Discrete Time Signal Processing Solution Manual

You are on page 1of 474

1

Solutions - Chapter 2
Discrete-Time Signals and Systems
3

2.1. (a) T(:z:[n])= g[nl*l


• Stable: Let 1*11 $ M then IT[*] I :;; lg[n]IM. So, it is stable if lg[n]l is bounded.
• Causal: y,[n] = g[n]z 1 [n] and 112[n] = g[n]:z:2[n], so if :z:,[n] = :z:2[n] for all n < no, then
y,[n] = 112[n] for all n <no, and the system is causal.
• Linear:

T(a:z: 1 [n] + b:z:2[n]) = g[n](= 1 [n] + bz2[n]


= og[n]:z: 1 [n] + bg[n]:z:2[n]
= oT(:z:1 [n]) + bT(z2[n])
So tbis is linear.
• Not time-invariant:

T(:z:[n - no]) = g[n]:z:[n - no]


cF y[n - no] = g[n - no]z[n - no]
which is not TI.
• Memoryless: y[n] = T(z[n]) depends only on the n" value of z, so it is memoryless.
(b) T(:z:[n]) =I::._ :z:[k]
• Not Stable: l:z:[n]l:;; M-+ IT(:z:[n])l :;; L::=,.,
l:z:[k]l :;; In- noiM. A$. n-+ oo, T-+ oo, so not
stable.
• Not Causal: T(:z:[n]) depends on tbe future values of :z:[n] when n < no, so tbis is not causal.
• Linear:

T(o:z: 1 [n] + bz2 [n]) = L: =•lkJ + b:z:.[kJ
t=no
• n

= 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:

T(z[n- no]) = L• :z:[k-no]


......
•-no
=
....L z[k]

·-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]

# aT(:t 1 [n]) + bT(z.[n])


This is not linear.
= =
• TI: T(:t[n -no]) e•l•-nol 11[n - no], so this is TI.
• Memory]ess: 11[n] depends on the n••
value of :t only, so it is memoryless.
=
(f) T(:t[n]) az[n] + b
• Stable: IT(z[n])l = [az(n] + bl ~ t>[MI + [b(, wbicb is stable for finite a aDd b.
• Causal: This doesn't use future values of z[n], so it is causal.
• Not linear:
T(e:t 1 [n] + d:t 2 [n]) = acz 1 [n] + t>d:t•[n] + b
# eT(:t 1 [n]) + d7'(:t2(n])
This is not linear.
5

• Tl: T(:z:(n -noll = c:z:(n - no] + b = y(n -no]. It is Tl.


• Memoryless: y(n] depends on the n'h value of :z:(n] only, so it is memoryless.
(g) T(:z:(n]) = :z:(-n]
• Stable: IT(:z:(n])l :>l:z:(-n]l $ M, so it is stable.
• Not causal: For n < 0, it depends on the future value of :z:(n], so it is not causal.
• Linear:

T(az 1 (n] + b:r 2(n]) = c:z:,[-n] + b:r2:z:[-n]


= aT(z 1 (n]) + bT(z2[n])
This is linear.
• Not Tl:

T(z(n - no]) = z( -n- no]


# y(n- no] = z[-n +no]
This is not Tl.
• Not memoryless: For n # 0, it depends on a value of z other than the n'h value, so it is not
memoryless.

(h) T(z(n]) = z[n] + u(n + 1]


• Stable: IT(z(n])l $ M + 3 for n;::: -1 and IT(:z:(n])l $ M for n < -I, so it is stable.
• Causal: Since it doesn't use future values of z(n], it is causal.
• Not linear:

T(az,(n] + b:r2(n]) = a:r,[n] + b:r2(n] + 3u(n + 1]


# aT(:z: 1 (n]) + bT(r2(n])
This is not linear.
• Not Tl:

T(:r(n- no] = :r(n -no]+ 3u(n + 1]


= y(n- no]
= r(n- no]+ 3u(n- no+ 1]
This is not TI.
• Memoryless: y(n] depends on then"' value of :r only, so this is memoryless.

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

(a) Since h[k] # 0, for (No $ n $ N 1 ),


N,
y(n] = L h[k]z(n - k]
t=No

The input, r(n] # 0, for (N2 $ n $ N3), so

r[n- k] # 0, for N2 $ (n- k) ::; N 3


6

Note that the minimum value of (n - l:) is N2. Thus, the lower bound on n, which occurs for
k =No is

Using a similar argument,


N• = N1 +N,.
Therefore, the output is nonzero for

(No +N2):,; n:,; (N, + N,).

(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.

2.3. We desire the step response to a system whose impulse response is

h[n] =4-nu[-n], for 0 < 4 < 1.


The convolution sum:
00

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

2.4. The difference equation:


3 1
y(n]- 411!n- 1] + Sy[n- 2] = 2l:(n- 1]
To solve, we take the Fourier transform of both sides.

The system function is given by:

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:

y[n]- 5y[n- 1] + 6y[n- 2] =0


Taking the Z-transform,
1- sz-l +6z-• = 0
(1- 2z- 1 )(1- 3z- 1 ) = 0.
The homogeneous solution is of the form

ll•[n] = A 1 (2)" + A,(3)".

(b) We take the z-transform of both sides:

Y(z)[1- sz- 1 + 6z- 2 ] = 2z- 1 X(z)


Thus, the system function is
Y(z)
H(z) = X(z)

= 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

h(n) = -2(2)"u(n] + 2(3)"u(n].


8

(c) Let z[n] = u[n] (unit step), then


1
X(z) = 1 z 1

and

Y(z) = X(z) · H(z)


= (1- z 1 )(1- 2z ')(I- 3z- 1f
Partial fraction expansion yields
1 4 3
Y(z) = 1 -z ' -:--7=+
l-2z 1 l-3z ,.
The inverse transform yields:
y[n] = u[n] - 4(2)"u[n] + 3(3)"u[n].

2.6. (ar The difference equation:


1
y[n] - 211[n- 1] = z[n] + 2z[n- 1] + z[n- 2]
Taking the Fourier transform of both sides,

Y(eiw)[l- !e-jw] = X(.,.iw)[l + 2e-)w +.-;:a...].


2
Hence, the frequency response is

1 + 2e-;... +.-;:a...
=
1- !e-iw

(b) A system with frequency response:

H(eiw)

cross multiplying,

Y(eiw)[l + ~·-jw + ~e-i2wJ = X(eiw)[l - ~e-iw + e-i3w],


and the inverse transform gives
1 3 1
y[n] + v[n- 1] + 11[n- 2] = z[n] - z[n- 1] + z[n - 3].
2 4 2
2.7. z[n] is periodic with period N if z[n] = :z:[n + NJ for some integer N.
(a) :z:[n] is periodic with period 12:
.;< fn) = .;< f )(n+N) =.;< fn+. .k)
~ 2?rk =iN, for integers k,N
Making k = 1 and N = 12 shows that z[n] has period 12.
9

(b) :r[n] is periodic with period 8:

eJ'( 'i'nJ = eJ'( 'i' )(n+N) = eJ'( 'i'n+2d)


3
= 27Tk = : N, for integers k, N

= N = ~k,for integers k,N


The smallest k for which both k and N are integers are is 3, resulting in the period N being 8.
(c) :r[n] = [sin(..n/5)]/(7Tn) is not periodic because the denominator term is linear inn.
(d) We will show that :t[n] is not periodic. Suppose that :t[n] is periodic for some period N:

,i!?,n) =,i!?,)(n+N) =,i!?,n+2••J


= 2..-k = ~N,for integers k,N
= N = 2v'2k, for some.integers k, N
There is no integer k for which N is an integer. Hence :r[n] is not periodic.

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]

2.9. (a) First the frequency response:

Y(eJ'w) _ ~.-iwy(eJ'w) + .!, 0 -2iwy(e"w) = .!..-2jw X(e'w)


6 6 3

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

For the step response s[n]:


00

s[n] = L h[k]u[n - A:]


= 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:

ll>[n] =A,(~)"+ A,(~)"


(c) The total solution is the sum of the homogeneous and particular solutions, with the particular
solution being the impulse response found in part (a):
y[n] = !l>[n] + y9 [n]
= A,(~)"+ A2 (~ )" + -2(~ )"u[n] + 2(~ )"u[n]
Now we use the constraint y[O] = y[1] = 1 to solve for A1 and A,:
y[O] = A 1 + A, - 2 + 2 = 1
y[1] = A,/2 + A,f3- 2/3 + 1 = 1
A1 +A2 = 1
A,f2+ A,/3 = 2/3
With A, = 2 and A,= -1 solving the simultaneous equations, we find that the impulse response
is
1 1 1 1
lf[n] = 2(-)"- (- )" + -2(- )"u[n] + 2(- )"u[n]
2 3 3 2
2.10. (a)
y[n] = h[n]• z[n]
00

= L a•u[-k- 1]u[n- A:]


lo=-oo

=
{ 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,

y(n] = u(n - 4] • u(n]


= w{n- 4]
{ 2n-3 n$3
= 1,
0

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.

y[n] = r[n] • h[n]


= (u[n]- u[n- 10]) • u[n]
= w[n]- w[n - 10]
= (2n+lu[-(n + 1)] + u[n])- (2n- 9 u[-(n- 9)] + u[n- 10])
2(n+l) - 2(n-1). n $ -2
= 1- 2(n-l),
{ 0,
-1 $ n $8
n~9

2.11. First we re-write :r[n] as a sum of complex e:rponentials:

Since comple¥ exponentials are eigenfunctions of LTI systems,


H(.J•I<).J"'i'- H(e-i•l')e-iwn/4
y[n] = 2j

Evaluating the frequency response at w = ±7:/4:

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

2.12. The difference equation:


1/[n] = n11[n - 1] + :[n]
Since the system is causal and satisfies initial-rest conditions, we may recursively find the response to
any input.

(a) Suppose z[n] = 6[n]:


y[n] = 0, for n <0
y[O] = 1
y[1] = 1
y[2] =2
y[3] = 6
y[4] = 24

y[n] = h[n] = n!ujn]


(b) To determine if the system is linear, consider the input:

:[n] = a<l[n] + bO[n]


performing the recursion,
y[n] = 0, for n < 0
y[O] =a+ b
y[1] = Q + b
y[2] = 2(a +b)
y[3] = 6(a + b)
y[4] = 24(a +b)

Because the output of the superposition of two input signals is equivalent to the superposition of
the individual outputs, the system is LINEAR.
(c) To determine if the system is time-invariant, consider the input:

:[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),

h[n- 1] = (n- l)!u[n- 1]# y[nJI•Jn]=l{n-1]


Conclude: NOT TIME INVARIANT.
13

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:

:z:,(n] =6(n] ==> y.(n] =(41 )"u(n]


. 1
:t2(n] =6(n- 1] ==> 112[n] = (4)"- 1u(n]
Even though :z: 2[n] = :z:1(n- 1], 112[n] "# y1[n- 1] = (i)"- 1u[n- 1]
(b) No. Consider the input/output pair :z:2(n] and 112[n] above. :z: 2(n] = 0 for n < 1, but y2(0] # 0.
(c) Yes. Since h[n] is stable and multiplication with u[n] will not cause any sequences to become
unbounded, the entire system is stable.
1!.16. (a) The homogeneous solution y,[n] solves the difference equation when :z:(n] = 0. It is in the form
Y•(n] = 2: A(c)", where the c's solve the quadratic equation
1 1
c'- -c+- = 0
4 8
So for c = 1/2 and c = -1/4, the general form for the homogeneous solution is:
y,(n] = A1 ( _!2 )" + A2(_ _!4 )"
(b) Taking the z-transform of both sides, we find that

Y(z)(1- .!.-•- .!z- 2 )


4 8
= 3X(z)
and therefore
Y(z)
H(z) = X(z)
3
= 1 - 1/4z 1 - 1/8z-2
3
= (1 + 1j4z-1)(1-1/2z ')
1 2
= :--~,.,-...,. + ~-;-:~...,.
1 + 1/4z 1 1 - 1/2z 1
14

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

h..e[n] = -((-1/4)" + 2(1/2)")u[-n -1]


(c) h<{n] is absolutely SUIDDlable, while h..,[n] grows without bounds.
(d)

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

L:• Jsin(.-n/3)1 =2h


n=O
• and summing Jh[n]l over all positive n therefore grows to oo.
(e) Stable. Notice that Jh[n]l is upperbounded by (3/4)1nl, which is absolutely summahle.
(f) Stable.
2 -5<n<-1
h[n] =
{
1: :o::;-;, ::;4
0, 1 otherwise

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

h(n] = (1/a)"- 1 u(n- 1]

(b) h(n] is absolutely S1lii1Jilable if 11/al < 1 or if Jal >1


16

2.21. For an arbitrary linear system, we have

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:

T{x[n] + x,[n]} = T{x[n]} + T{x 1 (n]}


= y[n] + !11 [n]
Since x[n] is zero for all n,
T{x[n] + x,[n]} = T{x,[n]} = y 1 [n]
Hence, y[n] must also be zero (or all n.

2.22. We use the graphical approach to compute the convolution:

y{n] = x{n]• h[n]


00

= L x[k]h[n - k]
k=-oo

(a) y{n] = x{n] • h{n]


y[n] = o[n- 1]• h[n] = h{n- 1]

. . .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

(d) y[n] = x{n] • h{n]


17
3 3

0123456 n

2.23. The ideal delay system:


11[n] = T{z[n]} = z[n- n 0]

Using the definition of linearity:


T{u 1 [n] + bo: 2 [n]} = az,[n- n.J + bo:2 [n- n.]
= oy 1 [n] + by2 [n]
So, the ideal delay system is LINEAR.
The moving average system:

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]

= oy1 [n] + i>y,[n]


Conclude, the moving average is LINEAR.

2.24. The response of the system to a delayed step:


Jl[n] = z[n] • h[n]
00

= 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

2.25. The output is obtained from the convolution sum:

y[n] = o:[n]• h[n] 00

= 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:

y[n] = O,for n < 0


n
y[n] = :La•
.,..,
1 - a<n+l)
= ,forO::;n::;N,
1 -a

N,
y[n] = :La•
>=<>
1_ 0 (N1+ll
= 1-a
, for N 1 < n < N2

N, n

y[n] = L alt. + ~ a<'=-N2)


k=O I:=N2
1_ 4 (N1+t) 1_ 4 (n+l)
= 1-a + 1-a
2_ 4
(N1+1) _
4
(n+t)
= 1
_a , (or N2 $ n::; (N1 + N2)

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

(a) z[n] = 5"u[n]:


00

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

Becuase the summation depends on n, z[n] is NOT AN EIGENFUNCTION.


(b) z[n] = e'""'":
00

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

y[n] = L h[k]eiw(•-•1 + L h[k]ei""'(•-•1


lt=-oo

L
00

L
00

= ,;wn h[k]e-;wt + eJ:Z..n h[k]e-i:Z..t

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

y[n] = L: h[k]s<•-•1 ,;:z..<•-•1

=
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:

ei<'f•l = ei<'fH•+NJ = .;<'f•+•••l


2
= 2.. k = ; N,for integers k,N

Making k = 1 and N = 5 shows that :~:(n] has period 5.


(b) :~:(n] is periodic with period 38. Since the sin function has period of 27r:

z[n + 38] = sin(r(n + 38)/19} = sin(rn/19 + 21r} = z(n]


21

(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

=a:r:,(n2] + b:r:,(n2 ]) =ay1 (n] + t>y,[n]


T(a:r:1 (nj + b:r:,[n])
• It is not time-invariant. If y[n] = T(:r[n]) = :r[n 2], then T(:r[n- 1]) = :r[n 2 - 1] f. y[n- 1].
22

(c) First notice that


00

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

2: o:r,[kJ + b:z:,[kJ =a 2: :z:,[kl + b 2: :z:,[kJ


k=n-1 .t=n-1 l:=n-1

• It is time-invariant. Let
00

y[n] = T(x[n]) = L :r[k],


.t=n-1

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•[n] =A,(!)"+ A,(-!)"


3 5
(b) We use the z-transform, and use different ROCs to generate the causal and anti-causal impulses
responses:
H(z) _ 1 _ 5/11 + 6/11
- (1- iz- 1 )(1 + !z ')- 1-lz ' 1 + ~z-1

hc[n] = : ( ~ )"u[n] + : 1 (-~ )"u[n]


1
h [n] =
..
-~(!)"u[-n
113
-1]- .!.(-!)"u[-n -1]
115
(c) Since h,(n] is causal, and the two exponential bases in hc[n] are both less than 1, it is absolutely
summable. hoc(n] grows without bounds as n approaches -oo.
(d)

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

2.32. We first re-write the system function H(&w):

Let lit [n] =.:[n]• g[n], then


1m et•n/2 + e-j'fln/2
.:[n] = cos(2) = 2
G(ei•/2)&•n/2 + G(e-i•/2)e-i•n/2
llt[n] =
2

Evaluating the frequency response at w = :!:7r /2:

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

H(ei'f) = H(e-ii) = &'f


Therefore,
311" ~ 211" 311" 1111"
11[n] = cos(-n +- + -)
2 4 3
= cos(-n
2
+ -12)
2.34. (a) Notice that
r[n] =:.o[n - 2] + 2%0 [n- 4] + :.o[n- 6]
Since the system is LTI,

11[n] = 110(n - 2] + 2l!O(n - 4] + 110[n - 6],


and we get sequence shown here:
2

-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:

6[n] = zl[n]- :r2[n]


and
1
6[n] =
2:r 2 [n + 1]
to determine the impulse response. With the given information, we can only use shifted inputs.
2.36. (a) Suppose we form the impulse:
1 1
6[n] = 2"'' [n]- 2:r2[n] + :r,[n]

Since the system is linear,


L{6[n]} = 21 yl[n]- 21y2[n] + 113[n]

A shifted impulse results when:

The response to the shifted impulse

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:

. h(n] = U(n + 2] + 6{n + 1]- U(n] + U[n- 1]


+U(n- 2] +6[n- 3]
3

2 2

o 1 2 n

-2

2.37. For an LTI system, we use the convolution equation to obtain the output:
00

y(n] = L :r(n - k]h(kJ


k=-oo

Let n = m+ N:
00

y(m + NJ = L :r(m + N- k]h(k]

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

Jl(m + N] = L :r(m - k]h(k]


t=-oo
= y(m]
So, the output must also be periodic with period N.

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],

is obtained by setting the input (forcing term) to zero.

3 1
Jl(n]- 4y(n - 1] + sy(n - 2] = 0

Solving,
3 -1
1- 4-z
1 -2
+ 8-z = 0•
26

(1- ~z- 1 )(1- ~z- 1 ) = 0,


2 4
and the hom(lgeneous solution takes the form

ll•{n] =A,(~)"+ A,(~)",


for the constants A1 and A2 .
(b) Substituting the intial conditions,

11•{-1] = A,(~)- 1 + A,(~)- 1 = 1,


and
Y•{O] =A, +A, = 0.
We have
2A 1 +4A, =1
At +A2 =0
· Solving,
A1 = -1/2
and
A2 = 1/2.
(c) Homogeneous equation:
1
y(n]- y(n- 1] + -y[n- 2]
4
=0
Solving,
1 - z- 1 + ~z- 2 = 0'
4
(1- 1 z-1)( 1- 2z
2
1 -1) = 0,
and the homogeneous solution takes the form

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

ll•{n] =A,(~)" +nB,(~)",


substituting into the difference equation:
1
llh(n]- !lh{n- 1] + 411•[n- 2] = 0

A,(~)" +nB,(~)"- A 1 (~)n-l- (n -1)B1 (!)n-l


2 2 2 2
!A,(!)"- 2 +!(n-2)Bt(~)"- 2 = 0.
4 2 4 2
27

(e) Using the solution from part (d):

and the initial conditions


11•[-1] = 1
and
11•[0] = 0,
we solve for A1 and B 1 :
A,= 0
B 1 = -1/2.
2.39. (a) For z 1 [n] =6[n],
y 1 [0] = 1
111[1] = ay[O] =a
For z 2[n] = 6[n- 1],
y,[O] = 1
y,[1] = ay[O] + z2[1] = a+ 1 # y,[O]
Even though z 2[n] = z 1 [n- 1], y,[n] # y,[n- 1]. Hence the system is NOT TIME INVARIANT.
(b) A linear system has the property that
T{az,[n] + bz 2[n]} = aT{z,[n]} + bT{z.[n]}}
Hence, if the input is doubled, the output must also double at each value of n.
=
Because y[O] 1, always, the system is NOT LINEAR.
(c) Let %3 = az.(n] + tlz.[n].
For n?: 0:

ll>[n] = :2:3[n] + ay3[n- 1]


• = az,[n] + ,8z2[n] + a(z.[n- 1] + y,[n- 2])
n-1 n-1
= a :l::>•z,[n- k] + ,8 L a•z.[n- k]
.=0 1=0
= o(h[n]• z,[n]) + ,9(h[n]• z 2 [n])
= Qlll[n] + ,8y,[n].
For n < 0:
= .,- 1 (y,[n + 1]- z,[n])
Y>[n]
.. ..
= -a L: .;•z,[n- kJ- .e L: ..•.,.tn- kJ
•=-1 J=-1
= oy1 (n] + ,8y,(n].
Forn=O:
Y>[n] =Y>(n] =112[n] = 0.
Conclude,
Y>(n] = oy,[n] + PY.(n], for all n.
Therefore, the system is LINEAR. The system is >till NOT TIME INVARIANT.
28

2.40. For the input

: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

For large n, (-j/2)(n+l) -t 0. Thus, the steady-state response becomes

(-1)"
y[n] = 1 + j/2
cos{wn)
=
l+j/2"

2.41. The input sequence,

:t{n]
"'
= I: o[n + 16k},
k=-oo
has the Fourier representation

X (eiw) = L"' L"' .S(n + 16k]e-iwn


n=- oo b:: -oo

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):

IH(e -.}!) IX(ei (I) )I

-fl! ~ -1! 0 1! "" ~ (I)


16 16 16 16 16 16
29

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,

11[n] = (z(n] + z(n]• h1 (n]) • h2(n]


= (z(n]• (o[n] + h1 [n])) • h2[n].
Let h(n] be the impulse response of the overall system,

y(n] = z(n]• h[n].


Comparing with the above expression,

h(n] = (o[n] + h.(n]) • l>,(n]


= h,[n] + h.(n]• h2[n]
= a"u(n] + p<•-•lu[n- 1].

(b) Taking the Fourier transform of h(n] from part (a),


00
H(ei"') = }: h[n]e-;"'"
n=-oc
00 00
= }: o"u(n]e-:;"'" + P L a<n-llu[n- 1]e-;"'n
n=-oo n=-oo
00 00
= L 0 ne-i"'" +p L 0 (t-l)e-;'"'t,
n=O l=O

where we have used t = (n- 1) in the second sum.


1 ~-;..
H(ei"') = .,.-----,.,-, + .
1 -or'"' 1- ae-'"'
1+~-;..
= :-'-""'="r:·
1- ae-'"'
, for lal < 1.
Note that the Fourier transform of o"u(n] is well known, and t.he second term of h[n] (see part (a))
is just a scaled and shifted version of o"u(n]. So, we could have used tbe properties of the Fourier
transform to reduce the algebra.
(c) We have
Y(ei"')
H(ei"') = X(e;"')
1 + p.-;..
= 1- ae-ic.~ •
30

cross multiplying,
Y(.,;->)[1- ae-;"'1 = X(ei"')[1 + 13•-;"'1
taking the inverse Fourier transform, we have

y[n1- ay[n- 11 =:[n1 + j3:(n- 1).


(d) From part (a):
h[n1 =0, for n < 0.
This implies that the systtm is CAUSAL.
U the system is stable, its Fourier transform existo. Therefore, the condition for stability is the
same as the condition imposed on the frequency response of part (b). That is, STABLE, if Ia! < 1.
2.43. For (-1 <a< 0), we have
1
X(ei"')-
- 1- oe-iw

(a) real part of X(ti"'):

= !2 ·[X(ei"') + X"(ei"')1
1- aC06(w)
= 1- 2aC06(w) + a 2

(b) imaginary part:

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..

A(w) is a zero phase (real) function of w. Hence,

LX(.,;"') = -2w, -.- $ w $ .-


(d)

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

Taking the inverse transform, we have


I 1 1 1
2a[n + 2] + 2a[n- 2] = 2z[n + 4] + 2"'[n]

• •
•-112 -4
I
. I 0 I . . I 112T ••••
I

4
-112
32

2.45. Let :z:[nJ = o[n), then


X(ei"') =1
The output of the ideal lowpass filter:

W(ei"') = X(ei"')H(ei"') = H(ei"')


The multiplier:

causes a shih in the frequency domain:

The overall output:


y[n] =e-;•nw[n] + w[n]
Y(ei"') = H(eilw-•l) + H(ei"')
Noting that:

Y(ei"') = 1, thus y[nJ = o[n].


2.46. (a) We first perform a partial-fraction expansion of X(ei"'):

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)

y[nJ = :z:(nJ + 2:z:[n- 1) + :(n - 2]


= :z:[n) • h[nJ
= :[n] • {6(nJ + 26[n - 1) + o[n- 2))
h{n] = 6!nJ + 26[n - 1) + .S[n - 2]

(b) Yes. h[n] is finite-length and absolutely summable.


33

(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]

2.48. (a) Notice that


s[n] = 1 +cos(=) = 1 + ( -1)"
S(~w) = 27f L 6(w- h)

s[n]

..ttJJ
(b) Since y[n] = :[n]s[n],

= _.!._ f' S(.,.)X(~(w-Bl)dw


271' -·
34

= _!._
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]),

W(ei"') = Y(ei"') +.!. (ei"'Y(ei"') + .-;"'Y(ei"'))


2
= Y(ei"')(1 + cos(w))
(d) The following figure shows X(d"'), Y(d"'), and W(d"') for a< 2 and a> 2. Notice that

X(ei"') = { 1, lwl $ wfa,


•0, wfo $lwl 2::"

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].

X. Y. and W 1o< a<2 X.Y,andWforl>2


2r-----------------. 6r------------------,
1.5
.{1
iS
0.5

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

2.49. (a) We start by interpreting each clue.


(i) The system is causal implies
h[n] = 0 for n $0.
(ii) The Fourier transform is conjugate symmetric implies h[n] is real.
(iii) The DTFT of the sequence h(n + 1] is real implies h(n + 1] is even.
From the above observations, we deduce that h(n] has length 3, therefore it has finite duration.
35

(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

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

(c) Let a[n] = v[-n] and b[n] = w[-n], then:


+oo
a[n] • b[n] = L a[.I:Jb[n- .1:]

+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

(a) real part of X(ei"'):

= .!_ • [X(ei"') + X"(ei"')]


2
1- acos(w)
= 1 - 2a cos(w) + a2
(b) imaginary part:

X,(ei"') = ;i · [X(ei"')- X"(.!"')]


-asin(w)
= 1- 2a cos(w) + a2

(c) m~tude:

IX(ei"'JI = [X(ei"')X"(ei"')]t

= (1- 2aco!(w) + a2) t


(d) phase:
LX(ei"') =arctan ( -asin(w) )
1- acos(w)

2.52. :r[n] can be rewritten as:

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

y[n] is then given by:


y(n] = 3 nn(0.5.-(n- 5)) _ 3 nn(0.5.-(n- 6)) .
.-(n- 5) .-(n- 6)
2.54.

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:

..._ eiiei't- ., .-;t.-; ...


y[n] = .-,, 2 +e-' 2

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)

Jl[n] = h[n]• (e-;...nz[n])


+ao
= L ,-;...•z[k]h(n- k] .
.t=-oc
38

Let :t[n] = a:tt[n] + b:t2[n], then:


1/[n] = h[n]• (e-jwon(a:t,[n] + b:t,[n]))
+co
= L .,-;...•(a:r,[k] + b:t,[k])h[n- k]
•~-oo
+co +co
.,-iw••:r,[k]h[n- k] + b
= "L
.t=-cc
E
k=-oo
,-;wo•:r [k]h[n- k]
2

= "II• [n] + blf,[n]

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].

We thus conclude that system S is not time in..ariant.


(c) Since the magnitude of .->won is always bounded by 1 and h[n] is stable, a bounded input :t[n]
will always produce a bounded input to the stable LTI system and therefore the output y[n] will
be bounded. We thus conclude that system S is stable.
(d) We can rewrite y[n] as:

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].

System C should therefore be a multiplication by ,-;won.

2.57. (a) H 1 (eiw) corresponds to a &equency shifted version of H(eJw), specifically:


H,(.,Jw) = H(.,J(w-•lj.

We thus have:
39

H,(~"') = { 0 , lwl < 0.8.-


1 ' 0.8" $ lwl $ "·
This is a highpass filter.

1 .--

.....1.-.....L---.....L---.....L-.....L-... w
- .. -o.s.. 0 o.s.. 1(

(b) H2(e;"') corresponds to a frequency modulated version of H(~"'), speciDcally:

H2(~"') = H(~"') • (o(w- O.Sw) + o(w + O.S1r)) where lwl $ ".

We thus have:

lwl < 0.3lr


0.3" $ Jwl $ 0. 7.-
0.7" < lwl $ "·
This is a bandpass filter.

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:

where H(eJ"') is given by:


H(~"') ={ 01 , Jwl < O.l?r
, O.b $lwl $ ..
Carrying out the convolution, - get:

0.1 , lwl < O.l?r


H,(~"') = { •• -~ +0 • 15 ' 0.1.- $ lwl $ 0.3..
0 0.3.. < lwl $ ...
40

2.58. Note that X(dw) is real, and Y(eiw) is given by:

Y(eiw) = { -jX(~) O<w<lr


+jX(e'w) -1r<w<O.

w[n] = z[n] + jy[n], therefore:


W(eiw) = X(eiw) + jY(eiw).
Using the above, we get:

"Y(·;w) { X(dW) O<w<1r


) .- = -X(dw) -.- <w <0.
We thus conclude:

W(e'w) = { 2X(eiw~ 0<w<7r


-7T <"' < o.

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

(b) By carrying out the convolution, we get:

-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:

y[n] = y,;,[n] + y..,.[n]


where y,;.[n] is the zero input response and y,.,.[n] is the zero state response, the response to zero initial
conditions:

y,,,[n] = a(~)n where a is a constant determined by the initial condition.


1
y..,.[n] = (l)nu[n]• z[n].

Ail example of a bounded input that results in an unbounded output is:

z[n] = J[n + 1].


The output is unbounded and given by:
1 1 1
y[n] = (-t+'u[n+ 1]- -(-)n.
2 2 2
2.62. The definition of causality implies that the output of a causal LTI system may only be derived from
past and present inputs.
The convolution sum:
00

y[n] = 'E h!l:]z[n- 1:]


k=-ao
-1 00

= 'E h[k]z[n- k] + 'E h[k]z[n- k]


k=-ao

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

This can only be guaranteed if h[k]= 0 for n < 0.


Using reverse logic, we can show that if h[n] = 0 for n < 0,
00

'E h!l:Jz[n- k].


y[n] =
....
Since the convolution snm specifies that the input is formed from past and present input values, the
system is, by definition, causal.
2.63. The system could be LTI. A possible impulse response is:

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]:

11[0] = 2~ 1:• Y(~)dw


= ..!._ ~+~ e-1'-~e-1'-~12 dw
271' -·
2
= - 3w
-# 0.

This proves that the system is not causal


2.65. z,[n] is even-symmetric around n = 1.5, furthermore since I;zl[n] < 0 and we want A1(0) ~ 0, we
need to include a .- factor in the phase. An appropriate choice for 81 (w) is therefore:

81 (w)

0
-j

z2[n] is odd-symmetric around n = 3, therefore:

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:

(c) Taking the inverse transform of H (e'"'), we get:

h[n] = h,[-n]• h 1 [n].

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"')

H(e'"') is thns given by:

(b)
44

With the given choice of H 1 (.;"'),

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]

A simple counterexample is z 1 [n] = z 2 (n] =o[n]. Then:

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]

which is not a. linear system, therefore the systan is not LTl.


(b)

Y,(ei"') = H 1 (ei"')X(ei"')
Y.(ei"') = H2(ei"')X(e'"')
Y(ei"') = Y,(ei"') • Y2(ei"').

Using the above relationships, we get:

Y(ei"') = { nnspecmed , o < lwl < o.6"


0 0.6.. ~ lwl ~ ...

2. 70. The first diHerence:


y[n] = V(z(n]) = z(n]- z(n- 1].
(a) To determine if the system is linear:

V(az1[n] + lr.r2[n)) = az.[n) + lr.r2[n)- cu1[n- 1]-lr.r2[n- 1)


= a(:r1 (n)- :r1 [n- 1)) + b(:r2[n)- :r2[n- 1))
= V(cu1[n)) + V(az2[n)).

Therefore, the system is LINEAR.


To determine if the first difference is time invariant:

V(:r 1[n -1)) = :r[n -1)- :r[n- 2)


= y[n -1).
The system is TIME INVARIANT.
(b) The impulse response is obtained by setting the input to :r[n) = 6[n),
y[n) = h[n) = 6[n)- 6[n- 1)
(c) Taking the Fourier transform of the result of part (b), we find that the system function is

H(e'w) = 1- .-iw.
Thus the magnitude of the frequency response is

IH(e'w)l = .,/(1- e-iw)(1- e-N)


= .j2 - 2 cos(w ).

We calculate the phase of the frequency response:

H{e'w) = (1- oos{w)) + j sin(w)


Thus,
LH(eiw) =arctan ( sin(w) )
. 1- cos(w)
46

(d) In general,

V(z[n]) = z[n]• (6[n]- o[n- 1])


= z[n]- z[n- 1].
So, for z[n] = f[n]• g[n],
V(z[n]) = /[n]• g[n]• (o[n]- o[n- 1])
= /[n]• V{g[n])
= V(/[n]) • g[n].
Where we have used the commutivity of the convolution operator to obtain the last two equalities.
(e) We desire the inverse system, h;[n], such that

h;[n]• V(z[n]) = :r[n]


The inverse system must satisfy:
h;[n]• h[n] = o[n],
in the frequency domain,

Recall from part (c),

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

(a) Suppose the impulse response is h•[n],


oa
L h•[n]e-i~n
,..._oa = c~oa h(n].!~").
47

(b) We have

H•(ei"') = (J:oo h[n]e-;wn) •


00

= L h•[n]eiwn.
--oo
If h[n] is real,
00

H•(eiw) = L h[n]eiwn

Hence, the frequency response is conjugate symmetric.

2. 72. The analysis equation for the Fourier transform:


00

X (eiw) = L z[n]e-;wn
n=-oo

(a) The Fourier transform of z•[n],


00

L z•[n]e-jwn =
n.=-oo

(b) The Fourier transform of z•[-n],


00 00
L z•[-n]e-;wn = L z•[l]eiwl
n.=-oo 1=-ao

= (too z[l]e-jwl) •

2.73. From property 1:


00

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.

X(ei"') = XR(eiw) + jXr(~w)


X•(e-;w) = XR(<-;w)- jXr(e;w)
48

From property 7, X(ei"') =X"(e-i"') for o:[n] real. Thus,


XR(ei"') + jXJ(ei"') = XR(e-i"')- iXI(e-i"').
We may infer
property 8: XR(ei"') = XR(e-i"')
property 9: X1(ei"') = -X1 (e-i"')
X(ei"')= iX(ei"')ieiLX(<'")
X"(e-'"'l = IX(e-1"')1•-;Lx(.-;.)
From property 7:

So,
property 10: IX(ei"')! = IX(e-iw)l
property 11: LX(ei"') = -LX(e-i"').
2. 7 4. Theorem 1:
00 00 00

L (a:r,[n] + b:r2 [n])e-i"'n = L az,[nje·iwn + L b:r,[n]e·iwn


n=-oo n=-oo n.=-oc

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

y[n] = L z[kjh[n - kj.


1=-oo
Taking the Fourier transform,

Y(eiw) = ~f;oo (1: 00


z[k)h[n- k)) e-i~
= 1: z[kj Cf;oo·h[n- kje-iwn)
1:
00

= z[kje-i"> (_too h{n- k}e-iw(n->))


00

Hence,

2. 76. The Modulation theorem:

the time-domain representation,

v[n] = __!_, (' d8 (' dw X (ei 1 )W(e'(w-B>)e->w•


(21r) '-· L.
= 2... f' d8 X(ei1 )w[n]ei'•
21f )_ ..
= z[n]w[n]
2. 77. (a) The Fourier transform of y•[-n] is Y"(e'"), and X(ei")Y(e'") forms a transform pair with z[n] •
y[n]. So

and
g[n] = z[n] • y"[-n)
form a transform pair.
(b)
00

L (z[n] •11"[-nj) e-jwn


--co
00 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

(c) Using the result from part (b):

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
"

_,

Substituting into Eq. (P2.77-1):


00

2: :(n]y"(n] = 2_ (' X(ei"')Y"(e'"')dw


271" J_w
--=
2 5 6">]
2
= 2_
2.. (<!H!H
1
= 60

2. 78. X (e'"') is given by:

(a)

1/.(n] = { :(n], n even


0, nodd
= ~ (1 + eiwn) :(n]
51

which transforms to

(b)

Y•[n] = :r[2n]

Y.(~w) = ~ [x(~'l') + X(~!'!'+•ll]


= Y,(~'t)

(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

= L o:*(n + N]:r(n- N]e-iwn


n.=-oo
= +.(-N,-w).

(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

:r(n + N):r*(n- N)e-iwn

= +s(N,w).

We thus conclude that:


+.(N,w) = 1
2
f" X(ei<•+(w/2ll)X*(ei<•-!w/2HjeJ2•N dv.
1f -·
53

2.80.

w(n] = o:[n] + y(n]


The mean of w(n]:

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:

••• (ei"') = t.,(ei"') + ••• (ei"').


{b)

tl>•• (m] = E(o:[n]e(n + m])


= E((s(n] + e[n])e[n + m])
= E(s(n])E(e(n]) + tl>ce(m] siDce s(n] and e(n] are indepe11dmt and statiOIWY·
= tl>.._(m] where - assumed e(n] has zero mean.
Taking the Fourier transform of the above equatiOD, - get:

••• (ei"') =••• (ei"').


54

(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].

Taking the Fourier transform, we get:

+.. (.;w) = B(dw)H(e-jw)


1 1

Taking the Inverse Fourier transform, we get:


alnl
,p.. [m] = - -•.
1-a
(b) Using part (a), we get:

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:

y[n] = z[n] - z[n- 1].


(a)

t/>n[m] = E(y[n]y[n + m])


= E((z[n]- z[n- 1])(z[n + m]- z[n + m- 1]))
= E(z[n]z[n + mj)- E(z[n]z[n + m- 1])- E(z[n- 1]z[n + mj)
+E(z[n -1]z[n +m -1])
= t/>00 [m]- tl>•• [m- 1]- t/>.,[m + 1] + t/>•• [m]
= 21/>•• [m]- t/>•• [m- 1]- tl>•• [m + 1]
= 2<T~o[m] - cr~o[m- 1]- ~o[m + 1].

----~--~----~--
-1 0 1 m

To get the power spectrum, we take the Fourier transform of the autocorrelation function:

+yy(eiw) = ~ _ CT~e-;w _ ~e'w


= 2a!- 2a'!cos(w)
= ~(1 - cos(w)).

·-
(b) The average power of the output of the system is given by ti>.. [O]:

tl>.. [o] = 2<1!.


(c) The noise power increased by going through the first-bad:ward-dilference system. This tells us
that the first backward di1ference amplilies the noise of a signal.
56

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:

</>•• (n] = a!o(n].


Therefore,

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:

y(n] = z(n]• h[n]

= L
.. z[k]h[n- k].

Taking the expected value of both sides:

""' =
..L E{z[n]}h[n- k]

.... = 0, ifm. = 0.
So,

u! = E{y2 [n]}

= E {...,.i;.., h[m]z(n- m] >=f;oo h{k]z[n- k]}


=
..E L.. h[m]h[k]E{z[n- m]z[n- k]}
=
.. ..
I'JUII'-oct=-oo

L h[m]h[k]o[m - k]

,.. =
a!E
..
m=-oo&=-oc

a! L h'[m].
• m,.;:-oo
57

2.86. Using the solution to problem 2.85:


(a)
00

~ =~ :E h~[kJ

this statement is TRUE, because o:[n] is a white noise sequence.


(b) Since y[n] is not a white noise sequence, this statement is FALSE.
(c) Let
h, [n] = anu[n]
h2[n] = bnu[n].
These systems are cascaded:
h[n] = h1 [n] • hz[n]
= :Ea•bn-•,

=
....,
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

y[n] = L h[k]:r[n - k], n ?: 0


b-oo

E{y[n]} = E {.~.., h[k]o:[n - k]} , n?:O


n

= L h[k]E{z[n- k]}
11:=-oo

{ m, ~~-oo h[k], n?:O


= 0, n<O
(b)
tl>,.[n,, n2] = E{11[n,)11[n2]}.

= E {J.., h[k]z[n 1 - k] .I;.., h[m]z[n2 - m]} , n ?: 0


n, no
= L L h[A:)h[m)E{z[n, - A:)z[n2 - m]}
1:--oo m=-oo
,., "2
= L L h[k)h[m]t/>u[n 1 - k,n2 - m].
t=-oom=-co
58

(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.

When cr![n] is very small (approximately zero), then:

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)

t/>u(n] = t/>-(n]• h(n]• h(-n]

= cr; (G)" Grn u(n)• u(-n])

= cr;,p,,[n].
2.90. (a)

<P•• (n] = E{y(k]z(k- n]}

= E Ltoo h(r]z(k - r) Joo h(m)v(k - n - m)}

Note that t/>•• (n] = E{z(p)v(p- t]}, therefore:


00 00

tf>.,(n] = :L :L h(r)h(m]E_{z(p)v(p- (n + t- s)]}


r=-oom=-oo
= h[-n)• h(n)• t/>•• [n].
+,,(eiw) = JH(e'w)i"+•• (e'w).
(b) No, consider z(n] white and
v(n) = -z[n)
tl>•• (n) = -~6(n]
+ •• (e'w) =-cr;.
Noting that JH(eiw)l' is positive,

Hence, the cross power spectrum can be negative.


2.91. (a) Since f[n] = e(n]- e(n- 1],

+u(e'w) is given by:

+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:

<Pf/{m] =..~(2o[m]- o{m + 1]- o[m- l]l.


(c)

~.,(d"') = H2 (d"')H2 (e-;"')~ff(d"')


= { ~(2- 2cos(w)) , [wl <We
0 , We< lwl S .-.
._.--,--.-a---~

(d)


61

Solutions - Chapter 3

The z- Transform
63

3.1. (a)

z [( 21)" u[n]]= ~ (1)"


2 z-• =~ 2z1)" = 1 - 1 1
00 00

!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)

Z[G) n u[-nJ] = ioo (2z)" = 1! 2z lzl < 2


I

(d)

Z[c5[n]J = z0 = I all z
(e)

Z[c5[n- 1]] = ,-• I• I > 0


(f)

Z[c5[n + 1]] = z+ 1 0~ lzl < oo


(g)

[( 21)" (u[n]- u[n- ]= ~ 1)" = 1 (2:)-l


9
( )-1o
Z 10]) 2, 1 -=_ (2
I• I> 0

3.2.
n O<n<N-1
:r[n] = { N, N-$ n- = n u[n]- (n- N)u[n- N]

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)

z.(n] = alnl 0 < lal <I

-1 00

x.cz> = L a-"'z-" + La"z-"


ta=-ao n=O
00 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)

z,(n] = zo(n -I]• zo[n] <* X,(z) = z- 1Xo(z) · Xo(z)


2 2

X,(z) = z-1 ( zNzN1(~-I I) ) = z2N-1


I ( N
zz _:-I
I) z ¥ O,I

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.

X(z) = (1 + 2z}(1 + 3:- 1 }(1- z- 1 )


= 2z + S- 4z- 1 - 3z-2
= __L"" ..,
:z:(n]z-n

Therefore,
3.6. (a)

X(z)- _ _,_!.....,. 1
- 1 + lz-1 1•1 > 2
2

Partial fractions: one pole-+ inspection, :z:(n] = (-i)nu[n]


Long division:
1 - jz-1 + iz-2 + ...
1
1+iz- 1
1 + !z-1
- !z-t
-jz-1 - iz-2
+ lz 2

+ ~z-2 + jz-3

= :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'

=> :r[n) =- ( -D n u[-n- 1)


(c)
1- lz-1 1
X(z)- 2
- 1 + ~z 1 + !z-2 1•1 > 2
Partial Fractions:
• •
-3 4 1
X(z) 1•1 > 2
= 1 + t• 1 + 1 + !•
:r[n) = [-3(-~r +4(-~rJ ..[nJ
Long division:
1 + (-~- !l·- 1 + (- 136 + 1)z-2 + ...
1
1
C-l-!l•- 1

(-~- !Jz-1 + tH- !l·-·


H+W+!ll• 2

(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

:(n] = -a6[n]- (1- a2 )a-<•+>lu[n]


Long division:

-a.+z-1 I 1
·-'-•)
- ( --;y-z -· + ...

1 - a.z-1
(a- 1 - a)z 1

=- x[n] = -a6[n]- (1- a2 )a-<•+>lu[n]


3.7. (a)

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

H(z) causal~ ROC 1•1 > 1.


(b) Since one of the poles of X(z), which limited the ROC of X(z) to be less than 1, is cancelled by
the zero of H(z), the ROC of Y(z) is the region in the z-plane that satisfi.. the remaining two
!
constraints lzl > and lzl > 1. Hence Y(z) converg.. on 1•1 > 1.
(c)
1 I
Y(z) = 1 - --
lz-•
+ 3
1 + z-'
Iz I > 1
2

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

(a) h[n] causal ~

h[n] = - ( 43)n u[n]- (- 3)n-l


4
u[n -1]
(b)
_!z-1 3
Y(z)
= X(z)H(z) = (1 - i•_ 13)( 1 + {z-') i < 1•1
_.!. .!.
= 13 + 13
1- lz-1 1 + !z-1
3 •
Therefore the output is

y[n] = -138(1)n 8(-i3)n u[n]


3 u[n] + 13
(c) For h[n] to be causal the ROC of H(z) must be f < lzl which includes the unit circle. Therefore,
· h[n] absolutely summable.
3.9.
1 + z- 1 2 1
H(z)- --~-..,.,.
- (1- ~z-')(1 + tz-') - (1- ~z-1) (1 + tz-1)
(a) h[n] causal~ ROC outside 1•1 = ~ ~ lzl > !·
(b) ROC includes 1•1 = 1 =>-stable.
(c)

y[n] = - 1 ( - 1)" u[n]- 4(2)"u[-n- 1]


3 4 3
-1 !
Y(z) = 1 + lz- 1
+ 3
1- 2z-1
• 1 + .-1 1
=
(1 + t• 1)(1- 2z- 1
)
4 < I• I< 2
Y(z) _ (1- jz- 1)
X(z) =
H(z) - (1- 2z-1) 1•1 < 2
z[n] = -(2)"u[-n- 1] + ~(2)"- 1 u[-n]
2
(d)

h[n]=2G)" u[n]- (-Dn u[n]


3.10. (a)

z[n] = G)" u[n _ 101 + G)" u[n _ 101


= Gr Gr u[n] + u[n]

-[( Gr (ir) + (u[nJ-u!n -nn]


The last tern> is finite length and cooverges everywhere except at z = 0.
Therefore, ROC outside largest pole < 1•1- i
69
(b)

:z:[n] = { 1, -10.$ ." .$ 10


0, othefWISO

Finite length but has positive and negative powers at z in its X (z ). Therefore the ROC is 0 <
1•1 <(X).
(c)

:z:[n] = 2"u[-n] = C)-·


2 u[-n]
:z:[-n] ... X(1/z)

G) •u[n] => ROC is


1
lzl > 2

(d)
cr·
2 u[-n] => ROC is lzl < 2

:z:[n] =[ (D •+<- (e''/ 3 )"] u[n- 1]

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] = u[n + 10]- u[n + 5]


= { 1, -10 .$ n .$ -6
01 otherwise

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] = (21)n-1 u[n] + (2 + 3j)"- u[-A- 1] 2

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

:z:[n] causal => X(z) = L:z:[n]z-•


n=O

· 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

The pole is at -2, and the zero is at 1/2.


(b)
1- lz-1
x.(z) = (1 + ~z-1 )3(1- jz-1)
The poles are at -1/2 and 2/3, and the zero is at 1/3. Since :r2 [n] is causal, the ROC is extends
from the outermost pole: 1•1 > 2/3.

(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.

G(z) = sin(z- 1 )(1 + 3z-2 + 2z-4 )


1 z-3 z-5 z-7 -2 -4
= (z- - -
3!
+ -5! - -)(1
7!
+ 3z + 2z )

= Lg[n]z-"
ft

g[ll] is simply the coefficient in front of z- 11 in this power series expansion of G(z ):

g[llJ = -w1 + 9!-


3 2
11!.

3.14.
1
B(z) = 1- !z- 2
• 1
=
(1- !z- 1 )(1 + !• ')
0.5 0.5
= 1-lz- 1 + I+lz
2 2

Taking the inverse z. Transform:


1 1 I I
h{n] = -(-)"u[n] + -(-- )"u[n]
2 2 2 2
So,
I 1
A,= 2; 0:2 = --;
2
3.15. Using long division, we get
I -10
1- iOfiZ
B(z) =
1 - ~=-l

Taking the inverse z-transform,

h[nJ={ (l)", n=0,1,2, ... ,9


0, otherwise

Since h[n] is 0 for n < 0, the system is causal.


3.16. (a) To determine H(z), we first find X(z) md Y(z):
1 1
X(z) = 1- fZ I 1- 2z I

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

The pole-zero plot of H(z) is plotted below.

H(z)

(b) Taking the inverse z..transform of H(z), we get

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

Y(z)(l- iz- 1 + z- 2 ) =X(z)(l- z-1 ),


and thus
Y(z) 1- z- 1
H(z)
= X(z) = 1- ~z-1 + z-2
73

1- z- 1
= (1-2z 1 )(1-iz- 1 )

= 2/3 + 1/3
1- 2z 1
1- ~z

lithe ROC is
(a) izl < i:

(b) t < izi < 2:

(c) izl > 2:


h[n] = 32 2nu[n] + J(2)nu[n]
1 1

= 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

Taking the inverse 2>-transform:

h[n] = -2o[n] + 31 (- 21 )nu[n] + 3u[n].


8

(b) We use the eigenfunction property of the input:


y[n] = H(~•l 2 )z[n],
where
1 •
= -2+
1+
3
fe-J•/2
+ 3
1- e-i•/ 2
1 1
= -2+-1--1
l·+-L-1.
- 'j) +J
-2j
= -~-··
,+f
Putting it together,
74

3.19. The ROC(Y(z)) includes the intersection of ROC(H(z)) and ROC(X(z)).


(a)

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)

The ROC is 1•1 > f·


3.20~ 1n both cases, the ROC of H(z) has to he chosen such that ROC(Y(z)) includes the intersection of
ROC(H(z)) and ROC(X(z)).
(a)
1- 4~z- 1
H(z)-
- 1 + lz-1
3

The ROC is 1•1 > ~­


(b)
1
H(z) = 1- !z-1

3 •2 1. (a) The ROC is 1•1 > t·
y[n] = 0 n <0
n " 1- a-<n+l} 1- a"+l
y[n] = L:;:r[k}h[n- k) =I;an-t =a" 1
_a 1 =
1
_a 0 :=; n < N- 1
A:=O .b:O
N-1 N-1 1 -N 1 -N
y[n] = L :r[k}h[n- k] = L an-t =a" 1=. :-1 = o"H : ~1 , n ~ N
.t:=O .t:=O

(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

4 n+l ( 1-0-N) n?. N


•-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

h[n] = -40[n]- 2 G) n u[n] + 7 (D n u[n]


76

(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)

1- z-• 1•1 > 1

....
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

Obtain a proper fraction:


77

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]

3.26. (a) :r(n] is right-sided and


1 - !.z- 1
X(z) = 1 + lz '
Long division:
igZ -2 + ...
1 +
1+ ~z- !1 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)

X(z) = 1- 11 z- 3 (z( > (3) -l => causal


3
By long division:
1 + jz-3 + lz-• + ...
1- tz- 3 1
1 - iz-3
+ !z
3
3

+ 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,

1 (-1)"+1 S8 (-1)" 1568 2700


z(n] = ~ (n+1) 2
5 u(n+l]+ ( )2
35 2 u(n]+ (3S) 2 (2)"u(-n-l]- (3S) 2 (3)"u(-n-l]
(h)
-1 z- 2 z- 3 z- 4
X(z) = e' = 1 + z- 1 + 2! + 3! + T! + ...

Therefore, z(n] = n!1 u(n].


(c)
3
z - 2z 2
X(z) =- - = z2 + 2z + =--=--=-
z-2 1-2z
[z[ < 2

Therefore,
z(n] = o(n + 2] + U(n + 1]- 2(2)"u(-n- 1]
3.28. (a)
d
nz(n] ~ -z dz X(z)

z(n- no]~ z-no 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)

X(z) = sin(z) = L..


t=O
(-1).
(2k + 1)!
z'l+l ROC includes izl = 1

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.

X(z) = e' +e'i• z~0


0
.. 1 .. 1 (1)" 1 .. 1 1
X(z) = L --;z" + L--; - = L (- )··-· + L ,.-· =
n=o n. n. z
n=() a=-oo n . n. n=O
:t[n] = -1I'+ o[n]
n .

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] = {cos ;2n) u(n]- en• {cos ~n) u[n- 1]


en•

= .n• (cos ~n) (u[n]- u(n- 1]) =o(n]


Therefore, X(z) = 1 for alllzl.
3.32. From the pole-zero diagram
3
lzl > 4

y(n] = :z:(-n + 3]= :z:[-(n- 3)]


z- 3 z- 1
~ Y (z) = z -• X (z - 1 ) = .,-,----,,..-....,..,.--,.--,..,.
(z 2- z-1 + iHz-1 + ~)
8/3
= z(2- 2z + z2)(~ + z)
81

Poles at 0,-~,1 ± j, zeros at oo

z[n] causal =0> z[-n + 3] is left-sided =0> ROC is 0 < jzj < 4/3.

-4/3

3.33. From pole-zero diagram


X(z) = z2 + 11
z-2
(a)

y[n] = (-21)n z[n] =0> Y(z) =X(2z) =-


4z2+1
-
2z- 2 1

zeros
poles

Y(z)

(b)

w[n] =cos(";') z[n] = ~(ei•n/ 2 + e-i•nl )z[n]


2

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

h[n] stable => h[n] = SD[n]- 2nu[-n- 1]- 3 G) n u[n]

(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

u[n]- 2n+1u[-n -1]


2

(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-<

u[n- 1]- L6[n- 4- 41:]


1•1 > 1

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)

H(z) = Y(z) = 4z(1 ~ z- 1) 1•1 > 2


1
X(Z) 1- :;z 1

(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

(c) The ROC of H (z) includes lz I = 1 => stable.


(d) From part (b) we see that h[n] starts at n = -1 =>not causal
3.37.
1 l
X(z) = 1-lz-1
3 4
+ 1-2z-1
2

bas poles at z =! and z = 2.


Since the unit circle is in the region of oonvergence X(z) and z[n] have both a causal and an anticausal
part. The causal part is "outside" the pole at !.
The anticausal part is "inside" the pole at 2, therefore,
z[OJ is the sum of the two parts ·

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

Therefore using a contour C that lies outside of jzj = 1 we get

y[l] = 1 j 2(z + l)z"dz


27rj !c (z- !Hz+ ~)(z- 1)
= 2(i +I)(!) 2(-t + 1)(-f) 2(1 + 1)(1)
<!+!H!-1) + H-!>H-ll + (1-i)(I+!l
18 2
= -s-- 5 +6=2
3.39. (a)

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:

V(z) = X(z)- W(z)


W(z) = V(z)H(z) + E(z)

we solve for W(z):


• H(z) 1
W(z) = l+ H(z) X(z) + 1 + H(z) E(z)
(b)
_,
H1(z) - H(z) - ~ =z-1
- l+H(z)-1+ 1 :~~~

= 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 &ltered 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-

FUrther, since the system is causal, h(n] = 0 for n < 0. Therefore,


10
h(O] + La 1 h[-k] .= h(O] = 6[0] = I.
k=l

(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

h[n] = (0.9)" cos( n)u[n],


..
4
we only need a second order difference equation:

The z-transform of h(n] can be found from the z-transform table:


1 o.o
H(z)- - 7>
- (1-0.9ei•/•z 1)(1-0.9.-J•/<z ')

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

h(n] = (43)n u(n]- 2 (3)n-l


4 u(n- 1]

(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

The ROC is i < lzl < 2.


(b) The following figure shows the pole-zero plot of Y(z). Since X(z) has poles at 0.5 and 2, the poles
at 1 and -0.5 are due to H(z). Since H(z) is causal, its ROC is lzl > 1. The ROC of Y(z) must
contain the intersection of the ROC of X(z) and the ROC of H(z). Hence the ROC of Y(z) is
1 < 1•1 < 2.

(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

Taking the inverse z-transform, we find


2 2 1
h[nj = 6[nj + 3u[n]- 3(-2)"u[n]

(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

X(oo) = lim L:;z[n]z-n


•~oo
= z[O]
n=O

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

~ z[n] = (D n+l u[n + 1] => right-sided but non-caosal.


89
3.49.

z[n] = o[n] + ..O(n- N) lal < 1


X(z) = 1 +az-N
a2z-2N sz-3N
X(z) = logX(z) = log(1 + az-N) = az-N- +4 - ...
2 3
Therefore,

:[nJ = z:
00

1<=1
(-1)"+1
k a••rn - kNJ

3.50. (a)

:t[n] =:t[-n] =* X(z) =X G)


Therefore,

i.e., 1/zc is also a zero of X(z).


(b)

:r[n] real= :r[n] = :r"[n] = X(z) = X"(z")


Therefore
X(zc) = 0 = X(z,j)
i.e., z0 is also a zero and by part (a) so is 1/zf,.
3.51. (a)

Z[:r"[n]] = n=~oo X"[n]z-n =


00 (
n~oo
00
z[n](z")-n ) " = X"(z")
(b)
00 00

Z[:r[-n]] =L z[-n]z-n =L z[n](z- 1)-n = X(z- 1)


n=-oo n.=-oo

(c)

Z[&{z(n]}] = Z [ z[n] ~ :r"[n]] =~ [X(z) + X"(z")]


(d)

Z[Im{:r[n]}] = Z ["'[n] ;;"'"[n]] = ;j [X(z)- X"(z")]

3.52.
00

z1(n) = (-Wz(n) '* X1(z) = L (-1)n:t(n)z-n = X(-z)


--oo
The poles and zeros are rotated 180 degrees about the origin.
90
N-l
3.53. (a) . - l::r(n]sin(nw}
9• (w} _ (Zm{X(e'w})) ~ ( } _ -;:::=<>;:=----
-tan
-l
7U{X(eiw)} = tanvz w - N-l
2: :r(n] coo(nw)
n=O

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

for N- 1 values of w, in the range 0 < Wt < "·


(b) :r(n] = 6(n] + U(n -1] + 36[n- 2] = X(z} = 1 + 2z- 1 + 3z-2
e.(w} = tan-• ( -2sin(w)- 3sin(2w) )
1 + 2 cos(w) + 3 cos(2w)

Consider the values 8: ( i) = 5: and 9s ( 23Jf) = 5


;, which give the equations

tane.G) + :r;od:t(lJ(tane.G)cos~+sin~)
+ :r(2] (tan e. G) cos .. + sin,.) J = o

tan e. c;) + :0 [:r(1] (tan B. c;) cos 2; 2


+sin ; )

+ :t(2] (tan B. c;) 4


cos ; +sin~)] = 0
1
1 + :r(O] (:r(1] · 1 + :r(2]· -1} = 0

- ~ + :r;O] (:r(1] + 2~ + ~) +:t(2] c~- ~) = 0

:r(O] + :r(1]- :t(2] = 0 } { :r(1] = 2:r(O]

-:r(O] + 2:<(1]- :r[2] = 0 = :t(2] = 3:r(OJ

Therefore
:r[n] = :r(0](6(n] + U[n- 1] + 36(n- 2]}
where :r(O] is undetermmed.

3.54. :r(n] = 0 for n < 0 implies:


00 00

lim X(z)
~-oo
= .z-oo~
lim ":r(n]z-• = :r(O] + lim ":r[n]z-• = :r(O]
;~-oo ~
n.-o n.=l
. 91

For the case :r:[n] = 0 for n > 0,


0 ...
•lim
....o X(z) ,..._.., ,....., L :r:(-n]z" = :r:(O]
lim "~ :r:(n]z-" =:r:(O] +lim
=.-o
n=1

a.ss. (&)
... 00

es.(n] = 2: :r:[A:J:r:(n + A:] = 2: :r:(-i;]:r:(n - A:] = :r:[-n]• :r:[n]

C.. (z) = X(z- 1)X(z) = X(•)X(z- 1)


X(z) has ROC: rll < 1•1 < r£ ud therefore X(z- 1) has ROC: r£ 1 < 1•1 < rji 1. Therefore c•• (z)
has ROC: maz{r£ 1,rll) < 1•1 < miD[rji 1,r£)
(b) :r:(n] = a"u[n] is st&ble if 1<>1 < 1. In this case
1 1
X(z) = 1- oz-1 lol < 1•1 ud X(z-1) = -1-az
-

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

This implies th&t


es.[n] = _!._, [a"u[n] + a-"u(-n- 11)
1-a
Thus, in SUIDI!l&rY, the poles &re &t a and a- 1 ; the zeros &re &t 0 and""; and tbe ROC of c•• (z)
1
is lal < izl < la- 1·

(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

X(z) = X(:r +ill)= u(.,,v) + iu(%,11)


then for the derivative to exist at z, we must have

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,

%(n] = ~if X(z)z•-Jdz does not exist.

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(z) = A(zo) + L"" -A•(zo)


n.
-(z- Zo)•
1
n=l

where A(zo) = 0. Thus

Res (X(z) at z = zo] = X(z)(z- zo)lz=ze = B(z)


A(z)
I
I
zsze
=· B(z)(z- Zo)
~ A•(Zo) (z- Zo)• •=ze
....4-t, n!

=
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)

4.2. The discrete-time seque~>ce


z[n] = coa(
4'"' )
results by samplillg the coatiDuoWI-time sigDal

Ze(t) = ms(o.,t).
Since w = nT aDd T = 1/1000 seconds, tbe signal &eque~>cy could be:

n.. = -.4 . 1000 = 25(hr


or possibly:
I n. = (2" + "4>. 1000 = 22so...
4.3. (a) Since z[n] = z.(nT),
..n
3 = 40()/)r.nT
l
T = 12000

(b) No. For example, since


cos( " n) 77f
3 = cos( 3n),
T can be 7/12000.
4.4. (a) Letting T = 1/100 gives
:r[n] = Ze(nT)
= sin ( 2().-n
1 ~) + cos ( 401rn 1 ~)
= sin(";) +cnsC;')
(b) No, anotber choice is T = 11/100:
z(n] = Ze(nT)

= sin (201m;~) + COl ( 40.-n 1~)


= m(ll;")+ms(22;")
= liD(";) +cos C;')
4.5. A plot of H(eJw) appears below.
96

• .(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

(c) '· = 625Hz

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

-4lt -2n 0 21t 47t (J)

(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

(b) Tbe desired respoDSe:


HUO) ={ 1 + ae-i•• 0 , for 101.5 f
0, · otheronse
Using w =OT, we obtain a discrete-time system which simulates the above respoDSe:
H(ei"') = 1 + ae-;!f<

(e) We need to take the inverse Fourier trauform of the discrete-time impulse respoDSe of pact (b).

h[n] = 2~ L: B(ei"')ei-' dul

== .!..f• (1 + ae-;!f<).,;- dul


2• -·
(i) CoDSider the ease wbeD T• =T:
h[n] = .!..f•
2w -•
(ei..,. +aeiw(ft-l))duJ

= _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

Hence, we may infer that the impulse response of the system


hjn) =T · u[n).
(c) We...., the expression for 11[n] as pven and eumlne tbe limit
.
lim y[n] = lim T · ""
£.... z[k]
•-oo

= T·
n-ao
..
L
t-=-ao

z[k]
tz:-oo

Recall the analysis equation for the Fourier transform:

X(~w) = L
.. :r[n]e-jwn
n.=-co

Hence,

(d) We...., the result from part (c). Noting tlw



X(~wl = f f: x.c;; + ;~1rr ).
r=-oo

Thus, we· have

From the pven information, we seek a ....Iue ofT such that:

= 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

4.9. (a) Since X(ei"') = X(ei<w-•1), X(ei"') is periodic with period r.


(h) Using the in'nllle DTFI",

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].

All odd samples of s(n] = 0, bec•nse s(n] = -s[n]. Hence s[3] = 0.


(c) Yes, y[n] COIItains all eVeD samples of s[n], &Dd all odd samples of s[n] are 0.

s[n] = { Jl[n/2], n" eVeD.


0, otheronse

4.10. Use z(n] = z.(nT), and simplify:


(a) s[n] = cos(2rn/3).
(b) z[n] = sin(bn/3) =- sin(2,-n/3)
(c) :z:[n J = "•i>•n/5!
•n/$000

4.11. (a) Pid: T such that

z[n] = s.(nT} = sin(10rnT) = sin(.-n/4) => T = 1/40


There are other choices. For example, by realizing that sin(.-n/4) = sin(!hm/4), we find T. = 9/40.
(b) Choose T = 1/20 to make z[n] = s.(nT). This is Wlique.
4.12. (a) Notice first that H(ei"') =10jw, -r :;:; w < r.
(i) After sampling,
3..
s[n] cos(
=
5 n),
y[n] = IH(ei't)icos(~; n + LH(ei't))
3.. ..
= 6.-cos(-n+ -)
5 2
. 3..
= -6.-IID(-n)
5
y.(t) = -6.. sin(61rt).

(ii) After sampling, s(n] =cos('; n) = cos('fn), 10 again, llc(t) = -6.-sin(61rt).


(b) llc(t) is what you would expect &om a diJrereatiator in the first cue but DOt in the second case.
This is becouse aliasing has occurred in the secoDd case.
101

4.13. (a)

z.(t) = sin(;t)

1/c{t) = sin( .!.(t- 5))


20
. (.. ..)
= liD -t--
20 4
rn ..
11[n] = sin(---)
2 4
(b) We get the same result as before:

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

4.16. (a} In the frequency domai.D,..., ha..o

r<· j(l))

• I _ lt/3I .c====;~. Slt/6


-lt/3
I (I)

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

•• l'T. (a) In the &equency cloJDain,


x(.,...J = { 0,
1, '"'' < 2r/3
21r/3 < lwl <'It
After the sampliDg rate change,

X4 (eiw) _ { 4/3, lwJ < 1r/2


- 0, 1r/2<lwl<"''
which leads to
:z:[n] = ~ sin(ffR/2)
3 1m

(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.

(a) Wo/2 S 7</3 ===> Wo ..... = 21r/3.


s
(h) Wo/3 1( /5 = ..,._..... = 37t /5.
(c) Since L > M, there is 110 cbance of aliasing. Hence Wo ....... = 1t .
•. 19. The nyquist sampliDg property must be satisfied: T S 1t /0.,.
4.20. (a) The Nyquist sampliDg property must be satisfied: T S ff/0., ===> F, 2: 2000.
(b) We'd have to sample 10 that X(eiw) lies between Jwl < 1t/2. So F, 2:4000.
4.21. (a) Keeping ill mmd that after sampliDg,"' = OT, the Fourier transform of :r[n] is

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

x(n] Bandpass X,._(t)


Filter

(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 ~.

1/c(t) = ~:"'• (~:~)


4. 24. The Fourier transform of lie ( t) is sketched below for each case.
105

(a)
I

-lt x5 X tal
(b)
I

(I) -2 lt X 5 X tal 21< X 5 X tal O


(c)

II X S X 1al O
(d)

(I)
-2 lt X 5 X tal 211 X 5 X ial O

4.25. (a) z,(l) = :<c(l)s(t) ~ X,UO) • sUO)

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"'):

:{n), n =Mk, =0, ±1, ±2


:,[n] ={ 0, otherwise
k

= (! ~ 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)

Additionally, X 4 (e'"') is simply X,(ei"') with the &equency am expanded by a factor of M:


00

X.(ei"') = L X,(Mn]e-;...•
---00

L :,(l)e-i(w/Jl)l
00
=
,__
107

(a) (i) X,(ei'-) ud X4(ei'-) are sketched below forM= 3, WR = r/2-


X5(ejW }
113

'' '' ''


"
'' ''

'' '' ''


'' "
'' ''
-21t -It It 21t (I)

{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.

4.27. Parseval's Theorem:

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:

Hence the amplitude of the Fourier trUISform does DOt chaa&e-


When we downsample, the downsampled sipal "'•[n] has less energy than the origiDal :r[n) because some
samples are discarded. Hence the amplitude or the Fourier trans£orm will change after downsampling.
108

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 &lteriag, 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:

X(e jm) X(e2j!ll 1 Xce2jm IHt (e j(l> ) yl (e j(l> )


tz H1(ej(l> ) '2

Y,(.;o-) is X(ri")H,(ei"') dOWilS&D1pled by 2:

Y1 (ei") = ~ { X(e';..f2)H, (ei"''J + X(e(2i(w-2<li')B,(ei(w-2•ll')}

= ~ { X(ei")H,(ei"''l + X(eilw-'-l)B, (eil't-•l)}

= ~ {H1 (ei"'') + H 1 (eil'f-•lj} X(ei")


= H2(ei")X(ei")
H,(e'") = H H,(ei"''> + H,(eil't-•l)}

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.

w1 =0 1T '* w, =2" · 500 · ..,\,. '*"-'1=-"3


2w
w, = O,T '* w, = 2" · 1000 · ~ '*""=- 3

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

4.32. (a) The highest frequency is •IT = 1t x 10000.

i ··1-------.. _____.......:_
Mlr-~--~~--~--r-~--~~--~-,

-----~

r;
r:
100

(b)
_..,
0

-· -u -u
~
1\..

__ -...cy,..,.._
~·~ 0
\
G.2
'" .. .
t)0.4
.

(c) To filter the 60Hz out,


1 3ft
""'= T!l =- - · 2.--60 =--
10,000 250
4.33.

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)

COS e; ft- i) = i(ei(\'-R-f) + .-j('fR-t)]


= !.-l(•/4)e/(h/2)R + !e~(•/4) 0 -;( .. /2)R
2 2
Since H(e"") is aD LTI system,"" caD fiDd the respoDSe to each oftbe two eigeDfunctioDS separately.

!l(n] = ~ 0 -i(•/4) H { e-fC.. /2)) eJC .. /2)R + ~ei(•/4) H { 0 -;(5•/2)) 0 -;(5•/2)R

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,

y(n] = ~·-i(•/4) H {ei<5•/2)) eJ("/2)R + ~ei(•/4) H (e-1(5•/2)) 0 -j(5•/2)n

= ~ (eii<5•/2)R-(•/2)} +e-i{( .. /2)R-(•/2)})'

= cos (5Jt n- ~) .
2 2 .

y[n)

0 n

·I ) (

4.35. The frequeDcy respoDSe H(e'"') "' H.(jOfT). FiDdiDg that

e.unJ "' UOJ 2 + !unl + 3·

1
H(ei"') = (10jw) 2 + 4{10jw) +3
;
1
-100w' + 3 + 40jw

4.36. (a) Since OT "' w, (2Jt · 100)T "' ~ => T = .:.,

!
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-·
.

4.38. X.(iO) is drawn below.


z.(t) is sampled at sampling period T, so there is no aliasing ill z[n).

.AA'A. -Jt
rr
• (I)

Inserting L - 1 zeros between samples compresses the frequency axis.

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 .-.

• _,. " 1. (j)


112

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

This requires only RL-1 multiplies, (assuming h(O] = 1.)


(d)
n+(ltL-1)
y(n] = L v[k]h[n - A:)
k-on-(ltL-1)

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)

ldcal Upsampler with gain


of I instead of L

Then we allalyze the system as follows:


:[n] = :r.(nT) DO aliasing assumed
1 T
tu{n] = r=c<nr> rate chan&e

v(n] = 111(n - 1] = 1 £"'< ( nrT - £


T) , delay at higher rate

y(n] = v[nL] = L"'< (nT - i)


113

4.41. (a) See figures below.


(b) From part(a), - see that

Y.(jO) = X.(j(O- ~)) + X.(j(O + ~ ))


Therefore,

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

T2~ ~ 2 X 2r(250) ~ T :S 500.


1

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"'):

X(ei"') = ~X.(jO) Inc, = s!ox.(jSOG-.I)


Y(ei"') = ~Y.(jO) lnaf = ~.-;ll/UMMl x.(jO) Ins,
= _1_._,...,. x.(j500.1J
500
= .-;..,• X(ei"')
114

Hence, we let
B(ei") = { 2eo, -;.. • Jwl < j
otherwise
Then, iD the following figure,

""' < j
otherwise

x[n] y[n)
H

(d) Yes, &om our analysis abow,


H2(ei") = .-jw/2

4.43. (a) Notice tim that

F.(jO)IB..,(iO)Ie-;0', 101 ~ 4001r


X.(jO) = E.(iO)JB..,(iO)Ie-;n', 4001r ~ 101 :S 8001r
{
0, otherwise

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

B (jO) = { B(ei"')l.,=nT. 101 ~ wfT


c 0, otherwtSe

FUrthermore, since Y.(iO) = H.(jO)X.(iO), the desired tranfer function is

Combining the two previous equatioiiS, we find

H(ei") = { ei(OOO...)', lwl :S 1r /2


0, 7</2 :S lwl :S r
(b) Some aliasing will occur if 21f /T < 16001<. Howewr, this is fine as long as the aliasing affects only
Eo{jO) and not F.(jO), as we show below:

' ,
,' ' ' , '
,' ''
'' '' I I ' '
' 'I
I "" ''
115

In order for the aliasing to DOt affect F.(jn), we require


2...
-T - 8001t >
- 400lr = -T->2..-
1200lt

The miDimum • il 1200lt. F« dlia choice, we get

H(ei'") = { ~l-l', ~/i ~· t1 S •


U4. (a) See the following figure:

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)

4.45. Notice first that siDce :r.(t) il time-llmited,

A= {o #ic(t)dt =J: :r.(t)dt = x.un)Jo,.o.


To estimak X.U · O) by DT pro •s!ng. "" need to aample only fast enollgh so that X.(j · 0) is Dot
aliued. Hence, ..., pick
2•/T = 2r x 10' = T =10-<.
116

The resulting spectrum satisfies

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:

X,. (I) • y[n] ~A=y[IOOOOO]


CID h[n) = Tu[n)

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

y,[n].,i f 3 1----+l•i Ejz)l t


(c) Yes, z[n] caD be recoDStrncted from 112[n] and W•[n]as demonstrated by the following figure:

x[n] wJn] y. (nl v, [n) •In1


H 3 (z) ~2 . t2
w.[n) y, (n) •. [n] s(n]
H 4 (z)
'2 t2 H 4 (z)
117

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

hi the fignre, .,[n] =:z:(2n], aDd hence,


_ { :z:[n], n eftll
,.[n] - 0, nodd
= :z:.[n]
Furthermore, i\ can be verified using \be IDFT \bat the impulse respoiiSe /4(n] corresponding w
H,(ei"') is
h.(n] = { -2/(jrn), n odd .
0, o\hennse
Notice in particular that every otner sample of the impulse respoiiSe /4[n] is zero. Also, from the
form of H,(ei"'), it is clear that H,(e'"')H,(e'"') = 1 a.nd hence h,[n] • /4[n] 6[n]. =
Therefore,

{ 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

¢ •••• (r) = E(:z:,(t)z;(t + r)).,. P•••• (n) = L: ¢ •••• (r)e-;'" dT

(a)

¢ .. [m] = E(z[n]:r:"[n + m]) = E(:z:,(nT)z;(nT + mT))


= ¢ •••• (mT), i.e., sampled antocovariance

(b) Since ¢u(m] is a sampled ¢ •••• (T)

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)

..... (.,.) = E(z.(t)z.(t + .,.)


4>•• (m) = E(z(n]z(n + m)) = E(z.(nT)z.(nT + mT))
= 4>.,.,(mT)
(b)

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

ifT= ~ ••• [m] = rm/0. = o,


sill'""

4.49. (a) Coasider the followillg plots.

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,

-1$(211) 1$(211) 'b

.1t 0 30(2>1) il -10(2Jr) 0 10(2Jr) b


(c)

:(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)

:,[n) = :.(n) • h..h(n]


h..,.[n] = { 1, 0 ~ n ~ L - 1
0, else

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.

.Pu(n] = z[n]• z[-n]


•-<.,;~) = X(ei~). x·<.,;~>

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,

111 (n] = h[n] • w[n] = w[n - k]h[k]

= lltll(n] + bou(n- 1] + cw(n- 2] + dw(n- 3] + ew[n- 4]


={ az(n/2] + <=[(n/2)- 1] + e((n/2)- 2], n even
O.:((n/2) - (1/2)] + dz((n/2) - (3/2)], n odd

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

h,(n/2] • z[n/2], n even


w,(n] = { 0, "odd
= { h2(0]z(n/2] + h2(1]:r[(n/2)- 1] + h,[2]:r[(nf2) - 2], n even
0, nodd
121

Comparing w,[n], tuo[n) with Jl[n) above:


w[n) tall give even samples if h 1 [0) = o, h 1 [1) = c,h,[2) =e. Similacly, w,(n] tall give the odd samples
if h,(n) deLays tuo(n) by ODe sample, i.e., h 3 (0) = O,h3 (1) = O,h,[2) = 0. Thus

tuo(n) = { h,(O)z[(n- 1)/2) + hz[1)z((n- 1)/2- 1) + hz(2]z((n- 1)/2- 2), n eveo


~ n~

h 2 (0) =b, 11,(1) =d, 11,[2) = 0

4.53. Sketches appear below.


(a) F"ust, X(ei"') is plotted.

The lowpass filter cuts otr at j.

• 1C/2 ftl2 I (I)

The downsampler expands the &equocy axis. Since Ro(eiw) is bandlimited to if, no aliasing
occurs.

~ It

The upsampler compresses the frequocy axis by a factor of 2.


It (I)

.~ -It It (I)

The lowpass filter cuts of at j ~ Yo(eiw) = Ro(ei"') as sketched above.


(b) Go(ei"') = l (X(eJW)B0 (ei"') + X(en"'+•l)B0 (en-•l))
122

(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)

== ~x(.,;wl [.Hl<.,;wl- S:<.,;wl]


+~X(ei<-l) [s.,(.,;w}Ho(ei<.-t,_...,...,..{ei"'->)]

The ali Mill& terms always c:aDcei. Y(ei"') Is propor&ioDaJ • X(«>") if il4(e"'}- JP.(.,...)\ is a
COI!St&llt.

X(.,;w) == 0, w/3 S lwl S w. :r:[n] C1111 be~ <i M oa co :W sipll.. De ...,..,adl is to


determiDe whether no is odd or eft11, dieD sample ., u.at no is -' ~ 1 1' ' _ . iiw; mter.
Thii reco+eu :r:[no].
4.54. (a) In the case where no is not blown, we "-mine whetheril •-.,.o.w as.~

i[n] = :r:[n] - M[s - ...,]


X(ei*') = X(e'w)-Ar.--
x<.,;wll... , == I::r:[n)H't
n
X(ei<•t2l) = -A(-j)""
If the result is real, no is even. If the rsult is imaginary, no a oiol
(b) If no is even, sample i[n) so that the even-numbered seque~~~e --..an.-. a...,.. l£,.,. io odd,
sample so the odd-numbered samples are oet to zero
(c) Filter the sampled seqnmce with a lowpass filter 'With cutoff&-. t -.fl. aDd~ 2 "IIislir an
exact procedure if ideal filters are used.
4.55. (a)

= { :r:1 [n/2], s. e¥en


w[n] :r 2[(n- 1)/2], ir. ....W
:r 1 [n] = w[2n]
:r2 [n] == w[2n + 1]

,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,

(c) The Fourier tr&Dsforms ..., sketched below.

AhrLPF

-r. JilL 11/L " (I)

LB/2T
After cosine
modula!ion
.. C\ 1C\ ..
-11 ~ (1)2 II Ill

LBI2T

.. fl 1 f\ ..
AhrHPF

-11 ~ (1)2 II Ill


):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

bound remained 2r · 10", L would become L "' 20 + M for M channels.


A branch oC the TDM demultiplexing system would be:

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

(d) ffn] = :[n]e[n]


E(/[n]) = E(z[n)e[n]) = E(z[n])E(e[n]) = 0
"' =E(/2 [n]) =E(:2 [n)e2 (n]) =E(:2 [nj)E(e2 [n]) =11!11!
rr[n, m) =E(z(n):[m)e[n)e[m)) = E(z[n):[m)) · E(e[n)e(m))

(e)

(f) Using ~e results of pan (c).


"c.-
2
- ..
2 (
I --
I-a
~.r.
1 ) ---
-1-o
4 4

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)

L =II file= IIIli M = 10

For System B, we need to resample such that


M T 2 x 10-< 1
T = T, = 1/55000 = 11

System C is simply the identity system.


4.59. The speech is first sampled at 44.1 klh, and we wish to resample it so that ~e samplillg rate is at S
kHz. There are 110 •li asing effects aaywhere in the sys~em. Hmce
L 44.1 441
-=---
M 8 80
We simply make L = 441, M =80, and "'• = "/441.
126

4.60. n,, and n, has to be chosen such that


(a) The region lfll :$; n, maps to lwl :$; fr/4:
.,
fl,T=4"==>0,=441r

(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)

V(z) = H 1 (z)(X(z)- Y(z))


U(z) = H,(z)(V(z)- Y(z))
Y(z) = U(z) + E(z)
H 1 (z)H2 (z) X 1 E( )
= 1 + H,(z)(1 + H,(z)) (z) + 1 + H 2 (z)(1 + H,(z)) z

Substituting H 1 (z) =1/(1- z- 1) and H2 (z) =z- /(1- z-


1 1 ), we fiDd
H.,(z) = .. -•
H.,(z) = (1- z- 1 ) 2

Hence the difference equation is y[n] = :r[n - 1] + f[n], where


f[n] = e[n]- 2e[n- 1] + e[n - 2].
(b)

P11 (ei") = .,:IH,,(ei"JI'2


= ..:1(1- .-;...) 12
= ..:(1- .-;")'(1 - e'")'
= .,:(2- 2cos(w)) 2
= o-:(4 sin2 (w/2)) 2
= 16o-! sin4 (w/2)
The total noise power cry is the autocorrelation of f[n] evaluated at 0:

o-} = E[(e[n]- 2e[n- 1] +e[n- 2]) 2 ]


= E[e2 [n]] + E[-2e 2[n- 1]] + E[e[n- 2]']
= fia!,
where we have used linearity of expectatioDS, and the fact that sincu[n] is white, E[e[n]e[n-.1:]] =0
for k ;,! 0.

ee
~--~~~~----·
-1t K Cll
127

(c) Since X(ei") is bancllimited, o:[n]• loo[n] = o:[n]. Bence,


w[n] = y[n]• ha[n] = (o:[n- 1] + /[n]) • h3[n] = o:[n- 1] + g[n],
where g[n] is the quantization noise in the region lwl < 1l/M.
(d) For a small angle o:, lin:"' o:. Therefore,

u: = b1 1•'.,,.,"' ,.:(21inwf2) 4 d..>

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

4.62. (a) (i) The transfer function from o:[n] to y.[n] is


.-•
r::?f
H sr (z ) = z-t = Z -1
1+ 1-.r-1

Hence y.[n] = o:[n- 1].


(ii) The transfer function from e[n] to lf,[n] is
1
H,,(z) = _, = 1-z-1
1+ ,!.-t
So

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

y(nJ =:(n- 1] + e[nJ- e[n- 1].


So here we apply the diJference equation to both &igm •-delta modulators:

lllc(nJ = e1[nJ- e1(n- 1]


,_[nJ = e1(n- 1] + e2[n]- e2[n- 1]
r 0 [nJ = lllc[n- 1]- (,_[nJ- "-["- 1])
= -e2[nJ + 2e2(n- 1]- e.[n- 2]
H ...(z) = -(1- z-1)2
P•• (w) = cr!(2- 2cosw)2

!
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)

Stable "* ROC is l $ 1•1 $ 3. Therefore,


h[n] = --81(1)ft-l
-3 u[n- 1]- 9
-(3)"- u[-n]
8
1

5.3.
1
11[n- 1] + 311[n- 2] = z[n]

z- 1 Y(z) + ~z-2 Y(z) = X(z)

Y(z) 1
H(z) = X(z) =
z-1 + l·-·
%
H(z) =
1 + !z-1
132

i) f < lzl, h[n] = (-l)n+ u{n + 1] ~ uswer (a)


1

ii) ! > 1•1.

h{n] = - ( -n n+l u[-n- 2]

= -(-D(-~fu[-n-2]
= 1( 1).
3 - 3 u[-n- 2] ~ aaswer (d)
5.4. (a)

z(n] = Gf u(n] + (2).u(-n -1]

1 z 1
X(z)= 1-lz-1- z-2' 2<l•l< 2
2

11[n] = 6 G) nu(n] - 6( ~) n u[n]


6 6 3
Y(z) = 1- tz 1 - 1- fz-1' 4 < 1•1
Y(z) - -lz- 1 (1- !z- 1)(1- 2z- 1) .:.1_-...:;2::..z-....,.1 3
H(z)- __ 2 2 _
- X(z)- (1- !z-1)(1- !z-1) -~z-1 - 1- iz 1' 1•1 > 4

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)

H(z) - Y(z) - 1 - 2z-1


- X(z) - 1- {z-1

Y(z)- ~z- 1 Y(z) = X(z)- 2z-1X(z)


11[n] - ~11[n - 1) =z(n] - 2z[n - 1]
4
i33
(d) The ROC is outside JzJ = i•
which includes the unit circle. Therefore the system is stable. The
h[n] we found in part (b) tells us the system is also uusal.
5.5.

11[n] =G) •u[n) + (U" u[n) +u[n)


1 1 I
Y(z) = + +-- JzJ > I
I -lz-1 1- iz-' I - z-1'
%[n) = u[n]
1 .
X(z) = I_ z-l, lzl >I
Y(z) 3 -ljz-1 + Jz- 2 I
H(z) - - - -:--r-.-:-+--. lzl > 3
- X(z) - 1- f.z-• + fiz-2'

(a) Cross multiplying &lid equating z-• with a delay in ti!J>e:


7 I 19 2
y[n- 2] = 3%[n]-
11[n]-
12
y[n- I]+
12 6 :r[n- 1] + J%[n- 2]
(b) Using partial fractions on H(z) we get:

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,

(I)" u[nJ- (l)n-1 u[n- I]+ (I)"


4 u[n]- (1)"-
1
h[n] = 3 3 4 u[n- I]+ 6[nJ

(c) Since tbe ROC of H(z) includeslzl =I tbe system is stable.


5.6. (a)

%[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)

:z(n] = 5u(n] * X(z) = 1 _5z- 1, 1•1 > 1

y(n] = (2 or +3 (-i)") u(n] *


Y(z) 1-z-1
Y(z) =1 _ lz- 1 + 1+ i.-
3
1, 1•1 > i
3

B(z) = X(z) = (1-lz-1)(1 + Jz-1)' l•l > i


lm

Re

(b)
7
1-z-1 -1 3
H(z) = (1- !z-1)(1 + tz-1) = (1- 1) + (1 + iz-1)' !• lzl > i

h(n] = - 2(1)" u(n] + 57(-.3)" u(n]


5 2
{c)

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

h{n] = ~ [(2)"- ( -D ") u[n]

(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]

z- 1 Y(z)- ~Y(z) + zY(z) = X(z)

Y(z)
H(z) = X(z)

=
.-•
1- ~z-1
2
+ z-2
.-•
= (1- 2z-1)(1- !z 1)
a
= 1- 2z-•

1 zeroatz~-

Three regiOIIS of couverseuce:


(a) 1•1 < ~:
h[n] 2
=--(2)"u[-n-
3
1] + -2(1)"
3 2
- u[-n- 1]

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:

(a) Yes, the system is allpaas, siDce it is of the appropriate form.


{b) No, the system is not allpaas, Iince tbe sero does Dot occur at the coajucate reciprocallocatioa of
the pole.
(c) Yes, the system is allpaas, Iince it is of the appropriate form.
{d) Yes, the system is allpaas. This system coasists of &D allpass system iD cascade with a pole at sero.
The pole at sero is simply a delay, ud does aot c:b&Dge the magDitude spectrum.

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.

H ...... (z) =2 G: t:=:)


(b) First, simply rellec:t tbe zero u z =
-3 to ita conjugate reciprocal location at z = -!· Then,
determil1e the sca.le fac\or. Tbis results iD

. ( ) _ (1 + tz- 1) {1 -jz-1)
1l,._z -3 ( )
z-1 1 + JZ I
1

The (1 + tz- 1) terms cancel, leaving

(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

B.....,(z) =3 (1- ~z- 1 )


(c) Simply rellec:t the zero u 3 to ita conjugate reciprocal locati011 at } and rellec:t tbe pole u f to its
conjugate reciprocalloc:&t.ioll u 1·Then, determine the ICAle factor.

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

grd [H,(ei")) = 2 grd [H.(ei"')) = 3


grd [H2 (ei"J) = 1.5 grd [H,(ei")) = 3
grd [Hs(ei")] = 2 grd [Ho(ei")] = 3.5

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.

5.21. h,,(n] is an ideallowpass &Iter with We =i


(a) y(n] = z(n] - z(n]• "',[n] ~ H{ei'"') = 1 - H1,.(ei'"')
This is a highpass &Iter.

1
.... . ..
-II -1114 0 7fl4 II (I)

(b) z(n] is first modulaled by "• 1owpass &ltered, 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--
••

-11-7111'8 7ltl8 It (I)

(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 • ••

-II -7112 0 lt (I)

5.22.
1- a- 1 z- 1 Y(z)
H(z) = 1 _ az-l = X(z), causal, so ROC is [z[ >a

(a) Cross multiplying and taking the inverse transform


1
Jl(n]- OJI(n- 1] =z[n]- ..-z(n- 1]
(b) Sinoe H(z) is causal, we know that the ROC is lzl >a. For stability, the ROC must include the
Wlit circle. So, H(z) is stable for Ia! < 1..
(c) a=! .

lzf > 112

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 ...-.;..

IB(..;..)I 2 = H(..;..)B"(ei") = 1 - .. - ·~;..


1 1 - .. -•..;...

1 - ... -... 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

cosO= 1, cos .. = -1, so there are no restrictioas.

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.

(b) Type I TypeD Type ill Type IV


Lowpus y y N N
Bandpass y y y y
High pass y N N y
Bandstop y N N N
Dilferenti&tor y N N y
143
5.24. (a) Taking the z-trausform ol both sides and reananging

B(z)- Y(z) _ + z-• -l


- X(z) - 1- tz-2

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

h[n] = 6[n + 1] + 6[n] + 46[n - 1] + 6[n - 2] + 6[n - 3]


This system has the frequency respcmse
H(eiw) = ei"" + 1 + 4e-;w + e-jl.wl + e-jlf.J
= .-JW(.,i,._ + ,,w + 4 + .-jw + .-j2w)
= .-1w(4 + 2cos(w) + 2cos(2w))
IH(eiw)l = 4 + 2cos(w) + 2 cos(2w)
LB(e'w) = -w
grd[B(,,w)] = 1

5.26. (a) A labeled pol~zero diagram appears below.

lm
1 zeroatz=-

Re
144

The table ol common z-transfarm pain ci...,. us


1
(r ,. llD""'n
. ) [ ] {rlin""')z-
u n ._. 1- (2r cos""')z-1 + r'z-2' lzl > r
which enables us to derift /a(n).

la{n) = (-.-1 - ) (r"lin""'n)u(n)


. liD""'
rz- 1 rz- 1
B(z) =1 {2r c:cs""')z-1 + r'z-• = (1- rz-1)2' lzl > r
Again, using a table lookup ci...,. us

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)

-It -rd2 0 It (I)


145

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

y(n] = 3u[n] + ( -D" u(n]

(ii)

:r:(t) =50+ 10cos(201rt) + 30cos(40rt)


1
T= t=nT
40
,.
:r:(n] = 50+ lOcos n "- 30c:os7n
2
= 50+ s.,i(n•/2) + Se-i(../2) + ts.,i•• + 15e-in•

Using the eigenfwlction property:


11[n] = 50H(ei0 )+s.:i<"•l2 >H(ei<•l2>)+5e-i<n•f21H(e-i<•I21)+1W""H(ei")+lSe-1"" H(e-1")
' 4
H(e'"') =1 - l - 1 ....,
1 e-JW - 1 e-J
H(e'0 ) = 6, H(ei<~l•>) = 7 (H)- iH. H(e-i<•/21) = 7 (H)+ iH.
H(e'") = 4, H(e-••) 4 =
y(n] =300 + 24v'2 cos (in - tan - 1 G)) + 120 cos ...n
I46

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)

h[n] = or u[n]- 28(2)"u[n]- 48(4)"u[-n -11

(b)

1 28
H 1(z) = 1- !z-1 - 1- 2z-1
48
H2(z) = _4z 1
1
5.30. (a)

H(z) = (z + ~~z- !l = z-(M-2) ( 1- ~z-2)

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]

Let v[n] = z[2n],


1
y[n] = v[n- (M- 2)/2]- 4v[n- (M/2)]
Therefore,
1
g[n] = 6{n- (M- 2)/2]- 46[n- (M/2)], M nen

G(z) _ z-<M-2)/2 _ !z-M/2


- 4
5.31. (a)

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

r[n] = 54(1)" 1 -1)- u[n)


2 u[n]- '5(3)"u[-n
(b) ROC includes z = oo so h{n] is caasal. Since both h[n) and :[n] are 0 for n < 0, we know that y[n)
isalsoOforn<O
H z - Y(z) - z-2
( ) - X(z) - 1- iz-1 + Jz-2

Y(z)- ~z- 1 Y(z) + ~z-2 Y(z) = z-2X(z)


7 3
y[n] = z[n- 2] + 2y[n- 1]- v[n- 2)
2
Since y[n] = 0 for n < 0, recursion can be done:

r[OJ =o, y[1) =o, v[2) =1


(c)

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)

X(z) = S(z)(1- .-aaz- 1 )


H 1 (z) = 1- .-••.-•
There are 8 zeros at z =.-•.; i• for k =0, ... , 7 and 8 poles at the origin.
lm

ftlordorpolo

(b)

Y(z) = H,(z)X(z) = H,(z)H1 (z)S(z)


H,(z) = H1~z) = 1- e!loz-•
1•1 > e-• Hable and causal, 1•1 < .-• not causal or Rabie
148

(c) Only the causal ho(n] is stable, therefore only it caD be used to recover •[n].

e-on, n = 0,8,16, ...


h[n 1= { 0, otherwise

(d)

z[n]• ho[n] = 6(n]- .-aoo[n- 8]


+ e-ao(6(n- 8]- .-ao.s(n- 16])
+ e- 1ao(6(n -16]- e-ao6(n- 32]) + · · ·
= 6(n]
5.U.

h(n] =G)" u(n] +G)" u(n]


(a)

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)

h,[n] = { h,[n], n < 100


0 100 n;
(c)
1

H(z) = ~~:~ = ~ h[m]z-m, N- 10


1
+1
m=O

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)

=oo :o- pole at ~<•1•1


20Jo&10 1H(~<•IIl)l

20Jo&,oiH(~<'-I•I)I = -oo :o- aero at ~(h/51


I!Monmo:e at "' .. 'f :o- pole Wide l1l1it clzcle here.
Since tbe impuloe respOiiiH is real, t.be poles aDd seroo must be ill CODjupte pairs. The remaining
2 zeros are at aero (the number of poles al-yo equalo the number of oeroo).
149

lm

Re

(b) Since H(z) has poles, wet:- h{n] is IIR.


(c) Since h{n] is causal and IIR, it cannot be symmetric, and thus c:annot have lineae phase.
{d) Since there is a pole at lzJ = 1, the ROC dOe. Dot include the unit circle. This means the system
is DOt stable.
5.36. (a)

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)

Cross multiplying and taking the inverse z-transform gives

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

Note that since h[n] is causal, ROC is lzl > 1.


(c)
150

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

H(z) = (1-0.!If:IO·hz- 1 )(1-0.9e-JO·••z- 1 )(1-1.~·••z- 1 )(1-1.25e-;o.a.z-•)


H,(z) = (0.9) 2 (1.25)2 (1- (10/9~-hz- 1 )(1- (10/9)e-JO·'•z- 1 ) •
(1- o.se>O·.... - 1 )(1- o.se-;o.a. .,- 1 )
Hz(z) = (0.9)2 (1- (10/9~·a.z- 1 )(1- (10/9)e-;o.a.z- 1 )(1-1.~·••z- 1 ) •
(1 - i.25e-JO.k .-•)
Hs(z) = (1.25)2 (1- O.!if:IO·'•z- 1 )(1- 0.9e-;o·•• z- 1 )(1 - o.se'O·'• z- 1 ) •
(1- o.ae-;o·•· .-•>
Hz(z) baa all its zeros outside the UDit circle, aDd is a maximum phase sequence. Hs (z) has all its
zeros imide the UDit circle, aDd thus is a minimum pbaae sequence.
(b)

H(z) = 1 + 2.5788z-• + 3.4975z-• + 2.5074z- 3 + 1.2656z4


h(n] = 6(n] + 2.57886(n- 1] + 3.49756(n- 2) + 2.50746(n- 3) + 1.26566(n- 4]
H1 (z) = 1.2656 + 2.5074z- 1 + 3.4975z-• + 2.5788z-• + z 4
h 1 (n] = 1.26566(n) + 2.50746(n- 1) + 3.49756(n- 2] + 2.57886(n- 3] + 6[n- 4]
Hz(•) = 0.81 + 2.1945z-• + 3.3906z- 2 + 2.8917z-• + 1.5625z-•
h.(n] = 0.816(n) + 2.19456(n -1) + 3.39066(n- 2) + 2.89176(n- 3) + 1.56256[n- 4]
H•(•) = 1.5625 + 2.8917·-· + 3.3906·-· + 2.1945·-· + 0.8lz 4
hs(n] = 1.56256(n] + 2.89176(n -1) + 3.39066(n- 2] + U9456(n- 3] + 0.816(n- 4)
(c)

n E(n) E,(n) E,(n) E,(n)


0 1.0 1.6 0.7 2.4
1 7.7 7.9 5.5 10.8
2 19.9 20.1 17.0 22.3
3 26.2 26.8 25.3 27.1
4 27.8 27.8 27.8 27.8
5 27.8 27.8 27.8 27.8

216.1 27.. %7..


26.2 %7.. %7,..
1!1.11 20.1

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

The plot ol E,[n] COITeSpODds ~ the miDimWII phase sequence..

5.S9. All zeros iDside the llllit circle meaDS the sequence is miDimWII phase. Since
Ill Ill

is true for all M, we can use M


phase leqlleJlCle.
- -
L lh.u,.[n]l ~ L lh!nJI"
2

= 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 &ltd 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.

A: h,[n]• ho[n]• h3[n] = (h1 [n]• ho[n]) • h3[n]


h.[n]• ho[n] is symmetric about n = 3, (-1 :5 n :57)
(h.[n]• ho[n]) • h•[n] is antisymmetric about n = 3, (-3 :5 n :59}

Thus, system A has generalized linear phase

B: (h,[n]• ho[n]) + h•[n]


h.[n] • h2 [n] is symmetric about n = 3, 11 we noted above. Adding h3 [n] ~this sequence will desuoy
all symmtery, 10 this does not have pneralized linear phase.
5.42. (a)

= 1, !wl <,.
A(ei")
f(w)= --, lwl < ..
153

A(J'>)
1
••• •••

••• • •• -· -all ........... .


& (I)

(b)
-· 0 • 01

h(n] = ..!_ (" .-;-~""dw = siD1r(n- a)


21r /_. r(n- a)

•••
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.

5.43. Type I: 5)'1DIIletric:, M Even, Odd Lezlcth

-
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

Type U: Symmetric, M Odd, EYeD Lencth

...
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))

(Jt~/ 2h[(M + 1)/2- n) cosw(n- (1/2)))


2
= e-iw(JI/2)

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

Type ffi: Alltisymmetric, M Even, Odd Length

= 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))

= .-;..(AI/2leil•/2) (% 2h[(M/2)- n]sinwn)

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

• Type IV: A.utisymmetric, M Odd, Even Length


AI
B(ei"') = L h(n)e-Jom
(AI-1)/2 AI
= L h[n]e-i"'" + L h[n)e-'"'"
n=O n=(JI+l)/2
(AI-1)/2 (AI-1)/2
= L h(n]e-i""' + L h{M- m]e-;...(AI-m)
n-o -
(AI-1)/2 (M-1)/2 · )
= .-,;...(AI/2) ~ h[m)ei"'((M/2)-•l _ ~ h[m)e-,;...((AI/2)-m)
(

= 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(..,)=

-· d[n]sm ...(n- (1/2)), 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:

Type IV -+ Type n C&DDOt be highpus:


157

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 to&ether gi...,. us

B(z) = (1 + 2z- 1 )(1 + o.sz- 1 )(1- o.Sei<•14 lz- 1 )(1- o.s.-i<•/41.-• l.


(1-1.25ei<•l4 lz- 1 )(1-1.25e-i(•/4 lz- 1 )(1 _ .-•)

5.47. The input :t[n] in the frequency domain loob like

!
159

X(a~'")
·~(10lt) (10lt)
5

-Q.Sit -Q.4lt 0 0.4lt O.Sn

while the c:ortespODdiDg output 11[n]looks like

Y(el'")
10e-j10..

-It -Q.31t 0 0.31t It (I)

Therefore, the filter must be

H(el'")

-It -Q.3lt 0 0.31t It (I)

In the time domain this is


h(n] = 2 sin(0.3r(n- 10)]
r(n- 10)

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)

sin2 (Sw/2) -;<w


= sin2 (w/2) e

(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,

is not necessarily linear. As a counter-ecample, consider the systems

h1 (n] = 5(n] + 5(n- 1]


h2(n] = 25(n]- 25(n- 1]
91(n] = =
h1 (n] + h2(n] 35[n]- 5(n- 1]
G,(dw) = =
3-e-;w 3-c:osw+ jsinw
LG,C&w) = tan-
1 ( sinw )
3 -cosw

Clearly, G,(e;w) does not have linear phase.


(b) This system must have generalized linear phase.

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

h,[n] = o[n] + o[n- 1]


h2[n] = o[n] + 26[n- 1] + o[n- 2]
93[n] = h1 [n]h2[n] = o[n] + U[n -1]
G,(ei") = 1 + 2e-1~ = 1 + 2cosw- j2sinw
LG,(eiw) = tan- 1 ( 1 +2sinw
2cosw
)

Clearly, G,(eiw) does not have linear phase.


5.51. False. Let h[n] equal

h[n] = sinw,(n- 4.3) +---+ H(ei'") = { ,-<.>w, lwl < w,


r(n- 4.3) . 0, otherwise

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

(z- 1 - O.s&•1 4 )(z- 1 - O.Be-1•14 )


H,(z) = (1- 0.8e1•/<z 1)(1-0.Se-1•/<z 1) 0 ••(•)

5.54. (a) To be rational, X(z) must be of the form

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

X(z) = (1- ~&•i•z-1) (1- ~ei3 •l•z- 1 ) ( 1- ~e-#l<z-1) ( 1- ~0-i••t•,- 1 )


Clearly, X(z) is rational.
(b) A sketch of the pole-zero plot for X(z) is shown below. Note that the ROC for X(z) is 1•1 > 0.
164

lm

4lh ordlf pole

(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

(1) :t(-1] + :t(O] + :[1] + :t[2] + :[3] + :[4] + :(5] 0 =


(2) -:t(-1] + :t(O]- :t(l] + :t(2]- :(3] + :t(4]- :t(S] = 0
{3) :t[2] = 0
(4) :t(-1] = -:t(S]
(5) =
:t[O] -:t[4]
(6) =
:t[1] -:t[3]
(7) :t[-1] 2 + z[0] 2 + :t[1] 2 + :[2] 2 + :[3] 2 + z[4]' + :t[5]' =28
(8) :t(-1] + :t(O] = 4
(9) :t[-1] + :t(O] + :t(1] = 6
(10) :t[5] = -3
:t[n] is easily obtained from solving the equations in the following order: (3},(10),(4),(8},(5),(9), and (6).

3
JC(nj
2
3 4 5
-1 0 1 2 n

-2
-3

5.56. (a) The LTI system 5 2 is characterized as a Jowpass filter.


The •·transform of h 1 [n] is found below.

· y[n]- y[n- 1] + ~y[n- 2] =


4
:t(n]
1
Y(z)- Y(z)z- 1 + 4Y(z)z- 2 = X(z)

Y(z) (1- z- + ~z- 2 )


1
= X(z)

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}

11[n] + a 1y[n- 1] + a 2 v[n- 2] = Poz(n]


Y(z) + a 1 Y(z}z-• + a2Y(z}z-2 = PoX(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

an appropriate value of Po must be found. Consider the case when z = 1. Then,

=
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

Taking the inverse :<-transform yields

2] - ~o[n- 1] + o[n]- ~o[n + 1] + !o[n + 2]


33
hs[n] = !o[n-
4 4 16 4 4
5.57. (a)
1 . 1 .
:t[n] = s[n] cos won= s[n]e-'won + s[n]e-Jwon
2 2
X(eiw) = !s(eJ!w-wol) + !s(ei<w..,.ol)
2 2

y[n] = !s[n]e-'(won-~) + !s[n]e-i(w.n-~)


2 2
= s[n] cos(won- t/>o)

(b) This time,

Y(eiw) = H(e-'w)X(e'w) = !.-J-•e-Jw••S(ei<w-wo)) + !ei~e-'w"'S(ei<w+wol)


2 2

y[n] = •rn _ n,]. G·rnJei<w··-~) + i·rnJ·-j(w,n-~))


= o[n - n,] • s[n] cos(won - t/>o)
= s[n - n•] cos(won - won• - t/>o)
Therefore, if ¢ 1 = t/>o + won 4 then
y[n] = s[n- n,]cos(won- ~)

for narrowband s[n].


(c)
d . d
Tgr = - dw arg[H(e-'w)J =- dw[-t/>o- wn,] = n4

Tph = _.!.arg[H(e-'w)] = _.!_[-1/>o -wn,] = t/>o- n•


w w w
y[n] = s[n- Tgr(wo}]cos[wo(n- rph(wo)))
(d) The effect would he the same as the following:
(i) Bandlimit interpolate the composite signal to aC-T signal with some rate T.
(ii) Delay the envelope by T · Tgr, and delay the c:.arrier by T · Tph.
(iii) Sample to a D-T signal at rate T
5.58. (a)

. m. = 0 ~ c;..(m] =r .,(m] ~ r .,(z) = +,..(z)

.;•• [m] = y[n] • y[-n] = :t[n] • :t[-n] • h[n] • h[-n]


168

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

n(1- d.z- )(1- d••>


t=l
1

(b) To "whiten" the signaly(n] we need a system:

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]:

h,(n]• h2(n] = h;[n]p[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

arg[G(ei")] = -tan- 1 ( 1 ~:~w)

= 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)

hi[n] + h,(n] ++ H1 (z) + H 2 (z)

xi[n] = (21)n u[n] +-+ 1iz-' = Xl(z)


1
_
171

z,[n] = 2 -2 (1)n u(n]+---+ 21z 1 = x,(z) 1- 0


Both of these are minimum phase, with a zero at z = 0 and a pole at z = ~.

X1(z) + X,(z) = 1- 31z 1


0
This is minimum phase, with the same pole and zero as X1(z) and X,(z).

(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

5.63. (a) Maximum phase systems are of the form


Jl
II<z- c,)
H (z) = =:'_:;:::1' - - - -

...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)

h,....(n] = 5[n- M] • hm;.[-n] = hm;.[-n + M]


5.64. (a) We desire IH(z)He(z)l = 1, where H.(z) is stable and causal and H(z) is not minimum phase.
So,
IH.,(z)Hmin(z)Hc(z)l = 1
Since jH.,(z)l = 1, we want
IHmu.(z)He(z)l =1
This means we have
1
He(z) = ll . ( )
''"" z
which will be stable and causal since all the zeros of Hm;n(z), which become the poles of H.(z),
are inside the unit circle.
173

(b) Since
1
H.(z) = H msn. (Z )
We have
G(z) = H•• (z)
(c)

H(z) = (1- o.wo·•• .-')(1 - o.Se-;o....-•)(1 - 1.2~· 1 • . - 1)(1 - 1.2e-i0 · 1• . - 1)

H,.,n(z) = (1.44)(1- O.Sei".>• z- 1 )(1 - 0.&-iO·•• z- 1 )(1 - (5/6)~.?. z- 1 )(1 - (S/6)e-JO.?• z- 1)


1
Hc(z) = (1.44)(1- O.Sei0-3•z-1 )(1- O.Se-iO.l•z-1 )(1- (S/6)ei0· 7 •z 1)(1- (S/6)e i 0 ·1 • z-1)
(z- 1 - (5/6)e-i0 ?o)(z-•- (S/6)ei0·?o)
G(z) = H •• (z) = (1- (5/6)ei0.1•z-1)(1- (S/6)e ,o.?oz 1)

Ae

41h order pole

Ae

C#'l order zero

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)

H(z) = Q(z)z- 1 - z0Q(z)

h(n] = q(n- 1]- z;q(n]


Hmin(z) = Q(z)- z•Q(z)z- 1

h...in[n] = q(nJ- ••q[n- 1]


(c)
n n

€ = L lh...u.[m]l 2 - L lh[m]l'
"'=0 ~
n

= L (lq[m]l 2 - ••q[m- ljq*[mJ- z;q*[m -ljq[mJ + l••l 2 lq[m- 1]1')


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)

€ = (1 -l••l')lq[n]( 2 ~ 0 Vn since I•• I< 1


Then
n n

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 ....

applies to the system.


5.68. (a)

g[n] = z[n]• h[n]


r[n] = g[-n] • h[n]
s[n] = r[-n] = g[n] • h[-n] = z[n]• (h[n]• h[-n])

hJ[n] = h[n]• h[ -n]


H,(ei"') = H(e'"')H"(ei"') = IH(ei"')l2
Since H 1 (ei"') is real, it is zero phase.
(b)

g[n] = z[n]• h[n]


r[n] = :r[-n]• h[n]

!f[n] = g[n] + r[-n] = :r[n]• h[n] + :r[n] • h[-n] = z[n] • (h[n] + h[-n])

h2[n] = h[n] + h[-n]


H 2 (ei"') = H(e'"') + H"(ei"')
= 21U{H(ei"')}
H2(ei"') is real, so it is also zero phase.

(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

5.69. False. Consider

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.

z,, magnitude = 0.5, angle = 153 degrees


zi' magnitude = 0.5, angle = 207 degrees
1/z., magnitude = 2, angle = 207 degrees
1/zi, magnitude = 2, angle = 153 degrees

(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

(o.5 ..,:1e) (12 samples)= 6 ms


(c) The zero locations used to create the foDowing plot were estinlated from the figure using a ruler
"!'d a protractor.

Estimate of Continuous-Time Magnitude Response


0

m -20 Zero Locations (0):


:!:!.
..,.,:::> -40 (0.0444)lr/T
( 0.128)1</T

.
'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,

h(n] = { g{n/2], n 0, ±2, ±4, ...=


0, otherunse
177

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

H( ) - liM. z-1- d• liM. (z-1- e;)(z-1- ••) I• IE II.


z - k=l
1 - d•z 1 t=l (1- ••• 1)(1 - e"•z 1)'

where II. is the ROC which includes the unit circle. Correspondingly, the associated inverse system is
1
H;(z) = H(z)

= liJI. 1- d.z-1 liM. (1- ••z- 1)(1- e;z-1)


bl
z-1- d• 1:=1
(z-1- ••")(z-1 ••)

= IT
>=I
z- 1 (z- d•)
z I - d•
IT
>=I
2
z- (z- ••H•- et)
(z-1- e;)(z I - ••)

= liM. z- d, liM, (z- e,)(z- et)


i:=l
1- d•z A:=l (1- e0z)(1- •••)
1
= H(1/z), lzl E II.
which in the time domain is
h;[n] = h[ -n]
5.74. We can model g[n] as

g[n] = z[n] + ao[n - no]

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

y[n] = (z[n] + ao[n- no]). h,,,[n]


= ah,,1[n- no]
= a(-1)(n-no)sin~(n-no)
.-(n- no)
y[n] looks similar to the picture of h,,1 [n] above except that it is scaled by a and shifted to no. Thus,

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,

H(z) = L"" h[nV"


n=O
N-1
= L h[N- 1 - n]z-n
n=O

Substituting m =N - i - n into the summation gives


0
H(z) = L h(m]zm-N+l
m=N-l
180
N-1
= L h[m]z"'z-N+ 1
-=<>
N-1
= z-N+l L h[m]z"'
.....0
= z-N+l H(z- 1 )

Thus, for such a filter,


H(l/z) = zN- 1H(z)

Ji zo is a zero of H(z), then H(zo) = 0, and

H(l/zo) = zf- 1H(zo) = o


Consequently, even-symmetric linear phase Fm filters have zeros that are reciprocal images.
181

Solutions - Chapter 6

Structures for Discrete-Time Systems


183
6.1. We proceed by obtaining the transfer functions for each of the networks. For network 1,

Y(z) = 2rcos8z- 1Y(z)- r 2 z- 2Y(z) + X(z)


or
H,(z) = Y(z) = 1
X(z) 1-2rcos8z '+r•z 2

For network 2, define W(z) as in the figure below:

W(z)

z[n]

-rsin8

rsin8
rcos8

y[n]
.-•

then
W(z) = X(z)- rsin8z- 1Y(z) + rcos8z- 1 W(z)
and

Eliminate W (z) to get


Y(z) rsin8z-•
H.(z) = X(z) = 1- 2rcos8z- 1 +r 2 z 2

Hence the two networks have the same poles.

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:

Y(z) =2X(z) + (~X(z)- ~Y(z) + ~Y(z}z- 1 }z-'-


184

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)

w[n] = z[n] + 3w[n- 1] + w(n- 2]

y[n] =w[n] + 11[n- 1] + 2y[n- 2]


(b)

W(z) = X(z) + 3z-1W(z) + z- 2W(z)


Y(z) = W(z) + z- 1Y(z) + 2z- 2Y(z)
So
Y(z)
X(z) = H(z)
1
= (1- z-1 - 2z-2)(1- 3z-1 - z-2)
1
= 1- u- 1 + 7z • + 2z-•.
(c) Adds and multiplies are circled above: 4 real adds and 2 real multiplies per output point.
(d) It is not possible to rednce the number of storage registers. Note that implementing H(z) above
in the canonical direct form n (minimum storage registers) also requires 4 registers.

6.6. The impulse responses of each system are shown below.


185

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
.

Therefore the direct form II is given by:

-1/4

x[n] ...___,.- - - - . •-~-::::-~---· y[n]


1 1/4

6.8. By looking at the graph, we get:

y[n] = 2y[n - 2] + 3x[n - 1] + x[n - 2].

6.9. The signal flow graph for the system is:

w1[n] w,(n]
x(n] y(n]

z-1
-1 2 w,(n]
w3(n]
z-1
4
186

(a) First we need to determine the transfer function. We have

w.(nj = :[nj - w3[nj + 4w4 [n- 1]


"'>[nj = IDt [nj
"'>[nj = ID2[n-1j
w4 [nj = 2w3[nj
y[nj = 102[nj + :[n- 1] + w,[nj.

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:

y[nj + y[n- 1]- Sy[n- 2] = :t[n] + 3:t[n- 1] + :[n - 2] - S:[n - 3].


6.10. (a)
1
w[nj = 2
y[nj + :t[nj
1
v[nj = 2y[n] + 2:t[nj + w[n- 1]
y[nJ = v[n- 1] + :t(n].

(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 )

We thus get the following cascade form:

-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 .

We thus get the following direct from ll ftow graph :


:[n] y[n]
rz-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

We thus have the following relationships:

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:

H(z) = Y(z) = -2 + 6z- 1 + 2z-2 •


X(z) 1- Sz 1

Taking the inverse Z -transform, we get the following difference equation:

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:

H(z) = (1- !z-1)( 1- 2z-1 + 3z-• ).


2 1- !z-2

Tbis confirms tbat both graphs have the same system function H(z).
6.17.

H(z) =1- ~z- 1 + ~z-2 + z-3 .


(a) Direct form implementation of this system:

(b) l!ansposed direct form implementation of the system:

6.18. The flow graph is just a cascade of two tr&DSposed direct form U structures, the system function is
thus given by:

Which can be rewritten as:


191

(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"

U a = -2 , the overall system function is:


1 - 2-z- 1
H(z) = 1 3 3 2.
1 + iz 1- iz

And finally if a = 0, the overall system function is:


1
H(z) = (1 + 2z- )(1 - ~z- )
1

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

6.20. The transfer function can he rewritten as:


{1 + 2z- 1 + !z-2 )
H(z)- •
- (1 + ~z-2)(1- !z 1 + z-2)

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.

h[n] =e'w. "u[n] ,.._.. H(z) = 1- eJ.,.z-1


0
l
. = --. Y(z)
X(z)

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:

y,[n] = :r[n] + coswoy.[n- 1]- sinWQy;[n -1]


y;[n] = sinWQy,[n- 1] + COSWQy;[n- 1].

: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

with the three types of 1st-order systems taken in various orders.


6.23. Causal LTI system with system function:

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

-1/5 1/4 z-1

-5/24 z-1

-1/12

(ii) Direct form IT.

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

-1/4 -1/5 1/2 z-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

(v) Transposed direct form II


We take the direct form II derived in part (ii) and reverse the arrows as well as exchange the
input and output. Then redrawing the flow graph, we get:

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) w1 [n) = z[n) + w,[n- 1)


1 1
(2) w,[n) = :(ll[n) + ws[n- 1)- sz[n]
5 1
(3) V13[n) = - 24y[n)- 12 11[n- 1]
(4} y[n) = w 1 [n].

Combining the above equations, we get:


1 5 . 1
y[n]- 4y[n -1] + 24y[n- 2] + 12y[n- 3] = z[n)- s1 * -1).
Taking the Z-transform of this equation and combining terms, we get the following transfer func-
tion:
1- lz-1
H(z)- •
- 1- lz-l + .l..z-2 + .l..z-3
4 24 12

which is equal to the initial transfer function.


6.24. (a}

1
H(z) = 1- az 1

y[n] z[n]

• •

y[n] = z[n) + cy[n- 1)


1
HT(z} = 1 -cz 1
= H(z}
(b)
1 -1
H(z) = 1 + iz
1- !• 1

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)

H(z) =" + bz- 1


+ cz- 2

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

6.26. (a) We can rearrange H(z) this way:


(1 + .-1)2
H (z) - ---'-:,...;...~'--:;
- 1- lz-1 + z-2
2
200

:z:[n] u[n] v[n] w[n] 0.2


y[n]

z-1 z-1 z-1 z-1


2 1/2 -1 2 2 2

z-1 z-1 z-1 z-1


-1 -1/2 -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)

y[n] = z[n] + abw[n] + bw[n- 1] + aby[n]


w[n] = -y[n].
Eliminate w[n]:

y[n] = z[n]- ablf[n]- by[n - 1] + aby[n]


y[n] = :[n]- by[n- 1]
So:
1
H(z) = 1 + 6z-1.

(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]

(d) (i) (c) has the most storage: 9 vs. 7.


(ii) (a) has the most arithmetic: 7 adds + 7 multiplies per sample, vs. 2 multiplies+ 2 adds per
sample.
6.30.

(a)
203

...·-r-----------~-==-~'=-·~-----------.

- ....
•• .
... . ..
·-r-----------~-=-='~-=·-~-----------,

. .
...
• •
• •
(b)

(c)

H(~w) = ..!_ 0 - ;7w [sin((1Sw)/2) _!e-Ns sin((1Sw)/2) _! eHr sin((1Sw)/2) ] .


15 sin(w/2) 2 sin((w- ~;)/2) 2 sin((w + (2.. )/15)/2)

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, .

1 r~, (1- .-jlSw) !~-- ..;"" (1- .-jlSw)


= 15 eif - e-if - ei""-s2J•'u' e-r·-fh/l.S) -
204

!ei""+<a;na> (1- e-ilW)]


eiw+C:I;/15) e-;•+!2;'UJ

= _.!._ [·-;...7(~"'¥- .-;...¥) -


15 2j sin 'f
i·-;...7.-;fr(~"'¥- .-;...¥)
2j sin ( w-(2;/1$))

!e-jw7 ~fr (~¥w _ .-;¥"')]


2j sin ( w+(2; /15))

= e-;w7 [sin(15w/2) _ !e-;fr sin{15w/2) _


15 sin(w /2) sin ( w-(2; /IS))

!~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

- + cos~- cos (:!!c + ~) z-


15 1
H(z) = 1-15z- - 1[
1-z-1
"
1-2cos 2•z-l+z 2
15 "
]

15

1/15
:tnJ y"fin]
z-15 z-1

-1

") /3

-1

where " = cos(2?rno/15), /3 = - cos(2lr(Ro + 1)/15), and 1 = 2 cos(2" /15).


6.31. (a)

G u[n]
<TnJ 1+r
!I[n]
-r r -1
1-r
,::-r wn] .-1

u[n] = Gz[n] + w[n - 1]


w[n] = -ru[n]- (1- r)y[n- 1]
y[n] = (1 + r)u[n] - ry[n- 1].

(b)

U(z) = GX(z) + z- 1W(z)


W(z) = -ru(z)- (1- r)z- 1 Y(z)
Y(z) = (1 + r)U(z)- rz- 1 Y(z).
206

Solve for U(z) in terms of X(z) and Y(z):

U(z) = GX(z)- (1- r)z- 2 Y(z)


1 +rz 1
Then
Y(z) = (1 + r) {GX(z)- (1- r)z-•Y(z)}- rz-'Y(z)
I+ rz 1
Y(z)(1 + rz- 1 ) = G(1 + r)X(z)- (1- r 2 )z- 2 Y(z)- rz-'Y(z)- r 2 z-2 Y(z)
Y(z)(1 + 2rz-• + z- 2 ) = G(1 + r)X(z)
H ( ) G(1 +r)
'z = 1+2rz '+z-•·
From the quadratic formula, the poles are at (-r + j~)- 1 and (-r- j~)- 1 • The
magnitude of ea.ch pole is 1. The angles are

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)

!lt[n] = (1 + r):z: 1 [n] + r:z:,[n]


112[n] = -r:z: 1 [n] + (1 - r):z:,[n].

(b)

111 [n] = (1 + a):z: 1 [n] + d:t,[n] (a= r =d)


112[n] = (1 + cd):z: 2 [n] + ab:t.[n] (c = d = -1).

(c)

111[n] = (1 + e):z:,[n] + e:z:2[n] (e = r)


112[n] = ej2:,[n] + (1 + ef):z:,[n] (! = -1).

(d) B and C preferred over A:


(i) coefficient quantization. If r is small, 1 + r may not be precisely representable even in floating
point. Also, network A has 4 multipliers that must be qnanti:ted, while B and C have only !.
(ii) computational complexity. Networks Band C require fewer multiplications per output sampl~
6.33.
.-•- 0.54
H(z) = 1-0.54z ,.

(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

(d) Yes, since there is only one "0.54" to quantize.


(e)

H(z) = u= ~~1 ) u= b~!.)


1

..
1

Cascading two sections like (c) gives


z-1 w[n] .-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]

I :·I . !·-: I· .I·-· 1/[n]



208

(f) Yes, same reason as part (d).

6.34. (a) We have:

w 1 (n] w,[n]
2 v(n]

:(n]

First, we lind the system function, we have:

(1) w1 (n] = :[n] + w2[n- 1]


(2) w,[n] = :(n] + w3[n- 1]
(3) w3 [n] = 2w,[n] + w4 [n- 1]
(4) v[n] = w3[n]
(5) w4 [n] = -v[n]- w,[n]

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.

(b) Now we have:

w,(n] ws[n]
2

:(n]

w,[n]

Let's lind the transfer function, we have:


209

(1) w1(n] = z(n] + w,{n- 1]


(2) w2[n] = z(n] + w3{n- 1]
{3) w3(n] = 2w1{n]
(4) w4 {n] = -w3{n- 1]- y(n]
(5) ws(n] = w3(n] + w4 {n- 1]
(6) y(n] = w.{n]

Taking the Z -transform of the above equations and combining terms, we get:

The system function is thus given by:

H(z) = Y(z) = 2(1 + z- 1)(1- z- 1 ).


X(z) 1- 2z-2

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

(a) Direct form 1:

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- ~z- 1 )Y(z) = (-~ + z- 1)X(z)

Inverse Z -transfoiming, we get:


1 I
y(n]- -y(n- 1] = --:r(n] + :z:(n- 1]
3 3
210

1
y[n] = :i(y[n- 1]- :z:[n]) + :z:[n- 1]

Which can be implemented with the foUowing Bow diagram:

.-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

:z:[n]-· -:_]-:_:-f_~3--·.'___.. _-. . ~:


-·] :J·~~. . . . . . . --·y[n]

-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

(1) w 1 (n] = 2:r(n] + w,(n]


(2) 1t12(n] = :r(n] + w 1 (n -1]
1
(3) w,(n] = - y(n] + 2y(n- 1]
2
(4) y(n] = w,(n] + y(n- 1]

Taking the Z -transform of the above equations, substituting and rearranging terms, we get:

(1- !,-t- 2z- 2 )Y(z)


2
=. (2z- 1 + 1)X(z).
Finally, inverse Z-transforming, we get the following difference equation:

1
y(n]- 2y(n- 1]- 2y[n- 2] = :r(n] + 2z(n - 1].

(c) From part (b), the system function is given by:

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)

y(2] = :r(2] + 2z(1] + !Y(1] + 2y(OJ


y(Oj = :r(Oj = 1
= =
y(1] :r[1] + 2z(O] + fy[O] ! + 2 + ! =3
Therefore,
1 3 19
y[2] = - + 1 + - + 2 = -.
4 2 4
6.37. (a)
212

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)

z-1 [<I- ,-NJ ii[kJIN ]


1- .Z:A;z- 1
= B[kJ z-1[I- ,-NJ. z- 1 [
N
1
l-z• .- 1
]

!
213

= iiJ:l (o(n]- o(n- N]) • (z:u(n])

= iiJ:l (z:u(n]- z:-N u(n - N)]

= iiJ:l z:{u(n]- u(n- N)}.


So
. h(n] = ( ~ ~ B(k].,;~.,.) (u(n]- u(n- N]).
(d) Note that, since (1- z;;.N) = 0,

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.

H(z) = (1 - z-N) [H[O]!N +


1- Z l
t
~1
1

1
H[k]/N + H(N/2]/N +
Ztz-l 1 ZN/ 2 z l
~
.£..-
~~~
H[l]/N ]
1- ZtZ l

= (1 _ z-N) [H[O]/N + H[N/2]/N +


1- z- 1+
t:- 1 z 1
1
H[k]!N +
.L- 1 - z.z- 1
t:- H[N- p]/N]
1

L., 1 - ZN-.,z- 1
A:=l p=l

z-N) [H[O]/N + H[N/2]/N + t:-


1
= (1 - ( H[k]/N + H[N- k]IN)]
} - Z+ l 1 Z l L., 1- ZtZ l 1- Z-A:Z l
•=1
= (1 - z-N) [H[O]/N + H[N/2]/N +
1- z- 1 1+ z 1

..!_
N
t:-
L.,
1
H[k](1 - Z-tz- 1 ) + H[N - k](1 - ztz- 1 )]
(1- Ztz- 1 )(1- Z-tZ 1)
~1

= (i _ z-N) [fi[O]/N + H[N/2]/N +


1- z- 1 1 + z-1

t:- 2IH[k] . cos(li[k])- z- cos(ii[k]- 2rk/N)]


~
1

N 2dz-l+z-2
l-2cos IV
1

~1

And since H[OJ =H(1), H[N/2] = H(-1),


_ ( _ -N) [H(1)/N
H( z
) -1z 1
1 -z
+ H(-1)/N
1 +z 1
+
214

t.' 2IH(e;~•l cos[9(2d:/N)]- .-• cos[9(2>roi:/N)- 21rk!N]]


{;:. N · 1- 2cos(21rk/N)z 1 + z- 2 •

H H[14] = O,then H[16 - 14] = k[2] = o also.

k[OJ/16

[n] y[n]

-z-16 .-•
lik[IJI
"
.-•
{3 Q

.-•
'Y

where u = cos(Q(1]), o = -cos(Q(1]- (2•"/16)), {3 = 2cos(2.. /16), and 'Y = -1.


6.38. (a)

xTnJ
h(O] h(1] h(2] h[10]

M:!

M (N + 1) multiplies per output sample; M N adds per output sample.


215
(b)

'f"n]
z-1

-
M:1 M:1 M:1

h[Oj h[1] h[10]

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.

6.39. Since each section is 3.4cm long, it takes

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

waves add at a boundary. Let


A. -A•
Otn= An+A~r:
(from A• into A.); then a•• = -a•• and we get:

y(n]

x(n] •
,...., = 1

6.40. (a) For rounding:

Ps(e)

1/A

-A/2 A/2

(b) For truncation:

Ps(•)

1/A

-A
218

m, = A .!..[ ede =.!.. e']"


-6 A 2 _6
=~
2

~=.!..[ e•de_A• = ~]o _A• =A•.


' A. -6 4 3A _6 4 12

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

y[n] = L h1 [k]:z: 1 [n- k] + L h2[k]:z:2[n- 1.:]

= !lt[n] + 112[n].

The correlation between y,[n] and 112[n] is

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),

or using the frequency domain formula,

..-} = 3u'-l-f
27rj
1 . _I_ dz
1- az- 1- az z1

= 3u'-l-f
21rj
dz
(z- a){l- az)
3u'

For network (a),


H(z) = bo + b,z-1
} - ClZ-l

h{n] = bo.S(n] + (bo + b, )a"u(n]


a
• Time domain calculation:
221

= 2<T'+<>'(b'+<abo+b,Y).
o 1-a2
• Frequency domain calculation:
1
~>'(n] = -21fJ /H(z)H(z- 1) dz
z
ft

= L (residues of H(z)~(z-') inside unit circle).

H(z)H(z- 1 ) (bo + b,z- 1 )(bo + b,z) z


z = (z:...a)(1-az);
(boz + b,)(bo + b,z)
= z(z- a)(I- az)

-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') .

~a+ I?, a+ b,bo + b,boa2 - b,bo + b,boa 2


= a(1 - a 2 )
~ + 1?, + 2bob,a
= 1 -a 2
= b' +(abo+ b,)'
o 1- a2
6.43. (a)
I I
y(n] = y(n- I]+
4 2, n ~ 0

1 n (I)' II-U}n+l
y(n] = 2L 4 = 2 !
i=O •

For large n, y[n] = (I/2)/(3/4) = 2/3.


(b) Working from the diHerence equation and qoa.ntizating after multiplication, it is easy to see that,
= =
in the quantized case, y(OJ I/2 and y(n] 5/8 for n ~ I. In the unquantized case, the output
monotonically approaches 2/3.

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

(c) The system diagram is direct form II:

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

z "' 0.52 + 0.84j and z "' 0.52 - 0.84j.

Note that
[z[ 2 "' 0.976 < 1.
The poles are inside the unit circle, therefore the system function is stable.
223

(b) If the coefficients are rounded to the nearest tenth, we have

1.04 -+ 1.0 and 0.98 -+ 1.0.

Now the poles are at


z= 1 -;-13 andz= 1 -;-13_
2 2
Note that now,
[z[• = 1.
The poles are on the unit circle, therefore the system is not stable.

6.45. The flow graphs for networks 1 and 2 respectively are:

· ·z-·D-a• · n.-·
z[n] ----~--w-i,ll-[n_J_ _ _

a"'-J
w..[n_J_ __.._ _ _ _ y[n]

·
z[n] .-• z-•

~
• I

1 1·: r 1·: r 1·: r I·: •


y[n]

(a) For Network 1, we have:

w1 [n] = z[n]- a8 z[n- 8]


w,[n] = ay[n- 1] + w,[n]
y[n] = w,[n]

Taking the Z -transform of the above equations and combining terms, we get:

Y(z}(1- az- 1 ) = (1- a•z-1 )X(z)

That is:
H(z) = 1 -a • z -· .
1-az- 1
For Network 2, we have:

y[n) =z[n) + az[n- 1) + a z[n- 2) + ... + a :r[n -7).


2 7
224

Taking the Z -transform, we get:

Y(z) = (1 + az- 1 + aZ z- 2 + ... + a7z-7)X(z).


So:

(b) Network 1:

z{n]

Network 2:

z{n] .-• .-• z-• z-1 z-• .-• z-•

1
., a'
••
a• a• a• a1
y{n]

(c) The nodes are circled on the fignres in part (b).


(d) In order to avoid overflow in the system, each node in the network must he constrained to have
a magnitude less than 1. That is if w.{n] denotes the value of the ktb node variable and h.{n]
denotes the impulse response from the input z{n] to the node nriahle w•{n] , a sullicient condition
for Jw.{n]J < 1 is
1
z...,. < ~-co Jh>{mJr
In this problem, we need to make sure overflow does not occur in each node, i.e. we need to take
the tighter bound on z,.... For network 1, the impulse response from wo{n] to y{n] is anu{n],
therefore the condition to avoid overllow from that node to the output is

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:

w1 {n] < zm.. (l- a 8 ) < 1-JaJ.


225

We thus conclude that to avoid overllow in network 1, 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

Filter Design Techniques


229

7.1. Using the partial fraction technique, we see

H.(s) = s+a = 0.5 _ + 0.5 _


(s + a) 2 + b2 s +a+ ]b s +a- Jb
Now we can use the Laplace transform pair
1
e-01 u(t) +---+ - -
s+a
to get

h,(t) = ~ ( .-(o+jb)l + .-(o-jb)l) u(t).


(a) Therefore,

h 1 [n]

H,(z) = jzi > .-•T


1
(b) Since

'
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

We can finally solve for H,(z)


H,(z) = S 2 (z)(1- z- 1 )
1- .-• 1- .-•
e-(•+i•)T z 1 + A •:z -:---==r==.,-"'"
= A, + A, -:-1 -:'::T.~=--:- 1 - e-<•-i6)T -1'
izl > .-•T
2
where A1 and A2 are as given above.

!
230

(c)

s,[n) = t h,[k) = ~ t( .-(o+jb)>T + .-(o-jb)•T)

=
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.

Taking the inverse z-transform of H2(z)


h2[n] = A16[n] + A2 [•-(o+ib)T•u[n)- .-(•+ib)T(n-llu[n- 11]
+ A2 [e-<•-ib)T•u[n] _ .-(o-;b)T(n-llu[n _ 1]
1

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.

7.3. We are given the digital filter constraints

1-6, :S IH(e'w)l :S I+ 6,, 0 :S lwl :S Wp


IH(e'w)l :S 0,, w, :S lwl :S"
and the analog filter constraints

1- J, :S IHc(iO)I :S 1, o s 101 so.


IHc(iOJI :S J., o, s 101
(a} If we divide the digital frequency specifications by (1 + 6,) we get
1-6,
1- cl, =
1 + 6,
20,
cl, =
1 +6,
0.
6. =
1 + 6,

(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

&, = 1- 0.89125 = 0.10875


6. = 0.17783

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

-1- +-+ ..,...._::.;.,--,


Td
l+a 1-e oT.,z-1
Therefore,

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-<>·•)

= (1+;_..) (s+·~)- (1+!-•• ) (.+·~)


Since the bilinear transform does not introduce any ambiguity, the representation is unique.
233

7.5. (a) We must use the minimum specifications!

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"')

-11 -0.62511 -0.311 0 0.311 0.62511 11 (I)

h.i[n] = sin(.625.-(n- 45))- sin(.3.-(n- 45))


.-(n- 45)

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

7.10. Using the bilinear transform frequency mapping eqnation,

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

7.14. Using the bilinear transform frequeocy mapping equation,

0, = f tan ( w, ~2 rk) , k an integer

= 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

IH.(jO)I < 0.04, 101 :S 2.-(51. 7126)


0.995 < IH.(jO)I < 1.005, 101 ~ 2.-(81.0935)
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 co;,vention can be obtained &om the above
specifications by scaling the magnitude response by nJos.
7.19. Using the relation w = llT,
w
T = n
.-j4
= 2lr(300)
= 417 J.IS
This choice ofT is unique. It is possible to find other values ofT that alias one of the given continuous-
time band edges to its corresponding discrete-time band edge. However, this is the only value ofT that
maps both band edges correctly.
7.20. True. The bilinear transform is a frequency mapping. The value of H(s) for a particular value of s
gets mapped to H ( e'~) at a particular value of w according to the mapping

• - .!._
- 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

7.22. (a) Applying the bilinear transform yields


H(z) =

1•1 > 1
which has the impulse response
h[n] Td (u[n] + u[n- 1J)
= 2"
238

(b) The difference equation is

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

-It 0 !tl2 It Ol -It 0 !tl2 lt 0

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

7.23. We start with IHc{iO)I,


240

10lt

-lt -0.11t 0.11t " (J)

Then, to get the overall system response we scale the frequency axis by T and bandlimit the result
according to the equation

IH..-,(jfl)l = { OIH' ,(&wT)I, 11"11 <;.


11"11 >;.

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-

Similarly, by taking the first derivative and evaluating at s = s0 _we get


r . d
~ [(s- sol' He(s)] = L:<r- k)At(s- ~o)<r-t-l) + ds [(s- sol'Ge(s)]
1=1
.
= (r- l)A1(s- ~o)•-• + (r- 2)A2 (s- ..,r-• + · 0· + Ar-l + 0 + dsd [(s- sol'Ge(•
A.,.-1 = ~ [(s- so)' Be(s)] J..,,,
This idea can be continued. By taking the (r - l:)-th derivative and evaluating at s = so we get
the the general form

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)

H we define Poly.(z) = nf.,


j"¢Jc
{I-e•;T•z- 1), we can note that all the roots of Poly.(z) are
inside the unit circle. Since the numerator of H(z) is a sum of A•Poly.(z) terms, we see
that there are no guarantees that the roots of the numerator polynomial are inside the unit
circle. In other words, the sum of minimum phase filters is not necessarily minimum phase.
By considering the specific example of
s+ 10
He(s) = (s + 1)(s + 2)'
and using T = 1, we can show that a minimum phase filter is transformed into a non-minimum
phase discrete time filter.
Bilinear 'Iransform: The bilinear transform maps a pole or zero at s = So to a pole or zero
(respectively) at Zo = - ..
,'+!••.
Thus,

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

H(e'w) =.~..,He (j (;. + 2;:))


The aliasing terms can destroy the all pass nature of the continuous-time filter.
243

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"')l .. =<~ = H,(jO)In=<~


Impulse lnvariance: Since impulse invariance may result in aliasing, we see that

H(ei") = H,(jO)
if and only if

or equivalently

which is generally not the case.


Bilinear Transform: Since, under the bilinear transformation, 0 = 0 maps tow= 0,
H(ei 0 ) = H,(jO)

for all H,(s).


(d) Answer: Only the bilinear transform design is guaranteed to create a bandstop filter from a
bandstop filter.
If H, (s) is a bandstop filter, the bilinear transform will preserve this because it just warps the
frequency axis; however aliasing (in the impulse invariance technique) can fill in the stop band.
(e) Answer: The property bolds under the bilinear transform, but not under impulse invariance.
Impulse Invariance: Impulse invariance may result in aliasing. Since the order of aliasing and
multiplication are not interchangeable, the desired identity does not hold. Consider Do, (s) =
=
H.,(s) .-•T/2.
Bilinear Transform: By the bilinear transform,

H(z) (2(1-z-'))
= H, T• 1 + z-1

- H" (:. G: :=:)) H" (:. G: :=:))


= H,(z)H,(z)

(f) Answer: The property holds for both impulse invariance and the bilinear transform.
Impulse Invariance:

H(ei"') = .fao H, (j (;. + ~: k))


= f
1:=-cc
He~(;(;.+ ~k)) + f H.,(;(;.+ ~k))
~=-CCI

= 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,

H, (e'w) = · f: H., (; (;. + 2;k))


A:=-oo
4

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,

H 1 (e'w) = He, (; :. tan(w/2))

Hz(e'w) = He, (; :. tan(w/2))


therefore,

7.26. (a) Since

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

-lt -w/2 0 7[ (1)

(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)

-lt -3lt!4 0 3lt!4 lt (1)


246
7.28. (a) We have

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)

=> w:P2 =wPl + 7r

(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.

7.29. (a) Substituting Z = ei' and z = &w we get,


ei' =
_.,..,
= ei<""'+•)
8-1f
w=-2-

(b)
247

H, (ei"')
ec ec
A

-It -1112 0 It (I)

(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=

where a is a nonuro integer and {3 is a real number.


(a) It is true for {3 > 0.
Proof:

• = {3 [1-
l+z
z-o]
0

S + $Z-a: = {3- {Jz-o


s-{3 = -Pz- sz-o 4
-

{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- 6, :s IH(e'~JI :S 1 + 6,, { '"'' :s ~} u { 34" < '"'' <or}


Note that the highpass region 3" f4 :S Jwl :S " is included because tan(w) is periodic with period

7.31. (a)

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

Therefore, if 1•1 < I

(.. +I)'+ o• < (.. -I)'+ o•


u < -u

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

0.95 $ IHc{iO)I S 1.05, 0 s 101 $ 1


IHc{iO)I S 0.01, v'3 $101
7.32. (a) By using Pa.rseval's theorem,

<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.

which is a rectangular window.


250

7.33. (a)

Hd(~w) =(I - 2u{w)J~(rf2-rw) for-" < w < 11"

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.

r,lnJ = { [k, n=O, ... ,~ -1


0, otherwise
••In] = rt(n- 1]
Using the known transform of a rectangular window we have

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

r 3 1n] = { .[b. n-O M-l


- , ... ,--y--
0, otherwise

= { .[b.
n-- 1, ... ,-2-
M-1
••In]
0, otherwise
In the frequency domain we have

w•. (eiw) = {2sin(w(M +1)/4) -iw(¥)


·~ VM sin(w /2) e

W •(ei"') = {2sin(w(M -1)/4) -jw(J¥+I)


- VM sin(wf2) e
Ws(ei"') = WR,(ei"')WR.(eiw)
= 2_ (sinlw(M + 1)/2]) (sinlw(M- 1)/2]) -iwM/ 2
M sin(w/2) sin(w/2) e
252

(b)

w[n] = [A+ Bcoo C:) + Ccoo (~n)] WR[n]


W(.;w) = {21rA6(w) +Blr [6 (w + !;) +6 (w- !;) ] +C+ (w + ;;) + 6 (w- ;;)]}
~ { sin(w(M + 1)/2))
21r sin(w/2)
.-jw/11/'}
where @ denotes periodic convolution.
(c) For the Hanning window A =0.5, B = -0.5, and. C =0.
Wftanning[n] = [o.5- 0.5coo c:)] w,[n]
Wftanning(.;w) = 0.5WR(.;w)- 0.25WR(.;w) ® [6 (w + !;) + 6(w- !;) l
= 0.5WR(.;w)- 0.25 [wR(.;<w+~l) + WR(e'(w-~lJ]

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.

Normalized Magnitude plot in dB

0 It

7.35. (a) The delay is ': = 24.


(b)

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

Therefore, in the time domain, hd[n) is (h1p 1 [n)) 2, or


2
= ('in:/4))

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,

sin• ("(n- N-1 l]


h[n) = •2
w[n)
.-•(n-N21)2

where
w[nJ = { 1, 0$ n$ N - 1
0, othei"W!Se
(d) The integral squared error <

can be reformulated, using Parseval's theorem, to


00

• = L Ja[n)- hd[n)J 2
-oo

Since

hd[n) _!!=1. < n < N-1


[I
an= { 0,
' 2 -
otherwise
- 2

-(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

h[n] =a6[n] + M[n- 1] + a6[n- 2]


Taking the DTFT of both sides gives
H(ei"') = a+ be-;..+ ... -;o..
= .-!w(cze!w + b + ...-;..)
= .-;"'(b + 2acosw)
A(ei"') = b+2acosw
257

=
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,

A( .,Jw) Jw=O = 1 +6p


A(eJw) Jw=•/3 = 1-6.
A(eJw) lw=•/2 = 6,
A(eJw) lw=• = -6,

Using A(&w) = b + 2acosw,


A(eJw) Jw=O = b+ 2a
A(e;w) lw=•/3 = b+a
A(eJw) Jw=•/2 = b
A(e;w) lw=• = b-2a
Solving these systems of equations for a and b gives
2
a = 5
2
b = 5
Thus, the optimal (in the minimax sense) causal 3-point lowpass filter with the desired passband and
stopband edge frequencies is
2 2 2
h[n] = 56[n] + 56[n - 1] + 56[n- 2]

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

If we let 8 = tan-• ( J':.) we see that


1 = cos(8)
J(l- u)2 + (nT) 2 1- (7

1 ( .• ..)
= 2(1- <7) e' + .-,
and thus the s-plane maps to the z-plane in the following manner

z = [2(1 1- u) (ei' + .-;•)] ei'


= 1 + 1 e'"211
2(1- <7) 2(1- <7)
=
2(1
1
<7)
+ 1
2(1- <7)
ei>--·ce.->
To find where the ;n axis of the s-plane maps, we let • =;n, i.e., u = 0 and find
1 1 _;2 ... "'(0T)
z =-
2
+ 2-~
259

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[n + ~- y[n] + cy[n] = Ax[nJ

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

lle(nT) = /.AT Ye (T)dT + 1/e(nT- T)


nT-T

"" [il. (nT)+ 'ile (nT- T)] ~ + 1/e(nT- T)


(b} Solving for i; < ( nT) in the difierential equation we get

il. (nT) = Az.(nT)- cy.(nT)


Substituting this into the answer &om part (a} yields
T
y.(nT) = (Az.(nT) - cy.(nT) + Aze(nT- T)- cy.(nT- T)] 2 + y,(nT- T)
(c) The difierence equation is

y[n] = (Az(nj- cy[n] + A.z[n- 1]- cy[n- 1]} T2 + y[n- 1]


y[nJ(1+c~) -y[n-1] (1-c~) =A~(z[n]+z[n-1]}
Therefore,
Y(z) [1 + c~]- Y(z)z- [1- c~] = A~X(z) (1 + z-
1 1
]

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

h[n] = l::. (1- e->aT•) (e-oT•f u[n]

(b) Since IH.(jn)l 2 =+.(in) and +(ei"') = H(ei"')H(e-;..} = IH(ei"')i', we see that since ¢[m] =
T.¢.(mT.), .

+(ei"') = J;,., •. (j (;: + 2;./c)).


Therefore, if +.(in) =:: 0 for 1n1 2! f;, then +(ei"') =::+.{if.;) and IH(ei"')i' =:: jH. (ii;) j' ·
(c) No. We could always cascade H(z) with an allpass filter. The new filter is di1ferent, but has the
same autocorrelation.
262

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

(b) Let Z = ei8 and z = ei"'. Then


z- 1 -a
z-1 = 1 az- 1
e-iw- Q
1- ae-Jw
e-;e- oe-i'e-iw = e-jw- Q

,-;"'(1 + ae-;8 ) = e-il + 0


e-;• +a
1 + ae i 6
e-i8 +a 1 + aei 1
= 1+ ae-i8 · 1 + aei 6
,-;• + 2a + a'ei'
= 1 + 2ocos8 + a 2
Using Euler's formula,

e-jw = cose- j sin9 + 2a + a 2 cos9 + ja2 sinO


1 + 2ocose + a 2
2o + (1 +a') cose + j[(a'- 1) sinO]
= 1+2ocos9+ci'
Noting that -w = tan-• [ Zt:/]'
-w = tan-• [ (a'- 1) sine ]
2o + (1 + a 2 ) cose
2
w = tan-• [ (1- a ) sine ]
2a + (1 + oZ) case
This relationship is plotted in the figure below for a = 0, ±0.5.

Warping of the frequency scale, LPF to LPF


(1)
263

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.

w _ tan-• [ (1 -a') sin(8,) )


9
- 2o + (1 + o2) cos(89 )
(c)

xjnl----r-------r---ytn]

xjn] ---+--~.---l----.l----+l y[n]


1 H(z)
H.,(Z) r
- o.p -
Looking at tbe 6ow graph for H(z) we see a feedback loop with no delay. This effectively makes
the current output, y[n], a function of itself. Hence, there is no computable difference equation.
(d) Ye.<, the 6ow graph manipulation would lead to a computable difference equation. The 6owgraph
of an FIR filter has a path without delays leading from input to output, but this does not present
any problems in terms of computation. Below is an example.

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

H(z) ·-· = H,9 (z


= H, 9 (Z)Iz=• t-.a- z--o)
1- - QZ

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

Using Euler's formula,

ei' = .,;--(1 +a')cosw- 2a+ j(l- a')sinw


l-2acosw+a'

Noting that 8 = w + tan-• [ ~H ],


8 ="'+tan _ 1 [ (1- a') sinw ]
(1 +a') cosw- 2a

(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.

7.46. (a) Since

y[n] = (2z[n]- h[n]• z[n]) • h[n]


= (2h[n]- h[n]• h[n]) • z[n]
the new transfer function is
g(n] =2h(n]- h[n]• h(n]
(i) It is FIR since the convolution of two finite length sequences results in a finite length sequence.
(ii) Note that the term h[n] • h[n] is symmetric since it is the convolution of two symmetric
sequences. Therefore, g(n] must he symmetric since it is the ciliference of two symmetric
sequences.
(b) The frequency response for G(&~) is
G(&w) = 2H(ei~) - H(.,;~)H(ei~)
265

G(ei"')

1~ '----------...1 -2S2 -f.1L----------


1~ 1 1~ ~ 0 \
H(ei"') H(ei"')
As shown above, if the passband of H(eiw) is the region [1- 6,1 +6,], then the passband of G(&w)
is in the region (I - ~, 1] which is a smaller band. However, the stop band gets bigger since it
maps to (-26,- 6i, 262- 5i].
Thus,

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

y(n] = (3z[n] - 2z(n] • h(n])• h(n] • h(n]


= (3h(n] • h(n] - 2h[n] • h(n] • h(n]) • z[n]
the new transfer function is

hsharp(n] = 3h(n] • h[n] - 2h[n] • h(n] • h[n]


and so

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

If 61 « 1 and 0:, « 1 then,

Maximum passband approximation error "" 0


Maximum stopband approximation error "" 0
266

(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

{0!:: lwl !:: wp} U{w, !:: lwl !:: 1f}


then we see we can have at most L + 3 alteruation frequencies.
If the transition hand has two of the local minim• or m•xim• of A.(ei"'), then only L- 3 can be in the
approximation bands. Even with all four endpoints of the approximation region as alternation points,
we can only have a maximum of L + 1 alteruation points. This does not satisfy the optimality condition
of the Alternation Theorem which requires ill least L + 2 alternation points. It follows tbal the transition
band cannot have more than two local minim• or maxim• of A.(ei"') either.
If the transition band only has one of the local minim• or m•xima of A.(ei"'), then the error will not
alternate between w7 and w, and they cannot both be alteruation frequencies. In this case, only L- 2
of the local minima or maxima of A. (ei"') are in the approximalion bands. If we add the maximum of
three band edges to the total count of alternation frequencies we get L + 1, which is again too low.
· Therefore, the transition band cannot have any local minima or m•xim• and must he monotonic.
7.48. (a) A.(ei"') has 7 alternations of the error. If the approximati'>D bands are of eqnaliength and the
weighting function is unity in both bands, why would the stopband have 1 extra al&ernation than
the passband? The answer is thal, lf it were an optimal filter, it would not. The optimal filter for
this set of specifications shonld have the ICmt number of alternations in each hand and therefore
requires an even number of alternations. Since the optimal approximation is unique, the one shown
in the figure cannot be optimal.
267

(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.,.(ei"') -1/2 = -[H.., (~<•-..)) -1/2]


H..,(e-i"') = -H.,. ( &<•-wl) + 1
h.,.[n] = -( -Wh.,.[-n] + 6[n]
This condition implies that

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,

H•(dw) ={ sin ~'f), lwl :S 0.2..


0, 0.4.. :S lwl :S "
To obtain equiripple behavior in H~fl), we need to weight the error so that the ripples grow
with H•(e'w). Then when we multiply by H0 (jfl) the ripples will be decreased to an equal size.
Therefore, we need
sin('f)
W(w)=
{
T"'
""' :S 0.2..
1, 0.4.- :S lwl :S .-
W(m)
11---

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.,

H•(dw) = { sin~~) IHr ~If) I' ""' < 0.2.,


0, O.b :S ""' :S "
This would take care of the distortion due to IH.(jfl)l but not of any phase distortion. The
weighting function will change in a similar manner.
269

7.50. (a) To avoid aliasing, we require

M $ "
"'•
So the maximum allowable decimation factor is
"
Mmax=-
"'•
(b)
V(e"")
11------..

0.9lt/100 n/100 (I)

Y(e"")
"
1/1001-------------------,

(c)

11-----.

00 0.9lt/100 (1)$1 (I)

"
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

Since we allow aliasing to occur in the transition bands, we have

SOw,, $ 1.55,.
w,, $ 0.031,.

(e) Using 6, = 0.01, 6, =0.001, Aw = 0.001.- we get


N = -10log10 {0.01 x 0.001)- 13 +
1
2.324(0.00l?r)
:::: 5069

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].

y[n] = 0.25z[n] + :r[n- 1] + 0.25:r[n- 2] (2 multiplies)


= :r(n- 1] + 0.25(:r[n] + :r(n- 2]) (1 multiply)
The procedure is similar for the other types of linear phase 6lters.
Thus, we need 2535 multiplies to compute each sample of the output.
(f) We have
-10log10 (0.01 x 0.001) -13
N, = 2.324(0.031 .. - 0.009.-) + I
271

"' 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

-1o log 10 (o.oo5 x o.001)- 13 +1


N, = 2.324(0.022.-)
251
"' -10log,.(0.005 X 0.001)- 13 +1
N• = 2.324(0.051<)
111
"'
Therefore, we have a total of 126 + 56 = 182 multiplies per output point.
(h) No. Since 6, « I we have ~ < 6, which means the stopband ripple is getting smaller. Thus, we
could actually increase the specifications.
(i) Performing a similar analysis on the other possibilities yields
M, M• Multiplies per output
50 2 182
25 . 4 156
20 5 172
10 10 291
5 20 557
4 25 693
2 50 1375
Thus, the choice M 1 = 25 and M 2 = 4 yields the minimum number of multiplications for this
example.
7.51. (a)

h,[n] = h[n] + .S,6[n- no]


H, (eiw) = H(eiw) + .S,e-;.,..
= A,(eiw)e-;"""' + .s,,-;w..
= (A.(eiw) + .S.] 0 -;wne
H,(eiw)

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)

wher~ c = ,/l+lh+lh~y't-lh+lh. Since 1- 61 $ Ae{ei'"') :S 1 + 01 in the passband and -62 :5


A,(e-'w) S 62 in the stopband, we have

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

where H 2 (z) is a minimum phase filter.


No. It will not work with a Type n linear phase filter. In this case no= M/2 is not an integer.
7.52. (a)

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

Since this last expression must equal


¥
L b(n] cosw(n-})

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

W(w) (Hd(eiw)- A(eiw)] = W(w) [ Hd(eiw)- ~ b(n]cosw ( n- D]


= W(w) [ Hd(eiw)- cos(w/2) ~ b(n] coswn]
= W(w) [H4(eiw)- tb[n]coswn]

where we have defined


M-1
L =
2
H.(eiW)
Hd(eiw) =
cos(w/2)
W(w) = W(w)cos(w/2)
We also see that
!pin
•fn)
{ma,cO} =min {max{·}}
wEF ~nJ wEF

(d) Type III filters:


M
H(eiw) = L: h\n]e-jwn
n=O
Jl/2-1 M
= L h(n]e-jwn + 0 + L h(n]e-jwn
n.=O n=M/2+1
M/2-1 M/2-1
= L h(n]e-jwn- L h(m]e-Jw(M-m)
....0 m=O
M/2-1
= 0 -jwM/2 L h(n] (e-jw(n-M/2) _ eiw(n-M/2))
n=O
M/2-1
= .-jwM/2 L (-2j)h(n] sinw(n- M/2)

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

.Y-1 .Y-1 .Y.-1


sinw 2: .:(nJcoswn = ~ 2: C!nJsinw(n + 11- ~ 2: .:(nJsinw(n -1)
=0 n=O n=O
275

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

where d[n] = 2jh((M + 1)/2- n] for n = 1, ... , (M + 1)/2. We C&I. find


2d[o]- dill
n=l
2
d[n] = d[n- 1]- d[n] 2<n<J!::!
2 - - 2

J{M;'J n -- M±l
-2-, 2

M-1
L "' 2
H.(~"')
fl.(~"') =
sinw/2
W(w) = W(w)sinw/2
F = F

7.53. (a) The fiow graph for A.(z) looks like

x{n] O.S(Z+Z- 1) O.S(z+Z-


1
) O.S(Z+Z-
1
} -
•• a, a2 ..._, ...
y(n)

(b) The filter length is 2L + 1. The causal version of the fiow graph looks like

x{n] 0.5(1+z..,) 0.5(1+Z-.2) 0.5(1+Z..z) -


•, a2 .,__, ...
••
.-· .-· .-· Y1n]

(c) The fiow graph for B,(z) 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

Do(z) = Z {ti0 {:o:[n)}} = X(z)


D 1 (z) = Z {ti 1 {:o:[n]}} = (z- z- 1 )X(z)
D 2 (z) = Z{ti2 {:o:[n]}}=(z-z- 1 ) 2 X(z)

Dt(z) = Z {Ll.•{:o:[n]}} = (z- z- 1 )• X(z)


279

(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"')

-lt ->t/6 0 ltl6

7.55. (a) Using DTFT properties,

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

W(ei"') = { ~fo,, 0$w$wp


w. :S w :S r
(b) Using DTIT properties,

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,

h 3[n] = h[n] • h(n]


B3(e'"') = B(ei"')B(ei"')
ln the passband, B 3(e'"') alternates about 1 + Of with a maximal error of 26~. ln the stopband,
B 3(ei"') alternates al2out ~/2 with a maximal error of ~/2. At first glance, it may appear that
H3(eJ"') is optimal. However, this is not the case. Counting alternations, we find that the original
filter H (ei"') has 8 alternations.
We know that since H(ei") is optimal, it must have at least L + 2 alternations. It is also possible
that H (eJ"') has L + 3 alternations, if it corresponds to the extraripple case. So L is either 5 or 6
for this filter. Consequently, the filter length of h[n], denoted as N, is either 11 or 13.

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)

/4(n] = h[n] - K 6[n]


B 4 (ei"') = B(ei"') - K
This filter is simply H(ei"') shifted down by K along the Ht(eJ"') axis. Consequently, this filter
satisfies the alternation theorem, and is optimal in the mjnimax sense.
281

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

grd (H(ei"')) = -! {arg (H(ei"')]}


= 12
> 0
282

7.57. (a) The desired tolerance scheme 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.

T!!Picai E(w) looks like :

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.

The inside of the unit circle of the Z-plane is

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,

cosS- j sinS+ 2a + a2 cosO+ ja2 sinS


= 1 + 2acos8 + a2
2a + (1 + a2) cosO+ j((a2 - 1) sinD]
= 1 + 2acos8 + a2

Noting that -w = tan- 1 [ ~~Jl,

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.

Warping of the frequency scale, LPF to LPF


(I)
285

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)

-2i sin((e,- "'•)/2]


= -2j sin((e. + "'•)/2]
sin((e,- w,)/2]
= sin((6, + "'•)/2]
(d) Using the equation for w found in part c, with e, = "/2,
w,=tan- 1 [1-
~ a2]

(i}
2
-1 [1- (-0.2679) ]
=
"'• tan 2( -0.2679}
-1 ( 0.9282 )
= tan -0.5358
= 2tr/3
(ii)

= tan-1 [I- (0)2]


"'• 2(0)
= tan- 1 (oo)

= 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]

(f) First, an equation for w is found in terms of 8 and a.

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

Warping of the frequency scale, LPF to HPF


(J)

For B, = "/2, this becomes

(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

••• • ••

-lt -7112 0 lt (I)

(iii)

_, rl- (0.4142) 2 "


"'• = tan 2(0.4142) 1
+
_, (0.8284)
= tan 0.8284 + ..
= 7r/4+tr
= 57f/4

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.

••• •••

-lt -31t/4 0 3lli4 lt (I)


289

Solutions - Chapter 8

The Discrete-Time Fourier Transform


291

8.1. We sample a. periodic continuous-time signal with a sampling rate:


n•. 1
F: = - = - = - H3 z
6
' 2.. T 10-
(a) The sampled signal is given by:
:o[n] = :o,(nT)
Expressed as a Discrete Fourier Series:

:o[n] = L' "•ei'f•n


1:-=-•
We note that, in accordance with the discussion of Section 8.1, the sampled signal is represented
by the summation of harmonically-related complex exponentials. The fundamental frequency of
this set of exponentials is 2.. /N, where N = 6.
Therefore, the sequence :t[n] is periodic with period 6.
(b) For any bandlimited continuous-time signal, the Nyquist Criterion may be stated from Eq. (4.14b)
as:
F,?. 2FN,
where F, is the sampling rate (Hz), and FN corresponds to the highest frequency component in
the signal (also Hz).
As evident by the finite Fourier series representation of :r,(t), this continuous-time signal is, indeed,
bandlimited with a maximum frequency of Fn = ,:.,
Hz.
Therfore, by sampling at a rate of F, =10
._, Hz, the Nyquist Criterion is violated, and aliasing
results.
(c) We use the analysis equation of Eq. (8.11):
N-1
X[k] =L x[n]e-i?,Hn
n=O

From part (a), x[n] is periodic with N = 6.


Substitution yields:

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

8.2. (a) Using the analysis equation of Eq. (8.11)


N-1
X[k] =L f[n]W~n

Since %[n] is also periodic with period 3N,


SN-1
X,[k] = L %(n]W3";J
n=O
N-1 2N-1 JN-1
= L f[n]W:;";J + L f[n]w;.-: + L f[n]W:;";J
n=O n=N n=2N

Performing a change of variables in tbe second and third summations of X3-[k],


N-1 N-1 N-1
X,[k] = L z[n]w;.-: + w;: L f[n + N]w;.-: + Wf~N L z[n + 2N]w;.-:
n=O n=O n=O

Since f[n] is periodic with period N, and W,tR = w~tln,


N-1
x,[kJ = (~+·-;•·(tl +•-;•·(>t') z::: z[nJw~tnJ
n=G

= (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:
~

i[n] = L z[n + rN], lal < 1


~-oc

(a) The Fourier transform of :r[n]:

X(eiw) = __L.,. z[n]e-j.m


~

=
..
L:a"e-;..,n
..... 1
= 1- ae-P.,' lal < 1

(b) The DFS of i[n]:


N-1
X[kj = L x[n]w~·
n=O
294

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)

:r:(n] = { 1, 0::; n::; ((N/2)- 1)


0, N/2 $ n $ (N- 1)
N-1
X[.i:J = L 0 $ k $ (N - 1)
..... :r:[n]Wg>,
(N/2)-1
=
....I: W'"
N

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

8.6. Consider the finite-length sequence


eiwon 0::; n::; (N- 1)
z[n] ={ ' otherwise
0,
(a) The Fourier transform of z[n]:
00

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)

= e-i(w-Wo)(N/2) (sin [(w- Wo)(Nf2)])


e-j(w-wo)/2 sin [(w- Wo)/2]
X(ei"') = -i(w-... )((N-1)/2) (sinf(w -Wo)(N/2)])
e sin ((w- Wo)/2]
(b) N-point DFT:
N-1
X(k] = L :r:(nJWtn, 0 $ k $ (N- 1)
=•
296
N-1
= L eJwonw;t
n=O
1_ .-;((2••/N)-wo)N
= 1- e j((blc/N) w 0 )

= . ((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

= -j(>•/NH•-.. )((N-1)/2) sin.-(k- ko)


• sin(.-(k- ko)/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,

X[k] = Jl1• + Wt + w;• + w;• + Wf" + W;", 0 :S k :S 4

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

8.8. Fourier transform of z[n] = (1/2)nu[n]:


00

X(eJw) = L z[n]e->wn
n=-CICI

= ~G)" .-jwn
1
=

Now, sample the frequency spectra of z[n]:

Y[k] = X(eJw)iw=2d/1D• 0$ k $ 9

We have the l~pt DFT:


I
Y[k] = 1_ le-i(2wA:flO)'
2

= L• y[n1w,•;
Recall:

So, we may infer:


(!j•
y[n] = 1- '<w·' 0 ::; n::; 9
8.9. Given a 2~pt finite-duration sequence z[n]:

(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

z,[n] = L z[n + 5r]


r=-oo

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,

:~:.[n] = :~:.[((n- 4))a]


From property 5 in Table 8.2,
DF"l'{:~: 1 [((n- 4))a]} = W:' X,[kj
Thus,

x,[kJ = w:• X.[kJ


= .-id X1[k]
x,[kJ = (-1)• x,[kJ

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

Taking the inverse DFT:


y[n] = -30[n]- 60[n- 1] + 30[n- 2] + 60[n- 3], 0 :S: n :S: 3
300

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

We thus conclude :r 3 (2] = 9.


8.15. y(n] is the linear convolution of :r,[n] and :z:2 (n] time-aliased toN = 4. Carrying out the 4-point circular
convolution, we get:

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] = z[((n + 3))s]·


Therefore:
z 1 [0] = z[3] = c.
We thus conclude that c = 2.
8.19. z,[n] and z[n] are related by a circular shift as can be seen from the plots. Using the properties of the
DFT and the relationship between X 1 [k] and X[k], we have:

= 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.

Using the properties of the DFT, we get:

z 1[n] = z[((n + 2))N]·


From the figures, we conclude that:
N=5.
This choice of N is unique.
8.21. (a) We seek a sequence y1 [n] such that
Y,[k] = x,[k]X,[k]
From the discussion of Section 8.2.5, y[n] is the result of the periodic convolution between f,[n]
and i 2 [n].
N-1
y1 [n] = Lf 1 [m]f 2 [n- m]

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

it:liiir ' Iiif


y,[n]
2 1
1
0
•8 '0. I n
-1 0 1 2 3 4 5 6 7 9101112 13 14 15

Using the analysis equation of Eq. (8.11), we may rigorously derive y,[n]:

X,[k]

= Lf,[n]W{"
n=<>
302

= 6 + swt + 4W;' + 3w.f• + 2w:• + w~·


X2[kj

= Li>[nJWt"
.....
= w;>
Y1 [k] = .X1 [k]X2 [k]
= 6W;' + sw:> + 4wt• + 3w~• + 2wt• + w;•
Noting that w;• = &"fml = 1 = W¥•, we use the synthesis equation of Eq. (8.12) to construct
y,[n]. Tbe result is identical to the sequence depicted above.
(b) The DFS of the signal illustrated in Fig. P8.21-2.is given by:

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:

z{((-n)).] = :r:[((-n + N))N]


8.23. We have :r:[n] for 0::; n $ P.
We desire to compute X(z)I•=•-"'""N' nsing one N-pt DFT.
(a) Suppose N > P (the DFT size is larger than the data segment). The technique nsed in this case
is often referred to as zero-padding. By appending zeros to a small data block, a larger DFT may
be nsed. Thus the frequency spectra may be more finely sampled. It is a common misconception
to believe that zero-padding enhances apectral resolution. The addition of a larger block of data
to a larger D FT would enhance this quality.
So, we append N • = N - P zeros to the end of the sequence as follows:

:'[n] = { :r:[n], 0::; n $ (P- 1)


0, P$n$N

: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

z[n] = L z[m] L o[m-n+rNJ


m=-oc r=-oo
00

z[n] = L z[n- rNJ


r=-oo

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

So, /m{X[k]} = 0, 'lk. However, /m{X(eiw)} # 0.


Note also that the size of the DFT plays a large role. For instance, consider taking the 3-pt DFT of
z[n] =
6[n- 1}, 0 ~ n ~ 2
X[k] = Wf, O~k~2
1, k =0
= .-;(b/3)' k=1
{ .-J(.. /l)' k =2
304

Now, Jm{X[k]} #- 0, for k = 1 or k = 2.

8.25. Both sequences z[n] and y(n] are of finite-length (N = 4).


Hence, no aliasing takes place. From Section 8.6.2, multiplication of the DIT of a sequence by a complex
exponential corresponds to a circular shift of the time-domain sequence.
Given Y[k] = w1• X[k], we have
y[n] = z[((n- 3)) 4 ]
We use the technique suggested in problem 8.28. That is, we temporarily extend the sequence such that
a periodic sequence with period 4 is formed.

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.

X[k] = 4+3Wf+2Wt•+wzt, Wf=e-;<><•t•>and0SkS5


W(k] = JU{ X[k]}
= ~ (X[k] + X"[k])
= ~ (4+3Wf +2Wf + w:• +4+3w.-• +2w,-•• + w,- 30 )
305

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,

w[n] = 40[n) + ~ ( o[n- 1) + o[n- SJ) + o[n- 2] + o[n- 4)


+ ~ ( o[n - 3] + o[n - 3])
3 3
w[n] = 40[n) + 2o[n- 1] + o[n- 2) + 6[n- 3) + 6[n - 4] + 26[n - 5),

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:

w[n) = ~ (:r[n) + x[N- nJ)


4
3


-1
I• • ' I I • •
0 1 2
1

3
2

4 5 6
:r[N- n], for N

7
n
=6

So, we observe that w[n] results as al>ove.


(c) The DIT is decimated by two. By taking altemate points of the DFT output, we have half as
many points. Tbe influence of this action in the tin>e domain is, as expected, the appearance of
aliasing. For the case of decimation by two, we shall find that an additional replica of x[n) surfaces,
since the sequence is now periodic with period 3.
306

From part (b):


X[k] = 4 + 3W." + 2Wi' + Wl•, 0$ k$ 5
Let Q[k] = X[2k],
Q[k] =4 + 3Wf + 2Wf' + w:•, 0 :S k :S 2
Noting that W:" = Jl1'•
q[n] = 56(n] + 36[n- I]+ 26{n- 2], 0 :S n :S 2
5

.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]

Note that :r[((0)) 4] =z[O], etc.


In the second approach, we work with the given signal. The signal is confined to
0 $ n $ 3; therefore, the circular nature must be maintained by picturing the signal on the circumfer-
ence of a cylinder.

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

Note that y[n] is a ten point sequence (N = 6 + 5- 1).


(a) For N = 6, the last four non-zero point (6 $ n $ 9) will alias to the first four points, giving us
Yt [n] = z, [nj(§)r2 [n]
308

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

:r'(n] = ! G ) (-1)", 0:0: n :0: 191

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

Since z[n] is of finite duration (N = 10), we assume:


z[n] = { nonzero, 0$n$9
0, otherwise

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

z 1{n] = z{n] ( I .;<•/IO) ) -•


2
8.32. We have a finite-length sequence, z{n] with N =
8. Suppose we interpolate by a factor of two.· That
is, we wish to double the size of :r{n] by inserting zeros at all odd values of n for 0 $ n $ 15.
Mathematically,
y{n] = { z{n/2], n even, 0$ n $ 15
0, n odd,

The 16-pt. DFT of y{n]:


15
Y{k] = LY{nJWt&",
n=O
0$ k $ 15

= 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

8.33. (a) Since


z[n], 0 :S n :S N- 1
:t2[n] = -z[n- N], N :S n :S 2N- 1
{
0, otherwise

If X[k] is known, z 2 [n] can be constructed by :

X[k) •I N-pt IDFT f--:~:[n')l-------,

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.

X1[k] ---.t·IL..__• __,~-----· -2_ X[k)


311

8.34. (a) The DFT of the even part of a real sequence:


H :z:[n] is of length N, then :z:.[n] is of length 2N- 1:

- { :[n]+t[-nJ, (-N+1)~n~(N-1)
:z:.(n] -
otherwise

x.(kJ = NL-1 (:z:[n] + :z:[-n]) w:>n


2 ZN-1, ( -N + 1) :S k :S (N- 1)
n=-N+l

~ :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

x.(kJ = ~ :z:(n]cos u:~"1)


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

So: DFT{:z:.(n]} # Re{X(k]}


(b)

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

8.36. We have the finite-length sequence:

z[n] = 20[n] + cS[n - 1] + cS[n - 3]

(i) Suppose we perform the 5-pt DFT:

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

Using the fact W!• = 141' = 1 and w:• = w;


Y[.I:J = 3 + 5Wt + w;• + 4Wi' + wt•, 0 :0 k ::; 5
313

(a) By inspection,

y(n] = 36(n] + 56(n- 1] + o(n- 2] + 46[n- 3] + o[n- 4], 0:::; n :::; 5

(b) This procedure performs the autocorrelation of a real sequence. Using tbe properties of tbe DFT,
an alternative method may be achieved with convolution:

y(n] =IDFT{X 2(k]} = z(n]• z[n]


The IDFT and DFT suggest that the convolution is circular. Hence, to ensure there is no aliasing, the
2: 2M- 1 where M is the length of z[n]. Since M = 3, N 2: 5.
size of the DFT must be N

8.37. (a)

g,[n] = z[N- 1 - n], 0:::; n:::; (N- 1)

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)

g,[n] = (-1)"z[n], 0:::; n:::; (N -1)

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)

{ :(n], O$n $ (N-1)


g,(n] = :[n - N], N $ n $ (2N -I)
0, otherwise

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

= L :(n] (I+ w;p) W2~


n=O
N-1

= (I+ w~tN/2)) L z[n]W~Im/2)


,..,..,
= {I+ (-l)t) X(eiw)lw=(d/N)
G3[k] = H3[k]
(d)

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)

go[n] = { :[n/2], n even, 0 ~ n ~ (2N -1)


0, nodd

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)

97[n] = x[2n], 0 ~ n ~ (N/2- 1)

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

x(n] =:((( -n))N]·


For 0 ~ n ~ (N- 1), this may be stated as,

x[n] = x(N- n], 0 ~ n ~ (N- 1)


316

For this case, N = 10, and

::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]

To satify both conditions, ::r[n] must be a constant for 0::; n ::; 9.

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

::r[n] @.[n] = I: y[n + rN], 0::; n::; (N -1)


r=-oo

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

Also, y[n] 5,= 0 ::; n ::; 4.


Using the above information:
317

n=O y(O) + y[SO) = 10


y(SO) = 5
n=4 !1[4) + 11[54) = 10
11[54) = 5
n=5 11[5) + 11[55) = 10

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) The linear convolution z[n]•11[n) is a 40 + 20- 1 =59 point sequence:

...-...---____
A•C B•C

----L-----------~--------~----------~--- n
10 28 40 58

Thus, z[n)•y[n) = w[n) is nonzero for 10 S n S 28 and 40 $ n $ 58.


318

(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:

w(n] = ~(n]• y(n], 40 $ n :$ 49

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:

h2(n] = h,(((n + 4))a]


Thus,
H2 [kJ = w.-•• n,[kJ
and
IH,[k]l = lw.-•• H,[k]t = IH,[k]l
So, yes the magnitudes of the S-pt OFTs are equal.
(b) h,[n] is nearly like (sinx)fx.
Since H 2 (k]= .,.. H,[k]. h,[n] is a better lowpass filter.
8.43. (a) Overlap add:
If we divide the input into sections of length L, each section will have an output length:

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:

65 OFTs and 64 IDFTs

(b) Overlap save:


We require 99 zeros to be padded in &om of the sequence. The first 99 points of the output of
each section will be discarded. Thus the length after padding is 10099 points. The length of each
=
section overlap is 256- 99 157 L. =
=
We require 65 x 157 10205 to get all10099 points. Because h(nj also requires OFT: a
66 OFTs and 65 IDFTs

(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

8.44. First we need to compute the values Q[O] and Q[3]:

Q[OJ = Xl(1} = Xl(eJW)lw=O


co
1
= I:x>fn]= - -1
1--
n=O 4
4
=
3
Q[3] = x.(-1} = x.(eJW)lw=r
co
= L r.[n]( -1}n
=<>
4
=
3

One possibility for Q[k], the six-point DFT, is:

4 4
Q[k] = 3"[k] + 3"fk - 3].

We then .find q[n], for 0 ~ n <6 :

q[n] = ~ L' Q[k]e'ftn


>=<>
= 6~(~3 + 3~(-1}n}
= ~(1
9
+ (-l}n)

otherwise it's 0. Here's a sketch of q[n]:

4/9 4/9 4/9


q[n]

l.l.I.
012345
n
320

8.45. We have:

DFIHz2 [n]} = X 2 [k] = "\"



L...z 2 [n]e-''
.....n.
0

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)

V[k] = X(z)l "'=2~tj( 2•\+•)


n=oc
= I: :r(n]z-nl z=2r' , ,--.---
n.=-oo
•••• 1

=•
= L x[n]z-nl , '·~··
..... ll=2e' 1

=
=•
L x(n)(2eii )-n,-j1fin
.....
=
=•
L v(n]e-;1fin.
n=O

We thus conclude that

(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).

We thus conclude that


w[n) = 2x[2n).
(d) Note that Y[k) can be written as:

Y[k] = X(k) + (-!)• X(k]


= X(k] + w:• X(k).
322

Using the DFT properties, we thus conclude that

y[n] = r[n] + z[((n - 4))s].

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,

Let M = 2, then we have

X(ei'f 0 ) = 1- 1 =0
X(d'f') = 1- I= 0

8.49. z 2 [n] is r,[n] time aliased to have only N samples. Since

z,[n] = (31 )"u[n],


We get:
--<11:._ n= O, ... ,N -1
z,[n] = { 1=1Tl"
0 otherwise

8.50. (a) Let n = 0, ... , 7, we can write x[n] as:


z[n] = I+ !(dtn + e-ifn)- !(dlf• + .-;lfn)
2 4
= 1 + !d'f"+
2
!d'f"7 - !d'fn3- !d'f"'
2 4 4
= !(s + 4d'f• + 4d'f•7 _ :W"fn3 _:lei¥"')
8
7
= ! 2: Xs[k]d'i"•
80=0

We thus get the following plot for Xs[k]:


323

8
Xa[k]

4 4

k
0 1 2 5 6 7

(b) Now let n = 0, ... , 15, we can write v[n] as:


1 1 ...,_ ....
v[n] = 1 + -(e', n + .-nn) _ -(e' , n + 0 - 1 , ")
"L ·•

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

We thus get the following plot for V16 (kj:

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

where X(&w) is the Fourier transform of z[n].


Note that z[n] can be expressed as:
z[n] = y[n]w[n]
where:
'"' 1 31m
y[n] = 1 +cos(-)-
4
-cos(-)
2 4
324

and w(n] is an eight-point rectangular window.


IX1o[k]l will therefore have as its even points the sequence JX8 (k]l .The odd points will correspond
to the bandlimited interpolation between the even-point samples. The values that we can find
exactly by inspection are thus:

IX1o[k]l = IXs(k/2]1 k = 0,2,4, ... ,14.

8
IX1o[k]l

I;>
4 4

2 2

"' "' "'


"'5 "' "' "'11
•. : k
0 1 2 3 4 6 7 8 9 10 12 13 14 15

8.51. We wish to verify the identity of Eq. (8.7):

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

Since m and n are integers;


.,J~(&-r)n = 1 , fork- r = mN
So,
N-1 N-1
2_ L .,J~(·-r)n I;1
= -Nn=<>
1
Nn=O
= 1 , for k - r = mN
(b)
N-1 ·~IN
1 " ' ·~In- 1-e'Y
N L.... e' - 1- eiif 1
n=O

This closed form solution is indeterminate for l mN only. =


For the case when l = mN, we use L 'Hopital 's Rnle to find:

lim
1 _ ei2-'
·~,
l-..mN 1- eJ-,r = [=:::;:] N l=mN
= N
325

(c) For the case when k- r # mN:


JV-1 1- ~2r(l-r)
~ L ~\f(l-r)n =
1 _ e>' \f(l-r)
n=O
= 0
Note that the denominator is nonzero, while the numerator will always be zero for k- r # mN.
8.52. (a) We know from Eq. (8.11) that if £ 1 [n] = f[n- m], we have:
JV-1
X,[k] = L f[n- m]W,t"
.....
If we substitute r =n - m into this equation, we get:
N-1-m
X1[k] = L f[r]W~(r+m)
r=-m
N-1-m
= w,tm L f[r]W,t'
r=-m

(b) We can decompose the summation from part (a) into

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

(c) Using the result from part (b):

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

2. The DFS of z'[-n]:


N-1
L
n=O
z'[-n]W~" = Lt+l z[IJW;J).
= X'[k]
3. The DFS of &{z[n]}:

~ z[n] + z'[n] w"" ~ ( X[kJ + X'[-kJ)


L- 2 N =
n=O
= x,[k]
4. The DFS of j/m{f[n]}:

~ z[n] - x' [n] w•n


L-
n=O
2 N = 21 (-X[k]- X'[-k]
- )

= Xo[k]
(b) Consider z[n] real:
1.
Re{X[k]} = X[k] + X'[k]
2
From part (a), if f[n] is real,

DFS {z[n]} = DFS {f'[n]}


DFS {z[-n]} = DFS {f'[-n]}

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

(i.e., the magnitude of X[k] is even)


4.

d[kJ = arctan (Im{~[k]})


~{X[k]}

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:

X(elw) = j;oo (~% X(k]W,Y•") .-'"'"


Rearranging the summations and combining terms:

The infinite summation is recognized as an impulse at w = (27rk/ N):

X(ei"') =!.. ~ X(k]6 (w- 2"k)


Nt=<> N
2. Since :r(n] corresponds to one period of f(n], we must apply a rectangular window (unit amplitude
and length N) to the periodic sequence. Thus, to extract one period from f[n]:
:r(n] = f(n]w[n]

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)).

X(eiw) = 2.1• d/1.!. ~


2.. -111'
k..
N ~
X[kJo ( 8 _ 2lrk)
N
sin [~(w- 8)] ,-;(¥ )<w-l)
-'n (!~!=!)
OMM2

Integration over - .. :::; 8 :::; r reduces to the summation (note the impulse train):

X(eiw) = ..!_ ~ X[kJsin ((Nw- 2rk)/2J,-;(¥)(w-'k")


N ~ sin((w- ~)/2]
Hence, the Fourier transform is obtained from the DFS via an interpolation formula.

8.55. The N-point DFT of the N-pt sequence, :r[n] is given by


N-1
X(kJ = L :r[n]W~n, 0:::; k:::; (N- I)
n=O
N-1
X[O] = L :r[n]
n=O

=
(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]

Since :r[n] = :r[N - I - n], then


z[O] = :r[N - I]
z[I] = z[N- 2]

Therefore, X[N/2]= 0.

8.56. (a) The conjugate-symmetric part of a sequence:

z,[n] = ~ (:r[n) +:r"[-nJ)


The periodic conjugate-symmetric part:

z,.[n] = ~ (:r[((n))N] +:r"[((-n))NJ), 0:::; n:::; (N -I)


329

Note that:

z[((n))t,] = z[n], 0 :0: n :0: (N- 1)


z"[((-n))N] = z"[-n + N] + z"[O]o[n]- z"[O]o[n- N]

Substituting into z,.[n]:

z,.[n] = ~ [z[n] + z"[-n + N] + z"[O]o[n]- z"[O]o[n- NJ], 0 :0: n :0: (N- 1)

Since,

z,[n] = H z[n] +z"[-nJ)

= ~ ( z[n] +z"[O]o[nJ), 0:0: n :0: (N -1)

andz,[n-N) = ~(z!n-N)+z"[N-nJ)
= ~ (- z"[O]o[n- N] + z"[N- nJ) 0$ n $ (N- 1)

'llf'e can combine to get:

z .. [n] = z,[n] + x,[n- N] 0 $ n :0: (N- 1)

The periodic conjugate-antisymmetric part is given as


z .. [n] = (z.{n] + z 0 [n- N]), 0 :0: n :0: (N- 1)

Recall that the odd part can be expressed as

z 0 [n] = ~ (z[n] - :t" [-n])

So,
z.[n- N] = ~ (z[n- N]- z"[N- nJ)
For 0 :5 n :5 (N - 1):

:t0 [n] = ~ ( z[n]- z"[O]o[nJ), 0 :0: n :0: N- 1

z 0 [n- N] = H z[O].![n- NJ- z"[N- nJ)

From the definition of z.. [n]:

z .. [n] = ~ ( z[((n))N]- z"[((-n))NJ), 0:0: n :0: (N -1)

= ~ (z[n]- z"[O]o[n] + z"[0]6[n- NJ- z"[N- nJ), 0 :0: n :5 (N- 1)

Recognizing the expressious for z.[n] and z.[n - N] in z ..[n], we have


z.. [n] =z.[n] + z.[n- N], 0 :0: n :0: (N- 1)
330

(b) z[n] is a sequence of length N; however,

z[n] = { "' 1 [n], 0 Sn S N/2


0, N/2SnSN-1

The even part: (assume N is even) '


z[n] z*[O]o[n] _< n _< N/ 2
2+ 2 • 0
:r.[n] = z•[-n]
- 2-. -N/2 <_ n S -1

0, otherwise

From part (a):


x.,.[n] = :r.[n] + :r.[n- N], 0 S n $ (N- 1)
Because :r[n] = 0 for jnj ?: N /2,
"'••[n] = z.[n], 0 $ n $ (N /2- 1)

Also, since x.[n] = z;[-n],

"'•[n] = .,;.[-n], -N/2 < n $ -1

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

From the synthesis equation:


N-1
:r[n] = ~ L X[k]WN""
•=<>
Hence,
N-1
z*[n] = ~ L X*[k]W~"
•=<>
substituting:
N-1
Ln=<l
\:rln1\'
331

= ! I: X'[k] ('f z[n]Wff)


1:=0 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

8.58. (a) This statement is TRUE:

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]

Expressed in the form

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

8.60. Ideally, the inverse system would be:


H;(z) = z- bz-•
Hence,
X(z) = (z- bz- 1 )Y(z)
and
:r[n] = y[n + 1]- by(n- 1], -oo ~ n ~ oo
H we use an N-pt block of v[n]:
y[n] : 0 ~ n ~ N,
then
v[kJ = (w,;v• -·bw~JY[kJ
and
v[n] = v[((n + 1))N]- by[((n- 1))N]

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

So, x.,[n] = x[n]e-;(wn/N).


(b)
N-l
x.,[N- k] = L x[n]e-;(rn/N).-;(.. /N)(N-•)•, 0 ~ k ~ (N- 1)
n=O
N-l
= L :r[n]e-;(rn/N)ei(2r/N)•
n=O
N-l
x.,[N- (k + 1)] = L :r[n]e-i(rn/N).-i(2r/N)(N-•-•)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)

x,.,[kJ = x1 ... [kJX, ... [kJ


N-1
x, ... [n] = L Z1M[r]x, ... [((n- r))N]
•=<>
From part (a):
:z:1M[n] = :z:1[n]e-il•n/N)
Z2M[n] = :z:,[n]e-i(•n/N)
Z3M[n] = z 3[n]e-;( .. /NI
So,
N-1
:z: 3(n] = ei<•n/N) L z.(r]z2[((n- r))N]e-ii•/N)[((n-r))H+r]
r=O
N-1
= L z.[r]:z:,[((n- r))N]e-i<•/N}[((n-r))H-(n-•)]
r=O

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:

ZtM[n] ~>M[n] = ZtM[n] • Z>M[n)


(i.e. no aliasing occurs.)

Z3M(nj = ZtM(n) * Z2M[n)

-
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):

let :r1 [n] = :r[n), 0 :5. n :5. 31


z,[n) = :r[n + 32], 0 :5. n :5. 30
h 1 [n) = h[n], 0 :5. n :5. 31
h 2 [n] = h[n + 32], 0 :5. n :5. 30

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.

y(n) = :r[n)• h(n)


= u,(n) + 112(n- 32) + y3(n- 32) + ll<[n- 64)
So
:r[nJ@).[n) = y(n) + y(n + 63), 0 :S n :S 62
The total computational cost is 12 DFTs of size N = 64.
(b) Using two 128-pt DFTs and one 128-pt inverse DFT:

y(n) = :r[n) @!n) = :r[n)• h(n)


The 63-pt circular convolution:

:r[nJ@.[n) = y[n) + y[n + 63), 0 :S n :S 62


(c) Using the 64-pt DFT method of part (a):

#mult = 4(12)(64log2 (64)) = 18432


Using 128-pt DFI's:
#mult = 4{3)(128log2{128)) = 10752
Direct convolution:
63
#mult =2 L n - 63 = 3969
=•
8.63. From each circular convolution, the first 49 points will be incorrect. Therefore, we get 51 good points
and the input must be overlapped by 100 - 51 = 49 points.

{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

h;[n] = t. f"" G o[n - kno]

The filter is IIR.


(c)
1 1
G[k] = H[k] = 1 _.,-;<•l•l•, 0 $ k $ (4no- 1)

The impulse response, g[n], is just h;[n] time-alia.sed by 4no points:

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)

y[n] = g[n] • h[n]


M 8 4 2 8
= 156[n] + 156[n- no]+ 156[n- 2no] + 156[n- 3no]- 156[n- no]
4 2 1
- o[n- 2no]- o[n- 3no]- o[n- 4no]
15 15 15
16 1
y[n] = 15
o[n]- o[n- 4no]
15
8.65. (a) We start by computing Xn[k + NJ:

N-1
Xn[k+ NJ = L i[n]HN[n(k+ N)]
n=O

= I; i[n](cos(2"(nk; nN)) + sin(2.-(nk + nN))


n=O N
N-1
= L
..,..
i[n](cos(21mk) + sin(2wnk ))
N N
N-1
= L i[n]HN[nk]
= Xn[k].
337

We thus conclude that the DHS coefficients form a sequence that is also periodic with period N.

(b) We have:

I N-1 I N-1 N-1


N
.,..L XH[k]HN[nk] = N L(L z(m]HN(mk])HN(nk]
1:=0 m=O
} N-1 N-l
= -
N
L z[m) L HN(mk]HN(nk]
m=O .t=O

= ~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))

= (CN[a]CN[b]- SN[a]SN[b]) + (SN[a]CN[b] + CN[a]SN[b])


= CN[b](CN[a] + SN[a]) + SN[b](-SN[a] + CN[a])
= CN[b](CN[a] + SN[a]) + SN[b](SN[-a] + CN[-a])
= CN[b]HN[a] + SN[b]HN[-a]
= CN[a]HN[b] + SN[a]HN[-b] ( s"mce HN[a + b] HN[b +a] ) =
Where we have used trigonometric properties.

(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]

Where we have used the periodicity of HN(nk] and i(n].

(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

This is the desired ·convolution property.

(f) Since the DFT of :r[n] is given by:


N-1
X[k] = L :r[n]e-1'iF
......
N- 1 2..kn 2,-kn
= L :r(n](cos(--~d + jsin(---:;.,--))
=0
339

~
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))

This allows us to obtain XH[k) from X[k).

(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

= ~(Xs[k] + Xs[((-k))N]l- j~(Xs[k]- Xs[((-k))N])


1 . 1 .
= (-- l)Xs[k] + (- + l)Xs[((-k))N]
2 2 2 2

This allows us to obtain X[k] from Xs[k].

8.66. (a) The DTFI' is given by:

X(ei") = . X(ei") + X(ei").-; .. N


= X(ei"')(l + e-i"'N)

The DIT is just samples of the DTFI':

X[kJ = .i(ei"ll ..=w


= X(ei2•12N)(l + (-l)t)

Therefore:

.i[kJ = { 2X[!~ k even


k odd

(b) The original system computes the following:

X[k]H[k] = { 2X[!JH[k~ ' keven


k odd

We thus want:

X[k]G[k] = 2X[k]H[2k] k = o, ... ,N -1


G(k] = 2H[2k]
2N-1

= 2 L h[n]e-;9 k =O, ... ,N -1


.....
g[n] = 2(h[n] + h(n + N])

System A time aliases and multiplies by 2.

For system B, we need:

Y[k] = { W(~~ k even


k odd
341
Thus:

y(n] = 1 w(n] , o::;n::;N-1


w[n-N]
0
N::;n::;2N-1
otherwise

System B regenerates the 2N -point sequence by repeating w(n].


8.67. (a) We have:

= { 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:

0::; k ::; i4N


otherwise
4N-~::;k-s_4N

that is:

&t• 0 < k < !f.


= o~he~
H(k]

1 eit•

(b) System A needs to perform the following operations:


0 :
'
4N- !f. < k < 4N
2 - -

X(k]H'(k] 0::; k::; ~


=
Y2(k]
1 X(k- 3N]H'(k- 3N}

Where H'(k] is the N -point DFT of h(n].


0 otherwise
4N-!f.<k<4N
2 - -

(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:

:r(n] = z 1 [n] for n = O,!, ... ,N -1


N-l

= N ~ 1 <2:: o[A:jxe•[k]cos(;:"l))
•=<>

where a[k] is given by:

o[k]= U k=OandN-I
I $.1: $N-2.

This completes the derivation.


8.69.

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:

V[k] = ~(X2(k] + X2[k + N])


= e'~ Re{Xfkje-;ifr} + .,i"'tlNl Re{X(k + N]e-; •<uNl}
= .,iii Re{X(k]e-;~} + eii,ifi Re{X(k + N]e-;~e-;i}
= .,;~ Re{X[k]e-;~} + ;ei~ Re{ -jX(k + N]e-;flt}
= ,ift(Re{X[k]e-;ffr} + jlm{X[k + N]e-;fi}.
343

Using the above expression, we get:

2Re{e-i~V[k]} = 2Re{e-i'* ei'* (&{X[k]e-ift} + jlm{X[k + NJ•-'*})}


= 2&{X[k]e-''*}
= X'2 [k] where we used equation (8.170).

Furthermore, we have:
N-1
2Re{e-i~V[k]} = 2&{e-i~ I; v[n]e-i"P}
n=O
N-1

= 2Re{ I; v[n]e-i< 'N'<•+lll}


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; :r[n]•_,zw .-''*}


=0
N-1

= 2Re{I; :r[n]e-iftl 2 •+1l}


n=O
N-1

= 2 I; Re{:r[n]e-iftl 2•+1l}
n=O
N-1
[ ] cos ("k(2nN+ 1)) .
2 I; :rn
= 2
n=O

From the results above, we conclude, for k = 0,1, ... , N - 1:

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:

z[n] = z,[n] for n = 0, 1, ... , N - 1


N-1
= ~ L P[k]X''[k]cos(""k(~';,+ 1)) O~n~N-1
.=0
where P[k] is given by:

P[k] = { ~ k=O
1~k~N-l.

This completes the deri\'B.tion.


8. 71. First we derive Parsev..l's theorem for the DFT.
Let x[n] be an N point sequence and define y[n] as follows:

y[n] = :r:[nJ®z*[((-n))N]·
Using the properties of the DFT, we have:

Y[k] = X[k]X*[k] = fX[k]f 2 •


·Note that:
y(O] = L fz[n]l 2
n

and using the DFT synthesis equation, we get:


N-1
1
y[O] = N
....L Y[k] .
345

Parseval's Theorem for the OFT is therefore:


N-1
L jz[n]l' = ~ L IX[k]l".
n A:=O

(a) Note that:


N-l N-1
L IX'1 [k]l 2 =L jX,[k]j 2
n=O

and, using equation (8.164):


2N-3 N-1
L JX,[k]j 2 =2 L IX"[k]j 2 -IX"[OJJ 2 -IX" IN -1]j
2

Using the OFT properties:


2N-3
L jz,[n]l' = 2Nl- 2 L JX,[k]J'
n '=0
and, using equation (8.161):

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'.

(b) Using equation (8.171).


N-1 /'rl-1
L IX'2 [k]l 2 =L jX,[k]j 2 .
n=O n=O

Note that, using equation (8.167):


2N-l N-1
L jX,fkJI' =2 L IX[k]l 2 -IX[O]j 2 ,

and, using equation (8.!66):


2N-l N-1
L jz,[n]l' =2 L lz[n]j 2 •
n=O n=O

Using the OFT properties:


2N-l l 2N-l
L
n=O
jz,[n]l' = 2N L
....
jX,[kJI'.

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

Computation of the Discrete-Time Fourier Transform


349

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

9.2. (a) The "gain" along the emphasio:ed path is -W}.


(b) In general, there is only one path between each input sample and each output sample.
(c) o:[O] to X[2]: The gain is I.
o:[l] to X[2]: The gain is W}.
o:[2] to X[2]: The gain is -WX, =-I.
x[3] to X[2]: The gain is -WX,W} -W}. =
x[4] to X[2]: The gain is WX, I. =
x[5] to X[2]: The gain is WX,W} = W}.
x[6] to X[2]: The gain is -WX,WX, =-I.
=
x[7] to X[2]: The gain is -WX,WX,W} -W}, as in Part (a).
Now
7
X[2] = L z[n]w;n
n=O

= z[O] + z[1]W/ + z[2JW: + z[3JW: + z[4JW: + z[S]W/ 0 + z(6]W/ 2


+ z[7JWi 4
= z(O] + z(1]W/ + z[2](-1) +z(3](-W/) + z(4](I) + z[SJWi
+ z(6]( -1) +z(7)(- Wf)
350
Each input sample contributes the proper amount to the output DFI' sample.

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]

Tbe output should then be extracted from D[r] in sequential order.

X[k] = D[k], k =0, ... , 7


(b) First, we find the DFT of (- WN )" for N = 8.

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

(c) First, the array D[r] is expressed in terms of C[r].

D[O] = C[O] + C[4]


D[l] = C[l] + C[S]WJ
D[2] = C[2] + C[6]Wi
D[3] = C[3] + C[7]Wi
351

D[4] = C[OJ - C[4]


D[S] = C[l]- C[S]Wi
D[6] = C[2]- C[6]Wi
D[7] = C[3]- C[7]Wl
Solving this system of equations for C[r] gives

C[O] = (D[O] + D[4])/2


C[l] = (D[l] + D[S])/2
C[2] = (D[2] + D[6])/2
C[3] = (D[3J + D[7])/2
C[4] = (D[O]- D[4])/2
C[S] = (D[l]- D[5])W8 1 /2
C[6] = (D[2]- D[6]) /2 w.-•
C[7] = (D[3]- D[7])W8- /2
3

for r = 0, 1, ... , 7. A sketch of C[r] is provided below.

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

1st stage: w,'


2nd stage: Wl, Wi
3rdstage: w,',W.',WJ',Wl
Here we have listed the different coefficients only. The values above correspond to the impulse
response
h[n] = w;.. u[n]
which can be generated by the recursion

y[n] = w.-y[n- 1] + :r[n]


Using this recursion, we only generate a sequence of length L = 2m-l, which consists of the different
coefficients. Then, the remaining If -
L coefficients are found by repeating these L coefficients.
(c) The difference equation from Part (b) is periodic, since

h[n] = Wi- u[n]


= .-j>•n/2- u[n]

has a period R = 2m. Thus, the frequency of this oscillator is


2"
w = 2""
352

9.5. Multiplying out the terms, we find that

(A- B)D+ (C -D)A =AD- BD +AC- AD= AC -BD =X


(A- B)D+ (C+ D)B =AD- BD +BC +BD =AD+ BC =Y
Thus, the algorithm is verified.
9.6. Answer 3
Decimation in Time: The figure is the basic hutter:fly with r = 2.
Decimation in Frequency: The figure is the end of one butter:fly and the start of a second with r = 2.
9. 7. The figure corresponds to the flow graph of a Sec:ond-order recursive system implementing Goertzel's
algorithm. This system finds X[k] for k = 7, which corresponds to a frequency of
14>< 7r
"'• = 32 = 16
9.8. This is an application of the causal version of the chirp transform with

N = 20 The length of z[n]


M = 10 The number of desired samples
2•
Wo = T The starting frequency
2•
f:>w = 21 The frequency spacing between samples

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

N = 12 The length of z[n]


M = 5 The number of desired samples

= ~; The starting frequency


""'
Aw = h
10 The distance in frequency between samples

Letting W = .-;&. we must have


r[n] = .-;wonwn'/2 = .-;'ftn.-;tfn'/2
9.11. Reversing the bits (denoted by -+) gives

0 = 0000 .... 0000 = 0


1 = 0001 .... 1000 = 8
2 = 0010 .... 0100 = 4
3 = 0011 .... 1100 = 12
4 = 0100 .... 0010 = 2
5 = 0101 .... 1010 = 10
6 = 0110 .... 0110 = 6
7 = 0111 .... 1110 = 14
8 = 1000 .... 0001 = 1
9 = 1001 .... 1001 = 9
10 = 1010 .... 0101 = 5
11 = 1011 .... 1101 = 13
12 = 1100 .... 0011 = 3
13 = 1101 .... 1011 = 11
14 = 1110 .... 0111 = 7
15 = 1111 .... 1111 = 15

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

9.15. Plugging in some values of N for the two programs, we find

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]

Y>fN] = :t[N] + W~z[N- 1) + · · · + W~(N-l):t[1) + Wf/' z[O)


N-1
= o+ L w~<N-ll:[lJ
l=O
N-1
= L w,v":[lJ
l=O
N-1
= L :[l]W~-·)l
l=O
= X[N-k]

(b) Using the figure, we lind the system function Y.(z).

Y.(z) = X(z) 1- w,v•z-•


1- 2z 'cos(~)+ z-2
1-w-•z-1
= X(z) N
(1- w,v•z-•)(1- w~z-')
X(z)

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

The contribution of all the Wt


's on the flow graph is 28 real multiplications and 20 real additions. The
butterflies contribute 0 real multiplications and 32 real additions per stage. Since there are four stages,
the butterfiies contribute 0 real multiplications and 128 real additions. In total, 28 real multiplications
and 148 real additions are required to implement the flow graph.
9.23. (a) Setting up the butterfly's system of equations in matrix form gives

Solving for

gives

which is consistent with Figure P9.6-2.


(b) The flow graph appears below.
359

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

Using this W[k), Y[k) = W[k]X[k].


(b) W[k] = G*l, where p[k] =the number of ones in the binary representation of index k.
(c) A procedure for finding r[n] is as foUows.
step 1: Form W'[k] = 1/W[k].
step 2: Take the inverse DFT of W'[k], yielding w'[n].
step 3: Let r[n] he the circular convolution of z[n] and w'(n].
U r[n] is input to the modi6ed FFT algorithm, then the output will be X(k], as shown below.

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

From the definition of the z-transform,


N-1
X(z&) = L z[n]z;"
n=O

Plugging in z•, and setting W = .-;t;.,,


N-1
X(z&) =L :z:(n](0.5)-•e-iwo•w""
n=O

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

y[n] = { z[n] + z[n + 256], 0 :S n :S 255


0, otherwu;e

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

(b) Evaluating the DIT at the odd-indexed samples gives us


N-1
X[2k + 1] = L :~:[n]e-;(2•/N)(2H1)n
.....
N/2-1 N-l
= L :<[n]e-i2•n/N .-;2•1n/(N/2) + L :<[n]e-i2•n/N .-i2•>n/(N/2)
n=O n=N/2
N/2-1
= DIT N/ 2 { :<[n]e-;(2•/N)n} + L :<[1 + (N/2)fe-i2•{1+(N/2)}/N 0 -;2d{I+(H/2)]/(N/2)

'"" N/2-1
= DITN/2 { :t[n]e-;(2•/N)n} + ( -1)( -1) L :t[l]e-;2•1/N 0 -;2rti/(H/2)

= DIT H/2 { 2:<[n]e-;(2r/N)n}

=
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]

Thus, the algorithm produces the desired results.


(b) Taking theM-point DIT Y[k], we find
.M-1 00
Y[k] = L L :t[n + rM]e-i2dn/M

oo M-1
= L L :t[n +rM]e-j2d(n+rM)/Me>"2•(rJt)t/JI
r=-oo n=O

Let I= n + rM. This gives


00

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

(:r[n]- z[n + (N/2)])e-i(.. /Nln, 0$ n::; (N/2) -1


y[n] ={ 0, otherwise

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]

Hence, XR(k) = XR{N- k) and Xr(k] = -Xr(N- k].


(b) In Part (a) it was shown that the DIT of a real sequence z[n] consists of a real part tbat has even
symmetry, and an imaginary part that has odd symmetry. We use this fact in the DIT of the
sequence g[nJ below.

G{k] = X 1{k] + jX2[k]


= (X1ER(k) + jX1or{k]) + j(X2ER[k) + jX2or(k])
= X1ER[k]- Xwr{k] +i (X1or[k] + x.ER[k])
real part imaginary part

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

Therefore, the even and real part of G[k] is

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].

X 1[k] = X1ER[k] + jX,or[k)


= GER[k] + jGor[k]
x.[k] = X2ER[k] + jX,or[k)
= GEr[k]- jGoR[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

X[k] = L z[n)e-j2wlm/N + L z[n)e-;2wlm/N


n even nodd
Substituting n = 2! for n even, and n = 2l + 1 for n odd, gives
(N/2)-1 (N/2)-1
X[k] = L :[2l)e-i2•1l/(N/2) + L :[2! + l)e-;2d(2l+l)/N
l=O l=O
(N/2)-1 (N/2)-1
= L :[2l]e-j2dl/(N/2) + ,-;2rt/N L :[2! + l)e-j2dl/(N/2)
l=O l=O

X1[k) + .-;•••IN X,[.l:),


={ X,[k- (N/2))- .-;•dfN X 2 [.1:- (N/2)],

(e) The algorithm is then


366

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

X,[k] = GsR[k] + jGor[kJ


X,'(k] = e-;2>•/N (Gsr(k]- jGoR[k])
which each have length Nf2.
step 5: Then, form
X[k] = X.[k] + X,'[k],
step 6: Finally, form
X[k] = X"[N- k],
Adding up the computational requirements for each step of the algorithm gives (approximately)
step 1: 0 real multiplications aad 0 real additions.
step 2: 2~ log2 ~ real multiplications aad 3~ log2 ~ real additions.
step 3: 2N real multiplications aad 2N real additions.
step 4: 2N real multiplications aad N real additions.
step 5: 0 real multiplications and N real additions.
step 6: 0 real multiplications and 0 real additions.
In total, approximately N log2 ~ + 4N real multiplications aad !N log2 ~ + 4N real additions are
required by this technique.
The number of real multiplications aad real additions required if X[k] is computed using one N-
point FFT computation with the imaginary part set to zero is 2N log2 N real multiplications aad
3N log2 N real additions.
9.32. (a) The length of the sequence is L + P- 1.
(b) In evaluating y[n] using the convolution sum, each nonzero value of h[n] is multiplied once with
every nonzero value of z[n]. This can be seen graphica.lly using the flip and slide view of convolution.
The total number of real multiplies is therefore LP.
(c) To compute y[n] = h(n] • z(n] using the DFT, we use the procedure described below.
step 1: Compute N point DFTs of z[n] and h[n].
step 2: Multiply them together to get Y[k] = H[k]X[k].
step 3: Compute the inverse DFT to get 11[n].
Since y[n] has length L + P - 1, N must be greater thaa or equal to L + P - 1 so the circular
convolution implied by step 2 is equivalent to linear convolution.
367

(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

DFT of z[n] (N/2)log2 N


DFT ol h[n] (N/2)1og2 N
Y[k] = X[k]H[k] N
Inverse DFT of Y[k] (N/2) log., N
(3N/2)log.,N +N

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.

N Direct Method FFI' method


2 1 20
4 4 64
8 16 176
16 64 448
32 256 1088
64 1024 2560
128 4096 5888
256 16384 13312

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)

lim 2"(v+1) : lim v+1


11-+CIO 2" - P +1 11-+IXI } + ;:)
: "
Thus, for P : 500 the direct method will be more efficient for " > 500.
(d) We want
.;,2:,.:"(,;_";.+..:.1~) < P.
2"-P+1-
Plugging in P: L/2: 2•- 1 gives
2"(v + 1) < ._ _
2 1
2• -2•- 1 + 1 -
As seen in the table below, the FFI' will reqnire fewer complex multiplications than the direct
method when " : 5 or P : 2' : 16.

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,

z(nt] = ~w-t't> CE G(k]w<•-tJ'I')

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

9.35. First note that

:r;[n]
= { :r[n], iL:;; n :S iL + 127,
0, otherwise

= { :r[n + iL], 0 :S n :S 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)

(b) The prw>ed Bow graph is shown below.


372

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

Number of multiplications = (Unpruned multiplications)+ (Pruned multiplications)


N .,_,
= 1'"2+ 2;2"'
1=1
373

9.37. (a) Starting with the equation

f[n] = :r[2n+ 1]- :r[2n -1] +Q, n = 0,1, ... , 2N -1,

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

f[n] = h[n]- h[n- 1] + Q, n = 0, 1, · · ·, 2N - 1.

Taking the N (2 point DFT of both sides gives


f-1
F[k] = H[k]- wt1,H[k] + Q L Wtj,

= H[k](1- Wf/) + ~ Qo[k]


So
F[O] = NQ
2
~
1
N2
= 2N .t.... :r[2n + 1]
n=O
f-1
= L x[2n+ 1]
n=O
= H[O]
Therefore,
X[k] = G[k] + WtH[k]
X[O] = G[O] + H[O] = G[O] + F[O]
X[N/2] = G[N/2] +w;12 H[N/2]
= G[O] + w; H[O]1'

= G(O]- H[O] = G[O] - F[O]

(b) The equation,


F[k] = H[k](1- Wf/) + ~ Q6(k]
for k # 0 becomes
Ffk} = Hfk){l - wJ,•)
= H[kJWt(wN•- Wt)
374

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

X[k] = G(k]- !; F[.l:] k "#- 0, N/2


2 sm(27rk/N)
requires (N - 2) /2 multiplications of the purely imaginary "twiddle factor" terms by the complex
coefficents of F[k] fork"#- 0, N/2. The number of multiplications were halved using the symmetry
sin(2,.(k + N/2)/N) = - sin(27rk/N) and the fact that F[k] is periodic with period Nf2. Since
multiplying a complex number by a purely imaginary number takes 2 real multiplies, we see that
the equation requires a total of (N- 2) real multiplies per stage.
We also need (N- 2) complex additions to add the G(k] and modified F[k] terms for k f. 0, N /2.
Since a complex addition requires two real additions, we see that the equation takes a total of
2(N - 2) real additions per stage.
Putting this all together with the fact that there are log2 N stages gives us the totals
Real Multiplications = (N - 2) log2 N
Real Additions = 2N log2 N .

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

+ z[n + (N/2)]w;+<Nf'>w;,•<•+<NI•» + z[n + (3N/4)Jw;+<•NJ•>w:,•<•+<•NI•»)


(N/4)-1
= L {(z[n]- z(n + (N/2)])- j(z[n + (N/4)]- z[n + (3Nf4)])}w;w~•·
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

+ j(z[n + (N/4)]- z[n + (3N/4)])}W~·w:,••, k = 0, 1, ... ,(N/4) -1.

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

g(n] = z[n] + z[n + (N /2)], n = 0, 1, ... , (N /2)- 1


f,(n] = z[n]- z[n + (N /2)], n = 0, 1, ... , (N/4)- I
/,[n] = z[n+(N/4)]-z(n+(3N/4)], n=O,!, ... ,(N/4)-1

A diagram for computing the values of X[k] looks like:

)
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

If we carry on applying the split-radix principle, we get the next diagram:


377

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)

9.39. (a) Noting that

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.

(b) An algorithm to compute 9, is described below.

Qo = +1
-B = ao9o
fori=1toM-1
if (8> 9)
Oi = -1
else
Qj = +1

end for

Using this algorithm, the sequence a 1 is found to be

IQ: I~ 1-~ I~ I~ 1-~ 1-~ I~ I; 1-~ 1-~ I~~ I


(c) Note that (X+ jY)(1 + ja;2-') =(X- a,Y2-') + j(Y + a;X2-'). Hence, the recursion is simply
multiplying by M complex numbers of the form (1 + ja;2-'). These can be represented in polar
form:
(1 + ja,2-') = ,,IJ + 2-2<.,Ja . ..-....(2-'l
= G;.tJail;

Multiplication of polar numbers produces a sum of the phases, a;9,.


M-1
B= L a,9,
i=O

(d) Multiplication of polar numbers produces a product of the magnitudes, G;.


M-1
GM = II v'1 +2-2i
i=O

9.40. (a) Starting with the definition of the DFT,


N-1
X[k] = L :z:(n]w;;•
n=O
380

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

= L (z[n] + z[n + N/3] +z{n + 2N/3JlWM3


n=O

Define the sequence


zi[n] =z[n] + :r[n + N/3] + :r[n + 2N/3]
The 3-point DFT of zi[n] is Xi[k] = X[3k].
(b) This part is similar to part (a). First, :r2 [n] is found. Starting again with the definition of the
DFT,
N-1
X[k] = L z[n]w~•
n={l

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

= L (z[n] + z[n + Nf3JW:I• + z[n + 2N/3JW;:'


13
)W~W,m3

Define the sequence


13
z 2 [n] = (z[n] + z[n + N/3JW::13 + z[n + 2N/3JW;:' )W~
The 3-point DFT of z 2 [n] is X2 [k] = X[3k + 1]. Next, z 3[n] is found in a similar manner. Starting
with the definition of the OFT,
N-1
X[k] = L z[n]~•
n=()

N-1
X[3k + 2] = L z[n]w,:;<•>+•>
n=()

N/3-1 2N/3-1 N-1


= L z[n]w,:;t•>+•> + L z[n]w,:;<>t+2l + L z[n]w,:;t>t+•>
n=O n=N/3 n=2N/3

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

Define the sequence


z 3[n] = (:r[n] + z[n + Nf3]W~N/l + z[n + 2N/3]w'~ 13 )Wl-"
=
Tbe 3-point OFT of :r3[n] is X3[k] X(3k + 2].
(c) To draw tbe radix-3 butterfly, it helps to derive the output of the buttertly first. From the definition
of the OFT,

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]

vi, xpl N=30FT XIS].

w: xp!
X(8]

(e) The system consisting entirely of N = 3 DFTs is drawn below .


384

x1(0(
lC(O) X) OJ

><(1) N=3 ~ N=3 X)3)


OFT OFT

><(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]

x[7J N•3 N=3 X[S]


OFT OFT

><(8) ----------ol X[8]

(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

Number of complex multiplications = (4


complex multiplications) (N 3-pt DFI's) (v stages)+
3-pt DFI' 3 stage
N twiddle factors ) . h .ddl facto )
( stage with twiddle factors (v - 1 stages WJt IW1 e rs
385

Replacing v with log, N, and simplifying this formula gives

Number of complex multiplications= ~Nlog, N- N


Note that this formula for a radix-3 FFT is of the form N log, N. The constant multiplier, ·isi,
significantly larger than that of a radix-2 FFT. This is because a radix-2 butterfly bas no complex
multiplications, while in part (c) we found that a radix-3 butterfly has 4 complex multiplications.
Also note that this formula is an upper bound, since some of the N twiddle factors in the v - 1
stages will be trivial. However, the formula is a good estimate.
9.41. (a)
N-1
X(k] = h"(k] ~)r[n]h"(n])h(k- n]

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-jw(1: 2 +2J:N+N 2 }jN

= 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]

(c) From the figure

xJkl
x{k) ---i(X}----!1 ~] y{k]

1\•
h [k]

we define the signals z1[k] and x2[k) to be

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,

y[k] = h"[k]:z:,[k] ke[N, ... ,2N-1]


N-1
= .-;d'IN L :z:[t]e-;•t'IN.,i•<•-tl'IN kE [N, ... ,2N- 1]
l..O
N-1
= L :z:[t]e-;•dtiN k [N, ... , 2N- 1] E

= '""'
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

= 3:: .,irr'IN .-•


.-=0 """

= 3::1 d"'' IN z-• r~1 (-e'"2r.IM z-M)t]

=
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

(e) The fiow graph for the system,

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,

Operation Complex Multiplications


x.[k] = x[k]i.•[k] N
Poles of H(z) fork= o, ... ,N -1 NM
Poles and branch exponentials of H(z) for k = N, ... , 2N- 1 2NM
y[k] = "'' [k]i.•[k] N
=3MN+2N
=3M3 +2M'
We are including the multiplicatinn of -1 and 1 &om the first branch here for simplicity.
Direct evaluation requires N' = M4 complex multiplications.
Note that since we are only interested in y[k] for k = N, N + 1, ... , 2N- 1, the initial delay of
2M2 is unneccessary; we have obtained all interestiDg output before the first delayed sample
appears.
388

Complex additions: The complex additions on the output side of each parallel branch do not
need to be computed until k ;:: N. Thus,

Operation Complex Additions


Poles of H(z) fork= o, ... ,N -1 NM
Poles and branch exponentials of H(z) fork= N, ... , 2N- 1 2NM
=3MN
=3M3
Direct evaluation requires N(N- 1) =M'(M2 - 1) complex additions.
9.42. (a) We separate the system function H(z) into two pieces; one corresponding to hR(n] = &{h(n]}
and another corresponding to hr(n] = /m{h(n]}.
1- W~z- 1
H(z) = 1- 2cos(27rk/N)z 1 + z-2
1- ws(21rk/N)z- 1
HR(z) = 1 - 2 ws(21rkjN)z 1 + z- 2
hR(n] = cos(27rlm/N)u(n]
sin(21rk/N)z- 1
Hr(z) = 1- 2ws(27rk/N)z-1 + z-2
hr(n] = sin(2.-kn/N)u(n]
Since :r(n] is real, the real and imaginary parts of X(k] = Y•(N] are computed using the following
fiowgraph.

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

cr~[NJ = <r., +<r~, :~:::>~(n]


n=O
389
N
crj[N] = .,;, + u!, L h~(n].
n=O

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:

<Tk = (N- 1)u2 = crj,


where ,.0 = 2- 28 /12. Fork= 0 there are no multiplies, and therefore u1. = 0 = crj.
~ 2 { 0 k=O
R = 171 = (N- 1)u' k#0

9.44. (a)

[X.,_,[q]HfNI = [X.,_,[qJIIHfNI = [X.,_,[q](


1 1
[X.,(p][ :5[X.,_,(p]( + (Xm-l(q]l < 2 + 2 = 1
implies that [Re{X,.(p]}l < 1 and [Im{X,.(p]}l <: 1. A similar argument holds for [X.,[q]l-
390

(b) The conditions are not suflicient to guarantee that overflow cannot occur.

IR.e{X~[p]}l :5 IR.e{X.,[p]}l + IRe{W.YX~_,[q]}l


:5 ~+ Ieos C;) Re{X.,_ 1 [q]}- sin C;) lm{Xm-t[q]}l

Consider the worst case, when r = 3Nf8. Then,


IR.e{X~[p]}l $ ~+I ~Re{X.,_.(q]} + ~Jm{Xm-t[q]}l
1 1 .
$2+.,!2,n.
Therefore, over&ow can occur.
9.45. (a) First, note that each stage has N /2 butterdies. In the first stage, all the multiplications are
~ = +1. In the second stage, half are +1 and the other half are wJ:I•
= -j. Successive stages
have half the number of the previous stage. In general,

Number of +1 multiplications in stage m =::.., m= l, ... ,v

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

Number of noise sources = (N /4) - 1, k odd


Number of noise sources < (N/4) -1, k even

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)~

~ ~ CT1 for large N


391

When the scaling is done at the input, an upper bound on the noise-to-signal ratio is found to be

t'(IF(k]l 2] < ~a~ 3 2 2 _ N 22- 28


t:[IX[k]I2J - Jlv = 4N as- 4
This equation is similar to Eq. 9.65, but scaled by ~. This implies that Eq. 9.65 is about 1 bit too
pessimistic. However, note that the N 2 dependence is still present.

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)~

9.46. The butterfly for decin3ation-in-frequency is drawn below.

X.,(q]

where

Xm(p] = Xm-l(p] + Xm-l(q]


Xm[q] = (Xm-1(p]- Xm-t[q])W,V

The statistical model is drawn below.


392

As in a decimation-in-time FFI', each output BOde connects to N - 1 buttedlies in a decimation-in-


frequency FFI', each oC which introduces a noise source whose variance is

Thus the output noise propagated through the 8owgraph is

£(1F[!:Jt'J = (N- 1)u1


since each noise source propagates along a unity gain path to the output nodes.
Thus, the results for the decimation-in-frequency FFT are identical to those for the decimation-in-time
FFT. This is true for both cases oC scaling either at the input of the FFT by 1/N, or at the input of
each stage of the FFT by 1/2.

9.47. Recall the following symmetry properties of the DFT:

:t[n] real = X[k] = X"[N- k]


:t[n] = :t•[N- n] = lm{X[k]}=O
:t[n] = -:t• [N - n] = Re{X[k]} = o
Thus, :t[n] real and even= X[k] real and even; and :t[n] real and odd= X[k] imaginary and odd.
(a)

111[n] = :t,[n] + :t3[n]

:t,[n] is real and even; x3[n] is real and odd.

Y,[k] = Re{Y1 [k]} + jlm{Y,[k]} = X,[k] + X 3[k]


Re{Y,[k]} is real and even; jlm{Y1 [A:]} is imaginary and odd; X 1 [k] is real and even; X3[k] is
imaginary and odd. It follows that

X 1 (!:] = Re{Y1 [A:]}


X,[A:J = jlm{Y,[A:]}

(b)

,.[n] = 111 (n] + ;,.[n]

Y3(k] = Y1 (k] + jY,(k] = [Re{Yi(A:]} + jlm{Y,[A:]}) + j(Re{Y,(k]} + jlm{Y.(k]}]


= Re{Yi(A:]} -1m{Y2 (k]} +j(Re{Y.(A:]} + !m{Y,[k]})
real odd even odd
Y3 (k] = (Ev{Re{Y,[k]}} + Od{Re{Y,(l:]}}] + j(Ev{lm{Y3 (A:]}} + Od{Im{Y3[k]} }]
393

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:]}]

+~[Im{Y,(k]}- Im{Y,[N- k]}J


= ~[Im{Y3 [k]} + Im{Y3 (N- k]}]
-~(Re{Y3 [k]}- Re{Y,[N- A:]}]
and from part (a)
1
X 1 (k] = 2[Re{Y,[k]} + Re{Y,[N- k]}]
1
x.[kJ = 2[Im{Y,[k]} + Im{Y,[N- k]}]
j
x,(kJ = 2[Im{Y,(k]} - Im{Y,(N- k]}]
-]
x.[kJ = T[Re{Y,[k]}- Re{Y,[N - k]}]

(c)

u,[n] = z,[((n + l))N]- z,[((n- 1))N]

u,[((N- n))N] = z,[((N- n + 1))N]- z,[((N- n- 1))N]


Since z,[n] = z,[((N- n))N],
u,[((N- n))N] = z,[((n- 1))N]- z,[((n + 1))N] = -u,[n]
For n = 0, we have u,(O] = -u,[O], which is only satisfied for u 3 (0] = 0.
(d) Using the circular shift property of the DFT we have

u,[k] = X3 (k].,;<••tNJ•- X3 [k]e-;<••tNJ•


= 2j sin(2rk/N)X3 (k]

(e) From part (a), since u,[n] is antisymmetric,

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

x1[A:J = 2I [R.e{Y>[k)} + Re{Y3 [N- k)})


1
x,[A:J = 2[lm{Ys[A:]} + Im{Y.[N- A:)})
x.[A:J = ~[lm{Ys[A:]}- Im{Y3[N- A:]}]/ sin(2.. k/N), k # O,N/2
x,[A:J = -~[R.e{Y3[k]}- Re{Y3[N- k)})/sin(2.-k/N), k # O,N/2
9.48. First, we find an expression for samples of the system function H(z ).

~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

Operation Complex Multiplications


FFTs of p(n] and q(n] 2[(L + M)/2] log2 (L + M) =(L + M) log,(L + M)
=
R(k] P[k]Q[k] L+M
Inverse FFT of R(k] [(L + M)/2]1og2 (L + M)
= (3(L + M)/2] log,(L + M) + (L + M)
Since a complex multiplication is computed using 4 real multiplications, the number of real multi-
plications required by this technique is 6(L+ M) log2 (L+ M) +4(L+ M). Plugging in some values
for (L + M) = 2", we find
L+M Direct FFT
2 4 20
4 16 64
8 64 176
16 256 «8
32 1024 1088
64 4096 2560
Thus, for (L + M) 2: 64, the FFT approach is more efficient.
(c) The binary integers u and v have corresponding decimal values, which are
L-1
"decimal = L ,..2•
M-1
vdecimal = L Vi2i
i=O

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

# real multiplications = 6(8000 + 1000) log,(8000 + 1000) + 4(8000 + 1000)


= 7.45 X lOS
In contrast, the direct computation requries 8.1 x 107 real multiplications.
396

(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 + N) = GN[a + N) + SN[a + N)


= cos(2.-a/N+2,-)+sin(2.-a/N+2.-)
= cos(21ta/N) +sin(2ra/N)
= CN(a] + SN[a]
= HN[a]

HN[a +b) = CN(a + b] + SN[a+b]


= cos(2,-afN + 2lrb/N) + sin(2..ajN + 21rbjN)
= cos(21tafN) cos(27rb/N) - sin(2,-a/N) sin(2lrb/N)
+ sin(21ta/N) cos(2.-b/N) + cos(2ra/N) sin(2,-b/N)

Grouping the terms in the last equation one way gives us

HN[a + b] = [cos(2lra/N) +sin(2,-a/N)]cos(2,-bfN)


+[cos( -21ta/N) +sin( -27ra/N)] sin(2,-b/N)
= HN[a)GN(bJ + HN(-a)SN(b]
while grouping the terms another way gives us

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

Now since HN[2rk] = HN;2 [rk], we have


(N/2)-1 (N/2)-1
XH(k] = L :t[2r}HN;2(rk] + L z{2r + l]HN;z[rk)CN(k]
r=O. r=O
(N/2)-1
+ L z{2r + l}HN;z[((-rk))N;2)SN(k)

= F(k] + G[k]CN(k] + G[(( -k))N;2)SN[k)


397

where
(N/2)-1
F[A:J = L z[2r]HNt 2 [rkJ
r=O

is the N /2-pomt DHT of the e\'ell-mdexed pomts and


(N/2)-1
G[A:] L
= ....., z[2r + 1]HN/2[rk]

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

where • denotes conjugation. Drawing the tlow graph, we get

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

F(lF, = 21 Fl1F2 = 2I FfF, = 21

T{'T, =I TfT2 =I

(1/N)Q 8 Q = (1/N) (F{IT{'FfTfFf) (F,T2F2T,F,)


= (1/N)(NI)
= I
where N = 8 in this case.
9.52. (a) First, we derive the circular convolution property of the DFT. We start with the circular convo-
lution of z[n] and h[n].
N-1
y[n] = L z[m]h[((n- m))N]
m=O
Taking the DIT of both sides gives
N-l /11-l
Y[k] = L L z[m)h(((n- m))N)W"'
...0 m=O
N-1 N-1
= L :z:[m) L h{((n- m))N)W"'
m=O .....
399

Using the circular shift property of the DFT,


N-1
Y[k] = L :r[m]H[k]WN'•
m=O
N-1
= H[k] L :r[m]WN'•
= H[k]X[k]
Next, the orthogonality of the basis vectors is shown to he a necessary requirement for the circular
convolution property. We start again with the circular convolution of :r[n] and h[n]
N-1
y[n] = L :r[m]h[((n- m))N]
Substituting the inverse DFT for :r[m] and h[((n- m))N] gives

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

With orthogonal basis vectors,


Y:
n=O
WN' = { N,
0,
k =0
k ;< 0
so the right-most summation becomes
N-1 N-1
L WN.tlmwt2m = 2: w;:2-i:J)m
m=O m=O
= No[k1- k,J
Therefore,
N-1 N-1
y[n] = ~ L L (X[k1]H[k Jw;••• N o[k1 - k
2 2 ])
il=Ol:2=0
N-1
= ~ L X[k,]H[k1]w_N•··
•·=<>
N-1
= L :r[m]h[((n- m))N]
m=O

Therefore, the circular convolution property holds as long as the basis vectors are orthogonal.
(b)
400

k = 0: ((I+ I+ I+ I))17 = ((4)),7 = 4


k = . I: ((I+ 4 + I6 + 64))17 = ((85))17 = 0
k = 2: ((I+ I6 + 256 + 4096)),7 = ((4369))17 = 0
k = 3: ((I + 64 + 4096 + 262I44)),7 = ((266305))17 = 0
The relation holds for P =I7, N =4, and WJV =4.
(c)

X[k] = ( (~ z[n]4n>))
17
= ((I· I+ 2 · 4• + 3 · I6• + 0 · 64•))17
= ((I +2·4• +3·16•))17

Using this formula for X(k),

X[O) = ((I+2+3)h7 = ((6)h7 = 6


X[1] = ((1 + 2. 4 + 3. I6))17 = ((57))17 = 6
X[2) = ((I+ 2 · I6 + 3 · 256))17 = ((801))17 = 2
X[3) = ((1 + 2. 64 + 3. 4096)),7 = ((124I7)),7 = 7

H[kJ (t
= ( h[nW•))
17
= ((3·I + 1·4· +0·I6. +0·64·))17
= ((3+4.)h7
Using this formula for H[k),

H[O] = ((3+I))17 = ((4))17 = 4


H[I] = ((3 + 4))17 = ((7))17 = 7
H[2) = ((3 + I6))17 = ((I9)h7 = 2
H[3] = ((3 + 64))17 = ((67)),7 = I6

Multiplying terms Y[k] = X[k]H[k] gives

Y[O] = ((24)),7 = 7
Y[1] = ((42)),7 = 8
Y[2] = ((4)h7 = 4
Y(3] = ((112))17 = IO

(d) By trying out different values,

N- 1 = 13
w;;' = 13
((N- 1 N)h7 =((W;i WJVJl11 =((13 · 4))17 =((52)),7 = 1
1
401
(e)

y(n] = ( ( 13 ~ Y(k] 13,..))


17
= ((13(7 ·1 + 8 .13n + 4 · 169n + 10 · 2I9r])),7

Using this formula for y(n],

y(O] = ((13(7 + 8 + 4 + 10]))17 = ({377)),7 = 3


y(l] = ((13(7 ·1 + 8. 13 + 4. 169 + 10. 2197])),7 = ((29584I)h7 = 7
y(2] =
((13(7. 1 + 8 ·169 + 4. 28561 + 10. 4826809])),7 = ((628988009))17 = 11
y(3] = ({13(7 . 1 + 8 . 2197 + 4 . 4826809 + 10. 10604499373]) h7 = ({1378836141137)) 17 = 3
Performing manual convolution y(n] = z(n] • h(n] gives
y(OJ = 3
y(l] = 7
y(2] = 11
y(3] = 3
The results agree.
9.53. (a) The tables below list the values for n and k obtained with the index maps.

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

As shown, the index maps only produce n = 0, ... , 5 and k = 0, ... , 5.


(b) Making the substitut>on we get

X(k] = X(k 1 + 2k2 ]


= L• z[nJwi••+•••Jn
2 I
= L L :r:{3nl + n2]Wi.kt+2l::z)(3na+n:z)
n:z=Ont=O

(c) Expanding out the We terms we get

(d) Grouping the terms we get

The interpretation of this equation is as follows


402

(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

(e) The signal flow graph looks like

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.,

n = n 1 + 2n2 for n 1 = 0, 1; n2 = 0,1,2


k=3k,+k2 for .1: 1 = 0, 1; lt2 = 0, 1, 2
405

Solutions - Chapter 10

Fourier Analysis of Signals

Using the Discrete Fourier Transform


407

10.1. (a) Using the relation


2..k
0 • =NT'
we find that the index k = 150 in X[k] corresponds to a continuons time frequency of
- 2.. (150)
- (1000)(10 4 )
= 2.. (1500) rad/s
=
(b) For this part, it is important to realize that the k 800 index corresponds to a negative continuous-
time frequency. Since the DFT is periodic in k with period N,
2.. (800 - 1000)
= 1000(10-4)
= -2 .. (2000) rad/s

10.2. Using the relation


n. = 2.-k
NT
or
k
f• =NT
we find that the equivalent analog spacing between frequencies is
I
t:.j =NT
Thus, in addition to the constraint that N is a power of 2, there are two conditions which must be met:
j. > 10, 000 Hz (to avoid aliasing)
* < 5Hz
These conditions can be expressed in the form
(given)

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

L = ( 16, ooo sa:les) (20 x 1o-• sec)


= 320 samples
(b) The frame rete is the number of frames of data processed per second, or equivalently, the num-
ber of DFT computations done per second. Since the window is advanced 40 samples between
computations of the DFT, the frame rate is

frame rate = ( 16 000 samples) ( 1 frame processed)


' sec 40 samples
= 400~
sec
408

(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

(b) Since an N -point DFT is periodic in k with period N, we know that


X[SOO] = 1 +j
implies that
X[-200] =1 + j
Using the relation

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

10.5. (a) After windowing, we have


z(n] = cos(l1oTn)
= ~ [.,iCioTn + .-;floTn]
= ~ [ei~<"l!:r)n + .-;lf<"l!:r)n]
for n = 0, ... , N - 1 and z(n] = 0 outside this range. Using the DIT properties we get

X(kj = ~ o(({k- N~:T))N] + ~ o(({k + NV;T))N]


If we choose
T=~ko
Nl1o
then
X[k] = ~ o[k - kol + ~ o[k- (N- ko)J,
whlch is nonzero for X[ko] and X[N- ko], but zero everywhere else.
(b) No, the choice forT is not unique since we can choose the integer ko.
10.6. Since :r(n] is real, X[k] must be conjugate symmetric.
X[k] = X'(((-k))N]
X(((-k))N] = X'(k]
Therefore,
X(((-900)hooo] = (!)'
X(!OOJ = I
X(((-420)hooo] = (5)'
X(580] = 5
Note that the k = 900 and k = 580 correspond to negative frequencies of !l. Since the DIT is periodic
in k with period N, we use k = 900 - 1000 = -100 and k = 580 - 1000 = -420, respectively, in the
equations below.

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

10.9. The rectangular window's mainlobe is


411" 4.-
6"'"" = T = 64
radians wide. The difference in frequency between each cosine must be greater than this amount to be
resolved. If they are not separated enough, the mainlobes from each cosine will overlap too much and
only a single peak will be seen. The separation of the cosines for each signal is

.o.w, = I~- ::1


1
= ~
.o.w, =1~- ::1=~ 2

.o.w, = li- ::1 = ~ 2

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

10.12. The equivalent frequency spacing is


1
t.J =NT
Thus, the minimum DFI' length N such that adjacent samples of X[k] correspond to a frequency spacing
of 5 Hz or less in the original continuous-time signal is

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

Xo[k] = 1/o[((k- 3))..] + 1/o[((k + 3))..]


= lSO[k - 3] + lSO[k - 33]

\\'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

X,[k] = ~"o[((k- 9lh•l + 1/o[((k + 9)) ..]


= lSO[k- 9] + lSO[k- 27]

Putting it all together we get

18(o[k- 3] + o[k- 33]), r=O


=
x,[kJ
1 18(o[k- 9] + .S[k- 27)), r=l
0, otherwise

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.

10.15. The instantaneous frequency of the chirp signal is

w;[n] = w0 + >.n
This describes a line with slope >. and intercept Wo· Thus,

>. = t.y = (0.5.-- 0.257r) = 41 34 10-· rad


l!.:t {19000 O) . X

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

10.17. We should choose Method 2.


414

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]

The rectangular window's transform is

W(eiw) = sin{16w) .-;w3l/2


sin(w/2)
In order to label V(eiw) correctly, we must find the mainlobe height, strongest sidelobe height, and the
first nulls of W(e'w).

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

Strongest Sidelobe height = 0.2239(32) "' 7.2

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

>.[n] = ! ["; + 1000sin(~)]


= 4" + scos
" (>rn)
8000
>.[n]
271" = !+.!.cos(~)
8 16 8000
Once A[n]/2,. is known, it is simple to sketch the spectrogram, shown below.

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)

X(ei"') = f f (;f + ;2;)


T=-CIC
Xc

Over one period, as5llllrlng no aliasing, this is

for -:If S. w :s; "


which is equivalent to
X(ei"') = { +x. (i'f), for 0 S w < .-
+x. (j~), for.- :s;w < 2"

Since the DFT is a sampled version of X(ei"'),

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

Method 2: This time we plug in values of N = 4000, and T = 50J.IS to find


x.[l:J = { 20, ooox, U2.-. 5 ·A:), foro 5 A: 5 t999
20, OOOX, U2.- · 5 · (.I: - 4000)) , for 2000 $ l: $ 3999
Therefore, we see this does provide the desired samples. A slretch is provided below.

---1 f- Af=5Hz

Method 3: Noting that Z3[n] = z2[n] + z2 (n- ~].we get


x.[kJ = x.[kJ + (-t)• x.[A:J
X [l:] = { 2X2 [k], fork e~en
3
0, otherwiSe

40, OOOX, U2.. · 5 ·A:), for k even, and 0 $ k $ 1999


X,[.l:] = 40,000X, U2.. · 5 · (l:- 4000)), fork even, and 2000 $ k $ 3999
{ 0, otherwise
This system provides the desired samples only for k an even integer. A sketch is provided below.

---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

Over one period, assuming no aliasing, this is

for -1r ~ w :5 1r

which is equivalent to

Since the DFT is a sampled version of X(eiw),

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].

The effective frequency spacing is


2>r
~n = NT
2lf
= (1000)(1/20,000)
= 2"(20) radfs
(b) Next, we determine if the designer's assertion that

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

250 500 750 999k

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(~) = !w(~l 2 ) + !w(ei<w-••>1•)


2 2
which looks like
422

Y(ei"')

~~~~~~~~~~~~~~~~~w

It 2lt 31t 4lt

Y[k] is equal to samples of the Y(eiw)


Y[k]= Y(eiw)lw=..k/N
= ~w (ei'lH) + ~w (e>~C¥l)
Now putting all that we know together, we see that fork = 0, 1, ... , 500, Y[k] is related to X,(jO)
as follows.
i-tX,(j2" · 10 · k), k even, k y! 500
=
Y[k]
1j.X,(j2,. · 10 · k),
~W(eid/Nj + olJ.W(ej•(k-N)/N)
k =
500
k odd
In other words, the even indexed DFT samples are not aliased, but the odd indexed values (and
k = 500) are aliased. The designer's assertion is not correct.

10.25. (a) Starting with definition of the time-dependent Fourier transform,


00

Y[n, ~) = L y[n + m]w[m]e-;•m


m=-oo

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

Y[n, >.) = .-i>•l~ h[k]X[n- k, >.)]


= e-i>n l~ h[k]ei(•-•)> X[n- k,>.)]
M
= L h[k]e-i» X[n - k, >.)
>=0

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

/(w) = L~ m=~- 1 ) [~ v[n]v[n + m]] e-1wm


L-1 L-1
= L~ L v[n] L v[n + m]e-iwm
n=O m.=-(L-1)

Let l = n + m in the second summation. This gives


l L-1 n+(L-1)
I(w) = LU L v[n] L v[l]e-iw(l-n)
n=O l=n-(L-1)
L-1 n+(L-1)
= L~ L v[n]ei"'" L v[l]e-iwl
A=O t=n.-(L-1)

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

10.28. (a) Let


1
wR(m] = v'M(u[n]- u[n- M])

be a scaled rectangular pulse. Then we can write the aperiodic autocorrelation as,

ws(m] = L"" wx[n]wx[n + m]


n.=-oo

""
= L wx(k- m]wx[k]

= L"" wx[k]wx[-(m- k)]


.t=-oo
= wx[m]• wx[-m]
The convolution above is the triangular signal described by the symmetric Bartlett window formula.
This is shown graphically below for a few critical cases of m.
Consider m = -(M- 1). This is first value of m for which the two signals overlap.

m =-(M- 1) case

0 WR(k) M-112

•. wR[k+(M-1))

k
T
k=-(M-1)

At m = 0, all non-zero samples overlap.


m=Ocase

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

The final result of the aperiodic autocorrelation is

T m
m=-(M-1)

Stated mathematically, this is

1-lmi/M, lmi:SM-1
I J
wsm ={ 0, otherwise

(b) The transform of the causal scaled rectangular pulse WR(n] is

W (.Jw) = 1 sin(wM/2) ,-,w(M-1)/2


R ..fMsin(w /2)
426

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].

w[n] = L"" :[m]:[n + m]

The signal w[n] will have a no1111egative Fourier transform.


10.29. (a) Rectangular: The Fourier transform of the rectangular window is given by
M-1
WR(el"') = L (1)e-jwm
m=-(M-1)

Let n = m + (M- 1). Then, m = n- (M -1), and


2(Jl-1)
WR(el"'} = L .-;. .(n-(M-1)]
n=O
2(Jl-1)
= e-i'"'<.U-1) L e-;~.~n
n=O

Using the relation

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

which is the convolution of a rectangular signal,


x[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

that it is scaled by 7i; and has a length 2M; 1 + 1 = M. Therefore,

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

HannjngfHammjng: Starting with

wH[m] = (o + ,8 cos[nn/(M- 1)]) wR[m)


UIH(mj = (o + ~.,;•m/(M-1) + ~e-i"""/(M-1)) WR(m)

We take the Fourier transform to find

WH(ei"') = oWR(ei"') + ~ {WR(eilw-•/(M-1)]) + WR(eilw+•/(M-lllJ)


= 0
sin[w (M- tlJ ~ [sin((w- -.f:r)(M- !JJ]
sin(w/2) +2 sin((w- ...:,J/2)

+ ~ [sin((w + Jf.:r)(M- lll]


2 sin((w + Ji_ 1 )/2]

-
(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

W (eiw) = sin[w(M- ~)]


R sin(w/2)
The numerator becomes zero when the argument of its sine term equals rn.
(2M -1)w
2 = '"'
2'"'
w = 2M-1
Plugging in n = 1 gives us half the mainlobe bandwidth.
21r
~ Mainlobe bandwidth = 2M-1
41r
Mainlobe bandwidth =
2M-1
2..
Mainlobe bandwidth :::
M
(M-1)
1
F = Q L 2
w (m]
m:-(M-1)
1
= -(2M -1)
Q
2M
:::
Q
Bartlett {triangular): The approximate mainlobe width, and the approximate variance ratio,
F, for the Bartlett window are found below for large M.

In part (a), we found the Fourier transform of the Bartlett window as


2
W (eiw) = _!_ [sin(wM/2)]
B M sin(w/2)
The numerator becomes zero when the argument of its sine term equals ..n.
wM
2 = '"'
2m
w = M
Plugging in n = 1 gives us half the mainlobe bandwidth.
1 2r
Mainlobe bandwidth = M
2
4..
Mainlobe bandwidth
= M
To compute F, we use the relations
M-1
M(M -1)
:[m =
2
M-1
M(M- 1)(2M- 1)
:[m• = 6
m=O
430

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,

~ [sin[(w- 1f=;)(M- !>J]


2 sin[(w- u"_ 1 )/2]

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+ Jl=r)(M -!)J]


2 sin{(w + u"_ 1 )/2]
is also 2wfM. Note tbat the peak 'Value for this term occurs at a frequency w:: -rfM.

Finally, 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.

Thus, for large M, the mainlobe bandwidth is bounded by


2.. . 41<
M < Mainlobe bandwtdtb < M

Therefore, a reasonable approximation for the mainlobe bandwidth is


. • 31<
Mainlobe bandwtdtb o:: M

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 )]

Using the relation


1 1
cos2 8 =- +- cos28
2 2
we get

F = -1
Q
[
L
Al-l

=-<M-1)
o 2 +2o/3
m=-(M-1)
L
Al-l
cos
(
~1
M-
)

[j• Al-l {J' Al-l ( 21m1 )]


+2 L:
=-(Al-l)
<1>+2 L:
--(Al-1)
cos M=I .

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]e-j(O./T)•m + L• :r[m + 7]e-i(2•/7)>me-i2d


m=O m=O


= 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

we see that :[m] + :r[m + 7] = I for 0 :S m :S 6. Thus,

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=?

= L• (:(n + m) + :(n + m + 7J)e-iC••fT)>m


m=O

With n ;o: 0 we have·:r[n + m) + :[n + m + 7] = I for 0 :S m :S 6, and so


X[n,k) = l>J'T{l}
= 7o(kJ
Therefore, for 0 $ n $ co we have

I:x[n,k] =

I:7o[kJ
.1:-=0 A:=O
= 7
10.31. (a) Sampling the continuous-time input signal
:z:(t) = e"(h/l)tO't
with a sampling period T = w-• yields a discrete-time signal
:z:[n] = :z:(nT) = e"3•n/l
In order for X,.[k] to be nonzero at exactly one value of k, it is necessary for the frequency of the
complex exponential of :z:[n] to correspond to that of a OFT coefticent, '"• = 27rkfN. Thus,
3.- 2.-k
8 = N
N = l6k
3
The smallest value of k for which N is an integer is k = 3. Thus, the smallest value of N such that
Xw[k] is nonzero at exactly one value of k is
N = 16
(b) The rectangular windows, w,[n] and w,[n], dil!er only in their lengths. w,[n] has length 32, and
w,[n] has length 8. Recall that compared to that of a longer windov.·, the Fourier transform of a
shorter window has a larger mainlobe width and higher sidelobes. Since the OFT is a sampled
version of the Fourier transform, we might try to look for these features in the two plots. We notice
that the second plot, Figure Pl0.31-3, appears to have a larger mainlobe width and higher sidelobes.
As a result, we conclude that Figure Pl0.31-2 corresponds to w1 [n], and Pl0.31-3 corresponds to
w,[n].
(c) A simple technique to estimate the value of Wo is to find the value of k at which the peak of jX,.[k]l
occurs. Then, the estimate, is

The Corresponding value of fio is


- 2.-k
l1o =NT
This estimate is not exact, since the peak of the Fourier transform magnitude jX,.(e'w)l might
occur between two values of the OFT magnitude IX,.[k]l, as shown below.

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,

sin((21rk/N- .,.)N/2] -;(0.4/N-wo)(N-1)/2


X,.[k] = sin[(27rk/N- .,.)/2] e

Note that the phase of X,.[k], using the above equation, is

(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

Re-expressing the equation for LX.,[k], we find


2(LX,.[k] - m.r) 2.-k
wo= N-1 +N
Let X1,.[k] be the OIT of the 32-point sequence :r1.,[n] = :r[n]w.[n], and let X2.,[k] be the OFT of
the S-point sequence :r,,.[n] = :r[n]w2 [n]. Note that the kth OFT ooeflicient of X,,.[k] corresponds
to x ... [4k]. Thus, we can relate the 8 OIT coeffients of X 2,.[k] to 8 of the DIT ooeflicients in
x ... [k]. Using the k = Oth OFT coefficient for simplicity, we find
= 2(LX,..(OJ- m.r) _ 2(LX.,,[O]- p1r)
""' 32-1 - 8-1
LX,.1[0]- m.- LX,.,[OJ- p1r
= 15.5 - 3.5
A solution that satisfies these equations, with m and p integers, will yield a precise estimate of
"'<>· We can accelerate solving these equations by determining which values of m and p to check.
This is done by looking at the peak of IX.,(k]l in a procedure similar to Part (c). Suppose that
the indices for two largest values of IX,.[k]l are k.n;n and k,... Then, we know that the peak of
IX(e'w)l will occur in the range
2,- k,;n 21r k,..
--<·-~<---
N --u- N
By re-expressing the equation for LX,.[k], we see that

m,.;n = ;1 [ 2LX,.1 [k.nin J + 2ll"k,.;n Wo


( ~- • ) (N -1) ]

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

X[n, >.) = L :r(n + m]w[m]e-i>m.


m=-oc

(a) Linearity: using :r[n] =az1(n] + b:t2 [n],


X[n, >.) = L :r[n + m]w(m]e-i>m

00

= L (az.(n + m] + b:t,[n + m])w[m]e-i>m


m=-oc
00

:r.[n + m]w[m]e-i'"' + b L :r,[n + m]w[m]e-i>m


m=-oo
= aX1 [n, >.) + bX,[n, >.)
(b) Shifting: using y[n] = :r[n- no],

Y[n, >.) = L
.. y[n + m]w[m]e-i-'m
m=-oo
436
00

= L :[n - fl<l + m)w[m)e-;>m

= X[n - fl(), ).)

(c) Modul4tion: using y[n) = .;... n:[n),


y[n + m) = .;wo<n+ml:[n + m)
00

Y(n, ).) = L y[n + m)w[m)e-;>m


m=-oo

L
00
= .;wo(n+mlz[n + m)w[m)e-;>m
m.=-oo

L .;... n:[n + m)w[m]e-;<>-wo)m


00
=
m=-oo

(d) Conjugate Symmetry: for z[n) a.nd w[n) real,


00

X[n, ).) = L z[n + m)w[m)e-;>m

= L.~oo :[n + m]w[m]ei'"'] •


= [X[n, -).)]'
= X'[n, -).)
10.33. (a) We are given that <Pc(T) = C {zc(t)zc(t + T)}. Since z[n] = Zc(nT),
¢[m] = C {:[n]z[n + m]}
= C {zc(nT)z0 (nT + mT)}
= <Pc(mT)
(b) P(w) a.nd Pc(fl) are the transforms of ¢[m] a.nd <Pc(T) respectively. Since ¢[m] is a sampled version
of <Pc(T), P(w) a.nd Pc(fl) are related by

(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

(a) Computing the autocorrelation function,

¢,,[m] = £ {:(n]:(n + m]}


= £{(A cos(won +e)+ e(n]) (A cos(wo(n + m) +e)+ e(n + m])}
= £ {A2 cos(won+ e) cos(wo(n + m) +e)}
+ £ {Ae(n] cos(wo(n + m) +e)}+£ {Ae(n + m] cos(won +e)}
+ £ {e(n]e(n + m]}
= A2 £ {cos(won + 8)cos(wo(n + m) + 8)}
+A£ {e(n]} £ {cos(wo(n + m) + 8)} +A£ {e(n + m]}£ {cos(won+ e)}
+ £ {e(n]e(n + m]}
First, note that
cos( a) cos( b) = 21 cos( a+ b)+ 21 cos(a- b)
Therefore, the fust term can be re-expressed as

A2 £ {~cos (2won + w0 m + 211) + ~ cos(wom)}


Next, note that
£ {e(n]} = 0
As a result, the two middle terms drop out. Finally, note that since e(n] is a sequence of zer<>-mean
variables that are uucorrelated with each other,

£ {e(n]e(n + m]} = cr~6(m], where cr~ = £{ e2 [nl}


Putting this together, we get

Since 2~ 1:• cos(2won + wom + 211)d8 = 0, we have


A•
¢,,[m] =
2 cos(wom) + cr~o(m]

(b) Since the Fourier transform of cos(wom) is ..O(w- wo) + .-o(w + wo) for !w! ::; "•

10.35. (a) Plugging in the equation


!(.1:] = I(w•) = L1 JV[k)l 2
into the relation
var(!(w)]::: .P;,(w)
we lind that

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

P•• (w) = <r.


This gives
var (IX(k]J 2 ] = N 2u!
(b) The cross-correlation is found below.
N-1 N-1
C {X(k]X.(r]} = L L C {:z:(nl]:z:(n•J} W~"'W,Y""'
n1=0 nt=O
N-1 N-l
= L L <r.o(nl - n2]w;n• W,Y""'
n1cOn,=O
N-1
""' -'w<t-r)n
= ~Ui
...... N

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

Q = 10 seconds . 20, 000 samples


second
Q = 200,000 samples

(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

Using the relation

we find

E {¢[ml} = ~ ~· [-1-1• (6)C~~(_,j(.,-Bl)d8]


2" -• 2" LU -•
P•• .,;"'mdw

= _1_1. [.!..f"
2" LU -•
P•• (e)
2r -•
c_(.,;I.,-Bl)_,j"'mdw] d8

Substituting w' =w - 6 in the inner integral yields

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

E {¢[m]} = _1_1• P•• (e).,;•m [.!..j• c.~(e'"'')_,j"''mt~w'] d8

=
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

which is a time aliased version of E {~•• [ml}.


(c) N should he chosen so that no time aliasing occurs. Since ~•• [m] is 2L -1 points long, we should
choose N ~ 2L.
10.38. (a) For 0 :S m :S M,
Q-m-1
~.. (m] = ~ L z(n]:r(n + m]
.....
I [M-1 2M-1 KM-1 ]
= Q L :r(n]:r(n + m] + L z(n]:r(n + m] + ... + L :r(n]z(n + m]
'J\=0 A=M n=(K -l)M

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.

Note that the circular correlation of z,[n] with !f<[n],


N-1
c_.[m] = L z,[n]y,[((n + m))N]
n=O

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

Thus, the minimum value of N is 2M - 1.


(c) A procedure for computing ~•• [m] is described below.
step 1: Compute X,[k] and Y;(k], which are the N ;<: 2M- 1 point DFTs of :r,[n] and y,[n].
=
step 2: Multiply x,[k] and Y;"[k] point by point, yielding C,[k] G\[k] x,[k]Y,•[k]. =
step 3: Repeat the above two steps for all data (K times), then compute
. 1 K-1
••• [k] = Q L
C,[k] forO~k~N-1
i=O

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

2 · ~ log2 N · K = KN1og2 N, for step 1


KN, for step 2
N, for divide by Q operation in step 3
If log2 N for step 4
So the total number of complex multiplications is {K + f)Nlog 2 N + {K + l)N.
(d) The procedure developed in part {c) would compute the cross-correlation estimate~•• without any
major modifications. All we need to do is redefine 1/<[n] as

y,[n] = y[n + iM],


and :t;[n] is the same as it was before, namely

[n1= { z[n + iMJ,


O$n$M-1
~~
0, M$n$N-1

Note that form < 0, ~•• [m] = ~.. [-m].


(e) For N =2M,

yi[n] = z[n + iMJ, for 0 S n S 2M - I


= z[n + iM](u[n]- u[n- MJ) + z[n + iM](u[n- M]- u[n- 2M])
= z[n + iMJ(u[n]- u[n- MJ) + z[n- M + (i + 1)M](u[n- MJ- u[n- 2M])
= :t;[n] + Zi+t[n- M]
Taking the DFT of this expression yields

Y;[k] = X;[k] + (-1)•x,+l[k]


A procedure for computing~•• for 0 S m S M- 1 is described below.
step 1: Compute theN point DFT X;[k] fori= 0,1, ... ,K.
step 2: Compute Y;[k] = X;[k] + (-1)• X;+l[k] fori= 0, 1, ... , K- 1.
step 3: Let Ao[k] = 0 and compute

A;[k] = A;- 1 [k] + X;[k]Y;"[k], i = 1, ... , K- I


step 4: Deline V[k] = AK-l[k]. Compute 11[m], theN point inverse DFT of V[k].
step 5: Compute
• 1
4J•• [m] = "Q"[m]
Assuming that a rad.ix-2 FFT, requiring ~ log2 N complex multiplications is used to compute the
forward and inverse DFTs, the number of complex multiplications is
(K + 1)~1og2 N, for step I
0, for step 2
(K -l)N, for step 3
~log2 N, for step 4
N, for the divide by Q in step 5
So the total number of complex multiplications is Kfl N log2 N + K N. Note that for large N and
K, this procedure requires roughly half the number of complex multiplications as the procedure
described in part (c).
10.39. (a) Using the relations,

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>.

= ~ /_: X[n, >.)X[n, ->.)ei'md>.


= ;.. i: (~00 x[n + l]w[l]e-;>J) Ct- 00
x[n + r]w[r]ei'r) ei'md>.

= f. f. x[n + l]w[l]x[n + r]w[r] ( 2~


l=-oor=-oo
/_: .-i>J.,J>rei>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>.

Using the Fourier transform relation,


o[n- no];..--; .-;wn,
we find
00 00

c[n, m] = L L x[n + l]w[l]x[n + r]w[r]o[m -I+ r]


l=-oor=-oo

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

c[n, m] = L :r[n + m + r]w[m + r]:r[n + r]w[r]


r=-oo
00

= L
r=-oc
x[n + r]w[r]:r[n + m + r]w[m + r]

(b) First, note that

IX[n, >.)1 2 = X[n, ->.)X[n, .>.)


= IX[n, ->.)J'
Starting with the definition of c[n, m],

..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>.

we substitute .>.' = ->.to get

c{n,-m] = __!_ /.-•IX[n, ->.')12 .,J>'md>.'


2.- •
= 2.1• 211' -r
IX[n,->.')1 2 ei>'md>.'

= 2.1• IX[n, >.')12 .,i>'md>.'


271'" -·
= c(n,m]
Thus, the time-dependent autocorrelation function is an even function o' Jll fur n fixed. Next, we
use this fact to obtain the equivalent expression for c{n, m].
00

c[n, m] = L :r[n + r]w[r]z[m + n + r]w[m + r]


00

= L :r[n + r]w[r]:r[-m + n + r]w[-m + r]


r=-oo

Substituting r' = n + r gives


00

= L :r[r']w[r' - n]:r[r'- m]w[(r'- m)- n]

00

= L :r[r']:r[r' - m]w[r' - n]w[-(m + n- r')]


'1""'=-oo

= L"" :r[r'J:r[r' - m]h.n[n - r'J


r';-cc

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

w[-r] = { 0,a' ' r;::o


r<O
into hm{r] = w[-r]w[-(m + r)J, we find
r ~O,r ~ -m
h,.[r] = { 0, otherwise
Taking the z-transform of this expression gives

Hm(z) = L"" h,.{r]z-•

00
= L:o2r+mz-r
..... 00

= am L (a z- 2 1
)'
......,
445

Again we have assumed that m is positive. H [zl > a2, then


a'"
H,.(z) =
1- a2 z 1

h,.[r] = a'"6[r] + a 2 h,.[r- 1]


Using this in the equation for c[n, m] gives
00

c[n, m] = L z[r]z[r - m]h,.[n - r]


r=-oo
00

= L
r=-oo
z[r]z[r- m] (c'"6[n- r] + c 2 h,[n- r - 11)
00

= a'"z[n]z[n- m] + a 2 L z[r]z[r - m]h,[n- r - 1]


r=-oo
= c'"z[n]z[n- m] + c 2 c[n- 1, m]
A block diagram of this system appears below.

x{n] ----.,----~----\ir----.,-------+ c[n,m]

z-m

(e) Next, consider the system

[ ] -{ ra', r;o:O
w-r-
0, r <0

h,.[r] = {ra' u[r]} {(r + m)a'+'"u[r + m]}


r 2:: O;r 2:: -m

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

Again we have assumed that m is positive. Thus,


446

a"'+2 z- 1 (1 + a 2 z- 1 + m - ma2 z'- 1 )


= ~1- ,.zz-1)3
a"'+2 (1 + m)z- 1 + a"'+4(1- m)z- 2
= 1-3a2 z- 1 +3a•z-•-a6 z 3
Cross-multiplying and talring the inverse z-tra.nsform gives
h,.[r]- 3a2 h,.[r- 1] + 3a4 h,.[r- 2]- a6 h,.[r- 3] = a"'+2(1 + m)6[r- 1] + a'"+4(1- m)6[r- 2]
h,.[r] = 3a2 h,.[r- 1]- 3a4 h,.[r- 2] + a6 h,.[r- 3] + a"'+ 2 (1 + m)6[r- 1] + a'"+4(1- m)6[r- 2]
Using this relation for h,[r] in
00

c{n, m] = L
r=-oo
:r[r]:r[r- m]h,[n- r]

we get
00

c{n, m] = L :r[r]:r[r- m] (3a2 h,[n- r - 1]- 3a4 h,[n- r - 2] + a6 h,.[n- r - 3])


r=-oo
00

+ L z[r]:r[r- m] (a"'+2 (1 + m)6[n- r - 1] + a'"+<(1- m)6[n- r - 2])


r=-oo
= 3a2 c{n- 1, m] - 3a4 c[n- 2, m] + a6 c[n- 3, m]
+a'"+ 2 (1 + m):r[n - 1]z[n- 1 - m] + a"'+4(1 - m)z[n- 2]z[n- 2-m]
A block diagram of this system appears below_

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

10.40. (a) Looking at the figure, we see that


X[n, >.) = {(:r[n]e-;"") • ho[nl} .,;>.n
= [ f: :r[n- m]e-;>.(n-m)ho[m]] .,;>.n
--oo
00

= L :r[n - m]ho(m].,;>.m
m.-=-oc

Let m' = -m. Then,


-oo
X[n, >.) = L :[n + m1ho[-m'Je-;>.m'
m•=oo
447
00
= L z[n + m']ho[-m1e-;>m'
m'=-oo
= X[n,>.)
if ho[-m] = w[m]. Next, we show that for>. fixed, X[n,>.) behaves as a linear, time-invariant
system.
Linear: Inputting the signal oz 1 (n] + bz 2[n] into the system yields
00

L (oz,[n + m] + bz2[n + m]) ho[-m]e-;>m =


--00
00 00

L oz,[n + m]ho[-m]e-;>m + L bz2[n + m]ho[-m]e-;>m = aX1 [n, .>.) + bX2[n, .>.)


m=-co m=-oo

The system is linear-


Time invariant: Shifting the input x[n] by an amount I yields
00
L x[n + m + l]ho[-m]e-;>m = X[n +I,.>.)
m=-oo

which is the output shifted by I samples. The system is time-invariant.


Next, we find the impulse response and frequency response o£ the system. To find the impulse
response, denoted as h[n], we Jet z[n] o[n]. =
00

h[nJ = 2: •In + mJwlmJ·-;•m


m=-oo
= w[-n]ei'"
= ho[n]ei'"
Taking the DTFT gives the frequency response, denoted as H(eJ~).
H(eJw) = Ho(eJ(w-1))
(b) We find S(eiw) to be

s[n] = (z[n]e-i'") • w[-n]


S(eJW) = X (ei<w+>)) W(e-jw)
S(eJW) = X (ei<w+>)) Ho(eJw)

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

Y>[n] = L z[n + mjw[mje-j.\om


m=-oc
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]---{

In the frequency domain, the input to the decimator is

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)

The output of the expander is


R-1
~~X (.;<w+>•-2•1/R)) Ho (.;<w-2•1/R))
449

The output Y• (ei") is then


R-1
Y.(eiw) =~~Go (ei(w->,)) X (ei<w-2•1/R)) Ho (ei<w->•-2•1/R))

The overall system output is formed hy summing these terms over k.


N-1
Y(eJw) = L
.... Yk(eiw)
R-lN-1
= ~ L L Go (ei<w-~•)) X (ei<w-2•1/R)) Ho (ei<w-~,-2.,/R))
h:O 1=0

To cancel the aliasing, we rewrite the equation as follows:


N-1
Y(elw) = X(ei")~ L Ho (ei<w-~•)) Go (ei<w-~•))

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

Therefore, we require the following relations to he satisfied so that y(n] = :r(n]:


N-1
L Go (ei<w-~·)) Ho (ei<w-~·- 2 •'/R)) = 0, v w, and l = l, ... ,R- l
1=0
N-1
L Ho (ei<w-~•)) Go (ei<w-~•)) = R, Vw
k=O

(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

X[n, >.•) = L :r[m]ho[n- m]e-;>,m


--co
450

or, using>..= 21rkjN,


co
X[n, k] = L :~:(m]ho(n- m]e-i•••mfN
m=-oo
Therefore, the output of the decimator is
co
X(Rn, k] = L :~:[m]ho(Rn- m]e-i2•>m/N
m=-oo

Recall that in general, the output of an expander with expansion factor R is


co
"'• (n] = L :~:(l]o(n - lR]
l=-oo

This relation is given in chapter 3. Therefore, the output of the expander is


co
L X[Rl, k]o(n -lR]
t•-oo

This signal is then convolved with go(n], giving


co co co
L L X[Rl, k]o[m -lR]go[n- m] =L X[Rl, k]g0 [n- lR]
m=-ool=-oo l=-oo

Therefore,

Y>[n] = ,t;.., go[n -lR] (J.., :~:[m]ho[RI- m]e-i••>mfN) ei 2•h/N

y[n] = ~,f.;.., go(n -lR] C~.., :~:[m]ho[RI _ m]e-i2dm/N) eJ2•h/N


N-l co co
= L L go(n -lR] L :~:[m]ho[RI- m]e-jbk(m-n)/N
A:=O l=-oo m=-oo
co co N-1

= L L go[n- lR]ho[RI- m]:~:[m] L eJ2•k(n-m)/N


l=-oo m=-CIO A:=O

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

where r is an integer. Therefore,


co co co
y[n] = L L go(n- lR]ho[lR- m]:~:(m]N L 6fn- m- rN]
t=-oom=-oo r=-oo
co 00

= N L L go[n-lR]ho[lR-n+rNJ:~:fn-rNJ
l:-oor=-oo
co 00

= N L :~:[n- rNJ L go[n -lRjho[tR + rN- n]


r= -co l= -=
451

..
Therefore, if we want y[n] = z[n], we require

L go[n -lR]ho[lR + rN- n] = .S[r].


l=-oc

for all values n.


(h) Intuitively, we see that it is possible since we are keeping the necessary number of samples. H
go[n] = o[n] find that
00

·2: .S[n - lR]ho[lR + r N - n] = ho[r N]


l=-oo
= .S[r]
since ho[r J\1 is zero for all values of r, except r = 0, where it is equal to 1. Thus, the condition
derived in Part (g) is satisfied.
(i) See Rabiner and Crochiere or Portnoll. (Hint: consider an overlap and add FFT algorithm.)
10.41. Note that h[n] is real in this problem.
(a) First, we express y[n] as the convolution of h[n] and z[n].
00

y[n] =L h[k]z[n- k]
k=-oo

The autocorrelation of y[n] is then

9'>,.[m] = £ {y[n + m]y[n]}

= £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

Since z[n] is white noise, it has the autocorrelation function

q,.. [l + m- k] =.,.;o[l + m- k]
Substituting this into the expression for q,,.[m] gives
00 00

q,•• [m] = .,.; L 2: h[k]h[l]o[l + m- k]


00

= .,.; L h[l + m]h[I]


l=-oo

Note that 00

4>•• [m] = .,.; L h[l - m]h[I]


1=-oo

is also a correct answer, since 4>.. [m] = 4>.. [-m].


452

(b) Taking the DTFT of 1/>n[m] will give the power density spectrwi. t,.(w).

t .. (w) = .. J;. {~ ..t..


oc oc
h[l+m]h[ij}·--

= ~ L: h[ij L: h[t + mJ•_,.....


1=-CIO m=-oo

Substituting k = I + m into the second summation gives


oc oc
t .. (w) = u! L h[ij L h[k]e-jw(t-1)
1=-oo l=-oc
oc oc
= ~ L: h[ij.;"'' L: h£"1·-;...
l=-oo 1:=-oo
oc oc
= ..~ L: h[-n·-;..• L: h[kJ.-;...
l=-oo l:=-oo
= ..~w(.;"')H(.;"'l
= ..~ IH(.;"'ll
2

(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

so y[n] is nonzero for 0 :5 n :5 M. Note that the autocorrelation sequence,


oc
1/>.,[m] = L y[n + m]y[n]
n.=-oo

can be re-written as a convolution


oc
1/>,.[m] = L g[m- n]y[n]

where g[n] = y[-n]. Therefore,


1/>,.[n] = 11[-n]•y[n]
Since y[ -n] is nonzero for - M :5 n :5 0, and 11[n] is nonzero for 0 :5 n :5 M, we see that their
convolution 1/>n[m] is nonzero only in the interva!Jml :5 M.
Z-transform domain: Note that
t,.(z) = ~H(z)H.(z)
If all the at's = 0, then
Jl
Lb•z-•
H(z) =
....
Jl Jl
t,.(z) = Lbtz-•L:b;z'
l:=O tao
453

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

(e) For an AR process, with bo = 1,

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

The ¢ •• [OJ term is

¢.,[0] = £ {:z:[n]y[n]}

= £ { :r:[n] (t. 4t11[n - k] + z[n]) }


N
= L "•£ {z[n]y[n- /.:]} + £ {z[n]z[n]}
1:=1
N
= L "•£ {z[n]y[n - /.:]} +.;
t=l

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

since ¢yy[k] = ¢ .. (-k]. Form 2: 1,


N
¢.. (m] = L a•¢.. (m- k] + ¢.,(m]
i=l
N
= L a.¢.,(m- k]
lr-1

since ¢ •• (m] is zero for all m 2: 1.


(f) By symmetry of the autocorrelation sequence, we know that

¢yy(m - k] = ¢n(k - m]
= ¢.,[1m- kl]
Thus,
N N
L a.¢.. (1m- kl] = L a.¢.. (m- kj
bl .k=l

Using the result from part (e), we get


N
L a•¢•• [1m- kl] = ¢.. (m]
t=l

form= 1, 2, .. . ,N.

10.42. (a) Sampling x,(t) we get

x(n] = x,(nT)

= :6 I:• (~ )1•1 ei<2•/16)0n


t=-4

Define the periodic sequence X(k] to be

1/2
••• •••
1/4

Then we see that we can write :r(n] in terms of X[k]:

z(n] = ..!._ L• X(k]ei<••l••lOn


16 -=-·
7
= ..!._ L X(k]ei<••/llltn
16 ..,_,
= IDFS{X(k]}
456

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.,

Xo[k] ={ X[A:], k=0, ... ,15


0, otherwise

Using this information we can now find G[k]


G[k] = DIT{g[n]}
= DIT{:r[n](u[n]- u[n- 16])}
= DIT{:r[n]}
= DIT{IDIT{ Xo[k]}}
= Xo[k]
Thus, G[ kJ looks like

G[kj

112 112

114 1/4

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 k

(b) We want to find a sequence Q[k] such that

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

x••,[32] = aX[31] + X[32] + aX[33]


we must therefore have

!
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.,,(e'"') = LWR (ei<w+2•/L)) + -i;wR(ei"') + IWR (e><w-2•/L))


458

(e) Taking the inverse DTFT of w••,(&w) gives w••,[n].

w ••,[n]

A sketch of '"••• [nJ is provided below.

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

-lt -lli6 0 " {I)

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

M(k,- k.:..) + 256 :S n :S M(k, + k") + 256


k~ - lc~ $ n $ k; + k6

where

k; = MJ;.+256
k~ = Mk.:..

(d) A typical sketch of X(eiw) and XN[k]look like.

-It It (I)
460

~[k)

After periodically replicating and upsampling by M we have a signal that looks lilre

~[n] alter upsampling

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

Thus, we need to extract the points

M(k.- k4) $n $ M(k. + k4)


A:~-A:~ :Sn:5 l:~+k~
461

where

k'c
k;,.
Solutions - Chapter 11

Discrete Hilbert Transforms


465

11.1. Using the fact that :t,[n] is the inverse transform of'R£ we get
JU{X(.;w)} = 2 _ ,..;w _ .u-iw

:z:,(n] = 26(n]- 4<5(n + 1]- 4<5(n- 1]

Since :z:(n] is causal, we can recover it from :,[n]

:z:(n] = 2:t,(n]u(n]- :z:,(O]o(n] = 26(n]- W(n -1]


This implies that
:z: 0 (n] = :z:(n] - 2z(-n] = d(n + 1] - 4<5[n- 1]
and since jim{X(&w)} is tbe transform of :z: 0 (n] we find

Im{X(.;w)} = 2osinw

11.2. Taking the inverse transform of ?U{X(.;"')} =5/4- wsw, we get


5 1 1
:z:,(n] = o(n] - o[n + 1] - o(n- 1]
4 2 2
Since x(n] is causal, we can rewver it from :z:,(n]

x[n] = 2:t,(n]u(n]- :z:,(O]o(n] = 45 o[n]- 6(n- 1],


11.3. Note that

:z:(n] = 6(n]- 21 o(n- 1]


but this does Dot satisfy the coDditioDS OD z(n] giveu in the problem statemeut.

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

11.4. Take the DTFT of :z:.(n] to get


1 1 1
:z:.(n] = 26(n] - o(n + 2] - o(n - 2]
4 4
. 1 1
X.(e"") = 2- 2cos2w.
where X.(~"')= !(X(~) +X"(~)] is the conjugau qmmetric part of X(~). Since X(~"')= 0 for
-.- :S w < 0 we have
= { 2X.(~"'), 0 :S w < ..
0, otherwise
= { 1-cos2w, O:Sw<O"
0, otherwise

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

H(~) = { -:j, 0<w<7r,


-lf<w<O.
. ],

(a) First, take the transform of :z:.(n]


X.(~"') = .-6(w- wo) + .-6(w + wo).
Now, filter with H(~"') and take the inverse transform to get :z:;[n]

X;(~"') = H(~"')Xr(~"')
= -j.-6(w- wo) + j1r6(w + wo)
:z:;(n] = sinwon

(b) Similarly, :z:;(n] =-cos won.


(c) z.[n] is the ideal low pass filter

[ sin(w.n) { 1, lwl :S "'•


"'• n J= '"' +---+ 0, "'• < lwl $ ..

After filtering with the Hilbert tranformer - get

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

11.6. Using Euler's identity,

jX1 (t!"') = j(2sinw- 3sin 4w)


= 2 ( ei"' -2 .-,.,) - 3 ( .;•· -2 .-;... )

= -~.!"" + .J•- .-;· + ~.-;<w


2 2
Since z.[n] is the inverse transform of jX,(tJ"') we get

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.,

z[n] = 2zo[n]u[n] + z[O]o[n]


= z[O]o[n]- 26[n- 1] + 30[n- 4]
The extra information given to us allows us to find z[O],

6 = X(tJ"ll.,=<>
""
I: z[nje'"(OJ
= n.=-oc
= z[O]- 2 + 3
Plugging this into our equation for z[n] we find

z[n] =SJ[n]- 26[n- 1] + 30[n- 4]


11.7. (a) Given the imaginary part of X(ei"'), we can take the inverse DTFT to find the odd part of z[n],
denoted z 0 [n].

Im {X(ei")} = sinw + 2sin2w


= .!..J"- .!..-;. + ~ .......,- ~.-;"'"'
2j 2] j j

= yei"'"' + 2~ ....-- ;/-;.- 7·-j""'


z.[n] = 'DF/ 1 (;Im {X(tJ"l}]

= 'DFT-' [&""' + ~&· - ~.-;.- .-;""']


1 1
= o[n + 2] + 2o[n + 1]- 2o[n- 1]- o[n- 2]

Using the formula z[n] = 2z0 [n]u[n] + z[O]o[n], we find


z[n] =-o[n- 1]- 26[n- 2] + z[0]6[n]
Any z[O] will result in a correct solution to this problem. Setting z[O] = 0 gives the result
z[n] = -<l[n- 1]- 26[n - 2]
468

(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].

z 0 [n] = 'D:F/ 1 (jim {X(ei"')}]


= 'D:FJ' [!.;
2
... + ~~-
2
~.-;.. - !.-;...]
2 2
1 3 3 1
= 2
o[n + 3]+ iln+ 1]- 26(n -1]- o[n- 3]
2
Using the formula z[n] = 2%0 (n]u(n] + z[O]o[n], we find
z(n] = -36[n- 1] - 6(n - 3] + :r(O]o[n]
Taking the DTIT of z[n] gives

X(ei"') = -3e-'-jw- .-j3w + z[Oj


469

Evaluating this at w = " gives


X(ei~)i..,=• = -Je-i• - .-ib + z(O] =3
3+l+z(O] = 3
z(OJ = -1
Therefore,
z(n] = -30(n- l]- o[n- 3]- o(n]
(h) Yes, the answer to part (a) is unique. The specifu:ation of X(eiw) at w = 1r allowed us to find a
unique z(n].
11.10. Factoring the magnitude squared response we get

jH(ei"')!2 = •4- cosw = I - cosw + 4' {I- !e-i"')(l - ~ei"')


5+4cosw l+4cosw+4 (I+ :u.->"')(1 + 2ei"')

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]

11.11. Note that z;[n] can be written as

z;(n] = -46(n + 3] + 46(n - 3]


Taking the DTFT of z;(n] gives
X;(ei"') = --4ei.., + 4e_,..,
= -4(2jsin3w)
= -8jsin3w
Since X(ei"') = 0 for -1r $ w < 0, we cao lind X(ei") using the relation

X(ei") = { 2jX;(e'"), 0 < w < "


0, -11' :sw < 0
470

Thus,
X(.;•) = { 16sin3w, 0 < w < ,.
0, -1r$W < 0
Therefore, the real part of X (&"') is

X,(.;"') = ~[X(&"')+ X"(e-i"'J)


= { 8sin3w, 0 < "' < ..
-8sin3w, -.- :5 w < 0

11.12. (a) Factoring the magnitude squared response we get

_;"')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]

satisifies the given conditions.


11.13. Expressing XR(ei"') in terms of complex exponentials gives

XR(e'"') = 1 + cosw + sinw- sin2w


= 1 + ~e'"' + ~.-;.. + ..!.,.;.. - ..!.,.-;.. - ..!_,.;,., + .!,.-j2w
2 2 2] 2] . 2] 2;
= _..!_,.;,., + ~e'"' + ..!_,.;., + 1 + ~.-;.. - ..!.,.-;.. + .!,.-j2w
2j 2 2j 2 2j 2j

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]

Using the relation x(n] = 2:r,[n]u[n]- :r,[O]o[n],

:r[n] = o[n] + o[n- 1] + jo[n- 1]- j6[n- 2]


471

We then find the conjugate-antisymmetric part, :t0 [n] as

:t0 [n] = !2 (z[n]- :r•[-n])


= !2 (o[n] + o[n- 1] + jo[n- 1]- jo[n- 2]- o[n]- o[-n- 1] + jo[-n- 1]- jo[-n- 2])
1
= 2 (o[n - 1] + jo[n- 1]- jo[n- 2]- o[n + 1] + jo[n + 1]- jo[n + 2])
I . .
= -2 (o[n + 1]- o[n- I])+~ (o[n + 1] + o[n- 1])- ~ (o[n + 2] + o[n- 2])
Taking the DTFT of z 0 [n] gives jXr(&"').

= _!2 {&w- e-jw) + t2 (&w + e-jw) - t2 (&"" + .-i"")


= -jsinw+jcosw-jcos2w
So

11.14. First note that,

(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)

.,,[n] = o[n] + ~ t. G)· (6[n + k] + o[n _ k])

Using (c) and then taking the odd part we get,

z[n] = 2z,[n]u[-n]- z,[0]6[n]

= o[n] +

z[n]- z[-n]
'Ecr
1=1
00

2 6[n + k]

=
cr
z 0 [n]
2

= 2I 'E 2 00 (o[n + k]- o[n- 1.:])


i=l
472

Now taking the DTFT and using (b),

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

:z: 0 (n] = v:rr-' (jim {X(ei"')})


= vn-' [~e>"'- ~·-'"']
1 1
=
2o(n + 1]- 2o(n- 1]
Using the formula z[n] = 2z.[n]u[n] + z[O]o[n),
z[n] = -o[n- 1) + z(O)o[n)
Since
L"" z[n] =3
n=-oo

-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

Using this information in the second condition we find

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

z.,[n] = ! txR[k]~2 •""12


2-
1
= ! L:xR[kJHl""'
2-
474

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

jX1[1] = (o- ~ +0- n = -j


jX1 (2J = (o+~+o-D =o
jXJ(3] = (o+~+O+n =i
Thus,
iX1[kJ = -io(k- 11 + io(k- 3]
11.20. As the following shows, the second condition implies :(OJ= 1.

: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

jX,,(kJ = ~ ( ,-i(2r/6)<• _ ei(2r/6)<•) _ ~ ( e-j(2r/6)5> _ ei!2r/6).. )

= - ~j sin(4rk/3) + ~j sin(5rk/3)
= j_!_ (-6[k- 2] + 6[k- 4])
../3

jX,,[k] = ~ (e-j(2r/6)• -ei(2r/6l•) _ ~ (e-j(2r/6)2> -ei(2r/6)2>)

= -~jsin(rk/3) + ~jsin(2.. k/3)


= j_!_(-6[k-2]+6[k-4])
../3
Thus, both x 2 [n] and :t 3 [n] are consistent with the information given.
11.21. (a) Method 1:
We are given

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,

x~w) = U(p,w) + jV(p,w)


= 1 + p- 1ocos""- jp- 1asinw
= 1 + op- 1(cosw- j sinw)
= 1 +op- 1•-;..
X(z) = 1 + oz- 1

(b) Method 2: Since XR(ei"') is the transform of z,[n] we have

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,

x(n] = z,(n] + z 0 [n]


= o(n] + oo(n- 1]
X(z) = 1 + oz- 1

Note that we could have obtained x(n] directly from z,(n] as follows

z(n] = 2z,[n]u(n]- z,[O]o(n]


= (U[n] + oo(n + 1] + ao(n- 1])u[n] - o(n]
= o(n] + oo(n- I]

11.22. Taking the z-transform of uN(n] we get

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 is even but k # 0 we see that iiN[k] = 0. For k odd, we get


2+ 2e-;2d/N
l-e-i2d/N
2e-;d/N(&d/N + .-;d/N)
= e-j.-lcfN (eid/N _ e jrrlc/N)
= -2j cot(11kjN)

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

Putting it all together gives us the desired answer

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),

2:r,.[n] n=1, ... ,Nf2-1


:r[n) = :r,.[n]uN[n] = :r,.[n], n=O,N/2
{ 0, otherwise
479

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

k=O, ... ,N-1


1- e j2d/H

N, k=O,
= -2j cot(1rk/N},
{ 0,
0 < k < N- 1, k odd
otherwise

11.24. We are given


HR(eiw) = HER(eiw) + HoR(eiw)
H1(eiw) = HEJ(eiw} + Hoi(eiw)

h,.(n] +--+ H,(eiw) = H,.(eiw) + jHs(eiw)


h;[n] +--+ H;(eiw) = Hc(eiw) + iHD(eiw)
where h,.[n], h.[n], HR(eiw), HI(eiw), HER(eiw), HoR(eiw), HEI(eiw), and Hoi(&w) are real.
Begin by breaking H(eiw) into its real and imaginary parts HR(eiw) and H1(eiw)

H(~) = HR(eiW) + iHI(ei"')


= [HER(ei"') + HoR(ei"')] + j[HEI(ei"') + Hoi(ei"')]
Now solve for the conjugate symmetric and conjugate antisymmetric parts of H(e'"')
H(eiw) + H"(e-i"')
= 2
[HER(ei"') + HoR(eJ"')] + j[HEI(ei"') + HOI(ei"')]
= 2
(HER(ei"')- HoR(eiw)]- j[HEI(ei"')- Hoi(ei"')]
+ 2
= HER(ei"') + jHoi(ei"')
H(ei"')- H"(e-i"')
H;(ei"') = 2j
= [HER(ei"') + HoR(ei"')] + j[HEI(ei"') + Hoi(ei"')j
2j
[HER(ei"')- HoR(ei"'))- ;[HEI(ei"')- HOI(&"')]
2j

Thus,
H,.(ei"') = HER(ei"') Hc(ei"') = HEI(ei"')
H 8 (ei"') = Hoi(ei"') HD(ei"'} = -HoR(ei"')
480

11.25. (a) By inspection,


H(~"') = j(2Hlp(~!w+fl)- 1)
. 1_ jH(~!w-fl)
Hlp(e"") = 2

(b) Find h[n1:


Taking the inverse DTFT of H (~.,) yields

h[n] = j [2·-j(•/•Jn hlp[n1- o[nJ]

= j [2cos(lm/2)hlp[n1- j2sin(.-n/2)hlp[n1- 6[n1]


= 2sin(.-n/2)hlp[n]

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,

i.[n1 = h[n - n•1w[n1


= 2sin(.-(n- n•)/2) hlp[n- n•1w[n1
= 2sin( ..(n- nd)/2) i.lp[n1

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"'}

-2x ' / -7t (I)

, .... '.... 31t 6lt

=
(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.

Then decimatiDg by (M + 1) produces the desired result.


11.27. Yes, it is possible to always uniquely recover the system iDput from the system output. Although
483

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

Z(eiw) = Y(e'w)- a.l(w}- M(w- .-}


We can construct a conjugate symmetric X(eiw) from Y(eiw) and Z(e'w) as

In the time domain, this is


z[n] = y[n] + z•[n]
Or, since
z[n] = y[n]-.!!...- b(- 1}"
2" 21r
z[n] = y[n] + y•[n]-.!!...- b(- 1)"
21r 21r
11.28. Since H(z) corresponds to a real anticausal sequence h[n], F(z) = H(1fz) corresponds to a real,
stable, causal sequence f[n]. We can apply the equation developed in the book for causal sequences to
F(z).
F(z) = - . 1 FR(v) (~) dv, izi ~ 1.
1
2WJ Tc z- v v
where v = ei 6 is the integration -variable; i.e., the closed contour C is the unit circle of tlie v-plane. Now
find H(z) ·

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

Let /[n] = 1l{z[n]} and g•[n] = z[n]. Then

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] is au ideal Hilbert 'Iraosformer:

h[n]

H(ei"') = { j,-j, 0 <w <r


_,.. <w<O

(a) In the frequency domain,

Therefore, <l>z;z,[m] = tl>•••• [m].


485
-

(b) The cross-<:orrelation between input and output is just the convolution of ¢ •••• (m] and h(m],
00

¢•••• (m] = L h[k]¢•••• [m- k]


r--oo
The following shows that it is an odd function of m:
00

¢ ••.,[-m] = L h(k]¢•••• [-m- k]


00

L h(-l]¢•••• (-m + l]
l=-cc
00

= L h(-l]¢•••• (m -l]
l=-oo
00

= - L h(l]¢•••• (m -l]

= -¢•••• (m]

since h[n] = -h[-n] and ¢ •••• [m] = ¢ •••• [-m].


(c) Starting from the definition of the autocorrelation and using the linearity ofthe expectation oper-
ator we get

¢zz (m] = £(:r[n]x"(n + m]]


= £[(:r,[n] + j:r;[n])(:r,[n + m]- i:r;[n + m])J
= ¢•••• [m] + ¢., •• [m] + j(¢•••• [m]- ¢ •••. [m])
= u •.•. (m]- 2j¢•••• [m]
The last line was found using the results from parts (a) and {b).
(d) Taking the transform of both sides of the equality from part (c) we find

P,,{w) = 2t •••• (eiw)- 2jt••• ,(eiw)


= 2(t•••• {eiw)- jH{eiw)tz.s.{e'"')J
= 2t•••• (eiw){l- jH(eiw)j
Since

l _ "H{eiw) = { 0, 0 < W < lr


] 2, -7r < w < 0

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

' '- " OJ

1 yc(jO) =sc(j!l)

(b) In the frequency domain


H;(ei"')
'lf <w<71"
= { 0,5, otherwise
Note that H;(e'"') = SG(ei<"'-47/ 5 !), where
1, lwl < f
= { o. otherwise
487
sin(lm/5)
g(n] =

We can therefore write h;(n] as

h;(n] = 5sin(jn) eiY.n

=
""
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

11.32. (a) We know


X(z) = log(X(z)]
When X(z) has a zero or a pole, the term log(X(z)] goes to negative infinity or infinity respectively.
Therefore, X(z) has a pole at these locations.
If :t[n] is causal, X (z) has a region of convergence that is the outside of a circle correspon~ to its
largest pole. However, we require the region of convergence to include the unit circle, i.e., X (dw)
is defined. These two conditions imply that the poles of X(z) must be inside the unit eircle.
But the pole locations for X(z) correspond to the pole and zero locations for X(z). We conclude
the poles and zeros of X (z) must be inside the unit eircle, i.e., o;[n] must be minimum phase.
(b) This argument is similar to the last, but in reverse. Start with the fact that z(n] is minimum phase
and must have its poles and zeros inside the unit circle. Then, as shown in part {a), the poles of
X (z) must also be inside the unit circle. Becanse the region of convergence mnst include the unit
eircle, we know it lies outside the eirc1e defined by its ia<gest pole. Thus, z[n] mnst be causaL
(c)

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

(d) Using the power series expansion

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

We~ now take the inverse transform o:f X(z).

log{ A), n=O

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.

You might also like