Copyright: B. Boashash
Printed in Brisbane
Boashash, Boualem
Signals and linear systems.
Bibliography
ISBN 1 86435 436 4
1. Signal processing. 2. System analysis. 3. Linear time
invariant systems. I. Boashash, Boualem. II. Queensland
University of Technology. Signal Processing Research Centre.
III. Title.
621.3822
Contents
Preface vii
1 Introduction to Signals and Systems 1
1.1 The Basic Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Signal Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 System Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2 System Analysis in the Time Domain 15
2.1 The Convolution Integral for Linear Shiftinvariant Systems . . . . . . . . . . . . 15
2.2 Properties of the Convolution Integral . . . . . . . . . . . . . . . . . . . . . . . . 17
2.3 Step Response of a Linear System . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.4 Stability of Linear Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3 Spectral Representation of Signals 23
3.1 The Fourier Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.1.2 Trigonometric Fourier Series Representations for Periodic Signals . . . . . 24
3.1.3 The Exponential Fourier Series . . . . . . . . . . . . . . . . . . . . . . . . 25
3.1.4 Properties of the Fourier Series Coecients . . . . . . . . . . . . . . . . . 25
3.1.5 Representation of the Fourier Coecients in the Frequency Domain . . . 28
3.1.6 Systems with Periodic Inputs . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.1.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.2 The Fourier Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.2.1 Denition of the Fourier Transform . . . . . . . . . . . . . . . . . . . . . . 32
3.2.2 Convergence of the Fourier Transform . . . . . . . . . . . . . . . . . . . . 33
3.2.3 Fourier Transforms in the Limit . . . . . . . . . . . . . . . . . . . . . . . . 34
3.2.4 Energy Spectral Density . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.2.5 Properties of the Fourier Transform . . . . . . . . . . . . . . . . . . . . . 36
3.3 The Laplace Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.3.1 Existence of the Laplace transform . . . . . . . . . . . . . . . . . . . . . . 39
3.3.2 Relation Between Fourier Transform and Laplace Transform . . . . . . . 41
3.3.3 Conditions for the Realisability of a Linear System or Filter . . . . . . . . 42
3.3.4 Frequency Response of Transfer Function using the Poles and Zeros of the
Laplace Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.3.5 Poles and Zeros . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.4 Time{Varying Spectra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.4.1 The Hilbert Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.4.2 Analytic signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
i
ii CONTENTS
3.4.3 Time{Frequency Relationships . . . . . . . . . . . . . . . . . . . . . . . . 48
3.4.4 Time{Frequency Distributions . . . . . . . . . . . . . . . . . . . . . . . . 50
4 System Analysis in the Frequency Domain 53
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.2 An Example of Frequency Domain System Analysis . . . . . . . . . . . . . . . . 53
4.3 Transfer Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
4.4 Steady{State System Response . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
4.5 Ideal Filters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
4.6 Fourier Transforms of Periodic Signals and Sampling . . . . . . . . . . . . . . . . 59
5 Introduction to Discrete{time Signal Processing 63
5.1 Analog to Digital Conversion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
5.1.1 Sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
5.1.2 Periodic Sampling: a tutorial . . . . . . . . . . . . . . . . . . . . . . . . . 67
5.1.3 Reconstruction of Band{Limited Signals . . . . . . . . . . . . . . . . . . . 70
5.1.4 Quantising and Encoding . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
5.2 Digital Processing of ContinuousTime Signals Using the DTFT . . . . . . . . . . 77
5.2.1 Discrete{time Processing of Continuous{time Signals. . . . . . . . . . . . 78
5.2.2 The Discrete{Time Fourier Transform . . . . . . . . . . . . . . . . . . . . 79
5.2.3 Convergence of the DTFT . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
5.2.4 Properties of the DTFT. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
5.3 The z{Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
5.3.1 Denition of the z{Transform . . . . . . . . . . . . . . . . . . . . . . . . . 86
5.3.2 Properties of the z{Transform . . . . . . . . . . . . . . . . . . . . . . . . . 91
5.3.3 Inverse z{Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
5.4 Digital Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
5.4.1 Dierence Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
5.4.2 Discrete Convolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
Bibliography 98
v
vi LIST OF TABLES
2 The material
The material contained in this document has been developed as lecture notes over a period of
more than 10 years. I wrote the rst draft in 1984 at the University of Queensland using as a
basis a course developed at the University of Lyon, where I lectured in 1982{3. I then further
adapted the notes with the aim of establishing a strong link between theory and practice,
while keeping the highest possible standards, logic and rigour in the text. I made this possible
by emphasising the importance of concepts and the philosophy of signals processing, and the
inspiration provided by early applications and developments in the eld of geophysical signal
processing, where some of the early developments in DSP took place. (I witnessed rst hand the
power of DSP in my 3 year long career in industry, as a research engineer in the seismic signal
processing division of Elf{Aquitaine, one of the biggest European oil companies.)
Further, the process of learning was facilitated by the use of suitable cross{references within
the text to previously acquired knowledge, and generally speaking, the presentation of the
material as a natural extension of basic principles already mastered by the students.
The lecture notes were revised every year since 1984. A major revision occurred in 1991
when I moved to the Queensland University of Technology, with the task of establishing a
strong undergraduate signal processing program as well as a research centre of international
standard. Since 1991, the material continued to be upgraded and taught under my supervi
sion. A complementary booklet containing the exercises and examples covered in tutorials is in
preparation.
vii
viii Preface
This subject is the prerequisite for the subject \Digital Signal Processing" (DSP). I hope
that this eort will help students towards mastering these important concepts and techniques,
and I look forward to receiving any suggestions for further improvement of this material.
The material contained in these notes comes straight from the original lecture notes of the
author, except for some improvements in the quality of gures and some examples that have
been added.
Parts of this manuscript appeared in an earlier joint publication with Dr. Zoubir (B.
Boashash and A. Zoubir, Signals in Linear Systems, QUT, 1996) whom I wish to thank.
The author also acknowledges the valuable help provided by Dr B. Senadji who checked the
text and by the various research assistants who assisted me in the revision of the notes.
0 Warning and disclaimer: Due to the ongoing process of revision, this draft document may contain typos
and errors due to the editing process. These will be corrected during the lectures and in the next version.
X (f )  System, H (f )  Y (f ) = X (f ) H (f )

;T
2
T
2 t
Figure 1.4: Square pulse.
x(t)
Quantised, Discretetime and Digital Signals. A quantised signal is one whose values
may assume only a countable number of values, or levels, but the changes from level to level
may occur at any time. A discretetime signal is one discretised in time. A quantised and
discretetime signal is referred to as digital signal (discretised in both time and amplitude).
Examples of digital signals are computer data.
Periodic and Aperiodic Signals. A signal x(t) is periodic if and only if there exists a
constant T0 > 0 such that
x(t + T0 ) = x(t)
;1 < t < 1
where T0 is called the period. Any signal that does not satisfy this property is called aperiodic.
x(t)
t1 t2 t3 t4 t5 t
t
Ae
t T0 t
(a) (b)
Figure 1.7: Examples of aperiodic (a) and periodic, x(t) = sin(2t=T0 ) (b) signals.
Complex Signals. It is often convenient to represent real signals in terms of complex quan
tities. For example, given the real, sinusoidal signal, x(t) = A cos(2fo t + ), we refer to the
complex signal
z (t) = A ej(2f0 t+)
x(t) is obtained from z (t) through
x(t) = <fz(t)g:
These expressions for x(t) are called time{domain representations. An alternative represen
tation for x(t) is provided in the frequency domain: Since z (t) = A ej (2f0 t+) is completely
specied by the real numbers A and for a given value of f0 , this alternative frequency{domain
representation can take the form of two plots, one showing the amplitude A as a function of
frequency f , and the other as a function of f as depicted in Figure 1.8.
jz(t)j (z(t))
6 6
jAj + 2f0 t
 f  f
0 f0 0 f0
Figure 1.8: Representation of z (t) in the frequency domain.
Singularity Functions. An important class of signals very much used in signal processing
are the singularity functions.
The unit step function.
The unit step function, u(t) is dened as
u(t) = 0
t<0
1
t0
B. Boashash, Signals in Linear Systems
6 Chapter 1. Introduction to Signals and Systems
6u(t)
1
 t
0
Figure 1.9: The unit step function.
3
2
1

t
;11 3 5 7
;2
x(t) = 4u(t) ; 6u(t ; 2) + 3u(t ; 4) + 2u(t ; 5)
The unit ramp function.
The unit ramp function r(t) is dened as
r(t) = t
t0
0
t<0
6r(t)
1
j
 t
0 1
Figure 1.10: The unit ramp function
and
Zt
u( )d = r(t):
;1
6
(t)
6
 t
0
Figure 1.11: The unit impulse function.
The motivation for dening a function with such properties is due to the need for representing
phenomena that happen in time intervals that are short when compared with the resolution
capability of any measuring apparatus used, but which produce an almost instantaneous change
in a measured quantity.
When dening the Dirac or delta function by means of distributions, it is possible to consider
the derivative of a delta function, even though it does not exist in the usual sense as a function.
As
(t) is not a usual function, we dene it by considering the limit of a conventional function
as some parameter approaches zero. For example, consider the signal
(t) = 1
rect( t ) = jtj ""
"
1
" " " 0
jtj > 2
2
No matter how small " is,
" (t) always has unit area. If " ! 0, the height of
" (t) ! 1 such
that unity area lies between the function and the t{axis.
There exist other functions which converge to the impulse function as " ! 0. These include
for example
1 t 2
p1 (t) = " t sin "
which is depicted in Figure 1.13.
B. Boashash, Signals in Linear Systems
8 Chapter 1. Introduction to Signals and Systems
6
" (t)
1
"

;" 2
"
2 t
Figure 1.12: Square pulse approximate for the unit impulse function.
2
1.8
1.6
epsilon=1.0
1.4 epsilon=0.5
1.2
p1(t)
0.8
0.6
0.4
0.2
0
5 4 3 2 1 0 1 2 3 4 5
t
Denition 1.2.1
(t) is a function dened by an integral function which assigns the value x(0)
to any function x(t) continuous at the origin
Z1
x(t)
(t)dt = x(0)
;1
The more general case of the above equation is
Z1
x(t)
(t ; t0 )dt = x(t0 )
;1
Properties of
(t).
Assuming that x(t) is continuous at t = t0 , we have:
1.
(at) = 1
(t)
jaj
2.
(;t) =
(t)
Z1
3. x(t)
(t ; t0 ) dt = x(t0 )
;1
Z1
4. x()
(t ; ) d = x(t) important (convolution)
;1
5. x(t)
(t ; t0 ) = x(t0 )
(t ; t0 )
Z1 Z1
6.
;1
0 (t ; )x() d = ;1
(t ; )x0 () d,
0 t0 2 (1 + cos 2(!0 t + )) dt
A 2 A2 Z t0 +T0
t
= 2T t]t0 + 2T
0 +T0
cos 2(!0 t + )dt
0 t0
0
 {z }
=0
= A
2
2
1.3 System Models
We want to look at ways to represent the eects of systems on signals, i.e. we want to design
appropriate system models that represent correctly the interaction of signals and systems and
the relationship of causes and eects for that system.
Representations for Systems. We will express the dependence of the system output on the
input symbolically as
y(t) = T x(t)]
which is read: y(t) is the response of T to x(t).
The operator T identies the system and species the operation to be performed on x(t) to
produce y(t). For a multiple{input, multiple{ output system, x(t) and y(t) are vectors and T is
a matrix.
B. Boashash, Signals in Linear Systems
1.3. System Models 11
x1(t)   y1 (t)
x2(t)   y2 (t)
..
. System T ..
.
xp(t)   yq (t)
;1 ;1
which is known as the superposition integral or convolution. The function h() is called the
impulse response of the system and h() is the response of the system to a Dirac impulse applied
at t = 0, i.e. T
(t)] = h(t).
Properties of Systems
Continuous{time and Discrete{time systems
If the signals processed by a system are continuous{time signals, the system itself is referred
to as a continuous time system. If, on the other hand, the system processes signals that
exist only at discrete times, it is called a discrete{time system. If the signals are quantised
to a nite number of levels, the system is called a quantised system. If the quantised system
is also a discrete{time system, then it is called a digital system.
Fixed and time{varying systems
x(t) y(t)
System
t0 t1 t t0 t1 t
Denition 1.3.2 A continuous{time system is causal if and only if the condition x1(t) =
x2 (t) for t t0 implies the condition T x1 (t)] = T x2 (t)] for t t0 for any t0 , x1 (t) and
x2 (t).
Instantaneous and dynamic systems
A system for which the output is a function of the input at the present time only is said
to be instantaneous (or memoryless).
A dynamic system (or a system with memory) is one whose output depends on past or
future values of the input in addition to the present time. If the system is also causal, the
output of a dynamic system depends only on present and past values of the input. An
example of a causal and timeinvariant system is given in Figure 1.15.
Example 1.3.1 A resistor R is an instantaneous system,
i(t) R
v(t)
v(t) = R i(t)
Example 1.3.2 For an inductor L
i(t) L
v(t)
1 Zt
i(t) = L v(x)dx:
;1
This relation shows clearly that i(t) depends on past values of v(t). Thus an inductor L is
a dynamic causal system.
B. Boashash, Signals in Linear Systems
1.3. System Models 13
Linear and nonlinear systems
A linear system is a system for which the theorem of superposition is valid that is, if x1 (t)
and x2 (t) are two inputs, we must have for any arbitrary constants 1 and 2 and any t
T x (t) + x (t)]
1 1 2 2 = 1 T x1 (t)] + 2 T x2 (t)]
= 1 y1 (t) + 2 y2 (t):
The property
T x(t)] = T x(t)]
is called homogeneity and
T x (t) + x (t)] = T x (t)] + T x (t)]
1 2 1 2
is referred to additivity.
Example 1.3.3 (System dierential equation) Let a system be given in Figure 1.16. A
R
L
vi (t) C vo (t)
or equivalently
LC d dt
2 v (t)
o + RC dvo (t) + v (t) = v (t)
2 dt o i
x(t)  System T  ?
two dierent approaches to solve this problem that make use of signal theory1 .
1. Carry out a solution in the time domain using the convolution integral.
2. Use frequencydomain analysis by means of the Fourier Transform or the Laplace Trans
form.
These two methods are of common use, and will be discussed in this chapter for LSI systems.
In this chapter, we will examine the rst method. We will concentrate on linear systems
because many of the physical systems can be very well modelled as linear systems. A very
important operation in timedomain analysis techniques is the convolution integral.
Denition 2.1.1 The impulse response of a system is its response to a unit impulse at time
t = 0 with all initial conditions of the systems zero.
1 We coul also describe the interaction of the system and the signals by a dierential equation and solve it.
This is a dicult task in practice and is better replaced by signal processing theory.
15
16 Chapter 2. System Analysis in the Time Domain
(t) ;! h(t) read h(t) is the response of
(t)
because
(t ; ) ;! h(t ; ) for all 2 R shift{invariance
Z 1 )
(t ; )
x( ;! x()h(Zt ;1 ) for all x(), 2 R homogeneity
x()
(t ; )d ;! x()h(t ; )d additivity
;1 Z;11
which is equivalent to x(t) ;! x()h(t ; )d convolution integral Thus
;1
Z1 Z1
y(t) = x()h(t ; )d = h()x(t ; )d
;1 ;1
Denition 2.1.2 If x(t) and h(t) are two signals, the convolution of x(t) with h(t) is a new
signal y(t) given by the operation
Z1
y(t) = x()h(t ; )d
;1 < t < 1
;1
Alternatively, y(t) can be found by convolving h(t) with x(t) which is expressed as
Z1
y(t) = h()x(t ; )d
;1 < t < 1
;1
A useful symbolic notation is very often employed
y(t) = x(t) h(t) = h(t) x(t)
Example 2.1.1 Consider the convolution of the two rectangular pulse signals
t ; 5
x(t) = 2rect 2
t ; 2
h(t) = rect 4
Z1
x(t) h(t) = x()h(t ; )d
;1
Calculating x(t) h(t) is conceptually no more di
cult than ordinary integration when the
two signals are continuous for all t. Often, however, one or both of the signals is dened in a
piecewise fashion, and the graphical interpretation of convolution becomes especially helpful. We
list in what follows the steps of this graphical aid to computing the convolution integral. These
steps demonstrate how the convolution is computed graphically in the interval ti;1 t ti ,
where the interval ti;1
ti ] is chosen such that the product x()h(t ; ) has the same analytical
form. The steps are repeated as many times as necessary until x(t) h(t) is computed for all t.
Step 1 For an arbitrary, but xed value of t in the interval ti;1
ti], plot x(), h(t ; ), and the
product x()h(t ; ) as a function of . Note that h(t ; ) is a folded and shifted version
of h() and is equal to h(;) shifted by t seconds.
Step 2 Integrate the product x()h(t ; ) as a function of . Note that the integrand depends on
t and , the later being the variable of integration, which disappears after the integration
is completed and the limits are imposed on the result. The integration can be viewed as
the area under the curve represented by the integrand. The convolution of x(t) and h(t) is
graphically presented in Figure 2.3.
Z1
x(t)
(t) =
;1
x()
(t ; ) = x(t)
Z1 Zt
x(t) u(t) =
;1
x()u(t ; )d =
;1
x()d integrator]
Z 1
Z 1
jy(t)j =
;1 x()h(t ; )d
;1 jx()jjh(t ; )j d
We require that the input is bounded, thus
jx()j M < 1
6x()
2

4 5 6
6h()
1

4
6h(;)
1

;4
6h(t ; )
1

t;4 t
Figure 2.2: The plot of x(), h(), h(;), and h(t ; ) as a function of .
6 case 1

t;4 4 t 6
6 case 2

t;4 4 6 t
6 case 3

4 t;4 6 t
6 x(t) h(t)
A
A
A
A
A
A
A
A
A 
4 6 8 10 t
Figure 2.3: Convolution of x(t) and h(t).
k
l=k
l 6= k
a
= Ek
l;k
where
k = 10
kelsewhere
=0
is Kronecker's delta function. If the constant Ek = 1, the i (t) are said to be orthonormal
signals.
Orthogonal signals are useful in that they lead to a series representation of signals in a
relatively simple fashion. Let i (t) be an orthonormal set of signals on an interval a t b
and let x(t) be a given signal with nite energy over the same interval. We can represent x(t)
in terms of f i (t)g by the series
1
X
x(t) = Xi i (t)
i=;1
where
Zb
Xk = x(t) k (t)dt
k = 0
1
2
: : :
a
Below, we investigate a method for expressing periodic signals in terms of \harmonically re
lated"1 complex exponentials. The choice of complex exponentials as an orthogonal basis is
appropriate since the complex exponentials are periodic, relatively easy to manipulate mathe
matically, and the result has meaningful physical interpretation.
1 Harmonics will be dened later in the text
23
24 Chapter 3. Spectral Representation of Signals
3.1.2 Trigonometric Fourier Series Representations for Periodic Signals
Let x(t) be a periodic signal with period T0 with f0 = 1=T0 . The general form of the Fourier
series representation of x(t) is the following
x(t) = a0 + a1 cos(2f0 t) + a2 cos(4f0 t) +
+ b1 sin(2f0 t) + b2 sin(4f0 t) +
;1 < t < 1
which can be rewritten as
1
X 1
X
x(t) = a0 + an cos(2nf0 t) + bn sin(2nf0 t)
;1 < t < 1 (3.1)
n=1 n=1
f0 : fundamental
2f0 : 1st harmonic
..
.
nf0 : (n ; 1)th harmonic
The problem is to nd a0 , an , and bn for a given x(t) and T0 . Then we shall obtain the
representation of x(t). Using
Z T0
sin(m2f0 t) sin(n2f0 t)dt = 0
m 6= n
T0
m=n
Z 0T0 2
and
Z T0
sin(m2f0 t) cos(n2f0 t)dt = 0
for all m
n 2 N
0
and for m 6= 0
2 Z T0
am = T x(t) cos(m2f0 t)dt (3.3)
Z T0
0 0
b = 2
m T0 x(t) sin(m2f t)dt 0 (3.4)
0
This Fourier series will converge to the function itself if it satises the Dirichlet conditions.
Dirichlet conditions: If
(i) x(t) is bounded, absolutely integrable
Z h+T0
h
jx(t)jdt < 1
and of period T0 and
B. Boashash, Signals in Linear Systems
3.1. The Fourier Series 25
(ii) x(t) has at most a nite number of maxima and minima in one period and a nite number
of discontinuities,
then the Fourier series of x(t) converges to x(t) at all points where x(t) is continuous, and
converges to the average of the righthand and left hand limits of x(t) at each point where x(t)
is discontinuous, i.e.
where t0 is the time where x(t) is discontinuous. Note that the conditions (i) and (ii) are
sucient but not necessary.
where
1 ZT 2nt
Xn = T x(t) exp ;j T dt (3.6)
0 0
are complex constants. Each term of the series has a period T0 and a fundamental radian
frequency !0 . Hence, if the series converges, its sum is periodic with period T0 .2 Such a series
is called the complex exponential Fourier series and the Xn are called the Fourier coecients.
n=1
where
ZT
a0 = X0 = T1 x(t)dt
0
2 ZT 2nt
an = 2<fXn g = T x(t) cos T dt
2
0
ZT 2nt
bn = ;2=fXn g = T x(t) sin T dt
0
In terms of magnitude and phase of Xn , the realvalued signal x(t) can be expressed as
X1
x(t) = X0 + jXn j cos 2nt T + argXn (3.7)
n=1
X1 2nt
= X0 + An cos T + "n (3.8)
n=1
where
An = 2jXn j = 2jX;n j
and
"n = argXn
This representation is alternative to the previous one and is more compact and meaningful.
Thus, for real signals, the magnitude of the Fourier coecients is an even function of the index
n. (where n represents the frequency f)
If x(t) is real and even, i.e. x(t) = x(;t), the imaginary part of Xn is zero, because
x(t) sin(2f0 t) is an odd function that integrates over an interval symmetrically placed about
t = 0. Therefore:
If x(t) is real and even, Xn is real and even. (ie Im[Xn]=0)
Similarly, if x(t) is real and odd, Xn is imaginary and odd. (ie Re[Xn]=0)
Example 3.1.1 Consider the following signal
8
< 0
;T=2 < t < ;=2
x(t) = : K
;=2 < t < =2 x(t) = x(t + T )
0
=2 < t < T=2
B. Boashash, Signals in Linear Systems
3.1. The Fourier Series 27
Signals of this type can be produced by a pulse generator and are used extensively in radar and
sonar systems. Calculate the Fourier coe
cients of x(t).
1 Z T=2 2nt
Xn = T x(t) exp ;j T dt
;T=2
1 Z =2 K exp ;j 2nt dt
= T ;=2 T
K T j exp ;j 2n ; exp +j 2n
= T 2n 2T 2T
j K 2j sin 2n
= ; 2n 2T
= K sin n = KT sin n
n T nT T
= K sinc n
T T
where
sinc(x) = sin(x) :
x
Parseval's Theorem
The averaged normalised power of a periodic waveform x(t) is given by
1 Z 1 Z
Pav = T
0 <T0 >
jx(t)j dt = T
2
0 <T0 >
x(t)x (t)dt
R
where <T0 > denotes integration over interval of length T0 . If we replace x(t) by its Fourier
series representation, we get
Z 1 !
1 X
Pav = T x(t) N X e;j 2nf0 t
dt
0 <T0 >
n=;1
X1 1 Z
=
Xn T x(t)e ; j 2nf0 t
dt
n=;1 0 <T0 >
X1
= Xn Xn
n=;1
1
X
= jXnj2
n=;1
so that
Z 1
X
Pav = T1 jx(t)j dt =
2
jXnj :
2
0 <T0 > n=;1
The average power of a periodic signal x(t) is the sum of the powers in the components of its
Fourier series.
B. Boashash, Signals in Linear Systems
28 Chapter 3. Spectral Representation of Signals
Periodic Convolution
Let x(t) and y(t) be two periodic signals with same period fundamental period T0 . We dene
the circular convolution of x(t) and y(t) by
1 Z
z (t ) = T x( )y(t ; )d
(3.9)
0 <T0 >
where the signal is taken over one period T0 . One can show that
1 Z
z (t + T0) = T x( + T0 )y(t ; + T0 )d
0 <T0 >
1 Z
= T0 x( )y(t ; )d
<T0 >
= z (t)
and therefore z (t) is periodic with period T0 . The reader should show that the periodic convo
lution is commutative and associative. We can write z (t) in a Fourierseries representation with
coecients
1 Z T0
Z =n T0 z(t)e;jn2f0 tdt
1 Z0
T0 1 Z T0
T0 0 x( )y(t ; )e
= T ;j 2nf0 t ddt
Z T0
0 0
Z T0
= T1 x( )e;j2nf0 T1 y(t ; )e;j2f0 (t; ) dt d
1
0 0
Z T0 10 Z0 T0 ;
= T x( )e ; j 2nf0
T0 ; y(v)e
;j 2f0 v
(;dv) d
0 0
Z T0
= T1 x( )e;j2nf0 Ynd = Xn Yn
0 0
where Xn and Yn are the Fourier series coecients of x(t) and y(t), respectively.
3.1.5 Representation of the Fourier Coecients in the Frequency Domain
Line Spectra. We have seen that if x(t) is periodic with period T0, then
1
X
x(t) = Xn ej 2nf0t
n=;1
Z
Xn = T1 x(t)E ;j2nf0 t
0 <T0 >
and we have (for a real signal)
Xn = X; n
jXnj = jX;nj
"n = ;";n:
This observation allows a periodic signal to be characterised graphically in the frequency domain
by two plots
B. Boashash, Signals in Linear Systems
3.1. The Fourier Series 29
One showing amplitudes of the components versus frequency, which is known as the am
plitude spectrum of the signal.
The other showing the relative phase of each component versus frequency, which is called
the phase spectrum of the signal.
We see that these spectral components or lines, are present at both positive and negative
frequencies.
For a real signal, the amplitude spectrum is even and the phase spectrum is odd.
Example 3.1.2 (Pulsetrain signal.) Let
1
X t ; nT
x(t) = A rect
0
:
n=;1
x(t)
6

;T 0 ; 2
2 T0 t
Since x(t) is periodic with period T0 it may be represented as a Fourier series, with fundamental
period T0 and fundamental frequency f0
1
X
x(t) = Xn ej 2nf0t :
n=;1
Solving for the Fourier series coe
cients, we found in example 3.1.1
A
n
Xn = T sinc T :
0 0
1. We note that the width of the envelope of the amplitude spectrum increases as the pulse
width decreases (for xed T0 ). That is, the pulse width of the signal and its corresponding
spectral width are inversely proportional.
2. We also note that the separation between lines in the spectrum is T10 , and therefore the
density of the spectral lines with frequency increases as the period T0 of x(t) increases (for
xed ).
Example 3.1.3 We which to compute and plot the magnitude and phase of the Fourier series
coe
cients of the signal given in Figure 3.1.
x(t) = ;k
;1 < t < 0
k
0<t<1
We have x(t + 2) = x(t) and therefore !0 = 2=2 = .
8
< 2k
n odd
Xn = : jn
0
n even
B. Boashash, Signals in Linear Systems
30 Chapter 3. Spectral Representation of Signals
x(t)
6
k

0 1 t
which leads to
8 2k
<
jXnj = : jnj
n odd
0
n even
and
8
< ;2
n = 2m ; 1
m = 1
2
: : :
argXn = : 0
n = 2m
m = 0
1
2
: : :
2 n = ;(2m ; 1)
m = 1
2
: : :
s j j
6Xn s 2k
s s 2k
3
s s 2k
5
s s s s s 
0 1 2 3 4 5 n
j j
6arg Xn
s s s
2
s s s 1 2s 3 4s 5 
0 n
;
s
2
s s
which leads together with (3.10), and using the linearity property, to
1
X
y(t) = H (nf0) Xn ej2nf0 t :
n=;1
This equation indicates that the output signal is the summation of exponentials with coecients
H (nf0) Xn
Note that since H (nf0 ) is a complex constant for each n, it follows that the output is also
periodic with Fourierseries coecients H (nf0 ) Xn . Since the fundamental frequency of x(t) is
T0 = f0 , the period of y(t) is equal to the period of x(t). Hence, the response of an LTI system
1
3.2.1 De
nition of the Fourier Transform
We have seen in the previous section that the Fourier Series representation of a periodic signal
x~(t) is
1
X
x~(t) = Xn ej 2nf0 t (3.11)
n=;1
where
1 Z T0 =2
Xn = T x~(t)e;j 2nf0t dt (3.12)
0 ;T0 =2
To develop a spectral representation for nonperiodic signals (i.e. signals with period T0 !
1), we let the frequency spacing f0 = T0;1 approach zero such that f0 ! df , an innitesimally
small quantity, and the product nf0 approaches a continuous frequency variable f .
6 6
x~(t)
T0 
x(t)
 
t t
X (f ) = F fx(t)g
x(t) = F ;1 fX (f )g
or x(t) $ X (f )
3.2.2 Convergence of the Fourier Transform
Sucient conditions for the convergence of the Fourier Transform integral are that:
Z1
1.
;1
jx(t)jdt exists, indicating that signals are absolutely integrable. This property comes
from the fact that the Fourier transform must be bounded, i.e.
x(t)e; j 2 ft dt
jx(t)jdt < 1:
2. Any discontinuities in x(t) be nite, and the number of such discontinuities must be nite.
This property is referred to as x(t) is \well behaving". Except for impulses, most signals
of interest are well behaved and satisfy (3.15).
These conditions exclude a number of important signals such as sinusoids that are not abso
lutely integrable. By means of distributions (delta functions), signals of this type can be handled
using essentially similar methods as for nite energy signals.
Example 3.2.1 The Fourier transform of the rectangular pulse x(t) = rect(t= ) is
Z1
X (f ) = x(t)e;j2ft dt
;1
Z =2
= e;j2ft dt
;=2
j e;jf ; ejf
= 2f
= sinc(f )
= 2 1 ; cos(2ft)dt
0
= sinc2(f )
Example 3.2.3 The Fourier transform of the impulse function is
Z1
Ff
(t)g = ;1
e;j 2ftdt = 1
Z1
X (f ) = ej2f0t e;j 2ftdt
Z;1
1
= e;j 2(f ;f0 )t dt
;1
=
( f ; f0 )
x(t) X (f )
1. 1
(f )
2. u(t) 0:5
(f ) + 1=(j 2f )
3.
(t) 1
4.
(t ; t0 ) e;j2f0 t
5. rect(t= ) sin(f )=(f )
6. sin(2fB t)=(t) rect(f=2fB )
7. ej 2f0 t
(f ; f0)
1
X X1
8. anejn2f0 t an
(f ; f0)
n=;1 n=;1
9. cos(2f0 t) 0:5
(f ; f0 ) +
(f + f0 )]
10. sin(2f0 t) ;j0:5
(f ; f ) ;
(f + f )]
0 0
This is Parseval's theorem for Fourier transforms. jX (f )j2 is the energy spectral density of
x(t). Integration over all frequencies yields the total energy contained in a signal. Integration
over a nite range B of frequencies gives the part of energy contained in this range B . Likewise
for jx(t)j2 {known as the instantaneous power.
Example 3.2.5 Consider the onesided exponential signal
x(t) = e;t u(t)
From Table 3.1 we have
jX (f )j 2
= 1 + 412 f 2 :
The total energy in this signal is
Z1
E =
;1
jX (f )j df
2
Z1 1 df
=
;1 1 + 42 f 2
= 12
Integration: R;1
t x(t0 )dt0 $ (j 2f ); X (f ) + X (0)
(t) 1 1
2
Multiplication: x (t)
x (t) $ X (f ) X (f )
1 2 1 2
now
Z1
x1 (t ; )e;j 2ft dt = X1 (f )e;j 2f
;1
therefore
Z1
x1(t ; ) = X1 (f )e;j2fej2ft dt
;1
Z1Z1
) x (t) x (t) =
1 2
;1 ;1
X1 (f )x2 ()e;j2f dej 2ft dt
Z1
= X1 (f )X2 (f )ej2ft dt = F ;1 X1 (f )X2(f )]
;1
therefore
F x (t) x (t)] = X (f )X (f )
1 2 1 2
u(t) $ 1
(j2f ); +
(f ) because u(0) = 1 and u(t) = R;1
1 1
2
t
(t)dt (integration theorem).
t $ ;j sgn f (duality)
1
Example 3.2.6
1
X 1
X
y(t) =
(t ; mTs ) $ fs
(f ; mfs ) fs = Ts;1
m=;1 m=;1
Proof: this signal is a periodic train of impulses called the ideal sampling waveform. Its
Fourier series is
1
X
y(t) = Yn ej2nfs t
n=;1
R
with Yn = T1s ;TTs =s2=2
(t) e;j 2nfs t dt = fs
P
therefore, y(t) = 1 ;1 fs e
j 2nfs t
1
X 1
X
y(t) = fs ej 2nfs t $ fs
(f ; fs)
;1 ;1
Example 3.2.7 Application of the dierentiation theorem.
Consider the triangle signal:
$(t=T ) = 1 ; jTtj
jtj < T
0
elswhere
d ; T sgn(t)
1
jtj < T
dt $(t=T ) = 0
elsewhere
B. Boashash, Signals in Linear Systems
3.3. The Laplace Transform 39
d2 $(t=T ) = 1
(t + T ) ; 2
(t) + 1
(t ; T )
dt T T T
then
d2 $(t=T ) $ (j 2f )2
F $(t=T )] = 1 hej2fT ; 2 + e;j 2ft i
dt T
;! $(t=T ) $ T sinc fT 2
We shall also use the operator notation X (s) = Lfx(t)g and sometimes x(t) $ X (s).
When a Laplace transform is computed, there is a restriction on the value of for which
the transform is valid. This value determines the region in the complex splane in which the
integral converges.
B. Boashash, Signals in Linear Systems
40 Chapter 3. Spectral Representation of Signals
6
1 ^T
;;@@
; @
; @ t
;T T
d ^
dt T
1 6
T
t
;1
T d2^
dt2 T
6
1 1
T T
6 6 t
;2 ?
T
6j 2f
The complex splane

The value of is important in determining a path of integration that could be used for evaluating
the inverse Laplace transform
I 0+j1
x(t) = j 21 X (s)est ds
0 ;j 1
by contour integration in the complex plane. If 0 = 0, it then follows that the Fourier transform
of x(t) exists and can be obtained form X (s) by substituting s = j 2f .
j 2 f
Examples:
Lfu(t)g = 1s
<fsg > 0
Ffu(t)g = j 21f + 12
(f )
Ffej f tu(t)g
2 o
= j 2(f1; f ) + 12
(f ; f0 )
0
Lfej f t u(t)g
2 o
1
= s ; j 2f
0
H (s) = N (s)
D(s)
then the order of N (s) must be smaller than the order of D(s).
4. jH (s)j is nite for all values of f .
5. jH (s)j ! 0 as f ! 1 and jh(t)j has the same behaviour as f1n , n 1.
(s + 2)2 + 9
2 1 2 f
3 2 1 1 2 3
3.3.5 Poles and Zeros of the Laplace Transform and its Region of Conver
gence
In many applications, the Laplace transform (or some part of it) can often be expressed in terms
of a rational polynomial function of the form
Y
N
(s ; zk )
X (s) = N (s) k=1
D(s) = Y
M
M N
(s ; pk )
k=1
The roots of N (s) are the zeros of X (s), since they represent the points in the Laplace transform
plane where X (s) = 0. Similarly, the roots of D(s) are the poles of X (s) since X (s) ! 1 at
such values of s. The poles and zeros of the Laplace transform completely determine the ROC.
Example 3.3.3
x(t) =
(t) ; 34 e;t u(t) + 13 e2t u(t)
; 1)2
H (s) = (s +5s1)(
+ 13 = A + B
s + 5) s + 1 s + 5
B. Boashash, Signals in Linear Systems
3.3. The Laplace Transform 45
j2 f
ROC
1 1 2
= s + 1 + s3 s+ 7+s210+s 15
+ 10
2
s+9
By inspection, we can nd the rst root of the denominator as s1 = ;1. To nd the remaining
roots, we perform the following division.
s3 + 7s2 + 15s + 9 = s2 + 6s + 9 = (s + 3)2
s+1
Thus
s3 + 7s2 + 15s + 9 = (s + 1)(s + 3)2
and
A + B + C
H~ (s) = s + 1 s + 3 (s + 3)2
B. Boashash, Signals in Linear Systems
46 Chapter 3. Spectral Representation of Signals
A and C is determined as before
A = H~ (s)
(s + 1)js=;1 = 1 ; 104 + 10 = 41
C = H~ (s)
(s + 3)2 js=;3 = 9 ; 30
;2
+ 10 = 11
2
For determining B , we take s = 0, which leads to
1 + B + 11 1 = 10 :
4 3 29 9
Thus B = 34 . Finally,
1 1 Z 1 x( )
Hfx(t)g = x(t) t = ;1 t ; d (in principle value)
0 f
f0 0 f0 f
p(t) = jxk(xt)kj 0
2
real
p(t) has the dimensions of a probabilty density. The \time interval where it is concentrated"
is given by $t such that equivalent duration:
1 Z1
($t) = kxk
2
t2 jx(t)j2 dt
;1
Similarly in the frequency domain with $y equivalent bandwidth.
B. Boashash, Signals in Linear Systems
3.4. Time{Varying Spectra 49
R(f)
f0 0 f0 f
Figure 3.10:
Timefrequency inequality
R1 2
t jx(t)j2 dt
($t) = ;1
2
R1 equivalent duration
;1 jx(t)j dt
2
R1 2
f jx(f )j2 df
($f ) = ;1
2
R1 equivalent bandwidth
;1 jx(f )j df
2
Z1 Z1
! ;1
jx_ (t)j2
dt =
;1
42 f 2 jx(f )j2 df
Z 1 21 1 Z 1 12
! $t $f = ;1 jt x(t)j dt
2 ;1 jx_ (t)j dt
kxk;2
2 2
In comunications the bandwidth required to transmit a message increases when the speed
of information transmission is increased.
B. Boashash, Signals in Linear Systems
50 Chapter 3. Spectral Representation of Signals
There is an analogy with quantum mechanics when the position and velocity of a particle
obeys:
If we use another denition of bandwidth and duration, we will still obtain an inequality
but with a dierent limit.
Given a linear lter h(t) $ H (f ), $t represents the memory of the lter, $f is its
bandwidth ! $t $f 41 .
Case of random process: Rx ( ) $ Sx (f ). $ t represents the statistical memory on
\correlation time" and f is the spectral bandwidth.
$ $f 41
t2
x(t) = k1 e; 2 $ k e;r2 2f 2
2
Signal Transform
x(t) X (s)
X
N X
N
n xn(t) n Xn(s)
n=1 n=1
x(t ; t0 ) X (s)e;t0 s
es0 tx(t) X (s ; s0 )
x(t) 1X s
dx(t) sX (s) ; x(0; )
Z t dt 1 X (s)
;
x( )d s
0
tx(t) ; dXds(s)
x(t) cos(2f0 ) 1 X (s ; j 2f ) + X (s + j 2f )]
2 0 0
fc f
X (f )  B (f )  Y (f )
(selected radio station)
Y (f ) = X (f )
B (f )
53
54 Chapter 4. System Analysis in the Frequency Domain
Practical systems may be described and realised in the frequency domain. Since the Fourier
transform allows convolution of signals and systems to be represented as multiplications in
frequency, it is eective to analyse signals and implement systems in the frequency domain.
x(t) C y(t)
d d
f 0 0 f0 f
/2
/2
0tT
: A exp t;T
; RC ; exp ; RCt
t>T
The result is plotted in gure 4.4 for several values of RC T together with jX (f )j and jH (f )j.
T
The parameter RC is proportional to the ratio of the 3 dB frequency of the lter to the spectral
width (T ;1 ) of the pulse. When the lter bandwidth is large compared with the spectral width of
the input pulse, the input is essentially undistorted by the system. For 2f0 =T 1, the system
distorts the input signal spectrum and the output does not resemble the input.
abs[X(f)] abs[X(f)]
abs[H(f)] abs[H(f)]
Input pulse
T/RC=10
T/RC=0.5
T 2T t
we can nd the Fourier transform of x(t) in terms of the Fourier coecients Xn , namely
1
X
X (f ) = Xn
(f ; nf0):
n=;1
Then
1
X
Y (f ) = H (f )
X (f ) = Xn H (f )
(f ; nf0):
n=;1
Using the relationship H (f )
(f ; nf0) = H (nf0)
(f ; nf0), we have
1
X
Y (f ) = Xn H (nf0)
(f ; nf0 )
n=;1
Therefore
1
X
y(t) = Xn H (nf0 )e+j2f0 nt:
n=;1
Writing Xn and H (nf0 ) in terms of magnitude and phase,
1
X
y(t) = jXnjjH (nf )j ej nf0t
0
2 +
n +
H (nf0 )
n=;1
We note that the nth spectral component of the input, Xn , appears at the output with amplitude
attenuated, or amplied by the amplitude response function jH (nf0 )j.
Example 4.4.1
jH (f )j
6
f
;B2
B
2
jH (f )j
6
k
 f
;f 0 f0
Figure 4.5: Transfer function of an ideal bandpass lter.
jH (f )j
6
 f
; B
2
B
2
Either representation (Fourier series or Fourier transform) contains the same information about
x(t), and either result can be used to plot the spectra of a signal.
To obtain x(t), we can consider it as the result of convolving the ideal sampling waveform
with a pulse{type signal p(t). The signal p(t) is an energy signal of limited duration such that
p(t) = 0
jtj T2 with T Ts :
1
X 1
X
xp(t) = $(t=3)
(t ; 6n) ) Xp (f ) = F f$(t=3)g 16
f ; n6
n=;1 n=;1
B. Boashash, Signals in Linear Systems
4.6. Fourier Transforms of Periodic Signals and Sampling 61
X1 X1 n n
Xp (f ) = 3sinc2 (3f ) 61
f ; n6 = 12 sinc2
2
f;6
n=;1 n=;1
5.1.1 Sampling
To sample a continuous{time signal x(t) at a discrete number of points, t = nT , where T is the
sampling period (time between samples), and n is an integer that establishes the time position
of each sample.
To extract samples of x(t), the sampling switch closes every T seconds. This yields:
a value of x(t) when the switch is closed
0 when the switch is open.
A useful sampling process requires that x(t) can be reconstructed from the samples.
Reconstruction of x(t) from its samples
The sample signal xs(t) is written:
xs(t) = x(t)p(t)
where p(t), called the sampling function, models the action of the sampling switch (see
gure 5.2). p(t) is periodic ! it can be represented by its Fourier series:
1
X
p(t) = Cn ejn2fst (5.1)
n=;1
63
64 Chapter 5. Introduction to Discrete{time Signal Processing
Samples of x(t)
t
0 T 2T 3T 4T 5T 6T 7T
Samples of a waveform
Sampling device
Z
Cn = T1
T
2
p(t)ejn2fst dt (5.2)
;2
T
In equations 5.1 and 5.2, fs is the fundamental frequency of p(t), which is the sampling
frequency, and is given by
fs = T1 hertz
x (t) xs (t)
p(t)
t
0 T 2T 3T 4T 5T 6T 7T
Figure 5.4: The waveform p(t). Ideally, p(t) should be a periodic set of delta functions
Since xs(t) is the product of x(t) and p(t), we have (using eq.5.1)
1
X
xs(t) = Cn x(t)ejn2fs t
n=;1
The Fourier transform of xs (t) is dened by
Z1 X
1
Xs (f ) = Cnx(t)ejn2fs te;j 2ft
;1 n=;1
Interchanging the order of integration and summation yields (as per Fubini's theorem)
1
X Z1
Xs(f ) = Cn x(t)e;jn2(f ;nfs)t dt
n=;1 ;1
From the denition of the Fourier transform,
Z1
X (f ; nfs) = x(t)e;jn2(f ;nfs)t dt
;1
Thus the Fourier transform of the sampled signal can be written
1
X
Xs(f ) = Cn X (f ; nfs)
n=;1
The spectrum of the sampled continuous{time signal, x(t), is composed of the spectrum of
x(t) plus the spectrum of x(t) translated to each harmonic of the sampling frequency. Each of the
translated spectra is multiplied by a constant, given by the corresponding term in the Fourier
series expansion of p(t).
If the sampled signal is ltered by the reconstruction lter, the output of the lter is, in the
frequency domain C0 X (f ) and the time domain signal is C0 x(t).
Important: Sampling in the time domain , periodising in the frequency domain.
B. Boashash, Signals in Linear Systems
66 Chapter 5. Introduction to Discrete{time Signal Processing
X(f)
 fn 0 fn f
Xs(f)
fs =1/T, fs > 2 fn
0f f
 fs  fn n fs
H(f)
Lowpass filter
fn < fc < fs  fn
 fc 0 fc f
Y(f)
 fn 0 fn f
Figure 5.5: Exact recovery of a continuous{time signal from its samples using an ideal low{pass
lter
fs ; fh fh
or
fs 2fhHz
Sampling Theorem:
A band limited signal x(t), having no frequency components above fh hertz, is completely
specied by samples that are taken at a uniform rate greater than 2fh hertz. Then, the time
between samples is no greater than 1/2fh seconds.
The frequency 2f is called the Nyquist rate.
In a practical situation, a signal cannot be strictly band limited, because it is time limited.
However, in all practical signals, there is some frequency beyond which the energy is negligible.
! denition of bandwidth.
Another source of error is the nonexistence of ideal reconstruction lters. Practically, we
need to sample at a higher frequency than the Nyquist rate.
p(t)
The system G in Figure 5.6 represents a system that converts an impulse train into a discrete
time sequence. This is often referred to as a zero{order hold (for more deatils, see 5] p. 520).
The modulation signal p(t) is a periodic impulse train,
1
X
p(t) =
(t ; nT )
n=;1
where
(t) is called the Dirac delta function. Consequently, we have
1
X 1
X 1
X
xs (t) = xc(t)
p(t) = xc(t)
(t ; nT ) = xc(nT )
(t ; nT ) = x(n)
(t ; nT ):
n=;1 n=;1 n=;1
The Fourier transform of xs (t)
Xs (f ), is obtained by convolving Xc (f ) and P (f ). The Fourier
transform of a periodic impulse train is a periodic impulse train, i.e.,
1 1
X k 1 1
X
P (f ) = T k=;1
(f ; ) = T
(f ; kf ):
T k=;1 s
Since
Xs(f ) = Xc(f ) P (f )
it follows that
1
X 1
X
Xs(f ) = T1 Xc j f ; Tk = T1 Xc (jf ; kjfs ) :
k=;1 k=;1
Figure 5.7 depicts the frequency domain representation of impulse train sampling. Figure
5.7(a) represents a bandlimited Fourier transform where the highest nonzero frequency com
ponent in Xc (f ) is at fB . Figure 5.7(b) shows the periodic impulse train P (f ), Figure 5.7(c)
shows Xs (f ), which is the result of convolving Xc (f ) with P (f ). From Figure 5.7(c) it is evident
that when
fs ; fB > fB
or fs > 2fB
the replicas of Xc (f ) do not overlap, and therefore xc (t) can be recovered from xs(t) with an
ideal low pass lter. However, in Figure 5.7(d) the replicas of Xc (f ) overlap, and as a result
the original signal cannot be recovered by ideal low pass ltering. The reconstructed signal is
related to the original continuoustime signal through a distortion referred to as aliasing.
If xc (t) is a bandlimited signal with no components above fB Hz, then the sampling fre
quency has to be equal to or greater than 2fs Hz. This is known as the Nyquist criterion and
2fs is known as the Nyquist frequency.
(a) X c (f a )
1
f f fa
B B
(b)
P(fa )
1
T
2 f s fs f s 2 f s fa
(c)
1 X s (f a )
a
T
fB fs fa
(d)
X s (f a )
1
T
fs fa
Figure 5.7: Eect in the frequency domain of sampling in the time domain: (a) Spectrum of
the continuoustime signal, (b) Spectrum of the sampling function, (c) Spectrum of the sampled
signal with fs > 2fB , (d) Spectrum of the sampled signal with fc < 2fB .
bandwidth
B fh
B fs ; fh
Reconstruction ltering:
If x(n) is the input to an ideal low pass lter with frequency response Hr (f ) and impulse
response hr (t), then the output of the lter will be
1
X
xr (t) = x(n)h(t ; nT ) :
n=;1
where T is the sampling interval. The reconstruction lter commonly has a gain of T and a
cuto frequency of fc = fs =2 = 1=2T . This choice is appropriate for any relationship between
fs and fB that avoids aliasing, i.e., so long as fs > 2fB .
1
X Ideal Reconstruction
Filter
x(nT )
(t ; nT ) ! T
j f j
H (f ) = 0
s elsewhere 0 :5fs ! x(t)
n=;1
Reconstruction ltering
The impulse response of such a lter is given by:
h (t) = sin (t=T ) :
r t=T
xr (t) =
1
X t ; nT )=T ]
x(n) sin(t(;
n=;1 nT )=T
with xc (nT ) = x(n).
B. Boashash, Signals in Linear Systems
5.1. Analog to Digital Conversion 71
sample of x (t)
x (t)
Example 5.1.1
x(t) = 6 cos 2(50)t
We consider this signal samples at 70 and 140Hz. The Nyquist Frequency is 100Hz.
X (f ) = 3
(f ; 50) + 3
(f + 50)
Spectrum of the sampled signal:
1
X
X (f ) = 3fs
(f ; 50 ; nfs) +
(f + 50 + nfs)]
n=;1
The assumed reconstruction lter is an ideal low{pass lter with a bandwidth of 0.5fs and
an amplitude response of T .
Example 5.1.2 We consider a nonperiodic signal for x(t) so that X (f ) has a real, continuous
spectrum.
The highest frequency is 50Hz ! the minimum acceptable sampling frequency is 100Hz.
Again, we assume sampling frequencies of 70 and 140Hz.
For a sampling frequency of 70Hz aliasing occurs. For fs = 140 Hz, there is no spectral
overlap since fs > 100Hz.
The reconstruction lter is an ideal low{pass lter with an amplitude response of T and a
bandwidth of 0.5fs .
The output spectrum of the reconstruction lter is shown for f = 70Hz and fs = 140Hz. The
impact of aliasing is clear.
X (f)
Spectrum of x (t)
50 0 50 f
Reconstruction filter
Xs(f)
transfer function
fs = 70 Hz
0 50 90 190 f
Xs(f)H (f) Xs(f)H (f)
20 0 20 f 50 0 50 f
Output of reconstruction Output of reconstruction
filter with sampling frequency filter with sampling frequency
equal to 70 Hz. equal to 140 Hz.
X (f)
Spectrum of x (t)
50 0 50 f
35 0 35 f 50 0 50 f
7 111 S
6 110
5 101
4 100
3 011
2 010
1 001
0 000
t
0 T 2T 3T 4T
q = 2n
In the above example, we have 8 levels, and each level is uniquely specied by a three{bit
word.
Transfer t characteristic:
X
xQ(t) = ms(s (x ; ms)
m
xQ
s
with nullarea
3s/2 s/2
s
X
xQ(t) = (ms + 2s )(s (x ; ms ; 2s )
m
xQ
without nullarea
s/2
x (t)
s s
s/2
nx(x) = x ; xQ
Center of
quantization
S
level
t
t 1 0 t1
a) Quantizing
(t)
S/2
t 1
t
0 t1
S/2
b) Quantizing error
2
(t)
2
S /4
t
t 1 0 t1
F F
density of the noise n(t) is in general constant in the bandwidth ; 2s
2s , where Fs = T1 .
s
We dene D = A=D converter dynamic range
S = 2Dn = D
2;n
which yields
E = D12
2;2n
2
In calculating the signal power, the assumption is usually made that the signal power at the
output of the A=D converter is equal to the signal power at the input of the A=D converter.
Example 5.1.3 We want to sample and quantise the signal
SNR = AE=2 = 32
22n
2
SNR(dB) = 10
log10 SNR = 10
log10 23 + 20n
log10 2
SNR = 1:176 + 6:02n
Thus, the signal{to{noise ratio at the output of the A=D converter increases by approximately
6 dB for each added bit of word length.
! importance of using a large word length.
We obtain a similar result with any test signal: the bias, here 1:76, will change with the
shape of the test signal, but in general, the SNR increases 6:02dB for each increment in the
word length.
50
25
n
4 8 12
6H (f ) Processed
ya (t) Signal
 

fB 0 fB f
s ;fs=2
where fs is the sampling frequency. For fs = 1 Hz which will be assumed in the sequel unless
other specied, we have Z 1=2
x(n) = X (f )ej2fndf :
;1=2
X (f ) is called the discrete time Fourier transform of x(n). Conversely, x(n) is the inverse Fourier
transform of X (f ). Sometimes it is useful to consider the equation for X (f ) as an operator that
transforms the sequence into a function, and we will refer to the Fourier transform operator as
1
X
Ffx(n)g = x(n)e;j 2fn = X (f ):
n=;1
Consequently x(n) will be obtained as
x(n) = F ;1 fX (f )g:
X (f ) is in general complex and is a function of a continuous variable f . X (f ) is always periodic
with period of 1, i.e., X (f ) = X (f + r), for all r 2 Z because it is the FT of a sampled signal.
This may be veried as follows:
1
X
X (f + r) = x(k)e;j 2fk e;j 2r = X (f )
k=;1
using the fact that ej 2r = 1 for all integer r.
We dene a sequence x(n) to be an eigenfunction of a system T if Tx(n)] = kx(n), for
any scalar k. Clearly, ej 2fn is the eigenfunction of any LSI system, and the scaling factor k
is H (f ), the Fourier transform of the system impulse response h(n). This concept is shown in
Figure 5.18, where the scaling factor H (ej! ) is called the transfer function of the system or the
frequency response function.
B. Boashash, Signals in Linear Systems
80 Chapter 5. Introduction to Discrete{time Signal Processing
x(n) = ej 2fn  LSI  H (f )
ej2fn
h(n)
In other words the error jX (f ) ; XM (f )j may not approach zero at each f as M ! 1, but the
total energy in the error does.
Example 5.2.1 Consider an ideal low pass lter with cut o frequency fc. Its transfer function
is given by
H (f ) = 1
jf j < fc
0
fc < jf j 1=2 :
The impulse response of the lter is given by
Z fc
h(n) = ej 2fndf
;fc
= j 21n ej 2fc n ; e;j 2fc n ]
h(n) = sin(2nfc n) ;1<n<1
B. Boashash, Signals in Linear Systems
5.2. Digital Processing of ContinuousTime Signals Using the DTFT 81
This is the impulse response of a non casual lter because h(n) 6= 0 for n < 0. Also, h(n) is not
absolutely summable since
1 sin(2f n)
X c e;j 2fn
n=;1 n
does not converge uniformly for all values of f . However, because these functions have such an
important role in digital ltering, it is accepted as a valid Fourier transform pair.
Example 5.2.2 Consider the sequence x(n) = 1 for all n. This sequence is neither absolutely
nor square summable. However it is useful to assign the periodic impulse train
1
X
X (f ) =
(f + k)
k=;1
where
(f ) is Dirac's impulse function to be its Fourier transform. One can check that indeed
this result is valid if X (f ) is substituted in
Z 1=2
x(n) = X (f )ej 2fn df:
;1=2
Convolution: 1
X
x(n) ? h(n) = x(k)h(n ; k) ! X (f )
H (f )
k=;1
B. Boashash, Signals in Linear Systems
82 Chapter 5. Introduction to Discrete{time Signal Processing
which results from the fact that
1 X
X
Ffx(n) ? h(n)g = x(k)h(n ; k)e;j2fn
n=;1 k
X X
= x(k) h(n ; k)e;j2fn
k n
X X
= x(k) h(u)e;j2f (u+k)
u
X X
= x(k)e j2fk
h(u)e;j2fu
;
k u
= X (f )H (f )
Parseval's theorem :
1
X Z
x(n)
y(n) = 1=2X (f )
Y (f ) df
n=;1 ;1=2
which simplies to
1
X Z
jx(n)j 2
=
;1=2
1=2jX (f )j2 df
n=;1
This result holds because
1
X 1 Z
X Z
x(n)y(n) = 1=2X (f )ej 2fn df 1=2Y (ej 2
) e;j 2
nd
n=;1 n=;1 ;1=2 ;1=2
Z 1=2 Z 1=2 " 1 #
X
= X (f )Y ( )
( ; f + k) dfd
;1=2 ;1=2 k=;1
Z
= 1=2X (f )Y (f ) df
;1 =2
x(;n) X (;f )
x(n) real and even X (f ) real and even
x(n) real and odd X (f ) pure imaginary and odd
Parseval's Theorem
1
X Z 1=2
jx(n)j = ; = jX (f )j d!
2 2
n ;1
= 1 2
Z =
X1
x(n)
y(n) = X (f )
Y (f )df
1 2
n=;1 ;= 1 2
1
X
xs(t) = x(t)
(t ; nT )
n=;1
Taking the Laplace Transform
Z 1X
1
Xs(s) = x(nT )
(t ; nT )e;st dt
0 n=0
Interchanging integration and summation
1
X Z1
Xs(s) = x(nT )
(t ; nT )e;st dt
n=0 0
s = + j! ! z = eT ej!T
Stable functions of the Laplace variable s must not have any poles in the right{half s{plane
or on the jw{axis. We can then derive a condition for the z {transform.
The magnitude of z is given by: jz j = eT
The right{half s{plane, > 0, corresponds to jzj > 1.
The left{half s{plane, < 0, corresponds to jzj < 1.
Im
z=e j
Re
Example 5.3.2
The unit pulse sequence is dened by the sample values
x(nT ) = 1
n = 0 4
=
(n)
0
n 6= 0
Its z {Transform is then:
X (z ) = 1 + 0
z;1 + 0
z;2 +
Thus
X (z ) = 1
B. Boashash, Signals in Linear Systems
88 Chapter 5. Introduction to Discrete{time Signal Processing
1
3 2 1 0 1 2 3 n
Important Result: The unit pulse sequence plays the same role in discrete{time systems that
the unit impulse function plays in analog systems. For example, the sequence p(n) in Figure
5.21 can be expressed as
p(n) = a;1
(n + 1) + a1
(n ; 1) + a3
(n ; 3) :
Any sequence x(n) can be represented as a linear combination of delayed unit sample sequences,
6p(n)
s
a;1 a1 s a3 s
n
;1 1 3
Figure 5.21: The sequence p(n).
1
X
x(n) = x(k)
(n ; k)
k=;1
as shown in Figure 5.22.
Example 5.3.3
The unit step sample sequence is dened by the sample values
X (nT ) = 1
n0
= 0
otherwise
Its z {Transform is then:
1
X
X (z ) = z;n
n=0
We know that for jxj < 1
1
X
xn = 1 ;1 x
n=0
B. Boashash, Signals in Linear Systems
5.3. The z{Transform 89
6x(n)
t
t t
t t t t t t t t t

1 n
t t t t
6 x(k)
(n ; k )
t t t t t t t t t

n k
3 2 1 0 1 2 3 4 5 6 7 n
Thus,
X1
X (z ) = z ;n = 1 ;1z ;1
jzj > 1
n=0
The unit step sample sequence is often denoted u(n). The sequence u(n) is related to
(n)
by
X
n
u(n) =
(k )
k=;1
or alternatively
1
X
u(n) =
(n ; k ) :
k=0
Conversely, the impulse sequence can be expressed as the rst backward dierence of the unit
step sequence, i.e.,
X
n ;1
nX
(n) = u(n) ; u(n ; 1) =
(k) ;
(k )
k=;1 k=;1
x(nT ) = e;nT
> 0
n 0
x(n)
X (z) = 1 ; e;1T z ;1
jz j > e;T
jzj > e;T denes;the
T
ROC: region of convergence.
Given and T , e is a constant k.
Then
X (z) = 1 ; k1
z;1
jzj > k
Example 5.3.5 x(n) = an u(n).
The z transform of x(n) is given by
1
X 1
X 1 ;
X
X (z) = an u(n)z ;n = anz ;n = az;1 n :
n=;1 n=0 n=0
The series converges if jz j > jaj. Then,
X (z ) = 1 ; 1az ;1 = z ;z a
jzj > jaj :
B. Boashash, Signals in Linear Systems
5.3. The z{Transform 91
The Fourier transform of x(n) converges only if
1
X
janu(n)j < 1
n=;1
which requires that
1
X
jajn < 1
n=0
which is the case for jaj < 1. For jaj = 1, we have x(n) = u(n) and X (z ) = z;z 1 , jz j > 1, as
discussed in the previous example. A typical ROC is shown in Figure 5.25. Outside the ROC,
i.e., Rx+ < jz j < Rx; , X (z ) does not converge and does not exist.
Im
R x+
Re
R x
X (z ) = 1 ;kz;1 + G(z)
where all the poles of G(z ) lie inside the unit circle ! x(1) = k
If X (z ) is in this form, the sample values, x(nT ), can be determined by inspection. This
form is easily obtained by long division when X (z ) is expressed as a ratio of polynomials in z .
Example 5.3.6
X (z ) = (z ; 1)(zz ; 0:2)
2
= 1 + 1:21z ;;11+:2z1:24z+;02:+
;1 2z ;2
1:248z ;3
! X (z) = 1 + 1:2z;1 + 1:24z;2 + 1:248z;3 +
we write:
X (z ) = z = k1 + k2
z (z ; 1)(z ; 0:2) z ; 1 z ; 0:2
where:
Thus,
25z ; 0:25z = 1:25 ; 0:25
X (z ) = 1z:; 1 z ; 0:2 1 ; z ;1 1 ; 0:2z ;1
We nd the inverse transform of X (z ) by using the list of transforms1 . The rst term has
k = 1 ! it's a unit step sample sequence. The second term has k = 0:2 which is a sampled
exponential.
Thus x(nT ) = 1:25 ; 0:25(0:2)n
n 0
The values of x(nT ) are: x(0) = 1:25 ; 0:25(0:2)0 = 1
x(T ) = 1:25 ; 0:25(0:2)1 = 1:2
x(2T ) = 1:25 ; 0:25(0:2)2 = 1:24
x(3T ) = 1:25 ; 0:25(0:2)3 = 1:248
Delay Operator Delay Operation is very important in digital systems. If the time sequence
fx(nT )g is delayed by k sample periods, the eect in the z{domain is to multiply X (z) by z;k .
Proof 5.3.3 X (z) = P1n=0 x(nT )z;n
The z {Transform of the sequence fx(nT ; kT )g is:
1
X
x(nT ; kT )z;n
n=0
P1
Let m = n ; k ! m=;k x(mT )z ;m;k
As x(mT ) is assumed zero for m < 0, we have:
1
X
z ;k x(mT )z;m = z ;k X (z)
m=;k
1 Gabel and Roberts, Signals in linear systems, 2nd edition, p212 and 213
dy + a y(t) = b x(t)
dt
x(t) ! System ! y(t)
The output y(t) can be expressed as:
Zt Zt
y(t) = b x()d ; a y()d
;1 ;1
This equation tells us thatRthe present value of the system output, y(t), is a function of all
t
R t of the input, b ;1 x()d, and is also a function of all past values of the system
previous values
output, b ;1 y()d.
5.4.1 Dierence Equations
A linear digital system operates in the same way: the present output, y(nT ), is computed using
the present input x(nT ), past inputs x(nT ; kT ), and past system outputs y(nT ; kT ).
The general dierence equation for this processor is:
y(nT ) = Lox(nT ) + L1 x(nT ; T ) + L2x(nT ; 2T ) +
;kmy(nT ; mT )
The processor generates an output by weighting the present input, the past n inputs, and
the past m outputs. The processor is illustrated in the Fig. 5.26.
The analysis problem is usually a problem of determining the system output given the system
input and a specication of the system. The most convenient method is to specify the coecients
of the dierence equation.
Another problem is the synthesis of digital lters in which the problem is to determine the
coecients of the dierence equation in order to perform some specied task.
The rst step is to solve the dierence equation. We can write:
y(nT ) + k1 y(nT ; T ) + k2 y(nT ; 2T ) +
+ km y(nT ; mT )
= Lo x(nT ) + L1 x(nT ; T ) + L2 x(nT ; 2T ) +
+ Lr x(nT ; rT )
B. Boashash, Signals in Linear Systems
5.4. Digital Systems 95
Now, we z {transform both sides using the time delay operator.
Z x(nT ; kT )] = z;k X (z)
Y (z ) + k z ; Y (z) + k z; Y (z ) +
+ km z;mY (z)
1
1
2
2
= Lo X (z ) + L z ; X (z ) + L z ; X (z ) +
+ Lr z ;r X (z )
1
1
2
2
That is:
Y (z )(1 + k1 z ;1 + k2 z;2 +
+ km z;m )
= X (z )(Lo + L1 z ;1 + L2 z ;2 +
+ Lr z ;r )
It follows:
H (z) = L1o++kLz1;z 1 ++kLz2;z2 ++
++k Lzr;z m
;1 ;2 ;r
1 2 m
H (z) = XY ((zz)) is called the pulse transfer function.
Relation input{output in the z {domain: Y (z ) = H (z )X (z ).
We have seen that the z {Transform of a unit pulse is unity. Thus, if a digital system has
a unit pulse input, the system output is: Y (z ) = H (z ). The inverse z {Transform yields:
y(nT ) = h(nT ).
f h(nT ) g is called the unit pulse response. In digital systems, it is the equivalent of the
impulse response in analog systems.
These systems are shift invariant since a time shift of the input time shifts the output without
changing the shape of the output.
If h(nT ) is zero for n < 0, the system is called a causal system.
5.4.2 Discrete Convolution
Formulation
Let us determine the time{domain equivalent of
Y (z) = H (z )X (z)
By denition, we have:
and
where x(iT ) = ith value of the system input and h(iT ) = ith value of the unit pulse response.
The three preceding equations yield:
Y (z ) = x (0) + x(T )z;1 + x(2T )z ;2 +
h(0) + h(T )z;1 + h(2T )z;2 +
] z;1 +
+ x(0)h(2T ) + x(T )h(T ) + x(2T )h(0) +
] z ;2
We note that: if a term multiplies z ;n , the sum of the arguments of x and h is nT .
Thus, the general term is:
+ y(nT )z;n +
Denition
Example 5.4.2
See Fig. 5.27.
We shall use discrete convolution to compute the output of a discrete processor whose input
x(nT ) and unit pulse response h(nT ) are shown in Fig. 5.27.
X
n
y(nT ) = x(mT )h(nT ; mT )
m=0
Example 5.4.3 To compute the relationship between the input and the output with a transversal
lter, h(;k) is rst plotted against k h(;k) is simply h(k) reected or \ipped" around k = 0,
as shown in Figure 5.27. Replacing k by k ; n, where n is a xed integer,
P leads to a shift in the
origin of the sequence h(;k) to k = n. y(n) is then obtained by the x(k)h(n ; k). See Figure
5.27.
t t t
t
3
1
2
 1 t 2

k k
6y(n) 7 t t
6h(;k) 6
4 t
2 t t
1 1 t
 
k n
For n suciently large, the terms h(nT ) of the unit pulse response are negligible ! the
steady{state response has been reached. The term in brackets is then a complex function
depending only on the input frequency !.
B. Boashash, Signals in Linear Systems
98 Chapter 5. Introduction to Discrete{time Signal Processing
We denote it
X
n
G(!) = h(mT )e;j!mT
m=0
! y(nT ) = G(!)Aej (!nT +)
We know that:
1
X
H (z) = h(mT )z;m
m=0
Comparing the two equations G(!) and H (z ), it follows:
;
G(!) = H ej!t
The steady{state response is easily obtained from the pulse transfer function by replacing z
in the pulse transfer function with ej!t .
Summary:
previous result
importance of the ROC.
99