University of Trento

Digital Signal Processing

for Mechatronics
examination rules, examples of examination questions
and exercises with solutions

A.Y. 2021-2022

January 4, 2022
Contents

1 Examination rules and instructions 2

1.1 General written test structure and rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Exceptions and further rules due to COVID-19 emergency . . . . . . . . . . . . . . . . . . . . . 3
1.3 Online examination rules and instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3.1 Online written test instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3.2 Online oral exam instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

2 Open questions examples 6

3 Exercises with solutions 8

3.1 Exercise 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
3.2 Exercise 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.3 Exercise 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3.4 Exercise 4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.5 Exercise 5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.6 Exercise 6 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.7 Exercise 7 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.8 Exercise 8 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.9 Exercise 9 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.10 Exercise 10 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.11 Exercise 11 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.12 Exercise 12 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.13 Exercise 13 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.14 Exercise 14 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.15 Exercise 15 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.16 Exercise 16 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.17 Exercise 17 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.18 Exercise 18 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.19 Exercise 19 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.20 Exercise 20 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.21 Exercise 21 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.22 Exercise 22 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3.23 Exercise 23 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.24 Exercise 24 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.25 Exercise 25 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

4 Multiple–choice and short questions 45

1
Chapter 1

Examination rules and instructions

Two alternative examination modes will be arranged for the course Digital Signal Processing for Mechatronics,
i.e., the express examination mode and the examination mode with project. In general, students are supposed
to take the exam in person in the University premises. However, exceptions and specific rules may apply
for people subjected to restrictions due to the COVID-19 pandemic. Further information on this point are
provided in the last two Section of this introductory chapter. In normal conditions, the following rules apply.

ˆ In the express examination mode, Students are required to take just the written test followed by an
optional oral exam (consisting of a pair of extra questions). The final grade results from the algebraic
sum of the scores of the answers to the written test questions, possibly rounded to the closest integer.
The minimum score to pass the written test (as well as to be admitted to the optional oral
exam) is 18. Round-up oral exams to reach sufficiency are usually not allowed. Usually, the optional
oral exam will change the written test score by no/more than about 10%.

ˆ In the examination mode with project, Students have to choose, develop and submit a Matlab project
following the instructions reported in the introduction to the project list file uploaded in the DOL/Moodle
community of the course. The project and the related report must be submitted by the date
of the written test. During the written test students can choose just one of the theoretical open
questions. However in this case, the oral exam is mandatory and at least one oral question will certainly
be on the project.
The grade to pass the written test and to be admitted to the oral has to be strictly greater
than 50% of the total written test score after excluding one theoretical question.
With this examination mode, the project and the related question(s) will account up to 30% of the final
grade, while the remaining 70% will be based on the results of the written test and on the answers to
the other oral questions.

1.1 General written test structure and rules

The exam is structured in order to evaluate whether and to what extent a student has reached the expected
educational objectives as well as a suitable level of understanding of the course contents. The evaluation of
the written test is based on the following criteria: completeness, correctness, clarity and brevity of answers
and solutions. The written test will consist of three sections structured as follows

ˆ Section 1: 2 open questions on theoretical aspects;

ˆ Section 2: 1 exercise;

2
 ˆ Section 3: from 5 to 10 multiple-choice or short-answer questions (a slight penalty shall be applied to
wrong answers).
In the rest of this document, some examples of questions for each Section are reported along with their
solutions. Of course, the real examination questions, although similar, will differ from those
reported in this file.
During the exam the use of own lecture notes, books, smartphones, tablets, laptops or other off-line or online
resources is not allowed. Just the use of basic scientific calculators (i.e., not provide with file storage capabilty)
is permitted.
The duration of the written test shall be
ˆ 2 hours for the Students opting for the express examination mode;
ˆ 1.5 hours for the Students opting for the examination mode with project. This reduction is motivated
by the fact that these Students have to answer to just one of the questions of Section 1.
As soon as the written test results are published, all Students with a sufficient grade that are
willing to take the optional oral exam are required to inform me immediately by email, to allow
me to schedule the oral exams with no risk of overcrowding.
Upon request, the registration of an unsatisfactory (i.e. too low) pending grade can be postponed to the
following examination session (no more than one). In this case, a Student can take the written test again,
but he/she will lose the pending grade as soon as he/she delivers the test. If instead he/she withdraws, the
pending grade can be retained and shall be registered.
IMPORTANT: The grades of all students that opted for the express examination mode and
that do not ask to take the oral exam or that do not explicitly decline or postpone the decision
on grade acceptance to the following session will be registered in ESSE3.

1.2 Exceptions and further rules due to COVID-19 emergency

As explained above, both written tests and oral exams are supposed to be taken in person following the general
access rules to UNITN premises. However, online exams will be arranged under the following exceptional
circumstances, i.e.
ˆ for everybody, in the case of local lockdown and/or in the case of strong mobility restrictions imposed
by local Authorities due to pandemic worsening;
ˆ for the Students that fall in one of the following categories:
1. Certified vulnerable people or Students with a physical disability;
2. Students with a positive Covid 19 test and/or under quarantine;
3. Students living in areas subjected to travel restrictions.
4. If strictly needed, Students that are supposed to take the exam on the day before, the same day,
or the day after a scheduled COVID-19 vaccination.
All Students of one of the categories specified above have to follow the procedure explained here:
https://www.unitn.it/en/coronavirus.
Any official online exam request must be forwarded to me by the interested Students at least
three days before the date of the written test. Please bear in mind that:
– Late requests will be declined even if well motivated;
– Extemporary, last-minute requests due to sudden health problems will be declined.
– The online exam requests shall be subject to verification. If not adequately supported by some
clear evidence or certification, the pending requests shall be declined.

3
1.3 Online examination rules and instructions
If, due to the reasons and in the cases explained above, the exam has to be taken online, the two questions of
Section 1 of the written test shall be replaced by a MANDATORY full oral exam.
As a consequence, the duration of the online written test shall be 1 hour regardless of the exami-
nation mode. However, the duration of the oral exam is expected to be much longer (at least 3-4 questions).
Again, for the students that opted for the examination mode with project, one of the oral question shall
certainly be about it.
In the case of online exam, the evaluation criteria will be changed as follows:
ˆ for the Students that opt for the express examination mode, 50% of the final grade will be provided by
the written test score, while the remaining 50% will depend on the answers to the oral questions;
ˆ for the Students that opt for the examination mode with project, 50% of the final grade will be provided
by the written test score, 30% will be based on the project and the remaining 20% will depend on the
answers to the other oral questions.

1.3.1 Online written test instructions

Online Students are required to abide by the rules listed below or they will be excluded from the exam.
1. A couple of days before the written test, every student that is authorized to take the exam online will
receive an email with the Zoom link to the examination virtual room as well as an invitation to a
nominal and strictly personal Gdrive folder where he/she will have to to upload the scanned copy of
the sheet where he/she writes the answers to the examination questions. The use of a proper scanner
is preferable, if possible. Otherwise, a smartphone with an ad-hoc App can be used to this purpose. In
any case, STUDENTS ARE WARMLY RECOMMENDED TO TEST THE SETUP FOR DOCUMENT
SCANNING WELL IN ADVANCE, TO BE SURE THAT IT WORKS PROPERLY AND THAT THE
FILES ARE READABLE. In the case of submission problems or unreadable files due to
Student’s negligence, the Student will be excluded from the exam.
2. On the examination day, before starting, every Student must show his/her own ID to the Examiner.
3. Students must be alone in the room and their working desk should be placed in front of a wall, if possible.
4. The desk/table used for the written test must be kept clear of any unnecessary object. Only a plain
scientific calculator, 1-2 sheets of blank paper and one pen can lie on the table.
5. Two electronic devices have to be used to make remote surveillance possible.
ˆ The webcam of a PC or a laptop should be placed quite in front of the Student (at a distance of
about 2 m) so that the Examiner can FULLY and CLEARLY see the Student’s workplace (i.e. not
only the Student’s face, but also his/her hands writing on the sheets and a vast part of the desk).
ˆ Another device (e.g. a smartphone) should be placed next to the back side of the Student in such a
way that the Examiner can see both the Student’s desk and the screen of the first device reasonably
well.
The cameras of both devices and one microphone only (to avoid excessive audio returns) must be kept
on during the whole exam.
6. All the other electronic devices different from those used for the connection (tablets, smartphones, etc...)
must be switched off.
7. Students should always stay in view during the whole test. It is not allowed to leave the room, unless
this is agreed with the Examiner.

4
1.3.2 Online oral exam instructions
The Zoom link to the virtual room to take the oral exam as well as the examination schedule will be reported
in the result sheets uploaded in the DOL/Moodle community of the course, whenever ready.
During the oral exam:

1. Every student must show his/her own ID to the Examiner before the beginning of the exam.

2. Just one electronic device is needed for remote surveillance. The camera should be placed quite in front
of the student (at a distance of about 2 m) so that the Examiner can FULLY and CLEARLY see the
Student’s workplace (i.e. not only the Student’s face, but also his/her hands writing on the sheets and
a vast part of the desk). Camera and microphone must be kept on during the whole exam.

3. Students might be asked to show to the Examiner the workplace and the room. Also, students have
to prove to be alone in the room and that all nearby electronic devices (excluding the one used for the
Zoom connection) are off. On the desk, Students can keep just some blank sheets and a pen. No books
or lecture notes are allowed.

4. During the oral, Students might be asked to write formulas or sketch diagrams. These activities have to
be monitored through the camera. At the end, all the notes written during the exam have to be scanned
and uploaded into the very same Gdrive folder in which the written test was previously uploaded. The
exact submission rules will be explained at the examination time.

5
Chapter 2

Open questions examples

1. Derive the continuous-time Fourier transform of AM and PM signals.

2. Provide the definition of strict-sense stationary, wide-sense stationary, cyclostationary and ergodic pro-
cesses. Explain the relationship between them and why the ergodic properties are particularly useful in
practice.

3. Prove that the Fourier transform of the convolution of two signals is given by the product of the trans-
forms of the two signals both in the continuous-time and in the discrete-time case.

4. Compute the mean value, the autocorrelation and the power spectral density (PSD) of the WSS process
y(n) at the output of a discrete-time LTI system with impulse response h(n), when it is stimulated by
a WSS process x(n) with mean µx and autocorrelation ϕx (n).

5. Show why the convolution operator is associative, commutative and distributive with respect to the
addition. As a result, derive the impulse response of the cascade and the parallel of M linear systems.

6. Spectral estimation of deterministic signals: methodology, and main uncertainty sources.

7. Provide a definition of spectral resolution and explain how it can be determined. Describe the tradeoff
between time vs. spectral resolution in the case of nonstationary signals and explain the meaning of
“uncertainty principle”.

8. Show the effect of down-sampling/decimation on a generic signal both graphically and mathematically.
What is the main risk of down-sampling? How can we solve the problem?

9. Periodograms and modified periodograms: definitions and statistical properties in terms of bias and
variance. Explain why and under which assumptions the average periodogram can improve spectral
estimation.

10. FIR filter design based on the windowing method: general idea and motivation, role and features of
different windows and filter design steps. Briefly explain pros and cons of the algorithmic FIR filter
design techniques compared to the classic windowing method.

11. Describe and explain in detail the design steps of a band-pass IIR filter.

12. Describe and explain the structure of an ideal C/D converter and a real Analog-to-Digital Converter
(ADC). Sketch the schematic and explain the principle of operation of a flash ADC circuit.

13. Describe and explain the structure of an ideal D/C converter and a real Digital-to-Analog Converter
(DAC). Sketch the schematic and briefly explain the principle of operation of an R − 2R DAC.

6
14. Provide the definition of system linearity, causality, time-invariance and BIBO stability. Also, prove why
in an LTI discrete-time system asymptotic stability is guaranteed if and only if all poles of the transfer
function lie within the unit circle.

15. Describe the decimation-in-time FFT algorithm over N = 2ν samples and explain why algorithm com-
plexity is O(N log2 N ).

16. Explain the relationship between the DIT FFT, the DIF FFT algorithms and the respective inverse Fast
Fourier Transforms. Compute and compare the number of real-valued operations of a classic FFT with
respect to a standard DFT when N = 1024 points are considered.

17. Show why an FIR system of type I exhibits a linear phase response. Where are the zeros and the poles
of its transfer function located in the z complex plane?

18. Explain the effect of up-sampling and interpolation both graphically and analytically. What is the main
problem of ideal interpolators and how can we mitigate it?

19. Considering an ideal RC analog low-pass filter, show how this system can be discretized and described
by a constant coefficient linear difference equation.

20. Provide the definition of deterministic autocorrelation in the continuous-time and discrete-time domain
and compute the corresponding Fourier transforms.

21. What properties should the impulse response have to make an LTI system BIBO stable and causal?
Justify your answer analytically.

22. Prove and explain why a filter with a linear phase response does not introduce any phase distortion.
Can we design IIR filter without phase distortion? Why?

23. Provide the definition of random power and energy signals. Prove why the SSS and WSS random
processes can just be power signals.

24. Describe the main sources of vibrations in a gearbox and sketch qualitatively how the spectrum should
look like.

25. Explain and prove why the discretization of an LTI system based on the bilinear transform preserves
system stability and how the analog frequency Ω is mapped/transformed into the normalized one ω.

7
Chapter 3

Exercises with solutions

3.1 Exercise 1
Design a low–pass discrete–time filter with the following specifications:
ˆ magnitude of frequency response constant within 0.5dB for |ω| ≤ π5 ;
ˆ magnitude of frequency response at ω = 0 equal to 1;
ˆ attenuation in the stopband of at least 20 dB for ω ≥ 0.48π.
After general considerations about the complexity and the order of Butterworth, Chebyshev and elliptic filters
(2 points), use the bilinear transformation method with T = 2 to design a Butterworth filter (6 points).

Solution
a) In general, for given filter specifications, the order of a Butterworth filter is larger then a Chebyshev one,
which is in turn larger than an Elliptic one, i.e.
Nbutter > Ncheby > Nelliptic
As a result, for the same specifications, the computational burden of a Chebyshev filter is smaller than a
Butterworth and larger than the corresponding elliptic filter.
b) The first step of an IIR filter design relies on the transformation of the specifications from the digital to
the analog domain. Since we are dealing with a low–pass filter, if we use the bilinear transformation, the
pass–band and the stop–band edges become:
 π

 Ω = tan = 0.325
2 ω   p
10
Ω = tan ⇒ (3.1)
T 2 

 Ω = tan(0.24π) = 0.939
s

After transforming the maximum approximation errors from dB to a linear scale, i.e.
(
20 log10 (1 − δ1 ) ≥ −0.5dB → 1 − δ1 ≥ 0.944
(3.2)
20 log10 (δ2 ) ≥ −20dB → δ2 ≤ 0.1
the order of the filter and the cut-off frequency result respectively from
 h .  i 
log δ12 − 1 1
(1−δ )2 − 1
N = 
2
 1  = [3.157] = 4
 (3.3)
 Ω 
2 log Ωsp

8
and
Ω
Ωc = r s  = 0.529 (3.4)
1
2N
δ22
−1

The poles of the analog Butterworth filter are given by:

sk = Ωc ejϕk (3.5)

where
π π
ϕk = + (2k + 1) ∀k = 0, 1, . . . , N − 1 (3.6)
2 2N
Therefore, in this case
(
ϕ0 = 58 π → s0,3 = Ωc cos(ϕ0 ) ± jΩc sin(ϕ0 ) = −0.202 ± j0.489
ϕ1 = 87 π → s1,2 = Ωc cos(ϕ1 ) ± jΩc sin(ϕ1 ) = −0.489 ± j0.202

We know that the poles of a Butterworth filter are located on a semicircle centered in the origin of the complex
plane and with radius Ωc . Thus, the transfer function is:
H0
H(s) = QN −1 (3.7)
k=0 (s − sk )

where
N
Y −1
H(0) = sk = ΩN 4
c = Ωc = 0.078 (3.8)
k=0

In the case considered (3.7) becomes

Ω4c 0.078
H(s) = 2 2 2 2
= 2
(s − 2ℜ{s0 }s + |s0 | )(s − 2ℜ{s1 }s + |s1 | ) (s + 0.404s + 0.28)(s2 + 0.978s + 0.28)

Finally, by replacing
2 1 − z −1
s= (3.9)
T 1 + z −1
into (3.7), the digital filter with transfer function

0.078(1 + z −1 )4
H(z) = ...
[(1 − z −1 )2 + 0.404(1 − z −1 )(1 + z −1 ) + 0.28(1 + z −1 )2 ]

1 (3.10)
... =
[(1 − z −1 )2
+ 0.978(1 − + z −1 )(1
+ 0.28(1 +z −1 ) z −1 )2 ]
−1
0.02056(1 + z ) 4
=
1 − 1.494z −1 + 1.1993z −2 − 0.4465z −3 + 0.072z −4
finally results.

9
3.2 Exercise 2
Let x(n) be a white noise with zero mean and variance equal to 2 that is filtered by a discrete–time system
with impulse response h(n) = δ(n) − 2δ(n − 3) − 3δ(n − 5). Derive mean, autocorrelation function and power
spectrum fo the filter output process w(n) (5 points). Then the process w(n) is used as input of a 4-bit
quantizer Q[·], with full scale range [−1, 1]. Assuming that the quantizer performs rounding compute mean
2 /σ 2 (3
and autocorrelation of the quantization error e(n) and compute the signal/quantization noise ratio σw e
points).

Solution
The impulse response of the system is

h(n) = δ(n) − 2δ(n − 3) − 3δ(n − 5)

and the input is a white noise µx = 0 and σx2 = 2. The frequency response of h(n) is

H(ejω ) = 1 − 2e−jω3 − 3e−jω5 (3.11)

a) We know that if E{x(n)} = 0 then E{w(n)} = 0. Moreover, the PSD of the output WSS process is given
by

Φw (ejω ) = |H(ejω )|2 Φx (ejω ) = σx2 |1 − 2e−jω3 − 3e−jω5 |2

= 2[(1 − 2 cos(3ω) − 3 cos(5ω))2 + (2 sin(3ω) + 3 sin(5ω))2 ] (3.12)
= 28 − 8 cos(3ω) − 12 cos(5ω) + 24 cos(2ω)

Since, the autocorrelation of a WSS process results from the inverse Fourier transform of (3.12), it follows
that:

ϕω (n) = F −1 [σx2 (1 − 2e−jω3 − 3e−jω5 )(1 − 2ejω3 − 3ejω5 )]

(3.13)
= 28δ(n) − 4δ(n + 3) − 4δ(n − 3) − 6δ(n + 5) − 6δ(n − 5) + 12δ(n + 2) + 12δ(n − 2)

where Z π
1
σx2 ϕ(0) = Φ(ejω )dw = Φ(ejω ) =
2π −π

b) Since the full scale range of the quantizer is [−1, 1], the rounding error is uniformly distributed within
[−1/23 , 1/23 ]. Therefore, E{e(n)} = 0 and

∆2 1
σe2 = = = 1.3 · 10−3 (3.14)
12 12 · 26
Moreover, assuming that the quantization noise is white, it results that

ϕe (m) = σe2 δ(m) (3.15)

and the SNR becomes

2
 
σm ϕω (0) 28 28
= = → 10 log10 = 43 dB (3.16)
2
σe σe2 1.3 · 10−3 1.3 · 10−3

10
3.3 Exercise 3
The following three discrete–time filters are identified specifying their non–zero poles and zeros:

ˆ 2nd –order IIR with poles in a and b and zeros in c and 1/c, with c being real;

ˆ 2nd –order IIR with poles in b and b∗ and zero in -1;

ˆ 2nd –order IIR with a double pole in 1/a and zeros in e and e∗ , with e being complex and not real.

For each of the three filters derive:

ˆ the transfer functions and the corresponding data-flow graphs (2 points);

ˆ the filter(s) to be put in cascade to obtain a global linear–phase FIR filter (3 points);

ˆ the data-flow graph of the resulting FIR filter with a minimal computational complexity (3 points).

Solution
a) The transfer functions of the three systems considered are listed below and result immediately from the
given specifications. The corresponding data–flow graphs are shown in Figs. 3.1-3.3.

1.


(1 − cz −1 ) 1 − 1c z −1 1 − c + 1c z −1 + z −2
H1 (z) = = (3.17)
(1 − az −1 )(1 − bz −1 ) 1 − (a + b)z −1 + abz −2

z −1
 

a+b − c + 1c

z −1
-ab 1

Figure 3.1: Data-flow graph of H1 (z).

2.
1 + z −1 1 + z −1
H2 (z) = = (3.18)
(1 − bz −1 )(1 − b∗ z −1 ) 1 − 2ℜ{b}z −1 + |b|2 z −2

3.
(1 − ez −1 )(1 − e∗ z −1 ) 1 − 2ℜ{e}z −1 + |e|2 z −2
H3 (z) =
2 = (3.19)
1 − a1 z −1 1 − a2 z −1 + a12 z −2

11
 1

z −1

2<{b} 1

z −1
−|b|2

Figure 3.2: Data-flow graph of H2 (z).

1

z −1
2
a −2<{e}

z −1
− a12 |e|2

Figure 3.3: Data-flow graph H3 (z).

b) The transfer function of the overall cascaded system including the unknown one is

Htot (z) = H1 (z)H2 (z)H3 (z)Hx (z) (3.20)

Since Htot (z) must be a linear-phase FIR system, then it must include only zeros which are symmetric with
respect to the unit circle, and all poles have to be compensated, i.e.
    
1 1 1
Htot (z) = (1 − cz −1 ) 1 − z −1 (1 + z −1 )(1 − ez −1 )(1 − e⋆ z −1 ) 1 − z −1 1 − ∗ z −1
c e e

Therefore, the transfer function of the unknown system is:

Htot (z)
Hx = =
H1 (z)H2 (z)H3 (z)
    
−1 −2

−1 2 −2

2 −1 1 −2 1 −1 1 −2
= 1 − (a + b)z + abz 1 − 2ℜ{b}z + |b| z 1 − z + 2z 1 − 2ℜ ∗ z + z
a a e |e|

The resulting FIR system is of type II, since M = 7. Indeed, we have a zero in -1.
       
1 −1 −2 −1

−1 2 −2

1 −1 1 2 −2
Htot(z) = 1 − c + z +z 1+z 1 − 2ℜ{e}z + |e| z 1 − 2ℜ z + z
c e e
= 1 + α1 z −1 + α2 z −2 + +α3 z −3 + +α4 z −4 + +α5 z −5 + +α6 z −6 + z −7

c) Finally, the data-flow graph of the FIR filter with minimal computational complexity is shown in Fig. 3.4.

12
 z −1 z −1 z −1 z −1
x(n)

z −1

α1 α2 α3

z −1 z −1 z −1 z −1

y(n)

Figure 3.4: Structure of the total FIR filter with minimal computational complexity.

3.4 Exercise 4
Given an analog system with transfer function
s + s1
Ha (s) = (3.21)
(s + s1 )2 + s22
where s1 and s2 are two real constant coefficient, determine the transfer function of the corresponding discrete–
time system using the impulse invariance method with sampling period equal to T (4 points). Find the values
of s1 , s2 and and T for which the system is stable (1 point). Finally, draw the signal flow graph direct form
I, II and II transposed (3 points).

Solution
a) First of all, the transfer function of the system can be rewritten using a partial fraction decomposition, i.e.
A1 A2
Ha (s) = + (3.22)
s + (s1 + js2 ) s + (s1 − js2 )
where
s + s1 1
A1 = lim =
s→−s1 −js2 s + s1 − js2 2
(3.23)
s + s1 1
A2 = lim =
s→−s1 +js2 s + s1 + js2 2
Therefore, the impulse response of the continuous-time system is
1  −(s1 +js2 )t 
h(t) = e + e−(s1 −js2 )t u(t) (3.24)
2
By applying the impulse invariance method, it follows that
T  −(s1 +js2 )nT 
h(n) = e + e−(s1 −js2 )nT u(n) (3.25)
2
 
T 1 1 T − T e−s1 T cos(s2 T )z −1
H(z) = + = (3.26)
2 1 − e(−s1 +js2 )T z −1 1 − e(−s1 −js2 )T z −1 1 − 2e−s1 T cos(s2 T )z −1 + e−2s1 T z −2
b) Observe that the discretized system is certainly stable for s1 > 0 and ∀T , regardless of the value of s2 .

c) The data-flow graphs of direct forms I, II and II transposed are shown in Figs. 3.5-3.7

13
 T

z −1 z −1

e−s1T cos(s2T ) 2e−s1T cos(s2T )

z −2
−e−2s1T

Figure 3.5: Direct form I.

z −1

2e−s1T cos(s2T ) e−s1T cos(s2T )

z −1
−e−2s1T

Figure 3.6: Direct form II.

Figure 3.7: Transposed form II.

3.5 Exercise 5
1
Let x(n) be an instance of a discrete-time random process with power spectral density Φx (ω) = 1+ω 2 . If the
signal is quantized with quantization step ∆, and the error e(n) is uniformly distributed between [−∆/2, ∆/2],
find the mean and the autocorrelation of the quantization noise (1 point). Afterwards, compute the signal–
to–noise ratio σx2 /σe2 (3 points), and

the power

nof
 the output noise if e(n) passes through a discrete–time filter
n
with impulse response h(n) = 21 13 + − 13 u(n) (4 points).

Solution
a) The mean of the quantization noise is approximately zero. The autocorrelation is
(
0 m ̸= 0
ϕe (m) = E{e(m)e(n + m)} = 2
(3.27)
σe m = 0

14
 ∆2 2−2B
where σe2 = 12 = 12 .

b) The SNR at the output of the system is

σx2 σx2
SN R = = (3.28)
σe2 ∆2 /12

where Z π
1 1 1
σx2 = 2
dω = [atan(ω)]π−π = 1.464 (3.29)
2π −π 1+ω 2π
c) Finally, the output noise power is given by:
∞
X Z π
σ2
σt2 = σe2 |h(n)| = e 2
H(ejω )2 dω (3.30)
n=−∞
2π −π

where (

1 n
2 9 n = 2k
|h(n)| = (3.31)
0 n = 2k + 1
Therefore, the output noise power results from
∞ ∞  n  n  !
X σe2 X 1 1 σ2 1 1 81 ∆2
σt2 = σe2 2
|h(n)| = + − = e 1 + 1 = (3.32)
2 9 9 2 1− 9 1+ 9
80 12
n=0 n=0

15
3.6 Exercise 6
Consider the filter described by the following linear constant coefficient difference equation:
 
1 1 1 1
y(n) = y(n − 1) + y(n − 2) + x(n) + √ + j √ x(n − 1) (3.33)
2 3 2 2
ˆ Derive the transfer function of the system (1 point);
ˆ Draw the signal flow graph of both direct and transposed form II structures (2 points);
ˆ Identify the minimum–order FIR filter that, used in cascade to the system described above, globally
implements a linear–phase FIR filter (4 points);
ˆ Finally, draw the data-flow graph of the resulting linear–phase FIR filter (1 point).

Solution
a) To compute the transfer function of the system, we can just apply the Z−transform to (3.33), i.e.
 
1 1 1 1
Y (z) = z −1 Y (z) + z −2 Y (z) + X(z) + √ + j √ z −1 X(z) (3.34)
2 3 2 2
Therefore, the transfer function of the system is:
 
Y (z) 1+ √1 + j √1
2 2
z −1
H(z) = = (3.35)
X(z) 1 − 21 z −1 − 13 z −2
b) The data-flow graphs of the system implemented using the direct form of type II and the transposed form
of type II are shown in Figs. 3.8-3.9, respectively.

x(n) y(n)

z −1
1
2 √1 + j √12
2

z −1
1
3

Figure 3.8: Implementation of the filter based on the direct form II.

c) In order to have a linear-phase FIR filter, the system cascaded to the system based on (3.33) (see Fig. 3.10)
should be able to: i) cancel the poles of (3.35); ii) ensure that the zeros of the global system are properly
symmetric (in pairs or in quadruples) with respect to the unit circle. Given that only the complex zero
z0 = − √12 (1 + j) appears in (3.35) and since |z0 | = 1, the transfer function of the wanted system is
    
1 1 1 1
H1 (z) = 1 + √ − j √ z −1 1 − z −1 − z −2 (3.36)
2 2 2 3

16
 x(n) y(n)

z −1
√1 + j √12 1
2 2

z −1
1
3

Figure 3.9: Implementation of the filter based on the transposed form II.

H(z) H1(z)

HF IR

Figure 3.10: Cascade of the given and the unknown systems.

As a result, the transfer function of the global FIR filter is:

  
1 −1 1 −1
√
HFIR (z) = 1 + √ (1 + j)z 1 + √ (1 − j)z = 1 − 2z −1 + z −2 (3.37)
2 2
d) Finally, the data-flow graph (tapped delay line) of the FIR filter is shown in Fig. 3.11.

x(n) z −1 z −1

√
2
y(n)

Figure 3.11: Data-flow graph of the global FIR filter.

3.7 Exercise 7
Design a low–pass, discrete–time filter with the magnitude of the frequency response in the passband constant
within 1.2 dB for ω ≤ 0.22π and attenuation in the stopband greater or equal to 12 dB for ω ≥ 0.38π. Derive
the order N and the frequency Ωc of a Butterworth filter fulfilling such specifications, using the bilinear trans-
formation method with T = 2 s. Identify the position of the poles in the s plane and the transfer function of

17
the continuous–time filter. Write the transfer function of the discrete–time filter (8 points).

Solution
The maxim approximation errors of the filter can be rewritten on a linear scale as

20 log10 (1 − δ1 ) ≥ −1.2 ω ≤ 0.22π (passband) (1 − δ1 ) = 0.87

(3.38)
20 log10 (δ2 ) ≤ −12 ω ≥ 0.38π (stopband) (δ2 ) = 0.251

By applying the bilinear transformation for T = 2 it follows that

2 ω 
p
Ωp = tan = 0.36rad/s
T 2 (3.39)
2 ω 
s
Ωs = tan = 0.68rad/s
T 2
So the order of the filter and the cut-off frequency are given respectively by
 h .  i 
log δ12 − 1 1
(1−δ ) 2 − 1
N = 
2
 1  = [3.024] = 4
 (3.40)
 Ω 
2 log Ωps

and
Ω
Ωc = r s  = 0.485rad/s (3.41)
1
2N
δ22
−1

The poles of the analog Butterworth filter are given by:

sk = Ωc ejϕk (3.42)

where ϕk = π2 + (2k + 1) 2N
π
for k = 0, 1, . . . , N − 1. We know that the poles of a Butterworth filter are located
on a semicircle centered in the origin of the complex plane and with radius Ωc (see Fig. 3.12). Thus, in the
case considered the poles of the analog filter are:

s0,3 = Ωc cos(ϕ0 ) ± jΩc sin(ϕ0 ) = −0.184 ± j0.44

s1,2 = Ωc cos(ϕ1 ) ± jΩc sin(ϕ1 ) = −0.44 ± j0.184 (3.43)

and the transfer function is

H0 0.053
H(s) = QN −1 = 2 (3.44)
k=0 (s − sk )
(s + 0.368s + 0.2304)(s2 + 0.88s + 0.2304)

where
N
Y −1
H(0) = sk = ΩN 4
c = Ωc = 0.053 (3.45)
k=0
Finally, by replacing
2 1 − z −1
s= (3.46)
T 1 + z −1
into (3.44) of the continuous-time filter, the transfer function of the discrete-time system is

(1 + z −1 )4 C0 C1
H(z) = (3.47)
(1 − A0 z −1 + B0 z −2 ) (1 − A1 z −1 + B1 z −2 )

18
 k=0

k=1

k=2

k=3

Figure 3.12: Position of the poles of the analog filter in the complex plane
.

where for each given second-order polynomial at the denominator of (3.44), we have that s2 +αk s+βk 1−z −1
=
s= T2
1+z −1
1−Ak z −1 +Bk z −2
(1+z −1 )2 Ck
and for T = 2 s it results that

 1

 Ck =

 1 + αk + βk
 2 − 2βk
Ak = (3.48)

 1 + αk + βk

 βk − αk + 1

 Bk =
1 + αk + βk

19
3.8 Exercise 8
Consider the discrete–time LTI system described by the following finite difference equation
y(n) − 5y(n − 1) = x(n) − 2x(n − 1) (3.49)
ˆ Derive the transfer function of the system and find its region of convergence (2 points);
ˆ State if the system is FIR or IIR, if it is stable or not and compute its Fourier frequency transform (if
applies) (2 points);
ˆ Find the impulse response of the system and the result of the convolution between such an impulse
response and a rectangular input sequence x(n):
(
1 n = 0, . . . , 4
x(n) = (3.50)
0 otherwise

Solution
a) To compute the transfer function, we can just apply the Z–transform to (3.49)
1 − 2z −1
Y (z) − 5z −1 Y (z) = X(z) − 2z −1 X(z) ⇒ H(z) = (3.51)
1 − 5z −1
Assuming initial rest conditions with a single pole in z = 5, then the region of convergence is
ROC = {z |z| > 5} (3.52)

b) The system is of type IIR and it is unstable, as the only pole available is not inside the unit circle. Since
the unit circle does not belong to the ROC, then the Fourier transform of h(n) does not exists.

c) The impulse response results from the inverse Z –transform of H(z) which is given by
1 2z −1 Z −1
H(z) = − −−−→ h(n) = 5n u(n) − 2 · 5n−1 u(n − 1)
1 − 5z −1 1 − 5z −1 (3.53)
n−1
= δ(n) + 3 · 5 u(n − 1)
Now it is possible to compute the convolution between the h(n) and the input sequence (3.50), i.e.
h i
y(n) = x(n) ∗ h(n) = x(n) + 3 5(n−1) u(n − 1) ∗ x(n) = (3.54)


 0 n<0



 δ(n) n=0

  n

 1



 1−

 1 + 3 · 5(n−1) 5
1≤n≤4
= 1 (3.55)

 1 −

   5

 1 5

 1 −

 5

 3 · 5(n−1) n>4

 1

 1−
5
Observe that, since the system is not BIBO stable, the convolution sum tends to diverge exponentially. For
instance y(0) = 1, y(1) = 4, y(2) = 19, y(3) = 94, y(4) = 469, y(5) = 2343, etc...

20
3.9 Exercise 9
Given the system represented by the data-flow graph shown in Figure (3.9)

ˆ Compute the transfer function H(z) = Y (z)/X(z) of the system (6 points);

ˆ Draw the flow graph of the same system in the direct and transposed form II (2 points).

Solution
a) The exercise relies on a data-flow graph analysis. Given the names of the nodes shown in Fig. 3.9, it follows
that
A(z) = X(z) + B(z) + F (z)
B(z) = C(z)z −1
C(z) = aB(z) + D(z)
D(z) = A(z)z −1 (3.56)
E(z) = D(z)
F (z) = E(z)z −1 + bE(z) = E(z)(b + z −1 )
Y (z) = D(z)

After some algebraic steps

z −1
B(z) = (aB(z) + D(z))z −1 = (aB(z) + Y (z))z −1 → B(z) = Y (z) (3.57)
1 − az −1

A(z) = X(z) + B(z) + (b + z −1 )A(z)z −1

z −1
Y (z)z = X(z) + Y (z) + (b + z −1 )Y (z)
1 − az −1
(3.58)
Y (z)(1 − az −1 ) = z −2 Y (z) + (bz −1 + z −2 )Y (z)(1 − az −1 ) + z −1 (1 − az −1 )Y (z)
Y (z) z −1 − az −2
=
X(z) 1 − (a + b)z −1 + (ab − 2)z −2 + az −3
b) The direct and transposed data-flow graphs associated with the same systems are shown in Fig. 3.13(a)-(b).

21
 x(n) y(n)

z −1 z −1
a+b y(n) x(n) a+b

z −1 z −1
2 − ab −a −a 2 − ab

z −1 z −1
−a −a

(a) Direct form II. (b) Transposed form II.

Figure 3.13: Data-flow graphs of the system equivalent to that shown in Fig. 3.9.

3.10 Exercise 10
Using the bilinear transformation method, design a discrete-time Butterworth filter with T = 1 and the
following specifications:
ˆ minimum passband gain −1 dB at ωp = 0.25π;
ˆ maximum stopband gain −10 dB at ωs = 0.3π;
Identify the position of poles in the s plane, compute the frequency response of the continuous-time filter and
explain the procedure to obtain the transfer function of the discrete-time filter (8 points).

Solution
The minimum passband magnitude response and the maximum stopband magnitude response of the filter on
a linear scale are (
20 log10 (1 − δ1 ) = −1 → (1 − δ1 ) = 0.89
(3.59)
20 log10 (δ2 ) = −10 → δ2 = 0.316
Now we use the bilinear transformation to obtain the pass-band and stop-band edge frequencies of the analog
filter for T = 1, i.e. (
ω

Ωp = T2 tan 2p = 0.828 rad/s

(3.60)
Ωs = T2 tan ω2s = 1.019 rad/s
Therefore, the order and the cut-off frequency of the filter are given respectively by:
     
1 1
log −1 −1
 δ22 (1−δ1 )2 
N =     = [8.52] = 9
 (3.61)
 Ωs 
2 log Ω
 p 

and
Ωs
Ωc = s  = 0.9019 rad/s (3.62)
1
2N
−1
δ22

22
The poles of the analog Butterworth filter are given by:
sk = Ωc ejϕk (3.63)
where
π π
ϕk = + (2k + 1) ∀k = 0, 1, . . . , N − 1
2 2N
5
k
 = 0, 8 ⇒ 9π



 ⇒ 2
π 2k + 1 k = 1, 7 3π (3.64)
ϕk = + π= k 7
= 2, 6 ⇒ 9π
2 18 


 k = 3, 5 ⇒ 8

 9π

k =4 ⇒ π
Therefore, in this case


 ϕ0,8 = 59 π → s0,8 = Ωc cos(ϕ0 ) ± jΩc sin(ϕ0 ) = −0.16 ± j0.89



 2
 ϕ1,7 = 3 π → s1,7 = Ωc cos(ϕ1 ) ± jΩc sin(ϕ1 ) = −0.45 ± j0.78
ϕ2,6 = 97 π → s1,7 = Ωc cos(ϕ1 ) ± jΩc sin(ϕ1 ) = −0.69 ± j0.58



 ϕ3,5 = 98 π → s1,7 = Ωc cos(ϕ1 ) ± jΩc sin(ϕ1 ) = −0.85 ± j0.31



ϕ4 = −0.9109
We know that the poles of a Butterworth filter are located on a semicircle centered in the origin of the complex
plane and with radius Ωc . Thus, the transfer function is:
ΩNc
H(s) = (3.65)
(s2 − 2ℜs0 + |s0 |)(s2 − 2ℜ{s1 } + |s1 |2 )(s2 − 2ℜ{s2 } + |s2 |2 )(s2 − 2ℜ{s3 } + |s3 |2 )(s − s4 )
Finally, by replacing
2 1 − z −1
s= (3.66)
T 1 + z −1
into (3.7), the transfer function of the IIR digital filters finally results, i.e.

H(z) = H(s) 1−z −1

s= T2
1+z −1
(3.67)
(1 + z −1 )9 C0 C1 C2 C3 C4
=
(1 − A0 z −1 + B0 z −2 )(1 − A2 z −1 + B2 z −2 )(1 − A3 z −1 + B3 z −2 )(1 − A4 z −1 )
where for polynomials s2 − 2ℜ{sk }s + |sk |2 for k = 0, 1, 2, 3 then

 8 − 2|sk |
 Ak
 =

 4 − 2ℜ{sk } + |sk |






4 + 2ℜ{sk } + |sk |2
Bk = (3.68)

 4 − 2ℜ{sk } + |sk |2







 1
 Ck =
4 − 2ℜ{sk } + |sk |2
1−z −1
On the contrary, for k = 4, it ensues from s − s4 = 2 1+z −1 − s4 that
 1

 Ck =

 2 − sk
(3.69)

 2 + sk

 Ak =
2 − sk

23
3.11 Exercise 11
Let y(n) = x(n) + e(n) be the sequence generated by a 12–bit, bipolar ADC with a full scale F S = 2 V, where
x(n) is the ideal (infinite precision) input signal, while e(n) is a zero–mean, uniformly distributed quantization
noise.

1. Compute the ideal quantization step Q of the ADC (1 point).

2. Compute the ENOB of the ADC, assuming that the input signal is a full-scale and uniformly distributed
random sequence and the power of the additional acquisition noise (i.e. not due to quantization) is 20
times larger than e(n) (1 point).

3. If the maximum frequency of an input full-scale sinewave to be acquired is about 10 kHz, how large
should the ADC sampling frequency be to avoid sample–and–hold amplifiers (recall that a sample–and–
hold device is not needed if the maximum variation of the signal within a sampling period is smaller
than the quantization step)? (2 points).

4. Assume that the quantization noise e(n) is filtered by a discrete-time filter with the following impulse
response:  n  n 
1 1 1
h(n) = + − u(n) (3.70)
2 5 5
compute the variance of the noise at the output of the filter (4 points).

Solution
a) The quantization step of the quantizer is

2F S 4 1
Q= B
= 12 = 10 (3.71)
2 2 2
Q2
Therefore, the mean value and the variance of the quantization noise are respectively µe = 0 and σe2 = 12 .
b) From the definition of effective number of bits it follows that

1 σ2
ENOB = B + log2 2 e 2 = 9.8 (3.72)
2 σe + σω
|{z}
2
20σE

c) Recalling the condition max( dx(t)

dt )Ts < Q, where x(t) is the input signal, if x(t) = F S sin(2πf0 t + ϕ) with
f0 = 10 kHz we have that
2F S
2πf0 · F S · Ts < → fs > 2B 2πf0 > 257 MHz (3.73)
2B+1
d) Given the filter with impulse response (3.70), the variance of the noise at the output of the filter is
∞ ∞    n    !
X σe2 X 1 n 1 1 n σe2 1 1
2 2 2
σt = σe |h(n)| = + +2 − = 1 + 1 ≃ σe2 (3.74)
n=−∞
4 25 25 25 2 1 − 25 1 + 25
n=0

24
3.12 Exercise 12
Let us consider a discrete-time system with a sampling frequency fs = 10 kHz. Design a linear-phase FIR
system able to cancel possible disturbances at DC (i.e. at 0 Hz), and at all harmonic frequencies of fd , with
fd = 2.5 kHz (5 points). Determine the impulse response and the system type (I, II, III or IV). Finally, draw
the structure of the system (3 points).

Solution
a) The frequency response of the wanted system is shown in Fig. 3.14. To implement this system we can just
assume to have zeros at frequencies 2πfd /fs k, for k = 0, 1, 2, . . . . Thus, the transfer function of the wanted
FIR system is
π π
H(z) = (1 − z −1 )(1 + z −1 )(1 − e−j 2 z −1 )(1 − e−j 2 z −1 )
(3.75)
= (1 − z −2 )(1 + z −2 ) = (1 − z −4 )

Figure 3.14: Frequency response of the wanted system.

b) The filter exhibits an antisymmetric impulse response with M = 4. Therefore, it is a linear-phase FIR
system of type III. The corresponding data-flow graph is shown in Fig. 3.15.

Figure 3.15: Tapped delay line of the wanted linear-phase FIR system.

25
3.13 Exercise 13
Assuming a sampling period T = 1, design a Butterworth discrete-time low pass filter with the following
specifications (6 points):
ˆ Magnitude ripple of no more than 0.2 dB for |ω| ≤ π
6

ˆ Stopband attenuation of at least 25 dB for |ω| ≥ 0.45π

In addition, compute the order of a Chebyshev I filter with the same specification and explain why it is dif-
ferent from the Butterworth case (2 points).

Solution
a) The minimum passband magnitude response and the maximum stopband magnitude response of the filter
on a linear scale are respectively
20 log10 (1 − δ1 ) ≥ −0.2 → 1 − δ1 = 0.977
(3.76)
20 log10 (δ2 ) ≤ −25 → δ2 = 0.056
The passband and stopband edge of the corresponding analog filter are instead
2 ω  2 ω 
p s
Ωp = tan = 0.536 Ωs = tan = 1.708 (3.77)
T 2 T 2
Therefore, the order and the cut-off frequency of the filter are given respectively by:
 h .  i 
log δ12 − 1 1
(1−δ )2
− 1
N = 
2
 1  = [3.8] = 4
 (3.78)
 Ωs 
2 log Ωp

and
Ω
Ωc = r s  = 0.8321 (3.79)
1
2N
δ22
−1

The poles of the analog Butterworth filter are located on a semicircle centered in the origin of the complex
plane and with radius Ωc , i.e.
sk = Ωc ejϕk (3.80)
where
π π
ϕk = + (2k + 1) ∀k = 0, 1, . . . , N − 1 (3.81)
2 2N
In the case considered, we have that
(
ϕ0 ,3 = ± 58 π → s0,4 = Ωc cos(ϕ0 ) ± jΩc sin(ϕ0 ) = −0.3184 ± j0.7688
ϕ1 ,2 = ± 78 π → s1,3 = Ωc cos(ϕ1 ) ± jΩc sin(ϕ1 ) = −0.7688 ± j0.3184
Thus, the transfer function of the filter is:
H0 Ω5c
H(s) = QN −1 =
k=0 (s − sk )
(s − s0 )(s − s4 )(s − s1 )(s − s3 )
Ω5c (3.82)
= 2
(s − 2ℜ{s0 }s + |s0 | )(s2 − 2ℜ{s1 }s + |s1 |2 )
2

0.4795
= 2
(s + 0.6368s + 0.6924)(s2 + 1.5376s + 0.6924)

26
Finally, by applying the bilinear transform the transfer function of the discrete-time IIR Butterworth filter is
0.0103(1 + z −1 )4
H(z) = H(s)| 2 1−z−1 = (3.83)
T 1+z −1 1 − 1.9604z −1 + 1.7254z −2 − 0.7188z −3 + 0.1193z −4
b) The order of a Chebyshev I filter fulfilling the same specifications is given by:
 
h p p . i
 log 1 − δ22 + 1 − δ22 (1 + ϵ2 ) ϵδ2 
 
N =  r  ! =3 (3.84)

 Ω Ω
2 
 log Ωp +s s
−1 
 Ωp 
q
where ϵ = (1−δ1 1 )2 − 1. The order of this filter is lower than the Butterworth one, as Chebyshev filters in
general exhibit a better selectivity.

3.14 Exercise 14
Consider the LTI filter described by the following constant-coefficient difference equation:
1 2 1
y(n) − y(n − 1) − y(n − 2) − y(n − 3) = x(n) − 3x(n − 1) (3.85)
5 7 3
ˆ Derive the transfer function and the data-flow graph of the direct form I structure (2 points);

ˆ Find the minimum-order FIR filter, which cascaded to the filter described above, globally provides a
linear-phase FIR filter (3 points);
ˆ Compute the transfer function of the total linear-phase FIR filter and draw its transposed folded delay-
line structure (2 points);
ˆ Compute the group delay of the total linear-phase FIR filter (1 point).

Solution
a) To compute the transfer function of the system, we can just apply the Z−transform to (3.85), i.e.
1 2 1
Y (z) − z −1 Y (z) − z −2 Y (z) − z −3 Y (z) = X(z) − 3X(z)z −1 (3.86)
5 7 3
Therefore, the transfer function of the system is:
Y (z) 1 − 3z −1
H(z) = = (3.87)
X(z) 1 2 1
1 − z −1 − z −2 − z −3
5 7 3
The corresponding direct structure of type I is shown in Fig. 3.16.

b) The transfer function of the overall cascaded system (including the unknown one) is HF IR (z) = H(z)Hx (z).
Since HF IR (z) must represent a linear-phase FIR system, then it must include only zeros which are symmetric
with respect to the unit circle, and all poles have to be compensated, i.e.
1 1 − 3z −1
HF IR (z) = (1 − 3z −1 )(1 − z −1 ) = Hx (z) (3.88)
3 1 − 1/5z −1 − 2/7z −2 − 1/3z −3

27
 Figure 3.16: Direct form I of the system.

Therefore, the transfer function of the system to be cascaded to H(z) is

1 1 2 1
Hx (z) = (1 − z −1 )(1 − z −1 − z −2 − z −3 ) (3.89)
3 5 7 3
c) The transfer function of the overall FIR filter becomes
1 10
HF IR = (1 − 3z −1 )(1 − z −1 ) = 1 − z −1 + z −2 (3.90)
3 3
Observe that the corresponding impulse response exhibit an even symmetry with M = 2. Therefore, (3.90)
denotes a Type I FIR filter. Since this is a second-order system, the folded delay line structure has the same
complexity as a classic tapped delay line, which is shown in Fig. 3.17.

Figure 3.17: Tapped delay line of the global FIR filter.

10 −jω
d) The frequency response of the global FIR filter is HF IR (ejω ) = 1 − e + e−j2ω . Since this is a
3
M
linear-phase Type I system, the group delay is τg = = 1.
2

28
3.15 Exercise 15
Let x(n) denote a discrete-time random process obtained by sampling a continuous-time WSS zero-mean
random process xa (t) with period T = 1 s. The process x(n) passes through a filter with impulse response
h(n) = δ(n − 4) − δ(n − 6) Assume that the autocorrelation ϕaa (t) of xa (t) is known. Derive:

ˆ The relation between the power spectral densities of xa (t) and x(n) and state under which conditions the
spectrum of the discrete-time process is a faithful representation of the spectrum of the continuous-time
process (3 points)

ˆ The mean and the autocorrelation of the filter’s output process y(n) (3 points)

ˆ The ratio between the power spectral densities of y(n) and x(n) (2 points)

Solution
a) The PSD of WSS random process xa (t) by definition is
Z ∞
Φa (Ω) = E{xa (t)xa (t + τ )}e−jΩτ dτ (3.91)
−∞

If the signals are sampled with sampling period T , the PSD of the discrete-time resulting process becomes:
∞
X ∞  
jω −jωn 1 X ω 2πk
Φx (e ) = E{x(k)x(n + k)}e = Φa + ω = ΩT (3.92)
T T T
k=−∞ k=−∞

Therefore, the PSD of the discrete-time process x(n) represents faithfully the PSD of the continuous-time
π
process if and only if Φa (Ω) ≈ 0 for |Ω| > = π (since T = 1 s).
T
b) As known, the output sequence of an LTI system is given by y(n) = x(n) ∗ h(n) where, in this case,
h(n) = δ(n − 4) − δ(n − 6). The mean value and the autocorrelation of the output process are given,
respectively, by:
µy = H(ej0 )µx = 0 (3.93)
and

Φy (m) = Φx (m) ∗ (h(m) ∗ h(−m)) = Φx (m) ∗ (2δ(m) − δ(m + 2) − δ(m − 2))

= 2Φx (m) − Φx (m − 2) − Φx (m + 2) (3.94)

c) Since the transfer function of the system is H(z) = z −4 − z −6 , the ratio between the power spectral densities
of y(n) and x(n) is finally given by

Φy (ejω )
= |H(ejω )|2 = (e−j4ω − e−j6ω )(e−j4ω − ej6ω ) = 2 − ej2ω − e−j2ω = 2 − 2 cos 2ω = 4 sin2 ω (3.95)
Φx (ejω )

29
3.16 Exercise 16
Consider a digital stop-band filter with the following specifications:

ˆ Lower and upper cut-off frequency: ωc1 = 0.2π and ωc2 = 0.4π;

ˆ Ideal frequency response amplitude in the pass-band = 1;

ˆ Ideal frequency response amplitude in the stop-band = 0.

Find the ideal impulse response of the filter (4 points). Assuming to implement a FIR filter approximating the
ideal filter above using the windowing method, compute the order and the impulse response of the FIR filter
when its transition bandwidth ∆ω = 0.02π and the maximum approximation error δ = −20 dB (2 points).
Finally, explain if better results could be obtained using the Kaiser window and why (2 points).

Solution
a) The ideal impulse response of the specified filter results from

Z π Z 0.2π Z π Z 0.4π
1 jω jωn 1 jωn 1 jωn 1
hid (n) = H(e )e dω = e dω + e dω + ejωn dω =
2π −π 2π −0.2π 2π 0.4π 2π −π
 jωn 0.2π  jωn π  jωn −0.4π
1 e 1 e 1 e
= + + =
2π jn −0.2π 2π jn 0.4π 2π jn −π
1
= (ej0.2πn − e−j0.2πn + ejπn − e−jπn − ej0.4πn + e−j0.4πn )
2jπn


 0.8 n=0

= (3.96)

 2
 cos(0.3πn) sin(0.1πn) n ̸= 0
πn

b) Given that δ = −20 dB and the maximum approximation error using a rectangular window is -21 dB, this
is suitable to design the filter with the wanted specifications. As far as the order M of the filter is concerned,
this is given by:
4π 4π
≤ △ω → M≥ − 1 ≈ 199 (3.97)
M +1 △ω
In practice, it is better to set M = 200 to avoid the zero at π. Thus, the impulse response of the designed
filter is h(n) = hid (n − 100).

c) In this case, using a Kaiser window does not lead to significant differences. In general, the role of a Kaiser
window is to relax the approximation error, while ensuring a shorter impulse response. However, for δ ≤ −20
dB, the β coefficient of Kaiser window would be equal to 0. Therefore, the Kaiser window coincides with
rectangular one.

30
3.17 Exercise 17
Consider a LTI system with two poles (one in −1/2 and one in 1/3) and DC gain H(1) = 6.

ˆ Compute the impulse response h(n) of the system (3 points);

1
ˆ Given an input signal x1 (n) = u(n) − u(n − 1), compute the output sequence y1 (n) (2 points);
2
π
ˆ Given an input signal x2 (n) = 50+10 sin(n )+30 sin(nπ), compute the output sequence y2 (n)(2 points)
2

Solution
a) The transfer function of the system is
A
H(z) = (3.98)
1 −1 1
(1 + z )(1 − z −1 )
2 3
A
where H(1) = = 6. Therefore, A = 6. To compute the impulse response, first of all let us perform a
3 2
,
2 3
partial fraction decomposition, i.e.
A1 A1
H(z) = + (3.99)
1 1
1 + z −1 1 − z −1
2 3
where
6 6 18 6 6 12
A1 = lim = = A2 = lim = = (3.100)
z→−1/2 1 −1 2 5 z→1/3 1 −1 3 5
1− z 1+ 1+ z 1+
3 3 2 2
By computing the inverse Z-transform, the impulse response finally results, i.e.
   
18 −1 n 12 1 n
h(n) = u(n) + u(n) (3.101)
5 2 5 3
1
b) Given the input signal x1 (n) = u(n) − u(n − 1), the corresponding Z-transform is
2
1
1 z −1 1 − z −1
X1 (z) = − = 2 (3.102)
1 − z −1 2(1 − z −1 ) 1 − z −1
The Z-transform of the output can be rewritten as
1
6(1 − z −1 ) B1 B2 B3
Y (z) = 2 = + + (3.103)
1 1 1−z −1 1 1
(1 − z −1 )(1 + z −1 )(1 − z −1 ) 1 + z −1 1 − z −1
3 3 2 3
where
1
6(1 − z −1 )
B1 = lim 2 =3 (3.104)
z→1 1 −1 1
(1 + z )(1 − z −1 )
2 3

31
 1
6(1 − z −1 ) 12
B2 = lim 2 = (3.105)
z→−1/2 1 5
(1 − z −1 )(1 − z −1 )
3
1
6(1 − z −1 ) 3
B3 = lim 2 = (3.106)
z→1/3 1 5
(1 − z −1 )(1 + z −1 )
2
Therefore, by computing the inverse Z-transform of X1 (z), the corresponding output sequence is
 n  
12 1 3 1 n
y1 (n) = 3u(n) + − u(n) + u(n) (3.107)
5 2 5 3
π
c) Given the input sequence x2 (n) = 50 + 10 sin(n ) + 30 sin(nπ), the corresponding output sequence is
2

y2 (n) = 50H(ej0 ) + 10|H(ejπ/2 )| sin[nπ/2 + ∠H(ejπ/2 )] + 30|H(ejπ )| sin[nπ + ∠H(ejπ )] (3.108)

where
6
H(ejω ) = (3.109)
1 −jω 1
(1 + e )(1 − e−jω )
2 3
Thus, it can be easily found that
36 36 6
H(ej0 ) = 6, H(ejπ/2 ) = = √ ej atan(1/7) , H(ejπ ) = =9 (3.110)
7−j 5 2 1 1
(1 − )(1 + )
2 3

Ultimately, the output sequence is:

72 1
y2 (n) = 300 + √ sin[nπ/2 + atan( )] (3.111)
2 7

32
3.18 Exercise 18
Design a Butterworth high-pass filter with the following specifications: ωs = 0.2π, ωp = 0.5π, δ1 = δ2 = 0.1 (7
points). Sketch the data-flow graph of the filter using a structure of your choice (1 point).

Solution
z −1 + α
a) Let us apply the high-pass to low-pass frequency transformation z̃ = − to filter specifications.
1 + αz −1
Assuming that α = 0, then z̃ −1 −1
= −z . Therefore, θp = 0.5π and θs = 0.8π. By applying the bilinear
transformation with T = 1 s, the pass-band and stop-band edge frequencies of the low-pass prototype filter
become
θp θs
Θp = 2 tan = 2 rad/s Θs = 2 tan = 6.155 rad/s (3.112)
2 2
Therefore, the order and the cut-off frequency of the low-pass prototype filter become respectively
 h .  i 
log δ12 − 1 1
(1−δ )2
− 1
N = 
2
 1 =3
 (3.113)
 Θ 
2 log Θ s
p

and
Θs
Θc = s = 2.863 rad/s (3.114)
1
6
−1
δ22
The position of the poles of the low-pass prototype filter in the negative complex half-plane are shown in
Fig. 3.18. As usual, the angles of the poles are given by ϕk = π2 + (2k + 1) 2Nπ
∀k = 0, 1, . . . , N − 1.
Therefore, the poles of the low-pass prototype filter are:

s0,2 = Ωc [cos(π/2 + π/6) ± j sin(π/2 + π/6)] = −1.431 ± j2.478

s1 = −Ωc = −2.863 (3.115)

and the overall transfer function is

H0
HLP (s) = (3.116)
(s + 2.863)(s2 + 2.8635s + 8.196)
2 1−z̃ −1
where H0 = 23.44.By replacing s = T 1+z̃ −1 into (3.116) it follows that

23.44(1 + z̃ −1 )3
HLP (z̃) =
[2(1 − z̃ −1 ) + 2.863(1 + z̃ −1 )][(1 − z̃ −1 )2 + 2.863(1 − z̃ −1 )(1 + z̃ −1 ) + 8.196(1 + z̃ −1 )2 ]
0.2691(1 + z̃ −1 )3
= (3.117)
(1 + 0.1775z̃ −1 )(1 + 0.4678z̃ −1 + 0.3610z̃ −2 )

Finally, by applying the frequency transformation z̃ = −z −1 , the transfer function of the wanted IIR high-pass
filter results, i.e.

0.2691(1 − z −1 )3
HHP (z) = HLP (z̃) z̃ −1 =−z −1
= (3.118)
(1 − 0.1775z −1 )(1
− 0.4678z −1 + 0.3610z −2 )
A cascade structure of the designed filter is shown in Fig. 3.19.

33
 Figure 3.18: Position of the poles of the low-pass prototype filter in the complex plane.

Figure 3.19: Cascade structure of the designed high-pass filter.

3.19 Exercise 19
A WSS continuous-time random signal s(t) is sampled without aliasing, with sampling period T = 3 ms.
The resulting discrete-time sequence x(n) obtained is processed by a FIR system with impulse response
h(n) = δ(n) + δ(n − 5). The mean of s(t) is 0, while its auto-correlation function is ϕs (τ ) = 1/2(u(τ + 0.004) −
u(τ − 0.004)), where τ is expressed in seconds and u(t) denotes the unit step. If y(n) is the output sequence of
the filter, compute the mean (1 point), the autocorrelation function (4 points) and the power spectral density
of y(n) (3 points).

Solution
a) Since the input WSS has a zero-mean, then E {y(n)} = 0.

34
b) The autocorrelation of the discretized input process is
(
0.5 n = −1, 0, 1
ϕs (n) = (3.119)
0 otherwise

The autocorrelation sequence of the corresponding output process is given by

ϕy (n) = ϕs (n) ∗ [h(n) ∗ h(−n)] = ϕs (n) ∗ [δ(n + 5) + 2δ(n) + δ(n − 5)]

1 1
= [δ(n − 6)+δ(n − 5)+δ(n − 4)]+δ(n − 1)+δ(n)+δ(n + 1)+ [δ(n + 4)+δ(n + 5)+δ(n +(3.120)
6)]
2 2
c) By computing the DTFT of (3.120), the power spectral density of process y(n) finally results, i.e.
1 1 1
Φy (ω) = (ej6ω + e−j6ω ) + (ej4ω + e−j4ω ) + (ej5ω + e−j5ω ) + 1 + ejω + e−jω
2 2 2
= 1 + 2 cos ω + cos 4ω + cos 5ω + cos 6ω

35
3.20 Exercise 20
Consider a narrowband signal x(t) with a spectral content mainly located in the band [1290, 1310] Hz and
sampled at fs = 10 kHz.

ˆ Compute the impulse response of an ideal all-pass linear phase system with a group delay equal to 1/4
of the sampling period (i.e. 25 µs) and unit gain over the whole frequency axis (2 points).
ˆ Using this system and the Hanning window, compute the frequency response of the FIR filter (with
M = 4) approximating the ideal system and sketch its data-flow graph with the minimal computational
cost (3 points).
ˆ Determine the impulse response of another 1st order FIR linear system (i.e. M = 1) assuring a group
delay equal to 1/4 of the sampling period and unit gain only at 1300 Hz (3 points).

Solution
a)The ideal transfer function of the ideal all-pass filter is

Hid (ejω ) = e−jω/4 |ω| ≤ π (3.121)

Therefore, the corresponding ideal impulse response is
1 R π −jω/4 jωn 1 R π jω(n−1/4) 1 π 1
hid (n) = −π e e dω = −π (e )dω = [ejω(n−1/4) ] −π
= sinc[π(n − )]
2π 2π 2πj(n − 1/4) 4

b) The impulse response of an FIR filter approximating the ideal one with M = 4, using a Hanning win-
dow is given by

M 1 9 n
h(n) = hid (n − )w(n) = sinc[π(n − )][1 − cos(π )] n = 0, . . . , M (3.122)
2 2 4 2
Therefore h(0) = h(4) = 0
1 5π
h(1) = sinc(− ) = −0.09
2 4
π
h(2) = sinc(− ) = 0.90
4
1 3π
h(3) = sinc( ) = 0.15
2 4
Finally, the frequency response of the filter is:

H(ejω ) = −0.09e−jω + 0.90e−2jω + 0.15e−j3ω (3.123)

The data-flow graph of the FIR filter with minimal computational complexity is the classic tapped delay line
(the folded delay line is not possible in this case since pairs of coefficients are not equal).
c)The transfer function of a generic first-order FIR filter is

H(z) = a + bz −1 (3.124)
Assuming that the group delay is equal to 1/4 of the sampling period and that the gain is equal to 1 at 1300
Hz means that
H(z) z=ejω0 = e−jω0 /4 = a + be−jω0 (3.125)

36
 1300
where ω0 = 2π = 0.817. This condition can be rewritten as a system of nonlinear equations with a and
10000
b as unknowns, i.e.


 (a + b cos ω0 )2 + b2 sin2 ω0 = 1 → a2 + b2 + 2b cos ω0 = 1



 b sin(ω0 ) ω0

 atan = → b sin ω0 = tan(ω0 /4)(a + b cos(ω0 ))
a + bcos(ω0 ) 4

 
sin ω0
Observe that a = b − cos ω0 . Therefore, the first equation can be rewritten as:
tan ω0
 2  
2 sin ω0 2 sin ω0
b − cos ω0 + 2b cos ω0 − cos ω0 + b2 = 1 (3.126)
tan(ω0 /4) tan(ω0 /4)

which leads to the following solutions

v
u 1
b = ±u u 2   = ±0.278 (3.127)
t sin ω0 sin ω0
− cos ω0 + 2 cos ω0 − cos ω0 + 1
tan(ω0 /4) tan(ω0 /4)
 
sin ω0
a = b − cos ω0 = ±0.789 (3.128)
tan(ω0 /4)

Excluding the negative solutions (which correspond to the angle outside [−π/2, π/2]), we finally obtain a =
0.789, b = 0.278.

37
3.21 Exercise 21
a) Given a notch filter with transfer function

(1 − cz −1 )(1 − c∗ z −1 )
H(z) = A (3.129)
(1 − dz −1 )(1 − d∗ z −1 )
processing a signal x(t) = s(t) + sin(2πf0 t) with f0 = 20 kHz, sampled at a rate fs = 120 kHz, determine the
coefficients of the system so that (5 points):

1. the sinewave component at f0 is perfectly cancelled;

2. the DC gain of the filter is 1;
3. the filter gain at f1 = 60 kHz is 1.01 (hint: in a notch filter the angle of pairs of poles and zeros is the same).

Assuming to represent all coefficients as fixed-point Q7 numbers, compute the quantized coefficient over 8
bits. Does the coefficient quantization affect the position of poles significantly in a canonical structure? Jus-
tify your answer. (3 points)

Solution
f0 f1
Given the discrete-time frequencies ω0 = 2π = π/3 and ω1 = 2π = π the three design conditions specified
fs fs
in the exercice can be used to set up a system of equations,

 ±jπ/3 ) = 0
 H(e
H(ej0 ) = 1 (3.130)

 ±jπ
H(e ) = 1.01

where the unknowns are variables c = |c|ejθ , d = |d|ejθ and A. From the first equation, it results immediately
that (1 − |c|ej(±θ±π/3 ) = 0. Therefore, |c| = 1 and θ = ω0 = π/3. By replacing such values into (3.130), the
two remaining system equations can be rewritten as


 A(1 − ejπ/3 )(1 − e−jπ/3 )
 =1
(1 − d)(1 − d∗ ) (3.131)

 A(1 + ejπ/3 )(1 + e−jπ/3 )
 = 1.01
(1 + d)(1 + d∗ )
Thus, it follows that 
 1−2|d| cos ω +|d|2
 A = 2(1−cos ω0 0 )
1 − 2|d| cos ω0 + |d|2 1 − cos ω0 (3.132)

 = 1.01
1 + 2|d| cos ω0 + |d|2 1 + cos ω0
|d|
Since Re{d} = |d|cosω0 = , the latter equation can be rewritten as:
2 √
4.01 ± 0.24
3|d|2 − 3|d| + 3 − 1.01|d|2− 1.01|d| − 0.01 = 0 =⇒ 1.99|d|2
− 4.01|d| + 1.99 = 0 → |d| = .
3.98
In fact, of the two possible solutions (0.8845 and 1.13 only the former one is acceptable, because with the
latter the filter would be unstable. Finally, from the first equation of (3.132) it follows that A = 0.8978 and
the transfer function of the notch filter becomes
1 − z −1 − z −2
H(z) = 0.8978 (3.133)
1 − 0.8845z −1 + 0.78234z −2

38
b) If all numbers are represented in Q7 notation then the quantization errors are included in the interval
1
[−1, 1 − 2−7 ] and the quantization step is ∆ = 7 . The quantized coefficients of the polynomial at the
2
denominator of (3.133) are then given by
 
0.8845
aˆ1 = ∆ = 113∆ = 0.8828125 (3.134)
∆
 
−0.7823
aˆ2 = ∆ = −100∆ = −0.78125 (3.135)
∆

Therefore, the position of poles is not significantly affected by quantization. Indeed, we have a biquadrate
structure and the poles are quite far apart.

39
3.22 Exercise 22
Consider the following three discrete-time IIR systems identified by:

1. Poles in a and b, and zeros in c and 1/c, with a and b complex and c real;

2. Poles in b and b∗ and zero in -1;

3. A double pole in 1/a and zeros in e and e∗ (with e complex).

Derive:

ˆ The system to be cascaded to each filter to obtain a linear-phase FIR system as well as the transfer
functions of the resulting FIR systems (6 points);

ˆ Type and data-flow graphs with minimal complexity of each FIR system (2 points).

Solution
a) The transfer functions of the three IIR systems are:

(1 − cz −1 )(1 − 1c z −1 ) 1 + z −1 (1 − ez −1 )(1 − e∗ z −1 )
H1 (z) = H2 (z) = H3 (z) = (3.136)
(1 − az −1 )(1 − bz −1 ) (1 − bz −1 )(1 − b∗ z −1 ) (1 − a1 z −1 )2

As known, linear-phase FIR systems consist only of zeros which should be symmetric with respect to the unit
circle. Therefore, the systems to be cascaded to H1 (z), H2 (z) and H3 (z) should cancel the respective poles
and should include the zeros needed to ensure the aforementioned symmetry. Thus, the transfer functions of
such systems are:

H̃1 (z) = (1 − az −1 )(1 − bz −1 ) (3.137)

−1 ∗ −1
H̃2 (z) = (1 − bz )(1 − b z )
   2
1 1 1
H̃3 (z) = 1 − z −1 1 − ∗ z −1 1 − z −1
e e a

and the resulting FIR systems are:

 
1 1
H̃F IR1 (z) = (1 − cz −1 )(1 − z −1 ) = 1 − c + z −1 + z −2 (3.138)
c c
H̃F IR2 (z) = 1 + z −1
 
1 −1 1 −2
H̃F IR3 (z) = 1 − 2ℜ{ }z + 2 z (1 − 2ℜ{e}z −1 + |e|2 z −2 ) (3.139)
e |e|
     
1 −1 1 1 −2 1
= 1 − 2ℜ e + z + 2
+ 4ℜ{e}ℜ{ } + |e| z − 2ℜ e + z −3 + z −4 .
e |e|2 e e

b) Observe that the three FIR systems are of Type I, Type II and Type I, respectively, while the corresponding
optimal data-flow graphs (folded delay lines) are shown in Fig. 3.20(a)-(c).

40
 z-1
z-1 z-1

z-1 z-1 z-1 z-1

-(c+1/c)
-2Re{e+1/e} 1/|e|2+4Re{e}Re{1/e}+|e|2

(a) (b) (c)

Figure 3.20: Folded delay-line structure of the linear-phase FIR filters with transfer functions H̃F IR1 (z),
H̃F IR2 (z) and H̃F IR3 (z).

3.23 Exercise 23
Consider a sequence x(n) resulting from a white noise w(n) (with variance σw 2 = 1) filtered by a system

with impulse response h(n) = sin(π/8n), for n = 0, . . . , N − 1, where N = 256. Given the sequence y(n) =
x(n) + e(n), where e(n) is another white noise, uncorrelated to w(n) and with variance σe2 = 20,

ˆ Compute the first three samples of the autocorrelation of y(n), i.e. ϕy (n), for n = 0, 1, 2 (3 points);

ˆ Estimate the PSD of the WSS process y(n) by using periodograms over 50-sample windows (4 points).

ˆ Find the frequency bins where the two main spectral peaks are located (1 points).

Solution
a) The autocorrelation of y(n) is given by:

ϕy (n) = σe2 δ(n) + σw

2
[h(n) ∗ h(−n)] = σe2 δ(n) + σw
2
c(n) (3.140)

Since h(n) is a sinewave truncated by a rectangular window consisting of N samples, then c(n) can be rewritten
as ( P
N −|n|−1 π(k+n)
sin( πk
8 ) sin( ) |n| ≤ N − 1
c(n) = k=0 8 (3.141)
0 |n| > N − 1
Thus, for |n| ≤ N − 1 the first term of (3.141) can be rewritten as:
N −|n|−1   N −|n|−1  
1 X πn π(2k + n) N − |n| πn 1 X π(2k + n)
c(n) = cos( ) − cos( ) = cos( ) − cos (3.142)
2 8 8 2 8 2 8
k=0 k=0

However, the rightmost term of (3.142) can be neglected, because it is always smaller to the sum of the samples
of a half-cycle (i.e. ≤ 1). Therefore,

N − |n| πn
ϕy (n) ≈ σe2 δ(n) + σw
2
cos( ) (3.143)
2 8
which finally returns ϕy (0) = 148, ϕy (1) = 118 and ϕy (2) = 90.

41
b) Assuming that the periodogram of y(n) is estimated at frequency bins k = 0, ..., L − 1 with L = 50,
2
|Y (k)|2 1 sin(πk)
it is given by Φ̂y (k) = L = Φy (k) ∗ L sin(πk/L) where Y (k) is the DFT of y(n). Observe that sequence
N −|n|
wB (n) = 2 in (3.143) exhibits a triangular shape centered at time 0 (the same as a Bartlett window,
2
but with maximum amplitude N/2 instead of 1). Hence, the DTFT of wB (n) is WB (ejω ) = 12 sin(N ω/2)
sin(ω/2) .
Considering the frequency-shifting property of the DTFT due to the multiplication by the cosine sequence it
finally follows from (3.140) that at frequencies ω = 2πk/L the periodogram of y(n) is
 
2 k 1 2 k 1 2 2
σ 2 sin(πk) σ 2 sin(N π( L − 16 )) sin(N π( L + 16 ))
Φ̂y (k) = e + w +  ∗ sin(πk) . (3.144)
L sin(πk/L) 2L sin(π( Lk − 16
1
)) sin(π( Lk + 16
1
)) sin(πk/L)

c) Observe that, since N is quite larger than L, the magnitude of the main-lobe of the spectral terms associated
with WB (ejω ) in (3.144) is certainly much larger than the others. Moreover, due to frequency shifting, such
peaks are theoretically located at ±pi/8. Therefore, the indexes of the frequency bins where the periodogram
is maximum are k∗ = ±[L/16] = 3, where [·] in this case denotes the rounding operator.

42
3.24 Exercise 24
Given the generic transfer function
1
H(z) = with 0 < ω0 < π (3.145)
1 − 2 cos ω0 z −1 + z −2
compute h(n) and the corresponding constant-coefficient difference equation (5 points). Is the system stable?
Sketch the positions
√ of poles and zeros in the Z complex plane and justify your answer (1 point). Assuming
that cos ω0 = 3/2 and that the system output is connected to a DAC with sampling frequency fs = 100 kHz,
what is the maximum frequency of the generated analog signal? (2 points).

Solution
a) The constant-coefficient difference equation of the system results immediately from the (3.145), i.e.

y(n) = x(n) + 2 cos ω0 y(n − 1) − y(n − 2). (3.146)

Using the partial fraction decomposition, the transfer function of system (3.145) can be rewritten as:

1 A A∗
H(z) = = + (3.147)
1 − 2 cos ω0 z −1 + z −2 1 − ejω0 z −1 1 − e−jω0 z −1

1 1 ej(ω0 −π/2)
where A = 1−e−jω0 z −1 z=ejω0
= 1−e−j2ω0
= 2 sin ω0 . By computing the IDTFT of (3.147) it results that

ej(ω0 −π/2) jω0 n e−j(ω0 −π/2) −jω0 n 1

h(n) = e u(n) + e u(n) = sin(ω0 n + ω0 )u(n). (3.148)
2 sin ω0 2 sin ω0 sin ω0

b) Expression (3.148) shows clearly that the system at hand is a digital oscillator generating a discrete-time
sinewave of amplitude sin1ω0 and normalized frequency ω0 . Therefore, the system is critically stable. This is
confirmed by the position of poles that lie on the unit circle, i.e. they have unit magnitude and phase equal
to ±ω0 .

fs
√
c) The frequency of the analog signal generated by the oscillator at the output of the DAC is f0 = 2π arccos 3/2 =
fs
12 = 8.3 kHz.

43
3.25 Exercise 25
Consider the frequency response of an analog differentiator, Ha (f ) = j2πf . Compute the coefficients {h−1 , h0 , h1 }
of a third-order, noncausal FIR system whose frequency response approximates Ha (f ) for f << 1/Ts (Ts being
the sampling period). Sketch the positions of the zeros of the FIR system in the z complex plane (4 points).
Compute the DTFT and the DFT over 4 points of the impulse response of the FIR system (4 points).

Solution
a) The transfer function of the noncausal FIR system is simply H(z) = h−1 z + h0 + h1 z −1 . If f << 1/Ts , the
frequency response of the system approximating the analog differentiator is
ω
H(ejω ) = h−1 ejω + h0 + h1 e−jω ≈ j . (3.149)
Ts
Consider that, due to the odd symmetry of Ha (f ), we must have that h1 = −h−1 , i.e. the FIR system is
certainly of type III (although with a null phase response). Thus, (3.149) can be more compactly rewritten as
h0 − 2jh1 sin ω ≈ j Tωs . By equating the real part and the imaginary part of this equation, a trivial algebraic
system of two equations in two unknowns results. Hence, h0 = 0 (as expected for type III FIR systems) and
h1 ≈ − 2Ts ωsin ω ≈ − 2T1 s . Ultimately, the transfer function of the discrete-time differentiator is

z − z −1
H(z) = (3.150)
2Ts
that includes just two zeros in (−1, 0) and (1, 0), respectively. In fact, such zeros are mandatory for type III
systems.

b) The DTFT associated with (3.150) is simply

j
H(z)|z=ejω = H(ejω ) = sin ω. (3.151)
Ts
P
The samples of the DFT over N = 4 points can be computed from its definition, i.e. H(k) = 3n=0 h(n)W4nk
for k = 0, . . . , 3, where W4 = e−jπ/2 is the twiddle factor over 4 points. By rearranging this expression in a
matrix form, it finally results that
   0       
H(0) W4 W40 W40 W40 h(0) 1 1 1 1 0 0
H(1) W40 W41 W42 W43  h(1) 1 −j −1 j  −1/2  j 
         
H(2) = W40 W42 W44 W46  h(2) = 1 −1 1 −1  0  =  0  . (3.152)
H(3) W40 W43 W46 W49 h(3) 1 j −1 −j 1/2 −j

44
Chapter 4

Multiple–choice and short questions

1. The power spectral density (PSD) of a stationary, real-valued random process

(a) is complex and it exhibits even symmetry in magnitude and odd simmetry in phase.
(b) is complex and exhibits even symmetry in both magnitude and phase.
(c) is purely real since the autocorrelation of the process certainly exhibits even symmetry.
(d) is purely real since the autocorrelation of the process certainly exhibits odd symmetry.

2. The figure below (on the left side) shows the power spectral density (PSD) of four stationary discrete-
time random processes. The pictures on the right side shows instead a realization of the corresponding
noises, but they are not shown in the right order. Match each signal to its most likely PSD and justify
briefly your answer.

3. Considering an LTI system with an FIR response h(n) ̸= 0 for n = 0, . . . , L and an input sequence x(n)
of the same length, how many DFT/IDFT points N should to be used to avoid the risk of time aliasing?

45
4. The fact that a continuous-time or discrete-time signal is absolutely summable is
(a) a sufficient condition for the existence of the Fourier transform.
(b) a necessary condition for the existence of the Fourier transform.
(c) a necessary and sufficient condition for the existence of the Fourier transform.
(d) neither a necessary nor a sufficient condition for the existence of the Fourier transform: the Fourier
transform of some signals may exist in any case.
5. For each one of the following discrete-time systems, determine whether they are linear, causal, time-
invariant and stable or not.

6. In FIR filter design based on the windowing method, the maximum approximation ripple of the frequency
response obtained with a given window function
(a) is greater in the passband than in the stopband.
(b) is greater in the stopband than in the passband.
(c) is approximately the same both in the passband and in the stopband.
(d) is different in the passband and in the stopband, as it depends on filter specifications.

46
7. What is the normalized frequency of a discrete-time sinewave of frequency 100 Hz sampled at 20 kHz?

(a) ω0 = 0.005.
(b) ω0 = 0.0314.
(c) ω0 = 0.01.
(d) ω0 = 0.0628.

8. The figures below show 6 impulse responses, frequency responses and zero-pole diagrams. Associate the
impulse responses to both the frequency responses and the zero-pole diagrams.

9. To avoid phase distortion, the group delay of a causal filter has to be

(a) constant.
(b) linearly increasing.
(c) linearly decreasing.
(d) null.

47
10. Assume to stimulate an unknown system with a pure sinewave of given amplitude and adjustable fre-
quency in order to reconstruct its frequency response. If the spectrum of the output signal includes some
harmonics of the test input signal, it means that

(a) the system is linear, but it exhibits poor performances in the stopband.
(b) the system is linear, but it exhibits poor performances in the passband.
(c) the system is linear and its frequency response includes multiple passbands.
(d) the system is only apparently linear, but in fact it includes some nonlinearity.

11. If you could perform the spectral estimation of a periodic signal acquired in coherent sampling conditions,
which window function should be used to minimize the impact of spectral leakage on the visualized
spectrum?

(a) The rectangular one.

(b) The Hamming one.
(c) The Blackman one.
(d) We cannot decide a priori. The best window depends on the signal features and on the intended
application.

12. Assume that you want to design an IIR low-pass filter and that, for the application considered, it is of
crucial importance to avoid gain errors at DC (0 Hz). Which is the most suitable and safest kind of
filter that should be used in such conditions?

(a) Butterworth.
(b) Chebyshev I.
(c) Chebyshev II.
(d) Elliptic.
x 2
13. Let X be a continuous random variable with PDF fX (x) = √1 e− 2 and let Y = X 2 . The PDF of
2π
random variable Y is
y 4
(a) fY (y) = √1 e− 2 for y ≥ 0.
2π
√1 e − y2
(b) fY (y) = 2πy
for y ≥ 0.
y 2
(c) fY (y) = √ 1 e− 2 for y ≥ 0.
2πy
y 4
(d) fY (y) = √ 1 e− 2 for y ≥ 0.
2πy 2

14. A type II linear-phase system

(a) can have a zero equal to 1 and another one equal to -1.
(b) can have a zero equal to 1 and must have a zero equal to -1.
(c) must have a zero equal to 1 and can have a zero equal to -1.
(d) must have zeros both in 1 and -1.

48
15. When you increase the length of the observation interval for spectral estimation,

(a) the spectral resolution improves, but the time resolution degrades.
(b) the time resolution improves, but the spectral resolution degrades.
(c) both time and spectral resolution improve provided that the spectral estimates over subsequent
intervals are averaged.
(d) the spectral resolution improves, provided that the right window function is chosen.

16. Given two random variables X and Y with variances σx2 and σy2 and correlation coefficient ρ = −0.5, the
variance of the difference between such random variables is

(a) σx2 − σx2 + σx σy .

(b) σx2 − σx2 − σx σy .
(c) σx2 + σx2 − σx σy .
(d) σx2 + σx2 + σx σy .

17. Which one of the following statements is TRUE?

(a) The bandwidth of an impulse of finite duration grows if its duration decreases.
(b) The bandwidth of an impulse of finite duration decreases if its duration grows.
(c) The bandwidth of an impulse of finite duration increases if its duration grows.
(d) No one of the above statements is true, as nothing can be said about the spectrum of a signal if its
actual shape is unknown.

18. Which one of the following statements on FIR filters is FALSE?

(a) They can be used to approximate any kind of frequency response, including those of asymmetric
and multiband discrete-time LTI systems.
(b) They do not suffer from stability problems, since they do not have poles.
(c) Their zeros normally liew within the unit circle.
(d) Their order is usually high: the higher the number of coefficients, the better the approximation of
the wanted frequency response.

19. In order to estimate the delay between the data records of two distinct signals collected by two different
sensors monitoring the very same event (e.g., the same sound), you should

(a) compute the convolution of the two signals to find its peak.
(b) compute the cross-correlation of the two signals to find its peak.
(c) compute the covariance of the two signals to find its peak.
(d) compute the peak of the spectral mainlobes of both signals to find their periods in the frequency
domain.

20. The modified periodogram is

(a) a consistent and unbiased estimator.

(b) a consistent and biased estimator.
(c) an inconsistent and unbiased estimator.
(d) an inconsistent and biased estimator.

49
21. The Parks & McClellan algorithm for FIR filter design is very important because

(a) it converges to the only optimal solution minimizing the maximum absolute error between the
actual filter frequency response and the ideal one.
(b) it converges to the only optimal solution minimizing the mean square error between the actual filter
frequency response and the ideal one.
(c) it converges to the only optimal solution minimizing the maximum weighted error between the
actual filter frequency response and the ideal one.
(d) it converges to the only optimal solution minimizing the least square error between the actual filter
frequency response and the ideal one.

22. In your opinion the zero-pole plot shown below corresponds to

(a) a Butterworth bandpass-filter of order 5

(b) a Butterworth bandpass-filter of order 10
(c) a Chebyshev II bandpass-filter of order 5
(d) a Chebyshev II bandpass-filter of order 10

0.8

0.6

0.4
Imaginary Part

0.2
3
2 5
0

-0.2

-0.4

-0.6

-0.8

-1

-1 -0.5 0 0.5 1
Real Part

23. Which one of the following statements about Sample-and-Hold Amplifiers (SHAs) for ADCs is FALSE?

(a) SHAs partially remove the spectral replicas of the input analog signal due to an unintentional
filtering action.
(b) SHAs are needed to hold the sample values for time intervals long enough to perform quantization
and binary encoding.
(c) SHAs are needed because the ideal Dirac pulses cannot be generated.
(d) SHAs are needed to hold the sample values for time intervals long enough to let the sampling
transients expire.

24. Let x1 (n) and x2 (n) be two finite sequences consisting of L1 and L2 samples, respectively. If the sum of
such sequences is performed prior to computing the DFT, the number of DFT points should be

(a) N ≥ L1 + L2 − 1.
(b) N ≥ L1 + L2 .
(c) N ≥ L1 + L2 − 2.
(d) N ≥ max(L1 , L2 ).

50
25. If the number of DFT/FFT points of a signal acquired for spectral estimation is greater than the window
length

(a) the visualization of the signal spectrum improves and the uncertainty in peak detection decreases.
(b) the spectral resolution improves.
(c) the spectral leakage decreases.
(d) the time aliasing problem is mitigated.

26. The Fourier transform of a generic time-reversed signal x(−t) ∈ C is

(a) X ∗ (Ω).
(b) X ∗ (−Ω).
(c) X(−Ω).
(d) −X(Ω).

27. The actual signal-to-quantization noise ratio at the output of a B−bit quantizer

(a) is 6.02B dB.

(b) is 6.0 · 2B dB.
(c) depends on additional noise sources.
(d) depends not only on the number of bits, but also on the number of quantization levels that are
stimulated by the input signal.

28. Which one of the following statements on the Costantinides expressions for frequency transformation is
FALSE?

(a) They must be rational functions in which the order of the polynomial at the numerator must be
the same as the order of the polynomial at the denominator.
(b) The number of coefficients N depends on the number of frequency intervals (e.g., filter passbands)
to be transformed in the other frequency domain.
(c) The points of the image axis of the z plane have to be mapped onto the image axis of the z̃ plane.
(d) The points of the unit circle of the z plane have to be mapped onto the unit circle of the z̃ plane.

29. In the gearbox frequency spectrum usually the largest spectral components are

(a) the fundamental sinusoidal terms due to gear meshing, which depend on gear stiffness.
(b) the harmonics at gear meshing frequencies.
(c) the ghost and subharmonic sinusoidal components depending on gear rotation speed.
(d) the gear rattle due to aperiodic vibrations within the backlash range.

30. The Discrete-time Fourier transform of the s(n) = sign(n) function (i.e. equal to 1 for t ≥ 0 and -1 for
t < 0) is
1
(a) S(ejω ) = 1−e−jω
.
2
(b) S(ejω ) = 1−e−jω
.
1 P
(c) S(ejω ) = 1−e−jω
+ π k=−∞ ∞δ(ω − 2πk).
2 P
(d) S(ejω ) = 1−e−jω
+ π k=−∞ ∞δ(ω − 2πk).

51
Solutions
1. (c) - The autocorrelation of a real-valued random process is a real and even function. Therefore, the
PSD is purely real due to the simmetry properties of the Fourier transform.

2. PSD 1 → signal 3; PSD 2 → signal 1; PSD 3 → signal 4; PSD 4 → signal 2.

3. Since both h(n) and the input sequence x(n) consist of L + 1 nonzero samples, the DFT/IDFT should
be computed over N ≥ 2L + 1 samples.

4. (a) - Function absolute summability is just a sufficient condition for Fourier transform existence. A
theorem confirms this. It is not a necessary condition, because some functions exist (e.g. the sinc(·)
one), that have a Fourier transform even though they are not absolutely summable.

5. (a) yes, yes, no, no;

(b) no, yes, yes, yes;
(c) no, yes, yes, yes;
(d) no, yes, yes, yes;
(e) yes, no, yes, yes;
(f) yes, yes, yes, yes;
(g) yes, yes, yes, yes;
(h) yes, no, no, no;
(i) no, no, no, yes;
(j) yes, no, yes, no;
(k) yes, no, no, yes;
(l) no, yes, yes, yes.

6. (c) - The ripple size of the frequency responses obtained with the windowing method are mainly due to
the integral of the spectral sidelobes within the wanted passband(s). The result of this integral changes as
a function of the frequency shift of the Fourier transform of the window function and the ideal frequency
response. However, the peak values due to the Gibbs’ phenomenon are approximately the same in the
passband(s) and in the stopband(s) and they do not scale as window length grows.

7. (b) - ω0 = 2π ff0s = 0.0314.

8. 1) Type IV – Freq. resp. 3; Zero-pole diag. 2;

2) Type IV – Freq. resp. 2; Zero-pole diag. 4;
3) Type II – Freq. resp. 5; Zero-pole diag. 1;
4) Type II – Freq. resp. 6; Zero-pole diag. 5;
5) Type IV – Freq. resp. 4; Zero-pole diag. 6;
6) Type II – Freq. resp. 1; Zero-pole diag. 3;

9. (a) - The group delay is given by τg = − dθ(ω)

dω . Recalling that, to avoid phase distortion, θ(ω) = −βω,
the group delay must be constant.

10. (d) - The sinewaves are the eigenfunctions of linear systems. Therefore, if the spectrum of the output
signal contains sinusoidal components that were not present in the input signal, they result from the
Fourier series of a periodic signal due to the nonlinear distortion of the input sinewave.

11. (a) - In coherent sampling conditions, if the rectangular window is used, the spectral leakage is not visible
and it does not affect the estimated spectrum, because the spectral sidelobes of the Dirichlet kernel are
sampled at the zero crossing points.

52
12. (c) - For sure, Chebyshev I and Elliptic are risky solutions as neither one ensures that the ripple mag-
nitude at DC is 0. A Butterworth filter could be a safe option, but a Chebyshev II filter can provide
zero magnitude error at DC with a lower filter order for the same selectivity. Also, its phase distortion
is reasonable since all ripples are located in the stopband.

13. (b) - Function Y = g(X) = X 2 is differentiable everywhere, i.e. g ′ (X) = 2X. Since Y can just take
nonnegative values, the PDF of Y for y < 0 is certainly 0. Observe that equation y = x2 has two
√ √
solutions, i.e. x1 = y and x2 = − y. Recalling that the PDF of a differentiable function of the
random variable X is given by:

fX (x1 ) fX (x2 )
fY (y) = + =
|g ′ (x1 ))| |g ′ (x2 ))|
√ √
fX ( y) fX (− y)
= √ + √ =
2 y 2 y
√
fX ( y)
= √ (4.1)
y

solution (b) finally results.

14. (b) - Since H(z) = z −M H(z −1 ), if M is odd, we have that H(−1) = −1−M H(−1) = −H(−1) if and
only if H(−1) = 0.

15. (a) - The spectral resolution improves because the spectrum mainlobe width of the chosen window func-
tion is inversely proportional to the observation interval length. However, if the signal is not stationary,
choosing longer intervals prevents our ability to track possible ignal variations in the frequency domain.

16. (d) - E{(X −Y )2 }−E{(X −Y )}2 = σx2 +σy2 −2E{XY }−2E{X}E{Y } = σx2 +σy2 −2ρσx σy = σx2 +σy2 +σx σy .

17. (a) - Even if the exact Fourier transform of a pulse depends on its shape, any pulse of finite duration can
be also regarded as the product between an arbitrary function and a rectangle of duration τ . Recalling
that the spectral mainlobe width of the sinc(·) function is inversely proportional to τ , it is generally true
that if the pulse duration decreases, its spectral content is spread over a broader frequency interval.

18. (c) - The zeros of FIR filters can be spread over the whole z complex plane. For instance, in the case of
linear-phase filters, if some zeros lie within the unit cicle their reciprocal values must lie outside it.

19. (b) - The correlation function is the operator used to evaluate the similarity of two given signals. If
the signals are due to the same event or phenomenon, but an unknown delay exists between the start
of sensors’ signal recordings, the time at which the correlation peak occurs provides a good estimate of
such a delay.

20. (d) - The modified periodogram is just asymptotically unbiased and its variance does not converge to
zero even if the observation interval length grows.

21. (c) - The Parks & McClellan algorithm exploits the so-called alternation theorem, which ensures con-
vergence to the optimal solution in a minimax sense. A weight function is typically used to change the
size of the magnitude response ripples as a function of filter specifications.

22. (b) - The zero-pole plot cannot refer to a Chebyshev II one, otherwise we should have some out-of-band
zeros on the unit circles at the frequencies where the ripples’ magnitude is null. Since 5 zeros of the
transfer function are located in (1, 0) and other 5 in (−1, 0), and observing that we have 5 pairs of
complex conjugate poles, the filter must be a Butterworth bandpass filter of order 10.

53
23. (a) - The SHAs do not perform any filtering of the discrete-time signals at the output of an ADC because,
after quantization and binary coding, just sequences of samples over a normalized time axis are obtained,
as expected in the ideal case. The unintentional filtering action is instead performed in the DACs.

24. (d) - Since we are considering just the sum and NOT the convolution of two sequences, the DFT should
be computed over a number of points at least greater than or equal to the number of samples of the
longer sequence. The shorter one is zero-padded.

25. (a) - When the number of points of a DFT/FFT grows, the spectrum of the signal is sampled at a higher
rate. Therefore, it is more clearly visible. However, spectral resolution and spectral leakage do not
change, as they depend just on the type and length of the considered window. Also, the time aliasing
is not an issue in this case, because once the spectral analysis is done, no inverse DFT/FFT should be
performed.
R∞ R∞
26. (c) - F(x(−t)) = −∞ x(−t)e−jΩt dt = −∞ x(τ )e−jΩ(−τ ) dτ = X(−Ω).

27. (d) - Unlike the total SNR, by definition the SQNR depends just on the quantizer resolution. Therefore,
it grows exponentially with the number of bits (i.e. by about 6 dB per bit). However, the actual SQNR
depends also on the power of the input signal. If its amplitude is too small to stimulate the whole input
range of the ADC, the SQNR can be quite lower than 6.02B.

28. (c) - The Costantinides expressions are conceived for the transformation of normalized frequency values.
Therefore, the unit circle and not the imaginary axis of the complex plane has to be preserved.

29. (a) - The fundamental components of tooth meshing vibrations cause usually the largest spectral peaks,
which are certainly larger than those of their harmonics and than the broadband noise floor due to the
gear rattle. The ghost components are usually minor, unless some major fault or gear damage occurs.

30. (b) - Recalling that s(n) − s(n − 1) = 2δ(n), by applying the Fourier transform to both sides of the
previous equation, it follows that S(ejω ) − e−jω S(ejω ) = 2. Therefore, S(ejω ) = 1−e2−jω .

54