EE 477 Introduction to Digital Communications Fall 2020

Midterm

NAME : SOLUTION

• You have 100 minutes to complete this exam.

• There are 6 problems and 100 points in this exam. Points for the individual problems and
subproblems are marked in the problem statement.

• Good Luck!

Problem 1 (12.5 points) : . . . . . . . . . . . . . . .

Problem 2 (12.5 points) : . . . . . . . . . . . . . . .

Problem 3 (10 points) :. . . . . . . . . . . . . . .

Problem 4 (15 points) :. . . . . . . . . . . . . . .

Problem 5 (25 points) :. . . . . . . . . . . . . . .

Problem 6 (25 points) :. . . . . . . . . . . . . . .

TOTAL:. . . . . . . . . . . . . . .
1. [12.5 Points] TRUE/FALSE Questions with short answers. In this part, please answer each
question with a TRUE or FALSE answer with a brief justification. This can be i) a short
explanation, ii) an example or iii) a counterexample. No points will be given to an answer
that does not include any of these three.
IMPORTANT NOTE: Answers without any explanation or with irrelevant or
incorrect explanation did not get any credit, even if the “True/False” statements
were incidentially correct.

(a) If a noise process n(t) is said to be white, its samples are independent and identically
distributed.
FALSE. The noise being white means its spectrum is flat and auto-correlation func-
tion is unit impulse. This means the samples are uncorrelated. If the noise is
Gaussian this also implies independence but for any other process uncorrelatedness
does not imply independence. In the question, it is not mentioned that the noise is
Gaussian.
(b) 8-QAM has the same spectral efficiency (bits/symbol) as 8-PSK and has better symbol
error performance at the same SNR.
FALSE. You may remember from your mini-homework that for general rectangular
M -QAM where M = N K , the probability of error is given as:
v  v 
u 6E N-PAM u 6E K-PAM
u u
N − 1 t avg. K − 1 t avg.
PeM-QAM ≈PeN-PAM + PeK-PAM = 2 Q +2 Q
 
N N0 (N 2 − 1) K N0 (N 2 − 1)


v  v 
u M-QAM M-QAM
u u
N −1  N2 − 1   + 2 K − 1 Q  6Eavg. K2 − 1 
u
t 6Eavg. t
=2 Q 
N  N0 (N 2 − 1) N 2 + K 2 − 2  K  N0 (N 2 − 1) N 2 + K 2 − 2 
v 
K(N − 1) + N (K − 1) 
u
u M-QAM
t 6Eavg. 
=2 Q 
 N0 (N 2 + K 2 − 2) 
NK

q 
M-QAM Eavg.
When N = 4 and K = 2 (or vice versa), 8-QAM Pe = 2.5Q whereas
3N0
q 
2Eavg. 2
for 8-PSK Pe8-PSK = 2Q sin (π/8))N0 . If you evaluate these two BER’s at
different SNR values you will see that while 8 − QAM performance is better at high
SNR’s, at low SNR’s it is worse than that of 8-PSK. (Those answers which said
8-QAM performs better at high SNR and hence TRUE are considered true and
given credit.
(c) If modulation-A and modulaton-B require SNR’s of γA and γB , respectively, to reach to
the same BER performance and if γA < γB , it means that modulation-B always provides
better spectral efficiency.
FALSE. BPSK and BFSK comparison is a good example. BPSK performs bet-
ter than BFSK but they both send 1 bit/symbol and thus have the same spectral
efficiency.
(d) Two modulations with the same minimum distance have the same symbol error proba-
bility.
FALSE. The average number of neighbors is also a factor, PAM and QAM may have
the same minimum distance but QAM will have roughly twice as many points at
the minimum distance, hence have worse error performance.
(e) The power spectrum of a digitally modulated signal is determined solely by the pulse
shaping filter used at the transmitter.
FALSE. Remember that we derived the power spectrum of a digitally modulated
signal as:
σ2 2 µ2 X   m 2  m
Φ(f ) = i |G(f )| + i2 ∞ G  δ f−
 
T T m=−∞ T T

from which it is obvious that the spectrum also depends on the values of the signal
variance and mean and the symbolling period as well as the pulse shaping filter.

[12 minutes]
2. [12.5 Points] Consider a communication system that sends one sided PAM symbols; that is,
the m-th transmitted waveform is

sm (t) = Am g(t) cos(2πfc t)

where Am = m∆ for m = 0, 1, 2 . . . M − 1. g(t) is the the square-root raised cosine pulse

shaping filter with a roll-off factor β = 0.3. Evaluate and plot the power spectral density of
this modulation.

As in the first question, this is a referral to the power spectrum of a digitally

modulated signal that we derived in our lectures and that can be written as:

σi2 µ2 X   m 2  m
Φ(f ) = |G(f )|2 + i2 ∞ G  δ f− .
 
T T m=−∞ T T

So all we have to do is to derive the mean (µ) and variance (σ 2 ), for our PAM
constellation, which is a bit different from what we had.
M −1
1 X M (M − 1)∆ (M − 1)∆
µ= Am = =
M 2M 2
m=0
M −1
1 X (M − 1)M (2M − 1)∆2 (M − 1)(2M − 1)∆2
E[A2m ]= A2m = =
M 6M 6
m=0
(M − 1)(2M − 1)∆2 (M − 1)2 ∆2 (M 2 − 1)∆2
σ2= − =
6 4 12
Then plug some non-zero values in (you were not necessarily expected to calcu-
late the mean and the variance but just to show that there are non-zero values
there) and plot a copy of the raised-cosine at the baseband and an weighted im-
pulse train having an impulse at each multiple of 1/T .

NOTE: Those who wrote 0.5Am G(f − fc ) + 0.5Am G(f + fc ), you answered with the frequency
response of sm (t). The question asks for the power spectral density which is not the same
thing.
[13 minutes]
3. [10 Points] In this and the subsequent question you will design a digital communication system
under different performance requirements. Suppose the communication system is required to
have a bandwidth of B = 1 MHz around a carrier frequency of fc = 2 GHz. Square-root
raised cosine filters with roll-off factor α = 0.2 are to be employed both at the transmitter
and receiver. The transmitter only uses M-QAM for modulation where all M-QAM symbol
constellations are at its disposal.

Suppose we want a minimum data rate of Rb = 6 Mega bits per second (Mbps).
What is the minimum constellation size that satisfies this data rate requirement.
[10 minutes] We are using a square-root pulse shaping filter, so if you con-
sider the the spectrum of (square-root) raised cosine filter, you will see that the
bandwidth is equal to (1 + α)/2T .
1+α 1.2
B= → 106 = → T = 6 × 10−7 sec/symbol.
2T 2T
All we have to do is multiply this with the desired bit rate to find the minimum
number of bits per symbol.

Rb × T = 3.6 bits/symbol

This means that we have to use at least 16-QAM constellation or 4 bps to meet
the data rate requirement. The actual data rate is Rb = T4 = 6×10
4
−7 = 6.67 Mbps.

1+α
(Those who have considered the bandwidth as two sided and who computed B = T also
received full credit.)
4. [15 Points] Consider the same communication system as in the previous question. The sys-
tem in question uses a bandwidth of B = 1 MHz around a carrier frequency of fc = 2 GHz.
Square-root raised cosine filters with roll-off factor α = 0.2 are to be employed both at the
transmitter and receiver. The transmitter only uses M-QAM for modulation where all M-
QAM symbol constellations are at its disposal.

(a) [10 Points] Suppose now we require this system to have a symbol error probability
that is no worse than 10−3 at 16 dB SNR. What should be the constellation size that
satisfies this requirement. Is your answer different from or the same as the one you gave
previously. If the two answers are different, what is the corresponding data rate if you
were to use this constellation.
First of all γdB = 16 ↔ γ = 39.8. Let us start with the square 16-QAM constel-
lation from the previous part to see it satisfies the criterion.
s ! r 
16-QAM 4−1 3Eavg. γ
Pe ≈4 Q = 3Q = 7.2 × 10−2 .
4 N0 (15) 5

This does not satisy the requirement. Hence we need to reduce the constel-
lation size. Let us try 8-QAM. Notice that because 8-QAM is not a square
constellation, its error probability is expressed differently (as you derived in
your mini homework). You can also refer to question 1b.
r 
8-QAM 5 γ
Pe ≈ Q = 3.37 × 10−4 .
2 3

So in order to satisfy the SER requirement we have send signals with 8-QAM
instead of 16-QAM. The corresponding bit rate is Rb = T3 = 6×10
3
−7 = 5 Mbps.
This means in order to comply with the error performance requirement we
had to sacrify 25% from the bit rate.
(b) [5 Points] What would be your answer if the 10−3 symbol error rate was to be achieved
at 18 dB SNR. First of all γdB = 18 ↔ γ = 63.1. Let us again start with the
square 16-QAM constellation from the previous part to see it satisfies the
criterion. r 
16-QAM γ
Pe ≈ 3Q = 5.72 × 10−4 .
5
Notice that at this SNR both the bit rate and error performance require-
ments are satisfied. In other words, you need high SNR values to provide
required data rates while keeping the quality of service.

[15 minutes]
5. [25 Points] Consider a binary communication system where xk ∈ ±1 amplitudes are sent over
an AWGN channel where the signal model is as follows:

rk = αxk + ηk

where α is an unknown constant on all the transmitted symbols (think of it as the constant
path loss during transmission), ηk ’s are independent and identically distributed Gaussian
noise samples, i.e. ηk ∼ N (0, N0 /2).

(a) [5 Points] It is obvious that we need to estimate what α is in order to be able to detect
xk ’s. For this purpose the first bit in the transmission x1 = 1 is selected as a pilot bit
that is known by both the transmitter and receiver. The subsequent symbols are the
normal data bits which are not known by the receiver. Make a maximum-likehood (ML)
estimate of α based on your pilot transmission. If this estimate is denoted α̂, what is
the mean and variance of α̂?

We have covered the unknown variable estimation when we considered syn-

chronization. For the first symbol instant, we send a known bit x1 = 1, so
the the correspondingreceived signal is:

r1 = αx1 + η1 = α + η1 .

Then the ML estimate of α is:

α̂=argmax f (r1 |α) = argmax log f (r1 |α)

(r −α)2
1 − 1N 1 (r1 − α)2
=argmax √ e 0 = argmax − log(πN0 ) −
πN0 2 N0
2
=argmin (r1 − α) .

To minimize this all you have to do is to take derivative and find α̂ which
makes the derivative 0. It is easy to see that α̂ = r1 and by the signal model
α̂ ∼ N (α, N0 /2).
[10 minutes]
(b) [10 Points] Suppose we feel generous and send the first two symbols x1 = 1 and x2 = 1
as pilots, which are followed by the actual data bits. Make another ML estimate of the
unknown α in this case. Is your estimate better, worse or the same. In other words, is
there any advantage to using more pilots. Whatever the answer is justify your response.
(Hint: You can again evaluate the mean and the variance of your estimate.)

Now, the pilot signal model becomes:

r1 =αx1 + η1 = α + η1
r2 =αx2 + η1 = α + η2

Alternatively,
     
r1 α η
= + 1 ↔r=α+η
r2 α η2

Then the ML estimate of α is:

α̂=argmax f (r|a) = argmax log f (r|a)

kr − αk2
=argmax − log(πN0 ) −
N0
=argmin kr − αk = argmin r12 + r22 − 2α(r1 + r2 ) + 2α2 .
2

r1 +r2
Again it is easy to see that α̂ which makes the derivative 0 is, α̂ = 2 .

r1 + r2 η1 + η2 N0
=α+ ∼ N (α, ).
2 2 4
Notice that the second estimate has the variance halved, which means it is
a closer estimate to the true value α and therefore a better one. Using more
pilots would improve the estimation performance.

[10 minutes]
 (c) [10 Points] Now you do not have the true α but you have an estimate of it, α̂, which is
hopefully a close approximation. Your actual signal model will be

rk = α̂xk + ηk

where you can now try to detect your bits. Devise your maximum likelihood detection
rule for k = 2, 3, . . . and provide the probability of error. Will there be a difference
between this probability of error and the one that you would obtain if you were to know
α perfectly? (Hint: Notice that α̂ is a random variable, not a deterministic value. You do
not need to estimate the exact probability of error, just showing how it can be evaluated
is sufficient.)

We have no choice but to use the estimated α to use in the MK detection, so

x̂k =argmax f (rk |α̂, xk ) = argmax log f (rk |α̂, xk ) k = 2, 3, . . .

(r −α̂x )2
1 − k N k
=argmax √ e 0
πN0
1 (rk − α̂xk )2
=argmax − log(πN0 ) −
2 N0
2
=argmin (rk − α̂xk ) k = 2, 3, . . . .

Because we have a binary symbol set we can also say that all we have to do is to
compare (rk − α̂)2 − (rk + α̂)2 with the threhold 0 (if larger decide in favor of 1 and
otherwise in favor of −1). Notice that after simplificatios this again corresponds
to the sign detection rule, that is if rk > 0 we decide in favor of 1 and otherwise
−1.
This corresponds to the probability of error:

p 
P (e|α̂) = Q 2α̂2 γ .

where γ = Eb /N0 . Notice that this error probability is dependent on α̂ which is

a Gaussian random variable α̂ ∼ N (α, σα2 ) where the values of σα2 depends on the
number of pilot symbols used. In order to obtain the average probability of error
we need to take the expectation over α̂ which means
Z
1
Z p  − (x−α)2
Pe = P (e|α̂)fα̂ (x)dx = p Q 2x γ e 2σα2 dx.
2
πσα2

[10 minutes]
6. [25 Points] In class we have seen that digitally modulated signals may undergo some phase
and timing uncertainties as they travel through the channel. Suppose we transmit an M-ary
PAM signal where the amplitude A belongs to one of the M PAM symbols and the received
signal is n o
r(t) = Re ejφ Ag(t − τ )ej2πfc t + n(t) 0≤t≤T

where φ is the unknown carrier phase, τ is the unknown timing error, n(t) is the AWGN noise
process with zero mean and constant power spectral density of amplitude N0 /2 and g(t) is
the pulse shaping filter. Suppose the receiver is not equipped with a synchronization circuit
and therefore it processes the received signal as if there is no carrier phase offset or timing
delay. Derive the probability of error in this case? At which values of φ and τ , the probability
of error becomes 21 .

Because we are sending PAM signals, the received signal is expressed as

r(t) = Ag(t − τ ) cos(2πfc t − φ) + n(t).

At the receiver side we correlate with g(t) cos(2πfc t) (not with cos(2πfc t)!!!). You
can also normalize the function as well but it is not critical at this point.

Z T Z T
r= r(t)g(t) cos(2πfc t)dt=A g(t − τ ) cos(2πfc t − φ)g(t) cos(2πfc t)dt + n
0 0
=Aφgg (τ ) cos(φ) + η

where Z T
φgg (τ ) = g(t)g(t − τ )dt.
0
and if g(t)’s were to be aligned (τ = 0) it would have corresponded to the εg . Also,
if τ = kT , φgg (τ ) = 0. The probability of error would be:

r !
6γ
Pe ∝ Q φgg (τ ) cos(φ)
M2 − 1

Obiously if τ = T and/or φ = π/2, 3π/2, 5π/2, . . ., it becomes 1/2.

[20 minutes]