4)" + 2(1/2)")u[n]
The anti-causal impulse response corresponds to assuming that the region of convergence is inside
the innermost pole, making
Y(z) = X(z)H(z)
1 1
= 1- !z-1 · (1 + lz-')(1- !z-•)
1/3 2 2/3
= 1 + 1/4z-l + 1- 1/2z-• + 1- 1/2z-l
1 1
y[n] = ( )"u[n] + 4(n + 1)("21 )"+lu[n + 1] + 2 (1 )"u[n]
34 32
2.17. (a) We have
r(n] = { ~: forOSnSM
otherwise
Taking the Fourier transform
M
R(eJw) = I>-jw•
n=O
1- .-jw(M+l)
=
(b) We have
[J- { i{1 +cos( 'if), for05n5M
wn- O, otherwise
We note that,
Thus,
W(_,..)
15
(c)
IR(ei ro )I
2.18. h[n] is causal if h[n] =0 for n < 0. Hence, (a) and (b) are causal, while (c), (d), and (e) are not.
2.19. h[n] is stable if it is absolutely summable.
(a) Not stable because h[n] goes to oo as n goes to oo.
(b) Stable, because h[n] is non...ero only for 0::; n::; 9.
(c) Stable.
_, 00
L: Jh!nll = L:
n n=-oo
3" = 2::<113>"
n=l
= 112 < oo
(d) Not stable. Notice tbat
So L:ih[n]l = 15.
2.20. (a) Taking the dif£erence equation y[n] = (1/a)y[n- 1] + z[n- 1] and assuming h[O] = 0 for n < 0:
h[O] = 0
h[1] = 1
h[2] = 1/a
h[3] = (1/a) 2
y[n] = T{x[n]},
Let x[n] = 0 for all n.
y[n] = T{x[n]}
For some arbitrary x 1 [n], we have
y 1 [n] = T{x1[n]}
Using the linearity of the system:
= L x[k]h[n - k]
k=-oo
. . .rr .. 0 2 3 °
(h) y[n] = x[n]• h{n]
5
0
~2
(c) y[n] = x[n]• h[n]
s s s s
4 4 4
t4
3 l 3 3 ' 3
2 2
2
IT 'I
O I 2 3 4 S 6 7 8 9 10 II 12 13 14 15 16 17 18 19 20 n
0123456 n
y[n] = Tz[n] = M 1
M L"'· z[n- k]
l + 2 + 1 k=--MI
by linearity;
T{a2:,[n] + bx,[n]} = 1
M, + M, + 1 L"''
lt=-Ml
(az,[n] + bo:,[n])
l lrl2 l M2
= M, + M2 + 1
~-~
L az,[n] + M1 + M 2 + 1 L
k=-M1
ln,[n]
= L z[k]h[n - k]
1=-oo
00
= L u[k - 4]h[n - k]
1=-oo
00
y[n] = L h[n - k]
t=4
EVllluating the above Slll1lmation:
For n < 4: =
y[n] 0
For n = 4: 11[n] = h[OJ = 1
For n = 5: 11[n] = h[1] + h[O] = 2
=
For.n 6: 11[n] = h{2] + h[l] + h[O] = 3
For n = 7: 11[n] = h[3] + h[2] + h[1] + h[OJ = 4
For n = 8: =
y[n] h[4] + h(3] + h{2] + h[l] + h[O] = 2
For n ~ 9: y[n] = h(5] + h{4] + h[3] + h(2] + h[l] + h[O] =0
18
= L o:[k]h[n - k]
'==-co
00
= L o:[k]u[n-k]
1:=-CIO
The convolution may be broken into five regions over the range of n:
N,
y[n] = :La•
>=<>
1_ 0 (N1+ll
= 1-a
, for N 1 < n < N2
N, n
N1 N1+N2
y[n] = :La• + L a<•-N,)
b:::O k=N2
N,
= :La• + L N,a"'
N,
= 2La•
.,..,
1 _ a(N,+l))
= 2· ( _
1 4
, forn > (N1 +N2 )
2.26. Recall that an eigenfnnction of a system is an input sigDal which appears at the output of the systel
scaled by a complex constant.
19
11[n] = L:
k=-oo
h[k]z[n- k]
00
= }: h[k]5<•-•lu[n - k]
t=-oo
n
= 5" L: h[kJ5-•
t=-oo
y[n] = L:
1::=-oo
h[k]ei""'<•-•1
00
= eJ1wn L: h[k]e-;:z..•
.t=-oo
= e'""'" · H(ei""')
YES, EIGENFUNCTION.
(c) eiwn + ei""'•:
00 00
L
00
L
00
Since the input cannot be extracted from the above expression, the sum of complex exponentials
is NOT AN EIGENFUNCTION. (Although, separately the inputs are eigenfunctions. In general,
complex exponential signals are always eigenfunctions of LTI systems.)
(d) z[n] = 5":
00
y[n] = L:
i:=-oo
h[k]5<•-•l
00
= 5" }: h[k]s-•
.11:=-oc
YES, EIGENFUNCTION.
(e) z[n] = s•ei""'":
00
=
b-oo
sneJ':L.In
..L: h[k]s-•.-1""'•
t=-oo
YES, EIGENFUNCTION.
20
2.27. • System A:
1
:~:(n] = (2)n
This input is an eigenfunction of an LTI system. That is, if the system is linear, the output will
be a replica of the input, scaled by a complex constant.
=
Since y(n] (t)", System A is NOT LTL
• System B:
:~:(n] =e-i•18 u[n]
The Fourie< transform of :~:(n] is
X(e-iw) = L 00
~/ 8 u(n]e-;wn
--00
00
= L:e-j(w-l)n
n-o
1
= 1- ,-;(w f).
The output is y[n] = 2:z:[n], thus
Y(eiW) = 2 .
1- ,-;(w-f)
Therefore, the frequency response of the system is
Hence, the system is a linear amplliie<. We conclude that System B is LTI, and unique.
• System C: Since :~:(n] = ei•/8 is an eigenfunction of an LTI system, we would expect the output to
be given by
y(n] = -yei•l•,
where 7 is some complex constant, if System C were indeed LTI. The given output, y(n] = 2ei•l•,
indicates that this is so.
Hence, System C is LTI. However, it is not 11llique, since the only constraint is that
2.28. z(n] is periodic with period N if :~:(n] = :~:(n + N] for some intege< N.
(a) :~:(n] is periodic with period 5:
(c) This is not periodic because the linear term n is not periodic.
(d) This is again not periodic. e;w is periodic over period 2.-, so we have to find k, N such that
:r[n + N] = .;(n+N) = .;(n+2d)
Since we can make k and N integers a.t the same time, :r[n] is not periodic.
2.29.
(a)
(b)
(c)
(d)
(e)
e I e .,
2.30. (a) Since cos(.-n) only takes on values of +1 or -1, this transformation outputs the current value of
=
:r[n] multiplied by eitber ±1. T(:r[n]) (-l)n:r:[n].
• Hence, it is stable, because it doesn't change the magnitude of :t[n] and hence satisfies bounded-
in fbounded-<>ut stability.
• It is causal, because each output depends only on the current value of :r:[n].
= = = =
• It is linear. Let 111[n] T(:r:,[n]) cos(.-n):r: 1 [n], and y,[n] T(:r:,[n]) cos(lm):t,[n]. Now
T(a:r: 1 [n] + b:r:2 [n]) =cos(.-n)(a:r:,[n] + b:r:2 [n]) = ay1 [n] + i>y,[n]
• It is not time-inwriant. If 11[n] = T(:r[n)) = (-1)n:r:[n], then T(:r:[n- 1)) = (-1)n:r:[n - 1] f.
y[n- 1].
(b) This transformation simply "samples" :r[n] at location which can be expressed as k'.
• 'l'be system is stable, siDce if :r{nj is bounded, :r[n2] is also bounded.
=
• It is not cansal. For eumple, T.o:{4] :r:[16].
= = =
• It is linear. Let llt[n] T(:,[n)) :r,(n2 j, and 112[nj = T(:r:,[n)) :r:2(n2 ]. Now
L6[n- k]
...., =u[n]
So T(x[n]) = :z:[n]u[n]. This transformation is therefore stable, causal, linear, but not time-
invariant.
To see that it is not time invariant, notice that T(6[n]) = 6[n], but T(6[n + 1]) = 0.
(d) T(:r[n]) = I;:.,._, :r[k]
• This is not stable. For example, T(u[n]) = oo for all n i. 1.
• It is not causal, since it sums /OMJJard in time.
• It is linear, since
00 00 00
• It is time-invariant. Let
00
then
T(:r[n- no]) = L"" x[k] = y[n- no]
i:=n-no-1
2.31. (a) The homogeneous solution !lh[n] solves the difference equation when x[n] = 0. It is in the form
y,[n] =I; A(c)", wbere the c's solve tbe quadratic equation
1 2
C+ - c - -
15 5
=0
So for c = 1/3 and c = -2/5, the general form for the homogeneous solution is:
Y(z) = X(z)H(z)
1 1
= 1-Jz-1 (1- !z-')(1 + iz-1)
-25/44 55/12 27/20
= + +
1- 1/3z-l 1 + 2/5z-1 1 3/5z- 1
y[n] = ~(!)"u[n]+ 55 (-!)''u[n]+ 27 (~)"u[n]
44 3 12 5 20 5
23
Therefore,
and
y[n] = ef•l•yt[n] = &f•l• cos(~n-
2
~)
2
2.33. Since H(e-iw) = H"(e.>w), we can apply the results of Example 2.13 from the text,
To find H(ei'f), we use the fact that H(eiw) is periodic over 211", so
-2
24
(b) Since
!lo[n] = -l:l:o[n + 1] + zo[n -1] = :J:o[n]• (-o[n + 1] + o[n- 1]),
h[n] = -6[n + 1] + 6[n- 1]
2.35. (a) Notice that zl[n] = z 2[n] + z 3[n + 4], so if T{·} is linear,
T{z [n]} = T{z,[n]}+T{z3[n+4]}
1
= y,[n] + Y•[n + 4]
From Fig P2.4, the above equality is not true. Hence, the system is NOT LINEAR.
(b) To find the impulse response of the system, we note that
6[n] = z 3[n + 4]
Therefore,
T{6[n]} = y3 [n+4]
= 36[n + 6].+ 26[n + 5]
(c) Since the system is known to be time-invariant and not linear, we cannot use choices such as:
Since,
L{o[n]} ""L{6[n -1]}
The system is NOT TIME INVARIANT.
(b) An impulse may be formed:
1 1
6[n] = 2"'' [n]- z,[n] + zs[n]
2
since the system is linear,
1 1
L{6[n]} = 2y,[n]- 2113[n] + 1/s[n]
= h[n]
25
from the figure,
J11 (n] = -6(n + 1] + U[n] + U[n- 1] + 6(n- 3]
112[n] = -6(n + 1] + 6(n]- U(n -1]- 6(n ~ 3]
lls(n] =U(n + 2] + 6(n + 1]- U(n] + U(n - 2]
Combining:
2 2
o 1 2 n
-2
2.37. For an LTI system, we use the convolution equation to obtain the output:
00
Let n = m+ N:
00
00
= L
t=-oo
:r((m- k) + N]h(kJ
Since :r(n] is periodic, :r(n] = :r(n + rN] for any integer r. Hence,
00
2.38. (a) The homogeneous solution to the second order difference equation,
3 1
y(n]- 41/(n - 1] + sllln - 2] = 2:r(n - 1],
3 1
Jl(n]- 4y(n - 1] + sy(n - 2] = 0
Solving,
3 -1
1- 4-z
1 -2
+ 8-z = 0•
26
ll•(n) =A,(~)".
Invoking the intial conditions, we have
11•(-1] = 2At = 1
11•{0] =A, =0
Evident from the above contradiction, the initial conditions cannot be met.
(d) The homogeneous difference equation:
1
11[n]- y(n- 1] + y[n- 2]
4 =0
Suppose the homogeneous solution is of the form
:t[n] = cos(wn)u[nj
= (-l)"u{nj,
the output is
"'
y[n] = L (j/2)•u[k](-1)(n-llu[n- k]
f:=-ao
"
(-1)" I;U/2) 1 {-W 1
=
11=0
"
= {-1)" I;H/2) 1
= {- 1)
"c-(
1=0
-j /2)(n+l))
1+j/2
(-1)"
y[n] = 1 + j/2
cos{wn)
=
l+j/2"
:t{n]
"'
= I: o[n + 16k},
k=-oo
has the Fourier representation
Therefore, the frequency representation of the input is also a periodic impulse train. There are 16
frequency impulses in the range -.- ~ w ~ 1r.
We sketch the magnitudes of X(&w) and H(eiw):
From the sketch, we observe that the LTI system is a lowpass filter which removes all but three of tbe
frequency impulses. To these, it multiplies a phase factor .-;..,.
The Fourier transform of the output is
1 1 .... 2">
Y(ei"') = -o(w) + -•-'" o(w- -
16 16 16
1 ·u2">
+ 16e' o(w + 16
Thus the output sequence is
1 1 21m 3"
tl(n] =-+-cos(-+-).
16 8 16 8
2.42. (a) From the figure,
cross multiplying,
Y(.,;->)[1- ae-;"'1 = X(ei"')[1 + 13•-;"'1
taking the inverse Fourier transform, we have
= !2 ·[X(ei"') + X"(ei"')1
1- aC06(w)
= 1- 2aC06(w) + a 2
1- 2aC06(w) + a2
(c) magnitude:
IX(e'"')l = [X(ei"')X"(ei"')1S
= C-2a~(w)+a2 )!
(d) phase:
LX(ei"') =arctan ( -asin(w) )
1- aC06(w)
00
2.44. (a)
X(ei"')j.,=O = L :[n)e-;"'n!w=O
--oo ..
L :[n1
= ,..._..,
=
(b)
X(.,;->)1..=• =
6
L
.. :(n1e-;•n
--00
= __L..,
00
:r(nj(-1)"
=2
31
(c) Because z[n] is symmetric about n = 2 this signal has linear phase.
X(.,;"')= A(w)e-;:z..
forn=O:
i: X(.,;"')dw =2.-z[O] =4.-
(e) Let y[n] be the unknown sequence. Then
Y(e-i"') = X(e-i"')
= L z[n].,;"'"
n
= L z[-n]e-i"'n
n
= L y[n]e-i"'"
n
Hence y[n) = z[-n).
• •I -4 o
-I
(f) We have determined that:
X(.,;"')= A(w)e-i:Z..
XR(ei"') = "-{X(.,;"')}
= A(w) cos(2w)
= !A(w) (.,;:z.. +e-;:z..)
2
• •
•-112 -4
I
. I 0 I . . I 112T ••••
I
4
-112
32
1-a2
X(ei"') = {1-"" '"')(1- at'"')
1 aei"'
= . + .
1-M-'~~~ 1- ae'"'
:[n) = anu(n] + a-nu[-n- 1]
a in\
=
(b)
1, [
-• X(ei"')cos(w)dw = -
1 j" X(ei"') e jw + e-jw dw
2 21< -· 2
= !2 ( 2_ [ X(ei"')ei"'dw+ 2_ [ X(ei"')e-'"'dw)
2'1' -· 211' -·
1
= 2(:z:[n -1) + :z;{n + 1J)
= !(
2
4 \n-1\ + 4 \n+ll)
2.47. (a)
(c)
H(~w) = 1 + 2e-jw
+ .-2jw
= 2e-jw(,!_~w + 1 + .!..-jw)
2 2
= 2e-Jw(cos(w) + 1)
(d)
IH(~w)l = 2(cos(w) + 1)
LH(~w) = -w
p
Magnitude It
-It It Ol
-It
-It
(e)
h.[n] = _.!._
2?r
1
<2~r>
H,{~w)~w dw
= _.!._ 1
2"' <211'>
H(e'(w+•)~wndw
= _.!._
2~
1
<2•>
H(~(w)e'(w-•l•tJw
= 0 -j•n...!...
2w <211'>
1 H(~(w)~wndw
= -1"h[n]
= 6[n]- 26[n- 1] + 6[n- 2]
..ttJJ
(b) Since y[n] = :[n]s[n],
= _!._
27r
1• S(~ )X(ei<~-•>)dw
-'JI'
1
= X(d~) + X(ei<~-•l)
Y(eiw) contains copies of X(ei"') replicated at intervals of.-.
(c) Since w[n] = y[n] + (1/2)(y[n + 1] + y[n- 1]),
So, for a > 2, Y(ei"') contains two noiH>verlapping replications of X(d"'), whereas for a < 2,
"aliasing" occurs. When there is aliasing, W(ei"') is not at all close to X(d"'). Hence, a must be
greater than 2 for w[n] to be "close" to :r[n].
oLJ~------------~-J oL---~--------L----"
-2 0 2
"
-2
.. 0
6.--------------------.
2
3~====~~====rl
2.5
{
2::
2
1.5~
-
~:
oL---~~------~U---~
-2 2 -2
.
0 2
10
0~~~------------~~ 0~~-U~------~U---~
-2 0 2
., •
0
(b) From part (a) we know that h[n] is length 3 with even symmetry around h[1]. Let h[O] = h[2] =a
and h[1] = b, from (iv) and using Parseval's theorem, we have
2c2 +b2 = 2.
From (v), we also have
2a -b= 0.
Solving the above equations, we get
1
h[O] = v'3
2
h[1] = v'3
1
h[2] = v'3
or
1·
h[O] = -v'l
2
h[1] = -v'l
1
h[2] = - y'l"
2.50. (a) Carrying out the convolution sum, we get the following sequence q[n]:
4 4
3 3 3 3
q[n]
1 1 1 1
n
012345678910
(b) Again carrying out the convolution sum, we get the following sequence r[n]:
444444 r[n]
3 3
44 n
0 1 2 3 4 5 6 7 8 9 10 11 12 13l14 .15 16
-4
-8
-12
-20
- 16
36
+oo
= L v[-k]w{k- n]
1:=-oo
+oo
= L v[r]w[-n- rJ where r = -k
= q[-n].
We thus conclude that q[-n] = v[ -n] • w[ -n].
2.51. For (-1 <a< 0), we have
X(.!"')= 1
1- ae jw
(c) m~tude:
IX(ei"'JI = [X(ei"')X"(ei"')]t
5lm
:r[n] = cos(T)
= cos( '"' )
2
.}'? .-;y
= -2-+-2-.
We now use the fact that complex exponentials are eigenfunctions of LTI systems, we get:
. 37
y(n]
2.53. First z(n] goes through a lowpass filter with cutoff frequency 0.5.-. Since the cosine bas a frequency of
0.6.-, it will be filtered out. The delayed impulse will be filtered to a delayed sine and the constant will
remain unchanged. We thus get:
[l = 3 nn(0.5.-(n- 5)) _
wn r(n- 5) +2
z[n] = cos(~-~)
4 3
= cos(----)
1m
4
"
3
= cos(-+-)
1m
4
"
3
eii .,;...- e-ife-i·.....
= 2
+
2
Using the fact that complex exponentials are eigenfunctions of LTI systems, we get:
2.55. Since system 1 is memoryless, it is time invariant. The input, z[n] is periodic in w, therefore w[n] will
also be periodic in w. A1; a consequence, y[n] is periodic in w and so is A.
2.56. (a)
where y,[n] and 112[n] are the responses to :r 1 [n] and :r2 [n] respectively. We thus conclude that
system S is linear.
(b) Let :r,[n] =:t[n- no], then:
112[n] = h[n]•
+co
(e-jwon:r,[n])
= L ,-;wo(•-•>:r,[n- k]h[k]
••-oo
+co
L ,-iwo(•-•>:r[n- no- k]h[k]
= 1=-co
'# y[n- no].
y[n] = h[n]•
+co
(e-iwo•z[n])
L .,-;... (n-•>:r[n -
= r--oo k]h[k]
+co
= L k=-oo
.,-jwon.,Jwo•:z:[n- k]h[k]
+oo
= e-;won
E
•=-oo
ei-•:z:[n- k]h[k].
We thus have:
39
1 .--
.....1.-.....L---.....L---.....L-.....L-... w
- .. -o.s.. 0 o.s.. 1(
We thus have:
H.(~"')
w
• -1r -0. 7.- -0.3lr 0 0.3.. 0. 7w "
(c) H 3 (e;"') corresponds to a periodic convolution of H,,(e'"') with another lowpass filter, specifically:
W(eiW)
_,.
-+-----------L--~~----~~"'
0
"
2.59. (a) Using the change of variable: r = -k, we can rewrite J4[n] as:
R,[n] = L"" :r"[-r]:r[n- r] =:r"[-n]• z[n].
r=-oo
We therefore have:
g[n] = z"[-n].
(b) The Fourier transform of z"[-n] is X"(~). therefore:
]4(eiw) = x·c.,;w)x(.,;w) = IX(eiw)l 2 -
2.60. (a) Note that z 2 [n] = -I:!~ :r[n- k]. Since the system is LTI, we haw:
l'>[n] = -
....L
..., y[n - k] .
41
-1 , n =0,n=2
h[n] = -~ , n =1
{ o.w.
2.61. The system is not stable, any bounded input that excites the zero input response will result in an
unbounded output.
The solution to the difference equation is given by:
Note that the first summation represents a weighted sum of future values of the input. Thus, if the
system is causal, _,
'E h[k]z[n- k] = 0.
b-oo
h[n]
2.64. Let the input be z[n] = o[n -1], if the system is causal then the output, 11[n], should be zero for n < 1.
Let's e-valuate y[O]:
81 (w)
0
-j
Bo(w)
...
=-3w + 2 lwl < "·
e.(w)
"
0
-j
_,
2.66. (a)
E(~"') = H 1 (~"')X(~"')
F(~"') = E(e-i"')
= H,(e-i"')X(e-i"')
G(~"') = H 1 (~"')F(~"')
= H 1 (~"')H 1 (e-i"')X(e-i"')
Y(e'"') = G(e-1"')
= H 1 (e-i"')H 1 (e'"')X(~"').
(b) Since:
We get:
2.67. (a) Using the properties of the Fourier transform and the fact that ( -1)" = e''", we get:
V(~"') = X{~(w+<))
W(~"') = H,(~"')V(~"')
= H,(e'"')X(~<w+•)
Y(ei"') = W(e'(w-•))
= H, (ei<w-•l)X (ei"')
(b)
44
H(.;"')
2.68. If "'' [n] =_z2 [n], UJ1 [n] and W:z[n] will not he necessarily equal.
w1 [n] = z,[-n-2]
W:z[n] = z.[-n + 2]
oF z.[-n- 2]
w,[n] = o[n + 2]
W:z[n] = o[n- 2].
2.69. (a) The overall system is not guarameed to he an LTl systan. A simple counterexample is:
y,[n] =
z[n]
112[n] = z[n]
y[n] = y,[n]1f2[n] = z 2 (n]
Y,(ei"') = H 1 (ei"')X(ei"')
Y.(ei"') = H2(ei"')X(e'"')
Y(ei"') = Y,(ei"') • Y2(ei"').
H(e'w) = 1- .-iw.
Thus the magnitude of the frequency response is
(d) In general,
So,
H,(.!~) = 1
1-e Jw'
and
h;(n] = u[n].
Hence, the unit step is the inverse system for the first difference.
2.11. For impulse response h[n], the frequency response of an LTI system is given by
oa
H(.!~) = L h[n]e-;~n
ta=-00
(b) We have
= L h•[n]eiwn.
--oo
If h[n] is real,
00
H•(eiw) = L h[n]eiwn
X (eiw) = L z[n]e-;wn
n=-oo
L z•[n]e-jwn =
n.=-oo
= (too z[l]e-jwl) •
x·<·-;w> = L: .,·[nJ·-;wn
ns-oo
for z[n] real, z[n] = z•[n], so
= L:
00
x·<·-jw> z[n1.-jwn
= X(ei"').
Tbus, the Fourier tr~orm of a real input is conjugate symmetric.
So,
property 10: IX(ei"')! = IX(e-iw)l
property 11: LX(ei"') = -LX(e-i"').
2. 7 4. Theorem 1:
00 00 00
Theorem 2:
= 00
o:[l]e-jw(t-n,)
L:
n=:-oo
o:[n- n·l··jwn = :E
l=-CJO
00
= ~41lft4
L:
l=-oo
x[l]e- 1"''
= ei"'"' X(ei"')
Theorem 3:
L
00 00
L z[n]eJ"'o"'e-jwn = :t[nje·i(w-.,.)n
n.=-ac
Theorem 4:
n.=-oo l=-oo
Theorem 5:
L
00
n:&[nje·jwn
n.=-OCl
2. 75. The output of an LTI system is obtained by the convolution sum,
00
Hence,
and
g[n] = z[n] • y"[-n)
form a transform pair.
(b)
00
= L L z[k)y"[k- n)e-i""
a=-cob-cc
forn=O:
1
[ . X(ei")Y"(.,;w)dw =
221' -oc
f:
t=-oo
z[kjy"[k)
50
sin(rn/4)
:(n] =
2m
sin(lm/6)
y"(n] = 511'11
We recogn.i2e eacll sequence to be a pulse in the frequency domain:
X(ei"')
-+-------L---L---L------~~ w
"
_,
(a)
which transforms to
(b)
Y•[n] = :r[2n]
(c)
• [l _{ :r[n/2], n even
y,n- 0, nodd
Y.(~W) = X(~""')
~·~
--t/.~:.._---~-.1-.~C-----~-·
-j i 0
w
2.79. (a)
00
+.(-N,-w) = L :r[n-N):r'[n+N)~wn.
,..._..,
+;(N,w) = (L
.. :r[n + N):r'[n- N]e-,wnr
=-oo
52
00
= L (o:(n + N]:r*(n- N]e-jwn)•
n=-oo
00
(b)
00
+.(N,w) = L Aan+Nu[n + N]Aan-Nu[n- N]e-;wn
n=-oo
= A2 L00
a2ne-jvln
nEN
00
= A2 L (a•e-iw)n
n=N
= A2 (a•e-iw)N
1- a2e-Jw
0 2Ne-iwN
= A•-"--':-~
1- c2e-iw.
(c)
L
00
X(e-i<•+(w/2))) = :r(n]e-j(v+w/2)n.
L
00
X*(e-i(•-(w/2))) = :r*(n]ei(v-w/2)n_
--co
Let S = .L
2Jf
J"
-1f
X(ei(•+(w/2H)X*(ei<•-(w/ 2 ll)ei2 •N dv ' then·.
s = -1 f" L 00 .
o:(n]e-'<•+w/2)n L..
:r*(k]e-i<•-wl2l•e-i2•N dv
27r -· n=-oo 1:=-oo
= l,.e o L L
co :r(n]:r*(k]e-;•<•t•> f" ..;<•-n+2N) dv
2 A=-oo.t=-oo -tr
= _!_ f: f:
2.- n=-oo >=-co
:r(n]:r*(k]e-;-';+» 2sin(r(k- n + 2N))
k - n + 2N
00
= " (
£.J z n]:r* n- 2N)e-'
[ .... sh-:aNl
'
= L 00
= +s(N,w).
2.80.
m, = E{ w[n]}
= E{o:[n] + y(n]}
= E{o:(n]} + E{y(n]}
= m.+m,
The variance of w[n]:
a!, =E{(w[n]- m,) 2 }
= E{ w'[n]} - m!
= E{(o:(n] + y(n])2 } - m!
=E{.,'(n]} + 2E{o:[n]l![n]} + E{y'[n]}- m~- 2m.m,- m~
If :z:[n] and y[n] are uncorrelated:
a! =
-_ .
a2+u2
.
E{o:2 (n]}- m! + E{y2 [n]} - m!
2.81. Let e[n] he a white Doise sequeDce and E{s(n]e(m]} = 0 for all nand m.
E{y[n]y(n + m]} = E{s(n]e[n]s[n + m]e(n + m]}
= E{s[n]s[n + m]e[n]e(n + m]}
SiDce s[n] is uncorrelated with e[n]:
E{y(n]y[n + m]} = E{ s(n]s(n + m]} E{ e(n]e[n + m]}
= cr!a;o[m]
2.82. (a)
1/>•• (m] = E(o:(n]o:[n + m])
= E((s(n] + e(n])(s(n + m] + e[n + m]))
= E(s[n]s(n + m]) + E(e[n]e[n + m]) + E(s(n]e(n + m]) + E(e[n]s[n + m])
= 1/>,.[m] + tl>.. (m] + 2E(e[n])E(s(n]) siDce s(n] and e(n] are iDdepeDdeDt and statiOIWY·
= tj>,,(m] + 1/>•• [m] where we assumed e[n] has zero mean.
Taking the Fourier transform of the above equation, we get:
(c)
~. [m] = E(:z:[n]s[n + m])
= E((s[n] + e[n])s[n + m])
= ,P.,[m] + E(e[n])E(s[n]) since s[n] and e[nJ are independent and stationary.
= ,P., [m] where we assumed e[n] has zero mean.
Taking the Fourier transform of the above equation, we get:
+.,(.;w) = +.,(.;w).
2.83. (Throughout this problem, we will usume lal < 1.)
(a)
IP••[m] = h[m]• h[-m].
IB(.;"')i' = H(.;w)B'(dw)
= B(dw)B(e-J"') since h[n] is real
= •••(eJw)
1 1
= (1- ae->w) (1- aeJ"')
1 1
= --.<
1-a 1- ae_,.,. + 1- 1oe''"'. ).
(c) Using Parseval's theorem:
-
L
-
lh[n]i•
= L lal 2"u[n]
+oc
= :E<1al 2 )"
ft=<l
1
= 1-1412 •
2.84. The first-backward-difference system is given by:
----~--~----~--
-1 0 1 m
To get the power spectrum, we take the Fourier transform of the autocorrelation function:
·-
(b) The average power of the output of the system is given by ti>.. [O]:
2.85. (a)
E{z(n]y(n]} =
..
E{z(n] E
..,_.., h(n]z(n - k]}
= L
.. h(k]E{z(n]z(n- k]}
=
..
1:=-CXI
E h!kJ<I>.. !kl
>=-co
Because z(n] is a real, statioua:y white noise process:
E{z(n]y(n]} = a! L
.. h[k]o[k]
= a!h[O].
(b) The variance of the output:
u! = E{(y[n]- ,.)2 }
= E{y2 (n]} - m;.
When a zero-mean random process is input to a determistic LTI system, the output is also zero-
mean:
= L
.. z[k]h[n- k].
""' =
..L E{z[n]}h[n- k]
.... = 0, ifm. = 0.
So,
u! = E{y2 [n]}
L h[m]h[k]o[m - k]
,.. =
a!E
..
m=-oo&=-oc
a! L h'[m].
• m,.;:-oo
57
~ =~ :E h~[kJ
=
....,
n
bn ( 1 - (a/W+') u[n]
n;<:O
1 -(a/b)
w[n] = o:[n] • h[n].
Since o:[n] is zero-mean, mw = 0 also.
o-~ = E{ w2 [n]}
00
= 0"~ :E h•[kJ.
•=<>
2.87. (a) x[n] is a stationary white noise process.
00
= L h[k]E{z[n- k]}
11:=-oo
(c)
co
= m, L h[k]
k=-oo .t=-oo
= m.,.
lim
n1,ft2-oo
~
~
~
L.J
h[k]h[m]t/>•• [n 1 - k, n2 - m] =~
~
~
L..J h[k]h[m]t/>.. [k, m].
t=-oo m=-oo i=-oo m=-oo
(d)
h[n] =anu[n]
co
E{y[n]} = m, L anu[n]
m.
= 1-a
2.88. (a) No, the system is not linear. In the expression of y[n], we have nonlinear terms such as :r2 [n] and
divisions by :r[n], :r[n- 1] and :r[n + 1].
(b) Yes, the system is shift invariant. If we shift the input by no, m,[n] shifts by no as well as ~[n]
and cr:!n], therefore y[n] shifts by no and the system is thus shift invariant.
(c) If :r[n] is bounded, m.[n] is bounded so is u~[n] and U:!n]. ~ a result, y[n] is bounded and
therefore the system is stable.
(d) No, the system is not causal. Values of the output at time n depend on values ofthe input at time
n +I (through u~[n] and m,[n]). Since present values of the ouput depend of future values of the
input, the system cannot be causal.
(e) Wben cr![n] is very large, u;[n] is zero, therefore:
y[n] = m,[n]
l n+l
= 3 L: :r[k]
i:=n-1
which is the average of the previous, present and next value of the input.
y[n] = :r[n].
y[n] makes sense for these extreme cases, because in very small noise power, the ouput is equal to
the input since the noise is negligible. On the other band, in very large noise power, the input is
too noisy and so the output is an average ofthe input.
2.89. (a)
E{ :r[n]:r[n]} = t/>.,[0].
(b)
+n(ei"') = X(ei"')X"(~w)
= W(~W)H(~W)W"(~w)H"(~w)
= +-(ei"')IH(~)I 2
= ~ 1
•t-cos(w)+ 1/4"
59
(c)
= cr;,p,,[n].
2.90. (a)
+IJ(eiw) = HI(e'w)H,(e-iw)+,,(e'w)
= (1- e-iw)(l - e'w)a~
= a~(2 _ e'w _ e-iw)
= ~(2- 2 cos(w)) .
..•
·-
60
(b) <Pf/[m] is the inverse Fourier transform of ~ 11 (e;..). Using part (a), we get:
(d)
•
61
Solutions - Chapter 3
The z- Transform
63
3.1. (a)
!z- (
1 lzl > 2
(b)
z[-G)"u(-n-1)] = - L:
-1 (1)"
2 .-· = - L:(2zl"
00
n=-oo n=l
2z 1 1
= -1- 2z = 1- !z 1 lzl < 2
(c)
(d)
Z[c5[n]J = z0 = I all z
(e)
d d I
n :r[n] <* -z-X(z)
dz
=:> n u[n] <* -zdz
1-z-1
lzl > 1
z-1
n u[n) <* (l _ z 1 )2. lzl > 1
-N-1
z[n- no] <* X(z) · .-... =:> (n- N)u[n- N] <* (1-
z z-1)2 lzl >I
therefore
64
3.3. (a)
-1 00
= La"z" + Lo"z-"
a=l n=O
az 1 z(I- a 2 ) 1
= - - + .,..-....:.....,...
1-az 1-az-1 (I - az)(z- a)' lal < lzl < lal
Xcf.Z)
IIa
(b)
I, 0 $ n < N- I N-1 I -N N I
%11 ={ 0, N < n- · ~ Xo(z) = "' z-n = - z 1 = '7.•=-,..,,...-...::...,,.,
0, n 0 Z ~ I- z zN 1 (z- I)
(c)
XrJ.z)
65
3.4. The pole-zero plot of X(z) appears below.
X(z}
2 3
(a) For the Fourier transform of :z:(n] to exist, the z-transform of :z:(n] must have an ROC which includes
the unit circle, therefore, Iii < lzl < 121.
Since this ROC lies outside ~, this pole contributes a right-sided sequence. Since the ROC lies
inside 2 and 3, these poles contribute left-sided sequences. The overall :z:(n] is therefore two-sided.
(b) Two-sided sequences have ROC's which look like washers. There are two possibilities. The ROC's
corresponding to these are: 1!1 < lzl < 121 and 121 < lzl < 131.
(c) The ROC must be a connected region. For stability, the ROC must contain the unit circle. For
causality the ROC must be outside the outermost pole. These conditions cannot be met by any of
the possible ROC's of this pole-zero plot.
3.5.
Therefore,
3.6. (a)
X(z)- _ _,_!.....,. 1
- 1 + lz-1 1•1 > 2
2
= :z:(n] =( -D n u(n]
66
(b)
1 1
X(z) = 1 1 1•1 < 2
1 + ••
Partial Fractions: one pole-+ inspection, :r[n) = -( -!Jnu[-n- 1]
Long division:
2z - 4z2 + 8z3 + ...
lz- 1 +1
2 1
1 + 2z
- 2z
- 2z - 4z2
+ 4z2
+4z' + Sz'
(d)
=> z[n) = [-3(( -n -n n+ ( n-2] u[n)
1- lz-1 1
2
X(z) = 1•1 > 2
1-lz 2
Partial Fractions:
•
1-lz-1 1 1
X(z) = 1- L-
•
2 = "1-+--2....1 .~1 1•1> 2
z[n) = ( -~ru[nJ
67
Long division: see part (i) above.
(e)
X(z) = 1z - 1 -a
(1Z-1
lzl > la- 11
Partial Fractions:
1
X(z) =-a- a- ( 1 - a•) lzl > la- 11
1- a-lz 1
-a.+z-1 I 1
·-'-•)
- ( --;y-z -· + ...
1 - a.z-1
(a- 1 - a)z 1
x[n] = .. r-n-
-1
11 + G
1
r u[n]
1
X(z) = 1- z 1 + 1- lz-1 2 < 1•1 < 1
2
Now to find H(z) we simply use H(z) =Y(z)/X(z); i.e.,
Y(z) -lz- 1 . (1- z- 1)(1- !z-1) = 1- z- 1
H(z) = X(z) = (1- !z 1)(1 + z-1) -!z-' 1 + z-l
Therefore,
y[n] = --3I(1)" I
-2 u[n] + -(-l)"u[n]
3
3.8. The causal system has system function
1-.-1
H (z) - --.'---:-
- 1 + lz-1
•
and the input is :[n] = (!)" u[n] + u[-n- 1). Therefore the z-transform of the input is
1 1 -jz-1 1
X (z) = - 1- z-1 = -::--r---7:--;:---~ 3 < 1•1 < 1
1- jz-1 (1 ""- jz-1)(1- z-1)
68
Finite length but has positive and negative powers at z in its X (z ). Therefore the ROC is 0 <
1•1 <(X).
(c)
(d)
cr·
2 u[-n] => ROC is lzl < 2
z[n] is right-sided, so its ROC extends outward from the outermost pole e1•1 3 • But since it is
non-zero at n = -1, the ROC does not include oo. So the ROC is 1 < lzl < oo.
(e)
z[n] is finite-length and has only positive powers of z in its X(z). So the ROC is lzl < oo.
(f)
z[n] is two-sided, with two poles. Its ROC is the ring between the twn poles: ! < 1•1 < j,-.'31 j, or
! < lzl < j.;.
3.11.
00
· which means this summation will include no positive powers of z. This means that the closed form of
= =
X(z) must converge at z oo, i.e., z oo must he in the ROC of X(z), or Jim,~ 00 X(z)-! oo.
(a)
( 1- .-1)2
Jim =1 could be causal
•~oo(1- !z-1)
70
(b)
(z- 1) 2
lim
•~oo (z -ll -co
- could not be causal
(c)
(• 1)5
lim -. =0 could be causal
<-+oo(z- f)'
(d)
(z- l4 ) 6
lim =oo could 110t be causal
<-+oo(z -!)•
3.12. (a)
1-l.-1
Xl(z) = 1 + ~z 1
(c)
1+z- 1 -2z-•
X3(z) = 1- •13• 1 +. 2
The poles are at 3/2 a.od 2/3, a.od the zeros are at 1 and -2. Since :r 3 [n] is absolutely summable,
the ROC must include the Ullit circle: 2/3 < lzl < 3/2.
3/2
71
3.13.
= Lg[n]z-"
ft
g[ll] is simply the coefficient in front of z- 11 in this power series expansion of G(z ):
3.14.
1
B(z) = 1- !z- 2
• 1
=
(1- !z- 1 )(1 + !• ')
0.5 0.5
= 1-lz- 1 + I+lz
2 2
1
3 < 1•1 < 2
72
5 5
Y(z) = 1- !z-1 l-iz I
-Jz-1 2
= (1 - }z-1 )(1 - fZ I)' lzl > 3
Now
Y(z)
H(z) = X(z)
2
= 1- ~z- 1 lzl > 3
3
H(z)
h[n]
(c) Since
H(z) = Y(z) = 1- ~z- 1
,
• X(z) 1- 3 z I
we can write
Y(z)(l- ~z- 1 ) = X(z)(l- 2z- 1 ),
whose inverse z-transform leads to
2
y[n]- 311[n- 1] = :t[n]- 2z[n- 1]
(d) The system is stable because the ROC includes the unit circle. It is also c:a.usaJ since the impulse
response h[n] = 0 for n < 0.
3.17. We solve this problem by finding the system function H(z) ofthe system, and then" looking at the
different impulse responses whith can result &om our thoice of the ROC.
Taking the z..transform of the difference equation, we get
1- z- 1
= (1-2z 1 )(1-iz- 1 )
= 2/3 + 1/3
1- 2z 1
1- ~z
lithe ROC is
(a) izl < i:
= h[O] = l.
(d) izl > 2 or izi < ~:
3.18. (a)
1 + 2'-1 + z- 2
H(z) = (1 + !z-
1 )(1- z-•)
! !
= -2+ ~ 1 + ""1__...%-_..,.1
1+ ~z
1
Y(z) = (1 + ~z-1)(1- tz-1)
The intersection of ROCs of H(z) and X(z) is 1•1 > ~· So the ROC of Y(z) is 1•1 > l·
(b) The ROC of Y(z) is exactly the .intersection of ROCs of H(z) and X(z): < 1•1 < 2. !
(c)
I
Y(z) = (1- i• 1)(1 + l• 1)
(b)
H(z) =
..
La"z-" =1- 1
1•1 > lal
az-1
n=O
N-1 1-z-N
X(z) =
n=O
L:·-"-- 1- z- 1 1•1 > 0
Therefore,
1- z-N 1 z-N
Y(z)- - -
- (1 - az-1 )(1- z-1) - (1- az-1 )(1- z 1) (1- az 1 }(I - z-•) 1•1 > lal
Now,
So
1- Qn+l 1 _ 4 n-N+l
= 1 - a u[n] - 1- a u[n - N]
0 n <0
1-a•+l
y[n] = 1-G OSnSN-1
3.22. (a)
~
y[n] = I:
i=-oc
h[k]:r[n - k]
(b)
Y(z) = H(z)X(z)
3 1
= 1 + lz-1
3
1 - z-1
• l !
= +
1 + !z-1 1- z-1
•
y[n] = -3 ( --
1r u[n] + -u[n]
9
n")
4 3 4
= ~ ( 1 + ~ (- u[n]
= 49 ( 1- ( -3l+1) u[n]
3.23. (a)
1- lz-2
H(z) = 2
(1-lz- 1}(1-lz 1)
5+ !z-1
= -4+ 1 _ lz-1 + lz-2
2
• 7
•
= -4-
1- lz
2
1
+
1- lz-1
•
(b)
3 1 1
y(n]- -y(n- 1] + -11[n- 2] = z(n]- -z(n- 2]
4 8 2
3.24. The plots of the sequences are shown below.
(a) Let
co
a(n] = L 6(n - 4k],
i=-oo
Then
co
A(z) =L z-•n
l=-oo
(b)
b(n]
B(z)
....
cu •••
•..
·~~~~===~~~~
·10-4 ....... ~0 • 10
n
.
c
Z"
.. ... l i r.
• ... .. . ... .. 0
n
• • • •
3.25.
X (z) = (z -
z•
a)(z - b)
z'
= -z~2--"""(a_::.+_,.b,...)z_+__,ab
1
z2 - (4 + b)z + ab z2
z2 - (4 + b)z + ab
(4+ b)z - ab
( + b) z - ab (•+>!•-•• (o+O)o--oo
X(z) = 1+ a =1+ •-• + •-•
~-~~-~ z-4 z-b
L L 1 ( 4 2z-l b'z-1 )
= 1+ ~
Z-4
- ..!.=!...
z-b
=1+ - -
4-b 1-<>z 1 - ;-'--7-::;-
1-bz-l
42 ,..
z[n] = .S[n] + _ b 4n- 1 u(n- 1]- _ b bn-lu[n- 1]
4 4
= .S[n] + (-1-)
4-b
(4n+l - bn+l )u(n- 1]
+ lz 2
•
Therefore, z(n] = 2(-l)nu(n]- .S(n]
(b)
Poles at !.
and-~. :r(n] stable,=> izl > ! =>causal.
Therefore,
:r(n] = 4 G) n ..[n]- 4 ( -n n ..[n]
(c)
1
X(z) = ln(1- 4z) izl < 4
Therefore,
78
(d)
+ j.z-3 - iz-6
+ lz •
•
n = 0,3,6, ...
= z(n]= {
otherwise
3.27. (a)
1 1
X(z) =
!z 1)2(1- 2z 1)(1- 3z 1) 2 < [z[ < 2
(I+
i s +dis-~+~
=(l+lz-2)2 (l+lz 1) (1-2z- 1) (1-3z 1)
Therefore,
Therefore,
z(n] = o(n + 2] + U(n + 1]- 2(2)"u(-n- 1]
3.28. (a)
d
nz(n] ~ -z dz X(z)
X(z)
3z-•
= (1- tz-1)2 = 12z
_2 [
-z dz
d ( 1
1- tz 1)]
z(n] is left-sided. Therefore, X(z) corresponds to:
1)n-2
:r(n] = -12(n- 2) ( ;( u[-n + 1]
79
(b)
Therefore,
Which is stable.
(c)
z7 - 2 7 1
X(z) = _ z- 7 = z - _ z- 7 lzl > 1
1
X(z)
1
= z7- L
.. z-7n
n=O
Therefore,
:t[nJ = o[n +
..
11- :E o[n- 7kJ
n=O
3.29.
3.30.
1 1
X(z) = log2 (- - z)
2 izl < 2
(a.)
00
X(z) =log(!- 2z) =-"-•-.
(2 )' -I 1 1
1
(I) l
.l..J • =- ".l..J -(2z)-'
-l = ".l..J -l -2 z-'
s=l t=-oo t=-oo
Therefore,
:t[n] = ;;
1 C)"
2 u[-n- 1]
(b)
n:t[n] .,. dz
1
-z!._log(l-2z)=-z(--)(-2)=z- 1
1- 2z
( -
1
1 - !z-1
)•
1
1•1 < 2
2
n:t[n] = Gf u[-n-1]
:t[n] = ;;~cr
2 u[-n-1]
80
ll.31. (a)
:z:(n] = cnu(n] + bnu(n] + enu(-n- 1] lei < lbl < lei
1 1 1
X(z) = +
1-az 1 1-bz-1 1-ez
- lbl < lzl < lei
1- 2ez- 1 + (be+ ac- cb)z-•
X(z) = (1- cz 1)(1- bz 1)(1- ez-1) lbl < izl < lei
Poles: a, b, c,
Zeros: z1, z2, ex> where z1 and z2 are roots of numerator quadratic.
pole-zero pi t {a)
(b)
:z:(n] = n 2 cnu(n]
1
:,[n] = anu[n] ¢> X,(z) =1 az- 1
izl >a
d az- 1
: 2 (n] = nz.(n] = nanu[n] ¢> X2(z) = -z dzX1(z) = (1-az ') 2 izi >"
1
:(n] = nzo(n] = n 2 c nu[n J ¢> -z dz
d
X2 (z) = -z dz
d ( cz-
( 1 _ cz- 1)•
) lzl >a
-az- 1(1 + az- 1)
X(z) = (1 - az 1)3 1•1 >"
(c)
z[n] causal =0> z[-n + 3] is left-sided =0> ROC is 0 < jzj < 4/3.
-4/3
zeros
poles
Y(z)
(b)
W(z) = !x(e-i•l
2
2 z) + !x(ei•12 z) = !x(-iz) + !x(iz)
2 2 2
W(z) = -2
1(-z +1) + -21(-r+1)
2
-jz- 2
1
2
z -1
= 2(z +
jz-
1
2
2 1)
i
poles at ±ti
zeros at ±1
82
Y(z)
3.34.
1 2
H(z)= 3-7z- +5z- =S+-;---;;-1~ 3
1- 2~z '+ z-2 1 2z-1 1- !z-1
(a)
n
y[n] = h[n] • z[n] = L h[l:]
n
- I: 2. = -2n+l n<O
k=-oo
=
- I: 2• + s- I:3 ~
-1 n ( )k = 4- 3 1 - ~
(')n+l
= -2 + 3 (1)n n;::o
1 1 2
=
t-oo
-2u[n] + 3 Gr 11:=0
(b)
1 1 1 1 1
Y(z) = 1- z H(z) = -21 - z -1 + 21 - 2z -1 + 31 - I z - I 2 < 1•1 < 2
Gr
1 '
2
y[n] = _ 2u[n] _ 2(2)nu[ -n _ 1] + 3 u[n]
3.35.
3 3
1-z
- - z - 1 (1-z-
- -)
H(z) = -
1- •• - 1- .-• 1•1 > 1
1 z
u[n] ~
1 zl=z-1 1•1 >1
z- 1 - z-4
U(z)H(z) = (1- z-<)(1- z- 1 )
.-1 .--
u[n] • h[n]
=
=
1- Z I
..
-1-z-<
bO
83
3.36.
1
z[n] = u[n] ¢> X(z) =
1
_ .- 1 1•1 > 1
1) n-1 (1)n+l 4z 1
11[n] = ( u[n + 1] = 4 u[n + 1] ¢> Y(z) = _ !z 1 1•1 > 2
2 2 1
(a)
(b)
4z 4 1
H(z) = t-lz-1 t-lz-1 1•1 >-2
2 2
1
h[n] = 4Gr+ u[n+1J-4Gru[nJ
= 40[n+1J-2Gru[nJ
l lz 1 1
z[O] = &-+OO
lim 1 - 1 z-l +lim J=-. = - +0= -
z-o z - 2 3
2 3
3.38.
z- 1 + z-2 2
Y(z)-
- (1- ~z- 1 )(1 + lz-1) . 1- z-
- -1 1•1 > 1
84
z10
X (z) = .,.-,..,..,.-....,.,,.,..:::..,-.,.,-;;-;-....-;-;--;-;-
(z- iHz- ~)'O(z + ~)2(z + j)(z + ll
Stable=> ROC includes jzj = 1. Therefore, the ROC is ! < jzj < ~-
(b) :r[-8] = E[residues of X(z)z-• inside C], where Cis contour in ROC (say the unit circle).
:r[8] = E [residues of (z - 1)( 3) z 3) (
2 z - 2 10 (z + 2 2 z + 25
)( 7 ) inside unit circle]
z+2
First order pole at z = ! is only one inside the unit circle. Therefore
1 1
:r[-8] = <!- ~)10(! + ~i·<! + ~><! + i> = 96
3.40. (a) After writing the following equalities:
= 1 = 1- z-1
1 + 1-r'
·-·
(c) H(z) is not stable due to its pole at z =1, but H 1 (z) and H2{z) are.
3.41. (a) Yes, h[nJ is BffiO stable if its ROC includes the unit circle. Hence, tbe system is stable if rmin < 1
and Tm.cz > 1.
(b) Let's consider the system step by step.
=
(i) First, v[n] a-•x[nJ. By t.aking the z-transform of both sides, V{z) X(oz). =
=
(ii) Second, v[n] is <ered to get w(n]. So W(z) H(z)V(z) H(z)X(az). =
=
(iii) Finally, y(n] o"w[n]. In the z-transform domain, Y(z) = W(zfo) = H(zfo)X(z).
In conclusion, the system is LTI, with system function G(z) = H(zfo) and g(n] = o"h[n].
85
(c) The ROC of G(z) is ar,.,. < 1•1 < ar.,... We want r,.,. < 1/a and r=• > 1/a for the system
to he stable. -
3.42. (a) h(n] is the response of the system when o:(n] = o(n]. Hence,
10
h(n] + La .1:] = <l(n] + tl<l[n- 1],
.... 1 h[n-
(b) Atn=1,
h(1] + a,h(O] = 6(1] + tlo(O} == a1 = p ~~~[
1
] = P- h[1]
(c) How can we extend h(n] for n > 10 and still have it compatible with the difference equation for S?
Note that tbe difference equation can describe systems up to order 10. If we choose
3.43. (a)
1 1
X(z)=1-!z-1 1-2z-1' 2 < I• I< 2
2
6 6 3
Y(z) = 1- jz-1 - 1 -lz 1' 1•1 > 4
H(z)
86
•
....
1
L 1
"' •
(b)
_,
-...
~00511.52
(c)
3
y(n]- 4y(n- 1] = z(n]- 2z(n- 1]
(d) The system is stable because the ROC includes the unit circle. It is also causal since h(nj = 0 for
n < 0.
3.44. (a)
_! i
X(z)
.z
= 1- 13 I
+ 3
1- 2z- 1
(c)
Y(z)
--
-u'-:_,,_-4.5,.--+o-;;...,--,,-1:-:;•-•'"'
H(z) = X(z)
llz-t
(1 z-l)(l+ ,.-i)(l-2z-i)
= (l-12z-Ij(l 2z-l)
87
= (1 + z- 1 )(1- !z-1 )
(1- z 1)(1- !z-1)
= 1+ 3
2
-i
1-z 1 + -1-+_,lf'-z---:-1
2
(d) Since H(z) has a pole on the unit circle, the system is not stable.
3.45. (a)
ny[nj = z[n]
dY(z)
-z~ = X(z)
Y(z) = -I z- 1X(z)dz
(b) To apply the results of part (a), we let z[n] = u[n- 1], and w[n] = y[n].
W(z) = -
I 1
z-1
z- 1 _ z- 1 dz
= - I 1
z(z- 1)
dz
= - 1-1
- + - -1d z
z z -1
= ln(z) -ln(z- 1)
3.46. (a) Since y[n] is stable, its ROC contains the unit-circle. Hence, Y(z) converges for~ < lzl < 2.
• (b) Since the ROC is a ring on the z-plane, y[n] is a two-sided sequence.
(c) z[n] is stable, so its ROC contains the unit-circle. Also, it has a zero at oo so the ROC includes
oo. ROC: lzl > ~·
(d) Since the ROC of z[n] includes oo, X(z) contains no positive powers of z, and so z[n] = 0 for
n < 0. Therefore z[n] is causal.
(e)
z[OJ = X(z)l•=~
A(1- 1) tz-
=
(1 + ~z- )(1 -
1
!z 1)1·=~
= 0
(f) H(z) has zeros at -.75 and 0, and poles at 2 and oo. Its ROC is lzl < 2.
88
I 2
-
-u'-:_,---,.._.,.,..---,,f-..,.:-:--,,:--.,.,,s,---:-'2
....
(g) Since the ROC of h[n] includes 0, H(z) contains no negative powers of z, which implies that
h[n] = 0 for n > 0. Therefore h[n] is anti-causal.
3.47. (a)
00
X(z) = L:;z[n]z-n
n=O
00
Therefore, X(oo) = z[O] # 0 and finite by assumption. Thus, X(z) has neither a pole nor a :zero
at z = oo.
(b) Suppose X(z) has finite numbers of poles and zeros in the finite z-plane. Then the most general
form for X(z) is ·
M
oo IT<•-c•)
X(z) = L:;z[n]z-n = KzL~•;:;:''----
n=O IT (z- d•)
Kzl
where K is a constant and M and N are finite positive integers and Lis a finite positive or negative
integer representing the net number of poles (L < 0) or zeros (L > 0) at z = 0. Clearly, since
X(oo) = z[O] # 0 and < oo we must have L + M = N; i.e., the total number of :zeros in the finite
z-plane must equal the total number of poles in the finite z-plane.
3.48.
X( ) = P(z)
z Q(z)
where P(z) and Q(z) are polynomials in z. Sequence is absolutely swnmable ~ROC contains 1•1 = 1
and roots of Q(z) inside 1•1 = 1.
These conditions do not necessarily imply that z[n] is causal. A shift of a causal sequence would only
=
add more zeros at z 0 to P(z ). For example, consider
z' 1
X(z) = --,
.. - .
1•1>-
2
z 1
= --~ =z· 1
z- i 1- iz-1
:[nJ = z:
00
1<=1
(-1)"+1
k a••rn - kNJ
3.50. (a)
(c)
3.52.
00
N-l N-1
tanB.(w) 2: :r(n]cos(wn) =- 2: :r(n]sin(nw}
N-l
tanB.(w}:t(OJ + 2: :r(n](tan B.(w} cos(nw) + sin(nw)) = 0
=•
N-1
1
tanB.(w,) + :r(O] ~ :r(n](tan9.(wt) cos(nwt) + sin(nw,)) =0
tane.G) + :r;od:t(lJ(tane.G)cos~+sin~)
+ :r(2] (tan e. G) cos .. + sin,.) J = o
Therefore
:r[n] = :r(0](6(n] + U[n- 1] + 36(n- 2]}
where :r(O] is undetermmed.
lim X(z)
~-oo
= .z-oo~
lim ":r(n]z-• = :r(O] + lim ":r[n]z-• = :r(O]
;~-oo ~
n.-o n.=l
. 91
a.ss. (&)
... 00
Therefore
1 1 -oz-1
c•• (z) = 1- az- 1 1- az = (1- az- 1)11- a 1z-1)
1.!.s ~ 1
= 1- az 1 - l - a-1z-1 lal < 1:; < Ia- I
(c) Cleuly, :r1[n] = :r[-n) will have the same &UtocorrelatioD function. For ex.ample,
1 1 1
X1(z)=--
1-oz
1•1 < ,..- 11=- c.,•• (z) = 1 -GZ 1 -G.Z -1 =c•• (z)
(d) Also, any del&yed w:rsioll of :r[n] will have the same autoconela.tioD function; e.c., :r:,[n] = :r[n- m]
implies
Xo(z)
.-..
= 1 _ oz- 1
z-,. z•
lol < 1•1 =- c...,(z) = 1 - ... - 1 1- QZ =c•• (•)
92
S.56. In order to be a z·transform, X(z) must be analytic in some aDDular recK>n afthe z-plane. To determine
if X(z) = z• is a~?a~rtic we examine the existence ol X'(z) by the Canchy Riemann conditions. If
In our cue,
X("'+ ill)=,.- ill
and thus,
8u 8u
a,
8,. = 1 ¥ =-1
unless"' and 11 are zero. Thus, X'(z) exists only at z = 0. X(z) is not aulytic anywhere. Therefore,
S.57. If X(z) has a pole at z = zo then A(z) can be expressed as a Taylor's series about z = zo.
=
A'(Zo) +
B(z)
~ A•(Zo) (z- Zo)•-1
I•=,. = B(Zo)
A'(zo)
~ n!
....2
!
93
Solutions - Chapter 4
Sampling of Continuous-Time Signals
9
4.1.
%[n] = Ze(nT)
= sin ( 21r(100)11 ~)
= sin {in)
Ze(t) = ms(o.,t).
Since w = nT aDd T = 1/1000 seconds, tbe signal &eque~>cy could be:
• .(I)
(a)
z,(t) =0,
, 101 2: 2.. · 5000
The Nyquist rate is 2 times the highest frequency. ,. T 10
=
.:,.. sec. This avoids all aliasing in
the C/D CODverter.
(b)
1
T = 10kHz
."'
8
= TO1
= 10,0000<
o, = 2.- · 625rad/sec
1
T = 20kHz
,.."' = TO1
8
= 20,000°'
o, = 2.- · 1250rad/sec
'· = 1250Hz
4.6. (a) The Fourier transform of the filter impulse respcmse
• H,(jO) = L hc(t)e-m• dt
= 1. 010
a-a.te-;nt dt
1
= a+ ;n
So, we take the magnitude
llfc(jO)I
1/a
0
97
(b) Sampling the filter impulse response in (a), the cliscrete-time fii,J ia described by
Hd(eiw)
..
hd(n] = Te-•"T u[n]
= L;Te-•"T.-jwn
.....,
T
= 1 - e-.Te jfll
Taking the magnitude of this response
. T
IH.(e'W)I = .
(1- 2e-aT cos(w) + e-2oT)i
Note that the frequency response of the cliscrete-time filter is periodic, with period 2...
!%<• jCD)I
(c) The minimum occurs at w = ... The corresponding value of the frequency response magnitude is
T
(1 + 2e-•T + e-2•T) i
T
= 1 + e aT·
11\i(e jro)l
T -----------------------
112
1/a 2Ja 3/a
4. 7. The continuous-time signal contains an attenuated replica of the original signal with a delay of.,.•.
Zc(t) =Sc(t) + OSc(t- 'rd)
(a) Tailing the Fourier transform of the analog signal:
Xc{iO) =Sc(iO) · (1 + ae-fr•n)
Note that Xc{jO) is zero for 101 > ..fT. Sampling the continuous-time signal yields the discret<
time sequence, z[n]. The Fourier transform of the sequence is
1 ~ jw .br
X{.;w) = T f... Sc(T+1T)
.,._..,
+ f f:
~-00
Sc(~ + ; 2; )e-j••(f+'f.<l.
98
(e) We need to take the inverse Fourier trauform of the discrete-time impulse respoDSe of pact (b).
= _sin_(_,.,._) + =ac::sin~[..~(n~-..:1!.1.))
'"' ..(n- 1)
= 6[n) + a6[n- 1]
(ii) For T• = T /2:
h[n] = .!.. r <eiW" + ~<"-t,l dul
211' }_.
= sin(1m) + asin[.. (n- !JJ
'"' .. (n-!)
= cS[n] + a sin[.. (n - ! )]
.. (n- ll
4.8. A plot of X,Url) appears below.
(a) For :,(1) to be reeovetable from :[n], the trauform of the discrete signal must have no aliasing.
When sampling, the radian frequency is rela&ed to the analog frequency by
w=OT.
No aliasing will occur if the sampling intenai satisfies the Nyquist Criterion. Tbus, for the band-
limited signal, :,(I), we should aeJect T as: ·
1
T::;2x10'"
99
(b) Assuming that the system is linear and time-iDvariaDt, tbe convolution IUIII describes the input-
output rebUionship.
11[n) = L
.. :r[k]hjn - k)
o--oo
..
11[n) = T "E z[l:)
= T
•=-ao..
L :r[l:)u[n - l:)
lo=-oo
= T·
n-ao
..
L
t-=-ao
z[k]
tz:-oo
X(~w) = L
.. :r[n]e-jwn
n.=-co
Hence,
= x.(illlln..o
For the 6Dal equality to be true, there mast be no contribution from tbe t.erms for which r i' 0.
That is, we require 110 aliasiD& at n = o. Since we are only illtereRed iD preserving the spectral
component at ll = 0, we may sample at a rate which is lowe:r than tbe Nyquist rate. Tbe maximum
'lalue ol T to satisfy u..e CIDilditiODS is
1
T~ 17loi'
100
s(n] = 2...1
2.. (b)
X(ei"')ei""'dw
= 2...1
2.. (b)
X(ei(w-•l)ei""'dw
= 2..1
2.. (b)
X(ei"'}ei(w+o)ndw
= ..!...,;••
2r (b)
1 X (ei"'}eiwndw
= (-1)"s[n].
4.13. (a)
z.(t) = sin(;t)
z.(t) = sin(;0 t)
llc(t) = sin( 1~ (t- 2.5))
= llll(-t-
10
-)
4
. . .
. rn .,
Jl[n] = SUI(---)
2 4
(c) Tbe samplillg period Tis not limited by the continuous time system h.(t).
4.14. The<e is no loss of iDformatioD if X(..,;.-1 2 ) and X(eilw/Z-•l) do not overlap. Tbis is true for (b), (d),
(e).
4.15. The output :r,[n] = :r[n] if no aliasing occurs as result of dOWDS&IIlplillg. That is, X(e'w) = 0 for
7r/3 ~ '"'' ~ ...
(a) r[n] = cos(rn/4}. X (eiw) has impulses at w = ± .. /4, so there is DO aliasi.Dg. r,[n] = r[n].
(b) r[n] = cos(.-n/2). X(eiw} has impulses at"'= ±tr/2, so there is aliasi.Dg. r,[n] # r[n].
(c) A sketch of X(eiw} is shown below. Clearly there will be no aliasillg and r,[n] = r[n].
X(ej<ll )
-1t/4 !t/4
r<· j(l))
This is UDique.
102
(b) One choice is
M 1t/2 2
T = 3-.:/4 = 3
However, this is not llllique. We caD also write i 4 [n] = eos('fn), 10 another choice is
(b) UpsampliDg by 3 and low-pass filtering :z:[n] = sin(3.-n/4) results ill sin(1rn/4). Downsampling by
=
5 gives us i4[n] sin(51tR/4) =-
sin(31tR/4).
4.18. For the condition to be satisfied, we have to ensure that Wo/L :5 uilil(7tfL,7t/M), so that thelowpass
filtering does not cut out part of the spectrum.
X (ei"')
(b) A straigbt-forward application of the Nyquist criterion would lead to an incorrect conclusion that
the sampliDg rate is at least twice the III&Jimum frequency of :z:,(l), or 202 • However, since the
spectrum is bandpass, we only need to ensure that the replicatioiiS ill &equency which occur as a
result of sampliDg do not overlap with the OficiDal. (See the following figure of X,UO).) Therefore,
we only need to ensure
n2 - -2r
T < n, = T<-
2ft
.0.0
i03
(c) The block diagram alODg with the frequerocy response of h(t) is sbawn here:
.
~~··
x(n] convert x(t)
sequenc::e
to!mpulse h(t)
llaiD
4.22. (a)
• lD
w = !lT,
~
-It
LiZ
(b) To recover simply filter out the undesired pam of X(eiw).
I,
It CD
(c)
• -21tff
I xrr
I rI fiT
I
21tff -n
T <21r
-
-n.
4.23. In tbe frequency domain, we have
104
s 0 (1) = 0, 101 ~ ;,
Therefore, sinCe we ue sampllng this s.(t) at the Nyquilt frequacy s(n) will be fuJI band and nnaljaoed.
s[n) = z.(nT,)
1/c(t) is a band-limi1ed interpolation of z(n) at a di!"erent period. Since DO aliasing occun at z(n), the
spectrum of llc(t) will be a frequency axis sc.aliDg of the spectrum of z.(t) for T, > T2 or T, < T2. As
we show ill the ~.
(a)
I
-lt x5 X tal
(b)
I
II X S X 1al O
(d)
(I)
-2 lt X 5 X tal 211 X 5 X ial O
21trr o
(I)
106
(b) Since Hd(ei"') is an ideallowpass filter with "'• =f, we don't care about any signal aliasing that
occurs in the region i !': "' !': 11. We require:
211 1l
- - 211 . 10000 !!:
T 4T
1 8
"f !!: -7 ·10000
7 1o-•
T !': ·-X MC
8
Abo, once all of the sipallies in the ranee iwl !! i• the filter will he inelfective, i.e., f :$ T(2•x1o").
So, T !!: 12.5,.sec.
(c)
Iff
4.26. First we show that X,(ei"') is just a sum of shifted versiollS of X(ei"'):
= (! ~ eil2de/M)) :[n)
X,(ei"') =
--oo 1
00 Jl-1
= L M L :[n]eil...,./Mle-;"'"
-=-oo 11-0
= !L M'-1
L
00
i:=-0 na-oo
:{n]e-J1.,-(2d/Mll•
1 Jl-1
= M ~X (eil... -tad/Jtll)
X.(ei"') = L X,(Mn]e-;...•
---00
L :,(l)e-i(w/Jl)l
00
=
,__
107
{ii) X,(ei"') and x.(ei'-) are sketched below for M =3, WH = "/4.
)
It (I)
(b) From the definitiou or X,(ei"'), we see that there will be no aliasing if the signal is handlimited to
1r I M. In this problem, M = 3. Thus the maximum value or "'H is 1r /3.
When we upsample, the added samples are zeros, so the apsampled signal z.[n] has the same energy as
the origiDal z[n):
..L 2
)z[n)l =
...
L J:r.[n)l 2
,
-=-oo --oc
and by Paneval's theorem:
4.28. (a) Yes, the sywtem is liDea: because ea.cb of the subblocks is JiDear. The C/D step is defined by
z(n] = z,(nT), whidl is clearly liDear. The DT system is an LTI I)'SUm. The D/C atep coBSistS
of coavertiDg the .equeace to impttlsa and of CT LTI <eriag, both of which are li.Dear.
(b) No, the system is aot time-iD..naat.
For eumple, suppooe that h[n] = 6[n], T = 5 and z,(t) = 1 for -1 !> t ~ 1. Such a I}'SUm would
=
rault iD z[n] 6[n] &Dd 11c(t) = liDc(r/5). Now suppose we delay the iDput to be z.(t - 2). Now
z[n] = 0 &Dd lf,(t) = 0.
4.H. We can aaalyze the system iD the frequency domaiD:
4.30.
X,(jO) = 0 101 ~ 4000lr
Y(jO) = IOIX.(jO), 1000.. ~ 101 !> 2000lr
SiDce only half the frequency band of X,(jO) is needed, we can alias everythiDg past 0 = 20001r. Hence,
T =1/3000 s.
Now that T is set, figure out H (ei") band edges.
B(ei") = '···I
0.- i• -< ,_
'···I < '•
- T
{ O~lwl < f,1f <lwiSr
4.31.
x.uo> = o, 101> f
.
. 1
llr(t) =
1 -~
z,(r)dr = H.(jO) = 1"0
In discrete-time, we want
109
IH(e jm )I
1t 2Km
arg(H(e j(l) ))
(I)
-2x
-lt
i ··1-------.. _____.......:_
Mlr-~--~~--~--r-~--~~--~-,
-----~
r;
r:
100
(b)
_..,
0
-· -u -u
~
1\..
__ -...cy,..,.._
~·~ 0
\
G.2
'" .. .
t)0.4
.
ll(n] = r(n)
Y(eiw) = X(~) • X(eiw)
therefore, Y(~) will occupy twice the frequency bud thai X(~) does if DO a!jasjng Occurs.
U Y(ei"') "I 0, -• < w < •, tha X(~) "I 0, -j < w < j ud so X(jSl) = 0, 101 ~ 2•(1000).
Since w = _nT,
,.
2 ~ T · 2•(1000)
1
T :5 4000
110
4.S4. (a) Since there is DO aliasiDg involved in this process,"" may choose T to be uy value. Choose T = 1
for simplicitf· X.(jll) =
0, !Ill 2: "fT. Since Y.(jll) =
H.(jll)X.(jll), Y.,(jll) = 0, !Ill 2: 1t fT.
Therefore, there trill be DO aliasing problems in going &om 1/e(t) to 11[n].
=
Recall the relat.iODShip w OT. We caD simply use this in our system CDDversioD:
H(ei"') = .-,.,1•
H(jll) = .-;art•
; e-;n;z, T=l
Note that the choice ofT ud therefore H(jll) is Dot unique.
{b)
Since H(e'"') is defiDed for 0 5 lwl 5 " we must evaluate the frequeDcy at the basebaDd, i.e.,
=
S1r /2 => 51f /2 - 2" "/2. Therefore,
= cos (5Jt n- ~) .
2 2 .
y[n)
0 n
·I ) (
1
H(ei"') = (10jw) 2 + 4{10jw) +3
;
1
-100w' + 3 + 40jw
!
111
(b) The downsamp1er bas M =2. Since o:[n) is bandlimited to iJ, there will be no aliasing. The
frequency axis simply expa~~ds by a fador of 2.
*
For llc(t) = .:.(t) Y.(in) = x.(in).
Therefore nr r
~ 2". lOOT' ~ = •·
4.37. h both syatems, the speech was fikered first so that the subsequent sampling result.s ill no aliasing.
Therefore, going •[n) to •l[n) basically requires rb•ngi~~& the sampling rate by & fador of 31dizf5kllz =
3/5. This is cloae .mh the following syatem:
I-_,.,..;. cutoff
Igain"3
I
LPFI --.o..; ~ S
Digital=1tl31-·
.
•
z.(t) is sampled at sampling period T, so there is no aliasing ill z[n).
.AA'A. -Jt
rr
• (I)
The
~. -wi L wJL · 1D
filter H(ei'-') removes frequency components between"/ L and"·
-~·;·)
.I\ -.1 L
&T wJL
1\ ct
The multiplication by ( -1 )• shifts the - of the frequency band from 0 to .-.
The D/C conversion maps the raDge - .. to .. to the ranp - ..IT to ..fT.
•U9. (a)
h[n] = 0, lnl > (RL -1)
Therefore, for causal system delay by RL - 1 samples.
(b) GeDeral interpolator condition:
11(0] = 1
h{kL] = 0, k =%1, :!:2, ...
(c)
(RL->l AL-1
y(n] = L h[k]v[n- II:]= h(O]v(n] + L h[n](v[n- k] + v[n + kj)
t=-(RL-1) ksl
If n = mL (m an integer), then .., don't bave any multiplications since h(O] = 1 and the other
non-zero samples of v(A:] hit at tbe zeroo h(n]. Otherwise the impulse response spans 2RL - 1
samples of v{n], but only 2R of tb- are DOn-zero. Therefore, there are 2R multiplies.
4.40. Split H(e!w} into a lowpas$ and a delay.
x, (t)
C/D
x[n)
I
tL - H...,<~"' )
w[n)
I
e
-j., v[n)
tL ~t)
V\f\IJ\N
-Slt/4 -31114 -1114 1114 31114 51114 "'
4.42. (a) The Nyquist criteriOD states that :<0 (1) can be ncooaed as long as
In this case, T = 1/500, so the Nyquist cri1eriOD is satisfied, and :<0 (1) can be recovered.
(b) Yes. A delay ill t.ime does DOt c:bange the bandwidth ol the sipal. Hence, y.(t) has the sam
bandwidth and same Nyqw.t sampling rate as :r.(t).
(c) Consider first the follcrwing a::pressiolll far X(ei"') and Y(ei"'):
Hence, we let
B(ei") = { 2eo, -;.. • Jwl < j
otherwise
Then, iD the following figure,
""' < j
otherwise
x[n] y[n)
H
For the given T = 1/800, there is no aliasing &om the C/D conversion. Hence, the equivalent CT
transfer function H.(jO) can be written as
' ,
,' ' ' , '
,' ''
'' '' I I ' '
' 'I
I "" ''
115
X(e i'l)
(I)
(b) For this to be true, H(ei"') Deeds to filter out X(ei"') for •/3 :5 lwl :5 r. Hence let wo = r/3.
Furthermore, we WUlt
•/2
T =2..-(1000) =- T2 = 1/6000
2
(c) MatchiDg the followiDgligure of S(ei"') with the figure for R.,(jO), and remembering that 0 = wfT,
we get T3 = =
(2r/3)/(2000r) 1/3000.
S(e.i<O)
(I)
Further,
X(ei"') __..
= l:.. z(n].
=
Therefore, - pick h(n] Tu(n], whiCh males the system ""accumulAtor. Our estilllate A is the output
11[n] at n = 10/{10-<) =
10", whea all rJ the DOD·zoro Ample& of z(n] haft heeD added-up. This is
"" -a estimate P""" out aasumptiaa ai both haDe!- IUid time-limitedness. Sillce the aasumptiou caD
nner be exactly satisfied, however, this method only P"""""
approximate estimate for a.ctualsigllals.
The overall system is as follows:
f
T = 1/10000
4.46. (a) Notice that
vo[n] = :r[3n]
111 [n] = :r(3n + 1]
112[n] = z[3n + 2],
and therefore,
vo[n/3], n = 3k
z[n] ={ Wl[(n -1)/3], n = 3k + 1
112[(n- 2)/3], n = 3k + 2
(b) Yes. Since the halldwidth of the filters are 2• /3, there is no aliasing introduced by downsampling.
H;nce to recoDStruct z[n], - need the system shown in the following figure:
Yo[n]•! f3 •! Edz>l l
YJ[n].,, f 3 1----+l•l E t<z>~~l
hi \he following discussion, It\ "'• [n] deno\e the even samples of :z:[n], and z.(n] denote the odd
samples ot :z:(n]:
= { 0,:z:(n], n eftll
n odd
~inJ,
n even
= { nodd
{ ll<(n/2], n even
v,[n] = 0, n odd
{ tu4 (n], n even
= 0, n odd
{ (:r: • h.)[n], neven
= 0, nodd
= :z:0 (n] • h.[n]
where the last equali~ follows from the fact that h.[n] is non-zero only in the odd samples.
Now, s[n] = v,[n]•h.[n] = :z:0 (n]•h.(n]•h.[n] = .rrlnl. aDd since :[n] = :.[n]+z.[n], •[n]+v3[n] =
:[n].
4.47. Sampling random processes
(a)
P..(w) =T1 ~
L.J P.... ("'
T+T2rk)
K=-oo
(c) If
P•••• = 0, for lwl 2: r
then
P.. (w) = ~P•••• {f), lwl $r
118
.(.48. {a)
2Tr)
Therefore, we require that " ~
1 ... P.... ("'
P..(w) = T
no.
T+T
-- L
(e) For the JpeCtrum of F"JC P3.&-2 it is dear \hat if T = II; then the discrete-time power spectrum
will be wbi1e, as shown ill the figure above.
pxx ( m)
I I I
I
I '
' I
I
'' I '' I '' I ''
I
I
') I
'v I
'<
I
'<
I
I
''
I
I
I '
'' I
I '
'' I
I '
I '' '
I I
I
I I '
'
I
' ''
-411: -211: 211: It (I)
(d) For white discrete-time signal:> 4>,.[m) = 0, m # 0 but 4>•• [m] 4>•••• (mT). Therefore, auy =
analog signal whose autocorrelation func:t.iou bas zeros equally spaced at illtenals of T will yield a
white discrete-time leqllellce is sampled with sampling period T. For example, for Fig P3.&-l:
sill CloT sill ClomT
4>.,.,(.,.) =--:;;-- * 4>•• (m) = rmT
S)'Sial>l: SySiem2:
-~
- . /K.
;:vi\%\.
-
•
m~ •
-&tm m 'b
119
Jls(t) = !12(1): Convolution is a linear process. Aliasing is a linear process. Periodic convolution is
equivalent to convolution followed by aliasing.
lls(t) 'I r(t): System 2 at S&ep 1 shows x;uo).
This is clearly not YsUO). YsUO) is an aliased
version of x.UO)
(b) Now,
:(t) = A cos(30lrt)
3 1
:3(1) = 4Acos(30.-t) + Acos(3. 30.-t),
4
v[n] = ~A cos (~.-n) +~A cos (~.-n)
v[n] = :3[n]
Jl[n] = z(n]
We can see here that sometimes aliasing won't be destructive. When aliased sections do not overlap
they can be reconstruc:1ed.
(d) This is the inverse to part (c). Since multiplication in time corresponds to convolution in frequency,
a signalz2 (t) has at most two times the bandwidth of :(t). Therefore, : 112 will have at least the l
bandwidth of z(t). If we run our signal through a box that will raise it to the 1/M power, then
the sampling rate can be decreased by a factor of M.
4.50. (a)
H (ei"')
.,.
= sin(wL/2)
sin(w/2) e
-i!L-1>-o/2
(b) The impuloe response h,.,. (n] corresponds to the convolution of two rectangular sequences, as shown
below.
• -L
A' L •n
= IlL·
•
- ..I.:l
2
II
..I.:l
2
I
. *.
n
-
..I.:l
2
II
..I.:l
2
I
•n
120
2
H (ei"') _ .!_ (siD(wL/2))
•- - L sill(w/2)
(c) The frequency respoue of -...order-hold is ~latter ill the regiDD (-ff/L,ff/LJ, but achieves less
out-<lf-baDd attenuatioll.
4.51.
Tbe baDd...;dth of+ •• (.,;"') i.s no larger than the baDd.,.;dth of X(ei"'). Therefore, the outputs of the
sys\ems will be the same if H2(ei"') is an ideallowpau 6her ...;th a c:utolf of wf L.
4.52. The idea here i.s to uploit the fact that every other sample supplied to h(n] ill Fig 3.27-1 is zero. That
__L....,
is,
w 1 (n]
= { ht(n/2] • z(n/2], n ewn
0, n odd
= { h 1 (0]z(n/2] + h,[1]%{(n/2) -1] + h 1 (2j%((n/2)- 2], newn
0, n odd
The downsampler expands the &equocy axis. Since Ro(eiw) is bandlimited to if, no aliasing
occurs.
~ It
.~ -It It (I)
(c)
Yo(.,;w) == ~Ho(.,;w) ( X(.,;w)Ho(.,;w) + X(eA-"~.,;r..-+"'1})
Y,(.,;w) == ~H,(.,;w) (.x(.,;w)H,(.,;w) +X(~)B.~))
Y(.,;w) == Yo(.,;w)- Y,(.,;w)
The ali Mill& terms always c:aDcei. Y(ei"') Is propor&ioDaJ • X(«>") if il4(e"'}- JP.(.,...)\ is a
COI!St&llt.
,i-1
w[n]
2(a]
The system is linear, time-'f3Q'ill& (due to ciowDsaatpling), 11M-+ d (._ t.ol(&+lD, ..I ~~&Me.
(b)
.,
T == -ON == 21<
., ......
X 5000 == lv ·-.
123
To avoid aliasiJI& in llc(l):
L=22,
AhrLPF
LB/2T
After cosine
modula!ion
.. C\ 1C\ ..
-11 ~ (1)2 II Ill
LBI2T
.. fl 1 f\ ..
AhrHPF
•
):M"' .
• !:!!!J!:!!!I
T T
(d) To ge11eralize forM cbamlels,- would...., the same modalalors, but -would choooe a larger
value of L t.o make room for additional spectra above the lower frequency bound. If the 1 -
124
w[n]
4.56. Sillce we wan' W(ei"') kl eqaa1 X(eiw), ~en H(ei"') mun compeDSlW! for ~e drop otrs ill H.. Ufl).
~(ej~
10
1\ )
•
'
OPT 1t Q)
4.57. (a)
E(e) =
I ep(e)cie =...•1 1"''
-1>./2
2
ede = ~L"'
e l>./2
2
12 =0
e'
~ = E(e2 - 0) = 1:;.1 1"''2
-1>/2
2
e de=
3
1>/2
1:;. L~>.; 2
1:;. 2
=12
E(e[m])E(e[n]), m ¥- n
r[m, n) = E(e[m)e[n]) = { E(e2 [n]l, m =n
t;.2
r[n,m) =r[n- m) =l26[n- m)
(b)
SNR = ..-!
C7~
=1~2
2..-!
(c) Let e,[n) be~. ou'put noise.
e,[n) = L h[k)e[n - k)
•
The variance oC :[n) is .,..;ghted similarly so ~e SNR does no' change. SNRout =12~.•
125
(e)
Again,~ ariaDce ol :[n] is weighted by the same factor, so the SNR does not change.
12
SNRout = 32·
4.58. Fim, notice that since 1/e(t) = : 1 (t):,(t), Ye(iO) = /,;(X 1 (j0) • X,(jO)), and so Ye(jO) = 0 for
JOJ;::: 1l1r/2 x 10'. Hence the Nyquist rate T = 1/55000s.
Choose System A and B such that w1[n) = a:r 1 (nT) ud w,[n] = hzz(nT).
For System A, we need to resample such that
M T 2 X 10-· 10
T = T, = 1/55000 = li
x, [n) ·I I ·I I ·I
I m n ! I [n)
(b) No aJiasiDg ocaus ill the region l!ll :$; n, darillc sampliDc:
21r
T -n, = n, = n, = 2r(4·44l -44r = 308r
4.61. (a)
ee
~--~~~~----·
-1t K Cll
127
2 1•'"'
1 J'!(2w/2)• ..,
"' ,.. w/11
~"'"I·'"'
= 2.. 5 -•1"'
= ~"·
sM•-·-
(e) X,(;O) must be suflic:iently haDdlimited that X(ei") =
X,(;OT) is zero for lwl > tr/M. Hence
x.unl o = for 101 > "/MT.
Assuming that is satisfied, v.[n] = o:[Mn- 1] = :,(MTn- T).
Downsampling does Dot change the variance of the noise, and hence a! =a!·
P.. (ei") = P11 (ei" 1"')
= 1oo~ sin4 (wf2M)
I It Ill
P,.(w) = P,(w)H,,(ei")H.,e-;w
= ,.:(1- .-;..)(1- ei")
= ,.:(2- 2cos(w))
(h) (i) o:[n] contributes cmly to 111 [n], but DOt Wz[n]. Therefore
ll~a[n] = o:[n - 1]
r,[n] = o:[n- 2}
128
(ii) In pan(a), the cillference equation desaibing the sigma-delta noise-shaper is
!
129
Solutions - Chapter 5
Transform Analysis of Linear Time-Invariant Systems
131
5.1.
_{ 1, 0 $ n $ 10,
11[n 1- 0, otherwise
Therefore,
Y(ei"') = .-J.., ~sm•t"'
This Y(ei"') is full band. Therefore, since Y(~") = X(~")H(~"), the Oll!y possible z}n) ud "'" that
could produce 11(n] is z[n] = lf[n] &Dd "'• = 11.
5.2. We haft 11[n- 1] -Jflf[n] + 11[n + 1] =z[n] or z- Y(z)- !IY(z) + zY(z) =X(z). So,
1
1
H(z) =
z-1 -If+ z
z
= <•- !H•- 3)
_! !
= __J_+_L.
•-! z-3
(a)
1 zeroatz=-
Re 3
• (b)
5.3.
1
11[n- 1] + 311[n- 2] = z[n]
Y(z) 1
H(z) = X(z) =
z-1 + l·-·
%
H(z) =
1 + !z-1
132
= -(-D(-~fu[-n-2]
= 1( 1).
3 - 3 u[-n- 2] ~ aaswer (d)
5.4. (a)
1 z 1
X(z)= 1-lz-1- z-2' 2<l•l< 2
2
lm
(b)
1 2z- 1 3
H(z) = 1- !z-1 - 1- !z-1' 1•1 > 4
• •
(3)n u(n]- 2 (3)•-
1
h(n] = 4 4 u(n- 1)
(c)
I z- 1 I z- 1 I
H(z) = 1 - + - 1 1 +I, 1•1 > 3
1 - 3 z 1 1- iz-1 1 - ~z-1 1 - 4 z-
So,
%[n] = -3I(I)"
2 u[nJ- 34(2)"u[-n- I]
X(z) = I--f!z 1
+ 1 = (I- fz-1)(1-
I- 2z-1
I
2z- 1)'
1
2 < fzl < 2
(b)
1 _7:7.,-:-2-::-~
Y(z) - -;-:--,-"-
. - (1- !z-1)(1- 2z-1)
This has tbe same poles as tbe input, then!Core tbe ROC is still ! < Jzl < 2.
(c)
Y(z) _2
H(z) = X(z) = 1- z .,. h[n] = 6[n]- cf(n- 2]
134
5. 7. (a)
Re
(b)
7
1-z-1 -1 3
H(z) = (1- !z-1)(1 + tz-1) = (1- 1) + (1 + iz-1)' !• lzl > i
H z _ Y(z) _ 1 - z- 1
( ) - X(z) - 1 + !z-1- lz •
• • •
Y{z) + ~z- 1 Y(z)- ~z- 2 Y(z) = X(z)- z- 1X(z)
1 3
u[n] + -y(n- 1]- -u(n- 2] = z[n] - :z(n- 1]
4 8
5.8. (a)
3
y(n] = 2y[n- 1] + y[n- 2) + z(n- 1]
3
Y(z) = 2z- 1Y(z) + z-2 Y(z) + z- 1X(z)
Therefore,
Y(z) z- 1 z- 1
H(z)--- - lzl > 2
- X(z)- 1- Jz 1 - z- 2 - (1- 2z- 1)(1 + ~z-1)'
lm
1 zeroalz•-
Re 2
!
135
(b)
1•1 > 2
(c) Use ROC of! < 1•1 < 2 lillce the ROC m- i.Dclude lzl = 1 for a stable system.
h[n] =- 52(2)"u[-n- 1]- 52 ( -21)" u[n]
5.9.
5
y[n- 1]- 211[n] + y[n + 1] = z[n]
Y(z)
H(z) = X(z)
=
.-•
1- ~z-1
2
+ z-2
.-•
= (1- 2z-1)(1- !z 1)
a
= 1- 2z-•
1 zeroatz~-
2
h[n] = --(2)"u[-n- 1]- - -
3 3 2
2(1)" u[n]
!Deludes 1•1 = 1, so this is stable.
136
(c) I• I> 2:
h(n) 2
= l(2)"u(n)- J2(1)"
2 u[n]
ROC outside of largest pole, so this is c:aasal.
5.10. Figure P5.16 shows two zeros and three poles inside the unit circle. SiDc:e the number of poles must
equal the number of zeros, there must be an additioul zero at z = oo.
H (z) is c:aasal, 10 the ROC lies outside the largest pole and includes tbe unit circle. Therefore, the
system is also stable.
The inverse sywtem switches poles and zeros. The inverse system could have a ROC that includes 1•1 1, =
=
m.akiDg it stable. However, the zero at z oo of H(z) is a pole for H;(z), so the system H;(z) c:aDDot
be causal
5.11. (a) It connot be determined. The ROC micht or might not include the unit circle.
(b) It connot be determined. The ROC micht or might not include z = oo.
(c) FaJ.e. Given that the system is causal, we know that the ROC must be outside the outermost pole.
Since the outermost pole is outside the unit circle, th~ ROC will not include the unit circle, and
thus the system is not stable.
(d) True. H the system is stable, the ROC must include the unit circle. Because there are poles both
inside and outside the unit circle, any ROC including the unit circle must be a ring. A ring-shaped
ROC means that we have a two-sided system.
5.12. (a) Yes. The poles z = :!:j(0.9) are inside the unit circle so the system is stable.
(b) First, factor H(z) into two parts. The first should be minimum phase aad therefore have all its
poles and zeros inside the unit circle. The second part should contain the rtm•i•ing poles and
zeros.
1 +0.2z- 1 1- 9z-•
H(z) = 1 + O.Slz 2 I
minimum phase poles & zeros
outside Wlil circle
All pass systems have poles and zeros that occur in conjugate reciprocal pairs. H we include the
factor (1 - iz- 2 ) in both parts of the equation above the first part will remain minimum phase
and the second will become allpass.
H(z)
= (1 + 0.2z- 1 )(1- tz- 2 ) . 1- 9z- 2
I+ 0.81z-2 -lz-2
1
= H 1 (z)H,.(z)
5.13. An a.ride: Technically, this problem is not well defined, since a pole/zero plot does not uniquely
determine a system. That is, ma.ny system functillns caa have the same pole/zero plot. For example,
consider the systems
H.(z) = .-•
Hz(z) = 2z- 1
Both of these systems have tbe same pole/zero plot, namely a pole at zero and a zero at infinity.
Clearly, the system H 1 (z) is allpass, as it passes all frequencies with unity gain (it is simply a unit
delay). However, one could ask whether H 2 (z) is allpass. Looking at the standard definition of an
137
allpaas system provided iD this chapter, the amwer would be DO, Iince the system does not pass all
&eque~~cies with •nitll gain.
A broader de&DitioD of &D allpaas system would be a system for wbic:b the system magDitude respoue
jH(ei"')j =o, where a is a real coDSt&Dt. Such a system would pass all &equccies, ud scale the output
by a CODSt&Dt factor a. Ill a practical setting, this deliDitiOD of &D allpass system is satisfactDry. Uncle.
this defiDitiOD, both systems H 1(z) &Dd H 2(z) woaJd be CODSidered allpaas.
For this problem, it is assumed that DODe of the poles or seros sbowD iD the pole/sero plots are scaled,
so this issue of 'UiDg the proper defiDitiOD of u allpass system does Dot apply. The st&Ddard defiDition
of &D allpaas system is Died.
Solution:
5.14. (a) By the symmetry of z 1[n] we kaow it bas liDear phase. The symmetry is arouad n = 5 so the
continuous phase of X 1 (e''") is arg[X 1 (~w)] = -5.1. Thus,
. d . } d
grd[X,(e'w)] = - dw {arg[XI(e'w)] = - dw { -5.1} = 5
3
2
34.56789 n
(b) By the symmetry of z2[n] we kaow it bas IIDear phase. The symmetry is arouad n = 1/2 so we
kaow the phase of X 2 (~w) is arg[X2 (~] -w/2. Thus, =
.
grd[X,(e'w)] d { arg[X,(e'w)]
=- dw . } =- dw
d { -2
"'} = 21
312
314
••• T •••
-2 -1 0 1 2 3 n
5.15. (a) h[n] is symmetric about n =1.
H(~) = 2 + e-,;.. + 2e-2jw
= e-"'c~ + 1 + 2e-,;..)
= (1 + 4cosw)e-iw
138
=
A(w) 1 + 4a~~w, a= 1, IJ 0 =
Generalized Linear phase but Dot Linear Phase Iince A(w) is Dot always positive.
(b) This sequence bas no even or odd symmetry, 10 it does not possess &eneralised linear phase.
(c) h(n] is symmetric about n = 1.
H(ei'-) = 1 + :se-Jw + e-2Jw
= .-Jw(el" + 3 + .-;...)
= (3 + 2 coow)e-Jw
A(w) = 3 + 2 <XIIw, a= 1, IJ = 0
Generalized LiDeu phase k LiDear Phase.
(d) h(n] has even symmetry.
H(eJw) = 1 + .-;w
= .-j(1/2lw(eJ(l/2)w + .-j(l/2)w)
= 2coo(w/2)e-i(l/2lw
1
A(w) = 2 coo(w/2), a= 2' 1J = 0
Generalized Linear Phase but Dot Linear Phase Iince A(w) is Dot always positive.
(e) h(n] has odd symmetry.
H(ei"') = 1 - .-2jw
= .-i"'(ei"' _ .-iw)
= e-jw2j sin""
= (2siDw)e-iw+if
A(w) = 21iDw, a= 1, IJ = i
Generalized Linear Phase but Dot Linear Phase since A(w) is not always positive.
5.16. The causality of the syslem eaDilOt be determined &om the figure. A causal sys1em h 1 (n] that has a
linear phase respoD.Se LH(ei•) = -atD, is:
h 1 (n] = .S(n] + 26(n- 1] + 6(n- 2]
H 1 (ei"') = 1 + 2e-Jw + .-;:....
= 0 -'"'(eJw + 2 + 0 -iw)
= e-i"'(2 + 2cos(w))
IH,(ei"'ll = 2 + 2 cos(w)
LH,(ei"') = -w
AD example of a non-ausal system with the same phase response is:
h2[n] = 6[n + 1] + 6[n] + 46[n - 1] + 6(n - 2] + 6[n - 3]
H2(ei"') = eJw + 1 + ..-;... + .-;:.... + .-;s..
.-;..(~
= +eJw +4+.-;.. +.-;... )
= .-;..(4 +2coo(w) + 2coo(2w))
=
-
IH2(eJw)l 4+ 2 CXII(w) + 2 coo(2w)
LH2(ei"') =
Thus, both the causal sequence h,[n] ud the DOD-causal sequence h2(n] have a liDeM phase respoD.Se
LH(e'"') = -aw, where a= I.
139
5.17. A minimum phase system Is ooe which has all its poles and zeros illside the unit circle. It is causal,
stable, and has a causal and stable iD-.e.
(a) H 1 (z) has a a«o outside the llllit circle u z =2 10 it il aot m.iDimum phase.
(b) B2(z) is miDimum phase Iince ita poles and aro. are illside the unit circk.
(c) Bs(z) is miDimum phase Iince ita poles and_,. are iJiside the llllit circle.
(d) H.(z) has a a«o outside the unit cirde u z =
cc 10 it Ia DOl miDimum phase. Moreover, the
iDftr.se system would 110& be causal due to the pole u iDbity.
5.18. A minimum phase system with aa eqaivaleut magaitude spectrum ean be found by aaalyzing tbe system
fullcti011, and nllectia,g all poles are aro. that are outside the llllit circle to their c:oujup.ce reciprocal
locat.i0111. Tbis will lDOft tbem illlide the unit c:lrcle. TheD, all poles and seras for H,..,.(z) will be
inside the llllit circle. Note that a ICAle factor may be illtroduced wh• the pole or zero is rededed
inside tbe IIDit cirde.
(a) Simply rellec:t the a«o u z = 2to ita oaajup.ce reciprocalloc:atiOil u z = !· Tben, detenniDe the
scale factor.
. ( ) _ (1 + tz- 1) {1 -jz-1)
1l,._z -3 ( )
z-1 1 + JZ I
1
(1- tz-1)
B,.;.(z) = 3
z- 1
Note that the term zh
does Dot alfect the frequency respoase magaitude of tbe system. Con-
sequently, it can be remooed. Thus, the ,.......aining term has a sero inside the unit circle, and is
therefore minimum phase. Aa a result, - are left with the system
9 (I-lz- 1 ) {I-lz-1)
B,..,.(z) =- - J
4 (1- !z-1)
5.19. Due to the symmetry of the impulae ~. all the ays1em2 have generalized linear phase of
argfH(e"")] = {J-....., where n. il the point or symmetry iD the impulse respoase graphs. The group
delay is
To find each system's group delay - need only find the point or symmetry n. in each system's impulse
response-
140
5.20. (a) Ye.s. The system fwlc:tioll could be a generalized liDear phase system implemenled by a linear
const&Dkoellicient cWferential eq1Wion {LCCDE) with real coefficients. The seros come in a
set cl four: a zero, its conjugate, aDd the two CODjupte reciprocals. The pole-zero plot conld
correspond to a Type I FIR liDear phase system.
{b) No. This system fwlctioll could- be a gaes-alized liDear phase system implemaled by a LCCDE
with real coefficients. Since the LCCDE has real coefficient~, its poleo and seros must come in
conjugate pairs. Howe"l'er, the seros in this pole-zero plot do not have corresponding conjugate
seros.
{c) Yu. The system function could be a generalized liDear phase system implemenled by a LCCDE
· with real coefficients. The pole-zero plot could correspond to a Type D FIR liDear phue system.
1
.... . ..
-II -1114 0 7fl4 II (I)
(b) z(n] is first modulaled by "• 1owpass <ered, and demodulated by .-. Therefore, H,,(ei'"') filters
the high &equacy components cl X(ei").
This is a highpass &Iter.
1
••• •••
-11 -311/4 0 II Ql
(c) h1,(2n] is a clownsampled version of the &Iter. Therefore, the &equacy niiJIOIIII! will be "spread
out" by a factor of two, with a gain of f.
This is a lowpass &Iter.
!
141
••• 112 • ••
-II 0
(d) This system apsamples ~..In] by a factor ol two. Therefore, the frequency am will he compressed
by a factor of two.
This is a haDcbtop lilter.
•••
- 1
r--
••
(e) This system upsamples the input before passing it through h,,[nj. This effectively doubles the
frequency bandwidth ol e,,(.,;w).
This is a lowpass filter.
••• 112 • ••
5.22.
1- a- 1 z- 1 Y(z)
H(z) = 1 _ az-l = X(z), causal, so ROC is [z[ >a
Re 2
142
(d)
1 o-1 z-l
B(z) = 1-"" 1 1- u-1' lzl >"
(e)
..
h{n) = (cs)•u[n)- !(cs)•- 1u(n- 1)
. 1- ..-1.-jw
B(e"") = B(z)l- = 1 ...-.;..
= 1 + :!f - 1 COIW)
" cos"' i
IB(ei'-)1 ( 1 + tJ"2 - 2a
2
= ! (" + 1- 2cscosw) l
.. 1 + ... - 2cs cosw
1
= .
5.23. (a) Type I:
Jl/2
Type II:
-
A(w) ='E cs(n]coswn
(JI+l)/2 1
A(w) = L
-1
b(n)cosw (n- 2)
cooO = 1, coo (n.. - j) = 0. So H(ei•) = 0.
Type ill:
-
Jl/2
A(w) = L c(n) siDwn
sinO= 0, sinn.- = 0, so B(ei") = B(ei•) = 0.
Type IV:
(JI+l)/2 1
A(w) = L
-=1
d(n)sillw ( n- 2)
sinO= 0, sill (n.. - j) ~ 0, so just B(ei") = 0.
Since the poles and zeros {2 poles at z = ±1/2, 2 zeros at z = ±2} occur ill conjugate reciprocal
pairs the system is allpus. This property is easy to """""iu since, as ill the system above,
the c:oellicients ol the n111Df!rator aad denomilla&or z..polynomials get reversed (and ill general
conjugated).
(b) It is a property ol aiJpus systems that the output energy is equal to the illput energy. Here i.s the
proof.
..
-
N-l
L l11[n]l 2
--
= :E
..
= ..!.. [
l11[n]l 2
2
IY(eiw)l dw (by Parseoa.l's Theorem)
2• -·
= 2~ £: IH(,,w)X(eiw)l dw
2
= 2~ L: 2
!IH(eiw)l =I since h[n] is allpass)
..
IX(,,wJI' dw
= L
"-'= -oo
[z[n]l 2 (by Parseval's theorem)
-
N-l
= L [z[n]l 2
= 5
5.25. The statement is false. A non-causal system C&ll indeed have a positive collStallt group delay. For
example, collSider the non-ausal system
lm
1 zeroatz=-
Re
144
la(n) = nr"u[n)
lm
1 zeroatz=-
Re
5. 27. Making use of some DTFT properties can aide in the solution of this problem. First, note that
/a2 (n) = (-1)"h1 [n)
/a2 (n) = e-i•"h.[n]
Using the DTFT property that states that modulation in tbe time domain corresponds to a shift in the
&equeocy domain,
B2(~) = H,(ei(w+•lj
Consequently, B2 (~) is simply B 1 (ei"') shifted by ... Tbe ideal low pass &Iter has now become the
ideal high pass &Iter, as shown below.
-rd2 0 It Q)
5.28. (a)
A 1
H(z)- (1-}z-1)(1 + lz-1)' 1•1 > 2 h(n] causal
H(1) =6~A =4
(b)
4 1
H(z) = (1-lz-1)(1 + lz-1)' 1•1> 2
= <lf> + <I>
1 -lz-1 1 + lz-1
h(n] = -12cr
-
15 2
u(n] +- --
5
1r u(n]
3
8(
(c) (i)
1 1- lz- 1
:r:(n] = u(n]- 2u(n -1] # X(z) =
1
_ 2. - 1 , 1•1 >I
Y(z) = X(z)H(z)
1-2lz-1
_ ____ 4
= , -· . _:- ~·-')(l+!z- 1 )'
1•1 >I
4
=
(1- z- 1)(1 + iz- 1 )
3 1
= 1 - z- 1 + 1 + lz-1
3
(ii)
5.29.
2I
H(z) = (I- !z-1 )(I- 2z-1 )(I- 4z 1)
I 28 46
= I - lz 1 - I - 2.- 1 + I - u-1
2
SiDce we bow the leq1leiiCe is not stable, the ROC must not iDdude lzl = 1, and since it is two-sided,
the ROC must be a riDe- Tllia ._,. Clllly 011e poaible choice: the ROC is 2 < lzl < 4.
(a)
(b)
1 28
H 1(z) = 1- !z-1 - 1- 2z-1
48
H2(z) = _4z 1
1
5.30. (a)
M
M-2 M-1 n
-1/4
(b)
1
111[n] = z[n - (M - 2)]- 4zln - M]
1
v[n] = w[2n] = z[2n- (M- 2)]- 4z[2n- M]
z-•
H(z) = ( _ !z- 1 )( 1 _ Jz- 1 ), stable, so the ROC is l < lzl < 3
1
!
147
1
:[n] = u[n] ~ X(z) =
1
_ z- 1 , 1•1 > 1
1 1 1
Y(z) = X(z)H(z) = 1 _ !z-1 + 1 _ 3.- 1 - 1
_ z- 1, 1 < 1•1 <3
1
H,(z) = H(z) =z - 72• + 32'
2
ROC: entire z..plane
7 3
h;(n] = o[n + 2)- 2o[n + 1) + 26[n)
5.32. Since H(e"') has a zero on the unit circle, its inverse system will have a pole on the unit circle and
thus is not stable.
5.33. (a)
ftlordorpolo
(b)
(c) Only the causal ho(n] is stable, therefore only it caD be used to recover •[n].
(d)
1 1 2- lz- 1 1
H(z) =
1- 12 z-1
+ 1- jz-' = 1-
•
p-• + 11 z-2 , lzl > -2
Since h(n], z(n] = 0 for n < 0 we caD assume initial rest conditions.
5 1 5
y(n] = 611[n- 1]- 6y[n- 2] + 2>:[n]- 6>:[n- 1]
(b)
N-1
11[n] = L h(m]z(n- m]
.......
(d) For IIR., we have 4 multiplies md 3 adds per ontput point. This cnes us a total of 4N multiplies
md 3N adds. So, IIR grows with order N. For FIR. we have N multiplies md N - 1 adds for the
n" output point, so this collfiguration bas order /(l. ·
5.35. (a)
lm
Re
H(z)
= (1- 2z- 1 )(1 + lz-
1)(1 + 0.9z- 1)
(1- z-1)(1 +0.7;z-1)(1- 0.7;z-1)
1 - O.&z-1 - 2.35z- 2 - 0.9z-•
= 1 - z- 1 + 0.49z- 2 - 0.49z- 3
Y(z)
= X(z)
y[n]- y(n- 1] + 0.49y[n- 2]- 0.49y{n- 3] = :{n]- o.&:[n- 1] - 2.35:r[n- 2j- 0.9:r[n- 3]
(b)
lm
Re
I H(el"') I
~r------------,,-------------,
l
f
0
-It 0 1t
I
(I)
!t
'
(d) (i) The system is not stable since the ROC does not include 1•1 = 1.
(ii) Because h(nJ is not stable, h(nJ does not approach a coDStant as n -+ ex>.
(iii) We can see peaks at"'= ±j in the graph of IH(e1w)l shoWD in part (c), so this is false.·
(iv) Swapping poles and zeros gives:
lm
Re
There is a ROC that includes the unit circle (0.9 < lzl < 2). However, this stable system
would be two sided, so - must conclude the statement is false.
5.37.
(1- !z-1)(1- tz-
1)(1- !z) 6 (1- !z-1)(1- !z- 1)(1- sz- 1)
X(z)= (1-tzl • =s (1-~z- 1 )
ft [ J X( -1 )-~(1-!az-1)(1-faz-1)(1-5az-')
a :r n ~ a z - 5 · (1 6az-')
A minimum phase sequence has all poles and zeros inside the unit circle.
la/21 < 1 • lal < 2
la/41 < 1 • lal < 4
1
l5al<1 • lal< 5
1
16al<1 • lal<6
151
Therefore, a•:[n] is real aDd minimum phase iff a is real aDd ]a] < l·
5.38. (a) The eaUA1 systems haw conjugate zero pairs inside or outside the UDit circle. Therefore
012345n 012345n
27.8 27.1 %7.1 %7.8 %7.8
EJ.~ EJ.~
17
0.7
s..s
f ~
l.
10
012345n 012345n
152
5.S9. All zeros iDside the llllit circle meaDS the sequence is miDimWII phase. Since
Ill Ill
= 0 lAd just compute 112[0]. The largest result will he the minimWII
The aunr is F.
5.40.
(i) A zero phase sequence bas all its poles &Dd zeros in conjupt.e reciprocal pain. Generalized
linear phase sySiemS ue zero phase sySiemS with additiOD&i poles or oeroo at z = 0, oo, 1 or
-1.
(ii) A stable system's ROC includes the UDit circle.
(a) The poles aze DOt in conjugau reciprocal pain, 10 this does DOt have zero or generalized linear
phase. H;(z) bas a pole at z = 0 lAd perhaps z =co. Therefore, the ROC is 0 < 1•1 < oo, which
meaDS the iDvene is stable. If the ROC includes z = co, the inverse will also he causal.
(b) Since the poles aze not conjllple reciprocal pairs, this does DOt have rero or generalized linear
J
phase either. H;(z) bas poles illside thellllit circle, 10 ROC is 1•1 > ~match the ROC of H(z).
Therefore, the inverse is both stable <d causal.
(c) The oeroo occur in conjugate reciprocal pain, 10 this is a zero phase system. The inverse has poles
both inside and outside the UDit circle. Therefore, a stable non-causal inverse emu.
(d) The zerOs occur in conjupt.e reciprocal pairs, 10 this is a zero phase system. Since the poles of the
inverse system are 011 the llllit circle.& stable invene does 110t exist.
5.41. Convolving two symmetric sequences yields another symmetric sequence. A symmetric sequence con-
volved with an antisymmetric sequence gives an &Dtisymmetric sequence. If you convolve two antisym·
metric Sequences, you will get a symmetric sequence.
= 1, !wl <,.
A(ei")
f(w)= --, lwl < ..
153
A(J'>)
1
••• •••
(b)
-· 0 • 01
•••
11=3
•
0 •
1 •
2
l
3 4• e
5
e
6
.. .
n
•••
(l"'
0
b
3.5
l
1
2
I
II
3 4
5
I
y
6 n
•••
•••
(l:
0
6
3.25
1
2 l3
T
4
5
6
'i'
6
•••
n
(c) U a is"" integer, then h(n] is symmetric about about the point n =a. U a = ':,where M is odd,
then h(n] is symmetric about If,
which is not a point of the sequence. For a in general, h(n] will
not be symmetric.
-
N
H(ei"') = L h[n]e-;..n
(N-2)/2 N
= L h[n]e-- + L h[nje-;..n + h[M/2]•-;w(N/2)
- .-(JI+2)/2
(N-2)/2 (N-2)/2
= L h[n]e-;wn + L h[M- m]e-jw(N-,.J + h[M/2]•-;..(N/2)
"""' ......
154
(JI-2)/2 (JI-2)/2 )
= e-iw(JI/2) L h(m)el"((JI/2)-"') + L h(m)e-i"((JI/2)-M) + h(M/2)
( ......, .......
(:~
12
= e-iw(JI/2) 2h(m) cosw((M/2)- m) + h(M/2))
= e-MM/2l (~2/I((M/2)-n)cos""'+h(M/2))
Let
ca{n) ={ h(M/2), n =0
2/o[(M/2)-n), n= 1, ... ,M/2
Thea
...L
Jl/2
B(ei") = e-,;..(Jt/2) o(n) cos""'
ud.., have
...
Jl/2
M
.A(w) = L o(n) cos(wn), o=T, 13=0
...
Jl
H(el") = L ll(n)e-i""
(JI-1)/2 Jl
= L h(n]e-1"'" + L
ll(n)e-;""
..... ..-(J1+1)/2
-
(JI-1)/2 (JI-1)/2
= L ll[n)e-;wn + L
h[M- m)e-iw(JI-m)
.....
= .-,iw(Jt/2)
(
.?;,
(JI-1)/2
ll(m)e"'((JI/2)-,.) + l;,
(JI-1)/2
ll(m)e-i"'((M/2)-ml
)
2
= e-MM/2) (Jt,?;/ 2h(m) c:osw((M/2)- m))
Let
6{n) = 2h[(M + 'tJ/2- n), n = 1, ... , (M + 1)/2
Thea
(JI+l)/2
B(e"') = e-MM/2l L 6(n) cosw(n- (1/2))
-1
ud.., have
(JI+1)/2
M
.A(w) = L b[n) cosw(n- (1/2)), a=2• 13=0
-1
155
= L"' h(n)<!-Jom
B(ei"')
=
-
(M-2)/2
L
-
h(nJe-i""' + 0 +
M
L
(1./+2)/2
h(n)e-Jom
-
(M-2)/2 (M-2)/2
=
-L
= e-i.. (M/2)
h{n)e-;..." +
(
(M-2)/2 ·
L
....0
L
h(m]ei-'<(M/2)-m) _
h[M- m)e-i"'(M~m)
(M-2)/2
L
h[m]e-,;...((AI/2)-m)
m=O
)
= .-;...(AI/21 (i (AI~/ 2
2h[m]sinw((M/2) -m))
Let
c(n) = h((M/2) -n), n = 1, ... ,M/2
Then
M/2
B(ei"') = .-MAI/2lei1•1 2l L c(n)sinwn
•=1
and we have
Al/2
M
A("') =L c(n] sin(wn), o=-
2'
/J=-"
2
-=1
= e-,;...(M/2) (j (Aij;,/ 2
2h{m)sin..,((M/2)- m))
(JH1)/2
= e-;.. (M/2lei1•/2l L 2h((M + 1)/2- n] sinw(n- (1/2))
.... 1
156
Let
d[n] = 2h[(M + 1)/2- n], n = 1, ... , (M + 1)/2
Then
(Af+l)/2
udwehave
B(ei"') = e-;..(AI/2lei<•l2l
....L d[n] .m..,(n- {1/2))
(AI+l)/2 M ,.
L 2' P =2
5.44. Filter Types n aDd
A(..,)=
m c:aDIIOt be higbpass filters siDce they both must haw a .eo at z =1.
Type I -+ Type I could be higbpus:
I IIII
Type n -+ Type IV can be bighpus:
III III
Type ill -+ Type m CaDDot be bighpass:
5.45.
H(z) = (1- o.sz- 1)(1 + 2jz-1)(1- 2jz- 1)
(1 - o.az-1 )(1 + O.Sz- 1)
2j
1m
Re
-21
(a) A minimum phase sysleiD has all poles and zeros inside 1•1 = 1
H ( J= (1- o.sz- 1)(1 + iz- )
2
1
z (1- 0.64.- 2 )
lm
Re
Re
158
=
(b) A geoeralized linear phue system has zeros and poles at z 1, -1, 0 or oo or in conjugate reciprocal
pain.
· (1 - o.s.-• J
B,(z) = (1- 0.64z-2)(1 + tz-2)
lm
Re
3rd--
2j
lm
Re
-2i
5.46. (a) Minimum phue systems ha..., all poles and zeros inside 1•1 = 1. Allpass systems have pole-zero
pairs at conjugate reciprocal locations. Generalized linear phue systems ha..., pole pain and zero
pain in conjugate reciprocal JocatioiiS and at z = 0, 1,-1 and oo. This implies that all the poles
and zeros of B.,;n(•) are seconcH>rder. When the allpass filter llips a pole or zero outside the unit
circle, one is left in the conjugate reciprocal Jocati011, giving us linear phase.
(b) We know that h[n] is lqth 8 aad theftfore has 7 zeros. Since it is an even length generalized
linear phue filter with real c:oeflidents and odd syuunetry it must he a Type IV filter. It therefore
has the property that its zeros come in coajugate reciprocal pain stated mathematically as B (z) =
B(1/z"). The zero at z = -2 implies a zero at z =-!.while =
the zero at z O.Sei(•/4 1 implies
=
zeros at z = 0.&-i(•/4 1, z = 1.25ei<•14 l and z 1.25e-i(•/4 l. Becouae it is a IV filter, it also must
have a zero at z = 1. Putting all this toðer gi...,. us
!
159
X(a~'")
·~(10lt) (10lt)
5
Y(el'")
10e-j10..
H(el'")
5.48. (a)
Property Applies? Comments
Stable No For a stable, causal system, all poles must be
inside the unit circle.
IIR Yes The system has poles at locatioiiS other thaD
z=Oorz=oo.
FIR No rnr systems caD ollly have poles at • - 0 or
z- co.
Minimum No Minimum ptwe . , . . - have all poles and zeros
Phase loeated inside tbe unit circle.
Allpass No Allpua systems have poles and zeros in conjugate
reciprocal pain.
Generalized Linear PIWe No The causal geaeralised linear p!We systems
presented in this chapter are FIR.
Positive Grouo Delay for all w No This system is not in the appropriate form.
160
(b)
Property Applies? Collllllellts
St&ble Yes The ROC for this system fwlctiOII,
1•1 > 0, COIIWDs the llllit circle.
(Note there is 7th order pole at z = 0).
IIR No The system has poles ODiy at z = o.
FIR Yes The .,._ has poleo ODiy at • - o.
Minimum No By debiUoa., a Jllillimum phase system mast
Phase ha.e all its poles uad ...,. located
iMU the IIIIi& cirde.
Allpus No Noce lhat the 1er01 011 the llllit circle will
C&UI4! the mapitude opectrum to drop r.ero at
cerWD &equeDCies. Clearly, tbis syslem il
DOt allpus.
GeDeralizecl LiDear Phase Yes This illhe polefzero plot of a type n FIR
liDear phase system.
Posiu.e Group Delay for all w Yes This syslem il causal uad liDear phase.
CoDJeqUeDdy, its group delay il a positift
aliiStaD.t.
(c) .
Property Applies? CoiiUDellts
Stable Yes All poles ue iDSide the Wlit circle. Since
the system is causal, the ROC includes the
Wlit circle.
IIR Yes The system bas poles at locatioDS other than
• = 0 or z = oc.
Fffi No FIR IJS'ellll cua ollly ha.e poles at z - 0 or
z = oo.
Minimum No Minimum phase systems haft all poles uad zeros
Phase located inside the wUt circle.
All pass Yes The poles inside the llllit circle have
WiiespODdiDg leiOIIocated at CODjupte ·
reciprocallocatioas.
Generalized LiDear Phase No The causal generalizedlillear phase systems
preseated in this chapter are FIR.
Positive Group Delay for all w Yes Stable allpass systems have positift group delay
for all ....
5.49. (a) Yu. By the region of convergence we know there are DO poles at z = oc uad it therefore mast be
causal. Another way to see this is to use lollC divisiOil to write H,(z) u
H,(z) =1
1
=:-I =
-s
1 + .-• + ,-s + ,-• + z-< ,1•1 > 0
(b) h 1 [n] is a causal rectaJ1CU1ar puhe of ~el~Cth 5. If we COII'IO!ft h,[n] with uaolher causal recW1gUlar
pulse of length N we will pt a triUlplar pulse of lellClh N + 5 - 1 "' N + 4. The trWigular pulse
is symmetric aro1111d its apa uad thus has liDear phase. To make the triuacDlar pulse g[n] haft at
=
least 9 11011zer0 aamples we CUI c:hoc.e N 5 « let hs[n] h, [n]. =
Proof:
161
= [1-.-;... ]2
1- e Jw
2
= r·-jwS/2 (&wS/2 _ .-jwS/2)]
e-i~<~/2 (eiw/2 _ e-jc.>/2)
(c) The required values for h 3 [n] can intuitively he worked out using the flip and slide idea of conv~
lution. Here is a second way to get the answer. Pick h3 (n] to he the inverse system for h.(n] and
then simplify using the geometric series as follows.
1-z- 1
Hs(z) = 1- z-•
= (1- z- 1 ) [1 + z-• + z-•o + z-•• + .. ·)
= 1_ z-1 + z-5 _ 2 -6 + 2 -10 _ z-11 + 2 -15 _ 2 -16 + ...
This choice for h3 (n] will make q(n] = 5(n] for all n. However, since we only need equality for
0 S n S 19 truncating the infinite series will give us the desired result. The final answer is shown
below.
1
h:Jnl
0 5 10 15 n
-1
5.50. (a) This system does not necessarily have generalized linear phase. The phase response,
G2 (&w) = H 1 (&w)H.(&w)
jG.(&wll = jH,(&wliiH•(&w)j
LG2(dw) = LH,(&w) + LH2(&w)
The sum of two linear phase responses is also a linear phase response.
162
(c) This system does not necessarily have linear phase. Using properties t: ~ DTIT, the circular
convolution of H,(ei'") and H 2(ei'") is related to the product of h 1 [n] and h 2[n]. Consider the
systems
Proof: Although the group delay is constant ( grd (H(e1'")] = 4.3) the resulting M is not an integer.
h[n] = ±h(M-n]
H(eiw) = ±eJMw H(e-iw)
±e:i(M+4.3)~o~
e-;4.3w
= 1 [wl <w,
M = -8.6
5.52. The type II FIR system Hu(z) has generalized linear phase. Therefore, it can be written in the form
where M is an odd integer and A.(eiw) is a real, even, periodic function of w. Note that the system
=
G(z) (1- z- 1 } is a type lV generalized linear phase system.
G(eJw} = 1- e-jw
= .-)w/2(eJw/2 _ e-iw/2)
= ,-;w/2(2j sin(w/2))
= 2 sin(w /2)e-jw/i+jw/ 2
= A.(e'w}e-jw/2+jw/2
A.,(eiw)
LG(eiw)
= 2sin(w/2)
w
= --+-
2 2
.
The cascade of Hu(z) with G(z) results in a generalized linear phase system H(z).
H(eiw} = A.(eiw)A.,(ei")e-i"'M/2e-iw/2+jw/2
= A' o(eiw)eiwM' f2+jw/2
where A' 0 (eiw) is a real, odd, periodic function of wand M' is an even integer.
Thus, the resulting system H(eiW) has the form of a type m
FIR generalized linear phase system. It is
antisymmetric, has odd length (M is even), and has geoeralized linear phase.
163
5.53. For all of the following we know that the poles and zeros are real or occur in complex conjugate pairs
since each impulse response is real. Since they are causal we also know that none have poles at infinity.
(a) • Since h.{n] is real there are complex conjugate poles at z = 0.9e'*'1•1•.
o If z[n] = u[n]
H 1 (z)
Y(z) = H 1 (z)X(z) =
1 -z 1
We can perform a partial fraction expansion on Y(z) and find a term (1)nu[n] due to the pole
at z = 1. Since y[n] eventually decays to zero this term must be cancelled by a zero. Thus,
the filter must have a zero at z = 1.
• The length of the impulse response is infinite.
(b) • Linear phase and a real impulse response implies that zeros occur at conjugate reciprocal
locations SO there are zeros at Z :: Zlr 1/zl,Zi, 1/zi where Zl = Q.SeJrft ·
• Since h,[n] is both causal and linear phase it must be a Type I, ll, ill, or IV Fm filter.
Therefore the filter's poles only occur at z = 0.
• Since the arg {H2 (&")} = -2.5w we can narrow down the filter to a Type IT or Type IV filter.
This also tells us that the length of the impulse response is 6 and that there are 5 zeros. Since
the number of poles always equal the number of zeros, we have 5 poles at z = 0.
=
• Since 20log jH2 (&0 )j = -oo we must have a zero at z 1. This narrows down the filter type
even more from a Type IT or Type IV filter to just a Type IV filter.
With all the information above we can determine H,(z) completely (up to a scale factor)
(c) Since H 3 (z) is allpass we know the poles and zeros occur in conjugate reciprocal locations. The
impulse response is infinite and in general looks like
M
II (1 - <••-1)
X(z) = Go
bo .:::•=::'1~--
N
II (1- d.z- 1
)
•=1
Because x[n] is real, its zeros must appear in conjugate pairs. Consequently, there are two more
= =
zeros, at z !e-1•1•, and z ~·-''•I•. Since x[n] is zero outside 0 5. n 5. 4, there are only four
zeros (and poles) in the system function. Therefore, the system function can be written as
lm
(c) A sketch of the pole-zero plot for Y(z) is shown below. Note that the ROC for Y(z) is [zl > !·
lm
5.55. • Since z[n] is real the poles & zeros come in complex conjugate pairs.
• From ( 1) we know there are no poles except at zero or infinity.
• From (3) and the fact that z[n] is finite we know that the signal has generalized linear phase.
• From (3) and (4) we have a = 2. This and the fact that there are no poles in the finite plane
except the five at zero (deduced from (1) and (2)) tells us the form of X(z) must be
X(z) =z[-1]z + z[O] + z[1]z- 1 + z[2]z- 2 + z[3]z- 3 + z[4)z 4 + z[S)z-•
The phase changes by .. at w = 0 and .. so there must be a zero on the unit circle at z ±1. The =
= = =
zero at z 1 tells us ~>:[n] 0. Tb,e zero at z -1 tells us :[(-1)":r:[n) 0. =
We can also conclude z[n] must be a Type m filter since the length of z[n) is odd and there is a
=
zero at both z ±1. :r:[n] must therefore be antisymmetric around n 2 and z[2) 0. = =
• Fr~m (5) and Parseval's theorem we have L: [:r:[n][ 2 = 28.
• From (6)
y[OJ = 2_
27r
1•
-r
Y(eiw)dw =4
= :r:[n]• u[n) In=<> = :r:[-1) + :r:[O)
11[1] = 2_
2'JI'
1•-r
Y(eiw)ei"'dw =6
= :r:[n] ou[n) [..,., = z[-1) + :r:[O] + :r:[1]
• The conclusion from (7) that :[(-1)":r:[n) = 0 we already derived earlier.
=
• Since the DTIT {z.[n]} 1U {X(...... )} we have
z[5] + :r:[-5) 3
2
= -2
z[S] = -3+z[-SJ
:t(S] =
-3
Summarizing the above we have the following (dependent) equations
!
165
3
JC(nj
2
3 4 5
-1 0 1 2 n
-2
-3
1 1
= (1- z-• + tz-') = (1- !z-')'
H,(z)
This system function has a second order pole at z = i· (There is also a second order zero at z = 0).
Evaluating this pole-zero plot on the unit circle yields a low pass filter, as the second otder pole
boosts the low frequencies.
Since
H 2 (ei"') = H 1 (-ei"')
H 2 (z) = H 1 (-z)
If we replace all references to z in H 1 (z) with -z, we will get H 2 (z).
1
H,(z) = (I+ tz-' )2
Consequently, H 2 (z) has two poles at z = -!·
(There is also a second order zero at z = 0).
Evaluating this pole-zero plot on the unit circle yields a high pass filter, as the second order pole
now boosts the high frequencies.
166
(b) Tbe LTI system 5 3 is cbara.cterized as a highpass filter. H3 (~} is the inverse system of H 1 (eJw},
=
since H,(ei"')H,(eJw) 1. Consequently, H 3 (z}H1 (z) = 1.
~ shown in part (a), H, (z) has a second order pole at z =
~. and a second order zero at z = 0.
Thus, H,(z) has a second order zero at z = !,
and a second order pole at z = 0. Evaluating this
pole-zero plot on the unit circle yields a high pass filter, as the second order zero attenuates the
low frequencies.
S3 is a minimum phase filter, since its poles and zeros are located inside the unit circle. However,
because the zeros of S. do uot occur in conjugate reciprocal pairs, s.
cannot be classified as one
of the four types of FIR filters with generalized linear phase.
(c) First, we compute the system function H4 (z}
H4 (z) = f3o
1 + o 1 z- 1 + a 2 z- 2
s. is a stable and noncausal LTI system. Therefore, its poles must be located 01dsidt the unit
circle, and its ROC must be an interior region that includes the unit circle. We place a second
order pole at z = 2, whicb is the (conjugate} recip>;OC&llocation of the second order pole of H, (z)
at z = !- This gives
(1-2z-1)2
Po
= (1 - 4z- 1 + 4z- 2}
In order for
=
I(1- z- 11+ ~z 2) I
=
1(1-:+tll
= 4
The values a 1 = -4, a 2 = 4, and Po = 4 satisfy the criteria. Note that Po = -4 also is a valid
solution.
(d) If h,[n]• h 1 (n] is FIR, then the poles of H 1 (z) must be cancelled by zeros of Hs(z). Thus, we
expect a second order zero of H,(z) at z = l·Therefore, H,(z) will have the term (1- !z- 1 }2 •
=
In order for the filter ho(n] to be zero phase, it must satisfy the symmetry property h,[n] 11$[-n],
whicb means that H,(z) = H 0(z). For this property to be satisfied, we need two more zeros located
at z = 2. In addition, we want these zeros to correspond to a noncausal sequence. Therefore, H,(z)
will also have the term (1- ~z) 2 •
Combining tbese two results,
167
N M
11!nJ = :L "•ll!n - tl + :L o•=ln - tJ, 11o =1
.1=1 .....
N M
Y(z) = L"•Y(z)z-• + X(z) + Lb•X(z)z-•
t=l i=l
So,
Or equivalently,
n c.z- c.z)
M
(1- 1
)(1-
r ..,(z) = A u!::.";;;.''----------
2
H,.(z)H,.(z-
1
) = H(z)!(z ')
Therefore,
n N
1
(1- d.z- )(1- d·•>
H,.(z)H,.(z- 1 ) = ":";"1' - - - - - - - . . , . -
IT (1- c.z- )(1- c,z)
1
.1=1
The poles of H,.(z) are the zeros of H(z) aDd the zeros of H,.(z) are the poles of H(z). \_Ve musl
now decide which N of the 2N zeros of H,.(z)H,.(z- 1 ) to associate with H,.(z). The remaininG
N zeros and M poles will he reciprocals and will he usoci•ted with H,.(z- 1). In order for H,.(z)
to he stable, we must chose all its poles inside the nnit circle. Thus for a pair c., c; 1 we chose the
one which is inside the nnit circle.
169
{c) There is no real constraint on the zeros of H,.(z), so we can select either d• or 4, 1 • Thus, it is not
unique.
5.59. (a)
.11-1 1 -;wM
H(ei"') =~ .-;..n = - e .
~
..... 1-e-'"'
.
H;(e'"')
1-e-Jw
= 1- .M
.-,.,
.,. h;[n] = Loo o[n- kM]- o[n- kM- 1]
•=<>
1 1 1
•••
1 2M+1
0 n
-1 -1 -1
h;[n] has infinite length, so we can never get a result without infinite sums. Therefore, it is not a
real time filter. We can use the transform approach but we must have all the input data aVailable
to do this.
{b) The proposed system is a windowed version of h;[n]:
Where
n] = { 1, 0 ~ n ~ qM
Pl 0, othennse
:(n]• h(n]• h;(n]p{n] = w(n]
Therefore, if :[n] is shorter than qM points, we can recover it by looking at w(n] in the range
0 ~ n ~ qM -1.
{c)
q 1-z-qM
h 1 [n] =L 6(n- kM].,. H 1 (z) = -M
•=<> 1-z
Thus,
1 1-z-M
H 2 {z) = H(z) 1- z tM
Note that
1-z-<~M
bas M zeros and qM poles. Since H2(z) is causal, there are no poles at z = oo. If H(z) bas P
poles and Z zeros:
Z+M$P+qM
170
5.60. (a)
1 cz-1 a- z- 1
H(z) =z--=--
a a
= =-~
.u- 1
lm
, poleatz=-
Re 1/a
. . 1
H(e'") = e'"--
a
= cosw + j sinw--a1
arg[H(ei")] = tan-1 ( sinw I)
cosw--
•
{b)
lm
Re
. z 1
G(z) = -z- a = .,........;=--...,
1- az-1
G(ei")- 1
- 1 - OL!-;~.~
= I - a cosw1+ jasinw
= tan-1 (a::: 1)
= arg{H(ei")J
5.61. (a) Because hl[n], h2 [n] are minimWll phase sequences, all pales and zeros of their .transforms must
be inside the unit circle.
hi[n] • h,[n] ++ H 1 (z)Hz(z)
Since H 1 (z) and H,(z) have all their pales and zeros inside the unit circle, their product will alsa.
(b)
(1)n
z 1(n] = 6
2 u(n]+---+ 1 -lz- 1 = X1(z)
6
1)n
z,[n] = -6 (-3 u(n] ........
-6
1 -1 = x,(z)
1- JZ
X1(z) has a pole at z =!and a zero at z = 0. X 2 (z) has a pole at z =land a zero at z = 0.
z-1
X1(z) + X2 (z) = (1- 1~ _1)( 1 - 1 1)
2 3z
This has zeros at z = 0, oo an9 poles at z = ~~l· Therefore, it is not minimum phase.
5.62. (a)
r(n] =-4 -
3
(1)n u[n] + -(2)"u[-n- I]
2
4
3
R(z) = ~ ~
1- ~z 1 1 - 2z-1
-2z-l
= (1 - !z-1 )(1 - 2z-1)
1
= (1- !z 1
)(1- ~z)'
ROC: ! < lzl < 2
lm
1 zeroatz=-
Re 2
(b)
r(n] = h(n] • h[-n] ~ R(z) = H(z)H(z-1)
. 1
R(z) = -;:--,-....;,..,,..-..,...,.
(1- !z- )(1- !z)
1
We have two choices &om H(z). Suu:e h[n] is minimum phase we need the one which has the pole
at z ·= ~, which is inside the llllit circle.
±1
H(z) = (
1- 1 ,z 1 ), ROC: lzl >!
h[n] =±G) n u[n]
172
...II,(z- d,)
Since the poles are outside the unit circle, the only stable system will have a ROC of lzl <min jd,j.
This implies the poles will all contribule to the h(n] with terms of the form -(d,)nu[-n -1], which
are anticausal. The zeros are all positive powers of z, which means they are shifting to left, and
h(n] is still anticausal.
(b)
M
Hmin(z) = hmin(O] II (1- c,z- 1 )
k=l
z -1 -c~:• )
H.,(z) II
= k=l
Jl (
1 - c z-1
•
(c)
H,....(z) = hm;.[O] TI
Pl
(1 - c,z-
1
) TI (:=• -2~1 )
k=l C1c
M
= hm,.(O] II (z- 1
- ck)
M
= z- 11 hm;.[O] II (1- c0z)
J:=l
= z- 11 Hmin(z- 1 )
(d)
(b) Since
1
H.(z) = H msn. (Z )
We have
G(z) = H•• (z)
(c)
Ae
Ae
5.65.
z- 1 -a
H(z) = Hmin(z) , Jal < 1
1 -az 1
Thus,
1- az- 1
lim Hmin(z) = lim H(z)
.&-+oo .1-+oo z 1 -a
h,.;,[O] = - !h[O]
a
Therefore, Jhm;n[OJI > Jh[OJI since Jal < 1. This process can be repeated if more than one allpass system
is cascaded. In each case, the factor for each will be larger than unity in the limit. ·
5.66. (a) We use the allpass principle and place a pole at z = •• and a zero at z = .J,.
••
.
H(z)
174
(b)
€ = L lh...u.[m]l 2 - L lh[m]l'
"'=0 ~
n
- L (lq[m- 1]1 2
- z0q*[m- 1]q[m]- ••q[m- 1]q*(m] + l••l 2 lq[m]l 2 )
n
= (1 - 1••12 ) L (lq(m]l' -lq[m- 1]1')
= (1 -l••l 2 )lq[n]l 2
(d)
L lh...u.[m]l' - L lh[m]l 2 ~ 0
m=O m=O
n n
L jh[m]l' $ L lh...in[m]j 2 Vn
m=D m=O
5.67. (a) x[n] is real, minimum phase and x[n] = 0 for n < 0. Consider the system:
x(n}
~·1--+i H,.;.(z) H..,(z) f-+ylnl
x[n] is the impulse response of a minimum phase system. y[n] is the impulse response of a system
which has the same frequency response magnitude as that of x[n] but it is DOt minimUIII phase.
Therefore, the equation applies.
n n
:E l:r:[kJI' ~ :E jy[kJI 2
.... ....
Since h•• [n] is causal and :r:[n] is causal, y[n] is also causal, and these sums are meaningful.
175
(b) As discussed in the book, the group delay for a rational allpass system is always positive. That is,
Therefore, filtering a signal z[n] by such a system will delay the energy in the output y[n]. H we
require that z[n] is causal, then 11[nJ will be causal as well, and the equation
n n
:L lziA:JI• ::: :L ly[A:JI'
bO ....
!f[n] = g[n] + r[-n] = :r[n]• h[n] + :r[n] • h[-n] = z[n] • (h[n] + h[-n])
(c)
..J2 ······
31t/4
1 ······....-------. 0 fl/4 " (I)
-..J2 ·····················
0 3fl/4 " (I)
In general, method A is preferable since method B causes a magnitude distortion which is a function
of the (possibly non-linear) phase of h[n].
176
1 1
H(z) = (1- ~z-1 )(1- 2z-l) - 1- ~z- 1 + z 2
This system function bas poles at z = 1/2 and z = 2. However, as the foDowing shows it is a generalized
linear phase fiher.
1
l- •
H(.,;w) =
;>w
2eJw+e
.
•
eiw
= . .
e'"'-i+e JIM
= ( 1 ).,;w
2cosw-!
5. 70. (a) Since h[n] is a real causal linear phase filter the zeros must occur in sets of 4. Thus, if z1 is a zero
of H(z) then zi,lfz, and 1/zi must also be zeros. We can use this to find 4 zeros of H(z) from
the given information.
(b) There are 24 zeros so the length of h(n] is 25. Since it is a linear phase filter it bas a delay of
(L- 1)/2 = (25- 1)/2 = 12 samples. That corresponds to a time delay of
.
'E0>
::!
-60
-80
( 0.244)1</T
-100
0 0.2 0.4 0.6 0.8 1
n ·rht
5. 71. (a) There are many possible solutions to this problem. Tbe idea behind any solution is to have h(n]
be an upsampled (by a factor of 2) version of g(n]. That is,
Thus, h[n] will process only the even-indexed samples. One suC::.. system would be described by
h[n] = 1 +o[n- 2]
gfn] = 1+o[n-1]
H(z) = 1+z-•
G(z) = 1+z-1
(b) As in part a, there are many possible solutions to this problem. The idea behind any solution is
to choose an h[n] that cannot be an upsampled (by a factor of 2) version of g[n]. Clearly, choosing
h[n] to filter odd-indexed samples satisfies this criterion. One such h{n] would be
h[n] = 1 + o[n - 1] + o[n - 2]
H(z) = 1 + z- 1 + z-•
(c) 1n general, the odd-indexed samples of h[n] must be zero, in order for a g[n] to be found for which
=
r{n] Jl[n]. Thus, there must not be any odd powers of z- 1 in H(z).
(d) For the conditions determined in part c, g{n] is a downsampled (by a factor of 2) version of h[n].
That is,
g[n] = h{2n]
5.72. (a) No. You cannot uniquely recover h[n] from cu[l].
c.. [l] = h[~ • h{-1]
Cu(.,.iw) = H(.,.iw)H(e-iw) = jH(.,.iw)j2
Cu(z) = H(z)H.(1/z•)
Causality and stability put restrictions on the poles of H(z) (they must be inside the unit circle)
but not its zeros. We know the zeros of CM(z) in general occur in sets of 4. Here is why. A
complex conjugate pair of zeros occur in H(z) due to the fact that h[n] is real. These 2 zeros and
their conjugate reciprocals occur in CM(z) due to the formula above for a total of 4. Thus, H(z)
is not uniquely determined since we do not know which 2 out of these 4 zeros to factor into H(z).
This is illustrated with a simple example below.
Re
01/z
1
Let the above be the pole-zero diagram for c.. (z) and
H 1(z) = (1-z1z- 1)(1-ziz-1)
H(z)
2 = (1- : z- (1- :iz-
1
1
)
1
)
Since
C.,.(z) = H 1(z)Hi(1/z•) = B•(z)B0(1jz•)
we cannot determine whether h 1 [n] or h2 [n] generated c,..[l].
178
(b) Yes. The poles of c.. (z) must occur in sets of 4 for the same reasons outlined above for the zeros.
However, since the poles of h[n] must be inside the unit circle to be causal and stable we do not
have any ambiguity in determining which poles to group into h[n]. We always choose the complex
conjugate poles inside the unit circle. Here is an example
Let the above be the zero/pole diagram for c.. (z). Then, if h[n] is to be real, causal, and stable
H(z) must equal
1
H(z)-
- (1-p1z-1)(1-pjz-1)
5. 73. As shown in the book, the most general form of the system function of an allpass system with a
real-valued impulse response is
where II. is the ROC which includes the unit circle. Correspondingly, the associated inverse system is
1
H;(z) = H(z)
= IT
>=I
z- 1 (z- d•)
z I - d•
IT
>=I
2
z- (z- ••H•- et)
(z-1- e;)(z I - ••)
Now send the corrupted signal g[n] through a highpass filter hap/ [n] with a cutof of we = "/2.
179
112
• •• -1 1 •••
0 n
The highpa.ss filter completely filters out the lowpass signal :r[n]. The output y[n] is
a= 2y[no]
z[n] = g[n] - 2y[no]o[n - no]
(a) When no is odd, y[nJ = 0 at all odd values of n except n = no. This leads to a procedure to lind
:r[n] from g(n]: ·
• Filter g(n] with the highpass filter described above.
• Find the only nonzero value at an odd index in the output y[n]. This value is y[no].
• z[n] = g[n]- 2y[no]o[n- no]
(b) The only time three cousecutive nonzero samples occur in y[n] is at n = no. The procedure to lind
z[n] is
• Filter g[n] with the highpass filter described above.
• Look for three cousecutive nonzero output samples. The middle value is y[no].
• z[n] = g[n] - 2y[no]o[n- no]
5. 75. Looking at the z-transform of the FIR filter,
Solutions - Chapter 6
W(z)
z[n]
-rsin8
rsin8
rcos8
y[n]
.-•
then
W(z) = X(z)- rsin8z- 1Y(z) + rcos8z- 1 W(z)
and
6.2. The only input to the y[n] node is a unity branch connection from the z[n] node. The rest of the network
=
does not a1Iect the input-output relationship. The difference equation is y[n] z[n].
6.3.
2+ !z- 1
H(z)- 4
- 1 + lz-1 _ lz 2
• •
System (d) is recognizable as a transposed direct form ll implementation of H(z).
6.4. (a} From the flow graph, we have:
That is:
Y(z)(1 + ~z- 1 - ~z- 2 ) = X(z)(2 + ~z- 1 ).
The system function is thus given by:
1
H(z) = Y(z) = 2 + lz-
X(z) 1+ lz 1 -lz •·
(b) To get the difference equation, we just inverse Z-transform the equation in a. We get:
1 3 1
11[n] + -y[n- 1]- -y[n- 2] = 2:[n] + -z(n- 1].
4 8 4
6.5. The flow graph for this system is drawn below.
z[n] y[n]
(a)
4 4
3 3
1 1 1 1
(a) (b)
0
-1 -1
-2 -2
3 3 3
2 2 2 2
1 1 1 1
(c) (d)
0
-1 -1 -1 -1
6.7. We have
_! + z-2
H(z) = 1-• iz
1
2
.
-1/4
w1[n] w,(n]
x(n] y(n]
z-1
-1 2 w,(n]
w3(n]
z-1
4
186
Taking the Z -transform of the above equations, rearranging and substituting terms, we get:
+ .-• - s.-•
= 1 + 13z-
1
H(z)
+ z- 1 - Sz-2 .
The difference equation is thus given by:
y[nj + y[n- 1]- 8y[n- 2] = :[nj + 3:t[n- 1] + :[n- 2]- S:[n- 3].
The intpulse response is the response to an impulse, therefore:
h[nj + h[n- 1]- 8h[n- 2] = J[n] + 3J[n- 1] + 6[n- 2]- 8J[n- 3].
From the above equation, we have:
h[OJ =1
h[1] = 3- h[O] = 2.
(b) From part (a) we have:
(b) Using the Z-transform of the difference equations in part (a), we get the transfer function:
_ Y(z) _ 1 + 2z- 1 + z- 2
H( z ) - X( ) -
z 1- •1 .- 1 - ••
1 2·
We can rewrite it as :
H(z) = (1 + .-1)(1 + z-1)
(1 + !•
1 )(1- z- 1 )
-1/2
187
(c) The system function has poles at z = -! and z = 1. Since the second pole is on the unit circle,
the system is not stable.
6.11. (a) H(z) can be rewritten as:
z- 1 - 6z-• + 8z-•
H(z) = 1
1- 2 z
1 .
1/2
rz-1
-6
~-1
(b) To get the trausposed form, we just reverse the arrows and exchange the input and the ouput. The
graph can then be redrawn as:
:[n] y[n]
f-1
1/2
f-1
-6
~-1
6.12. We define the intermediate variables w,[n], w.[n] and w3 [n] as follows:
-1 w1[n] 2
:[n] ---r-------11'--'-----ll[n]
188
w,fn] =
-z[n] + w:o[n] + w,[n]
w,[n] = z[n- 1] + 2w,[n]
w,(n] = w,[n- 1] + y[n- 1]
ll(n] =
2w,(n].
Z -transforming the above equations and rearranging and grouping terms, we get:
6.13.
1- lz-2
B(z) = 1- 1 z _: - 1 -2 ·
4 8z
The direct form I inlplementation is:
z[n] y[n]
z-1 z-1
1/4
z-1 z-1
1/8
-1/2
6.14.
1+ §.z-1 + lz-2
H(z)- 6
- 1- iz-1- iiz-2.
The direct form II implementation is:
z[n] y[n]
z-1
1/2 5/6
z-1
1/2 1/6
189
6.15.
1 7 -1 + 1 -2
H(z) = - •• •• .
l+z l+lz
2
2
To get the transposed direct form II implementation, we first get the direct form II:
z[n] y[n]
.-1
-1 -7/6
.-1
-1/2 1/6
Now, we reverse the arrows and exchange the role of the input and the ouput to get the transposed
direct form II:
:r[n] y[n]
.-1
-7/6 -1
.-1
1/6 -I .
....'
6.16. (a) We just reverse the arrows and reverse the role of the input and the output, we get:
:r[n] y[n]
.-1 .-1
-1/2 -2
.-1
1/4 3
190
(b) The original system is the cascade of two tr&DSposed direct form n structures, therefore the system
function is given by:
H(z) = ( 1 - 2z-1 + 3z-2 )(1- !z-1).
1- !z-2 2
•
Tbe transposed graph, on the other band, is the cascade of two direct form n structures, therefore
the system function is given by:
Tbis confirms tbat both graphs have the same system function H(z).
6.17.
6.18. The flow graph is just a cascade of two tr&DSposed direct form U structures, the system function is
thus given by:
(1+2z- 1)(1-lz-
3
1)
H(z)-
- (1 + ~z-1 - jz-2)(1- az-1)
In order to implement this system function with a second-order direct form n signal flow graph, a
pole-zero cancellation has to occur, this happens if a = l, =
a -2 or a 0. U a = = f,
the overall system
function is: ·
_ 1 + 2z- 1
H< z ) - 1 + iz 1 1 3
- az
2"
1 + lz-1- !z-2
• •
6.19. Using partial fraction expansion, the system function can be rewritten as:
-8 1
H(z) = 1 1 -1 + 1 2 -1 + 9.
- JZ + 3Z
Now we can draw the Bow graph that implements this system as a parallel combination of first-order
transposed direct form n sections:
-8
:r(n] y(n]
z-1
1/3
z-1
-2/3
which can be implemented as the following cascade of second-order transposed direct form n sections:
192
z[n] y[n]
z-1 z-1
2 5/2
z-1 z-1
5/4 -1/4 -1
6.21.
So y[n] = =
e'w•y[n- 1] + :t[n]. Let y[n] llr[n] + jy;[n]. Then llr[n] + jy;[n] = (cosWo +; sinWQ)(y,[n-
1] + jy;[n- 1]) + :r[n]. Separate the real and imaginary parts:
:r[n] !lr[n]
-sinwo
COSWo
Jl;[n]
6.22.
(1 + z-1)2
H(z) = (I - tz-1 )(1 - !z-1 )"
1 1
1 + z- ) ( 1 + :z- )
H(z)"' ( 1- tz-1 1 -!z-1 ·
193
:r[n] y[n]
• • • • •
1 1/4
• F J J
1/2
FI
( 1 + z-1 ) ( 1+ z-1 )
H(z) = 1- iz-1 1 -lz-1 .
:r[n] y[n]
• • • • •
l LI l
1/2 1/4
!·-: I
Plus 12 systems of this form:
:r[n] y[n]
1/4 1/2
1- lz-1
5
H(z)-
- (1- ~z-1 + lz-')(1 + tz-1).
(a) (i) Direct form I.
so
. 1 1 5 1
bo = 1, bt =--and
5
At=-
4
, a2 = --
24
, 03 = --.
12
194
z[n] y[n]
z-1 z-1
-5/24 z-1
-1/12
z[n] y[ n]
.-1
1/4 -1/5
.-1
-5/24
.-1
-1/12
(iii) Cascade form using first and second order direct form IT sectious.
195
1
1 + !z-2).
3
So
bo1 =I , b11 = 1>,1 = 0, -! ,
bu.= I, b12 = 0, ~ = 0 and
au=-~, a21 = 0, au= i, on= -l·
:r:[n] v[n]
.-1 .-1
-1/3
(iv) Parallel form using first and second order direct form II sections.
We can rewrite the transfer function as:
27 98 .J!.. -1
H(z) = I+ffi!z 1 + 1 _m- 1uz
!z-1 - iz-2 ·
•
So
eot = t':s 'eu =0'
36
•- - .!!. e12--125'
- and
It
.. 11-
-
""U2-125'
-i1 I
..
,_21 -
-0 I
n _,a-
2
.,.12- -3·
1 22-
1
196
27/125
z-l
-1/4
(n] y[n]
z-•
1/2 -36/125
z-•
1/3
w 1 (n]
[n] y[n]
z-•
to:z(n]
-1/5 1/4
z-•
"'•(n]
-5/24
z-•
-1/12
(b) To get the difference equation for the flow graph of part (v) in (a), we first define the intermediate
variables: w,[n] , w,[n] and w3 [n] . We have:
197
1
H(z) = 1- az 1
y[n] z[n]
• •
y[n) z[n)
• •
1/2 1/4
198
1 1
y[n] = :t[n] + :~:[n- 1] + y[n- 1]
4 2
1 + lz-1
HT(z) =
1- ,z
1 1
= H(z)
(c)
y[n]
•
:r[n]
•
y[n] = 4:t[n] + bz[n - 1] + c:r[n - 2]
HT(z) = "+ bz- 1 + cz- 2 = H(z)
(d)
H rsinllz- 1
(z) = 1- 2rcosllz 1 + r 2 z 2
y[n]
.-1
w[n]
-rsinll
rsinll
rcosll :r[n]
199
v = X +z-1U
u = rcosiiV- rsiniiY
w = rsiniiV + rcosllz- 1 W
y = .-1w
y
=-
X =
Hr(z)
rsinez- 1
= 1- 2r cosBz-1 + r2z-2
= H(z).·
6.25. (a)
H(z)
1 - llz-1
•
+ •~z-2 - !z-3
8
(b)
9 9 11 7
y[n] = 2z[n] + 8't:[n- 1] + Sz[n- 2] + g-z[n- 3] + Sz[n - 4]
11 5 7
+ gll[n- 1] - 411[n- 2] + By[n- 3].
(c) Use Direct Form II:
z[n] y[ n]
.-1
11/8 9/8
.-1
-5/4 9/8
.-1
7/8 11/8
.-1
7/8
The solution is not unique; the order of the denominator 2ud-order sections may be rearranged.
(b)
1
u[n] = :z:[n] + 2:z:[n- 1] + :r[n- 2] + u[n- 1]- u[n- 2]
2
1
v[n] = u[n]- v[n- 1]- 2"[n- 2]
w[n] = v[n] + 2v[n- 1] + v[n- 2]
7
y[n] = w[n] + 2w[n- 1] + w[n- 2] + 2y[n- 1]- gll[n- 2].
-"If
-·
T 2• .. "'
(b) For H,(z) =H(-z), replace each z- 1 by -z-1 . Alternatively, replace each coeflicient of an
odd-delayed variable by its negative.
201
(c)
[n] II[n]
.-•
-1 2
.-•
2 -1
.-•
-2 1
.-•
1 -2
6.28.
z[n] y[n]
b
"
.-•
a w[n] -1
(a)
(b)
202
z(n] y(n]
• •
-b
6.29. (a)
z[n]
• I
y(n]
•
(b) From
it follows that
~n-n
7 1 -az
• -·
L-az = 1-az 1"
R=0
(c)
z[n]
•
1..
a
I.-: l . !-·:
•
y(n]
(a)
203
...·-r-----------~-==-~'=-·~-----------.
•
- ....
•• .
... . ..
·-r-----------~-=-='~-=·-~-----------,
. .
...
• •
• •
(b)
(c)
H(~w) = 1- e -jlSw [
1 . ~ -i~
+ 20 . + 1 e'·~ ]
15 1- .-,.., 1- e>fte-iw -1--....--"';if;;-e_-j,-w
When no = 15/2, .
!~n sin(15w/2)]
sin ( w+<2; /15))
- ·-
....... ...
..
When no= 0,
e-;w7 [sin{15w/2) !e-h'< sin(15w/2)
B(ei"') = -15- sin(w /2)
+ sin { w-("; /IS)) +
!~r. sin(15w/2J]
sin ( tot+(2;/15))
.....,.,_ .....
The system will have genesalized linear phase if the impulse sesponse has even symmetry (note it
cannot have odd symmetry), or alt.ematively, if the frequency sesponse can be expressed as:
B(ei"') =.-;•7 A.(ei"')
205
where A.(ei"') is a real, even, periodic function in w. We thus conclude that the system will have
generalized linear phase for no = Jtk, where k is an odd integer.
(d) Rewrite H(z) as
1/15
:tnJ y"fin]
z-15 z-1
-1
") /3
-1
G u[n]
<TnJ 1+r
!I[n]
-r r -1
1-r
,::-r wn] .-1
(b)
respectively.
(c) U(z) = z- 1 (GX(z) + W(z)), W(z) = -rU(z)- (1- r)Y(z), and Y(z) = z- 1 ((1 +r)U(z)- rY(z))
lead to
= 1 +G(12rz+ r)z-
2 _2
H2(z) 1 + z 2 =z H 1 (z).
6.32. (a)
(b)
(c)
(a)
207
•
z[n]
• •
11[n]
b c
H(z) = cdz-1 + 1d
1- bz-
= =
so set b 0.54, c -1.852, and d -0.54. =
(b) With quantized coefficients b, c, and d, &i "t 1 and d "t -bin general, so the resulting system would
DOt be alJpass.
(c)
.-1
•
z[n] •
11[n]
-1
..
1
-1 w[n -1] -1
The first delay in the second section has output w[n- 1] so we can combine with the second delay
of the first section,
.-1 -1
•
z[n]
w 1 (n] w,[n]
2 v(n]
•
:(n]
•
First, we lind the system function, we have:
Taking the Z-transform of the above equations and combining terms, we get:
(1- z- 1 )Y(z) + z- 1 Y(z) = (2 + z- 1 )X(z).
The system function is thus given by:
1
H(z) = Y(z) = 2 + z-
X(z) 1+z '-z-2
Since the system function is second order (highest order term is z- 2 ), we should be able to im-
plement this system using only 2 delays, this can be done with a direct form II implementation.
Therefore, the minimum number of delays required to implement an equivalent system is 2.
w,(n] ws[n]
2
:(n]
w,[n]
Taking the Z -transform of the above equations and combining terms, we get:
Since the transfer function is not the same as the one in part a, we conclude that system B does not
represent the same input-<:~utput relationship as system A. This should not be surprising since in
system B we added two unidirectional wires and therefore changed the input-<:~utput relationship.
6.35.
z-1-!
H(z) 1 !, .
= } - jZ
z-1/3
z(n] y(n]
.-• z-•
,.,.
z
From the graph above, it is clear that 2 delays and 2 multipliers are needed.
(b)
1
y[n] = :i(y[n- 1]- :z:[n]) + :z:[n- 1]
.-1
---....------..>-----.----•
:z:[n] • y[n]
(c)
l: Fl
-1
z- 1 - !
.-1
I
z- 1 - 2
H(z)=(1-lz 1)(1-2z- 1 ).
3
This can he implemented as the cascade of the Bow graph in part (b) with the foUowing Bow graph:
I, .
.-1
:r[n] • • y[n]
-1
However the above Bow graph can be redrawn as:
l: l .-1
-1 .-1
:r[n] •
. ,-·j
I
: l' I • y[n]
Now cascading tbe above Bow graph with the one from part (b) and grouping the delay element
we get the foUowiog system with two multipliers and three delays:
!
-1
-I
211
6.36. (a) Transpose = reverse arrows direction and reverse the inputfoutput, we get:
w,(n]
•
l' .
:r(n] • • y(n]
_,:,,~
(b) From part (a), we have:
l :, !·-:
"'' (nJ.
•
W3(n] z-l
Taking the Z -transform of the above equations, substituting and rearranging terms, we get:
1
y(n]- 2y(n- 1]- 2y[n- 2] = :r(n] + 2z(n - 1].
H(z) = ~ + 2z-l •.
1- 2z- 1 - 2z
It has poles at
8 8
z= -..:...,= and z
1 - -133
= - -..::....,=
1 + -133
which are outside the unit circle, therefore the system is NOT BIBO stable.
(d)
H[O]
.-•
...
ii[IJ
.-•
z,
1/N
:r[n] y[n]
' '
' '
'' '
'
''
' '
H[N- 1]
(b) Note that the •• 's are the zeros of (1- z-N). Then write H(z) over a common denominator:
H(z) =
n:-',;,1<1- ••• -• >
=
N-1
}: H; II
-[ j N-1
(1- z;z-').
k=O ....
·-·
Therefore, H(z) is the sum of polynomials in z- 1 with degree :S N -1. Hence, the system impulse
response has length :S N.
(c)
!
213
H(z.,) = (1 - z-N)ii(m]!N
1 - z,z- 1
I
z=.z: ..
a N •
= 4{(1- z- )H(m]/N}Iz=z-
1;{1 - Z,Z 1 }lz=z-
Nz;;.N-1H(m]/N
= z,z,•
= B(m]z;;.N
= H[m].
(e) If h(n] is real, IH(eiw)l =
IH(ei!2•-wl)l, and LH(ei"') -LH(ei< 2•-wl). H(ei 2rtfN) = = H[k] =
=
IH(k]lei81•j, so IH[k]l !H[N- k]l and B[k] = -ti[N- k], k = o, 1, ... , N- 1.
1
H[k]/N + H(N/2]/N +
Ztz-l 1 ZN/ 2 z l
~
.£..-
~~~
H[l]/N ]
1- ZtZ l
L., 1 - ZN-.,z- 1
A:=l p=l
..!_
N
t:-
L.,
1
H[k](1 - Z-tz- 1 ) + H[N - k](1 - ztz- 1 )]
(1- Ztz- 1 )(1- Z-tZ 1)
~1
N 2dz-l+z-2
l-2cos IV
1
~1
k[OJ/16
[n] y[n]
-z-16 .-•
lik[IJI
"
.-•
{3 Q
.-•
'Y
xTnJ
h(O] h(1] h(2] h[10]
M:!
'f"n]
z-1
-
M:1 M:1 M:1
vTnJ
N + 1 multiplies per output sample; N adds per output sample. The number of computations has
been reduced by a factor of M in both adds and multiplies.
(c)
M:1
[n] 1111n]
X -
z-•
I
1/"!!
The total computation can not be reduced because to compute the value of any given output
sample, the previous output value must be known.
216
(d)
nI
M:1
.-1 .-I
718 112
nu
M:1
.-1
112 I
7,8
M:1
.-1 .-1
I
1/2 7I 8
rIV>
M:1
.-1
718 112
Only direct form IT (ii) can be implemented more efliciently by commuting opefatiODS with the
doWDSalllplers.
3.4cm -10-·
3.4~ · 1o< - sec
to traverse one section. Since the sampling rate is 20ldh (T, = 0.5 · 10--.sec), it takes two sampling
intervals to traverse a section. The entire system is linear and so the forward going and backward going
217
y(n]
•
x(n] •
,...., = 1
Ps(e)
1/A
-A/2 A/2
Ps(•)
1/A
-A
218
6.41. Since the system is linear, y[n] is the sum or the outputs due to :z:1[n] and :z:2[n]. Therefore
00 00
= !lt[n] + 112[n].
E{y,(m]112[n]} = E Lt..,
00 00
h,[l]z 1 [m -l] · .t.., h2[k]:r2[n- k]}
= L L h,[l]h2[k]E{x,[m -l]:r2[n- k]}
t=-oo 1=-oo
If :z:,[n] and :z:2[n] are uncorrelated, E{:r 1 [m -l]:z:2[n- k]} = 0; hence, E{y,[m]y2[n]} = 0. Therefore,
y,[n] and 112[n] are uncorrelated.
6.42. (a) The linear noise model for each system is drawn below.
b, a
(2a 2 )
(c)
x[n] bo y[n]
• •
I
'
. .-·j
:b,.
I (3<7')
I a
l·-:
(b) Clearly (a) and (c) are different. Thus the answer is either (a) and (b) or (b) and (c). H we take
(b) apart, we get
220
bo
z-1 z-1
b1 a
• •
I:·-l r F I
b1 a
We see that the noise all goes through the poles. Note that the lei' source sees a system function
(1 - az- 1 )- 1 while the 2<1 2 source sees z- 1 /(1- az- 1 ). However, the delay (z- 1 ) does not affect
the average power. Hence, the answer is (b) and (c).
(c) For network (c),
..-} = 3u'-l-f
27rj
1 . _I_ dz
1- az- 1- az z1
= 3u'-l-f
21rj
dz
(z- a){l- az)
3u'
= 2<T'+<>'(b'+<abo+b,Y).
o 1-a2
• Frequency domain calculation:
1
~>'(n] = -21fJ /H(z)H(z- 1) dz
z
ft
-b,bo
residue (z = 0) = - a
-
~a+ I?, a+ b,bo + b1 boa
2
.d ( ) (boa+ b,)(bo + b,a)
res1 ue z =a = a( 1 - a') = o( 1 _ c') .
1 n (I)' II-U}n+l
y(n] = 2L 4 = 2 !
i=O •
ll
32
y[n] unqoa.ntized
----J---~--~L---J---~----L------ n
0 I 2 3 4 5
222
i
! ! 5 5 !
--l...-.,-f_.l..__i..L--l-i___._j_ . _ _ j •I: ~-
0 1 2 3 4 5
1
X(~w) _ 2
- 1 + e-Jt..~
So
=
which implies that y(n] (1/2)(1/4)", which approaches 0 as n grows large.
To find the quantized output (working from the difference equation): y(O] 1/2, y(1] = =1/8, and
=
y(n] 0 for n ;:=: 2.
1
2
fi
1 _, ...!...
128 1
512
2048 y(n] unqua.ntized
n
0 1 2 3 4 5
! !
•
y(n] quantized
n
0 1 2 3 4 5
6.44. (a.) To check for stability, we look at the poles loca.tion. The poles are Joca.ted at
Note that
[z[ 2 "' 0.976 < 1.
The poles are inside the unit circle, therefore the system function is stable.
223
· ·z-·D-a• · n.-·
z[n] ----~--w-i,ll-[n_J_ _ _
a"'-J
w..[n_J_ __.._ _ _ _ y[n]
·
z[n] .-• z-•
~
• I
Taking the Z -transform of the above equations and combining terms, we get:
That is:
H(z) = 1 -a • z -· .
1-az- 1
For Network 2, we have:
(b) Network 1:
z{n]
Network 2:
1
., a'
••
a• a• a• a1
y{n]
w,...<1-laJ.
Where we assumed tlw JaJ < 1. The transfer function from z{n] to w 1!nJ is 1- a•z-•, therefore
to avoid overflow at that node we need:
1-[a[
:z:m4%< - 1 ••
-a
Now, for network 2, the transfer function from input to output is given by 6[n] +acl[n -1] +a2 6[n-
2] + ... + a7 o[n -7], therefore to avoid averllow, we need:
1
:.,.... < 1+ Ia[+a2 + ... +,..
,_ 7·
1
(e) For network 1, the total noise power is ,~tr For network 2, the total noise power is 7~. For
network 1 to have less noise power than network 2, we need
~,,<7~.
1- a
That is:
5
[al < 7·
The largest value of [al such that the noise in network 1 is less than network 2 is therefore ~.
227
Solutions - Chapter 7
h 1 [n]
'
we get
s 0 (t) =
1-~
h,(T)dT +---+ - '-
.
H (s)
= Sc(s)
S _ s+a _ A, A2 A;
,(s)- •(s+a+jb)(s+a-jb)- s + s+a+jb + s+a-jb
where
a 0.5
At= a2 +til' A,=---
a+ jb
Though the system h2 [n] is related by step invariance to h,(t), the signals 2 [n] is related to s,(t) by
impulse invariance. Therefore, we know the poles ofthe partial fraction expansion of S,(s) above
must transform as Zl = e•• T, and we can find
s, (z) = 1 _Az1 1 + 1 _ e A2
(a+jb)T z-1
•·
+ .,--..,.-.,....,=--;-
1 _ e-<• ,.}T z-1
Now, since the relationship between the step response and the impulse response is
n oo
s,[n] = L h,[k] = L h,[k]u[n- k] = h2 [n]• u[n]
S,(z) = 1-
H,(z)
z 1
!
230
(c)
=
1
2
[1 _
1=-oo
.-(o+jb)(n+l)T
1 _ e (o+jb)T
1:=0
+
1_ .-(o-jb)(n+l)T]
1 _ e-<•-ib)T u[n]
= [Bt + B,e-C•+jb)Tn + B;e-(•-ib)Tn u[n) l
where
1- e-•T cosbT e-<o+j6o)T
B, = 1- 2e •T cosbT + e-24T' .8, = - 1 _ e-C•+ib)T
;.From this we can see that
B, B2 Bi
= I - z-1 + 1- e-(o+ji)T z-1 + 1 - e-(o-ji)T z
I S,(z)
since the partial fraction constants are different. Therefore, s 1 [n] # s2[n), the two step responses
are not equal.
where A1 and A2 are as defined earlier. By comparing h,[n] and h2[n) one sees that h,[n] I h2[n).
The overall idea this problem illustrates is that a filter designed with impulse invariance is different
from a filter designed with step invariance.
7.2. Recall that n = wfT•·
(a) Then
0.89125 ~ IH(jO)I ~ 1, o ~ 101 ~ 0.2.. IT•
IH(jO)I ~ 0.17783,
The plot of the tolerance scheme is
I H(jO) I
231
(b) As in the book's example, since the Butterworth frequency response is monotonic, we can solve
1
IHc(i0.2 ../TdJI' = 2N = (0.89125) 2
1+ ( 0.2 .. )
OcTd
1
IHc(i0.3.-/TdJI 2 = 2N = (0.17783) 2
1 + (0.3.. )
OcT4
to get OcTd = 0.70474 and N = 5.8858. Rounding uptoN= 6 yields OcTd = 0.7032 to meet the
specifications.
(c) We see that the poles of the magnitude-squared function are again evenly distributed around a
circle of radius 0.7032. Therefore, Hc(s) is formed from the left half-plane poles of the magnitude-
squared function, and the result is the same for any value of T4. Correspondingly, H(z) does not
depend on Td.
(b) Solving the equations in Part (a.) for 61 and 0,, we find
cl,
6, = 2- cl,
26,
0. = 2- cl,
In the example, we were given
Plugging in these values into the equations for 6, and 0., we find
o, = 0.0575
0. = 0.1881
232
The filter H'(z) satisfies the discrete-time filter specifications where H'(z) = (I+ ot)H(z) and
H(z) is the filter designed in the example. Thus,
1. [ 0.2871- 0.4466z- 1 -2.1428 + 1.1455z-1
H'(z) = 0575
1- 1.297lz-1 + 0.6949z-2 + 1- 1.0691z 1 + 0.3699:- 2
1.8557- 0.6303z- 1 ]
+ 1- 0.9972z-1 + 0.2570z-2
0.3036- 0.4723z- 1 -2.2660 + 1.2ll4z- 1
= 1- 1.297lz-1 + 0.6949z-2 + 1 - 1.0691z- 1 + 0.3699z-•
1.9624- 0.6665z-1
+ :-1--...;0::;.99=72='-z-_.;:1:.:+::0;:.2::;5~7=0-z--=-2
(c) Following the same procedure used in part (b) we find
[ 0.0007378(1 + z- 1)1
H'(z) = 575
l.0 (1-1.2686z 1 + 0.7051z 2)(1- 1.0106z- 1 + 0.3583:- 2)
x 1 - 0.9044z-! + 0.2155z- 2]
0.0007802(1 + z- 1)1
= (1- 1.2686z-1 + 0.7051z-2)(1- 1.0106z- 1 + 0.3583z 2)
1
x "'"1---o=-.""9044z,..,.,-_,_
1 -+""'0'"".2'"'1'"'5"'5-z"""•
7.4. (a) In the impulse invariance design, the poles transform as •• = e"T' and we have the relationship
He(s)
= 2/T• _ 1/T•
• + 0.1+ 0.2 •
1 0.5
= -----
•+0.1 •+0.2
The above solution is not unique due to tbe periodicity of z = eiw. A more general answer is
_ 2fT•
- s+ (0.1+i~) s+ (o.2+iW)
where k and I are integers.
(b) Using the inverse relationship for the bilinear transform,
1 + (Td/2)s
z=
1- (T./2)s
we get
2 1
He(s) = 1- • ...,.• (W.) 1- • ...,.• (W.)
2(•+ 1) (• + 1)
= s(1 + e-<>.2) + (1- e-<>·2) a(1 + e-<>·•) + (1- e-<>·•)
6 = O.Ql
l!.w = 0.05.-
A = -20log10 6 = 40
A-8
M + 1 = 2.2BS.<:l.w + 1 = 90.2 -+ 91
= 0.5842(A- 21)0 ·4 + 0.07886(A- 21) = 3.395
{3
(b) Since it is a linear phase filter with order 90, it has a delay of 90/2 = 4S samples.
(c)
Hd(ei"')
7.6. (a) The Kaiser formulas say that a discontinuity of height 1 produces a peak error of 6. If a filter has
a discontinuity of a different height the peak error should be scaled appropriately. This filter can
be thought of as the sum of two filters. This first is a lowpass filter with a discontinuity of 1 and
a peak error of 6. The second is a highpass filter with a discontinuity of 2 and a peak error of 26.
ln the region 0.37r::; lwl::; 0.4757r, the two peak errors add but must be less or equal to than 0.06.
c! + 26 ::; 0.06
6,.... = 0.02
A= -20log(0.02) =33.9794
{3 = 0.5842(33.9794- 21) 0 "4 + 0.07886(33.9794- 21) = 2.65
(b) The transition width can be
l:.w = 0.37r - 0.27r ar
l!.w = 0.525lf - 0.475lf
= O.llr rad = 0.05.- rad
We must choose the smallest transition width so l!.w,.... = 0.05.- rad. The corresponding value of
Mis
33.9794-8
M = 2.285(0.05") = 72.38 -+ 73
7.7. Using the relation w =fiT, the passband cutoff &eqnency, w9 , and the stopband cutotr frequency, w.,
are found to he
w, = 20'(1000)10-'
= 0.2.. rad
"'• = 2r(l100)10-'
0.22.- rad
=
234
Therefore, the specifications for the discrete-time frequency response H•(ei") are
0.99 $ IH•(ei")l $ 1.01, 0 :Siwl $ 0.201<
IH.(ei"ll $ O.Gl, 0.22 .. $ '"'' $ ..
7.8. Optimal Type I filters must have either L + 2 or L + 3 alternations. The filter is 9 samples long so its
order is 8 and L = M /2 = 4. Thus, to he optimal, the filter must have either 6 or 7 alternations.
Filter 1: 6 alternations Filter 2: 7 alternations
Meets optimal conditions Meets optimal conditions
7.9. Using the relation w =nT, the cutoff frequency w, for the resulting discrete-time filter is
We = flcT
= [2.. (1000)][0.0002]
= 0.4.. rad
w, = 2tan- 1 Cl;T)
= 2 tan -• C"(2000)(~.4 x w-•))
= 0.7589.- rad
7.11. Using the relation w = !lT,
w,
n, = T
1</4
= 0.0001
= 25001r
= 2.. (1250) rad
s
7.12. Using the bilinear transform frequency mapping equation,
0, = -tan
T2
-
("'')
2
= 2 ("'2)
0.001 tan 2
2000rad
= s
= 21<(318.3) rad
s
7.13. Using the relation w = OT,
T = "'•
o,
27t/5
= 21<(4000)
= 501'5
This value of T is unique. Although one can find other values of T that will alias the continuous-time
frequency n, = 2.-(4000) raD./s to the disaete-time frequency "'' = 2lr /5 rad, the resulting aliased filter
will not he the ideal lowpass filter.
235
= ftan(~')
T 2
= 2.-(300) tan (31r/5)
- 2- = 1.46 ms
The only ambiguity in the above is the periodicity in w. However, the periodicity of the tangent function
"cancels" the ambiguity and so T is unique.
7.15. This filter requires a maximal passband error of 0, = 0.05, and a maximal stopband error of o, = 0.1.
Converting these values to dB gives
op= -26 dB
0$ = -20 dB
This requires a window with a peak approximation error less than -26 dB. Looking in Table 7.1, the
Hanning, Hamming, and Blackman windows meet this criterion.
Next, the minimum length L required for each of these filters can be found using the "approximate
width of mainlobe" column in the table since the mainlobe width is about equal to the transition width.
Note that the actualleogth of the filter is L M + 1. =
Hanning:
8,.
O.l7r = M
M = 80
Hamming:
8,.
O.b = M
M = 80
Blackman:
12,.
O.llr = M
M = 120
7.16. Since filters designed by the window method inhereotly have 6, =0, we must use the smaller value for
o.
0 = 0.02
A = -20log10 (0.02) = 33.9794
{J =0.5842{33.9794- 21) ·' + 0.07886{33.9794- 21)
0
2.65 =
A- 8 33.9794 - 8
M = 2.285Cw = 2.285(0.65.-- 0.631r) =
180.95-+ 181
7.17. Using the relation w = OT, the speclfiations which should be used to design the prototype continuous-
time filter are
-0.02< H(jO) < 0.02, 0 :S 101 :S 2r(20)
0.95 < H(jO) < 1.05, 21r(30) :S 101 :S 2r(70)
-0.001 < H(jO) < 0.001, 2r(75) :S 101 :S 21r{100)
236
Note: Typically, a continuous-time filter's passband tolerance is between 1 and 1 - 6, since historically
most continuous-time filter approximation methods were developed for passive systems which have a
gain less than one. H necessary, specifications using this convention can be obtained &om the above
specifications by scaling the magnitude response by ,_:,..
7.18. Using the bilinear transform frequency mapping equation,
0, = !.
T
tan ("'•) =
2 2 X
2
10 3
tan (~)
2
= 2.-(51. 7126) rad
s
2
0, = !.tan("'')=
T 2 2 X I0-3
tan ( 0 ·23") = 2..(81.0935) rad
s
Thus, the specifications which should be used to design the prototype continuous-time filter are
• - .!._
- T•
(1-
1
·-'~)
+e-i~ ·
The continuous frequency axis gets warped onto the discrete-time frequency axis, but the magnitude
values do not change. H H(s) is constant for all s, then H(ei") must also be constant.
7.21. (a) Using the bilinear transform frequency mapping equation,
we have
(b)
237
K ················ . . . . . . . . . . . . . . . . . . . . . ·········· .. .
0+---------------------------------+
0
(c)
w, = 2tan-• ('"~;T•)
2t.an-• (n~·)
w, =
tow= w,- w, = 2[tan-• (n;T•)- t.an- 1
(n~T·)]
lt
ot-~~===================+
0
1•1 > 1
which has the impulse response
h[n] Td (u[n] + u[n- 1J)
= 2"
238
y(n] =
r. (:z:(n] + :z:(n- 1]) + y(n- 1]
2
This system is not implementable since it has a pole on the unit circle and is therefore not stable.
(c) Since this system is not stable, it does not strictly have a frequency response. However, if we ignore
this mathematical subtlety we get
= r. (l+e-j"')
2 1-e '"'
=· r. (d--'2 + .-;..'•)
2 e;../2- .-;..12
T•
= 2
j cot(w/2)
and since the Laplace transform evaluated along the j!l axis is the continous..time Fourier transform
we also have
I H(ei"') I
T(2
LH(ei"') L H (j.Q)
c
!tl2 !tl2
lt It
-It 0 Ol -lt 0 0
-!t12 -!t12
1n general, we see that we will not be able to approximate the high frequencies, but we can
approximate the lower frequencies if we chooser.= 4/'lf.
(d) Applying the bilinear transform yields
G(z)
JzJ > 1
which has the impulse response
g(n]
239
(e) This system does not strictly have a frequency response either, due to the pole on the unit circle.
However, ignoring this fact again we get
G(.,;w) = ;d [! ~ :=;:]
= 2 (.,;w/2 _ e-iw/2)
Tc~ eiw/2 +e jw/2
= i. tan(w/2)
2"
G(jO) = jO
~ I G(ei"') I
0
-II 0 1112 II
"' -II 1112
LG(ei"') LG (jO)
c
1112 7112
-II -II
0 II 0 II 0
"'
1112 -1112
Again, we see that we will not be able to approximate the high frequencies, but we can approximate
the lower frequencies if we chooser.= 4/7f.
(f) If the same value of T• is used for each bilinear transform, then the two gystems are inverses of
each other, since then
10lt
Then, to get the overall system response we scale the frequency axis by T and bandlimit the result
according to the equation
n
(b) Using the frequency mapping relationships of the bilinear transform,
0 = :.tan(~),
w = 2tan- ( ~·) '
0 1
we get
10lt
--
0.981t
-lt
Then, to get the overall system response we scale the frequency axis by T and bandlimit the result
according to the equation
H
I ....
(jfl)l - { IH,(&wT)I, 101 < f
- 0, 101 >;.
241
10lt
-98001t
7o24o (a) Expanding the sum to see things more clearly, we get
r A
= L (s- ••o)• + Ge(s)
bt
= - A- + ( A )2+000+( A. ) + GeS
1 2 ( )
•-.so s-.ao s-so r
Now multiplying both sides by (• - ~o)• we get
(•- •ol' He(•) = A1(•- •o)•-l + A•(•- ~or-• + 000+A.+ (•- •ol'Ge(s)
Evaluating both sides of the equal sign at s = so gives us
= (s- •ol' He(•) J..,,,
Ar
Note that (s- so)•Ge(s) =0 when s = 'o because Ge(•) bas at most one pole at s =So-
d'-·
At = (r _1 k)! ( cis•-• [(•- so)' Be(s))J,z,. )
(b) Using the following transform pair from a lookup table,
t>-1 1
(k _ 1)! ·-"'u(t) -+ (• + a)•, "IU{s} > -a
we get
242
7.25. (a) Answer: Only the bilinear transform design will guarantee that a minimum phase discrete-time
filter is created &om a minimum phase continuous-time filter. For the following explanations
remember that a discrete-time minimum phase system bas all its poles and zeros inside the
unit circle.
Impulse Invaria.nce: Impulse invariance maps left-half 6-plane poles to the interior of the z-plane
unit circle. However, left-half s-plane zeros will not necusarilN be mapped inside the z-plane
unit circle. Consider:
H(z)
1
+f..
l.zol = 11- -.•o I
Since He( s) is minimum phase, all the poles of He(s) are located in the left half of the s-plane.
Therefore, a pole s 0 = a+ jrl must have a < 0. Using the relation for 1 0 , we get
(1 + fa)• + (tfl)2
l.zol = (1- fa)• + (f0)2
< I
Thus, all poles· and zeros will be inside the z-plane unit circle and the discreu-time filter will
be minimum phase as well.
(b) Answer: Only tbe bilinear transform design will result in an allpass filter.
Impulse Invariance: In the impulse invariance design we have
Bilinear Transform: The bilinear transform only warps the in!quency axis. The magnitude
response is not affected. Therefore, an all pass filter will map to an all pass filter.
(c) Answer: Only the bilinear transform will guarantee
H(ei") = H,(jO)
if and only if
or equivalently
H(z) (2(1-z-'))
= H, T• 1 + z-1
(f) Answer: The property holds for both impulse invariance and the bilinear transform.
Impulse Invariance:
= H 1 (ei"') + H2 (e'"')
244
Bilinear Transform:
H\~) = H.\~R\-+·~'0)
= H.,(;. G::=:)) +H.,(:. G::=:))
= H1 (z) + Hz(z)
(g) Answer: Only the bilinear transform will result in the desired relationship.
Impulse Invariance: By impulse invariance,
Hz (e'w) = f:
11:=-oo
+ 2;•k))
H., (; (;.
We can clearly see that due to the aliasing, the phase relationship is not guaranteed to be
maintained.
Bilinear Transform: By the bilinear transform,
and we desire
we see that
requires
f: H. (i~k) = 0.
·---
.....
(b) Since the bilinear transform maps ll = 0 tow = 0, the condition will hold for any choice of H.Uil).
7.27.
245
(a)
h1 [n] = h[2n]
H, (e'w) = L"" h[2n]e'wn
n=-oo
= I:
n even
h[n]e"i"
= f.
n.=-oo
~ [h[n] + (-lth[n]]e'"f
I
= 2H(e' ¥ ) + 2H
I ( e>--r-
·•'")
H, (ei"')
1/2
(b)
h[n/2]e-Jwn
H2(dw) =
I:
n even
"" h[n]e-Jw2n
= I:
n.=-oc
= H (e'>w)
H (ei"')
2
- 1
-
-x-7lt!8 -lt/8 0 lt!8 7lt!8 7[ (1)
(c)
1- z- 1
• = 1 + z-1
1- e-jw
;n = 1 +e jw
eil!ll/2 _ e-Jw/2
= eJw/2 + e j,.;/2
n = tan(~)
n. =tan (1'-) .......... w,. = 2tan- 1 (0,)
(b)
1 + z- 1
• = 1- z- 1
1 + e-jw
jO = 1- e-iw
eJ~i~~/2+ e-iw/2
= ei,.,/2 _ e-iw/2
n = -cot G)
= tan ("'~w)
n. = -w)
tan ( w "'2
(c)
(d)
The even powers of z do not get changed by this transformation, while the coefficients of the odd
powers of z change sign.
Thus, replace A, C, 2 with -A, -C, -2.
(b)
247
H, (ei"')
ec ec
A
(c)
h[n] +-+ B(,;')
h 1 [n] +-+ B ( ,;<2w+•l)
In the frequency domain, we first shift by 1r and then we upsample by 2. In the time domain, we
can write that as
h [n]
1
= { (-l)nl2 h[n/2], for n even
O, for n odd
(d) In general, a filter
H(z) = bo + ~z-l + b,z-• + · · · + bM-IZM-l + bMz-M
Qo + O.tz-l + 0.2Z 2 + · · · + O.N-lZN l + O.Nz-N
will transform under Hl(z) = H(-z2 ) to
bo- b,z-2 + ~>,z-• +.--- b.v_,z>M-2 + bMz-•M
H 1 ( z) = Go-
~--"-'.::..__..-':-=-::-.-.:,--_..:;.:;c::..:.;:....,o-:;..:,..:=.:.=
O.tz-2 + a2z-4. + ... - 0.M-1Z2N-2 + O.Nz-2N
where we are assuming here that M and N are even. All the delay terms increase by a factor of
two, and the sign of the coellicient in front of any odd delay term is negated.
The given difference equations therefore become
g(n] = z[n] + a,g[n- 2]- b!/[n- 4]
f[n] = -a.g[n- 2]- b,f[n- 2]
y[n] = cd[n] + c2g[n - 2]
To avoid any possible confusion please note that the b, and a, in these difference equations are
not the same b, and a, sbown above for the general case.
7.30. We are given·
H(z) = He(s) I -/3[=]
1+·-·
I=
• = {3 [1-
l+z
z-o]
0
{3-. = z- ({3 + s)
0
{3-.
z-" = fJ+•
zo {3 +.
= fJ-.
!
248
The poles •• of a stable, causal, continuous-time filter satisfy the condition 1U {s} < 0. We want
these poles to map to the points z•
in the z-plane such that Jz•l < I. With a > 0 it is also true
that if lz•l < 1 then lzfl < 1. Letting •• = u + jw we see that
lz•l< 1
lzfl< 1
I.B+a + iOI < 1.8- u- iOI
(,8 + a) 2 + 0 2 < (,8- a) 2 + 0 2
2<7,8 < -2<7,8
But since the continuous-time filter is stable we have &{ ••} < 0 or a < 0. That leads to
-,8 < .8
This can only be true if ,8 > 0.
(b) It is true for ,8 < 0. The proof is similar to the last proof except now we have jz0 j > !.
(c) We have
z• =
~~:Ln
2
Jz 1 = 1
jzj = 1
Hence, the jO axis of the s-plane is mapped to the unit circle of z-plane.
(d) First, find the mapping between 0 and w.
1- e-i'lw
j{l = 1 + e-i2w
eJw- e-il,,
= e~w + e-iw
n = lan(w)
w = tan- 1(0)
Therefore,
1 + z- 1 s+1
s= 1 +--+z= - -
1-z s-1
Now, we evaluate the above expressions along the jO axis of the •-plane
jO+ I
z = jO -I
izl = I
249
(b) We want to show 1•1 < I if 'R.e{ s} < 0.
<T+jO+ I
z = <T+jO -I
,f(.. + I)'+ o•
1•1 = ,j(<T I)'+ 0 2
it must also be true that <T < 0. We have just shown that the left-half s-plane maps to the interior
of the z-plane unit circle. Thus, any pole of Hc(s) inside the left-half s-plane will get mapped to
a pole inside the z-plane unit circle.
(c) We have the relationship
1 + e-i""
jO = I-e 'w
eit.J/2 + e-Jlll/2
= eiw/2 - e-Jw/2
o = - cot(w/2)
10,1 = lcot(1t/6)!=v'3
10, I = Icot(•"/2)1 = o
l!l,.,l = lcot(,/4)1=1
Therefore, the constraints are
<2 = 2..1•
21f -11'
IE(e'w)l 2 dw
= L""
n.=-oc
le[n]l 2
where
h..ln], n <0,
e(n] = h4[n]- h(n], O$n$M,
{ h..(n], n>M
(b) Since we only have control over e[n] for 0 $ m S M, we get that <' is minimiud if h(n] = h..(n]
for 0$ n SM.
(c)
= { ~:
OSnSM,
w(n] otherwise.
7.33. (a)
IHd(~w)l = 1, 'lw
i- 'TW,
-j -Tw,
:----JW2
-It 0
-W2 r-------.:
(b) A Hilbert transformer of this nature requires the filter to have a zero at z = 0 which introduces the
180° phase diHerence at that point. A zero at z = 0 means that the sum of the filter coefficients
equals zero. Thus, only Types ill and IV fulfill the requirements.
(c)
For the -windowed Fm system to be linear phase it must be antisymmetric about '1·
Since the
ideal Hilbert transformer h..[n] is symmetric about n = T we should choose T = '1·
(d) The delay is M/2 = 21/2 = 10.5 samples. It is therefore a Type IV system. Notice the mandatory
zero at w = 0.
0!---------------------~It
251
(e) The delay is M /2 = 20/2 = 10 samples. It is therefore a Type ill system. Notice the mandatory
zeros at w = 0 and ,.. .
lH(ei"')J
0~--------------------~
"
7.34. (a) It is weU known that convolving two rectangular windows results in a triangular window. Specifi-
cally, to get the (M + 1) point Bartlett window forM even, we can convolve the foUowing rectangular
windows.
WR,(eiw) = [i"sin(wM/4),-iw(lf-!)
VM sin(w/2)
WR,(eiw) = {2 sin(wM/4) ,-,w( 'f+!)
VM sin(w/2)
Ws(eiw) = WR, (eiw)WR,(eiw)
2
= 2_ (sin (wM/4)) -jwJI/2
M sin(w/2) e
Note: The Bartlett window as defined in the text is zero at n = 0 and n = M. These points are
included in the M + 1 points.
For M odd, the Bartlett window is the convolution of
= { .[b.
n-- 1, ... ,-2-
M-1
••In]
0, otherwise
In the frequency domain we have
(b)
where
W (.;w) = sin(w(M + 1)/2)) -jwM/2
R sin(w/2) e
Below is a normalized sketch of the magnitude response in dB.
0 It
IHd(ei'")l
I I 1/2 1/2 I I
-It ~.67t ~.31t 0 0.371 0.67t II {I)
This can be viewed as the sum of two lowpass filters, one of which has been shifted in frequency
(modulation in time-domain) to w = ... The linear phase factor adds a delay.
h•[n] = sin(0.3lr(n- 24)) + !(-l)(n-••lsin(0.4.-(n- 24))
1r(n- 24) 2 .-(n- 24)
253
(c) To find the ripple values, which are all the same in this case since it is a Kaiser window design, we
first need to .determine A. Since we know p and A are related by
0.1102(A- 8.7), A> 50
{J = 3.68 = { 0.5842(A- 21) 0 ·4 + 0.07886(A- 21), 21 $ A $ 50
0, A< 21
we can solve for A in the following manner:
1. We know P = 3.68. Therefore, from the formulas above, we see that A~ 21.
2. If we assume A > 50 we find,
3.68 = 0.1102(A- 8.7)
A = 42.1
But, this contradicts our assumption that A > 50. Thus, 21 $ A $ 50.
3. With 21 :5 A :5 50 we find,
3.68 = 0.5842(A- 21) 0·' + 0.07886(A- 21)
A = 42.4256
With A, we can now calculate 6.
6 = 10-A/20
= 10-42.425<1/20
= 0.0076
The discontinuity of I in the first passband creates a ripple of 6. The discontinuity of 1/2 in the
second passband creates a ripple of 6/2. The total ripple is 36/2 = 0.0114 and we therefore have
6, = 0. = 6, = 0.0114
Now using the relationship between M, A, and t:.w
A-8
M = 2.285t:.w
42.4256-8
t:.w = 2 _28S(4S) =0.3139"' O.llr
Putting it all together with the information about Hd(eiw) we arrive at our final answer.
0.5
o+--------r~~=-==~~~----------r---
0
!
254
7.36. (a) Since H(ei0 ) ¥0 and H(ei•) ¥0, this must be a Type I filter.
(b) With the weighting in the stopband equal to 1, the weighting in the passband is ~-
W(Ol)
1.6f---------.
1
00~------------~~----~~--------------~----(1)
0.4lt 0.58lt It
(c)
IE(Ol)l
(d) An optimal (in the Parks-McClellan sense) Type I lowpa.ss filter can have either L + 2 or L + 3
alternations. The second case is true only when an alternation occurs at all band edges. Since this
filter does not have an alternation at w = ..
it should only have L + 2 alternations. From the figure,
= =
we see that there are 7 alternations so L 5. Thus, tbe filter length is 2L + 1 11 samples long.
(e) Since the filter is 11 samples long, it has a delay of 5 samples.
(f) Note the zeroes off the unit circle are implied by the dips in tbe frequency response at the indicated
frequencies.
Re
0
101h order pole
7.37. (a) The most straightforward way to find hc(n] is to recognize tbat H•(ei") is simply tbe (periodic)
convolution of two ideallowpa.ss filters witb cutoff frequency "'• = ../4. That is,
where
255
sin2 (.-n/4)
=
(b) h[n) must have even symmetry aronnd (N- 1)/2. h[n) is a type-1 FIR generalized linear phase
system, since N is an odd integer, and H(eiw) ¥- 0 for"'= 0. Type-1 FIR generalized linear phase
systems have even symmetry aronnd (N- 1)/2.
(c) Shifting the filter hd[n) by (N - 1)/2 and applying a rectangular window will result in a causal
h[n) that minimi:r.es the integral squared error •· Consequently,
where
w[nJ = { 1, 0$ n$ N - 1
0, othei"W!Se
(d) The integral squared error <
• = L Ja[n)- hd[n)J 2
-oo
Since
-(N-1)/2-1 (N-1)/2
= L 2
L "" Ja[n)- h•[n)J 2
< Ja[n)- hd[n)J +
-(N-1)/2
Ja[n)- hd[nll' + :E
(N-1)/2+1
-(N-ll/2-1 00
= L 2
Jhd[n)J + 0 + L Jhd[n)J
2
-oo (N-1)/2+1
By symmetry,
•=2 L"" Jhd[n)J
2
(N-1)/2+1
7.38. (a) A Type-llowpass filter that is optimal in the Parks-McClellan can have either L + 2 or L + 3
alternations. The second case is true only when an alternation OCCUIS at all bal!d edges. Since this
=
filter does not have an alternation at "' 0 it only has L + 2 alternations. From the figure we see
there are 9 alternations so L = 7. Thus, M = 2L = 2(7) = 14.
256
(b) We have
hHp(n] = -ei'"h£p(n]
HHP(ei"') = -HLP(ei(w-•l)
= -A.(ei<w-•l)e-j(w-•l¥
= A.(ei<w-•l)e-i'-'¥
= B.(ei"')e-;w\(
where
B.(ei"') = A.(ei<w-•l)
The fact that M = 14 is used to simplify the exponential term in the third line above.
(c)
(d) The assertion is correct. The original amplitude function was optimal in the Parks-McClellan
sense. The method used to create the new filter did not change the filter length, transition width,
or relative ripple sizes. All it did was slide the frequency response along the frequency axis creating
a new error function E'(~o~) = E(~o~- 1r). Since translation does not change the Chebyshev error
(max IE(~o~)IJ the new filter is still optimal.
7.39. For this filter, N = 3, so the polynomial order L is
L=N-1=1
2
Note that h[n] must be a type-1 FIR generalized linear phase filter, since it consists of three samples,
and H (ei"') t- 0 for "' = 0. h[n] can therefore be written iD the form
=
The filter must have at least L + 2 3 alternations, but no more than L + 3 = 4 alternations to satisfy
the alternation theorem, and therefore be optimal in the minimax sense. Four alternations can be
obtained if all four band edges are alternation frequencies such that the frequency response overshoots
at w = 0, undershoots at w = i, overshoots a.t w = ~, and undershoots at w = 1r.
Let the error in the passband and the stopband be 6p and 6,. Tben,
7.40. True. Since filter Cis a stable I1R filter it bas poles in the left half plane. The bilinear transform maps
the left half plane to the inside of the unit circle. Thus, the discrete filter B has to have poles and is
therefore an Iffi filter.
7.41. No. The resulting discret.>-time filter would not have a constant group delay. The bilinear trans-
formation maps the entire jfl axis in the s-plane to one revolution of the unit circle in the z..plane.
Consequently, the linear phase of the continuous-time filter will get nonlinearly warped via the bili-
nar tranSform, resulting in a nonlinear phase for the discrete-time filter. Thus, the group delay of the
discrete-time filter will not be a constant.
7.42. (a) Using the fact that H,(s) = }:\:\ and cross multiplying we get
H ( ) = Y;,(s) A
' s X,(s) = s+c
(s + c)Y;,(s) = AX,(s)
dy,(t) + cy,(t) = Azo(t)
dt
(b)
dy,(t)
dl.
I t=nT
= [Az,(l)- cy,(tlJJt=nT
= Az,(nT)- cy,(nT)
y,(nT) - y,(nT- T)
::: Az,(nT)- cy,(nT)
T
258
(c)
y[n] -y[n- 1]
= Az[n] - cy[n]
T
Az[n] = (c+ f) y[n]- fll[n -1]
AX(z) = (c +f) Y(z)- fY(z)z-•
H( ) = Y(z) A
z X(z) =
c+ +- +•-•
(d)
H,(s) J ,_,_,
~---.- = s~c~ -=-.-
•-·-•
A
= ~+c
= H(z)
(e) First solve for z
1- .-•
s =- T
-
1.
% = 1-sT
and then substitute s = "+ jn to get
1
z = 1- (u + ;n)T
=
1 ei--·us>
V(1- u)• + (nT) 2
1 ( .• ..)
= 2(1- <7) e' + .-,
and thus the s-plane maps to the z-plane in the following manner
Therefore, the jO-uis maps to a circle of radius 1/2 that is centered at 1/2 in the z-plane. We
also see that_ the region u < 0, i.e., the left half of the s-plane, maps to the interior of this circle.
s-plane
z-plane
If the continuous-time system is .table, it. poles are in the left half s-plane. As shown above, these
poles map to the interior of the unit circle and so the discrete-time system will also be stable. The
stability is independent of T.
Since the jO-uis does not map to the unit circle, the ~ete-time frequency response will not be
a faithful reproduction of the continuous-time frequency response. As T gets smaller, i.e., as we
oversample more, a larger portion of the jO-uis gets mapped to the region close to the unit circle
at w = 0. Although the frequency range becomes more compressed the shape of the two responses
will look more similar. Thus, as T decreases we improve our approximation.
(f) Substituting for the first derivative in the differential equation obtained in part (a) we get
y,(nT + T) - yo(nT) ( T)
T + cy, n = Ax,(nT)
Y(z) A
H(z) =X( z ) = T+c
1 = H,(s) 1,-•-•
_...,.....
z-1
s =
T
z = 1+sT
= I+ CT +jOT
To find where the jO axis of the s-plane maps, we let s = jO, i.e., " = 0 and find
z =I +jOT
Therefore, the jO-uis lies on the line ~{z} = 1. We also see that the region u < 0, i.e., the left
half of the s=plane, maps to the left of this line.
s-plane z-plane
!
260
If the continuous-time system is stable, its poles are in tbe left half •-plane. As shown above, these
poles can map to a point outside the unit circle and 10 the discrete-time system will not necessarily
be stable. There are cases where -varying T can tum au oostable oystem into a stable system, but
it is not true for the general case.
Since the jO-axis does not map to the unit circle, the discrete-time frequency respoDOe will not be
a faithful reproduction of the continuous-time frequency response. However, as T gets smaller our
approximation gets better for tbe same reasons outlined for the first backward difierence.
7.43. (a} Just doing the integration reveal$
AT
ile (T)dT + Ye(nT- T) = lle(T}I:t-T + 1/e(nT- T) = lle(nT)
/.nT-T
Using the area in the trapezoidal region to replace tbe integral above, we get
Y(z) Af(1 + z- 1 )
H(z) = X(z) = 1 + cf- z 1 + z-14
(d)
= 1- z 1 + cf(l + z 1}
= H(z)
7.44.
+.(jO) = H.(jO)H.(-jO)
+(z) = H(z)H(z- 1 )
261
(a} (i} Since He(•} has poles at ••• H.( -•} has poles at -••·
(ii} The material in this chapter shows that under impulse invariance
-
A• T•A•
- +---+ 1- ~ltT,.z-1 .
8 - Sj;
Thus, going from step 1 to step 2 means that the autocorrelation of the discrete-time system
is a sampled version of the autocorrelation of the continuous-time system.
(iii} Since +(z) = H(z)H(z- 1) we can choose the poles and zeros of H(z) to be all the poles inside
the unit circle, and that choice leaves all the poles and zeros outside the unit circle for H(z- 1 ).
Consider the following example using h.(t) = ,-a•u(t).
1 1
H.(s) =- -
s+Q
and H.(-s) =- -
-s+o
+.(s) = H.(s)H.(-•i
= [.~a] [-s~a]
1/2o _ 1/2o
= s+o s-o
+(z} =
if a> 0, then
(b) Since IH.(jn)l 2 =+.(in) and +(ei"') = H(ei"')H(e-;..} = IH(ei"')i', we see that since ¢[m] =
T.¢.(mT.), .
7.45. (a) Since the two 8ow diagrams are equivalent we have
z-1-a 1-az
z-1 =
1-az 1
z-a
=--- z-a
z = 1-az
.
H(z) = H,,(Z)Iz= t--a
·-· = H,, (z-a)
1- -
- - QZ
Although a warping of the frequency seale is evident in the figure, (except when o = 0, which
corresponds to z-•
= .-• ), if the original system bas a piecewise-constant lowpass frequency
response with cutofF frequency 89 , then the transformed system will likewise have a similar lowpass
response with cutofF frequency w9 determined by the choice of o.
xjnl----r-------r---ytn]
bo
xjnJ---~--=--,----Yfn]
iz-1
The transformation would destroy the linear phase of the FIR filter since the mapping between 8
and w is nonlinear. The only exception is the special case when o = 0, i.e., when 8 = w.
Since there are feedback terms in the transformed filter, it must be an IIR filter. It therefore has
an infinitely long impulse response. ·
(e) Since the two 6ow diagrams are equiva.lent we have
z-• -a az
z-• = z
-1
1- az- 1
=z
-11 -
--
z- a
z-a
Z = z--
1-az
Letting Z =& 8
and z = ei"' we have,
= eiwei""-~
1-aeJ"'
· ei"" - a 1 - ae-;v,
= eJW . .
1 - aeJw 1 ae-J111
.
· ei"' - 2o + a'e-;"'
= ~..,~~~----~~
1- 2ocosw+oZ
264
(I)"
2!t .......... .
-It
rc/2 It 9
-It
............ -211
We see from the plot of w ..,rsus 8 that a lowpass filter will not always transform into a lowpass
=
filter. Take, for example, the case when the originallowpass filter has a cutoff of 8 "f2. With
a = 0 it would transform into an allpass filter.
G(ei"')
A= (1-6?}
B = 1
c = -26, -6~
D = 26, -6i
If 61 « 1 and 0:, « 1 then,
Maximum passband approximation error "" 0
Maximum stopband approximation error "" 20:,
(c) Since
The new tolerance specifications can he found in a similar manner to the last section. We get,
A = I-36?-26~
B = 1
c = 0
D = 36i+Ul
(d) The order of the impulse response h(n] isM. Since it is linear phase it must therefore have a delay
of "f
samples. To convert the two systems we must add a delay in the lower leg of each network
to match the delay that was added by the first filter.
l(nl------;f,...------1~'---~-n)- T - h [ n _ J__,t--y(n]
2z......
l(n]---,tr---i h(n) IJ 2
j h[n) H.__hl_nJ_.~y(n]
3z....,
The restrictions on the filter that carry over from part (a) are that it have
(i) Even symmetry
(ii) Odd Length
Hence, Type I Fffi filters can be used.
The length of h[n] is 2L+ l. Since the term that is longest in the twicing system's impulse response
is the h[n] • h[n] term, the length of g[n] is 4L + 1. Since the term that is longest in the sharpening
system's impulse response is the h[n] • h[n] • h[n] term, the length of h5 harp[n] is 6L + 1.
7.47. We know that any system whose frequency response is of the form
L
A.(~"')= L:a•(cos(w))•
1=0
can have at most L - 1 local maxima and minim• in the open interval 0 < w < .- since it is in the form
of a polynomial of degree L.
If we include all endpoints of the approximation region
(b) A polynomial of degree L can have at most L- 1 local minima or maxima in an open interval.
Since A. (~"') has t4ree local extrema in the interval from 0 < w < .-, we know L ~ 4.
Note that the optimal filter is half wave anti-symmetric if you lower its frequency response by one
half, i.e.,
A.,.(~"') = -A.,. (~<•-wl)
=
where A• .,(~"') H.,.(e;"') - 1/2. Another way of saying this is to say that the optimal filter is
=
anti-symmetric around w .-/2 after lowering the response by 1/2. This property holds because the
optimal filter has symmetric bands with the same number of alternations. Plugging in A.,.(e-i"') =
H..,(ei"')- 1/2 into the above expression gives
h.,.[-n], n odd
h..,[n] ={ 0, neven,n-:/0
0.5, n=O
A sample plot of h.,.[n] appears below, for L = 6.
112
6
n
= =
Note that because h.,.[n] 0 for n even, n -1 0, a plot of h..,[n] for L 5 would have the same
=
nonzero samples, and therefore be equivalent. So the optimal filter with L 6 is really the same
= =
filter as the case of L 5, just as the optimal filter with L 4 is the same filter as the case with
L=3.
We know the filter non-optimal filter has 7 alternations. The optimal filter should be able to meet
the same specifications, but with a lower order. From part (a), we know the number of alternations
must be even. Thus, the optimal filter for these specifications will have 6 alternations.
An optimal lowpass filter has either L + 2 or L + 3 alternations which means L = 4 or L = 3.
However, we showed above that these are really the same filter. Since the optimal fi.lter has L = 4,
the fi.lter shown in the problem cannot have L = 4.
Putting it all together we find L > 4 for the fi.lter shown in the figure.
7.49. (a)
268
(b) The delay of tbe linear phase system is 51/2 = 25.5 samples since it is a linear phase system of
order 51. Therefore, the total delay is
Delay =
=
--
H(ei=)
25.5T + O.ST
26T
= 2.6 ms
HoUO)
(c) H(dflT) should cancel the effects of H0 (jO). However, to cancel the effects ofthe delay introduced
by Ho(jfl) would require a noncausal filter which is not practical in this situation. Using the
relation w = nT,
oo~~~~~----------~--
0.2lt 0.41t It "'
H(ei"')
0 0.2lt "'
It "'
(d) lf Hr(ifl) is also sloping across the band, 101 < 7r/T, we would combine its effects with those of
Ho(jfl) and compensate as in part (c), i.e.,
M $ "
"'•
So the maximum allowable decimation factor is
"
Mmax=-
"'•
(b)
V(e"")
11------..
Y(e"")
"
1/1001-------------------,
(c)
11-----.
"
W1(fi'>}
'~I
00
.~
0.45ll 5()(1)$1 "
(I)
270
V2(#>)
,~1
00 .~
0.4511: O.SK
"
Y(#>)
>0001
00
\
0.911 II
(d) After the first decimation by 50 is performed, W1 (&w) should look like the following:
Cll
w,<#i
SOw,, $ 1.55,.
w,, $ 0.031,.
In general, the number of multiplies required to compute a single output sample is just N. For
a linear phase 6lter, however, the symmetry in the coefficients allow us to cut the number of
multiplies (roughly) in half if implementing the 6lter with a difference equation. The following
is an example of how this is accomplished using the simple Type I linear phase 6lter h[n] =
0.256[n] + 6[n- 1] + 0.256[n- 2].
"' 232
-10log,.(O.Ol X 0.001)- 13 +1
N• = 2.324(0.51<- 0.45.. )
"' 103
H we again use linear phase filters we find
116 multiplies to get each sample of vt [n]
0 multiplies to get each w,
sample or [n] from v, [n]
52 multiplies to get each sample of v,[n] from w,[n]
0 multiplies to get each sample of y[n] from v,[n]
The total number of multiplies is 168.
(g) We have
H,(eiw) is real since A,(eiw) is real and 6, is real. It is nonnegative since A,(eiw) 2: -6,. Note
that H,(eiw) is an even function or w and is a zero-phase filter.
(b) H 3 (&w) is a zero-phase filter with real coeflicients. Thus, a zero at z• implies there must also be
zeros at z<, 1/z•, and 1/zt· In addition, a zero on the unit circle must be a double zero because
272
both its value and its derivative is zero. Note that this last property is true for H,(&w) but not
for A.(eiw). We can write H 3 (z) as
H,(z) = H,(z)H,(l/z)
where H,(z) contains all the complex conjugate zero pairs inside the unit circle and H,(I/z)
contains the corresponding complex conjugate zero pairs outside the unit circle. We factor one of
the double zeros on the nnit circle and its complex conjugate zero into H,(z). The other pair on
the nnit circle goes into H 2 (1/z).
Since H,(z) has its zeros on or inside the unit circle it is minimum phase ( - allow minimum
phase systems to have zeros on the unit circle in this problem). Since the zeros occur in complex
conjugate pairs, h2 [n] is real.
(c)
wE passband
wE stopband
Therefore,
The original filter h[n] has order M. Therefore, h1 [n] also has order M, but h2 [n] has order M/2
due to the spectral factorization. Since h,.;.[n] has the same order as h2 [n]we lind that the length
of h.un[n] is M /2 + 1.
(d) No. If we remove the linear phase constraint, then the zeros of H 3 (z) ue not distributed in
conjugate reciprocal quads. It then becomes impossible to express
M
H(eiw) = L h[n]e-.iwn
.....
273
(M-1)/2 M
= L h[n]e-jwn + L h[n]e-jwn
n=O n.=(M+l)/2
(M-1)/2 (M-1)/2
= L h[n]e-iwn + L h[M- m]e-iwM .;wm
..... m=O
(M-1)/2 (M-1)/2 ]
= .•-jwM/2 L h[n]eiw(M/2-n) + L h[n]e-jw(M/2-n)
[ ..... .....
(M-1)/2
= .-jwM/2 L 2h[n]cosw(M/2- n)
.....
(.11+1)/2
= .-jwM/2 L 2h[M{'- n]cosw(n- !l
=1
Then
(M+1)/2
H(eiw) = .-jwM/2 L b(n]cosw(n- 1/2)
=
where b[n] 2h[(M + 1)/2- n] for n = 1, ... , (M + 1)/2.
(b) Using the trigonometric identity
we get
--.-
M->
¥ ¥-
cos(w/2) L ii[n] coswn = 21"-
L.. b[n]cosw(n+ !l + 21"-
L..b[n]cosw(n- 21 )
n=O n=O n=O
¥ ¥
= 21"-
L.,.b[n-1]cosw(n-!)+ 1"-
2
L.,.b[n]cosw(n- 12 )
n=l n=)
1- 1-M 1
+ 2b[O]coswf2- 2b[+JcoswM/2
¥
= ~L (b[n] +ii{n -11) cosw(n- !l + }ii[O]cosw/2- !ii[M:[-I]coswM/
=1
we can just matcb up the multipliers in front of the cosine \ermS of the two expressions. We get
b[1] + 2ii{Oj
n=1
2
b{n]= ii{n] + ii[n - 1] 2 < n < .!!=!
2 1 - - 2
11 1
ii[ 2 1 n - ll±l
2 • - 2
274
(c) Consider
M/2
= .-iwM/2 L 2jh(M/2-m] sinwm
=1
Then
M/2
H(eiw) =.-jwM/2 L c(n] sinwn
=
where c(n] 2jh(M /2- n] for n = 1, ... , M /2.
If we Follow a similar analysis as the one in part (b) we get
llf llf
= 2 I:c[n -l]sinwn- 2'LC!n]sinwn
n=l n=l
+ ~C[O]sinw+ ~c[¥JsinwM/2
= ~ t
n=1
(C{n- 1]- c[n])sinwn + ~C[O] sinw + ~c[¥JsinwM/2
Matching terms we get
2C{OJ- c[IJ
n=l
2
C[n - I]- C[n]
c[n]=
2
cr¥- 11 n -- M
2
2
In a manner similar to that of part (c) we can find
M
L = --1
2
Hd(ei~)
if.(ei~) = sinw
W(w) = W(w) sinw
i' = F
Type ry filters:
M
H(ei~) = L h[nje-i•.n
n=O
(M-1)/2
= L h[n]e-Jwn +
J\==0 n=(Jl+l)/2
(M-1)/2 (M-1)/2
= L h[nje-i"'n- L h[mje-J"'(M-m)
(M-1)/2
= .-j~M/2 L h[nj ( ,-J~(n-M/2) _ eJw(n-M/2))
n=O
(M-1)/2
= .-jwM/2 L (-2j)h[n]sinw(n- M/2)
n=O
(M+1)/2
= .-iwM/2 L 2jh((M + 1)/2- m]sinw(m- 1/2)
m=1
Then
(M+l)/2
H(ei"') =.-JwM/> L d[n] sinw(n- 1/2)
-1
276
J{M;'J n -- M±l
-2-, 2
M-1
L "' 2
H.(~"')
fl.(~"') =
sinw/2
W(w) = W(w)sinw/2
F = F
(b) The filter length is 2L + 1. The causal version of the fiow graph looks like
x{n] -
••
"• 1
: O.S(Z+Z- ) r- a ..__,
-0-----+-----1---+y"{n]
The filter length is still 2L + 1. The modified fiow graph looks like
IC(n)
a,
JI 0.5(1+Z"")
"o a,
.-·
(d) Because Z = ei' and z = ei"' we have
ei' + ,-;•
2 = <to+<>, [ ei"' +2·-'"']
cos II = Qo + 01COSW
cosw =
cos II- <>o
"'(cos II- <>o)
' for ~cos~~""\ :S 1
_1
w = cos
"'
(e)
!
278
-I
·1·:.~
..8
a>
0
0 :··
-~-~------~----~
1 cos(ro) _,
It
The picture above shows the mapping for a 0 somewhere between 0 and 1. The top right plot is
the mapping of
=
ms8 ao + (1-ao)msw
We see that as ao increases, the transformation pushes the new passband further towards .-. The
new filter is not generally an optimal filter since we lose ripples or alternations while keeping L
fixed. (Note that some of the original filter does not map anywhere in the new filter).
(f) In a similar manner, this choice of ao will cause the new passband to decrease with decreasing Qo·
7.54. (a) Let Dt(z) be the z-transform of fl.(tl{:r[nj}. Then
(b) By taking the transform of both sides of the continuous-time difterential equation one gets (assum-
ing initial rest conditions)
N M
:~:>•••Y(s) = 2),s' X(s)
A:=O r=O
Solving for He(s)
Similarly,
N M
L at(z- z-l )•Y(z) =L b,(z- z- 1
)' X(z)
.... r=O
=> m(z) = z- z- 1
(c) First, map the continuous-time cutoff frequency into discrete-time and then make the sketch.
s = z- z- 1
j!l =.!"'- .-;~
eiw- e-jw
0 = 2sin(w) =I
j
w=-..6
H(ei"')
h1 [n] = h[-n]
Ht(ei~) = H(e-;~)
Since H(ei~) is symmetric about w = 0, H(e-;~) = H(ei"'). Thus, H 1 (ei~) = H(e-i-') = H(ei~).
H(e'~) is optimal in the minimax sense, so H1 (ei~) is optimal in minimax sense as well.
280
Bd(ei"') ={ ~: O$w$wp
w. :s w :s 1r
h2(n] = (-1)nh(n]
= (e-i•)nh(n]
B2(ei"') = B(ei!w+•l)
B2(ei"') is a high pass filter obtained by shifting B(ei"') by,. along the frequency axis. B2(e'"')
satisfies the alternation thereom, and is therefore optimal in the minimax sense.
0$ w :S 1r -w.
r-w,:Sw:SK
0$ w :s 71"- "'·
w-w,:Sw511'
(c) Using DTIT properties,
The filter h3[n] is the convolution of two length N sequences. Therefore, the length .of h3(n], de-
noted as N', is 2N- 1. Since N is either 11 or 13, N' must be either 21 or 25. It follows that
the polynomial order for h 3 (n], denoted as L', is either 10 or 12. For h 3 (n] to be optimal in the
mininlax sense, it must have at least L' + 2 alternations. Tbns, hs(n] must abihit at least 12 alter-
nations, for the non-extraripple case, or at least 14 alternations in the extraripple case to be optimal.
A simple counting ofthe alternations in H.(eJ"') reveals that there are 11 alternations, consisting
of the 8 alternations that were in H(eJ"') plus 3 where H(ei"') =
0. These are too few to satisfy
either the non~raripple case or the extraripple case. As a result, this filter is not optimal in the
mmwmax JeDSe.
(d)
O$w $w,
w. :S w :S 1r
O$w$w,
w. ::; w::; 7(
(e) hs(n] is h(n] upsampled by a factor of 2. In tbe frequency domain, upsampllng by a factor of 2 will
cause the frequency axis to get scaled by a factor of 1/2. Consequently, Hs(&w) will be a bandstop
filter that satisfies the alternation theorem, with twice as many alternations as H (&w). This filter
is optimal in the minimax sense.
0$ w $ w,/2
w,/2 $ w 5 "-w,/2
.- -w,/2 $ w $.-
0 5w 5 w,/2
w,/2$w$ .. -w,/2
.- -w,/2 $ w $.-
7 .56. We have an odd length causal linear phase filter with values from n = 0, ... , 24. It must therefore be
either a Type I or Type ill filter.
(a) True. We know either '
Type I Type ill
h(m] = h(24- m] or h(m] = -h(24- m]
for -oo < m < oo since the filter has linear phase. Substituting m = n + 12 we get
h(n + 12] = h(12- n] or h(n + 12] = -h(l2- n]
(b) Fwe. Since the filter is linear phase it either has zeros both inside and outside the unit circle or
it has zeros only on the unit circle.
If the filter has zeros both inside and outside the unit circle, its inverse has poles both inside and
outside the unit circle. The only region of convergence that would correspond to a stable inverse
would be the ring that includes the unit circle. The inverse would therefore be two-sided and not
causal.
If the filter only has zeros on the unit circle, its inverse has poles on the unit circle and is therefore
unstable.
(c) Insufficient /nformatioTL If it is a Type ill filter it would have a zero at z = -1 but if it is a Type
I filter this is not necessarily true.
(d) True. To minimize the maximum weighted approximation error is the goal of the Parks-McClellan
algorithm.
(e) True. The filter is Fm so there are no feedback paths in the signal flow graph.
(f) True. The filter has linear phase and
arg (H(eJw)) ={J- 12w
where {J = 0," for a Type I filter or {J = ../2, 3r/2 for a Type ill filter. Tbe group delay is
(b)
W(w) = !· ~
j;
(or~)
(or 1)
(or f.)
~:: :~
·(or 1)
O:S!wi:Swl
.... :Siwl $ "'>
"'•:Siwi:S ..
(c) From the Alternation Theorem, the minimum number of alternations is L + 2.
(d) The trigonometric polynomial (of degree L) can have at most L- 1 points of local minima or
maxima in the open interval between 0 and ... If these are all alternation points and, in addition,
all the band edges are alternatinn points, we find the maximum number of alternations is
L-1+6=L+5
(e) If M = 14, then L = M/2 = 7. The maximum number of alternations is therefore 7 + 5 = 12.
E(oo) o
0 ., .,
(f) ~ will be shown in part (g), the 3 band case can have maxima and minima in the transition
regions. It follows that we do not have to have an extremal frequency at w4 • Therefore, if we
started with an optin>al maximal ripple filter and jnst slid "'• over we may move a local minin>um
or maximum into the transition region, bnt there will still be enongh alternations left to satisfy
the alternation theorem. Thus, the maximum approximation error does not have to decrease.
(g) (i) If a point in the transition region has a local minimum or maximum then there is the possibility
that the surrounding points of maximum error do not alternate. Thus, we might lower the
number of alternations by two. B~, if...., started with L + 5 alternations this reduction
does not drop the number of alternations below the lower limit of L + 2 set by' the Alternation
Theorem. Therefore, local maxima and minin3a of A.(eiw) can occur in the transition regions.
Note that this is not true in the 2 band case.
(ii) If a point in the approximation bands is a local minimum or maximum, the surrounding points
of maximum error do not alternate. Thus, a local minimum or maximum in the approximation
bands implies that the total number of alternations is reduced by two. However, if we started
with L + 5 alternations this reduction does not drop the number of alternations below the
lower limit of L + 2 set by the Alternation Theorem. Therefore, we can have a local maximum
or minimum in the approximation bands. Note that in the 2-band case we drop &om L + 3 to
L + 1 which violates the Alternation Theorem.
283
7.58. (a) In order for condition 3 to hold, G(z- 1) must he an allpass system, since
z-t = G(z- 1)
.-jB = G(e-"")
= IG<·-jW)jeJLG(•-'"l
Clearly, !G(e-1wJI must equal unity to map the unit circle of the Z-plane onto the unit circle of
the z-plane.
(b) Consider one allpass term in the product, and note that a• is real.
0 $IZI <I
Or equivalently,
I<IZ-'I<oc
Substituting the allpass term for z-t gives
I < It~,a~;:, I
(I- a,z- 1)(1- a,z"- 1) < (z- 1 - a,)(z•-l- a,)
1- okz- 1 - aa:z•-l + o~z- 1 z•-l < z- 1z•-l- OA:.:z•-l- ocz- 1 + ai
(1- ail < z- 1 z"- 1 (1- ail
If (I - aiJ < 0, then
z-1 z•-1
1 >
I
I >
lzi'
1•1 > I
The inside of the unit circle of the Z-plane maps to the outside of the unit circle of the z-plane.
This is not the desired result. However, if (1- ail > 0, then
I < z-lz•-1
I
1 <
lzi'
1•1 < 1
The inside of the unit circle of the Z-plane maps to the inside of the unit circle of the z-plane. This
is the desired result. Thus, for condition 2 to be satisfied,
1- ai > 0
ja,j 2< 1
Ia•! < 1
This condition holds for the general case as well since the general case is just a product of the
simpler allpass terms.
!
284
(c) First, it is shown that G(z- 1) produces the desired mapping for some value of a. Starting with
G(z- 1 },
z-1 -a
z-• = 1-az I
e-jf.J -a
e-;• = 1- oe-Jw
e-il- ae-;•e-;w e-;.., -a
=
e-.;..(1 + ae-i1 ) = e-;• + 0
e-;161
e-i• +a
= 1 + ae-i'
e-il +a 1 +aei'
= 1 + ae ;• 1 + aei1
,-;• + 2a + a 2 ei'
= 1+2acos8+a2
Using Euler's formula,
2
-w = tan-• [ (a - 1} sinD ]
2a+(1 +a2 )cos8
2
w = tan-' [ (1- a ) sinO ]
2a+(1+a2 )cos8
This relationship is plotted in the figure below for dilferent values of a. Although a warping of the
frequency scale is evident in the figure, (except when a = 0, which corresponds to z-l = z-l }, if
the original system has a piecewise-constant lowpass frequency response with cutoff frequency e,,
then the transformed system will likewise have a similar lowpass response with cutoff frequency "'•
determined by the choice of a.
Next, an equation for a is found in terms of e. and "'•· Starting with G(z- 1
),
z-1-a
z-1 =
1-az 1
e-;w,.- o
1 ae-iw,
e-;w,.- Q
e-il, - e-;w,. = o(e-i(B,.~,.) - 1)
e-il,. _ e-;w,.
" = e i(l,.+w,.) - 1
e-i(l,.+w,.)/2(e-i(l,.-w,.)/2 _ ei(B,.-w,.)/2)
= e j(l,.+~,o~,.)f2(e j{B,.+w,.)/2 _ ei(B,.+w,.}/2)
(i}
2
-1 [1- (-0.2679) ]
=
"'• tan 2( -0.2679}
-1 ( 0.9282 )
= tan -0.5358
= 2tr/3
(ii)
= 1f/2
(iii)
2
-1 [1- (0.4142) ]
"'• = tan 2(0.4142)
= tan- 1 (1)
= 1f/4
(e) The first-order allpass system
G(z-1) =- z-1 +a
1 +az- 1
satisfies the criteria that the unit circle in the Z-plane maps to the unit circle in the z-plane, and
that e = 0 maps tow="· Next, a is found in terms of e, and "'•·
z- 1 +a
z-1 = 1 + QZ l
286
e-iw,. +0
= 1 + ae-Jw.,
-e-;B, - oe-i(w,.+l)
a(1 + e-i(w,+B,)) = -e-;1,. _ e-;w,.
e-il, + e-iw,.
a = 1 + e j(w,.+l,.)
e-i(w,.+l,.)/2(e-i(-w,.+l,.}/2 + e-j(w,.-1.. )/2)
= e-i(w,.+l,)/2(ei("',.+l,.}/2 + e-i(~oo~,.+l,.)/2)
cos[(wp- Bp)/2]
= cos[(wp + Bp)/2]
z- 1 +a
z-• = 1 +az 1
e-Jw + Q
= 1 +ae ;w
-e-i' - ae-i(w+B) = e-jw +Q
.-iw(l + ae-i') = -e-i'- a
e-il +0
= 1 + ae-1 8
e-il +o 1 + oei'
= 1 + ae-i' 1 + oei8
.-jl + 2a + ci'ei'
= 1 + 2acos8 + a 2
cos8- j sin8 + 2a + a 2 ccs8 + ja 2 sinS
= 1 + 2acosS + a 2
cosS + 2a + ci' cosS + j(- sinS+ a 2 sinS)
= 1 + 2acos8 +a>
Therefore,
2
(a -1)sin8 ]
-w + 7r = tan _ 1 [ 2a+(1+a 2 )cos8
w = tan-• [ (1-ci')sin8 ] +w
2a + (1 + a 2 )cos8
Note that this lowpass to highpass expression is the similar to the lowpass to lowpass expression
for w found in part (c). The only dilference is the additive " term, which shifts the lowpass filter
into a highpass filter. The frequency warping is plotted below.
287
(i)
_1 [I- (-0.2679) 2 ]
w, = tan 2( -0.2679) +"
-1 ( 0.9282 )
= tan -0.5358 + 71"
= 2tr/3+tr
= Sr/3
The right edge of the low pass filter gets warped to 57f /3, which is equivalent to -1r /3. The
frequency response of this filter appears below.
... ...
-ltl3 0
"
-It (I)
(ii)
w, = -1 r~- (O)'J
tan 2(0) + ..
= tan- 1 (oo) + .-
= .-f2+tr
= 37r/2
The right edge of the low pass filter gets warped to 3tr /2, which is eqaiYaleot to -ir /2. The
frequency response of this filter appears below.
288
••• • ••
(iii)
The right edge ofthe low pass filter gets warped to 57</4, which is equivalent to -3>rf4. The
frequency response of this filter appears below.
••• •••
Solutions - Chapter 8
X[k] = t (t
n=O tn=-8
a,.ei'fmn) ,-;'f•n
= L• L• a,.ei(2•/l)(m-•)n
A=Om=-t
We reverse the order of the summations, and use the orthognnality relationship from Example 8.1:
• 00
X[k]=6 La,. L
m=-1 r=-oc
o[m-k+rNJ
Taking the infinite summation to the outside, we recognize the convolution between a, and shifted
impulses (Recall a,. = 0 for JmJ > 9). Thus,
00
X[k] =6 L "•-•·
=-oo
Note that from X[k], the aliasing is apparent. ·
292
= (l+e-;b(ll+e-jb("fl)X[k]
= { 3X[k/3J, k = 3l
0, otherwise
(b) Using N = 2 and %[n] as in Fig P8.2-l:
N-1
X[kj = L f(n]W~n
= %(0] + %(1]e-;..
= 1+2(-1).
= {3, -1,
k=O
k = 1
Observe, from Fig. P8.2-1, that i(n] is also periodic with period 3N = 6:
3N-1
X3 (k] = L i(n]w;.-:
n=G
= L• %(n]t-jfk
n=G
= (1 + .-;'f• +e-i¥>)(1 + 2(-l)l)
= (1 + .-;'f• + .-;'f•)X[k/3]
= ( ~3. ::~
0, lc = 1,2,4,5.
293
8.3. (a) The DFS coefficients will be real if i[n] is even. Only signal B can be even (i.e., is[n] = %8 [-n];
if the origin is selected as the midpoint of either the nonzero block, or the zero block).
(b) The DFS coefficients will be imaginary if x[n] is even. None of the sequences in Fig P8.3-1 can be
odd.
(c) We use the analysis equation, Eq. (8.11) and the closed form expression for a geometric series.
Assuming unit amplitudes and discarding DFS points which are zero:
•
XA[k] = L;ei'fkn
=<>
1- eif.U:
= 1- eit•
1-(-1)•
= .••
1- ....
= 0, k = ±2, ±4, ...
2
Xs[kj = L;ei'f•·
=<>
1- eJtlk
= 1- eit.t
3
Xc[kj = :E e~>t>n- :E 7ei>t•·
n=O
3
=•
= L (eit•• _ e~f•ln+4))
n=O
= (1 - ei.. ) 1 - ..,:•
1-eJ'i.C
= 0, k = ±2, ±4,
8. 4. A periodic sequence is constructed from the sequence:
as follows:
~
=
..
L:a"e-;..,n
..... 1
= 1- ae-P.,' lal < 1
N-1 oo
= L L z{n+rNJW~n
N-l oo
= L L <>n+rNu{n+rNJW~n
n=O r=-oo
N-1 oo
= LLQn+rNw~
ft=O r=O
Rearranging the summations gives:
,lal < 1
,lal < 1
8.5. (a)
z[n] = o[n]
N-1
X[kj L o[n]W~".
= n=O 0 ::; k ::; (N - 1)
= 1
(b)
z[n] = c5[n- no], o::; no :5 (N -1)
N-1
X{k] = L o(n - no]W~n' 0$ k $ (N -1)
n=O
= w•no
N
(c)
{ 1, n even
z[n] = 0, nodd
N-1
X[kj = Ezfn]W~n, 0$ k $ (N -1)
n=O
(N/2)-1
= E
n=O
W]./'"'
1- e-i2•"
= 1- e j(••fN)
{ N/2, k=O,N/2
X[kj = 0, otherwise
295
(d)
l
1- e-id
= 1- e i(2d)fN
k=O
N/2,
X[k] = 1- e ~2•k/NJ' k odd
0, k even, 0::; k 5 (N- 1)
(e)
{an, 0 ::; n $ (N - 1)
:r:[n] = 0, otherwise
N-1
X[k] = L anwfy", O$k$(N-1)
n=<l
1- aNe-i2d
= 1 _ ae-i(2d)/N
1-aN
X[k] = 1_ ae-iC21fk)/N
X(ei"') = L z[n]e-;wn
n.=-oc
N-1
= L eit.~one-jwn
n=<>
1 _ e-i{w-wo)N
X(ei"') = 1- e j(w We)
= . ((2••
-;('N'--.... )(!!?)sm N" -Wo ) T
"]
e sin[(~- w,) /2]
Note that X[k] = X(&w)lw=(2d)/N
(c) Suppose Wo = (2do)/N, where ko is an integer:
1- e-i(A:-to)b
X[k] = 1_ e-i(•-.. )(2•)/N
8. 7. We have a six-point uniform sequence, :z:[n], which is nonzero for 0 :S n :S 5. We sample the Z-transform
of :z:[n] at four equally-spaced points on the unit circle.
X[k] = X(z)I,=•"·•J•>
We seek the sequence :1 [n] which is the inverse DFT of X[k]. Recall the definition of the Z-transform:
00
X(z) = L :z:[n]z-n
n=-oo
Since x[n] is zero for all n outside 0 :S n :S 5, we may replace the infinite summation with a finite
summation. Funhermore, after substituting z = ei( 2 ·d"/ 4 ), we obtain
X[k] =L
• :z:[nJWt", 0 :S k :S 4
n=O
Note that we have taken a 4-point DFT, as specified by the sampling of the Z-transform; however, the
original sequence was of length 6. As a result, we can expect some aliasing when we return to the time
domain via the inverse DFT.
Performing the DFT,
Taking the inverse DFT by inspection, we note that there are six impulses (one for each value of n
above). However,
W4•• -- ......
"'"
and w•• - w•
4. - 4'
so two points are aliased. Tbe resulting time-domain signal is
2 2
:z:,[n]
1 1
•
-1 0
l l
1
II
2 3
•4 •5 •6 n
297
X(eJw) = L z[n]e->wn
n=-CICI
= ~G)" .-jwn
1
=
Y[k] = X(eJw)iw=2d/1D• 0$ k $ 9
= L• y[n1w,•;
Recall:
(a) We wish to obtain X(&w)lw=h/• using the smallest DFT possible. A possible size of the DFT is
evident by the periodicity of eJW lw=h/•· Suppose we choose the size of the OFT to be M = 5.
The data sequence is 20 points long, so we use the time-aliasing technique derived in the previous
problem. Specifically, we alias z[n] as:
00
This aliased version of z[n] is periodic with period 5 now. The 5-pt DFT is computed. The desired
value occun at a frequency corresponding to:
2.-k 4...
N=s
For N = 5, k = 2, so the desired value may be obtained as X[k]l•= 2 .
(b) Next, we wish to obtain X(eJw)iw=l0./27·
The smallest DFT is of size L = 27. Since the DFT is larger than the data block size, we pad z[n]
with 7 zeros as follows:
z[n], 0 ::; n ::; 19
%2 [n J
{ 0,
= 20 ::; n ::; 26
We take the 27-pt DFT, and the desired value corresponds to X[kJ evaluated at k = 5.
298
8.10. From Fig P8.10-1, the two 8-pt sequences are related through a circular shift. Specifically,
8.11. We wish to perform the circular convolution between two 6-pt sequences. Since .,,[n] is just a shifted
impulse, the circular-convolution coincides with a circular shift of :~: 1 [n] by two points.
y[n] = :~:.[nJ®:t,[nj
= :~:.[nJ(§)[n - 2]
= :~: 1 [((n- 2)),]
.i); f
4 y[n]
8.12. (a)
-1 0 I 2 3
f
4
I
5
I
6
I
7
n
transforms to
3
X[kJ=:L;cos<";'lWf", 0SkS3
.....
The cosine term contributes only two non-zero values to the summation, giving:
(b)
X[kj =
-
-
1 - .-id,
1- w•• . 0SkS3
h[n] = 2", 0S nS 3
3
H[kj. = L2"Wf", 0SkS3
.....
= 1+2Wt+4Wf" +8Wf"
299
(c) Remember, circular convolution equals linear convolution plus aliasing. We need N ~ 3+4 -1 = 6
to avoid aliasing. Since N =
4,we expect to get aliasing here. First, find 11[n] z[n]• h[n]: =
6
y[n] = z[n]• h[n]
3
2
; T n
-1 0 1 2 3 4 5 6 7
-4
-8
For this problem, aliasing means the last three points (n = 4, 5, 6) will wrap-around on top of the
first three points, giving y[n] = z[nJ<Dh[n]:
6
y[n] = z[n]@)h[n]
3
n
-1 0 1 2 3 4 5
-3
-6
(d) Using the DFT values we calculated in parts (a) and (b):
Y[k] = X[k]H[k]
= 1 + 2Wf + 4w:• + BW]k - w:•- 2W]• - 4Wt' - 8W;'
Since Wt' = K1• and W;' = Wf
Y[k] = -3 - 6Wf + 3w;• + sw:•, o :s: k :s: 3
8.13. Using the properties of the OFT, we get y(n] = :z:(((n- 2)).], that is y(n] is equal to :z:[n] circularly
shifted by 2. We get:
2 2
y[n]
l.lii.
012345
n
8.14. :r,[n] is the linear convolution of :z: 1 (n] and :z:2 (n] time-aliased to N = 8. Carrying out the 8-point
circular convolution, we get:
9 9 9 9
8 8
7 7
i
0 1 2 3 4
i
5 6 7 8
n
1 24+ 1 a 2+a
y(n]
-1
•
0
llll
1 2 3 4
Matching the above oequence to the one given, we get a = -1, which is unique.
• •5 n
8.16. X,[k] is the 4-point OFT of :r[n] and :z:,[n] is the 4-point inverse OFT of X 1 [k], therefore :r 1 (n] is :r(n]
time alia.ed to N = =
4. In other words, :z:,[n] is one period of :f(n] :r[((n)) 4 ]. We thus have:
4 = b+ 1.
Therefore, b = 3. This is clearly unique.
8.17. Loo~ at the oequences, we see that :r, [n] • :r,[n] is non-zero for 1 $ n $ 8. The smallest N such that
.:r 1 [n]~ 2 (n] = :r,[n] o:r2 [n] is therefore N = 9.
301
8.18. Taking the inverse DFT of X 1 [k] and using the properties of the DFT, we get:
= z[((n- m)),].
z 1 [n]
m = 2 works, clearly this choice is not unique, any m = 2 + 61 , where I is an integer, would work.
8.20.
X, (k) = X(kje+j(2d2/NI.
Since f 2 [n] is a periodic in2pulse, shifted by two, the resultant sequence will be a shlfted (by two).
replica of i ,[n].
6 6
Using the analysis equation of Eq. (8.11), we may rigorously derive y,[n]:
X,[k]
•
= Lf,[n]W{"
n=<>
302
Xs[k] = L• ia[n]Wf"
.....
= 1+Wt•
Therefore:
Y,[kJ = .X,[k].X.[kJ
= .X,[kJ + w:> X,[kJ
Since the DFS is linear, the inverse DFS of Y2 [k] is given by:
!i2[nj = :t,[n] + %1[n - 4].
8.22. For a finite-length sequence z[n], with length equal toN, the periodic repetition of z[-n] is represented
by
z[((-n))N) =z[((-n+lN))N], l: integer
where the right side is justified since z[n) (and z[-n]) is periodic with period N.
The above statement holds true for any choice oft. Therefore, for l = 1:
:r:[n]
--~~·IL-__N_~_t_o_FT -J~---. __
X[k] (O $ k $ N)
303
(b) Suppose N > P, consider taking a DFT which is smaller than.the data block. Of course, some
aliasing is expected. Perhaps we could introduce time aliasing to offset the effects.
Consider the N-pt inverse DFT of X[k],
N-1
z[n] = N1 L X[k]W,V"", 0 ~ n ~ (N- 1)
lo=O
Suppose X[k] was obtained as the result of an infinite summation of complex exponents:
z(nj =~ ~
k=O
(f: m=O
z(mje-;(2d/N)m) W,V""
Rearrange to get:
z[n] = L
oo
z[m] (
~ L ,-;(2•/N)(m-n)i
N-1 )
m=-oc k=O
Using the orthogonality relationship of Example 8.1:
00 00
So, we should alias z[n] as above. Then we take the N-pt DFT to get X[k].
8.24. No. Recall that the DFT merely samples the frequency spectra. Therefore, the fact the /m{X[k]} 0 =
for 0 ~ k ~ (N - 1) does not guarantee that the imaginary part of the continuous frequency spectra is
also zero.
For example, consider a signal which consists of an impulse centered at n = 1.
z[n] ;, o[n- 1], 0~ n ~ 1
The Fourier transform is:
X(eiw) = ,-;w
Re{X(eiw)} = cos(w)
/m{X(eiw)} = -sin(w)
Note that neither is zero for ali 0 ~ w ~ 2. Now, suppose we take the 2-pt DFT:
X[kJ = w;, o ~ k ~ 1
= { 1, k=O
-1, k=1
1 1 1
I I rt I I rr I I r r
3/4 3/4 3/4
1/2 %(n]
n
-4 -3 -2 -1 0 1 2 3 4 5 6 7
Now, we shift by three (to the right), and set all values outside 0 S n S 3 to zero.
I 1
--~·~·~~I~t~2-~L~4~~·~··------~~~J
3/4
-2 -1 0 1 2 3 4 5
8.26. (a) When multiplying the DIT of a sequence by a complex exponential, the time-domain signal
undergoes a circular shift.
For this case,
Y[k] = W:' X[k], 0SkS5
Therefore,
y[n] = z[((n- 4))o], OSnSS
4
3
II • •
2 y[n]
1
•
-1 0
I t
1
•2 •3 4 5 6 7
n
(b) There are two ways to approach this problem. First, we attempt a solution by brute force.
Notice that
w~ = e-·;(2·lll:fN}
w-•
N = eJ(2d/N) = e-i(2r/N)(N-t) = w::-t
W[k] = 4+~
2' [w.• + w!-•] + [w•• + w•-••] + !2[w.,.
6 6 ' + w•-••]
6 6'
So,
Sketching w[n):
4
I
w[n)
3/2 1 1 1 3/2
• TT T T T• • n
-1 0 1 2 3 4 5 6 7
As an alternate approach, suppose we use the properties of the DIT as listed in Table 8.2.
W[k] = Xe{X[k)}
X[k] + X"[k)
= 2
w[n) = ~ IDFT{X[k)} + ~ IDIT{X"[k)}
= ~ ( :r[n) + x"[((-n))NJ)
For 0 :5 n :5 N - 1 and :r[n) real:
•
-1
I• • ' I I • •
0 1 2
1
3
2
4 5 6
:r[N- n], for N
7
n
=6
.li i. ..
-1 0 I 2 3 4 5
q[n]
8.27. (a) The linear convolution, z 1 [n] • z 2 [n] is a sequence of length 100 + 10- I = 109.
10 10
9 9
8 8
7 7 zl[n] • z,[n]
6 6
5 5
4 4
3 3
2 2
; T T; n
-1 0 I 2 3 4 5 6 7 8 9 99 100 101 102 103 104 105 106 107 108 109
(b) The circular convolution, z 1 (n] ~ 2 [n], can be obtained by aliasing the first 9 points of the linear
convolution above:
10 10 10 10 10 10 10 10 10 10 10
llllllllll L.
01234 56789 99
(c) Since N? 109, the circular convolution zl[nJ@,[n] will be equivalent to the linear convolution
of part (a).
8.28. We may approach this problem in two ways. First, the notion of modulo arithmetic may be simplified
if we utilize the implied periodic extension. That is, we redraw the original signal as if it were periodic
=
with period N 4. A few periods are sufficient:
i j i f i i i f ii i
i(n]
-4 -3 -2 -1 0 1 2 3 4 5 6
f7
n
307
To obtain z 1 [n] = z[((n- 2)) 4 ], we shift by two (to the right) and only keep those points which lie in
the original domain of the signal (i.e. 0 $ n $ 3):
• •
-2 -1 0
ii li 1 2 3
•4 •5
z 1 [n]
To obtain :r 2[n] = z[(( -n)) 4 ], we fold the pseudo-periodic version of :r[n] over the origin (time-reversal),
and again we set all points outside 0 $ n $ 3 equal to zero. Hence,
.. !iii.
-2 -1 0 1 2 3
I
4 5
:r2[n]
8.29. Circular convolution equals linear convolution plus aliasing. First, we find y[n] =:r,[n] • :r2[n]:
8
6 6
5
4 4 y[n]
3 3
2
T~ n
01234567891011
8
10 6
8
6 II• (n]
4
-1 0 1 2 3 4 5 6 7
"
(b) For N = 10, N ;::: 6 + 5- 1, so no aliasing occurs, and circular convolution is identical to linear
convolution.
8.30. We have a finite length sequence, whose 64-pt DFT contains only one nonzero point (for k = 32).
(a) Using the synthesis equation Eq. (8.68):
N-1
:r:[n] = N
1
L X[k]W,V"". 0 :0: n :0: (N- 1)
1:=0
Substitution yields:
32
:r:[n] = ! X(32JW6< "
.!_,.;j<(32)n
=
64
= _!_eJ .. n
64
:r:[n] = !<-1)", 0:0: n :0: (N -1)
The answer is unique because we have taken the 64-pt DFT of a 64-pt sequence.
(b) The sequence length is now N = 192.
:r:(n]
:r:[n] =
64 :0: " :0: 191
This solution is not unique. By taking only 64 spectral samples, :r(n] will be aliased in tin2e.
As an alternate sequence, consider
8.31. We have a 1~point sequence, :r[n]. We want a modified sequence, :r1 (n], such that the 1~pt. DFT of
:ri[n] corresponds to
X;(k] = X(z)l.=!•""••t••>+<•t>•ll
Recall the definition of the Z-transform of :r[n]:
00
X(z) = L :r[n]z-•
n.=-oc
309
Therefore,
X(z) = L• z[n]z-•
n=O
Substituting in z =!&1<•••/IO)+(•/IO)J:
9 ( 1 . (2. . /10)+(•/10)) )-n
X(z)],=!•W'"'"'>+<•I">l = L:;z[n] 2e'l
n=O
We seek the sigual z 1 {n], whose 1!)-pt. DFT is equivalent to the above expression. RecaJ.J the analysis
equation for the DFT:
X,{k)
•
= L;z.{n]Wt.;', 0$ k $9
n=O
Since wton = e-i( 2 " /lD)A:n' by comparison
= L;z{n]Wff"
n=O
7
Y{k] = L;z{n]Wf", 0$ k $ 15
n=O
Therefore, the 16-pt. DFT of the interpolated sigual contains two copies of the 8-pt. DFT of z{n]. This
is expected since Y{k] is now periodic with period 8 (see problem 8.1). Therefore, the correct choice is
c.
As a quick check, Y[OJ = X{O].
310
1 Scale by -1 1
I Shift by N I
L--{~co~n~ca~t~e~na~t~e}-.......,[~2~N~-p~t~D~FT~}--- X2[k]
(b) To obtain X[k] from X,[k], we might try to take the inverse DFT (2N-pt) of X,[k), then take the
N-pt DFT of :r,[n] to get X[k].
However, the above approach is highly inefficient. A more reasonable approach may be achieved if
we examine the DFT analysis equations involved. First,
2N-1
X,[k) = L :r,[n)W;'N, 0 :S k :S (2N - 1)
n=O
N-1
= L :r[n)W;';)
n=O
N-1
= L :r[n)w;:f2)n, 0 :S k :S (N- 1)
n=O
X,[k) = X[k/2), 0 :S k :S (N- 1)
Thus, an easier way to obtain X[k] from X1[k] is simply to decimate X1[k] by two.
- { :[n]+t[-nJ, (-N+1)~n~(N-1)
:z:.(n] -
otherwise
~ :z:[-n]W:._ ~ :z:[n]w:>n
= L..J - 2 - 2N-1 + L..J 2 2N-1
n=-N+l n=O
Letm= -n,
N-1 ( ) N-1 ( )
x.[kJ = "L..J !!'_
2 w:-l:n
2N-1 L..J ~
+ "'"" 2 w,kn
2N-1
n=O n.=O
Reeall
N-1
X[k] = L :z:[n]W~", 0 :S k :S (N -1)
n=O
and
Re{X[k]} = N-1 (2 k )
~ :z:[n] cos ';.. n
X[kJ + x·[kJ
Re{X[k]} = 2
N-1 N-1
1
2L 2L
1
= :z:(n]W~" + :z:(n]WN'"
n=O n=O
N-1
= ~L (:z:[n] + :z:[N- n])W~"
n=O
So,
Re{X(k]} =DFT { ~(:z:[n] + :z:[N- n])}
8.35. From condition 1, we can determine that the sequence is of finite length (N = 5). Given:
X(~"') = 1 +A, cosw + A2 cos2w
= 1 + A,(~"+ e-'"'l + A• (e'..., + e-i'"')
2 2
From the Fourier analysis equation, we can see by makhing terms tbat:
A1
:z:[n] = 6[n] + T(o[n - 1] + 6(n + l J) + 2(6[n-
A2 2] + 6(n + 2]}
312
Condition 2 yields one of the values for the amplitude constants of condition 1. Since z[n] • cS[n - 3] =
=
:r[n- 3] 5 for n = = =
2, we know z[-1] 5, and also that z[1] z[-1] 5. Knowing both these values =
tells us that A1 = 10.
For condition 3, we perform a circular convolution between z[((n-3))s] and w[n], a three-point sequence.
For this case, linear convolution is the same as circular convolution since N = 8 ?: 6 + 3 - 1.
We know z[((n- 3))s] = z[n- 3], and convolving this with w[n] from Fig P8.35-1 gives:
22
A2 + 15
lli
2
+ 11
1>- + 13
5+A>
lli
1>-
J
0 '
1 2 3 4 5 6 7 8
n
For n = 2, w[n] • z[n- 3] = 11 so A2 = 6. Thus, z[2] = z[-2] = 3, and we have fully specified z[n]:
5 5
l lI
3 3
z[n]
•
-3
I
-2 -1
T
1
0 1 2
I
3
n
X[k] = 2 + Wt + Wi', 0 :0 k :0 5
where w; = .-;('fl•.
(ii) Now, we square the DFT of z{n]:
Y[.I:J = X 2 [.1:J
= 2 + 2Wt +2Wff•
+2Wf + W;"+ wr
+ 2w1• + w:• + wt•. o ::; .1: ::; 5
(a) By inspection,
(b) This procedure performs the autocorrelation of a real sequence. Using tbe properties of tbe DFT,
an alternative method may be achieved with convolution:
8.37. (a)
I'l-l
G,[k] = L z[N- 1- n]W~n, 0 :S k :S (N- 1)
n=O
Letm=N-1-n,
/'1-1
G,[k] = L z[m]W~(/'1-l-m)
m=O
/'1-1
= w~<N-l) I: z[mJw,v•m
m=O
/'1-1
G,[k] = ei(2d/l'l) L z[m]ei(2dm/l'l)
m=O
= ei(>d/1'1 X(ei'"')l..,z(b>/1'1)
G,[kJ = H1[k]
(b)
1'1-1
G2[k] = L (-1)"z[n]W;', 0 :S k :S (N- 1)
n=O
1'1-1
= L z[n]w~fl·w~·
n=O
1'1-1
= L z[n]w:.,'+fl•
n=O
= X(ei'"')l..,=b(Hf)/N
q,[k] = He[k]
314
(c)
2N-1
G3(k] L :(n]W2~•
= ,..,.., 0$ k $ (N -I)
N-1 2N-l
L :fn]w;~ +
= ,..,.., L :(n - NJWt.V
=N
N-l N-l
= L :fn]Wf.Y + L :fm]W;~m+Nl
n=O m=O
N-l
z[n]+z[n+N/2], O$n$(N/2-I)
94 [nl
·
={ 0, otherwise
G4 [k] = 0$ k :S (N- I)
N/2-1 N/2-1
= L
n=O
z[nJWt/2 + L
n=O
z[n+N/2]W,t/2
N/2-1 N-1
= L z[nJWti 2 + L z[m]W~j';-N/ 2 )
A=O m=N/2
N-l
= L z[n]W~tn
n=O
= X(eiw)lw=(bt/N)
G4 [k] = Ha[k]
(e)
{ :[n], 0 :S n $ (N -I)
gs[n] = 0, N $ n $ (2N -I)
0, otherwise
315
2/11-1
Go[k] = L :[nJW:N, 0 ~ k ~ (N -1)
.....
N-1
L x[n]W,I'N
= .....,
= X(eiw)[w=(d/N)
Go(k] = H2[k]
(f)
2N-l
Go[k] = L :r[n/2]WfN, 0 ~ k ~ (N -1)
n:O
N-1
= L :r[n]W~n
n=O
= X (eiw)lw=(2d/N)
Go[k] = H1[k]
(g)
f-1
G1[k] = L 0 ~ k ~ (N- 1)
..... x[2nJWM2,
= ~ :r[n] c ~1)n) w~n
+
N-1 ( 1 + W(N/2)n)
= L:x[n] ; w~n
n=O
N-1
= ~ ~ :r[n] (w;:• + w~•+Nt•l)
= ~ [x<ei<••tNlJ + x(ei<••tNH>+Nt•lJ]
G1[k] = Ho[k]
8.38. From Table 8.2, the N-pt OFT of an N-pt sequence will be real-valued if
::r[1] = ::r[9]
::r[2] = ::r[S]
The Fourier transform of ::r[n] displays generalized linear phase (see Section 5.7.2). This implies that for
::r[n] .,< 0, 0 ::; n ::; (N- 1):
::r[n] = ::r[N - 1 - n]
For N = 10,
::r[O] = ::r[9]
::r[1] = ::r[S]
::r[2] = ::r[7]
8.39. We have two 100-pt sequences which are nonzero for the interval 0::; n ::; 99.
If x, [n] is nonzero for 10::; n S 39 only, the linear convolution
::r,[n] • :r2[n]
is a sequence of length 40 + 100- 1 = 139, which is nonzero for the range 10 ::; n S 139.
A 100-pt circular convolution is equivalent to the linear convolution with the first 40 points aliased by
the values in the range 100 ::; n $ 139.
Therefore, the 100-pt circular convolution will be equivalent to the linear convolution only in the range
40::; n ::; 99.
8.40. (a) Since ::r[n] is 50 points long, and h[n] is 10 points long, the linear convolution y[n] = :r[n] • h[n]
=
must be 50 + 10 - 1 59 pts long.
(b) Circular convolution= linear convolutin +aliasing.
If we let y[n] = ::rjn] • h[n], a more mathematical statement of the above is given by
00
For N =50,
::r[nJ@.[n] =11[;.] + y[n +50],
We are given: :r[n] @>.In] =10
Hence,
y[n] + y[n + 50] = 10, 0 ::; n ::; 49
11[55) = ?
n=8 y[8) + 11[58) = 10
11[58) = ?
n=9 11[9) = 10
n=49 11[49) = 10
To conclude, we can determine 11[n) for 9 S n S 55 only. (Note that 11[n) for 0 S n S 4 is given.)
8.41. We have
A B
----~------~-----------L------~--- n
0 9 30 39
n
10 19
...-...---____
A•C B•C
----L-----------~--------~----------~--- n
10 28 40 58
(b) The 40-pt circluar convolution can be obtained by aliasing the linear convolution. Specifically, we
alias the points in the range 40 $ n $ 58 to the range 0 $ n $ 18.
Since w(n] = ~(n]•y(n] is zero for 0 $ n $ 9, the circular convolution g(n] = ~[nJ@\t[n] consists
of ooly the {aliased) wlues:
Also, the points of g(n] for 18 $ n $ 39 will be equiwlent to the poi.nts of w(n] in this range.
To conclude,
w(n] = g(n], 18:$n:$39
w(n +40] - g(n], O$n:$9
8.42. (a) The two sequences are related by the circular shift:
L + 100 - 1 = L + 99
Thus, the required length is
L = 256 - 99 = 157
If we bad 63 sections, 63 x 157 = 9891, there will be a remainder of 109 points. Hence, we must
pad the remaining data to 256 and use another OFT.
Therefore, we require 64 OFTs and 64 IDFTs. Since h(n] also requires a OFT, the total:
(c) Ignoring the transients at tbe beginning and end of the direct convolution, each outpnt point
requires 100 multiplies and 99 adds.
overlap add:
# mult = 129(1024) = 132096
# add = 129(2048) = 264192
319
overlap save:
# mult = 131(1024} = 134144
#add = 131(2048) = 268288
direct convolution:
# mult = 100(10000} = 1000000
#add = 99(10000} = 990000
4 4
Q[k] = 3"[k] + 3"fk - 3].
l.l.I.
012345
n
320
8.45. We have:
Then:
1 •
z,[O] = 7Lx,[kJ
k=O
1 •
= 7 })Re{Xa[k]} + jlm{X,[k]})
k=O
1 •
= 7 L Re{Xo[k]} , since z 2 [0] is real.
k=O
= g[O].
To determine the relationship between z 2 [1] and g[1], we first note that since z 2 [n] is real:
X(ei"') = X"(e-i"').
Therefore:
X[k] = X"[N- kJ, k = 0, ... ,6.
\\'e thus have:
1 •
= 7 LRe{X,[k]}w,-•
g[1]
....
= !7 t, x,[kJ +2 x;[klw:-•
= !7 L
.
k=O
1:=0
[
x, klw-•
2 7
7
+7
k=O
.
! L x,[N- klw:-•
2 7
1 1 •
= 2"''[1] + 14 L x,[kJW,'
.=0
• X,[kJW,-u
= ~z,[1] + 1~ L
.=0
1
= 2(z,[1] + za[6])
1
= (z,[1] + 0)
2
1
= 2"''[1].
8.46. (i) This corresponds to :;[n] = zj[(( ..:n))N], where N = 5. Note that this is only true for z 2[n].
(ii) X;(e'"') has linear phase corresponds to z;[n] having some internal symmetry, this is only true for
x,[n].
321
(ill) The DFT has linear phase corresponds to %;(n) (the periodic sequence obtained from z;(n]) being
symmetric, this is true for z 1 [n] and x 2 [n] only.
8.47. (a)
(b)
=•
= L x[n]z-nl , '·~··
..... ll=2e' 1
=
=•
L x(n)(2eii )-n,-j1fin
.....
=
=•
L v(n]e-;1fin.
n=O
(c)
3
w(n] = ~ L W[k]w;•n
•=<>
3
= ~ L(X(k] + X(k + 4J)e+''f•n
1:=0
3 3
= ~L X(k]e+J'f•n + ~L X(k + 4Je+''f•n
k=O lo=O
3 7
= ~ L X[k]e+i'f>n + ~ L X(k]e+i'f•n
k=O 1:=4
7
= ! LX(k]e+i'f»n
4>=0
= 2x[2n).
8.48. (a) No. x[n] only has N degrees of freedom and we have M ~ N constraints which can only be satisfied
if r[n] = 0. Specifically, we want
X(d'ii') = DFTM{r[n]} = 0.
Since M ;:: N, there is no aliasing and x[n] can be expressed as:
1 Al-l
z[n] =M
....2: X[k]Wt;, n = O, ... ,M -1 .
Where X[k] is the M-point DFT of :r[n], since X[k] = 0, we thus conclude that r[n] = 0, and
therefore the answer is NO.
(b) Here, we only need to make sure that when time-aliased to M samples, :r[n] is all zeros. For
example, let
z[n] = o[n] - o[n - 2]
then,
X(ei'f 0 ) = 1- 1 =0
X(d'f') = 1- I= 0
8
Xa[k]
4 4
k
0 1 2 5 6 7
2 4
= 1 +-~16
1_;hn2 +-r.-lll
1 _;"'-nH --C"'H
1 _;><n6 --c-U
1 .>"-'n!O
2 2 4 4
= 2..(16 + 8eitin2
16
+ seitinH- (eitin6- 4e'ifn!O)
= _!_
••
~ 11,16 (k]e''fi-n
16 L..,
&=0
16
v,.[kJ
8 8
k
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
-4 -4
(c)
0:5k:515
8
IX1o[k]l
I;>
4 4
2 2
k- r = mN, m: integer
~ L
N-1
.,Ji;(&-r)n ={ 1,
n=O 0, otherwise
(a) For k- r = mN,
.,J~(&-r)n = .,J~(mN)n
eibmn
lim
1 _ ei2-'
·~,
l-..mN 1- eJ-,r = [=:::;:] N l=mN
= N
325
Using the fact that f[r] and W,t' are periodic with period N:
-1 -1
L f[r]W,t' = L f[r + N]W~(r+N)
r=-m r=-m
Substituting l =r + N -1 JV-1
L f[r]W~' = L f[l]W,t'
r=-m l=N-m
f I:
1
X1[k] = Wf,m f[r]Wff + JVY:- f[r]W~']
LN-m r=O
JV-1
= w,tm L f[r]Wff
r=O
= Wt/"X[k]
Hence, if :ii[n] = x[n- m], then X 1[k] = w~m X[k].
8.53. (a) 1. The DFS of i"[n] is given by:
JV-1
JV 1
( ~i[n]W,.V""
)"
:Ex"[n]W~" =
n.=O
= x·[-kJ
326
= Xo[k]
(b) Consider z[n] real:
1.
Re{X[k]} = X[k] + X'[k]
2
From part (a), if f[n] is real,
So,
X[kJ = X'[-kJ
X[-kJ = X'[k]
X'[k] + X[-k]
&{X[k]} =
2
= &{X[-k]}
(i.e. the real part of X[k] is even.)
2.
X[kJ- X'[kJ
/m{X[k]} = 2
x·[-k]- x[-kJ
= 2
= /m{X[-k]}
(i.e., the imaginary part of X[k] is odd.)
3.
IX[k]l = VX[k]X'[k]
= Vx•[-A:]X[-k]
= IX[-kJI
327
Im{X[-k]})
= arctan ( ~{X(-k]}
= -d[-k]
(i.e., the angle of X[k] is odd.)
8.54. 1. Let :r[n] (0 ~ n ~ N- 1) be one period of the periodic sequence f(n]. The Fourier transform of this
periodic sequence can be expressed as:
00
X(eJ"') = L f[n]e-i"'"
=-oo
Recall the synthesis equation, Eq. (8.12):
N-1
f[n] = }_ L
X[k]W,Y•"
Nt=<>
Substitution yields:
1 0 ~ n ~ (N -1)
w[n] = { '
0, otherwise
The window has a Fourier transform:
00
W(e'"') = L w(n]e-i"'"
.... -oo
N-1
= L e-Jwn
a=O
1- e-;wN
= 1- .-,., = .-j'f ei'f - .-j'f
= .-i.,(¥Jsin(w~)
sin(~)
328
3. Since :r[n] = z[n]w[n], the Fourier transform of :r[n] can be represented by the periodic convolution
(see Eq. (8.28)).
Integration over - .. :::; 8 :::; r reduces to the summation (note the impulse train):
=
(a) Suppose :r[n] -:r[N -I-n]. For N even, all elements of :r[n] will cancel with an antisymmetric
component. For N odd, all elements have a counterpart with opposite sign. However, :r[(N -I)/2]
must also be zero.
=
Therefore, for :r[n] -:r[N -I - n], X[OJ 0. =
(b) Suppose :r[n] = :r[N - I - n] and N even.
N-1
X[N/2] = L :r[nJWJ:'12 )n
n=O
N-1
= L :r[n](-I)n
n=O
= :r(O]- :r(I] + :r[2] - :r(3] + · · · + :r(N- 2] - :r[N - I]
Therefore, X[N/2]= 0.
Note that:
Since,
andz,[n-N) = ~(z!n-N)+z"[N-nJ)
= ~ (- z"[O]o[n- N] + z"[N- nJ) 0$ n $ (N- 1)
So,
z.[n- N] = ~ (z[n- N]- z"[N- nJ)
For 0 :5 n :5 (N - 1):
0, otherwise
To conclude:
x•• [n], 0 Sn S N/2
"'••[n] n= N/2
:r.(n] = 2 •
:r;.[-n], -N/2<n$-1
"';.[-n]
n = -N/2
2
8.57.
N-1 N-1
L jz[n]l = L :r[n]z*(n]
2
n=O n.=O
= ! L X'[k]X[k]
N-1
1=0
N-1 N-1
L lz[n]l 2 = ..!_
NI=O
L IX[k]l 2
n=O
X[k] = X(e-i"')L=•••fN
= Bc;k) &!2•/N)to
A[k] = Bc;k)
2.-a
= N
"'
(b) This statement is FALSE:
Suppose z[n] = o[n] + !o[n- 1],
X[k]
X[k] = A[kje''•,
1 •
A[k] = 1 + -(-1)
2
and"' = 0
The Fourier transform of :r[n] is X(e-i"') = 1 + !ei"', which cannot be expressed in the form
X(e-'"') = B(w)e-' 0 "'.
8.59. We desire 128 samples of X(ei"')Y(ei"').
Since z[n] and y[n] are 256 points long, the linear convolution, :r[n] • y[n], will be 512 points long.
We are given a 128-pt DFT only. Therefore, we must time-alias to get 128 samples. The most ellicient
implementation is:
:r[n] J I
I I I
I .I
m
I -, IV
I
R,[k]
y[n] J I
I I
L J
Total cost = 110.
332
Because the shift is circular, the points at n = 0 and n = (N- 1) will not be correct. Therefore, only
the points in the range 1 ~ n ~ (N - 2) are valid.
8.61. (a)
N-1
X[k] = L :r[n]e-i<>•/N)••, 0 ~ k ~ (N -1)
n=O
N-l
X.,[k] = L :r[n]e-i(2d/N+r/N)n
n=O
N-l
= E z[nje-j(wn/N)e-J(2,.../N)I:n
n=O
= L :r[n]e-;(rn/N)
N-l
0 -;(2r/N)(N-l)nei(2r/N).,.
n=O
N-l
= L :r[n]ei<•n/N)ei(2r/N).,.
. n=O
= X.it[k]
So,
X.,[k] = X.it[N- (k + 1)] , for 0 ~ k ~ (N- 1) and :r[n] real.
333
(c) (i) N -k-1 is odd when k is even. H R(k] = XM(2k], we may obtain XM(k] from R[k] as follows:
G { R[l:/2], k even
(k] = R•((N- (k + 1))/2], k odd
where we note that
R•((N- (k + 1))/2] = Xit(N- (k +I)]
fork odd.
(ii)
R(kJ = x... [2kJ
(N/2)-1
= L z[n]e-i(4d/N+•/N)n
n=O
N-1
= L z[n]e-,<•n/N)e-i('Mf)
(N/2)-1 N-1
= L z[n]e-11•n/Nie -i('-'ffl' )•+ L x(n]e-i<•n/N)e -i('-'ffl')
n.=O n=N/2
(N/2)-1
= L (z[n]e-il•n/N) +z[n+N/2Je-il•n/Nie-'l•/21)e-;('Mf)
n=O
So,
r(n] = (x[n] - jx(n + N f2])e-il•n/N), 0<- n <- (N2 -1)
(d)
Since,
n-r, n>r
((n-r))N ={ -
N+n-r, n<r
n ~r
((n- r))N - (n- r) = { 0,
N, n <r
334
then
.-j(•/N){((n-r))N-(n-r)) = sgn [n _ rj = { 1, n~r
-1, n <r
and
N-1
Z3 =L z,[r)z,[((n- r))N]sgn [n- r]
~·
(e) Suppose, that for n 2: N/2:
zut[n] = z 1 [n]e-;(•n/N) = 0
"'>M[n] = z,[n]e-;<•n/N) 0 =
then the modilied circular convolution is equivalent to the modified linear convolution:
-
N-1
= L ZtM[r)z2M[n- r)
Thus,
N-1
:z:,[n) = ei(on/N) L :z:,[n]z,[n- r)e-j(•n/Nie-J(•n/N)(n-r)
r=O
N-1
= L z,[r]:z:,[n- r)e-;(r/N)(n-r)
r=<>
So,
N-1
z,[n) =L :r,[r]:z:,[n- r)e-><•fN)(n-r) = :z:1 [n) • :r2[n)
8.62. (a) We wish to compute z[n)@j.[n):
z[n). h[n) = z,[n]. h,[n] + :r,[n). h,[n). o[n- 32] + z,[n]. h,[n). o[n- 32]
+ z,[n]. h,[n]. 6[n- 32]. o[n- 32]
Let
·~1·11'"'
!11 [n] = :r,[n] • h1 [n]
!f2[n] = :r,[n) • h,[n) = z 1 [n] 2[h)
!13[n] = z,[n) • h,[n) = :r,[n] ,[n]
!lt[n] = :r 2[n) • h2[n)= :r,[n) 64 ,[n]
335
We can compute each of the above circular convolutions with two 64-pt DFI's and one 64-pt inverse
DFT.
{a) V = 49
(b) M =51
{c) The points extracted correspond to the range 49 :S n :S 99.
Distorting filter: h[n) = o[n)- !6(n- "<>)
8.64. {a) The Z-transform of h[n)
00
H{z) = L h(nV"
n=-oo
H{z) = 1- !.-...
2
The N-pt DFT of h(n]: {N = 4no)
•no-1
H(k) = L h(n]w;,:-e, o :S k :S {4no - I)
n=O
= 1-!w;no
2 ...
H(k) = 1 - !e-i(•/>l•
2
336
(b)
H;(z) = 1
+ 1,_...,,
112
(1)-no for causality
lzl > 2
g[n] = (1+116+2!6+···)o[n]+G+;2+5~2+··}[n-no]
+ G~ + + 1:24 + .. -) o[n - 2no] "" G + 1!8 + ~8 + .. -) o[n - 3no]
16 8 4 2
g[n] = 15
o[n] + o[n- no]+ o[n- 2no] + o[n- 3no]
15 15 15
(d) Indeed,
G[k]H[k] = 1, 0 $ k $ (4no- 1)
However, this relationship is only true at 4no distinct frequencies. This fact does not imply that
for all w:
(e)
N-1
Xn[k+ NJ = L i[n]HN[n(k+ N)]
n=O
We thus conclude that the DHS coefficients form a sequence that is also periodic with period N.
(b) We have:
= ~z[n]N
= z[n].
Where we have used the fact that }:~,;;;,1 HN[mk]HN[nk] = N only if ((m))N = ((n))N, otherwise
it's 0.
This completes the derivation of the DHS synthesis formula.
(c) We have:
HN[a+NJ = cos(27f'{a;N))+s"m(27r(a;N))
2•a 271'a
= cos(N + 2") + s"m( N + 21r)
2.-a) . (2.-a)
= (
cosN +smN
= HN(a].
And:
HN[a+b] = cos(2.-(';.,+b))+s"m(2.-(';.,+b))
(d) We have:
N-1
DHS(z[n- no]) = I: z[n- no]HN[nk]
-=0
N-1-no
= L z(n)HN((n +no)k)
n.=-ftcl
338
N-1-no
= L i(n](HN[nk]CN(nok] + HN(-nk]SN(nok])
n=-no
N-1-no N-1-no
= CN(nok] L i(n]HN(nk] + SN(nok] L i(n]HN[-nk]
N-1 N-1
= CN[nok] L i(n]HN[nk] + SN[nok] L i[n]HN[-nk]
=<> =<>
= CN(nok]XH[k] + SN(nok]XH(-k]
(e) We have:
N-1
DHT{i,[n]} = DHT{L :r.(m]:r,[((n- m))N]}
m=O
N-1 N-1
XHl[k] = L (L :t1[m]:r,[((n- m))N])HN[nk]
N-1 N-1
= L :t1[m] L :r,[((n- m))N]HN(nk]
N-1
= L :r.(m]DHT{:r•[((n- m))N]}
m=O
N-1
= L :r.(m](XH2[k]CN[mk] + xH.[((-k))N]SN[mk]) (using P8.65-7)
N-1 N-1
= L :r![m]XH>[k]CN(mk] + L :r![m]XH2[(( -k))N]SN[mk]
m=O
= 1:
=0
:t1(m]XH>(k](HN(mk] + HN(-mk])
2
+ 1: :r![m]XH>[((-k))N](HN[mk] - HN[-mk])
m=O
2
1 1
=
2xH.[k](XH![kJ +XH![((-k))N]) + 2xH,[((-k))N](XH!(kJ- xH1[((-k))NJl
= 21 xH1[k](XH,[kJ + xH.[((-k))Nll + 21 xH.(((-k))N](XH•[kJ- xH.[((-k))NJl
~
1
2r kn . 2" kn
= ....,
.L.. o:[n](cos( """N) - j sm( """N))
N-1
-· L o:[nj(CN[kn)- jSN(kn])
=•
then:
N-l
L o:[n)CN[knj = ~(X[kj +X((( -k))N])
=D
N-1
L o:[n)SN[knj
...., = -
2
1
j (X[k)- X[(( -k))N))
We thus get:
N-1
XH(k) = L o:[n](CN[kn) + SN(kn))
(g) We have:
N-1
Xg(kj = L o:[n](CN[kn) + SN(kn))
=0
Therefore:
N-l
L o:[n]CN[kn] =
=D
N-l
L o:[n]SN[kn] =
=D
We thus get:
N-1
X(k] = I; o:[n]e-i'iP
=D
N-1
= L o:[n](CN[knj- jSN(kn))
=D
340
Therefore:
We thus want:
= { 1 , Jwl::; ~
0 , otherwise
Since h(n] is FIR, we assume it is non-zero over 0::; n ::; N. The phase of H(e 1 ~) should he set
such that h(n] is symmetric about the center of its range, i.e. ~- Therefore, the phase of H(ei~)
should he ei "f' . So one possible H(k] may be:
that is:
1 eit•
(c) It is cheaper to implement N -point DFTs than 4N -point DFTs, therefore the implementation in
Figure P8.67-2 is usually preferable to the one in Figure P8.67-1.
8.68. Substituting the expression for X 1 (k] from equation (8.164) into equation (8.165), we get:
2N-3
z 1 (n] = _1_ " X 1(kjei2•>n/(2N-2)
2N-2L....
1=0
N-l 2N-3
= 1 ( L X<l(kjei2•>n/(2N-2) + L X<l(2N- 2- k}ei2•1m/(2N-2l)
2N - 2 .=<1 •=N
342
Note that:
2N-3 N-2
L X"[2N _ 2 _ k]ei2dn/(2N-2) = L X"[r],i2r(2N-2-r)n/(2N-2)
I=N
N-2
= L xcl[k]e-;2<0../(2N-2)
1=1
therefore:
I N-1 N-2
z 1 (n] = 2N- 2 <2:: xc1(A:Jei2dn/(2N-2) + L xc1(k]e-;2 ..../(2N-2))
k=O 1=1
l
= 2N"=2(xc1[0] + xcl[N- l]ei2•n + L
N-2
Xcl[k](ei'"•"/(2N-2) + .-j2dn/(2N-2)))
N-2
•=•
1
= - -(Xc1 (0]
2N-2
+ Xc 1 (N -l]ei'" + "Xcl(k]2cos( dn ))
£.... N-1
1:=1
and:
= N ~ 1 <2:: o[A:jxe•[k]cos(;:"l))
•=<>
o[k]= U k=OandN-I
I $.1: $N-2.
v[n] = z2[2n]
therefore, for k =0, I, ... , N- 1:
I
V(k] = i(X2[k] + X2[k + N]).
Using equation (8.I68), we have:
Furthermore, we have:
N-1
2Re{e-i~V[k]} = 2&{e-i~ I; v[n]e-i"P}
n=O
N-1
= 2 I; Re{v[n]e-i<'W'I•+lll}
n=O
N- 1 27rk 1
= 2 I; v[n] cos( N(n + 4))
=0
= 2 I: [ J
vncos
2
+I))
("k(4nN .
n=O
and:
2ReV'~V[k]} = 2Re{X[k]e-i'*}
2N-l
= 2 I; Re{:r[n]e-iftl 2•+1l}
n=O
N-1
[ ] cos ("k(2nN+ 1)) .
2 I; :rn
= 2
n=O
8.70. Substituting the expression for X2[k] from equation (8.174) into equation (8.175), we get:
344
2N-1 .
= -1 L x,[k]e'...... ,<,N>
2N •=<>
1 N-1 2N-1
= 2
Ncx•2[o] + L x"[k]ei..t<•N>.,......,<,N> _ L xe2[2N _ k]ei .. t<•N>.,;• ....t<•"'>J
Cl k=N+l
1 N-1 · 2N-l
= u,;(X"[O] + L X''[k].,id(2n+l)/(2N) _ L X"[2N _ j;).,id(2n+l)/(2N))
k=l i=N+l
N-1 N-1
=
2
1
N(X"[O] + L X''[kj.,id(2n+1)/(2N) _ L X"[kj.,i•(2N--<o)(2o+1)/(2N))
1=-1 1=1
1 N-1 N-1
=
2
N(X"[O] + L
X"[k]eid(2n+1)/(2N) + L X"[k]e-j•l(..,+1)f(2N))
i=l i=l
N-1
1
= 2N(X''[O] + L x•'[k](ei••<••+1){(2N) + .-jd(2n+l)/(2N)))
i=l
N-1
1
= -2N (X'2[0] + ~ X' 2(k] cos( 1rk(2n + 1)))
~ 2N .
.f:=l
Furthermore:
P[k] = { ~ k=O
1~k~N-l.
y[n] = :r:[nJ®z*[((-n))N]·
Using the properties of the DFT, we have:
2N-3 N-1
L Jz,[n]l 2 =2 L jz[n]l' -lz[O]I 2 - )z/N -1]1
2
•
n=O n=O
We thus conclude:
u./- 2 (2
N-1
L
n=O
JX"fk]l' -IX''[OJI'- IX"[N- 1]1 2 ) = 2
N-1
L
n=O
jz[n]l'- lr[OJI'- lriN- !]I'.
We thus conclude:
2~(2 L
N-1
....
IX[k]J 2 -IX[O]j 2 ) =2 L
-
N-1
jz[n]j 2 •
347
Solutions - Chapter 9
9.1. There are several possible approaches to this problem. Two are presented below.
Solution #1: US.. the program to compute the DFT of X(k], yielding the sequence g[n].
N-1
g[n] =L X[k]e-j2dn/N
1=0
Then, compute
1
z[n] = Ng[((N- n))N]
for n = 0, ... , N- I. We demonstrate that this solution produces the inverse DFT below.
1
z[n] = Ng[((N- n))N]
N-1
= ~ L X[k]e-j2w1(N-n)/N
1=0
N-1
= ~ L X[kjd2dn/N
1=0
Solution #2: Take the complex conjugate of X[k], and then compute its DFT using the program,
yielding the sequence f[n].
N-1
/[n] =L
X"[kje-j2wln/N
i:=O
Then, compute
z[n] = ~/"[n]
We demonstrate that this solution produces the inverse DFT below.
o:[n] = ~f"[n]
N-1
= ~ L X[kj&••ln/N
1=0
9.3. (a) The input should be placed into A[r] in bit-reversed order.
A[O] = :r[O]
A[1] = :r[4]
A[2] = :r[2]
A[3] = :r[6]
A[4] = :r[l]
A[S] = :r[S]
A[6] = :r[3]
A[7] = :r[7]
X[k] '
= 2::<-Ws)"w,;•
n=O
'
= :L<-1J"w,;w,;•
n=O
'
= :L:<w.-•l"W;w,;•
n=O
'
= :L:w;<•-•)
n=O
1- w•-•
= 1- w;-•
= 86[k- 3]
A sketch of D[r] wis provided below.
0 1 2 3 4 5 6 7 r
0 1 2 3 4 5 6 7 r
9.4. (a) In any stage, N/2 butterflies must be computed. In the mth stage, there are 2m-l different
coefficients.
(b) Looking at figure 9.10, we notice that the coefficients are
We therefore have
y[n + 19] = X(eiw• ), n = 0, ... , 9
for Wn = wo +nD.w or
y(n] = X(eiw• ), n = 19, ... , 28
for Wn = wo + (n- 19)J:>w.
9.9. In this problem, we are using butter:fly flow graphS to compute a DFT. These computations are done in
=
place, in an array of registers. An example flow graph for aN = 8, (or v log2 8 3), decimation-in- =
time DFT is provided below.
353
-1 -1 -1
(a) The difference between l 1 and lo can be foood by using the figure above. For example, in the first
stage, the array elements A[4] and A[SJ comprise a butterfly. Thus, l 1 - lo = 5- 4 = 1. This
difference of 1 holds for all the other butterflies in the first stage. Looking at the other stages, we
find
stage m = 1: l 1 -lo = 1
stage m = 2: l1 - lo = 2
stage m = 3: l 1 -lo = 4
From this we find that the difference, in general, is
l, -lo =2m->, form= 1, ... ,v
(b) Again looking at the figure, we notice that for stage 1, there are 4 butterflies with the same twiddle
factor. The lo for these butterflies are 0, 2, 4, and 6, which we see differ by 2. For stage 2, there are
two butterflies with the same twiddle factor. Consider the butterflies with the W,l' twiddle factor.
The lo for these two butterflies are 0 and 4, which differ by 4. Note that in the last stage, there
are no butterflies with the same twiddle factor, as the four twiddle factors are unique. Thus, we
found
stage m = 1: t:l.lo = 2
stage m = 2: t:l.lo = 4
stage m = 3: n/a
From this, we can generalize the result
t:J.lo =2m, ·form= l, ... ,v -1
9.10. This is an application of the causal version of the chirp transform with
The new sample order is 0, 8, 4, 12, 2, 10, 6, 14, 1, 9, 5, 13, 3, 11, 7, 15.
9.12. False. It is possible by rearranging the order in which the nodes appear in the signal !low graph.
However, the computation cannot be carried out in-place.
9.13. Only them = 1 stage will have this form. No other stage of aN= 16 ra.dix-2 decimation-in-frequency
FFT will have a W1& term raised to an odd power. ·
9.14. The possible values of r for each of the four stages are
~=1, r=O
m=2, r=0,4
m=3, T =
0,2,4,6
m=4, r = 0,1,2,3,4,5,6, 7
355
N Program A Program B
2 4 20
4 16 80
8 64 240
16 256 640
32 1024 1600
64 4096 3840
Thus, we see that a sequence with length N = 64 is the shortest sequence for which Program B runs
faster than Program A.
9.16. The possible values for r for each oftbe four stages are
m= 1, r=O
m=2, r =0,4
m=3, r =0,2,4,6
m=4, r =0, 1,2,3,4, 5,6, 7
where WN is the twiddle factor for each stage. Since the particular butterBy shown bas r = 2, the stages
which have this butterfiy are
m=3,4
9.17. The FFT is a decimation-in-time algorithm, since the decimation-in-frequency algorithm bas only !Vf2
terms in the last stage.
9.18. lf the N 1 = 1021 point DFT was calculated using the convolution sum directly it would take Nf
multiplications. lf the N 2 = 1024 point DFT was calculated using the FFT it would take N2log2 N2
multiplications. Assuming that the number of multiplications is proportional to the calculation time
the ratio oftbe two times is
N'f 1021 2
N2log2 N2 = 1024log2 1024 = l01. 8 "' 100
which would explain the results.
9.19. X(e'••l•) corresponds to the k = 3 index of a length N = 8 DFT. Using the 6ow graph of the
second-order recursive system for Goertzel's algorithm,
a = 2cosC~k)
= 2 cos C"i3))
= -¥2
b = -W~
= -e-itnr/8
1 +j
=
V2
356
9.20. First, we derive a relationship between the X 1 (ei") and X(ei") using the shift and time reversal
properties of the DTFT.
o: 1 [n] = z[32- n]
X,(.,;w) = X(e-i")e-i3 ""'
Looking at the figure we see that calculating y[32] is jnst an application of the Goertzel algorithm with
=
k 7 and N = 32. Therefore,
y[32] = x,[7]
= X, (ei" ) 1.,= 'H'
= X(e-i")e-i"321 w=fi
= X(e-ilt)e-i<Ttl>2
= X(e-ilt)
Note that if we put z[n] through the system directly, we would be evaluating X(z) at the conjugate
location on the unit circle, i.e., at w = +77r/l6.
9.21. (a) Assume z[n] = 0, for n < 0 and n > N- 1. From the figure, we see that
Y>[n) = :t[n] + Wty.[n- 1]
Starting with n =0, and iterating this recursive equation, we lind
Y>[OJ = z[OJ
11•[1] = :[1] + wtz[O]
Y>[2] = i[2] + W~z[1] + WYz[O]
Therefore, y,[n] = z[n] + Wty,[n- 1]. This is the same dilference equation as ill part (a).
357
9.22. The flow graph for 16 point radix-2 decimation-in-time FFT algorithm is shown below.
x(O] .__....._:;::--+--::7..._+--.._----~-+--<r-----+-r-T-----+1X[O)
To determine the number of real multiplications and additions required to implement the Sow graph,
358
consider the number of real multiplications and additions introduced by each of the coefficients Wt:
W/'6 :0 real multiplications + 0 real additions <Wl'. = 1)
Wt, : 0 real multiplications+ 0 real additions (W140 (a + jb) = b- aj)
Wf, : 2 real multiplications+ 2 real additions (Wr6 (a + jb) = "!<a+ b)+ i"'(b- a))
Wt6 : 2 real multiplications+ 2 real additions similarly
Wf, : 4 real multiplications+ 2 real additions
Wl6 : 4 real multiplications+ 2 real additions
Wf, : 4 real multiplications + 2 real additions
W,:'6 : 4 real multiplications + 2 real additions
Solving for
gives
112
X{O] x(O]
W:t2
X{4] x(1]
1/2
Xl2! x(2]
W:t2
X{6] x(3]
112
X{1] x(4]
Ill' 12
X{S] e x(S]
112
X{3] x(6]
W:t2
X{7]
(c) The modification is made by removing all factors of 1/2, changing all Wj;:' toW,\., and ~labeling
the input and the output, as shown in the flow graph below.
W'
x(4) • X{1)
W'
x(6) • Xl31
W'
x(S] ' XIS)
(d) y..,, 1n general, for each decimation-ill-time FFT algorithm there ezists a decimation-ill-frequency
FFT algorithm that corresponds to interchanging the input and output and reversing the direction
of all the arrows in the flow graph.
360
9.24. (a) Using the figure, it is observed that each output Y[k] is a scaled version of X[k]. The scaling
factor is W[k], which is found to he
k =01234567
W[k] = 1 G G (fl G G2 G2 G3
Y(k] = W(k]X[k]
= W(k)X[kjW'(k]
= X(k]
9.25. Let z• be the z-plane locations of the 25 points uniformly spaced on an arc of a circle of radius 0.5
from -'If /6 to 271' /3. Then
where
71'
wo =
6
&; = C:)
5..
(;4)
=
144
This is similar to the expression for X(ei~) using the chirp transform algorithm. The oilly difference is
the (0.5)-• term. Setting
we get
N-'
X(z.) = w•'t• L g\n\w-(•-•1' , •
...0
using the result of the c:hirp transform algorithm. A procedure for computing X (z) at the points •• is
then
361
• Multiply the sequence x(n] by the sequence (0.5)-•e-i"'••w•'/2 to form g(n] .
• Convolve g(n] with the sequence w-•'1 2 •
• Multiply this result by the sequence w•'/2 to form X(z•)·
A block diagram of this system appears below.
g{n]
x[n]---!(x 1--~x}----+ X(~)
9.26.
2N-l
Y(k] = L y(n]e-il~l••
n=O
N-1 2H-l
= L 2
e-j(r/N)n e-j(2•/N)(k/2)n + L 2
e-i(•/N)n e-j(2 .. /N)(II:J2)n
n=O n=N
N-1
= L L
N-1
e-i(•/N)n 2 e-j(2•/N)(k/2)n + e-i(•/N)(l+N) 2 e-i(2wjN)(kf2)(l+N)
n=O l=O
N-1 N-1
= L 2
e-i{•/N)n e-i(2r/N)(t/2)n + e-i•lc L 2 2
e-j(wfN)(l +2Nl+N )e-i<2•fN}(i/2)l
n=O ~~
N-1 N-1
= L 2
e-j{'ff/N)n e-j{2•/N).(k/2)n + (-l)k L 2
e-i(rfN)l e-j(2wjN)(1/2)J
n=O l=O
N-1
= (1 + (-1)•) L 0 -j(r/N)n' 0 -j(2r/N)(•/2)n
n=O
= { 2X(k/2], k even
0, k odd
Thus,
k even
Y(k] = { 0, k odd
9.27. Let
y(n] = 0 -j2rn/027 z(n]
Then
Y(e'"') = X(eil"'+tM)
Let y'(n] = 2::=-oollfn +256m], 0 :5 n :5 255, and let Y'[k] be the 256 point DFT of y'(n]. Then
Y'(k] =X (ei(lii+iit-J)
See problem 9.30 for a more in-depth analysis of this technique.
362
9.28. (a) The problem states that the effective frequency spacing, !!.f, should be 50 Hz or less. This
constrains N such that
1
!!./ = NT :S 50
1
N "2:. SOT
~ 200
Since the sequence length L is 500, and N must be a power of 2, we might conclude that the
minimum -.alue for N is 512 for computing the desired samples of the z-transform.
However, we can compute the samples with N equal to 256 by using time aliasing. In this technique,
we would zero pad z[n] to a length of 512, then form the 256 point sequence
We could then compute 256 samples of the z-transform of y[n]. The effective frequency spacing of
these samples would be 1/(NT)"' 39Hz whlch is lower than the 50 Hz specification.
Note that these samples also correspond to the even-indexed samples of a length 512 sampled
z-transform of z[n]. Problem 9.30 discusses this technique of time aliasing in more detaiL
(b) l.<!t
y[n] = (1.2Wz[n]
Then, using the modulation property of the z-transform, Y(z) = X(0.8z) and so Y[k] = X(0.8e; 2r>fN).
9.29. (a) We offet two solutions to this problem.
Solution #1: Looking at the DFT of the sequence, we find
N-1
X[k] = L z[n]e-j2dn/N
=0
(N/2)-1 N-1
= L z[n]e-j2dn/N + L z[n]e-j2dn/N
n=O n.=N/2
(N/2)-1 (N/2)-1
= L z[n]e-;•••nJN + L z[r + (N/2)]e-;2d[•+(N/2)J/N
n=O r=O
(N/2)-1
= L z[n][l- (-1)•je-j2•h/N
=0
= 0, k even
Solution #2: Alternatively, we can use the circular shift property of the DFT tO lind
X[k] = -X[k]e-;<~>•<f>
= -(-t)•X[k]
= (-1)>+ 1 X[k]
When k is even, we have X(k] = -X[k] whlch can only be true if X(k] = 0.
363
'"" N/2-1
= DITN/2 { :t[n]e-;(2•/N)n} + ( -1)( -1) L :t[l]e-;2•1/N 0 -;2rti/(H/2)
=
for k 0, ... , N /2- 1. Thus, we can compute the odd-indexed DIT values using one N /2 point
DIT plus a small amount of extra computation.
9.30. (a) Note that we can write the even-indexed values of X[k] as X[2k] for k = 0, ... , (N/2) - 1. From
the definition of the D IT, we find
N-1
X[2k] = L :t[n]e-;2r(2t)n/N
.....
N/2-1
= L :t[n]e-;IPi,>n
.....
N/2-1
+ L :t[n + (N/2)je-;,Ji•>>ne-;,Ji,<Nf 2 >•
N/2-1
= L (:t[n] + :t[n + (Nf2)J)e-irPfu>n
.....
= Y[k]
oo M-1
= L L :t[n +rM]e-j2d(n+rM)/Me>"2•(rJt)t/JI
r=-oo n=O
Y[k] = L :t[l]e-;,..,/.11
l=-oo
= xcei2••fMJ
Thus, the result from Part (a) is a special case of this result if we let M =
N /2. In Part (a), there
are only two r terms for which y[n] is nonzero in the range n = 0, ... , (N /2) - 1.
364
(c) We can write the odd-indexed values of X[k] as X(2k + 1] fork = 0, ... , (N/2)- 1. From the
definition of the DFT, we find
N-1
X[2k + 1] = L z[n]e-i..(2H1)n/N
.....,
N-1
= L z[n]e-i2rn/Ne-i2r(2>)n/N
.....,
(N/2)-1 (N/2)-1
= L z[n]e-i2•n/N.-i,Jj,,>n + L z[n + (N/2)]e-j2r{n+(N/2)]/Ne-;;!r,>ln+(N/2))
n=O n=O
(N/2)-1
= L [C:r[n]- :r[n + (N/2)])e-i~n] .-;rM.r>n
n=O
Let
n _ { (z[n]- z[n + (N/2)])e-i< 2•/N)n, 0 $ n $ (N/2)- 1
y[J-0, oth "
ennse
Then Y[k] = X[2k + 1]. Thus, The algorithm for computing the odd-indexed DIT values is as
follows.
step 1: Form the sequence
step 2: Compute the N/2 point DFT of y[n], yielding the sequence Y[k].
step 3: 'l'he odd-indexed values of X[k] are then X[k] = Y[(k- 1)/2), k = 1, 3, ... , N- 1.
9.31. (a) Since x[n) is real, z[n) = x"[n], and X[k] is conjugate symmetric.
N-1
X[k] = L x"[n]e-i~>n
..,..
N-1 )"
= ~ z{n]ei~>ne-i~Nn
(
= X"[N-k]
In these expressions, the subscripts "E" and "0" denote even and odd symmetry, respectively, and
the subscripts "R" and "1" denote real and imaginary parts, respectively.
365
GER[k) = X1ER[k)
the odd and real part of G[k] is
GoR[k] = -X,or[k]
the even and imaginuy part of G[k] is
GEr[k] = X2ER[k)
and the odd and imaginary part of G[k] is
Gor[k] = X1or[k]
Having established these relationships, it is easy to come up with expressions for X1[k] and X2[k].
(c) An N = 2" point FFT requires (N/2) log2 N complex multiplications and N log2 N complex addi-
tions. This is equivalent to 2N log2 N real multiplications and 3N log2 N real additions.
{i) The two N-point FFTs, X 1 [k] and X 2 [k], require a total of 4Nlog2 N real multiplications and
6N log2 N real additions.
{ii) Computing theN -point FFT, G[k], requires 2N log2 N real multiplications and 3Nlog2 N real
additions. Then, the computation of GER[k], GEr[k], Gor[k], and GoR[k] from G[k] requires
approximately 4N real multiplications and 4N real additions. Then, the formation of X1 [.I:] and
X,[k] from GER[k], GEr[k], Gor[k], and GoR[k] requires no real additions or multiplications.
So this technique requires a total of approximately 2N log2 N + 4N real multiplications and
3N log2 N + 4N real additions.
(d) Starting with
N-1
X[k] =L z[n]e-j2dnfN
"'""
and separating :[n] into its even and odd numbered parts, we get
step 1: Form the sequence g[n] = z[2n] + jz[2n + 1], which has length N /2.
step 2: Compute G[k], the N/2 point DFT of g[n].
step 3: Separate G[k] into the four parts, fork= 1, ... , (N/2)- 1
1
GoR(kj = 2(GR(k]- GR((N/2)- k])
1
GsR(kJ = 2(GR[k] + GR[(N/2)- k])
1
Gor[k] = (Gr[k]- Gr[(N/2)- k])
2
Gsr[k] = ~(Gr[k] + G1 [(N/2)- k])
which each have length N /2.
step 4: Form
(d) For these signals, N is large enough so that circular convolution of z[n] and h[n] and the linear con-
volution of :z;[n] and h[n] produce the same result. Counting the number of complex multiplications
for the procedure in part (b) we get
Since there a.re 4 real multiplications for every complex multiplication we see that the procedure
takes 6Niog., N + 4N real multiplications. Using the answer from part (a.), we see that the direct
=
method requires (N /2)(N/2) N' /4 real multiplications.
The following table sbows that the sma.llest N = 2• for which the FFT method requires fewer
multiplications than the direct method is 256.
9.33. (a) For each L point section, P- 1 samples a.re discarded, leaving L - P + 1 output samples. The
complex multiplications a.re:
L point FFI' of input: (L/2) log2 L = v2• /2
Multiplication of 61ter and section DFI': L = 2•
L point inverse FFT: (L/2) log2 L = v2• /2
Total per section: 2•(v + 1)
Therefore,
Complex Multiplications 2• ( v + 1)
Output Sample
= 2• - P + 1
Note we assume here that H[k] ba.s been precalculated.
(b) The figure below plots the number of complex multipllcations per sample \'etS1I.i v. For v = 12,
the number of multiplies per sample reaches a. minimum of 14.8. In comparison, direct evaluation
of the convolution sum would require 500 complex multiplications per output sample.
368
.5125
f"'20
'S
CL
~ 15
8.
!! 10
S!
Is
'5
~ oL-~--~--L-~--~--L-~--~~~~
10 11 12 13 14 15 16 17 18 19 20
v
Although " : 9 is the first valid choice for overlap-save method, it is not plotted since the value is
so large (in the hundreds) it would obscure the graph.
(c)
I Overlap/Save I Direct I
I" I I 2"-1 I
,..~ .. +1~
2"' 2"'-1 1
1 2 1
2 4 2
3 6.4 4
4 8.9 8
5 11.3 16
9.34. This problem asks that" we find eight equally spaced inverse OFT coefticients nsing the chirp tranSform
. algorithm. The book derives the algorithm for the forward DFI'. However, with some minor tweaking,
it is easy to formulate an inverse DFI'. First, we start with the in?erSe DFI' relation
N-1
:r[n] = .!_ L X(kjel'ralo/N
NbO
N-1
:r[nt) : .!. L X{kjd2wn,./N
N>=O
369
Next, we define
.O.n = 1
nt = no + l.O.n
where l = 0, ... , 7. Substituting this into the equation above gives
z[nt] =
Defining
we find
N-1
z[nt] =~ L X(k]ei'•no>!Nw-tt
4=0
Using the relation
lk = ~[l' +k2 - (k -tl'l
we get
N-1
z[nt] = 2.
NL-
"X[k)ei2 •no. >tNw-t't'w-"'''w<•-tJ't>
.
•=<>
Let
Then,
From this equation, it is clear that the inverse DFI' ca.a be comptned using the chirp transform algorithm.
All we need to do is replace n by k, change the sign oC each oftl>e expoltelltial terms, and divide by a
factor of N. Therefore,
= ei2·d:no/Nw-~: /2
2
m1 (k)
m,[k) = "'-1:2/2
h(k) = N2.w•'t•
Using this system with no = 1020, and l = 0, ... , 7 will result in .a sequence y(n) which will contain the
desired sample:;. wilere
y[O] = z[1020)
y[1] = z[1021]
y[2] = z{1022]
y(3) = z(1023)
y[4) = z[O)
y(5] =
:t(l)
y(6) = :[2)
y(7) == z[3)
370
:r;[n]
= { :r[n], iL:;; n :S iL + 127,
0, otherwise
Using the above we can implement the system with the following block diagram.
x[n] ~
Shift
by -iL I FFT-1
w[n ] = u[n]-u[n-128)
Multiply . J 256-pt r-- IFFf-2
Multiply ~ 256-pt --+ y;[n]
FFT-1
h[n) ~
256-pt ---+
The FFT size was chosen as the next power of 2 higher than the length of the linear convolution. This
insures the circular convolution implied by multiplying DFTs corresponds to linear convolution as well.
Neonv = N., + Nh - 1
= 128+ 64- I
= 191
NpPT = 256
9.36. (a) The flow graph of a decimation-in-frequency radix-2 FFT algorithm for N = 16 is shown below.
371
x(OJ X(OJ
x(1J
w;, X(8J
x(2J X(4J
x(3J
w;. X(12]
x(4J X(2]
x(S]
w;, X(10]
x(6J X(6]
x[7J
w;, X(14]
x(8J X{1J
x(9J
w;, X{9J
xf10J X{ 51
x[11]
w;, X(13]
x[12] X{3J
x(13J
w;. X(11]
x(14] X[7J
x(15]
w;, X[15)
x!O] X(O]
>11]
w:. X[B]
X[4]
w:. X[12]
X[2]
w:. X(10]
X[6]
w:. X[~4]
X[1]
w:. X[9]
X!5]
w:. X[13]
w:. X(3]
w:. X[11]
X[7]
w:. X[15]
(c) The pruned butterflies can be used in ( 11 - I') stages. For simplicity, assume that N /2 complex
multiplies are required in each unpruned stage. Counting all W~ terms gives
where
2 f-1
Q=N L :r[2n+ 1],
n=O
we note that :r[2n + 1] = h[n], and :r[2n- 1] = h[n- 1] for n = 0, 1, ... , lf - 1. We then get
So
Therefore,
X{k] = G(k] + WtH{k]
= G(k]+ F(k]
w-•
N - w•
N
j F{k]
= G(k]- 2sin(21rk/N)
Clearly, we need to compute X[O] and X(N/2] with a separate formula since the sin(27rk/N) = 0
fork= 0 and k = Nf2.
(c) For each stage of the FFT, the equations
X(O] = G(OJ + F(OJ
X[N/2] = G[OJ - F(O]
require 2 real additions each, since the values G[O] and F[O] may be complex. We therefore require
a total of 4 real additions to implement these two equations per stage.
For a single stage, the equation
Note that this is approximately half the computation of that of the standard FFT.
(d) The division by sin(27rk/N) fork near 0 and N/2 can cause X[k] to get quite large at these values
of k. Imagine a signal zl[n], and signal z2[n] formed from z,[n] by adding a small amount of white
noise. Using this FFT algorithm, the two FFTs X 1 [k] and X 2 [k] can vary greatly at such values
of k.
9.38. (a)
N-1
X(2k] = L z{n]W~
.....
(N/2)-1
= L (z!nJW~ +z(n + (N/2)]w~•<n+(N/•ll)
-=0
(N/2)-l
= L (r[n) +z[n + (N/2Jll w~·n
.....
375
In the derivation above, we used the fact that W},N = 1. Since WJl" = W},'J 2, X[2k] has been
expressed as an N/2 point DFT of the sequence z[n] + z[n + (N/2)], n = 0, 1, ... ,(N/2)- 1.
(b)
N-1
X[4k + 1] = L z[n]W~...+1)n
n=O
(N/<)-1 .
= L (z[n]w;w~>n + z(n + (Nf4)]w;+<NJ•>w:,•<•+<NJ<))
n=O
In the derivation above, we used the fact that WJ:I< = -j, W;:'14 = j, Wf:/2 = -1, and W},N = 1.
Since wt,•• = Wf<'J4 , X[4k + 1] has been expressed as a N/4 point DFT. But we need to multiply
the sequence (z(n]- z[n + (N/2)])- j(z[n + (N/4)]- z[n + (3Nf4)]) by the twiddle factor W;,
0::; n ::; (N/4) -I before we compute the N /4 point DFT.
The other odd-indexed terms can be shown in the sa.me way to be
(N/4)-1
X[4k + 3] = L {(z[n]- z[n + (N/2)])
n=O
Parts (a) and (b) show that we can replace the computation of anN point DFT with the compu-
tation of one N /2 point DFT, two N /4 point DFTs, and some extra complex arithmetic.
(c) Assume N = 16 and define
)
376
g(O)
>!OJ X[O]
x(1 J X[S]
x(2) X[4]
>13) X[12)
8-point OFT
>141 X[2]
>IS] X[10]
>161 X[6]
x(7J X[14)
x(B]
~. X[1)
,
w,.
x(9] X[9]
x(10)
w.. X[S]
>111)
w:. X[13)
><!12]
~. X[3)
>113]
w:. X[11)
4;>oi'11 OFT
>114)
w,. X[7]
>115]
w,. X[15]
-j
X[OJ
~. X[8)
lC[4)
~. X[12)
X[2)
~. X[10)
lC[6)
~. X[14)
-1
lC[1)
~. X[9)
~. X[5)
~. X[13)
X[3)
~. X(11)
~. lC[7J
~. X[15)
(d) The fiow diagram for the regular radix-2 dec:im&tion-in-frequency algorithm is shown in the next
figure for N = 16. Not conming trivial multiplications by W~, we find that there are 17 complex
multiplications total. Of these 17 complex multiplications, 7 are multiplications by Wt, = -j.
Since a multiplication by - j can be done with zero real multiplications, and a complex multipli-
cation requires 4 real multiplications, we find that the total number of real multiplicationS for the
=
decimation-in-frequency algorithm to be (10)(4) 40.
Taking a look at the split-radix algorithm, we find again that there are 17 complex multiplications.
378
In this case, however, 9 of these are by wt. = -j. Thus, it takes a total of (8)(4) = 32 real
multiplications to implement this ftow graph.
lqO)
w:. lCI8J
X[ol)
w:. lq12)
lCI2J
w:. lq10)
X{6)
w:. lq14)
lq1)
w:. lq6)
X[5J
w:. X(13)
X(3)
w:. X(11)
X(7[
w:. X(15)
arctan(:r)
'"'' < 1
379
we notice that, to first order, the 9; and 9;+1 differ by a factor of 2. ·(Note tbat these formulae are
in radians). This approximate factor of 2 for sucessive 9; is confirmed by looking at some values
of 9;: 9o = 45°, 81 = 26.6°, 92 = 14.0°, 93 = 7.1°. So we have a set of angles whose values are
decreasing by about a factor of 2.
You can add and subtract these 8; angles to form any angle 0 < 8 < "/2. The error is bound by
9M = arctan(2-M), the angle tbat would be included next in the sum. H the error were greater
than 9M, then one of the a; terms must bave been incorrect. The inclusion of the Mth term must
bring the sum closer to 9.
Qo = +1
-B = ao9o
fori=1toM-1
if (8> 9)
Oi = -1
else
Qj = +1
end for
N-1
X[3k] = L z[nJW1,..
N/3-1 2N/3-l N-1
= L z[nJW1,.. + L z[nJW1nk + L z[nJW1nk
ft.=O a=N/3 ,_2N/3
Substituting m = n- N /3 into the second summation, a.nd m = n- 2N/3 into the third summation
gives
N/3-1 N/3-1 N/3-1
X[3k] = L z[nJW1,.. + L z[m + Nf3JW1"'•wg• + L z[m + 2Nf3JW1"'•w~•
n.=O m=O m=O
N/3-1
= L (z[n]+z[n+N/3]+z[n+2N/3])W1nk
"""'
N/3-1
N-1
X[3k + 1] = L z[nJw;<3>+1l
.....
N/3-1 2N/3-1 N-1
= L z[n]w;< 3>+ 1l + L z[n]w;< 3>+1l + L z[n]w;< 3>+1l
ft=O n.=N /3 n.=2N/3
Substituting m = n-N/3 into the second summation, a.nd m = n~2Nf3 into the third summation
gives
N/3-1 N/3-1
X[3k + 1] = L z[nJw;<3t+l) + L z[m + N/3JW1m+N/SH 3H 1)
n=() m=O
N/3-1
+ L z[m + 2N/3JwJ.."'+2N/3)(3HI)
m=O .
N/3-1 N/3-1
= L z[n]w;<St+IJ + L z[m + N /3Jw;<n+ 1 >wg•w~' 3
n=() m=O
N/3-1
+ L z[m + 2N/3Jw;<St+llwJt•wJ:13
......
N/3-1
= L (:r[n] + o:[n + N /3}w;'3 + o:[n + 2N/3JW:J'f3)w;:<3•+1l
n=()
381
N/3-1
N-1
X[3k + 2] = L z[n]w,:;<•>+•>
n=()
Substituting m = n- N/3 into the second summation, and m = n- 2N/3 into the third summation
gives
N/3-1 N/3-1
X[3k + 2] = L :r[n]w,:;t•>+•> + L z[m + Nf3JW~m+N/3)(3'+2)
n=O m=-if-
N/3-1
+ L :r(m + 2 N/ 3 ]W~m+2N/3)(3t+2)
N/3-1 N/3-1
= L: :r[n]w,:;< .. +•> + L: :r[m + Nf3]w;< .. +•>wff•w~Nt•
n=O m=O
N/3-1
+ L: :r[m + 2N/3Jw;<.. +•>w,V•w'~ 13
m=O
N/3-1
= L (z[n] + :r[n + Nj3]W~N/l + z[n + 2N/3]w'~' 3 )W,:;<3t+'l
n=O
N/3-l
= L (:r[n] + z[n + Nf3JW;:'
13
+ z[n + 2N/3]w'~ 13 )Wl-"WM 3
n=O
X[k] = L• z[n]W3"
n=()
X[O] =
z[O] + z[1] + z[2]
X[1] = z[OJ + z[1]WJ + z[2]Wf
X[2] =
z[O] + z[1]Wf + z[2JW: = z[O] + z[1]Wf + z[2JWJ
The butterfly for the 3 point OFT is drawn below.
382
x, _,IPJ
(d) Using the results from parts (a) and (b), the flow graph is drawn below.
383
x1(0}
X(OJ
x1(1}
X(3J
x,(2)
X(6]
vof. xJOI
X(1]
w'
• •,!11 N•3DFT X(4]
vi, x,f2J
X[7J
vof. xjOJ
X(2]
w: xp!
X(8]
•
384
x1(0(
lC(O) X) OJ
><(2)
~ X)6)
x1[1)
><(3) X)1)
w'
><(4) N=3
OFT
• x,f1l Na3
OFT
X)4)
><(5)
w: X[7)
><(6) X[2]
(f) A direct implementation of the 9 point DFI' equation requires 9" = 81 complex multiplications.
The system in part (e), in constrast, requires 4 complex multiplications for each 3 point DFI', and
an additional 4 from the twiddle factors, if we do not count the trival ~ multiplications. In total,
the system in part (e) requires 28 complex multiplications. In general, a r&dix-3 FFI' of a sequence
=
of length N 3" requires appro:r;imaUly
N-1
= e-id fN
2
L 2 2
z{n]e-jwn2/N ei•(t -2.Cn+n )/N
N-1
.....
= L x[n]e-;•.,.fN
"'""
(b)
N-1
X[k + NJ = h"[k + NJ L x[n]h"[n]h[k + N- n]
=0
2
h"(k + Nj = .-jr(t+N) /N
= e-id 2 /N e-ittN
= h"[k]e-,•N
So
N-1
X[k + NJ = h"[k]e-i•N L x[n]h"[n]h[k- n]ei•N
= X(k]
xJkl
x{k) ---i(X}----!1 ~] y{k]
1\•
h [k]
x,(k] = x[k]h"[k]
386
00
"'•lkJ = 2: ,.,[t]i.[k- ll
l=-oo
N-1
= L :z:[t]e-i•t'IN .,;•<•-tl' IN
....
k E [0, ... , 2N- 1]
Therefore,
= '""'
X[k+NJ kE[O, ... ,N-1J
= X[k]
{d)
2N-1
H(z) = L .,id'INz-•
.=0
M-12M-1
= L L 2
ei7t(r+lM) fNz-(r+lM)
r=O l=O
J4-12M-1
== L L ei'"2 jN ei2•rl/M ei•l2 z-r z-LM
r=O l::=O
1
r~ (.,i2rriM .-M)l(-1)'2]
1
= 3:: .,;rr'IN .-•
=<I '""'
1
r~ (d2r.IM .-M)'{-1)']
1
=
r=O
M-1 . ,
~
~
e1•r fN z-r
[!- l=O
ei2rr(2M)IM z-2M']
1 + ei2r./M z-M
=<I
~ .,;rr' IN -• [
1
= 1 - z-•M' ]
~ z 1 + ei(b/AI)rz-M
is drawn below.
387
x;tl
z_..,
z-1
-z--N- -1
,?A
z....,
z-1
~
ef's/M
z....,
z-1
~M--1)/M
(f) Complex multiplications: Since we are only interested in y[k] for k = N, N + 1, ... , 2N- 1 we
do not need to calculate the complex multiplications on the output side of each parallel branch
until k <': N. Thus,
Complex additions: The complex additions on the output side of each parallel branch do not
need to be computed until k ;:: N. Thus,
x[r] Re{X[k]}
z-1 .s>...
p>" e2
2 cos"\
lm{X[k]}
sin (J)k
z-1
e1
e3
-1
(b)
Re{X(k]} = Re{y.(N]}
Im{X(k]} = lm{yo[N]}
Since the output of interest is the Nth sample, we need only consider the .ariance at time N.
The noise e,[n] is input to both hR(n] and hr(n]. Using the techniques from chapter 6, we find the
variance of the noise is
N
Let 8=2dfN.
N
2
L:cos 8n
n=O
N
= ~ L(eifn + e-ifn)2
n=O
N
= !. L(ei12n + 2 + .-if2n)
4n=O
1 ( 1 _ ei21(N+I) 1 _ e-i21(N+l))
= 4 1- ei28 + 2(N + 1) + 1- e-;28
= ~(1 + 2(N + 1) + 1)
= N +1
2
Similarly, L~=<> h1{n] = N f2.
Therefore,
2-28
uh(n] = !2(1 + (N/2) + 1)
2-28
= l2(N +4)/2
8
= ~;
2
u~[n] (1 + (N/2))
2-28
= l2(N +2)/2
9.43.
N-1 N-1
X(k] =L :r[n]cos(2.. kn/N)- j L :r(n]sin{2.-kn/N).
n=O n-o
For k # 0, there are N - 1 multiplies in the computation of the real part and the imaginary part:
9.44. (a)
(b) The conditions are not suflicient to guarantee that overflow cannot occur.
and
0 m= 1
Number of - j multiplications in stage m ={ •
N/2"', m=2, ... ,v
(b) If we assume that all the + 1 and - j multiplications are done noiselessly, then the noise variance
will be different at each output node. This is easily seen by looking at Figure 9.10, where we see
for example that X[O] will be noise-free, while X[1] will not be noise-free. Thus, a noise analysis
would be required for each output node separately. A somewhat simpler approach would be to
assume that since the first two stages consist of only + 1 and -j multiplications, these two stages
can be performed noiselessly. Each output node is connected to all N/2 butterdies in the first stage
and to N /4 butterfiies in the second stage. Thus, if the first two stages are performed noiselessly,
a better estimate of the number of independent noise sources contributing to the output is
N N N
N- 1- 2- 4 = 4 - 1.
Note that all the odd indexed outputs will have exactly (N/4)- 1 of these noise sources, while
the even indexed outputs will have less. In fact, X[O], X[N/4], X[N/2], X[3N/4] will be noiseless.
X[N/8], X[3N/8], X[SN/8], and X[7N/8] will have one noise source. It is possible to continue this
analysis for all X[k], but clearly, a complicated formula would be required to describe tbe number
of noise sources for all even k. We have shown that
Thus, th<O number of noise sources is upper bounded ( N I 4) - 1. Using this bonnd, we can get a
"<W\'t~~"'-~~,~~-
N
£[1F[k]I'J $ (4-1)~
When the scaling is done at the input, an upper bound on the noise-to-signal ratio is found to be
Another approach to this noise analysis is to compute the average noise at the output by using the
average number of noisy butterflies connected to an output node. This style of analysis is used in
Weinstein.
(c) Now assume as before that the first two stages are noiseless. Tbus, equation 9.67 would not include
the first two stages.
•-1
t:(IF(k]l 2] < a~ L 2(v-m)(~)2v-2m-2
m=2
•-1
= a~ L (~ )v-m-2
m=2
•-3
= 2a~ D~l·
= 2a2
B
c-
i::O
(!)•-2)
2
1-!2
= ~(1- 2 (!t- 2)
2 4)
= 4as(1--
N
Thus,
4
:::. 12Na~N-
t'[IF[k]l']
t:[IX[k]l 2] N
<
12(N -4)~
X.,(q]
where
(b)
Thus we have
Re{Y,[k]} = Ev{Re{Y,[k]}}
Re{Y2[k]} = Ev{Im{Y3 [k]}}
Im{Y1 [k]} = Od{Im{Y,[k)}}
Im{Y2[k]} = -Od{Re{Y3 [k)}}
and so
1
= 2
[Re{Y,[k]} + Re{Y,[N- A:]}]
(c)
x,[A:J =Re{Y,(k]}
u,(A:J =;rm{Y,(A:]}
and so
X (k] = Im{Y,(k]} k-" 0 N
s 2siD if r , 2
Note that X3 (0] cannot be recovered using this technique, and if N is even, neither can X,[N/2].
394
(f) In part (b) replace X3 and X, with U3 and U, and use the re3Ult of (d) to give
~M b -r
H(z) = -'-~ ,z
1 - I:t=1 a,z-l
~M b ,-;2•></N
H(d'••l") = -"-'-=<•<:=<>T·~---=
~" a,e-;2d.l/N
1- L...t=l
Now assume N, M ~ 511. Let b[n) = bn and
1
a[n) ={ '
a,.,
Let B[k], A[k] be the 512 pt DFTs of b[n), and a[n]. Then
H(d'd./512) = B[k]
A[k]
9.49. (a) It is interesting to note that (linear) convolution and polynomial multiplication are the same
operation. Many m&thematical software tools, like Matlab, perform polynomial multiplication
using convolution. Here, we replace
L-1 Jl-1
p(z) = L a;z', q(z) =L b;z'
;.:() ;.:()
with
L-1 Jl-1
p[n] = L a;c5[n - i], q[n] = L
..., b;6[n - i)
;.:()
Then,
r[n) = p[n] • q[n).
The coefficients in r[n) will be identically equal to those of r(z). We caD compute r[n] with circular
convolution, inste.ad of linear convolution, by zero padding p[n) and q[n] to a length N = L+ M -1.
This zero padding ensures that linear convolution and circular convolntion will give the same result.
(b) We can implement the circular convolution of p[n] and q[n] using the following procedure.
step 1: Take the DFTs of p[n] and q[n] using the FFT program. This gives P{A:] and Q[k).
step 2: Multiply to get R[A:J = P[A:JQ[A:J.
step 3: Take the inVerse DFT of R(A:] using the FFT program. This gives r[n].
395
Here, we assumed that the FFT program also computes inverse DFTs. If not, it is a relatively
simple matter to modify the input to the program so that its output is an inverse DFT. (See
problem 9.1 ).
While it may seem that this procedure is more work, for long sequences, it is actually more efficient.
The direct computation of r(n] requires approximately (L + M) 2 real multiplications, since a; and
b, are real. Assorning that a length L + M FFT computation takes [(L + M)/2] log2 (L + M)
complex multiplications, we count the complex multiplications required iD the procedure described
above to be
Note the resemblance to p(x) and q(x) of part (a). We form the signals
L-1
u(n]
v[n]
=
=
-....
L u;o[n- i]
J t-1
L v;o[n - i]
and use the procedure described in part (b). This computes the product u · v in binary. For
L = 8000 and M = 1000, this procedure requires approximately
(d) For the (forward and inverse) FFI's, the mean-square value of the output noise is (L + M)o-~.
While ~ will be sma.ll, as there are 16 bits, the noise can be significant, since L + M is a large
number.
9.50. (a) Using the definition of the discrete Hartley transform we get
HN[a+bJ = [cos(21tb/N)+sin(2,-b/N))cos(2.-a/N)
+[cos( -27rb/N) +sin( -21rb/N)]sin(2,-a/N)
= HN(b}GN(aJ + HN[-b}SN(a)
(b) To obtain a fast algorithm for computation of the discrete Hartley transform, we can proceed as
in the decimation-in-time FFI' algorithm; i.e.,
(N/2)-1 (N/2)-1
XH(k] = L :t[2r]HN[2rk] + L :t[2r + l)HN((2r + l)k)
r=O r=O
(N/2)-1 (N/2)-1
= L :t[2r)HN[2rkj + :L :t(2r + l}HN[2rk]CN[k}
r=O r=O
(N/2)-1
+ L :t(2r + l)HN[((-2rk))N)SN(kj
r=O
where
(N/2)-1
F[A:J = L z[2r]HNt 2 [rkJ
r=O
is the N /2-pomt DHT of the odd-mdexed pomts. As m the derivation of the decimation-in-time
FFT algorithm, we can continue to divide the sequences m half if N is a power of 2. Thus the
mdex'mg will be exactly the same except that we have to access G[(( -A:))N/2] as well as G[k] and
F[k]; i.e., the "butterlly" is slightly more complicated. The fast Hartley transform will require
N log2 N operations as in the case of the DFT, but the multiplies and adds will be real instead of
complex.
9.51. (a)
1 1 0 0 0 0 0 0 1 0 0 0 0 0 0 0
1 -1 0 0 0 0 0 0 0 1 0 0 0 0 0 0
0 0 I 1 0 0 0 0 0 0 141 0 0 0 0 0
0 0 I -1 0 0 0 0 0 0 0 Wi 0 0 0 0
0 0 0 0 1 I 0 0 0 0 0 0 1 0 0 0
0 0 0 0 1 -1 0 0 0 0 0 0 0 1 0 0
0 0 0 0 0 0 1 I 0 0 0 0 0 0 w~ 0
0 0 0 0 0 0 1 -1 0 0 0 0 0 0 0 Wf
1 0 I 0 0 0 0 0 1 0 0 0 0 0 0 0
0 I 0 I 0 0 0 0 0 1 0 0 0 0 0 0
1 0 -1 0 0 0 0 0 0 0 1 0 0 0 0 0
0 1 -1 0 0 0
0 0 0 0 0 1 0 0 0 0
F• = 0 0 0 0 I 0 I 0
T• = 0 0 0 0 141 0 0 0
0 0 0 0 0 I 0 1 0 0 0 0 0 WJ 0 0
0 0 0 0 1 0 -1 0 0 0 0 0 0 0 Wf 0
0000010 -1 0 0 0 0 0 0 0 Wi
1000100 0
01000 I 0 0
00100 01 0
0001000 1
F3 = I 0 0 0 -1 0 0 0
0 1 0 0 0 -1 0 0
0 0 1 0 0 0 -1 0
0 0 0 1 0 0 0 -1
(b)
QH = Ff/TfF:TfF:
= F 1TiF•T2F3
398
xfO] )iOJ
xf1] )i1]
-1
xf2] )i2]
xf3]
xf4] )i4]
xf5]
xf6] y(6]
x(7J
This structure is the decimation in frequency FIT with the twiddle factors conjugated and therefore
calculates
N · IDIT{z[n]}
(c) Knowing that Q calculates the DIT and 1tQ
8 calculates the IDFT, we should realize that cas-
cading the two should just return the original signal. More formally we have
T{'T, =I TfT2 =I
y[n] = Y: (~ Y:
m=O 1: 1=0
X[k1JW..;•·m) (~ ~ H[k ]w;,;••«n-m))N)
I:2=0
2
= Y: (~ Y:
m=O .k1=0
X[k1Jw..;••m) (~ ~ H[k ]w;,;••<•-m))
.t 2:0
2
= ~2 ~ ~ (x[k1]H[k2]W_N••• ~ w..;••mw~•m)
t,=O &2=0 m=O
Therefore, the circular convolution property holds as long as the basis vectors are orthogonal.
(b)
400
X[k] = ( (~ z[n]4n>))
17
= ((I· I+ 2 · 4• + 3 · I6• + 0 · 64•))17
= ((I +2·4• +3·16•))17
H[kJ (t
= ( h[nW•))
17
= ((3·I + 1·4· +0·I6. +0·64·))17
= ((3+4.)h7
Using this formula for H[k),
Y[O] = ((24)),7 = 7
Y[1] = ((42)),7 = 8
Y[2] = ((4)h7 = 4
Y(3] = ((112))17 = IO
N- 1 = 13
w;;' = 13
((N- 1 N)h7 =((W;i WJVJl11 =((13 · 4))17 =((52)),7 = 1
1
401
(e)
n, k,
I 0 \I 12 0 1 2
n, 0 I (• I i2 k, 0 0 2 4
1 j3 j 4j5 1 1 3 5
(i) Let G[k,, n2] be the N = 2 point DFTs of the inner parenthesis; i.e.,
1
G(k 1, n 2] = L z(3n 1 + n 2JW:' "',
n,=O
This calculates 3 DFTs, one for each column of the index map associated with n. Since the
DFT size is 2, we can perform these with simple butterflies and use no multiplications.
(ii) Let G[k,, n2] be the set of 3 column DFTs multiplied by tbe twiddle factors.
0 ~ k, ~ 1,
{ O~n 2 ~2.
(iii) The outer sum calculates two N = 3 point DFTs, one for each of the two -values of k,.
2
X(k, +2k2] =L 1
G(k1,n2]W3""',
n:a=D
The only complex multiplies are due to the twiddle factors. Therefore, there are 10 complex
multiplies. The direct implementation requires N 2 = 62 = 36 complex multiplies (a little less if
you do not count multiplies by 1 or -1).
403
(f) The alternate index map can be found be reversing the roles of n and It; i.e.,
Solutions - Chapter 10
1
10,000 < T < 5N
The minimal N = 2" that satisfies the relationship is
N = 2048
for which
10, ooo Hz < TI < 10, 24o Hz
Thus, Fmin = 10,000 Hz, and Fmax = 10,240 Hz.
10.3. (a) The length of a window is
(c) The most straightforward solution to this problem is to say that since the window length Lis 320,
we need N 2: Lin order to do the DFT. Therefore, a value of N = 512 meets the criteria of N 2: L,
N = 2". However, since the windows overlap, we can find a smaller N.
Since the window advances 40 samples between computations, we really only need 40 114lid samples
for each DFT in order to reconstruct the original input signal. If we time alias the windowed data,
we can use a smaller DFT length than the window length. With N =256, 64 samples will be time
aliased, and remaining 192 samples will be valid. However, with N = 128, all the samples will be
aliased. Therefore, the minimum size of N is 256.
(d) Using the relation
1
A!= NT'
the frequency spacing for N = 512 is
AJ = 16• OOO = 31.25 Hz
512
and for N = 256 is
A! = 16256
•
000
= 62.5 Hz
10.4. (a) Since :z:[n] is real, X[k] must be conjugate symmetric.
X(kj = X"[((-k))N)
We can use this conjugate symmetry property to find X[k] for k = 200.
X[((-k))N] =. X"[k]
X[((-800))10oo] = (1 + j)"
X[200] = 1- j
we find
-211"(200)
n-200 = (1000)(1/20, 000)
= -211"(4000) rad/s
2.-(200)
nooo = (1000)(1/20, 000)
= 2.-(4000) rad/s
Consequently,
1+j
X<(j n) ln=-••<4000) = 20,000
l-j
x.u n) ln=••<toOOl = 20,000
Note that both expressions for X.(j n) have been multiplied by the sampling period T = 1/20,000
because sampling the continuous-time signal :z:.(t) involves multiplication by 1/T.
409
Starting with
2.-k
n. =NT'
we find
211"(-100)
n_,.., =
(1000)(1/10,000)
= -2.. (1000) rad/s
2.-{100)
n,.., = (1000)(1/10, 000)
= 2.. (1000) rad/s
2r(-420)
n_... = (1000)(1/10,000)
= -2r(4200) rad/s
211"(420)
0... = (1000)(1/10,000)
= 2r(4200) rad/s
410
Consequently,
1
Xe(j 0) lns-2r(l000) = 10,000
1
X.(j 0) ln=2r(lOOO) = 10,000
1
x.(j 0) ln=-2r(4200) = 2,000
1
x.(j 0) ln=2r(4200) = 2,000
Note that all expressions for Xe(j 0) have been multiplied by the sampling period T = 1/10,000 because
sampling the continuous-time signal ze(t) involves multiplication by 1/T.
10. 7. The Hamming window's mainlobe is ll.wml = 1~, radians wide. We want
t..w.... < "
s..
~
..
100
L-1 100
L ~ 801
Because the window length is constrained to be a power of 2, we see that
Lmin = 1024
10.8. All window• expect the Blackman 5ati5fy the criteria. Using the table, and noting that the window
length N =
M + 1, we find
Rectangular:
4..
M+1
4.-
= 256
= ~<~rad
64- 25
The resolution of the rectangular window satisfies the criteria.
Bartlett, Hanning, Hamming:
ll.wml = sM..
= 255
. .
= 31.875 ~ 25 rad
The resolution of the Bartlett, Hanning, and Hamming windows satisfies the aiteria.
Blackman:
121'
='M
12..
= 255
= -"-~~rad
21.25 25
The Blackman window does not have a frequency resolution of at least ,. /25 radians. Therefore,
this window does not satisfy the criteria.
411
Clearly, the cosines in z 1 [n] are too closely spaced in frequency to produce distinct peaks.
In z,[n], we have a small amplitude cosine which will be obscnrred by the large sidelobes from the
rectangular window. The peak will therefore not be visible.
The only signal from which we would expect to see two distinct peaks is z2[n].
10.10. The equivalent continuous-time frequency spacing is
1
.O.f = NT
Thus, to satisfy the criterion that the frequency spacing between consecutive DFT samples is 1 Hz or
less we must have
.O.f s 1
1
NT s 1
1
T ?:
N
1
T ?: 1024 sec
However, we must also satisfy the Sampling Theorem to avoid aliasing. We therefore have the addition
restriction that,
1
f ?: 200Hz
1
T S sec
200
Putting the two constraints together we find
_1-<T<-1-
1024- - 200
1
=
Tmin 1024 sec
10.11. The equivalent frequency spacing is
.o.n = =
2.-
NT
2.-
(Sl92)(50ps) = 15.34 rad/s
or
.O.f = -.o.n
2lr
= 2.44 Hz
412
t.J $ 5
1
$ 5
NT
1
N ?: 5T
8000
?: 5
?: 1600 samples
10.13. Since w[n] is the rectangular window and we are using N = 36 we have
JS
X,[k] = L z[rR + m]e-i(b/l&)•m
m=O
= DFI'{z[rR + n]}
Because :r[n] is zero outside the range 0 $ n $ 71, X,[k] will be zero except when r = 0 orr= 1.
V.'hen r = 0, the 36 points in the sum of the DFI' only include the section
ef(if)3n + e-i(if)3n
cos(1mj6) = 2
of :r[n]. Therefore, we can use the properties of the DFI' to find
\\'hen r = 1, the 36 points in the sum of the DFI' only include the section
ef(if)On + e-i(JT)9n
COS(7m/2) =
2
of :r[n]. Therefore, we can use the properties of the DFI' to find
10.14. The signals :r2 [n], :r3 [n], and za[n] could be z[nJ, as described below.
413
Looking at the figure, it is clear that there are two nonzero OFT coefficients at k = 8, and k = 16.
These correspond to frequencies
(2w)(8)
WI =
128
= ~rad
8
(2w)(16)
=
"" 128
~ rad
=
4
Also notice that the magnitude of the DFT coefficient at k = 16 is about 3 times that of the OFT
coeflicient at k = 8.
• z,[n): The second cosine term has a frequency of .26.- rad, which is neither 1r /8 rad or 1r /4 rae!.
Consequently, z 1 [n] is not consistent with the information shown in the figure.
• :t2[n]: This signal is consistent with the information shown in the figure. The peaks occur at the
correct locations, and are scaled properly.
• z,[n]: This signal is consistent with the information shown in the figure. The peaks occur at the
correct locations, and are scaled properly.
• :r,[n]: This signal bas a cosine term with frequency .-/16 rad, which is neither 1f/8 rad or .-;4 rad.
Consequently, z 4 [n] is not consistent with the information shown in the figure.
• :ts[n]: This signal has sinusoids with the correct frequencies, but the scale factors on the two terms
are not consistent with the information shown in the figure.
• :to[n]: This signal is consistent with the information shown in the figure. Note that phase infor-
mation is not represented in the DIT magnitude plot.
w;[n] = w0 + >.n
This describes a line with slope >. and intercept Wo· Thus,
Wo = 0.25.- rad
10.16. Using
1
l!.f= NT
and assuming no aliasing occured when the continuous-time signal was sampled, we find that the fre-
quency spacing between spectral samples is
1
l!.f = (1024)(1/10, 000)
= 9.771b
or
t.fl =21ft./= 61.4 radfs
Method 1: This doubles the number of samples we take of the frequency wria.ble, but does not change
the frequency resolution. The size of the main lobe from the window remains the same.
Method 2: Thi3 improves the fre'luency resolution ,;...,. the ,...;, lobe from the window get. mudler.
Method 3: This increases the time resolution (the ability to distinguish events in time), but does not
affect the frequency resolution.
Method 4: This will decrease the frequency resolution since the main lobe from the window increases.
This is a strange thing to do since there are samples of x[n] that do not get used in the transform.
Method 5: This will only improve the resolution if we can ignore any problems due to sidelobe leakage.
For example, changing to a recta.ngu1ar window will improve our ability to resolve two equal
amplitude sinusoids. In most eases, however, "!' need to worry about sidelobe levels. A large
sidelobe might mask the presence of a low amplitude signal. Since we do not know ahead of time
the nature of the signal we are trying to analyze, changing to a rectangular window may actually
make things worse. Thus, in general, changing to a rectangular window will not necessarily increase
the frequency resolution.
10.18. No, the peaks will not have the same height. The peaks in V2(&"') will he larger than those in Vi (e;"').
First, note that the Fourier transform of the rectangular window bas a higher peak than that of the
Hamming window. H this is not obvious, consider Figure 7.21, and recall that the Fourier transform of
an L-point window tu[n], evaluated at DC (w = 0), is
L-l
W(&O) = L w{n]
"""
Let the rectangular window he WR[n], and the Hamming window he WH(n]. It is clear from the figure
(where M =
L+1) that
L-l L-1
L WR{n] > L WH{n]
Therefore,
""" .....
WR(&O) > WH(.,.i 0 )
Thus, the Fourier transform of the rectangular window has a higher peak than that of the Hamming
window.
Now recall that the multiplication of two signals in the time domain corresponds to a periodic convolution
in the frequency domain. So in the frequency domain, V1 (&"') is the convolution of two scaled impulses
from the sinusoid, with the Fourier transform of the L-point Hamming window, WH(&"'). This results
in two scaled copies of WH(ei-'), centered at the frequencies of the sinusoid. Similarly, V2 (&"') consists
of two scaled copies of W R ( ei"'), also centered at the frequencies of the sinusoid. The scale factor is the
same in both cases, resulting from the Fourier transform of the sinusoid.
Since the peaks of the Fourier transform of the rectangular window are higher than those of the H•mming
window, the peaks in V2 (&"') will be larger than those in V1 (ei"').
10.19. Using the approximation given in the chapter
_ 24.-(A., + 12)
L- 155Am~ + 1
we lind for A., = 30 dB and A,., = ~ rad,
L :: 24.-(30 + 12)
155(.. /40) +1
= 261.1 .... 262
415
10.20. (a) The best sidelobe attenuation expected nnder these constraints is
24.-(A,, + 12)
L :::
155.0.,.. +1
24r(A., + 12)
512 :::
155(.-/100) +1
A,, ::: 21 dB
(b) The two sinusoidal components are separated by at least ..-{50 radians. Since the largest allowable
mainlobe width is .-/100 radians, we know that the peak of the DFT magnitude of the weaker
sinusoidal component will not be located in tbe mainlobe of the DFT magnitude of the stronger
sinusoidal component. Thus, we only need to consider the sidelobe height of the stronger compo-
nent.
Converting 21 dB attenuation back &om dB gives
-21 dB = 20log10 m
m = 0.0891
Since the amplitude of the stronger sinusoidal component is 1, the amplitude of the weaker sinu·
soidal component must be greater than 0.0891 in order for the weaker sinusoidal component to be
seen over the sidelobe of the stronger sinusoidal component.
10.21. We have
v[n] = cos(2,.n/5)w[n]
ei2•n/• + .-;..n/5]
= [ 2
w[n]
Mainlobe Height of W(eiw): The peak height is at w = 0 for which we can use l'hopital's rule to find
W(ei 0 ) = 32cos(16w) I
cos(w/2) w=O
= 32
Strongest Sidelobe height of W(eiw): The strongest sidelobe height for the rectangular window is
13 dB below the main peak height. Therefore, since 13 dB = 0.2239 we have
First Nulls of W(eiw): The first nulls can be found be noting that W(eiw) = 0 when sin(16w) = 0
Thus, the first nulls occur at
2.-
w=±-
32
Therefore, JV(eiw)llooks like
416
16 I V(J">} I
-2KIS + 2Jrl'32
2><15 + 2lri32
_,
" Ill
Note that the numbers used above for the heights are not exact because we are adding two copies of
W(e'w) to get V(eiw) and the exact values for the heights will depend on relative phase and location of
the two copies. However, they are a very good approximation and the error is small.
10.22. The 'instantaneous frequency' of :r[n], denoted as >.[n], can be determined by taking the derivative
with respect to n of the argument of the cosine term. This gives
0.5
0.45
0.4
0.35
0.3
..,"
~ 0.25
0.2
0.15
0.1
..
0.05 ·--
0
2000 4000 6000 8000 10000 12000 14000 16000
Sample number (n)
417
Here, we see a cosine plot shifted up the frequency (>./27r) axis by a coustant. A$ is customary in a
spectrogram, only the frequencies 0 ::; >.f2" :S: 0.5 are plotted.
10.23. In this problem, we relate the DFT X(k] of a discrete-time signal z[n] to the continuous-time Fourier
transform X.(jO) of tbe continuous-time signal z.(t). Since z(n] is obtained by sampling z.(t),
z(n] = z.(nT)
we find
'Xc: (·2d)
'f" 1 "FiT ' forO:S:k< 'f
X[k] = { 1X ( ·2•<•-Nl)
~ c: J NT I for~:S:k:S:N-1
Breaking up the DFT into two terms .l.ilo< this is nnececessaJuy to relate the negative frequencies of X.(jO)
'f
to the proper indicies S k S N - I in X(k].
Method 1: Using the above equation for X(k], and plugging in values of N = 4000, and T = 251JS, we
find
x,[k] = { 40,ooox. (;2" ·10 · k), foro::; k s 1999
40, ooox. (j21f · 10 · (k- 4000)), for 2000 :s; k S 3999
Therefore, we see this does not provide the desired samples. A sketch is provided below, for a
triangular-shaped X.(jO).
-1 ...... .t.t=10Hz
418
---1 f- Af=5Hz
---1 f- Af=5Hz
(19
10.24. (a) In this problem, we relate the DFT X[k] of a discrete-time signal :t[n] to the continuous-time
Fourier tranform X.(j!l) of the continuous-time signal :z:.(t). Since :z:[n] is obtained by sampling
:r.(t),
=
:r[n] :tc(nT)
X(ei~) = ffr=-oc
X. (if+ i ;r)
2
for -1r ~ w :5 1r
which is equivalent to
forO:Sk:SN-1
we find
'x(·2••)
,.c JW'f ' forO:Sk<~
X[k) = {
1 X ( ·2•~TN!) for~:Sk:SN-1
,. c J '
Breaking up the DFT into two terms like this is necessary to relate the negative frequencies of
lf
X,(jn) to the proper indicies :S k :S N- 1 in X[k].
Y[k] = aX,(j2,. · 10 · k)
is correct. To understand the effect of each step in the procedure, it helps to draw some frequency
domain plots. Assume the spectrum of the original sit,nal :t,(t) looks like
420
Q
-~~(,~~~~)---------+--------~~~,~~)
Sampling this continuous-time signal will produce the discrete-time signal z[n], with a spectrum
X(~ .
o X{k] = Samples of X( el "')
Next, we form
X[k], O$k$250
W[k] = 0, :s; k :s; 749
251
{
X[k], 750 :s; k :s; 999
and find w[n] as the inverse DFT of W[k].
421
W[k]
1/T
Before going on, we should plot the Fourier transform, W(&~), of w[n]. It will look like
W(r!")
1/T
1-----....:>.I.~.D.I.~.D.I.~.D.i.:__ _ _ _ .,
21t
"
W(ei~) goes through the DFT points and therefore is equal to samples of X,(jO) at these points
for 0 ::; k ::; 250 and 750 ::; k ::; 999, but it is not equtJ to X,(jl1) between those frequencies.
Furthermore, W(e'~) = 0 at the DFT frequencies for 251 :=; k ::; 749, but it is not zero between
those frequencies; i.e. we can not do ideallowpass filtering using the DFT.
Now we define
¥[n] = { w[2n], 0 ::; n ::; 499
0, 500 ::; n ::; 999
and let Y[k] be the DFT of y[n]. First note that Y(~) is
Y(ei"')
~~~~~~~~~~~~~~~~~w
we plug in
M
=L
to get
y[n + m]
.... h(k]z[n + m - k]
oo M
Y[n, ~) = L L h[k]z[n + m- k]w[m]e-;>m
M oo
= L h(k] L z[n + m- k]w[m]e-;•m
le-o --oo
M
= L h[k]X[n -
,... k, ~)
= h[n] • X[n, ~)
where the convolution is for the variable n.
(b) Starting with
we find
H the window is long compared toM, then a small time shift in X[n, >.) won't radically alter the
spectrum, and
X[n- k, >.) ::: X[n, >.)
Consequently,
M
Y[n,>.) ::: L h[k]e-i'" X[n, >.)
::: H(ei')X[n, >.)
10.26. Plugging in the relation for c,.[m] into tbe equation for /(w) gives
Note that for all values of 0 ~ n ~ L- 1, the second summation will be over all non-zero values of v[l]
in the range 0 s; l s; L - 1. M. a result,
L-1 . L-1 _
= -U L v[n]e'wn L v[l]e-'"''
1
I(w)
L n=O l=<l
1
= - V"(ei"')V(ei"')
LU
1
= - [V(ei"')[ 2
LU
Note that in this analysis, we have assumed that v[n] is a real sequence.
424
10.27. (a) Since z[n] has length L, the aperiodic function, c,,[m], will he 2£- 1 points long. Therefore,
in order for the aperiodic correlation function to equal the periodic correlation fuction, C..[m], for
0 $ m $ L- 1, we require that the inverse DFT is not time aliased. So, the minimum inverse
DFT length N min is
Nmin 2L-1 =
(b) H we require M points to he unaliased, we can have L- M aliased points. Therefore, for C.,[m] =
c..[m] for 0 $ m::;; M- 1, the minimum inverse DFT length Nmin is
Nmin = 2£- 1 - (L - M)
= L+M-1
be a scaled rectangular pulse. Then we can write the aperiodic autocorrelation as,
""
= L wx(k- m]wx[k]
m =-(M- 1) case
0 WR(k) M-112
•. wR[k+(M-1))
k
T
k=-(M-1)
Consider m = (M- 1). This is last value of m for which the two signals overlap.
m=(M-1)case
0 WR(k) M-112
• WR(k-(M-1)]
T k
k=M-1
T m
m=-(M-1)
1-lmi/M, lmi:SM-1
I J
wsm ={ 0, otherwise
From part (a), we know that the Bartlett window can be found by convolving WR[m] with WR[-m].
In the frequency domain, we therefore have,
Ws(ei"') = WR(el"')WR(e-i"')
= [ 1 sm(wM/2) -jw(M-1112] [ 1 sin(-wM/2)eiw(M-1)/2]
,fM sm(w/2) • ,fM sin(-w/2)
2
= _!._ [sm(wM/2)]
M sm(w/2)
(c) The power spectrum, defined as the Fourier transform of the aperiodic autocorrelation sequence,
is always no1111egative. Thus, any window that can be represented as an aperiodic autocorrelation
sequence will have a DOilllegative Fourier transform. So to generate other finite-length window se-
quences, w[n], that have no1111egative Fourier transforms, simply take the aperiodic autocorrelation
of an input sequence, :[n].
we lind
. 1_ -;... (2(M-1)+1]
= e'w(M-1) 0
1- e-Jw
.
. 1- e-jw(2M-l)
= e'w(M-1) .
1 e ''"'
eiw(Jt-1) - e-jwltl
= 1- e-iw
0 -;w/2[eiw(M-1/2) _ 0 -;...(M-1/2)]
= 0 -jw/2[eiw/2 _ 0 -;.../2]
.;w(M-1/2) _ 0 - ;...(Jl-1/2)
= eiw/2 _ e-jw/2
= 2jsin{w(M- j)]
2j sm(w/2)
= sm[w(M- t)]
sm(w/2)
427
or
. sin[w2Al-1J
w,(e'"') = sin(wi2)
where 2M- I is the window length. A sketch of W,(ei"') appears below.
-21t/(2M-1) 21t/(2M-1)
Bartlett (triangular): Ws(.,;"') is the Fourier transform of a triangular signal,
w~ m]
1
-(M-1)
G>i
(
0
II
Ii
;>
i9
M-1
m
...- 112
m
-(M-1)/2 0 (M--1)/2
with itself. Tha.t is, wa[m] = :[m] • :[m].
Above, we found the Fourier tranform of a rectangular window, as
where 2M - I was the length of the window. We can use this result to find the Fourier
transform of :[m]. The signal :~:[m] is similar to the rectangular window, the difterence being
428
X(ei"') = 1 sin(wM/2)
.,fM sin(w/2)
=
The time domain convolution, w8 (m] o:(m]•o:(m] corresponds to a multiplication, Ws(ei"') =
(X(ei"')] 2 in the frequency domain. As a result,
2
Ws(ei"') = (X(ei"'))
2
= ..!_ [sin(wM/2)]
_M sin(w/2)
A sketch of Ws(ei"') appears below.
0 It
-
(I)
-It It
429
(b) Rectangular: The approximate mainlobe width, and the approximate variance ratio, F, for the
rectangular window are found below for large M.
In part (a), we found the Fourier transform of the rectangular window as
1
1 (M- ) ( lml) 2
F = Q L
--CM-1)
1-M
= .!.
Q.
[2 'f: (1- !!:)' -1]
1
,...., M
1 [ M-14 2 M-1 M-1 ]
:E 1 -M....,
= -Q 2 m=O - :Em+- :E m• -1
M',....
= .!_ [2M_ 4(M- 1}M 2(M- 1}M(2M- I) _
Q 2M + 6M2
t]
:: h [2M-2M+ 2~]
2M
::
3Q
HanningfHamming-. We can approximate the mainlobe bandwidth by analyzing tbe Fourier
transform derived in Part (a). Looking at one of the terms from this expression,
we note that tbe numerator is zero whenever the its argument equals wn, or
mr "
"' = M- {1/2) + M- 1
nw .-
:: -
M+M
-
O'(n +I)
::
M
So the mainlobe bandwidth for this term is
4Mainlobe bandwidth :: ~
2
Mainlobe bandwidth :: ""
ii
Note tbat tbe peak value for this term occurs at a frequency w:: 1r/M.
A similar analysis can be applied to tbe other terms in Fourier transform derived in Part {a).
The mainlobe bandwidth for the term
sin[w (M- ! )]
a .
SlD(w/2)
431
is also 2" fM. Note that the peak value for this term occun at a frequency w =0.
A sample plot of these three terms, for {J = 2o and large M is shown below.
2
1 .,_, ( ( ""' ))
F = Q L
m=-(M'-1)
o+{Jcos M-1
.!.[ ~ a +2o{J ~
1 1
=
Q m=-(M-1)
2
m=-(Al-1)
cos(::\)+{J
2
'f cos•(;:
m=-(Al-1)
1 )]
F = -1
Q
[
L
Al-l
=-<M-1)
o 2 +2o/3
m=-(M-1)
L
Al-l
cos
(
~1
M-
)
Noting that
Jl-1 ( )
L cos M":1 = -1
=-(Al-l)
•
Jl-l ( 2nn )
L: cos M-1 = 1
m=-{AI-1}
432
we conclude
2 2
F = .!_ [(2M- I)Q2 - 2Q/3 + /3 (2M- I)+ /3
Q 2 2
]
~ 2:(Q·+~·)
10.30. (a) Using the definition of the time-dependent Fourier trausform we find
13
X(O,k] = L :[m]e-j(2•/7)•m
m=O
• 13
= L :(m]e-;(2•/T)•m + L:(l]e-j(2•/1)"'
l=1
•
= L (:[m] + :[m + 7])e-j(2•/7)•m
m=O
By plotting :[m]
-'{mj
•••
-1 0 1 2 3 4 5 6 7 8 9 10 n
X[O,k] = L• (I)e-j(>•/T)•m
m=O
= l>J'T{l}
= 7o(kJ
(b) If we follow tbe same procedure we used in part (a) we find
13
X(n, k] = L :(n + m)e-;(2•/T)>m
m=O
• 13
= L :(n + m)e-;(2•/T)>m + I:=!n + ije-;(2•/7)>~
m=O l=?
DTFT: IX (ei"')l
o OFT:IX}JI
434
The maximum possible error, Omax error, of the frequency estimate is one half of the frequency
resolution of the DFI'.
1 2.-
flmax error =
=
.
2NT
NT
For the system parameters of N = 32, and T = 10-4 , this is
Omax error =982 rad/s
(d) To develop a procedure to get an exact estimate Of flo, it helps to derive X,.[k]. First, let's find
the Fourier transform of :t,.[n] = :t[n]w[n], where w[n] is anN-point rectangular window.
N-l
x.(~'"') = L eJwone-'"'"
n=O
N-1
= L e-j(w-wo)n
......
Let w' =w- wo. Then,
N-1
X.,(eiw) = L .-jw'n
......
1- e-iw'N
= 1-e jw'
{eiw' N/2 _ .-;w' N/2)e-jw' N/2
= (e jw'/2- e-iw'/2)e '""''
= sin.(w'N/2) -;w'(N-1)/2
sin(w' /2) •
= sin[(w- .,.)N/2] -j(w-wo)(N-1)/2
sin[(w- .,.)/2] •
Note that X,.(eiw) has generalized linear phase. Having established this equation for X,.(eiw),
we now find X.,[k]. Recall that X,.[k] is sintply the Fourier transform X,.(eiw) evaluated at the
frequencies w = 27fk / N, for k = 0, ... , N - 1. Thus,
(21rk/N- .,.)(N- 1)
LX ,. [kl = 2
+m"
where the m,. term comes from the fact that the term
lin{(2d:/N -.,.)N/2]
lin{(2.,kfN ""')/2]
can change sign (i.e. become negative or positive), and thereby offset the phase by .- radians. In
addition, this term accounts for wrapping the phase, so that the phase stays in the range [-1r, .-].
435
m,.... = ;(ux...[k.....J+C";=-Wt!)(N-1)]
In these equations, W. is the estimate found in Part (c). So we would look for values of m in the
range [lm..,nJ, fm,....l]. Similar expressions bold for p.
Once WQ is known, we can find flo using the relation flo =WQ /T.
10.32. For each part, we use the definition of the time-dependent Fourier transform,
00
00
Y[n, >.) = L
.. y[n + m]w[m]e-i-'m
m=-oo
436
00
L
00
= .;wo(n+mlz[n + m)w[m)e-;>m
m.=-oo
(c) The condition is that no aliasing occurs when sampling. Thus, we require that Pc(fl) = 0 for
101 ;::: f so that
1 .
P(w) = TPc (f), lwl <"
10.34. In this problem, we are given
• z[n] = A cos(Won + 8) + e[n]
• 8 is a nniform random variable on 0 to 2.-
• e[n] is a.n independent, zero mean random variable
437
(b) Since the Fourier transform of cos(wom) is ..O(w- wo) + .-o(w + wo) for !w! ::; "•
ftr [I JV(k]! 2
] ::: .P;.(w)
2
var [!V(k)! ] ::: L 2 .P;.(w)
438
This equation can be used to find the approximate variance of JX(k]J 2 • We substitute the signal
X(k] for V(k], the OFT length N for L, and use the power spectrum
l _ WN(t-r)]
= 0'2
z [ 1- "'
w~A:-r)
= Nu~6(k- r]
Note that the cross-correlation is zero everywhere except when k = r. This is what one would
expect for white noise, since samples for which k # r are completely uncorrelated.
10.36. (a) The length of the data record is
(b) To achieve a 10 Hz or less spacing between samples of the power spectrum, we require
1
NT ::::
10Hz
1
N ?
lOT
20,000
? 10
? 2, 000 samples
Since N must also be a power of 2, we choose N = 2048.
(c)
Q
K = L
200,000
= 2048
= 97.66 segments
H we zero-pad the last segment so that it contains 2048 samples, we will have K = 98 segments.
(d) The key to reducing the variance is to use more segments. Two methods are discussed below.
Note that in both methods, we want the segments to be length L = 2048 so that we maintain the
frequency spacing.
439
(i) Decreasing the length of the segments to /oth their length, and then zer<>-padding them to
=
L 2048 samples will increase K by a factor of 10. Accordingly, the variance will decrease by
a factor of 10. However, the frequency resolution will be reduced.
(ii) U we increase the data record to 2,000,000 samples, we can keep the window length the same
and increase K by a factor of 10. ·
10.37. (a) Taking the expected value of
~[m] = 21 ~· I(w)_,j"'mdw
gives
" -·
E {¢[m]} = E{
2~ L I(w).,;"'md<u}
= 2
1
, f. E {I(w)} ei'-dw
we find
= _1_1. [.!..f"
2" LU -•
P•• (e)
2r -•
c_(.,;I.,-Bl)_,j"'mdw] d8
Note we can change the limits of integration of the inner integral to be [-", "] because we are
integrating over the whole period. Doing this gives
=
2" LU -•
21r~U L: P.. (e).,;•m
2" -•
{~~[m]}d8
= L~c_[m] ( 2~ £: 1
P•• (e)_,j md8]
1
= w<-[m];•• [m]
(b)
¢p[m] ! ....
N-1
= 2: I[k]e'"'.."''N
440
~p(m]
..
By applying the sampling theorem to Fourier transforms, we see that
= L ~.. (m+rNJ
E ~p[ml} =
..
r=-oo
L E ~•• (m + rNJ}
1 ..
= LU L c,.,.(m+rNJ¢,,(m+rNJ
~-oo
I [M-1
= Q L :r(n]:r(n + m]
.....
M-1 M-1 ]
+ L :r(n + MJ:r(n + M + m] + ... + L :r(n + (K- 1}MJ:r(n + (K- 1}M + m]
.....
K-1 JJ-1
.....
= ~ LL :r(n + iM):r(n + iM +m]
i=O n=O
K-1
= .!. L e;(m]
Q...,
where
M-1
e;(m] = L :r(n + iM]:r(n + iM + m] forO:Sm:SM-1
.....
(b) We can rewrite the expression for e;(m] from part (a) as
M-1
e;(m] = L :r(n + iMJ:r(n + iM + m]
.....
Jl-1 /11-1
= L :r(n+iM]z[n+iM +m]+ L O·z[n+iM +m]
....0 =M
/01-1
= L :z:;(n]ll;(n +m]
.....
where
.r;[n] = { .r(n + iM], O:Sn:SM-1
0, M:Sn:SN-1
441
and
y,[n] = z[n + iM] forO~n~N-1
Thus, the correlations e;[m] can be obtained by computing N-point l•ntJJr correlations. Next, we
show that for N ;<: 2M - 1, circular correlation is equivalent to linear correlation.
can be expressed as
Cy.[m] = ~.[-m]
N-1
= L z,[((n- m)),,jy,[n]
=<>
N-1
= :E :ra((m- n))N]Y,[n]
-=0
where :r'[n] = :r[-n]. Note that this is a circular convolution of :r,[-n] with y,[n]. Thus, we have
expressed the circular correlation of z,[n] with y,[n] as a circular convolution of :r,[-n] with y,[n].
Now recall from chapter 8 that the circular convolution of two M point signals is equivalent to
their linear convolution when N ;<: 2M - 1. Since we can express the circular correlation in terms
of a circular convolution, this result applies to circular correlation as well. Therefore, we see that
if N ;<:2M -1,
=
e;[m] C.[m] forO~m~M-1
step 4: Take theN point inverse DFI' of +•• [k] to get ~•• [m].
Assuming that a radix-2 FFI', requiring ~log, N complex multiplications is used to compute the
forward and inverse DFTS, the number of complex multiplications is
c[n,m]
X[n, .\) =
..
L z[n + m]w[m)e-iM•d.\
-=-oo
443
we find
c(n,m] = 2.1•
2• -·
IX[n, A)j 2 ei'md>.
=
00
L L
l=-ocr=-oo
00 x[n + l]w[l]x[n + r]w[r] ( -1
21r
1" .
-'II'
)
.-J>(-I+rlei>md>.
The 6[m - I + r] term is zero everwhere except when m - l + r = 0. Therefore, we can replace the
two sums of land r with one sum over r, by substituting l == m + r.
00
= L
r=-oc
x[n + r]w[r]:r[n + m + r]w[m + r]
..1"1'-·
1
c(n, m] = IX[n, .>.)1 2 .,J>m d>.
2
c[n, -m] = -1
2
. -· IX[n, >.)J 2 .-J.,md>.
00
where
h,.{r] = w[-r]w[-(m + r)]
(c) To compute c[n,m] by causal operations, we see that
h,.{r] =w[-r]w[-(m + r)]
requires that w[r] must be zero for
-r < 0
r > 0
and w[r] must be zero for
-(m+r) < 0
m+r > 0
r > -m
Thus, w{r] must he zero for r > min(O, -m). If m is positive, then w[r] must he zero for r > 0.
This is equivalent to the requirement that w[-r] must he zero for r < 0.
(d) Plugging in
00
= L:o2r+mz-r
..... 00
= am L (a z- 2 1
)'
......,
445
= L
r=-oo
z[r]z[r- m] (c'"6[n- r] + c 2 h,[n- r - 11)
00
z-m
[ ] -{ ra', r;o:O
w-r-
0, r <0
To get the z-transform H,.(z), recall the z-transform property: rz[r] +-> -z•~•l. Using this
property, we find
ra2'u[r] ~
(1-a•z 1)2
a 2 z- 1(1 +a2z-1)
(1-a•z 1)3
c{n, m] = L
r=-oo
:r[r]:r[r- m]h,[n- r]
we get
00
y(n]
x[n] ~ + + c[n,m)
-1 z-1
z am+2(m+ 1)z-m 3a2
z-1
z-1 -3a4
X
z-1 am+4(1-m)z-m a6
z-1
= L :r[n - m]ho(m].,;>.m
m.-=-oc
Note that most typical window sequences are lowpass in nature, and are centered around a fre-
= =
quency of w 0. Since H 0 (eiw) W(e-iw) is the Fourier transform o£ a window which is lowpass
in nature, the signal S (e)w) is also lowpass.
The signal s[n] = X[n, .>.) is mnltiplied by a complex exponential ei'". This modulation shifts the
frequency response of S(eiw) so that it is centered at w = >..
= s(n]ei'"
h{n]
H(eJw) = S ( ei(w->))
Since S(eiw) is lowpass filter centered at w = 0, the overall system is a bandpass filter centered at
w = >..
448
(c) First, it is shown that the individual outputs Y•[n] are samples (in the,\ dimension) of the time-
dependent Fourier transform.
00
= L z[n + m]w[mje-jbkm/N
m=-oo
= X[n, A)J.=2d/N
Next, it is shown that the overall output is y[n] = Nw[O]z[n].
N-1
y[n] = L Y•[n]
k=O
N-1 oc
= L L z[n + m]w[m]e-j2•>m/N
i:=O m.=-oc
CICI N-1
= L L x[n + m]w[m]e-j2dmfN
N6JmJ
= Nw[Ojx[n]
(d) Consider a single cha.nnel,
decimator expander
x[n]---{
X ( ei<"'+>•l) Ho(ei"')
so the output of the decimator is
R-1
~~X (ei<<w-2•1)/R+>•l) Ho (ei<w->wi)/R)
R-1
.... N-1
+ LX (ei<w-2•1/R)) ~ L Go (ei(w->•)) Ho (ei<w->•->•1/R))
l=l k=O
Aliasing Component
(e) Yes, it is possible. G0 (eJw) = NH0 (&w) will yield exact reconstruction.
(f) See chapter 7 in "Multirate Digital Signal Processing" by Crochiere and Rabiner, 1983.
(g) Once again, we consider a single channel,
decimator expander
From Part (a), we know that the output of the filter ho(n] is
00
Therefore,
Now recall that I::~· ei2<k(n-m)/N = No[((n- m))N]. by considering it as a Fourier series ex-
pansion, or as an inverse DFT of Ne-;2 wmJ:fN. Thus,
N-1 co
L eJ2d(n-m)/N =N L 6[n- m- rNJ
k=O r=-oo
= N L L go[n-lR]ho[lR-n+rNJ:~:fn-rNJ
l:-oor=-oo
co 00
..
Therefore, if we want y[n] = z[n], we require
y[n] =L h[k]z[n- k]
k=-oo
= £Ct... 00
h[k]z[n + m- k] ,t. h[l]z[n- 1]}
= L L h[k]h[l]£{ z[n + m - k]z[n - I]}
.b:-oo I= -oo
.. 00
= L L h[k]h[/]9'>•• [1 + m- k]
~-ool=-oo
q,.. [l + m- k] =.,.;o[l + m- k]
Substituting this into the expression for q,,.[m] gives
00 00
Note that 00
(b) Taking the DTFT of 1/>n[m] will give the power density spectrwi. t,.(w).
(c) This problem can be approached either in the time domain or the z-transform domain.
Time domain: Since all the ai: 's are zero for a MA process,
Jl
y[n] =L bto:[n - k]
t=O
The relation for ~ .. (z) above is found by multiplying two polynomials in z. The highest power
= =
of z in ~ .. (z) is zM which arises from tbe multiplication of the k 0 and I M coefficients.
The smallest power of z in ~,.(z) is z-M wbich arises from the multiplication of the k = M
and I= 0 coeflicents. Thus, ,.,(m] is nonzero only in the interval 1m! SM.
(d) For an AR process,
H(z) =
= IIi':, (1 - a•z-')
Since
~ .. (z) = ~H(z)H"(z)
~ •• (z)
--
-------~~·~----~
= rrl=1(1-
N
o.z- 1)(1- o;z)
Thus, the poles for ~•• (z) come in conjugate reciprocal pairs. A sample pole-zero diagram appears
below.
X X
lm
Re
Nth order zero X
Nth order zero at z = ~
X X
By performing a. partial fraction expansion on ~ •• (z) we find that each pole pair contributes a.
sequence of the form A.ol;' 1
~c{ml
~·.
~ M
and therefore
N
,,,[m] = L A.al"'t
l=l
454
H(z) = Y(z) = 1
X(z) 1- ~ 1 ....-•
which means that
N
y[n] = L 4tll[n - kJ + :r:[n]
0=1
The autocorrelation function is then
¢.,[m] = ¢.,[-m]
= £ {y[n - m]y[n]}
= £ { y[n- m] (t.aw[n- kJ + :z:[n])}
N
= L at£ {y[n - m]y[n- k]} + £ {y[n - m]:z:[n]}
t=t
N
= L 4t9\yy[m- k] + 9'>,.[-m]
t=l
N
= L ...;.. [m- k] + ¢,.[m]
1:=1
Form= 0,
N
q,.. [o] = L ...¢.,[-/.:] + ¢•• [o]
bt
¢.,[0] = £ {:z:[n]y[n]}
Note that :r:[n] is uncorrelated with the y[n- k], for k = 1, ... , N. Therefore,
¢.,[0] =.;
Thus,
N
q,..roJ = 1:...q,..r-~.:J +.;
lost
N
= 1: ...q,..r~.:J +<r!
t=l
455
¢yy(m - k] = ¢n(k - m]
= ¢.,[1m- kl]
Thus,
N N
L a.¢.. (1m- kl] = L a.¢.. (m- kj
bl .k=l
form= 1, 2, .. . ,N.
x(n] = x,(nT)
1/2
••• •••
1/4
However, since the period we use in the sum of the IDFS is unimportant we ·can also write
,.
z[n] = _!_ L X[k]~(2•/16)•n
16 ....
= IDFS{X[k]}
= IDIT{X0 [k]}
where Xo[k] is the period of X[k] starting at zero, i.e.,
G[kj
112 112
114 1/4
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 k
q[nj = =c (n:n
= ~ 2:• (~ )1•1 ~<••/32)..
k=-•
We can apply the same idea as we did in part (a), except now the DFS and DIT siu should he
32 instead of 16. Going through the same steps will lead us to the sequence Q[k] that looks like:
Q{kj
1/2 112
1/4
12 20 31 k
0 4 8 16 24 28
457
(Here we have assumed a = 1). We see that we can interpolate in the time domain by zero padding
in the middle of the DIT samples.
10.43. (a) Using the relation,
• O$.k$.N/2
f• = { 11'f•
·-N
ll'r• N/2$.k$.N
where N is the DIT length and T is the sampling period, the continuous-time frequencies corre-
sponding to the DIT indices k = 32 and k = 231 are
32
/32 = (25Q){1/20, 000)
= 2500Hz
231-256
!731 = (256)(1/20, 000)
= -1953Hz
{b) Since
z(nj = z(n]wR(nj
the DTFT of z[n] is simply the periodic convolution of X(ei"') with WR(e'"').
(c) Multiplication in the time domain corresponds to periodic convolution in the frequency domain, as
shown in pa.rt (b). To evaluate this periodic convolution at the frequency w32 = 27r(32)/L, (where
L = N = 256) corresponding to the k = 32 DIT coeflicient, we first shift the window w.,,(ei"')
to w, 2. Then, we multiply the shifted window with X(ei"'), and integrate the result. In order for
!
1, "'= 0
w.,,(e'"')= a, w=±21rfL
0, 21fkjL, fork= 2,3, ... ,£- 2
Note that we are ouly specifying w.,,(ei"') at the DFT frequencies w = 21fk/ L, fork = 0, ... , L-1.
(d) Note that the L point DIT ofa rectangular window of length Lis
L-1
WR[k] =
=
-~)1)e-i2d/L
l-ej2•IIL
= U[k]
w.,,(ei"') is only specified at DFT frequenciesw = 2rkfL, and it can take on other values between
these frequencies. Therefore, the DTIT of w.,,(ei"') can be written in terms of WR(ei"') and two
shifted versions of WR(e'"').
w ••,[n]
21256
11256
0~~----~------~------~----~~ n
0 64 128 192 256
10.44. (a) After the lowpass filter, the highest frequency in the signal is /:w. To avoid aliasing in the
downsampler we must have
twM ~
",..
M ~
6w
N
~
2kA
N
Mmax=--
2kA
(b) The fourier transform of x,[n] looks like
so M = 6 is the largest M we can use that avoids aliasiag. With this choioe of M the fourier
transform of x,[n] looks like
459
-It 0 It (I)
Taking the OFT of z,[n] gives usN samples of X,(eiw) spaced 27f/N apart in frequency. By
examining the figures above we see that these samples correspond to the desired samples of X(e;w)
which will he spaced 2.0.W/N apart inside the region -.0,., < w < t:.w.
Note that after downsampling the endpoints of the region alias. Therefore, we cannot trust the
values our new OFT provides at those points. However, the way the problem is set up we already
know the values at the endpoints from the original OFT.
(c) The system p{n] periodically replicates XN[n] to create XN[n]. Then, the upsampler inserts M -1
zeros in betweeen each sample of XN[n]. Thus, the samples k, - k.:.. and k, + k.:.. which border
the zoom region in the original OFT map to M(J;.- k.:..) and M(k, + k.:..). The system h[n] then
interpolates between the nonzero points filling in the "missing" samples. Since the linear phase
= =
filter is length 513 it adds a delay of M/2 512/2 256 samples so the desired samples of XN M[n]
now lie in the region
where
k; = MJ;.+256
k~ = Mk.:..
-It It (I)
460
~[k)
After periodically replicating and upsampling by M we have a signal that looks lilre
M-1 zeros
. .. •••
0
1 ) n
Filtering by h[n] then interpolates between the samples. XNM[n] is shown below if we assume that
h[n] is the ideal zero phase filter. The points with an x correspond to the interpolated points.
~[n]
interpolated points
.. . ~ •••
0 M(N-1) n
where
k'c
k;,.
Solutions - Chapter 11
11.1. Using the fact that :t,[n] is the inverse transform of'R£ we get
JU{X(.;w)} = 2 _ ,..;w _ .u-iw
Im{X(.;w)} = 2osinw
1
:o(n] =o(n- 1]- -6(n-
2
2]
wbicb satisfies all the coDStraiDts. Tbe idea bebiDd this choice is that casc•diDg a sigual with an allpass
system does Dot change ~be maguitude squared fespoDSe.
ADotber choice that works is X(.;w) = !{1- 2e-'"'Je-i~ for which we get
1
:z:(n] = 26(n- 1]- 6(n- 2]
Tbe idea bebiDd this choice was to flip tbe zero to its zeciprocallocati<m outside tbe unit circle. This
has the same maguitude squared respouse up to a scaliDg factor; heuce, the l term.
466
Thus,
and
Im{X(~"')} = 0.
About Notation: XR(~"') with a capital R is the real part of X(ei"'). X.(e'"') with a small r is the
conjugate symmetric part of X(~"') which is complex-valued in general.
11.5. The Hilbert transform can be viewed as a filter with frequency response
X;(~"') = H(~"')Xr(~"')
= -j.-6(w- wo) + j1r6(w + wo)
:z:;(n] = sinwon
0 $ "' :S "'•
-we:Sw$0
"'• :S lwl :S ,.
Taking the inverse transform yields
'2~-wc
1
"'. [n J = - ,.,. r. ._. . . _. r· ._;. . .
""' - -1
2r 0
,.,. ~ = 1-cosw.n
""'
wn
467
3 3
z.[n] = -26[n + 4] + o[n + 1]- o[n- 1] + 26[n- 4]
Because z[n] is real and causal we can recover most of x[n], i.e.,
6 = X(tJ"ll.,=<>
""
I: z[nje'"(OJ
= n.=-oc
= z[O]- 2 + 3
Plugging this into our equation for z[n] we find
(b) No, the answer to part (a) is not unique, since any choice for z(O] will resiilt in a correct solution.
11.8. Using Euler's identity and the fact that z 0 (n] is the inverse transform of jXJ(~"') we find
jX1(~) = 3jsin2w
= 3(~""'-;·-j""')
3
z.(n] = 2(6(n + 2]- o(n- 2])
Because z(n] is real and causal we can recover all of :r(n] except at n =0,
z(n] = 2%0 (n]u(n] + z(O]o(n]
= -36(n - 2] + z(O]o[n]
Therefore,
z[n] + z(-n]
:r,[n] =
2
(-36(n - 2] + z[O]o[n]) + (-36[n + 2] + z[O]o[n])
= 2
3 3
= - o[n + 2] + z[O]o(n] - 2 o[n - 2]
2
Using the fact that XR(e;"') is the transform of z,[n] we find
. 3 .... 3 ....
XR(e'"') = -2e' + :r(Oj- 2e-'
= z[O]- 3cos2w
Thus, XR2(~"') and XR3(ei"') are possible if :r(O] = -1 and z[O] = 0 respectively.
11.9. (a) Given the imaginary part of X(ei"'), we can take the inverse DTIT to find the odd part of z[n],
denoted x0 [n].
JH(z)l
2 = (1- !z-1
)(1- !z)
(I+ 2z-l)(I + 2z)
= H(z)H"(lfz•)
Since h(n] is stable and causal and has a stable aod causal inverse, it must he a minimum phase system.
It therefore has all its poles and zeros inside the unit circle which allows us to uniquely identify H(z)
from JH(z)J 2 • ·
H(z) = 1+2z
lzl > ~
I ( l)(n-1)
h(n] = -26(n- I]+ -
2
u[n- 1]
Thus,
X(.;•) = { 16sin3w, 0 < w < ,.
0, -1r$W < 0
Therefore, the real part of X (&"') is
_;"')J2
IH( e- = -109 - -32 cosw = 1 - -2 cosw + -1 = ( 1 - -e
3 9
1 - '"' ) ( 1- -e'
3
1 ·.,)
3
= H(&"')H"(&"')
Thus, one choice for H(ei"') and h[n] is
H(.;"') = 1- ~.-;..
3
1
h[n] = o(n] - 3o[n- 1]
(b) No. We can fiud a new system by taking the zero from the original system and ftipping it to its
reciprocal location. This only changes the magnitude squared response by a scaling factor. If we
compensate for the scaling factor the two magnitude squared responses will be the same. Thus, we
find
= ~(1- 3e-i"')
3
1
h(n] = 3
o[n] - 30[n - 1]
Taking the inverse DTFT of XR(ei"') gives the conjugate-symmetric part of :r[n], denoted as :r,(n].
1 1 1 1 1 1
:r,[n] = - 2/(n + 2] + 2o[n + 1] + 2/[n + 1] + o[n] + 2o[n- 1]- 2/(n- 1] + 2/[n- 2]
(a) The inverse transform of XR(e'w) is z,[n], the even part of z[n]. This is true for any sequence
whether it is causal, anticausal, or neither.
(b) jX 1 (eiw) is the transform of z 0 [n], the odd part of z[n]. This is true for any sequence whether it
is causal, anticausa.l, or neither.
(c) For an anticausal sequence
z[n] = 2z,[n]u[-n]- z,[O)o[n]
Using Euler's identity and (a),
XR(&w) = f: G) t cos(kw)
t. or
1=0
= 1+ ~ (&•w + .-jkw)
= o[n] +
z[n]- z[-n]
'Ecr
1=1
00
2 6[n + k]
=
cr
z 0 [n]
2
jXJ(ei"') = ~ t. 0r ~,
(ei .... _ .. ... )
= "" (1).
2 sin(kw)
j {;
= "" (1).
iL 2 sin(kw)
4=0
Thus,
. = L"" (1).
XJ(e'"')
....
2 sin(kw)
11.15. Given X;(ei"'), we can take the inverse DTFT of jX;(ei"') to find the odd part of z(n], denoted z.(n].
Im {X(ei"')} = sinw
= ..!..eJ"' _1_e-iw
2j 2j
-1 + z[O] = 3
z[O] = 4
Therefore,
z(n] = 40(n]- o[n- 1]
11.16. Using Euler's identity and the fact that z.[n] is the inverse transform of XR(ei"') we have
XR(ei"') = 2- 4cos(3w)
= 2- 2(ei"' + ·-'"')
z.[n] = -20(n + 3] + 20(n] - 20[n- 3]
Since z(n] is real and causal, it is fnlly determined by its e?eD part z.(n],
z[n] = 2z.(n]u[n] - z.(0]6(n]
= 40(n]- 40(n- 3] - 20[n]
= 20[n] - 4o[n- 3]
473
X(~w)lw=•
"' :r[n]~•n
= L:
=-"'
"'
= L :r[n](-W
n.=-oc
= 2+4
'F 7
Thus, there is no real, causal sequence that satisfies both conditions.
11.17. There is more than one way to solve this problem. Two solutions are presented below.
Solution 1: Yes, it is possible to determine :r[n] uniquely. Note that X(k], the 2 point DFT of a real
signal :r(n], is also real, as demonstrated below.
1
X(k] = L :r(n]e-i'•""'l2
n=O
1
X(k] = L::rfn](-:W•
n=O
Thus,
X[O] = :r(O] + z[l]
X(l] = z[O]- z(l]
Clearly, if z[n] is real, then X[k] is real. Therefore, we can conclude that the imaginary part XI(k]
is zero.
Therefore, the inverse DFT of XR(k] is z[n], computed below.
1
z(n] = ~ L XR[k]~2·""' 1 '
k=O
1 1
z(n] = 2 L XR(k](-1)""'
k=O
1
z[O] = 2{XR(Oj + XR(l])
= -1
1
z[l] = 2{XR(Oj - XR(1])
= 3
Thus,
z(n] = -o(n] + 35[n- 1]
Solution 2: Start by making the assumption that X(k] is complex, i.e., X1[k] is nonzero and XR[k] =
U[k] - 45(k- 1]. Then, because z.,[n] is the inverse DFT of XR[k] we find
and
1
= 2(XR[Oj + XR[1])
= -1
1
= 2(XR[Oj- XR[1])
= 3
z,.[n] = -o[n] + 36[n- 1]
Because z[n] is real and causal, we can determine it from z .. [n]
z .. [n], n=O
2:t.. [n], 0 < n<N/2
z[n] =
z .. [N/2], n=N/2
0, otherwise
With N = 2 we have
z[n] = -o[n] + 36[n- 1]
If we began by making the assumption that X[k] was real, i.e., Xr[k] = 0 and X[k] = XR[k] =
26[k]- 46[k- 1] than by taking the inverse transform we find that
z[n] = z ..[n] = -o[k] + 36[k- 1]
This is the same answer we got before. Since there was no ambiguities in our determination of
z[n], we conclude that z[n] can be uniquely determined.
The next problem shows that when N > 2, we cannot necessarily uniquely determine z[n] from
XR[k] unless we make additional assumptions about z[n] such as periodic causality. When N > 2
the two assumptions we used above leads t.o two different sequences with the same XR[k].
11.18. Sequence 1: Fork= 0, 1,2 we have
XR(k] = 90(kj + 66[k- 1] + 66[((k + 1)),)
and XR[k] = 0 for any other k. Using the DFT properties and taking the inverse DFT we find for
n =·1,2,3
z,.[n] = 3 + 2 ( e><>•/3)n + .-j(2•/3)n)
= 3 + 4 cos(2,.n/3)
= 76[n] + o[n- 1] + o[n- 2]
If we let x[n] = z,.[n] we have the desired sequence.
Sequence 2: If we assume z[n] is periodically causal, we can nse the foUowing property t.o solve for
x[n] from z,.[n]:
z ..[o], n=O
z[n] = 2:tq[n], 0 < n < ~
{
0, otherwise
Note that this is only true for odd N. For even N, we would also need t.o handle the n = N /2
point as shown in the chapter. We have
z ..[o], n=0
z[n] = 2:tq[n], n = 1
{ 0, otherwise
= 76[n] + 26[n- 1]
475
11.19. Given the real part of X[k], we can take the inverse DIT to find the even periodic part of z[n],
denoted Zcp[n].
Using the inverse DIT relation,
N-1
zq[n] = N
1
L XR(k]W_,..
1=0
we find
1
zcp(Oj = 4(4+1+2+1)=2
%cp[1] = !(Hj-2-j)=!
4 . 2
1
%cp[2] = -(4-1+2-1) = 1
4
Zcp(3] = -I (4 - ). - 2 +).) =-1
4 2
Thus,
1 1
zcp[n] = U[n] + o(n- 1] + o[n- 2] + o(n- 3]
2 2
Next, we can relate the odd periodic and even periodic parts of :(n] using
zcp(n], 0 < n < N/2
:t.,.[n] = -:tcp{n], N/2 < n :S N- 1
{ 0, otherwise
Performing this operation gives
1 1
z.,.[n] = -o(n-
2
1]- -o(n- 3]
2
Taking the DIT of z.,.{n] yields jXJ(k]. Using the DIT relation,
N-1
iXI(k] =L :.,.[n]W"•
=<>
we find
jXJ(O] = (0 + ~ + 0- D= 0
:t(OJ. = !
61=0
t X(k]ei(>•l•l""l
......
= ~ LX(k]
•
1=0
= 1
. 476
This condition eliminates all choices except :t 2 (n] and :t3 [n].
The odd periodic· parts of :t 2 (n] and :t 3 (n] for n = 0, ... , 5 are
:to(n]- :t2[((-n))•]
:t., [n] = 2
= !3 (6(n-. 4]- 6(((n + 4))6]) - !3 (6(n- 5]- 6[((n + 5)).])
:t 3 [n]- :t3[(( -n)) 6 ]
:t..,(n] = 2
1 1
= 3 (6(n- 1]- 6(((n + 1))6 ]) -
3 (6(n- 2]- 6(((n + 2)) 6 ])
For n < 0 or n > 5, these sequences are zero. Since the transform of :top(n] is jXI(k] we find for
k = 0, ... ,5
= - ~j sin(4rk/3) + ~j sin(5rk/3)
= j_!_ (-6[k- 2] + 6[k- 4])
../3
XR(peiw) = U(p,w)
= 1+p- 1 acosw
Since au
7ii = !p&!
av we have
'
av =
8w
v = -ap-• sinw + K(p)
Since ~ = - ~ ~ we have,
ap- 2 sinw + K'(p) = ap- 2 sinw
Thus,
K'(p) = 0
K(p) = C
477
Since z[n] is real V(p,w) is an odd function of w. Hence, V(p,O) = 0, implying that C = 0.
Therefore,
XR(ei"') = 1 + QCOSW
= Q
1 + -t!w
2
. Q
+ -e-Jw
2
.
Q Q
z,(n] = o(n] + 2o(n + 1] + 2o(n- 1]
Because x(n] is real and causal, we can recover z 0 (n] from z,[n] as follows
{ z,(n], n>O
x 0 (n] = 0, n=O
-z,[n], n<O
Q Q
= --o(n + 1] + -o(n- 1]
2 2
Thus,
Note that we could have obtained x(n] directly from z,(n] as follows
2 2Z -N/2
- 1 +z -N/2
1-z 1 1- Z I
1- z-N/2 + z-1 _ z-1-N/2
= 1-z- 1 1•1 "I 0
Sampling this we find
478
When k = 0 we get 0/0 which, if the function was continuous, you would use I'Hopital's rule. In this
case the function is discrete so that is not available to us. One route to the answer is to use the definition
oftbe DFS
UN[O] = t iiN[n]e-;~•nl
1:=0 k=O
N
= :~:.:UN(n)
1=0
= N
N, k = 0,
UN[k) = -2j cot(rk/N), k odd,
{ 0, keven,k#O
11.23. (a) Because :r,,[n) is the inverse DFT of X11(k) we ha.ve for n = 0, ... , N- 1 and k = 0, ... , N- 1
X[k)+ X"[k)
XR[k] = 2
:r[n) +x"[((-n))N)
=
2
or equivalently, if we periodically extend these sequences with period N
_ [ J f[n) + %(-n)
%en = 2
Note that since the signal is real f"[-n] = f[-n).
The first period of f[n) is zero from n = M ton= N- 1. II N = 2(M- 1) there is no overlap of
f[n) and f[-n] except at n = 0 and n = Nf2. We can therefore recover f(n) from z,[n] with the
following:
_ { ~,[n], n=1, ... ,N/2-1
:r[n] = :r,[n], n = 0, N /2
0, n=M, ... ,N-1
II we tried to make N any smaller, the overlap of f[n] and %(-n] would prevent the recovery of :r[n).
Consequently, the smallest value of N we can use to recover X[k] from XR[k] is N = 2(M- 1).
(b) II N = 2(M- 1),
where
2, n = 1,2, ... ,N/2-1
1, n=O,N/2
{ 0, otherwise
= 2u[n]- 2u[n- N/2]- 6[n] + 5[n- N/2]
Taking the DFT of :r[n] we find
where
N, k=O,
= -2j cot(1rk/N},
{ 0,
0 < k < N- 1, k odd
otherwise
Thus,
H,.(ei"') = HER(ei"') Hc(ei"') = HEI(ei"')
H 8 (ei"') = Hoi(ei"') HD(ei"'} = -HoR(ei"')
480
The simplification in the last step used the fact that hlp[n1 = '""(;;121 is zero for even nand equals
1/2 for n = 0.
Find h~p[n1 :
Taking the inverse DTFT of Hlp(~"') yields
Using the fact that h[n] is zero for n = 0 and n even we can reduce this to
sin(.-n/2) 1
hlp[n] =
2
h[n] + 26[n1
(c) The linear phase causes a delay of n• = M/2 in the responses. H nd is not an integer, then we
interpret hlp[n1 and h[n1 as
sin(.-(n-n•)/2)
= .-(n nd)
2 sin2 (.-(n- n•)/2)
= ,.. (n nd)
Then,
where i.[n1 and i.lp[n1 are the causal FIR approrimations to h[n1 and hlp[n1. Similarly,
•
hlp[n1 ={ sin(,..(n- n•)/2)h·[ 1 !o[ -
2
1 [ 1
• n + 2 n "- w n • Meven
sin(,.-(n- n•)/2)h{n1, Modd
(d) The lowpass filter c:ornsponding to the first filter in the example looks like
481
IHlp(ei"')l
1.2
1
0.8 M =18,jl =2.629
0.6
0.4
0.2
00 0.21t 0.81t It
Cll
The lowpass filter corresponding to the second filter in the example looks like
1.2
11--------.....
0.8
0.6
0.4
0.2
0 ~o------~o~.21t~----~o~21t~----~o.~6~n=---~o~.8~n------~"
11.26. (a) The example shown here samples at the Nyquist rate ofT= tr/(fl, + <lfl) as in the chapter's
example, but the bandpass signal is such that <lfl/(fl, + <lfl) = 3/5. Then, 27</(M!T) = 10/3.
482
1 sc(j!l)
nc '"nc+.!10c n
1/T
s,<~
' / -x 2x
2/T S(ei"'}
=
(b) U 2"/(~!lT) M + e, where M is an iDteger and e some fraction, then using the Nyquist rate of
2,- JT = 2(!le + ~!l) will force decimation by M. M. jnst shown, this choice for T causes S4(ei")
to have iDtervals of zero. Instead, choose T such that 2"/(~nT) is the next highest iDteger
21r
~!lT =M+l.
Y(efw) contains roughly half the frequency spectrum as X(eiw), we can reconstruct X(efw) from Y(eiw).
We can accomplish this by recognizing that since z[n] is real, X(efw) must be conjugate symmetric.
The output of the system, y[n], has a Fourier transform Y(eiw) that is the product of X(eiw) and
H(eiw). Therefore, Y(eiw) will correspond to
At first glance, it may seem like X(efw) = Y(eiw) + y•(e-iw). This is close to the right answer, but it
doesn't take into consideration the fact that Y(efw) is non-zero at w = 0 and w = .-. Thus, the solution
X(eiw) = Y(efw) + y•(e-iw), will be incorrect at w = 0 and w = •, since Y(eiw) and y•(eiw) will
overlap at these frequencies. It is necessary to pay special attention to these frequencies to get the right
answer. Let
Z(eiw) = { 0, _ w = O,w = .-
Y(e'w), otherwise
Alternatively, we can express Z(eiw} with the constants a an~ b defined as
00
a = Y(eiwlL=O = L: y[n]
n:.:-oo
00
b = Y(efw)L=• = L: y[n](-1)"
n=-oo
H(z) = F(1/z)
= ....!.., 1 HR(v- 1 ) (z-• + v)
2.-; Tc z-> - v
dv,
v
izi ~ 1
where HR(v) = "IU{H(ei')}.
11.29. (a) We have
1l{z[n]} = z[n] • h[n]
1l{1l{z[n]}} = z[n]• h[n] • h[n]
We need to show that h[n]•h[n] = -6(n]. Alternatively, we need to show that H(eiw)H(eiw) = -1,
which is easily seen from
H(ei"') = { -:i 0 < w < ,.
J - r <w <0
(b) In Parseval's theorem,
f:
a=-oc
/[n]g•[n] =
2~ { ..
F(ei"')G•(ei"') dw
f:
n=-oc
1l{z[n]}z[n] =
2~ { •
H(ei"')X(ei"')X(e-;"') dw
where
H(ei"') = { -:i
3
0 < w <,.
- r <w < 0
but the integral = 0 since the integrand is au odd function over the symmetric interval.
(c) Since 1l{z[n]} = z[n] • h[n]
1l{z[n]• y[n]} = (z[n]• y[n]) • h[n]
= (z[n]• h[n]) • y[n]
= z[n]• (y[n]• h[n])
by the commutativity and associativity of convolution.
11.30.
h[n]
z,[n] z;[n]
h[n]
(b) The cross-<:orrelation between input and output is just the convolution of ¢ •••• (m] and h(m],
00
L h(-l]¢•••• (-m + l]
l=-cc
00
= L h(-l]¢•••• (m -l]
l=-oo
00
= - L h(l]¢•••• (m -l]
= -¢•••• (m]
we get
0, 0 < w <"
P,.(w) = { 4t,_..(.;w), -lr < w < 0
11.31. (a) As shown in the figure below, tbe system retODStructs the original bandpass signal. As in the
example, T = lr/(0, +Art) and M = 5.
486
21(5T)
2fT
-lt
1/T Y,(fi'>J
1 yc(jO) =sc(j!l)
=
""
5 cos(•• n) sin( !:n)
6 5 +j
5sin( lln) sin( .!n)
$ $
wn ""
h,.;(n] h;;(n]
(e) Using the information from part (b) we find
11[n] = II• (n]• h;(n]
= (!lre(n] + illo.[n]) • (h,.;(n] + jh;;(n])
= (1/•• (n]• h,.;(n]-llo.(n]• h;;(n]) +j(y.. (n]• h,.;(n] + llr•(n]• h;;[n))
y,(n]
We can now redraw the figure using only real operations:
kleal
y,JnJ ----; tM
Y,.ln]
hJnJ
Y)nJ
0/C
Converter
r--+
-1
----; tM
Y.,(n]
I>Jnl -
(d) From comparing the top and bottom figures in the answer to part (a), it is evident that the desired
complex system response is given by:
H(,.;w)={ 1, -1r<w<O
0, 0::; w ::; 1f
X(z)
488
Jl; Jl~
= log{A) + ~)og(1-ll>z- 1 ) + ~)og(1- b•z)
k=l k=l
II; N.
- :E Jog{1 - •••- 1
l- :E log{1 - d•• l
bl 1=1
log(1- z) = - L:
oo
n=1
n
we find
log{1- Ctz- 1) = - f: nz
a" -n
n=1
1 1•1 > lal
""pn
log{1- .Bz) = -L-;-z", 1•1 > 1.8- 1 1
n=1
-1 p-n
-n
= :E
n.=-oo
-n z'
1•1 > 1.8- 1 1
From the equations above we can identify the following z-transform pain;
a"
--u[n -1] +-+ log(1- az- 1 ),
n
1•1 > lal
p-n
-u[-n- 1] +-+ log{1- .Bz),
n
1•1 > 1.8- 1 1
n>O
i[n] =
n<O
(e) From the results of part {d), we see if i(n] is causal, all the b• and d1 terms must be zero. But the
expression for X (z) shows these terms correspond to the zeros and poles outside the unit circle.
We conclude that all the zeros and poles o:f X(z) are inside the unit circle, i.e., :z:[n] is a minimum
phase sequence.