ECE 380: Control Systems

Course Notes: Winter 2014

Prof. Shreyas Sundaram

Department of Electrical and Computer Engineering

University of Waterloo
ii

c Shreyas Sundaram
Acknowledgments

Parts of these course notes are loosely based on lecture notes by Professors
Daniel Liberzon, Sean Meyn, and Mark Spong (University of Illinois), on notes
by Professors Daniel Davison and Daniel Miller (University of Waterloo), and
on parts of the textbook Feedback Control of Dynamic Systems (5th edition) by
Franklin, Powell and Emami-Naeini. I claim credit for all typos and mistakes
in the notes.
The LATEX template for The Not So Short Introduction to LATEX 2ε by T. Oetiker
et al. was used to typeset portions of these notes.

Shreyas Sundaram
University of Waterloo

c Shreyas Sundaram
iv

c Shreyas Sundaram
Contents

1 Introduction 1
1.1 Dynamical Systems . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 What is Control Theory? . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Outline of the Course . . . . . . . . . . . . . . . . . . . . . . . . 4

2 Review of Complex Numbers 5

3 Review of Laplace Transforms 9

3.1 The Laplace Transform . . . . . . . . . . . . . . . . . . . . . . . 9
3.2 The Inverse Laplace Transform . . . . . . . . . . . . . . . . . . . 13
3.2.1 Partial Fraction Expansion . . . . . . . . . . . . . . . . . 13
3.3 The Final Value Theorem . . . . . . . . . . . . . . . . . . . . . . 15

4 Linear Time-Invariant Systems 17

4.1 Linearity, Time-Invariance and Causality . . . . . . . . . . . . . . 17
4.2 Transfer Functions . . . . . . . . . . . . . . . . . . . . . . . . . . 18
4.2.1 Obtaining the transfer function of a differential equation
model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
4.3 Frequency Response . . . . . . . . . . . . . . . . . . . . . . . . . 21

5 Bode Plots 25
5.1 Rules for Drawing Bode Plots . . . . . . . . . . . . . . . . . . . . 26
5.1.1 Bode Plot for Ko . . . . . . . . . . . . . . . . . . . . . . . 27
q
5.1.2 Bode Plot for s . . . . . . . . . . . . . . . . . . . . . . . 28
5.1.3 Bode Plot for ( ps + 1)−1 and ( zs + 1) . . . . . . . . . . . . 29

c Shreyas Sundaram
vi CONTENTS

 ±1
5.1.4 Bode Plot for ( ωsn )2 + 2ζ( ωsn ) + 1 . . . . . . . . . . . 32
5.1.5 Nonminimum Phase Systems . . . . . . . . . . . . . . . . 34

6 Modeling and Block Diagram Manipulation 37

6.1 Mathematical Models of Physical Systems . . . . . . . . . . . . . 37
6.1.1 Mechanical Systems . . . . . . . . . . . . . . . . . . . . . 37
6.1.2 Electrical Systems . . . . . . . . . . . . . . . . . . . . . . 38
6.1.3 Rotational Systems . . . . . . . . . . . . . . . . . . . . . . 39
6.2 Block Diagram Manipulation . . . . . . . . . . . . . . . . . . . . 41
6.2.1 Systems with Multiple Inputs and Outputs . . . . . . . . 45

7 Step Responses of Linear Systems 47

7.1 Step Response of First Order Systems . . . . . . . . . . . . . . . 47
7.1.1 Rise time . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
7.1.2 Settling time . . . . . . . . . . . . . . . . . . . . . . . . . 49
7.2 Step Response of Second Order Systems . . . . . . . . . . . . . . 50
7.2.1 Underdamped and Critically Damped Systems (0 ≤ ζ ≤ 1) 50
7.2.2 Overdamped System (ζ > 1) . . . . . . . . . . . . . . . . 52
7.2.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . 52

8 Performance of Second Order Step Responses 55

8.1 Performance Measures . . . . . . . . . . . . . . . . . . . . . . . . 55
8.1.1 Rise time (tr ) . . . . . . . . . . . . . . . . . . . . . . . . . 56
8.1.2 Peak value (Mp ), Peak time (tp ) and Overshoot OS . . . 56
8.1.3 Settling Time (ts ) . . . . . . . . . . . . . . . . . . . . . . 57
8.2 Choosing Pole Locations to Meet Performance Specifications . . 58
8.3 Effects of Poles and Zeros on the Step Response . . . . . . . . . . 60
8.3.1 Effect of a Zero on the Step Response . . . . . . . . . . . 61
8.3.2 Effect of Poles on the Step Response . . . . . . . . . . . . 63

9 Stability of Linear Time-Invariant Systems 65

9.1 Pole-zero cancellations and stability . . . . . . . . . . . . . . . . 65
9.2 Stability of the Unity Feedback Loop . . . . . . . . . . . . . . . . 67
9.3 Tests for Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . 68

c Shreyas Sundaram
CONTENTS vii

9.3.1 A Necessary Condition for Stability . . . . . . . . . . . . 68

9.3.2 A Necessary and Sufficient Condition: Routh-Hurwitz Test 70
9.3.3 Testing For Degree of Stability . . . . . . . . . . . . . . . 72
9.3.4 Testing Parametric Stability with Routh-Hurwitz . . . . . 73

10 Properties of Feedback 75
10.1 Feedforward Control . . . . . . . . . . . . . . . . . . . . . . . . . 76
10.2 Feedback Control . . . . . . . . . . . . . . . . . . . . . . . . . . . 77

11 Tracking of Reference Signals 81

11.1 Tracking and Steady State Error . . . . . . . . . . . . . . . . . . 82

12 PID Control 89
12.1 Proportional (P) Control . . . . . . . . . . . . . . . . . . . . . . 89
12.2 Proportional-Integral (PI) Control . . . . . . . . . . . . . . . . . 91
12.3 Proportional-Integral-Derivative (PID) Control . . . . . . . . . . 91
12.4 Implementation Issues . . . . . . . . . . . . . . . . . . . . . . . . 93

13 Root Locus 95
13.1 The Root Locus Equations . . . . . . . . . . . . . . . . . . . . . 97
13.1.1 Phase Condition . . . . . . . . . . . . . . . . . . . . . . . 98
13.2 Rules for Plotting the Positive Root Locus . . . . . . . . . . . . . 100
13.2.1 Start Points and (Some) End Points of the Root Locus . 100
13.2.2 Points on the Real Axis . . . . . . . . . . . . . . . . . . . 100
13.2.3 Asymptotic Behavior of the Root Locus . . . . . . . . . . 101
13.2.4 Breakaway Points . . . . . . . . . . . . . . . . . . . . . . . 105
13.2.5 Some Root Locus Plots . . . . . . . . . . . . . . . . . . . 108
13.2.6 Choosing the Gain from the Root Locus . . . . . . . . . . 110
13.3 Rules for Plotting the Negative Root Locus . . . . . . . . . . . . 111

14 Stability Margins from Bode Plots 115

15 Compensator Design Using Bode Plots 123

15.1 Lead and Lag Compensators . . . . . . . . . . . . . . . . . . . . 123
15.2 Lead Compensator Design . . . . . . . . . . . . . . . . . . . . . . 126
15.3 Lag Compensator Design . . . . . . . . . . . . . . . . . . . . . . 130

c Shreyas Sundaram
viii CONTENTS

16 Nyquist Plots 135

16.1 Nyquist Plots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
16.2 Drawing Nyquist Plots . . . . . . . . . . . . . . . . . . . . . . . . 140
16.2.1 Nyquist Plots For Systems With Poles/Zeros On The Imag-
inary Axis . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
16.3 Stability Margins from Nyquist Plots . . . . . . . . . . . . . . . . 146

17 Modern Control Theory: State Space Models 151

17.1 State-Space Models . . . . . . . . . . . . . . . . . . . . . . . . . . 151
17.2 Nonlinear State-Space Models and Linearization . . . . . . . . . 156
17.2.1 Linearization via Taylor Series . . . . . . . . . . . . . . . 157
17.3 The Transfer Function of a Linear State-Space Model . . . . . . 159
17.4 Obtaining the Poles from the State-Space Model . . . . . . . . . 160
17.5 An Overview of Design Approaches for State-Space Models . . . 162

c Shreyas Sundaram
Chapter 1

Introduction

1.1 Dynamical Systems

For the purposes of this course, a system is an abstract object that accepts inputs
and produces outputs in response. Systems are often composed of smaller com-
ponents that are interconnected together, leading to behavior that is more than
just the sum of its parts. In the control literature, systems are also commonly
referred to as plants or processes.

Input Output
System

Figure 1.1: An abstract representation of a system.

The term dynamical system loosely refers to any system that has an internal
state and some dynamics (i.e., a rule specifying how the state evolves in time).
This description applies to a very large class of systems, from automobiles and
aviation to industrial manufacturing plants and the electrical power grid. The
presence of dynamics implies that the behavior of the system cannot be en-
tirely arbitrary; the temporal behavior of the system’s state and outputs can be
predicted to some extent by an appropriate model of the system.
Example 1. Consider a simple model of a car in motion. Let the speed of
the car at any time t be given by v(t). One of the inputs to the system is the
acceleration a(t), applied by the throttle. From basic physics, the evolution of
the speed is given by
dv
= a(t). (1.1)
dt
The quantity v(t) is the state of the system, and equation (1.1) specifies the
dynamics. There is a speedometer on the car, which is a sensor that measures

c Shreyas Sundaram
2 Introduction

the speed. The value provided by the sensor is denoted by s(t) = v(t), and this
is taken to be the output of the system.

As shown by the above example, the inputs to physical systems are applied via
actuators, and the outputs are measurements of the system state provided by
sensors.
Other examples of systems: Electronic circuits, DC Motor, Economic Sys-
tems, . . .

1.2 What is Control Theory?

The field of control systems deals with applying or choosing the inputs to a
given system to make it behave in a certain way (i.e., make the state or output
of the system follow a certain trajectory). A key way to achieve this is via the
use of feedback, where the input depends on the output in some way. This is
also called closed loop control.

Input Output
System

Mapping from
output to input

Figure 1.2: Feedback Control.

Typically, the mapping from outputs to inputs in the feedback loop is performed
via a computational element known as a controller, which processes the sensor
measurements and converts it to an appropriate actuator signal. The basic
architecture is shown below. Note that the feedback loop typically contains
disturbances that we cannot control.

Desired Control
Output Input Output
Controller System
−

Figure 1.3: Block Diagram of a feedback control system.

Example 2 (Cruise Control). Consider again the simple model of a car from
Example 1. A cruise control system for the car would work as follows.

c Shreyas Sundaram
1.2 What is Control Theory? 3

• The speedometer in the car measures the current speed and produces
s(t) = v(t).

• The controller in the car uses these measurements to produce control sig-
nals: if the current measurement s(t) is less than the desired cruising
speed, the controller sends a signal to the throttle to accelerate, and if
s(t) is greater than the desired speed, the throttle is asked to allow the
car to slow down.

• The throttle performs the action specified by the controller.

The motion of the car might also be affected by disturbances such as wind
gusts, or slippery road conditions. A properly designed cruise control system
will maintain the speed at (or near) the desired value despite these external
conditions.

Example 3 (Inverted Pendulum). Suppose we try to balance a stick vertically

in the palm of our hand. The sensor, controller and actuator in this example
are our eyes, our brain, and our hand, respectively. This is an example of a
feedback control system. Now what happens if we try to balance the stick
with our eyes closed? The stick inevitably falls. This illustrates another type
of control, known as feedforward or open loop control, where the input to the
system does not depend on the output. As this example illustrates, feedforward
control is not robust to disturbances – if the stick is not perfectly balanced to
start, or if our hand moves very slightly, the stick will fall. This illustrates the
benefit of feedback control.

As we will see later, feedback control has many strengths, and is used to achieve
the following objectives.

• Good tracking. Loosely speaking, feedback control allows us to make

the output of the system follow the desired reference input (i.e., make the
system behave as it should).

• Disturbance rejection. Feedback control allows the system to maintain

good behavior even when there are external inputs that we cannot control.

• Robustness. Feedback control can work well even when the actual model
of the plant is not known precisely; sufficiently small errors in modeling
can be counteracted by the feedback input.

Feedback control is everywhere; it appears not only in engineered systems (such

as automobiles and aircraft), but also in economic systems (e.g., choosing the
interest rates to maintain a desired rate of inflation, growth, etc.), ecological sys-
tems (predator/prey populations, global climate) and biological systems (e.g.,
physiology in animals and plants).

c Shreyas Sundaram
4 Introduction

Example 4 (Teary Eyes on a Cold Day). Blinking and tears are a feedback
mechanism used by the body to warm the surface of the eyeball on cold days –
the insides of the lids warm the eyes. In very cold situations, tears come from
inside the body (where they are warmed), and contain some proteins and salts
that help to prevent front of eyes from freezing.

1.3 Outline of the Course

Since control systems appear in a large variety of applications, we will not
attempt to discuss each specific application in this course. Instead, we will
deal with the underlying mathematical theory, analysis, and design of control
systems. In this sense, it will be more mathematical than other engineering
courses, but will be different from other math courses in that it will pull together
various branches of mathematics for a particular purpose (i.e., to design systems
that behave in desired ways).
The trajectory of the course will be as follows.

• Modeling: Before we can control a system and make it behave in a

desired manner, we need to represent the input-output behavior of the
system in a form that is suitable for mathematical analysis.

• Analysis: Once we understand how to model systems, we need to have a

basic understanding of what the model tells us about the system’s response
to input signals. We will also need to formulate how exactly we want the
output to get to its desired value (e.g., how quickly should it get there, do
we care what the output does on the way there, can we be sure that the
output will get there, etc.)

• Design: Finally, once we have analyzed the mathematical model of the

system, we will study ways to design controllers to supply appropriate
control (input) signals to the system so that the output behaves as we
want it to.

We will be analyzing systems both in the time-domain (e.g., with differential

equations) and in the frequency domain (e.g., using Laplace transforms). We
will start by reviewing some relevant mathematical concepts.

c Shreyas Sundaram
Chapter 2

Review of Complex
Numbers

Consider the polynomial f (x) = x2 + 1. The roots of the polynomial are the
values of x for which f (x) = 0, or equivalently x2 = −1. Clearly, there are no
real numbers that satisfy this equation. To address this problem, let us define
a new “number” j, such that j 2 = −1. Since this number does not belong to
the set of real numbers, we will call it an imaginary or complex number. With
this number in hand, we can actually generate an infinite set of other complex
numbers.1

Definition 1 (Complex Numbers). A complex number s is

of the form s = σ + jω, where σ and ω are real numbers.
The number σ is called the real part of s, and is denoted
by σ = Re(s). The number ω is called the imaginary part
of s and is denoted by ω = Im(s). The set of all complex
numbers is denoted by C.

A complex number σ + jω can alternatively be viewed simply as a pair of real

numbers (σ, ω). Thus, we can plot complex numbers in a two-dimensional plane
called the complex plane. The horizontal axis is called the real axis and the
vertical axis is called the imaginary axis. The real axis represents all complex
numbers s such that Im(s) = 0; in other words, it contains all real numbers, and
thus real numbers are a subset of the complex numbers. The imaginary axis
represents all complex numbers s such that Re(s) = 0. The following regions of
the plane will be useful to our discussions.
1 An excellent (and intuitive) perspective on complex numbers can be found

in a New York Times essay written by Steven Strogatz; it is available at

http://opinionator.blogs.nytimes.com/2010/03/07/finding-your-roots/.

c Shreyas Sundaram
6 Review of Complex Numbers

• All complex numbers s satisfying Re(s) < 0 are said to lie in the Open
Left Half Plane (OLHP).
• All complex numbers s satisfying Re(s) ≤ 0 are said to lie in the Closed
Left Half Plane (CLHP).
• All complex numbers s satisfying Re(s) > 0 are said to lie in the Open
Right Half Plane (ORHP).
• All complex numbers s satisfying Re(s) ≥ 0 are said to lie in the Closed
Right Half Plane (CRHP).

Figure 2.1: A complex number s = σ + jω in the complex plane.

An alternative representation of complex numbers is the polar form. Specifically,

given a complex number s = σ + jω, note that s is also specified uniquely by its
distance r from the origin, and the angle θ that s (as a vector from the origin)
makes with the positive real axis:
p ω
r = σ 2 + ω 2 , θ = arctan .
σ
The number r is called the magnitude of s, and is denoted by r = |s|. The
number θ is called the phase of s (in radians), and is denoted by θ = ∠s.
Conversely, we have σ = r cos θ and ω = r sin θ.
Example. What is the polar form representation of s = 3 − j4?

Note from Euler’s equation that ejθ = cos θ +j sin θ, and so the complex number
s can be denoted by s = rejθ .
Given the complex number s = σ + jω, its complex conjugate is defined as the
complex number s∗ = σ − jω. Note that
ss∗ = (σ + jω)(σ − jω) = σ 2 + ω 2 = |s|2 = |s∗ |2 .
In the geometric representation, the complex conjugate of a complex number is
obtained by reflecting the vector about the real axis.
As we will see throughout the course, many interesting properties of control
systems are related to the roots of certain polynomials. The following result
explains why complex numbers are so useful.

c Shreyas Sundaram
 7

Theorem 1 (Fundamental Theorem of Algebra). Consider

any polynomial f (x) = an xn + an−1 xn−1 + · · · + a1 x + a0 ,
where the coefficients an , an−1 , . . . , a0 are complex numbers.
Then the polynomial has n complex roots. Furthermore, if all
coefficients are real, then all non-real roots of f (x) appear in
complex conjugate pairs.

Decomposing a Real Polynomial into Factors

Let f (x) be a polynomial of degree n with all real coefficients. Suppose that
f (x) has q roots at x = 0, r nonzero real roots at locations z1 , z2 , . . . , zr , and 2p
complex conjugate roots at c1 , c∗1 , c2 , c∗2 , . . . , cp , c∗p (we know that these complex
roots appear in complex conjugate pairs). Now, note that

(x − ci )(x − c∗i ) = x2 − (ci + c∗i )x + ci c∗i = x2 − 2Re(ci )x + |ci |2 .

If we define ωi = |ci | and

Re(ci )
ζi = −
ωi
we can write
(x − ci )(x − c∗i ) = x2 + 2ζi ωi x + ωi2 .
We will be using this notation frequently throughout the course. The polynomial
f (x) can therefore always be decomposed as

f (x) = K0 xq (x − z1 )(x − z2 ) · · · (x − zr )(x2 + 2ζ1 ω1 x + ω12 ) · · · (x2 + 2ζp ωp x + ωp2 )

for some constant K0 .

c Shreyas Sundaram
8 Review of Complex Numbers

c Shreyas Sundaram
Chapter 3

Review of Laplace
Transforms

3.1 The Laplace Transform

Suppose f (t) is a function of time. In this course, unless otherwise noted, we
will only deal with functions that satisfy f (t) = 0 for t < 0.

The Laplace Transform of the function f (t) is defined as

Z ∞
L{f (t)} = f (t)e−st dt .
0

This is a function of the complex variable s, so we can write L{f (t)} = F (s).
Note: There are various conditions that f (t) must satisfy in order to have
a Laplace Transform. For example, it must not grow faster than est for some
s. In this course, we will only deal with functions that have (unique) Laplace
Transforms.1
Example. Find the Laplace Transform of f (t) = e−at , t ≥ 0, where a ∈ R.
Solution.
Z ∞
F (s) = L{f (t)} = e−at e−st dt
0
Z ∞
= e−(s+a)t dt
0
1 Uniqueness may be lost at points of discontinuity. More specifically, if f (t) and g(t) are
piecewise continuous functions, then F (s) = G(s) for all s implies that f (t) = g(t) everywhere
except at the points of discontinuity.

c Shreyas Sundaram
10 Review of Laplace Transforms

1 ∞ 1
=− e−(s+a)t  = (if Re(s + a) > 0) .

s+a 0 s+a

Example. Find the Laplace Transform of the unit step function 1(t), defined
as 
1 if t ≥ 0,
1(t) =
0 otherwise.

Figure 3.1: The unit step function 1(t).

Solution.

Example: The Dirac delta function (or impulse function). Consider the
function δ (t) defined as
1
if 0 ≤ t ≤ ,
δ (t) = 
0 otherwise,
where  is a small positive number.

Figure 3.2: (a) The function δ (t). (b) The impulse function δ(t).
R∞
Note that δ (t)dt
−∞ 
= 1. Now suppose that we let  → 0; we still have
Z ∞
lim δ (t)dt = 1 .
→0 −∞

Note that as  gets smaller, the function gets narrower, but taller. Define the
impulse function 
∞ if t = 0,
δ(t) = lim δ (t) =
→0 0 otherwise.
Some properties of δ(t):

c Shreyas Sundaram
3.1 The Laplace Transform 11

R∞
• −∞
δ(t)dt = 1.
R∞
• Let f (t) be any function. Then −∞
δ(t − τ )f (t)dt = f (τ ). This is called
the sifting property.

What is the Laplace Transform

R∞ of δ(t)?
Solution. L{δ(t)} = 0 δ(t)e−st dt = e0 = 1.
Transforms of other common functions can be obtained from Laplace Transform
tables:

Table 3.1: Sample Table of Laplace Transforms.

f (t), t ≥ 0 F (s)
δ(t) 1
1
1(t) s
e−at 1
s+a
1
t s2
ω
sin ωt s2 +ω 2
s
cos ωt s2 +ω 2
.. ..
. .

Note: The functions f (t) in this table are only defined for t ≥ 0 (i.e., we are
assuming that f (t) = 0 for t < 0).

Some properties of the Laplace Transform

1. Linearity. L{αf1 (t) + βf2 (t)} = αF1 (s) + βF2 (s).

Example. What is L{5e−3t − 2 cos 4t}?
Solution.

2. Time Delay. Consider a delayed signal f (t − λ), λ ≥ 0 (i.e., a shifted

version of f (t)).

Z ∞ Z ∞
L{f (t − λ)} = f (t − λ)e−st dt = f (t − λ)e−st dt
0
Zλ∞
= f (τ )e−s(τ +λ) dτ (letting τ = t − λ)
0
−sλ
=e F (s) .

c Shreyas Sundaram
12 Review of Laplace Transforms

Figure 3.3: The time-delayed signal f (t − λ).

Example. What is L{1(t − 5)}?

Solution.

3. Differentiation. L{ df
dt } = sF (s) − f (0). More generally,

dm f m m−1 m−2 df dm−1 f

L{ } = s F (s) − s f (0) − s (0) − · · · − (0) .
dtm dt dtm−1
d −at
Example. What is L{ dt e }?
Solution.

Rt
4. Integration. L{ 0 f (τ )dτ } = 1s F (s).
Rt
Example. What is L{ 0 cos τ dτ }?
Solution.

5. Convolution. The convolution of two signals f1 (t) and f2 (t) is denoted

by Z t
f1 (t) ∗ f2 (t) = f1 (τ )f2 (t − τ )dτ .
0

The Laplace Transform of the convolution of two signals is given by

L{f1 (t) ∗ f2 (t)} = F1 (s)F2 (s) .

c Shreyas Sundaram
3.2 The Inverse Laplace Transform 13

This is a very important property of Laplace Transforms! Convolution of

two signals in the time-domain (which could be hard to do) is equivalent
to the multiplication of their Laplace Transforms in the s domain (which
is easy to do).
Example. What is L{1(t) ∗ 1(t)}?
Solution.

3.2 The Inverse Laplace Transform

Given the Laplace Transform F (s), we can obtain the corresponding time-
domain function f (t) via the Inverse Laplace Transform:
Z σ+j∞
1
f (t) = L−1 {F (s)} = F (s)est ds .
2πj σ−j∞
This is quite tedious to apply in practice, and so we will not be using it in this
class. Instead, we can simply use the Laplace Transform tables to obtain the
corresponding functions.
Example. What is L−1 { s(s+1)1
}?
Solution.
1 1 1
L−1 { } = L−1 { − }
s(s + 1) s s+1
1 1
= L−1 { } − L−1 { }
s s+1
(by linearity of the (inverse) Laplace Transform)
= 1(t) − e−t , t ≥ 0 .

1
In the above example, we “broke up” the function s(s+1) into a sum of simpler
functions, and then applied the inverse Laplace Transform (by consulting the
Laplace Transform table) to each of them. This is a general technique for
inverting Laplace Transforms.

3.2.1 Partial Fraction Expansion

Suppose we have a rational function
bm sm + bm−1 sm−1 + · · · + b1 s + b0 N (s)
F (s) = = ,
sn + an−1 sn−1 + · · · + a1 s + a0 D(s)
where the ai ’s and bi ’s are constant real numbers.

c Shreyas Sundaram
14 Review of Laplace Transforms

Definition 2. If m ≤ n, the rational function is called

proper. If m < n, it is strictly proper.

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

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

Definition 3 (Zeros and Poles). The complex numbers

−z1 , −z2 , . . . , −zm are the roots of N (s) and are called the
zeros of F (s). The complex numbers −p1 , −p2 , . . . , −pn are
the roots of D(s) and are called the poles of F (s).

Note: Remember these terms, as the poles and zeros of a system will play a
very important role in our ability to control it.
First, suppose each of the poles are distinct and that F (s) is strictly proper.
We would like to write

k1 k2 kn
F (s) = + + ··· + ,
s + p1 s + p2 s + pn

for some constants k1 , k2 , . . . , kn , since the inverse Laplace Transform of F (s)

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

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

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

ki = (s + pi )F (s)|s=−pi .

s+5
Example. What is the partial fraction expansion of F (s) = s3 +3s2 −6s−8 ?

c Shreyas Sundaram
3.3 The Final Value Theorem 15

Solution.

The partial fraction expansion when some of the poles are repeated is obtained
by following a similar procedure, but it is a little more complicated. We will
not worry too much about this scenario here. One can also do a partial fraction
expansion of nonstrictly proper functions by first dividing the denominator into
the numerator to obtain a constant and a strictly proper function, and then
applying the above partial fraction expansion. The details will be covered in a
homework problem.

3.3 The Final Value Theorem

Let F (s) be the Laplace Transform of a function f (t). Often, we will be in-
terested in how f (t) behaves as t → ∞ (this is referred to as the asymptotic
or steady state behavior of f (t)). Can we obtain this information directly from
F (s)?

Theorem 2 (Final Value Theorem). If all poles of sF (s)

are in the open left half plane, then

lim f (t) = lim sF (s) .

t→∞ s→0

c Shreyas Sundaram
16 Review of Laplace Transforms

1
Example. F (s) = s(s+1) . What is limt→∞ f (t)?
Solution.

Why do we need the poles of sF (s) to be in the OLHP in order to apply the
Final Value Theorem? First, note from the partial fraction expansion of F (s)
that
k1 k2 k3 kn
F (s) = + + + ··· +
s s + p2 s + p3 s + pn
⇔ f (t) = k1 1(t) + k2 e−p2 t + k3 e−p3 t + · · · + kn e−pn t .

Note that if F (s) does not have a pole at s = 0, then the constant k1 will simply
be zero in the above expansion. Based on the above expansion, if one of the
poles has a positive real part, the corresponding exponential term will explode,
and thus f (t) will have no final value! On the other hand, if all poles have
negative real parts, all of the exponential terms will go to zero, leaving only the
term k1 1(t) (corresponding to ks1 in the partial fraction expansion). Thus, the
asymptotic behavior of f (t) is simply k1 , and this is obtained by calculating
lims→0 sF (s).
One must be careful about applying the Final Value Theorem; if the signal f (t)
does not settle down to some constant steady state value, the theorem might
yield nonsensical results. For example, consider the function f (t) = sin t, with
Laplace transform
1
F (s) = 2 .
s +1
The function sF (s) does not have all poles in the OLHP (it has two poles on the
imaginary axis). However, if we forgot to check this before applying the Final
Value Theorem, we would get

lim sF (s) = 0,
s→0

and would mistakenly assume that f (t) → 0. Clearly f (t) has no steady state
value (it constantly oscillates between −1 and 1).

c Shreyas Sundaram
Chapter 4

Linear Time-Invariant
Systems

4.1 Linearity, Time-Invariance and Causality

Consider the system:

Input Output
System

The system is said to be linear if the Principle of Superposition holds.

Definition 4 (Principle of Superposition). Suppose that the

output of the system is y1 (t) in response to input u1 (t) and
y2 (t) in response to input u2 (t). Then the output of the
system in response to the input αu1 (t) + βu2 (t) is αy1 (t) +
βy2 (t), where α and β are arbitrary real numbers. Note that
this must hold for any inputs u1 (t) and u2 (t).

The system is said to be time-invariant if the output of the system is y(t − τ )

when the input is u(t − τ ) (i.e., a time-shifted version of the input produces an
equivalent time-shift in the output).
The system is said to be causal if the output at time t depends only on the
input up to time t. In particular, this means that if u(t) = 0 for t < τ , then
y(t) = 0 for t < τ .

c Shreyas Sundaram
18 Linear Time-Invariant Systems

Figure 4.1: The Principle of Superposition.

Figure 4.2: The Time-Invariance Property

In this class, we will primarily be dealing with the analysis of linear time-
invariant causal systems (if the system is nonlinear, we will linearize it). We
will be interested in how the system responds to certain types of inputs. The
impulse response of a system is the output of the system when the input to
the system is δ(t), and is denoted by the signal h(t). The step response is the
output of the system when the input to the system is 1(t).
Note: The impulse response for causal systems satisfies h(t) = 0 for t < 0.

4.2 Transfer Functions

R∞
Recall from the sifting property of the delta function that u(t) = 0 u(τ )δ(t −
τ )dτ . Note that this can be interpreted in the following way: the signal u(t) is
the sum of an infinite series of weighted and shifted delta functions. The delta
function at t = τ is weighted by the value u(τ ). Now, the time-invariance of

c Shreyas Sundaram
4.2 Transfer Functions 19

Figure 4.3: (a) A causal system. (b) A non-causal system.

the system says that the output of the system due to the input δ(t − τ ) will be
h(t − τ ). Using the Principle of Superposition, we see that the output of the
system when u(t) is the input will be the infinite sum (or integral) of weighted
and shifted impulse responses (when all initial conditions are equal to zero):
Z ∞
y(t) = u(τ )h(t − τ )dτ .
0

Note that the output of the system is just the convolution of the signals u(t)
and h(t)! This is a well known relationship for linear time-invariant systems.
Applying Laplace Transforms to both sides of the above equation, we obtain
Y (s) = H(s)U (s), or equivalently,

Y (s)
H(s) = .
U (s)

The function H(s) is the ratio of the Laplace Transform of the output to the
Laplace Transform of the input (when all initial conditions are zero), and it is
called the transfer function of the system. Note that this transfer function is
independent of the actual values of the inputs and outputs – it tells us how any
input gets transformed into the output. It is a property of the system itself. We
will frequently represent systems in block diagrams via their transfer functions:

Input Output
U (s) Y (s)
H(s)

Note: The transfer function is the Laplace Transform of the impulse response
of the system.

c Shreyas Sundaram
20 Linear Time-Invariant Systems

4.2.1 Obtaining the transfer function of a differential equa-

tion model
We will be dealing with systems whose input and output are related via a
differential equation of the form

y (n) + an−1 y (n−1) + · · · + a1 ẏ + a0 y = bm u(m) + bm−1 u(m−1) + · · · + b1 u̇ + b0 u .

Taking Laplace Transforms of both sides (assuming that all initial conditions
are zero), we get

(sn +an−1 sn−1 +· · ·+a1 s+a0 )Y (s) = (bm sm +bm−1 sm−1 +· · ·+b1 s+b0 )U (s) ,

from which we obtain the transfer function

Y (s) bm sm + bm−1 sm−1 + · · · + b1 s + b0
H(s) = = .
U (s) sn + an−1 sn−1 + · · · + a1 s + a0

The impulse response of this system is given by h(t) = L−1 {H(s)}.

Example. Consider the system given by the differential equation ẏ+3y = 2u(t).
What is the transfer function of this system? What is the impulse response?
What is the step response?
Solution.

c Shreyas Sundaram
4.3 Frequency Response 21

4.3 Frequency Response

Suppose that the input to the system is u(t) = es0 t , t ≥ 0, for some s0 = σ + jω.
Recall that
Z ∞ Z ∞
y(t) = u(τ )h(t − τ )dτ = u(t − τ )h(τ )dτ .
0 0

Since both u(t) and h(t) are zero for t < 0, the above expression becomes
Z t Z t Z t
y(t) = u(t − τ )h(τ )dτ = e s0 (t−τ )
h(τ )dτ = e s0 t
e−s0 τ h(τ )dτ .
0 0 0
Rt
The quantity 0 e−s0 τ h(τ )dτ looks a lot like the Laplace Transform of the signal
h(τ ) evaluated at s = s0 , except that the upper limit on the integration is t
instead of ∞. Suppose that we examine what happens when t becomes very
large (i.e., as t → ∞). In this case, if the integral exists, we can write

y(t) = H(s0 )es0 t .

Thus, the asymptotic response (if H(s0 ) exists) to a complex exponential input
is that same complex exponential, scaled by the transfer function evaluated at
s0 . This gives us one potential way to identify the transfer function of a given
“black box” system: apply the input es0 t for several different values of s0 , and
use that to infer H(s).
Problem: If Re(s0 ) > 0, then es0 t blows up very quickly, and if Re(s0 ) < 0, it
decays very quickly. What about s0 = jω? This would solve the problem. How
do we apply ejωt in practice?
jωt −jωt
Solution: Sinusoids. Recall the identity cos ωt = e +e 2 . Using the
Principle of Superposition and the property derived above:

• The response of the system to the input u1 (t) = 21 ejωt is

1
y1 (t) = H(jω)ejωt .
2

• The response of the system to the input u2 (t) = 21 e−jωt is

1
y2 (t) = H(−jω)e−jωt .
2

Note that H(jω) is just a complex number, and so we can write it in the polar
form H(jω) = |H(jω)|ej∠H(jω) , where |H(jω)| is the magnitude of H(jω) and
∠H(jω) is the phase of H(jω) (they will both depend on the choice of ω).
Similarly, H(−jω) is just the complex conjugate of H(jω),1 and so we can
1 You should be able to prove this by using the fact that the numerator and denominator

of H(s) are polynomials with real coefficients.

c Shreyas Sundaram
22 Linear Time-Invariant Systems

write H(−jω) = |H(jω)|e−j∠H(jω) . Using these identities, and the Principle of

Superposition, the output of the system when the input is u(t) = u1 (t)+u2 (t) =
cos ωt is
1 1
y(t) = y1 (t) + y2 (t) = H(jω)ejωt + H(−jω)e−jωt
2 2
1 1
= |H(jω)|ej∠H(jω) ejωt + |H(jω)|e−j∠H(jω) e−jωt
2 2
1  
= |H(jω)| ej(ωt+∠H(jω)) + e−j(ωt+∠H(jω))
2
= |H(jω)| cos(ωt + ∠H(jω)) .

In other words, the (steady-state) response to the sinusoid cos ωt is a scaled and
phase-shifted version of the sinusoid! This is called the frequency response
of the system, and will be a useful fact to identify and analyze linear systems.
Later, we will be plotting the magnitude and phase of the system as we sweep
ω from 0 to ∞; this is called the Bode plot of the system.
As an example, consider a linear system with transfer function

ωn2
H(s) = ,
s2 + 2ζωn s + ωn2

where ζ and ωn are some real numbers. We will study systems of this form in
more detail later in the course. Since the denominator of this transfer function
has degree 2, it is called a second order system. The magnitude of this function
at s = jω is given by

ωn2
|H(jω)| =
| − ω 2 + 2ζωn ωj + ωn2 |
ωn2 1
=p =q ,
2 2 2 2
(ωn − ω ) + 4ζ ωn ω2 2
(1 − ( ωωn )2 )2 + 4ζ 2 ( ωωn )2

and the phase is given by

ωn2
∠H(jω) = ∠
−ω 2 + 2ζωn ωj + ωn2
 
1 2ζ ωωn
=∠ = − tan−1  2  .
 
−( ωωn )2 + 2ζ( ωωn )j + 1

ω
1− ωn

Since these quantities are a function of ωωn , we can plot them vs ωωn for vari-
ous values of ζ. Note that in the following plots, we used a logarithmic scale
for the frequency. This is commonly done in order to include a wider range
of frequencies in our plots. The intervals on logarithmic scales are known as

c Shreyas Sundaram
4.3 Frequency Response 23

decades.

We can label the following important characteristics of the magnitude plot:

c Shreyas Sundaram
24 Linear Time-Invariant Systems

• The magnitude of the transfer function for low frequencies (i.e., near ω =
0) is called the low frequency gain or the DC gain.
• The bandwidth of the system is the frequency at which the magnitude
drops to √12 times the DC gain, and is denoted by ωBW . For the sec-
ond order system considered above, the plot shows that the bandwidth is
approximately equal to ωn .
• The resonant peak is the difference between maximum value of the fre-
quency response magnitude and the DC gain, and is denoted by Mr .

• The resonant frequency is the frequency at which the resonant peak

occurs, and is denoted by ωr .

The concepts of bandwidth and DC gain will play an important role in this
course, and you should be comfortable manipulating and deriving these quan-
tities. While we identified these metrics using the magnitude plot of a second
order system, these quantities can be used to discuss the frequency response
of any transfer function. For simple first order systems, we can calculate these
quantities explicitly, as shown in the following example.
Example. Find the bandwidth and DC gain of the system with transfer func-
b
tion H(s) = s+a .
Solution. The DC gain is the magnitude of the transfer function when s = jω,
with ω = 0. In this case, the DC gain is
b
|H(0)| = .
a
The bandwidth is the frequency ωBW at which the magnitude |H(jωBW )| is
equal to √12 of the DC gain. In this case, we have

1 b b b
√ = |H(jωBW )| = =p .
2 a |jω BW + a| 2 2
a + ωBW

Cancelling out the b and squaring both sides, we obtain

2a2 = a2 + ωBW
2
,

which means that ωBW = a.

Next, we will develop some ways to systematically draw the magnitude and
phase for general transfer functions.

c Shreyas Sundaram
Chapter 5

Bode Plots

A Bode plot is a plot of the magnitude and phase of a linear system, where
the magnitude is plotted on a logarithmic scale, and the phase is plotted on
a linear scale. Specifically, consider the linear system with transfer function
H(s) = N (s)
D(s) . For the moment, assume that all poles and zeros of the transfer
function are real (to avoid cumbersome notation), and write
(s + z1 )(s + z2 ) · · · (s + zm )
H(s) = K .
(s + p1 )(s + p2 ) · · · (s + pn )
When working with Bode plots, we will find it more convenient to write this
system as:
( s + 1)( zs2 + 1) · · · ( zsm + 1)
H(s) = Ko zs1 ,
( p1 + 1)( ps2 + 1) · · · ( psn + 1)
where Ko = K zp11zp22...z
...pn . This is called the Bode form, and the reason for doing
m

this is that the DC gain of the above transfer function can now immediately be
obtained as Ko , which will be useful when drawing Bode plots. We will handle
more general transfer functions after the following discussion.
The magnitude of H(jω) is given by
|j zω1 + 1||j zω2 + 1| · · · |j zωm + 1|
|H(jω)| = |Ko | .
|j pω1 + 1||j pω2 + 1| · · · |j pωn + 1|

Now if we take the logarithm of both sides (any base is acceptable, but base 10
is conventional), we get
m n
X ω X ω
log |H(jω)| = log |Ko | + log |j + 1| − log |j + 1| .
i=1
zi i=1
pi

In other words, when viewed on a logarithmic scale, the magnitudes of each

of the individual terms in the transfer function add together to produce the

c Shreyas Sundaram
26 Bode Plots

magnitude of the overall transfer function. This is quite useful, and the reason
for introducing the logarithm. In keeping with convention, we will multiply both
sides of the above equation by 20, and work with the units in decibels; this only
scales the magnitudes, but does not change the additivity due to the logarithm.
Note that the phase of H(jω) already satisfies the additivity property:
m n
X ω X ω
∠H(jω) = ∠Ko + ∠(j + 1) − ∠(j + 1) ,
i=1
zi i=1
pi

and thus it suffices to consider the phase on a linear scale.

Note that we can always draw Bode plots for any transfer function by simply
evaluating the magnitude and phase for each value of ω and then plotting these
values. However, will want to come up with some quick rules to sketch these
plots.

5.1 Rules for Drawing Bode Plots

Now that we understand the general motivation behind drawing the log-magnitude
and phase of Bode plots, we will study how to draw the Bode plots of general
transfer functions (potentially involving complex conjugate terms as well). We
will assume for now that the transfer function H(s) has all zeros and poles in the
CLHP (we will discuss the more general case later). Recall from Chapter 2 that
a polynomial with real coefficients can always be decomposed into a product of a
constant, first order terms (corresponding to real roots) and second order terms
(corresponding to complex conjugate roots). Thus, we can assume (without loss
of generality) that H(s) is composed of four different types of factors:

• Ko (a constant)
• sq (corresponding to zeros at the origin if q is a positive integer, or poles
at the origin if q is a negative integer)
• ( ps + 1)−1 and ( zs + 1) (corresponding to real poles and zeros)
 ±1
• ( ωsn )2 + 2ζ( ωsn ) + 1 (corresponding to complex conjugate zeros if the
exponent is 1, and complex conjugate poles if the exponent is −1).
2
Example. Consider H(s) = 3 s3(s+2)(s +2s+4)
(s+1)(s2 +3s+4) . Write this in Bode form, and
write the logarithm of the magnitude of H(jω) in terms of the logarithm of the
magnitudes of each of the factors. Also write the phase of H(jω) in terms of
the phases of each of the factors.

c Shreyas Sundaram
5.1 Rules for Drawing Bode Plots 27

Since the log-magnitude and phase are obtained by simply adding together the
log-magnitudes and phases of the individual factors, we can draw the Bode plot
for the overall system by drawing the Bode plots for each of the individual
factors, and then adding the plots together.

5.1.1 Bode Plot for Ko

The Bode plot of constant terms is very easy to draw. The log-magnitude is
simply 20 log |Ko |, and the phase is

0 if Ko > 0
∠Ko =
π if Ko < 0.

Both the magnitude and phase are just horizontal lines.

1
Example. Draw the Bode plot of 10 .
Solution.

c Shreyas Sundaram
28 Bode Plots

Example. Draw the Bode plot of −10.

Solution.

5.1.2 Bode Plot for sq

The log-magnitude of the factor sq for s = jω is given by

20 log |(jω)q | = 20q log |jω| = 20q log |ω| .

On a log scale, this simply a straight line with slope 20q, going through the
point 0 when ω = 1.
The phase of sq at s = jω is
π
∠(jω)q = q∠jω = q ,
2
which is just a horizontal line at q π2 .
Example. Draw the Bode plot of s2 .
Solution.

c Shreyas Sundaram
5.1 Rules for Drawing Bode Plots 29

1
Example. Draw the Bode plot of s3 .
Solution.

5.1.3 Bode Plot for ( ps + 1)−1 and ( zs + 1)

A system zero at s = −z will correspond to the factor ( zs + 1) in the Bode form

of the transfer function, and a system pole at s = −p will correspond to the
factor ( ps + 1)−1 , where we are assuming that z > 0 and p > 0. Let’s deal with
the zero first. The log-magnitude of the factor ( zs + 1) at s = jω is given by

r
ω ω
20 log |j + 1| = 20 log 1 + ( )2 .
z z

If ω  z, we have 20 log |j ωz + 1| ≈ 0, and thus for values of ω less than z, the

magnitude is simply a horizontal line at 0. If ω  z, we have 20 log |j ωz + 1| ≈
20 log( z1 ) + 20 log(ω). On a log scale, this is a line of slope 20, going through
the point 0 when ω = z. These two rules together produce a magnitude plot

c Shreyas Sundaram
30 Bode Plots

that looks like this:

The point ω = z is called the breakpoint. These straight lines were derived
based on values of ω that were much smaller or much larger than the breakpoint,
and thus they are called asymptotes. Note that they are only approximations
to the shape of the actual magnitude plot – for example, the actual value of
20 log(j ωz + 1) at ω = z is equal to 3 dB. However, the approximations will
suffice for us to obtain some general intuition about Bode plots.

The phase of (j ωz + 1) is given by

ω ω
∠(j + 1) = tan−1 .
z z

For ω  z, this is approximately equal to 0, and for ω  z, this is approximately

equal to π2 . At the breakpoint ω = z, the phase is tan−1 (1) = π4 . The phase
curve transitions smoothly from 0 to π2 through π4 ; for convenience, we will draw
the transition as a line starting at ω = 0.1z and ending at ω = 10z (i.e., one
decade before and one decade after z):

The magnitude and phase plots for a factor corresponding to a pole follow the
same rules as the plot for the zero, except that everything is negated. Specifi-

c Shreyas Sundaram
5.1 Rules for Drawing Bode Plots 31

cally, for a factor of the form (j ωp + 1)−1 , we have

r r
ω ω ω
20 log |(j + 1) | = 20 log( 1 + ( )2 )−1 = −20 log( 1 + ( )2 ) ,
−1
p z p
ω −1 −1 ω
∠(j + 1) = − tan .
p p

The Bode plot for the factor (j ωp + 1)−1 looks like this:

s+10
Example. Draw the Bode plot for H(s) = 10 (s+1)(s+100) .
Solution.

c Shreyas Sundaram
32 Bode Plots

 ±1
5.1.4 Bode Plot for ( ωsn )2 + 2ζ( ωsn ) +1

We have already seen what the magnitude and phase plots look like for a second
order system of this form. To derive general rules for drawing this, note that
the magnitude of this function at s = jω is given by
  r
 jω 2 ω  ω ω
20 log ( ) + 2ζ( )j + 1 = 20 log (1 − ( )2 )2 + 4ζ 2 ( )2 .
 
ωn ωn ωn ωn
 
For ω  ωn , we have ωωn ≈ 0, and so 20 log ( ωjωn )2 + 2ζ( ωωn )j + 1 ≈ 0. For
 
ω  ωn , we have
s 
r 4  2
ω 2 2 2
ω 2
ω ω
(1 − ( ) ) + 4ζ ( ) ≈ = ,
ωn ωn ωn ωn
 
and so 20 log ( ωjωn )2 + 2ζ( ωωn )j + 1 ≈ 40 log ω − 40 log ωn . This is a line of slope
 
40 passing
 through the point  0 when ω = ωn . The general magnitude curve for
 jω 2 ω
20 log ( ωn ) + 2ζ( ωn )j + 1 thus looks like:


 
The phase of ( ωjωn )2 + 2ζ( ωωn )j + 1 is given by
 
2ζ ωωn
 
jω 2 ω
∠ ( ) + 2ζ( )j + 1 = tan−1  2  .
 
ωn ωn

ω
1− ωn

For ω  ωn , the argument of the arctan function is almost 0, and so the phase
curve starts at 0 for small ω. For ω = ωn , the argument is ∞, and so the phase
curve passes through π2 when ω = ωn . For ω  ωn , the argument of the arctan
function approaches 0 from the negative side, and so the phase curve approaches
π for large values of ω. Just as in the first order case, we will take the phase
curve transitions to occur one decade before and after ωn . This produces a

c Shreyas Sundaram
5.1 Rules for Drawing Bode Plots 33

phase curve of the form:

 −1
The Bode plot for the factor ( ωjωn )2 + 2ζ( ωωn )j + 1 looks just like the Bode
 
jω 2
plot for the factor ( ωn ) + 2ζ( ωωn )j + 1 , except that everything is flipped:

(s+1)(s2 +3s+100)
Example. Draw the Bode Plot of H(s) = s2 (s+10)(s+100)

c Shreyas Sundaram
34 Bode Plots

Solution.

5.1.5 Nonminimum Phase Systems

So far we have been looking at the case where all zeros and poles are in the
CLHP. Bode plots can also be drawn for systems that have zeros or poles in
the RHP – however, note that for systems that have RHP poles, the steady
state response to a sinusoidal input will not be a sinusoid (there won’t even be a
steady state response, as the output will blow up). This does not change the fact
that the transfer function will have a magnitude and phase at every frequency
(since the transfer function is simply a complex number at every frequency ω).
Transfer functions with zeros in the right half plane are called nonminimum
phase systems, and those with all zeros and poles in the CLHP are called
minimum phase systems.

To gain intuition about how the Bode plot of a nonminimum phase system
compares to that of a minimum phase system, let us see how the Bode plots of
the terms H1 (s) = s + 1 and H2 (s) = s − 1 compare. First, note that

p
|H1 (jω)| = |jω + 1| = 1 + ω 2 = |H2 (jω)|,

and thus the magnitude plots of the two terms are identical. To compare the
phase contribution, it is useful to examine the complex number representation

c Shreyas Sundaram
5.1 Rules for Drawing Bode Plots 35

of H1 (jω) and H2 (jω):

From this, we see that ∠H2 (jω) = ∠(jω − 1) = π − ∠H1 (jω). Thus, the phase
plots of the two terms look like this:

An alternative method to draw the phase plot of H2 (s) = s−1 is to first convert
it to Bode form to obtain H2 (s) = −1(−s + 1), where we now have a gain of −1
in front. This gain contributes nothing to the log-magnitude, but it contributes
a phase of π. The phase of −s+1 is the negative of the phase of s+1 (since they
are complex conjugates), and once again, we obtain that the phase of H2 (jω)
is π − ∠H1 (jω).
Example. Draw the Bode plots for the systems
s+1 s−1
H1 (s) = 10 , H2 (s) = 10 .
s + 10 s + 10

Solution.

c Shreyas Sundaram
36 Bode Plots

As we can see from the above example, the magnitudes of the two transfer
functions do not depend on whether the zero is in RHP or the LHP. However,
the phase plots are quite different. Based on the above analysis, we see that the
phase contribution of a zero in the right half plane is always at least as large (in
absolute terms) as the phase contribution of a zero in the left half plane – this
is the reason for calling systems with such zeros (or poles) nonminimum phase.
Note that for minimum phase systems, the magnitude plot uniquely determines
the transfer function, but for nonminimum phase systems, we need both the
magnitude plot and the phase plot in order to determine the transfer function.

c Shreyas Sundaram
Chapter 6

Modeling and Block

Diagram Manipulation

With the mathematical foundations from the previous chapters in hand, we are
now ready to move on to the modeling of control systems.

6.1 Mathematical Models of Physical Systems

The first task of control system design is to obtain an appropriate mathematical
model of the plant. In many applications, this can be difficult to do, as ex-
perimental data is often noisy, and real-world systems are often quite complex.
Thus, we must frequently come up with an approximate model, maintaining a
tradeoff between complexity and how accurately it captures the actual physical
plant. In general, we try to follow Einstein’s rule: “Make things as simple as
possible, but no simpler.”
If the system is known (or approximated) to be linear, then one way to identify
a model is to apply sinusoidal inputs of appropriate frequencies to the system,
and then try and fit an appropriate transfer function to the resulting Bode plot.
An alternative (and complementary) method is to use physical laws to obtain
the model. We will now study some examples of this approach.

6.1.1 Mechanical Systems

The key equation governing the model of many mechanical systems is Newton’s
Law: F = ma. In this equation, F represents the vector sum of all the forces
acting on a body, m represents the mass of the body, and a represents the
acceleration of the body. The forces acting on the body can be generated by an

c Shreyas Sundaram
38 Modeling and Block Diagram Manipulation

outside entity (as an input to the system), or by springs and dampers attached
to the body.
Example: Mass-Spring-Damper System.

6.1.2 Electrical Systems

Standard components in electrical systems include resistors, capacitors and
inductors, connected together in various ways. The quantities of interest in
electrical systems are voltages and currents.

Figure 6.1: (a) Components of Electrical Systems. (b) An Electrical System.

The main modeling technique for electrical systems is to use Kirchoff ’s Laws.

c Shreyas Sundaram
6.1 Mathematical Models of Physical Systems 39

• Kirchoff ’s Voltage Law (KVL): The algebraic sum of the voltages

around a closed loop is zero.

• Kirchoff ’s Current Law (KCL): The algebraic sum of the currents

coming into a node is zero.

We will see an example of an electric circuit model in the context of a DC motor

below.

6.1.3 Rotational Systems

When the system involves rotation about a point, the system dynamics are
governed by a modified form of Newton’s Law: τ = Jθ̈. Here, τ is the sum
of all external torques about the center of mass, J is the moment of inertia of
the body, and θ is the angular position of the body (so that θ̈ is the angular
acceleration).
Example: A DC motor consists of an electrical component and a rotational
component. The input voltage to the electrical component induces a current,
which then provides a torque to the motor shaft via a magnetic field. This
torque causes the shaft to rotate. In turn, this torque also induces a voltage
drop (called the back emf) in the electrical circuit. Derive the overall system
equations. and find the transfer function from the input voltage to the angular
velocity (ω = θ̇) of the DC motor.

c Shreyas Sundaram
40 Modeling and Block Diagram Manipulation

c Shreyas Sundaram
6.2 Block Diagram Manipulation 41

All of the above examples considered above yielded models that involved linear
differential equations of the form

y (n) + an−1 y (n−1) + · · · + a1 ẏ + a0 y = bm u(m) + bm−1 u(m−1) + · · · + b1 u̇ + b0 u .

As we saw in Section 4.2, the transfer function of such systems is given by

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

H(s) = = .
U (s) sn + an−1 sn−1 + · · · + a1 s + a0

In practice, many systems are actually nonlinear, and there is a whole set of
tools devoted to controlling such systems. One technique is to linearize the
system around an operating point, where the nonlinearities are approximated
by linear functions. In the rest of the course, we restrict our attention to lin-
ear systems, and assume that these linearization techniques have been applied
to any nonlinearities in the system. We will now study how to manipulate
interconnections of linear systems.

6.2 Block Diagram Manipulation

Control systems typically involve several smaller systems (or components) that
are interconnected together in various ways – the output of one system will be
the input to other systems. For example, remember from the first lecture that
the basic block diagram of a feedback control system looks like this:

We will frequently want to manipulate block diagram representations of systems

in order to find the overall transfer function of the system in terms of the transfer
functions of the individual components or subsystems. There are three types of
interconnections that we will be studying.
Series Connection. In this case, the output of one system feeds directly into
the input of another system.

Figure 6.2: Two Systems Connected in Series.

c Shreyas Sundaram
42 Modeling and Block Diagram Manipulation

The overall transfer function from U (s) to Y (s) can be obtained as follows:

Y (s) = H2 (s)Y1 (s)
⇒ Y (s) = H2 (s)H1 (s)U (s) ,
Y1 (s) = H1 (s)U (s)
Y (s)
and thus the overall transfer function is H(s) = U (s) = H2 (s)H1 (s).

Parallel Connection. In this case, two (or more systems) obtain the same
input U (s), and their outputs are summed together to produce the output of
the overall system.

Figure 6.3: Two Systems Connected in Parallel.

The overall transfer function from U (s) to Y (s) can be obtained as follows:

Y1 (s) = H1 (s)U (s) 
Y2 (s) = H2 (s)U (s) ⇒ Y (s) = (H1 (s) + H2 (s)) U (s) ,
Y (s) = Y1 (s) + Y2 (s)


Y (s)
and thus the overall transfer function is H(s) = U (s) = H1 (s) + H2 (s).

Feedback Connection. In this case, the output of one system feeds into the
input of a second system, and the output of this second system feeds back into
the input of the first system (perhaps in conjunction with another signal).
The overall transfer function from R(s) to Y (s) can be obtained by noting that

Y (s) = H1 (s)E(s)
Y2 (s) = H2 (s)Y (s)
E(s) = R(s) − Y2 (s),

which yields

Y (s) = H1 (s)(R(s) − Y2 (s)) = H1 (s)R(s) − H1 (s)H2 (s)Y (s).

c Shreyas Sundaram
6.2 Block Diagram Manipulation 43

Figure 6.4: Feedback Connection of Two Systems.

Thus the overall transfer function is

Y (s) H1 (s)
H(s) = = .
R(s) 1 + H1 (s)H2 (s)

This is an important transfer function; the numerator is called the “forward

gain”, and the denominator is described as “1 + the loop gain”.
Note that in the above feedback configuration, the output of the system H2 (s)
is subtracted from the reference signal R(s) to produce the error E(s). This con-
figuration is thus called a negative feedback configuration. If the signal Y2 (s)
is instead added to the reference signal, the configuration is called a positive
feedback configuration. In this case, the transfer function would be given by

Y (s) H1 (s)
H(s) = = .
R(s) 1 − H1 (s)H2 (s)

Note: The basic feedback control system shown at beginning of this section is
a special case of the negative feedback configuration shown above, with H1 (s) =
P (s)C(s), and H2 (s) = 1. The basic feedback control system is thus said to be
in “unity feedback” configuration.
Based on the above configurations, we can derive the following rules to modify
block diagrams.

c Shreyas Sundaram
44 Modeling and Block Diagram Manipulation

Example. Compute the transfer function from U (s) to Y (s).

c Shreyas Sundaram
6.2 Block Diagram Manipulation 45

6.2.1 Systems with Multiple Inputs and Outputs

We will frequently encounter interconnected systems that have multiple inputs

(e.g., control signals, disturbances, etc.) and outputs (e.g., different sensor
measurements, etc.). For linear systems, we can simply use the Principle of
Superposition to find the transfer function from the inputs to the outputs. More
specifically, if we have a set of inputs U1 (s), U2 (s), . . . , Um (s) and a set of outputs
Y1 (s), Y2 (s), . . . , Yp (s), then we find the transfer function Hij (s) from input
Ui (s) to output Yj (s) by simply removing all of the other inputs and outputs
from the system and calculating the resulting transfer function as above. By
linearity, each output can then be written as

Yi (s) = H1i (s)U1 (s) + H2i (s)U2 (s) + · · · + Hmi (s)Um (s), i ∈ {1, 2, . . . , p}.

c Shreyas Sundaram
46 Modeling and Block Diagram Manipulation

Example. Write the output Y (s) of the following system in terms of the inputs
R(s) and D(s).

D(s)
R(s) + Y (s)
+
C(s) P (s)
−

c Shreyas Sundaram
Chapter 7

Step Responses of Linear

Systems

Consider a system with transfer function of the form

bm sm + bm−1 sm−1 + · · · + b1 s + b0
H(s) = .
sn + an−1 sn−1 + · · · + a1 s + a0
The degree (n) of the denominator polynomial is called the order of the system.
As we have already seen, systems of order one and two arise frequently in prac-
tice; even when the order of a system is higher than two, one can sometimes
approximate the system by a first or second order system in order to obtain
intuition about the system’s behavior. It is therefore important for us to study
the responses of first and second order systems to various types of inputs (and
in particular, to step inputs). Recall that for step inputs u(t) = 1(t), we have
U (s) = 1s .
Note that if all poles of H(s) are in the OLHP, the Final Value Theorem states
that the steady-state value of the step response is given by
1
lim sY (s) = lim sH(s)U (s) = lim sH(s) = lim H(s) = H(0) .
s→0 s→0 s→0 s s→0
The quantity H(0) is called the DC gain of the system. This is true regardless
of the order of the system.

7.1 Step Response of First Order Systems

Consider a first order system with transfer function
b0
H(s) = , a0 6= 0.
s + a0

c Shreyas Sundaram
48 Step Responses of Linear Systems

This system has a single pole at s = −a0 . The step response of this system is
obtained by calculating

b0 1
Y (s) = H(s)U (s) =
s + a0 s
 
b0 1 1
= − ,
a0 s s + a0

which produces
b0 b0
y(t) = 1(t) − e−a0 t , t ≥ 0. (7.1)
a0 a0
Consider two cases:

• If a0 > 0, then the pole of the transfer function is in the OLHP, and
e−a0 t → 0 as t → ∞. The step response thus reaches a steady-state value
of ab00 (which is the DC gain of the system). The response is said to be
stable.

• If a0 < 0, the pole of the transfer function is in the ORHP, and e−a0 t → ∞
as t → ∞. The step response therefore goes to ∞, and there is no steady
state value. The response is said to be unstable.

Figure 7.1: Step Response of First Order System. (a) Pole in OLHP. (b) Pole
in ORHP.

We will now study two measures of performance that can be used to evaluate
the step-response.

7.1.1 Rise time

The rise time tr of a step response is defined as the amount of time required for
the response to go from 10% of its final value to 90% of its final value.

c Shreyas Sundaram
7.1 Step Response of First Order Systems 49

b0
The rise time of a first order system with transfer function H(s) = s+a0 is

ln 9
tr =
a0

(you should be able to easily prove this). Note that the larger a0 is, the smaller
the rise time becomes. This can also be seen from the actual step-response
(7.1): a larger value of a0 means that the term e−a0 t dies out faster, leading the
response to get to its steady state quicker.

7.1.2 Settling time

Another measure of performance is the settling time of the response, defined as

the time required for the response to get to within a certain percentage of its
final value and stay there. We will take this percentage to be 2%, but sometimes
the settling time is defined in terms of 1% or 5% as well. By equating (7.1) to
0.98 ab00 and solving for t, we find that the 2% settling time for the first order
step response is
3.91
ts ≈ .
a0

1
Note: The quantity τ = a0 is called the time-constant of the system, and (7.1)
is commonly written as

b0 b0 t
y(t) = 1(t) − e− τ .
a0 a0

A larger time-constant means that the system takes longer to settle to its steady
state. To find the time-constant in practice, note that at t = τ we have

b0 b0
y(τ ) = (1 − e−1 ) ≈ 0.63 .
a0 a0

Since the steady state value of y(t) is ab00 (the DC gain of the system), we see that
the time-constant is the point in time where the output of the system reaches
63% of its final value.
One can also readily verify that the bandwidth of the system H(s) occurs at
s = a0 (i.e., |H(ja0 )| = √12 |H(0)|). Thus, we get the following rule of thumb.

Fast response ⇔ Small τ ⇔ Large a0 ⇔ Large bandwidth

c Shreyas Sundaram
50 Step Responses of Linear Systems

7.2 Step Response of Second Order Systems

Consider a second order system of the form
b1 s + b0
H(s) = .
s2 + a1 s + a0
In order to obtain intuition about these systems, we will be focusing on a par-
ticular form of second order system:
ωn2
H(s) = , ωn > 0 .
s2 + 2ζωn s + ωn2
The poles of this
p transfer function are obtained from the quadratic formula as
s = −ζωn ± ωn ζ 2 − 1. The location of these poles in the complex plane will
vary based on the value of ζ. We analyze three different cases:

• 0 ≤ ζ < 1: The system has two complex poles in the CLHP (they will be
in the OLHP if ζ > 0). The system is said to be underdamped.
• ζ = 1: The system has two repeated poles at s = −ωn . The system is said
to be critically damped.
• ζ > 1: The system has two poles on the negative real axis. The system is
said to be overdamped.

Figure 7.2: Location of Poles in Complex Plane for Different Ranges of ζ. (a)
0 ≤ ζ < 1. (b) ζ = 1. (c) ζ > 1.

7.2.1 Underdamped and Critically Damped Systems (0 ≤

ζ ≤ 1)
Theppoles of the transfer function in this case are given by s = −ζωn ±
jωn 1 − ζ 2 . To simplify the notation, define
p
σ = ζωn , ωd = ωn 1 − ζ 2 ,

c Shreyas Sundaram
7.2 Step Response of Second Order Systems 51

in which case the poles are at s = −σ ± jωd . The transfer function can be
written as
ωn2 ωn2
H(s) = = .
(s + σ + jωd )(s + σ − jωd ) (s + σ)2 + ωd2
The Laplace Transform of the step response is given by Y (s) = H(s) 1s , and
using a Laplace Transform table, we see that
 
−σt σ
y(t) = 1 − e cos ωd t + sin ωd t . (7.2)
ωd
• When ζ = 0, we have σ = 0 and ωd = ωn , and the response becomes
y(t) = 1 − cos ωn t, which oscillates between 0 and 2 for all t.
• When ζ = 1, we have σ = ωn and ωd = 0, and the response becomes
y(t) = 1 − e−ωn t (1 + ωn t); one can verify that this does not have any
oscillations at all, and asymptotically reaches the steady state value of 1
(this is the DC gain H(0)).

The response for intermediate values of ζ falls in between these two extremes.
The behavior of y(t) for different values of ζ (with a fixed value of ωn ) is shown
in Fig. 7.3.

Figure 7.3: Step response as a function of ζ, for 0 ≤ ζ ≤ 1.

As we can see from the responses for various values of ζ, a larger value of ζ
corresponds to fewer oscillations and less overshoot, and thus ζ is said to rep-
resent the amount of damping in the response (a higher value of ζ corresponds

c Shreyas Sundaram
52 Step Responses of Linear Systems

to more damping). ζ is called the damping ratio. Examining Fig. 7.2, we see
that the angle between the imaginary axis and the vector connecting the pole
to the origin is given by sin−1 ζ. The magnitude of the vector is given by ωn .
Thus, as ζ increases with a fixed ωn , the poles move on the perimeter of a circle
toward the real axis.
The quantity ωn is called the undamped natural frequency, since it repre-
sents the location of the pole on the imaginary axis when there is no damping
(i.e., when ζ = 0). The quantity ωd is called the damped natural frequency,
since it represents the imaginary part of the pole when there is damping. When
ζ = 0, we have ωd = ωn . If ζ > 0 and we increase ωn , the poles move further to
the left in the OLHP. This causes the term σ = ζωn to increase, which causes
the term e−σt to die out faster in the step response (equation (7.2)). Thus,
increasing ωn has the effect of making the oscillations die out faster (and, in
general, making the system respond faster). Recall from the magnitude plot
of the second order frequency response from Section 4.3 that the bandwidth of
the system is approximately equal to ωn . Thus, we obtain the following rule of
thumb for the underdamped system.

Larger ωn ⇔ Larger bandwidth ⇔ Faster response

We will see this more formally in the next chapter.

7.2.2 Overdamped System (ζ > 1)

Thep poles of the transfer function in this case are given by s = −ζωn ±
ωn ζ 2 − 1, which are both on the negative real axis (as ζ → ∞, one pole
approaches s = 0, and the other pole approaches s = −∞). The transfer func-
tion can be written as
ωn2
H(s) = p p .
(s + ζωn + ωn ζ 2 − 1)(s + ζωn − ωn ζ 2 − 1)
The Laplace Transform of the step response is given by Y (s) = H(s) 1s , and
using partial fraction expansion, we see that the step response will have the
form √ 2 √ 2
y(t) = 1 − k1 e(−ζωn −ωn ζ −1)t − k2 e(−ζωn +ωn ζ −1)t ,
for some constants k1 and k2 . This response has no oscillations; a sample
response is shown in Fig. 7.4.

7.2.3 Discussion
Based on the above analysis, we can come to the following general conclusions
about how the poles of second order systems affect their step responses.

c Shreyas Sundaram
7.2 Step Response of Second Order Systems 53

Figure 7.4: Step response of an overdamped system (ζ > 1).

• If the poles are complex, the step response will have oscillations and over-
shoot.
• As the poles move toward the real axis while maintaining a fixed distance
from the origin, the amount of oscillation decreases (leading to less over-
shoot) – this corresponds to the damping ratio ζ increasing (with a fixed
ωn ),
• If ωn increases, the poles move further left in the OLHP, and the oscilla-
tions die out faster.
• If all poles are on the negative real axis, there will be no oscillation and
no overshoot.

Note that if the transfer function has one or more poles in the open right half
plane, the step response will contain a term that goes to ∞ as t → ∞, and thus
there is no steady state response.
Based on these general rules, the step responses for various pole locations are
shown below:

c Shreyas Sundaram
54 Step Responses of Linear Systems

c Shreyas Sundaram
Chapter 8

Performance of Second
Order Step Responses

We have seen that the step response of a second order system with transfer
function
ωn2
H(s) = 2
s + 2ζωn s + ωn2
will have different characteristics, depending on the location of the poles (which
are a function of the values ζ and ωn ). We will now introduce some measures
of performance for these step responses.

8.1 Performance Measures

We will be studying the following metrics in order to evaluate the performance
of second order step responses.

• Rise time (tr ): The time taken for the response to first get close to its
final value. This is typically measured as the time taken for the response
to go from 10% of its final value to 90% of its final value.

• Settling time (ts ): The time taken for the response to stay close to its
final value. We will take this to be the time after which the response stays
within 2% of its final value. Other measures of “closeness” can also be
used (e.g., 1% instead of 2%, etc.).

• Peak value (Mp ) and overshoot OS: This is largest value of the step
response. One can also calculate the overshoot as the maximum amount
that the response overshoots its final value, divided by the final value
(often expressed as a percentage).

c Shreyas Sundaram
56 Performance of Second Order Step Responses

• Peak time (tp ): This is the time at which the response hits its maximum
value (this is the peak of the overshoot).

Figure 8.1: Step response of second order system, with tr , Mp , tp and ts shown.

8.1.1 Rise time (tr )

An explicit formula for the rise time is somewhat hard to calculate. However,
we notice that rise time increases with ζ and decreases with ωn . The best linear
fit to the curve gives us the approximation

2.16ζ + 0.6
tr ≈ ,
ωn

which is reasonably accurate for 0.3 < ζ < 0.8. A cruder approximation is
obtained by finding a best fit curve with ζ = 0.5, yielding

1.8
tr ≈ .
ωn

Keep in mind that these are only approximations, and iteration may be required
if design specifications are not met.

8.1.2 Peak value (Mp ), Peak time (tp ) and Overshoot OS

Recall that the step response will have overshoot only if the poles are complex
– this corresponds to the case 0 ≤ ζ < 1, and the corresponding system is called
underdamped. The overshoot is defined as

Mp − y(∞)
OS = ,
y(∞)

c Shreyas Sundaram
8.1 Performance Measures 57

where Mp is the peak value of the step response, and y(∞) is the final (steady
state) value of the step response. Note that if we want to express the overshoot
as a percentage, we simply multiply OS by 100, and denote this by %OS.
We can explicitly calculate the peak value of the step response (and the corre-
sponding time) as follows. First, note that at the peak value, the derivative of
the step response is zero. Recalling that the step response is given by
 
σ
y(t) = 1 − e−σt cos ωd t + sin ωd t ,
ωd
dy
calculate dt and set it equal to zero to obtain
dy σ
0= = σe−σt (cos ωd t + sin ωd t) − e−σt (−ωd sin ωd t + σ cos ωd t)
dt ωd
σ 2 −σt
= e sin ωd t + ωd e−σt sin ωd t
ωd
σ2
=( + ωd )e−σt sin ωd t .
ωd
From this, we obtain sin ωd t = 0, which occurs for t = 0, ωπd , ω2πd , · · · . The first
peak therefore occurs at the peak time
π
tp = .
ωd
Substituting this into the expression for y(t), we obtain the value of the step
response at the peak to be
π −σ π σ
Mp = y( ) = 1 − e ωd (cos π + sin π)
ωd ωd
−σ ωπ
=1+e d .
p
Substituting y(∞) = 1, σ = ζωn , ωd = ωn 1 − ζ 2 , and the above expression
for Mp into the definition of overshoot, we obtain
−σ ωπ
1+e d −1 −ζωn √π
1−ζ 2
OS = =e ωn
1
− √ πζ
=e 1−ζ 2 .

8.1.3 Settling Time (ts )

The settling time ts is the time taken for the step response to stay within a
certain percentage of its final value. This percentage will vary depending on the
application. In this course, we will take 2% as a good measure. To calculate ts ,
we note once again that the step response is
 
σ
y(t) = 1 − e−σt cos ωd t + sin ωd t .
ωd

c Shreyas Sundaram
58 Performance of Second Order Step Responses

We can obtain a rough approximation of ts by calculating the time at which the

exponential term e−σt becomes equal to 0.02 (since the cosine and sine terms
oscillate forever between two fixed values):
4 4
e−σts ≈ 0.02 ⇒ −σts ≈ ln 0.02 ⇒ ts ≈ = .
σ ζωn

Summary. The expressions for the various performance metrics are:

2.16ζ + 0.6 1.8
tr ≈ or tr ≈
ωn ωn
− √ πζ
Mp = 1 + e 1−ζ 2 (for systems with DC gain of 1)
πζ
−√
OS = e 1−ζ 2

π π
tp = p =
ωn 1 − ζ 2 ωd
4
ts ≈ .
ζωn

Note that the expressions for tr and ts are only approximations; more detailed
expressions for these quantities can be obtained by performing a more careful
analysis. For our purposes, the above relationships will be sufficient to obtain
intuition about the performance of second order systems. In particular, we note
the following trends:

• As ωn (bandwidth) increases (with ζ fixed), tr de-

creases, OS stays the same, ts decreases and tp de-
creases.
• As ζ increases (with ωn fixed), tr stays approximately
constant, OS decreases, ts decreases and tp increases.

We would typically like a small OS, small tr and small ts .

8.2 Choosing Pole Locations to Meet Perfor-

mance Specifications
As we will see later in the course, we can design a controller for a given system so
that the overall closed loop transfer function has poles at any desired locations
in the complex plane. We will need to choose these pole locations so that the
step response of the system achieves certain performance specifications. For

c Shreyas Sundaram
8.2 Choosing Pole Locations to Meet Performance Specifications 59

example, suppose that we want the step response to have overshoot less than
some value OS, ¯ rise time less than some value t̄r , settling time less than some
value t̄s , and peak time less than some t̄p . From the expressions given at the end
of the previous section, these constraints on the step response can be translated
into constraints on ζ and ωn :

− √ πζ
e 1−ζ 2 ¯
≤ OS ⇒ Find upper bound for ζ from this.
1.8 1.8
≤ t̄r ⇒ ωn ≥
ωn t̄r
4 4
≤ t̄s ⇒σ≥
σ t̄s
π π
≤ t̄p ⇒ ωd ≥ .
ωd t̄p

Noting that σ is the real part of the poles of the transfer function, ωn is the
distance of the complex poles from the origin, ωd is imaginary part of the poles,
and sin−1 ζ is the angle between the imaginary axis and the vector joining the
origin to the pole, we can formulate appropriate regions in the complex plane
for the pole locations in order to satisfy the given specifications.

Figure 8.2: Pole Locations in the Complex Plane. (a) Overshoot. (b) Rise time.
(c) Settling time. (d) Peak time. (e) Combined specifications.

c Shreyas Sundaram
60 Performance of Second Order Step Responses

Example. We would like our second order system to have a step response with
overshoot OS ≤ 0.1 and settling time ts ≤ 2. What is the region in the complex
plane where the poles can be located?
Solution.

8.3 Effects of Poles and Zeros on the Step Re-

sponse

So far, we have been studying second order systems with transfer functions of
the form
ωn2
H(s) = .
s2 + 2ζωn s + ωn2

This transfer function produced a step response y(t), with Laplace Transform
Y (s) = H(s) 1s .
Note: All of the analysis so far also holds if we consider the transfer function:

ωn2
H(s) = K ,
s2 + 2ζωn s + ωn2

for some constant K. In this case, the step response is simply scaled by K,
but the time characteristics (such as rise time, settling time, peak time and
overshoot) of the response are not affected.
We will now consider what happens when we add zeros or additional poles to
the system.

c Shreyas Sundaram
8.3 Effects of Poles and Zeros on the Step Response 61

8.3.1 Effect of a Zero on the Step Response

Suppose that we modify the second order transfer function given above by
adding a zero at s = −z, for some z. The modified transfer function is given by

( z1 s + 1)ωn2
Hz (s) = .
s2 + 2ζωn s + ωn2

Note that the reason for writing the zero term as z1 s + 1 instead of s + z is to
maintain a DC gain of 1 for the transfer function (just so that we can compare
it to the original transfer function). We can split the above transfer function
into the sum of two terms:
ωn2 1 ωn2 1
Hz (s) = 2 2
+ s 2 2
= H(s) + sH(s) ,
s + 2ζωn s + ωn z s + 2ζωn s + ωn z

where H(s) is the transfer function of the original system (without the zero).
Denote the Laplace Transform of the step response for this system by Yz (s) =
Hz (s) 1s . Using the above decomposition of Hz (s), we obtain
 
1 1 1 1 1 1
Yz (s) = H(s) + sH(s) = H(s) + sH(s) = Y (s) + sY (s) .
z s s z s z

Noting that the inverse Laplace Transform of sY (s) is ẏ, we obtain

1
yz (t) = y(t) + ẏ(t) .
z
Thus the step response of the second order system with a zero at s = −z is
given by the step response of the original system plus a scaled version of the
derivative of the step response of the original system. A sample plot for z > 0
(i.e., corresponding to the zero being in the OLHP) is shown in Fig. 8.3.

Figure 8.3: Step response of second order system with transfer function Hz (s) =
2
( z1 s+1)ωn
s2 +2ζωn s+ωn 2 , z > 0.

c Shreyas Sundaram
62 Performance of Second Order Step Responses

Note that as z increases (i.e., as the zero moves further into the left half plane),
the term z1 becomes smaller, and thus the contribution of the term ẏ(t) decreases
(i.e., the step response of this system starts to resemble the step response of the
original system). From the above figure, the effect of a LHP zero is to increase
the overshoot, decrease the peak time, and decrease the rise time; the settling
time is not affected too much. In other words, a LHP zero makes the step
response faster. One can also see this by thinking about the effect of the zero
on the bandwidth of the system. Since the presence of the term ( z1 s + 1) in
the numerator of the transfer function will only increase the magnitude of the
Bode plot at frequencies above ω = z, we see that adding a zero will generally
increase the bandwidth of the system. This fits with our rule of thumb that a
larger bandwidth corresponds to a faster response.
Now consider what happens if z is negative (which corresponds to the zero being
in the ORHP). In this case, the derivative ẏ(t) is actually subtracted from y(t)
to produce yz (t). A sample plot is shown in Fig. 8.4. Note that the response
can actually go in the opposite direction before rising to its steady state value.
This phenomenon is called undershoot.

Figure 8.4: Step response of second order system with transfer function Hz (s) =
2
( z1 s+1)ωn
s2 +2ζωn s+ωn 2 , z < 0.

Recall from Section 5.1.5 that zeros in the right half plane are called nonmini-
mum phase – this is due to the fact the phase of the system has a large swing
from its maximum and minimum values (as compared to the phase plot of the
system with the same magnitude plot, but with all zeros in the OLHP). The
effect of a RHP zero is to slow down the system, and perhaps introduce under-
shoot. However, the magnitude plot of the system is affected in the same way
as with a LHP zero: the bandwidth increases. Thus, for a nonminimum phase

c Shreyas Sundaram
8.3 Effects of Poles and Zeros on the Step Response 63

zero, our rule of thumb about a larger bandwidth implying a faster response no
longer holds.
It is also worth noting that if z = 0, the above analysis no longer holds directly.
Adding a zero at s = 0 produces the transfer function Hz (s) = sH(s), and the
step response of this system is purely the derivative of the step response of the
original system. However, the steady state value of this step response is zero,
not 1 (note that this agrees with the DC gain of the new system).

8.3.2 Effect of Poles on the Step Response

A general n–th order system will have n poles in the complex plane. As we have
already discussed, if any of these poles are in the RHP, the step response will
be unstable (i.e., it will go to infinity). Consider the case of all poles being in
the OLHP; a sample distribution is shown in Fig. 8.5.

Figure 8.5: Sample distribution of poles in OLHP, with dominant poles encir-
cled.

From the partial fraction expansion, we know that a pole −p contributes a term
of the form e−pt to the step response. If p is large, this exponential term dies out
quickly. Suppose that the set of poles for a system can be divided into a cluster
of poles that are closest to the origin, and another cluster that are very far away
in comparison (e.g., 5 times further away). The poles that are closest to the
origin are called the dominant poles of the system. The exponential terms in
the step response corresponding to the far away poles will die out very quickly
in relation to the exponential terms corresponding to the dominant poles. Thus,
the system effectively behaves as a lower order system with only the dominant
poles. This is one way to approximate a high order system by a lower order
system (such as a first or second order system).
Since each additional pole contributes an additional exponential term that must
die out before the system reaches its final value, each additional pole increases
the rise time of the system. In other words, adding a pole to the system makes

c Shreyas Sundaram
64 Performance of Second Order Step Responses

the step response more sluggish. Again, one can see this by looking at the
bandwidth of the system: including an additional term of the form ( p1 s + 1) in
the denominator of the transfer function causes the Bode plot to decrease faster
after ω = p, which generally decreases the bandwidth of the system.
Example. Plot the step response of the system with transfer function
4
H(s) = ,
( p1 s + 1)(s2 + 2s + 4)

for p = 0.1, 2, 10, 100.

Solution.

Summary. Based on our discussions, we can make the following observations:

• Adding a LHP zero to the transfer function makes the step response faster
(decreases the rise time and the peak time) and increases the overshoot.
The bandwidth of the system is increased.

• Adding a RHP zero to the transfer function makes the step response
slower, and can make the response undershoot.1 The bandwidth of the
system is increased.
• Adding a LHP pole to the transfer function makes the step response slower.
The bandwidth of the system is decreased.
• If the system has a cluster of poles and zeros that are much closer (5 times
or more) to the origin than the other poles and zeros, the system can be
approximated by a lower order system with only those dominant poles and
zeros.

1 Actually, one can show that the step response for a given (stable) linear time-invariant

system will have undershoot if and only if the transfer function has an odd number of zeros in
the ORHP. See the paper “On Undershoot and Nonminimum Phase Zeros” by M. Vidyasagar
(IEEE Transactions on Automatic Control, vol. 31, no. 5, May 1986, p. 440) for the
derivation of this result.

c Shreyas Sundaram
Chapter 9

Stability of Linear
Time-Invariant Systems

Recall from our discussion on step responses that if the transfer function contains
poles in the open right half plane, the response will go to infinity. However, if
all poles of the transfer function are in the open left half plane, the response
will settle down to the DC gain of the transfer function. To describe these
characteristics of systems, we define the following terms:

• A signal f (t) is bounded if there exists some constant M such that

|f (t)| ≤ M for all t. It is called unbounded if it is not bounded.
• A system is bounded-input bounded-output (BIBO) stable if every
bounded input leads to a bounded output. We will also refer to such
systems simply as stable.
• A system is unstable if there is at least one bounded input that produces
an unbounded output.

A linear time-invariant system is stable if all the poles of the transfer

function are in the OLHP, and it is unstable if at least one pole is in
the CRHP.

9.1 Pole-zero cancellations and stability

Consider the linear time-invariant system given by the transfer function
bm sm + bm−1 sm−1 + · · · + b1 s + b0 N (s)
H(s) = = .
sn + an−1 sn−1 + · · · + a1 s + a0 D(s)

c Shreyas Sundaram
66 Stability of Linear Time-Invariant Systems

Recall that this system is stable if all of the poles are in the OLHP, and these
poles are the roots of the polynomial D(s). It is important to note that one
should not cancel any common poles and zeros of the transfer func-
tion before checking the roots of D(s). Specifically, suppose that both of the
polynomials N (s) and D(s) have a root at s = a, for some complex (or real)
number a. One must not cancel out this common zero and pole in the transfer
function before testing for stability. The reason for this is that, even though the
pole will not show up in the response to the input, it will still appear as a re-
sult of any initial conditions in the system, or due to additional inputs entering
the system (such as disturbances). If the pole and zero are in the CRHP, the
system response might blow up due to these initial conditions or disturbances,
even though the input to the system is bounded, and this would violate BIBO
stability.
To see this a little more clearly, consider the following example. Suppose the
transfer function of a linear system is given by
s−1 N (s)
H(s) = = .
s2 + 2s − 3 D(s)
Noting that s2 + 2s − 3 = (s + 3)(s − 1), suppose we decided to cancel out the
common pole and zero at s = 1 to obtain
1
H(s) = .
s+3
Based on this transfer function, we might (erroneously) conclude that the system
is stable, since it only has a pole in the OLHP. What we should actually do is
look at the original denominator D(s), and correctly conclude that the system
is unstable because one of the poles is in the CRHP. To see why the pole-
zero cancellation hides instability of the system, first write out the differential
equation corresponding to the transfer function to obtain
ÿ + 2ẏ − 3y = u̇ − u .
Take the Laplace Transform of both sides, taking initial conditions into account:
s2 Y (s) − sy(0) − ẏ(0) + 2sY (s) − 2y(0) − 3Y (s) = sU (s) − u(0) − U (s) .
Rearrange this equation to obtain
s−1 s+2 1 1
Y (s) = U (s)+ 2 y(0)+ 2 ẏ(0)− 2 u(0) .
s2 + 2s − 3 s + 2s − 3 s + 2s − 3 s + 2s − 3
| {z }
H(s)

Note that the denominator polynomial in each of the terms on the right hand
sides is equal to D(s) (the denominator of the transfer function). For simplicity,
suppose that y(0) = y0 (for some real number y0 ), ẏ(0) = 0 and u(0) = 0. The
s+2
partial fraction expansion of the term s2 +2s−3 y0 is given by
 
s+2 y0 1 1
y 0 = + ,
s2 + 2s − 3 4 s+3 s−1

c Shreyas Sundaram
9.2 Stability of the Unity Feedback Loop 67

and this contributes the term y40 e−3t + et , t ≥ 0, to the response of the sys-

tem. Note that the et term blows up, and thus the output of the system blows
up if y0 is not zero, even if the input to the system is bounded.
If all poles of the transfer function are in the OLHP (before any pole-zero
cancellations), all initial conditions will decay to zero, and not cause the output
of the system to go unbounded.
The above example demonstrates the following important fact:

Stability of a transfer function must be checked without canceling

any common poles and zeros from the transfer function. In particular,
systems with unstable pole-zero cancellations are unstable.

9.2 Stability of the Unity Feedback Loop

Consider the standard feedback structure shown below.

R(s) E(s) Y (s)

C(s) P (s)
−

n (s)
Suppose that we write P (s) = dpp(s) and C(s) = ndcc(s) (s)
for some polynomials
np (s), dp (s), nc (s), dc (s). The transfer function from r to y is given by

np (s)nc (s)
P (s)C(s) dp (s)dc (s) np (s)nc (s)
H(s) = = np (s)nc (s)
= .
1 + P (s)C(s) 1 + dp (s)dc (s) dp (s)dc (s) + np (s)nc (s)

The denominator of the above transfer function is called the characteristic poly-
nomial of the closed loop system, and we have the following result.

The unity feedback system is stable if and only if all roots

of the characteristic polynomial dp (s)dc (s) + np (s)nc (s) are
in the OLHP.

Note that the above test captures unstable pole/zero cancellations: if there is
an unstable pole s = a in dp (s) or dc (s), and that same root appears in either
nc (s) or nd (s), then s = a would be a root of dp (s)dc (s) + np (s)nc (s) and would
thus cause the characteristic polynomial to fail the test for stability.

c Shreyas Sundaram
68 Stability of Linear Time-Invariant Systems

9.3 Tests for Stability

Consider a general transfer function of the form

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

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

We would like to determine whether all poles of this transfer function are in the
OLHP. One way to do this would be to actually find the poles of the system
(by finding the roots of D(s)). However, finding the roots of a high-degree
polynomial can be complicated, especially if some of the coefficients are symbols
rather than numbers (this will be the case when we are designing controllers, as
we will see later). Furthermore, note that we do not need the actual values of
the poles in order to determine stability – we only need to know if all poles are
in the OLHP. Can we determine this from the coefficients of the polynomial?

9.3.1 A Necessary Condition for Stability

Suppose that the polynomial D(s) has roots −p1 , −p2 , . . . , −pn (for simplicity,
assume that these are all real, but the following analysis also holds if some of
the roots are complex). The polynomial can then be written as

D(s) = sn + an−1 sn−1 + · · · + a1 s + a0 = (s + p1 )(s + p2 ) . . . (s + pn ) .

How do the coefficients a0 , a1 , . . . , an−1 relate to p1 , p2 , . . . , pn ? To see the

pattern, consider the following cases:

(s + p1 )(s + p2 ) = s2 + (p1 + p2 )s + p1 p2
(s + p1 )(s + p2 )(s + p3 ) = s3 + (p1 + p2 + p3 )s2 + (p1 p2 + p1 p3 + p2 p3 )s
+ p1 p2 p3
(s + p1 )(s + p2 )(s + p3 )(s + p4 ) = s4 + (p1 + p2 + p3 + p4 )s3
+ (p1 p2 + p1 p3 + p1 p4 + p2 p3 + p2 p4 + p3 p4 )s2
+ (p1 p2 p3 + p1 p2 p4 + p1 p3 p4 + p2 p3 p4 )s
+ p1 p2 p3 p4
..
.

Based on the above examples, we see that:

• an−1 is the sum of all the pi ’s.

• an−2 is the sum of all products of the pi ’s taken two at a time.

• an−3 is the sum of all products of the pi ’s taken three at a time.

c Shreyas Sundaram
9.3 Tests for Stability 69

..
.

• a0 is the product of all the pi ’s.

Now, suppose that all roots are in the OLHP (i.e., p1 > 0, p2 > 0, . . . , pn > 0).
This means that a0 > 0, a1 > 0, . . . , an−1 > 0 as well. This leads us to the
following conclusion.

The polynomial D(s) = sn + an−1 sn−1 + · · · + a1 s + a0 has all roots in

the OLHP only if all of the coefficients a0 , a1 , . . . , an−1 are positive.

Note that the above condition is necessary, but it is not sufficient in general.
We will see this in the following examples.
Examples.

• D(s) = s3 − 2s2 + s + 1:

• D(s) = s4 + s2 + s + 1:

• D(s) = s3 + 2s2 + 2s + 1:

• D(s) = s3 + 2s2 + s + 12:

In the last example above, the polynomial can be factored as D(s) = (s +

3)(s2 − s + 4), and the second factor has roots in the CRHP. This shows that
a polynomial having all coefficients positive does not mean that all of the roots
are in the OLHP. Only the converse is true: if all of the roots are in the OLHP,
then all of the coefficients are positive.
Although the test for positive coefficients allows us to immediately determine
whether the roots of a polynomial are not stable, we need a way to conclusively
determine whether all of the roots are stable.

c Shreyas Sundaram
70 Stability of Linear Time-Invariant Systems

9.3.2 A Necessary and Sufficient Condition:

Routh-Hurwitz Test
The problem of determining how many roots of the polynomial D(s) = an sn +
an−1 sn−1 + · · · + a1 s + a0 are in the ORHP was studied by E. J. Routh in the
late 1800’s. He devised the following technique for solving this problem (we will
assume that the polynomial D(s) has no roots on the imaginary axis, so it is
sufficient to check for poles in the ORHP, as opposed to CRHP).

First, construct a Routh array:

sn an an−2 an−4 ···

n−1
s an−1 an−3 an−5 ···
sn−2 b1 b2 b3 ···
sn−3 c1 c2 c3 ···
.. .. .. .. ..
. . . . .
s2 ∗ ∗
s1 ∗
s0 ∗

The first two rows of this array contain the coefficients of the polynomial. The
numbers b1 , b2 , b3 , . . . on the third row are defined as:
 
1  an an−2  an−1 an−2 − an an−3
b1 = − = ,
an−1 an−1 an−3  an−1
 
1  an an−4  an−1 an−4 − an an−5
b2 = − = ,
an−1 an−1 an−5  an−1
 
1  an an−6  an−1 an−6 − an an−7
b3 = − = ,
an−1 an−1 an−7  an−1
..
.
Notice the pattern: the i–th element on the third row is obtained by taking
the negative determinant of the matrix consisting of the first column and the
(i+1)–th column in the first two rows, divided by the first element in the second
row. The third row will have one less element than the first two rows.
Similarly, the numbers c1 , c2 , c3 , . . . on the fourth row are defined as:
 
1 an−1 an−3  b1 an−3 − an−1 b2
c1 = −  = ,
b1 b1 b2  b1
 
1 a an−5  b1 an−5 − an−1 b3
c2 = −  n−1 = ,
b1 b1 b3  b1
 
1 a an−7  b1 an−7 − an−1 b4
c3 = −  n−1 = ,
b1 b1 b4  b1
..
.

c Shreyas Sundaram
9.3 Tests for Stability 71

Again, notice the pattern: the i–th element on the fourth row is obtained by
taking the negative determinant of the matrix consisting of the first column and
the (i + 1)–th columns of the preceding two rows, divided by the first element
in the immediately preceding row.
We continue this process until the (n + 1)–th row (corresponding to s0 in the
Routh array), which will have only one entry. After this process is complete,
we can use the following result1 to check stability of D(s).

The number of sign changes in the first column of the Routh array
indicates the number of roots that are in the ORHP. All roots are in
the OLHP if and only if there are no sign changes in the first column
(i.e., either all entries are positive, or all are negative).

Note: The above algorithm requires us to divide certain determinants by the

first entry in each row of the Routh array. If this entry is zero, but all other
entries in the row are nonzero, we can replace the zero by a small positive value
, and continue as before (the above result still holds in this case). On the other
hand, if an entire row of the Routh array is zero, this signifies the potential
existence of poles on the imaginary axis, and a slightly more complex procedure
will have to be followed. Be aware that these special cases exist, but we will not
deal with them in this course.
Example. Determine whether the polynomial D(s) = s4 + 3s3 + 2s2 + 2s + 9
has all roots in the OLHP.
Solution.

Example. Determine whether the polynomial D(s) = s4 + 5s3 + 9s2 + 5s + 2

1 A relatively simple proof of this result can be found in the paper “Elementary Proof of

the Routh-Hurwitz Test” by G. Meinsma (Systems and Control Letters, vol. 25, no. 4, 1995,
pp. 237-242).

c Shreyas Sundaram
72 Stability of Linear Time-Invariant Systems

has all roots in the OLHP.

Solution.

9.3.3 Testing For Degree of Stability

There may be times when we would like to test a polynomial to see if the
real parts of all of its roots are less than a certain value; so far, we have been
considering this value to be zero (i.e., we have been testing to see if all roots lie
in the OLHP). Suppose that we would like to see if all roots of a polynomial
D(s) have real part less than λ. Consider the polynomial D̄(s) = D(s + λ). It
is easy to see that the roots of D̄(s) are the roots of D(s) shifted by λ: if s = a
is a root of D(s), then s = a − λ is a root of D̄(s). Thus, all roots of D(s) have
real parts less than λ if and only if all roots of D̄(s) are in the OLHP, and we
can use the Routh-Hurwitz test on D̄(s) to see whether all roots of D(s) lie to
the left of λ.

Example. Determine whether the polynomial D(s) = s3 + 3s2 + 3s + 9 has all

roots with real parts less than −1.

c Shreyas Sundaram
9.3 Tests for Stability 73

Solution.

9.3.4 Testing Parametric Stability with Routh-Hurwitz

In control system design, we will frequently run across cases where some of the
coefficients of the polynomial are parameters for us to design or analyze (such
as control gains, or unknown values for system components). We can use the
Routh-Hurwitz test to determine ranges for these parameters so that the system
will be stable.
Example. In the feedback control loop shown below, determine the range of
values for K for which the closed loop transfer function (from R(s) to Y (s)) will
be stable.

R(s) 1
Y (s)
− K (s+6)(s+3)(s−1)

Solution.

c Shreyas Sundaram
74 Stability of Linear Time-Invariant Systems

Example. In the feedback loop shown below, determine the values of K and a
for which the closed loop system is stable.

R(s) K(s+a) 1
Y (s)
− s+1 s(s+2)(s+3)

Solution.

c Shreyas Sundaram
Chapter 10

Properties of Feedback

With the tools to analyze linear systems in hand, we now turn our attention
to feedback control. Recall the unity feedback control loop that we looked at
earlier:

D(s)
R(s) + Y (s)
+
C(s) P (s)
−

Figure 10.1: Feedback Control.

In the above figure, P (s) is the plant, C(s) is the controller, r is the reference
signal, y is the output of the system, and d is a disturbance affecting the
control loop (e.g., wind on an airplane, noise in a sensor, faults in an electrical
grid, etc.).
Recall from Chapter 1 that there are three main properties that a good control
system should have:

• Tracking. The output of the system should behave like the reference
signal. This property is studied by examining the transfer function from
the reference input to the output.
• Disturbance Rejection. The disturbances should affect the output as
little as possible. This property is studied by examining the transfer func-
tion from the disturbance to the output.
• Robustness. The output should track the reference signal even if the
plant model is not exactly known, or changes slightly. This property is
studied by examining the sensitivity of the transfer function to perturba-
tions in the plant, as we show below.

c Shreyas Sundaram
76 Properties of Feedback

Sensitivity. Let Try (s) denote the transfer function of the control system from
the reference r to the output y. Now suppose that we allow the plant P (s) to
change by a small amount δP (s) to become the new plant P̄ (s) = P (s) + δP (s).
This will cause the transfer function Try (s) to also change by a small amount
δTry (s), to become T̄ry (s) = Try (s)+δTry (s). The question is: how does δTry (s)
compare to δP (s)? More specifically, the sensitivity is defined as the fractional
(or percentage) change in the transfer function as related to the fractional change
in the plant model:

δTry (s)
Try (s) δTry (s) P (s)
S(s) = δP (s)
= .
δP (s) Try (s)
P (s)

δT (s)
Note that for small perturbations δP (s) and δTry (s), the expression δPry(s) is
the derivative of Try (s) with respect to P (s). In order to have good robustness,
we want the sensitivity to be as small as possible.

We will analyze the tracking, disturbance rejection and robustness properties for
both a feedforward control configuration and a feedback control configuration,
and see why feedback is an important concept for control system design.

10.1 Feedforward Control

Recall from the first lecture that feedforward control does not make use of
measurements of the system output in order to supply the input. The block
diagram for feedforward control is shown in Fig. 10.2.

D(s)
R(s) + Y (s)
+
C(s) P (s)

Figure 10.2: Feedforward Control.

We can examine how well feedforward control satisfies the properties listed
above.

• Tracking. The transfer function from the reference to the output is ob-
tained by assuming that the disturbance is not present (i.e., take d = 0),
and is given by
Y (s)
Try (s) = = P (s)C(s) .
R(s)

c Shreyas Sundaram
10.2 Feedback Control 77

We can potentially get perfect tracking by choosing C(s) = P (s)−1 , in

which case we have Y (s) = R(s). Note, however, that “inverting” the
plant in this way may not always be possible in practice. Setting the
implementation details aside for now, we note that feedforward control
can potentially provide perfect tracking.
• Disturbance Rejection. The transfer function from the disturbance to
the output is obtained by assuming that the reference signal is not present
(i.e., r = 0), and is given by
Y (s)
Tdy (s) = =1 .
D(s)
Thus, the output is completely affected by the disturbance. Furthermore,
the feedforward control scheme provides us with no way to reduce this
influence. Thus, feedforward control is bad at rejecting distur-
bances.
• Robustness. To examine how robust feedforward control is to variations
in the plant, we examine the sensitivity
δTry (s) P (s)
S(s) = .
δP (s) Try (s)
Since Try (s) = P (s)C(s), we have
P (s)
S(s) = C(s) =1 .
P (s)C(s)
This shows that the transfer function is 100% sensitive to changes in the
plant, which is bad for robustness.

We can summarize the above findings as follows.

Feedforward control is potentially good for tracking, bad for

disturbance rejection and bad for robustness.

10.2 Feedback Control

Now consider the feedback control scheme shown in Fig. 10.1. We will study
how well this control scheme satisfies the properties described earlier.

• Tracking. The transfer function from the reference to the output is ob-
tained by assuming that the disturbance is not present (i.e., take d = 0),
and is given by
Y (s) P (s)C(s)
Try (s) = = .
R(s) 1 + P (s)C(s)

c Shreyas Sundaram
78 Properties of Feedback

We would like to make this transfer function as close to 1 as possible, so

that we can get Y (s) = R(s). One way to accomplish this would be to
choose C(s) = K for some very large constant K (this is called “high gain
control”). In this case, the magnitude of P (s)C(s) will be large for all s,
and thus Try (s) is approximately equal to 1. Thus, feedback control
can provide good tracking.
• Disturbance Rejection. The transfer function from the disturbance to
the output is obtained by assuming that the reference signal is not present
(i.e., r = 0), and is given by

Y (s) 1
Tdy (s) = = .
D(s) 1 + P (s)C(s)

We would like to make this transfer function as small as possible (so

that the disturbance does not affect the output). Once again, suppose
we choose C(s) = K for some very large constant K. In this case, the
denominator of Tdy (s) becomes very large, and so Tdy (s) becomes small.
Thus, feedback control can be good at rejecting disturbances.
• Robustness. To examine how robust feedback control is to variations in
the plant, we examine the sensitivity

δTry (s) P (s)

S(s) = .
δP (s) Try (s)
P (s)C(s)
Since Try (s) = 1+P (s)C(s) , we have

C(s) P (s) 1
S(s) = = .
(1 + P (s)C(s))2 P (s)C(s) 1 + P (s)C(s)
1+P (s)C(s)

Once again, choosing C(s) = K, for a large constant K, makes S(s) very
small. Thus, feedback control is robust to variations in the plant.

We can summarize the above findings as follows.

Feedback control is good for tracking, good for disturbance

rejection and good for robustness.

In addition to the above benefits, we will see later that feedback control can
also stabilize an unstable plant (whereas feedforward control cannot). Also, it
is worth noting that although we used high gain control to show that feedback
provides good tracking, disturbance rejection and robustness, it is not the only
option (and sometimes not even the best option). In practice, high gains are

c Shreyas Sundaram
10.2 Feedback Control 79

hard to implement in hardware (due to saturation effects and physical limi-

tations). Furthermore, in some cases, high gain can even destabilize a stable
system (we will see this when we study root-locus methods). We will be study-
ing ways to design more sophisticated feedback controllers to get around these
problems.

c Shreyas Sundaram
80 Properties of Feedback

c Shreyas Sundaram
Chapter 11

Tracking of Reference
Signals

Recall the basic feedback control loop that we looked at earlier:

R(s) E(s) Y (s)

C(s) P (s)
−

In the above figure, P (s) is the plant, C(s) is the controller, r is the reference
signal, y is the output of the system, and e is the error (i.e., e = r − y). We will
neglect disturbances for now. Note that the signal y is directly subtracted from
the reference signal in this feedback loop, so this configuration is called unity
feedback. The transfer function from r to y is
Y (s) P (s)C(s)
Try (s) = = .
R(s) 1 + P (s)C(s)
Recall that the product P (s)C(s) is called the forward gain of the feedback
loop. We can always write P (s)C(s) as
a(s)
P (s)C(s) = ,
sq b(s)
for some polynomials a(s) and b(s), and some nonnegative integer q (this is the
number of poles at the origin in the product P (s)C(s)). The polynomial b(s)
has no roots at s = 0 (otherwise this root can be grouped into the term sq );
this means that b(0) 6= 0. The transfer function then becomes
a(s)
sq b(s) a(s)
Try (s) = = .
1 + sa(s)
q b(s)
sq b(s) + a(s)

c Shreyas Sundaram
82 Tracking of Reference Signals

12(s+2)
Example. Suppose P (s) = s+1 s2 and C(s) = s(s+4) . What are a(s), b(s) and
q? What is the transfer function?
Solution.

We’ll assume that C(s) is chosen so that the closed loop system is stable (other-
wise, we do not have any hope of tracking any signal). This means that all roots
of sq b(s) + a(s) are in the OLHP. We will now examine how well this feedback
loop tracks certain types of reference signals.

11.1 Tracking and Steady State Error

We will start by examining the steady state error of the above feedback loop
to reference signals of the form r(t) = tm , t ≥ 0, where m is some nonnegative
integer. The reference r(t) is a step input when m = 0, a ramp input when
m = 1, and a parabolic input when m = 2. Note from the Laplace Transform
m!
table that the Laplace transform of r(t) = tm , t ≥ 0 is R(s) = sm+1 .

Figure 11.1: The reference input r(t) = tm , t ≥ 0, for m = 0, 1, 2.

c Shreyas Sundaram
11.1 Tracking and Steady State Error 83

To determine how well the system tracks (or follows) the reference inputs, we
will examine the error signal e = r − y. Specifically, the Laplace Transform of
e(t) is given by:
E(s) = R(s) − Y (s) = R(s) − P (s)C(s)E(s) ,
from which we obtain
1 1
E(s) = R(s) = m! .
1 + P (s)C(s) (1 + P (s)C(s))sm+1
Recall that the Final Value Theorem states that if all poles of sE(s) are in the
OLHP, then the signal e(t) settles down to some finite steady state value. From
the above expression for E(s), we have
1
sE(s) = m! .
(1 + P (s)C(s))sm
Note that, as written, sE(s) seems to have m poles at the origin. Note from our
earlier discussion, however, that P (s)C(s) has q poles at the origin, so that we
can write P (s)C(s) = sa(s)
q b(s) , for some polynomials a(s) and b(s). The expression

for sE(s) can then be written as

1 b(s)
sE(s) = m! = m! .
(1 + a(s) m (sq b(s) + a(s))sm−q
sq b(s) )s

Recall that the polynomial sq b(s) + a(s) has all roots in the OLHP (by our
assumption of stability). We thus only have to check for poles at the origin
(given by the quantity sm−q ). We consider three different cases:

• If q > m, the function sE(s) will have all poles in the OLHP, and q − m
zeros at the origin. The steady state error is obtained from the Final Value
Theorem as
0q−m b(0)
ess = lim e(t) = lim sE(s) = m! =0 .
t→∞ s→0 0q b(0)
+ a(0)
Thus, if q > m, we have perfect steady state tracking (i.e., the steady
state error is zero).
• If q = m, the function sE(s) has all poles in the OLHP, and no zeros at
the origin. From the Final Value Theorem, we have
b(s)
ess = lim e(t) = lim sE(s) = lim m!
t→∞ s→0 s→0 sq b(s) + a(s)
1
= lim m! a(s)
s→0 sq + b(s)
1
= lim m! ,
s→0 sq + sq P (s)C(s)

c Shreyas Sundaram
84 Tracking of Reference Signals

which is some nonzero constant.1 Note that if q = 0, the steady state

1
error is ess = lims→0 m! 1+P (s)C(s) , and if q > 0, the steady state error
1
is ess = lims→0 m! sq P (s)C(s) . Thus, we have steady state tracking to
within a constant finite error.

• If q < m, the function sE(s) will have m − q poles at the origin. Thus
the signal e(t) blows up as t → ∞, and there is no steady state value. In
other words, the system output does not track the reference input
at all.

Summary. Suppose P (s)C(s) has q poles at the origin (q is a non-

negative integer), and suppose the reference input to the unity feedback
loop is r(t) = tm , t ≥ 0, where m is a nonnegative integer.
• If q > m, the output y(t) will track the reference r(t) perfectly in
steady state (i.e., the steady state error will be zero).
• If q = m, the output y(t) will track the reference r(t) to within a
constant (finite) steady state error.
• If q < m, the output y(t) will not track the reference r(t) at all.

Note that if a linear system can track a signal tm , t ≥ 0, then it can also track
any polynomial of degree m or less (by linearity).
System type. The above results indicate that the number of poles at the origin
in P (s)C(s) determines the type of reference inputs that the closed loop system
can track. Thus, the integer q is called the system type. Specifically, a system
of type q can track reference signals that are polynomials of degree q or less to
within a constant finite steady state error.
Note: It does not matter whether the poles in P (s)C(s) come from the plant
P (s) or from the controller C(s). The only thing that matters is how many poles
their product has. We can therefore use this fact to construct controllers with a
certain number of poles in order to track certain types of reference signals, even
if the plant does not have the required number of poles. We will see this in the
next lecture.

1 Do not try to memorize this. Instead, always just derive the tracking error using first

principles, first calculating E(s) in terms of Try (s) and R(s), and then applying the final
value theorem.

c Shreyas Sundaram
11.1 Tracking and Steady State Error 85

Example. Consider the unity feedback loop with C(s) = 12(s+2) s(s+4) and P (s) =
s+1
s2 . What is the system type? What is the steady state tracking error for the
signals r(t) = 1(t), r(t) = t1(t), r(t) = t2 1(t), r(t) = t3 1(t), and r(t) = t4 1(t)?
Solution.

c Shreyas Sundaram
86 Tracking of Reference Signals

Example. Suppose that C(s) = K (for some positive constant K), and P (s) =
1
s+1 . What is the system type? What is the steady state tracking error for the
signals r(t) = 1(t), r(t) = t1(t) and r(t) = t2 1(t)?
Solution.

c Shreyas Sundaram
11.1 Tracking and Steady State Error 87

To see why having an integrator guarantees perfect tracking for a step input (if
the closed loop system is stable), rearrange the closed loop system as follows:

R(s) E(s) 1
W (s) Rest of Y (s)
s P (s)C(s)
−

In the above diagram, we have simply pulled the integrator out of the product
P (s)C(s), and denoted the signal at the output of the integrator by W (s) (or
w(t) in the time-domain). Now suppose that the closed loop system is stable
(note that this is a necessary assumption); in this case, all of the signals in the
system will settle down to some steady state values when r(t) is a unit step
input. This includes the signal w(t), which is related to e(t) as
Z t
1
W (s) = E(s) ⇔ w(t) = e(τ )dτ.
s 0

If e(t) settles down to some nonzero value, the above integral will become un-
bounded, and thus w(t) will not settle down to some steady state value, con-
tradicting the fact that the system is stable. Thus the only way for all signals
to have settled to a steady state value (which is guaranteed by stability) is if
e(t) → 0. An alternative way to see this is to note that if w(t) settles down to
a steady state value, then ẇ(t) = e(t) = 0.
This is an example of what is known as the internal model principle: if we
wish to perfectly track a signal of a certain form, we should include a model of
that signal inside our feedback loop.

c Shreyas Sundaram
88 Tracking of Reference Signals

c Shreyas Sundaram
Chapter 12

PID Control

R(s) E(s) Y (s)

C(s) P (s)
−

So far, we have examined the benefits of feedback control, and studied how
the poles of P (s)C(s) affect the ability of the control system to track reference
inputs. We will now study a type of controller C(s) that is commonly used in
practice, called a proportional-integral-derivative (PID) controller. To
develop this controller, we will assume that the plant is a second order system
of the form
b0
P (s) = 2 .
s + a1 s + a0
Note that the transfer function from r to y for the above feedback loop is given
by
P (s)C(s)
Try (s) = .
1 + P (s)C(s)

12.1 Proportional (P) Control

We start with a simple controller of the form C(s) = KP , where KP is a
constant that we will choose. In this case, the input to the plant is simply
u(t) = KP e(t), which is proportional to the error. Thus, this type of controller
is called a proportional controller.
With this controller, the transfer function from r to y in the feedback control
system becomes
KP b0
Try (s) = 2 .
s + a1 s + (a0 + KP b0 )

c Shreyas Sundaram
90 PID Control

Recall that the poles of this transfer function dictate how the system behaves
to inputs. In particular, we would like to ensure that the system is stable (i.e.,
all poles are in the OLHP). Since the gain KP affects one of the coefficients in
the denominator polynomial, it can potentially be used to obtain stability.
1
Example. Suppose P (s) = s2 +3s−1 . Can we stabilize this plant with propor-
tional control?
Solution.

1
Example. Suppose P (s) = s2 −3s−1 . Can we stabilize this plant with propor-
tional control?
Solution.

The above examples demonstrate that simple proportional control can stabilize
some plants, but not others.
Another benefit of proportional control is that it can potentially be used to
speed up the response of the system. Recall the standard second order system
had a denominator of the form s2 + 2ζωn s + ωn2 , and the larger ωn is, the faster
the system responds. In the closed loop transfer function Try (s) above, the term
ωn2 is given by a0 + KP b0 , and thus we can potentially make ωn very large by
choosing KP to be very large, thereby speeding up the system.

c Shreyas Sundaram
12.2 Proportional-Integral (PI) Control 91

Now let’s consider tracking. Recall from the previous lecture that in order to
track a step input perfectly (i.e., with zero steady state error), the system must
be of type 1 (a type 0 system would track a step within a finite steady state
error). If C(s) = KP and P (s) = s2 +ab10s+a0 , the system would only be of type
0 (if a0 is not zero), and thus we will not be able to track a step perfectly. To
rectify this, we will have to add an integrator to the controller in order to make
the system type 1.

12.2 Proportional-Integral (PI) Control

To obtain perfect tracking for step inputs, we will introduce an integrator into
the controller (i.e., we will add a pole at the origin) in order to ensure that the
system will be of type 1. The controller thus becomes

KI
C(s) = KP + .
s
In the time-domain, this corresponds to the input to the plant being chosen as
Z t
u(t) = KP e(t) + KI e(τ )dτ ,
0

and thus this is called a proportional-integral controller. With this con-

troller, the transfer function from r to y is
b0 KP s+KI
s2 +a1 s+a0 s b0 (KP s + KI )
Try (s) = = .
1 + s2 +ab10s+a0 KP s+K
s
I s3 + a1 s2 + (a0 + KP b0 )s + KI b0

Note that we now have a third order system. Two of the coefficients of the
denominator polynomial can be arbitrarily set by choosing KP and KI appro-
priately. Unfortunately, we still have no way to stabilize the system if a1 < 0
(recall that for stability, all coefficients must be positive). Even if the system
is stable with the given value of a1 , we might want to be able to choose better
pole locations for the transfer function in order to obtain better performance.
To do this, we add one final term to the controller.

12.3 Proportional-Integral-Derivative (PID) Con-

trol
Consider the PI controller from the last section, and add a term that corresponds
to the derivative of the error. The controller with this additional term has the
form
KI
C(s) = KP + + KD s .
s

c Shreyas Sundaram
92 PID Control

In the time-domain, the input to the plant due to this controller is given by
Z t
u(t) = KP e(t) + KI e(τ )dτ + KD ė(t) ,
0

and so this controller is called a proportional-integral-derivative controller.

The transfer function from r to y is

b0 KP s+KI +KD s2
s2 +a1 s+a0 s
Try (s) = 2
1 + s2 +ab10s+a0 KP s+KsI +KD s
b0 (KP s + KI + KD s2 )
= .
s3 + (a1 + KD b0 )s2 + (a0 + KP b0 )s + KI b0

Note that we are now able to arbitrarily set all coefficients of the denominator
polynomial, via appropriate choices of KP , KI and KD . Thus we can now guar-
antee stability (only for a second order plant, though), good transient behavior,
and perfect tracking!
1
Example. Consider the plant P (s) = s2 −3s−1 . Design a PID controller so that
the closed loop system has perfect tracking for a step input, and has poles at
s = −5, −6, −7.
Solution.

c Shreyas Sundaram
12.4 Implementation Issues 93

12.4 Implementation Issues

PID control is extremely common in industry, as it is easy to design – we just
have to “tune the knobs” corresponding to the gains KP , KI and KD appropri-
ately. However, it is worth noting that it may not be desirable to implement
the controller as given above; in practice, all signals will contain high frequency
noise, and differentiating noise will once again create signals with large mag-
nitudes. To avoid this, the derivative term KD s is usually implemented in
1
conjunction with a low pass filter of the form τ s+1 , for some small τ . This has
the effect of attenuating the high frequency noise entering the differentiator,
and produces the controller
KI KD s
C(s) + KP + + .
s τs + 1
We will be seeing controllers of this form frequently in the second half of the
course.

c Shreyas Sundaram
94 PID Control

c Shreyas Sundaram
Chapter 13

Root Locus

Based on our discussion so far, we know that the response of a linear system to
an input is dictated by the location of the poles of the transfer function. For
example, the response will be unstable if there are any poles in the CRHP, or
may contain oscillations if the poles appear in complex conjugates. We have also
seen that feedback control can be used to move the poles of a closed loop sys-
tem: by choosing the controller gain appropriately, one can potentially stabilize
unstable systems (and perhaps even destabilize stable systems). In this section
of the course, we will examine in more detail how the poles of a transfer function
vary in the complex plane in response to changes in a certain parameter. We
will begin with some examples.
1
Example. Consider the unity feedback loop with C(s) = K and P (s) = s−3 .
How does the pole of the closed loop system vary with K (for K ≥ 0)?
Solution.

c Shreyas Sundaram
96 Root Locus

1
Example. Consider the unity feedback loop with C(s) = K and P (s) = s2 +2s .
How do the poles of the closed loop system vary with K (for K ≥ 0)?
Solution.

1
Example. Consider the closed loop transfer function Try (s) = s2 +bs+1 . How
do the poles of the system vary with b (for b ≥ 0)?
Solution.

While we are able to easily draw the locations of the poles for first and second

c Shreyas Sundaram
13.1 The Root Locus Equations 97

order systems, we cannot do the same for higher order systems (because it
becomes difficult to explicitly calculate the roots). Furthermore, we would like
to obtain some intuition about how the poles of the system will be affected by
our choice of controller. We would thus like to come up with a way to sketch
how the poles of a system behave in response to a change in a parameter. Since
the poles are given by the roots of the denominator polynomial, such a sketch is
called a root locus. The trajectory of each root in the plane is called a branch
of the root locus.

13.1 The Root Locus Equations

Consider the unity feedback loop:

R(s) Y (s)
K L(s)
−

The transfer function L(s) could represent the plant, or it could represent some
composite system (such as the combination of a controller and plant). We will
write
N (s) sm + bm−1 sm−1 + bm−2 sm−2 + · · · + b1 s + b0
L(s) = = n ,
D(s) s + an−1 sn−1 + an−2 sn−2 + · · · + a1 s + a0
where N (s) and D(s) are polynomials in s. As usual, the degree of N (s) is m,
the degree of D(s) is n, and we assume that n ≥ m (i.e., L(s) is proper). The
transfer function from r to y is given by
KL(s) KN (s) KN (s)
Try (s) = = = .
1 + KL(s) D(s) + KN (s) ∆(s)
The polynomial ∆(s) = D(s)+KN (s) is called the characteristic polynomial
of the system. Note that the roots of D(s) + KN (s) are the closed loop poles,
the roots of D(s) are the open loop poles, and the roots of N (s) are the open
loop zeros. When we plot these elements graphically, we will use × to denote
poles, and ◦ to denote zeros.

The root locus is a graph of how the roots of D(s) + KN (s) vary with
K. Equivalently, the root locus is the set of all solutions s to the
1
equation L(s) = − K .

Note that the root locus can actually be used to find how the roots of any
polynomial vary with a single parameter, and thus it is a very general tool. For

c Shreyas Sundaram
98 Root Locus

example, aside from analyzing how the poles of a unity feedback loop vary with
a controller gain K, the root locus can also be used to analyze how the poles
vary in response to a change in one of the system parameters. We saw this in
the third example above, and we will see more of it later.
To start, we will only focus on the case where K varies from 0 to ∞ – this is
called the positive root locus. We will deal with the negative root locus later.
We will also assume that both N (s) and D(s) are monic (i.e., the coefficient
corresponding to the highest power in both polynomials is equal to 1). This
is not a strict assumption, because we can always divide the entire polynomial
D(s) + KN (s) by the leading coefficient of D(s), and then absorb the leading
coefficient of N (s) into K to define a new gain K̄. After plotting the root locus,
we can then map the gain K̄ back to K.

13.1.1 Phase Condition

To obtain an alternative (but equivalent) characterization of the root locus,

consider the equation
1
L(s) = − .
K
Since we are focusing on the case where K ≥ 0, we see that L(s) must be
a negative real number, and the root locus consists of all points satisfying this
condition. In particular, since the phase of a negative real number is 180 degrees
(or π radians), we obtain the following characterization of the root locus.

The positive root locus is the set of all points s in the complex plane for
which ∠L(s) = (2l + 1)π radians (where l is any integer).

We can obtain a geometric interpretation of the above condition as follows. Let

(s + z1 )(s + z2 ) · · · (s + zm )
L(s) = ,
(s + p1 )(s + p2 ) · · · (s + pn )

where −z1 , −z2 , . . . , −zm are the open loop zeros, and −p1 , −p2 , . . . , −pn are
the open loop poles. The phase of L(s̄), for some point s̄ in the complex plane,
is given by

∠L(s̄) = ∠(s̄ + z1 ) + ∠(s̄ + z2 ) + · · · + ∠(s̄ + zm )

− ∠(s̄ + p1 ) − ∠(s̄ + p2 ) − · · · − ∠(s̄ + pn ) .

c Shreyas Sundaram
13.1 The Root Locus Equations 99

Note that the phase of the point s̄+zi is given by the angle between the positive
real axis and the vector from −zi to s̄:

The same holds for the phase of the point s̄+pi . The phase of L(s̄) can therefore
be obtained by summing these angles, and this will allow us to determine if s̄
is on the root locus.

s+4
Example. Consider L(s) = s((s+1) 2 +1) . Is the point s = −3 on the root locus?

Is the point s = −4 + j on the root locus?

Solution.

c Shreyas Sundaram
100 Root Locus

While we can use the above method to test if specific points are on the root locus,
it is quite cumbersome to determine all points in the complex plane that are on
the root locus in this way. What we need are some general rules for sketching
the root locus for a given L(s) (or equivalently, for the equation D(s) + KN (s)).

13.2 Rules for Plotting the Positive Root Locus

We will now develop some rules for plotting the positive root locus for a given
polynomial D(s) + KN (s).

13.2.1 Start Points and (Some) End Points of the Root

Locus
Consider the equation D(s) + KN (s) = 0. Since this is an n–th degree polyno-
mial, it will have n roots, and thus the root locus will have n branches. Let’s
start by considering what happens for the extreme values of K:

• When K = 0, the roots of this equation are simply the roots of D(s),
which are the open loop poles.

• When K → ∞, we consider the equivalent root locus equation N (s) 1

D(s) = − K .
As K → ∞, the right side goes to zero, and thus the root locus consists
of the points that cause N (s)
D(s) to be zero. This will be true for the m roots
of N (s) (i.e., the open loop zeros), and thus m of the branches end at the
open loop zeros. We will see what happens to the other n − m branches
later.

This brings us to the following rule:

Rule 1. The n branches of the root locus begin at the open loop poles
(when K = 0). Of the n branches, m branches end at the open loop
zeros (when K = ∞).

13.2.2 Points on the Real Axis

Recall that the positive root locus contains all points s̄ such that

∠(s̄ + z1 )+∠(s̄ + z2 )+· · ·+∠(s̄ + zm )−∠(s̄ + p1 )−∠(s̄ + p2 )−· · ·−∠(s̄ + pn )

= (2l + 1)π radians .

c Shreyas Sundaram
13.2 Rules for Plotting the Positive Root Locus 101

Consider a point s = s̄ on the real axis. Each real pole or zero to the right
of s̄ contributes −π radians or π radians to the angle. Each pair of complex
conjugate poles or zeros contributes nothing to the angle of a s̄ (since the angles
of the complex conjugate poles or zeros will sum to zero). Each pole or zero to
the left of s̄ will also contribute nothing (i.e., 0 radians) to the angle.

Thus, in order for s̄ to satisfy the above condition, there must be an odd number
of zeros or poles to the right of s̄.

Rule 2. The positive root locus contains all points on the real axis
that are to the left of an odd number of zeros or poles.

(s+3)(s+7)
Example. Consider L(s) = s2 ((s+1) 2 +1)(s+5) . Determine the portions of the

real axis that are on the root locus.

Solution.

13.2.3 Asymptotic Behavior of the Root Locus

We have already seen that m out of the n branches end up at the open loop
zeros as K → ∞. To see what happens to the other n − m branches, consider
1
the root locus equation L(s) = − K . As K → ∞, the right side goes to zero,

c Shreyas Sundaram
102 Root Locus

and thus the root locus consists of the points that cause L(s) to be zero. This
will be true for the m open loop zeros – are there any other choices of s that
will make the left hand side equal to zero? To answer this, write L(s) as
sm + bm−1 sm−1 + · · · + b1 s + b0
L(s) =
sn + an−1 sn−1 + · · · + a1 s + a0
1 1 1 1
n−m + bm−1 sn−m+1 + · · · + b1 sn−1 + b0 sn
= s 1 1 1 .
1 + an−1 s + · · · + a1 sn−1 + a0 sn
If n > m, we see that the numerator goes to zero if |s| → ∞. Thus, the system
L(s) = N (s)
D(s) is said to have n − m zeros at infinity, in addition to the m finite
zeros (i.e., the roots of N (s)). We can thus conclude that m of the n branches
go to the open loop zeros, and the remaining n − m branches go off to infinity
as K → ∞. The question is, how do these branches approach infinity? The
following rule characterizes the behavior of these branches.

Rule 3. Of the n branches in the root locus, n − m of the branches go

to infinity, and asymptotically approach lines coming out of the point
s = α with angles Φl , where

Σ open loop poles − Σ open loop zeros (2l + 1)π

α= , Φl = , (13.1)
n−m n−m
for l = 0, 1, 2, . . . , n − m − 1.

The asymptotes for a few values of n − m are shown below:

We will now go over a sketch of the proof of this result.1 First, note that the
1 This proof is borrowed from Prof. Daniel Liberzon at the University of Illinois.

c Shreyas Sundaram
13.2 Rules for Plotting the Positive Root Locus 103

transfer function L(s) has n poles and m zeros at certain locations. Now, if we
were to consider some point s with a very large magnitude (i.e., very far away
from the other poles and zeros), the poles and zeros would essentially look like
they were clustered at one point; let’s call this point α. So, for large |s|, we
would like to find a good value of α so that we can approximate the transfer
function as

sm + bm−1 sm−1 + bm−2 sm−2 + · · · + b1 s + b0 1

L(s) = ≈
sn + an−1 sn−1 + an−2 sn−2 + · · · + a1 s + a0 (s − α)n−m
sn + an−1 sn−1 + an−2 sn−2 + · · · + a1 s + a0
⇔ m ≈ (s − α)n−m .
s + bm−1 sm−1 + bm−2 sm−2 + · · · + b1 s + b0

Note that we are taking the exponent to be n − m, because to the point s, the
m zeros look like they are directly on top of n poles, and thus they ‘cancel’ each
other out. To see what value of α to choose, let’s start dividing the denominator
on the left hand side into the numerator to obtain the first couple of terms of
the quotient:

Thus, we can write

(s − α)n−m ≈ sn−m + (an−1 − bm−1 )sn−m−1 + · · · ,

and since (s−α)n−m = sn−m −α(n−m)sn−m−1 +· · · , if we compare the second

term in the two expansions, we obtain

bm−1 − an−1
α= . (13.2)
n−m
Next, suppose that
m
X
N (s) = (s + z1 )(s + z2 ) · · · (s + zm ) = sm + zi sm−1 + · · ·
i=1
n
X
D(s) = (s + p1 )(s + p2 ) · · · (s + pn ) = sn + pi sn−1 + · · · .
i=1
Pm Pn
Thus, we have bm−1 = i=1 zi and an−1 = i=1 pi , which we substitute into
(13.2) to produce α in (13.1).

c Shreyas Sundaram
104 Root Locus

To derive the asymptote angles in (13.1), consider the root locus of (s−α)−(n−m) .
Recall that the root locus of a function is the set of all s such that

X X
∠s + zi − ∠s + pi = (2l + 1)π (13.3)

for some integer l, where zi are the zeros and pi are the poles. In this case, the
function
P (s − α)−(n−m) has no zeros, and all n − m poles at α. Thus, we have
∠s + pi ≈ (n − m)∠s − α, and the above expression becomes

(2l + 1)π
∠s − α = .
n−m

Note that the negative sign in front of the pole angles in (13.3) contributes an
angle of −π, which can just be absorbed into the term (2l + 1)π. There are
n − m different possibilities for the angle on the right hand side of the above
equation, corresponding to l = 0, 1, . . . , n − m − 1 (after this, the angles start
repeating), and thus there are n − m different asymptotes leading out from the
poles at α with the angles specified by (13.1).
1
Example. Consider L(s) = s(s+2) . Draw the portions of the real axis that are
on the positive root locus, and determine the asymptotes.
Solution.

1
Example. Consider L(s) = s((s+1) 2 +1) . Draw the portions of the real axis that

are on the positive root locus, and determine the asymptotes.

c Shreyas Sundaram
13.2 Rules for Plotting the Positive Root Locus 105

Solution.

s+6
Example. Consider L(s) = s((s+1) 2 +1) . Draw the portions of the real axis that

are on the positive root locus, and determine the asymptotes.

Solution.

Since we are only interested in polynomials D(s) + KN (s) that have purely real
coefficients, the roots of the polynomial will either be real or appear as complex
conjugate pairs. This produces the following important fact.

The root locus is always symmetric about the real axis.

13.2.4 Breakaway Points

We have already seen some examples of root loci where two or more points
come together and then bounce off each other in opposite directions. Such
points are called breakaway points. Note that at such points, the characteristic

c Shreyas Sundaram
106 Root Locus

polynomial ∆(s) = D(s) + KN (s) will have multiple roots at the breakaway
point. Let the breakaway point be s = s̄. Then we can write

∆(s) = (s − s̄)q D̄(s) ,

where q ≥ 2 is the multiplicity of the root s̄, and D̄(s) is some polynomial. This
means that ∆(s̄) = 0 and d∆ ds (s̄) = 0. Substituting ∆(s) = D(s) + KN (s), we
have

D(s̄) + KN (s̄) = 0
dD dN
(s̄) + K (s̄) = 0 .
ds ds

D(s̄)
Solving the first equation, we get K = − N (s̄) , and substituting this into the
second equation, we come to the following rule.

Rule 4. The root locus will have multiple roots at the points s̄ for
which both of the following conditions are satisfied.
• N (s̄) dD dN
ds (s̄) − D(s̄) ds (s̄) = 0.

D(s̄)
• −N (s̄) = K is a positive real number.

s+6
Example. Draw the positive root locus for L(s) = s(s+2) .
Solution.

c Shreyas Sundaram
13.2 Rules for Plotting the Positive Root Locus 107

Example. Verify that the branches in the positive root locus for L(s) =
1
s((s+1)2 +1) never intersect.
Solution.

c Shreyas Sundaram
108 Root Locus

There are various other rules that we could derive to draw root locus plots, but
they tend to be cumbersome to apply by hand. The above rules will be sufficient
for us to get intuition about many systems.

13.2.5 Some Root Locus Plots

1
Example. Consider a control system in unity feedback with P (s) = s2 and
s+1
C(s) = K s+12 . Draw the positive root locus.
Solution.

c Shreyas Sundaram
13.2 Rules for Plotting the Positive Root Locus 109

1
Example. Consider a control system in unity feedback with P (s) = s2 and
C(s) = K s+1
s+4 . Draw the positive root locus.
Solution.

c Shreyas Sundaram
110 Root Locus

1
Example. Consider a control system in unity feedback with P (s) = s2 and
s+1
C(s) = K s+9 . Draw the positive root locus.
Solution.

The above examples show that as the pole of the controller moves in closer to
the root locus, it tends to push the branches of the locus to the right. From our
discussion so far, we can state the following rules of thumb: poles repel, and
zeros attract.

13.2.6 Choosing the Gain from the Root Locus

The positive root locus tells us how the poles of the closed loop system vary as
we increase K from 0 to ∞. Once we’ve plotted the root locus, we may wish to
select points on the branches so that the closed loop system will have certain
desired properties (such as rise time, overshoot, peak time, etc). We then need
to find the gain K so that the closed loop poles will be at the desired locations.
To find the gain, note that
1 1
L(s) = − ⇔K=− .
K L(s)
Since K is a positive real number, we can examine the magnitude of both sides
of the above equation to obtain

c Shreyas Sundaram
13.3 Rules for Plotting the Negative Root Locus 111

1 |s + p1 ||s + p2 | · · · |s + pn |
K= = .
|L(s)| |s + z1 ||s + z2 | · · · |s + zm |

Specifically, if we have a desired point s̄ on the root locus, we can find the gain
K that produces a pole at s̄ by multiplying and dividing the lengths of the
vectors from each of the poles and zeros to s̄, according to the above equation.
1
Example. Suppose L(s) = s2 +2s . Find the gain K that results in the closed
loop system having a peak time of at most 2π seconds.
Solution.

Note that MATLAB is an extremely useful tool for doing this in practical con-
troller design. Once one has plotted the root locus for a given system in MAT-
LAB (using the rlocus command), one can simply click on the root locus branch
at any desired location to find the value of the gain at that point.

13.3 Rules for Plotting the Negative Root Locus

In the last section, we saw how to plot the root locus for positive values of the
gain K. We can now complete the root locus by considering negative values of
K – this is called the negative root locus. The only difference in this case
stems from the phase condition. Recall that the root locus consists of all points
1
s satisfying L(s) = − K . If K is a negative real number, L(s) is a positive real
number, and thus we can state the following.

The negative root locus is the set of all points s in the complex plane for
which ∠L(s) = 2lπ radians (where l is any integer).

c Shreyas Sundaram
112 Root Locus

All of the rules for plotting the positive root locus translate directly once we
consider this new phase condition:

• Rule 1. The n branches of the root locus begin at the open loop poles
(when K = 0). Of the n branches, m branches end at the open loop zeros
(when K = −∞).
• Rule 2. The negative root locus contains all points on the real axis that
are to the left of an even number of zeros or poles.
• Rule 3. Of the n branches in the root locus, n − m of the branches go to
infinity, and asymptotically approach lines coming out of the point s = α
with angles Φl , where
Σ open loop poles − Σ open loop zeros 2lπ
α= , Φl = ,
n−m n−m
for l = 0, 1, 2, . . . , n − m − 1.
• Rule 4. The root locus will have multiple roots at the points s̄ for which
both of the following conditions are satisfied.

N (s̄) dD dN
ds (s̄) − D(s̄) ds (s̄) = 0.
D(s̄)
−N (s̄) = K is a negative real number.

Note that the gain for a particular point s on the negative root locus is given
by
1 |s + p1 ||s + p2 | · · · |s + pn |
K=− =− .
|L(s)| |s + z1 ||s + z2 | · · · |s + zm |

The asymptotes for the negative root locus look like this:

c Shreyas Sundaram
13.3 Rules for Plotting the Negative Root Locus 113

1
Example. Determine the negative root locus for L(s) = s((s+1)2 +1) , and then

draw the complete root locus (both positive and negative).

Solution.

c Shreyas Sundaram
114 Root Locus

c Shreyas Sundaram
Chapter 14

Stability Margins from

Bode Plots

The last chapter showed how to analyze and understand the closed loop system
from a root locus perspective. We will now study the use of Bode plots to
analyze closed loop systems, complementing the root locus techniques. In the
next chapter, we will use these ideas to design controllers (building on our study
of PID controllers).
Suppose we’re given the Bode plot for the transfer function L(s), and we would
like to study properties of the following feedback loop:

R(s) Y (s)
K L(s)
−

In other words, we would like to infer some things about the closed loop
system based on the open loop Bode plot. Remember that we also did this
when we studied root locus plots: we studied the locations of the closed loop
poles by starting with the open loop poles and zeros.
Recall the root locus equation 1 + KL(s) = 0 (the closed loop poles are the
values s that satisfy this equation). When K is a positive real number, this
means that |KL(s)| = 1 and ∠KL(s) ≡ π (modulo 2π). A point s = jω on the
imaginary axis (for some ω) will be on the positive root locus if |KL(jω)| = 1 and
∠KL(jω) ≡ π. Since we have access to |KL(jω)| and ∠KL(jω) from the Bode
plot, we should be able to determine the imaginary axis crossings by finding the
frequencies ω (if any) on the plot that satisfy the conditions |KL(jω)| = 1 (or
20 log |KL(jω)| = 0) and ∠KL(jω) ≡ π.
1
To develop this further, suppose that L(s) = s
s(s+1)( 100 +1) . The Bode plot of

c Shreyas Sundaram
116 Stability Margins from Bode Plots

KL(s) (with K = 1) is given below:

We’ll be using the following terminology.

• Gain crossover frequency: This is the frequency ωcg such that

|KL(jωcg )| = 1 (or equivalently, 20 log |KL(jωcg )| = 0).

• Phase crossover frequency: This is the frequency ωcp such that

∠KL(jωcp ) ≡ π.

For the above example, we have ωcg ≈ 1 and ωcp ≈ 10. The phase at ωcg
is approximately − 3π
4 , and so the feedback configuration with K = 1 does
not have any closed loop poles on the imaginary axis. Is there another value
of K for which the closed loop system will have poles on the imaginary axis
(and thus cross the boundary from stability to instability)? To determine this
from the Bode plot, note that 20 log |KL(jω)| = 20 log K + 20 log |L(jω)| and
∠KL(jω) = ∠L(jω) (for K > 0). Thus, K has no effect on the phase, and it
affects the magnitude plot by shifting it up or down by 20 log K. For example,
when K = 10, the entire magnitude gets shifted up by 20 log 10 = 20 dB (when

c Shreyas Sundaram
 117

the vertical axis denotes 20 log |KL(s)|); this is shown on the above Bode plot
by the dashed lines. Based on this plot, we see that changing K has the effect of
changing the gain crossover frequency (but not the phase crossover frequency).
In order to find the value of K that causes some closed loop poles to lie on
the imaginary axis in the above example, we need to find out how to make the
gain crossover frequency and the phase crossover frequency coincide. Examining
the magnitude plot, we see that 20 log |L(j10)| ≈ −40, and thus the magnitude
curve needs to be shifted up by approximately 40 dB in order to set ωcg = ωcp ,
which can be accomplished by setting 20 log K ≈ 40, or K ≈ 100. Thus, we can
conclude that the closed loop system will have an imaginary axis crossing when
K ≈ 100. One can easily see from the Bode plot that this is the only positive
value of K for which this will happen.
We can verify this result by examining the positive root locus of L(s):

As expected, the branches cross the imaginary axis only once (other than the
trivial case where K = 0). To find the locations where the branches cross the
imaginary axis, we note that

1 + KL(s) = 0
1
⇔ 1+K s =0
s(s + 1)( 100 + 1)
⇔ s3 + 101s2 + 100s + 100K = 0 .

We use the Routh-Hurwitz test to determine the region of stability as 0 < K <
101. Thus, we have a potential imaginary axis crossing at K = 101. To find the
points on the imaginary axis where this happens, we set s = jω and K = 101
and solve the equation

(jω)3 + 101(jω)2 + 100(jω) + 100(101) = 0

⇔ − jω 3 − 101ω 2 + 100ωj + 100(101) = 0 .

Setting the imaginary and real parts to zero, we find that ω = 10. Thus, we
have an imaginary axis crossing at s = ±10j when K = 101. Note that this

c Shreyas Sundaram
118 Stability Margins from Bode Plots

agrees with the analysis from the Bode plot (the Bode plot actually told us
K ≈ 100, since we approximated the Bode plot with straight lines).
While we could determine imaginary axis crossings by looking at the Bode
plot, we didn’t necessarily know which direction the branches were going –
are we going from stability to instability, or instability to stability? We could
determine this information by looking at the root locus, but we will later develop
a completely frequency domain approach to characterizing the stability of the
closed loop system. For now, we will assume that the closed loop system is
stable with a given value of K, and investigate ways to design controllers using
a frequency domain analysis in order to improve the stability of the closed loop
system.
Stability Margins
We will define some terminology based on the discussion so far. Consider again
the Bode plot of KL(s) with K = 1:

c Shreyas Sundaram
 119

Assuming that the closed loop system is stable, we can ask the question: How
far from instability is the system? There are two metrics to evaluate this:

• Gain margin: This is the amount by which K can be multiplied

before |KL(jωcp )| = 1 (i.e., the gain crossover frequency and phase
crossover frequencies coincide).
• Phase margin: This is the amount by which the phase at ωcg
exceeds −π; more specifically, it is defined as

P M = ∠L(jωcg ) + π .

In general, we would like to have large gain and phase margins in order to
improve the stability of the system. In the above example with K = 1, the gain
margin is approximately 100, and the phase margin is approximately π4 . Let us
consider some more examples, just to be clear on the concept of gain and phase
margins.
1
Example. What is the gain margin and phase margin for KL(s) = s(s+1)2 ?
Solution.

c Shreyas Sundaram
120 Stability Margins from Bode Plots

s+1
Example. What is the gain margin and phase margin for KL(s) = s
s2 ( 10 +1) ?
Solution.

In the above example, we noticed that the gain margin is ∞, since the phase only
hits −π at ω = ∞. However, note that as we increase K, the gain crossover
frequency starts moving to the right, and the phase margin decreases. If we
examine the root locus for the system L(s) = s2 (s+1s , we see that we have a
10 +1)
set of poles that move vertically in the plane as K → ∞, and thus the damping
ratio ζ for these poles decreases as K → ∞. This seems to indicate that there
might be some relationship between the phase margin and the damping ratio ζ.
We will now derive an explicit relationship between these two quantities.
2
ωn
Consider the system L(s) = s(s+2ζωn ) , which is placed in the feedback configu-
ration

R(s) Y (s)
L(s)
−

c Shreyas Sundaram
 121

ω2
The closed loop transfer function is Try (s) = s2 +2ζωnn s+ω2 , which is the standard
n
second order system with damping ratio ζ. The phase plot of L(jω) is given by

This shows that the gain margin is ∞ (and this is easily verified by looking at
the root locus). Next, let’s examine the phase margin. By setting the magnitude
of L(jω) equal to 1, one can verify that the Bode plot of L(s) has gain crossover
frequency equal to qp
ωcg = ωn 1 + 4ζ 4 − 2ζ 2 .
Using the fact that P M = ∠L(jωcg ) + π, we obtain (after some algebra)
 
2ζ
P M = tan−1  qp  .
1 + 4ζ 4 − 2ζ 2

Notice that the phase margin is a function of ζ and not ωn . Interestingly, this
seemingly complicated expression can be approximated fairly well by a straight
line for small values of ζ:

c Shreyas Sundaram
122 Stability Margins from Bode Plots

For 0 ≤ ζ ≤ 0.7, the phase margin (in degrees) and damping ratio are
related by

P M ≈ 100ζ .

While we derived the above expression for the standard second order system, we
can also use it as a general rule of thumb for higher order systems. Specifically,
as the phase margin decreases, the system becomes less stable, and might exhibit
oscillatory behavior. We can use the above relationship to design control systems
in the frequency domain in order to obtain certain time-domain characteristics
(such as meeting overshoot specifications).
Another useful rule-of-thumb that is generally adopted for the closed loop band-
width ωBW is as follows:

In general, we have ωcg ≤ ωBW ≤ 2ωcg , and ωBW ≈ ωn , which leads to

the approximation

ωBW ≈ ωcg ≈ ωn .

Note that our discussions here have assumed a typical Bode plot that has large
magnitude at low frequencies, and low magnitude at high frequencies, with a
single gain crossover frequency. We can deal with more complicated Bode plots
by generalizing our discussions, but we’ll focus on these typical Bode plots for
now.
In order to obtain fast transient behavior, we typically want a large gain crossover
frequency, but this would come at the cost of decreasing the phase margin. Fur-
thermore, in order to obtain better steady state tracking, we would typically
want to increase the gain K in order to boost the low frequency behavior, but
this would again move the gain crossover frequency to the right and decrease
the phase margin. Therefore, we must consider more complicated controllers
(other than just a simple proportional controller K) in order to obtain a good
phase margin, a good gain crossover frequency, and good steady state tracking.
This will be the focus of the next part of the course.

c Shreyas Sundaram
Chapter 15

Compensator Design Using

Bode Plots

We now turn our attention to designing dynamic controllers (also called com-
pensators), building on our earlier study of PID controllers. We have seen so far
that the phase margin of a given system is representative of the system’s stabil-
ity, and is directly related to the damping of the system – a larger phase margin
makes the system more stable, and increases the damping. Given a system, we
thus want to design a controller that improves the phase margin. In certain sys-
tems, one way to do this would be to decrease the gain of the system, so that the
gain crossover frequency moves to the left (in the direction of increasing phase).
However, we have seen that the low frequency gain of the system is related to
how well the system tracks reference inputs – a larger low frequency gain corre-
sponds to better tracking. Another metric is the gain crossover frequency: since
the gain crossover frequency is approximately equal to the bandwidth and the
natural frequency of the system, a larger gain crossover frequency corresponds
to faster response, but also leads to smaller phase margins. Therefore, we would
like to design more sophisticated controllers in order to keep the low frequency
gain large (in order to meet tracking specifications), or to increase the gain
crossover frequency (in order to obtain faster transients), and also to increase
the phase at the gain crossover frequency (in order to boost the phase margin).
In this chapter, we will study the design of lead and lag compensators using
Bode plots. We will start by introducing the form of these controllers.

15.1 Lead and Lag Compensators

In our discussion on root locus plots, we came up with the following rule of
thumb: poles repel and zeros attract. In particular, we saw that adding a zero to
the system via the controller can produce stability, when a simple proportional

c Shreyas Sundaram
124 Compensator Design Using Bode Plots

controller could not. Recall that this was the same conclusion that we reached
when we were studying PID control.
1
Example. For the unity feedback loop with P (s) = s2 , draw the positive root
locus when C(s) = K and C(s) = K(s + 1).
Solution.

However, we also discussed the fact that the PD controller C(s) = KP + KD s is

not physically implementable, since it is not proper, and it would differentiate
high frequency noise, thereby producing large swings in output. In order to
avoid this, we replaced the PD controller with a controller of the form C(s) =
sp
KP + KD s+p ; the larger the value of p, the better the controller approximates
P D control, but the more susceptible it is to high frequency noise. The modified
PD controller can be written as

ps KP (s + p) + KD ps
C(s) = KP + KD =
s+p s+p
(KP + KD p)s + KP p
=
s+p
KP p
s+ KP +KD p
= (KP + KD p) .
s+p

KP p
If we let K = KP + KD p and z = KP +KD p , we obtain the dynamic controller

s+z
C(s) = K .
s+p

This controller is called a lead controller (or lead compensator) if z < p and
a lag controller (or lag compensator) if z > p. To see where this terminology
comes from, recall that if we applied a sinusoidal input cos(ωt) to the controller
C(s), the output would be |C(jω)| cos(ωt + ∠C(jω)) in steady state. The phase
of C(jω) is given by ∠C(jω) = ∠(jω + z) − ∠(jω + p), and if z < p, we have
∠C(jω) > 0 (i.e., the output leads the input). On the other hand, if z > p, we

c Shreyas Sundaram
15.1 Lead and Lag Compensators 125

have ∠C(jω) < 0 (i.e., the output lags the input).

As described above, the approximation to the PD controller has a large pole

p, and thus is a lead compensator. As one might expect, a lag compensator is
then an approximation to a PI controller. Specifically, consider a PI controller
of the form
KI
KI KP s + KI s+ K P
C(s) = KP + = = KP .
s s s
Recall that the reason for considering PI control was that it guaranteed perfect
tracking for a step (if the closed loop system is stable). It did this by introducing
a pole at the origin in P (s)C(s), which is the same as saying that it increased
the DC gain of P (s)C(s) to infinity. In general, the higher we can make the
DC gain of P (s)C(s), the better it can track reference inputs. One way to do
this would be to add a PI controller to the loop, but in some circumstances,
we might be interested in using a stable controller (whereas the PI controller by
itself is unstable). Thus, suppose we approximate the PI controller by
s+z
C(s) = K ,
s+p
KI
where K = KP , z = K P
, and p is a small positive number. The smaller we make
p, the better this controller approximates the PI controller. In particular, we
are interested in using this controller to boost the DC gain of P (s)C(s), and so
we would like to make z > p. As we discussed earlier, this is a lag controller.
Lag controllers are frequently used to improve the steady state error after we
have already satisfied the transient specs (potentially with some other con-
troller). The objective is typically to boost the DC gain of the controller, while
trying to keep the poles of the closed loop system from changing too much.
One example of this is lead-lag compensator design: use a lead compensator
to stabilize the system, and then use a lag compensator to boost the tracking
accuracy. We will discuss this later.
To summarize:

• Lead compensation approximates PD control, and is used to stabilize the

system and improve the transient characteristics (by moving the locus to
the left and improving the phase margin).
• Lag compensation approximates PI control, and is used to boost DC gain.

We are now ready to design lead and lag compensators.

c Shreyas Sundaram
126 Compensator Design Using Bode Plots

15.2 Lead Compensator Design

First, consider the standard unity feedback loop:

R(s) E(s) Y (s)

C(s) P (s)
−

s+z
A lead controller will have the form C(s) = Kc s+p , where p > z. Since p > z,
we can write z = αp for some 0 < α < 1. The Bode form of the above controller
is then given by
s s
s + αp αp( αp + 1) αp + 1
C(s) = Kc = Kc s = Kc α s .
s+p p( p + 1) |{z} p +1
K | {z }
Cl (s)

The phase margin of the closed loop system can be obtained by examining the
Bode plot of C(s)P (s), which is obtained by simply adding together the Bode
plots of Cl (s) and KP (s) (since the magnitude is on a log scale, and the phases
inherently add). The gain K of the compensator can be first be chosen to meet
steady state error specifications, or to obtain a certain crossover frequency. Once
that is done, let’s see what the lead compensator contributes to the system by
examining the Bode plot of Cl (s):

We see that the phase plot of Cl (s) has a bump, and we can use this positive
phase contribution to increase the phase of KP (s). Specifically, we would like
to choose α and p so that the bump occurs near the crossover frequency of
KCl (s)P (s), thereby increasing the phase margin of the system. To see how to
choose the pole and zero, note that the phase of Cl (jω) is given by
ω ω
∠Cl (jω) = tan−1 ( ) − tan−1 ( ) .
αp p

c Shreyas Sundaram
15.2 Lead Compensator Design 127

From the phase plot, we note that the maximum phase occurs halfway between
the zero and the pole (on a logarithmic scale). If we denote the frequency where
the maximum phase occurs as ωmax , we have
1 p
log ωmax = (log(αp) + log(p)) = log αp2 ,
2
√
from which we obtain ωmax = αp. If we denote φmax = ∠Cl (jωmax ), we obtain
(after some algebra)
1−α
sin φmax = .
1+α
These expressions are important, so let’s restate them:

The maximum phase of the lead compensator with zero at αp√and pole
at p is denoted by φmax and occurs at the frequency ωmax = αp. The
maximum phase satisfies the equation
1−α 1 − sin φmax
sin φmax = or equivalently, α= .
1+α 1 + sin φmax

The idea will be to choose the pole and zero of the compensator such that ωmax
lies on the crossover frequency of KP (s), with the hope of contributing an extra
φmax degrees of phase margin. Let’s try an example to see how this works.
1
Example. Consider KP (s) = s(s+1) . Draw an approximate Bode plot for
Cl (s)KP (s) when the pole and zero of the compensator are such that the max-
imum compensator phase occurs at the gain crossover frequency of KP (s).
Solution.

c Shreyas Sundaram
128 Compensator Design Using Bode Plots

From the above example, we see that although the compensator does contribute
φmax to the phase at the gain crossover frequency of KP (s), the gain crossover
frequency of Cl (s)KP (s) actually shifts to the right due to the positive magni-
tude contribution of Cl (s). Thus, the phase margin of KCl (s)P (s) is actually
a little less than the phase margin of KP (s) plus φmax . In order to still get our
desired phase margin, we should therefore make φmax a little larger than we need
(usually about 10◦ extra is enough), so that the phase margin of KCl (s)P (s)
will meet the specification.

The complete design procedure for lead compensators is as follows.

Lead Compensator Design

1. Choose K to meet a steady state error specification, or to meet a

gain crossover frequency specification (in the latter case, one can
choose K so that the crossover frequency of KP (s) is a little less
than desired, since the lead compensator will shift the frequency
to the right a little bit).
2. Find the phase margin of KP (s) (from the Bode plot).

3. Find how much extra phase is required in order to meet the phase
margin spec. Set φmax to be this extra phase plus 10◦ .
1−sin φmax
4. Find α = 1+sin φmax .

5. Set ωmax to be the gain crossover frequency

√ of KP (s). From this,
we can calculate p = ω√max
α
and z = αωmax The compensator is
.
given by
s
αp + 1
C(s) = K s .
p +1

6. Check if the compensator achieves the specifications. If not, iter-

ate or add another lead compensator.

Note: The lead compensator also sometimes appears as Clead (s) = Gain1 1+aT s
1+T s .
Comparing this to the lead compensator given above, we have Gain1 = K,
T = p1 and a = α1 .
1
Example. Consider P (s) = s(s+1) . Design a lead compensator so that the
closed loop system has a steady state tracking error of 0.1 to a ramp input, and
overshoot less than 25%.

c Shreyas Sundaram
15.2 Lead Compensator Design 129

Solution.

c Shreyas Sundaram
130 Compensator Design Using Bode Plots

15.3 Lag Compensator Design

Let us now consider lag controller design using the Bode plot method. Recall
that lag controllers are approximations to PI controllers, and are used to boost
the DC gain (in order to improve the steady state tracking error). Here, we will
see how to design lag compensators using the Bode plot method to improve the
phase margin and satisfy steady state tracking specs. The Bode plot analysis
will give us a perspective on why the pole and zero of the lag compensator are
usually chosen to be very small.
s+z
A lag controller will have the form C(s) = Kc s+p , where z > p. Since z > p,
we can write z = βp for some β > 1. The Bode form of the above controller is
then given by
s s
s + βp βp( βp + 1) βp + 1
C(s) = Kc = Kc = K c β .
s+p p( ps + 1) s
p +1
| {z }
Cg (s)

Note that we are interested in the gain-boosting properties of the lag controller,
and so we will group the DC gain β with the dynamics of the controller in the
term Cg (s) (this is in contrast to the lead controller, where we grouped the DC
gain α with the gain Kc ). In this case, we will be using the gain Kc to obtain a

c Shreyas Sundaram
15.3 Lag Compensator Design 131

desired phase margin, and the controller Cg (s) to boost the DC gain. Since the
Bode plot of C(s)P (s) is obtained simply by adding together the Bode plots of
Cg (s) and Kc P (s), let us examine the Bode plot of Cg (s):

Note from the magnitude plot that Cg (s) will add to the magnitude of Kc P (s)
at low frequencies, thereby reducing the steady state tracking error:

Furthermore, we see that the phase plot of Cg (s) has a dip between ω = p and
ω = z, which will reduce the phase of Kc P (s) in that frequency range. This
is generally bad, because a lower phase might lead to a reduced phase margin.
The idea will be to choose the pole and zero very small, so that the dip in phase
will occur at very low frequencies (far away from the gain crossover frequency).

The design procedure for lag compensators can be summarized as follows.

c Shreyas Sundaram
132 Compensator Design Using Bode Plots

Lag Compensator Design

1 Choose Kc to meet the phase margin specification (with about

10◦ of buffer to accommodate the phase lag induced by the lag
controller) by moving the gain crossover frequency to the left.
2 Find the low frequency gain of Kc P (s), and determine how much
extra gain β should be contributed by Cg (s) in order to meet the
tracking specification.
3 Choose the zero z of the compensator to be about one decade
below the gain crossover frequency of Kc P (s).
z
4 Choose the pole p of the compensator as p = β. The compensator
is given by
s
βp + 1
C(s) = Kc β s .
p +1

5 Check if the compensator achieves the specifications. If not, iter-

ate or add another compensator.

1
Example. Consider P (s) = s(s+1) . Design a lag compensator so that the
closed loop system has a steady state tracking error of 0.1 to a ramp input, and
overshoot less than 10%.
Solution.

c Shreyas Sundaram
15.3 Lag Compensator Design 133

c Shreyas Sundaram
134 Compensator Design Using Bode Plots

Note that we can use either a lead compensator or a lag compensator to satisfy
the specs here. The difference is that the lag compensator increases the phase
margin by reducing the gain crossover frequency, whereas the lead compensator
increases the phase margin by adding more phase to the system. Therefore, the
response of the system with the lead compensator will generally be faster than
that of the same system with a lag compensator. However, the lag compen-
sator is capable of boosting the DC gain of the system without substantially
moving the gain crossover frequency or reducing the phase margin. Thus, a lag
compensator is often used in order to improve the tracking characteristics of an
existing controller, without affecting the other performance metrics too much.
The choice of controller will generally depend on the application requirements
and constraints.

c Shreyas Sundaram
Chapter 16

Nyquist Plots

So far, we have studied root locus methods and Bode plot methods for analyzing
the behavior of closed loop systems from the open loop transfer functions. The
root locus allows us to see how the poles of the transfer function change when
we vary a certain parameter, and allows us to visualize the effect of adding
additional poles and zeros. However, the root locus is not capable of handling
delays in the feedback loop (because a delay of τ contributes a term e−sτ to
the transfer function, which does not have a nice zero/pole interpretation).
Furthermore, the root locus cannot handle general uncertainties in the model
(it can, however, tell us something about the locations of the poles when a single
parameter is allowed to change slightly).

Bode plots are able to capture uncertainties and delay, and we have seen how to
use them to design controllers and analyze properties of the closed loop system.
However, up to this point, we have been assuming that the closed loop system
is stable when we put P (s) in the unity feedback loop with a certain controller
gain K. Under this condition, we have seen how to use the Bode plot of the
open loop system KP (s) to determine how much we can boost K before the
closed loop poles cross the imaginary axis. We have also seen how to use Bode
plots to design lead and lag controllers in order to meet certain performance
specifications. We will now study Nyquist plots, which complement Bode
plots to provide us with frequency response techniques to determine the stability
of the closed loop system (i.e., we will not have to assume initial stability, as
we did in the Bode plot analysis). Furthermore, the Nyquist plots will provide
us with an alternative mechanism to evaluate the robustness of the system (via
the gain margin and phase margin). To develop these concepts, we will need
the notion of contours in the complex plane.

Contour: A contour is a piecewise smooth path in the complex plane. The

contour is closed if it starts and ends at the same point. A contour is simple

c Shreyas Sundaram
136 Nyquist Plots

if it doesn’t cross itself at any point.

Suppose we have a system with transfer function

(s + z1 )(s + z2 ) · · · (s + zm )
H(s) = ,
(s + p1 )(s + p2 ) · · · (s + pn )

and we consider a simple closed clockwise contour C in the complex plane. At

each point s̄ on the contour C, H(s̄) is simply some complex number. If we
evaluate H(s) at all points on C, we get a closed (but not necessarily simple)
contour which we will denote by H(C):

Let’s focus on a particular point s̄ on the contour C. The complex number H(s̄)
has a magnitude and a phase; the latter is given by
m
X n
X
∠H(s̄) = ∠(s̄ + zi ) − ∠(s̄ + pi ) .
i=1 i=1

Note that H(s̄) can be represented as a vector from the origin with magnitude
|H(s̄)| and angle ∠H(s̄), as shown in the above figure. We will be interested in
seeing how the phase of H(s̄) changes as the point s̄ moves around the contour
C. To do this, we see from the above expression for ∠H(s̄) that we can examine
how each of the quantities ∠(s̄ + zi ) and ∠(s̄ + pi ) vary as s̄ moves around the

c Shreyas Sundaram
 137

contour C. Suppose the contour C and the distribution of poles and zeros looks
like this:

The quantity ∠(s̄ + zi ) is given by the angle between the vector from −zi to
s̄ and the positive real axis, and the quantity ∠(s̄ + pi ) is given by the angle
between the vector from −pi to s̄ and the positive real axis. Now, consider a
zero −zi that is outside the contour C. As s̄ moves around the contour C and
comes back to its starting point, the vector s̄ + zi swings up and down, but it
does not swing all the way around. As a result, the net change in ∠(s̄ + zi ) is
0◦ . The same analysis holds for a pole outside the contour C.
Now consider a zero −zj inside the contour C. As s̄ moves around C, the vector
s̄ + zj turns all the way around, and the net change in ∠(s̄ + zj ) is therefore
−360◦ . Similarly, if we consider a pole −pj inside C, the net change in ∠(s̄ + pj )
is also −360◦ .
If we put this all together, we see that every zero and pole inside the contour
C induces a net phase change of −360◦ as s̄ moves around C, and every zero
and pole outside the contour C induces a net phase change of 0◦ . Let Z denote
the number of zeros of H(s) inside the contour C, and let P denote the number
of poles of H(s) inside the contour C. From the earlier expression for ∠H(s̄),
we see that ∠H(s̄) undergoes a net change of −(Z − P )360◦ as s̄ moves around
the contour C. Since each net change of −360◦ means that the vector from the
origin to H(s̄) swings clockwise around the origin for one full rotation, a net
change of −(Z − P )360◦ means that the contour H(C) must encircle the origin
in the clockwise direction Z − P times. This leads us to the following principle.

The Principle of the Argument: Let C be a simple closed clockwise

contour, and consider the contour H(C) which is generated by evaluating
a function H(s) on C. The contour H(C) will encircle the origin in a
clockwise direction Z − P times, where Z is the number of zeros and P is
the number of poles of H(s) that are contained inside C.

Note: The reason for calling this the Principle of the Argument is that the phase

c Shreyas Sundaram
138 Nyquist Plots

of a complex number is also sometimes called the argument of the number, and
the above principle is derived by considering how the phase of H(s̄) changes as
s̄ moves around the contour C.
Also note that we assume that the contour C does not pass through any of the
poles or zeros of the transfer function (because the phase contribution of a zero
or pole is undefined if we evaluate the contour at that point). Similarly, the
above argument only applies if the contour H(C) does not pass through the
origin; the number of encirclements of the origin is undefined otherwise.

16.1 Nyquist Plots

We will now use the Principle of the Argument to study the stability of a closed
loop system. Consider a system L(s) placed in the following feedback loop:

R(s) Y (s)
L(s)
−

The transfer function from r to y is given by

L(s)
Try (s) = ,
1 + L(s)
and the closed loop poles of this system are the set of all values s such that
1 + L(s) = 0. Define H(s) = 1 + L(s) (note that H(s) is a transfer function).
Specifically, if we denote L(s) = N (s)
D(s) , we obtain

N (s) D(s) + N (s)

H(s) = 1 + = .
D(s) D(s)
Note that the poles of H(s) are actually the poles of L(s) (which are the open-
loop poles). The zeros of H(s) are the poles of the transfer function Try (s),
and thus they are the closed loop poles of the system. To determine how many
closed loop poles lie in the CRHP, we use the Principle of the Argument. First,
define the contour C as follows:

This contour encloses the entire right half plane. Part C1 contains points of the

c Shreyas Sundaram
16.1 Nyquist Plots 139

form s = jω as ω ranges from 0 to ∞, and part C2 contains points of the form

s = −jω, as ω ranges from ∞ to 0. Parts C1 and C2 form the left boundary
of the right half plane. Part C3 is a semicircle of infinite radius, and can be
described by the points s = ejθ , where  → ∞, and θ starts at 90 degrees and
ends at −90 degrees. Suppose we evaluate H(s) on C to produce the contour
H(C); for this choice of C, H(C) is called the Nyquist Plot of H. If we let N
denote the number of times H(C) encircles the origin clockwise, the Principle
of the Argument tells us that

N =Z −P ,

where

• Z is number of zeros of H(s) enclosed by the contour C (which corresponds

to the number of closed loop poles in the right half plane),

• P is the number of poles of H(s) enclosed by the contour C (corresponding

to the number of open loop poles in the right half plane).

Since P is known (for example, by knowing L(s), or by assuming that L(s) is

stable), and since N can be determined by looking at the contour H(C), we
can figure out Z, and thus determine the stability of the closed loop system.
We will see an example of a Nyquist plot in the next example, and then we will
discuss how to draw these plots by hand.
10
Example. Suppose L(s) = (s+1) 3 . Determine the stability of the closed loop

system by examining the Nyquist plot of H, where H(s) = 1 + L(s), and C is

the contour containing the right half plane.
Solution.

In order to apply the above technique, we needed to draw the Nyquist plot of
H (which is the contour H(C)). We can relate the Nyquist plot of H to the
Nyquist plot of the open loop system L(s) by noting that H(s) = 1 + L(s). The
contour H(C) is thus obtained by shifting the contour L(C) one unit to the

c Shreyas Sundaram
140 Nyquist Plots

right in the complex plane:

Encirclements of the origin by H(C) thus correspond to encirclements of the

point −1 on the real axis by L(C). This means that we can simply focus on
drawing the Nyquist plot of L, and seeing how many times it encircles the −1
point. Next, we will see how to draw the Nyquist plot of L.

16.2 Drawing Nyquist Plots

Consider again the contour C:

The Nyquist plot of L(s) is obtained by combining the contours L(C1 ), L(C2 )
and L(C3 ), where C1 , C2 and C3 are the three portions of the contour C. We
will now examine how to draw each of these contours.
Contour C1
Note that the contour C1 is made up of points of the form s = jω, as ω ranges
from 0 to ∞. Each point on the contour L(C1 ) is then of the form L(jω), which
is just a complex number with magnitude |L(jω)| and phase ∠L(jω). We have
access to these quantities from the Bode plot of L(s), and so we can draw the
contour L(C1 ) by drawing the magnitude and phase plots from the Bode plot
together in the complex plane.

c Shreyas Sundaram
16.2 Drawing Nyquist Plots 141

10
Example. Consider L(s) = (s+1)2 . Draw the contour L(C1 ).
Solution.

10
Example. Consider L(s) = (s+1)3 . Draw the contour L(C1 ).
Solution.

c Shreyas Sundaram
142 Nyquist Plots

Contour C2
Now that we have drawn the contour L(C1 ), let us turn our attention to the
contour L(C2 ). Note that the points on C2 are of the form s = −jω, as ω ranges
from ∞ to 0. The points on the contour L(C2 ) are thus of the form L(−jω),
which is the complex conjugate of L(jω). The magnitude of L(jω) and L(−jω)
are the same, but the phases are negatives of each other. This means that the
contour L(C2 ) is simply a mirrored version of L(C1 ) about the real axis. The
contour L(C2 ) can now be added to the plots in the examples above.

Contour C3
The contour C3 is described by points of the form s = ejθ , where  → ∞,
and θ ranges from 90◦ to −90◦ . The contour L(C3 ) is made up of points of the
form L(ejθ ), and each of these points can be evaluated by substituting ejθ into
L(s). Specifically, note that since  is taken to be very large (infinite, in fact),
this term will dominate every factor that it appears in. Thus, if L(s) is strictly
proper, L(ejθ ) will simply evaluate to zero (and thus L(C3 ) is a single point at
the origin). If L(s) is nonstrictly proper, then L(ejθ ) will be some constant.
This will become clearer by evaluating L(C3 ) for the previous two examples,
and also from the following additional example.
s+1
Example. Draw the Nyquist plot of L(s) = s+10 .
Solution.

c Shreyas Sundaram
16.2 Drawing Nyquist Plots 143

16.2.1 Nyquist Plots For Systems With Poles/Zeros On

The Imaginary Axis
The Principle of the Argument assumes that the contour C does not pass
through any of the zeros or poles of the function H(s) (because the angle con-
tribution of a pole or zero is undefined when we evaluate the contour at that
point). However, there may be cases when we want to draw the Nyquist plot
of a system L(s) that has poles or zeros on the imaginary axis. In this case,
we modify the standard contour C to take a small detour around the poles or
zeros:

For example, the new portion C4 on the contour is described by points of the
form s = ejθ where  → 0, and θ ranges from −90◦ to 90◦ . We can evaluate
L(C4 ) by substituting s = ejθ into L(s), and examining what happens as  → 0
(similarly to what was done for the portion C3 ). A few examples will make this
clear.
1
Example. Consider L(s) = s(s+1) . Draw the Nyquist Plot of L(s).
Solution.

c Shreyas Sundaram
144 Nyquist Plots

1
Example. Consider L(s) = s2 (s+1) . Draw the Nyquist Plot of L(s).
Solution.

c Shreyas Sundaram
16.2 Drawing Nyquist Plots 145

s(s+1)
Example. Consider L(s) = (s+10)2 . Draw the Nyquist Plot of L(s).
Solution.

c Shreyas Sundaram
146 Nyquist Plots

16.3 Stability Margins from Nyquist Plots

Now that we have seen how to establish stability of a closed loop system by
drawing the Nyquist plot of the open loop system L(s), we can return to the
topic of stability margins. Specifically, when we were looking at Bode plots, we
considered gain and phase margins as indicators of how stable the closed loop
system was. Since the Nyquist plot tells us about the stability of the system, it
can also provide us with the gain and phase margins.
To see this, consider the closed loop system:

R(s) Y (s)
K L(s)
−

The gain K simply scales the Nyquist plot (since the magnitude on every point
of the contour KL(C) simply gets multiplied by K). In other words, increasing
K serves to push all of the points on the Nyquist plot further away from the
origin. We will see this from the following example.
Example. Draw the Nyquist plot of KL(s) for K = 1 and K = 200, where
1
L(s) = s(s+1)(s+10) .
Solution.

c Shreyas Sundaram
16.3 Stability Margins from Nyquist Plots 147

Note that the system is stable for small K, but unstable for large K (this can
be confirmed from the root locus of L(s)):

Recall that the gain margin is defined as the factor by which K can be increased
before the closed loop system becomes unstable (we will define the gain margin
only for systems that are closed loop stable for an initial value of K). In the
Nyquist plot, this corresponds to scaling the plot so that the number of encir-
clements of the −1 point changes. Note that this is in complete accordance
with the Bode plot analysis. Specifically, the point −1 in the complex plane
is a complex number with magnitude 1 and phase −180◦ . When we looked
at Bode plots, we showed that imaginary axis crossings occur when the gain
crossover frequency and phase crossover frequency coincide (which corresponds
to the case where KL(jω) = −1).
Similarly, the phase margin is defined as the amount by which the angle of
L(jωcg ) exceeds −180◦ , where ωcg is the point where |L(jωcg )| = 1. In the
Nyquist plot, this can be obtained in the following way. First, draw a line
from the origin to the point where the Nyquist plot crosses a circle of radius
1 centered at the origin. This crossing point corresponds to |L(jω)| = 1. The
phase margin is then the angle between this line and the negative real axis:

Example. Identify the gain margin and phase margin for the example given
1
above (with L(s) = s(s+1)(s+10) and K = 1).

c Shreyas Sundaram
148 Nyquist Plots

Example. Investigate the stability (and associated margins) for the system
KL(s) = 10(s+1)
s(s−1) .
Solution.

c Shreyas Sundaram
16.3 Stability Margins from Nyquist Plots 149

Nyquist Plots With Uncertainty and Delay

One of the benefits of Nyquist plots and Bode plots is that they provide a means
to deal with uncertainty in the system. Most systems are susceptible to high
frequency noise, and it is usually hard to get a good indication of the magnitude
and phase of L(jω) for large values of ω. In this cases, we usually have a region
of uncertainty in the Nyquist plots and Bode plots for high frequencies:

The gain margin and phase margin are clearly useful concepts for dealing with
this issue. Typically, we would like to design the system to have sufficiently
high margins so that the closed loop system is stable even with the worst case
uncertainty.
Another benefit of Nyquist and Bode plots is that they can readily handle delays
in the system. For example, consider the system ÿ(t) + 2ẏ(t) + y(t) = u(t − T ).
In this system, the output at time t is a function of the input at time t − T .
Since the Laplace transform of a delayed signal u(t−T ) is e−sT U (s), the transfer
function for this system is given by

Y (s) e−sT
L(s) = = .
U (s) (s + 1)2

When s = jω, the term e−jωT is a complex number with magnitude 1 and phase
−ωT . This term has the effect of subtracting ωT radians from the phase at each
1
frequency ω on Bode plot of (s+1) 2:

c Shreyas Sundaram
150 Nyquist Plots

The Nyquist plot of L(s) thus looks like this:

Once again, the phase margin and gain margin come in handy, as they give us
an indication of the maximum delay that the system can tolerate before going
unstable. More specifically, note that a large delay can cause the Nyquist plot
to rotate enough that the number of encirclements of −1 changes, indicating
that the closed loop system becomes unstable. In other words, a large delay can
destabilize a feedback loop that is otherwise stable!

c Shreyas Sundaram
Chapter 17

Modern Control Theory:

State Space Models

Up to this point, we have been analyzing and designing control systems by us-
ing a transfer-function approach (which allowed us to conveniently model the
system, and use techniques such as root-locus, Bode plots and Nyquist plots).
These techniques were developed and studied during the first half of the twen-
tieth century in an effort to deal with issues such as noise and bandwidth issues
in communication systems. Transfer function methods have various drawbacks
however, since they cannot deal with nonlinear systems, are not very convenient
when considering systems with multiple inputs and outputs, and are difficult to
use for formulating ‘optimal’ control strategies. Starting in the 1950’s (around
the time of the space race), control engineers and scientists started turning to
state-space models of control systems in order to address some of these is-
sues. These are purely time-domain ordinary differential equation models of
systems, and are able to effectively represent concepts such as the internal state
of the system, and also present a method to introduce optimality conditions into
the controller design procedure. This chapter will provide an introduction to
the state-space approach to control design (sometimes referred to as “modern
control”).

17.1 State-Space Models

State-space models are simply a set of differential equations defining a system,
where the highest derivative in each equation is of order 1. To derive these
models, it is easiest to start with an example. Consider the system that is given
by the differential equation

ÿ + 3ẏ + 2y = 4u .

c Shreyas Sundaram
152 Modern Control Theory: State Space Models

The transfer function for this system can be readily found to be H(s) = YU (s)
(s)
=
4
s2 +3s+2 . To represent this model in state-space form, we first draw an all-
integrator block diagram for this system. Specifically, the all-integrator
block diagram is simply a set of integrator blocks that are chained together
according to the constraints imposed by the system. To obtain the diagram for
this system, we first solve for the highest derivative:

ÿ = −2y − 3ẏ + 4u .

Starting from the highest derivative (ÿ), we need to somehow obtain the lower
derivatives y and ẏ. This can be done by integrating ÿ twice, so we chain together
two integrator blocks. From this, we can easily obtain the all-integrator block
diagram as:

Each integrator block in this diagram can be viewed as representing one of the
internal states of the system. Let us assign a state variable to the output of
each integrator in order to represent the states. In this case, we will use the
state variables x1 and x2 defined as

x1 = y, x2 = ẏ .

Using the all integrator block diagram, we can differentiate each of the state
variables to obtain

ẋ1 = ẏ = x2
ẋ2 = ÿ = −2y − 3ẏ + 4u = −2x1 − 3x2 + 4u
y = x1 .

The above first-order differential equations form the state-space model of

the system. We can represent this model more concisely in matrix-vector form
as
      
ẋ1 0 1 x1 0
= + u
ẋ2 −2 −3 x2 4
| {z } | {z } | {z } |{z}
ẋ A x B
 
  x1
y= 1 0 .
| {z } x2
C | {z }
x

c Shreyas Sundaram
17.1 State-Space Models 153

General linear state-space models are thus given by

ẋ = Ax + Bu
(17.1)
y = Cx .

• The vector x is called the state vector of the system. We will denote
the number of states in the system by n, so that x ∈ Rn . The quantity
n is often called the order of the system. In the above example, we have
n = 2.
• In general, we might have multiple inputs u1 , u2 , .. . , um to the system.
0 In
this case, we can define an input vector u = u1 u2 · · · um (the
notation M0 indicates the transpose of matrix M). In the above example,
m = 1.

• In general, we might have multiple outputs y1 , y2 , . . . , yp . In this case, we

 0
can define the output vector y = y1 y2 · · · yp . Note that each of
these outputs represents a sensor measurement of some of the states of
the system. In the above example, we have p = 1.

• The system matrix A is an n × n matrix representing how the states of

the system affect each other.
• The input matrix B is an n × m matrix representing how the inputs to
the system affect the states. The inputs are applied by m actuators.

• The output matrix C is a p × n matrix representing the portions of the

states that are measured by the outputs. The outputs are provided by p
sensors.

Let us consider another example with multiple inputs and outputs.

Example. Derive the state-space model for the following mass-spring-damper
system. The inputs to the system are the forces F1 and F2 , and we would like
to measure the positions of each of the masses.
Solution.

c Shreyas Sundaram
154 Modern Control Theory: State Space Models

c Shreyas Sundaram
17.1 State-Space Models 155

The states in state-space models often represent physical quantities in the sys-
tem. For example, one common model1 for an F-8 aircraft contains four states:

• V : the horizontal-velocity deviation in feet/second.

• γ: the flight-path angle in radians.

• α: the angle of attack in radians.

• q: the pitch rate in radians/second.

Furthermore, the input to the system is applied via a deflection in the elevator
angle, and is denoted by δe . These quantities are shown visually below:

The state-space model for this system is given by

−1.357 × 10−2
      
V̇ −32.2 −46.3 0 V −0.433
 γ̇  
 = 1.2 × 10−4 0 1.214 0   γ   0.1394 
 +
 α̇   −1.212 × 10−4   α   −0.1394  δe

0 −1.214 1
q̇ 5.7 × 10−4 0 −9.01 −6.696 × 10−1 q −0.1577
| {z } | {z } | {z } | {z }
ẋ A x B
 
0 0 0 1
y= x.
1 0 0 0
| {z }
C

Note that real aircraft dynamics are more complicated than this, and are non-
linear; however they can be approximated by choosing the dominant states of
the system, and linearizing the dynamics. We will look at this in more detail
next.
1 See the paper Linear Regulator Design for Stochastic Systems by a Multiple Time-Scales

Method by Teneketzis and Sandell, IEEE Transactions in Automatic Control, vol 22, no. 4,
Aug 1977, pp. 615-621, for more details.

c Shreyas Sundaram
156 Modern Control Theory: State Space Models

17.2 Nonlinear State-Space Models and Lineariza-

tion
Although we have been focusing on linear systems so far in the course, many
practical systems are nonlinear. Since state-space models are time-domain
representations of systems, they can readily capture nonlinear dynamics.
Example: Pendulum.

While we can represent nonlinear systems using state-space models, we have

much better analysis techniques and tools for linear systems. Thus, nonlinear
systems are frequently approximated by linear systems through a process known
as linearization. For example, suppose that we are interested in the model of
the pendulum when it is close to vertical (i.e., when θ is close to zero). In this
case, note that sin θ ≈ θ for θ ≈ 0:

The second state equation becomes ẋ2 = − gl x1 + ml1 2 Te , and thus the nonlinear
pendulum model can be approximated by the linear model
   
0 1 0
ẋ = x+ Te
− gl 0 1
ml2
 
y= 1 0 x .

c Shreyas Sundaram
17.2 Nonlinear State-Space Models and Linearization 157

Next, we will see how to do this linearization more systematically.

17.2.1 Linearization via Taylor Series

Consider a general nonlinear differential equation of the form

ẋ = f (x),

where x is a real-valued signal and f (·) is some nonlinear function. In order to

linearize general nonlinear systems of this form, we will use the Taylor Series
expansion of functions. Suppose that x̄ is a point such that f (x̄) = 0. In this
case, the point x̄ is called an equilibrium point of the above system, since we
have ẋ = 0 when x = x̄ (i.e., the system reaches an equilibrium at x̄). Recall
that the Taylor Series expansion of f (x) around the point x̄ is given by

1 d2 f  1 d3 f 
  
df  2
f (x) = f (x̄) + (x − x̄) + (x − x̄) + (x − x̄)3 + · · · .
dx x=x̄ 2 dx2 x=x̄ 6 dx3 x=x̄

This can be written as


df 
f (x) = f (x̄) + (x − x̄) + higher order terms.
dx x=x̄
| {z }
a

For x sufficiently close to x̄, these higher order terms will be very close to zero,
and so we can drop them to obtain the approximation

f (x) ≈ f (x̄) + a(x − x̄) .

Since f (x̄) = 0, the nonlinear differential equation ẋ = f (x) can be approxi-

mated near the equilibrium point by

ẋ = a(x − x̄) .

To complete the linearization, we define the perturbation state (also known

as delta state) δx = x − x̄, and using the fact that δ ẋ = ẋ, we obtain the
linearized model
δ ẋ = aδx .
Note that this linear model is valid only near the equilibrium point (how “near”
depends on how nonlinear the function is).

Extension To Functions of Multiple States and Inputs

The extension to functions of multiple states and inputs is very similar to the
above procedure. Suppose the evolution of state xi is given by

ẋi = fi (x1 , x2 , . . . , xn , u1 , u2 , . . . , um ) ,

c Shreyas Sundaram
158 Modern Control Theory: State Space Models

for some general function fi . Suppose that the equilibrium points are given by
x̄1 , x̄2 , . . . , x̄n , ū1 , ū2 , . . . , ūm , so that
fi (x̄1 , x̄2 , . . . , x̄n , ū1 , ū2 , . . . , ūm ) = 0 ∀i ∈ {1, 2, . . . , n} .
Note that the equilibrium point should make all of the functions fi equal to
zero, so that all states in the system stop moving when they reach equilibrium.
The linearization of fi about the equilibrium point is then given by
n  m 
X ∂fi  X ∂fi 
fi (x1 , . . . , xn , u1 , . . . , um ) ≈ (x j − x̄ j ) + (uj − ūj ) .
j=1
∂xj xj =x̄j j=1
∂uj uj =ūj

If we define the delta states and inputs δxj = xj − x̄j (for 1 ≤ j ≤ n) and
δuj = uj − ūj (for 1 ≤ j ≤ m), the linearized dynamics of state xi are given by
n  m 
X ∂fi  X ∂fi 
δ ẋi = δxj + δuj .
j=1
∂xj xj =x̄j j=1
∂uj uj =ūj

Note: Sometimes the “δ” notation is dropped in the linearized equation, with
the implicit understanding that we are working with a linearized system.
Example. Linearize the nonlinear state-space model
ẋ1 = x21 + sin x2 − 1
ẋ2 = −x32 + u
y = x1 + x2
around the equilibrium point x̄1 = 1, x̄2 = 0, ū = 0.
Solution.

c Shreyas Sundaram
17.3 The Transfer Function of a Linear State-Space Model 159

17.3 The Transfer Function of a Linear State-

Space Model
Since we can generally convert nonlinear models to a linear model (in a small
region around the equilibrium point), we will focus on linear state-space models
of the form
ẋ = Ax + Bu, x ∈ Rn , u ∈ R m , y ∈ R p
y = Cx .
for the rest of the chapter. Since this model represents a linear system, we can
ask how the matrices A, B and C relate to the transfer function of the system.
To see this, take the Laplace Transform of the above state space equations:
sX(s) − x(0) = AX(s) + BU(s)
Y(s) = CX(s) .
Note that this includes the initial conditions of all the states. The first equation
can be rearranged to solve for X(s) as follows:
(sI − A)X(s) = x(0) + BU(s) ⇔ X(s) = (sI − A)−1 x(0) + (sI − A)−1 BU(s) .
The term I represents the n × n identity matrix. Substituting this into the
equation for Y(s), we obtain
Y(s) = C(sI − A)−1 x(0) + C(sI − A)−1 BU(s) .

The transfer function of the state-space model ẋ = Ax + Bu, y = Cx

(when x(0) = 0) is

H(s) = C(sI − A)−1 B .

c Shreyas Sundaram
160 Modern Control Theory: State Space Models

Note that H(s) is a p × m matrix, and thus it is a generalization of the transfer

function for standard single-input single-output systems. In fact, it is a matrix
where entry i, j is a transfer function describing how the j–th input affects the
i–th output. When p = 1 and m = 1, we get the transfer function that we
studied in the first part of the course.
Example. Calculate the transfer function for the state space model
   
0 1 0  
ẋ = x+ u, y = 1 0 x .
−2 −3 4 | {z }
| {z } |{z} C
A B

Solution.

Note that the above solution agrees with the transfer function at the beginning
of the section.

17.4 Obtaining the Poles from the State-Space

Model
In the last section, we saw that the transfer function of the linear system with
state-space model
ẋ = Ax + Bu
y = Cx
is given by H(s) = C(sI − A)−1 B. Now, note that inverse of any matrix M is
given by M −1 = det(M
1
) adj(M ), where det(M ) is the determinant of M and

c Shreyas Sundaram
17.4 Obtaining the Poles from the State-Space Model 161

adj(M ) is the adjoint matrix corresponding to M (this is just a matrix whose

entries are polynomials obtained from the entries of M ). This means that the
denominator of every entry in matrix (sI − A)−1 is det(sI − A), and so the poles
of the transfer function H(s) are the values of s for which det(sI − A) = 0.
Next, recall the definition of the eigenvalues and eigenvectors of an n × n square
matrix A. Specifically, a complex number λ and a n × 1 vector v are said to be
an eigenvalue and eigenvector, respectively, of A if they satisfy the equation

Av = λv.

In other words, multiplying A by v produces a scaled version of v, and the

scaling factor is λ. Rearranging this equation, we get (λI − A)v = 0. Recall
that this equation has a nonzero solution for v if and only if the matrix λI−A has
linearly dependent columns, which is equivalent to saying that det(λI − A) = 0.
Thus every possible λ that causes the determinant to be zero is an eigenvalue of
A, and as noted above, these are also the poles of the transfer function of the
linear system.

The poles of the transfer function H(s) = C(sI − A)−1 B are exactly
the eigenvalues of matrix A. In other words, the poles are the values s
that satisfy det(sI − A) = 0.

   
0 1 1
Example. Find the poles of the system with A = , B = ,
  −20 −9 1
C= 2 3 .
Solution.

   
1 1 −2 1
Example. Find the poles of the system with A =  0 −9 3 , B = 1,
  0 0 2 0
C= 1 0 1 .

c Shreyas Sundaram
162 Modern Control Theory: State Space Models

Solution.

17.5 An Overview of Design Approaches for State-

Space Models
There are various properties of systems that one can analyze in the state-space
domain. For the plant given by the state-space equations in (17.1), the following
definitions are commonly used.

• The system (17.1) is said to be controllable if there exists an input signal

u(t) such that the state x(T ) achieves any desired value at some time-step
T.

• The system (17.1) is said to be observable if it is possible to recover the

entire state x(t) by looking at the outputs y(t) of the system over some
period of time.

Controllability is about making the system state (not just the output) behave
how we want it to by applying proper inputs. Observability is about whether we
can determine what the internal state of the system is doing by looking at the
outputs. Both of these concepts play central roles in modern control systems
design.
When it comes to stabilization of systems in the state-space domain, one com-
monly considers the use of linear state feedback. Specifically, suppose that
we have access to the entire state x(t) of the system for all time; this is unreal-
istic, since we only have access to y(t), which measures only a few of the states,
but let’s just assume it for now. Linear state feedback control applies an input
of the form
u(t) = −Kx(t),

c Shreyas Sundaram
17.5 An Overview of Design Approaches for State-Space Models 163

for some matrix K that we will choose. The closed loop system is then

ẋ = Ax + Bu = Ax − BKx = (A − BK)x.

The dynamics of the closed loop system are given by the matrix A − BK;
specifically, as discussed in the previous section, the poles of this system are
given by the eigenvalues of this matrix. Thus, in order to obtain a stable system,
we have to choose K so that all eigenvalues of A − BK are stable (i.e., in the
OLHP). It turns out that it is possible to do this if the system is controllable.
The above feedback mechanism assumed that we have access to the entire state.
Since we only have access to the measurements of a few state variables (provided
by the output y(t)), one strategy would be to try to reconstruct the entire state,
based on the measurements available. This is possible if the system is observable,
in which case one can construct a state-estimator that provides an estimate
of x(t) to be used with the linear state feedback input described above. The
architecture of state feedback control with a state estimator looks like this:

The details of these topics, along with issues such as choosing the inputs op-
timally, dealing with noise, etc., are treated in more advanced undergraduate
and graduate courses. Hopefully this course has piqued your interest in control
systems, and motivated you to learn more about this subject in future courses!

The End.

c Shreyas Sundaram