ECE 2321 SIGNALS AND SYSTEMS

Lecture notes for BSc Electrical and Telecommunication Engineering

First Edition
Martin Wafula
Multimedia University of Kenya
2
 PREFACE

These notes for ECE 2321: Signals and Systems are designed to bridge your existing
knowledge of differential equations (from MTE 2220) with the fundamental tools used to
analyze and understand both continuous and discrete signals. Here, you will learn how to
apply Fourier and Laplace transforms to represent and examine signals and how to use these
techniques to study and design systems that handle these signals.
We will begin by classifying signals and systems, exploring the time and frequency domains,
and identifying the characteristics that distinguish linear, time-invariant systems from others.
In moving from the basics to more advanced methods, you will see how mathematical concepts
translate into practical engineering approaches—such as understanding how a given input
affects a system or how to synthesize a system that meets certain performance requirements.
To reinforce theory, laboratory exercises and MATLAB and Python simulations will pro-
vide hands-on experience. These activities will let you visualize signal properties, practice
sampling procedures, and investigate the effects of system parameters, ultimately giving you
a well-rounded skill set that combines conceptual understanding with practical know-how.

3
4
 CONTENTS

Preface 3
Contents 5
List of Figures 9
List of Tables 11

I First part: Continous and Discrete Signals 13

1 Introduction to Signals and Systems 15
1 Signals and Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2 Outline of This Course . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2 Properties of Signals and Systems 19
1 Signal Energy and Power . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2 Transformations of Signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3 Periodic, Even and Odd Signals . . . . . . . . . . . . . . . . . . . . . . . . . 22
4 Exponential and Sinusoidal Signals . . . . . . . . . . . . . . . . . . . . . . . 22
4.1 Continuous-Time Complex Exponential Signals . . . . . . . . . . . . 22
4.2 Discrete-Time Complex Exponential Signals . . . . . . . . . . . . . . 23
5 Impulse and Step Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
5.1 Discrete-Time . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
5.2 Continuous-Time . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
6 Properties of Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
6.1 Interconnections of Systems . . . . . . . . . . . . . . . . . . . . . . . 28
6.2 Properties of Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3 Analysis of Linear Time-Invariant Systems 35
1 Discrete-Time LTI Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2 Continuous-Time LTI Systems . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3 Properties of Linear Time-Invariant Systems . . . . . . . . . . . . . . . . . . 38
3.1 The Commutative Property . . . . . . . . . . . . . . . . . . . . . . . 38
3.2 The Distributive Property . . . . . . . . . . . . . . . . . . . . . . . . 39
3.3 The Associative Property . . . . . . . . . . . . . . . . . . . . . . . . 40
3.4 Memoryless LTI Systems . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.5 Invertibility of LTI Systems . . . . . . . . . . . . . . . . . . . . . . . 41
3.6 Causality of LTI Systems . . . . . . . . . . . . . . . . . . . . . . . . . 42
3.7 Stability of LTI Systems . . . . . . . . . . . . . . . . . . . . . . . . . 43

5
6 CONTENTS

3.8 Step Response of LTI Systems . . . . . . . . . . . . . . . . . . . . . . 44

4 Differential and Difference Equation Models for Causal LTI Systems . . . . . 44
5 Differential and Difference Equation Models for Causal LTI Systems . . . . . 45
5.1 Linear Constant-Coefficient Differential Equations . . . . . . . . . . . 45
6 Differential and Difference Equation Models for Causal LTI Systems . . . . . 47
6.1 Linear Constant Coefficient Difference Equations . . . . . . . . . . . 47
7 Block Diagram Representations of Linear Differential and Difference Equations 49
4 Fourier Series Representation of Periodic Signals 51
1 Applying Complex Exponentials to LTI Systems . . . . . . . . . . . . . . . . 52
2 Fourier Series Representation of ContinuousTime Periodic Signals . . . . . . 54
3 Fourier Series Representation of Continuous-Time Periodic Signals . . . . . . 55
4 Calculating the Fourier Series Coefficients . . . . . . . . . . . . . . . . . . . 56
4.1 A Vector Analogy for the Fourier Series . . . . . . . . . . . . . . . . . 58
5 Properties of Continuous-Time Fourier Series . . . . . . . . . . . . . . . . . . 62
5.1 Linearity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
5.2 Time Shifting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
5.3 Time Reversal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
5.4 Time Scaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
5.5 Multiplication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
5.6 Parseval’s Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
6 Fourier Series for Discrete-Time Periodic Signals . . . . . . . . . . . . . . . . 66
6.1 Finding the Discrete-Time Fourier Series Coefficients . . . . . . . . . 67
5 The Continuous-Time Fourier Transform 73
1 The Fourier Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
1.1 Existence of Fourier Transform . . . . . . . . . . . . . . . . . . . . . 76
2 Fourier Transform of Periodic Signals . . . . . . . . . . . . . . . . . . . . . . 77
3 Properties of the Continuous-Time Fourier Transform . . . . . . . . . . . . . 78
3.1 Linearity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
3.2 Time-Shifting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
3.3 Conjugation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
3.4 Differentiation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
3.5 Time and Frequency Scaling . . . . . . . . . . . . . . . . . . . . . . . 81
3.6 Duality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
3.7 Parseval’s Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
3.8 Convolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
3.9 Multiplication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
6 The Discrete-Time Fourier Transform 87
1 The Discrete-Time Fourier Transform . . . . . . . . . . . . . . . . . . . . . . 88
2 The Fourier Transform of Discrete-Time Periodic Signals . . . . . . . . . . . 90
3 Properties of the Discrete-Time Fourier Transform . . . . . . . . . . . . . . . 91
3.1 Periodicity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
3.2 Linearity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
3.3 Time and Frequency Shifting . . . . . . . . . . . . . . . . . . . . . . 92
3.4 First Order Differences . . . . . . . . . . . . . . . . . . . . . . . . . . 92
CONTENTS 7

3.5 Conjugation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
3.6 Time-Reversal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
3.7 Differentiation in Frequency . . . . . . . . . . . . . . . . . . . . . . . 94
3.8 Parseval’s Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
3.9 Convolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
3.10 Multiplication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
7 Sampling 99
1 The Sampling Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
2 Reconstruction of a Signal From Its Samples . . . . . . . . . . . . . . . . . . 101
2.1 Zero-Order Hold . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
2.2 First-Order Hold . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
3 Undersampling and Aliasing . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
4 Discrete-Time Processing of Continuous-Time Signals . . . . . . . . . . . . . 103
8 The Laplace Transform 107
1 The Laplace Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
2 The Region of Convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
3 The Inverse Laplace Transform . . . . . . . . . . . . . . . . . . . . . . . . . 112
3.1 Partial Fraction Expansion . . . . . . . . . . . . . . . . . . . . . . . . 113
4 Some Properties of the Laplace Transform . . . . . . . . . . . . . . . . . . . 115
4.1 Convolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
4.2 Differentiation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
4.3 Integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
5 Finding the Output of an LTI System via Laplace Transforms . . . . . . . . 116
6 Finding the Impulse Response of a Differential Equation via Laplace Transforms117

II Second part: Graphs Signals 119

1 Graphs Signals and Systems 121
1 Graph Terminology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
2 Signals Defined on Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
3 Graph Convolutional Filters . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
3.1 Convolutional Filters . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
3.2 Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
4 Graph Fourier Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
5 Frequency Response . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131

III Third part: An Introduction to Multidimensionsal Signals

and Systems 133
1 Introduction to Multidimensional Signals and Systems 135
1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
Bibliography 137
8 CONTENTS
 LIST OF FIGURES

1.1 An abstract representation of a system. . . . . . . . . . . . . . . . . . . . . . 15

1.2 Block Diagram of a feedback control system. . . . . . . . . . . . . . . . . . . 17

3.1 A series interconnection of systems. . . . . . . . . . . . . . . . . . . . . . . . 40

3.2 Caption . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3.3 An equivalent series representation of the interconnection shown in Fig. 3.1 . 41
3.4 A system in series with its inverse. . . . . . . . . . . . . . . . . . . . . . . . 41
3.5 Integrator and Delay blocks. . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.6 Chained integrator and delay blocks. . . . . . . . . . . . . . . . . . . . . . . 50

5.1 A series interconnection of systems. . . . . . . . . . . . . . . . . . . . . . . . 84

7.1 The frequency spectrum of the signal x(t) and the signal xp (t). . . . . . . . . 101
7.2 The impulse response of a zero-order-hold (left) and a first-orderhold (right). 102
7.3 Sampling x(t) = cos(t) at a frequency of ωs = 4. . . . . . . . . . . . . . . . . 103
7.4 The frequency response of a bandlimited differentiator. . . . . . . . . . . . . 105

1.1 The graph convolution as a shift register. Highlighted are the nodes that reach
node 1 on each consecutive shift; that is, the nodes j whose signal value xj
contributes to S x i . The resulting summary of each communication Sk x is
k


correspondingly weighted by a filter parameter hk . For each k, the parameter

hk is the same for all nodes. . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
1.2 Discrete-time periodic signals as graph signals over a directed cycle graph.
Each node Vc = {1, . . . , 6} is a time instant with adjacencies captured in the
matrix Ac . The temporal signal forms the graph signal x = [x1 , . . . , x6 ]⊤ and
the shift Ac x acts as a delay operation that moves the signal to the next time
instant node. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127

9
10 List of Figures
LIST OF TABLES

11
12 List of Tables
 Part I

First part: Continous and Discrete

Signals

13
 CHAPTER 1
INTRODUCTION TO SIGNALS AND
SYSTEMS

1 Signals and Systems

Loosely speaking, signals represent information or data about some phenomenon of interest.
This is a very broad definition, and accordingly, signals can be found in every aspect of the
world around us.
For the purposes of this course, a system is an abstract object that accepts input signals
and produces output signals in response.
Examples of systems and associated signals:

• Electrical circuits: voltages, currents, temperature,...

• Mechanical systems: speeds, displacement, pressure, temperature, volume, ...

• Chemical and biological systems: concentrations of cells and reactants, neuronal activ-
ity, cardiac signals, ...

• Environmental systems: chemical composition of atmosphere, wind patterns, surface

and atmospheric temperatures, pollution levels, ...

• Economic systems: stock prices, unemployment rate, tax rate, interest rate, GDP, ...

• Social systems: opinions, gossip, online sentiment, political polls,...

• Audio/visual systems: music, speech recordings, images, video, ...

Figure 1.1: An abstract representation of a system.

15
16 Introduction to Signals and Systems

• Computer systems: Internet traffic, user input, ...

From a mathematical perspective, signals can be regarded as functions of one or more

independent variables. For example, the voltage across a capacitor in an electrical circuit
is a function of time. A static monochromatic image can be viewed as a function of two
variables: an x-coordinate and a y-coordinate, where the value of the function indicates the
brightness of the pixel at that (x, y) coordinate. A video is a sequence of images, and thus
can be viewed as a function of three variables: an x-coordinate, a y-coordinate and a time
instant. Chemical concentrations in the earth’s atmosphere can also be viewed as functions
of space and time.
In this course, we will primarily be focusing on signals that are functions of a single
independent variable (typically taken to be time). Based on the examples above, we see that
this class of signals can be further decomposed into two subclasses:

• A continuous-time signal is a function of the form f (t), where t ranges over all real
numbers (i.e., t ∈ R ).

• A discrete-time signal is a function of the form f [n], where n takes on only a discrete
set of values (e.g., n ∈ Z ).

Note that we use square brackets to denote discrete-time signals, and round brackets to
denote continuous-time signals. Examples of continuous-time signals often include physical
quantities, such as electrical currents, atmospheric concentrations and phenomena, vehicle
movements, etc. Examples of discrete time signals include the closing prices of stocks at the
end of each day, population demographics as measured by census studies, and the sequence of
frames in a digital video. One can obtain discrete-time signals by sampling continuous-time
signals (i.e., by selecting only the values of the continuous-time signal at certain intervals).
Just as with signals, we can consider continuous-time systems and discrete time systems.
Examples of the former include atmospheric, physical, electrical and biological systems, where
the quantities of interest change continuously over time. Examples of discrete-time systems
include communication and computing systems, where transmissions or operations are per-
formed in scheduled time slots. With the advent of ubiquitous sensors and computing tech-
nology, the last few decades have seen a move towards hybrid systems consisting of both
continuous-time and discrete-time subsystems - for example, digital controllers and actu-
ators interacting with physical processes and infrastructure. We will not delve into such
hybrid systems in this course, but will instead focus on systems that are entirely either in
the continuous-time or discrete-time domain.
The term dynamical system loosely refers to any system that has an internal state and
some dynamics (i.e., a rule specifying how the state evolves in time).
This description applies to a very large class of systems, including individual vehicles,
biological, economic and social systems, industrial manufacturing plants, electrical power
grid, the state of a computer system, etc. The presence of dynamics implies that the behavior
of the system cannot be entirely arbitrary; the temporal behavior of the system’s state and
outputs can be predicted to some extent by an appropriate model of the system.
Example : 1
Consider a simple model of a car in motion. Let the speed of the car at any time t be
Introduction to Signals and Systems 17

Figure 1.2: Block Diagram of a feedback control system.

given by v(t). One of the inputs to the system is the acceleration a(t), applied by the
throttle. From basic physics, the evolution of the speed is given by
dv
= a(t)
dt
The quantity v(t) is the state of the system, and equation (1.1) specifies the dynamics.
There is a speedometer on the car, which is a sensor that measures the speed. The value
provided by the sensor is denoted by s(t) = v(t), and this is taken to be the output of the
system. ■
Much of scientific and engineering endeavor relies on gathering, manipulating and un-
derstanding signals and systems across various domains. For example, in communication
systems, the signal represents voice or data that must be transmitted from one location to
another. These information signals are often corrupted en route by other noise signals, and
thus the received signal must be processed in order to recover the original transmission. Sim-
ilarly, social, physical and economic signals are of great value in trying to predict the current
and future state of the underlying systems. The field of signal processing studies how to
take given signals and extract desirable features from them, often via the design of systems
known as filters. The field of control systems focuses on designing certain systems (known
as controllers) that measure the signals coming from a given system and apply other input
signals in order to make the given system behave in an desirable manner. Typically, this is
done via a feedback loop of the form
Example : Inverted Pendulum
Suppose we try to balance a stick vertically in the palm of our hand. The sensor, controller
and actuator in this example are our eyes, our brain, and our hand, respectively, which
communicate using signals of various forms. This is an example of a feedback control
system. ■

2 Outline of This Course

Since the concepts of signals and systems are prevalent across a wide variety of domains,
we will not attempt to discuss each specific application in this course. Instead, we will deal
with the underlying mathematical theory, analysis, and design of signals and systems. In
this sense, it will be more mathematical than other engineering courses, but will be different
from other math courses in that it will pull together various branches of mathematics for a
particular purpose (i.e., to understand the nature of signals and systems).
The main components of this course will be as follows.
18 Introduction to Signals and Systems

• Signal and systems classifications: develop terminology and identify useful properties
of signals and systems

• Time domain analysis of LTI systems: understand how the output of linear time-
invariant systems is related to the input

• Frequency domain analysis techniques and signal transformations (Fourier, Laplace,

z-transforms): study methods to study signals and systems from a frequency domain
perspective, gaining new ways to understand their behavior

• Sampling and Quantization: study ways to convert continuous-time signals into discrete-
time signals, along with associated challenges

The material in this course will lay the foundations for future courses in signal processing
(ECE438, 445), Digital Communications (ECE 2414), Digital Filters Design (ECE 2413),
control theory (ECE 2415, EEE 2422) and Digital Image Processing (ECE 2528).
 CHAPTER 2
PROPERTIES OF SIGNALS AND SYSTEMS

We will now identify certain useful properties

1 Signal Energy and Power . . . 20
and classes of signals and systems. Recall
2 Transformations of Signals . . 21
that a continuous-time signal is denoted by
3 Periodic, Even and Odd Signals 22
f (t) (i.e., a function of the real-valued vari-
4 Exponential and Sinusoidal
able t ) and a discrete-time signal is denoted
Signals . . . . . . . . . . . . . 22
by f [n] (i.e., a function of the integer-valued
5 Impulse and Step Functions . 26
variable n ). When drawing discrete-time
6 Properties of Systems . . . . . 28
signals, we will use a sequence of dots to in-
dicate the discrete nature of the time vari-
able.

19
20 Properties of Signals and Systems

1 Signal Energy and Power

Suppose that we consider a resistor in an electrical circuit, and let v(t) denote the voltage
signal across the resistor i(t) denote the current. From Ohm’s law, we know that v(t) = i(t)R,
where R is the resistance. The power dissipated by the resistor is then

v 2 (t)
p(t) = v(t)i(t) = i2 (t)R =
R
Thus the power is a scaled multiple of the square of the voltage and current signals.
Since the energy expended over a time-interval [t1 , t2 ] is given by the integral of the power
dissipated per-unit-time over that interval, we have
Z t2 Z t2 Z t2
2 1
E= p(t)dt = R i (t)dt = v 2 (t)dt
t1 t1 R t1

The average power dissipated over the time-interval [t1 , t2 ] is then

Z t2 Z t2
1 1 1 2 R
E= v (t)dt = i2 (t)dt
t2 − t1 t2 − t1 R t1 t2 − t1 t1

We will find it useful to discuss the energy and average power of any continuoustime
or discrete-time signal. In particular, the energy of a general (potentially complex-valued)
continuous-time signal f (t) over a time-interval [t1 , t2 ] is defined as
Z t2
E[t1 ,t2 ] ≜ |f (t)|2 dt
t1

where |f (t)| denotes the magnitude of the signal at time t.

Similarly, the energy of a general (potentially complex-valued) discrete-time signal f [n] over
a time-interval [n1 , n2 ] is defined as
n2
X
E[n1 ,n2 ] ≜ |f [n]|2
n=n1

Note that we are defining the energy of an arbitrary signal in the above way; this will end
up being a convenient way to measure the "size" of a signal, and may not actually correspond
to any physical notion of energy.
We will also often be interested in measuring the energy of a given signal over all time. In
this case, we define
Z ∞
E∞ ≜ |f (t)|2 dt
−∞

for continuous-time signals, and

∞
X
E∞ ≜ |f [n]|2
n=∞
Properties of Signals and Systems 21

for discrete-time signals. Note that the quantity E∞ may not be finite.
Similarly, we define the average power of a continuous-time signal as
Z T
1
P∞ ≜ lim |f (t)|2 dt
T →∞ 2T −T

and for a discrete-time signal as

N
1 X
P∞ ≜ lim |f [n]|2
N →∞ 2N + 1
n=−N

Based on the above definitions, we have three classes of signals: finite energy (E∞ < ∞),
finite average power (P∞ < ∞), and those that have neither finite energy nor average power.
An example of the first class is the signal f (t) = e−t for t ≥ 0 and f (t) = 0 for t < 0. An
example of the second class is f (t) = 1 for all t ∈ R. An example of the third class is f (t) = t
for t ≥ 0. Note that any signal that has finite energy will also have finite average power,
since

E∞
P∞ = lim =0
T →∞ 2T

for continuous-time signals with finite energy, with an analogous characterization for
discrete-time signals.

2 Transformations of Signals
Throughout the course, we will be interested in manipulating and transforming signals into
other forms. Here, we start by considering some very simple transformations involving the
time variable. For the purposes of introducing these transformations, we will consider a
continuous-time signal f (t) and a discretetime signal f [n].
Time-shifting: Suppose we define another signal g(t) = f (t − t0 ), where t0 ∈ R. In
other words, for every t ∈ R, the value of the signal g(t) at time t is the value of the signal
f (t) at time t − t0 . If t0 > 0, then g(t) is a "forward-shifted" (or time-delayed) version of
f (t), and if t0 < 0, then g(t) is a time-advanced version of f (t) (i.e., the features in f (t)
appear earlier in time in g(t)). Similarly, for a discrete-time signal f [n], one can define the
time-shifted signal f [n − n0 ], where n0 is some integer.
Time-reversal: Consider the signal g(t) = f (−t). This represents a reversal of the
function f (t) in time. Similarly, f [−n] represents a time-reversed version of the signal f [n].
Time-scaling: Define the signal g(t) = f (αt), where α is some real number. When
0 < α < 1, this represents a stretching of f (t), and when α > 1, this represents a compression
of f (t). If α < 0, we get a time-reversed and stretched (or compressed) version of f (t).
Analogous definitions hold for the discrete-time signal f [n].
The operations above can be combined to define signals of the form g(t) = f (αt+β), where
α and β are real numbers. To draw the signal g(t), we should first apply the time-shift by β
to f (t) and then apply the scaling α. To see why, define h(t) = f (t + β), and g(t) = h(αt).
Thus, we have g(t) = f (αt+β) as required. If we applied the operations in the other order, we
22 Properties of Signals and Systems

would first get the signal h(t) = f (αt), and then g(t) = h(t + β) = f (α(t + β)) = f (αt + αβ).
In other words, the shift would be by αβ rather than β.
Examples of these operations can be found in the textbook (OW), such as example 1.1.

3 Periodic, Even and Odd Signals

A continuous-time signal f (t) is said to be periodic with period T if f (t) = f (t + T ) for all
t ∈ R. Similarly, a discrete-time signal f [n] is periodic with period N if f [n] = f [n + N ] for
all n ∈ Z. The fundamental period of a signal is the smallest period for which the signal is
periodic.
A signal is even if f (t) = f (−t) for all t ∈ R (in continuous-time), or f [n] = f [−n] for all
n ∈ Z (in discrete-time). A signal is odd if f (t) = −f (−t) for all t ∈ R, or f [n] = −f [−n]
for all n ∈ Z. Note that if a signal is odd, it must necessarily be zero at time 0 (since
f (0) = −f (0) ).
Given a signal f (t), define the signals
1 1
e(t) = (f (t) + f (−t)), o(t) = (f (t) − f (−t)).
2 2
It is easy to verify that o(t) is an odd signal and e(t) is an even signal. Furthermore,
x(t) = e(t) + o(t). Thus, any signal can be decomposed as a sum of an even signal and an
odd signal.

4 Exponential and Sinusoidal Signals

4.1 Continuous-Time Complex Exponential Signals
Consider a signal of the form

f (t) = Ceat
where C and a are complex numbers. If both C and a are real, there are three possible
behaviors for this signal. If a < 0, then the signal goes to zero as t → ∞, and if a > 0, the
signal goes to ∞ as t → ∞. For a = 0, the signal is constant.
Now suppose f (t) = ej(ω0 t+ϕ) for some positive real number ω0 and real number ϕ (this
corresponds to C = ejϕ and a = jω0 in the signal given above). We first note that

f (t + T ) = ej(ω0 (t+T )+ϕ) = ej(ω0 t+ϕ) ejω0 T

If T is such that ω0 T is an integer multiple of 2π, we have ejω0 T = 1 and the signal is
periodic with period T . Thus, the fundamental period of this signal is
2π
T0 =
ω0
Note that if ω0 = 0, then f (t) = 1 and is thus periodic with any period. The fundamental
period is undefined in this case. Also note that f (t) = e−jω0 T is also periodic with period T0 .
Properties of Signals and Systems 23

The quantity ω0 is called the fundamental frequency of the signal.

Note that periodic signals (other than the one that is zero for all time) have infinite energy,
but finite average power. Specifically, let
Z T0
Ep = |f (t)|2 dt
0
be the energy of the signal over one period. The average power over that period is then
Pp = ET0p , and since this extends over all time, this ends up being the average power of the
signal as well. For example, for the signal f (t) = ej(ω0 t+ϕ) , we have
Z T Z T
1 2 1
P∞ = lim |f (t)| dt = lim 1dt = 1
T →∞ 2T −T T →∞ 2T −T

Given a complex exponential with fundamental frequency ω0 , a harmonically related set

of complex exponentials is a set of periodic exponentials of the form

ϕk (t) = ejkω0 t , k ∈ Z
In other words, it is the set of complex exponentials whose frequencies are multiples of the
fundamental frequency ω0 . Note that if ejω0 t is periodic with period T0 (i.e, ω0 T0 = 2πm for
some integer m ), then ϕk (t) is also periodic with period T0 for any k ∈ Z, since

ϕk (t + T0 ) = ejkω0 (t+T0 ) = ejkω0 T0 ejkω0 t = ejkm2π ϕk (t) = ϕk (t)

Although the signal f (t) given above is complex-valued in general, its real part and imag-
inary part are sinusoidal. To see this, use Euler’s formula to obtain

Aej(ω0 t+ϕ) = A cos (ω0 t + ϕ) + jA sin (ω0 t + ϕ)

Similarly, we can write
A j(ω0 t+ϕ) A −j(ω0 t+ϕ)
A cos (ω0 t + ϕ) = e + e
2 2
i.e., a sinusoid can be written as a sum of two complex exponential signals.
Using the above, there are two main observations. First, continuous-time complex expo-
nential signals are periodic for any ω0 ∈ R (the fundamental period is ω2π0 for ω0 ̸= 0 and
undefined otherwise). Second, the larger ω0 gets, the smaller the period gets.
We will now look at discrete-time complex exponential signals and see that the above two
observations do not necessarily hold for such signals.

4.2 Discrete-Time Complex Exponential Signals

As in the continuous-time case, a discrete-time complex exponential signal is of the form

f [n] = Cean
where C and a are general complex numbers. As before, let us focus on the case where
C = 1 and a = jω0 for some ω0 ∈ R in order to gain some intuition, i.e., f [n] = ejω0 n .
24 Properties of Signals and Systems

To see the differences in discrete-time signals from continuous-time signals, recall that a
continuous-time complex exponential is always periodic for any ω0 . The first difference
between discrete-time complex exponentials and continuous time complex exponentials is
that discrete-time complex exponentials are not necessarily periodic. Specifically, consider
the signal f [n] = ejω0 n , and suppose it is periodic with some period N0 . Then by definition,
it must be the case that

f [n + N0 ] = ejω0 (n+N0 ) = ejω0 n ejω0 N0 = f [n]ejω0 N0 .

Due to periodicity, we must have ejω0 N0 = 1, or equivalently, ω0 N0 = 2πk for some integer
k. However, N0 must be an integer, and thus we see that this can be satisfied if and only
of ω0 is a rational multiple of 2π. In other words, only discrete-time complex exponentials
whose frequencies are of the form

k
ω0 = 2π
N
for some integers k and N are periodic. The fundamental period N0 of a signal is the
smallest nonnegative integer for which the signal is periodic. Thus, for discrete-time complex
exponentials, we find the fundamental period by first writing

ω0 k
=
2π N
where k and N have no factors in common. The value of N in this representation is then
the fundamental period.
2π 3π
Example : Consider the signal f [n] = ej 3 n +ej 4 n . Since both of the exponentials
have frequencies that are rational multiples of 2π, they are both periodic. For
the first exponential, we have
2π
3 1
=
2π 3
which cannot be reduced any further. Thus the fundamental period of the first
exponential is 3 . Similarly, for the second exponential, we have
3π
4 3
=
2π 8
Thus the fundamental period of the second exponential is 8 . Thus f [n] is
periodic with period 24 (the least common multiple of the periods of the two
signals).

■
The same reasoning applies to sinusoids of the form f [n] = cos (ω0 n). A necessary condi-
tion for this function to be periodic is that there are two positive integers n1 , n2 with n2 > n1
such that f [n1 ] = f [n2 ]. This is equivalent to cos (ω0 n1 ) = cos (ω0 n2 ). Thus, we must either
have
Properties of Signals and Systems 25

ω0 n2 = ω0 n1 + 2πk
or

ω0 n2 = −ω0 n1 + 2πk
for some positive integer k. In either case, we see that ω0 has to be a rational multiple of
2π. In fact, when ω0 is not a rational multiple of 2π, the function cos (ω0 n) never takes the
same value twice for positive values of n.
The second difference from continuous-time complex exponentials pertains to the period of
oscillation. Specifically, even for periodic discrete-time complex
exponentials, increasing the frequency does not necessarily make the period smaller. Consider
the signal g[n] = ej(ω0 +2π)n , i.e., a complex exponential with frequency ω0 + 2π. We have

g[n] = ejω0 n ej2πn = ejω0 n = f [n]

i.e., the discrete-time complex exponential with frequency ω0 + 2π is the same as the
discrete-time complex exponential with frequency ω0 , and thus they have the same funda-
mental period. More generally, any two complex exponential signals whose frequencies differ
by an integer multiple of 2π are, in fact, the same signal.
This shows that all of the unique complex exponential signals of the form ejω0 n have frequen-
cies that are confined to a region of length 2π. Typically, we will consider this region to be
0 ≤ ω0 < 2π, or −π ≤ ω0 < π. Suppose we consider the interval 0 ≤ ω0 < 2π. Note that

ejω0 n = cos (ω0 n) + j sin (ω0 n)

As ω0 increases from 0 to π, the frequencies of both the sinusoids increase. 1 For those ω0
between 0 and π that are also rational multiples of π, the sampled signals will be periodic,
and their period will decrease as ω0 increases.
Now suppose π ≤ ω0 < 2π. Consider the frequency 2π − ω0 , which falls between 0 and π.
We have

ej(2π−ω0 )n = e−jω0 n = cos (ω0 n) − j sin (ω0 n)

Since ejω0 n = cos (ω0 n) + j sin (ω0 n), the frequency of oscillation of the discretetime com-
plex exponential with frequency ω0 is the same as the frequency of oscillation of the discrete-
time complex exponential with frequency 2π − ω0 . Thus, as ω0 crosses π and moves towards
2π, the frequency of oscillation starts to decrease.
To illustrate this, it is again instructive to consider the sinusoidal signals f [n] = cos (ω0 n)
for ω0 ∈ 0, π2 , π, 3π . When ω0 = 0, the function is simply constant at 1 (and thus its period

2
is undefined). We see that the functions with ω0 = π2 and ω0 = 3π 2
have the same period (in
fact, they are exactly the same function). 2

The following table shows the differences between continuous-time and discretetime signals.
01
As we will see later in the course, the signals cos (ω0 n) and sin (ω0 n) correspond to continuous-time
signals of the form cos (ω0 t) and sin (ω0 t) that are sampled at 1 Hz . When 0 ≤ ω0 < π, this sampling rate
is above the Nyquist frequency ωπ0 , and thus the sampled signals will be an accurate representation of the
26 Properties of Signals and Systems

ejω0 t ejω0 n
Distinct signals for different values of ω0 Identical signals for values of ω0 sepa-rated by 2π
Periodic only if ω0 = 2π Nk for some in-
Periodic for any ω0
tegers k and N > 0.
Fundamental period undefined for ω0 =
Fundamental period: undefined for
ω0 = 0 and ω2π0 otherwise
0 and k ω2π0 otherwise
Fundamental frequency ω0 Fundamental frequency ω0
k

As with continuous-time signals, for any period N , we define the harmonic family of
discrete-time complex exponentials as
2π
ϕk [n] = ejk N n , k∈Z
This is the set of all discrete-time complex exponentials that have a common period N ,
and frequencies whose multiples of 2π N
. This family will play a role in our analysis later in
the course.

5 Impulse and Step Functions

5.1 Discrete-Time
The discrete-time unit impulse signal (or function) is defined as

0 if n ̸= 0

δ[n] = .
1 if n = 0
The discrete-time unit step signal is defined as
(
0 if n < 0
u[n] =
1 if n ≥ 0
Note that by the time-shifting property, we have

δ[n] = u[n] − u[n − 1]

X∞
u[n] = δ[n − k]
k=0

In other words, the unit step function can be viewed as a superposition of shifted impulse
functions.
Suppose we are given some arbitrary signal f [n]. If we multiply f [n] by the time-shifted
underlying continuous-time signal.
2
Note that cos (ω0 n) = cos ((2π − ω0 ) n ) for any 0 ≤ ω0 ≤ π. The same is not true for sin (ω0 n). In fact,
one can show that for any two different frequencies 0 ≤ ω0 < ω1 ≤ 2π, sin (ω0 n) and sin (ω1 n) are different
functions.
Properties of Signals and Systems 27

impulse function δ[n − k], we get a signal that is zero everywhere except at n = k, where
it takes the value f [k]. This is known as the sampling or sifting property of the impulse
function:

f [n]δ[n − k] = f [k]δ[n − k]
More generally, for any signal f [n], we have
∞
X
f [n] = f [k]δ[n − k]
k=−∞

i.e., any function f [n] can be written as a sum of scaled and shifted impulse functions.

5.2 Continuous-Time
The continuous-time unit step function is defined by
(
0 if t < 0
u(t) =
1 if t ≥ 0

Note that u(t) is discontinuous at t = 0 (we will take it to be continuous from the right).
To define the continuous-time analog of the discrete-time impulse function, we first define
the signal

0 if t < 0

δϵ (t) = 1ϵ if 0 ≤ t ≤ ϵ
0 if t > ϵ


R∞
where ϵ ∈ R>0 . Note that for any ϵ > 0, we have −∞ δϵ (t)dt = 1. As ϵ gets smaller, the
width of this function gets smaller and the height increases proportionally. The continuous-
time impulse function is defined as the limit of the above function as ϵ approaches zero from
the right:

δ(t) = lim δϵ (t)

ϵ↓0

This function is drawn with an arrow at the origin (since it has no width and infinite
height). We will often be interested in working with scaled and timeshifted versions of the
continuous-time impulse function. Just as we did with discrete-time functions, we can take
a continuous-time function f (t) and represent it as
Z ∞
f (t) = f (τ )δ(t − τ )dτ
−∞

In other words, if we take an infinite sequence of shifted impulse functions, scale each of
them by the value of the function f (t) at the value of the time-shift, and add them together
(represented by the integration), we get the function f (t). For instance, we have
28 Properties of Signals and Systems

Z ∞ Z ∞
u(t) = u(τ )δ(t − τ )dτ = δ(t − τ )dτ
−∞ 0

Just as the discrete-time impulse function could be viewed as a difference of the discrete-
time unit step and its time-shifted version, the continuous-time impulse function can be
viewed as the derivative of the continuous-time unit step function.

6 Properties of Systems
As we discussed during the first lecture, a system can be viewed as an abstract object
that takes input signals and produces output signals. A continuous-time system operates
on continuous-time signals, and discrete-time systems operate with discrete-time signals.
Examples of the former include many physical systems such as electrical circuits, vehicles,
etc. Examples of the latter include computer systems, a bank account where the amount of
money is incremented with interest, deposits and withdrawals at the end of each day, etc.

6.1 Interconnections of Systems

We will often be interested in connecting different systems together in order to achieve a
certain objective. There are three basic interconnection patterns that are used to build more
complicated interconnections.
The first is a serial connection of systems:

An example of a series interconnection of systems occurs in communication systems; the

signal to be transmitted is first passed through an encoder, which transforms the signal
into a form suitable for transmission. That transformed signal is then sent through the
communication channel (the second system in the chain). The output of the communication
channel is then passed through a decoder, which is the third system in the chain. The output
of the decoder is an estimate of the signal that entered the encoder.
The second type of interconnection is a parallel interconnection:

The third type of interconnection is a feedback interconnection:

Feedback interconnections form the basis of control systems; in this case we are given a specific
system (System 1) that we wish to control (or make behave in a certain way). The second
system (System 2) is a controller that we design in order to achieve the desired behavior.
This controller takes the current output of the system and uses that to decide what other
inputs to apply to the original system in order to change the output appropriately.
Properties of Signals and Systems 29

6.2 Properties of Systems

Systems With and Without Memory
A system is memoryless if the output each time-instant (either in discretetime or continuous-
time) only depends on the input at that time-instant. For example, the system

y[n] = cos(x[n])

is memoryless, as the output at each time-step n ∈ Z only depends on the input at that
time-step.
However, the system whose input and output are related by
Z t
y(t) = x(τ )dτ
−∞

is not memoryless, as the output depends on all of the input values from the past.
Systems with memory are often represented as having some sort of state and dynamics, which
maintains the necessary information from the past. For example, for the system given above,
we can use the fundamental theorem of calculus to obtain

dy(t)
= x(t)
dt
30 Properties of Signals and Systems

where the state of the system is y(t) (this is also the output), and the dynamics of the
state are given by the differential equation above. Similarly for the
discrete-time system
n
X
y[n] = x[k]
k=−∞

we have
n−1
X
y[n] = x[k] + x[n] = y[n − 1] + x[n]
k=−∞

which is a difference equation describing how the state y[n] evolves over time.

Invertibility
A system is said to be invertible if distinct inputs lead to distinct outputs. In other words,
by looking at the output of a system, one can uniquely identify what the input was.
An example of an invertible system is y(t) = αx(t), where α is any nonzero real number.
Given the output y(t), we can uniquely identify the input as x(t) = α1 y(t). However, if α = 0,
then we have y(t) = 0 regardless of the input, and there is no way to recover the input. In
that case, the system would not be invertible.
Another example of a noninvertible system is y(t) = x2 (t), as the sign of the input is lost
when converting toP the output.
The system y[n] = nk=−∞ x[k] is invertible; to see this, we use the equivalent representation
y[n] = y[n − 1] + x[n] to obtain x[n] = y[n] − y[n − 1] for all n ∈ Z.

Causality
A system is causal if the output of the system at any time depends only on the input at that
time and from the past. In other words, for all t ∈ R, y(t) depends only on x(τ ) for τ ≤ t.
Thus, a causal system does not react to inputs that will happen in the future. For a causal
system, if two different inputs have the same values up to a certain time, the output of the
system due to those two inputs will agree up to that time as well. All memoryless systems
are causal.
There are various instances where we may wish to use noncausal systems. For example, if
we have time-series data saved offline, we can use the saved values of the signal for k > n to
process the signal at a given time-step n (this can be used for music and video editing, for
example). Alternatively, the independent variable may represent space, rather than time. In
this case, one can use the values of the signal from points on either side of a given point in
order to process the signal.
Example :
The signal y[n] = x[−n] is noncausal; for example, y[−1] = x[1], and thus the output at
negative time-steps depends on the input from positive time-steps (i.e., in the future).
The signal y(t) = x(t) cos(t + 1) is causal; the t + 1 term does not appear in the input,
and thus the output at any time does not depend on values of the input at future times.
Properties of Signals and Systems 31

Stability
The notion of stability is a critical system property. There are many different notions of
stability that can be considered, but for the purposes of this course, we will say that a
system is stable if a bounded input always leads to a bounded output. In other words, for
a continuous-time system, if there exists a constant B1 ∈ R≥0 such that the input satisfies
|x(t)| ≤ B1 for all t ∈ R, then there should exist some other constant B2 ∈ R≥0 such that
|y(t)| ≤ B2 for all t ∈ R. An entirely analogous definition holds for discrete-time systems.
Loosely speaking, for a stable system, the output cannot grow indefinitely when the input is
bounded by a certain value.
Example :
The system y(t) = tx(t) is memoryless and causal, but not stable. For example, if x(t) = 1
for all t ∈ R, we have y(t) = t which is not bounded by any constant.
Similarly,
Pn the system y[n] = y[n − 1] + x[n] is not stable. This is seen by noting that
y[n] = k=−∞ x[k]. So, for example, if x[n] = u[n], we have y[n] = (n + 1) if n ≥ 0, which
is unbounded.
An example of a stable causal memoryless system is y(t) = cos(x(t)). Another example of
a stable and causal system is
(
0 if n < 0
y[n] =
αy[n − 1] + x[n] if n ≥ 0
where α ∈ R satisfies |α| < 1. Specifically, if |x[n]| ≤ B1 for all n ∈ Z, then we have
B1
|y[n]| ≤ 1−|α| for all n ∈ Z. To see this, we prove by induction. Clearly |y[n]| ≤ 1−|α|
B1
for
n ≤ 0. Suppose that |y[n]| ≤ 1−|α| for some n ≥ 0. Then we have
B1

|y[n + 1]| = |αy[n] + x[n + 1]| ≤ |α||y[n]| + |x[n + 1]|

B1
≤ |α| + B1
1 − |α|
B1
= .
1 − |α|
Thus, by induction, we have |y[n]| ≤ B1
1−|α|
for all n ∈ Z. ■

The above notion of stability is known as Bounded-Input-Bounded-Output (BIBO)

stability. There are also other notions of stability, such as ensuring that the internal state
of the system remains stable as well. One of the main objectives of control systems is to
ensure that the overall system remains stable, as you will see in your control systems courses
in later semesters.
Time-Invariance A system is said to be time-invariant if the system reacts in the same
way to an input signal, regardless of the time at which the input is applied. In other words,
if y(t) is the output signal when the input signal is x(t), then the output due to x (t − t0 )
should be y (t − t0 ) for any time-shift t0 . Note that for a system to be time-invariant, this
should hold for every input signal.
32 Properties of Signals and Systems

Example :
The system y(t) = cos(x(t)) is time-invariant. Suppose we define the input signal w(t) =
x (t − t0 ) (i.e., a time-shifted version of x(t) ). Let yw (t) be the output of the system when
w(t) is applied. Then we have

yw (t) = cos(w(t)) = cos (x (t − t0 )) = y (t − t0 )

and thus the output due to x (t − t0 ) is time-shifted version of y(t), as required.
An example of a time-varying system is y[n] = nx[n]. For example, if x[n] = δ[n], then
we have the output signal y[n] = 0 for all time. However, if x[n] = δ[n − 1], then we have
y[n] = 1 for n = 1 and zero for all other times. Thus a shift in the input did not result in
a simple shift in the output. ■

Linearity
A system is linear if it satisfies the following two properties.

1. Additivity: Suppose the output is y1 (t) when the input is x1 (t), and the output is
y2 (t) when the input is x2 (t). Then the output to x1 (t) + x2 (t) is y1 (t) + y2 (t).

2. Scaling: Suppose the output is y(t) when the input is x(t). Then for any complex
number α, the output should be αy(t) when the input is αx(t).

Both properties together define the superposition property: if the input to the system is
α1 x1 (t) + α2 x2 (t), then the output should be α1 y1 (t) + α2 y2 (t). Note that this must hold for
any inputs and scaling parameters in order for the system to qualify as linear. An entirely
analogous definition holds for discrete-time systems.
For any linear system, the output must be zero for all time when the input is zero for all
time. To see this, consider any arbitrary input x(t), and let the corresponding output be
y(t). Then, using the scaling property, the output to
αx(t) must be αy(t) for any scalar complex number α. Simply choosing α = 0 yields the
desired result that the output will be the zero signal when the input is the zero signal.
Example :
The system y(t) = tx(t) is linear. To see this, consider two arbitrary input signals x1 (t)
and x2 (t), and two arbitrary scalars α1 , α2 . Then we have

t (α1 x1 (t) + α2 x2 (t)) = α1 tx1 (t) + α2 tx2 (t) = α1 y1 (t) + α2 y2 (t)

where y1 (t) and y2 (t) are the outputs due to x1 (t) and x2 (t), respectively.
The system y[n] = x2 [n] is nonlinear. Let y1 [n] = x21 [n] and y2 [n] = x22 [n]. Consider the
input x3 [n] = x1 [n] + x2 [n]. Then the output due to x3 [n] is

y3 [n] = x23 [n] = (x1 [n] + x2 [n])2 ̸= x21 [n] + x22 [n]
in general. Thus the additivity property does not hold, and the system is nonlinear.
The system y[n] = Re{x[n]} is nonlinear, where Re{·} denotes the real part of the
argument. To see this, let x[n] = a[n] + jb[n], where a[n] and b[n] are real-valued signals.
Consider a scalar α = j. Then we have
Properties of Signals and Systems 33

y[n] = Re{x[n]} = a[n]

However, Re{jx[n]} = Re{ja[n] − b[n]} = −b[n] ̸= jy[n]. Thus, scaling the input signal
by the scalar j does not result in the output being jy[n], and so the scaling property does
not hold.
The system y(t) = 2x(t) + 5 is nonlinear; it violates the fact that the all-zero input
should cause the output to be zero for all time. One can also verify this by applying two
different constant input signals and checking that the output due to the sum of the inputs
is not equal to the sum of the corresponding outputs. ■
We will be focusing almost entirely on linear time-invariant systems in this course; in
practice, many systems are nonlinear and time-varying, but can often be approximated by
linear time-invariant systems under certain operation conditions.
34 Properties of Signals and Systems
 CHAPTER 3
ANALYSIS OF LINEAR TIME-INVARIANT
SYSTEMS

In this part of the course, we will fo-

1 Discrete-Time LTI Systems . 36
cus on understanding the behavior of lin-
2 Continuous-Time LTI Systems 37
ear time-invariant (LTI) systems. As we
3 Properties of Linear Time-
will see, the linearity and time-invariance
Invariant Systems . . . . . . . 38
properties provide a nice way to understand
4 Differential and Difference
the input-output relationship of the system.
Equation Models for Causal
To develop this, let us start by considering
LTI Systems . . . . . . . . . . 44
discrete-time LTI systems.
5 Differential and Difference
Equation Models for Causal
LTI Systems . . . . . . . . . . 45
6 Differential and Difference
Equation Models for Causal
LTI Systems . . . . . . . . . . 47
7 Block Diagram Representa-
tions of Linear Differential
and Difference Equations . . . 49

35
36 Analysis of Linear Time-Invariant Systems

1 Discrete-Time LTI Systems

Consider a discrete-time system with input x[n] and output y[n]. First, define the impulse
response of the system to be the output when x[n] = δ[n] (i.e., the input is an impulse
function). Denote this impulse response by the signal h[n].
Now, consider an arbitrary signal x[n]. Recall from the sifting property of the impulse
function that
∞
X
x[n] = x[k]δ[n − k]
k=−∞

i.e., x[n] can be written as a superposition of scaled and shifted impulse functions.
Since the system is time-invariant, the response of the system to the input δ[t − k] is
h[t − k]. By linearity (and specifically the scaling property),
P the response to x[k]δ[n − k] is
x[k]δ[n − k]. By the additivity property, the response to ∞ k=−∞ x[k]δ[n − k] is then

∞
X
y[n] = x[k]h[n − k]
k=−∞

The above is called the convolution sum; the convolution of the signals x[n] and h[n] is
denoted by
∞
X
x[n] ∗ h[n] = x[k]h[n − k]
k=−∞

Thus we have the following very important property of discrete-time LTI systems: if x[n]
is the input signal to an LTI system, and h[n] is the impulse response of the system, then
the output of the system is y[n] = x[n] ∗ h[n].
Example :
Consider an LTI system with impulse response
(
1 if 0 ≤ n ≤ 3
h[n] =
0 otherwise
Suppose the input signal is
(
1 if 0 ≤ n ≤ 3
x[n] =
0 otherwise
Then we have
∞
X
y[n] = x[n] ∗ h[n] = x[k]h[n − k]
k=−∞

Since both x[k] = 0 for k < 0 and h[n − k] = 0 for k > n,

Analysis of Linear Time-Invariant Systems 37

n
X
y[n] = x[k]h[n − k]
k=0

Thus y[n] = 0 for n < 0. When n = 0 we have

0
X
y[0] = x[k]h[−k] = x[0]h[0] = 1
k=0

When n = 1 we have
1
X
y[1] = x[k]h[1 − k] = x[0]h[1] + x[1]h[0] = 2.
k=0

Similarly, y[2] = 3, y[3] = 4, y[4] = 3, y[5] = 2, y[6] = 1 and y[n] = 0 for n ≥ 7. ■

Example :
Consider an LTI system with impulse response h[n] = u[n]. Suppose the input signal is
x[n] = αn u[n] with 0 < α < 1. Then we have
∞
X
y[n] = x[n] ∗ h[n] = x[k]h[n − k]
k=−∞

Since both x[k] = 0 for k < 0 and h[n − k] = 0 for k > n, we have
n n
X 1 − αn+1
X
k
y[n] = x[k]h[n − k] = α =
k=0 k=0
1−α
for n ≥ 0, and y[n] = 0 for n < 0. ■

2 Continuous-Time LTI Systems

The analysis and intuition that we developed for discrete-time LTI systems carries forward
in an entirely analogous way for continuous-time LTI systems. Specifically, recall that for
any signal x(t), we can write
Z ∞
x(t) = x(τ )δ(t − τ )dτ
−∞

The expression on the right hand side is a superposition of scaled and shifted impulse
functions. Thus, when this signal is applied to an LTI system, the output will be a superpo-
sition of scaled and shifted impulse responses. More specifically, if h(t) is the output of the
system when the input is x(t) = δ(t), then the output for a general input x(t) is given by
Z ∞
y(t) = x(τ )h(t − τ )dτ
−∞

This is the convolution integral and is denoted by y(t) = x(t) ∗ h(t).

38 Analysis of Linear Time-Invariant Systems

Example :
Suppose x(t) = e−at u(t) with a ∈ R>0 and h(t) = u(t). Then the output of the LTI system
with impulse response h(t) is given by
Z ∞
y(t) = x(t) ∗ h(t) = x(τ )h(t − τ )dτ
−∞
Z ∞
= e−aτ u(τ )u(t − τ )dτ
−∞
Z t
= e−aτ dτ
0

if t ≥ 0, and y(t) = 0 otherwise. Evaluating the above expression, we have

(
1
(1 − e−at ) if t ≥ 0
y(t) = a
0 if t < 0
■
Example :
Consider the signals

if 0 < t < T if 0 < t < 2T

 
1 t
x(t) = , h(t) =
0 otherwise 0 otherwise
where T > 0 is some constant. The convolution of these signals is easiest to do graph-
ically and by considering different regions of the variable t. The result is


 0 t<0

1 2
2t 0<t<T



y(t) = T t − 12 T 2 T < t < 2T
 1 2 3 2
−2t + T t + 2T 2T < t < 3T





0 t > 3T
■

3 Properties of Linear Time-Invariant Systems

In this section we will study some useful properties of the convolution operation; based on
the previous section, this will have implications for the input-output behavior of linear time-
invariant systems ( h[n] for discrete-time systems and h(t) for continuous-time systems).

3.1 The Commutative Property

The first useful property of convolution is that it is commutative:

x[n] ∗ h[n] = h[n] ∗ x[n]

x(t) ∗ h(t) = h(t) ∗ x(t)
Analysis of Linear Time-Invariant Systems 39

To see this, start with the definition of convolution and perform a change of variable by
setting r = n − k. This gives
∞
X ∞
X
x[n] ∗ h[n] = x[k]h[n − k] = x[n − r]h[r] = h[n] ∗ x[n]
k=−∞ r=−∞

The same holds for the continuous-time convolution. Thus it does not matter which of
the signals we choose to flip and shift in the convolution operation.

3.2 The Distributive Property

The second useful property of convolution is that it is distributive:

x[n] ∗ (h1 [n] + h2 [n]) = x[n] ∗ h1 [n] + x[n] ∗ h2 [n]

x(t) ∗ (h1 (t) + h2 (t)) = x(t) ∗ h1 (t) + x(t) ∗ h2 (t)
This property is easy to verify:
∞
X
x[n] ∗ (h1 [n] + h2 [n]) = x[k] (h1 [n − k] + h2 [n − k])
k=−∞
X∞ ∞
X
= x[k]h1 [n − k] + x[k]h2 [n − k]
k=−∞ k=−∞

= x[n] ∗ h1 [n] + x[n] ∗ h2 [n].

The distributive property has implications for LTI systems connected in parallel:

Let h1 [n] be the impulse response of System 1 , and let h2 [n] be the impulse response for
System 2. Then we have y1 [n] = x[n] ∗ h1 [n] and y2 [n] = x[n] ∗ h2 [n]. Thus,

y[n] = y1 [n] + y2 [n] = x[ n ∗ h1 [n] + x[n] ∗ h2 [n] = x[n] ∗ (h1 [n] + h2 [n])
The above expression indicates that the parallel interconnection can equivalently be viewed
as x[n] passing through a single system whose impulse response is h1 [n] + h2 [n] :
40 Analysis of Linear Time-Invariant Systems

Figure 3.1: A series interconnection of systems.

Figure 3.2: Caption

3.3 The Associative Property

A third useful property of convolution is that it is associative:

x[n] ∗ (h1 [n] ∗ h2 [n]) = (x[n] ∗ h1 [n]) ∗ h2 [n]

x(t) ∗ (h1 (t) ∗ h2 (t)) = (x(t) ∗ h1 (t)) ∗ h2 (t).

In other words, it does not matter which order we do the convolutions. The above rela-
tionships can be proved by manipulating the summations (or integrals); we won’t go into the
details here.
Just as the distributive property had implications for parallel interconnections of systems,
the associative property has implications for series interconnections of systems. Specifically,
consider the series interconnection shown in Fig. 3.1.
We have

y[n] = y1 [n] ∗ h2 [n] = (x[n] ∗ h1 [n]) ∗ h2 [n] = x[n] ∗ (h1 [n] ∗ h2 [n])

Thus, the series interconnection is equivalent to a single system with impulse response
h1 [n] ∗ h2 [n], as shown in Fig. 3.2.
Further note that since h1 [n] ∗ h2 [n] = h2 [n] ∗ h1 [n], we can also interchange the order of
the systems in the series interconnection as shown in Fig. 3.3, without changing the overall
input-output relationship between x[n] and y[n].
Analysis of Linear Time-Invariant Systems 41

Figure 3.3: An equivalent series representation of the interconnection shown in Fig. 3.1

Figure 3.4: A system in series with its inverse.

3.4 Memoryless LTI Systems

Let us now see the implications of the memoryless property for LTI systems. Specifically, let
h[n] (or h(t) ) be the impulse response of a given LTI system. Since we have
∞
X ∞
X
y[n] = x[k]h[n − k] = x[n − k]h[k]
k=−∞ k=−∞

we see that y[n] will depend on a value of the input signal other than at timestep n unless
h[k] = 0 for all k ̸= 0. In other words, for an LTI system to be memoryless, we require
h[n] = Kδ[n] for some constant K. Similarly, a continuous-time LTI system is memoryless
if and only if h(t) = Kδ(t) for some constant K. In both cases, all LTI memoryless systems
have the form

y[n] = Kx[n] or y(t) = Kx(t)

for some constant K.

3.5 Invertibility of LTI Systems

Consider an LTI system with impulse response h[n] (or h(t) ). Recall that the system is
said to be invertible if the output of the system uniquely specifies the input. If a system is
invertible, there is another system (known as the "inverse system") that takes the output of
the original system and outputs the input to the original system, as shown in Fig. 3.4.
Suppose the second system is LTI and has impulse response hI [n]. Then, by the associative
property discussed earlier, we see that the series interconnection of the system with its inverse
is equivalent (from an input-output sense) to a single system with impulse response h[n]∗hI [n].
In particular, we require

x[n] = x[n] ∗ (h[n] ∗ hI [n])

for all input signals x[n], from which we have

h[n] ∗ hI [n] = δ[n]

42 Analysis of Linear Time-Invariant Systems

In other words, if we have an LTI system with impulse response h[n], and another LTI
system with impulse response hI [n] such that h[n] ∗ hI [n] = δ[n], then those systems are
inverses of each other. The analogous statement holds in continuous-time as well.

Example :
Consider the LTI system with impulse response h[n] = αn u[n]. One can verify that this
impulse response corresponds to the system
n
X
y[n] = x[k]αn−k = αy[n − 1] + x[n]
k=−∞

Now consider the system yI [n] = xI [n] − αxI [n − 1], with input signal xI [n] and output
signal yI [n]. The impulse response of this system is hI [n] = δ[n]−
αδ[n − 1]. We have

h[n] ∗ hI [n] = αn u[n] ∗ (δ[n] − αδ[n − 1])

= αn u[n] ∗ δ[n] − (αn u[n]) ∗ (αδ[n − 1])
= αn u[n] − α αn−1 u[n − 1]

= αn (u[n] − u[n − 1])

= αn δ[n]
= δ[n]
Thus, the system with impulse response hI [n] is the inverse of the system with impulse
response h[n]. ■

3.6 Causality of LTI Systems

Recall that a system is causal if its output at time t depends only on the inputs up to (and
potentially including) t. To see what this means for LTI systems, consider the convolution
sum

∞
X ∞
X
y[n] = x[k]h[n − k] = x[n − k]h[k]
k=−∞ k=−∞

where the second expression follows from the commutative property of the convolution. In
order for y[n] to not depend on x[n + 1], x[n + 2], . . ., we see that h[k] must be zero for k < 0.
The same conclusion holds for continuous-time systems, and thus we have the following: A
continuous-time LTI system is causal if and only if its impulse response h(t) is zero for all
t < 0. A discrete-time LTI system is causal if and only if its impulse response h[n] is zero for
all n < 0.
Note that causality is a property of a system; however we will sometimes refer to a signal
as being causal, by which we simply mean that its value is zero for n or t less than zero.
Analysis of Linear Time-Invariant Systems 43

3.7 Stability of LTI Systems

To see what the LTI property means for stability of systems, consider again the convolution
sum

∞
X
y[n] = x[k]h[n − k]
k=−∞

Note that

∞
X ∞
X
|y[n]| = x[k]h[n − k] ≤ |x[k]h[n − k]|
k=−∞ k=−∞
X∞
= |x[k]||h[n − k]|
k=−∞

Now suppose that x[n] is bounded, i.e., there exists some B ∈ R≥0 such that |x[n]| ≤ B
for all n ∈ Z. Then the above expression becomes

∞
X
|y[n]| ≤ B |h[n − k]|
k=−∞

Thus, if ∞k=−∞ |h[n − k]| < ∞ (which means that h[n] is absolutely summable), then
P
|y[n]| will
P∞also be bounded for all n. It turns out that this is a necessary condition as
well: if k=−∞ |h[n − k]| = ∞, then there is a bounded input that causes the output to be
unbounded.
The same conclusion holds in continuous-time
R∞ as well. Thus, we have: A continuous-time
LTI system isP stable if and only if −∞ |h(τ )|dτ < ∞. A discrete-time LTI system is stable
if and only if ∞ k=−∞ |h[k]| < ∞.

Example :
Consider the LTI system with impulse response h[n] = αn u[n], where α ∈ R. We have
∞ ∞
(
X X 1
if |α| < 1
|h[k]| = |α|k = 1−|α|
k=−∞ k=0
∞ if |α| ≥ 1
Thus, the system is stable if and only if |α| < 1.
Similarly, consider the continuous-time LTI system with impulse response h(t) = eαt u(t),
where α ∈ R. We have
(
∞
− α1 if α < 0
Z ∞ Z ∞
ατ 1 ατ
|h(τ )|dτ = e dτ = (e ) =
−∞ 0 α 0 ∞ if α ≥ 0
Thus, the system is stable if and only if α < 0. ■
44 Analysis of Linear Time-Invariant Systems

3.8 Step Response of LTI Systems

Just as we defined the impulse response of a system to be the output of the system when the
input is an impulse function, we define the step response of a system to be the output when
the input is a step function u[n] (or u(t) in continuous-time). We denote the step response
as s[n] for discrete-time systems and s(t) for continuous-time systems.
To see how the step response is related to the impulse response, note that
∞
X ∞
X n
X
s[n] = u[k]h[n − k] = u[n − k]h[k] = h[k]
k=−∞ k=−∞ k=−∞
This is equivalent to s[n] = s[n − 1] + h[n]. Thus, the step response of a discretetime LTI
system is the running sum of the impulse response.
Note that this could also have been seen by noting that δ[n] = u[n] − u[n − 1]. If the
impulse is applied to an LTI system, we get the impulse response h[n]. However, by the
linearity property, this output must be the superposition of the outputs due to u[n] and
u[n − 1]. By the time-invariance property, the output due to u[n − 1] is s[n − 1], and thus
for LTI systems we have h[n] = s[n] − s[n − 1], which corroborates what we obtained above.
For continuous-time systems, we have the same idea:
Z ∞ Z ∞ Z t
s(t) = u(τ )h(t − τ )dτ = u(t − τ )h(τ )dτ = h(τ )dτ
−∞ −∞ −∞
Differentiating both sides and applying the fundamental theorem of calculus, we have
ds
= h(t)
dt
i.e., the impulse response is the derivative of the step response.

4 Differential and Difference Equation Models for Causal

LTI Systems
As we have already seen in a few examples, many systems can be described using differential
equation (in continuous-time) or difference-equation (in discrete time) models, capturing the
relationship between the input and the output. For example, for a vehicle with velocity v(t)
and input acceleration a(t), we have
dv
= a(t)
dt
If we included wind resistance or friction (which produces a force that is proportional to
the velocity in the opposite direction of travel), we have
dv
= −αv(t) + a(t)
dt
where α > 0 is the coefficient of friction. Similarly, given an RC circuit, if we define the
voltage across the capacitor as the output, and the source voltage as the input, then the
input and output are again related via a differential equation of the above form.
Analysis of Linear Time-Invariant Systems 45

5 Differential and Difference Equation Models for Causal

LTI Systems
In discrete-time, consider a bank-account where earnings are deposited at the end of each
month. Let the amount in the account at the end of month n be denoted by s[n]. Then we
have

s[n] = (1 + r)s[n − 1] + x[n]

where r is the interest rate and x[n] is the new amount deposited into the account at the
end of month n.
Since such differential and difference equations play a fundamental role in the analysis of
LTI systems, we will now review some methods to solve such equations.

5.1 Linear Constant-Coefficient Differential Equations

To illustrate the solution of linear differential equations, we consider the following example.
Example :
Consider the differential equation
dy
+ 2y(t) = x(t)
dt
where x(t) = Ke3t u(t) ( K is some constant). The solution to such differential equations
is given by y(t) = yh (t) + yp (t), where yp (t) is a particular solution to the above equation,
and yh (t) is a homogeneous solution satisfying the differential equation
dyh
+ 2yh (t) = 0
dt
The above differential equation is called homogeneous as it has no driving function x(t).
Let us first solve the homogeneous equation. For equations of this form (where a sum
of derivatives of yh (t) have to sum to zero), a reasonable guess would be that yh (t) takes
the form

yh (t) = Aemt
for some m ∈ C. Substituting this into the homogeneous equation gives

mAemt + 2Aemt = 0 ⇒ m + 2 = 0 ⇒ m = −2
Thus, the homogeneous solution is yh (t) = Ae−2t , for any constant A.
Next, we search for a particular solution to the equation
dyp
+ 2yp (t) = Ke3t u(t)
dt
It seems reasonable to try yp (t) = Be3t , for some constant B. Substituting and evalu-
ating for t > 0, we have
46 Analysis of Linear Time-Invariant Systems

K
3Be3t + 2Be3t = Ke3t ⇒ B =
5
Thus, the particular solution is given by yp (t) = 5 e for t > 0.
K 3t

Together, we have y(t) = yh (t) + yp (t) = Ae−2t + K5 e3t for t > 0. Note that the coefficient
A has not been determined yet; in order to do so, we need more information about the
solutions to the differential equation, typically in the form of initial conditions. For exam-
ple, if we know that the system is at rest until the input is applied (i.e., y(t) = 0 until x(t)
becomes nonzero), we have y(t) = 0 for t < 0. Suppose we are given the initial condition
y(0) = 0. Then,
K K
y(0) = A + =0⇒A=−
5 5
Thus, with the given initial condition, we have y(t) = 5 (e3t − e−2t ) u(t).
K

The above example illustrates the general approach to solving linear differential equa-
tions of the form
N M
X dk y X dk x
ak k = bk k
k=0
dt k=0
dt
First find the homogeneous solution to the equation
N
X dk y h
ak k = 0
k=0
dt
by hypothesizing that yh (t) = Aemt for some m ∈ C. If there are N different values of m,
denoted m1 , m2 , . . . , mN for which the proposed form holds, then we take the homogeneous
solution to be yh (t) = A1 em1 t + A2 em2 t + · · · + AN emN t , where the coefficients A1 , . . . , AN
are to be determined from initial conditions. If there are not N different values of m, then
further work is required; we will see a more general way to solve these cases later in the
course.
Next, find a particular solution to the equation
N M
X dk yp X dk x
ak = bk k
k=0
dtk k=0
dt
where x(t) is some given function. The idea will be to make yp (t) a linear combination
of terms that, when differentiated, yield terms that appear in x(t) and its derivatives.
Typically this only works when x(t) involves terms like et , sin(t), cos(t), polynomials in t,
etc. Let’s try another example. ■
Example :
Consider the differential equation

y ′′ (t) + y ′ (t) − 6y(t) = x′ (t) + x(t)

where x(t) = e4t u(t).
Analysis of Linear Time-Invariant Systems 47

We first search for a homogeneous solution yh (t) = Aemt satisfying

yh′′ (t) + yh′ (t) − 6yh (t) = 0 ⇒ m2 Aemt + mAemt − 6Aemt = 0 ⇒ (m2 + m − 6) = 0. ■

6 Differential and Difference Equation Models for Causal

LTI Systems
This yields m = −3 or m = 2. Thus, the homogeneous solution is of the form

yh (t) = A1 e−3t + A2 e2t

for some constants A1 and A2 that will be determined from the initial conditions.
To find a particular solution, note that for t > 0, we have x′ (t) + x(t) = 4e4t + e4t = 5e4t .
Thus we search for a particular solution of the form yp (t) = Be4t for t > 0. Substituting into
the differential equation (3.1), we have
yp′′ (t) + yp′ (t) − 6yp (t) = x′ (t) + x(t) ⇒ 16Be4t + 4Be4t − 6Be4t = 5e4t ⇒ B = 14
5
.
Thus, yp (t) = 14 e for t > 0 is a particular solution.
5 4t

The overall solution is then of the form y(t) = yh (t) + yp (t) = A1 e−3t + A2 e2t + 14 5 4t
e
for t > 0. If we are told that the system is at rest until the input is applied, and that
y(0) = y ′ (0) = 0, we have

5
y(0) = A1 + A2 + =0
14
20
y ′ (0) = −3A1 + 2A2 + =0
14
Solving these equations, we obtain A1 = 17 and A2 = − 12 . Thus, the solution is
 
1 −3t 1 2t 5 4t
y(t) = e − e + e u(t)
7 2 14

6.1 Linear Constant Coefficient Difference Equations

The same general idea that we used to solve differential equations in the previous section
apply to solving difference equations of the form
N
X M
X
ak y[n − k] = bk x[n − k]
k=0 k=0

The overall solution will be of the form y[n] = yh [n] + yp [n], where yh [n] is a homogeneous
solution to
N
X
ak yh [n − k] = 0
k=0

and yp [n] is a particular solution satisfying the difference equation (3.2) for the given
function x[n]. In this case, we seek homogeneous solutions of the form yh [n] = Aβ n for some
48 Analysis of Linear Time-Invariant Systems

A, β ∈ C, and seek particular solutions that have the same form as the quantities that appear
in x[n]. Let’s do an example.

Example :
Suppose we have the difference equation
1
y[n] − y[n − 1] = x[n]
2

n
with x[n] = 31 u[n].
The solution to this difference equation will be of the form y[n] = yh [n] + yp [n], where
yh [n] is the homogeneous solution satisfying
1
yh [n] − yh [n − 1] = 0
2
and yp [n] is a particular solution to the given difference equation with the given input
signal x[n].
To find the homogeneous solution, we try yh [n] = Aβ n for some constants A and β.
Substituting into the homogeneous difference equation, we obtain
1 A 1
yh [n] − yh [n − 1] = Aβ n − β n−1 = 0 ⇒ β = .
2 2 2
Thus, the homogeneous solution is yh [n] = Aβ for some A that we will identify based
n

on initial conditions (after we have found the particular solution).

To find the particular solution, we
attempt to mimic the input. Thus, we seek a particular
n
solution of the form yp [n] = B 13 for n ≥ 0 and some constant B. Substituting this into
the difference equation (3.3), we have
 n  n−1
1 B 1 1n
B − =
3 2 3 3
for n ≥ 1 (note that we don’t look at n = 0 here because we have not defined yp [−1]).
1 n
Solving this, we get B = −2. Thus the particular solution is yp [n] = −2 3 for n ≥ 0.


1 n 1 n
Now, we have y[n] = yh [n] + yp [n] = A 2 − 2 3 for n ≥ 0. Suppose we are told that

the system is at rest for n < 0, i.e., y[n] = 0 for n < 0. Looking at equation (3.3), we have
1
y[0] − y[−1] = 1 ⇒ y[0] = 1
2
Substituting the expression for y[n], we have

1 = y[0] = A − 2 ⇒ A = 3
Thus, the solution is given by
  n  n 
1 1
y[n] = 3 −2 u[n]
2 3
■
Analysis of Linear Time-Invariant Systems 49

An alternative method to solve difference equations is to write them in recursive form,

and then iteratively solve, as shown by the following example.
Example :
Suppose
1
y[n] − y[n − 1] = x[n].
2
We can rewrite this as
1
y[n] = y[n − 1] + x[n]
2
Suppose that x[n] = δ[n] and the system is initially at rest (i.e., y[n] = 0 for n < 0).
Then we have
1
y[0] = y[−1] + δ[0] = 1
2
1 1
y[1] = y[0] + δ[1] =
2 2
1 1
y[2] = y[1] + δ[2] =
2 4
..
.
 n
1
y[n] =
2

n
Thus, the impulse response is h[n] = 12 u[n].1 ■

7 Block Diagram Representations of Linear Differential

and Difference Equations
It is often useful to represent linear differential and difference equations using block dia-
grams; this provides us with a way to implement such equations using primitive computa-
tional elements (form form the components of the block diagram), and to derive alternative
representations of systems. Here, we will focus on differential and difference equations of the
form

dN y(t) dN −1 y(t)
+ a N −1 + · · · + a0 y(t) = b0 x(t)
dtN dtN −1
y[n + N ] + aN −1 y[n + N − 1] + · · · + a0 y[n] = b0 x[n]
Drawing block diagrams for more general differential and difference equations (involving
more than just x[n] on the right hand side) is easier using Laplace
01
One can also calculate this using the homogeneous and particular solutions; in this case, the particular

n
solution would have the form yp [n] = Bδ[n] and B would be found to be zero, so that y[n] = yh [n] = A 12 .
Under the condition of initial rest and y[0] = 1 (obtained from the difference equation), we obtain A = 1,
thus matching the impulse response calculated recursively above.
50 Analysis of Linear Time-Invariant Systems

Figure 3.5: Integrator and Delay blocks.

Figure 3.6: Chained integrator and delay blocks.

and z-transform techniques, and so we will defer a study of such equations until then.
For the above equations, we start by writing the highest derivative of y (or the most advanced
version of y ) in terms of all of the other quantities:
$$

dN y(t) dN −1 y(t)
= −aN −1 − · · · − a0 y(t) + b0 x(t)
dtN dtN −1
y[n + N ] = −aN −1 y[n + N − 1] − · · · − a0 y[n] + b0 x[n]
$$
Next, we use a key building block: the integrator block (for continuous-time) or the delay
block (for discrete-time). Specifically, the integrator block is a system whose output is the
integral of the input, and the delay block is a system whose output is a delayed version of the
input. Thus, if we feed dydt
into the integrator block, we get y(t) out, and if we feed y[n + 1]
into the delay block, we get y[n] out, as shown in Fig. 3.5.
To use these blocks to represent differential and difference equations, we simply chain a
sequence of these blocks in series, and feed the highest derivative into the first block in the
chain, as shown in Fig. 3.6.
This series chain of integrator (or delay) blocks provides us with all of the signals needed
N
to represent (3.4) and (3.5). Specifically, from equation (3.4), we see that d dty(t)
N is a lin-
dN −1 y(t) N
ear combination of the signals dtN −1 , · · · , y(t), x(t). Thus, to generate the signal d dty(t)
N ,

we simply take the signals from the corresponding integrator blocks, multiply them by the
coefficients, and add them all together. The same holds true for the signal y[n + N ] in (3.5).
 CHAPTER 4
FOURIER SERIES REPRESENTATION OF
PERIODIC SIGNALS

In the last part of the course, we decom-

1 Applying Complex Exponen-
posed signals into sums of scaled and time-
tials to LTI Systems . . . . . 52
shifted impulse functions. For LTI systems,
2 Fourier Series Representation
we could then write the output as a sum
of ContinuousTime Periodic
of scaled and time-shifted impulse responses
Signals . . . . . . . . . . . . . 54
(using the superposition property). In this
3 Fourier Series Representation
part of the course, we will consider alternate
of Continuous-Time Periodic
(and very useful) decompositions of signals
Signals . . . . . . . . . . . . . 55
as sums of scaled complex exponential func-
4 Calculating the Fourier Series
tions. As we will see, such functions exhibit
Coefficients . . . . . . . . . . 56
some nice behavior when applied to LTI sys-
5 Properties of Continuous-
tems. This particular chapter will focus on
Time Fourier Series . . . . . . 62
decomposing periodic signals into complex
6 Fourier Series for Discrete-
exponentials (leading to the Fourier Series),
Time Periodic Signals . . . . . 66
and subsequent chapters will deal with the
decomposition of more general signals.

51
52 Fourier Series Representation of Periodic Signals

1 Applying Complex Exponentials to LTI Systems

Recall that a complex exponential has the form x(t) = est (in continuous-time), and x[n] = z n
(in discrete-time), where s and z are general complex numbers. Let’s start with a continuous-
time LTI system with impulse response h(t). When we apply the complex exponential x(t) =
est to the system, the output is given by
Z ∞ Z ∞
y(t) = x(t) ∗ h(t) = h(τ )x(t − τ )dτ = h(τ )es(t−τ ) dτ
−∞ −∞
Z ∞
=e st
h(τ )e−sτ dτ
−∞

Let us define
Z ∞
H(s) = h(τ )e−sτ dτ
−∞

If, for the given complex number s, the above integral exists (i.e., is finite), then H(s) is
just some complex number. Thus, we see that for an LTI system, if we apply the complex
exponential x(t) = est as an input, we obtain the quantity

y(t) = H(s)est
as an output. In other words, we get the same complex exponential out of the system,
just scaled by the complex number H(s). Thus, the signal est is called an eigenfunction of
the system, with eigenvalue H(s).
The same reasoning applies for discrete-time LTI systems. Consider an LTI system with
impulse response h[n], and input x[n] = z n . Then,
∞
X ∞
X
y[n] = x[n] ∗ h[n] = h[k]x[n − k] = h[k]z n−k
k=−∞ k=−∞
X∞
=z n
h[k]z −k .
k=−∞

Let us define
∞
X
H(z) = h[k]z −k .
k=−∞

If this sum converges for the given choice of complex number z, then H(z) is just some
complex number. Thus, we see again that for a discrete-time LTI system with the complex
exponential x[n] = z n as an input, we obtain the quantity

y[n] = H(z)z n
as an output. In this case z n is an eigenfunction of the system, and H(z) is the eigenvalue.
So, to summarize, we have the following:
Fourier Series Representation of Periodic Signals 53

• If the signal x(t) = est is applied to anR LTI system with impulse response h(t), the
∞
output is y(t) = H(s)est , where H(s) = −∞ h(τ )e−sτ dτ (assuming the integral exists).

• If the signal x[n] = z n is applied to anPLTI system with impulse response h[n], the
output is y[n] = H(z)z n , where H(z) = ∞ k=−∞ h[k]z
−k
(assuming the sum converges).

As we will see later in the course, the quantities H(s) and H(z) are the Laplace Transform
and z-Transform of the impulse response of the system, respectively.
Note that the above translates to superpositions of complex exponentials in an natural
way. Specifically, if the input is x(t) = i=1 ai e for some complex numbers a1 , . . . , an and
Pn si t

s1 , . . . , sn , we have
n
X
y(t) = ai H (si ) esi t
i=1

An essentially identical relationship is true for discrete-time systems.

Example :
Consider the signal x(t) = cos(ωt). We can write this as
1 1
x(t) = ejωt + e−jωt
2 2
Thus, the output will be
1 1
y(t) = H(jω)ejωt + H(−jω)e−jωt .
2 2
Now, suppose that the impulse response of the system is real-valued (i.e., h(t) ∈ R for
all t ). Then, we have
Z ∞ Z ∞
∗ jωτ ∗
h(τ )e−jωτ dτ = H(jω)


H(−jω) = h(τ )e dτ =
−∞ −∞

Thus, for real-valued impulse responses, we have H(−jω) = H(jω)∗ . We can equiva-
lently write these in polar form as

H(jω) = |H(jω)|ej∠H(jω) , H(−jω) = |H(jω)|e−j∠H(jω) .

Thus,
1 1
y(t) = H(jω)ejωt + H(−jω)e−jωt
2 2
1 1
= |H(jω)|e e + |H(jω)|e−j∠H(jω) e−jωt
j∠H(jω) jωt
2 2
= |H(jω)| cos(ωt + ∠H(jω))
■
Example :
Consider the system y(t) = x (t − t0 ), where t0 ∈ R. The impulse response if this system
is h(t) = δ (t − t0 ), and thus
54 Fourier Series Representation of Periodic Signals
Z ∞ Z ∞
−st
H(s) = h(t)e dt = δ (t − t0 ) e−st dt = e−st0
−∞ −∞

Suppose we apply the signal x(t) = cos (ω0 t) to the system. We expect the output
to be cos (ω0 (t − t0 )), based on the definition of the system. Let’s verify this using the
identities we derived earlier. We have

|H (jω0 )| = e−jω0 t0 = 1, ∠H (jω0 ) = −ω0 t0

Thus, when we apply the input cos (ω0 t), the output is given by

y(t) = |H (jω0 )| cos (ω0 t + ∠H (jω0 )) = cos (ω0 t − ω0 t0 )

matching what we expect. ■

2 Fourier Series Representation of ContinuousTime Peri-

odic Signals
Consider the complex exponential signal

x(t) = ejω0 t

Recall that this signal is periodic with fundamental period T = ω2π0 (assuming ω0 > 0
). Based on this complex exponential, we can define an entire harmonic family of complex
exponentials, given by

ϕk (t) = ejkω0 t , k ∈ Z

In other words, for each k ∈ Z, ϕk (t) is a complex exponential whose fundamental frequency
is kω0 (i.e., k times the fundamental frequency of x(t) ). Thus, each of the signals ϕk (t) is
periodic with period T , since

ϕk (t + T ) = ejkω0 (t+T ) = ejkω0 t ejkω0 T = ejkω0 t ejk2π = ejkω0 t

Note that T may not be the fundamental period of the signal ϕk (t), however.
Since each of the signals in the harmonic family is periodic with period T , a linear combination
of signals from that family is also periodic. Specifically, consider the signal

∞ ∞
2π
X X
jkω0 t
x(t) = ak e = ak ejk T t

k=−∞ k=−∞

The terms corresponding to k = 1 and k = −1 are known as the first harmonic of the
signal x(t). The terms corresponding to k = 2 and k = −2 are known as the second harmonic
and so forth.
Fourier Series Representation of Periodic Signals 55

3 Fourier Series Representation of Continuous-Time Pe-

riodic Signals
Suppose we are given a certain periodic signal x(t) with fundamental period T . Define
ω0 = 2π
T
and suppose that we can write x(t) as
∞ ∞
!
2π
X X
x(t) = ak ejk T t = ak ejkω0 t
k=−∞ k=−∞

for some sequence of coefficients ak , k ∈ Z. Then the above representation is called the
Fourier Series representation of x(t). The quantities ak , k ∈ Z are called the Fourier Series
coefficients.
Example :
Consider the signal x(t) = cos (ω0 t), where ω0 > 0. We have
1 1
x(t) = ejω0 t + e−jω0 t
2 2
This is the Fourier Series representation of x(t); it has only first harmonics, with coef-
ficients a1 = a−1 = 12 .
Similarly, consider the signal x(t) = sin (ω0 t). We have
1 jω0 t 1
x(t) = e − e−jω0 t .
2j 2j
Once again, the signal has only first harmonics, with coefficients a1 = 2j1 and a−1 = − 2j1 .
Suppose that we have a periodic signal that has a Fourier Series representation
∞
X
x(t) = ak ejkω0 t
k=−∞

Now suppose that x(t) is real, i.e., x∗ (t) = x(t). Taking the complex conjugate of both
sides of the above expression, we have
∞
X
x∗ (t) = a∗k e−jkω0 t
k=−∞

Equating the expressions for x(t) and x∗ (t), we have

∞
X ∞
X
ak e jkω0 t
= a∗k e−jkω0 t
k=−∞ k=−∞

Comparing the terms, we see that for any k ∈ Z, the coefficient of ejkω0 t is ak on the
left hand side, and is a∗−k on the right hand side. Thus, for real signals x(t), the Fourier
Series coefficients satisfy

a−k = a∗k
56 Fourier Series Representation of Periodic Signals

for all k ∈ Z. Substituting this into the Fourier Series representation (4.1) we have
∞
X
ak ejkω0 t + a−k e−jkω0 t


x(t) = a0 +
k=1
∞
X
ak ejkω0 t + a∗k e−jkω0 t


= a0 +
k=1
X∞
2 Re ak ejkω0 t

= a0 +
k=1

where Re is the real part of the given complex number. If we write ak in polar form as
rk ejθk , the above expression becomes
∞
X ∞
X
j(kω0 t+θk )

x(t) = a0 + 2 Re rk e = a0 + 2 rk cos (kω0 t + θk )
k=1 k=1

This is an alternate representation of the Fourier Series for real-valued signals (known
as the trigonometric representation). ■

4 Calculating the Fourier Series Coefficients

Suppose that we are given a periodic signal x(t) with period T , and that this signal has a
Fourier Series representation
∞
X
x(t) = ak ejkω0 t
k=−∞

We will soon see conditions under which a signal will have such a representation, but for
now, suppose that we are just interested in finding the coefficients ak , k ∈ Z. To do this,
multiply x(t) by e−jnω0 t , where n is some integer. This gives
∞
X ∞
X
−jnω0 t jkω0 t −jnω0 t
x(t)e = ak e e = ak ej(k−n)ω0 t
k=−∞ k=−∞

Now suppose that we integrate both sides of the above equation from t0 to t0 + T for any
t0 :
Z t0 +T Z t0 +T ∞
X ∞
X Z t0 +T
−jnω0 t j(k−n)ω0 t
x(t)e dt = ak e dt = ak ej(k−n)ω0 t dt
t0 t0 k=−∞ k=−∞ t0

Now note that if n = k, we have

Z t0 +T Z t0 +T
j(k−n)ω0 t
e dt = 1dt = T
t0 t0

Otherwise, if n ̸= k, we have
Fourier Series Representation of Periodic Signals 57

Z t0 +T 0 t +T
1
ej(k−n)ω0 t dt = ej(k−n)ω0 t
t0 j(k − n)ω0 t0
1
ej(k−n)ω0 (t0 +T ) − ej(k−n)ω0 t0


=
j(k − n)ω0
=0

Thus, we have
Z t0 +T ∞
X Z t0 +T
−jnω0 t
x(t)e dt = ak ej(k−n)ω0 t dt = an T
t0 k=−∞ t0

or equivalently,
Z t0 +T
1
an = x(t)e−jnω0 t dt
T t0

where t0 is any arbitrary starting point. In other words, we obtain the Fourier coefficient
an by multiplying the signal x(t) by e−jnω0 t and then integrating the resulting product over
any period.
Example :
Consider the signal

T
0 − 2 ≤ t < T1

x(t) = 1 −T1 ≤ t ≤ T1

0 T1 < t < T2


where T1 ≤ T and x(t) is T -periodic.

Define ω0 = 2π
T
. We have
T
Z Z T1
1 2 1 2T1
a0 = x(t)dt = x(t)dt =
T − T2 T T1 T
For k ̸= 0, we have
T T
Z Z T1 1
1 2
−jkω0 t 1 1 1
ak = x(t)e dt = e−jkω0 t dt = e−jkω0 t
T − T2 T −T1 T −jkω0 −T1
2
= sin (kω0 T1 )
T kω0
sin (kω0 T1 )
=
kπ
where we used the fact that T ω0 = 2π.
Note that in this case, ak is real for all k ∈ Z, and satisfies ak = a−k . Thus we can also
write the Fourier series as
58 Fourier Series Representation of Periodic Signals

∞
X ∞
X
ak ejkω0 t + e−jkω0 t = a0 + 2


x(t) = a0 + ak cos (kω0 t)
k=1 k=1

4.1 A Vector Analogy for the Fourier Series

In the derivation of the Fourier series coefficients, we saw that
Z t0 +T
ejkω0 t e−jnω0 t dt = 0
t0

if k ̸= n, and is equal to T otherwise. The functions ejkω0 and ejnω0 (for k ̸= n ) are said
to be orthogonal. More generally, a set of functions ϕk (t), k ∈ Z, are said to be orthogonal
on an interval [a, b] if
Z b
ϕk (t)ϕ∗n (t) = 0
a

if k ̸= n, and nonzero otherwise. Note that ϕ∗n (t) is the complex conjugate of ϕn (t).
We then derived the expressions for the coefficients by using the orthogonality property.
However, that derivation assumed that the signal could be written as a linear combination of
the functions in the harmonic family, and then derived the coefficient expressions. Here will
justify this by first trying to approximate a given signal by a finite number of functions from
the harmonic family and then taking the number of approximating functions to infinity. We
will start by reviewing how to approximate a given vector by other vectors, and then explore
the analogy to the approximation of functions.

Review of Approximation of Vectors

The above definition of orthogonality of functions is exactly analogous to the definition of
orthogonal vectors in a vector space. Recall that two vectors v1 , v2 ∈ Rn are said to be
orthogonal if v1′ v2 = 0. Suppose we are given the vectors
     
1 1 0
x =  2  , v1 =  0  , v2 =  1 
3 0 1
Note that v1′ v2 = 0 and thus v1 and v2 are orthogonal. Suppose we wish to approximate
the vector x using a linear combination of the vectors v1 and v2 . In other words, we wish to
find coefficients a and b so that the approximation

x̂ = av1 + bv2
is "close" to x. A typical metric of "closeness" is taken to be the square of the approxi-
mation error. Specifically, the approximation error is given by

e = x − x̂ = x − av1 − bv2
Fourier Series Representation of Periodic Signals 59

Note that e is a vector, where the i-th component is the approximation error for the i-th
component of x. We try to minimize e21 + e22 + e23 , which is given by

e21 + e22 + e23 = e′ e = (x − av1 − bv2 )′ (x − av1 − bv2 )

=x′ x − ax′ v1 − bx′ v2 − av1′ x + a2 v1′ v1 + abv1′ v2
− bv2′ x + abv2′ v1 + b2 v22
Noting that v1 and v2 are orthogonal, we have

e′ e = x′ x − ax′ v1 − bx′ v2 − av1′ x + a2 v1′ v1 − bv2′ x + b2 v22

This is a convex function of the scalars a and b. If we wish to minimize e′ e, we take the
derivative with respect to these scalars and set it equal to zero. This yields

∂e′ e
= −x′ v1 − v1′ x + 2av1′ v1 = 0
∂a
v′ x
⇒ a = ′1
v1 v1
′
∂e e
= −x′ v2 − v2′ x + 2bv2′ v2 = 0
∂b
v2′ x
⇒b= ′
v2 v2
where we used the fact that x′ v1 = v1′ x and x′ v2 = v2′ x (since these quantities are all
scalars). In terms of the vectors given above, we obtain
1 5
a= = 1, b= .
1 2

Application to Approximation of Functions

Entirely analogous ideas hold when we are trying to approximate one function as a linear
combination of a set of orthogonal functions (as in the Fourier series). Given a set of orthog-
onal functions ϕk (t) over an interval [a, b], suppose we wish to approximate a given function
x(t) as a linear combination of some finite number of these functions. Specifically, suppose
that we are given some positive integer N , and wish to find the best coefficients αk ∈ C such
that the estimate
N
X
x̂(t) = ak ϕk (t)
k=−N

is "close" to x(t). Mathematically, we will use the notion of squared error to measure
"closeness." Specifically, the approximation error at any given point in time t is given by
N
X
e(t) = x(t) − x̂(t) = x(t) − ak ϕk (t)
k=−N

and the squared error over the entire interval [a, b] is then defined as
60 Fourier Series Representation of Periodic Signals

Z b
|e(t)|2 dt
a

Here, we will allow e(t) to a general complex valued function, so the absolute value in the
integral is interpreted as the magnitude of the complex number e(t), i.e., the square error
over the interval [a, b] is given by

Z b
e∗ (t)e(t)dt
a

Consider the harmonic family ϕk (t) = ejkω0 t , and suppose that we wish to find the best
approximation of a given T -periodic signal x(t) as a linear combination of ϕk (t) for −N ≤
k ≤ N , i.e.,

N
X
x̂(t) = ak ejkω0 t
k=−N

with error

N
X
e(t) = x(t) − x̂(t) = x(t) − ak ejkω0 t
k=−N

We evaluate the squared error over any interval of length T (since the functions ϕk (t) are
orthogonal over such intervals):

Z t0 +T Z t0 +T
Squared Error = 2
|e(t)| dt = e∗ (t)e(t)dt
t0 t0
N
! N
!
Z t0 +T X X
= x∗ (t) − a∗k e−jkω0 t x(t) − ak ejkω0 t dt
t0 k=−N k=−N
N N
!
Z t0 +T X X
= x∗ (t)x(t) − x∗ (t) ak ejkω0 t − x(t) a∗k e−jkω0 t dt
t0 k=−N k=−N
N N
!
Z t0 +T X X
+ a∗k e−jkω0 t ak ejkω0 t dt
t0 k=−N k=−N
N N
!
Z t0 +T Z t0 +T X X
= |x(t)|2 dt + −x∗ (t) ak ejkω0 t − x(t) a∗k e−jkω0 t dt
t0 t0 k=−N k=−N
N N
!
Z t0 +T X X
+ a∗k an e−jkω0 t ejnω0 t dt
t0 k=−N k=−N
Fourier Series Representation of Periodic Signals 61

Z t0 +T N
X Z t0 +T N
X Z t0 +T
∗
= 2
|x(t)| dt − ak x (t)e jkω0 t
dt − a∗k x(t)e−jkω0 t dt
t0 k=−N t0 k=−N t0

N N
!
X X Z t0 +T
+ a∗k an e−jkω0 t ejnω0 t dt
k=−N n=−N t0
Z t0 +T N
X Z t0 +T N
X Z t0 +T
∗
= 2
|x(t)| dt − ak x (t)e jkω0 t
dt − a∗k x(t)e−jkω0 t dt
t0 k=−N t0 k=−N t0

N
X
+T |ak |2 ,
k=−N
R t +T
where we used the fact that t00 e−jkω0 t ejnω0 t dt = 0 if k ̸= n and T otherwise.
Our job is to find the best coefficients ak , −N ≤ k ≤ N to minimize the square error. Thus,
we first write ak = bk +jck , where bk , ck ∈ R, and then differentiate the above expression with
respect to bk and ck and set the result to zero. After some algebra, we obtain the optimal
coefficient as

1 t0 +T
Z
ak = bk + jck = x(t)e−jkω0 t dt
T t0
which is exactly the same expression we found for the Fourier series coefficients earlier.
Note again why we bothered to go through this exercise. Here, we did not assume that a signal
x(t) had a Fourier series representation; we simply asked how to best approximate a given
signal by a linear combination of complex exponentials, and found the resulting coefficients.
These coefficients match exactly the coefficients that we obtained by assuming that the signal
had a Fourier series representation, and lends some justification for the validity of the earlier
analysis.
As N gets larger, the approximation error will get smaller and smaller. The question is
R t +T
then: will t00 |e(t)|2 dt go to zero as N goes to infinity? If so, then the signal would, in
fact, have a Fourier series representation (in the sense of having asymptotically zero error
between the true signal and the approximation). It turns out that most periodic signals of
practical interest will satisfy this property.

When Will a Periodic Signal Have a Fourier Series Representation?

There are various different sufficient conditions that guarantee that a given signal x(t) will
have a Fourier series representation. One commonly used set of conditions are known as the
Dirichlet conditions stated as follows.
A periodic signal x(t) has a Fourier series representation if all three of the following
conditions are satisfied:

• The signal is absolutely integrable over one period:

Z t0 +T
|x(t)|dt < ∞
t0
62 Fourier Series Representation of Periodic Signals

• In any finite interval of time x(t) has bounded variation, meaning that it has only a
finite number of minima and maxima during any single period of the signal.

• In any finite interval of time, there are only a finite number of discontinuities, and each
of these discontinuities are finite.

We won’t go into the proof of why these conditions are sufficient here, but it suffices to
note that signals that violate the above conditions (the last two in particular) are somewhat
pathological. The first condition guarantees that the Fourier series coefficients are finite,
since R t +T R t +T R t +T
|ak | = T1 t00 x(t)e−jkω0 t dt ≤ T1 t00 x(t)e−jkω0 t dt = T1 t00 |x(t)|dt.
Thus if the signal is absolutely integrable over a period, then |ak | < ∞ for all k ∈ Z.

Gibbs Phenomenon
If x(t) has a Fourier series representation, then that representation will exactly equal x(t)
at all points t where x(t) is continuous. At points of discontinuity in x(t), the value of the
Fourier series representation will be equal to the average of the values of x(t) on either side of
the discontinuity. One particularly interesting phenomenon occurs at points of discontinuity:
the Fourier series typically overshoots the signal x(t). The height of the overshoot stays
constant as the number of terms N in the approximation increases, but the width shrinks.
Thus, asymptotically, the error goes to zero (although technically the two signals are not
exactly the same at the discontinuity).

5 Properties of Continuous-Time Fourier Series

We will now derive some useful properties of the Fourier series coefficients. Throughout we
will use the notation
FS
x(t) ←→ ak
to denote that the T -periodic signal x(t) has Fourier series coefficients ak , k ∈ Z.

5.1 Linearity
Suppose we have two signals x1 (t) and x2 (t), each of which is periodic with period T . Let
FS FS
x1 (t) ←→ ak , x2 (t) ←→ bk .
For any complex scalars α, β, let g(t) = αx1 (t) + βx2 (t). Then
FS
g(t) ←→ αak + βbk .
The above property follows immediately from the definition of the Fourier series coefficients
(since integration is linear).
Fourier Series Representation of Periodic Signals 63

Example :
Consider x1 (t) = cos (ω0 t) and x2 (t) = sin (ω0 t). We have
α jω0 t α −jω0 t β β
g(t) = α cos (ω0 t) + β sin (ω0 t) = e + e + ejω0 t − e−jω0 t
2 2 2j 2j
   
α β α β
= + ejω0 t + − e−jω0 t
2 2j 2 2j
Thus, we see that each Fourier series coefficient of g(t) is indeed given by a linear
combination of the corresponding Fourier series coefficients of x1 (t) and x2 (t). ■

5.2 Time Shifting

Define g(t) = x(t − τ ), where τ ∈ R is some delay. Let the Fourier series coefficients of g(t)
be given by bk , k ∈ Z. Then
Z t0 +T Z t0 +T
1 −jkω0 t 1
bk = g(t)e dt = x(t − τ )e−jkω0 t dt
T t0 T t0

Define t̄ = t − τ , so that dt dt̄

= 1. Then we have
t0 −τ +T R t −τ +T
1 −jkω0 (t̄+τ )
dt̄ = e−jkω0 τ T1 t00−τ x(t̄)e−jkω0 t̄ dt̄ = e−jkω0 τ ak .
R
bk = T t0 −τ x(t̄)e
Thus
FS
x(t − τ ) ←→ e−jkω0 τ ak .
Example :
Consider x(t) = cos (ω0 t), and let g(t) = x(t − τ ) We have
g(t) = cos (ω0 (t − τ )) = 12 ejω0 (t−τ ) + 21 e−jω0 (t−τ ) = 12 e−jω0 τ ejω0 t + 21 ejω0 τ e−jω0 t .
Thus, we see that the coefficient of ejω0 t is 21 e−jω0 τ , and the coefficient of e−jω0 t is 12 ejω0 τ ,
as predicted by the expressions derived above. ■

5.3 Time Reversal

Define y(t) = g(−t). Let the Fourier series coefficients of g(t) be given by bk , k ∈ Z. Note
that
∞
X ∞
X
−jkω0 t
g(t) = x(−t) = ak e = a−k ejkω0 t
k=−∞ k=−∞

Thus we have
FS
x(−t) ←→ a−k ,
i.e., the Fourier series coefficients for a time-reversed signal are just the timereversal of the
Fourier series coefficients for the original signal. Note that this is not true for the output of
LTI systems (a time reversal of the input to an LTI system does not necessarily mean that
the output is a time-reversal of the original output).
64 Fourier Series Representation of Periodic Signals

Example :
Consider x(t) = sin (ω0 t) and define g(t) = x(−t). First, note that
1 jω0 t 1
x(t) = sin (ω0 t) = e − e−jω0 t = a1 ejω0 t + a−1 e−jω0 t
2j 2j
We have
1 −jω0 t 1
g(t) = sin (−ω0 t) = e − ejω0 t = b1 ejω0 t + b−1 e−jω0 t
2j 2j
and thus we see that b1 = − 2j1 = a−1 and b−1 = 1
2j
= a1 . ■

5.4 Time Scaling

Consider the signal g(t) = x(αt), where α ∈ R>0 . Thus, g(t) is a time-scaled version of x(t).
Note that the period of g(t) is T2 , where T is the period of x(t). We have
∞
X
g(t) = x(αt) = ak ejkω0 αt
k=−∞

Thus, g(t) has the same Fourier series coefficients as x(t), but the Fourier series repre-
sentation has changed: the frequency is now ω0 α rather than ω0 , to reflect the fact that the
harmonic family is in terms of the new period Tα rather than T .
Example :
Consider x(t) = cos (ω0 t), which has series representation
1 1
x(t) = ejω0 t + e−jω0 t .
2 2
Then we have
1 1
g(t) = x(αt) = cos (ω0 αt) = ejω0 αt + e−jω0 αt .
2 2
■

5.5 Multiplication
Let x1 (t) and x2 (t) be T -periodic signals with Fourier series ak and bk respectively. Consider
the signal g(t) = x1 (t)x2 (t), and note that g(t) is also T -periodic. We have
∞
X ∞
X ∞
X ∞
X
jlω0 t jnω0 t
g(t) = x1 (t)x2 (t) = al e bn e = al bn ej(l+n)ω0 t .
l=−∞ n=−∞ l=−∞ n=−∞

Define k = l + n, so that
∞ ∞ ∞ ∞
!
X X X X
g(t) = al bk−l ejlω0 t = al bk−l ejkω0 t
l=−∞ l=−∞ k=−∞ l=−∞

Thus the Fourier series coefficients of g(t) are given by

Fourier Series Representation of Periodic Signals 65

∞
FS
X
g(t) = x1 (t)x2 (t) ←→ al bk−l .
l=−∞

In other words, the Fourier series coefficients of a product of two signals are given by the
convolution of the corresponding Fourier series coefficients.
Example :
Consider the signals x1 (t) = cos (ω0 t) and x2 (t) = sin (ω0 t). Define g(t) = x1 (t)x2 (t). The
Fourier series representations of x1 (t) and x2 (t) are given by
1 1 1 jω0 t 1
x1 (t) = ejω0 t + e−jω0 t , x2 (t) = e − e−jω0 t .
2 2 2j 2j
Denote the Fourier series coefficients of x1 (t) by the sequence ak , with a−1 = a1 = 12 ,
and ak = 0 otherwise. Similarly, denote the Fourier series coefficients of x2 (t) by the
sequence bk , with b−1 = − 2j1 , b(1) = 2j1 , and bk = 0 otherwise. Denote the Fourier series
coefficients of g(t) by ck , k ∈ Z. Then we have
∞
X
c k = ak ∗ b k = al bk−l .
l=−∞

Convolving the two sequences given above, we see that c−2 = − 4j1 , c2 = 1
4j
, and ck = 0
otherwise. Thus
1 j2ω0 t 1
g(t) = x1 (t)x2 (t) = cos (ω0 t) sin (ω0 t) = e − e−j2ω0 t
4j 4j
Noting that cos (ω0 t) sin (ω0 t) = 1
2
sin (2ω0 t), we see that the above expression is, in
fact, correct. ■

5.6 Parseval’s Theorem

Consider a T -periodic signal x(t), and let ak , k ∈ Z be its Fourier series coefficients. Now let
us consider the average power of x(t) over one period, defined as
Z t0 +T
1
|x(t)|2 dt
T t0

Substituting |x(t)|2 = x∗ (t)x(t) and the Fourier series for x(t) we have

∞
! ∞
!
Z t0 +T Z t0 +T
1 1 X X
|x(t)|2 dt = a∗k e−jkω0 t a∗n ejnω0 t dt
T t0 T t0 k=−∞ n=−∞
∞ ∞ Z t0 +T
X X 1
= a∗k an e−jkω0 t ejnω0 t dt
k=−∞ n=−∞
T t0

Using the fact that ejkω0 t and ejnω0 t are orthogonal for k ̸= n, we have
66 Fourier Series Representation of Periodic Signals

Z t0 +T ∞ ∞
1 X X
2
|x(t)| dt = a∗k ak = |ak |2
T t0 k=−∞ k=−∞

This leads to Parseval’s Theorem: for a T -periodic signal x(t) with Fourier series coeffi-
cients ak , k ∈ Z, we have
Z t0 +T ∞
1 X
|x(t)|2 dt = |ak |2
T t0 k=−∞

Example :
Consider the signal x(t) = cos (ω0 t), with Fourier coefficients a1 = a−1 = 1
2
and ak = 0
otherwise. We have

Z t0 +T Z t0 +T Z t0 +T
1 2 1 2 1 1 1
|x(t)| dt = cos (ω0 t) dt = (1 + cos (2ω0 t)) dt =
T t0 T t0 T t0 2 2

We also have
∞
X 1 1 1
|ak |2 = a21 + a2−1 = + =
k=−∞
4 4 2
which agrees with the direct calculation of average power, as indicated by Parseval’s
Theorem. ■

6 Fourier Series for Discrete-Time Periodic Signals

We now turn our attention to the discrete-time Fourier series. Specifically, consider the
discrete-time signal x[n], n ∈ Z, and suppose it is N -periodic. Let
ω0 = 2πN
, and note that the signal ϕk [n] = ejkω0 n is also N -periodic for any k ∈ Z. Thus, as
in the case for continuous-time signals, we would like to write x[n] as a linear combination
of signals from the harmonic family ϕk [n], k ∈ Z, i.e.,
∞ ∞
2π
X X
jkω0 n
x[n] = ak e = ak ejk N n
k=−∞ k=−∞

At this point, we encounter the first main difference between the discrete-time and continuous-
time Fourier series. Recall that a discrete-time complex exponential with frequency ω0 is the
same as a discrete-time complex exponential with frequency ω0 + 2π. Specifically, for any
k ∈ Z, consider

ϕk+N [n] = ej(k+N )ω0 n = ejkω0 n ejN ω0 n = ejkω0 n

since N ω0 = 2π. Thus, there are only N different complex exponentials in the discrete-
time harmonic family for the fundamental frequency ω0 , and so we have the following.
Fourier Series Representation of Periodic Signals 67

The discrete-time Fourier series for an N -periodic signal

x[n] is given by
+N −1
n0X +N −1
n0X
2π
jkω0 n
x[n] = ak e = ak ejk N n
k=n0 k=n0

where n0 is any integer.

In other words, the discrete-time signal can be written in terms of any N contiguous
multiples of the fundamental frequency ω0 . Since there are only N coefficients ak in this
representation, and since it does not matter which contiguous N members of the harmonic
family we choose, the Fourier series coefficients are N -periodic as well, i.e., ak = ak+N for all
k ∈ Z.

6.1 Finding the Discrete-Time Fourier Series Coefficients

To find the Fourier series coefficients in (4.2), we use a similar trick as in the continuous-time
case. Specifically, first multiply both sides of (4.2) by e−jrω0 n , where r is any integer, and
sum both sides over N terms. This gives
+N −1
n1X +N −1 n0X
n1X +N −1
−jrω0 n
x[n]e = ak ej(k−r)ω0 n
n=n1 n=n1 k=n0

where n1 is any integer. Interchanging the summations, we have

+N −1
n1X +N −1
n0X +N −1
n1X
−jrω0 n
x[n]e = ak ej(k−r)ω0 n
n=n1 k=n0 n=n1

Now suppose that k − r is a multiple of N . In this case we obtain

+N −1
n1X +N −1
n1X
ej(k−r)ω0 n = 1=N
n=n1 n=n1

On the other hand, if r − k is not a multiple of N , we use the finite sum formula to obtain
+N −1
n1X
ej(k−r)ω0 n1 − ej(k−r)ω0 (n1 +N )
ej(k−r)ω0 n = =0
n=n1
ej(k−r)ω0 − 1

Thus, we have the following.

The discrete-time Fourier series coefficients are given by
+N −1
n1X
1
ak = x[n]e−jkω0 n
N n=n1

where n1 is any integer. The Fourier series coefficients are

N -periodic, i.e., ak = ak+N .
68 Fourier Series Representation of Periodic Signals

Example 4.11. Consider the N -periodic signal x[n], which is equal to 1 for −N1 ≤ n ≤ N1
and zero otherwise (modulo the periodic constraints).
We have
N1
1 X
ak = e−jkω0 n
N n=−N
1

If k = 0, we have a0 = 2N1 +1
N
. For k ∈ {1, 2, . . . , N − 1}, we have (via the finite sum
formula)
1 1
1 e−jkω0 2 ejkω0 (N1 + 2 ) − ejkω0 (N1 + 2 )
1
1 ejkω0 N1 − ejkω0 (N1 +1)
ak = =
N 1 − e−jkω0 N e−jkω0 12 1
ejkω0


2 − e
−jkω0 21

1 sin kω0 N1 + 21
=
sin k ω20


N
It is of interest to note that we do not have to worry about convergence conditions for
discrete-time Fourier series, as we did in the continuous-time case. Specifically, for an N -
periodic discrete-time signal x[n], we only require N numbers to completely specify the
entire signal. The Fourier series coefficients ak , k ∈ {0, 1, . . . , N − 1} thus contain as much
information as the signal itself, and form a perfect representation of the signal. In other
words, for discrete-time signals, the Fourier series representation is just a transformation of
the signal into another form; we do not encounter discrepancies like the Gibbs phenomenon
in discrete-time.

4.5.2 Properties of the Discrete-Time Fourier Series

The discrete-time Fourier series has properties that can be derived in almost the same way
as the properties for the continuous-time Fourier series (linearity, time-shifting, etc.). Here
we will just discuss the multiplication property and Parseval’s theorem.

Multiplication of Discrete-Time Periodic Signals

First, let’s start with an example. Consider two 2-periodic signals x1 [n] and x2 [n]. We
know that the Fourier series respresentations can be uniquely specified by two coefficients.
Specifically,

x1 [n] = a0 + a1 ejω0 n , x2 [n] = b0 + b1 ejω0 n

Now consider the product

g[n] = x1 [n]x2 [n] = a0 + a1 ejω0 n b0 + b1 ejω0 n

= a0 b0 + (a0 b1 + a1 b0 ) ejω0 n + a1 b1 ej2ω0 n

Now, note that ω0 = 2π
2
, and thus ej2ω0 n = 1. This gives

g[n] = (a0 b0 + a1 b1 ) + (a0 b1 + a1 b0 ) ejω0 n

Fourier Series Representation of Periodic Signals 69

Thus, the Fourier series coefficients of g[n] are given by c0 = a0 b0 +a1 b1 and c1 = a0 b1 +a1 b0 .
We can write these in a unform way as follows:

1
X
c0 = a0 b0 + a1 b1 = a0 b0 + a1 b−1 = al b−l
l=0
1
X
c 1 = a0 b 1 + a1 b 0 = al b1−l
l=0

where we used the fact that b1 = b−1 by the periodic nature of the discrete-time Fourier
series coefficients. The above expressions show that the Fourier series coefficients of the
product of the two signals are given by a form of convolution of the coefficients of those
signals; however, the convolution is over a finite number of terms, as opposed to over all
time-indices.
Let us generalize this to functions with a larger period. Let x1 [n] and x2 [n] be two N -periodic
discrete-time signals, with discrete-time Fourier series coefficients ak and bk , respectively. We
have

N
X −1 N
X −1
jkω0 n
x1 [n] = ak e , x2 [n] = bk ejkω0 n
k=0 k=0

Define the signal g[n] = x1 [n]x2 [n]. We have

−1
N
! N −1
!
X X
g[n] = x1 [n]x2 [n] = al ejlω0 n br ejrω0 n
l=0 r=0
N
XN−1 X −1
= al br ej(l+r)ω0 n
l=0 r=0

where we have used l and r as the indices in the Fourier series in order to keep the terms
in the two sums distinct.
Define the new variable k = l + r. Substituting into the above expression, this gives
70 Fourier Series Representation of Periodic Signals

N
X −1 l+N
X−1
g[n] = al bk−l ejkω0 n
l=0 k=l
−1 −1 X−1
N N l+N
!
X X
jkω0 n jkω0 n
= al bk−l e + al bk−l e
l=0 k=l k=N
−1 −1
N N l−1
!
X X X
= al bk−l ejkω0 n + al bk+N −l ej(k+N )ω0 n
l=0 k=l k=0
−1 −1
N N l−1
!
X X X
= al bk−l ejkω0 n + al bk−l ejkω0 n
l=0 k=l k=0
N
X −1 N
X −1
= al bk−l ejkω0 n
l=0 k=0
−1 N −1
N
!
X X
= al bk−l ejkω0 n
k=0 l=0

Thus, the Fourier series coefficients of g[n] are given by

N
X −1
ck = al bk−l
l=0

for 0 ≤ k ≤ N − 1. The above convolution is known as the periodic convolution of two

periodic signals; for any k, there are only N terms in the summation. We saw an example
of this with the 2-periodic signals that we had above. In essence, the above calucations are
simply multiplying together the discrete-time Fourier series representations of the two signals
(each of which has N terms), and then using the periodicity of the discrete-time complex
exponentials in the frequencies to combine terms together. Note that we can actually perform
the sum over any continguous N values of l, since all of the signals involved are periodic.

Parseval’s Theorem for Discrete-Time Signals

Let x[n] be an N -periodic discrete-time signal, with Fourier series coefficients ak , 0 ≤ k ≤
N − 1. The average power of x[n] over one period is

+N −1
n0X
1
|x[n]|2
N n=n0

where n0 is any integer. Parseval’s theorem for discrete-time signals states the following.

+N −1
n0X N −1
1 X
|x[n]|2 = |ak |2
N n=n0 k=0
Fourier Series Representation of Periodic Signals 71

The following example illustrates the application of the various facts that we have seen
about the discrete-time Fourier series.
Example 4.12. Suppose we are told the following facts about a discrete-time signal x[n].

• x[n] is periodic with period N = 6.

•
P5
n=0 x[n] = 2.

•
P7 n
n=2 (−1) x[n] = 1.

• x[n] has the minimum power per period of all signals satisfying the preceding three
conditions.

The above facts are sufficient for us to uniquely determine x[n]. First, note that the
Fourier series representation of x[n] is
5 5
π
X X
jkω0 n
x[n] = ak e = ak ejk 3 n ,
k=0 k=0

where we used the fact that the signal is 6 -periodic.

From the second fact, we have
N −1
1 X 2 1
a0 = x[n] = = .
N k=0 6 3
π
To use the third fact, note that (−1)n = e−jπn = e−j3 3 n . Thus, the third fact seems to be
related to the Fourier series coefficient for k = 3. Specifically, we have
7
1 X 1
a3 = x[n]e−jπn =
N k=2 6
To use the last fact, note from Parseval’s theorem that the power of the signal over one
period is given by
5 5
1 X X
2
|x[n]| = |ak |2
N n=0 k=0

= |a0 |2 + |a1 |2 + |a2 |2 + |a3 |2 + |a4 |2 + |a5 |2

We are told that x[n] has the minimum average power over all signals that satisfy the other
three conditions. Since the other three conditions have already set a0 and a1 , the average
power is given by setting all of the other Fourier series coefficients to 0 . Thus, we have
5
X π 1 1
x[n] = ak ejk 3 n = a0 + a3 ejπn = + (−1)n
k=0
3 6
72 Fourier Series Representation of Periodic Signals
 CHAPTER 5
THE CONTINUOUS-TIME FOURIER
TRANSFORM

In the last part of the course, we decomposed

1 The Fourier Transform . . . . 74
periodic signals into superpositions of com-
2 Fourier Transform of Periodic
plex exponentials, where each complex expo-
Signals . . . . . . . . . . . . . 77
nential is a member of the harmonic family
3 Properties of the Continuous-
corresponding to the fundamental period of
Time Fourier Transform . . . 78
the signal. We now turn our attention to the
case where the signal of interest is not peri-
odic. As we will see, the main idea will be to
view an aperiodic signal as a periodic signal
whose period goes to ∞.

73
74 The Continuous-Time Fourier Transform

1 The Fourier Transform

Suppose that we are given a signal x(t) that is aperiodic. As a concrete example, suppose
that x(t) is a solitary square pulse, with x(t) = 1 if −T1 ≤ t ≤ T1 , and zero elsewhere.
Clearly x(t) is not periodic.
Now define a new signal x̃(t) which is a periodic extension of x(t) with period T . In other
words, x̃(t) is obtained by repeating x(t), where each copy is shifted T units in time. This
x̃(t) has a Fourier series representation, which we found in the last chapter to be

2T1 2 sin (kω0 T1 )

a0 = , ak =
T kω0 T
for k ∈ Z.
Now recall that the Fourier series coefficients are calculated as follows:
Z T
1 2
ak = x̃(t)e−jkω0 t dt
T − T2

However, we note that x(t) = x̃(t) in the interval of integration, and thus
Z T
1 2
ak = x(t)e−jkω0 t dt
T − T2

Furthermore, since x(t) is zero for all t outside the interval of integration, we can expand
the limits of the integral to obtain

1 ∞
Z
ak = x(t)e−jkω0 t dt
T −∞
Let us define
Z ∞
X(jω) = x(t)e−jωt dt
−∞

This is called the Fourier transform of the signal x(t), and the Fourier series coefficients
can be viewed as samples of the Fourier transform, scaled by T1 , i.e.,
1
ak = X (jkω0 ) , k ∈ Z
T
Now consider the fact that
∞ ∞
X
jkω0 t 1 X
x̃(t) = ak e = X (jkω0 ) ejkω0 t
k=−∞
T k=−∞

Since ω0 = 2π
T
, this becomes
∞
1 X
x̃(t) = X (jkω0 ) ejkω0 t ω0
2π k=−∞
The Continuous-Time Fourier Transform 75

Now consider what happens as the period T gets bigger. In this case, x̃(t) approaches
x(t), and so the above expression becomes a representation of x(t). As T → ∞, we have
ω0 → 0. Since each term in the summand can be viewed as the area of the rectangle whose
height is X (jkω0 ) ejkω0 t and whose base goes from kω0 to (k + 1)ω0 , we see that as ω0 → 0,
the sum on the right hand side approaches the area underneath the curve X(jω)ejωt (where
t is held fixed). Thus, as T → ∞ we have
Z ∞
1
x(t) = X(jω)ejωt dω
2π −∞
Thus we have the following.
Given a continuous-time signal x(t), the Fourier Transform of the signal is given by
Z ∞
X(jω) = x(t)e−jωt dt
−∞
The Inverse Fourier Transform of the signal is given by
Z ∞
1
x(t) = X(jω)ejωt dω
2π −∞
The Fourier transform X(jω) is also called the spectrum of the signal, as it represents the
contribution of the complex exponential of frequency ω to the signal x(t).
Example :
Consider the signal x(t) = e−at u(t), a ∈ R>0 . The Fourier transform of this signal is
Z ∞ Z ∞
−jωt
X(jω) = x(t)e dt = e−at e−jωt dt
−∞ 0
∞
1
=− e−(a+jω)t
a + jω 0
1
= .
a + jω
To visualize X(jω), we plot its magnitude and phase on separate plots (since X(jω) is
complex-valued in general). We have
1 −1 ω
 
|X(jω)| = √ , ∠X(jω) = − tan .
a2 + ω 2 a
The plots of these quantities are show in Fig. 4.5 of the text.
■
Example :
Consider the signal x(t) = δ(t). We have
Z ∞
X(jω) = δ(t)e−jωt dt = 1
−∞

In other words, the spectrum of the impulse function has an equal contribution at all
frequencies. ■
76 The Continuous-Time Fourier Transform

Example :
Consider the signal x(t) which is equal to 1 for −T1 ≤ t ≤ T1 and zero elsewhere. We have
Z T1
1 jωT1
2 sin (ωT1 )
X(jω) = e−jωt dt = e − e−jωT1 = .
−T1 jω ω
■
Example :
Consider the signal whose Fourier transform is
(
1, |ω| ≤ W
X(jω) =
0 |ω| > W
We have
Z ∞ Z W
1 jωt 1
x(t) = X(jω)e dω = ejωt dω
2π −∞ 2π −W
W
1 1 jωt
e =
2π jt −W
sin(W t)
=
πt
The previous two examples showed the following. When x(t) is a square pulse, then
X(jω) = 2 sin(ωT
ω
1)
and when X(jω) is a square pulse, x(t) = sin(Wπt
t)
. This is an example of
the duality property of Fourier transforms, which we will see later.
Functions of the form sin(W
πt
t)
will show up frequently, and are called sinc functions. Specif-
ically

sin(πθ)
sinc(θ) =
πθ
Thus 2 sin(ωT1 ) ωT1
and sin(W t) W Wt
.



ω
= 2T1 sinc π πt
= π
sinc π
■

1.1 Existence of Fourier Transform

Just as we saw with the Fourier series for periodic signals, there are some rather mild con-
ditions under which a signal x(t) is guaranteed to have a Fourier transform (such that the
inverse Fourier transform converges to the true signal). Specifically, there are a set of suffi-
cient conditions (also called Dirichlet conditions) under which a continuous-time signal x(t)
is guaranteed to have a Fourier transform:
R∞
1. x(t) is absolutely integrable: −∞
|x(t)|dt < ∞.

2. x(t) has a finite number of maxima and minima in any finite interval.

3. x(t) has a finite number of discontinuities in any finite interval, and each of these
discontinuities is finite.
The Continuous-Time Fourier Transform 77

If all of the above conditions are satisfied, x(t) is guaranteed to have a Fourier transform.
Note that this only a sufficient set of conditions, and not necessary.
An alternate sufficient condition is that the signal have finite energy (i.e., that it be square
integrable):
Z ∞
|x(t)|2 dt < ∞
−∞

For example, the signal x(t) = 1

t
is square integrable, but not absolutely integrable.
u(t−1)
Thus the finite energy condition guarantees that x(t) will have a Fourier transform, whereas
the Dirichlet conditions do not apply.

2 Fourier Transform of Periodic Signals

The Fourier transform can also be applied to certain periodic signals (although such signals
will not be absolutely integrable or square integrable over the entire time-axis). A direct
application of the Fourier transform equation to such signals will not necessarily yield a
meaningful answer, due to the fact that periodic signals do not die out. Instead, we will
work backwards by starting with a frequency domain signal and doing an inverse Fourier
transform to see what pairs arise.
Thus, consider the signal x(t) whose Fourier transform is

X(jω) = 2πδ (ω − ω0 )
i.e., the frequency domain signal is a single impulse at ω = ω0 , with area 2π. Using the
inverse Fourier transform, we obtain
Z ∞ Z ∞
1 jωt
x(t) = X(jω)e dω = δ (ω − ω0 ) ejωt dω
2π −∞ −∞
jω0 t
=e
Thus, the Fourier transform of x(t) = ejω0 t is X(jω) = 2πδ (ω − ω0 ). Similarly, if
∞
X
X(jω) = ak 2πδ (ω − kω0 )
k=−∞

then an application of the inverse Fourier transform gives

∞
X
x(t) = ak ejkω0 t
k=−∞

In other words, if x(t) is a periodic signal with Fourier series coefficients ak , then the Fourier
transform of x(t) consists of a sequence of impulse functions, each spaced at multiples of ω0 ;
the area of the impulse at kω0 will be 2πak .
Example :
Consider the signal x(t) = cos (ω0 t). The Fourier series coefficients are a1 = a−1 = 12 .
Thus, the Fourier transform of this signal is given by
78 The Continuous-Time Fourier Transform

X(jω) = a1 2πδ (ω − ω0 ) + a−1 2πδ (ω + ω0 ) = πδ (ω − ω0 ) + πδ (ω + ω0 )

■
Example :
Consider the periodic signal
∞
X
x(t) = δ(t − nT )
n=−∞

The Fourier series coefficients for this signal are given by

Z T ∞
1 2 X 1
a0 = δ(t − nT )dt =
T − T2 n=−∞ T
Z T X ∞
1 2 1
ak = δ(t − nT )e−jkω0 t dt =
T − T2 n=−∞ T
Thus, the Fourier transform of this signal is given by
∞ ∞  
X 2π X 2kπ
X(jω) = ak 2πδ (ω − kω0 ) = δ ω− .
k=−∞
T k=−∞ T
■
Thus, if x(t) is an impulse train with period T , its Fourier transform is also an impulse
train in the frequency domain, except with period 2π T
. Once again, we see that if T increases
(i.e., the period increases in the time-domain) we obtain a time-shrinking in the frequency
domain.

3 Properties of the Continuous-Time Fourier Transform

We will now discuss various properties of the Fourier transform. As with the Fourier series,
we will find it useful to introduce the following notation. Suppose x(t) is a time-domain
signal, and X(jω) is its Fourier transform. We then say

X(jω) = F{x(t)}
x(t) = F −1 {X(jω)}
We will also use the notation
F
x(t) ←→ X(jω)
to indicate that x(t) and X(jω) are Fourier transform pairs.

3.1 Linearity
The first property of Fourier transforms is easy to show:
The Continuous-Time Fourier Transform 79

F {αx1 (t) + βx2 (t)} = αF {x1 (t)} + βF {x2 (t)}

which follows immediately from the definition of the Fourier transform.

3.2 Time-Shifting
Suppose x(t) is a signal with Fourier transform X(jω). Define g(t) = x(t − τ ) where τ ∈ R.
Then we have
Z ∞ Z ∞
−jωt
G(jω) = g(t)e dt = x(t − τ )e−jωt dt = e−jωτ X(jω)
−∞ −∞
Thus

F{x(t − τ )} = e−jωτ X(jω)

Note the implication: if we time-shift a signal, the magnitude of its Fourier transform is
not affected. Only the phase of the Fourier transform gets shifted by −ωτ at each frequency
ω.

3.3 Conjugation
Consider a signal x(t). We have
Z ∞ ∗ Z ∞
∗ −jωt
X (jω) = x(t)e dt = x∗ (t)ejωt dt
−∞ −∞
Thus,
Z ∞
∗
X (−jω) = x∗ (t)e−jωt dt = F {x∗ (t)}
−∞
The above is true for any signal x(t) that has a Fourier transform. Now suppose addition-
ally that x(t) is a real-valued signal. Then we have x∗ (t) = x(t) for all t ∈ R. Thus

X ∗ (−jω) = F {x∗ (t)} = F{x(t)} = X(jω)

Based on the above relationship between X ∗ (−jω) and X(jω) for real-valued signals, we
see the following. Write X(jω) in polar form as

X(jω) = |X(jω)|ej∠X(jω)
Then we have

X(−jω) = X ∗ (jω) = |X(jω)|e−j∠X(jω)

Thus, for any ω, X(−jω) has the same magnitude as X(jω), and the phase of X(−jω) is
the negative of the phase of X(jω). Thus, when plotting X(jω), we only have to plot the
magnitude and phase for positive values of ω, as the plots for negative values of ω can be
easily recovered according to the relationships described above.
80 The Continuous-Time Fourier Transform

Example :
Consider again the signal x(t) = e−at u(t); we saw earlier that the Fourier transform of this
signal is
1
X(jω) =
a + jω
It is easy to verify that
1
X(−jω) = = X ∗ (jω)
a − jω
as predicted. Furthermore, we can see from the plots of the magnitude and phase of
X(jω) that the magnitude is indeed an even function, and the phase is an odd function.
Suppose further that x(t) is even (in addition to being real-valued). Then we have x(t) =
x(−t). Then we have
Z ∞ Z ∞ Z ∞
X(−jω) = jωt
x(t)e dt = jωt
x(−t)e dt = x(t)e−jωt dt
−∞ −∞ −∞
= X(jω)
This, together with the fact that X(−jω) = X ∗ (jω) for real-valued signals indicates
that X(jω) is real-valued and even.
Similarly, if x(t) is real-valued and odd, we have X(jω) is purely imaginary and odd.
■
Example :
Consider the signal x(t) = e−a|t| , where a is a positive real number. This signal is real-
valued and even. We have
Z ∞ Z 0 Z ∞
−jωt at −jωt
X(jω) = x(t)e dt = e e dt + e−at e−jωt dt
−∞ −∞ 0
1 1
= +
a − jω a + jω
2a
= 2 .
a + ω2
As predicted, X(jω) is real-valued and even. ■

3.4 Differentiation
Consider the inverse Fourier transform
Z ∞
1
x(t) = X(jω)ejωt dω
2π −∞

Differentiating both sides with respect to t, we obtain

Z ∞
dx(t) 1
= jωX(jω)ejωt dω
dt 2π −∞
The Continuous-Time Fourier Transform 81

Thus, we see that

 
dx(t)
F = jωX(jω)
dt

3.5 Time and Frequency Scaling

Let a be a nonzero real number and consider the signal g(t) = x(at) (i.e., a time-scaling of
x(t)). We have
Z ∞
F{g(t)} = x(at)e−jωt dt
−∞

If we perform the substitution τ = at, we have

( R∞ ω
1
a −∞
x(τ )e−j a τ dτ, a>0
F{g(t)} = 1 ∞
R −j ω τ
− a −∞ x(τ )e a dτ, a < 0
This can be written in a uniform way as

1  ω
F{x(at)} = X j .
|a| a
Thus we see again that shrinking a signal in the time-domain corresponds to expanding it
in the frequency domain, and vice versa.

3.6 Duality
We have already seen a few examples of the duality property: suppose x(t) has Fourier
transform X(jω). Then if we have a time-domain signal that has the same form as X(jω),
the Fourier transform of that signal will have the same form as x(t). For example, the square
pulse in the time-domain had a Fourier transform in the form of a sinc function, and a sinc
function in the time-domain had a Fourier transform in the form of a square pulse.
We can consider another example. Suppose x(t) = e−|t| . Then one can verify that

2
F e−|t| =

1 + ω2
Specifically we have
Z ∞
−|t| 1 2
e = e−jωt dω
2π −∞ 1 + ω2
If we multiply both sides by 2π and interchange ω and t, we obtain
Z ∞
−|ω| 2
2πe = 2
e−jωt dt
−∞ 1 + t

Thus, we have
82 The Continuous-Time Fourier Transform

 
2
F = 2πe−|ω|
1 + t2
Duality also applies to properties of the Fourier transform. For example, recall that differ-
entiation in the time-domain corresponds to multiplication by jω in the frequency domain.
We will now see that differentiation the frequency domain corresponds to multiplication by
a certain quantity of t in the time-domain. We have
Z ∞
dX(jω)
= x(t)(−jt)e−jωt dt
dω −∞

Thus, differentiation in the frequency domain corresponds to multiplication by −jt in the

time-domain.

3.7 Parseval’s Theorem

Just as with periodic signals, we have the following.
Z ∞ Z ∞
2 1
|x(t)| dt = |X(jω)|2 dω
−∞ 2π −∞

To derive this, note that

Z ∞ Z ∞ Z ∞ Z ∞
∗ 1
2
|x(t)| dt = x(t)x (t)dt = x(t) X ∗ (jω)e−jωτ dωdt
−∞ −∞ 2π −∞ −∞
Z ∞ Z ∞
1 ∗
= X (jω) x(t)e−jωt dtdω
2π −∞ −∞
Z ∞
1
= X ∗ (jω)X(jω)dω
2π −∞
Z ∞
1
= |X(jω)|2 dω
2π −∞

3.8 Convolution
Consider a signal x(t) with Fourier transform X(jω). From the inverse Fourier transform,
we have
Z ∞
1
x(t) = X(jω)ejωt dω
2π −∞
This has the interpretation that x(t) can be written as a superposition of complex expo-
nentials (with frequencies spanning the entire real axis). From earlier in the course, we know
that if the input to an LTI system is ejωt , then the output will be H(jω)ejωt , where
Z ∞
H(jω) = h(t)e−jωt dt
−∞
The Continuous-Time Fourier Transform 83

In other words, H(jω) is the Fourier transform of the impulse response h(t). This, together
with the LTI property of the system, implies that
Z ∞ Z ∞
1 jωt 1
x(t) = X(jω)e dω ⇒ X(jω)H(jω)ejωt dω = y(t).
2π −∞ 2π −∞
Thus, we see that the Fourier transform of the output y(t) is given by

Y (jω) = H(jω)X(jω)
In other words:

The Fourier transform of the output of an LTI system is

given by the product of the Fourier transforms of the input
and the impulse response.
This is potentially the most important fact pertaining to LTI systems and frequency
domain analysis. Let’s derive this another way just to reinforce the fact.
Suppose that we have two signals x(t) and h(t), and define
Z ∞
y(t) = x(t) ∗ h(t) = x(τ )h(t − τ )dτ
−∞

We have
Z ∞ Z ∞ Z ∞ 
−jωt
Y (jω) = y(t)e dt = x(τ )h(t − τ )dτ e−jωt dt
−∞
Z−∞
∞
−∞
Z ∞ 
−jωt
= x(τ ) h(t − τ )e dt dτ
−∞ −∞
Z ∞
= x(τ )e−jωτ H(jω)dτ
−∞
Z ∞
= H(jω) x(τ )e−jωτ dτ
−∞
= H(jω)X(jω)
In the third line, we used the time-shifting property of the Fourier transform. Thus we
see that convolution of two signals in the time-domain corresponds to multiplication of the
signals in the frequency domain, i.e.,

F{x(t) ∗ h(t)} = F{x(t)}F{h(t)}.

One thing to note here pertains to the existence of the Fourier transform of h(t). Specifi-
cally, recall that the LTI system is stable if and only if
Z ∞
|h(t)|dt < ∞
−∞

This is precisely the first condition in the Dirichlet conditions; thus, as long as the system
is stable and the impulse response satisfies the other two conditions (which almost all real
84 The Continuous-Time Fourier Transform

Figure 5.1: A series interconnection of systems.

systems would), the Fourier transform is guaranteed to exist. If the system is unstable, we
will need the machinery of Laplace transforms to analyze the input-output behavior, which
we will defer to a later discussion.
The convolution - multiplication property is also very useful for analysis of interconnected
linear systems. For example, consider the series interconnection shown in Fig. 5.1.
We have

y(t) = y1 (t) ∗ h2 (t) = (x(t) ∗ h1 (t)) ∗ h2 (t) = x(t) ∗ (h1 (t) ∗ h2 (t)) .
Taking Fourier transforms, we obtain

Y (jω) = X(jω)H1 (jω)H2 (jω)

This reinforces what we saw earlier, that the series interconnection of LTI systems can
be lumped together in a single LTI system whose impulse response is the convolution of
the impulse responses of the individual systems. In the frequency domain, their Fourier
transforms get multiplied together.
One of the important implications of the convolution property is that it allows us to
investigate the effect of systems on signals in the frequency domain. For example, this
facilitates the design of appropriate filters for signals, as illustrated in the following example.
Example :
Consider a signal v(t) which represents a measurement of some relatively low frequency
content (such as a voice signal). Suppose that we measure this signal via a microphone,
whose output is given by

x(t) = v(t) + n(t)

where n(t) is high-frequency noise. Note that X(jω) = V (jω) + N (jω). We would like
to take the measured signal x(t) and remove the noise; unfortunately we do not have access
to n(t) to subtract it out. Instead, we can work in the frequency domain. Suppose we
design a filter (an LTI system) whose impulse response as the following Fourier transform:
(
1 |ω| ≤ W
H(jω) =
0 |ω| > W
where W is the highest frequency of the underlying signal v(t). If we feed x(t) into this
filter, the output will have Fourier transform given by

Y (jω) = X(jω)H(jω) = V (jω)H(jω) + N (jω)H(jω)

If all of the frequency content of the noise n(t) occurs at frequencies larger than W ,
then we see that N (jω)H(jω) = 0, and thus
The Continuous-Time Fourier Transform 85

Y (jω) = V (jω)H(jω) = V (jω)

In other words, we have recovered the voice signal v(t) by passing x(t) through the
low-pass filter.
Recall that the inverse Fourier transform of the given H(jω) is

sin(W t)
h(t) =
πt
However, there are various challenges with implementing an LTI system with this im-
pulse response. One is that this is noncausal, and thus one must potentially include a
sufficiently large delay (followed by a truncation of the signal) in order to apply it. An-
other problem is that it contains many oscillations, which may not be desirable for an
impulse response.
Instead of the above filter, suppose consider another filter whose impulse response is

h2 (t) = e−at u(t)

This filter can be readily implemented with an RC circuit (with the input signal being
applied as an input voltage, and the output signal being the voltage across the capacitor).
The Fourier transform of this impulse response is
1
H2 (jω) =
jω + a
The magnitude plot of this Fourier transform has content at all frequencies, and thus
this filter will not completely eliminate all of the high frequency noise. However, by tuning
the value of a, one can adjust how much of the noise affects the filtered signal. Note that
this filter will also introduce phase shifts at different frequencies, which will also cause
some distortion of the recovered signal. ■

3.9 Multiplication
We just saw that multiplication in the time domain corresponds to convolution in the fre-
quency domain. By duality, we obtain that multiplication in the frequency domain corre-
sponds to convolution in the time-domain. Specifically, consider two signals x1 (t) and x2 (t),
and define g(t) = x1 (t)x2 (t). Then we have
Z ∞ Z ∞
−jωt
G(jω) = g(t)e dt = x1 (t)x2 (t)e−jωt dt
−∞ −∞
Z ∞ Z ∞
1
= x2 (t) X1 (jθ)ejθt dθe−jωt dt
2π −∞ −∞
Z ∞ Z ∞
1
= X1 (jθ) x2 (t)e−j(ω−θ)t dtdθ
2π −∞ −∞
Z ∞
1
= X1 (jθ)X2 (j(ω − θ))dθ
2π −∞
Thus,
86 The Continuous-Time Fourier Transform

Z ∞
1 1
F {x1 (t)x2 (t)} = (X1 (jω) ∗ X2 (jω)) = X1 (jθ)X2 (j(ω − θ))dθ
2π 2π −∞

Multiplication of one signal x1 (t) by another signal x2 (t) can be viewed as modulating the
amplitude of one signal by the other. This plays a key role in communication systems.
Example :
Consider a signal s(t) whose frequency spectrum lies in some interval [−W, W ]. Consider
the signal p(t) = cos (ω0 t). The Fourier transform of p(t) is given by

P (jω) = πδ (ω − ω0 ) + πδ (ω + ω0 )
Now consider the signal x(t) = s(t)p(t), with Fourier transform given by
Z ∞
1
X(jω) = S(jθ)P (j(ω − θ))dθ
2π −∞
1 ∞
Z
= S(jθ)δ (ω − θ − ω0 ) dθ
2 −∞
1 ∞
Z
+ S(jθ)δ (ω − θ + ω0 ) dθ
2 −∞
1 1
= S (j (ω − ω0 )) + S (j (ω + ω0 ))
2 2
Thus, multiplying the signal s(t) by p(t) results in a signal x(t) whose frequency spec-
trum consists of two copies of the spectrum of s(t), centered at the frequencies ω0 and
−ω0 and scaled by 21 . ■
The above example illustrates the principle behind amplitude modulation (AM) in commu-
nication and radio systems. A low frequency signal (such as voice) is amplitude modulated to
a higher frequency that is reserved for that signal. It is then transmitted at that frequency to
the receiver. The following example illustrates how the receiver can recover the transmitted
signal.
Example :
Consider the signal x(t) = s(t)p(t) from the previous example. Its frequency spectrum
has two copies of the spectrum of s(t), located at ±ω0 . We want to recover the original
signal s(t) from x(t). To do this, suppose we multiply x(t) by cos (ω0 t) again, to obtain

y(t) = x(t) cos (ω0 t)

As above, we have
1 1
Y (jω) = X (j (ω − ω0 )) + X (j (ω + ω0 ))
2 2
By drawing this, we see that the frequency spectrum of y(t) contains three copies of
the spectrum of s(t) : one copy centered at ω = 0 (and scaled by 21 ), one copy at 2ω0
scaled by 14 , and one copy at −2ω0 , scaled by 41 . Thus, if we want to recover s(t) from
y(t), we simply apply a low pass filter to it (and scale it by 2 ). ■
 CHAPTER 6
THE DISCRETE-TIME FOURIER
TRANSFORM

We now turn our attention to discrete-time

1 The Discrete-Time Fourier
aperiodic signals. We saw that the Fourier
Transform . . . . . . . . . . . 88
transform for continuous-time aperiodic sig-
2 The Fourier Transform of
nals can be obtained by taking the Fourier
Discrete-Time Periodic Signals 90
series of an appropriately defined periodic
3 Properties of the Discrete-
signal (and letting the period go to ∞ ); we
Time Fourier Transform . . . 91
will follow an identical argument for discrete-
time aperiodic signals. The differences be-
tween the continuous-time and discrete time
Fourier series (e.g., that the latter only in-
volves a finite number of complex expo-
nentials) will be reflected as differences be-
tween the continuous-time and discrete-time
Fourier transforms as well.

87
88 The Discrete-Time Fourier Transform

1 The Discrete-Time Fourier Transform

Consider a general signal x[n] which is nonzero on some interval −N1 ≤ n ≤ N2 and zero
elsewhere. We create a periodic extension x̃[n] of this signal with period N (where N is large
enough so that there is no overlap). As N → ∞, x̃[n] becomes equal to x[n] for each finite
value of n.
Since x̃[n] is periodic, it has a discrete-time Fourier series representation given by
N
X −1
x̃[n] = ak ejkω0 n
k=0

where ω0 = 2π
N
. The Fourier series coefficients are given by
+N −1
n0X
1
ak = x̃[n]e−jω0 kn
N n=n0

where n0 is any integer. Suppose we choose n0 so that the interval [−N1 , N2 ] is contained in
[n0 , n0 + N − 1]. Then since x̃[n] = x[n] in this interval, we have
+N −1
n0X ∞
1 1 X
ak = x[n]e−jω0 kn = x[n]e−jω0 kn
N n=n0
N n=−∞
Let us now define the discrete-time Fourier transform as
∞
X
jω
x[n]e−jωn .


X e ≜
n=−∞

From this, we see that ak = , i.e., the discrete-time Fourier series coefficients are
1 jkω0


e N
X
obtained by sampling the discrete-time Fourier transform at periodic intervals of ω0 . Also
note that X (ejω ) is periodic in ω with period 2π (since e−jωn is 2π-periodic).
Using the Fourier series representation of x̃[n], we now have
N −1 N −1 N −1
X
jkω0 n 1 X jkω0

jkω0 n 1 X
X ejkω0 ejkω0 n ω0


x̃[n] = ak e = X e e =
k=0
N k=0 2π k=0
Once again, we see that each term in the summand represents the area of a rectangle of
width ω0 obtained from the curve X (ejω ) ejω . As N → ∞, we have ω0 → 0. In this case,
the sum of the areas of the rectangles approaches the integral of the curve X (ejω ) ejωn , and
since the sum was over only N samples of the function, the integral is only over one interval
of length 2π. Since x̃[n] approaches x[n] as N → ∞, we have
Z
1
X ejω ejωn dω


x[n] =
2π 2π
This is the inverse discrete-time Fourier transform, or the synthesis equation.
The main differences between the discrete-time and continuous-time Fourier transforms
are the following. (1) The discrete-time Fourier transform X (ejω ) is periodic in ω with
The Discrete-Time Fourier Transform 89

period 2π, whereas the continuous-time Fourier transform is not necessarily periodic. (2)
The synthesis equation for the discrete-time Fourier transform only involves an integral over
an interval of length 2π, whereas the one for the continuous-time Fourier transform is over
the entire ω axis. Both of these are due to the fact that ejωn is 2π-periodic in ω, whereas the
continuous-time complex exponential is not.

Since the frequency spectrum of X (ejω ) is only uniquely specified over an interval of length
2π, we have to be careful about what we mean by "high" and "low" frequencies. Recalling
the discussion of discrete-time complex exponentials, high-frequency signals in discrete-time
have frequencies close to odd multiples of π, whereas low-frequency signals have frequencies
close to even multiples of π.

Example :
Consider the signal

x[n] = an u[n], |a| < 1

We have
∞
X ∞
X
X ejω = x[n]e−jωn = an e−jωn


n=−∞ n=0
∞
X
n
= ae−jω
n=0
1
=
1 − ae−jω
If we plot the magnitude of X (ejω ), we see an illustration of the "high" versus "low"
frequency effect. Specifically, if a > 0 then the signal x[n] does not have any oscillations
and |X (ejω )| has its highest magnitude around even multiples of π. However, if a < 0,
then the signal x[n] oscillates between positive and negative values at each time-step; this
"high-frequency" behavior is captured by the fact that |X (ejω )| has its largest magnitude
near odd multiples of π. See Figure. 5.4 in OW for an illustration. ■
Example :
Consider the signal

x[n] = a|n| , |a| < 1

We have
90 The Discrete-Time Fourier Transform

∞
X ∞
X
X ejω = x[n]e−jωn = a|n| e−jωn


n=−∞ n=−∞
−1
X ∞
X
−n −jωn
= a e + an e−jωn
n=−∞ n=0
X∞ ∞
X
= n jωn
a e + an e−jωn
n=1 n=0
jω
ae 1
= +
1 − aejω 1 − ae−jω
2
1−a
=
1 − 2a cos(ω) + a2
■

2 The Fourier Transform of Discrete-Time Periodic Sig-

nals
In the last chapter, we saw that if we take the Fourier transform of a continuoustime periodic
signal, we obtain scaled impulses located at the harmonic frequencies. We will see something
similar here for discrete-time periodic signals.
First, consider the signal

x[n] = ejω0 n .
We claim that the Fourier transform of this signal is
∞
X
X ejω =


2πδ (ω − ω0 − 2πl)
l=−∞

i.e., a set of impulse functions spaced 2π apart on the frequency axis. To verify this, note
that the inverse Fourier transform is given by
Z
1
X ejω ejωn dω


2π 2π
The integral is only over an interval of length 2π, and there is at most one impulse function
from X (ejω ) in any such interval. Let that impulse be located at ω0 + 2πr for some r ∈ Z.
Then we have

Z Z
1 jω jωn 1
2πδ (ω − ω0 − 2πr) ejωn dω = ej(ω0 +2πr)n = ejω0 n


X e e dω =
2π 2π 2π 2π

Thus consider a periodic discrete-time signal x[n], with Fourier series

N
X −1
x[n] = ak ejkω0 n = a0 + a1 ejω0 n + a2 ej2ω0 n + · · · + aN −1 ej(N −1)ω0 n
k=0
The Discrete-Time Fourier Transform 91

where ω0 = 2πN
. The Fourier transform of each term of the form ak ejkω0 n is a set of impulses
spaced 2π apart, with one located at ω = kω0 . Furthermore each of these impulses is scaled
by ak 2π. Since ak = ak+N l for any l (by the periodicity of the discrete-time Fourier series
coefficients), when we add up the Fourier transforms of all of the terms in the Fourier series
expansion of x[n], we obtain
∞  
jω

X 2πk
X e = 2πak δ ω −
k=−∞
N

Thus, the Fourier transform of a discrete-time periodic signal is indeed a sequence of

impulses located at multiples of ω0 , with a period of 2π.
Example 6.3. Consider the impulse train
∞
X
x[n] = δ[n − kN ]
k=−∞

The Fourier series coefficients of x[n] are given by

N −1
1 X 1
ak = x[n]e−jkω0 n =
N k=0 N

Thus the Fourier transform of x[n] is

∞   ∞  
jω

X 2πk 2π X 2πk
X e = ak 2πδ ω − = δ ω−
k=−∞
N N k=−∞ N

3 Properties of the Discrete-Time Fourier Transform

3.1 Periodicity
The Fourier transform of a signal x[n] is periodic in frequency, with period 2π :

X ejω = X ej(ω+2π)

This comes out of the fact that discrete-time complex exponentials are periodic in fre-
quency with period 2π.

3.2 Linearity
It is easy to see that

F {αx1 [n] + βx2 [n]} = αX1 ejω + βX2 ejω

92 The Discrete-Time Fourier Transform

3.3 Time and Frequency Shifting

We have

F {x [n − n0 ]} = e−jωn0 X ejω

and

F ejω0 n x[n] = X ej(ω−ω0 )



The first property is easily proved using the inverse Fourier transform equation, and the
second property is proved using the Fourier transform equation.
Example :
Consider a discrete-time low-pass filter, whose Fourier transform Hlp (ejω ) is a square pulse
centered at even multiples of π. Now consider the high pass filter Hhp (ejω ) which consists
of square pulses centered at odd multiples of π. We see that Hhp (ejω ) = Hlp ej(ω−π) .

Thus we have

hhp [n] = ejπn hlp [n] = (−1)n hlp [n].

■

3.4 First Order Differences

Consider the discrete-time analog of differentiation, which is to take the differences between
subsequent samples of the signal. By applying the linearity and time-shifting properties, we
have

F{x[n] − x[n − 1]} = X ejω − e−jω X ejω = 1 − e−jω X ejω .

3.5 Conjugation
For any discrete-time signal (that has a Fourier transform), we have

F {x∗ [n]} = X ∗ e−jω

Furthermore, if x[n] is real, we have x∗ [n] = x[n] and thus

X ejω = X ∗ e−jω

3.6 Time-Reversal
Consider the time-reversed signal x[−n]. We have
Z 2π Z 2π
1 −jωn 1
jω
X e−jω ejωn dω



x[−n] = X e e dω =
2π 0 2π 0
The Discrete-Time Fourier Transform 93

which is obtained by performing a change of variable ω → −ω (note that the negative sign
introduced by this change is canceled out by the reversal of the bounds of integration that
arise because of the negation). Thus, we have

F{x[−n]} = X e−jω

Together with the conjugation property, we see that for real-valued even signals (where
x[n] = x[−n] ), we have

X ejω = F{x[n]} = F{x[−n]} = X e−jω

Thus, the Fourier transform of real, even signals is also real and even.

6.3.7 Time Expansion

Recall that for a continuous-time signal x(t) and a scalar a ̸= 0, we had
 
1 jω
F{x(at)} = X
|a| a
Thus, an expansion in the time-domain led to a compression in the frequency domain and
vice versa.
In discrete-time, expansion and contraction of time-domain signals is not achieved simply
by scaling the time-variable. First, since the time index must be an integer, it does not make
sense to consider the signal x[an], where a < 1. Similarly, if we consider integer values of a
larger than 1 , then the signal x[an] only considers the values of x[n] at integer multiples of
a, and all information between those values is lost.
A different expansion of signals that preserves all their values is as follows. For a given
signal x[n] and positive integer k, define the signal
(
x[n/k] if n is a multiple of k
x(k) [n] =
0 otherwise

Thus, the signal x(k) [n] is obtained by spreading the points of x[n] apart by k samples and
placing zeros between the samples. We have
∞
X ∞
X
−jωn
x(k) [rk]e−jωrk

F x(k) [n] = x(k) [n]e =
n=−∞ r=−∞

since x(k) [n] is nonzero only at integer multiples of k. Since x(k) [rk] = x[r], we have
∞
X
x[r]e−jωrk = X ejkω


F x(k) [n] =
r=−∞

Example :
Consider the signal
94 The Discrete-Time Fourier Transform

1 n ∈ {0, 2, 4}

x[n] = 2 n ∈ {1, 3, 5}
0 otherwise



We note that we can write x[n] as

x[n] = g[n] + 2g[n − 1]

where
(
1 n ∈ {0, 2, 4}
g[n] =
0 otherwise
This g[n] can be viewed as an expansion of the signal
(
1 0≤n≤2
h[n] =
0 otherwise
Specifically, g[n] = h(2) [n]. The Fourier transform of h[n] is given by
2 −3jω 3


X 1 − e sin ω
H ejω = e−jωn = = e−jω 2



−jω ω
n=0
1 − e sin 2

Thus, the Fourier transform of g[n] is given by

sin(3ω)
G ejω = H ej2ω = e−2jω



sin(ω)
Finally,

sin(3ω)
X ejω = G ejω + 2e−jω G ejω = e−2jω 1 + 2e−jω




sin(ω)
■

3.7 Differentiation in Frequency

Suppose x[n] has Fourier transform X (ejω ). Then we have
∞ ∞
d d X X
X ejω = x[n]e−jωn = (−jn)x[n]e−jωn


dω dω n=−∞ n=−∞
Thus

dX (ejω )
F{nx[n]} = j
dω

3.8 Parseval’s Theorem

We have
The Discrete-Time Fourier Transform 95

∞ Z
X 1 2 2
X ejω


|x[n]| = dω
n=−∞
2π 2π

3.9 Convolution
Just as in continuous time, the discrete time signal ejωn is an eigenfunction of discrete-time
LTI systems. Specifically, if ejωn is applied to a (stable) LTI system with impulse response
h[n], the output of the system will be H (ejω ) ejωn .
Thus consider a signal x[n] written in terms of its Fourier transform as
Z
1
X ejω ejωn dω


x[n] =
2π 2π

This is a linear combination of complex exponentials (where the scaling factor on the
complex exponential ejωn is 2π
1
X (ejω ). By the LTI property, we thus have
Z Z
1 jω jωn 1
X ejω H ejω ejωn dω = y[n]




x[n] = X e e dω →
2π 2π 2π 2π

The expression on the right hand side is the output y[n] of the system when the input is
x[n]. Thus we have

Y ejω = H ejω X ejω

As in the continuous-time case, convolution in the time-domain is given by multiplication

in the frequency domain.
Example 6.6. Consider the system shown in Fig. 5.18a in OW. Let us analyze the rela-
tionship between y[n] and x[n] for that system.
First, we have w1 [n] =
(−1)n x[n] = ejπn x[n]. By the frequency shifting property, we see that
W1 (ejω ) = X ej(ω−π) . Next, we have

W2 ejω = Hlp ejω W1 ejω = Hlp ejω X ej(ω−π)

The signal w3 [n] is given by w3 [n] = (−1)n w2 [n], and thus W3 (ejω ) = W2 ej(ω−π) .

Putting this together with the expression for W2 (ejω ), we obtain

W3 ejω = Hlp ej(ω−π) X ej(ω−2π) = Hlp ej(ω−π) X ejω

From the bottom path, we have W4 (ejω ) = Hlp (ejω ) X (ejω ). Thus, we have

Y ejω = W3 ejω + W4 ejω = Hlp ejω + Hlp ej(ω−π) X ejω

Recall that Hlp ej(ω−π) is a high-pass filter centered at π. Thus, this system acts as a

bandstop filter, blocking all frequencies in a certain range and letting all of the low and high
frequency signals through.
96 The Discrete-Time Fourier Transform

3.10 Multiplication
Consider two signals x1 [n] and x2 [n], and define g[n] = x1 [n]x2 [n]. The discretetime Fourier
transform of g[n] is given by
∞
X
jω
x1 [n]x2 [n]e−jωn


G e =
n=−∞

Replacing x1 [n] by the synthesis equation

Z
1
X1 ejθ ejθn dθ


x1 [n] =
2π 2π
where we simply replaced the dummy variable ω with θ to avoid confusion, we obtain
∞ Z
X 1
jω
X1 ejθ ejθn dθx2 [n]e−jωn



G e =
n=−∞
2π 2π
Z ∞
1
X
= X1 e jθ
x2 [n]e−j(ω−θ)n dθ
2π 2π n=−∞
Z
1
X1 ejθ X2 ej(ω−θ) dθ



=
2π 2π
This resembles the typical convolution of the signals X1 (ejω ) and X2 (ejω ), except that the
integral is over only an interval of length 2π as opposed over the entire frequency axis. This
is called the periodic convolution of the two signals. Recall that we saw the same thing when
we considered the discrete-time Fourier series of the product of two periodic discrete-time
signals.
Example :
Consider the two signals

sin π2 n sin π4 n



x1 [n] = , x2 [n] =
πn πn
The Fourier transforms of these signals are square pulses, where the pulse centered at
0 extend from − π2 to π2 (for X1 (ejω ) ) and from − π4 to π4 (for X2 (ejω ) ). The Fourier
transform of g[n] = x1 [n]x2 [n] is given by
Z
jω 1
X1 ejθ X2 ej(ω−θ) dθ




G e =
2π 2π
Since we can choose any interval of length 2π to integrate over, let’s choose the interval
[−π, π) for convenience. We also only need to determine the values of the Fourier transform
for values of ω between −π and π, since the transform is periodic. Depending on the value
of ω, there are different cases that occur:

• If −π ≤ ω < − 3π , then there is no overlap in the signals X1 ejθ and X2 ej(ω−θ) ,

4
and thus G(jω) is zero.
The Discrete-Time Fourier Transform 97

• If − 3π4
≤ ω < − π4 , then there is partial overlap in the signals; the product is
a
rectangle with support from − π2 to ω + π4 , and thus G(jω) evaluates to 2π
1
ω + 3π
4
.

• If − π4 ≤ ω < π4 , there is full overlap and G(jω) is 2π

1 π
= 4.

1
2

• If π4 ≤ ω < 3π , there is partial overlap and is 1 3π

.


4
G(jω) 2π 4
− ω

• If 3π
4
≤ ω < π, there is no overlap and G(jω) is zero.

Note that since we are only integrating over θ between −π and π, the values of X ejθ

outside of that interval does not matter. Thus, we could also create a new signal jθ


X̂ 1 e
which is equal to X1 ejθ over the interval [−π, π) and zero everywhere else. The Fourier

transform can then be written as

Z Z ∞
jω 1 jθ j(ω−θ) 1
X̂1 ejθ X2 ej(ω−θ) dθ






G e = X1 e X2 e dθ =
2π 2π 2π −∞

i.e., it is the usual convolution of the signals X̂1 (ejω ) and X2 (ejω ). ■
98 The Discrete-Time Fourier Transform
 CHAPTER 7
SAMPLING

Thus far, we have considered continuous-

1 The Sampling Theorem . . . . 100
time signals and discrete-time signals (and
2 Reconstruction of a Signal
their associated Fourier transforms) as two
From Its Samples . . . . . . . 101
parallel tracks. In this part of the course, we
3 Undersampling and Aliasing . 103
will bring these two threads together and re-
4 Discrete-Time Processing of
late their frequency spectra. The main tool
Continuous-Time Signals . . . 103
that we will leverage is sampling continuous-
time signals to yield discrete-time signals.
In particular, it is often desirable to process
signals using digital systems (e.g., comput-
ers or embedded devices). Thus, we take a
continuous-time signal, sample it at a suf-
ficiently fast rate, process it using a digital
filter, and then convert the processed signal
back to a continuous-time signal.

99
100 Sampling

1 The Sampling Theorem

Consider a continuous-time signal x(t). A sampled version of this signal is obtained by
considering the values of the signal only at certain discrete points in time. In particular,
periodic or uniform sampling occurs when we pick some positive real number Ts , and consider
the values x (nTs ) , n ∈ Z. We will often denote this discrete sequence as x[n] (dropping the
sampling period Ts ), which was the notation that we used in our analysis of discrete-time
signals. The sampling frequency is denoted by ωs = 2π Ts
.
One can always sample a signal this way. However, the main question is how fast one
needs to sample (i.e., how small Ts needs to be) in order for the samples to be a faithful
representation of the underlying continuous-time signal. We will study this question here.
First, it will be useful for us to have a mathematical representation of the sampled signal.
Define the impulse train
∞
X
p(t) = δ (t − nTs )
n=−∞

Then, sampling a signal x(t) at a sampling period of Ts can be represented as multiplying

that signal by an impulse train, i.e.,
∞
X ∞
X
xp (t) = x(t)p(t) = x(t)δ (t − nTs ) = x (nTs ) δ (t − nTs ) .
n=−∞ n=−∞

Note that the values of the signal x(t) are irrelevant outside of the points where the impulse
functions in p(t) occur (i.e., at the sampling instants). Let us consider the frequency spectra
of these signals. Specifically, by the multiplication property of Fourier transforms, we have
Z ∞
1
Xp (jω) = X(jθ)P (j(ω − θ))dθ
2π −∞
Furthermore, since p(t) is periodic, we saw that the Fourier transform of p(t) will be given
by
∞
2π X
P (jω) = δ (ω − nωs )
Ts n=−∞

as the Fourier series coefficients of p(t) are each 1

Ts
. Thus,
∞ Z ∞
1 X
Xp (jω) = X(jθ)δ (ω − θ − nωs ) dθ
Ts n=−∞ −∞
∞
1 X
= X (j (ω − nωs ))
Ts n=−∞

Thus, the frequency spectrum of xp (t) consists of copies of the frequency spectrum of x(t),
where each copy is shifted (in frequency) by an integer multiple of the sampling frequency
ωs and scaled by T1s (see Fig. 7.1).
Sampling 101

Figure 7.1: The frequency spectrum of the signal x(t) and the signal xp (t).

If we want to be able to reconstruct x(t) from its sampled version xp (t), we would like to
make sure that there is an exact copy of X(jω) that can be extracted from Xp (jω). Based
on the above discussion, we see that this will be the case if no two copies of X(jω) overlap
in Xp (jω). Looking at Fig. 7.1, this will occur as long as

ωs − ωM > ωM
or equivalently,

ωs > 2ωM
where ωM is the largest frequency at which x(t) has nonzero content. This leads to the
sampling theorem.
If the sampling frequency ωs is larger than twice the largest frequency of the signal x(t),
then we can reconstruct the signal x(t) from its sampled version xp (t) by passing xp (t) through
an ideal low-pass filter, with cutoff ωc = ω2s .
The frequency 2ωM is called the Nyquist rate.

2 Reconstruction of a Signal From Its Samples

In general, it is not possible to implement an ideal low-pass filter: obtaining sharp cutoffs is
difficult, and furthermore, an ideal low-pass filter is noncausal (as it corresponds to a sinc
function in the time-domain). There are various other options that are frequently used to
reconstruct sampled signals.

2.1 Zero-Order Hold

The simplest option to reconstruct a signal is to simply hold the value of the signal constant
at the value of the previous sample. This is called a zero-orderhold (ZOH). To compare
this strategy to the ideal low-pass-filter, let’s consider the transfer function of the ZOH.
Specifically, note that if we put an impulse function into the ZOH , the output h0 (t) (i.e.,
impulse response) will be a square pulse from t = 0 to t = Ts (because the ZOH will keep
the value of the sample at t = 0 constant at 1 until the next sample at t = Ts , after which
102 Sampling

Figure 7.2: The impulse response of a zero-order-hold (left) and a first-orderhold (right).

point all samples are zero). This is shown in Fig. 7.2. It is easy to check that the transfer
function is given by

∞
sin ω T2s
Z

−jω T2s
H0 (jω) = h0 (t)e−jωt dt = e ω
−∞ 2

This has magnitude Ts at ω = 0 (like the ideal reconstructor), and the first frequency
at which it is equal to zero is at ωs (unlike the ideal reconstructor that cuts off at ω2s .

Furthermore, this frequency spectrum is not bandlimited, and thus the copies of X(jω) in
the spectrum of Xp (jω) will leak into the reconstructed signal under the ZOH .

2.2 First-Order Hold

A slightly more sophisticated reconstruction mechanism is to create a continuoustime signal
by joining each consecutive pair of samples by a line. This is called a first-order-hold (FOH).
The impulse response h1 (t) of an FOH is shown in the right plot of Fig. 7.2 (again, imagine
how the FOH reacts to an impulse coming into it: it sees a value of 0 at t = −Ts , a value of
1 at t = 0 and a value of 0 at t = Ts ). The transfer function is given by
Z ∞
H1 (jω) = h1 (t)e−jωt dt
−∞
Z 0   Z Ts  
1 −jωt 1
= t+1 e dt + − t + 1 e−jωt dt.
−Ts Ts 0 Ts

After some algebra (integration by parts, etc.), we obtain

!2
1 sin ω T2s 1
H1 (jω) = ω = |H0 (jω)|2
Ts 2
Ts

The magnitude of this filter is smaller than that of H0 (jω) outside of ω2s , although it is
still not bandlimited. Furthermore, the FOH is noncausal, but can be made causal with a
delay of Ts .
Higher order filters are also possible, and can be defined as a natural extension of zero
and first order holds.
Sampling 103

Figure 7.3: Sampling x(t) = cos(t) at a frequency of ωs = 4.

3 Undersampling and Aliasing

If the sampling frequency ωs is not strictly larger than twice the largest frequency, we will not
be able to perfect reconstruct the original signal. To illustrate this, it is easiest to consider
sampled sinusoids.
Consider the signal x(t) = cos(t) which has frequency ω0 = 1. The sampling theorem
indicates that as long as the sampling frequency ωs is larger than
2ω0 = 2, we can reconstruct x(t) from its samples. Let us choose ωs = 4. The frequency
spectrum of xp (t) = x(t)p(t) is shown in Fig. 7.3.
Now consider another signal x1 (t) = cos (ω1 t), and suppose that we sample this signal at
ωs = 4. Let xp1 (t) be the resulting (continuous-time) sampled signal. For what value of ω1
will the frequency spectrum Xp1 (jω) look exactly the same as Xp (jω) ?
To answer this, note that X1 (jω) has impulses located at ±ω1 , and Xp1 (jω) will have
impulses located at kωs ± ω1 , for k ∈ Z. Looking at the frequency spectrum of Xp (jω) in
Fig. 7.3 , we see that ω1 should be odd (otherwise, Xp1 (jω) will have impulses at some
even frequencies, whereas all of the impulses are at odd frequencies in Xp (jω)). Suppose we
try ω1 = 3. Then Xp1 (jω) will have impulses at ±3, which matches two of the impulses in
Xp (jω). We should check the copies of the signals in Xp1 (jω) as well. Specifically, there will
be a copy centered at ωs = 4, with one impulse three units to the left (at ω = 1 ) and one
impulse three units to the right (at ω = 7 ). Similarly, the copy centered at 2ωs will have one
impulse at 5 and one impulse at 11 . The same is true for the negative frequencies. Thus,
we see that if ω1 = 3, then Xp1 (jω) looks exactly the same as Xp (jω), and thus the signals
x(t) = cos(t) and x1 (t) = cos(3t) look exactly the same if sampled at ωs = 4.

4 Discrete-Time Processing of Continuous-Time Signals

Let’s take a closer look at taking a signal from continuous-time, operating on it, and con-
verting it back to discrete-time. Specifically, given a signal x(t), let xp (t) = x(t)p(t) be the
continuous-time representation of the sampled signal, and let xd [n] = x (nTs ) be the sequence
of samples.
We have
Z ∞ ∞
X ∞
X
Xp (jω) = x(t)δ (t − nTs ) e−jωt dt = x (nTs ) e−jωnTs
−∞ n=−∞ n=−∞

One the other hand, if we take the discrete-time Fourier transform of the sequence xd [n],
104 Sampling

we have
∞
X ∞
X
−jωn
jω
x (nTs ) e−jωn


Xd e = xd [n]e =
n=−∞ n=−∞

Comparing the two expressions, we see that

 
jω

ω
Xd e = Xp j
Ts
In other words, the frequency spectrum of xp (t) (given by the continuous-time Fourier
transform) is just a frequency-scaled version of the frequency spectrum of xd [n] (given by the
discrete-time Fourier transform). Specifically, Xp (jω) is ωs = 2π
Ts
periodic, whereas Xd (ejω )
is 2π periodic. This scaling is essentially due to the fact that the discrete-time signal xd [n]
is "normalized" with respect to the sampling period; it only operates on the sequence of
samples, and does not explicitly consider how far apart those samples are. However, xp (t)
explicitly contains the sampling period Ts , as the impulses are spaced that far apart.
The above result has the following implication for the digital processing of signals. Suppose
that we wish to implement a filter that has a continuous-time Fourier transform H(jω), but
using a discrete-time system. Suppose H(jω) is bandlimited, with highest frequency  ωM.
Then we simply design the discrete time filter to have frequency response Hd (ejω ) = H j Tωs
for −ωM Ts ≤ ω ≤ ωM Ts (and Hd (ejω ) being 2π-periodic otherwise). The inverse Fourier
transform of Hd (ejω ) can then be found to obtain the impulse response of the digital filter.
After the sampled signal is processed with this digital filter, it can then be transformed back
into continuous-time via a ZOH, FOH, etc.
Example :
Consider a bandlimited differentiator
(
jω |ω| ≤ ωc
H(jω) =
0 otherwise
The magnitude and phase are shown in Fig. 7.4
To implement this in discrete-time, we create a discrete-time filter with transfer function
Hd (ejω ) to have the same shape as H(jω) (except for the fact that Hd is periodic), with
frequencies scaled by Ts , i.e.,
  ( ω
ω j Ts |ω| ≤ ωc Ts
Hd ejω = H j


=
Ts 0 ωc Ts < |ω| ≤ 2π
The impulse response of the corresponding filter is
(
(−1)n
nTs
n ̸= 0
hd [n] =
0 n=0
■
Sampling 105

Figure 7.4: The frequency response of a bandlimited differentiator.

106 Sampling
 CHAPTER 8
THE LAPLACE TRANSFORM

Thus far, we have seen ways to take time-

1 The Laplace Transform . . . . 108
domain signals and transform them into
2 The Region of Convergence . 110
frequency-domain signals, by identifying the
3 The Inverse Laplace Transform 112
amount of contribution of complex expo-
4 Some Properties of the
nentials of given frequencies to the signal.
Laplace Transform . . . . . . 115
Specifically, for periodic signals, we started
5 Finding the Output of an LTI
with the Fourier series representation of a
System via Laplace Transforms 116
signal in terms of its harmonic family. For
6 Finding the Impulse Response
more general absolutely integrable signals,
of a Differential Equation via
we generalized the Fourier series to the
Laplace Transforms . . . . . . 117
Fourier transform, where the signal is rep-
resented in terms of complex exponentials of
all frequencies (not just those from the har-
monic family).

107
108 The Laplace Transform

1 The Laplace Transform

To develop this, first recall that complex exponentials of the form est are eigenfunctions of
LTI systems, even when s is a general complex number. Specifically, if x(t) = est is the input
to an LTI system with impulse response h(t), we have
Z ∞ Z ∞
y(t) = x(t) ∗ h(t) = x(t − τ )h(t)dτ = e st
h(t)e−sτ dτ
−∞ −∞

Based Ron the above, we see that the output is the input signal est , multiplied by the
∞
quantity −∞ h(t)e−sτ dτ . We will call this the Laplace transform of the signal h(t).
The Laplace transform of a signal x(t) is given by
Z ∞
X(s) = x(t)e−st dt
−∞

where s ∈ C. We will also denote the Laplace transform of

x(t) by L{x(t)}.
Note that the limits of the integration go from −∞ to ∞, and thus this is called the
bilateral Laplace transform. When the limits only go from 0 to ∞, it is called the unilateral
Laplace transform. For the purposes of this course, if we leave out the qualifier, we mean
the bilateral transform. Note that when s = jω, then X(s) is just the Fourier transform of
x(t) (assuming the transform exists). However, the benefit of the Laplace transform is that
it also applies to signals that do not have a Fourier transform. Specifically, note that s can
be written as s = σ + jω, where σ and ω are real numbers. Then we have
Z ∞
X(s) = X(σ + jω) = x(t)e−σt e−jωt dt
−∞

Thus, for a given s = σ + jω, we can think of the Laplace transform as the Fourier
transform of the signal x(t)e−σt . Even if x(t) is not absolutely integrable, it may be possible
that x(t)e−σt is absolutely integrable if σ is large enough (i.e., the complex exponential can
be chosen to cancel out the growth of the signal in the Laplace transform).
Example :
Consider the signal x(t) = e−at u(t) where a is some real number. The Laplace transform
is given by
Z ∞ Z ∞
−st
X(s) = x(t)e dt = e−(s+a)t dt
−∞ 0
∞
1 −(s+a)t
=− e
s+a 0
1
=
s+a
as long as Re{s + a} > 0, or equivalently Re{s} > −a. Note that if a is positive, then
the integral converges for Re{s} = 0 as well, in which case we get the Fourier transform
The Laplace Transform 109

X(jω) = jω+a1
. However, if a is negative, then the signal does not have a Fourier transform
(but it does have a Laplace transform for s with a sufficiently large real part). ■
Example :
Consider the signal x(t) = −e−at u(−t) where a is a real number.
We have
Z ∞ Z 0
−st
X(s) = x(t)e dt = − e−(s+a)t dt
−∞ −∞
0
1 −(s+a)t
= e
s+a −∞
1
=
s+a
as long as Re{s + a} < 0, or equivalently, Re{s} < −a. ■
Comparing the above examples, we notice that both the signals e−at u(t) and −e−at u(−t)
had the same Laplace transform s+a 1
, but that the ranges of s for which each had a Laplace
transform was different.
Consider a signal x(t). The range of values of s for which the Laplace transform integral
converges is called the Region of Convergence (ROC) of the Laplace transform.
Thus, in order to specify the Laplace transform of a signal, we have to specify both the
algebraic expression (e.g., s+a
1
) and the region of convergence for which this expression is
valid. A convenient way to visualize the ROC is as a shaded region in the complex plane. For
example, the ROC Re{s} > −a can be represented by shading all of the complex plane to
the right of the line Re{s} = −a. Similarly, the ROC Re{s} < −a is represented by shading
the complex plane to the left of the line Re{s} = −a.
Example :
Consider the signal x(t) = 3e−2t u(t)−2e−t u(t). It is easy to see that the Laplace transform
is a linear operation, and thus we can find the Laplace transform of x(t) as a sum of the
Laplace transform of the two signals on the right hand side.
The Laplace transform of 3e−2t u(t) is s+2
3
, with ROC Re{s} > −2. The Laplace transform
of −2e u(t) is − s+1 , with ROC Re{s} > −1. Thus, for the Laplace transform of x(t)
−t 2

to exist, we need s to fall in the ROC of both of its constituent parts, which means
Re{s} > −1. Thus,
3 2 s−1
X(s) = − = 2
s+2 s+1 s + 3s + 2
with ROC Re{s} > −1. ■
In the above examples, we saw that the Laplace transform was of the form

N (s)
X(s) =
D(s)
where N (s) and D(s) are polynomials in s. The roots of the polynomial N (s) are called
the zeros of X(s) (since X(s) will be zero when s is equal to one of those roots), and the
110 The Laplace Transform

roots of D(s) are called the poles of X(s) (evaluating X(s) at a pole will yield ∞). We can
draw the poles and zeros in the s-plane using ◦ for zeros and × for poles.
Example :
Consider the signal
4 1
x(t) = δ(t) − e−t u(t) + e2t u(t)
3 3
The Laplace transform of δ(t) is
Z ∞
L{δ(t)} = δ(t)e−st dt = 1
−∞

for any value of s. Thus the ROC for δ(t) is the entire s-plane. Putting this with the
other two terms, we have

4 1 1 1 (s − 1)2
X(s) = 1 − + =
3s+1 3s−2 (s + 1)(s − 2)
with ROC Re{s} > 2. ■

2 The Region of Convergence

Let us dig a little deeper into the region of convergence for Laplace transforms. Recall that
for a given signal x(t), the ROC is the set of values s ∈ C such that the Laplace transform
integral converges. More specifically, writing s = σ + jω, we see that
Z ∞ Z ∞
−st
x(t)e−σt e−jωt dt


x(t)e dt =
−∞ −∞
Thus, as long as x(t)e −σt
is absolutely integrable, this integral exists. Note that this does
not depend on the value of ω. Thus, we have the following fact about the ROC.
Property 1. The ROC consists of strips parallel to the jω-axis in the s-plane.
For the next property of the ROC, suppose that the signal x(t) has a Laplace transform
given by a rational function. We know that the poles of this function are the set of complex
s such that X(s) is infinite. Since X(s) is given by the Laplace transform integral, we see
that the ROC cannot contain any poles of X(s).
Property 2. For rational Laplace transforms, the ROC does not contain any
poles of X(s).
The third property pertains to signals that are of finite duration (and absolutely inte-
grable). Specifically, suppose that x(t) is nonzero only between two finite times T1 and T2 .
Then we have
Z T2
X(s) = x(t)e−st dt
T1
which is finite for any finite s. Thus we have the following.
Property 3. If x(t) is of finite duration and absolutely integrable, then the
ROC is the entire s-plane.
The Laplace Transform 111

Another way to think of the above property is as follows. No matter what σ we pick, the
signal x(t)e−σt will be absolutely integrable as long as x(t) is of finite duration and absolutely
integrable. The fact that x(t) is of finite duration allows us to overcome the fact that the
signal e−σt may be growing unboundedly outside of the interval [T1 , T2 ].
While the previous property considered the case where the signal is of finite duration, we
will also be interested in signals that are only zero either before or after some time. First,
a signal x(t) is right-sided if there exists some T1 ∈ R such that x(t) = 0 for all t < T1 . A
signal x(t) is left-sided if there exists some T2 ∈ R such that x(t) = 0 for all t > T2 . A signal
x(t) is two-sided if it extends infinitely far in both directions.
Property 4. If x(t) is right-sided and if the line Re{s} = σ0 is in the ROC, then
the ROC contains all values s such that Re{s} ≥ σ0 .
To see why this is true, first note that since x(t) is right-sided, there exists some T1 such
that x(t) = 0 for all t < T1 . If s with Re{s} = σ0 is in the ROC, then x(t)e−σ0 t is absolutely
integrable, i.e.,
Z ∞
|x(t)|e−σ0 t dt < ∞
T1

Now suppose that we consider some σ1 > σ0 . If T1 > 0, then e−σ1 t is always smaller than
e over the region of integration, and thus x(t)e−σ1 t will also be absolutely integrable. If
−σ0 t

T1 < 0, then
Z ∞ Z 0 Z ∞
−σ1 t −σ1 t
|x(t)|e dt = |x(t)|e dt + |x(t)|e−σ1 t dt
T1 T1 0
Z 0 Z ∞
−σ1 t
≤ |x(t)|e dt + |x(t)|e−σ0 t dt
T1 0

The first term is finite (since it is integrating some signal of finite duration), and the
second term is finite since x(t)e−σ0 t is absolutely integrable. Thus, once again, x(t)e−σ1 t is
absolutely integrable, and thus s with Re{s} ≥ σ0 also falls within the ROC of the signal.
The same reasoning applies to show the following property.
Property 5. If x(t) is left-sided and if the line Re{s} = σ0 is in the ROC, then
the ROC contains all values s such that

Re{s} ≤ σ0 .
If x(t) is two-sided, we can write x(t) as x(t) = xR (t) + xL (t), where xR (t) is a right-sided
signal and xL (t) is a left-sided signal. The former has an ROC that is the region to the right
of some line in the s-plane, and the latter has an ROC that is the region to the left of some
line in the s-plane. Thus, the ROC for x(t) contains the intersection of these two regions (if
there is no intersection, x(t) does not have a Laplace transform).
Property 6. If x(t) is two-sided and contains the line Re{s} = σ0 in its ROC,
then the ROC consists of a strip in the s-plane that contains the line Re{s} = σ0 .
Example :
Consider the signal x(t) = e−b|t| . We write this as
112 The Laplace Transform

x(t) = e−bt u(t) + ebt u(−t)

Note that we modify the definition of u(t) in this expression so that u(0) = 12 , so that
x(0) = 1 as required. As this modification is only at a single point (of zero width and
finite height), it will not make a difference to the quantities calculated by integrating the
signals.
The signal e−bt u(t) has Laplace transform
1
L e−bt u(t) =

s+b
with ROC Re{s} > −b. The signal ebt u(−t) has Laplace transform
−1
L ebt u(−t) =

,
s−b
with ROC Re{s} < b. If b ≤ 0, then these two ROCs do not overlap, in which case x(t)
does not have a Laplace transform. However, if b > 0, then x(t) has the Laplace transform
1 1
L{x(t)} = − ,
s+b s−b
with ROC −b < Re{s} < b. ■
As we will see soon, a rational Laplace transform X(s) can be decomposed into a sum of
terms, each of which correspond to an exponential signal. The ROC for X(s) consists of the
intersection of the ROCs for each of those terms, and since none of the ROCs can contain
poles, we have the following property.
Property 7. If X(s) is rational, then its ROC is bounded by poles or extends
to infinty.
Example :
Consider
1
X(s) =
s(s + 1)
There are three possible ROCs for this Laplace transform: the region to the right of
the line Re{s} = 0, the region between the lines Re{s} = −1 and Re{s} = 0, or the region
to the left of the line Re{s} = −1. ■

3 The Inverse Laplace Transform

Consider again the Laplace transform evaluated at s = σ + jω :
Z ∞
X(σ + jω) = x(t)e−σt e−jωt dt
−∞

Since this is just the Fourier transform of x(t)e−σt , we can use the inverse Fourier transform
formula to obtain
The Laplace Transform 113

Z ∞
−σt 1
x(t)e = X(σ + jω)ejωt dω.
2π −∞

If we multiply both sides by eσt , we get

Z ∞
1
x(t) = X(σ + jω)e(σ+jω)t dω
2π −∞
Doing a change of variable s = σ + jω, we get
Z σ+∞
1
x(t) = X(s)est dω
2π σ−j∞

This is the inverse Fourier transform formula. It involves an integration over the line in the
complex plane consisting of points satisfying Re{s} = σ. There are actually simpler ways to
calculate the inverse Fourier transform, using the notion of partial fraction expansion, which
we will consider here.
Example :
Consider X(s) = 1
s(s+1)
. First, we note that
1 1 1
= −
s(s + 1) s s+1
Now each of these terms is of a form that we know (they correspond to complex expo-
nentials). So, for example, if the ROC for X(s) is the region to the right of the imaginary
axis, since the ROC consists of the intersection of the ROCs of both of the terms, we know
that both terms must be right-sided signals. Thus,

x(t) = u(t) − e−t u(t)

Similarly, if the ROC is between the lines Re{s} = −1 and Re{s} = 0, the first term is
left-sided and the second term is right-sided, which means

x(t) = −u(−t) − e−t u(t)

Finally, if the ROC is to the right of the line Re{s} = −1, both terms are left-sided
and thus

x(t) = −u(−t) + e−t u(−t)

■
In the above example, we "broke up" the function s(s+1)
1
into a sum of simpler functions,
and then applied the inverse Laplace Transform to each of them. This is a general technique
for inverting Laplace Transforms, which we now study.

3.1 Partial Fraction Expansion

Suppose we have a rational function
114 The Laplace Transform

bm sm + bm−1 sm−1 + · · · + b1 s + b0 N (s)

X(s) = n n−1
=
s + an−1 s + · · · + a1 s + a0 D(s)

where the ai ’s and bi ’s are constant real numbers.

Definition 3.1: If m ≤ n, the rational function is called proper.

If m < n, it is strictly proper.

By factoring N (s) and D(s), we can write

(s + z1 ) (s + z2 ) · · · (s + zm )
X(s) = K
(s + p1 ) (s + p2 ) · · · (s + pn )

Recall that the zeros of X(s) are given by −z1 , −z2 , . . . , −zm , and the poles are −p1 , −p2 , . . . , −pn .
First, suppose each of the poles are distinct and that X(s) is strictly proper. We would like
to write

k1 k2 kn
X(s) = + + ··· +
s + p1 s + p 2 s + pn

for some constants k1 , k2 , . . . , kn , since the inverse Laplace Transform of X(s) is easy in
this form. How do we find k1 , k2 , . . . , kn ?
Heaviside’s Cover-up Method. To find the constant ki , multiply both sides of the expansion
of X(s) by (s + pi ) :

k1 (s + pi ) k2 (s + pi ) kn (s + pi )
(s + pi ) X(s) = + + · · · + ki + · · · + .
s + p1 s + p2 s + pn

Now if we let s = −pi , then all terms on the right hand side will be equal to zero, except
for the term ki . Thus, we obtain

ki = (s + pi ) X(s)|s=−pi
Example :
Consider X(s) = s+5
s3 +3s2 −6s−8
. The denominator is factored as

s3 + 3s2 − 6s − 8 = (s + 1)(s − 2)(s + 4)

We would thus like to write
s+5 k1 k2 k3
X(s) = = + +
(s + 1)(s − 2)(s + 4) s+1 s−2 s+4
Using the Heaviside coverup rule, we obtain
The Laplace Transform 115

4 4
k1 = (s + 1)X(s)|s=−1 = =−
(−3)(3) 9
7 7
k2 = (s − 2)X(s)|s=2 = =
(3)(6) 18
1 1
k3 = (s + 4)X(s)|s=−4 = =
(−3)(−6) 18
■
The partial fraction expansion when some of the poles are repeated is obtained by following
a similar procedure, but it is a little more complicated. We will not worry too much about this
scenario here. One can also do a partial fraction expansion of nonstrictly proper functions by
first dividing the denominator into the numerator to obtain a constant and a strictly proper
function, and then applying the above partial fraction expansion.

4 Some Properties of the Laplace Transform

The Laplace transform has various properties that are quite similar to those for Fourier
transforms (linearity, time-shifting, etc.) We will focus on two important ones here.

4.1 Convolution
Consider two signals x1 (t) and x2 (t) with Laplace transforms X1 (s) and X2 (s) and ROCs R1
and R2 , respectively. Then

L {x1 (t) ∗ x2 (t)} = X1 (s)X2 (s),

with ROC containing R1 ∩ R2 . Thus, convolution in the time-domain maps to multiplica-
tion in the s-domain (as was the case with Fourier transforms).
Example :
Consider the convolution u(t) ∗ u(t). Since L{u(t)} = 1
s
with ROC Re{s} > 0, we have
1
L{u(t) ∗ u(t)} =
s2
with ROC containing the region Re{s} > 0.
■

4.2 Differentiation
Consider a signal x(t), with Laplace transform X(s) and ROC R. We have
Z
1
x(t) = X(s)est ds
2π
Differentiating both sides with respect to t, we have
Z
dx(t) 1
= sX(s)est ds
dt 2π
116 The Laplace Transform

Thus, we see that

 
dx
L = sX(s)
dt
with ROC containing R. More generally,
 m 
d x
L m
= sm X(s)
dt

4.3 Integration
Rt
Given a signal x(t) whose Laplace transform has ROC R, consider the integral −∞
x(τ )dτ .
Note that
Z t
x(τ )dτ = u(t) ∗ x(t)
−∞
and thus using the convolution property, we have
Z t 
1
L x(τ )dτ = X(s)
−∞ s
with ROC containing R ∩ {Re{s} > 0}.

5 Finding the Output of an LTI System via Laplace Trans-

forms
The Laplace transform properties (time-domain convolution corresponding to frequency-
domain multiplication, in particular) are very useful in analyzing the output of LTI systems
to inputs. Specifically, consider an LTI system with impulse response h(t) and input y(t).
We know that

Y (s) = H(s)X(s)
assuming all Laplace transforms exist. Using the expressions for H(s) and X(s), we can
thus calculate Y (s) (and its ROC), and then use an inverse Laplace transform to determine
y(t).
Example :
Consider an LTI system with impulse response h(t) = e−2t u(t). Suppose the input is
x(t) = e−3t u(t). The Laplace transforms of h(t) and x(t) are
1 1
H(s) = , X(s) =
s+2 s+3
with ROCs Re{s} > −2 and Re{s} > −3, respectively. Thus we have
1 1
Y (s) = H(s)X(s) =
s+2s+3
The Laplace Transform 117

with ROC Re{s} > −2. Using partial fraction expansion, we have
1 1
Y (s) = −
s+2 s+3
and thus y(t) = e−2t u(t) − e−3t u(t). ■
Example :
Consider an LTI system with impulse response h(t) = −e4t u(−t) and input x(t) = e2t u(t),
where we interpret u(−t) as being 1 for t < 0. The Laplace transforms are
1 1
H(s) = , X(s) =
s−4 s−2
with ROCs Re{s} < 4 and Re{s} > 2, respectively. Since there is a nonempty inter-
section, we have
1 1 1 1 1 1
Y (s) = H(s)X(s) = = −
s−4s−2 2s−4 2s−2
with ROC 2 < Re{s} < 4. Thus, y(t) is two-sided, and given by
1 1
y(t) = − e4t u(−t) + e2t u(t)
2 2
■

6 Finding the Impulse Response of a Differential Equa-

tion via Laplace Transforms
The differentiation property of Laplace transforms is also extremely useful for analyzing
differential equations. Specifically, suppose that we have a constant coefficient differential
equation of the form
n m
X dk y(t) X dk x(t)
ak = bk
k=0
dtk k=0
dtk
Taking Laplace transforms, we obtain
n
! m
!
X X
k k
ak s Y (s) = bk s X(s)
k=0 k=0

or equivalently
Pm
bk s k
Y (s) = Pnk=0 k
X(s).
k=0 ak s
| {z }
H(s)

Thus, the impulse response of the differential equation is just the inverse Laplace transform
of H(s) (corresponding to an appropriate region of convergence).
118 The Laplace Transform

Example :
Consider the differential equation

d3 y(t) d2 y(t) dy(t)

+ 2 − − 2y(t) = x(t)
dt3 dt2 dt
Taking Laplace transforms, we have

Y (s) 1 1
H(s) = = 3 2
=
X(s) s + 2s − s − 2 (s − 1)(s + 1)(s + 2)
In this case, we have the partial fraction expansion
1 1 1 1 1 1
H(s) = − +
6s−1 2s+1 6s+2
Suppose we are told the impulse response is causal (which implies it is rightsided).
Thus, the ROC would be to the right of the furthest pole and we have
 
1 t 1 −t 1 −2t
h(t) = e − e + e u(t)
6 2 6
■
 Part II

Second part: Graphs Signals

119
 CHAPTER 1
GRAPHS SIGNALS AND SYSTEMS

Graphs are mathematical objects used to

1 Graph Terminology . . . . . . 122
represent the structured medium generat-
2 Signals Defined on Graphs . . 124
ing the information (such as power or trans-
3 Graph Convolutional Filters . 125
portation networks) or to represent prox-
4 Graph Fourier Transform . . . 129
imities between datapoints (such as corre-
5 Frequency Response . . . . . 131
lations). In either case, graphs often fa-
cilitate the downstream localized processing
and learning tasks. We follow the presenta-
tion in the paper Isufi et al., “Graph filters
for signal processing and machine learning
on graphs”.

121
122 Graphs Signals and Systems

1 Graph Terminology
We denote a weighted graph by G = (V, E, W), where V = {1, . . . , N } is the set of nodes,
E ⊆ V × V is the set of edges such that (i, j) ∈ E if and only if there is an edge from node i
to node j, and W : E → R+ is a weight function. If all edge weights equal one, the graph is
said unweighted.
Graphs can be directed or undirected. In an undirected graph, there is no orientation
in the edges in E, W(i, j) = W(j, i), and the neighboring set for a node i is denoted by
Ni = {j ∈ V : (i, j) ∈ E}. In a directed graph (or digraph), an edge (i, j) ∈ E has an
orientation starting from node i and ending at node j. We say that node j is an out-neighbor
of i (and i an in-neighbor of j ). The out-neighboring set of node i is denoted by Niout =
{j ∈ V : (i, j) ∈ E} and, likewise, the in-neighboring set by Niin = {j ∈ V : (j, i) ∈ E}.
We represent graph G via the weighted adjacency matrix A, which is an N × N sparse
matrix with nonzero elements [A]ji = aji = W(i, j) > 0 representing the strength of edge
(i, j) ∈ E. Matrix A is symmetric aij = aji for an undirected graph and asymmetric aji ̸=
aij for a directed one. For undirected graphs, another widely used matrix is the graph
Laplacian L = D − A, where D = diag(A1) is a diagonal matrix whose P i th diagonal
element [D]ii = dii is the sum of all edge weights incident to node i, i.e., dii = j∈Ni aij [7].

Definition 1.1: Graph shift operator (GSO).

We represent the structure of a graph G with a generic matrix S ∈ RN ×N called the
graph shift operator matrix. The only requirement for S to be a valid GSO is that

[S]ji = sji = 0 whenever (i, j) ∈

/ E for i ̸= j. (1.1)

Both matrices A and L are special cases for S [5], [21]. Other candidate GSOs include the
normalized adjacency matrix An = D−1/2 AD−1/2 , the normalized Laplacian matrix Ln =
D−1/2 LD−1/2 or the random walk Laplacian Lrw = D−1 L.
A graph and associated GSO can represent:

1. Physical networks: Here, nodes and edges physically exist. For example, in a sensor
network, nodes are sensors and edges are communication links [32]. A directed edge
indicates the communication direction and the edge weight captures the communication
channel properties. Other examples include:

(a) multi-agent robot systems where nodes are robots and edges are communication
channels [33];
(b) power networks where nodes are buses and edges are power lines [34];
(c) telecommunication networks where nodes are transceivers and edges are channels
[35], [36];
(d) water networks where nodes are junctions and edges are pipes [37], and;
(e) road networks where nodes are intersections and edges are roads [38].
Graphs Signals and Systems 123

2. Abstract networks: These graphs typically represent dependencies between the dat-
apoints. Consider N datapoints, each described by a feature vector fi ∈ RF , and let
dist (fi , fj ) be a distance measure (e.g., Euclidean) between datapoints i and j. Each
datapoint is considered as a node and two datapoints could be connected based on [39]:

(a) ε- neighborhood, where the edge weight is

(
f (dist (fi , fj ) ; θ) if dist (fi , fj ) ≤ ε
aij = (1.2)
0 otherwise

where f (·; θ) is a parametric function (e.g., a Gaussian kernel

f (dist (fi , fj ) ; θ) = exp (− dist (fi , fj ) /2θ2 )) and ε > 0 is a constant controlling the
edge sparsity;
(b) k-nearest neighbor, where each node is connected only to the k closest datapoints
with respect to f (dist (fi , fj ) ; θ), which can be again a Gaussian kernel or a Pearson
correlation; iii) weighted fully connected graph, which is the particular case of (1.2)
for ε → ∞.

The above approaches build undirected abstract networks based on similarities. Al-
ternatives for directed or causal dependencies are also possible; see [40]-[42]. In fact,
identifying the most suitable abstract network from the data is a wide research area by
itself.

Abstract networks are encountered, for example, in:

1. recommender systems, where two items are connected, e.g., if their Pearson correlation
is greater than some value [43];

2. brain networks, where the nodes are brain regions and the edges are given, e.g., by the
cross-correlation or mutual information of electroencephalography (EEG) time series
in the different regions [44];

3. social networks, where nodes are users and edge weights may represent the number of
interactions between them; iv ) economic networks, where nodes are different economic
sectors and the edges maynrepresent the input and output production going from one
sector to another [45].

Since abstract networks represent dependencies between datapoints, they can be manipulated
by recomputing edge weights, clustering, or pruning to facilitate representation. However,
this is not typically the case for physical networks, as they often represent the medium
with respect to which processing is performed. Graph filters can leverage such structure for
distributed processing.
124 Graphs Signals and Systems

2 Signals Defined on Graphs

Data is often represented as a graph signal, as features, or embeddings associated to each
node.

Definition 2.1: Graph signal.

A graph signal x is a function from the node set to the field of real numbers; i.e.,
x : V → R. We can represent a graph signal as a vector x ∈ RN , where the i th entry
[x]i = xi is the signal value at node i[19].

We denote the space of all graph signals defined on graphs with N nodes by XN = {x :
V → R, |V| = N }. An example of a graph signal is a recording in a brain network, i.e., each
brain area corresponds to a node, two nodes share a link based on structural connectivity,
and the brain EEG measurement is the signal of a particular node. We may want to process
such a signal to understand, e.g., how different individuals have mastered a specific task [46].
Processing and learning tasks with graph signals include:

1. Signal reconstruction, including interpolation and denoising: We often observe a cor-

rupted version of the graph signal only at a subset of nodes. This may be the case of
noisy measurements in sensor [32], power [34], or water networks [47], or when a few
ratings are available in a recommender system [43]. We want to denoise the signal or
interpolate the missing values by relying on the neighboring signal values and the graph
structure.

2. Signal compression: When graph signals feature nice patterns such as having similar
values at neighboring vertices, it is possible to compress the signal (with or without
loss of information) by developing representations that require fewer coefficients, and
storing those coefficients rather than the original signal.

3. Signal classification: This task consists of classifying different graph signals observed
over a common underlying graph. One such example is classifying patients based on
their brain recordings, as discussed above [46].

4. Node classification: This task consists of classifying a subset of nodes in the graph
given the class labels on another subset. Depending on the available information, there
are two ways to approach this problem. First, when node features are available we
can treat them as a collection of graph signals and leverage their coupling with the
underlying connectivity to infer the missing labels. The state-of-the-art for this task is
achieved by GNNs, which, as we shall see in Section VIII rely heavily on graph filters
[23]. Second, when node features are unavailable, we treat the available labels as graph
signals and transform node classification into a label interpolation task that can be
solved with graph filters [22].

5. Graph classification / regression: These tasks start with a collection of different graphs
and (optionally) graph signals.
Graphs Signals and Systems 125

Figure 1.1: The graph convolution as a shift register. Highlighted are the nodes that reach
node 1 on each consecutive shift; that is, the nodes j whose signal value xj contributes to
S x i . The resulting summary of each communication S x is correspondingly weighted by
k k

a filter parameter hk . For each k, the parameter hk is the same for all nodes.

The classification task assigns a label to the whole graph; e.g., classifying molecules
into different categories such as soluble vs. non-soluble, whereas the regression task
assigns a continuous number to each graph (e.g., the degree of solubility) [48].

6. Link prediction: Here, the goal is to infer if two nodes are connected given the current
structure and the respective graph signals [49]. This is the case of friend recommenda-
tion in social networks, whereby leveraging the friendship connectivity between users
based on their feature signals (e.g., geo-position, workplace) we can infer missing links.

7. Graph identification: This task extends the link prediction to that of inferring the whole
graph structure given only the graph signals [41]. Here graph filters play a central role
in modeling the relationships between candidate graph structures and the observed
signals. We detail this problem in Sec. IX-C.

8. Distributed processing: Here the graph topology represents the structure of a sensor
network and we want to distributively solve a task related to graph signals. Graph filters
lend themselves naturally to this setup because they rely only on local information. We
shall discuss in Sec. IX-E their use for different tasks such as distributed denoising,
reconstruction, and consensus.

3 Graph Convolutional Filters

The convolution is a key operation in both SP and ML. In SP, convolution helps under-
standing linear, time-invariant systems and defines filtering operations. In ML, convolu-
tional filters are the basic building blocks of CNNs, and their computational efficiency and
parameter-sharing property tackle the curse of dimensionality. Convolutions also leverage
the symmetries in the domain (such as translations in time or space) and allow for a de-
gree of mathematical tractability with respect to domain perturbation [50]. Following first
principles, we generalize here the convolutional filter to the graph domain with the goal of
inheriting the above properties.
126 Graphs Signals and Systems

3.1 Convolutional Filters

A convolutional filter is defined by a shift-and-sum operation of the input signal [51]. While a
shift in time implies a delay, a graph signal shift requires taking into account the underlying
topological structure.

Definition 3.1: Graph signal shift.

A graph signal shift is a linear transformation S : XN → XN obtained from applying a
GSO S to a signal x, i.e. S(x) = Sx. The shifted signal at
node i is computed as
N
X X
[Sx]i = [S]ij x = sij xj , (1.3)
j=1 j∈Niin ∪{i}

which is a local linear combination of the signal values at neighboring nodes.

If the GSO is the adjacency matrix A, the shifted signal represents a one-step propagation.
Instead, if the GSO is the graph Laplcaian
P L, the shifted signal is a weighted difference of
the signals at neighboring nodes [Lx]i = j∈Ni aij (xi − xj ).
Upon defining the signal shift, a graph convolutional filter is simply a weighted sum of
multiple shifted signals.

Definition 3.2: Graph convolutional filter.

Given a set of parameters h = [h0 , . . . , hK ]⊤ , a graph convolutional filter of order K is
a linear mapping H : XN → XN comprising a linear combination of K shifted signals
K
X
H(x) = hk Sk x = H(S)x (1.4)
k=0

where H(S) = hk Sk is the N × N polynomial filtering matrix.

PK
k=0

The output signal at node i is yi = h0 xi + h1 [Sx]i + . . . + hK SK x i , which is a linear

 

combination of signal values located at most up to K-hops away. This is because Sk ji ̸= 0

 

implies that there exists at least one path of length k between nodes i and j through which
the signals can diffuse. These signals are shifted repeatedly over the graph as per (1.3); see
also 1.1. The term convolution for (1.4) is rooted in the algebraic extension of the convolution
operation [20] and the discrete-time counterpart can be seen as a particular case over a cyclic
graph; see the box of 1.2 Because of this analogy, the filter in (1.4) is also referred to as a
finite impulse response graph filter of order K.
Graphs Signals and Systems 127

Discrete-time circular convolution. The graph signal shift (1.3), the graph convo-
lutional filter (1.4), and their spectral equivalents in Sec. IV generalize the respective
concepts developed for discrete-time periodic signals.

Figure 1.2: Discrete-time periodic signals as graph signals over a directed cycle graph.
Each node Vc = {1, . . . , 6} is a time instant with adjacencies captured in the matrix
Ac . The temporal signal forms the graph signal x = [x1 , . . . , x6 ]⊤ and the shift Ac x
acts as a delay operation that moves the signal to the next time instant node.

We can represent an arbitrary discrete-time periodic signal as a graph signal x =

[x1 , . . . , xN ]⊤ residing over the vertices of a directed cyclic graph Gc = (Vc , Ec ) in which
each node is a time instant and directed edges connect adjacent time instances of the
form (n, 1 + n mod N ), n = 1, . . . , N , as shown in 1.2 The adjacency matrix of this
graph is a cyclic matrix Ac such that [Ac ]1+n mod N, n = 1 and zeros everywhere else.
Signal shift: Setting the GSO S = Ac , operation (1.3) shifts the signal cyclically and
acts as a delay operation, i.e., [Ac x]1+n mod N = xn .
Convolutional filter : The graph convolutional filter (1.4) for graph Gc reduces to the
circular convolution, i.e., the output at the temporal node i is yn = [H (Ac ) x]n =
k=0 hk [x]1+(n−k+1) mod N .
PK
Signal variation: Using the total variation in (8), we measure how much the signal
changes from its delayed version. This is a key building block for developing filters in
standard signal processing [52].
Fourier transform: The cyclic adjacency matrix can be eigendecomposed √ as Ac =
DFTN diag(λ)DFT−1 N with eigenvectors [DFT] ]
N kn = (1/ N ) expj2π(k−1)(n−1)/N

forming the discrete Fourier transform (DFT) matrix and eigenvalues λ =

[exp(−j2π0/N, . . . , −j2π(N − 1)/N )] containing the frequencies. The DFT for sig-
nal x is x̃ = DFTN x, which coincides with the graph Fourier transform (GFT) for this
particular graph.

3.2 Properties
Graph convolutional filters satisfy the following properties.
Property 1 (Linear). The convolution is linear in the input signal. For two inputs x1 and
x2 and filter H(S) it holds that

αH(S)x1 + βH(S)x2 = H(S) (αx1 + βx2 )

128 Graphs Signals and Systems

where α, β are scalars; i.e., the linear combination of the outputs equals the output of the
inputs’ linear combination.
Property 2 (Shift invariance). The graph convolution is invariant to shifts, i.e.,
SH(S) = H(S)S. This implies that given two filters H1 (S) and H2 (S) with respective
parameters h1 and h2 operating over the same graph and input signal x, it holds that we
can switch the filters order

H1 (S)H2 (S)x = H2 (S)H1 (S)x

Property 3 (Permutation equivariance). Graph convolutions are equivariant to per-
mutations in the graph support. Specifically, denote the set of permutation matrices by

P = P ∈ {0, 1}N ×N : P1 = 1, P⊤ 1 = 1


Then for a graph with GSO S and P ∈ P, the permuted graph has the GSO Ŝ = P⊤ SP,
which describes the same topology but with a reordering of the nodes. Likewise, the permuted
signal corresponding to the ordering in Ŝ is x̂ = P⊤ x. Permutation equivariance for filter (1.4)
implies

H(Ŝ)x̂ = P⊤ H(S)x
i.e., the filter output operating on the permuted graph Ŝ with the permuted signal x̂ is
the permuted output of the same filter operating on the original graph S with the original
signal x.
Thus, graph convolutions are independent of the arbitrary ordering of the nodes. And
as temporal and spatial convolutions exploit symmetries in the Euclidean domain by being
translation equivariant, graph convolutions exploit symmetries in the graph domain by being
permutation equivariant. This is key to their success in learning input-output mappings from
a few training samples [4].
Property 4 (Parameter sharing). All the nodes share the parameters  K among them. For
two nodes i, j, the respective outputs are yi = h0 xi +h1 [Sx]i +. . .+hK S x i and yj = h0 xj +
h1 [Sx]j + . . . + hK SK x j , which shows that the k-shifted signal Sk x is weighted by the same
 

parameter hk .
Props. 364 imply that graph convolutions are inductive processing architectures. That is,
they can be designed or trained over a graph G and transferred to another graph G (with
possibly a different number of nodes) without redesigning or retraining for this specific graph
[53]. This is particularly useful, e.g., when using graph filters for distributed SP tasks, as
the physical channel graph may change. In Sec. IV (Prop. 8), we characterize the degree of
transference for specific GCFs.
Property 5 (Locality). Graph convolutions are local architectures. To see this, set
z(0) = S0 x. The one shifted signal z(1) = Sx = Sz(0) is local
by definition. The k > 1 shift
z(k) = Sk x can be computed recursively as z(k) = S S(k−1) x = Sz(k−1) , which implies that
the (k − 1) st shift z(k−1) needs to be shifted locally to the neighbors. Hence, to compute the
filter output, each node exchanges locally with neighbors all K shifts z(0) , . . . , z(K−1) .
Locality of computation makes the graph convolutional filters readily distributable, as we
discuss in Sec. IX-E
Graphs Signals and Systems 129

Property 6 (Linear computation cost). Graph convolutions have a computational

complexity of order O(K|E| + KN ); i.e., linear in the number of edges in the graph. Each
signal shift z(k) = Sz(k−1) incurs a complexity order of O(|E|) and each needs to be multiplied
with a scalar hk , incurring an additional cost of order O(N ). Since these operations have to
be computed K times, the cost of obtaining the intermediate outputs z(1) , . . . , z(K) needs to
be scaled by K.
Props. 4/6 imply that graph convolutions are effective solutions to tackle the curse of
dimensionality in large graphs. Their parameter sharing property makes them suitable ar-
chitectures
to learn input-output mappings from a few training samples, irrespective of the graph di-
mensions (i.e., O(K) number of parameters); their locality allows them to extract patterns
in the data in the surrounding of a node, and; their linear computational complexity facili-
tates scalability. These properties and the flexibility to learn from multiresolution neighbors
make graph convolutions suitable linear solutions for ML tasks. In Sec. VIII we discuss how
to learn more expressive representations via neural network solutions while preserving these
benefits in a form akin to CNNs for time series data and images.

4 Graph Fourier Transform

The DFT can be seen as the projection of a temporal signal onto the eigenvectors of the
cyclic graph adjacency matrix; box Fig. 2. We define the graph Fourier transform (GFT) by
following a similar strategy.
Graph Fourier transform (GFT). Consider the eigendecomposition of the GSO
S = V diag(λ)V−1 with eigenvectors V = [v1 , . . . , vN ] and the corresponding, possibly
complex, eigenvalues λ = [λ1 , . . . , λN ]. The GFT of a signal x is defined as the signal
projection onto the GSO eigenspace

x̃ = V−1 x
and likewise, the inverse GFT is x = Vx̃.
In the definition of GFT, we are assuming the GSO S is diagonalizable. While definitions
of GFT for nondiagonalizable GSOs exist [22], [54], [55], we hold to the diagonalizability
assumption for a consistent and simple exposition.
The eigenvectors of S in the columns of V serve as the basis expansion for the GFT. In the
discrete-time case, the complex exponentials fulfill this role. The coefficients x̃ are the weights
each of these eigenvectors contribute to represent the signal. Following again this analogy,
the vector λ contains the so-called graph frequencies. Interpreting these graph frequencies λ
and the respective GFT coefficients x̃ requires understanding how the signal varies over the
graph. In turn, measuring variability requires defining a criterion that accounts for the graph
structure. Here, we review two basic criteria used for undirected [19] and directed graphs
[21], [22].
Undirected graphs. The local signal difference at two connected nodes i and j can be
√
computed as ∆xij = aij (xi − xj ), which is higher when the signal in two strongly connected
nodes is substantially different. The local signal variability at a node i accounts for the
variabilities in its neighboorhood
130 Graphs Signals and Systems

sX ! 21
X
∆xi = (∆xij )2 = aij (xi − xj )2
j∈Ni j∈Ni

which is higher if the signal at node i differs from that of its neighboors. Then, the
variability of a signal x over an undirected graph is the squared-sum of all local variabilities
1X
TV2 (x) = (∆xi )2 = x⊤ Lx
2 i∈V
The TV2 (x) is the Laplacian quadratic form for signal x and quantifies how much the
signal varies over the graph. In fact, the constant graph signal x = c1 has a zero variability.
We can use the quadratic form (6) to interpret the variability of the Laplacian eigenvectors
L = V diag(λ)V⊤ . Treating each eigenvector vi as a graph signal, we have

TV2 (vi ) = vi⊤ Lvi = λi

Thus, we can sort eigenvectors based on their variability 0 = TV2 (v1 ) ≤ TV2 (v2 ) ≤ . . . ≤
TV2 (vN ), which implies that the respective eigenvalues 0 ≤ λ1 ≤ λ2 ≤ . . . ≤ λN carry the
notion of frequency in the graph setting. We refer to the eigenvalues λi close to 0 as low
frequencies and to those λi ≫ 0 as high frequencies. The lowest graph frequency is λ1 = 0
which corresponds to a constant eigenvector. Accordingly, the GFT coefficient x̃i indicates
how much eigenvector vi contributes to the variability of signal x over G.
Directed graphs. To measure the signal variability for directed graphs, we use again the
analogy with the cyclic graph representing time signals; Fig. 2 . I.e., we measure how much
the diffused signal Sx has changed w.r.t. signal x as

TV1 (x) = ∥x − Sx∥1 := x − |λmax |−1 Ax 1

−1
where the shift operator is set to S = |λmax | A with λmax being the maximum eigenvalue
1
Expression (8) is known as the total variation of a graph signal and attains a high value if
the shifted signal differs more from the original one. However, unlike the quadratic measure
for undirected graphs (6), the total variation in 8 is non-zero for constant signals.
We can now measure the variability of the adjacency matrix eigenvectors A = V diag(λ)V−1 .
We have that TV1 (vi ) ≤ TV1 (vj ) iff |λmax − λi | < |λmax − λj |. That is, the eigenvector as-
sociated to the largest eigenvalue has the slowest variability, while the eigenvector associated
to the eigenvalue farthest from λmax has the highest variability. Since the eigenvalues may
be complex, the distances have to be computed in the complex plane. The order of the
eigenvalues according to increasing variability is Re {λ1 } ≥ Re {λ2 } ≥ · · · ≥ Re {λN }, see
[22. Figs. 2, 3]. In this case, the eigenvalues located (in a complexplane sense) closest to
the largest real eigenvalue are the ones corresponding to lower frequencies, while eigenvalues
located farthest from it, are the highest frequency ones.
Either on a directed or an undirected graph, the variability of a graph signal x can be
often expressed by only a few K ≪ N GFT basis vectors v1 , . . . , vN . In this case, we say
the graph signal is K-bandlimited and expand it as

x = VK x̃K
Graphs Signals and Systems 131

with VK = [v1 , . . . , vK ] and x̃K ∈ RK . Without loss of generality, we assume the first K
⊤
eigenvectors express the signal variability, i.e., the GFT coefficients are x̃ = x̃⊤ ⊤
.

K , 0N −K

Fig. 3: The frequency response of the filter 12], given in the black solid line, is completely
characterized by the values of the filter parameters h. Given a graph, this frequency response
gets instantiated on the specific eigenvalues of that graph, determining the effect the filter
will have on an input 11.

5 Frequency Response
By subsituting the eigendecomposition S = V diag(λ)V−1 into (3), we can write the filter
output as

K
X K
X
k
hk V diag λ⊙k V−1 x


y= hk S x =
k=0 k=0

where λ⊙k ∈ CN : λ⊙k i := λki . Using then (4) and defining the GFT of the output
 

ỹ := V−1 y, we can write the filter inputoutput spectral relation as

K
X
hk diag λ⊙k x̃


ỹ =
k=0

Convolution theorem for graph filters. It follows from (11) that any shift-and-sum convolu-
tional graph filter of the form (3) operates in the spectral domain as a pointwise multiplication
ỹi = h̃ (λi ) x̃i between the input signal GFT x̃ = V−1 x and the filter frequency response

K
X
h̃(λ) = hk λk
k=1

01
Normalizing the adjacency matrix is made for pure technical reasons to prevent excessive scaling of the
shifted signal [22].
132 Graphs Signals and Systems

Such a result is reminiscent of the convolution theorem [52], whereby the convolution in the
graph domain, translates into a multiplication in the frequency domain. The filter frequency
response is an analytic polynomial in λ and it is independent of the graph. That is, for fixed
filter parameters, the filter effect on all graphs is already set. The specific filter effect on a
given graph is on the positions where the frequency response is instantiated; see Fig. 3. This
allows anticipating how the frequency response will behave for a wide range of graphs.
In this context, graph convolutional filters satisfy the following properties in the spectral
domain.
Property 7 (GFT of the filter). Eq. 11) can be rewritten as

ỹ = diag(h̃)x̃ with h̃ = Λh
where Λ ∈ CN ×(K+1) is a Vandermonde matrix such that [Λ]ik = λk−1 i . The vector
h̃ ∈ CN is known as the GFT of the filter parameters. Unlike traditional DSP, the convolution
operation is not commutative, and thus the input signal is mathematically (and conceptually)
different from the filter parameters. This becomes evident in the fact that the GFT of the
signal depends on the eigenvectors of the GSO, while the GFT of the filter on the eigenvalues.
Property 8 (Lipschitz continuity to changes in S). Let S, Ŝ ∈ RN ×N be two GSOs, po-
tentially corresponding to different graphs with the same number of nodes N . Define the
relative difference of Ŝ with respect to S as

d(Ŝ; S) = min ∥E∥

E∈R(Ŝ;S)
n o
for R(Ŝ; S) = E : P⊤ ŜP = S + (ES + SE), P∈P the set containing all the relative
difference matrices E. Let the frequency response of the filter [cf. (12]] satisfy λh̃′ (λ) ≤ C
for some C < ∞ and where h̃′ (λ) is the derivative of (12). Then, it holds that
√
∥H(Ŝ)x − H(S)x∥2 ≤ ε(1 + 8 N )C∥x∥2 + O ε2

for ε > 0 such that d(Ŝ; S) ≤ ε. Thus, if the relative difference between two GSOs is
small, the output of the filters to the same input signal will also be small.
Filters whose frequency response satisfies λh̃′ (λ) ≤ C are known as integral Lipschitz
filters. These filters may exhibit high variability for low values of λ (because its derivative
can be high), but they have to be approximately constant for high values of λ (because its
derivative has to be small). An example is shown in Fig. 3. For finite graphs with finite edge
weights, all convolutional filters [cf. (3)] are integral Lipschitz within the spectrum interval of
interest but the constant C may be large. This constant depends only on the filter parameters
and, thus, filters can be designed or learned to have a small value of C guaranteeing a tighter
bound; see [56] for details.
 Part III

Third part: An Introduction to

Multidimensionsal Signals and Systems

133
135
136 Introduction to Multidimensional Signals and Systems

CHAPTER 1
INTRODUCTION TO
MULTIDIMENSIONAL SIGNALS AND
SYSTEMS

The theory of signals and systems is a core

1 . . . . . . . . . . . . . . . . . 136
subject in engineering, providing tools to an-
alyze and design complex systems, model
natural processes, and describe economic in-
terdependencies. Its strength lies in math-
ematical abstractions like differential equa-
tions, impulse responses, and transfer func-
tions, enabling efficient system analysis and
signal processing. Traditionally, it focuses
on one-dimensional, time-dependent signals,
relevant to fields like communications, con-
trol, and biology. However, many signals de-
pend on multiple variables, such as images
(two spatial dimensions), video (space and
time), and wave propagation (spatial dimen-
sions and time).
Multidimensional systems require advanced
mathematical tools like gradients, diver-
gence, and integrals, which underpin physics
and engineering. While early applications
used analog methods (e.g., photography,
television), digital computing has expanded
multidimensional signal processing across ar-
eas like imaging, video, and control systems.
Despite its broad scope, literature on multi-
dimensional systems is fragmented into spe-
cialized fields and lacks the coherence of one-
dimensional theory.
This section introduces multidimensional
signals and systems, building on one-
dimensional concepts and covering discrete
and continuous systems, with methods for
practical implementation. While theory-
 BIBLIOGRAPHY

Rabenstein et al.: Multidimensional Signals and Systems: Theory and Founda-

tions rabenstein2023multidimensional
Rudolf Rabenstein and Maximilian Schäfer. Multidimensional Signals and Systems: Theory
and Foundations. Springer Nature, 2023.

Isufi et al.: Graph filters for signal processing and machine learning on graphs
isufi2024graph
Elvin Isufi et al. “Graph filters for signal processing and machine learning on graphs”. In:
IEEE Transactions on Signal Processing (2024).

137