EPFL - EE-406 - Fundamentals of Electrical

Circuits and Systems I

Prof. Jean-Philippe Thiran
Lecture notes — Fall 2025
2
Contents

1 Introduction 7
1.1 Preamble . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.2 Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.3 Practical Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2 Signals and Systems 9

2.1 Signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.1.1 Continuous-time versus Discrete-time . . . . . . . . . . . . . . . . . . 9
2.1.2 Continuous-amplitude versus Discrete-amplitude . . . . . . . . . . . . . 9
2.1.3 Periodic Signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.1.4 The Energy of a Signal . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.1.5 The Power of a Signal . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.2 Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.3 Basic Properties of Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.3.1 Linearity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.3.2 Time-Invariance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.3.3 Memory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.3.4 Invertibility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.3.5 Causality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.3.6 Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

3 Linear Time-Invariant Systems 15

3.1 The Impulse Response (discrete-time) . . . . . . . . . . . . . . . . . . . . . . 15
3.2 The Impulse Response (continuous-time) . . . . . . . . . . . . . . . . . . . . . 17
3.3 The Convolution Operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.3.1 Convolution with the Delta Function . . . . . . . . . . . . . . . . . . . 18
3.3.2 The Commutative Property . . . . . . . . . . . . . . . . . . . . . . . . 19
3.3.3 The Distributive Property . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.3.4 The Associative Property . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.3.5 Techniques to Evaluate Convolution Sums and Integrals . . . . . . . . 20
3.4 Composition of LTI Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.4.1 Parallel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.4.2 Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

3
4 CONTENTS

3.5 Properties of LTI Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

3.5.1 Memory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.5.2 Invertibility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.5.3 Causality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.5.4 Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.6 Systems Modeled by Differential Equations . . . . . . . . . . . . . . . . . . . . 22
3.A Exchanging Integration and Summation Order . . . . . . . . . . . . . . . . . . 23

4 Fourier Methods for Stable LTI Systems 25

4.1 Continuous-time Signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
4.1.1 The Fourier Transform . . . . . . . . . . . . . . . . . . . . . . . . . . 26
4.1.2 Integrating a Complex Exponential . . . . . . . . . . . . . . . . . . . . 26
4.1.3 Properties of the Fourier Transform . . . . . . . . . . . . . . . . . . . 27
4.1.4 LTI Systems and the Fourier Transform . . . . . . . . . . . . . . . . . 27
4.1.5 LTI Systems and the Fourier Transform: Differential Equations . . . . . 28
4.1.6 Delta Functions and the Fourier Transform . . . . . . . . . . . . . . . 29
4.2 Discrete-time Signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
4.2.1 The Discrete-time Fourier Transform . . . . . . . . . . . . . . . . . . . 30
4.2.2 Properties of the Discrete-time Fourier Transform . . . . . . . . . . . . 30
4.2.3 LTI Systems and the Discrete-time Fourier Transform . . . . . . . . . . 31
4.2.4 LTI Systems and the Discrete-time Fourier Transform: Difference Equations 31
4.2.5 Delta Functions and the Discrete-time Fourier Transform . . . . . . . . 32
4.3 Convergence Issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
4.A Continuous-Time Fourier Transform : Properties . . . . . . . . . . . . . . . . . 34
4.B Continuous-Time Fourier Transform : Pairs . . . . . . . . . . . . . . . . . . . 35
4.C Discrete-Time Fourier Transform : Properties . . . . . . . . . . . . . . . . . . 36
4.D Discrete-Time Fourier Transform : Pairs . . . . . . . . . . . . . . . . . . . . . 37
4.E Continuous-time Periodic Signals . . . . . . . . . . . . . . . . . . . . . . . . . 38
4.E.1 The Fourier Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
4.E.2 Properties of the Fourier Series . . . . . . . . . . . . . . . . . . . . . . 38
4.E.3 LTI Systems and Fourier Series . . . . . . . . . . . . . . . . . . . . . . 39
4.F Continuous-Time Fourier Series : Properties . . . . . . . . . . . . . . . . . . . 40
4.G Continuous-Time Fourier Series : Pairs . . . . . . . . . . . . . . . . . . . . . . 41

5 The Frequency Response of Stable LTI Systems 43

5.1 Continuous-time Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
5.1.1 The Frequency Response . . . . . . . . . . . . . . . . . . . . . . . . . 43
5.1.2 Interpretation of the Frequency Response . . . . . . . . . . . . . . . . 44
5.1.3 Basic Properties of the Frequency Response . . . . . . . . . . . . . . . 45
5.1.4 Why do we consider complex-valued signals? . . . . . . . . . . . . . . 45
5.1.5 Composition of Systems . . . . . . . . . . . . . . . . . . . . . . . . . . 46
5.2 Discrete-time Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
5.2.1 The Frequency Response . . . . . . . . . . . . . . . . . . . . . . . . . 47
5.2.2 Interpretation of the Frequency Response . . . . . . . . . . . . . . . . 47
CONTENTS 5

5.2.3 Basic Properties of the Frequency Response . . . . . . . . . . . . . . . 47

5.2.4 Composition of Systems . . . . . . . . . . . . . . . . . . . . . . . . . . 48
5.3 Sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
5.3.1 Impulse-Train Sampling . . . . . . . . . . . . . . . . . . . . . . . . . . 49
5.3.2 The Sampling Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . 50
5.3.3 Sampling a Complex Exponential . . . . . . . . . . . . . . . . . . . . . 50
5.3.4 Signal Reconstruction and Aliasing . . . . . . . . . . . . . . . . . . . . 51

6 The Transfer Function and The Laplace Transform 53

6.1 The Transfer Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
6.2 The Laplace Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
6.2.1 Key Example 1: Single Real-Valued Pole . . . . . . . . . . . . . . . . . 55
6.2.2 Key Example 2: Complex-Valued Pole . . . . . . . . . . . . . . . . . . 58
6.2.3 Key Example 3: Multiple Poles . . . . . . . . . . . . . . . . . . . . . . 60
6.3 LTI Systems and The Laplace Transform . . . . . . . . . . . . . . . . . . . . . 62
6.4 LTI Systems and The Laplace Transform : Composition . . . . . . . . . . . . . 63
6.5 LTI Systems and The Laplace Transform : Differential Equations . . . . . . . . 64
6.6 Control Systems : Stability and Causality . . . . . . . . . . . . . . . . . . . . 65
6.A Laplace Transform : Properties . . . . . . . . . . . . . . . . . . . . . . . . . . 68
6.B Laplace Transform : Pairs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
6 CONTENTS
Chapter 1

Introduction

1.1 Preamble
This document serves as support material for the EPFL course titled Fundamentals of
electrical circuits and systems I (EE-406).
Most important warning: These lecture notes serve only as a foundation so that
we all have the same definitions and main theorems. They do not contain explanations or
examples, which will be provided in class.
We will closely follow the reference book by Oppenheim and Willsky [1]. If you wish to
access more details and a deeper treatment of this subject, you may refer to [2].
This course is one of the important pillars for building your understanding and further
exploration of many areas of electrical and electronic engineering and energy.

1.2 Acknowledgements
This document was initially written in English by M. Gastpar and edited by Y. Shkel, to
whom I extend my sincere thanks for their help.

1.3 Practical Information

Instructor: Prof. Jean-Philippe Thiran, [Link]@[Link], EPFL - Signal
Processing Laboratory (LTS5) - Office: ELD 240

Teaching Assistants and Student Assistants: See slides and Moodle

Course Web Page: We will use [Link]

7
8 CHAPTER 1. INTRODUCTION
Chapter 2

Signals and Systems

2.1 Signals
2.1.1 Continuous-time versus Discrete-time
As you have seen in your earlier classes, we can model time in two rather different ways, as
continuous or as discrete.
A continuous-time signal is a function x(t) that is defined for all times t ∈ R. By
constrast, for a discrete-time signal, we model time as discrete ticks — for example, every
second, every nanosecond, or every hour. A discrete-time signal is then a sequence x[n]
defined for all integers n. In this class, we will use the square brackets x[n] to denote
discrete-time signals.
We should also emphasize that while we often refer to t and n as “time,” they may
equally well represent space (or any other data indexing).

2.1.2 Continuous-amplitude versus Discrete-amplitude

In most of this class, we will model signals as having continuous values. That is, the signal
x(t) or x[n] is assumed to be real- or complex-valued.
Of course, in practice, we also sometimes encounter signals that are known to only attain
values in a certain fixed, discrete set (for example, the signal might be only either 0 or 1).
This is referred to as a discrete-amplitude signal.

2.1.3 Periodic Signals

Periodic signals play a key role for many reasons. From a technological perspective, there
are many ways to generate such signals.
A continuous-time signal x(t) is called periodic with period T (where T > 0) if for all
times t, it satisfies
x(t) = x(t + T ). (2.1)
The fundamental period of a periodic signal x(t) is the smallest value of T (where T > 0)
such that x(t) = x(t + T ) holds for all times t.

9
10 CHAPTER 2. SIGNALS AND SYSTEMS

A discrete-time signal x[n] is called periodic with period N (where N is a positive

integer) if for all times n, it satisfies

x[n] = x[n + N ]. (2.2)

The fundamental period of a periodic signal x[n] is the smallest positive integer N such that
x[n] = x[n + N ] holds for all times n.

2.1.4 The Energy of a Signal

For any continuous-time signal x(t), we define its energy by
Z ∞
E = |x(t)|2 dt. (2.3)
−∞

If this integral is finite, then x(t) is referred to as an “energy signal.”

Similarly, for a discrete-time signal x[n], we define its energy by
∞
X
E = |x[n]|2 . (2.4)
n=−∞

If this infinite sum is finite, then x[n] is referred to as an “energy signal.”

2.1.5 The Power of a Signal

Clearly, for every non-zero periodic signal, the energy must be infinite. For those signals, a
more interesting quantity would be the energy per period. More generally, for any signal, it
is interesting to calculate the (average) energy per time unit, which in physics is called the
power.
For general signals x(t), we formally define the power as
Z T
1
P = lim |x(t)|2 dt. (2.5)
T →∞ 2T −T

If this limit is finite (and strictly larger than zero), then the signal x(t) is referred to as a
“power signal.”
Similarly, for a discrete-time signal x[n], we define its power by

N
1 X
P = lim |x[n]|2 . (2.6)
N →∞ 2N + 1
n=−N

If this limit is finite (and strictly larger than zero), then x[n] is referred to as a “power
signal.”
2.2. SYSTEMS 11

x(t) or x[n] y(t) or y[n]

System H
input signal output signal

Figure 2.1: A System.

2.2 Systems
A “system” takes a signal as its input and outputs a new signal. We will generally express
this as

y(t) = H{x(t)} or y[n] = H{x[n]}, (2.7)

where x(t) (or x[n]) is the input signal and y(t) (or y[n]) is the output signal.
We should alert the reader that both the picture (Figure 2.1) and the notation are
slightly misleading: To the untrained eye, it might look as if the particular output sample
at time n, which we denote by y[n], is obtained merely from the input sample at that
same point in time, denoted by x[n]. Obviously, this would not make for an interesting
framework! What is meant is that the input is a very long signal {x[k]}∞ k=−∞ , and the
system H is allowed to use the entire signal in order to produce the output y[n].

2.3 Basic Properties of Systems

The next goal is to classify systems. There are many properties that we could isolate. The
following six properties are classically considered to be the most powerful.

2.3.1 Linearity
A system H{·} is linear if for all signals x1 (t) and x2 (t) and for all constants a1 and a2 , it
is true that

H{a1 x1 (t) + a2 x2 (t)} = a1 H{x1 (t)} + a2 H{x2 (t)}, (2.8)

or, for the discrete-time case,

H{a1 x1 [n] + a2 x2 [n]} = a1 H{x1 [n]} + a2 H{x2 [n]}. (2.9)

Otherwise, the system is called non-linear.

Examples:
R∞ −|τ | x(t
• y[n] = 2x[n] and y(t) = −∞ e − τ )dτ are linear systems.
• y(t) = x2 (t) and y[n] = x[n] + 1 are non-linear systems.
12 CHAPTER 2. SIGNALS AND SYSTEMS

2.3.2 Time-Invariance
A system H{·} is time-invariant if for every signal x(t) and for every real number τ, it is
true that

if system input x(t) produces system output y(t)

then system input x(t − τ ) produces system output y(t − τ ).

In discrete time, it has to hold that for every signal x[n] and every integer number n0 ,
it is true that

if system input x[n] produces system output y[n]

then system input x[n − n0 ] produces system output y[n − n0 ].

Otherwise, the system is called time-variant.

Examples:
• The system y(t) = x(t) + 21 x(t − 23 ) + 13 x(t − 3) is time-invariant.
• The system y(t) = x(2t) is time-variant.

2.3.3 Memory
A system is called memoryless if the current system output only depends on the current
system input (for any given point in time).
More formally, a system H{·} is memoryless if for all signals x(t) and for every time
instant t, the output of the system at time t, y(t) = H{x(t)}, is independent of the input
signal values x(τ ) for all values of τ ̸= t.
By the same token, discrete-time systems are called memoryless if for all signals x[n] and
for every time instant n, the output of the system at time n, y[n] = H{x[n]}, is independent
of the input signal values x[k] for all values of k ̸= n.
Otherwise, the system is said to have memory.

Examples:
• y(t) = x2 (t) cos(4t − 3) is memoryless.
R∞
• The system y(t) = −∞ e−|τ | x(t − τ )dτ has memory.

It is often interesting to also quantify the amount of memory that a system has. For
discrete-time systems, this is easy to do. Consider the system y[n] = (x[n] − 12 x[n −
1])2 cos(x[n − 2]). Here, a good definition is to say that the system has memory 2 : At any
given point in time, the system holds the two previous inputs “in storage.”

2.3.4 Invertibility
A system is called invertible if distinct inputs lead to distinct outputs.
More formally, a system H{·} is invertible if there exists a system G{·} such that for all
signals x(t), we have G{H{x(t)}} = x(t). For discrete-time systems, the definition is exactly
2.3. BASIC PROPERTIES OF SYSTEMS 13

the same, with t replaced by n.

Examples:
• The system y(t) = 2x(t) is invertible, and the inverse system is z(t) = 21 y(t).
• The system y[n] = x2 [n] is not invertible.

2.3.5 Causality
A system is causal if its output signal only depends on present and past inputs, but not on
future inputs.
More formally, a system H{·} is causal if for all signals x(t) and for all times τ, the
output y(t) = H{x(t)} evaluated at time τ, y(τ ), only depends on the input signal x(t) for
times −∞ < t ≤ τ.
In discrete time, it has to hold that for all input signals x[n] and all times n0 , the output
y[n] = H{x[n]} evaluated at time n0 , y[n0 ], only depends on the input signal x[n] for times
−∞ < n ≤ n0 .
Otherwise, the system is called non-causal.

Examples:
• The system y(t) = x(t) + 21 x(t − 23 ) + 13 x(t − 3) is causal.
• The system y(t) = x(t) + 21 x(t + 23 ) + 13 x(t − 3) is not causal.

Anti-Causality
In engineering, it is also convenient to have a notion of anti-causal systems. In such a
system, for all signals x(t) and for all times τ, the output y(t) = H{x(t)} evaluated at time
τ, y(τ ), only depends on the input signal x(t) for times τ ≤ t < ∞.

Examples:
• The system y(t) = x(t) + 21 x(t + 23 ) + 13 x(t + 3) is anti-causal.
• The system y(t) = x(t) + 21 x(t + 23 ) + 13 x(t − 3) is not anti-causal (nor causal).

2.3.6 Stability
To define stability, we first need the notion of a bounded signal: A signal x(t) is called
bounded if there exists a positive constant B < ∞ such that |x(t)| ≤ B for all times t. For
discrete-time systems, the definition is exactly the same, with t replaced by n.
A system H{·} is stable if for all bounded input signals x(t), the corresponding output
signal y(t) = H{x(t)} is also bounded. For discrete-time systems, the definition is exactly
the same, with t replaced by n.
Otherwise, the system is called unstable.
This notion of stability is often referred to as “BIBO Stability” (bounded-input bounded-
output) and will be the only notion of stability discussed in this class.

Examples:
14 CHAPTER 2. SIGNALS AND SYSTEMS

• The system y(t) = 1010 x(t) is stable

• The system y(t) = 1/x(t) is not stable.
Chapter 3

Linear Time-Invariant Systems

For the remainder of the class, we will restrict attention to systems that are linear and
time-invariant, often referred to as LTI systems.

3.1 The Impulse Response (discrete-time)

In this section, we derive the fact that LTI systems can be characterized in a convenient
way via their so-called impulse response. Let us first consider a discrete-time LTI system
with input x[n]. In full generality, the output can be expressed as y[n] = H{x[n]}. Now we
want to exploit the fact that the system is assumed to be both linear and time-invariant in
order to simplify this. We start with a trick. For each integer k, we introduce the function

1, for n = k,
δk [n] = (3.1)
0, otherwise.

Then, if we denote xk = x[k], we can express the signal x[n] equivalently as

∞
X
x[n] = xk δk [n]. (3.2)
k=−∞

−2 −1 1 2 3 4 k−1 k k+1

Figure 3.1: The Kronecker-delta function δk [n].

15
16 CHAPTER 3. LINEAR TIME-INVARIANT SYSTEMS

Now, we can exploit the linearity of the system, as follows:

( ∞ )
X
y[n] = H xk δk [n] (3.3)
k=−∞
∞
(∗) X
= xk H {δk [n]} , (3.4)
k=−∞

where step (∗) is due to the linearity. For the next step, it is better to switch to a more
standard notation. Let us define the so-called Kronecker-delta function:

1, for n = 0,
δ[n] = (3.5)
0, otherwise.

With this, we can express δk [n] = δ[n − k]. Therefore, we can equivalently express Equa-
tion (3.2) as
∞
X
x[n] = x[k]δ[n − k]. (3.6)
k=−∞

and Equation (3.4) as

∞
X
y[n] = x[k]H {δ[n − k]} . (3.7)
k=−∞

To proceed, we next define the signal

h[n] = H {δ[n]} , (3.8)

which is simply the system response when the input is the Kronecker-delta function δ[n].
Therefore, the signal h[n] is called the impulse response of the system H{·}. Next, we use
the fact that the system is time-invariant. Specifically, time invariance means that the
system reponse to a shifted Kronecker-delta function δ[n − k] is precisely

H {δ[n − k]} = h[n − k]. (3.9)

This means that we can equivalently characterize the system output signal as
∞
X
y[n] = x[k]h[n − k]. (3.10)
k=−∞

The fundamental upshot is that any LTI system is uniquely characterized by its impulse
response h[n]. Moreover, given the impulse response, it is easy to determine the system
output for any input signal x[n] using the formula given in Equation (3.10). This formula
is important enough to have its own name: It is referred to as the convolution sum (or the
convolution operation).
3.2. THE IMPULSE RESPONSE (CONTINUOUS-TIME) 17

3.2 The Impulse Response (continuous-time)

Now, let us turn to continuous-time systems. The key step is to express the input signal
in the form given in Equation (3.6). That is, we are looking for a “function”R∞ δ(t) that
satisfies, for any signal x(t) that is continuous at t, the identity x(t) = −∞ x(τ )δ(t − τ )dτ.
Unfortunately, this is not quite as straightforward as in the discrete-time case. For example,
defining δ(t) = 1 for t = 0, and δ(t) = 0 otherwise (like in the discrete-time case), clearly
does not satisfy the desired identity. In fact, in a similar fashion, it is easy to show that no
“standard” function can hope to satisfy the desired identity.
By generalizing the notion of functions, one can indeed define a mathematical object
δ(t), called the (Dirac) delta function, satisfying
Z ∞
x(t) = x(τ )δ(t − τ )dτ (3.11)
−∞

for all t at which x(t) is continuous. Regarding δ(t) itself, we can say that it has to be equal
to zero whenever t ̸= 0. At t = 0, its value is “infinite” in the sense that Equation (3.11)
ends up holding.
Using the Dirac delta function, we can again proceed as follows:
Z ∞ 
y(t) = H x(τ )δ(t − τ )dτ (3.12)
−∞
Z ∞
(∗)
= x(τ )H {δ(t − τ )} dτ, (3.13)
−∞

where step (∗) is again due to the linearity. To proceed, we assume that it is fine to define
the signal

h(t) = H {δ(t)} , (3.14)

which is simply the system response when the input is δ(t). Next, we use the fact that the
system is time-invariant. Specifically, time invariance means that the system’s reponse to
a shifted δ(t − τ ) is precisely

H {δ(t − τ )} = h(t − τ ). (3.15)

This means that we can equivalently characterize the system output signal as
Z ∞
y(t) = x(τ )h(t − τ )dτ. (3.16)
−∞

This is the powerful result that we are after: It again shows that any LTI system can
be characterized by its impulse response h(t), and that for any input signal x(t), the corre-
sponding system output signal y(t) can be found using the formula given in Equation (3.16).
This formula is referred to as the convolution integral (or the convolution operation).
The remaining problem is to make the steps leading up to the fundamental Equa-
tion (3.16) precise. An important observation is that the delta function δ(t) is no longer
18 CHAPTER 3. LINEAR TIME-INVARIANT SYSTEMS

present in this final expression. In order to fully accomplish the derivation, a mathematical
theory has been developed under the heading of “generalized functions” or “distributions.”
We encourage the mathematically inclined reader to study this in more detail. In class, we
present a more direct argument using the mathematical concept of limits.

3.3 The Convolution Operation

As we have seen, LTI systems are fully described by their impulse response: The output
signal is simply given by the convolution of the input signal with the impulse response.
In this section, we explore some aspects of convolution. In fact, convolution is important
enough to have its own symbol. In many texts, it is denoted as

∞
X
(x ∗ h)[n] = x[k]h[n − k], (3.17)
k=−∞

or
Z ∞
(x ∗ h)(t) = x(τ )h(t − τ )dτ, (3.18)
−∞

and in this class, we will follow this convention. However, in other texts, you will also see
convolution denoted as x[n] ∗ h[n] or x(t) ∗ y(t).

3.3.1 Convolution with the Delta Function

Let us now consider the important special case (x ∗ δ)[n], i.e., the convolution of a signal
x[n] with the Kronecker-delta function δ[n]. We find

∞
X
(x ∗ δ)[n] = x[k]δ[n − k] = x[n], (3.19)
k=−∞

since the function δ[n − k] is zero except if k = n, when it is one. That is, the convolution
of any signal with the delta function simply returns back that signal.
In continuous time, we can similarly study (x ∗ δ)(t), i.e., the convolution of a signal
x(t) with the Dirac delta function. We find
Z ∞
(x ∗ δ)(t) = x(τ )δ(t − τ )dτ = x(t), (3.20)
−∞

which is of course simply Equation (3.11) that we have studied earlier (and thus, it comes
with the same limitations; in particular, the last equality holds whenever x(t) is continuous).
That is, the convolution of any signal with the delta function simply returns back that signal.
3.3. THE CONVOLUTION OPERATION 19

3.3.2 The Commutative Property

A first interesting question is whether we can swap the order of the two signals, i.e., whether
we have that (x ∗ h)[n] = (h ∗ x)[n]. Let us write out:
∞ −∞
X (a) X
(x ∗ h)[n] = x[k]h[n − k] = x[n − k̃]h[k̃] (3.21)
k=−∞ k̃=∞
X∞
= x[n − k̃]h[k̃] = (h ∗ x)[n], (3.22)
k̃=−∞

where for the summation variable substitution in step (a), we have defined k̃ = n − k (and
we recall that n is a constant). Hence, the answer is positive: we can swap the order.
For the continuous-time case, the argument is essentially the same, as follows:
Z ∞ Z −∞
(b)
(x ∗ h)(t) = x(τ )h(t − τ )dτ = x(t − τ̃ )h(τ̃ )(−dτ̃ ) (3.23)
−∞ ∞
Z ∞ Z ∞
= − x(t − τ̃ )h(τ̃ )(−dτ̃ ) = x(t − τ̃ )h(τ̃ )dτ̃ (3.24)
−∞ −∞
= (h ∗ x)(t), (3.25)

where for the integration variable substitution in step (b), we have defined τ̃ = t − τ (and
we recall that t is a constant). Hence, again, the answer is positive: we can swap the order.

3.3.3 The Distributive Property

The next question is whether we have that

(x ∗ (h1 + h2 ))[n] = (x ∗ h1 )[n] + (x ∗ h2 )[n] (3.26)

or, in continuous time,

(x ∗ (h1 + h2 ))(t) = (x ∗ h1 )(t) + (x ∗ h2 )(t). (3.27)

For both cases, the answer is positive and can be verified in a straightforward way, as we
will show in class.

3.3.4 The Associative Property

Finally, we consider whether we have that

((x ∗ h1 ) ∗ h2 )[n] = (x ∗ (h1 ∗ h2 ))[n], (3.28)

or, in continuous time,

((x ∗ h1 ) ∗ h2 )(t) = (x ∗ (h1 ∗ h2 ))(t). (3.29)

20 CHAPTER 3. LINEAR TIME-INVARIANT SYSTEMS

That is, the question is whether multiple convolutions can be executed in an arbitrary
order. Here, it turns out that the answer is somewhat more P subtle. The property
P holds if
all three
P involved signals are absolutely summable (i.e., n∈Z |x[n]| < ∞, |h
n∈Z 1 [n]| <
∞, n∈Z |h2 [n]| < ∞) or, in the continuous-time case, absolutely integrable. This is a
sufficient condition for associativity to hold, but it is not necessary — associativity may
also hold in more general cases. But there are also simple cases where associativity does
not hold, see Appendix 3.A for one such example.

3.3.5 Techniques to Evaluate Convolution Sums and Integrals

There are several techniques to evaluate convolution. For a thorough understanding of LTI
systems, it is important to develop a certain intuition for several of them. We will do this
in class and in the Problem Sets. The techniques are:

• “graphical” convolution (also called “flip-and-drag”)

• analytical, i.e., simply by solving the integral (or the sum) in closed form

• numerical (i.e., using numerical software).

3.4 Composition of LTI Systems

A further reason why LTI systems are such powerful models is because when multiple LTI
systems are combined in several natural ways, the new overall system is again an LTI
system. The most common compositions will be discussed in the sequel.

3.4.1 Parallel
For two systems H1 and H2 in parallel composition (Figure 3.2), we can express the overall
system G simply as

y(t) = G{x(t)} = H1 {x(t)} + H2 {x(t)}. (3.30)

If both H1 and H2 are LTI systems, then G is also an LTI system. Moreover, the impulse
response g(t) (or g[n]) of the system G is simply given by

g(t) = h1 (t) + h2 (t) or g[n] = h1 [n] + h2 [n]. (3.31)

3.4.2 Series
For two systems H1 and H2 in series composition (Figure 3.3), we can express the overall
system G simply as

y(t) = G{x(t)} = H2 {H1 {x(t)}}. (3.32)

3.5. PROPERTIES OF LTI SYSTEMS 21

System H1

System H2

Figure 3.2: Parallel Composition of two systems.

If both H1 and H2 are LTI systems, then G is also an LTI system. Moreover, if both systems
are stable, the impulse response g(t) (or g[n]) of the system G is simply given by
Z ∞ ∞
X
g(t) = h1 (τ )h2 (t − τ )dτ, or g[n] = h1 [k]h2 [n − k]. (3.33)
−∞ k=−∞

System H1 System H2

Figure 3.3: Series Composition of two systems.

3.5 Properties of LTI Systems

We are studying systems that are both linear and time-invariant. In this section, we consider
the remaining three fundamental system properties. As we have seen, LTI systems are fully
described by their impulse response h(t) (or h[n]). Hence, our discussion will connect the
fundamental system properties to the particular shape of the impulse response.

3.5.1 Memory
An LTI system is memoryless if and only if, for some constant a, we have

y(t) = ax(t) or y[n] = ax[n]. (3.34)

What does this imply for the impulse response? Let us first take up the discrete-time case.
Here, using the definition of the Kronecker-delta function δ[n], we can see that we must
have

h[n] = aδ[n]. (3.35)

22 CHAPTER 3. LINEAR TIME-INVARIANT SYSTEMS

The situation is a little more involved in continous time. But using the concept of the delta
function δ(t) as discussed above, we can see that for a memoryless LTI system, we must
have

h(t) = aδ(t). (3.36)

3.5.2 Invertibility
An LTI system with impulse response h(t) is invertible if and only if there exists a function
g(t) such that (g ∗ h)(t) = δ(t). For discrete-time systems, the definition is exactly the same,
with t replaced by n.

3.5.3 Causality
An LTI system is causal if and only if the impulse response function is identically zero for
negative lags: h(t) = 0 for t < 0 (or h[n] = 0 for n < 0).

3.5.4 Stability
An LTI system is stable if and only if the impulse response is absolutely integrable (or
summable), i.e., if and only if
Z ∞ ∞
X
|h(t)|dt < ∞ or |h[k]| < ∞. (3.37)
−∞ k=−∞

A proof of this result will be developed in class.

3.6 Systems Modeled by Differential Equations

Many systems are naturally described by differential equations (for example, using the
laws of physics). A fundamental insight is that any system that is described by a linear
differential equation (with constant coefficients) constitutes an LTI system. To be more
precise, we also have to add that this is only true if the initial conditions for the system are
all zero. (If they are not zero, the system is still almost linear, and many of the techniques
that we will see in this class can be modified to be applied to these systems.)
We will start our discussion with the consideration of discrete-time systems. In this
case, differential equations are usually referred to as difference equations. Here is a simple
example of such a linear difference equation with constant coefficients:
1
y[n] + 3y[n − 1] = x[n] − x[n − 1]. (3.38)
2
Our goal is to argue that indeed, this describes an LTI system. Unfortunately, there is a
small hiccup that we need to discuss first. In particular, let us suppose that the input signal
x[n] is zero for all n < 0. Then, for n = 0, we obtain the relationship y[0] + 3y[−1] = x[0],
and we see that we also explicitly need to specify y[−1]. A tempting choice is to simply set
3.A. EXCHANGING INTEGRATION AND SUMMATION ORDER 23

y[−1] = 0. In this case, it is easy to verify that the resulting system is indeed both linear
and time-invariant. By the same reasoning, it can be seen that any system described by a
linear difference equation with constant coefficients and where all initial conditions are set
to zero describes an LTI system. For continuous-time systems, the analogous consideration
concerns differential equations like the ones you have seen in your Analysis class. As you
remember, to solve those equations, it was necessary to specify initial conditions. Again, if
the differential equation is a linear differential equation with constant coefficients and the
initial conditions are set to “rest” (meaning that all signals and derivatives that appear in
the equation are initially set to zero), then this describes an LTI system.
Now that we know that Equation (3.38) describes an LTI system, it is natural to ask
what its impulse response is. This can be done explicitly by solving the difference (or
differential) equation. However, in this class, we will see a number of much more powerful
and efficient tools to solve this problem.

Appendix 3.A Exchanging Integration and Summation Order

For multiple stacked sums (or integrals), it is often convenient to arrange the summation
(or integration) order as one wishes. For example, in Subsection 3.3.4, we observed that
“in many cases,” convolution is associative, and we stated a sufficient condition.1 In this
appendix, we give a simple explicit counterexample, that is, a case where one cannot swap
the summation order. Consider the following three signals, defined for −∞ < n < ∞:

 1, n=0
d[n] = −1, n=1 (3.39)
0, otherwise.


e[n] = 1, ∀n (3.40)

1, n≥0
u[n] = (3.41)
0, otherwise.

Let us compute both

(u ∗ (d ∗ e))[n] and ((u ∗ d) ∗ e)[n], (3.42)

in the specific order as indiciated by the brackets. For the left version, we first calculate

∞
X ∞
X
(d ∗ e)[i] = d[m]e[i − m] = d[m] = 0 (3.43)
m=−∞ m=−∞

1
A more general sufficient condition to guarantee that the integration/summation order can be swapped is
the so-called Fubini theorem, which can be found in any textbook on Analysis, see for example [3, Theorems
8.6 and 8.8], but note that no deeper understanding of these theorems is required for our class.
24 CHAPTER 3. LINEAR TIME-INVARIANT SYSTEMS

hence (u ∗ (d ∗ e))[n] = 0 for all n. On the other hand, for the version on the right, we first
calcuate
∞
X ∞
X
(u ∗ d)[l] = u[k]d[l − k] = d[l − k] (3.44)
k=−∞ k=0
(
1 l=0
= (3.45)
0 otherwise

hence ((u ∗ d) ∗ e)[n] = e[n] = 1 for all n. As a final remark, recall the simple sufficient
condition stated in Subsection 3.3.4: If all involved signals are absolutely summable (or
integrable), then convolution is associative. This condition is clearly not satisfied in the
above example: Neither e[n] nor u[n] are absolutely summable. Hence, the above example
does not contradict our sufficient condition. (However, let us again stress that the condition
of absolute summability is not a necessary condition for swapping summation order.)
Chapter 4

Fourier Methods for Stable LTI

Systems

In the previous chapter, we have seen that an output of an LTI system for a given input
signal could be computed via convolution between the input signal and the impulse response
of the LTI system. This characterization of LTI systems is based on representing signals
as linear combinations of shifted impulses. In this and the following chapters, we explore
an alternative representation for signals and LTI systems as linear combinations of complex
exponentials. As we will see, this new representation will gives us new and powerful tools
for the analysis of LTI systems.
The next task is to discuss how we can represent signals as linear combinations of
complex exponentials. These techniques are due to Joseph Fourier (1768-1830), and in his
honor, are called Fourier transform and Fourier series.
A more subtle question concerns the existence of such representations. Although many
signals of interest can be expressed as linear combinations of complex exponentials, there
are exceptions and technicalities. Many of these were resolved by other mathematicians
later on, including Cauchy and Dirichlet. You see a more detailed account of these issues
in your Analysis III-IV classes.
A good answer for the purpose of engineering is that Fourier representations are well
defined for

• Periodic signals

R∞
• Absolutely integrable signals, i.e., signals for which −∞ |x(t)|dt is finite. This class
of functions is often referred to as L1 .

R∞ 2 dt
• Finite energy signals, i.e., signals for which −∞ |x(t)| is finite. This class of func-
tions is often referred to as L2 .

25
26 CHAPTER 4. FOURIER METHODS FOR STABLE LTI SYSTEMS

4.1 Continuous-time Signals

4.1.1 The Fourier Transform
For a signal x(t), we define the Fourier transform as
Z ∞
X(ω) = x(t)e−jωt dt, (4.1)
−∞

if this integral converges. In your Analysis III-IV courses, you will study this question
in more detail. A sufficient
R ∞ condition for convergence is that the signal x(t) is absolutely
integrable, i.e., we have −∞ |x(t)|dt < ∞. The inverse Fourier transform is defined as
Z ∞
1
x(t) = X(ω)ejωt dω, (4.2)
2π −∞

again if this integral converges.

X(ω) is called the spectrum of the signal x(t). Intuitively, the (complex) number X(ω)
determines “how much” of every frequency ω is present in the signal x(t). When we talk
about the Fourier transform, X(ω), of a signal, we say that we consider the signal in the
frequency-domain rather than x(t) which is the representation of the signal in the time-
domain.
A key question is the following one : If we start with a signal x(t), apply the Fourier
transform to obtain X(ω), and then apply the inverse Fourier transform, do we get back
the same signal x(t)? (In other words, is the Fourier transform invertible?) For the purpose
of our class, the answer is yes — we will always use the Fourier transform in this sense.
However, strictly speaking, this invertibility only applies to well-behaved signals and has
a good number of footnotes and restrictions, some of which you will encounter in your
Analysis III-IV courses. We include a few brief comments in Section 4.3.

4.1.2 Integrating a Complex Exponential

For the remainder of this class, one simple integral (that you know very well) will play a
key role. To be on the safe side, we include it here. Namely, for any complex number c, we
have
Z ∞
1
e−ct dt = ,
0 c

if Re{c} > 0. Otherwise, the integral does not converge. Note that by a change of integration
variable τ = −t, this also implies
Z 0 Z 0 Z ∞
1
ecτ dτ = e−ct (−dt) = e−ct dt = ,
−∞ ∞ 0 c

again if Re{c} > 0. (Again, otherwise, the integral does not converge.)
4.1. CONTINUOUS-TIME SIGNALS 27

It is assumed that you are fully familiar with the case where c is a real number. To
derive the case when c is complex-valued, you may proceed as follows: Write c = cR + jcI
and observe
Z ∞ Z ∞ Z ∞
−ct −cR t
e dt = e cos(cI t)dt − j e−cR t sin(cI t)dt
0 0 0
cR cI
= −j 2 ,
c2R + c2I cR + c2I

where the second expression involves only standard real-valued integrals that can be looked
up in your favorite integral table. Combining,
cR cI c∗ 1
−j 2 = = .
c2R + c2I cR + c2I |c|2 c

4.1.3 Properties of the Fourier Transform

Our keen interest will now be devoted to exploring the various properties of the Fourier
transform. Tables with many of these properties are given at the end of the chapter in
Appendices 4.A and 4.B.
At this point, we consider only one very special function, often referred to as the “box
function:”

1, |t| ≤ 1,
b(t) = (4.3)
0, otherwise.

Its Fourier transform is found by direct evaluation:

Z 1
1 jω 1 −jω 2 sin(ω)
B(ω) = e−jωt dt = e − e = , (4.4)
−1 jω jω ω
1
where for the last step, we have used the identity sin(x) = 2j (ejx − e−jx ).
Since this last function appears frequently, engineers have introduced a special symbol:
sin(πx)
sinc(x) = , (4.5)
πx
called the “sinc-function” (pronounced essentially like “sink-function”). Using this, we can
express Equation (4.4) equivalently as

B(ω) = 2 sinc(ω/π). (4.6)

4.1.4 LTI Systems and the Fourier Transform

The most fundamental insight of our class is that for any LTI system, we can express the
input-output relationship as
Z ∞
y(t) = x(τ )h(t − τ )dτ, (4.7)
−∞
28 CHAPTER 4. FOURIER METHODS FOR STABLE LTI SYSTEMS

which was Equation (3.16). But then,

Z ∞
Y (ω) = y(t)e−jωt dt
−∞
Z ∞ Z ∞ 
= x(τ )h(t − τ )dτ e−jωt dt. (4.8)
−∞ −∞

In this chapter, we assume that the system is stable, which is the same as assuming that
h(t) is absolutely integrable. Moreover, we assume that the Fourier transform of x(t) exists.
Then, we may swap integration order to obtain
Z ∞ Z ∞ 
−jωt
Y (ω) = x(τ ) h(t − τ )e dt dτ. (4.9)
−∞ −∞

Let us define
Z ∞
H(ω) = h(t)e−jωt dt (4.10)
−∞

to be the Fourier transform of the impulse response h(t). We call H(ω) the frequency
response of the system and will look at it in more detail in the next chapter. For now, the
inner integral in (4.9) can be solved for example using the “shift in time” property, leading
to e−jωτ H(ω). Thus, we are left with
Z ∞
Y (ω) = H(ω) x(τ )e−jωτ dτ
−∞
= H(ω)X(ω). (4.11)

That is, if we are analyzing LTI systems, then it is an excellent idea to consider signals in the
frequency-domain rather than the time-domain: In the time domain, the system operation
is given by the convolution, which, as we saw, is a little tedious to evaluate in many cases.
By contrast, in the frequency domain, the system operation is simply characterized by
(point-wise) multiplication, a much simpler operation!
This insight is also often referred to as the convolution property of the Fourier transform,
stating that

(h ∗ x)(t) ◦−• H(ω)X(ω), (4.12)

where we use the shorthand symbol ◦−• to mean that on the open circle side is the time-
domain representation and on the closed bullet side is the corresponding frequency domain
representation.

4.1.5 LTI Systems and the Fourier Transform: Differential Equations

In many cases, continuous-time LTI systems are conveniently described by differential equa-
tions. For those, one could first solve the differential equation assuming that the input is
4.1. CONTINUOUS-TIME SIGNALS 29

the delta function, thus finding the impulse response h(t). However, there is a much eas-
ier approach in this case. To see this, we first observe that if a certain signal x(t) has
d
Fourier transform X(ω), then the signal dt x(t) must have Fourier transform jωX(ω). In
our notation,

d
x(t) ◦−• jωX(ω). (4.13)
dt
But then, we observe that if two signals x(t) and y(t) satisfy

d d2 1d
y(t) + 3 y(t) + 2 2 y(t) = x(t) + x(t), (4.14)
dt dt 2 dt
then their Fourier transforms X(ω) and Y (ω) must satisfy

1
Y (ω) + 3jωY (ω) + 2(jω)2 Y (ω) = X(ω) + jωX(ω). (4.15)
2
First, we trivially rearrange this equation to
 
2

1
Y (ω) 1 + 3jω + 2(jω) = X(ω) 1 + jω . (4.16)
2

To see why this insight is highly valuable, we have to recall from Equation (4.11) that
Y (ω) = H(ω)X(ω), which implies that H(ω) = Y (ω)/X(ω). That is, we find

1 + 12 jω
H(ω) = . (4.17)
1 + 3jω + 2(jω)2

In other words, we found the frequency response of the system in a couple of very easy
steps, without solving any differential equations! (If needed, we can now take the inverse
Fourier transform of H(ω) to recover the impulse response h(t).)
We should also note that this approach does not just work for the example differential
equation above; it should be clear that it works for any constant-coefficient linear differential
equation whatsoever.

4.1.6 Delta Functions and the Fourier Transform

Fourier methods are useful even in cases where the Fourier transform (4.1) does not converge.
For example, let us pretend that a certain signal z(t) has Fourier transform given by

Z(ω) = 2πδ(ω − ω0 ), (4.18)

for a certain (fixed) value of ω0 . To find the signal z(t), we have to apply the inverse Fourier
transform, Formula (4.2),
Z ∞
1
z(t) = 2πδ(ω − ω0 )ejωt dω = ejω0 t , (4.19)
2π −∞
30 CHAPTER 4. FOURIER METHODS FOR STABLE LTI SYSTEMS

where for the last step, we have used the fundamental property of the Dirac-delta function
in Equation (3.11) and the fact that ejωt is a continuous function of ω. In particular, we also
point out that for the special choice ω0 = 0, we find that the corresponding time domain
signal is z(t) = 1 for all t, −∞ < t < ∞.
The key observation is that while the Fourier transform Z(ω) may be a tricky object
to deal with, the corresponding time-domain signal is a perfectly well-behaved function,
making the overall argument very useful in practice.
By analogy, we can also consider the time-domain signal y(t) = δ(t − t0 ), which is a
Dirac-delta at time t0 . Its Fourier transform is found to be
Z ∞
Y (ω) = δ(t − t0 )e−jωt dt = e−jωt0 . (4.20)
−∞

As we will see, this “generalized” Fourier transform turns out to be very useful in
practice. Nevertheless, we should end this discussion on a cautionary note: These arguments
must be used with some care. To be precise, correctness must be argued on a case-by-case
basis.

4.2 Discrete-time Signals

4.2.1 The Discrete-time Fourier Transform
For a discrete-time signal x[n], we define the discrete-time Fourier transform as
∞
X
X(ω) = x[n]e−jωn , (4.21)
n=−∞

if this sum converges.

The inverse discrete-time Fourier transform is defined as
Z π
1
x[n] = X(ω)ejωn dω. (4.22)
2π −π
Again, there is a somewhat more subtle question concerning the invertibility of the discrete-
time Fourier transform, but for the purpose of our class, and for the (well-behaved) signals
that are of interest to us, invertibility will always hold, making the discrete-time Fourier
transform an important tool.

4.2.2 Properties of the Discrete-time Fourier Transform

A simple (but important) fact is that the discrete-time Fourier transform is 2π-periodic.
That is, for every integer k, we have
∞
X ∞
X ∞
X
X(ω + k2π) = x[n]e−j(ω+k2π)n = x[n]e−jωn e|−j2πkn
{z } = x[n]e−jωn = X(ω).
n=−∞ n=−∞ =1 n=−∞

Tables with many of these properties are given at the end of the chapter in Appendices 4.C
and 4.D.
4.2. DISCRETE-TIME SIGNALS 31

4.2.3 LTI Systems and the Discrete-time Fourier Transform

The most fundamental insight of our class is that for any discrete-time LTI system, we can
express the input-output relationship as
∞
X
y[n] = x[k]h[n − k], (4.23)
k=−∞

which was Equation (3.10). But then,

∞
X
Y (ω) = y[n]e−jωn
n=−∞
∞ ∞
!
X X
= x[k]h[n − k] e−jωn . (4.24)
n=−∞ k=−∞

In this chapter, we assume that the system is stable, hence, that h[n] is absolutely summable.
Moreover, we assume that the discrete-time Fourier transform of x[n] exists. Then, we may
swap summation order to obtain
∞ ∞
!
X X
−jωn
Y (ω) = x[k] h[n − k]e . (4.25)
k=−∞ n=−∞

The inner sum can be solved for example using the “shift in time” property, leading to
e−jωk H(ω) (where again, the frequency response H(ω) is the discrete-time Fourier transform
of the impulse response). Thus, we are left with
∞
X
Y (ω) = H(ω) x[k]e−jωk
k=−∞
= H(ω)X(ω). (4.26)

Again, we see that if we switch from time-domain to frequency domain, the convolution
operation becomes a simple point-wise multiplication. Exactly like in the continuous-time
case, this insight is also often referred to as the convolution property of the discrete-time
Fourier transform, stating that

(h ∗ x)[n] ◦−• H(ω)X(ω). (4.27)

4.2.4 LTI Systems and the Discrete-time Fourier Transform: Difference Equa-
tions
In many cases, discrete-LTI systems are conveniently described by difference equations.
For those, one could first solve the difference equation assuming that the input is the
Kronecker-delta function, thus finding the impulse response h[n]. However, there is a much
easier approach in this case. To see this, we first observe that if a certain signal x[n] has
32 CHAPTER 4. FOURIER METHODS FOR STABLE LTI SYSTEMS

discrete-time Fourier transform X(ω), then the signal x[n − 1] must have Fourier transform
e−jω X(ω). In our notation,
x[n − 1] ◦−• e−jω X(ω). (4.28)
But then, we observe that if two signals x[n] and y[n] satisfy
1 2 3
y[n] − y[n − 1] = x[n] + x[n − 1] − x[n − 2], (4.29)
2 3 5
then their discrete-time Fourier transforms X(ω) and Y (ω) must satisfy
1 2 3
Y (ω) − e−jω Y (ω) = X(ω) + e−jω X(ω) − e−j2ω X(ω). (4.30)
2 3 5
First, we trivially rearrange this equation to
   
1 −jω 2 −jω 3 −j2ω
Y (ω) 1 − e = X(ω) 1 + e − e . (4.31)
2 3 5
To see why this insight is highly valuable, we have to recall from Equation (4.26) that
Y (ω) = H(ω)X(ω), which implies that H(ω) = Y (ω)/X(ω). That is, we find
1 + 32 e−jω − 35 e−j2ω
H(ω) = . (4.32)
1 − 12 e−jω
In other words, we found the frequency response of the system in a couple of very easy
steps, without solving any difference equations! (If needed, we can now take the inverse
discrete-time Fourier transform of H(ω) to recover the impulse response h[n].)
We should also note that this approach does not just work for the example difference
equation above; it should be clear that it works for any constant-coefficient linear difference
equation whatsoever.

4.2.5 Delta Functions and the Discrete-time Fourier Transform

The discrete-time Fourier transform is useful even in cases where the expression in For-
mula (4.21) does not converge. For example, consider a signal z[n] whose discrete-time
Fourier transform Z(ω) has a Dirac-delta of amplitude 2π at frequency ω0 , where |ω0 | < π.
At this point, it is important to recall that Z(ω) must be 2π-periodic. In other words, if
there is a Dirac-delta at frequency ω0 , then there must be Dirac-deltas at all frequencies
ω0 + 2πℓ, for all integers ℓ. One can express such a “train” of Dirac-deltas as
∞
X
Z(ω) = 2πδ(ω − ω0 − 2πℓ). (4.33)
ℓ=−∞

We can now use the inverse discrete-time Fourier transform (Formula (4.22)) to determine
the corresponding time-domain signal z[n] as
∞
Z π !
1 X
z[n] = 2πδ(ω − ω0 − 2πℓ) ejωn dω. (4.34)
2π −π
ℓ=−∞
4.3. CONVERGENCE ISSUES 33

In this doubly infinite sum, only a single Dirac-delta function falls into the interval from
−π to π (namely, the one where ℓ = 0). But since the integral over ω only concerns this
interval, we find
Z π
z[n] = δ(ω − ω0 )ejωn dω = ejω0 n (4.35)
−π

where for the last step, we have used the fundamental property of the Dirac-delta function
in Equation (3.11) and the fact that ejωn is a continuous function of ω. An interesting
special case is the choice ω0 = 0. Now, the Dirac-delta in Z(ω) is located at frequency zero
(plus multiples of 2π), and the corresponding time-domain signal is the all-ones sequence
z[n] = 1.

4.3 Convergence Issues

As you see in more detail in your Analysis III-IV classes, a fundamental question is the
convergence of the Fourier integrals. One way in which we can ask this question is: If we
start with a signal x(t) and calculate the Fourier transform
Z ∞
X(ω) = x(t)e−jωt dt, (4.36)
−∞

and then invert this to obtain

Z ∞
1
x̃(t) = X(ω)ejωt dω, (4.37)
2π −∞

do we really have x(t) = x̃(t)? So far, we have just pretended that this is the case. But
there are some subtleties.
A first case where one can make
R ∞ general statements concerns all those signals x(t) that
are absolutely integrable (i.e., −∞ |x(t)|dt < ∞). For those signals, we have x(t) = x̃(t)
for all values of t except for those t for which x(t) is not continuous. At these points of
discontinuity, the interesting and important Gibbs phenomenon occurs. (If you are curious,
enter this search term into your internet search engine of preference and you will find a
plethora of documents and animations.)
R ∞ A second case concerns the finite-energy signals, which are those signals x(t) for which
2 dt < ∞. For these signals, a successful perspective comes in the shape of a
−∞ |x(t)|
mathematical concept called Hilbert space.
34 CHAPTER 4. FOURIER METHODS FOR STABLE LTI SYSTEMS

Appendix 4.A Continuous-Time Fourier Transform : Properties

Z ∞ Z ∞
1
x(t) = X(ω)e jωt
dω ←→ X(ω) = x(t)e−jωt dt
2π −∞ −∞

Property Signal Fourier transform

Linearity αx(t) + βy(t) αX(ω) + βY (ω)

Shift in time x(t − t0 ) e−jt0 ω X(ω)

Shift in frequency ejω0 t x(t) X(ω − ω0 )

1 ω 
Scaling in time and frequency x(at) X
|a| a
Time reversal x(−t) X(−ω)

dn
Differentiation in time x(t) (jω)n X(ω)
dtn
dn
Differentiation in frequency (−jt)n x(t) X(ω)
dω n
Z t
1
Integration in time x(τ )dτ X(ω), assuming X(0) = 0.
−∞ jω

Convolution in time (x ∗ y)(t) X(ω)Y (ω)

1
Convolution in frequency x(t)y(t) (X ∗ Y )(ω)
2π

Conjugate x∗ (t) X ∗ (−ω)

Conjugate, time-reversed x∗ (−t) X ∗ (ω)

Conjugate symmetry x(t) real-valued X(ω) = X ∗ (−ω)

which implies |X(ω)| = |X(−ω)|
x(t) real and even X(ω) real and even
i.e., x(t) = x(−t) i.e., X(ω) = X(−ω)

Z +∞ Z +∞
1
Parseval’s Equality |x(t)|2 dt = |X(ω)|2 dω
−∞ 2π −∞
4.B. CONTINUOUS-TIME FOURIER TRANSFORM : PAIRS 35

Appendix 4.B Continuous-Time Fourier Transform : Pairs

Signal Fourier transform

Dirac delta function x(t) = δ(t) X(ω) = 1

x(t) = δ(t − t0 ) X(ω) = e−jt0 ω

+∞ +∞  
X 2π X 2πk
Dirac comb x(t) = δ(t − nT ) X(ω) = δ ω−
n=−∞
T T
k=−∞

Constant function x(t) = 1 X(ω) = 2πδ(ω)

Harmonics x(t) = ejω0 t X(ω) = 2πδ(ω − ω0 )
x(t) = cos(ω0 t) X(ω) = π (δ(ω − ω0 ) + δ(ω + ω0 ))
π
x(t) = sin(ω0 t) X(ω) = (δ(ω − ω0 ) − δ(ω + ω0 ))
j


1, t ≥ 0 1
Step function u(t) = U (ω) = + πδ(ω)
0, t < 0. jω

1
One-sided exponential x(t) = e−at u(t) X(ω) =
a + jω
with Re(a) > 0
tn−1 −at 1
for integers n ≥ 2 x(t) = e u(t) X(ω) =
(n − 1)! (a + jω)n

2a
Two-sided exponential x(t) = e−a|t| with Re(a) > 0 X(ω) =
a2 + ω2

r ( q
2π
ω0 |ω| ≤ 21 ω0
ω 
Sinc function x(t) = sinc
0
t X(ω) = ω0 ,
2π 2π 0, otherwise.
where sinc(x) = sinπxπx

(
√1 , |t| ≤ 12 t0 , √
 
t0
t0
Box function b(t) = B(ω) = t0 sinc ω
0, otherwise. 2π
36 CHAPTER 4. FOURIER METHODS FOR STABLE LTI SYSTEMS

Appendix 4.C Discrete-Time Fourier Transform : Properties

Z +∞
1 X
x[n] = X(ω)ejωn dω ←→ X(ω) = x[n]e−jωn
2π 2π n=−∞

Property Signal Fourier transform

Linearity αx[n] + βy[n] αX(ω) + βY (ω)

Shift in time x[n − n0 ] e−jωn0 X(ω)
Shift in frequency ejω0 n x[n] X(ω − ω0 )

Time Reversal x[−n] X(−ω)

dX(ω)
Differentiation in Frequency nx[n] j
dω

Convolution in time (x ∗ y)[n] X(ω)Y

Z (ω)
1
Circular convolution in fre- x[n]y[n] X(θ)Y (ω − θ))dθ
2π 2π
quency

Conjugate x∗ [n] X ∗ (−ω)

Conjugate, time-reversed x∗ [−n] X ∗ (ω)

Conjugate symmetry x[n] real-valued X(ω) = X ∗ (−ω)

which implies |X(ω)| = |X(−ω)|
x[n] real and even X(ω) real and even
i.e., x[n] = x[−n] i.e., X(ω) = X(−ω)

Parseval’s Relation for Aperiodic Signals

+∞ Z
X 1
|x[n]|2 = |X(ω)|2 dω
n=−∞
2π 2π
4.D. DISCRETE-TIME FOURIER TRANSFORM : PAIRS 37

Appendix 4.D Discrete-Time Fourier Transform : Pairs

Signal Fourier transform

Kronecker delta δ[n] 1

δ[n − n0 ] e−jn0 ω

+∞
X
Constant x[n] = 1 2π δ(ω − 2πl)
l=−∞
+∞
X
Harmonics x[n] = ejω0 n 2π δ(ω − ω0 − 2πl)
l=−∞

 +∞
1, n ≥ 0 1 X
Step function u[n] = + πδ(ω − 2πk)
0, otherwise. 1 − e−jω
k=−∞

αn , n ≥ 0

1
One-sided x[n] =
0, otherwise. 1 − αe−jω
exponential with |α| < 1

nαn , n ≥ 0 αe−jω

“Arithmetic- x[n] =
0, otherwise. (1 − αe−jω )2
-geometric” with |α| < 1

r ( q
2π
ω0 |ω| ≤ 12 ω0
ω 
Sinc sequence sinc
0
n X(ω) = ω0 ,
2π 2π 0, otherwise.
where sinc(x) = sinπxπx

(
√1 , n0
|n| ≤ 21 (n0 − 1),


n0 √ sinc 2π ω

Box sequence x[n] = n0 1
0, otherwise. sinc 2π ω
where n0 is odd
38 CHAPTER 4. FOURIER METHODS FOR STABLE LTI SYSTEMS

Appendix 4.E Continuous-time Periodic Signals

Although we will not have time to cover the material of this Appendix in class, we include
it in the lecture notes since you have seen Fourier series in your Analysis classes.

4.E.1 The Fourier Series

Consider a signal x(t) with fundamental period T that can be expressed as
∞
X 2π
x(t) = Xk ej T kt
, (4.38)
k=−∞

for arbitrary complex coefficients Xk . The so-called fundamental frequency of this signal is
ω0 = 2πT . As you learn in your Analysis III-IV courses, there is actually a small miracle
happening here: All well-behaved T -periodic functions can be expressed in this fashion! (In
your Analysis classes, you learn more about which signals qualify as “well-behaved.”)
The coefficients Xk for −∞ < k < ∞ are referred to as the Fourier series of the periodic
signal x(t). For any T -periodic signal x(t), these coefficients are simply given by
Z T
1 2π
Xk = x(t)e−j T kt
dt. (4.39)
T 0

You see a proof of this in your Analysis classes.

4.E.2 Properties of the Fourier Series

The Fourier series coefficients have many interesting properties, some almost trivial, others
rather intricate. A table of these properties is given in Appendices 4.F and 4.G.
In these notes, let us only pull out a single property, usually referred to as the Parseval
identity. As we saw, for a periodic signal, the power can be calculated easily as
Z T
1
P = |x(t)|2 dt. (4.40)
T 0

A question of obvious engineering interest is whether we can also calculate the power directly
from the Fourier series Xk (i.e., without first retrieving the signal x(t) via Equation (4.38)
and then integrating the square). Indeed, the Parseval identity asserts that
Z T ∞
1 X
|x(t)|2 dt = |Xk |2 . (4.41)
T 0 k=−∞

This has an intuitively pleasing intuition: |Xk |2 characterizes the power of the signal x(t) at
frequency k · 2π
T . Parseval’s relation then says that the overall power of the signal is simply
the sum of the powers at the individual frequencies.
4.E. CONTINUOUS-TIME PERIODIC SIGNALS 39

To prove this is surprisingly simple: We write out

∞ 2
1 T 1 T X
Z Z
2 j 2π kt
|x(t)| dt = Xk e T dt
T 0 T 0
k=−∞
Z T ∞
! ∞
!∗
1 X 2π X 2π
= Xk ej T kt Xm ej T mt dt
T 0 m=−∞
k=−∞
∞ ∞ Z T
∗ 1 2π 2π
X X
= Xk Xm ej T kt e−j T mt dt. (4.42)
m=−∞
T 0
k=−∞

The key now is to observe that

1 T j 2π (k−m)t
Z 
1, if k = m,
e T dt = (4.43)
T 0 0, otherwise,
which you can verify easily by recalling that ejx = cos(x) + j sin(x). In words, this says that
2π
the functions √1T ej T kt are orthonormal. Using this in Equation (4.42) completes the proof.

4.E.3 LTI Systems and Fourier Series

It is immediately clear that if we pass a signal of the form of Equation (4.38) through an
LTI system with frequency response H(ω), the output is
( ∞ )
j 2π
X
kt
y(t) = H Xk e T
k=−∞
∞
X n 2π o
= Xk H ej T kt
k=−∞
∞  
X 2π 2π
= Xk H j k ej T kt , (4.44)
T
k=−∞

where the second equality follows by linearity, and the last equality is Equation (5.2). By
analogy to Equation (4.38), the Fourier series representation of the signal y(t) is given by
∞
X 2π
y(t) = Yk ej T kt
. (4.45)
k=−∞

Comparing this with Equation (4.44), we see that

 
2π
Yk = H k Xk . (4.46)
T
That is, when we are studying periodic input signals and LTI systems, it is advantageous to
represent the input signal by its Fourier series and the LTI system by its frequency response.
In this way, rather than having to solve a convolution integral, we can perform the system
operation in the frequency domain according to Equation (4.46) by a simple point-wise
multiplication, separately for each frequency.
40 CHAPTER 4. FOURIER METHODS FOR STABLE LTI SYSTEMS

Appendix 4.F Continuous-Time Fourier Series : Properties

+∞
X +∞
X
x(t) = Xk ejkω0 t = Xk ejk(2π/T )t
k=−∞ k=−∞
Z Z
1 −jkω0 t 1
Xk = x(t)e dt = x(t)e−jk(2π/T )t dt
T T T T

Let x(t) and y(t) both be periodic signals of fundamental period T.

Property Periodic Signal Fourier Series Coefficients

Linearity Ax(t) + By(t) AXk + BYk

Time-Shifting x(t − t0 ) Xk e−jkω0 t0 = Xk e−jk(2π/T )t0
Frequency-Shifting ejM (2π/T )t x(t) Xk−M
Conjugation x∗ (t) ∗
X−k
Time Reversal x(−t) X−k
Time Scaling x(αt),
Z α>0 Xk
Periodic Convolution x(τ )y(t − τ )dτ T Xk Yk
T
+∞
X
Multiplication x(t)y(t) Xl Yk−l
l=−∞
dx(t) 2π
Differentiation jkω0 Xk = jk Xk
Z dt
t   T 
1 1
Integration x(τ )dτ Xk = Xk
−∞  jkω0 ∗ jk(2π/T )

Xk = X−k
Re(Xk ) = Re(X−k )


Conjugate Symmetry
x(t) real Im(Xk ) = −Im(X−k )
for Real Signals
|X | = |X−k |

 k



arg Xk = − arg X−k
Real and Even Signals x(t) real and even Xk real and even

Real and Odd Signals x(t) real and odd Xk purely imaginary and odd

Parseval’s Relation for Periodic Signals

Z +∞
1 X
|x(t)|2 dt = |Xk |2
T T
k=−∞
4.G. CONTINUOUS-TIME FOURIER SERIES : PAIRS 41

Appendix 4.G Continuous-Time Fourier Series : Pairs

Signal Fourier series coefficients

+∞
X
Xk ejkω0 t Xk
k=−∞

X1 = 1
ejω0 t
Xk = 0, otherwise

X1 = X−1 = 21
cos ω0 t
Xk = 0, otherwise

1
X1 = −X−1 = 2j
sin ω0 t
Xk = 0, otherwise

 0 = 1, Xk = 0, k ̸= 0
X 
x(t) = 1 this is the Fourier series represen-
tation for any choice of T > 0

Periodic
 square wave
1, |t| < T1  
x(t) = ω0 T1 kω0 T1 sin kω0 T1
0, T1 < |t| ≤ T2 sinc =
π π kπ
and
x(t + T ) = x(t)

+∞
X 1
δ(t − nT ) Xk = for all k
n=−∞
T
42 CHAPTER 4. FOURIER METHODS FOR STABLE LTI SYSTEMS
Chapter 5

The Frequency Response of Stable

LTI Systems

As we saw in previous chapter, the frequency-domain characterization of an LTI system

in terms of its frequency response is an alternative to the time-domain characterization
through convolution and impulse response. It is particularly convenient to analyze LTI
systems in the frequency-domain because differential equations, difference equations, and
convolution operations in the time domain all become algebraic operations in the frequency
domain. In this chapter, we will continue to develop the frequency-domain perspective by
studying frequency-selective filtering and sampling. The culmination of the tools developed
so far will be the sampling theorem, a surprising and elegant result which states that under
certain conditions a continuous-time signal could be represented with discrete-time samples
without any loss of information!

5.1 Continuous-time Systems

5.1.1 The Frequency Response
Let us suppose that the input to our stable LTI system is given by

x(t) = ejω0 t , (5.1)

which is also called a complex exponential of (angular) frequency ω0 . Note that it is a periodic
signal of fundamental period T = ω2π0 . Then, the output is given by
Z ∞ Z ∞
y(t) = ejω0 (t−τ ) h(τ )dτ = ejω0 t e−jω0 τ h(τ )dτ
−∞
| −∞ {z }
H(ω0 )
jω0 t
= H(ω0 )e , (5.2)

as long as the
R ∞integral converges. Now, if we assume that the LTI system is stable, which
means that −∞ |h(t)|dt converges, then it is easy to show that the integral in the definition

43
44 CHAPTER 5. THE FREQUENCY RESPONSE OF STABLE LTI SYSTEMS

of H(ω0 ) converges, too. Recall, we call H(ω0 ) the frequency response of our LTI system at
frequency ω0 .
It is important to pause at this point and fully digest this result: For any LTI system,
if the input is a complex exponential of frequency ω0 , then the output is again a complex
exponential of exactly the same frequency ω0 . The only thing that the LTI system does
to the complex exponential is to change its amplitude (namely, by |H(ω0 )|) and its phase
(namely, by arg(H(ω0 ))).
One powerful interpretation of this is that complex exponentials are eigenfunctions of all
LTI systems. To understand this, recall from your class on linear algebra that the matrix-
vector product Ax is, in general, cumbersome to evaluate — except if x is an eigenvector
of the matrix A. In this case, we have that Ax = λx, where λ is the eigenvalue correspond-
ing to the eigenvector x. To return to LTI systems, the eigenvalue corresponding to the
“eigenvector” ejω0 t is simply the frequency response evaluated at ω0 , namely, H(ω0 ), as a
comparison with Equation (5.2) reveals.

5.1.2 Interpretation of the Frequency Response

The frequency response is a powerful system representation in part because it admits very
natural insights into the behavior of the system.
The first interpretation concerns the magnitude |H(ω)| of the frequency response. A
simple plot of this magnitude against the frequency ω already gives a good idea of what
kind of a system we are dealing with. For example, some systems will have a large value
of |H(ω)| for small frequencies ω, but a much smaller value (perhaps even zero) for large
frequencies. Such a system is referred to as a low-pass filter, because it only passes the
low frequencies, but significantly weakens (or even fully blocks) the high frequencies. In
a similar fashion, there are high-pass filters, band-pass filters (these pass only a certain
intermediate range of frequencies, but block both low and high frequencies), band-stop (or
notch) filters (these block a certain intermediate range of frequencies, while passing both
low and high frequencies), and so on. It should be noted that this filter terminology is not
of the rigorous mathematical kind; rather, it is approximate engineering parlance (but it
turns out to be very useful).
The second interpretation concerns the phase arg H(ω). Again, if the input signal is the
complex exponential x(t) = ejωt , the output is

y(t) = H(ω)ejωt
= |H(ω)|ej arg H(ω) ejωt
= |H(ω)|ejω(t+arg H(ω)/ω) . (5.3)

That is, the phase can be thought of as a delay by − arg H(ω)/ω. It should be noted that
each frequency will be delayed by a different time shift (except in the practically interesting
case of so-called linear phase, which means that arg H(ω) = −t0 ω, where the delay is exactly
t0 irrespective of the frequency).
5.1. CONTINUOUS-TIME SYSTEMS 45

5.1.3 Basic Properties of the Frequency Response

1. We consider both positive and negative values of ω0 since clearly, the main insight we
obtained in Equation (5.2) holds for both cases.
At the same time, it might be a little confusing that we are considering “negative”
frequencies. To clarify things, let us consider the frequency −ω0 (where ω0 itself is a
positive number). That is, we consider the signal

x(t) = e−jω0 t . (5.4)

If the goal is to gain insight and interpretability, it is often a good idea to express
complex numbers in terms of their real and imaginary parts, which gives:

x(t) = cos(−ω0 t) + j sin(−ω0 t). (5.5)

But remembering from calculus that cos(−ω0 t) = cos(ω0 t) and sin(−ω0 t) = − sin(ω0 t),
this is exactly the same as

x(t) = cos(ω0 t) − j sin(ω0 t), (5.6)

and this expression only contains positive frequencies. Thus, negative frequencies
do not need to be interpreted in a physical way, they are only a tool to make the
mathematical expressions more compact.

2. When the impulse response h(t) of the system is real-valued, the frequency response
satisfies

H(ω0 ) = H ∗ (−ω0 ), (5.7)

for every ω0 , where ∗ denotes the complex conjugate. One often says that in this case,
the frequency response is conjugate-symmetric.

5.1.4 Why do we consider complex-valued signals?

Why did we use a complex exponential as the test input? — We do not really have to do
this. Instead, we could have considered the real-valued test input x(t) = cos(ω0 t). Here,
too, something nice happens. But as we will see, formulas get a little more cumbersome,
which is why the complex exponentials in the end are more convenient to deal with.
More precisely, if the system input is x(t) = cos(ω0 t), the output is equally simple
to figure out. To see this, it is enough to recall the Euler formula: x(t) = cos(ω0 t) =
1 jω0 t
2e + 12 e−jω0 t . Since the system is linear, the output is just the sum of the outputs
corresponding to the two summands, hence y(t) = 12 H(ω0 )ejω0 t + 21 H(−ω0 )e−jω0 t .
How to interpret this formula? First of all, since we are considering a real-valued input,
we probably also want to assume that the system impulse response h(t) is real-valued. But
then, the system output y(t) must be real-valued, simply from Formula (3.10). At a first
glance, the expression we found for the output signal y(t) does not appear to be real-valued.
How can we understand that in spite of appearance, it actually is real-valued?
46 CHAPTER 5. THE FREQUENCY RESPONSE OF STABLE LTI SYSTEMS

To see this, we exploit the property that if h(t) is real-valued, we must have that
H(ω0 ) = H ∗ (−ω0 ). Using polar coordinates, we can write H(ω0 ) = |H(ω0 )|ej arg H(ω0 ) and
thus, H ∗ (ω0 ) = |H(ω0 )|e−j arg H(ω0 ) . But then, we observe
1 1
y(t) = H(ω0 )ejω0 t + H(−ω0 )e−jω0 t
2 2
1 1
= H(ω0 )e jω0 t
+ H ∗ (ω0 )e−jω0 t
2 2
e j arg H(ω 0 ) ejω0 t + e−j arg H(ω0 ) e−jω0 t
= |H(ω0 )|
2
= |H(ω0 )| cos(ω0 t + arg H(ω0 )), (5.8)

where, for the last step, we have again used the Euler formula. What this says is that if
the system input is x(t) = cos(ω0 t), then the output is again a cosine of exactly the same
frequency. However, it no longer has amplitude 1, but rather, the amplitude is given by
|H(ω0 )|. Moreover, it is also phase-shifted by arg H(ω0 ) (meaning that there is no longer a
peak at t = 0, but there now is a peak at t = − arg H(ω
ω0
0)
).
An alternative derivation is found by considering cartesian (rather than polar) coordi-
nates, which is left as an exercise for the reader.
In summary, by comparing Equation (5.8) to the much simpler expression in Equa-
tion (5.2), it should be clear that dealing with complex exponentials is much more convenient
than dealing with cosines.

5.1.5 Composition of Systems

As we have seen earlier, there are two natural ways of composing two systems: parallel and
series. In both cases, it is a simple matter to find the frequency response of the composed
system from the frequency responses of the individual systems.
In particular, let us consider two systems with frequency responses H1 (ω) and H2 (ω),
respectively, composed in parallel exactly as in Figure 3.2. Then, the overall system has
frequency response

G(ω) = H1 (ω) + H2 (ω) (5.9)

since integration is linear. If instead these systems are composed in series like in Figure 3.3,
then we can use Equation (3.33) to infer that
Z ∞ Z ∞ 
−jω0 t
G(ω) = e h1 (τ )h2 (t − τ )dτ dt
−∞ −∞
= H1 (ω)H2 (ω), (5.10)

which follows from the usual tricks (inverting integration order and change of variable inside
the integral). Recall that this is also known as the “convolution property” of the Fourier
transform .
5.2. DISCRETE-TIME SYSTEMS 47

5.2 Discrete-time Systems

5.2.1 The Frequency Response
Let us suppose that the input to our stable LTI system is given by

x[n] = ejω0 n , (5.11)

which is called a complex exponential of (angular) frequency ω0 . Then, the output is given
by
∞
X ∞
X
y[n] = jω0 (n−k)
e h[k] = e jω0 n
e−jω0 k h[k]
k=−∞ k=−∞
| {z }
H(ω0 )

= H(ω0 )ejω0 n , (5.12)

as long as the (doubly) infinite

P∞ sum converges. Now, if we assume that the LTI system is
stable, which means that k=−∞ |h[k]| converges, then it is easy to show that the sum in
the definition of H(ω0 ) converges, too. Recall that we call H(ω0 ) the frequency response of
our LTI system at frequency ω0 .
Just like in the continuous-time case, this powerful formula says that if the input to
any LTI system is a complex exponential of frequency ω0 , then so is the output. The only
change introduced by the LTI system concerns the amplitude and the phase of the complex
exponential.
Hence, in the same fashion as in the continuous-time case, one can think of the complex
exponentials ejω0 n as eigenfunctions of all (discrete-time) LTI systems.
We should also add a short note concerning the somewhat unusual notation H(ω0 ).
This notation conveniently stresses that the discrete-time Fourier transform is a 2π-periodic
function. Indeed, you can easily verify in Equation (5.12) that if instead of ω0 , you consider
the frequency ω0 + ℓ2π for any integer ℓ, you obtain exactly the same answer: H(ω0 ) =
H(ω0 + ℓ2π)).

5.2.2 Interpretation of the Frequency Response

The interpretation of the frequency response proceeds along the same lines as in the
continuous-time case, but keeping in mind that only the frequency interval (−π, π] needs
to be considered (since the frequency response is 2π-periodic).

5.2.3 Basic Properties of the Frequency Response

1. The frequency response is 2π-periodic.
This fact is a simple consequence of the frequency response being the discrete-time
Fourier transform of the impulse response.
48 CHAPTER 5. THE FREQUENCY RESPONSE OF STABLE LTI SYSTEMS

2. When the impulse response h[n] of the system is real-valued, the frequency response
satisfies

H(ω0 ) = H ∗ (−ω0 ), (5.13)

for every ω0 , where ∗ denotes the complex conjugate. One often says that in this case,
the frequency response is conjugate-symmetric.

3. Following up on this insight, we can again show that if the system’s impulse response
h[n] is real-valued and if the input to the system is x[n] = cos(ω0 n), then the corre-
sponding output is given by

y[n] = |H(ω0 )| cos(ω0 n + arg(H(ω0 ))), (5.14)

using the same exact steps as in Subsection 5.1.4.

5.2.4 Composition of Systems

As we have seen earlier, there are two natural ways of composing two systems: parallel and
series. In both cases, it is a simple matter to find the frequency response of the composed
system from the frequency responses of the individual systems.
In particular, let us consider two systems with frequency responses H1 (ω) and H2 (ω),
respectively, composed in parallel exactly as in Figure 3.2. The overall frequency response
is given by

G(ω) = H1 (ω) + H2 (ω) (5.15)

since the overall sum can be broken into two separate sums. If instead these systems are
composed in series like in Figure 3.3, then we can use Equation (3.33) to infer that
∞
X ∞
X
G(ω) = e−jω0 n h1 [k]h2 [n − k] (5.16)
n=−∞ k=−∞
= H1 (ω)H2 (ω), (5.17)

which follows from the usual tricks (inverting summation order and change of variable
inside the integral). Again, this is the “convolution property” of the discrete-time Fourier
transform.

5.3 Sampling
We will use the tools developed so far to analyze how to represent a continuous-time sig-
nal with equally spaced discrete-time samples. A somewhat surprising result known as
the sampling theorem states that under certain conditions a continuous-time signal can be
completely recovered from a sequence of its samples. The sampling theorem is extremely
important and useful since processing discrete-time signals is more flexible and is often
preferable to processing continuous-time signals.
5.3. SAMPLING 49

5.3.1 Impulse-Train Sampling

One way in which the sampling operation may be represented mathematically is with
impulse-train sampling. This is done by multiplying the continuous-time signal x(t) that
we wish to sample by a periodic impulse train. In the time domain,

xp (t) = x(t)p(t) (5.18)

where
∞
X
p(t) = δ(t − nT ). (5.19)
n=−∞

The periodic impulse train p(t) is referred to as the sampling function, the period T is the
sampling period, and fundamental frequency of p(t), ωs = 2π T , as the sampling frequency.
Combining (5.18) with (5.19) we see that xp (t) is an impulse train with the amplitudes
of the impulses equal to the samples of x(t) at intervals spaced by T . That is,
∞
X
xp (t) = x(nT )δ(t − nT ). (5.20)
n=−∞

We emphasize that (5.18) represents a continuous-time system with an input x(t) and output
xp (t). A discrete-time signal of samples could also be constructed by letting x[n] = x(nT ).
The values of the discrete-time signal x[n] are the same as the amplitudes of the impulses of
xp (t) at nT . We can think of xp (t) as an equivalent representations of a sampled signal x[n];
in other words, xp (t) and x[n] contain the same amount of information about the original
signal x(t).
To understand the impact of sampling on the continuous-time signal x(t) we will keep
analyzing the continuous-time system (5.18) in the frequency domain. From the convolution
in frequency property, we know that
1
Xp (ω) = (X ∗ P )(ω). (5.21)
2π
From Appendix 4.B using the Dirac comb Fourier pair,
∞
2π X
P (ω) = δ(ω − kωs ). (5.22)
T
k=−∞

Combining (5.21) with (5.22) and using the distributive property of convolution we compute,
∞
1 X
Xp (ω) = X(ω − kωs ). (5.23)
T
k=−∞

That is, Xp (ω) is a periodic function of ω consisting of a superposition of shifted replicas of

X(ω), scaled by T1 . Suppose that there is a frequency, ωM , such that X(ω) = 0, whenever
50 CHAPTER 5. THE FREQUENCY RESPONSE OF STABLE LTI SYSTEMS

|ω| > ωM . If ωM < (ωs − ωM ), or equivalently, ωs > 2ωM , then there is no overlap between
the shifted replicas of X(ω), whereas with ωs < 2ωM , there is overlap. Consequently, if
ωs > 2ωM , x(t) can be recovered exactly from xp (t) using a lowpass filter with gain T and
a cutoff frequency greater than ωM and less than ωs − ωM .
To fully understand this, it is key to draw a picture — which the reader is encouraged
to do, and which we will do in lecture.

5.3.2 The Sampling Theorem

By modeling the sampling operation with impulse-train sampling we arrived at the sampling
theorem. Succinctly, the sampling theorem can be stated as follows.
Let x(t) be a band-limited signal with X(ω) = 0 for |ω| > ωM . Then x(t) is uniquely
determined by its samples x(nT ), n = 0, ±1, ±2, . . . , if

ωs > 2ωM , (5.24)

where
2π
ωs = . (5.25)
T
Given these samples, we can reconstruct x(t) by generating a periodic impulse train in
which successive impulses have amplitudes that are successive sample values. This impulse
train is then processed through an ideal lowpass filter with gain T and cutoff frequency
greater than ωM and less than ωs − ωM . The resulting output signal will exactly equal x(t).
The frequency 2ωM , which, under the sampling theorem, must be exceeded by the
sampling frequency, is commonly referred to as the Nyquist rate. (The frequency ωM cor-
responding to one-half the Nyquist rate is often referred to as the Nyquist frequency.)

5.3.3 Sampling a Complex Exponential

In previous sections we derived the sampling theorem by representing the sampling operation
as a multiplication by an impulse train. In this section, we consider what happens as
we sample a continuous-time complex exponential. Before we do this, it is important to
recall the following helpful fact about discrete-time complex exponentials: a discrete-time
exponential is 2π-periodic as a function of ω. That is

ejωn = ejωn ej2πn = ej(ω+2π)n (5.26)

holds for every integer n. More generally, this holds for any multiple of 2π. That is,

ejωn = ejωn ej2πln = ej(ω+2πl)n (5.27)

for l = 0, ±1, ±2, . . . .

Suppose now that we sample a continuous-time signal

x(t) = ejω0 t (5.28)

5.3. SAMPLING 51

with sampling frequency ωs . The resulting discrete-time signal is

x[n] = ejω0 nT , (5.29)

where T = 2π
ωs . By defining ω̃ = 2π ωω0s we can rewrite x[n] as
ω0
x[n] = ej2π ωs n = ej ω̃n . (5.30)

In light of our discussion above about periodicity of discrete-time complex exponentials we

can also write

x[n] = ej(ω̃+2πl)n , for l = 0, ±1, ±2, . . . (5.31)

But, the fact that we can rewrite the same signal x[n] in all these different ways has an
important implication to our discussion about sampling. It implies that if we sample any
continuous-time signal of the form

x̃(t) = ej(ω0 +ωs l)t (5.32)

with sampling frequency ωs the resulting discrete-time signal will always be the same x[n].
At this point, we may wonder which continuous-time complex exponential is the correct
reconstruction given samples x[n]? A reasonable answer is to pick the one with the smallest
magnitude frequency. That is, we could find the value of l that minimizes the quantity
|ω0 + ωs l|. For the impulse-train sampling discussed in the previous section, this is exactly
what a low-pass filter reconstruction procedure does.
Going back to (5.28), we see that if |ωs | > 2|ω0 |, then |ω̃| < π and picking the signal
with lowest magnitude frequency would recover x(t). If |ωs | < 2|ω0 |, then |ω̃| > π, and the
reconstruction procedure would produce a different complex exponential.

5.3.4 Signal Reconstruction and Aliasing

As we have seen, for a band-limited signal, if the sampling instants are sufficiently close,
then the signal can be reconstructed exactly by using an ideal low-pass filter. In the time
domain, the impulse response of this lowpass filter is
ωc T sin ωc t
h(t) = (5.33)
πωc t
where T is the gain of the filter and ωc is the cutoff frequency. Letting xr (t) denote the
reconstructed signal we obtain

xr (t) = (xp ∗ h)(t) (5.34)

or
∞
X
xr (t) = x(nT )h(t − nT ). (5.35)
n=−∞
52 CHAPTER 5. THE FREQUENCY RESPONSE OF STABLE LTI SYSTEMS

The reconstruction procedure in the time domain becomes

∞
X ωc T sin ωc (t − nT )
xr (t) = x(nT ) . (5.36)
n=−∞
π ωc (t − nT )

The reconstruction using the impulse response of an ideal lowpass filter is commonly referred
to as band-limited interpolation, since it implements exact reconstruction if x(t) is band
limited and the sampling frequency satisfies the conditions of the sampling theorem.
So far we assumed that the sampling frequency was sufficiently high that the conditions
of the sampling theorem were met. What happens if this is not the case? The spectrum of
the sampled signal consists of scaled replication of the spectrum of x(t), and this forms the
basis of the sampling theorem. When ωs < 2ωM , the spectrum of x(t) overlaps in Xp (ω)
and thus is no longer recoverable by lowpass filtering. This effect, in which the individual
terms overlap, is referred to as aliasing.
Chapter 6

The Transfer Function and The

Laplace Transform

As we saw, every LTI system can be characterized by its impulse response. If the system is
stable, then there is a very insightful characterization in terms of the frequency response.
This admits many intuitive interpretations.
What if the system is not stable? As we will now see, we can still apply a similar trick as
before. In this way, we obtain a “generalized” frequency response, often referred to as the
transfer function or system function. The latter does not quite have the same physically
intuitive significance as the frequency response, but it will turn out to be an extremely
versatile tool to understand general LTI systems.
The tool introduced in this section is the Laplace transform, which bears many similar-
ities to the Fourier transform but uses a general complex number s rather than frequency.

6.1 The Transfer Function

Let us suppose that the input to our LTI system is given by

x(t) = est , (6.1)

for an arbitrary complex-valued constant s. Then, the output is given by

Z ∞ Z ∞
y(t) = es(t−τ )
h(τ )dτ = e st
e−sτ h(τ )dτ
−∞
| −∞ {z }
H(s)
st
= H(s)e , (6.2)

as long as the integral converges.

The function H(s) is called the transfer function (or sometimes the system function) of
the considered LTI system. The fundamental insight in Equation (6.2) is of a very similar
nature to the consideration in the case of the frequency response: The general exponential

53
54CHAPTER 6. THE TRANSFER FUNCTION AND THE LAPLACE TRANSFORM

signal est is essentially left unchanged by the LTI system! The only change occurs in the
magnitude and phase.
A second important observation is that if instead of an arbitrary complex number s, we
consider a purely imaginary number s = jω, then we obtain exactly the frequency response.
In this sense, the frequency response is a special case of the transfer function, obtained as
H(ω) = H(s = jω). For this reason, we use exactly the same symbol H(·) (that is, the
capital letter H) both for the frequency response and for the transfer function.
The convergence discussion relating to Equation (6.2) will be a major part of the next
section. As we recall, in the case of the frequency response, the convergence only depended
on the impulse response h(t). By contrast, here, convergence depends both on the impulse
response h(t) and on the value of s. It is precisely this dependence that makes the Laplace
transform a very useful tool to understand the stability of LTI systems.

6.2 The Laplace Transform

For a time-domain signal x(t), the Laplace transform is defined as
Z ∞
X(s) = x(t)e−st dt, (6.3)
−∞

for those values of s for which this integral converges. That is, the transfer function H(s)
of a system is exactly the Laplace transform of its impulse response h(t).
We observe that by only considering s of the form s = jω, that is, by evaluating the
Laplace transform only along the imaginary axis in the complex s-plane, we obtain exactly
the Fourier transform. In this sense, the Laplace transform is a strict generalization of the
Fourier transform.
The convergence discussion relating to Equation (6.3) will be the subject of the next
few subsections. The main insight is that for many cases of interest, the integral in Equa-
tion (6.3) will converge for certain values of s and will diverge for other values of s. As we
will see, it is very instructive to sketch the region of those values of s for which the integral
converges into the full complex s-plane. This is called the Region of Convergence (ROC).
To invert the Laplace transform, i.e., given X(s) (and its corresponding region of con-
vergence), find back the time-domain signal x(t), we can formally write
Z
x(t) = X(s)est ds, (6.4)

but this involves complex integration (over the variable s). Due to the time constraints
imposed by this class, we will not get to cover this approach. Instead, we will invert the
Laplace transform using the tables provided at the end of this chapter, in Appendices 6.A
and 6.B
6.2. THE LAPLACE TRANSFORM 55

6.2.1 Key Example 1: Single Real-Valued Pole

To begin to understand and appreciate the Laplace transform, it is instructive to juxtapose
the following two examples:
 −at 
e , t ≥ 0, 0 t > 0,
hc (t) = (6.5) ha (t) = −at (6.7)
0, t < 0, −e , t ≤ 0,

where a is a real number. In the engineering where a is a real number. In the engineer-
literature, such a signal is sometimes called ing literature, such a signal is sometimes
right-sided because the graph of the signal called left-sided because the graph of the sig-
lives on the “right side” in the Euclidean nal lives on the “left side” in the Euclidean
plane. It is also sometimes called a causal plane. It is also sometimes called a anti-
signal since if you interpret this signal as the causal signal since if you interpret this signal
impulse response of an LTI system, then this as the impulse response of an LTI system,
system would be causal. then this system would be anti-causal.
Moreover, this signal is well-behaved (more Moreover, this signal is well-behaved (more
precisely, absolutely integrable) if a > 0. In precisely, absolutely integrable) if a < 0. In
the engineering literature, such a signal is the engineering literature, such a signal is
also sometimes referred to as a stable signal also sometimes referred to as a stable signal
since if we interpret the signal as the impulse since if we interpret the signal as the impulse
response of an LTI system, then this system response of an LTI system, then this system
would be stable. would be stable.
By contrast, if a ≤ 0, the signal is not abso- By contrast, if a ≥ 0, the signal is not abso-
lutely integrable, and we could refer to this lutely integrable, and we could refer to this
as an unstable signal. as an unstable signal.
We find We find
Z ∞ Z ∞
Hc (s) = e−sτ h(τ )dτ Ha (s) = e−sτ h(τ )dτ
−∞ −∞
Z ∞ Z 0
= e−(s+a)τ dτ = − e−(s+a)τ dτ
0 −∞
1 1
= , (6.6) = , (6.8)
s+a s+a
as long as Re(s) + a > 0, i.e., as long as as long as Re(s) + a < 0, i.e., as long as
Re(s) > −a. Otherwise, the exponential in- Re(s) < −a. Otherwise, the exponential in-
side the integral leading to Equation (6.2) side the integral leading to Equation (6.2)
diverges, and hence, the integral is not fi- diverges, and hence, the integral is not fi-
nite. nite.
56CHAPTER 6. THE TRANSFER FUNCTION AND THE LAPLACE TRANSFORM

We observe that in both cases, the Laplace transform evaluates to the formula H(s) =
1
s+a . This is often referred to as a pole at −a.
At this point, we should have doubts about the usefulness of the Laplace transform:
1
Apparently, one and the same Laplace transform, in our example H(s) = s+a , can represent
two very different time domain signals! As it stands, the Laplace transform cannot be
inverted.
In order to have a meaningful transform, we have to add more information: We have to
specify the range of values of s for which the formula is valid. This is usually referred to as
the region of convergence (ROC).
1
For the example at hand, that is H(s) = s+a , it turns out that there are only two
interesting choices of the region of covergence:

ROC(H(s)) = {s : Re(s) > −a} or ROC(H(s)) = {s : Re(s) < −a}

⇓ ⇓
e−at ,
 
t ≥ 0, 0, t > 0,
h(t) = h(t) = (6.9)
0, t < 0, −e−at , t ≤ 0.

By far the best way to use the concept of ROC is graphical. For the example at hand, this
is illustrated in Figure 6.1 for a = 2 and in Figure 6.2 for a = −2.
Let us summarize the observations from these example figures concerning the ROC:

• The signal is “causal” (that is, right-sided) if the ROC extends indefinitely to the
right.

• The signal is “anti-causal” (that is, left-sided) if the ROC extends indefinitely to the
left.

• The time-domain signal is “stable” (that is, absolutely integrable) if and only if the
ROC includes the imaginary axis. This is pleasing: As we observed, the Fourier
transform is simply the Laplace transform, evaluated along s = jω, that is, along the
imaginary axis. Hence, whenever the ROC includes the imaginary axis, this implies
that the Fourier transform of the time-domain signal also exists. But as we have seen,
the Fourier transform only exists for integrable time-domain signals.

As we will show, these are fundamental properties of the ROC in full generality, not just
for the examples considered here.
6.2. THE LAPLACE TRANSFORM 57

Im{s} 3 Im{s} 3

2 2

1 1

−3 −2 −1 1 2 3 −3 −2 −1 1 2 3

Re{s} Re{s}

−1 −1

−2 −2

−3 −3

e−2t , t ≥ 0,
 
0, t > 0,
(a) Hence, h(t) = (b) Hence, h(t) =
0, t < 0, −e−2t , t ≤ 0,
a “stable and causal” signal an “unstable and anti-causal” signal
1
Figure 6.1: The two regions of convergence for H(s) = s+2 (that is, a = 2).

Im{s} 3 Im{s} 3

2 2

1 1

−3 −2 −1 1 2 3 −3 −2 −1 1 2 3

Re{s} Re{s}

−1 −1

−2 −2

−3 −3

e2t , t ≥ 0,
 
0, t > 0,
(a) Hence, h(t) = (b) Hence, h(t) =
0, t < 0, −e2t , t ≤ 0,
an “unstable and causal” signal a “stable and anti-causal” signal
1
Figure 6.2: The two regions of convergence for H(s) = s−2 (that is, a = −2).
58CHAPTER 6. THE TRANSFER FUNCTION AND THE LAPLACE TRANSFORM

6.2.2 Key Example 2: Complex-Valued Pole

So far, we have considered a real-valued exponential. Let us now consider a complex-valued
version, which we can write for example as
 −(a+jω )t
e 0 , t ≥ 0,
g(t) = (6.10)
0, t < 0,

where a and ω0 are real numbers. This signal is complex-valued, which is not a problem,
but here we prefer to consider the slightly more general case
 −(a+jω )t
e 0 + e−(a−jω0 )t = 2e−at cos(ω0 t), t ≥ 0,
hc (t) = (6.11)
0, t < 0,

which, as you can easily convince yourself, is again real-valued. Let us again observe that
this is a well-behaved signal if a > 0, but it is not integrable if a ≤ 0 (and, in fact, blows
up exponentially fast if a < 0). For this signal, we find the Laplace transform
1 1 2(s + a)
Hc (s) = + = . (6.12)
s + (a + jω0 ) s + (a − jω0 ) (s + (a + jω0 ))(s + (a − jω0 ))
Hence, this Laplace transform has two (complex-conjugate) poles as illustrated in Figure 6.3
for the example a = 2, ω = 1, and in Figure 6.4 for the example a = −2, ω = 1. Moreover,
this transfer function also has a zero at s = −a, that is, we have H(s = −a) = 0. This is
often illustrated by a circle in the complex s-plane, leading to the so-called pole-zero plot.
But consider now the signal

0, t > 0,
ha (t) = −(a+jω )t −(a−jω )t −at (6.13)
−e 0 −e 0 = −2e cos(ω0 t), t ≤ 0.

By analogy to the real-valued pole example, one can easily calculate the corresponding
Laplace transform as
2(s + a)
Ha (s) = , (6.14)
(s + (a + jω0 ))(s + (a − jω0 ))
exactly the same as before.
So again, this Laplace transform represents two very different signals. In order to decide
which of the two signals is meant, we need additional information in the shape of the ROC.
Interestingly, the ROC only depends on the value of a. It does not hinge on the value of ω0 .
More specifically, there are only two ROCs of interest, namely, ROC(H(s)) = {s : Re(s) >
−a} and ROC(H(s)) = {s : Re(s) < −a}. In the first case, the ROC extends indefinitely
to the right side, and the corresponding time-domain signal turns out to be a causal signal.
In the second case, the ROC extends indefinitely to the left side, and the corresponding
time-domain signal turns out to be an anti-causal signal.
The stability can also directly be read out of the ROC: whenever the ROC includes
the imaginary axis, then the corresponding time-domain signal is stable (that is, absolutely
integrable).
6.2. THE LAPLACE TRANSFORM 59

Im{s} 3 Im{s} 3

2 2

1 1

−3 −2 −1 1 2 3 −3 −2 −1 1 2 3

Re{s} Re{s}

−1 −1

−2 −2

−3 −3

2e−2t cos(t),
 
t ≥ 0, 0, t > 0,
(a) Hence, h(t) = (b) Hence, h(t) =
0, t < 0, −2e−2t cos(t), t ≤ 0,
a “stable and causal” signal an “unstable and anti-causal” signal
2(s+2)
Figure 6.3: The two regions of convergence for H(s) = (s+(2+j))(s+(2−j)) (that is, a = 2 and
ω0 = 1).

Im{s} 3 Im{s} 3

2 2

1 1

−3 −2 −1 1 2 3 −3 −2 −1 1 2 3

Re{s} Re{s}

−1 −1

−2 −2

−3 −3

2e2t cos(t), t ≥ 0,
 
0, t > 0,
(a) Hence, h(t) = (b) Hence, h(t) =
0, t < 0, −2e2t cos(t), t ≤ 0,
an “unstable and causal” signal a “stable and anti-causal” signal
2(s−2)
Figure 6.4: The two regions of convergence for H(s) = (s−(2−j))(s−(2+j)) (that is, a = −2
and ω0 = 1).
60CHAPTER 6. THE TRANSFER FUNCTION AND THE LAPLACE TRANSFORM

6.2.3 Key Example 3: Multiple Poles

More generally, in many interesting cases, the Laplace transform is the ratio of two poly-
nomials:
P (s)
H(s) = , (6.15)
Q(s)

where P (s) and Q(s) are polynomials. Not surprisingly, this transfer function again repre-
sents several different time-domain signals. How many signals? What types (well-behaved
or exploding / right-sided or left-sided)? This is best tackled by looking at the poles in the
complex s-plane, and hence, by directly discussing the possible ROCs.
We will study one particular example to gain insight. In particular, let
3 1 1
H(s) = = − . (6.16)
(s + 2)(s − 1) s−1 s+2

Here, we have two poles, one at s = −2 and one at s = 1. The three possible ROCs are
shown in Figure 6.5. Let us now specifically determine the corresponding impulse responses.
First, for the case illustrated in Figure 6.5a, we find

0 t > 0,
h(t) = t −2t (6.17)
−e + e , t ≤ 0,

which is anti-causal and unstable (due to the term e−2t ). Note again that you can easily
verify this result: Simply consider the integral
Z ∞
H(s) = e−sτ h(τ )dτ, (6.18)
−∞

and plug in the claimed h(t). It is straightforward to observe that in this case, the integral
converges if and only if Re(s) < −2, which is exactly the ROC illustrated in Figure 6.5a.
3
In this regime, the integral indeed evaluates to H(s) = (s+2)(s−1) .
Second, for the case of Figure 6.5b, we find

e − e−2t , t ≥ 0,
 t
h(t) = (6.19)
0, t < 0,

which is causal and unstable (due to the term et ). Again, the result is easily verified by
direct evaluation of the integral.
Finally, the most interesting choice of ROC is shown in Figure 6.5c, which is

−e−2t , t ≥ 0,

h(t) = (6.20)
−et , t < 0,

which is a stable system. But note that it is not causal. Rather, its impulse response h(t)
is non-zero for all times t. By analogy, the result can be verified by direct evaluation of the
integral.
6.2. THE LAPLACE TRANSFORM 61

Im{s} 3 Im{s} 3

2 2

1 1

−3 −2 −1 1 2 3 −3 −2 −1 1 2 3

Re{s} Re{s}

−1 −1

−2 −2

−3 −3

(a) anti-causal and unstable (b) causal and unstable

Im{s} 3

−3 −2 −1 1 2 3

Re{s}

−1

−2

−3

(c) two-sided and stable

3
Figure 6.5: The three possible regions of convergence for two real poles: H(s) = (s+2)(s−1) .
62CHAPTER 6. THE TRANSFER FUNCTION AND THE LAPLACE TRANSFORM

Summary of the properties of the ROC

When the transfer function is the ratio of two polynomials in s, the first step is to draw out
the pole plot (a pole “landscape”) in the complex s-plane much like in all the examples we
have discussed, except that there will generally be many poles in arbitrary locations. The
next step is to identify all possible choices of the ROC. As we have already seen, ROCs
must satisfy certain conditions to be valid. To summarize the main requirements:

1. The ROC consists of strips parallel to the jω-axis in the s-plane.

2. The ROC is bounded by poles or extends to infinity. It cannot contain any poles of
H(s).

3. If the ROC includes the imaginary axis, then the signal is stable. (This also has an
intuitively pleasing additional reason: On the imaginary axis, we have s = jω, that is,
on the imaginary axis, the Laplace transform is exactly equal to the Fourier transform.
But as we have already seen, a signal has a Fourier transform if and only if it is stable
(i.e., absolutely integrable).)

4. For a causal (i.e., right-sided) signal, the ROC must extend unboundedly to the right
(to Re(s) = ∞). By analogy, for an anti-causal (i.e., left-sided) signal, the ROC must
extend unboundedly to the left (to Re(s) = −∞).

6.3 LTI Systems and The Laplace Transform

Just like in the Fourier domain, we again have the convolution property in the Laplace
domain. It should be clear that if we pass a signal of the form of Equation (6.4) through
an LTI system with transfer function H(s), the output is simply
Z 
st
y(t) = H X(s)e ds
Z
X(s)H est ds

=
Z
= X(s)H(s)est ds, (6.21)

where the second equality follows by linearity, and the last equality is Equation (6.2). The
key insight from Equation (6.21) is that the Laplace transform Y (s) of the output signal
y(t) is simply given by

Y (s) = H(s)X(s). (6.22)

Again, this insight is also often referred to as the convolution property of the Laplace trans-
form, stating that

(h ∗ x)(t) ◦−• H(s)X(s). (6.23)

6.4. LTI SYSTEMS AND THE LAPLACE TRANSFORM : COMPOSITION 63

1
H(s) = s+a a>0 a<0

ROC(H(s)) = {s : Re(s) > −a} stable unstable

that is, interpreted

as a causal system impulse
 −atresponse
e , t ≥ 0,
h(t) =
0, t < 0.

ROC(H(s)) = {s : Re(s) < −a} unstable stable

that is, interpreted

as an anti-causal system impulse
 response
0 t > 0,
h(t) = −at
−e , t ≤ 0.

Table 6.1: All possible interpretations for a single real-valued pole, that is, for the transfer
1
function H(s) = s+a , where a is a real number.

Here, we use the shorthand symbol ◦−• to mean that on the open circle side is the time-
domain representation and on the closed bullet side is the corresponding Laplace domain
representation.
Hence, a system can be characterized by its transfer function H(s). But as in our
discussion of the Laplace transform, this is not yet enough information. We also have to
specify the region of convergence belonging to H(s) in order to have a full characterization
1
of the system. For the simple example of a single pole, i.e., the case of H(s) = s+a , this is
fully described in Table 6.1.

6.4 LTI Systems and The Laplace Transform : Composition

As in the case of the frequency response, the Transfer Function has very desirable properties
with respect to certain compositions of LTI systems. The most important are the parallel
and the series composition introduced in Section 3.4.
For the parallel composition of LTI systems with transfer functions H1 (s) and H2 (s),
respectively, we find an overall transfer function of

H(s) = H1 (s) + H2 (s), (6.24)

simply by the linearity of the Laplace transform. To understand the properties of the
64CHAPTER 6. THE TRANSFER FUNCTION AND THE LAPLACE TRANSFORM

composed system (in terms of causality and stability), it suffices to study the poles of the
composed system H(s) (as long as we assume that both H1 (s) and H2 (s) are rational
functions of s, as we do throughout this class).
Similarly, for the series composition of LTI systems with transfer functions H1 (s) and
H2 (s), respectively, we find an overall transfer function of
H(s) = H1 (s)H2 (s), (6.25)
which follows from the convolution property of the Laplace transform. Again, to understand
the properties of the composed system (in terms of causality and stability), it suffices to
study the poles of the composed system H(s).

6.5 LTI Systems and The Laplace Transform : Differential Equa-

tions
In many cases, continuous-time LTI systems are conveniently described by differential equa-
tions. For those, one could first solve the differential equation assuming that the input is
the delta function, thus finding the impulse response h(t). However, there is a much easier
approach in this case. To see this, we first observe that if a certain signal x(t) has Laplace
d
transform X(s), then the signal dt x(t) must have Laplace transform sX(s). In our notation,
d
x(t) ◦−• sX(s). (6.26)
dt
But then, we observe that if two signals x(t) and y(t) satisfy
d d2 1d
y(t) + 3 y(t) + 2 2 y(t) = x(t) + x(t), (6.27)
dt dt 2 dt
then their Laplace transforms X(s) and Y (s) must satisfy
1
Y (s) + 3sY (s) + 2s2 Y (s) = X(s) + sX(s). (6.28)
2
First, we trivially rearrange this equation to
 
2

1
Y (s) 1 + 3s + 2s = X(s) 1 + s . (6.29)
2
To see why this insight is highly valuable, we have to recall from Equation (6.22) that
Y (s) = H(s)X(s), which implies that H(s) = Y (s)/X(s). That is, we find
1 + 12 s
H(s) = . (6.30)
1 + 3s + 2s2
In other words, we found the transfer function of the system in a couple of very easy steps,
without solving any differential equations! (If needed, we can now take the inverse Laplace
transform of H(s) to recover the impulse response h(t).)
We should also note that this approach does not just work for the example differential
equation above; it should be clear that it works for any constant-coefficient linear differential
equation whatsoever.
6.6. CONTROL SYSTEMS : STABILITY AND CAUSALITY 65

6.6 Control Systems : Stability and Causality

The analysis and design of control systems is one of the key applications of the Transfer
Function and the Laplace transform. To fully appreciate this is beyond the scope of this
class, but we will at least try to give a sense for the issues at hand by the aid of a few
examples.
As a first example, let us consider an automobile suspension system. This example is
taken from [1, p.473 ff.]. Here, we assume that the road is essentially flat at a reference
elevation (no mountains). It has minor deviations from this reference elevation due to
defects, potholes, and the like. These deviations are the system input x(t). The suspension
system should be designed so that the passengers do not feel the variation in road surface.
Therefore, the system output y(t) is the deviation of the chassis position from the desired
position above the reference elevation of the road. A classical suspension system has two
components: a spring (French: ressort; German: Feder), with some characteristic spring
constant k and a dashpot (French: amortisseur; German: Stossdämpfer), again with its own
characteristic constant b. A simplified differential equation for the resulting overall system
is

d d2 d
ky(t) + b y(t) + M 2 y(t) = kx(t) + b x(t), (6.31)
dt dt dt
where M is the mass of the vehicle. Note that the main point of our class is not to come
up with such differential equations — this is what you learn in physics (nevertheless, for
this example, we will briefly discuss how to obtain the differential equation). For our class,
the main point is the next step: What kind of a system is described by this differential
equation, and how is the system behavior influenced by the various parameters in the
differential equation?
The Transfer Function and the Laplace transform permit you:

• To analyze and understand the system behavior and the impact of the choice of the
parameters k and b.

• To design the overall system, i.e., select k and b in an appropriate fashion.

To see how to do this, we first have to find the transfer function of the system under study.
This is a simple matter:

k + bs
H(s) = . (6.32)
M s2 + bs + k

This is often referred to as a second-order system since it has two poles. Since second-order
systems appear frequently, there is a standard form for them: We express the denominator
as s2 + 2ζωn s + ωn2 . If we do this here, we obtain

ωn2 + 2ζωn s
H(s) = , (6.33)
s2 + 2ζωn s + ωn2
66CHAPTER 6. THE TRANSFER FUNCTION AND THE LAPLACE TRANSFORM

where
k b
ωn2 = and 2ζωn = . (6.34)
M M
The poles for this system are at
 p   p 
ωn −ζ + ζ 2 − 1 and ωn −ζ − ζ 2 − 1 . (6.35)

Clearly, our system has to be causal. This means that the ROC must extend unboundedly
to the right (to Re(s) = ∞). The question is whether it is stable. As we have seen, for it
to be stable, the ROC must include the imaginary axis. In summary, this means that our
control system is stable if and only if there are no poles in the right half-plane Re(s) ≥ 0.
Looking back at the formulas for the two poles, it can indeed be verified that the system is
always stable as long as ωn is positive (meaning that b is positive). More precisely, in the
interesting regime where |ζ| < 1, we have two complex-conjugate poles in the left half plane
(i.e., with Re(s) < 0) and thus, the system is stable.
There are further insightful interpretations for this classical second-order system. As
one can show:

• the parameter ωn characterizes how quickly the system responds to a change in the
input signal (in our example, this could be a pothole), and

• the parameter ζ, often referred to as the damping ratio, characterizes the shape of the
response (steep curve of reaction with overshoot versus shallow curve of reaction with
no overshoot). This can be interpreted as being related to how close the poles are to
the imaginary axis: As we know, if the poles are on the imaginary axis, the system
is unstable. In extension of this, if the poles are very close to the imaginary axis, the
system is “almost unstable,” which shows up by a vigorous overshooting. Pictures
will be shown in class for illustration.

As a side note, the parameter ωn is often referred to as the undamped response frequency
of the system. To understand this terminology, suppose that ζ = 0. Then, the system has
two poles on the imaginary axis, located at jωn and −jωn . Hence, its impulse response is a
cosine/sine of frequency ωn . Clearly, for the automobile suspension system, this would not
be a good choice. (It corresponds essentially to a broken dashpot, and can be observed on
certain old cars by leaning on the chassis and then releasing.)
A more general view of control systems is the one captured in Figure 6.6. The idea is
that the output of a certain base system (the system H in the figure) is being measured by
a sensor (the system B in the figure). Ideally, this measurement corresponds to the precise
system output, that is, ỹ(t) = y(t). The systems engineer has a reference trajectory x(t) in
mind that the system should follow. (For example, the trajectory that a drone is supposed
to take.) There is an error e(t) between the actual system output (the actual trajectory
of the drone) and the intended trajectory x(t), hence, e(t) = x(t) − ỹ(t). The controller is
a cleverly designed system that uses this error signal e(t) in such a way as to generate a
system input z(t) that will make the system H get closer to the desired trajectory.
6.6. CONTROL SYSTEMS : STABILITY AND CAUSALITY 67

System
Reference Error Input System Output
x(t) - + ie(t)- Controller z(t) - LTI System y(t) -
G H
−6

Measured Output
ỹ(t) Sensor 
B

Figure 6.6: A generic view of control systems.

With the tools we have learned so far, we can make a few very easy observations. For
example, if we assume that all systems are LTI systems, then we can study transfer functions
and Laplace transforms. We immediately make the following two observations:

E(s) = X(s) − Ỹ (s) = X(s) − B(s)Y (s) (6.36)

Y (s) = H(s)G(s)E(s). (6.37)

But by plugging in, we obtain

Y (s) = H(s)G(s)(X(s) − B(s)Y (s)), (6.38)

which we can rearrange into

Y (s)(1 + H(s)G(s)B(s)) = H(s)G(s)X(s), (6.39)

or more interestingly

Y (s) H(s)G(s)
= . (6.40)
X(s) 1 + H(s)G(s)B(s)

In other words, we have found the end-to-end transfer function of the full system in Fig-
ure 6.6:
H(s)G(s)
Hend−to−end (s) = . (6.41)
1 + H(s)G(s)B(s)

As earlier, we can now discuss various control strategies (i.e., various systems G(s)) in terms
of their causality and stability properties. A particularly interesting consideration is how
the poles of the end-to-end system move in the s-plane as we change the controller G(s).
Several elements are yet missing from Figure 6.6. Perhaps the most interesting concerns
additional external inputs to the system H. For the drone example, this could include the
wind and air pressure situation. Often, it is instructive to model these external disturbances
as random processes. You will encounter models of random processes soon in other classes.
68CHAPTER 6. THE TRANSFER FUNCTION AND THE LAPLACE TRANSFORM

Appendix 6.A Laplace Transform : Properties

Property Signal Transform ROC

x(t) X(s) R
x1 (t) X1 (s) R1
x2 (t) X2 (s) R2

Linearity ax1 (t) + bx2 (t) aX1 (s) + bX2 (s) At least R1 ∩ R2

Shift in time x(t − t0 ) e−st0 X(s) R

Shift in the es0 t x(t) X(s − s0 ) Shifted version of R

s-Domain (i.e., s is in the ROC
if (s − s0 ) is in R)

1 s
Scaling in time x(at) X “Scaled” ROC
|a| a
(i.e., s is in the ROC
if (s/a) is in R)

d
Differentiation x(t) sX(s) At least R
dt
in time

d
Differentiation −tx(t) X(s) R
ds
in the s-Domain
Z t
1
Integration x(τ )d(τ ) X(s) At least R ∩ {Re(s) > 0}
−∞ s
in time

Convolution x1 (t) ∗ x2 (t) X1 (s)X2 (s) At least R1 ∩ R2

Conjugation x∗ (t) X ∗ (s∗ ) R

Conjugate symmetry x(t) real-valued X(s) = X ∗ (s∗ )

6.B. LAPLACE TRANSFORM : PAIRS 69

Appendix 6.B Laplace Transform : Pairs

Signal Transform ROC

Dirac delta function δ(t) 1 All s

δ(t − T ) e−sT All s


1, t ≥ 0 1
Step function u(t) = Re(s) > 0
0, otherwise. s
1
−u(−t) Re(s) < 0
s

tn−1 1
u(t) Re(s) > 0
(n − 1)! sn

tn−1 1
− u(−t) Re(s) < 0
(n − 1)! sn

1
One-sided e−αt u(t) Re(s) > −Re(α)
s+α
exponential
1
−e−αt u(−t) Re(s) < −Re(α)
s+α

tn−1 −αt 1
e u(t) Re(s) > −Re(α)
(n − 1)! (s + α)n

tn−1 −αt 1
− e u(−t) Re(s) < −Re(α)
(n − 1)! (s + α)n

s
One-sided Cosines [cos ω0 t]u(t) Re(s) > 0
s2 + ω02
and Sines
ω0
[sin ω0 t]u(t) Re(s) > 0
s2 + ω02

s+α
[e−αt cos ω0 t]u(t) Re(s) > −Re(α)
(s + α)2 + ω02
ω0
[e−αt sin ω0 t]u(t) Re(s) > −Re(α)
(s + α)2 + ω02
70CHAPTER 6. THE TRANSFER FUNCTION AND THE LAPLACE TRANSFORM
Bibliography

[1] A. V. Oppenheim and A. S. Willsky, with S. Hamid Nawab, Signals and Systems. Upper
Saddle River, NJ: Prentice Hall, 2nd ed., 1996.

[2] M. Vetterli, J. Kovacevic, and V. Goyal, Foundations of Signal Processing. Cambridge

University Press, 2014.

[3] W. Rudin, Real and Complex Analysis. Cambridge, MA: Boston, MA, third ed., 1987.

71