E4702 HW#4-5 solutions
1. (P3.17) Midtread type uniform quantizer (figure 3.10(a) in Haykin)
Gaussian-distributed random variable with zero mean and unit variance is applied to this quantizer input
(a) the probability that the amplitude of the input lies outside the range -4 to 4
(Solution.)
This problem can be described more clearly with the complementary error function. (see eqn.(4.29) in
Haykin) Z ∞
2 2
erfc(u) = √ e−z dz
π u
For the quantizer input, m, with the normal distribution,
Z ∞ √
1 2
P [|m| > 4] = 2 √ e−t /2 dt = erfc(2 2) = 6.334 · 10−5
4 2π
(b) output signal-to-noise ration of the quantizer ( input located between [4,-4])
(Solution.)
The average power of the message signal is
Z 4 Z ∞
2 1 −x2 /2 1 2
P = x √ ·e dx ≈ x2 √ · e−x /2 dx = 1
−4 2π −∞ 2π
We can evaluate the above integral by using Matlab with the following commands:
>> syms x
>> eval(int(x^2*exp(-x^2/2)/sqrt(2*pi),-4,4))
Then, we will get 0.9989 as a result. Here, we will use the approximated average power.(i.e. P = 1)
Putting this and mmax = 4 into equation(3.32) in Haykin,
2 1 2 −2R 16 −2R
σQ = 4 2 = 2
3 3
The signal-to-noise ratio is
1 3 · 22R
(SNR)O = 16 −2R =
3 2
16
In dB scale,
(SNR)O(dB) = 10 log10 (3/16) + 10 ∗ log10 2 ∗ 2R = −7.27 + 6.02R
We can compare this result to that of sinusoidal modulating signal, which is given in Example3.1 of
Haykin. (SNR)O(dB) of the sinusoidal modulating signal is 1.8 + 6R, which is 9dB greater than Gaussian
input.
The reason can be directly observed from the average power: Psinusoidal = A2m /2 = 8 where Am = 4
and PGaussian = 1. This means that more number of signals in the Gaussian input signal come with
lower amplitude compared to the sinusoidal signal. (This can be also seen on the bell-shaped curve of
Gaussian distribution.)
As a following result, in case of Gaussian input signal, non-uniform quantizer - dividing the lower
amplitude part in a finer scale - would outperform the uniform quantizer.
1
2. (P3.19) mean-square quantization error for a nonuniform quantizer
- ∆i : ith step size of the given nonuniform quantizer
- pi : probability that the input signal amplitude lies within the ith interval
- ∆i is small compared with the excursion of the signal
(Solution.)
The expected mean-square quantization error can be represented in terms of the quantization error in each
interval.
L
X
2 2
σQ = σQ,i pi
i=1
where L is the number of quantization levels.
2
In the ith interval, the quantization error for input signal , qi , and the mean-square quantization error, σQ i
,
is defined as follows.
mi + mi+1
qi = m −
2
2
£ ¤
σQ,i = Em (qi − E[qi ])2
From the condition that the quantization interval is small compared with the excursion of the signal, we can
assume that the input signal is uniformly distributed in an interval. Then,
E[qi ] = 0
Z ∆i /2
1 ∆2
2
σQ,i = qi2 dqi = i
−∆i /2 ∆i 12
Thus,
L
2 1 X 2
σQ = ∆ pi
12 i=1 i
3. (P3.11(d)) Power spectral density of bipolar return-to-zero signals
(Solution.)
From the lecture, power spectral density is described as follows.
|G(f )|2 X
Ss (f ) = R(k)ejk2πTb (1)
Tb
−∞<k<∞
PN
where sN (t) = −N an g(t − nTb ).
Tb
For the bipolar RZ, g(t) = rect( Tbt/2 ) and G(f ) = 2 sinc(f Tb /2)
The probabilistic description of an is:
0, with P (an = 0) = 1/2
an = A, with P (an = 1) = 1/4
−A, with P (an = −1) = 1/4
The autocorrelation function of an is:
R(k) = E[an an+k ]
2
• R(0) (
0, with p = 1/2 (when an = 0)
an an
A2 , with p = 1/2 (when an = A or − A)
1 1 A2
Thus, R(0) = 0 · 2 + A2 · 2 = 2
• R(1)
Here, notice that
P (an+1 = m|an = 0) = P (an+1 = m), m ∈ {0, A, −A}
1/2, m = 0
P (an+1 = m|an = A or − A) = 1/2, m = −an
0, m = an
Thus, (
0, with p = 3/4 (when an = 0 or an+1 = 0)
an an+1
−A2 , with p = 1/4 (otherwise)
2
3 1
R(1) = 0 · 4 + (−A2 ) · 4 = − A4
• R(k), k > 1
an an+k (k > 1) has non-zero value with the probability of 14 only when both bits are non-zero. Given
Pn+k−1
that an an+k is non-zero, it is either 1 or -1. Here, we define bn,k = j=n+1 aj which is zero when the
number of non-zero bits between an+1 and an+k−1 is even.
1 1
P (an an+k = 1) = P (bn,k = 0) , P (an an+k = −1) = P (bn,k 6= 0)
4 4
Notice here that P (an an+k = 1) = P (an an+k = −1).
Applying the equation (2) to be given in (P3.23),
1
P (an an+k = 1) = P (an an+k = −1) =
2
Thus,
R(k) = E[an an+k ]
= 1 · P (an an+k = 1) + (−1) · P (an an+k = −1) + 0 · P (an an+k = 0)
= P (an an+k = 1) − P (an an+k = −1)
=0
For the negative index of k, remind that R(k) is an even function.
By putting G(f ) and R(k) into (1),
· ¸
| T2b sinc(f Tb /2)|2 2 1 1 ¡ −j2πf Tb ¢
Ss (f ) = A − e + ej2πf Tb
Tb 2 4
· ¸
A2 Tb 1 1
= sinc2 (f Tb /2) − 2 cos(2πf Tb )
4 2 4
A2 Tb 1 − cos(2πf Tb )
= sinc2 (f Tb /2)
4 µ ¶ 2
A2 Tb f T b
= sinc2 sin2 (πf Tb )
4 2
3
4. (P3.23) a chain of (n-1) regenerative repeaters with n sequential decisions made on a binary PCM wave
repeaters
1-p1
p1
p1
1-p1
source destination
0 1 2 n-2 n-1
p1 : probability of error on each decision procedure
(Solution.)
(a) The error probability on the nth stage can be represented in a recursive form.
pn+1 = pn (1 − p1 ) + (1 − pn )p1 = p1 + (1 − 2p1 )pn
This equation can be expanded as
pn+1 = p1 + (1 − 2p1 )pn
= p1 + p1 (1 − 2p1 ) + (1 − 2p1 )2 pn−1
= p1 + p1 (1 − 2p1 ) + p1 (1 − 2p1 )2 + (1 − 2p1 )3 pn−2
n−1
X
= p1 (1 − 2p1 )k + (1 − 2p1 )n p1
k=0
1 − (1 − 2p1 )n
= p1 + (1 − 2p1 )n p1
1 − (1 − 2p1 )
1
= [1 − (1 − 2p1 )n+1 ]
2
Thus,
1
pn = [1 − (1 − 2p1 )n ] (2)
2
(b)
n µ ¶
X n
(1 − 2p1 )n = (−2p1 )k
k
k=0
If p1 is very small, we can approximate this equation by taking only the constant term and the first
polynomial term.
(1 − 2p1 )n ≈ 1 − 2np1
1
pn ≈ [1 − (1 − 2np1 ] = np1
2
5. (P3.31) A one-step linear predictor
f0
- input : sin(2π 10f 0
n) = sin(2π · 0.1n)
4
x[n] A sin(2S 0.1n)
Z 1
w1
x[n] w1 A sin(2S 0.1(n 1))
(a) the optimum value of w1 for the minimizing the prediction error variance
(Solution.)
The prediction error is defined as:
e[n] = x[n] − x̂[n]
The index of performance (or the prediction error variance) is defined as:
J = E[e2 [n]]
Then,
J = E[x2 [n]] + E[x̂2 [n]] − 2E[x[n]x̂[n]]
= A2 E[sin(2π0.1n)2 ] + w12 A2 E[sin(2π0.1(n − 1))2 ] − 2A2 w1 E[sin(2π0.1n) sin(2π0.1(n − 1))]
1 1
= A2 + w12 A2 + w1 A2 E[cos(4π0.1n − 2π0.1n)] − w1 A2 E[cos(2π0.1)]
2 · 2 ¸
1 1
= A2 w12 − cos(2π0.1)w1 +
2 2
Taking derivative of J over w1 ,
dJ
= A2 w1 − A2 cos(2π0.1)
dw1
J has its minimum when this derivative is zero. Thus,
w1 = cos(2π0.1) = 0.809
(b) the minimum value of the prediction error variance
(Solution.)
when w1 = 0.809,
Jmin = 0.1727A2