Introduction to Feedback Control

Christopher Nielsen
Control Systems Group
Department of Electrical & Computer Engineering
Acknowledgments
These notes are for introductory undergraduate courses on control systems at the University of Waterloo. They
are partially based on notes written by Prof. John Thistle, Prof. Daniel Miller, Prof. Daniel E. Davison at
the University of Waterloo and Prof. Bruce Francis, Prof. Manfredi Maggiore, Prof. Raymond Kwong at the
University of Toronto and Prof. Andrew Lewis at Queen’s University. All errors, typographical or otherwise,
are solely mine.

i
Contents

Notation iv

Acronyms and Initialisms v

1 Introduction 1
1.1 Familiar examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 What is control engineering? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Control engineering design cycle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.4 Feedback control in software and computer systems . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.5 Very brief history . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.6 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

2 Mathematical models of systems 12

2.1 General comments on modelling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2 Block diagrams . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.3 Modelling based on first principles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.4 State-space models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.5 Linearization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.6 Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.7 The Laplace transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
2.8 Transfer functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
2.9 Block diagram manipulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
2.10 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
2.A Appendix: Linear functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
2.B Appendix: The impulse and convolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57

3 Linear system theory 60

3.1 Initial-state response . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
3.2 Input-response . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
3.3 Total response . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
3.4 Stability of state-space models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
3.5 Bounded-input bounded-output stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
3.6 Steady-state gain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
3.7 Frequency response . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
3.8 Graphical representations of the frequency response . . . . . . . . . . . . . . . . . . . . . . . . . 76
3.9 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
3.A Appendix: Complex numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
3.B Appendix: Time domain, s-domain and frequency domain . . . . . . . . . . . . . . . . . . . . . . 91

ii
4 First and second order systems 93
4.1 Prototype first order system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
4.2 Prototype second order system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
4.3 General characteristics of a step response . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
4.4 The effect of adding poles and zeros . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
4.5 Dominant poles and zeros . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
4.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113

5 Feedback control theory 114

5.1 Closing the loop . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
5.2 Stability of feedback systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
5.3 The Routh-Hurwitz criterion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
5.4 Steady-state performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
5.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135

6 Root-locus method 137

6.1 Basic root-locus construction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
6.2 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
6.3 Non-standard problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
6.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149

7 PID control 150

7.1 Classical PID controller . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
7.2 Empirical tuning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
7.3 Pole placement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
7.4 Integrator windup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
7.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164

8 Frequency domain methods for stability analysis 166

8.1 Cauchy’s principle of the argument . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
8.2 Nyquist stability criterion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
8.3 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
8.4 Stability margins . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
8.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186

9 Introduction to control design in the frequency domain 187

9.1 Loop shaping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
9.2 Performance specifications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
9.3 Lag compensation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
9.4 Lead compensation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198
9.5 Lead-lag compensation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
9.6 Loop shaping theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
9.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211

Bibliography 213

Index 216

iii
Notation

R[s] Ring of polynomials in the complex variable s with real coefficients. 10

:= Equal by definition. 10
δ(t) Impulse function. 10
+
C Closed right half complex plane. 10
Kgm Gain margin. 181
Klgm Lower gain margin. 181
Kugm Upper gain margin. 180
Im(z) Imaginary part of z ∈ C. 90
N The set of natural numbers, {1, 2, . . .}.
Nk The set {1, 2, . . . , k}.
%OS Percentage overshoot. 105
Φlpm Lower phase margin. 177
Φpm Phase margin. 177
Φupm Upper phase margin. 177
Re (z) Real part of z ∈ C. 90
R The set of real numbers.
R(s) The set of rational functions in s ∈ C with coefficients in R. 48
Tp Time to peak. 107
Tr Rise time. 108
Ts Two percent settling time. 107
adj(·) Adjoint matrix. 71
det(·) Determinant of a square matrix. 67
||·||∞ Least upper bound of a signal. 68
C The set of complex numbers.
C+ Open right half complex plane. 10
C− Open left half complex plane. 10
Z The set of integers.
1(t) Unit step function 10
F (x) Fourier transform of a signal x(t). 45
L (f ) Laplace transform of a signal f (t). 40
ωgc Gain crossover frequency. 176
ωpc Phase crossover frequency. 180
z Complex conjugate of z ∈ C. 90
yss Steady-state value of a signal y(t). 73
g∗u Convolution
√ of signals g(t) and u(t). 64
j −1. 90
jR Imaginary axis. 10
sm Stability margin. 184

iv
Acronyms and Initialisms

B.F.L. Buffer Fill Level. 9

B.I.B.O. Bounded-Input Bounded-Output. 68

C.C.W. counterclockwise. 167

C.R.H.P. Closed Right Half Complex Plane. 10

C.W. clockwise. 165

E.M.F. Electromotive Force. 26

F.T. Fourier Transform. 46

F.V.T. Final-Value Theorem. 73

K.C.L. Kirchhoff’s Current Law. 23

K.V.L. Kirchhoff’s Voltage Law. 23

L.T. Laplace Transform. 7

L.T.I. Linear Time Invariant. 7

O.D.E. Ordinary Differential Equation. 4

O.L.H.P. Open Left Half Complex Plane. 10

O.R.H.P. Open Right Half Complex Plane. 10

P.D.E. Partial Differential Equation. 15

P.I. Proportional-Integral. 9

P.I.D. Proportional-Integral-Derivative. 149

R.E.D. Random Early Detection. 9

R.O.C. Region of Convergence. 40

S.I.S.O. Single-Input Single-Output. 9

T.C.P. Transmission Control Protocol. 8

T.F. Transfer Function. 7

v
Chapter 1

Introduction

This course is about the analysis and design of control systems. Control systems are ubiquitous, essential, and
largely invisible to the general public. Without control systems there could be no manufacturing, no vehicles,
no computers, no regulated environment — in short, no technology. Control systems are what make machines,
in the broadest sense of the term, function as intended [Doyle et al., 1990].

Contents
1.1 Familiar examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 What is control engineering? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Control engineering design cycle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.4 Feedback control in software and computer systems . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.5 Very brief history . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.6 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

1.1 Familiar examples

A familiar example of a control system is the one we all have that regulates our internal body temperature at
37◦ C. As we move from a warm room to the cold outdoors, our body temperature is maintained. Nature has
many other interesting examples of control systems: blood sugar regulation in the human body; the population
of a species, which varies over time, is a result of many dynamic interactions with predators and food supplies.
And human organizations are controlled by regulations and regulatory agencies.
Our homes are full of control systems. There are thermostats to regulate the temperature of the house,
the refrigerator, the dishwasher, the clothes washer and dryer, the oven, and the hot water heater. The level
of water in the toilet tanks and the sump pump is also automatically controlled. Many household controllers
are less obvious. These include the automatic controllers that regulate the flow of fuel to the furnace, those
inside electronic devices in the house and those that ensure that our coffee maker and microwave oven work
properly [Society, 2015].
The modern automobile is also full of automatic controllers. Pollution is reduced, fuel is saved, and comfort
is enhanced by automatic controllers in the car. Stability and control augmentation systems make nearly
every modern vehicle easier for the human to control and safer. Still, in 2012, over thirty thousand Americans
died in car accidents [Association, 2015]. Some estimate that over 90% of fatalities are caused by human
error [Smith, 2015]. Hence the control systems needed for autonomous driving have the potential to save many
lives. Other examples of control systems:

• autofocus mechanism in cameras,

• noise cancelling head phones,

1
 1.2. What is control engineering? 2

• cruise control system in cars,

• auto-pilot on a plane,

• flow control in data networks,

• bioengineering applications, e.g., prosthetics and artificial pancreas [Cobelli et al., 2009], [Bequette, 2005],

• economic modelling and policy development [Seierstad and Sydsaeter, 1986],

• Cyber-physical systems, e.g., smart grid, traffic control systems [Kim and Kumar, 2012],

• operating systems [Hellerstein et al., 2004], [Pothukuchi et al., 2018]

• database privacy control [Ny and Pappas, 2014],

• opinion dynamics in social networks [Anderson and Ye, 2019] [Jia et al., 2015],

• middleware (application servers, database management systems, email servers) [Hellerstein et al., 2004].

• self-adaptive software systems [Filieri et al., 2015], [Patikirikorala et al., 2012].

1.2 What is control engineering?

A system maps an input to an output; for dynamical systems these are functions of time. If we use the
symbol P to denote a system that maps an input signal u to an output signal y, then the relationship is denoted
by y = P (u). Figure 1.1 illustrates this relationship. The direction of the arrows indicate whether a signal is

Figure 1.1: Basic block diagram of a system.

an input or an output of the system P . In control engineering we are given a system called the plant. We are
interested in making the output of the plant behave in a desirable manner in the presence of model uncertainty
and despite uncontrolled influences called disturbances. We change the behaviour of the plant by connecting
it to another system called the controller or compensator1 . The essential picture is given in Figure 1.2.
In Figure 1.2 the reference signal r(t) represents the desired output. The external disturbance is d(t).
The controller affects the behaviour of the plant through the input signal to the plant u(t) which is physically
implemented by actuators. The basic feedback loop of (i) sensing, (ii) computation and (iii) actuation is the
central concept in control.

Example 1.2.1. (Helicopter) In helicopter flight control, the plant is the helicopter itself plus its sensors
and actuators. The controller is implemented as a program running in an on-board computer. The disturbance
might be the force of the wind acting on aircraft. The design of a flight control system for a helicopter requires
first the development of a mathematical model of the helicopter dynamics. This is a very advanced subject,
well beyond the scope of this course. We must content ourselves with much simpler plants. N

1
The term compensation is used for the process of designing a controller, reflecting the fact that one is “compensating” for
deficiencies in the innate properties of the plant. The controller itself is sometimes called a compensator.

Version 1.2.2816, September 1, 2020

3 CHAPTER 1. INTRODUCTION

Figure 1.2: Basic block diagram of a control system.

Example 1.2.2. (Open-Loop Control) Suppose we want the plant output to track the reference signal, i.e.,
we want the difference r(t) − y(t) to approach zero as time passes. One way to do this is to use an open-loop
architecture. In this case, one omits the sensors (saving money) and hence removes the part of the diagram in
Figure 1.2 with dashed lines. The controller reads the reference signal r(t) and uses this information to decide
on the input u(t) to the plant.
The open-loop control design has inherent problems. First, if d(t) is unknown then the disturbance may
cause the plant output to deviate from the reference signal. Another problem is that the plant model used to
decide on the input signal can never be a completely accurate representation of a real physical system. Since
there is no feedback, the controller is oblivious to these issues and will not take any corrective action when the
plant output inevitably deviates from the reference signal. N

A powerful way to solve these problems is to use feedback. In this architecture the plant output is read
by sensors which may themselves be considered a system. The sensor output is then compared to the reference
signal and an error signal e(t) is formed. The error is sent to the controller2 which decides on the corrective
action to be taken.

Example 1.2.3. (Cruise Control) A classic example is the cruise control system in a car. The plant is the
car itself. The output of the plant is the car’s speed. The speed is read by a speedometer whose values are sent
to an embedded processor in the car. The control algorithm running on the processor compares the speedometer
reading to the desired speed set by the driver and then decides whether to increase or decrease the throttle. In
this example disturbances may include the pitch of the road, ice and wind. N

Example 1.2.4. (Automated Shopping Cart) Consider an automated shopping cart that follows a person
around the grocery store while they shop. The cart must maintain a predefined safe distance from the shopper
while not colliding with shelves. In this example the plant is the shopping cart. The actuators are motors
attached to the wheels. The cart is equipped with distance sensors, e.g., a camera, that measure how far the
cart is from the shelves and the shopper. The sensors send the distance information to an embedded computer.
The control algorithm running on the computer compares the distance measurements to the desired distances
and sends appropriate command signals to the motors. The disturbances in this example include other shoppers
and carts. N

1.2.1 Detailed introductory example

A common actuator used in control systems to provide rotary motion is the permanent magnet DC motor.
DC motors are found in robotic manipulators, ground and aerial vehicles and many other applications. Suppose
2
Of course the controller may have access to the signals r(t) and s(t) separately. We’ll generally stick to the simpler case where
the controller reads e(t).

Version 1.2.2816, September 1, 2020

 1.2. What is control engineering? 4

we have a DC motor whose output is the angular velocity ω(t) of its shaft, the input is a voltage u(t), and
there is a disturbance torque T (t) resulting from, for example, an unknown external load being applied to the
shaft of the motor. This external torque is something we cannot alter. In Chapter 2 we’ll derive the governing
differential equation for this system to be
dω 1
= − ω(t) + kE u(t) + kT T (t).
dt τ
Here τ , kE and kT are positive physical constants. This system can be represented by the block diagram in
Figure 1.3a.
12

10

0
0 2 4 6 8 10 12 14 16 18 20
time (s)

(a) Block diagram. (b) kE = 2, u(t) = 3, τ = 2, T (t) = 0.

Figure 1.3: Block diagram of DC motor speed model and open-loop step response.

Suppose that we want the motor shaft to spin at a constant desired speed of ω0 radians per second. This
constant represents our reference signal, i.e., ωref (t) = ω0 . Our first naı̈ve attempt to control this system is to
use an open-loop scheme. We start applying constant voltages to observe the resulting speed of the motor’s
shaft. From this we’ll generate a rule for our control signal u(t). If u(t) = u0 (constant) and there is no
disturbance torque T (t) = 0 then the system model becomes

dω 1
= − ω(t) + kE u0 , t ≥ 0.
dt τ
If we assume, as is reasonable, that the motor isn’t moving when we apply the voltage then ω(0) = 0. The
solution to the Ordinary Differential Equation (O.D.E.) is then
 t

ω(t) = kE u0 τ 1 − e− τ , t ≥ 0.

Figure 1.3b gives a plot of this signal for kE = 2, u0 = 3, τ = 2. From either the above expression for ω(t) or
by running more experiments (with different values of u0 ) on the motor to obtain enough data like Figure 1.3b,
we can figure out that if we want the motor to run at a speed of ω0 radians per second, then we should set the
voltage to
ω0
u(t) = . (1.1)
kE τ
The block diagram for this open-loop strategy is depicted in Figure 1.4. As discussed in Example 1.2.2, this
strategy has problems. First, suppose that the disturbance torque is a non-zero constant T (t) = −T0 . The
ODE becomes
dω 1
= − ω(t) + kE u0 − kT T0 , t ≥ 0.
dt τ
and its solution, again assuming ω(0) = 0, is
 t

ω(t) = τ (kE u0 − kT T0 ) 1 − e− τ , t ≥ 0.

Version 1.2.2816, September 1, 2020

5 CHAPTER 1. INTRODUCTION

Figure 1.4: Open-loop control of a DC motor.

With the open-loop control strategy (1.1) the ultimate speed of the motor is ω0 − τ kT T0 . Figure 1.5a shows
the response using the open-loop control when ω0 = 5, kT = 1 and T0 = 2. We can see that the final speed
(≈ 1 rad/sec) is way below the desired speed of 5. The other problem is that, due to modelling errors, we are
unlikely to have exact numerical values for τ and kE . Figure 1.5b shows the response of the open-loop system
when the actual value of τ is 1.6 instead of our estimated value of 2. Again the motor fails to reach the desired
speed of 5.
1 4

0.9
3.5

0.8
3
0.7

2.5
0.6

0.5 2

0.4
1.5

0.3
1
0.2

0.5
0.1

0 0
0 2 4 6 8 10 12 14 16 18 20 0 2 4 6 8 10 12 14 16 18 20
time (s) time (s)

(a) Response with constant disturbance. (b) Response with modelling error.

Figure 1.5: Open-loop control of a DC motor in the presence of disturbances and modelling error.

As mentioned in our introductory discussion, a powerful way to solve the problems we’ve just observed is
to use feedback in our decision making. To that end, let’s suppose instead that the motor has a tachometer
which measures its shaft speed. The tachometer takes an angular velocity and returns a voltage. We’ll model
the tachometer as a scaling factor kS . The voltage from the tachometer is available for feedback. Our controller
converts this voltage back to the physically meaningful unit of radians per second ωmeas (t) and then compares
the measured speed to the reference3 . This is illustrated in Figure 1.6.
In Figure 1.6 the controller computes the tracking error, i.e., the difference between the desired speed ωref (t)
and the measured speed ωmeas (t). Then it multiplies the error by a constant Kp which is the gain of the
controller. The differential equation governing this system is
dω 1

= − ω(t) + kE Kp ωref (t) − k̄S kS ω(t) + kT T (t).
dt τ
The only independent signals in the closed-loop system are the reference signal and the disturbance torque.
Figure 1.7a shows the response of the closed-loop system, for various values of the gain Kp , when ωref (t) = 5,
k̄S kS = 1, there is no disturbance torque T (t) = 0 and we assume perfect knowledge of the physical constants.
Compared to the open-loop response (cf. Figure 1.3b) we see that the closed-loop system doesn’t achieve
the desired speed but that the response is much faster. We can try to increase the gain of the controller to see
3
Ideally in Figure 1.6 we have k̄S kS = 1 so that ωmeas (t) = ω(t). For this reason the sensor blocks in the feedback path are often
omitted in which case we have a unity feedback system.

Version 1.2.2816, September 1, 2020

 1.3. Control engineering design cycle 6

Figure 1.6: Closed-loop control of a DC motor.

if this helps. Figure 1.7a shows the response for Kp ∈ {5, 10, 15, 20, 25}. As the gain Kp increases the error
decreases and the response becomes faster. High gain is a very important concept in control but we have to be
careful using it. Figure 1.7b shows the applied voltage u(t) corresponding to Kp ∈ {5, 10, 15, 20, 25}. As the
5 0

4.5
-20
4

3.5 -40

3
-60

2.5

-80
2

1.5
-100

-120
0.5

0 -140
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 0 1 2 3 4 5 6 7 8 9 10
time (s) time (s)

(a) Angular velocity. (b) Applied voltage.

Figure 1.7: Effect of increasing gain on closed-loop response of DC motor.

gain increases so does the applied voltage. If the applied voltage is too large we will damage the motor so we
need to be judicious when using high gain4 .
A major benefit of the closed-loop architecture is in handling disturbances and modelling error. Figure 1.8
shows the response of the closed-loop system (cf. Figure 1.5) with T (t) = −2 (Figure 1.8a) and τ = 1.6
(Figure 1.8b). In these scenarios the closed-loop system gets closer to the desired speed of ωref (t) = 5. It’s
actually possible to design a controller so that we exactly track the desired speed. We’ll learn how to do this as
the course progresses. In summary, it is hoped that this simple introductory example helps convince you that
feedback is a good thing.

1.3 Control engineering design cycle

The typical components of a modern computer controlled system are shown in Figure 1.9. The typical control
engineering design cycle involves the following steps:

1. Study the system to be controlled and decide what types of sensors and actuators will be used and where
they will be placed.
4
Another problem with high gain is that it can lead to instability. We’ll see this in detail as the course progresses.

Version 1.2.2816, September 1, 2020

7 CHAPTER 1. INTRODUCTION

5 5

4.5 4.5

4 4

3.5 3.5

3 3

2.5 2.5

2 2

1.5 1.5

1 1

0.5 0.5

0 0
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
time (s) time (s)

(a) Response with constant disturbance. (b) Response with modelling error.

Figure 1.8:4 Closed-loop control of a DC motor in the presence of CHAPTER

disturbances and modelling error.
1. INTRODUCTION

noise external disturbances noise

Output
! Actuators System Sensors !

Process

Clock

D/A Computer A/D Filter

Controller

operator input
Figure 1.3: Components of a computer-controlled system. The upper dashed box represents
Figure 1.9: Components of theaprocess
computer controlled
dynamics, systemthe[Åström
which include and
sensors and Murray,
actuators 2019].toThe
in addition upper dashed box represents the
the dynamical
process dynamics, which include the sensors
system being controlled.and actuators
Noise in addition
and external to can
disturbances theperturb
dynamical systemofbeing
the dynamics the controlled. Noise and
external disturbances canprocess.
perturbThe thecontroller
dynamicsis shown in the
of the lower dashed
process. The box. It consists
controller is ofshown
a filter in
andthe
analog-to-
lower dashed box. It consists
of a filter and analog-to- digital
digital(A/D)
(A/D) and and
digital-to-analog (D/A) converters,
digital-to-analog (D/A) as well as a computer
converters, as wellthat asimplements
a computer that implements the
control algorithm. A systemthe control
clockalgorithm.
controlsAthe system clock controls
operation the controller,
of the operation of the controller, synchronizing
synchronizing the A/D, D/A and computing
the A/D, D/A and computing processes. The operator input is also fed to the computer as an
processes. The operator input is also fed to the computer as an external input.
external input.

2. Model the resulting system to be controlled.

A modern controller senses the operation of a system, compares it against the
• By modeldesiredwe mean a mathematical
behavior, modelactions
computes corrective of thebased
plant.on For a continuous-time
a model of the system’s dynamic system
a model response
is typically
to external inputs and actuates the system to effect the desiredmay
one or more differential equations. Experiments be needed to obtain
change.
numericalThis basicoffeedback
values loop of sensing,
some parameters. Thecomputation
processes of anddeducing
actuationaismodel
the central
fromcon-
experimental data is
cept inidentification.
called system control. The key issues in designing control logic are ensuring that the
dynamics of the closed loop system are stable (bounded disturbances give bounded
3. Simplify the model errors)ifand
necessary
that theysohave
that it is tractable.
additional desired behavior (good disturbance attenua-
tion, fast responsiveness to changes in operating point, etc). These properties are
• Classical established
control, theusingsubject of this course, uses a transfer function to model the plant. Linear Time
a variety of modeling and analysis techniques that capture the
Invariantessential
(L.T.I.)dynamics
systems,ofand only LTI
the system and systems,
permit thehave transfer
exploration functions.
of possible The Transfer Function
behaviors
(T.F.) of inanthe LTI system is defined to be the ratio of the
presence of uncertainty, noise and component failure. Laplace Transform (L.T.) of the output to
the Laplace A transform of the input where the LTs are taken with zero initial conditions 5.
typical example of a control system is shown in Figure 1.3. The basic elements
of sensing, computation and actuation are clearly seen. In modern control systems,
4. Analyze the resulting model; determine its properties.
computation is typically implemented on a digital computer, requiring the use of
5 analog-to-digital
If the plant is nonlinear, (A/D)
which virtually andsystem
every digital-to-analog (D/A)
is, then we first converters.
need to linearizeUncertainty
its model toenters
obtain a transfer function.
the system through noise in sensing and actuation subsystems, external disturbances
that affect1,the
Version 1.2.2816, September underlying system operation and uncertain dynamics in the system
2020
(parameter errors, unmodeled effects, etc). The algorithm that computes the control
 1.4. Feedback control in software and computer systems 8

5. Decide on performance specifications.

6. Decide on the type of controller to be used.

7. Design a controller to meet the specs, if possible; if not, modify the specs or generalize the type of controller
sought.

• In this course the controller itself is a transfer function. It corresponds to an ODE that relates the
controller output (input signal to plant) to the controller input (usually tracking error).

8. Simulate the resulting controlled system, either on a computer or in a pilot plant.

9. Repeat from step 1 if necessary.

10. Choose hardware and software and implement the controller.

11. Tune the controller on-line if necessary.

1.4 Feedback control in software and computer systems

The tools of this course find applications in networking [Hellerstein et al., 2004, Keshav, 2012, Åström and Murray, 2019]6
and even algorithm design [Lessard et al., 2016] as the next few examples illustrate.

Load control of a web server

The following example is borrowed from [Keshav, 2012, Example 8.2]. Consider a web server that responds to
GET and POST queries from web browsers. The server uses a buffer to hold pending requests before they are
served so that requests arriving to a busy server are not lost. Requests arrive at different time instants and are
served after spending some time in the buffer.
The difference between arrivals and departures at any momment in time represents the number of queries
in the buffer (queue length) at that time. Denote this quantity by y(t). The control objectives are

(i) Don’t let queue length get too big thereby exceeding the buffer capacity causing requests to be lost.

(ii) Don’t let queue length get to zero – causing a buffer underflow and idling the web server.

Taken together, these specifications suggest that we would like to keep the number of queries in the buffer
y(t) equal to some reference value r(t) that is enough to not greatly impact the server response time but large
enough that if the web server were to suddenly obtain additional capacity, the buffer would not underflow.
The problem here is that the rate at which the server handles requests is unknown – it depends on various
things including the number of clients. We can model the uncertain service rate as a disturbance d(t). The
controller’s goal is therefore to decide on the request rate u(t) so that the number of bits in the buffer y(t)
remains close to r(t) despite the disturbance d(t). The only information available to the controller is r(t) and
y(t).

Detecting router overloads

The following example is borrowed from [Hellerstein et al., 2004, §1.6.4]. A central element of the Internet
is Transmission Control Protocol (T.C.P.) which provides end-to-end communication across network nodes.
The designers of TCP were concerned about regulating traffic flows in the presence of network congestion. One
way in which this regulation occurs is at routers that direct packets between endpoints. In particular, routers
6
The references [Keshav, 2012, Hellerstein et al., 2004] contain various other computing examples.

Version 1.2.2816, September 1, 2020

9 CHAPTER 1. INTRODUCTION

have finite-size buffers. Thus, to prevent buffer overflows during congestion, routers may discard packets (which
results in their later re-transmission as part of the TCP protocol).
Unfortunately, by the time that buffer overflows occur, it may be that the network is already congested.
The idea behind Random Early Detection (R.E.D.) is to take action before congestion becomes severe. RED
measures how much of critical buffers are consumed, a metric that is referred to as the Buffer Fill Level (B.F.L.).
As depicted in Figure 1.10, RED introduces the capability of randomly dropping packets even if buffer capacity
is not exceeded. If BFL is small, no packets are dropped. However, as BFL grows larger, progressively more
packets are dropped.

Figure 1.10: Router operation under random early detection. Incoming packets are dropped based on an adjustable drop
probability.

RED has been shown to reduce network congestion and to improve network throughput. However, one
challenge with using RED in practice is specifying its configuration parameters, especially the BFL at which
packets start being dropped and the maximum drop probability. One approach to tuning these parameters
has been to view RED as a feedback control system. Figure 1.11 depicts this perspective. Control theory is
used to study the impact on stability and other properties for different settings of RED configuration parame-
ters. Proportional-Integral (P.I.) control has been used for management of the accept queues of network routers.

Figure 1.11: Feedback control of buffer fill level in a router. The reference input is the desired BFL, and the control input
is the package drop probability.

1.5 Very brief history

1. Ancient control (pre-1940): ad hoc control, almost no analysis tools. Examples: water clock (2nd century
BCE). More recently, in 1769 a feedback control system was invented by James Watt: the flyball governor
for speed control of a steam engine. Stability of this control system was studied by James Clerk Maxwell.

2. Classical control (1940’s-1950’s): this constitutes the main topics covered in this course. Deals mostly
with Single-Input Single-Output (S.I.S.O.) linear systems. The pioneering work of Bode, Nyquist, Nichols,
Evans and others appeared at this time. This resulted in simple graphical tools for analyzing single-input
single-output feedback control problems which are now generally known by the generic term Classical
Control Theory. Specifications are based on closed-loop gain, bandwidth, and stability margin. Design
is done using Bode plots. Controllers are typically tuned by hand. Examples: anti-aircraft guns (WWII),
chemical plants, nuclear reactors, cruise control in cars, motor control.

Version 1.2.2816, September 1, 2020

 1.6. Notation 10

3. Modern control (1960’s - 1970’s): The approach to control analysis and design, covered in ECE488, is
in the time domain and uses state-space models instead of transfer functions. This period followed the
work of Wiener, Kalman (and others) on optimal estimation and control. Specifications may be based on
closed-loop eigenvalues, that is, closed-loop poles. This approach is known as the “state-space approach.”

4. Robust and adaptive control (1980’s): frequency domain tools and thinking but state-space computational
methods; robust control including the effect of model error on the performance of feedback controllers;
optimal control (where the controller is designed by minimizing a mathematical function). In this context
classical control extends to H∞ optimization and state-space control extends to Linear-quadratic Gaus-
sian (LQG) control. Discrete-event systems (DES). Examples: advanced aircrafts, large power systems,
distributed control.

5. “Post modern” control (1990’s - present): Striking developments in the control of nonlinear systems;
hybrid systems; distributed robotics; sensor networks. Examples: unmanned vehicles, social networks
and dynamics, cooperative control, advanced robotics including humanoids, medical applications, control
over networks.

1.6 Notation
Generally, signals are written lower case: e.g., x(t). Their transforms are capitalized: X(s) or X(jω). The
impulse is δ(t). In signals and systems the unit step is denoted u(t), but in control u(t) denotes a plant input
so we’ll denote the unit step by 1(t). We will use the following notation for parts of the complex plane

C− := {s ∈ C : Re (s) < 0} (Open Left Half Complex Plane (O.L.H.P.))

C+ := {s ∈ C : Re (s) > 0} (Open Right Half Complex Plane (O.R.H.P.))
jR := {s ∈ C : Re (s) = 0} (Imaginary axis)
+
C := {s ∈ C : Re (s) ≥ 0} (Closed Right Half Complex Plane (C.R.H.P.)).

We denote by R[s] the ring of polynomials in the (complex) variable s with coefficients in R. The symbol :=
means equal by definition. It is convenient to write vectors sometimes as column vectors and sometimes as
n-tuples, i.e., ordered lists. For example
 
x1
x= , x = (x1 , x2 ).
x2

We’ll use both.

1.7 Summary
In this chapter we introduced the essential idea of feedback control. The key points that you should understand
are the following.
1. Feedback control is used in all sorts of engineered systems that we use everyday and is found in many
natural phenomena.

2. Control engineering is about making the output of a system (the plant) behave in a desirable manner
despite disturbances acting on the system and despite modelling uncertainty. The essential picture is
given in Figure 1.2 and you should understand what all the signals and systems in the figure represent.

3. In Section 1.2.1 we studied the problem of controlling the speed of a DC motor. We used this example to
introduce the deficiencies of open-loop control and to motivate the use of feedback.

Version 1.2.2816, September 1, 2020

11 CHAPTER 1. INTRODUCTION

4. The “big picture” of the control engineering design cycle was presented at a high level in Section 1.3. The
remainder of this course is spent filling in the details of this process. The components of a moden control
system were shown in Figure 1.9.

5. Examples of feedback control applied to problems in computer hardware and software were presented.
These examples demonstrate the power of the ideas covered in this course and the breadth of their
application areas.

6. A brief history of the subject and notation used in this course were presented.

Version 1.2.2816, September 1, 2020

Chapter 2

Mathematical models of systems

Before doing any controller design, we must first find a good mathematical model of the system we want to
control. A good model is simple but accurate. The system model serves as the basis on which a controller is
designed. Modelling is an important part of the control engineering design process discussed in Chapter 1.
In principle, there are two different ways in which mathematical models can be obtained: from prior knowl-
edge — e.g., in terms of physical laws — or by experimentation on a process. When attempting to obtain a
specific model, it is often necessary and beneficial to combine both approaches. In this chapter we focus on the
first approach to modelling, i.e., modelling based on physical principles.

Contents
2.1 General comments on modelling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2 Block diagrams . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.3 Modelling based on first principles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.4 State-space models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.5 Linearization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.6 Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.7 The Laplace transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
2.8 Transfer functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
2.9 Block diagram manipulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
2.10 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
2.A Appendix: Linear functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
2.B Appendix: The impulse and convolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57

2.1 General comments on modelling

For some problems, it is both possible and feasible to find precise controller settings by simple trial and error.
However, many problems preclude this approach due to complexity, efficiency, cost or even danger. Also, a trial
and error approach cannot answer, before trial, questions such as:

• Given a plant and control objective, what controller can achieve the objective? Can it be achieved at all?

• Given plant and controller, how will they perform in closed-loop?

• Why is a particular loop behaving the way it is? Can it be done better? If so, by which controller?

• How would the loop performance change if the system parameters were to change, or disturbances were
larger, or a sensor were to fail?

12
13 CHAPTER 2. MATHEMATICAL MODELS OF SYSTEMS

To answer these questions systematically, we need a means to capture the behaviour of the system in such
a way that it can be manipulated outside the constraints of physical reality. An appropriate means to achieve
this goal is to use a mathematical model [Goodwin et al., 2001]. A mathematical model is a set of equations
that describe how a system transforms input signals into output signals.
There is a trade-off between model complexity and model accuracy. Typically, increasing the accuracy of
a model also increases its complexity. The goal is to develop a model that is adequate for the problem at
hand without being overly complex. Models are used for both control design and simulation. The models used
for control design are usually simpler than simulation models. A model, no matter how detailed, is never a
completely accurate representation of a real physical system.
One may observe that a general method of representing a system is by using a table of all possible inputs to
the system and the possible responses (outputs) for these inputs. However, only very rarely will such a crude
technique be useful. One major reason for paying so much attention to linear systems is that for such systems
the table of input-output pairs can be drastically abbreviated [Kailath, 1980].

2.1.1 Examples of mathematical models

Example 2.1.1. (Mass-Spring-Damper) A common class of mathematical models for dynamical systems1 is
ordinary differential equations. In mechanics, one of the simplest differential equations is that of a mass-spring
system with damping
M q̈ + c(q̇) + Kq = u. (2.1)
Figure 2.1 illustrates this system. The variable q(t) ∈ R represents the position of a box with mass M with

Figure 2.1: Mass-spring-damper system. The position of the mass M is q with q = 0 corresponding to the rest position of
the spring. The forces acting on M are due to a linear spring with constant k, a damper whose force depends on velocity
q̇ and an applied force u.

q = 0 corresponding to the position of the mass when the spring is at rest. The notation q̇ represents the time
derivative of q, i.e., the velocity of the mass. The symbol q̈ represents the second derivative with respect to
time, i.e., the acceleration. The spring is assumed to be linear, i.e., it satisfies Hooke’s law, which says that
the force exerted by the spring is proportional to the displacement of the mass. The friction element, called a
damper, is taken as a possibly nonlinear function of the velocity q̇. It can model effects such as viscous drag or
static friction. The force u is an externally applied force which we treat as the input. This system is second
order because the highest derivative appearing in the differential equation (2.1) is the second derivative of q.
If the damper’s force is a nonlinear function of q̇, then the differential equation (2.1) is nonlinear. If the
damper function is a linear function of q̇, i.e., it has the form c(q̇) = bq̇ for some real constant b, then the
differential equation is linear. N

Example 2.1.2. (Simplified Atmospheric Model) Consider a rectangular slice of air heated from below
and cooled from above with it sides kept at constant temperatures. This is our atmosphere in its simplest
1
Roughly speaking a dynamical system is one in which the effects of actions do not occur immediately. For example the speed
of a car does not change instantaneously when the gas pedal is pushed.

Version 1.2.2816, September 1, 2020

 2.1. General comments on modelling 14

description. The bottom is heated by the earth and the top is cooled by the void of outer space. Within
this slice, warm air rises and cool air sinks. In the model, as in the atmosphere, convection cells develop,
transferring heat from bottom to top. The state of the atmosphere in this model can be completely described
by three variables, namely the convective flow x, the horizontal temperature distribution y, and the vertical
temperature distribution z; by three parameters, namely the ratio of viscosity to thermal conductivity σ, the
temperature difference between the top and bottom of the slice ρ, and the width to height ratio of the slice β,
and by three differential equations describing the appropriate laws of fluid dynamics

ẋ = σ(y − x), ẏ = ρx − y − xz, ż = xy − βz.

These equations were introduced by E.N. Lorenz [Lorenz, 1963], to model the strange behaviour of the atmo-
sphere and to justify why weather forecast can be erroneous. The Lorenz equations are still at the basis of
modern weather forecast algorithms. In these equations there is no external influence, i.e., no input and the
resulting system is said to be autonomous. It is also nonlinear because of the terms xz and xy in, respectively,
the ẏ and ż equations. N

Example 2.1.3. (Google Page Rank) The initial success of Google can be attributed in large part to the
efficiency of its search algorithm which, linked with a good hardware architecture, creates an excellent search
engine. Originally, a key part of the search engine was PageRank™, a system for ranking web pages developed
by Google’s founders Larry Page and Sergey Brin at Stanford University.
The main idea of the algorithm is the following. The web can be represented as an oriented (and sparse)
graph in which the nodes are the web pages and the oriented paths between nodes are the hyperlinks. The
basic idea of PageRank is to walk randomly on the graph assigning to each node a vote proportional to the
frequency of return to the node. If xi [k] denotes the vote of website i at time k, the updated vote is given by
X xj [k]
xi [k + 1] = .
nj
j∈N (i)

Here nj is the number of nodes (websites) connected to node (website) j and N (i) is the index set of all
the nodes that link to node i. If we let N ≈ 4 × 1010 denote the total number of webpages on the internet,
then i ∈ {1, . . . , N }. The graph, representing the internet, is not strongly connected, therefore to improve the
algorithm for computing the vote, one considers a random jump, with probability p (typically 0.15) to another
node (another page). As a result

X xj [k] N
X xj [k]
xi [k + 1] = (1 − p) +p .
nj N
j∈N (i) j=1

If we collect the variables xi as a vector x ∈ RN we can express the equations for the vote in the form

x[k + 1] = Ax[k]

for an N × N matrix A of real constants. This is a linear, time-invariant difference equation. One can show,
using the tools of ECE481, that
lim x[k] = x̄ (a constant).
k→∞

where all the elements x̄i of the constant vector x̄ are non-negative and bounded. The vector x̄ (after a certain
normalization) is the Google Page Rank and x̄i is the rank of page i. N

Example 2.1.4. (Heated Rod) Consider a thin rod of length L made of a homogeneous material, see Fig-
ure 2.2. Assume its sides are thermally insulated, while its ends are kept at constant temperatures T1 and T2 .
Let q0 : [0, L] → R be a differentiable function representing the temperature distribution in the rod at time 0,

Version 1.2.2816, September 1, 2020

15 CHAPTER 2. MATHEMATICAL MODELS OF SYSTEMS

Figure 2.2: Heated Rod. The rod has length L, is insulated except for its ends which are kept at constant temperatures T1
and T2 .

i.e, q0 (x) is the temperature of the rod at the point x ∈ [0, L] and at time t = 0. Then, the evolution of the
temperature as time progresses is described by the Partial Differential Equation (P.D.E.)

∂q ∂2q
= k 2, k > 0.
∂t ∂x
This is the one-dimensional heat equation. The unique solution to this PDE is the differentiable function
q(t, x) : R × [0, L] → R such that, for all x ∈ [0, L] and all t ∈ R,

q(0, x) = q0 (x), q(t, 0) = T1 , q(t, L) = T2 .

If the temperature T1 at one end of the rod is allowed to be a variable input that can be freely assigned, then
the heat propagation process turns into a control system. N
Although these examples are quite different, they share important fundamental properties. They all describe
a dynamical system, i.e., a process that evolves with time. They illustrate the power of mathematical models to
help us understand simple systems (like the mass-spring-damper) to very complex systems (like the atmosphere).
In obtaining the mathematical model for a given physical system, we make use of a storehouse of known
physical cause-and-effect relations for individual elements–a storehouse built up through innumerable physical
experiments, performed (sometimes at great expense) by many research teams. It is unrealistic for an engineer
to be an expert on all classes of system models, this requires domain specific knowledge. Instead, the aim of
this chapter is to give an introduction to basic modelling and linearization.

2.2 Block diagrams

The importance of block diagrams in control engineering can’t be overemphasized. One could easily argue that
you don’t understand your system until you have a block diagram of it. We shall take the point of view that
a block diagram is a picture of a function. We can draw a picture of the function y = f (x) like in Figure 2.3.
Thus a box represents a function and the arrows represent variables; the input is the independent variable, the

Figure 2.3: A block diagram is a picture of a function.

output the dependent variable.

Example 2.2.1. (Simple Cart) The simplest vehicle to control is a cart on wheels illustrated in Figure 2.4.
Figure 2.4a is a schematic diagram, not a block diagram, because it doesn’t say which of u, y causes the
other.
Assume the cart has mass M and can only move in a straight line on a flat surface. Assume a force u is
applied to the cart and let y denote the position of the cart measured from a stationary reference position.
Then u and y are functions of time t and we could indicate this by u(t) and y(t). We regard the functions u
and y as signals.

Version 1.2.2816, September 1, 2020

 2.2. Block diagrams 16

(a) Schematic diagram. (b) Block diagram.

Figure 2.4: Schematic and block diagrams of a simple cart.

Newton’s second law tells us that there’s a mathematical relationship between u and y, namely, M ÿ = u.
We take the viewpoint that the force can be applied independently of anything else, that is, it’s an input. Then
y is an output. We represent this graphically by the block diagram in Figure 2.4b.
Suppose the cart starts at rest at the origin at time 0, i.e., y(0) = ẏ(0) = 0. Then the position depends only
on the force applied. However y at time t depends on u not just at time t, but on past times as well. So we
can write y = f (u), i.e., y is a function of u, but we can’t write y(t) = f (u(t)) because the position at time t
doesn’t depend only on the force at that same time t. N

We may need to allow a block to have more than one input as in Figure 2.5. This means that y is a function
of u and v, y = f (u, v). Block diagrams may also have summing junctions as in Figure 2.6.

Figure 2.5: Block with multiple inputs.

(a) Stands for y = u + v (b) Stands for y = u − v.

Figure 2.6: Summing junctions.

Example 2.2.2. (Ball and Beam) This example concerns a beam balanced on a fulcrum as shown in Fig-
ure 2.7. This system is used in the lab for ECE481. Suppose a torque τ is applied to the beam. Let θ denote
the angle of tilt and d the distance of roll. Then both θ and d are functions of τ . The block diagram could be

valve

Figure 2.7: Ball and Beam. A ball sits on a beam that can be rotated by a torque applied at the centre of the beam. The
position of the ball is denoted by d, the angle of the beam is θ and the applied torque is τ .

Version 1.2.2816, September 1, 2020

17 CHAPTER 2. MATHEMATICAL MODELS OF SYSTEMS

as in Figure 2.8a or, as in Figure 2.8b, or initial conditions, if they are not fixed, could be modelled as inputs
too. N

(a) Block diagram. (b) Alternate block diagram.

Figure 2.8: Block diagrams for ball and beam system.

2.3 Modelling based on first principles

The systems we model in this course are simple and are generally governed by linear differential equations. In
this section we review examples of devices whose behaviour is reasonably well-modelled by such equations.

2.3.1 Mechanical devices

Consider a mass M restricted to move in a single direction. Let q ∈ R denote its position. Then Newton’s
equations give X
M q̈ = applied forces
X (2.2)
= Fi .
The same sort of thing happens with rotary devices like the one in Figure 2.9. Let J denote the moment of

Figure 2.9: Rotor on a shaft. A rotor or disk is attached to a shaft. The rotation of the shaft causes the rotor to spin.
The position of the disk is its angle θ with respect to some fixed reference. Its velocity is θ̇ and its acceleration is θ̈.

inertia of the rotor about its axis of rotation. Let θ denote the rotor’s angular position. Then the governing
equation for this system is X
J θ̈ = applied torques
X (2.3)
= τi .

The forces and torques2 that appear on the right-hand side of the above summations are usually modelled by
various types of simple components. These simple components form the building blocks of mechanical models.
Next we review the equations for them.
2
A torque is sometimes called a moment of force or moment for short.

Version 1.2.2816, September 1, 2020

 2.3. Modelling based on first principles 18

Linear springs
For linear translational springs the force due to the spring is given by

Fk (q) = kq, k > 0, (2.4)

Figure 2.10: Mass-Spring: M q̈ = −Fk (q) = −kq.

where k is the spring constant. The torque due to a rotational or torsion spring obeys a rotational version of
Hooke’s law. It is often used to model flexible shafts and its equation is

τk (θ) = kθ, k > 0. (2.5)

Figure 2.11: Torsion-Spring: J θ̈ = −τk (θ) = −kθ.

Linear dampers
A linear damper is usually used to model viscous friction. For translational motion the equation is

Fb (q̇) = bq̇ b > 0, (2.6)

Figure 2.12: Mass-Damper: M q̈ = −Fb (q) = −bq̇.

where b is the coefficient of damping. The torque due to a rotational or inertia damper is given by

τb (θ) = bθ̇, b > 0. (2.7)

Figure 2.13: Inertia-Damper: J θ̈ = −τb (θ) = −bθ̇.

We follow the procedure from [Cannon Jr., 2003] for deriving the equations of motions for mechanical
systems.

Version 1.2.2816, September 1, 2020

19 CHAPTER 2. MATHEMATICAL MODELS OF SYSTEMS

1. Define a reference frame from which to measure distance.

2. Choose a set of coordinates that determine the configuration of the system.

3. Separate the system into its mechanical components. Each component should be either a single point
mass or a single rigid body.

4. For each component determine all external forces and moments (torques) acting on it.

5. For each component, express the position of the centre of mass in terms of the chosen coordinates.

6. Apply Newton’s equations for translational motion to each mechanical component.

7. Apply Newton’s equations for rotational motion to each mechanical component. For each component, the
sum of moments about a point that is either (a) the centre of mass or (b) in the component but stationary,
equals the moment of inertia of the component about that point multiplied by the angular acceleration.

We follow this procedure in the next example though it is too simple to be really representative.

Example 2.3.1. (Car Suspension) Find the governing differential equations for the simplified quarter model
of a car suspension system shown in Figure 2.14. Here M1 represents 1/4 of the mass of the car chassis while
M2 represents half the axle mass.

Figure 2.14: Simplified model of a car suspension system.

1. In this example the reference frame is already suggested by Figure 2.14. From the figure we see that up
is taken as the positive direction.

2. The figure is already suggesting the coordinates q1 , q2 , q3 so we’ll use them.

• We assume that the origin of our coordinates, i.e., the zero position (q1 , q2 , q3 ) = (0, 0, 0), is the
position at which all the springs are balanced by the force gravity. In this way we can ignore gravity.
We can do this when we have linear springs.
• The variable q1 represents how far M1 has moved from its rest position. The variable q2 represents
how far M2 has moved from its rest position. The variable q3 represents the distance from the road
to M2 . It is an exogenous signal, i.e., it comes from the “outside world.” We can view it as a
disturbance to the suspension system.

3. In this example there is no rotational motion. We have two masses M1 and M2 which we treat as point
masses.

Version 1.2.2816, September 1, 2020

 2.3. Modelling based on first principles 20

(a) Mass 1. (b) Mass 2.

Figure 2.15: Free body diagrams for Example 2.3.1.

4. Free body diagrams are very useful to accomplish this step. Figure 2.15 shows the free body diagrams for
both masses.

5. Since this example is so simple, we already did this when we chose the set of coordinates in step 2.

6. Apply Newton’s equations for translational motion to each mechanical component.

• For mass 1 we have X

M1 q̈1 = applied forces
M1 q̈1 = −k1 (q1 − q2 ) − b1 (q̇1 − q̇2 )
⇒ M1 q̈1 + b1 q̇1 + k1 q1 = b1 q̇2 + k1 q2 .
• For mass 2 we have X
M2 q̈2 = applied forces
M2 q̈2 = k1 (q1 − q2 ) + b1 (q̇1 − q̇2 ) − k2 (q2 − q3 )
⇒ M2 q̈2 + b1 q̇2 + (k1 + k2 )q2 = k1 q1 + b1 q̇1 + k2 q3 .
7. There is no rotational motion in this example so we can skip this step.

In summary, the differential equations governing this mechanical system are

M1 q̈1 + b1 q̇1 + k1 q1 = b1 q̇2 + k1 q2

M2 q̈2 + b1 q̇2 + (k1 + k2 )q2 = k1 q1 + b1 q̇1 + k2 q3 .

The next example follows the same procedure for a system with both translational and rotational motion.

Example 2.3.2. (Anti-lock Brake System) An Anti-lock Braking System (ABS) can greatly improve the
safety of a vehicle in extreme circumstances, as it maximizes the longitudinal tire-road friction while maintaining
large lateral forces, which guarantee vehicle steerability. We’d like to obtain a model for a rolling wheel on a flat
road as shown in Figure 2.16. In this figure r is the wheel’s radius, the term τb represents the braking torque,
i.e., the input signal, and F (q̇) is the force due to friction with the road. We follow the same general procedure
as in the previous example.

1. Define a reference frame from which to measure distance.

• We take the left direction as positive translational motion and the counterclockwise rotation of the
wheel as the positive direction of rotation. This choice is already suggested by Figure 2.16.

2. Choose a set of coordinates that determine the configuration of the system.

Version 1.2.2816, September 1, 2020

21 CHAPTER 2. MATHEMATICAL MODELS OF SYSTEMS

Figure 2.16: A rolling wheel.

• Let q denote the position of the contact point of the wheel and let θ denote the angle that a point
on the wheel makes with the road.

3. Separate the system into its mechanical components.

• In this example there is one rigid body, the wheel.

• Let M denote the mass of the wheel.
• Let J denote the moment of inertia of the wheel about its axis of rotation.

4. For each component determine all external forces and moments (torques) acting on it.

• Figure 2.16 already shows the relevant forces.

• For translational motion, the only force is F (q̇) pointing in the negative q direction.
• For rotational motion, the torques are rF (q̇) in the positive θ direction and the braking torque τb in
the negative θ direction.

5. For each component, express the position of the centre of mass in terms of the chosen coordinates.

• Since this example is so simple, we already did this when we chose the set of coordinates. The centre
of mass has position q. We ignore the vertical coordinate since we are assuming a flat road surface.

6. Apply Newton’s equations for translational motion to each mechanical component.

• We have
M q̈ = −F (q̇)

7. Apply Newton’s equations for rotational motion to each mechanical component.

• In this case we have

J θ̈ = rF (q̇) − τb .

In summary, the differential equations governing this mechanical system are

M q̈ = −F (q̇)
J θ̈ = rF (q̇) − τb .

Version 1.2.2816, September 1, 2020

 2.3. Modelling based on first principles 22

If we model the road friction as a linear damper, then F (q̇) = bq̇. When simulating an ABS system, the friction
model is more sophisticated than just a linear damper. However, for the purpose of controller design, the linear
model of friction is a reasonable starting point. N

Example 2.3.3. (Physical Pendulum) Consider the physical pendulum shown in Figure 2.17a. It consists
of a rod made of a homogenous material of length ` and mass M hanging from a pivot point. Its center of mass
is at the midpoint `/2. A motor connected to the rod at the pivot point provides a torque τ . Assume there is
no friction.

(a) Schematic diagram. (b) Picking coordinates.

Figure 2.17: A physical (rod) pendulum.

1. It is standard practice to let counterclockwise rotation correspond to positive angular displacement. There
is no translational motion so we don’t need to define any other reference frames.

2. Following Figure 2.17b we choose the downward position of the pendulum to correspond to the angle
θ = 0. By our choice above, θ increases as the pendulum rotates counterclockwise.

3. In this example there is only rotational motion. The moment of inertia about the rotational axis is3
`2

2
J = M 12 + M 2` = 13 M `2 . You are not expected to know how to compute J.

4. For each component determine all external forces and moments (torques) acting on it.

• Figure 2.17b shows a free body diagram of the moments acting on the pendulum.

5. The position of the centre of mass is completely determined by θ.

6. In light of the previous point, we can ignore translational motion in this example.

7. Apply Newton’s equations for rotational motion to each mechanical component about the pivot point.

• Using basic trigonometry on the force due to gravity

`
J θ̈ = τ − M g sin (θ).
2
In summary, the equation of motion is
1 `
M `2 θ̈ = τ − M g sin (θ)
3 2
3 g
⇒ θ̈ = τ − 1.5 sin (θ).
M `2 `
In this example the governing equations are nonlinear. N
3
Apply the parallel axis theorem.

Version 1.2.2816, September 1, 2020

23 CHAPTER 2. MATHEMATICAL MODELS OF SYSTEMS

2.3.2 Electrical devices

The fundamental principals of circuit analysis are Kirchhoff ’s voltage law and Kirchhoff ’s current law. Kirch-
hoff’s Voltage Law (K.V.L.) states that the sum of voltage drops around a closed loop must be zero and Kirch-
hoff’s Current Law (K.C.L.) states that the sum of the currents entering a node must be zero. The voltages
and currents that must be summed in these laws depend on the electrical components in the circuit.

Resistors
The current-voltage characteristics of a resistor R are static. Namely, the voltage v(t) across a resistor at any
fixed time t only depends on the current i(t) at that particular time. The equations are

v(t) = Ri(t), R > 0. (2.8)

Figure 2.18: A linear resistor.

Equivalently, using complex impedances in the Laplace domain (see Section 2.7),

V (s) = RI(s). (2.9)

Inductors
The current-voltage characteristics of an inductor L are dynamic. Namely, the current i(t) passing through an
inductor at any fixed time t depends on the voltage across the inductor v(τ ) for τ ≤ t. The equations are

Zt
1
i(t) = v(τ )dτ
L
0 (2.10)
di(t)
v(t) = L .
dt
Figure 2.19: An inductor.

In the Laplace domain, if we assume zero initial conditions, equation (2.10) becomes

V (s) = sLI(s). (2.11)

Capacitors
The current-voltage characteristics of a capacitor C are also dynamic. Namely, the voltage v(t) across a capacitor
at any fixed time t depends on the current passing through the capacitor i(τ ) for τ ≤ t. The equations are

Zt
1
v(t) = i(τ )dτ
C
0 (2.12)
dv(t)
i(t) = C .
Figure 2.20: A capacitor dt

Version 1.2.2816, September 1, 2020

 2.3. Modelling based on first principles 24

In the Laplace domain, if we assume zero initial conditions, equation (2.12) becomes
1
V (s) = I(s). (2.13)
sC

Operational amplifiers
Operational amplifiers, often called op-amps, are among the most widely used electronic devices. We only
consider ideal op-amps in this course. The circuit symbol for an ideal op-amp is shown in Figure 2.21. KCL
and KVL are used to analyse op-amp circuits under two fundamental assumptions which are valid for ideal
op-amps and approximately valid for more sophisticated models.

1. There is no current into, or out of, either the

positive or negative terminals.

2. The voltage at the positive terminal v+ (t)

equals the voltage at the negative terminal
Figure 2.21: Ideal op-amp.
v− (t).

Example 2.3.4. (Low Pass Filter) Consider the RC low pass filter in Figure 2.22 with input voltage u(t)
and output voltage y(t) — the voltage across the capacitor. This circuit only has one loop so that Kirchhoff’s

Figure 2.22: An RC filter.

voltage law gives

−u(t) + vR (t) + y(t) = 0
where vR (t) is the voltage across the resistor. Substituting in the device equations for the resistor and capacitor
we get
Zt
1
−u(t) + Ri(t) + i(τ )dτ = 0.
C
0
If we note that i(t) = C ẏ(t), then the above can be written as an ODE
RC ẏ(t) + y(t) = u(t).
N

Example 2.3.5. Consider the RLC circuit in Figure 2.23. Apply KVL to this circuit to get
−u(t) + vR (t) + vC (t) + vL (t) = 0
where vR (t), vC (t), vL (t) are the voltages across, respectively, the resistor, capacitor and inductor. We substitute
the device equations and use the definition of y(t) to obtain
Zt
1
−u(t) + Ry(t) + Lẏ(t) + y(τ )dτ = 0.
C
0

Version 1.2.2816, September 1, 2020

25 CHAPTER 2. MATHEMATICAL MODELS OF SYSTEMS

Figure 2.23: RLC circuit.

This is not a differential equation because of the integral. Take the time derivative of this equation to obtain
an ODE model
1
−u̇ + Rẏ + Lÿ + y = 0.
C
N

Example 2.3.6. (Pure Gain Op-Amp) An ideal op-amp circuit is shown in Figure 2.24. Determine the
governing equations of this circuit. With the ideal op-amp assumptions in mind, i.e.,

Figure 2.24: A pure gain op-amp circuit

• There is no current into, or out of, either the positive or negative terminals,

• The voltage at the positive terminal v+ (t) equals the voltage at the negative terminal v− (t),

we perform KCL at the negative terminal

u(t) − v− (t) y(t) − v− (t)

+ = 0.
R1 R2
From the ideal op-amp assumptions we have v+ (t) = v− (t). Without loss of generality we can measure all
voltages relative to the positive terminal so that v+ (t) = 0. Therefore

R2 u(t) + R1 y(t) = 0.

This is a static system since the value of y at time t only depends on the value of u at time t, i.e., y(t) =
−R2 /R1 u(t). For this reason the circuit is called a pure gain system. N

2.3.3 Electromechanical devices

Example 2.3.7. (DC Motor) A common actuator used to provide rotary motion in control systems is the
permanent magnet DC motor which was discussed in Section 1.2.1. In this example we perform a detailed
derivation of its equations of motion.

Version 1.2.2816, September 1, 2020

 2.3. Modelling based on first principles 26

A DC motor works on the principle that a current-carrying conductor in a magnetic field experiences a
force. The motor consists of a fixed stator and a movable rotor (also called the armature) that rotates inside
the stator.
The stator produces a magnetic field. DC motors can be classified according to the way that the magnetic
field is produced. Here we are discussing a permanent magnet motor whose stator consists of a permanent
magnet. The key physical cause-and-effect relations are:

1. If the stator produces a magnetic field and a current i passes through the armature, then there will be a
torque on the armature causing it to rotate.

2. Whenever a conductor moves inside a magnetic field, a voltage is created across its terminals that is
proportional to the velocity of the conductor in the field. This voltage is called the back emf .

The “motor equations” model these two phenomena and provide the link between the electrical and mechanical
subsystems of the DC motor. They give the torque τ (t) on the rotor in terms of the armature current i(t) and
express the back emf voltage4 e(t) in terms of the shaft’s rotational velocity θ̇(t)

τ (t) = Kτ i(t)
(2.14)
e(t) = Ke θ̇(t).

Here Kτ and Ke are physical constants.

To model a DC motor we first draw a functional block diagram in Figure 2.25. In this figure v(t) is the
voltage applied to the rotor circuit, i(t) is the resulting current in the circuit. This current acts as an input to
the mechanical subsystem through the first motor equation (2.14). The signal θ(t) is the angular position of

electrical mechanical
subsystem subsystem

Figure 2.25: Functional block diagram of a DC motor.

the motor’s shaft while θ̇(t) is its angular velocity which acts as an input to the electrical subsystem via the
second motor equation (2.14).
Next we model each block in the functional diagram. Start by drawing the equivalent circuit for the electrical
subsystem and the free body diagram for the mechanical subsystem, the rotor, as shown in Figure 2.26. We
denote the rotor shaft’s inertia with the symbol J and assume there is viscous friction, modelled using a
rotational damper with coefficient b.

Figure 2.26: DC Motor: Circuit and free body diagram for the rotor / shaft.
4
Since the generated Electromotive Force (E.M.F.) works against the applied armature voltage, it is called back emf.

Version 1.2.2816, September 1, 2020

27 CHAPTER 2. MATHEMATICAL MODELS OF SYSTEMS

Application of Newton’s laws to the shaft, along with the motor equations (2.14), yield

J θ̈ = −bθ̇ + τ
(2.15)
= −bθ̇ + Kτ i.
Notice that this equation assumes that the only source of inertia in the motor is from to the shaft itself. If we
connect a load to the motor then this is no longer true and the load can be modelled as a disturbance torque
on the right-hand side of (2.15). The electrical equation is found using KVL
di
− v + Ri + L +e=0
dt (2.16)
di
⇒ −v + Ri + L + Ke θ̇ = 0.
dt
In summary, (2.15) and (2.16) are the equations of motion governing a DC motor. N
Remark 2.3.1. If in Example 2.3.7 the response of the circuit is much faster than that of the mechanical
system, e.g., if the inductance L ≈ 0, then v(t) = Ri(t) + Ke θ̇(t) and the equations (2.15), (2.16) reduce to
 
Kτ Ke Kτ
J θ̈ = − b + θ̇ + v. (2.17)
R R
If we now define the constants
1 K K Kτ
ω(t) := θ̇(t), := b + τ e , kE := ,
τ R R
then (2.17) recovers the model used for motor speed control in Section 1.2.1. 
Our last example, adapted from [Gopal, 1963], provides a somewhat complete treatment of how to model
the connection of mechanical components to DC motors.
Example 2.3.8. (Geared Drive) Electric motors generally produce their maximum power (= torque × angu-
lar velocity) at high speeds. Therefore at high speeds, they generally produce small torques. As a consequence,
gears are needed to drive large loads (that require high torque) at high speeds.
Consider the setup in Figure 2.27. A DC motor as in Example 2.3.7 is connected through a gear train to
the physical pendulum from Example 2.3.3. The gear with N1 teeth is called the primary gear and the gear
with N2 teeth is called the secondary gear. We denote the angular displacement of the motor shaft by θ1 and
the angular displacement of the pendulum by θ2 . The coordinates are such that θ1 = 0 when θ2 = 0. Let J1
denote the inertia of the motor shaft and the gears. Let τ2 denote the torque on the pendulum.
To analyse this system we first need to understand the effect of the gears. Ideal gears have no dynamics but
they impose algebraic constraints5

N1 θ1 = N2 θ2
τ1 τ2 (2.18)
= .
N1 N2

Figure 2.28: Gears

5
In reality, there is always a certain amount of backlash (free play) between coupled gears. On the one hand, excessive backlash
results in a nonlinear relationship between θ1 and θ2 and can lead to undesirable sustained oscillations and positional inaccuracy.
On the other hand, keeping the backlash small inevitably increases the friction between the gears causing the teeth to wear out
faster. We assume ideal gear equations (2.18).

Version 1.2.2816, September 1, 2020

 2.4. State-space models 28

Figure 2.27: DC Motor connected to a physical pendulum.

From Example 2.3.7 the motor’s dynamics are

N1
J1 θ̈1 = −bθ̇1 + Kτ i − τ2
N2
di
L = −Ri − Ke θ̇1 + v.
dt
The torque from the pendulum appears as a disturbance to the motor’s angular acceleration (cf. Section 1.2.1).
Notice that if N1 /N2  1 then the disturbance’s effect is reduced. From Example 2.3.3 the pendulum’s dynamics
are
`
J2 θ̈2 + M g sin (θ2 ) = τ2 .
2
If we replace the expression for τ2 into the above equation and use the gear equations (2.18) we have
  
N1 N1 ` N1
J1 θ̈1 = −bθ̇1 + Kτ i − J2 θ̈1 + M g sin θ1
N2 N2 2 N2
di
L = −Ri − Ke θ̇1 + v.
dt
Simplifying the above we get the model
 2 !  
N1 N1 ` N1
J1 + J2 θ̈1 = −bθ̇1 + Kτ i − M g sin θ1
N2 N2 2 N2
di
L = −Ri − Ke θ̇1 + v.
dt
N

2.4 State-space models

State-space models are a way of expressing mathematical models in a standard form.

Version 1.2.2816, September 1, 2020

29 CHAPTER 2. MATHEMATICAL MODELS OF SYSTEMS

Figure 2.29: A simple cart subject to air resistance.

Example 2.4.1. (Cart with Air Resistance) Consider a cart on wheels, driven by a force u and subject to
air resistance as in Figure 2.29. Typically air resistance creates a force depending on the velocity ẏ; let’s say
this force is a possibly nonlinear function D(ẏ). Assuming M is constant, Newton’s second law gives

M ÿ = u − D(ẏ).

We are going to put this in a standard form by defining two so-called state variables, in this example position
and velocity:
x1 := y, x2 := ẏ.
Then
ẋ1 = x2
1 1
ẋ2 = u− D(x2 )
M M
y = x1 .
These equations have the form
ẋ = f (x, u) (state equation)
(2.19)
y = h(x) (output equation)
where  
x1
x := , u := applied force
x2
 
x2
f : R2 × R → R2 , f (x1 , x2 , u) =  
1 1
Mu − M D(x2 )

h : R2 → R, h(x1 , x2 ) = x1 .
The function f is nonlinear if D is; h is linear. Equation (2.19) constitutes a state-space model of the system,
and x is called the state or state vector. Here the plant is a possibly nonlinear system, u (applied force) is
the input, y (cart position) is the output, and
 
cart position
x=
cart velocity

is the state of the system. (We’ll define state later.)

As a special case, suppose the air resistance is a linear function of velocity

D(x2 ) = dx2 , d is a real constant.

Then f is linear function of x, u:

   
0 1 0
f (x, u) = Ax + Bu, A := , B := .
0 −d/M 1/M

Version 1.2.2816, September 1, 2020

 2.4. State-space models 30
 
Defining C := 1 0 , we get the linear state model
      
ẋ1 0 1 x1 0
= d + 1 u
ẋ2 0 −M x2 M
  (2.20)
  x1
y= 1 0 .
x2

In other words, when D(ẏ) is linear, the state model has the structure

ẋ = Ax + Bu, y = Cx.

This model is of a linear, time-invariant system.

N

Generalizing this example, we can say that an important class of models is

ẋ = f (x, u), f : Rn × Rm → Rn
(2.21)
y = h(x, u), h : Rn × Rm → Rp .

This model is nonlinear, time-invariant. The input u has dimension m, the output y dimension p, and the state
x dimension n. An example where m = 2, p = 2, n = 4 is shown in Figure 2.30. In Figure 2.30 we have

Figure 2.30: A multiple-input multiple-output example.

u = (u1 , u2 ), y = (y1 , y2 ), x = (y1 , ẏ1 , y2 , ẏ2 ).

The LTI special case of (2.21) is

ẋ = Ax + Bu, A ∈ Rn×n , B ∈ Rn×m

(2.22)
y = Cx + Du, C ∈ Rp×n , D ∈ Rp×m .

In this course we mainly deal with single-input single-output systems. Therefore we restrict ourselves to the
case when m = 1 and p = 1.
Now we turn to the concept of the state of a system. Roughly speaking, x(t0 ) encapsulates all the system
dynamics up to time t0 , that is, no additional prior information is required. More precisely, the concept is this:
For any t0 and t1 , with t0 < t1 , knowing x(t0 ) and knowing {u(t) : t0 ≤ t ≤ t1 }, we can compute x(t1 ), and
hence y(t1 ).

Example 2.4.2. (Choosing State Variables) Consider the cart system in Figure 2.31 with no applied forces
and no air resistance. If we were to try simply x = y, then knowing x(t0 ) without ẏ(t0 ), we could not solve the
initial value problem for the future cart position. Similarly x = ẏ won’t work. Since the equation of motion,
M ÿ = 0, is second order, we need two initial conditions at t = t0 , implying we need a 2-dimensional state vector.
In general for mechanical systems it is customary to take the vector x to consist of positions and velocities of
all masses. N

Version 1.2.2816, September 1, 2020

31 CHAPTER 2. MATHEMATICAL MODELS OF SYSTEMS

Figure 2.31: A simple cart with no air resistance, no applied force.

Even though this course is about continuous-time systems, it is instructive to look at the state of a simple
discrete-time system.
Example 2.4.3. (Digital Filter) Consider a weighted averaging algorithm defined recursively by the difference
equation
y[k] = 3y[k − 1] + 2y[k − 2] + u[k − 2].
Here y[k] ∈ R is the output of the algorithm at the kth iteration and the above equation represents the update
rule. We can recursively solve for the output sequence {y[k]}k≥0 as long as we are given the initial values

y[−1], y[−2]

and the input sequence {u[k]}k≥−2 . The concept of a state extends to this system: The state vector x[k0 ]
encapsulates all the system dynamics up to time k0 . For any integers k0 and k1 , with k0 < k1 , knowing x[k0 ]
and knowing {u[k] : k0 ≤ k ≤ k1 }, we can compute x[k1 ], and hence y[k1 ]. As in the previous example, neither
y[k0 − 1] nor y[k0 − 2] on their own contain enough information to compute future values of y; they aren’t valid
state vectors. You should check that a valid choice of state vector for this system is x[k] = (x1 [k], x2 [k]) :=
(y[k], y[k + 1]) and that the corresponding state-space model is

x1 [k + 1] = x2 [k]
x2 [k + 1] = 2x1 [k] + 3x2 [k] + u[k]
y[k] = x1 [k].
N
Why are state models useful? First, there’s a rich control theory for them. This is covered in ECE488. You
may have heard of the Kalman filter; it is based on a state model. Second, they give a convenient data type to
store a linear model, namely, the matrices A, B, C, D which makes state models useful for computations. And
third, they give a simple way to linearize, namely, compute Jacobians. More on this in Section 2.5.
Example 2.4.4. (Mass-Spring-Damper) Consider the system in Figure 2.32. The dynamic equation is6

M ÿ = u + M g − K(y − y0 ) − bẏ.

It’s appropriate to take the position and velocity as state variables

x := (x1 , x2 ), x1 := y, x2 := ẏ

and then we get the equations

ẋ1 = x2
1 K K b
ẋ2 = u+g− x1 + y0 − x2
M M M M
y = x1 .
6
Here we are assuming that the spring is unstretched/uncompressed when y = y0 (cf. Example 2.3.1).

Version 1.2.2816, September 1, 2020

 2.4. State-space models 32

Figure 2.32: Another mass-spring-damper.

This has the form

ẋ = Ax + Bu + c
y = Cx,

where      
0 1 0 0
A= K , B= 1 , c= K , C = [1 0].
−M − Mb M g+ M y0
The constant vector c is known, and hence is taken as part of the system rather than as a signal. This system
is nonlinear. Can you see why? N

Example 2.4.5. (RLC Circuit) Recall the RLC circuit from Example 2.3.5.
1
−u̇ + Rẏ + Lÿ + y = 0.
C
It is natural to take the state variables to be voltage drop across C and current through L because together
these variables characterize the energy stored in the circuit. Let
Zt
1
x1 := capacitor voltage = y(τ )dτ
C
0
x2 := inductor current = y.
Then the KVL equation becomes
−u + Rx2 + Lẋ2 + x1 = 0
and the capacitor equation becomes
1
ẋ1 = x2 .
C
Thus
1
      
ẋ1 0 C x1 0
= + 1 u.
ẋ2 − L1 −RL x2 L
Since this is an LTI system, the state model of this system has the structure
1
   
0 C 0
ẋ = Ax + Bu, A= , B = 1
− 1 −R L
 L  L
y = Cx + Du, C= 0 1 , D = 0.
N

Version 1.2.2816, September 1, 2020

33 CHAPTER 2. MATHEMATICAL MODELS OF SYSTEMS

Example 2.4.6. Recall the simplified car suspension system from Example 2.3.1
M1 q̈1 + b1 q̇1 + k1 q1 = b1 q̇2 + k1 q2
M2 q̈2 + b1 q̇2 + (k1 + k2 )q2 = k1 q1 + b1 q̇1 + k2 q3 .
For mechanical systems, we take the position and velocities of the masses as our state variables. Namely, let

x := (x1 , x2 , x3 , x4 ), x1 := q1 , x2 := q̇1 , x3 := q2 , x4 := q̇2 .

Let the input to the system be the road height u := q3 . Let the output of the system be the position of the
chassis y := q1 . We get the equations

ẋ1 = x2
1
ẋ2 = (−k1 x1 − b1 x2 + k1 x3 + b1 x4 )
M1
ẋ3 = x4
1
ẋ4 = (k1 x1 + b1 x2 − (k1 + k2 )x3 − b1 x4 + k2 u)
M2
y = x1 .

This has the form

ẋ = Ax + Bu
y = Cx,

where
0 1 0 0
  

0
k1
 −M b1 k1 b1
−M M1 M1
  0   
A=
 0
1 1 , B=
 0 ,
 C= 1 0 0 0 .
0 0 1 
k1 b1 k2
M2 M2 − k1M+k
2
2 b1
M2 M2
N

Example 2.4.7. (DC Motor) We return to the DC motor from Example 2.3.7. The equations of motion are

J θ̈ = −bθ̇ + Kτ i
di
− v + Ri + L + Ke θ̇ = 0.
dt
For state variables we take the position and velocities for the mechanical subsystem and the inductor current
for the electrical subsystem
x := (x1 , x2 , x3 ), x1 := θ, x2 := θ̇, x3 := i.
Let the input to the system be the applied voltage u := v. Let the output of the system be the angular position
of the motor’s shaft y := θ. We get the equations

ẋ1 = x2
1
ẋ2 = (−bx2 + Kτ x3 )
J
1
ẋ3 = (−Ke x2 − Rx3 + u)
L
y = x1 .

This has the form

ẋ = Ax + Bu
y = Cx,

Version 1.2.2816, September 1, 2020

 2.5. Linearization 34

where   

0 1 0 0
A =  0 − Jb KJτ  ,
 
B =  0 , C= 1 0 0 .
1
0 − KLe − R
L L
N

2.5 Linearization
Virtually all physical systems are nonlinear in nature. We saw that nonlinear systems can be modelled by
equations of the form (2.21) which we re-write here

ẋ = f (x, u)
y = h(x, u).

There might be disturbance signals present, but for now we suppose they are lumped into u. This state-space
model includes the linear version (2.22) as a special case, i.e., f (x, u) = Ax + Bu, h(x, u) = Cx + Du, but
also includes a much larger class of systems that includes robotic manipulators, mobile and humanoid robots,
population dynamics, economic systems etc.

Example 2.5.1. (Physical Pendulum) Recall the governing ODE of the physical pendulum from Exam-
ple 2.3.3
3 g
θ̈ = 2
τ − 1.5 sin (θ).
M` `
For state variables we take the position and velocity of the pendulum

x := (x1 , x2 ), x1 := θ, x2 := θ̇.

The input is the applied torque u := τ and the output is taken to be the pendulum’s angle y := θ. We get the
equations

ẋ1 = x2
g 3
ẋ2 = −1.5 sin (x1 ) + u
` M `2
y = x1 .

This has the form (2.21) where

 
x2
f (x, u) = , h(x, u) = x1 .
−1.5 g` sin (x1 ) + 3
M `2
u

This system is nonlinear because f (x, u) contains the term sin (x1 ); you can’t find constant matrices A, B such
that f (x, u) equals Ax + Bu. N

There are techniques for controlling nonlinear systems, but that’s an advanced subject. However, many
systems can be linearized about an equilibrium configuration. In other words, although almost every physical
system contains nonlinearities, oftentimes its behaviour within a certain operating range of an equilibrium
configuration can be reasonably approximated by that of a linear model. The linear approximation is then used
as the basis for control design. One reason for approximating the nonlinear system (2.21) by a linear model is
that, by doing so, one can apply rather simple and systematic linear control design techniques. Keep in mind,
however, that a linearized model will only be a good approximation to the nonlinear model when the system
operates in a sufficiently small range around an equilibrium configuration. In this section we see how to make
these approximations. The idea is to use Taylor series.

Version 1.2.2816, September 1, 2020

35 CHAPTER 2. MATHEMATICAL MODELS OF SYSTEMS

Example 2.5.2. To review Taylor series, let’s linearize the function y = f (x) = x3 about the point x = 1. The
Taylor series expansion is
∞
X f (n) (x)
f (x) = cn (x − x)n , cn =
n!
n=0
f 00 (x)
= f (x) + f 0 (x)(x − x) + (x − x)2 + · · · .
2
Taking only terms n = 0, 1 gives
f (x) ≈ f (x) + f 0 (x)(x − x),
that is, letting y := f (x) denote the value of f at the nominal point x,

y − y ≈ f 0 (x)(x − x).

Defining δy := y − y, δx := x − x, we have the linearized function δy = f 0 (x)δx, or δy = 3δx in this case.

Obviously, this approximation gets better and better as |δx| gets smaller and smaller. Figure 2.33 depicts the
linearization in this example. N

2
y = f(x)

-2

-4

-6

-8
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2
x

Figure 2.33: Linearizing y = x3 at x = 1.

Taylor series extends to multivariable functions f : Rn → Rm .

Example 2.5.3. Consider the function

   
f1 (x1 , x2 ) x1 + 1
f : R2 → R3 , f (x1 , x2 ) =  f2 (x1 , x2 )  =  x1 x2  .
f3 (x1 , x2 ) x21 − x2

We’ll linearize the equation y = f (x) about a nominal point, say x = (1, −1). Let
 
2
y := f (x) =  −1 
2

denote the value of f at the nominal point x.

Version 1.2.2816, September 1, 2020

 2.5. Linearization 36

The multivariable Taylor series expansion of f at the point x = (1, −1) is

∂f
f (x) = f (x) + (x − x) + higher order terms
∂x x=x

where
∂f
= Jacobian of f evaluated at x
∂x x=x
 ∂f ∂f

1 1
∂x ∂x2
 ∂f21 ∂f2 
=  ∂x 1 ∂x2 
∂f3 ∂f3
∂x1 ∂x2 x=x
 
1 0
=  x2 x1 
2x1 −1 x=(1,−1)
 
1 0
=  −1 1 .
2 −1

Taking only the lower order terms in this expansion we have that

∂f
f (x) ≈ f (x) + (x − x) .
∂x x=x

In other words
∂f
y−y ≈ (x − x) .
∂x x=x
Defining the perturbations
δx := x − x, δy := y − y
we have the linearized function
δy = Aδx
where  
1 0
∂f
A= =  −1 1 .
∂x x=x 2 −1
In what sense does this approximate the equation y = f (x) near x = x? The exact equation is

δy = f (x + δx) − f (x)

and the linear approximation is

δya = Aδx
so the error is kδy − δya k. It can be proved that if f is continuously differentiable, then there exists M > 0
such that
kδy − δya k ≤ M kδxk.
Thus the error kδy − δya k is arbitrarily small if kδxk is sufficiently small. N

Version 1.2.2816, September 1, 2020

37 CHAPTER 2. MATHEMATICAL MODELS OF SYSTEMS

By direct extension, if f : Rn × Rm → Rn , then

∂f ∂f
f (x, u) ≈ f (x, u) + δx + δu.
∂x (x,u) ∂u (x,u)

This idea extends to differential equations

ẋ = f (x, u). (2.23)

Definition 2.5.1. Consider a system in state-variable form (2.23). A constant pair (x, u) is called an equi-
librium configuration of (2.23) if f (x, u) = (0, 0, . . . , 0). The constant x is called an equilibrium point.

Example 2.5.4. (Equilibria of the Physical Pendulum) Back to the pendulum example. Suppose we turn
off the motor and apply no torque, i.e., we set u = u = 0. Let’s use the definition above to find all corresponding
equilibria. To do this we must solve the equation f (x, u) = 0 for x. In other words, we must solve

x2 = 0
g
−1.5 sin (x1 ) = 0
`
for x1 , x2 . Thus the equilibria (x, u) of the pendulum with u = 0 are given by
 
kπ
x= , k is an integer.
0

Physically, this means that the pendulum is at equilibrium whenever the angle θ is either 0 (pointed downward)
or π (pointed upward), and the angular velocity θ̇ is zero.
Now suppose that we turn on the motor in such a way that it produces the constant torque u = u 6= 0. The
corresponding equilibria satisfy f (x, u) = 0, i.e.,

x2 = 0
g 3
−1.5 sin (x1 ) + u = 0.
` M `2
This yields the equilibria "   #
2u
arcsin M g` + 2πk
x= , k is an integer.
0
If we pick a position x1 ∈ (−π, π] and set u = u = 0.5M g` sin (x1 ), then the state
 
x1
x=
0

is an equilibrium point of the pendulum. Physically, this means that by imparting a suitable constant torque
to the pendulum one can make the pendulum be at rest at any desired angle x1 . For instance, by imparting
the torque u = u = 0.5M g`, the configuration x1 = π/2, x2 = 0 is an equilibrium of the pendulum. N

Version 1.2.2816, September 1, 2020

 2.5. Linearization 38

Exercise 2.1. Find an equilibrium configuration (x, u) for system from Example 2.5.4 at which its output equals 3π/4.
First, assume there is an equilibrium configuration (x, u) of (2.23), that is, a constant solution x(t) =
x, u(t) = u. Now consider the perturbations from the equilibrium configuration:
δx(t) = x(t) − x, δu(t) = u(t) − u, δx(t), δu(t) small.
We have
ẋ(t) = f (x(t), u(t))
= f (x, u) + Aδx(t) + Bδu(t) + higher order terms
where
∂f ∂f
A := , B := .
∂x (x,u) ∂u (x,u)

Since f (x, u) = 0, we have the linearized equation to be

˙ = Aδx + Bδu.
δx
Similarly, the output equation y = h(x, u) linearizes to
δy = Cδx + Dδu,
where
∂h ∂h
C := , D := .
∂x (x,u) ∂u (x,u)

Summary
Linearizing ẋ = f (x, u), y = h(x, u): Select, if one exists, an equilibrium configuration (x, u). Compute the
four Jacobians, A, B, C, D, of f and h at the equilibrium point. Then the linearized system is
˙ = Aδx + Bδu,
δx
δy = Cδx + Dδu.
Under mild conditions (sufficient smoothness of f and h), this linearized system is a valid approximation of the
nonlinear one in a sufficiently small neighbourhood of the equilibrium configuration.
Example 2.5.5. (Physical Pendulum) Suppose that we want to control the physical pendulum
ẋ1 = x2
g 3
ẋ2 = −1.5 sin (x1 ) + u
` M `2
y = x1
near its upright position, i.e., ȳ = π. In this example we compute a linear model that approximates the
pendulum near this position. First we find the equilibrium configuration (x, u) corresponding to the desired
output ȳ. Using the results of Example 2.5.4 we get the equilibrium configuration (verify!)
  
π
(x, u) = ,0 .
0
To get the linearization at (x, u) we have to compute the four matrices A, B, C, D:
   
∂f 0 1 ∂f 0
A= = , B= =
∂x (x,u) 1.5 g` 0 ∂u (x,u) 3
M `2
∂h   ∂h
C= = 1 0 , D= = 0.
∂x (x,u) ∂u (x,u)

Version 1.2.2816, September 1, 2020

39 CHAPTER 2. MATHEMATICAL MODELS OF SYSTEMS

The linearized model is    

˙ = 0 1 0
δx δx + δu
1.5 g` 0 3
M `2
 
δy = 1 0 δx
where δx = x − x, δu = u − u and δy = y − y measure how far the pendulum is from equilibrium. N

2.6 Simulation
Concerning the model
ẋ = f (x, u), y = h(x, u),
simulation involves numerically computing x(t) and y(t) given an initial state x(0) and an input u(t). If the
model is nonlinear, simulation requires an ODE solver, based on, for example, the Runge-Kutta method. Scilab
and MATLAB have ODE solvers and very nice simulation GUIs called, respectively, Scicos and SIMULINK.

Example 2.6.1. Recall the model of the physical pendulum from Example 2.3.3. The MATLAB code below
simulates this model for twenty seconds in the case M = 1, ` = 1, g = 9.8, θ(0) = 0, θ̇(0) = 0 and u(t) = 1(t).
After solving the ODE, the code also plots two figures to visually display the results of the simulation.

1 options = odeset('RelTol',1e−12, 'abstol', 1e−12); % set options for ODE solver

2 t span = [0 20]; % amount of time we will simulate for
3
4 x0 = [0;0]; % Set the initial conditions
5 M = 1; ell = 1; g = 9.8; % model parameters
6
7 % Now simulate the system
8 [t, x] = ode45(@examplesys, t span, x0, options, M, g, ell); %% medium accuracy ODE ...
solver, Runge−Kutta method
9
10 figure;
11 plot(t, x(:, 1), t, x(:, 2));
12 legend('$\theta(t)$', '$\dot{\theta}(t)$', 'Interpreter','latex', 'FontSize', 14);
13 xlabel('time (s)', 'Interpreter','latex', 'FontSize', 17);
14 grid on;
15
16 figure;
17 plot(x(:, 1), x(:, 2));
18 xlabel('$x 1(t) = \theta(t)$', 'Interpreter','latex', 'FontSize', 17);
19 ylabel('$x 2(t) = \dot{\theta}(t)$', 'Interpreter','latex', 'FontSize', 17);
20 grid on;
21
22 function dx = examplesys(t,x, M, g, ell)
23
24 x1 = x(1); %% theta
25 x2 = x(2); %% theta dot
26
27 u = 1; % control signal (applied torque)
28
29 f = [x2; −1.5*g/ell*sin(x1) + 3/M/ellˆ2*u]; % system vector field
30
31 dx = f; % The system model
32 end

Version 1.2.2816, September 1, 2020

 2.7. The Laplace transform 40

2.7 The Laplace transform

The Laplace transform is the fundamental tool used in control systems to get to the frequency domain.7 You
have already met Laplace transforms in previous courses so we only have to give a brief review here.
In signal processing, the two-sided Laplace transform (LT) is used, but in control only the one-sided LT is
used. Let f (t) be a signal (a real-valued function of time) defined for all t or just for t ≥ 0. The one-sided
Laplace transform of f (t) is
Z∞
F (s) = f (t)e−st dt. (2.24)
0

Here s is a complex variable. We use L (f ) to denote the Laplace Transform of f . Normally, the integral
converges for some values of s and not for others. That is, there is a region of convergence. It turns out
that the Region of Convergence (R.O.C.) is always an open right half-plane, of the form {s ∈ C : Re (s) > a}.
Within the ROC F (s) has no poles.

Example 2.7.1. The unit step: 

1 , t≥0
f (t) = 1(t) =
0 , t < 0.
Actually, the precise value at t = 0 doesn’t matter. The LT is
Z∞ ∞
−st e−st 1
F (s) = e dt = − =
s 0 s
0

and the ROC is

ROC : Re (s) > 0.
The same F (s) is obtained if f (t) = 1 for all t, even t < 0. This is because the one-sided LT is oblivious to
negative time. Notice that F (s) has a pole at s = 0 on the western boundary of the ROC. N

The LT exists provided f (t) satisfies two conditions. The first is that it is piecewise continuous on t ≥ 0.
This means that, on any time interval (t1 , t2 ), f (t) has at most a finite number of jumps, and between these
jumps f (t) is continuous. A square wave has this property for example. The second condition is that it is of
exponential order, meaning there exist constants M, c such that |f (t)| ≤ M ect for all t ≥ 0. This means that if
f (t) blows up, at least there is some exponential that blows up faster. For example, exp(t2 ) blows up too fast.

Example 2.7.2. Some other examples: An exponential:

1
f (t) = eat , F (s) = , ROC : Re (s) > a.
s−a
A sinusoid:
1 jωt
e + e−jωt


f (t) = cos (ωt) =
2
s
F (s) = , ROC : Re (s) > 0.
s2 + w2
N

The LT thus maps a class of time-domain functions f (t) into a class of complex-valued functions F (s). The
mapping f (t) 7→ F (s) is linear.

Version 1.2.2816, September 1, 2020

41 CHAPTER 2. MATHEMATICAL MODELS OF SYSTEMS

Figure 2.34: Time domain signal for Example 2.7.3.

Example 2.7.3. Let’s use linearity to find the LT of the signal in Figure 2.34. We can write f = f1 + f2 ,
where f1 (t) = t is the unit ramp starting at time 0 and f2 the ramp of slope −1 starting at time 1. By linearity,
F (s) = F1 (s) + F2 (s). We compute that
1
F1 (s) = , Re (s) > 0
s2
1
F2 (s) = −e−s , Re (s) > 0.
s2
Thus
1 − e−s
F (s) = , Re (s) > 0.
s2
N
There are tables of LTs. So in practice, if you have F (s), you can get f (t) using a table like Table 2.1. If

Table 2.1: Important (one-sided) Laplace transforms.

Description Time domain f (t) s-Domain F (s)

1
Unit step 1(t) s
Impulse δ(t) 1
1
Ramp t s2
1
Exponential eat s−a
ω
Sine sin (ωt) s2 +ω 2
s
Cosine cos (ωt) s2 +ω 2
n!
General exponential tn eat (s−a)n+1
ω
Generalized sine eat sin (ωt) (s−a)2 +ω 2
s−a
Generalized cosine eat cos (ωt) (s−a)2 +ω 2
2ωs
Sine with linear growth t sin (ωt) (s2 +ω 2 )2
s2 −ω 2
Cosine with linear growth t cos (ωt) (s2 +ω 2 )2

F (s) is a rational function, i.e.,

N (s)
F (s) =
D(s)
7
This is in contrast to communication theory where the Fourier transform dominates. Control problems frequently involve
unstable systems, unbounded signals, and transient response requirements–all absent in communication problems.

Version 1.2.2816, September 1, 2020

 2.7. The Laplace transform 42

where N , D ∈ R[s] are polynomials and the degree of N is strictly less than the degree of D, then we can use
partial fraction expansions and the linearity of the LT to obtain f (t).

Example 2.7.4. Given F (s) = 3s+17

s2 −4
, let us find f (t). We don’t need the ROC to find f (t), but actually we
know what it is. Since we’re using the one-sided LT the ROC must be a right half-plane, and because F (s)
must be analytic within its ROC, the ROC of F (s) must be Re (s) > 2. We have
3s + 17 c1 c2 23 11
F (s) = 2
= + , c1 = , c2 = −
s −4 s−2 s+2 4 4
and therefore
f (t) = c1 e2t + c2 e−2t , t ≥ 0.
We do not know if f (t) equals zero for t < 0. N

Example 2.7.5. Given

s−2
F (s) = ,
(s − 1)(s + 2)
let us find f (t). Once again, we don’t need the ROC to find f (t), but we can deduce that it is the right half-plane
Re (s) > 1. We have
s−2 c1 c2 1 4
F (s) = = + , c1 = − , c2 =
(s − 1)(s + 2) s−1 s+2 3 3
and therefore
1 4
f (t) = − et + e−2t , t ≥ 0.
3 3
N

Example 2.7.6. Given

s+1
F (s) =
s(s + 2)2
we write
s+1 c1 c2 c3
F (s) = 2
= + + .
s(s + 2) s s + 2 (s + 2)2
This implies
s + 1 = c1 (s + s)2 + c2 s(s + 2) + c3 s.
Comparing coefficients we get three linear equations. You can verify that
1 1 1
c1 = , c2 = − , c3 = .
4 4 2
Therefore
1 1 −2t 1 −2t
f (t) = − e + te , t ≥ 0.
4 4 2
N

One use of the LT is in solving initial-value problems involving linear, constant-coefficient differential equa-
tions. In control engineering we do this by simulation. But let us look briefly at the LT method. We first
observe that if we have the LT pair
f (t) ←→ F (s)
and f is continuously differentiable at t = 0, then

f˙(t) ←→ sF (s) − f (0). (2.25)

Version 1.2.2816, September 1, 2020

43 CHAPTER 2. MATHEMATICAL MODELS OF SYSTEMS

To prove this formula, start with the LT of f˙(t) and integrate by parts:
Z∞ Z∞
−st −st ∞
e f˙(t)dt = e f (t) 0
+s e−st f (t)dt.
0 0

Now s is such that e−st f (t) converges to 0 as t goes to ∞. Thus the right-hand side of the preceding equation
becomes
−f (0) + sF (s).
This proves (2.25).

Example 2.7.7. Consider the initial-value problem

ẏ − 2y = t, y(0) = 1.

The range of t for the differential equation isn’t stated. Let us first assume that y(t) is continuously differentiable
at t = 0. This implies the differential equation holds at least for −ε < t < ∞ for some positive ε, that is, it
holds for a little time before t = 0. Then we can apply (2.25) to get
1
sY (s) − 1 − 2Y (s) = .
s2
Solving for Y (s) we get
s2 + 1 1 1 5
Y (s) = 2
=− − 2 + .
s (s − 2) 4s 2s 4(s − 2)
Therefore
1 1 5
y(t) = − − t + e2t . (2.26)
4 2 4
Since we used the one-sided LT, which is oblivious to negative time, we can assert that (2.26) holds at least for
t ≥ 0. Note that it satisfies y(0) = 1.
On the other hand, suppose instead that the problem is posed as follows:

ẏ − 2y = t, t > 0; y(0) = 1.

That is, the differential equation holds for positive time and the initial value of y is 1. We aren’t told explicitly
that y(t) is continuously differentiable at t = 0, but we are justified in making that assumption, since any other
solution, for example the one satisfying y(t) = 0 for t < 0, satisfies (2.26) for t ≥ 0.
N

Example 2.7.8. Let us take the same system as in the previous example but modify the forcing term:

ẏ − 2y = 1, y(0) = 1.

We assume y(t) is continuously differentiable at t = 0 and the solution goes like this:
1
sY (s) − 1 − 2Y (s) =
s
s+1 1 3
Y (s) = =− +
s(s − 2) 2s 2(s − 2)
1 3
y(t) = − + e2t , t ≥ 0.
2 2
N

Version 1.2.2816, September 1, 2020

 2.7. The Laplace transform 44

If f (t) is twice continuously differentiable at t = 0, then

f¨(t) ←→ s2 F (s) − sf (0) − f˙(0).

Example 2.7.9. The equation

ÿ + 4ẏ + 3y = et , y(0) = 0, ẏ(0) = 2
can be solved as follows. We assume y(t) is twice continuously differentiable at t = 0. Then
1
s2 Y (s) − 2 + 4sY (s) + 3Y (s) = .
s−1
So
2s − 1
Y (s) =
(s − 1)(s + 1)(s + 3)

1 1 3 1 7 1
= + −
8s−1 4s+1 8s+3
1 3 7
y(t) = et + e−t − e−3t , t ≥ 0.
8 4 8
The same solution would be arrived at for t > 0 if, instead of assuming y(t) is twice continuously differentiable
at t = 0, we were to allow jumps at t = 0. N
The LT of the product f (t)g(t) of two functions is not equal to F (s)G(s), the product of the two transforms.
Then what operation in the time domain does correspond to multiplication of the transforms? The answer is
convolution. Let f (t), g(t) be defined on t ≥ 0. Define a new function
Zt
h(t) = f (t − τ )g(τ )dτ, t ≥ 0.
0

We say h is the convolution of f and g. Note that another equivalent way of writing h is
Zt
h(t) = f (τ )g(t − τ )dτ, t ≥ 0.
0

We also frequently use the star notation h = f ∗ g or h(t) = f (t) ∗ g(t).

Theorem 2.7.1. The LT of f ∗ g is F (s)G(s).
Proof. Let h := f ∗ g. Then the LT of h is
Z∞
H(s) = h(t)e−st dt.
0

Substituting for h we have

Z∞ Zt
H(s) = f (t − τ )g(τ )e−st dτ dt.
0 0
Now change the order of integration (draw the picture):
Z∞ Z∞
H(s) = f (t − τ )e−st dt g(τ ) dτ.
0 τ

Version 1.2.2816, September 1, 2020

45 CHAPTER 2. MATHEMATICAL MODELS OF SYSTEMS

In the inner integral change variables, r = t − τ :

Z∞ Z∞
 

H(s) =  f (r)e−sr dr e−sτ g(τ )dτ.

0 0

Thus
Z∞
H(s) = F (s)e−sτ g(τ )dτ.
0

Pull F (s) out and you get H(s) = F (s)G(s).

Properties of Laplace transforms: Let f (t) and g(t) be real-valued functions, continuously differentiable
at t = 0, and let a be a real constant.

(i) L {f + g} = L {f } + L {g}

(ii) L {af } = aL {f }
n o
(iii) L dfdt = sL {f } − f (0)

(iv) L {f ∗ g} = L {f } L {g}
t 
(v) L f (τ )dτ = 1s L {f }
R
0

(vi) L {f (t − T )} = e−sT L {f } , T ≥ 0.

2.7.1 Fourier Transform

The Fourier transform of x(t) is
Z∞
X(jω) = x(t)e−jωt dt.
−∞

We use F (x) to denote the Fourier Transform of x. The inversion formula is

Z∞
1
x(t) = X(jω)ejωt dω.
2π
−∞

If the absolute value of x(t) is integrable, that is,

Z∞
|x(t)|dt < ∞
−∞

then X(jω) is a continuous function of ω. An example is x(t) = e−|t| cos (t). If the absolute value of x(t) is only
square integrable, that is,
Z∞
|x(t)|2 dt < ∞,
−∞

Version 1.2.2816, September 1, 2020

 2.7. The Laplace transform 46

then X(jω) may not be a continuous function of ω. An example is sin (t)/t.

The constant signal that equals one for all time is neither absolutely integrable nor square integrable.
Its Fourier Transform (F.T.) is defined to be 2πδ(ω). We can convince ourselves of this by noting that the
inversion formula is an instance of the sifting formula
Z∞
1
2πδ(ω)ejωt dω = 1.
2π
−∞

The forward FT is defined to be

Z∞
e−jωt dt = 2πδ(ω)
−∞

and it has no meaning in the sense of ordinary functions. Likewise, the sinusoidal signal ejω0 t is neither absolutely
integrable nor square integrable. Its FT is defined to be 2πδ(ω − ω0 ).
In general, FTs where either x(t) or X(jω) is not a function (i.e. has an impulse) must be treated with care
to make sure that the result is correct.

2.7.2 Laplace vs Fourier

Here’s an interesting question: Suppose you have a real-valued signal g(t). Assume g(t) = 0 for t < 0. Let G(s)
be the Laplace transform of g(t) and then substitute jω for s in G(s), so now you have G(jω). Is G(jω) the
Fourier transform of g(t)? The answer is: not necessarily.
To illustrate this, for this section only, denote the LT of g(t) as GLT (s) and the FT of g(t) as GF T (jω).
Let’s look at two examples.
Example 2.7.10. Consider the step function u(t) = 1(t). The LT and ROC are
1
ULT (s) = , ROC = {s ∈ C : Re (s) > 0}.
s
That is, the ROC is the open right half complex plane and ULT (s) has a pole on the boundary of that region.
Next consider the causal signal g(t) = e−t 1(t). Then
1
GLT (s) = , ROC = {s ∈ C : Re (s) > −1}.
s+1
Notice that the ROC contains the imaginary axis. N

Example 2.7.11. Let’s turn to FTs and take the FT of the two signals from Example 2.7.10. The FT of the
unit step is
Z∞
1
UF T (jω) = e−jωt dt = πδ(ω) + .
jω
0
Thus the FT is a distribution having an impulse. Notice that the FT and LT are not equivalent
UF T (jω) 6= ULT (jω).
Indeed, setting s = jω in ULT (s) requires some justification since the imaginary axis is not in the ROC.
For g(t) = e−t 1(t) we have
1
GF T (jω) = .
jω + 1
In this case the FT is a function. Furthermore, the FT and LT are equivalent
GF T (jω) = GLT (jω).
In general, the LT and FT are equivalent when the ROC of the LT includes the imaginary axis. N

Version 1.2.2816, September 1, 2020

47 CHAPTER 2. MATHEMATICAL MODELS OF SYSTEMS

2.8 Transfer functions

Linear time-invariant systems, and only LTI systems, have transfer functions. The transfer function of an
LTI system is defined to be the ratio Y (s)/U (s) where the LTs are taken with zero initial conditions.

Example 2.8.1. (Time-Varying System) A system governed by the differential equation

ẏ + (sin t)y = u.

does not have a transfer function—it isn’t time invariant. N

Example 2.8.2. (Low Pass Filter) Recall the RC filter from Example 2.3.4

RC ẏ + y = u.

Apply Laplace transforms with zero initial conditions:

sRCY (s) + Y (s) = U (s).

Therefore the TF is
Y (s) 1
= .
U (s) RCs + 1
This transfer function is rational, a ratio of polynomials. N

Exercise 2.2. Re-derive the TF from Example 2.8.2 using the voltage-divider rule and complex impedances.

Example 2.8.3. (Mass-Spring-Damper) A linear mass-spring-damper:

M ÿ = u − Ky − bẏ.

We get
Y (s) 1
= 2
.
U (s) M s + bs + K
This transfer function also is rational. N

Example 2.8.4. (RLC Circuit) In Example 2.3.5 we derived the governing ODE for an RLC circuit to be
1
−u̇ + Rẏ + Lÿ + y = 0.
C
Apply Laplace transforms with zero initial conditions:
1
−sU (s) + sRY (s) + s2 LY (s) + Y (s) = 0.
C
The resulting TF is
Y (s) Cs
= 2
.
U (s) LCs + RCs + 1
N

Version 1.2.2816, September 1, 2020

 2.8. Transfer functions 48

Example 2.8.5. (DC Motor) In Remark 2.3.1 we found the simplified model of a DC motor, assuming L ≈ 0,
to be  
Kτ Ke Kτ
J θ̈ = − b + θ̇ + v.
R R
If we now define  
1 Kτ Ke Kτ
a1 := b+ , b0 := ,
J R JR
then the TF from the voltage v to the shaft angle θ is
Θ(s) b0
= .
V (s) s(s + a1 )
The TF from the voltage v to the shaft velocity ω is
Ω(s) b0
= .
V (s) s + a1
N
Table 2.2 lists other examples of transfer functions.
Table 2.2: Common Transfer Functions.

Description Governing Equation Transfer Function

Pure gain y(t) = u(t) 1

1
Integrator ẏ(t) = u(t) s
1
Double integrator ÿ(t) = u(t) s2
Ideal differentiator y(t) = u̇(t) s
Time delay y(t) = u(t − T ), T > 0 e−sT (not rational)
Kωn2
Prototype second order system ÿ(t) + 2ζωn ẏ(t) + ωn2 y(t) = Kωn2 u(t) s2 +2ζωn s+ωn
2

Rt Ki
Proportional-integral-derivative controller y(t) = Kp u(t) + Ki u(τ )dτ + Kd u̇(t) Kp + s + Kd s
0

Definition 2.8.1. (i) A transfer function G(s) is (real) rational if it is the quotient of two polynomials
bm sm + · · · + b1 s + b0
G(s) =
sn + an−1 sn−1 + · · · + a1 s + a0
where the coefficients ai , bi are real constants. The numbers m, n are the degrees of the numerator and
denominator polynomials. Let R(s) denote the set of rational functions in s ∈ C with coefficients in R.

(ii) A rational transfer function is proper if n ≥ m. This is equivalent to the condition

lim G(s) exists in C.

s→∞

(iii) A rational transfer function is strictly proper if n > m. This is equivalent to the condition

lim G(s) = 0.
s→∞

(iv) A rational transfer function is improper if it is not proper.

Version 1.2.2816, September 1, 2020

49 CHAPTER 2. MATHEMATICAL MODELS OF SYSTEMS

It is understood, once and for all, that the ratio of polynomials

s+1 (s + 1)(s − 1)
,
s(s + 2) s(s + 2)(s − 1)

represent the same rational function. In the first function the numerator and denominator polynomials are
coprime8 while in the second function these polynomials are not coprime. They represent the same rational
5
function in the same way that the expressions 10 and 12 represent the same rational number.

Definition 2.8.2. A complex number p ∈ C is a pole of a transfer function G(s) if

lim |G(s)| = ∞.
s→p

A complex number z ∈ C is a zero of a transfer function G(s) if

lim G(s) = 0.
s→z

If G(s) is rational and proper and if its numerator and denominator polynomials are coprime, then the roots
of the denominator are the poles of G(s) and the roots of the numerator are the zeros of G(s). For example the
functions
s+1 (s + 1)(s − 1)
,
s(s + 2) s(s + 2)(s − 1)
have poles at s = 0 and s = −2 and zeros at9 s = −1. In the second expression, s = 1 is a root of the
denominator but isn’t a pole since the numerator and denominator aren’t coprime.

2.8.1 Obtaining a transfer function from a state model

Let’s find the transfer function for the LTI state model
ẋ = Ax + Bu
y = Cx + Du.

Take Laplace transforms with zero initial conditions10

sX(s) = AX(s) + BU (s)

Y (s) = CX(s) + DU (s).
Eliminate X(s)
(sI − A)X(s) = BU (s)
⇒ X(s) = (sI − A)−1 BU (s)
⇒ Y (s) = C(sI − A)−1 B + D U (s).


| {z }
transfer function G(s)
8
Two polynomials are coprime if they have no common roots.
9
Since both functions approach zero as |s| → ∞ we sometimes say that the TFs have zeros at infinity. Unless otherwise stated,
we only consider finite poles and zeros.
10
Here X(s) denotes the component wise LT of the vector x(t). That is, if x(t) = (x1 (t), . . . , xn (t)) then X(s) = (X1 (s), . . . , Xn (s)).

Version 1.2.2816, September 1, 2020

 2.8. Transfer functions 50

Remark 2.8.3. The transfer function obtained from a state model is always real rational and proper. We can
see this if we write (sI − A)−1 using Cramer’s rule as
1
(sI − A)−1 = adj (sI − A)
det (sI − A)

where adj (·) is the classical adjoint of a matrix. The (i, j) entry of adj (sI − A) is (−1)i+j times the determinant
of the matrix obtained from (sI − A) by eliminating its jth row and ith column. The determinant det (sI − A)
is an element of R[s] with degree n. The adjoint is formed by cofactors of the elements of the matrix. Therefore
adj (sI − A) is a matrix whose entries are elements of R[s]; each of which has degree strictly less than n.
Therefore (sI − A)−1 is a matrix in which each entry is a strictly proper element of R(s). It follows that
G(s) = C(sI − A)−1 B + D ∈ R(s) is proper. 

Exercise 2.3. Use Cramer’s rule to show that the TF associated with a state-space model can also be computed using
 
sI − A B
det
Y (s) −C D
= .
U (s) det (sI − A)

This construction naturally leads to the converse question known as the realization problem: Given G(s),
find A, B, C, D such that
G(s) = C(sI − A)−1 B + D.
A solution exists if and only if G(s) is rational and proper. The solution is never unique.

Summary
Let us recap our procedure for getting the transfer function of a system:

1. Apply the laws of physics or first principles to get differential equations governing the behaviour of the
system. Put these equations in state form. In general these are nonlinear.

2. Find an equilibrium, if there is one. If there is more than one equilibrium, you have to select one. If there
isn’t even one, this method doesn’t apply.

3. Linearize about the equilibrium point.

4. Take Laplace transform of the linearized system with zero initial state.

5. Solve for the output Y (s) in terms of the input U (s).

The transfer function from input to output satisfies

Y (s) = G(s)U (s).

In general G(s) is a matrix. In the SISO case, G(s) is a scalar-valued transfer function.

Example 2.8.6. (TF of the Linearized Pendulum) In Example 2.5.1 we found the nonlinear state model
of a physical pendulum (step 1). In Example 2.5.5 we found the pendulum’s equilibrium configuration corre-
sponding to being upright (step 2). In Example 2.5.5 we also linearized about the equilibrium configuration to
obtain a LTI state model (step 3)
   
˙ 0 1 0
δx = δx + δu
1.5 g` 0 3
M `2
 
δy = 1 0 δx.

Version 1.2.2816, September 1, 2020

51 CHAPTER 2. MATHEMATICAL MODELS OF SYSTEMS

We now find the system’s transfer function (step 4, step 5) as

G(s) = C(sI − A)−1 B + D

1
= C adj (sI − A)B + D
det (sI − A)
  
1   s 1 0
= 2 1 0
s − 1.5 g` 1.5 g` s 3
M `2
3 1
= .
M ` s − 1.5 g`
2 2

2.9 Block diagram manipulations

We take the point of view that a block diagram is a picture of a function (see Section 2.2). With this in
mind, manipulating block diagrams becomes pretty straightforward. In this section we assume that each block
represents a transfer function and that the inputs and outputs to each block are the LTs of the appropriate
signal as in Figure 2.35. In that figure Y (s) = G(s)U (s) so Y (s) is a linear function of U (s).

Figure 2.35: A block diagram of a transfer function.

Series
Consider a system that is the cascade connection of systems with transfer functions G1 (s) and G2 (s), as
shown in Figure 2.36. Let V (s) denote the LT of the signal between the blocks. Then

Figure 2.36: Cascade connection of transfer function.

V (s) = G1 (s)U (s), Y (s) = G2 (s)V (s) =⇒ Y (s) = G2 (s)G1 (s)U (s).

The transfer function of the series connection is thus G(s) = G2 (s)G1 (s). The order of the individual transfer
functions is due to the fact that we place the input signal on the right-hand side of this expression, hence we
first multiply by G1 and then by G2 . Unfortunately, this has the opposite ordering from the diagrams that
we use, where we typically have the signal flow from left to right, so one needs to be careful. The ordering is
important if either G1 or G2 is a vector-valued transfer function. If G1 (s) and G2 (s) are scalar TFs, then the
order doesn’t matter.

Version 1.2.2816, September 1, 2020

 2.9. Block diagram manipulations 52

Figure 2.37: Parallel connection of transfer function.

Parallel
Consider a system that is the parallel connection of systems with transfer functions G1 (s) and G2 (s), as shown
in Figure 2.37. Let V1 (s) denote the LT of the output of the top branch and let V2 (s) denote the LT of the
output of the bottom branch. Then we have

V1 (s) = G1 (s)U (s), V2 (s) = G2 (s)U (s) =⇒ Y (s) = V1 (s) + V2 (s) = (G1 (s) + G2 (s))U (s)

and the transfer function for a parallel connection is G(s) = G1 (s) + G2 (s).

Negative feedback
Consider the negative feedback connection of systems with transfer functions G1 (s) and G2 (s), as shown in
Figure 2.38. We have

Figure 2.38: Negative feedback connection of transfer function.

E(s) = U (s) − G2 (s)Y (s), Y (s) = G1 (s)E(s) =⇒ Y (s) = G1 (s) (U (s) − G2 (s)Y (s))
=⇒ (1 + G1 (s)G2 (s)) Y (s) = G1 (s)U (s)
G1 (s)
=⇒ Y (s) = U (s).
1 + G1 (s)G2 (s)
Thus the transfer function for a negative feedback connection is G(s) = G1 (s)/(1 + G1 (s)G2 (s)). These three
basic interconnections can be used as the basis for computing transfer functions for more complicated systems.
Example 2.9.1. (Proportional error feedback for a first order plant) Consider the block diagram in
Figure 2.39. This is a negative feedback interconnection in which the TF G1 (s) is the series connection of Kp
and K/(τ s + 1) and G2 (s) = 1. Therefore the TF from r to y is
K
Y (s) Kp τ s+1 Kp K
= K
= .
R(s) 1 + Kp τ s+1 τ s + 1 + Kp K

Version 1.2.2816, September 1, 2020

53 CHAPTER 2. MATHEMATICAL MODELS OF SYSTEMS

Figure 2.39: Proportional error feedback on a first order plant.

This TF has a pole at s = (−1 − Kp K)/τ and therefore, by picking the proportional gain Kp appropriately, we
can place the pole anywhere along the real axis. N

Moving blocks past summing junctions

It’s not hard to convince yourself, by writing the expressions for Y (s), that the two block digrams in Figure 2.40
are equivalent.

Figure 2.40: Changing the order of summing junctions and transfer functions.

Example 2.9.2. Find the transfer function from U (s) to Y (s) for the system in Figure 2.41. Our approach is

Figure 2.41: Example 2.9.2.

to move blocks around until common configurations are revealed. We start by moving the input to G1 (s) past
the block G2 (s). This yields the system in Figure 2.42. The feedback portion in Figure 2.42 has TF

G2
.
1 + G2 G3
The series portion has TF G1 /G2 . Therefore the parallel portion has TF

G1
1+ .
G2

Version 1.2.2816, September 1, 2020

 2.9. Block diagram manipulations 54

Figure 2.42: Rearranging the block diagram from Example 2.9.2.

The final TF from U to Y is the series connection of these TFs so we have

  
G2 G1 G1 + G2
Y (s) = 1+ U (s) = U (s).
1 + G2 G3 G2 1 + G2 G3
N
The approach used in Example 2.9.2 is ad-hoc; it would be hopeless for a really complicated interconnection
of systems (and these do occur in practice). Since all of the functions in our block diagrams are linear, there
is a more systematic way to do block diagram reduction. The systematic approach turns the block reduction
problem into the problem of solving a set of linear equations.
1. Introduce new variables {v1 , v2 , . . .} representing the output of every summing junction.
2. Write down the expression for the inputs to each summing junction in terms of the signals {u, y, v1 , v2 , . . .}
and the TF blocks.
3. Write the equation of each summing junction using the expressions from the previous step.
4. Eliminate the variables {v1 , v2 , . . .}.

Example 2.9.3. We re-do Example 2.9.2 using the above procedure. Consider the system in Figure 2.41.

1. Let v1 be the output of left-most summing junction. We don’t need a variable for the summing junction
on the right since it is y.
2. The inputs to the left-most summing junction are u and G2 G3 v1 . The inputs to the right-most summing
junction are G1 v1 and G2 v1 .
3. The equations are
v 1 = u − G2 G3 v 1
y = G 1 v1 + G 2 v1 .
Or, in matrix notation,     
1 + G2 G 3 0 v1 u
= .
G1 + G2 −1 y 0
4. We can solve for y easily since this is such a small problem but I’ll use Cramer’s rule since it scales better
to larger problems. We have
 
1 + G2 G 3 u
det
G 1 + G2 0 −G1 − G2 G1 + G2
Y (s) =  = U (s) = U (s).
1 + G2 G3 0 −1 − G G
2 3 1 + G2 G3
det
G1 + G2 −1

Version 1.2.2816, September 1, 2020

55 CHAPTER 2. MATHEMATICAL MODELS OF SYSTEMS

Example 2.9.4. (2019 Midterm) Find the transfer function from r to y for the system shown in Figure 2.43.

Figure 2.43: A complicated block diagram.

1. Label, from left to right, the outputs of the summing junctions as v1 , v2 and v3 .

2. Working from left to right, the inputs to the left-most summing junction are r and y. The inputs to the
next summing junction are v1 and G2 v3 . The inputs to the next summing junction are G1 v2 and y. The
inputs to the right-most summing junction are G4 v2 and G3 G2 v3 .

3. The equations are

v1 = r − y
v2 = v1 − G 2 v3
(2.27)
v 3 = G1 v 2 − y
y = G4 v 2 + G3 G 2 v 3 .
Or, in matrix notation,     
1 0 0 1 v1 r
1 −1 −G2 0  v2  0
   =  .
0 G1 −1 −1 v3  0
0 G4 G3 G2 −1 y 0

4. In principle we can now solve for y using Cramer’s rule except that no one likes finding the determinant
of a 4 × 4 matrix by hand. To make the problem tractable for hand calculations, let’s make some simple
substitutions in (2.27). Substitute the v1 expression into v2

v 2 = r − y − G2 v 3 .

Now substitute this expression for v2 into the v3 and y equations

v3 = G1 (r − y − G2 v3 ) − y
y = G4 (r − y − G2 v3 ) + G3 G2 v3 .

In matrix notation     
1 + G1 G2 1 + G1 v3 G1 r
=
G 4 G2 − G3 G 2 1 + G 4 y G4 r

Version 1.2.2816, September 1, 2020

 2.10. Summary 56

Now we can easily apply Cramer’s rule:

 
1 + G1 G 2 G1 R
det
G4 G2 − G3 G2 G4 R G4 + G1 G2 G3
Y (s) =  = R(s).
1 + G1 G2 1 + G1 1 + G 4 + G 2 (G1 − G4 + G3 + G1 G3 )
det
G 4 G2 − G3 G 2 1 + G 4

2.10 Summary
In this chapter we introduced the very important topic of modelling. The key things you should know after
reading this chapter are as follows.
1. The usefulness of mathematical models and the key role they play in all branches of engineering. We
presented various examples showing that there are different types of models (ODEs, PDEs, discrete-time).

2. No model is perfect.

3. If you can’t draw the block diagram of a system, then you don’t understand the system very well.

4. The basic building blocks of linear models for mechanical and electrical systems were introduced in
Section 2.3. You should be comfortable modelling (deriving the governing ODEs) simple linear systems
including those with electrical and mechanical subsystems.

5. In Section 2.4 we introduced state models which provide a standard way of expressing mathematical
models. You should know what the state of a system is and how to convert an ODE model into a state
model. You should know the difference between a nonlinear state model and a linear state model.

6. We learned how to linearize nonlinear state models in Section 2.5. You must be able to find equilibrium
configurations of nonlinear state models and then find the system’s linear state space approximation.

7. You must know the basic properties of a Laplace transform and be comfortable going from the time
domain to the s-domain and vice versa.

8. Section 2.8 is very important because most of the systems we study hereafter are modelled by transfer
functions. You should know what a rational transfer function is, what poles and zeros are, and what it
means for a TF to be proper. You should be able to obtain a TF model for a nonlinear state model near
an equilibrium configuration and you must be able to convert an LTI state model into a TF model.

9. You should understand why TF models are also called input-output models.

10. You should be able to reduce complex block diagrams into simpler forms using the basic rules in Section 2.9.

2.A Appendix: Linear functions

In this course we deal only with linear systems. Or rather, whenever we get a nonlinear system, we linearize
it as soon as possible. So we had all better be very certain about what a linear function is. Let us recall even
what a function is. If X and Y are two sets, a function from X to Y is a rule that assigns to every element
of X an unambiguous element of Y —there cannot be two possible different values for some x in X. The terms
function, mapping, and transformation are synonymous. The notation

f :X→Y

Version 1.2.2816, September 1, 2020

57 CHAPTER 2. MATHEMATICAL MODELS OF SYSTEMS

means that f is a function from X to Y . We typically write y = f (x) for a function. To repeat, for each x
there must be one and only one y such that y = f (x); y1 = f (x) and y2 = f (x) with y1 6= y2 is not allowed.
Let f be a function R → R. This means that f takes a real variable x and assigns a real variable y, written
y = f (x). So f has a graph in the (x, y) plane. To say that f is linear means the graph is a straight line through
the origin; there’s only one straight line that is not allowed—the y-axis. Thus y = ax defines a linear function
for any real constant a; the equation defining the y-axis is x = 0. The function y = 2x + 1 is not linear—its
graph is a straight line, but not through the origin.
In your linear algebra course you were taught that a linear function is a function f from a vector space X
to another (or the same) vector space Y having the property

f (a1 x1 + a2 x2 ) = a1 f (x1 ) + a2 f (x2 )

for all vectors x1 , x2 in X and all real numbers11 a1 , a2 . If the vector spaces are X = Rn , Y = Rm , and if f is
linear, then it has the form f (x) = Ax, where A is an m × n matrix. Conversely, every function of this form is
linear.
This concept extends beyond vectors to signals. For example, consider a capacitor, whose constitutive law
is
dv
i=C .
dt
Here, i and v are not constants, or vectors—they are functions of time. If we think of the current i as a function
of the voltage v, then the function is linear. This is because

d(a1 v1 + a2 v2 ) dv1 dv2

C = a1 C + a2 C .
dt dt dt
On the other hand, if we try to view v as a function of i, then we have a problem, because we need in addition
an initial condition v(0) (or some other initial time) to uniquely define v, not just i. Let us set v(0) = 0. Then
v is a linear function of i. You can see this from the integral form of the capacitor equation:

Zt
1
v(t) = i(τ )dτ.
C
0

2.B Appendix: The impulse and convolution

Now we take a little time to discuss the problematical object, the impulse δ(t). The impulse, also called the
Dirac delta function, is not really a legitimate function, because its “value” at t = 0 is not a real number. And
you can’t rigorously get δ as the limit of a sequence of ever-narrowing rectangles, because that sequence does
not converge in any ordinary sense. Yet the impulse is a useful concept in system theory and so we have to make
it legitimate. The French mathematician Laurent Schwartz worked out a very nice, consistent way of dealing
with the impulse, and more general “functions.” The impulse is an example of a distribution, sometimes
called a generalized function.
The main idea is that δ(t) is not a function, but rather it is a way of defining the linear transformation
φ 7→ φ(0) that maps a signal φ(t) to its value at t = 0. This linear transformation should properly be written
as δ(φ) = φ(0) (i.e., δ transforms φ to φ(0)) but historically it has been written as
Z∞
φ(t)δ(t)dt = φ(0).
−∞
11
This definition assumes that the vector spaces are over the field of real numbers.

Version 1.2.2816, September 1, 2020

 2.B. Appendix: The impulse and convolution 58

You know this as the “sifting formula.” Let us emphasize that the expression
Z∞
φ(t)δ(t)dt (2.28)
−∞

is not intended to mean integration of the product φ(t)δ(t) of functions—δ isn’t a function; rather, the expression
is that of an operation on φ whose value is defined to be φ(0). The expression is defined to be valid for all
functions φ(t) that are smooth at t = 0; smooth means continuously differentiable up to every order.
Needless to say, we have to be careful with δ; for example, there’s no way to make sense of δ 2 because the
expression (2.28) is not valid for φ = δ, again, because δ isn’t a function, let alone a smooth one. Because the
unit step is not smooth at t = 0, 1(t)δ(t) is undefined too. However, (2.28) does apply for φ(t) = e−st , because
it is smooth:
Z∞
e−st δ(t)dt = 1.
−∞

Thus the LT of δ equals 1.

Initial-value problems involving δ, such as

ẏ + 2y = δ,

or worse,
ẏ + 2y = δ̇,
require more advanced theory, because y(t) cannot be continuously differentiable at t = 0. Instead of pursuing
this direction, since initial-value problems are of minor concern in control we turn instead to convolution
equations.
As you learned in signals and systems, linear time-invariant systems are modelled in the time domain by a
convolution equation:
Z∞
y(t) = g(t − τ )u(τ )dτ.
−∞

If g(t) is a smooth function for all t, then expression (2.28) applies and we get
Z∞
g(t) = g(t − τ )δ(τ )dτ.
−∞

Thus g(t) equals the output when the input is the unit impulse. We call g(t) the impulse-response function,
or the impulse response. The case where g itself equals δ isn’t covered by what has been said so far, but
distribution theory can be used to justify δ ∗ δ = δ.

Example 2.B.1. Consider a low pass RC filter:

1
G(s) = .
RCs + 1
The inverse LT of G(s) is
1 − t
e RC .
RC
Since the filter is causal, to get the impulse response we should take that time-function that is zero for t < 0:
1 − t
g(t) = e RC 1(t).
RC

Version 1.2.2816, September 1, 2020

59 CHAPTER 2. MATHEMATICAL MODELS OF SYSTEMS

For the high pass filter,

RCs
G(s) =
RCs + 1
1 − t
g(t) = δ(t) − e RC 1(t).
RC
N

In general, if G(s) is strictly proper, then g(t) is a regular function. If G(s) is proper but not strictly proper,
then g(t) contains an impulse.
As a final comment, some texts (Dorf and Bishop [Dorf and Bishop, 2011]) give the LT pair

f˙(t) ←→ sF (s) − f (0− )

indicating that f (0) and f (0− ) may be different, that is, f (t) may be discontinuous at t = 0. If that is the case,
then f˙ has an impulse and distribution theory needs to be invoked. For more details see [Lundberg et al., 2007].

2.B.1 Summary
Rigorously speaking, the impulse is not a function. Instead, it is defined only indirectly via the sifting formula.
However, in the subject of control systems, and other signals and systems subjects, the impulse is used just as
though it were a function. Usually no harm is done, i.e., you get the right answer. But sometimes questions
arise that require a rigorous way to answer them; distribution theory is that way. For example, the derivative
of the impulse, δ̇(t), can be defined rigorously, and its LT is s. In this way we can obtain the impulse response
of the differentiator y = u̇—it is g(t) = δ̇(t). The transfer function is G(s) = s.

Version 1.2.2816, September 1, 2020

Chapter 3

Linear system theory

In the preceding chapter we saw nonlinear state models and how to linearize them about an equilibrium point.
A linearized system has the form (dropping the δ notation)

ẋ = Ax + Bu
y = Cx + Du.

or, using transfer functions,

Y (s)
= G(s) = C (sI − A)−1 B + D.
U (s)
In this chapter we study such models.
The main concepts are internal stability, bounded-input bounded-output stability, steady-state gain and
frequency response. We start with the time domain (state equations), then move into the s-domain (transfer
functions) and conclude with the frequency domain.

Contents
3.1 Initial-state response . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
3.2 Input-response . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
3.3 Total response . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
3.4 Stability of state-space models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
3.5 Bounded-input bounded-output stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
3.6 Steady-state gain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
3.7 Frequency response . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
3.8 Graphical representations of the frequency response . . . . . . . . . . . . . . . . . . . . . . . . . 76
3.9 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
3.A Appendix: Complex numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
3.B Appendix: Time domain, s-domain and frequency domain . . . . . . . . . . . . . . . . . . . . . . 91

3.1 Initial-state response

Consider an LTI state-space model where the input is set to zero

ẋ = Ax, x(0) = x0 , A ∈ Rn×n .

Recall two facts

1. If n = 1 then A is a scalar and the solution to the ODE is x(t) = eAt x0 .

60
61 CHAPTER 3. LINEAR SYSTEM THEORY

2. The Taylor series expansion of et at t = 0 is

t2
et = 1 + t + + ···
2
and this series converges for all t. Therefore, when A is a scalar,
A2 t2
eAt = 1 + At + + ··· .
2
The second fact suggests the following definition.

Definition 3.1.1. The matrix exponential of A ∈ Rn×n is

∞
A
X 1 k 1 1
e := A = I + A + A2 + A3 + · · · .
k! 2! 3!
k=0

It can be proved that the matrix exponential converges for every matrix A. If A is n × n then eA is too; if
A is not square, then eA is undefined.
Example 3.1.1.    
0 0 A 1 0
A= , e = .
0 0 0 1
Notice that eA is not obtained by taking the exponent of each element of A. N

Example 3.1.2. (Diagonal Matrix)

e3 0
   
3 0 A
A= , e = .
0 1 0 e

When A is a diagonal matrix then eA is obtained by taking the exponent of each diagonal element of A. N

Example 3.1.3. (Nilpotent Matrix)  

0 1 0
A =  0 0 1 .
0 0 0
This matrix has the property that A3 = 0, i.e., A is nilpotent1 . Therefore the power series of eA has only a
finite number of non-zero terms
1 1 12
 
1
e A = I + A + A2 =  0 1 1  .
2
0 0 1
N

Exercise 3.1. Compute eA for  

0 1
A= .
0 0
Using the matrix exponential we can define the matrix valued function R → Rn×n given by t 7→ etA . It has
the properties
1
A square matrix A is nilpotent if there exists an integer k ≥ 0 such that Ak = 0. It is nilpotent if, and only if, all of its
eigenvalues are zero.

Version 1.2.2816, September 1, 2020

 3.1. Initial-state response 62

1. etA t=0
=I (identity matrix).

2. et1 A et2 A = eA(t1 +t2 ) . Note that etA1 etA2 6= et(A1 +A2 ) . Equality holds if and only if A1 and A2 commute,
i.e., A1 A2 = A2 A1 .

−1
−1
3. eA = e−A which implies that eAt = e−At .

4. A and eAt commute.

d tA
5. dt e = AetA = etA A.
Theorem 3.1.2. The unique solution to ẋ = Ax, x(0) = x0 , A ∈ Rn×n is x(t) = eAt x0 .
If we take the Laplace transform of the equation ẋ = Ax without setting the initial conditions to zero, we
get that2
X(s) = (sI − A)−1 x0 .
Comparing this expression to the time-domain solution from Theorem 3.1.2 we have that eAt and (sI − A)−1
are Laplace transform pairs
L eAt = (sI − A)−1 .


This fact can be used to compute eAt by hand for small problems (n ≤ 3).
Example 3.1.4. Calculate etA for the matrix
 
0 1
A= .
−2 −3
n o
From above we know that etA = L −1 (sI − A)−1 . We proceed to perform the necessary calculations:
 
s −1
sI − A = ,
2 s+3

inverting a 2 × 2 matrix is easy

 
−1 1 s+3 1
(sI − A) = 2
s + 3s + 2 −2 s
 
1 s+3 1
= .
(s + 1)(s + 2) −2 s

Now use partial fractions on each entry of the above matrix

 2 1 1 1 
− s+2 −
(sI − A)−1 = s+1 −2 2
s+1
−1
s+2
2 .
s+1 + s+2 s+1 + s+2

Find the inverse Laplace transform of each entry in (sI − A)−1 to obtain

2e−t − e−2t e−t − e−2t

 
tA
e = .
−2e−t + 2e−2t −e−t + 2e−2t
N

Exercise 3.2. Compute eAt using the Laplace transform for the A matrix in Exercise 3.1.
Another way to compute eAt is by using eigenvalues and eigenvectors. We won’t cover the general theory
and instead present an example.
2
Here X(s) = L (x) is the component wise LT of the vector x. That is, if x(t) = (x1 (t), . . . , xn (t)) then X(s) = (X1 (s), . . . , Xn (s))

Version 1.2.2816, September 1, 2020

63 CHAPTER 3. LINEAR SYSTEM THEORY

Example 3.1.5.  
0 1
A= .
−2 −3
The MATLAB command [V, D] = eigs(A) produces
   
1 −1 −1 0
V = , D= .
−1 2 0 −2

The eigenvalues of A are the diagonal elements of D. The columns of V are the corresponding eigenvectors.
So, for example,    
1 1
A = −1 .
−1 −1
It follows that AV = V D (verify this) and therefore that eAt V = V eDt (prove this). The nice thing about eDt
is that it is easy to compute since D is a diagonal matrix (cf. Example 3.1.2). In this case
 −t 
Dt e 0
e = .
0 e−2t

Then
e−t 2e−t − e−2t e−t − e−2t
     
At Dt −1 1 −1 0 2 1
e =Ve V = = .
−1 2 0 e−2t 1 1 −t
−2e + 2e −2t −e−t + 2e−2t
N

The approach in Example 3.1.5 works when A has n linearly independent eigenvectors so that V is invertible.
Otherwise the theory is more complicated and requires the Jordan form of the matrix A.

Remark 3.1.3. Examples 3.1.4 and 3.1.2 present two different methods of computing the function eAt . For
large problems, both methods are tedious to apply by hand. Programs like MATLAB, Mathematica and Maple
can do these calculations. The command in MATLAB is expm. 

3.2 Input-response
Now set the initial state to zero and consider the response from the input:

ẋ = Ax + Bu, x(0) = 0. (3.1)

We can use the properties of the matrix exponential to solve this equation. Here’s a derivation of the solu-
tion [Chen, 1999]. Write
ẋ − Ax = Bu.
Multiply by e−tA :
e−tA ẋ − Ae−tA x = e−tA Bu.
d
e−tA x(t) . Thus
 
The left-hand side equals dt

d  −tA
x(t) = e−tA Bu.

e
dt
Integrate from t = 0 to t:
Zt
−tA
e x(t) − x(0) = e−τ A Bu(τ )dτ.
|{z}
=0 0

Version 1.2.2816, September 1, 2020

 3.2. Input-response 64

Multiply by etA
Zt
x(t) = e(t−τ )A Bu(τ )dτ. (3.2)
0

This is the solution of (3.1). It gives the value of the state and at time t as a function of u(τ ), 0 ≤ τ ≤ t when
the initial state equals zero. Substituting the solution into the output equation leads to
Zt
y(t) = CeA(t−τ ) Bu(τ )dτ + Du(t). (3.3)
0

Special case: Single-input single-output

In the special case when dim u = dim y = 1, i.e., the system is single-input, single-output, then D a scalar and
so is y(t) in (3.3). If u = δ, the unit impulse, then

y(t) = CeAt B1(t) + Dδ(t),

where 1(t) denotes the unit step. We conclude that the impulse response of the system is

g(t) = CeAt B1(t) + Dδ(t) (3.4)

and equation (3.3) is a convolution equation

y(t) = (g ∗ u)(t).

Example 3.2.1. Consider a single-input single output system in state-space form

ẋ(t) = −3x(t) + 2u(t)

y(t) = 4x(t).

According to the development in this section, this system’s input-response is given by the integral (convolution)
equation (3.3) which for this example is

Zt
y(t) = 8e−3(t−τ ) u(τ )dτ.
0

If we apply a unit step input u(t) = 1(t), then

Zt
y(t) = 8 e−3(t−τ ) dτ
0
Zt
−3t
= 8e e3τ dτ
0
8
= e−3t e3t − 1


3
8
1 − e−3t .


=
3
N

Version 1.2.2816, September 1, 2020

65 CHAPTER 3. LINEAR SYSTEM THEORY

3.3 Total response

Consider the state equation forced by both an initial state and an input:

ẋ = Ax + Bu, x(0) = x0 .

The system is linear in the sense that the state at time t equals the initial-state-response at time t plus the
input-response at time t
Zt
x(t) = e x0 + eA(t−τ ) Bu(τ )dτ.
At

0
Similarly, the output is
Zt
At
y(t) = Ce x0 + CeA(t−τ ) Bu(τ )dτ + Du(t).
0
These two equations constitute a solution in the time domain.

Summary
We began with an LTI system modelled by a differential equation in state form:

ẋ = Ax + Bu, x(0) = x0
y = Cx + Du.

We solved the equations to get

Zt
At
x(t) = e x0 + eA(t−τ ) Bu(τ )dτ
0
(3.5)
Zt
y(t) = CeAt x0 + CeA(t−τ ) Bu(τ )dτ + Du(t).
0

These are integral (convolution) equations giving x(t) and y(t) explicitly in terms of x0 and u(τ ), 0 ≤ τ ≤ t.
In the SISO case, if x0 = 0 then
Zt
y = g ∗ u = g(t − τ )u(τ )dτ
0

where g(t) = CeAt B1(t) + Dδ(t).

Example 3.3.1. (Mass-Spring) We illustrate these calculations in an example. We stress that normally time
domain solutions are obtained through simulations. These calculations are useful for gaining intuition about
the structure of the solution (3.5) and especially for understanding the stability arguments in the next section.
Recall the mass-spring-damper system of Example 2.1.1 with linear damping

M q̈ + bq̇ + kq = u.

Suppose M = 1, k = 4 and b = 0, i.e., no damping. The state equation is

ẋ1 = x2
ẋ2 = −4x1 + u

Version 1.2.2816, September 1, 2020

 3.4. Stability of state-space models 66

where (x1 , x2 ) := (q, q̇). We are interested in the solution to this system when u(t) = 1(t) and x1 (0), x2 (0) are
arbitrary. We compute
( −1 ) s 1
cos (2t) 12 sin (2t)
   
−1 −1 −1 s −1 −1
e =L
At
=L =L
2 +4 2 +4
 s s
(sI − A) = .
4 s − s24+4 s2s+4 −2 sin (2t) cos (2t)

We can now compute

   1
  Zt  1
 
x1 (t) cos (2t) 2 sin (2t) x1 (0) cos (2(t − τ )) 2 sin (2(t − τ )) 0
= + dτ
x2 (t) −2 sin (2t) cos (2t) x2 (0) −2 sin (2(t − τ )) cos (2(t − τ )) 1(τ )
0
 1
  Zt  1

cos (2t) 2 sin (2t) x1 (0) 2 sin (2(t − τ ))
= + dτ
−2 sin (2t) cos (2t) x2 (0) cos (2(t − τ ))
0
 1
  Zt  1

cos (2t) 2 sin (2t) x1 (0) 2 sin (2λ)
= + dλ (change of variable λ := t − τ )
−2 sin (2t) cos (2t) x2 (0) cos (2λ)
0
1 1
    
cos (2t) 2 sin (2t) x1 (0) 4 (1 − cos (2t))
= + 1 .
−2 sin (2t) cos (2t) x2 (0) 2 sin (2t)

Figure 3.1 shows simulation results for this system when x1 (0) = 0 and x2 (0) = 0. Figure 3.1a is a plot of
(x1 (t), x2 (t)) as they evolve in the plane. It is called a phase plot and the curve is called a phase curve.
Figure 3.1b plots x1 and x2 versus time. If we are interested in controlling the position of the mass then we
0.5
x 1(t)
x 2(t)
0.5

0.4

0.4

0.3

0.3

0.2

0.2

0.1
0.1

0
x 2 (t)

-0.1
-0.1

-0.2 -0.2

-0.3 -0.3

-0.4
-0.4

-0.5
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 -0.5
x 1 (t) 0 2 4 6 8 10 12 14 16 18 20

(a) Phase curve. (b) Position (x1 , black) & velocity (x2 , grey) vs. time.

Figure 3.1: Time domain response of a mass-spring system with zero initial conditions and unit step input.

take x1 as the system output while if we were interested in velocity control we take x2 as the output. N

3.4 Stability of state-space models

In this course there are two different notions of stability. The one we study in this section concerns state-space
models. The stability question is: For the system

ẋ = Ax (3.6)

will x(t) converge to 0 as t goes to infinity, for every x(0)?

Version 1.2.2816, September 1, 2020

67 CHAPTER 3. LINEAR SYSTEM THEORY

Definition 3.4.1. System (3.6) is asymptotically stable if x(t) → 0 as t → ∞ for every initial condition
x(0).

As we saw before, the trajectory of (3.6) is specified by x(t) = eAt x(0). So asymptotic stability is equivalent
to the condition that eAt → 0 as t → ∞.
Example 3.4.1. Recall the matrix eAt for the mass-spring system from Example 3.3.1
cos (2t) 21 sin (2t)
 
At
e = .
−2 sin (2t) cos (2t)
Since eAt does not approach zero as t → ∞ this system is not asymptotically stable. Physically this corresponds
to the mass oscillating back and forth indefinitely if it is initialized with a non-zero initial position or velocity
and no applied force. Our intuition suggests that friction should prevent such a phenomenon. Indeed, if we set
b = 1 so that there is a damper to model friction, then (cf. Example 3.3.1)
 
0 1
A=
−4 −1
and t √ √ √ √ √
e− 2
 
15 cos ( 15t/2) + √15 sin (√15t/2) 2√15 sin (√15t/2)
eAt = √ .
15 −8 15 sin ( 15t/2) 15 cos ( 15t/2) − 15 sin ( 15t/2)
In this case eAt → 0 as t → ∞ and the system is asymptotically stable. This is consistent with our physical
intuition that friction should cause energy loss in the system eventually resulting in the mass returning to its
rest position. N
Fortunately, we do not need to compute eAt to determine whether or not (3.6) is asymptotically stable.
Asymptotic stability only depends on the eigenvalues of A.
Proposition 3.4.2. eAt → 0 as t → ∞ if, and only if, every eigenvalue of A has negative real part.
Proof. If A has n linearly independent eigenvectors, then it can be diagonalized by a similarity transformation,
so that
eAt = V eDt V −1 .
Then
eAt → 0 ⇐⇒ eDt → 0
⇐⇒ eλi t → 0, ∀i
⇐⇒ Re λi < 0, ∀i.
The proof in the general case is more complicated but not more enlightening.

Example 3.4.2. Re-visiting the systems from Example 3.4.1, when there is no damping
 
0 1
A=
−4 0
and the eigenvalues of A are found by solving det (sI − A) = 0 to yield s = ±j2. The eigenvalues have zero
real part and therefore by Proposition 3.4.2 the system is not asymptotically stable.
When there is friction  
0 1
A=
−4 −1
√
and the eigenvalues are − 12 1 ± j 15 . By Proposition 3.4.2 the system is asymptotically stable. These

conclusions are consistent with the observations made in Example 3.4.1. N

Version 1.2.2816, September 1, 2020

 3.5. Bounded-input bounded-output stability 68

3.5 Bounded-input bounded-output stability

Consider an LTI system with a single input, a single output, and a strictly proper rational transfer function.
The model is therefore y = g ∗ u in the time domain, or Y (s) = G(s)U (s) in the s-domain. We ask the
question: Does a bounded input (BI) always produce a bounded output (BO)? First, let’s define precisely what
boundedness of a signal means.

Definition 3.5.1. Let u(t) be a real-valued signal defined for t ≥ 0. We say u is bounded if there exists a
constant b such that, for all t ≥ 0, |u(t)| ≤ b.

Familiar bounded signals are steps u(t) = 1(t) and sinusoids u(t) = sin (t), but not ramps u(t) = t nor the
exponential u(t) = et . The least upper bound of a signal u is denoted kuk∞ . You can think of3 kuk∞ as
maxt≥0 |u(t)|.

Definition 3.5.2. A linear time-invariant system with input u and output y is Bounded-Input Bounded-
Output (B.I.B.O.) stable if every bounded input u produces a bounded output y, i.e.,

kuk∞ finite ⇒ kyk∞ finite.

The system is unstable if it is not BIBO stable.

Example 3.5.1. Consider a system with transfer function

1
G(s) = .
s+2
Is this system BIBO stable? The impulse response of this system is

g(t) = L −1 {G(s)} = e−2t .

Then
y(t) = (g ∗ u)(t)
Zt
= e−2τ u(t − τ )dτ.
0
3
To be more precise, kuk∞ := supt≥0 |u(t)| where sup is the supremum.

Version 1.2.2816, September 1, 2020

69 CHAPTER 3. LINEAR SYSTEM THEORY

This implies
Zt
|y(t)| = e−2τ u(t − τ )dτ
0
Zt
≤ e−2τ u(t − τ ) dτ
0
Zt
= e−2τ |u(t − τ )| dτ
0
Zt
≤ e−2τ dτ kuk∞
0
1
≤ kuk∞ .
2
In other words
1
kyk∞ ≤ kuk∞
2
which means that every bounded u produces a bounded y. N

Example 3.5.2. Consider a system with transfer function

1
G(s) = .
s−2
Is this system BIBO stable? The impulse response of this system is

g(t) = L −1 {G(s)} = e2t .

Then
y(t) = (g ∗ u)(t)
Zt
= e2τ u(t − τ )dτ.
0

If we pick u(t) = 1(t) then

Zt
1 2t
e2τ dτ =


y(t) = e −1 , t≥0
2
0
which is unbounded. N
Remark 3.5.3. BIBO stability asks that every bounded input produce a bounded output. If you pick one
bounded input and the output happens to be bounded, you cannot conclude that the system is BIBO stable.
Sometimes you have to be clever about how you choose bad inputs. In the above example the input
s−2
U (s) = ,
s(s + 1)
which corresponds to the bounded signal

u(t) = −2 + 3e−t , t ≥ 0,

produces a bounded output. 

Version 1.2.2816, September 1, 2020

 3.5. Bounded-input bounded-output stability 70

Fortunately, we do not need to compute convolution integrals to determine if a system is BIBO stable.
BIBO stability only depends on the poles of G(s).
Theorem 3.5.4. Assume G(s) is strictly proper, rational. Then the following three statements are equivalent:
1. The system is BIBO stable.
R∞
2. The impulse-response function g(t) is absolutely integrable, i.e., |g(t)|dt < ∞.
0

3. Every pole of the transfer function G(s) has negative real part.

Example 3.5.3. (Low Pass Filter) Consider the RC filter from Example 2.8.2
1 1 −t/RC
G(s) = , g(t) = e 1(t).
RCs + 1 RC
1
The transfer function G(s) has one pole at s = − RC . Since R is a resistance and C is a capacitance, they are
both positive and so the pole has negative real part. According to the theorem, we expect that every bounded
u produces a bounded y. Let’s see
Zt
|y(t)| = g(τ )u(t − τ )dτ
0
Zt
≤ |g(τ )||u(t − τ )|dτ
0
Zt
≤ |g(τ )|dτ kuk∞
0
Z∞
≤ |g(τ )|dτ kuk∞
0
Z∞
1 −τ /RC
= e dτ kuk∞
RC
0
= kuk∞ .
Thus kyk∞ ≤ kuk∞ which confirms, as expected, that every bounded input produces a bounded output. N

Example 3.5.4. (Integrator) The integrator has

1
G(s) = , g(t) = 1(t).
s
According to the theorem, the system is not BIBO stable; there exists some bounded input that produces an
unbounded output. For example the unit step u(t) = 1(t) is bounded, but it produces the output y(t) = t1(t),
which is an unbounded ramp. It is not true that every bounded input produces an unbounded output, only
that some bounded input does. For example, if the input is u(t) = (sin t)1(t), then the output is bounded. N
1 1 1 1 1
Using Theorem 3.5.4 we conclude that s+1 , (s+1)2 , s2 +2s+1 are BIBO stable, but not s nor s−1 .

The theorem can be extended to the case where G(s) ∈ R(s) is only proper (and not strictly proper). Then
write
G(s) = G1 (s) + c, G1 (s) strictly proper, c ∈ R.

Version 1.2.2816, September 1, 2020

71 CHAPTER 3. LINEAR SYSTEM THEORY

Then the impulse response has the form

g(t) = g1 (t) + cδ(t).
Theorem 3.5.4 remains true with the second statement changed to say that g1 (t) is absolutely integrable. On
the other hand, the next result shows that improper systems are never BIBO stable.

Theorem 3.5.5. If G(s) ∈ R(s) is improper, then G(s) is not BIBO stable.

Proof. Write the transfer function G(s) as

G(s) = G1 (s) + G2 (s)

where G1 (s) ∈ R(s) is strictly proper and G2 (s) ∈ R[s] has degree at least 1. For any input u the output y is,
by linearity, y = y1 + y2 where

Y1 (s) = G1 (s)U (s), Y2 (s) = G2 (s)U (s).

If G1 (s) has poles with non-negative real parts, then the result follows from Theorem 3.5.4. So let’s assume
G1 (s) has all its poles in C− and consider the bounded input u(t) = sin (t2 ). Since G1 has all poles in C− , the
signal y1 is bounded. We’ll now show that y2 is not bounded. Any derivative of u will involve terms polynomial
in t and such terms will not be bounded as t → ∞. But y2 is a linear combination of u and its derivatives so
the result follows.

Theorem 3.5.5 shows that the differentiator system G(s) = s is not BIBO stable.

3.5.1 Stability of state-space models and BIBO stability

Consider a single-input single-output system modelled by

ẋ = Ax + Bu
y = Cx + Du.
or
Y (s)/U (s) = G(s)
= C (sI − A)−1 B + D
1
= C adj (sI − A)B + D.
det (sI − A)
From the last expression we see that the poles of G(s) are contained in the eigenvalues of A. Thus

asymptotic stability of state-space model =⇒ BIBO stability.

Usually the poles of G(s) are identical to the eigenvalues of A, that is, the two polynomials

det (sI − A), C adj (sI − A)B + D det (sI − A)

have no common roots, i.e., they’re coprime. In these cases, the two stability concepts are equivalent.

Example 3.5.5. (Mass-Spring) The A matrix for the mass-spring system from Example 3.3.1 was
 
0 1
A= .
−4 0

The eigenvalues of A are found by solving

det (sI − A) = s2 + 4 = 0

Version 1.2.2816, September 1, 2020

 3.5. Bounded-input bounded-output stability 72

to get s = ±j2. By Proposition 3.4.2 this system is not asymptotically stable. The system’s transfer function
is given by G(s) = C(sI − A)−1 B + D. The matrices C, D depend on what we want to chose as the output. If
we take the position of the mass as the output then
 s 1  
  s 2 +4 s 2 +4 0 1
G(s) = 1 0 = 2 .
− s24+4 s2s+4 1 s +4

By Theorem 3.5.4 the system is not BIBO stable. In this case the two notions of stability are equivalent because
the polynomials
det(sI − A) = s2 + 4
and   
  s 1 0
C adj (sI − A)B + D det (sI − A) = 1 0 =1
−4 s 1
are coprime. N

Example 3.5.6. Consider the parallel interconnection in Figure 3.2. If we take the state vector to be x =

Figure 3.2: Parallel connection of first order systems.

(x1 , x2 ), then the state-space model is    

1 0 1
ẋ = x+ u
0 1 1
 
y = 1 −1 − u.
The eigenvalues of the A matrix are {1, 1} 1 so that by Proposition 3.4.2 this system is not asymptotically
stable. The system’s transfer function is
 1  
  s−1 0 1
G(s) = 1 −1 1 − 1 = −1.
0 s−1 1

By Theorem 3.5.4 the system is BIBO stable. In this case the two notions of stability are not equivalent because
the polynomials
det(sI − A) = (s − 1)2
and   
s−1 0 1
− (s − 1)2 = 1
 
C adj (sI − A)B + D det (sI − A) = 1 −1
0 s−1 1
have roots in common. N

Exercise 3.3. Show that the two notions of stability presented are not equivalent for the system
      
ẋ1 0 1 x1 0
= + u
ẋ2 1 0 x2 1
 
  x1
y = 1 −1 .
x2

Version 1.2.2816, September 1, 2020

73 CHAPTER 3. LINEAR SYSTEM THEORY

3.6 Steady-state gain

In this section we show that if we apply a constant input to a BIBO stable system then the steady-state
output is also constant. The ratio of the steady-state output to the constant input is also a constant, i.e., it
doesn’t depend on the magnitude of the input. The fundamental result used in this section is the Final-Value
Theorem (F.V.T.).
Theorem 3.6.1 (Final-value theorem). Let f (t) be a signal defined for t ≥ 0 and let its Laplace transform F (s)
be rational and proper.
(a) If F (s) has all its poles in C− , then f (t) converges to zero as t → ∞.
(b) If F (s) has all its poles in C− except for a simple pole at s = 0, then
lim f (t) = lim sF (s). (3.7)
t→∞ s→0

(c) In all other cases, f (t) does not approach a constant as t → ∞.

Remember, you have to know that f (t) has a final value, by examining the poles of F (s), before you can
apply the final value theorem. Equation (3.7) is valid if sF (s) has no poles in the closed right half complex
+
plane C .
Theorem 3.6.2. If G(s) is BIBO stable and u(t) = b1(t), b a real constant, then
yss := lim y(t) = bG(0).
t→∞

Proof. If u(t) = b1(t) then

Zt Zt
y(t) = g(τ )u(t − τ )dτ = b g(τ )dτ
0 0
where g(t) = L −1 {G(s)} is the impulse response. This implies that
Z∞
yss = b g(τ )dτ = bG(0).
0

In the last deduction we used the definition of the Laplace Transform (2.24) with s = 0. We are allowed to
make the substitution s = 0 in (2.24) because G(s) is BIBO stable which means, by Theorem 3.5.4, that all its
poles are to the left of the imaginary axis and hence s = 0 is in the region of convergence of (2.24).

Exercise 3.4. Re-prove Theorem 3.6.2 using the final-value theorem.

Definition 3.6.3. Let G(s) be a BIBO stable transfer function and let u(t) = b1(t), b a real constant. The
steady-state gain of G(s) is
yss
= G(0).
b

Exercise 3.5. Using MATLAB, plot the unit step responses of

20 −20s + 20
G1 (s) = , G2 (s) = .
s2 + 0.9s + 50 s2 + 0.9s + 50
They have the same steady-state gains and the same poles, but notice the big difference in transient response.

Version 1.2.2816, September 1, 2020

 3.7. Frequency response 74

3.7 Frequency response

If you apply a sinusoidal input to a BIBO stable LTI system, then, in steady-state, the output is also a sinusoid
of the same frequency with a possible change in magnitude and phase. This is illustrated in Figure 3.3.

Figure 3.3: Steady-state output of a BIBO stable LTI system excited by a sinusoidal input.

To understand the above fact, consider a single-input, single-output LTI system. It will then be modelled
by
y =g∗u or Y (s) = G(s)U (s).
Let us assume G(s) is rational, proper, and has all its poles in C− . Then the system is BIBO stable by
Theorem 3.5.4. The first fact we want to see is this: Complex exponentials are eigenfunctions4 .

Proof.
Z∞
jωt
u(t) = e , yss = g(t − τ )u(τ )dτ.
−∞

Notice that the convolution integral used in this proof goes from t = −∞ to t = +∞. That is, the input was
applied starting at t = −∞. If the time of application of the sinusoid is t = 0, there is a transient component
in y(t) too. Taking the limits as above removes the transient terms. Continuing we have
Z∞
yss = g(τ )ejω(t−τ ) dτ
−∞
Z∞
= g(τ )ejωt e−jωτ dτ
−∞
Z∞
 

= g(τ )e−jωτ dτ  ejωt

−∞
= G(jω)ejωt .

Thus, if the input is the complex sinusoid ejωt , then the steady-state output is the complex sinusoid

G(jω)ejωt = |G(jω)|ej(ωt+]G(jω)) . (3.8)

Now consider the input5

ejωt + e−jωt
u(t) = cos (ωt) = .
2
4
The equation g ∗ ejωt = G(jω)ejωt has the same form as the equation Ax = λx where A is a square matrix, x is a vector and λ
is a complex number. We say that λ is an eigenvalue of A and x is an eigenvector. Likewise, in the preceding equation we say that
sinusoids are eigenfunctions of LTI systems and Fourier transforms G(jω) are eigenvalues.
5
Here we are using Euler’s formula ejx = cos (x) + j sin (x).

Version 1.2.2816, September 1, 2020

75 CHAPTER 3. LINEAR SYSTEM THEORY

By the above analysis and linearity of the system, the steady-state output is6
1 1
y(t) = G(jω)ejωt + G(−jω)e−jωt
2  2
jωt
= Re G(jω)e .

Expressing G(jω) in polar form

G(jω) = |G(jω)| e]G(jω)
we get that
y(t) = |G(jω)| cos (ωt + ]G(jω)).
In conclusion, we have shown the following.

Theorem 3.7.1. Assume G(s) is rational, proper, and has all its poles in C− . Then the steady-state response
to the input u(t) = aejωt is
y(t) = aG(jω)ejωt . (3.9)
In particular

• The steady-state response to the input u(t) = a cos (ωt) is

y(t) = a |G(jω)| cos (ωt + ]G(jω)). (3.10)

• The steady-state response to the input u(t) = a sin (ωt) is

y(t) = a |G(jω)| sin (ωt + ]G(jω)). (3.11)

• The steady-state response to the input u(t) = a1(t) is

y(t) = aG(0). (3.12)

Therefore the response of a BIBO stable LTI system to a continuous-time sinusoid input of frequency ω is
also a continuous-time sinusoid of the same frequency. The amplitude of the output sinusoid is |G(jω)| times
the input amplitude, and the phase of the output is shifted by ]G(jω) with respect to the input phase. This
justifies the picture in Figure 3.3 and leads to the following definition.

Definition 3.7.2. Assume G(s) is rational, proper, and has all its poles in C− .

(a) The function R → C, ω 7→ G(jω) is the frequency response of G.

(b) The function R → R, ω 7→ |G(jω)| is the amplitude or magnitude response of G.

(c) The function R → (−π, π], ω 7→ ]G(jω) is the phase response of G.

Hence, the frequency response of a BIBO stable continuous-time LTI system G(s) is determined by setting
s = jω and varying ω. Here’s the important point: Under the stated assumptions about the system (G(s)
is rational, proper, and all poles in C− ), the region of convergence of the Laplace transform includes the
imaginary axis. Therefore it is legitimate to set s = jω in G(s) to get G(jω), which, by the way, equals the
Fourier transform of the impulse response g(t).
6
Here we are using the fact that if G(s) ∈ R(s), then G(−jω) = G(jω). See Remark 3.8.1.

Version 1.2.2816, September 1, 2020

 3.8. Graphical representations of the frequency response 76

3.8 Graphical representations of the frequency response

One of the reasons why frequency response is a powerful tool is that it is possible to succinctly understand its
primary features by means of plotting functions or parameterized curves. In this section we see how this is
done.
Example 3.8.1. Consider a system with transfer function
−10
G(s) = .
s + 10
Its frequency response is the function
−10
G(jω) = .
jω + 10
The magnitude and phase responses are plotted in Figure 3.4 for ω ∈ [−100, 100]. The basic plots can be
generated in MATLAB using the script below.

1 w = −100:0.01:100;
2 s = 1i.*w;
3 G = −10./(s+10);
4 plot(w, abs(G), 'LineWidth', 2); grid on;
5 figure; plot(w, rad2deg(angle(G)), 'LineWidth', 2); grid on;

200
0.9

150
0.8

0.7
100

0.6
50

0.5

0.4

-50

0.3

-100

0.2

-150
0.1

0 -200
-100 -80 -60 -40 -20 0 20 40 60 8 00 -100 -80 -60 -40 -20 0 20 40 60 80 100

(a) Magnitude response. (b) Phase response.

Figure 3.4: A plot of the magnitude and phase response for Example 3.8.1.

N
Remark 3.8.1. From now on, when we graphically represent the functions from Definition 3.7.2 we will only
consider non-negative values of ω. One may think that we are losing information by not considering negative
values of frequency. This is not the case because when G(s) ∈ R(s) (real rational)7 , then its magnitude is
symmetric about ω = 0, i.e., |G(jω)| = |G(−jω)| and its phase is anti-symmetric about ω = 0, i.e., ](G(jω)) =
−](G(−jω)). See Figure 3.4. 

3.8.1 Polar plots

The idea here is to plot the frequency response G(jω) as a curve in the complex plane parametrized by ω ∈ [0, ∞),
i.e., we draw Im (G(jω)) vs Re (G(jω)) as ω ranges from 0 to ∞. The resulting curve is called a polar plot of
the frequency response. While polar plots aren’t used too much for design, they provide insight into the shape
of a Bode plot and are fundamental to Nyquist plots (see Chapter 8). In anticipation of asymptotic Bode plots
(Section 3.8.2), we present the polar plots for four basic transfer functions.
7
The remark remains true when the transfer function has a time-delay, i.e., is of the form e−sT G(s) with T > 0 and G(s) ∈ R(s).

Version 1.2.2816, September 1, 2020

77 CHAPTER 3. LINEAR SYSTEM THEORY

Pure gain
Consider G(s) = K. Then G(jω) = K and the frequency response is constant. Figure 3.5 shows the polar
plots.

(a) K > 0. (b) K < 0.

Figure 3.5: Polar plot of frequency response for G(s) = K.

Zero at the origin

Consider G(s) = sn . Then G(jω) = (jω)n . When n = 1, G(jω) = jω and is therefore an imaginary number.
When n = 2, G(jω) = −ω 2 is a negative real number. Continuing this way for different values of n we get the
polar plot in Figure 3.6.

Figure 3.6: Polar plot of frequency response for G(s) = sn .

Real zero
Consider G(s) = τ s ± 1 with τ > 0. For τ s + 1 the zero is located at s = − τ1 and lies in C− . For τ s − 1 the
zero is located at s = τ1 and lies in C+ . The frequency response is G(jω) = jτ ω ± 1. Figure 3.7 shows the polar
plots for both cases.

Complex conjugate zeros

Consider the case when G(s) has complex conjugate zeros
 2 
2 2 2 s ζ
G(s) = s + 2ζωn s + ωn = ωn +2 s+1 , ζ ∈ [0, 1), ωn 6= 0. (3.13)
ωn2 ωn

Version 1.2.2816, September 1, 2020

 3.8. Graphical representations of the frequency response 78

(a) jτ ω − 1. (b) jτ ω + 1.

Figure 3.7: Polar plot of frequency response for G(s) = τ s ± 1, τ > 0.

We consider three sub-cases. The first is when ζ = 0 so that

 2 
2 s
G(s) = ωn +1 , ζ = 0, ωn 6= 0
ωn2
has purely imaginary roots. Then
ω2
 
G(jω) = ωn2 1− 2
ωn
is a real number and its polar plot is shown in Figure 3.8. Note that when ω = ωn , G(jω) = 0 and when
ω changes sign the phase of G(jω) changes abruptly from 0◦ to 180◦ . The second sub-case is when G(s) has

Figure 3.8: Polar plot of frequency response for G(s) = s2 + ωn2 , purely imaginary roots.

complex conjugate roots in C−

s2
 
ζ
G(s) = ωn2 +2 s+1 , ζ ∈ (0, 1), ωn > 0.
ωn2 ωn
In this case
ω2
  
ζ
G(jω) = ωn2 1 − 2 + j2 ω .
ωn ωn
The polar plot of G(jω) is shown in Figure 3.9a. The last sub-case is when G(s) has complex conjugate roots
in C+  2 
2 s ζ
G(s) = ωn − 2 s + 1 , ζ ∈ (0, 1), ωn > 0
ωn2 ωn

Version 1.2.2816, September 1, 2020

79 CHAPTER 3. LINEAR SYSTEM THEORY

for which
ω2
  
ζ
G(jω) = ωn2 1 − 2 − j2 ω .
ωn ωn
The polar plot of G(jω) is shown in Figure 3.9b. Note that when making a polar plot, the thing one loses is

(a) Complex conjugate roots in C− . (b) Complex conjugate roots in C+ .

Figure 3.9: Polar plot of frequency response for G(s) = s2 ± 2ζωn s + ωn2 for ζ ∈ (0, 1), ωn > 0.

frequency information. That is, one can no longer read from the plot the frequency at which, say, the magnitude
of the frequency response is maximum.
Example 3.8.2. (Non-Minimum Phase) Consider the two rational transfer functions
1+s 1−s
G1 (s) = , G2 (s) = .
s2 +s+1 s2 +s+1
√

On the one hand, both transfer functions are BIBO stable since they both have poles at s = − 12 1 ± 3 ∈ C− .
On the other hand, G1 (s) has a zero at s = −1 ∈ C− while G2 (s) has a zero at s = 1 ∈ C+ . For reasons that
will become apparent shortly, systems with zeros in C− are called minimum phase while systems with zeros
in C+ are called non-minimum phase. The polar plots of their frequency responses
1 + jω 1 − jω
G1 (jω) = , G2 (jω) =
−ω 2 + jω + 1 −ω 2 + jω + 1
are shown in Figure 3.10. As the names suggest, the minimum phase system undergoes a smaller phase change
as we move along its parameterized polar curve. N
0.4

0.2
0

-0.2

-0.2

-0.4
-0.4

-0.6

-0.6

-0.8

-0.8 -1

-1.2

-1

-1.4

-1.2 -1.6
0 0.2 0.4 0.6 0.8 1 1.2 1.4 -1.5 -1 -0.5 0 0.5 1

(a) Polar plot of G1 (jω). (b) Polar plot of G2 (jω).

Figure 3.10: Polar plots of minimum and non-minimum phase transfer functions in Example 3.8.2.

Version 1.2.2816, September 1, 2020

 3.8. Graphical representations of the frequency response 80

Definition 3.8.2. A rational, proper transfer function is minimum phase if all its zeros have Re (s) ≤ 0.
The transfer function is non-minimum phase if it is not minimum phase.

3.8.2 Bode plots

Control design is typically done in the frequency domain using Bode plots. A Bode plot is simply a plot of the
amplitude and phase response functions of G, see Definition 3.7.2. These plots, however, are done in a very
particular way: we plot 20 log |G(jω)| vs ω and ]G(jω) vs ω with ω plotted on log scales (all logarithms are
base 10 in this course). The two plots (i) 20 log |G(jω)| vs log (ω) and (ii) ]G(jω) vs log (ω) together are called
the Bode plot of G(s). The units of the plot of 20 log |G(jω)| are called decibels (dB).
A Bode plot tells us how the system G(s) responds in steady-state to sinusoidal inputs. Since the Fourier
series can be used to represent a large class of signals, Bode plots can be used to tell us about a system response
to almost all input signals.

Example 3.8.3. (RLC Circuit) In this example we’ll use software to create the Bode plot of the RLC circuit
-20

-25

-30

-35
Magnitude (dB)

-40

-45

-50

-55

-60

-65

-70
90

45
Phase (deg)

-45

-90
10 -3 10 -2 10 -1 10 0 10 1 10 2 10 3
Frequency (rad/s)

Figure 3.11: Bode plot of RLC circuit with R = 10, L = 1, C = 1.

from Example 2.3.5. Later we’ll discuss how to sketch them by hand. The governing equation for the circuit is
1
−u̇ + Rẏ + Lÿ + y = 0.
C
The TF of this system is (verify)
Y (s) s
= G(s) = 1
U (s) Ls2 + Rs + C
and the frequency response is
jω
G(jω) = 1 .
−jLω 2 + jRω + C
Figure 3.11 shows the Bode plot for this system and was generated using the following script in MATLAB

Version 1.2.2816, September 1, 2020

81 CHAPTER 3. LINEAR SYSTEM THEORY

1 s = tf('s');
2 R =10; L = 1; C = 1;
3 G = s/(L*sˆ2 + R*s + 1/C);
4 bode(G);
5 grid on;

Example 3.8.4. (Low Pass Filter) Recall the TF of the so-called low pass RC filter from Example 2.8.2
0

-5

-10
Magnitude (dB)

-15

-20

-25

-30

-35

-40
0

-30
Phase (deg)

-60

-90
2 3 4 5 6
10 10 10 10 10
Frequency (rad/s)

Figure 3.12: Bode plot of RC circuit with RC = 10−4 .

Y (s) 1
= G(s) = .
U (s) RCs + 1
The associated frequency response is
1
.
G(jω) =
jRCω + 1
Figure 3.12 shows the Bode plot for this system with RC = 10−4 . The Bode plot provides justification for the
name low-pass filter. From the plot of the magnitude response we see that when we apply sinusoidal inputs
with frequencies above 104 rad/s, the amplitude of the steady-state output signal is heavily attenuated. For
instance, if the input has amplitude one and frequency ω = 105 , then 20 log |G(jω)| ≈ −20dB and therefore the
amplitude of the output is |G(jω)| ≈ 10−1 . N
The above example helps to motivate the notion of bandwidth for control systems. The term bandwidth
is used by engineers in many different contexts like (i) bandwidth of filters (ii) bandwidth of communication
channels (iii) bandwidth of control systems. In feedback control, where the plant is often a low pass filter, the
higher the bandwidth the better because, as we will see in Chapter 4, the higher the bandwidth of a control
system, the faster its response.

Definition 3.8.3. Let G(s) be a rational, proper transfer function with all its poles in C− . Let Gmax :=
supω∈[0,∞) |G(jω)| denote the maximum magnitude of G(jω). The bandwidth of G is the width of the frequency
range in [0, ∞) in which, for every ω in this range
1
|G(jω) ≥ √ Gmax .
2

Version 1.2.2816, September 1, 2020

 3.8. Graphical representations of the frequency response 82

In terms of decibels, for every ω in this frequency range, 20 log |G(jω)| − 20 log Gmax ≥ −3dB.

Example 3.8.5. The Bode plot in Figure 3.11 of Example 3.8.3 shows that the RLC circuit has a bandwidth
of 9.9 radians/second. Input signals with frequencies outside of the interval [0.1, 10] are heavily attenuated. N

Example 3.8.6. The Bode plot in Figure 3.12 of Example 3.8.4 shows that the RC circuit has a bandwidth of
104 radians/second. Input signals with frequencies outside of the interval [0, 104 ] are heavily attenuated. N
When, as in Example 3.8.6, the maximum magnitude of G(jω) occurs at ω = 0, we will often say that a
system’s bandwidth is simply ωBW with the understanding that this refers to the width of the interval [0, ωBW ].
For such systems, the bandwidth ωBW indicates that the system will essentially pass all signals in the band
[0, ωBW ] without much attenuation. The bandwidth is easily obtained from the amplitude response plot of G(s)
as illustrated in Figure 3.13.
Magnitude (dB)

Frequency (rad/s)

Figure 3.13: Obtaining the bandwidth of a system from its amplitude response.

√
Remark 3.8.4. The significance of the ratio 1/ 2 (-3dB) in Definition 3.8.3 has its origins in the notion of
signal power. The (instantaneous) power in a signal u(t) is typically proportional to the amplitude of u2 (t).
If we apply an input √signal of frequency ωBW to a system with G(0) = 1, the output signal will have power
proportional to (u(t)/ 2)2 = u2 /2. In other words, the output will have half the power of the input signal. 

3.8.3 Drawing Bode plots by hand

We consider only rational G(s) with real coefficients, i.e., G(s) ∈ R(s). Then G(s) has a numerator and
denominator, each of which can be factored into terms of the following forms (cf. Section 3.8.1):
1. Pure gain: K
2. Pole or zero at s = 0: sn .
3. Real nonzero pole of zero: τ s ± 1, τ > 0.
s2
4. Complex conjugate pole or zero: 2
ωn
+ 2 ωζn s + 1, ζ ∈ [0, 1), ωn 6= 0.

Example 3.8.7.
1
s2


40s2 (s − 2) (40)(2) 2s − 1
G(s) = 2
= 1

1 2 1

.
(s + 5)(s + 4s + 100) (5)(100) 5s +1 100 s + 25 s +1
N

Version 1.2.2816, September 1, 2020

83 CHAPTER 3. LINEAR SYSTEM THEORY

Once G(s) has been factored as above, we find the frequency response by setting s = jω. Using the basic
properties of logarithms and complex numbers it is easy to show that 20 log |G(jω)| and ](G(jω)) consist of
sums and differences of the four basic terms. This is best illustrated by example.
Example 3.8.8. Continuing from Example 3.8.7 consider
80 1 1 1 2 1
20 log |G(s)| = 20 log + 20 log s2 + 20 log s − 1 − 20 log s + 1 − 20 log s + s+1
500 2 5 100 25
and        
80 2

1 1 1 2 1
](G(s)) = ] +] s +] s−1 −] s+1 −] s + s+1 .
500 2 5 100 25
N
Example 3.8.8 suggests that if we are able to plot the Bode plot for the four basic factors then we can add
and subtract these plots to obtain the Bode plot of G(s). This is the main idea of this section.

Pure gain
Consider G(s) = K. Then G(jω) = K and the frequency response is constant. The Bode plots are shown in
Figure 3.14. It is instructive to compare the Bode plots to the polar plots in Figure 3.5.
Bode Diagram Bode Diagram
Magnitude (dB)

Magnitude (dB)
Phase (deg)

Phase (deg)

10 2 10 3 10 4 10 5 10 6 10 2 10 3 10 4 10 5 10 6
Frequency (rad/s) Frequency (rad/s)

(a) K > 0. (b) K < 0.

Figure 3.14: Bode plots of G(s) = K.

Zero at the origin

Consider G(s) = sn . Then G(jω) = (jω)n and

20 log |G(jω)| = 20 log |(jω)n | = 20 log |ω|n = n20 log |ω|

and 


0◦ n ∈ {0, 4, 8, . . .}
90◦

n ∈ {1, 5, 9, . . .}
](G(jω)) = ](jω)n =


180◦ n ∈ {2, 6, 10, . . .}
−90◦

n ∈ {3, 7, 11, . . .} .
It is useful to compare these expressions, especially for the phase response, to the polar plots in Figure 3.6.
Notice that when ω = 1 = 100 then 20 log |G(jω)| = 0; when ω = 10 = 101 then 20 log |G(jω)| = n20; when
ω = 100 = 102 then 20 log |G(jω)| = n40; etc. In other words every time the frequency increases by a factor of
10 (called a decade increase), the magnitude response increases by n20 dB. Thus on a 20 log |G(jω)| vs log (ω)
plot the magnitude response is a straight line with slope n20 dB which passes through 0dB at ω = 100 . The
Bode plots for n ∈ {1, 2, 3, 4} are shown in Figure 3.15.

Version 1.2.2816, September 1, 2020

 3.8. Graphical representations of the frequency response 84

160

140

120
Magnitude (dB)

100

80

60

40

20

0
0 1 2
10 10 10

200

150

100
Phase (deg)

50

-50

-100
10 0 10 1 10 2
Frequency (rad/sec)

Figure 3.15: Bode plot of G(s) = sn for n ∈ {1, 2, 3, 4}.

Real zero
Consider G(s) = τ s + 1 with τ > 0. Then G(jω) = jτ ω + 1 and

20 log |G(jω)| = 20 log |jτ ω + 1|, ](G(jω)) = ](jτ ω + 1) = arctan2 (1, τ ω).

We now make some key observations used to draw Bode plots by hand. The polar plot in Figure 3.7b is helpful
for understanding these observations.
1
1. When ω  τ then:

(a) 20 log |jτ ω + 1| ≈ 20 log |1| = 0.

(b) ](jτ ω + 1) = arctan2 (1, τ ω) ≈ arctan2 (1, 0) = 0◦ .
1
2. When ω  τ then:

(a) 20 log |jτ ω + 1| ≈ 20 log |ω|.

(b) ](jτ ω + 1) = arctan2 (1, τ ω) ≈ arctan2 (0, ω) = 90◦ .
1
3. When ω = τ then:
√
(a) 20 log |jτ ω + 1| = 20 log | 2| = 3dB.
(b) ](jτ ω + 1) = arctan2 (1, 1) = 45◦ .

The frequency τ1 is called the break frequency of τ s + 1. Based on the above observations, we make the
following simplifications for drawing the Bode plot of τ s + 1 by hand:
1. For ω < τ1 , |jτ ω + 1| ≈ 1.

Version 1.2.2816, September 1, 2020

85 CHAPTER 3. LINEAR SYSTEM THEORY

2. For ω > τ1 , |jτ ω + 1| ≈ |ω|.

3. For ω < 0.1

τ , ]jτ ω + 1 ≈ ]1 = 0◦ .

4. For ω > 10
τ , ](jτ ω + 1) ≈ ]jω = 90◦ .

The lines obtained using the above approximations are called asymptotes and the resulting approximation of
the Bode plot is called an asymptotic Bode plot. Figure 3.16 shows the asymptotic Bode plot when τ = 1
as well as the exact Bode plot. The maximum error between the magnitude plots occurs at ω = τ1 .
Bode Diagram
40

35

30

25
Magnitude (dB)

20

15

10

0
90

60
Phase (deg)

30

0
10 -2 10 -1 10 0 10 1 10 2
Frequency (rad/s)

Figure 3.16: Exact and asymptotic Bode plots of τ s + 1 with τ = 1.

Now consider G(s) = τ s − 1 with τ > 0. Then G(jω) = jτ ω − 1. The magnitude plot is exactly as before.
The only difference is in the phase plot. Once again it is instructive to observe the polar plot in Figure 3.7a. In
this case we have that when ω  τ1 , ]G(jω) ≈ 180◦ and when ω  τ1 , ]G(jω) ≈ 90◦ . The asymptotic Bode
plot is drawn in the same manner. The plots are shown in Figure 3.17.

Complex conjugate zeros

Lastly we consider the case when G(s) has complex conjugate zeros

s2 ζ
G(s) = + 2 s + 1, ζ ∈ [0, 1), ωn 6= 0.
ωn2 ωn

For asymptotic Bode plots we make the approximation that ζ = 1, i.e.,

2
s2
  
1 s
G(s) ≈ 2
+2 s+1 = +1
ωn ωn ωn

In this approximation we treat the complex conjugate zeros as two real, repeated zeros at s = −ωn . The
advantage of doing this is that we can generate asymptotic Bode plots in the same fashion as the previous
section using a break frequency of |ωn |. The disadvantage is that the approximation is quite crude and can lead
to large errors between the asymptotic and exact Bode plots. The error is entirely dependent on the actual value

Version 1.2.2816, September 1, 2020

 3.8. Graphical representations of the frequency response 86

Bode Diagram
40

35

30

25
Magnitude (dB)

20

15

10

0
180

150
Phase (deg)

120

90
10 -2 10 -1 10 0 10 1 10 2
Frequency (rad/s)

Figure 3.17: Exact and asymptotic Bode plots of τ s − 1 with τ = 1.

of ζ. The closer ζ is to 1, the more accurate the plot. To make the asymptotic Bode plot more accurate you
can correct it by computing the true value of |G(jω)| and ]G(jω) near ω = ωn . Most introductory textbooks
have tables of correction factors for different values of ζ.
Figure 3.18 shows a plot for complex conjugate zeros in C− , i.e., ωn > 0, ζ ∈ (0, 1). Figure 3.19 shows the
plot for complex conjugate zeros in C+ , i.e., ωn < 0, ζ ∈ (0, 1). Once again it is instructive to compare these
Bode plots to the polar plots in Figure 3.7.
Bode Diagram
80

70

60

50
Magnitude (dB)

40

30

20

10

-10
180

135
Phase (deg)

90

45

0
10 -2 10 -1 10 0 10 1 10 2
Frequency (rad/s)

s2
Figure 3.18: Exact and asymptotic Bode plots of 2
ωn + 2 ωζn s + 1 with ωn = 1.

Version 1.2.2816, September 1, 2020

87 CHAPTER 3. LINEAR SYSTEM THEORY

Bode Diagram
80

70

60

50
Magnitude (dB)

40

30

20

10

-10
0

-45
Phase (deg)

-90

-135

-180
-2 -1 0 1 2
10 10 10 10 10
Frequency (rad/s)

s2
Figure 3.19: Exact and asymptotic Bode plots of 2
ωn + 2 ωζn s + 1 with ωn = −1.

A special case is when ζ = 0. In this case G(s) has imaginary roots at s = ±jωn . When ζ = 0 and ω = ωn
we have G(jω) = 0. Hence the magnitude bode plot approaches negative infinity as ω → ωn . This is illustrated
in Figure 3.20. Compare with the polar plot in Figure 3.8. In such a case the asymptotic Bode plot has infinite
error at the break frequency ω = ωn .
Bode Diagram
40

20

-20
Magnitude (dB)

-40

-60

-80

-100

-120

-140

-160
0

-45
Phase (deg)

-90

-135

-180
10 -1 10 0 10 1
Frequency (rad/s)

s2
Figure 3.20: Exact and asymptotic Bode plots of 2
ωn + 1 with ωn2 = 1.

Version 1.2.2816, September 1, 2020

 3.8. Graphical representations of the frequency response 88

Exercise 3.6. Using the principles developed so far, sketch the asymptotic Bode plot of

40s2 (s − 2)
G(s) = .
(s + 5)(s2 + 4s + 100)

Examples

Example 3.8.9. (Minimum Phase) A minimum phase TF

1 1
G1 (s) = = 0.1 .
s + 10 0.1s + 1
The frequency response is
1
G(jω) = 0.1 .
j0.1ω + 1
It follows that
20 log |G(jω)| = 20 log |0.1| − 20 log |j0.1ω + 1| = −20 − 20 log |j0.1ω + 1|
](G(jω)) = ]0.1 − ](j0.1ω + 1) = −](j0.1ω + 1).
Figure 3.21 shows the polar plot and asymptotic Bode plot for this frequency response. Make sure you under-
stand the asymptotic Bode plot in Figure 3.21b and that you can plot it by hand. N
-20

0
-25

-30
-0.005

-35
Magnitude (dB)

-0.01
-40

-45
-0.015

-50

-0.02
-55

-60
-0.025
0

-0.03

-30
-0.035
Phase (deg)

-0.04
-60

-0.045

-90
-0.05 -1 0 1 2 3
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1 10 10 10 10 10

(a) Polar plot. (b) Asymptotic Bode plot.

Figure 3.21: Polar and asymptotic Bode plots for Example 3.8.9.

Example 3.8.10. (All-Pass) An all-pass TF

1−s
G2 (s) = .
1+s
The frequency response is
jω − 1
G2 (jω) = −1 .
jω + 1
It follows that
20 log |G2 (jω)| = 20 log |−1| + 20 log |jω − 1| − 20 log |jω + 1| = 0dB.
](G2 (jω)) = ]−1 + ](jω − 1) − ](jω + 1) = 180◦ + ](jω − 1) − ](jω + 1).

Figure 3.22 shows the polar plot and asymptotic Bode plot for this frequency response. As before, make sure
you understand the Bode plot in Figure 3.22b and know how to plot it by hand. N

Version 1.2.2816, September 1, 2020

89 CHAPTER 3. LINEAR SYSTEM THEORY

-0.1

-0.2
1

0.8

-0.3 0.6

0.4

Magnitude (dB)
0.2
-0.4
0

-0.2

-0.5 -0.4

-0.6

-0.6 -0.8

-1
0

-0.7
-45

-0.8

Phase (deg)
-90

-0.9
-135

-1 -180
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1 10 -2 10 -1 10 0 10 1 10 2

(a) Polar plot. (b) Asymptotic Bode plot.

Figure 3.22: Polar and asymptotic Bode plots for Example 3.8.10.

Example 3.8.11. (Non-Minimum Phase) A non-minimum phase TF

1−s 1
G3 (s) = G2 (s)G1 (s) = .
1 + s s + 10
Using basic properties of complex numbers we have that

|G3 (jω)| = |G2 (jω)||G1 (jω)| = |G1 (jω)|

so that the magnitude response of G3 is that same as that of G1 from Figure 3.21b. Furthermore the frequency
response is
]G3 (jω) = ]G2 (jω) + ]G1 (jω).
Figure 3.23 shows the polar plot and asymptotic Bode plot for this frequency response. As before, make sure
you understand the Bode plot in Figure 3.23b and know how to plot it by hand. From the polar plot we again
-20

0.05
-25

-30

-35
Magnitude (dB)

-40

-45

0
-50

-55

-60
180

135

90
-0.05
45
Phase (deg)

-45

-90

-135

-180
-0.1
-0.1 -0.08 -0.06 -0.04 -0.02 0 0.02 0.04 0.06 0.08 0.1 10 -2 10 -1 10 0 10 1 10 2 10 3

(a) Polar plot. (b) Asymptotic Bode plot.

Figure 3.23: Polar and asymptotic Bode plots for Example 3.8.11.

see (cf. Example 3.8.2) that the phase change for G3 is greater than the phase change for G1 . Of all TFs
having the same magnitude response |G1 (jω)|, G1 has the minimum phase change. Hence the terminology of
Definition 3.8.2. N
Remark 3.8.5. This section has dealt with graphical methods done by hand on paper. Why did we bother?
Why didn’t we just ask MATLAB to draw the Bode plots? The answer is that by learning to draw the graphs,

Version 1.2.2816, September 1, 2020

 3.9. Summary 90

you acquire understanding. In addition, there will come times when MATLAB will fail you. For example,
poles on the imaginary axis: MATLAB isn’t smart enough to properly draw Nyquist plots which we cover in
Chapter 8. 

3.9 Summary
This chapter covered some of the basic results in linear system theory. Parts of this chapter should have been
review from your signals and systems course. After reading this chapter you should know the following.

1. You should know how to compute the matrix exponential eAt using Laplace transforms for small problems
(n ≤ 3).
2. Understand the structure of the total response of an LTI state-space model.
3. Know what it means for an LTI system to be asymptotically stable and how to test for asymptotic stability
using eigenvalues.
4. Know the definition of BIBO stability and how to test for it by checking the poles of a TF.
5. Understand the relationship between asymptotic stability and BIBO stability of LTI systems.
6. Be able to apply the final-value theorem and compute the steady-state gain of a BIBO stable system.
7. Understand the physical interpretation of a system’s frequency response.
8. Be able to plot asymptotic Bode plots by hand.

3.A Appendix: Complex numbers

We define C, the complex numbers, to be a set of ordered pairs (x, y) where x and y are real and where
addition and multiplication are defined by:
(x1 , y1 ) + (x2 , y2 ) := (x1 + x2 , y1 + y2 )
(x1 , y1 )(x2 , y2 ) := (x1 x2 − y1 y2 , x1 y2 + x2 y1 ).

We adopt the standard way of writing z = (x, y) ∈ C as z = x + jy where j · j = j 2 = −1 and x, y ∈ R. This

is called the Cartesian form of z. With this convention, addition and multiplication of complex numbers can
be done using the ordinary rules of algebra
(x1 + jy1 ) + (x2 + jy2 ) = (x1 + x2 ) + j(y1 + y2 )
(x1 + jy1 )(x2 + jy2 ) = (x1 x2 − y1 y2 ) + j(x1 y2 + x2 y1 ).
For the complex number z = x + jy we write Re (z) := x and Im (z) := y for the real part and imaginary part,
respectively, of z. Note that both the real part and the imaginary part of a complex numbers are real numbers.
The complex number −2 + j3 has real part −2 and imaginary part 3. Complex numbers of the form z = x + j0
are called purely real while complex numbers of the form z = 0 + jy are called purely imaginary. The
complex conjugate of z = x + jy is defined to be the complex number z = x − jy. It is easy to prove that
1 1
Re (z) = (z + z) , (z − z) .
Im (z) =
2 2j
p
By |z| we denote the magnitude or modulus of z and it equals x2 + y 2 . You can verify that
1 z
|z| = |z|, |z|2 = zz, = 2.
z |z|

Version 1.2.2816, September 1, 2020

91 CHAPTER 3. LINEAR SYSTEM THEORY

For z, w ∈ C, you can also verify the following basic properties of magnitudes
z |z|
|zw| = |z||w|, = .
w |w|
By ]z we denote the the angle between the positive real axis and the line segment from 0 to z in the complex
plane. This angle is called the argument of z and we write ]z = arg (z). Since we can always add integer
multiples of 2π, arg (z) is not single-valued and therefore “arg” is not a function. To avoid confusion, we always
use the principle value of the argument.

Definition 3.A.1. The principle value of the argument (or principle argument) of a complex number
z is the value of arg (z) greater than −π and less than or equal to π.

The principle argument of z = z + jy can be computed as arctan2 (x, y) where arctan2 : R2 \ {0} → (−π, π]
is the four quadrant arctan. Unless otherwise stated, we always use the principle value of the argument in this
course so that −π < ]z ≤ π.
Example 3.A.1. The argument of −1 − j is any number −3π/4 + 2kπ where k is an arbitrary integer. The
principle argument of −1 − j is −3π/4. N
Euler’s equation states that for all x ∈ R, ejx = cos (x) + j sin (x). Notice that |ejx | = 1 and that we can
always represent x ∈ R as x = θ + 2πk where θ ∈ (−π, π] and k is an integer. Using this representation, the
principle argument of ejx equals θ and we can write any non-zero complex number z = x + jy in polar form
z = |z|ej]z .
It is common in engineering to drop the dependence on e and simply write |z|]z. This is called phasor
notation. It is easy to show that   
r1 ejθ1 r2 ejθ2 = r1 r2 ej(θ1 +θ2 ) .
It follows that for any z, w ∈ C
z
]zw = ]z + ]w, ] = ]z − ]w.
w

3.B Appendix: Time domain, s-domain and frequency domain

In this course we work in three domains: (1) the time-domain, (2) the s-plane and (3) the frequency domain. In
each domain there is one mathematical object that captures the essence of system behaviour: (1) the impulse
response g(t) (2) the transfer function G(s), and (3) the frequency response G(jω). Anything that we say about
one of these three objects should be reflected in the others. We will move with ease between these three domains
for the remainder of the course and you should understand the relationship between them. This relationship is
illustrated, in the case of a BIBO stable system, in Figure 3.24.
We now review the justification for the various arrows in Figure 3.24. In Section 2.7 we discussed the
conditions under which the LT L of g(t) exists. Since g is assumed to be causal we can uniquely recover g(t)
from G(s) without being told the ROC for G(s). This justifies the arrows between g(t) and G(s).
In Section 2.7.1 we discussed the conditions under which the FT F of g(t) exists. Since the system in
Figure 3.24 is assumed BIBO stable, it follows from Condition 2 of Theorem 3.5.4 that F (g) exists (see also
Section 3.7). This justifies the arrows between g(t) and G(jω).
Since G(s) is BIBO stable, Condition 3 of Theorem 3.5.4 tells us that the ROC of G(s) includes the imaginary
axis and the arrow from G(s) to G(jω) is justified (see Sections 2.7.2 and 3.7). We are left to justify the arrow
from G(jω) to G(s). Note that it is not obvious that one should be able to recover G(s) from G(jω). After all,
the frequency response function only gives us data on the imaginary axis.

Version 1.2.2816, September 1, 2020

 3.B. Appendix: Time domain, s-domain and frequency domain 92

time domain L / s-domain

(impulse response g(t)) o (transfer function G(s))
L −1
g 7
F s=jω

F −1 (3.14)
' w

frequency domain
(frequency response G(jω))

Figure 3.24: Relationship between impulse response, s-domain and frequency response of a BIBO stable LTI system.

Proposition 3.B.1. Let G(s) ∈ R(s) be a strictly proper BIBO stable system with impulse response g(t) and
frequency response G(jω). Then, for all s ∈ C+
Z∞
1 G(jω)
G(s) = dω. (3.14)
2π s − jω
−∞

Proof. As discussed above and in Section 3.7, under the assumption of BIBO stability, the frequency response
of a system equals the Fourier transform of g(t) and therefore
Z∞
1
g(t) = F −1 {G(jω)} = G(jω)ejωt dω.
2π
−∞

On the other hand, G(s) is the LT of g(t) and so for any s in its ROC, which by BIBO stability includes C+ ,
we have
Z∞
G(s) = g(t)e−st dt
0
Z∞ Z∞
1
= G(jω)ejωt e−st dω dt
2π
0 −∞
Z∞ Z∞
1
= G(jω)ejωt e−st dt dω
2π
−∞ 0
Z∞
1 G(jω)
= dω.
2π s − jω
−∞

Remark 3.B.2. State models fit into the above discussion via the impulse response (3.4) given by g(t) =
CeAt B1(t) + Dδ(t) or via the transfer function G(s) = C(sI − A)−1 B + D. 

Remark 3.B.3. This discussion can be extended to unstable systems and systems with poles on the imaginary
axis. The treatment in [Dullerud and Paganini, 2000] is somewhat advanced but rather complete. 

Version 1.2.2816, September 1, 2020

Chapter 4

First and second order systems

In the previous chapter we studied properties of linear time-invariant systems in a somewhat general manner.
We showed that a single-input, single-output LTI system can be modelled by state equations

ẋ = Ax + Bu, y = Cx + Du, u, y ∈ R

or by input-output models
y =g∗u or Y (s) = G(s)U (s).
In this chapter we study the important special cases when G(s) ∈ R(s) is given by

K
G(s) = (prototype first order system),
τs + 1
Kωn2
G(s) = 2 (prototype second order system).
s + 2ζωn s + ωn2

First and second order systems are important because they are easy to analyze which makes them useful for
developing intuition that can be applied to more complicated systems, e.g., higher order systems, systems with
zeros. Also, the behaviour of these more complicated systems can sometimes be well approximated by first and
second order models. Our main objective in this chapter is to relate the position of the poles of G(s) in the
s-plane to the corresponding time-domain responses. This relationship is a key to the analysis and design of
LTI feedback controllers.

Contents
4.1 Prototype first order system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
4.2 Prototype second order system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
4.3 General characteristics of a step response . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
4.4 The effect of adding poles and zeros . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
4.5 Dominant poles and zeros . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
4.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113

4.1 Prototype first order system

Definition 4.1.1. The prototype first order system with input u and output y is governed by the differential
equation
τ ẏ + y = Ku, K, τ ∈ R. (4.1)

93
 4.1. Prototype first order system 94

The associated transfer function is

Y (s) K
= . (4.2)
U (s) τs + 1

System (4.1) in Definition 4.1.1 can of course be given a state model. Define x := y to obtain (verify)
ẋ = ax + bu, y = cx + du, x, y, u ∈ R
and
1 K
a = − , b = , c = 1, d = 0.
τ τ
For the remainder of this section we study how the values of the constants K and τ affect the time-domain
response of the prototype first order system.
Example 4.1.1. (Heating a room) Consider the problem of heating up a room as shown in Figure 4.1.
Suppose that:

Figure 4.1: Heating a room.

• u is the heat provided by the heater; the input.

• T is the inside temperature. This is the variable we want to control; the output.
• Te is the exterior (outside) temperature. We assume that Te is constant.
• Q is the heat flow out of the room.
• Qa is the heat accumulating within the system.
By conservation of energy we have that
Qa = u − Q.
If Qa = β dT
dt and Q = α(T − Te ), where β and α are real constants, then the system model is

dT u α
= − (T − Te ). (4.3)
dt β β
This model is nonlinear (Do you see why? Hint: see Example 2.4.4.) with input u and output T . The governing
equation (4.3) is first order so we only need one state variable. The natural choice is x := T so that the resulting
nonlinear state model is
α 1 αTe
ẋ = f (x, u) := − x + u +
β β β
y = h(x, u) := x.

Version 1.2.2816, September 1, 2020

95 CHAPTER 4. FIRST AND SECOND ORDER SYSTEMS

Suppose that we want to control the temperature of the room to maintain an inside temperature of T > Te ,
i.e., the room should be T − Te degrees warmer than the outside.
Exercise 4.1. Show that (x, u) = (T , α(T − Te )) is the unique equilibrium configuration corresponding to the desired temperature
x = T (see Definition 2.5.1).
∂f ∂f ∂h ∂h
Exercise 4.2. Compute Jacobians a = ∂x (x,u)
, b= ∂u (x,u)
, c= ∂x (x,u)
, d= ∂u (x,u)
and verify that a = −α/β, b = 1/β,
c = 1, d = 0.
Finally, let δx := x − x = T − T and δu := u − u = u − α(T − Te ) denote deviations from the equilibrium
configuration to obtain the linearized model

˙ = − α δx + 1 δu
δx
β β
δy = δx.

Let Y (s) := L {δy} and U (s) := L {δu}. The resulting transfer function of the linearized model is
1 1
Y (s) β α
= α = β
.
U (s) s+ β αs +1
β
This system has the form of a prototype first order system (4.2) with τ = α and K = α1 . N

By Theorem 3.5.4 the prototype first order systems is BIBO stable if τ > 0 and unstable if τ < 0.
Exercise 4.3. Is the prototype first order system (4.2) BIBO stable when τ = 0? What if K = 0?
Exercise 4.4. Is the state-space model of the prototype first order system asymptotically stable when u = 0 and τ < 0? Hint:
Use Proposition 3.4.2.
Table 4.1 lists key properties of the prototype first order system (4.2).

Table 4.1: Properties of prototype first order system (4.2).

Pole location Zero location Steady-state gain Bandwidth

s = − τ1 None K 1
τ rad/s.

Exercise 4.5. Verify that Table 4.1 is correct using Definitions 2.8.2, 3.6.3 and 3.8.3.
Using the ideas of either Section 3.3 or Section 2.7 (Examples 2.7.7 and 2.7.9) it is easy to show that the
impulse response of the prototype first order system is given by
K −t
g(t) = e τ , t ≥ 0. (4.4)
τ

Exercise 4.6. Verify that (4.4) is the impulse response of (4.2).

Suppose that we sample the impulse response (4.4) every τ seconds starting from t = 0 as shown in
Figure 4.2a. Doing so we obtain the sampled values
K
g(0) = , g(τ ) = e−1 g(0) ≈ 0.37g(0), g(2τ ) = e−2 g(0) ≈ 0.14g(0), g(3τ ) = e−3 g(0) ≈ 0.05g(0)
τ
which shows that the smaller τ > 0 is, the faster the impulse response decays to zero. Next we obtain the step
response by setting u(t) = 1(t) so that

K 1 K K
Y (s) = G(s)U (s) = = −
1 + τs s s s + τ1

Version 1.2.2816, September 1, 2020

 4.1. Prototype first order system 96
1
1

0.9
0.9

0.8
0.8

0.7
0.7

0.6
0.6

0.5
0.5

0.4
0.4

0.3
0.3

0.2
0.2

0.1
0.1

0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 0

(a) Impulse response. (b) Step response.

Figure 4.2: Responses of a prototype first order system with K = τ = 1 sampled every τ seconds.

and, assuming zero initial conditions,

 t

y(t) = L −1 {Y (s)} = K 1 − e− τ , t ≥ 0. (4.5)

Again, suppose that we sample the step response (4.5) every τ seconds starting from t = 0 as shown in
Figure 4.2b. We obtain the sampled values

y(0) = 0, y(τ ) = K 1 − e−1 ≈ 0.63yss , y(2τ ) = K 1 − e−2 ≈ 0.86yss ,

y(3τ ) = K 1 − e−3 ≈ 0.95yss , y(4τ ) = K 1 − e−4 ≈ 0.98yss .

After 4τ seconds the step response has reached 98% of its steady-state value. The smaller τ > 0 is, the faster the
step response converges to its steady-state value. The parameter τ determines the rate at which e−t/τ decays
to zero which motivates the subsequent definition.

Definition 4.1.2. The time constant of the prototype first order system is the parameter τ in (4.2).

In summary, the time constant completely determines the transient response of the first order system. As
τ > 0 approaches zero, the system pole moves towards −∞ along the real axis of the s-plane, the system’s
bandwidth approaches infinity and the system’s time-domain response gets faster.

Example 4.1.2. In Example 2.9.1 proportional error feedback was applied to a prototype first order system.
The resulting TF from the reference input to the plant output was found to be
Y (s) Kp K
= .
R(s) τ s + 1 + Kp K

The time constant of the closed-loop system is τ /(1 + Kp K) and the closed-loop steady-state gain is Kp K/(1 +
Kp K). The step response (4.5) is
 
Kp K t(1+Kp K)
y(t) = 1 − e− τ , t ≥ 0.
1 + Kp K

By choosing large positive values of Kp , i.e., high gain control, the time constant of the closed-loop system can
be made arbitrarily small and the steady-state gain arbitrarily close to one. However large values of Kp will
result in large control signals and may not be physically implementable. N

Version 1.2.2816, September 1, 2020

97 CHAPTER 4. FIRST AND SECOND ORDER SYSTEMS

4.2 Prototype second order system

Definition 4.2.1. The prototype second order system with input u and output y is governed by the
differential equation
ÿ + 2ζωn ẏ + ωn2 y = Kωn2 u, K, ζ, ωn ∈ R. (4.6)
The associated transfer function is
Y (s) Kωn2
= 2 . (4.7)
U (s) s + 2ζωn s + ωn2

To put (4.6) into state model form, let x := (y, ẏ) to obtain (verify!) ẋ = Ax + Bu, y = Cx + Du with
   
0 1 0  
A= 2 , B= 2 , C = 1 0 , D = 0.
−ωn −2ζωn Kωn
The prototype second order system is very important because it is easy to analyse but it also has far richer
dynamics than the first order system. This makes it a workhorse for obtaining simple models of many engineering
systems.
Example 4.2.1. (Mass-Spring-Damper) The mass spring damper in Figure 4.3 from Examples 2.1.1
and 3.3.1 has transfer function

Figure 4.3: Mass-spring-damper system.

Y (s) 1
= 2
.
U (s) M s + bs + k
Comparing this transfer function to the prototype second order system (4.7) we deduce
r
k 1 b
ωn = , K= , ζ= √ .
M k 2 kM
N

Example 4.2.2. (RLC Circuit) Consider the series RLC circuit in Figure 4.4 (cf. Examples 2.3.5 and 2.4.5).
The system TF is (verify!)
1
Y (s) 1 LC
= G(s) = 2 = 2 R 1
.
U (s) s LC + sRC + 1 s + L s + LC
Comparing this transfer function to the prototype second order system (4.7) we deduce
r
1 R C
ωn = √ , K = 1, ζ= .
LC 2 L
N

Version 1.2.2816, September 1, 2020

 4.2. Prototype second order system 98

Figure 4.4: RLC circuit with voltage across capacitor taken as output.

The prototype second order system’s TF is real rational and it has no zeros so its poles are the roots of its
denominator, i.e., the roots of s2 + 2ζωn s + ωn2 . These are easily computed using the quadratic formula
p  p 
s = −ζωn ± (ζωn )2 − ωn2 = ωn −ζ ± ζ 2 − 1 . (4.8)

We observe from (4.8) that

• When ζ > 1 the system has two real, distinct, poles.

• When ζ = 1 the system has two real, repeated, poles.

• When 0 < ζ < 1 the system has two complex conjugate poles.

• When ζ ≤ 0 the system is unstable.

Figure 4.5 shows how the poles of the prototype second order system (4.7) vary on the s-plane as a function
of ζ for fixed ωn > 0.

(a) ζ ranging from 0 to +∞. (b) ζ ranging from −∞ to 0.

Figure 4.5: Pole locations of the prototype second order system as a function of ζ with fixed ωn .

Just as the time constant characterizes the prototype first order system, the parameter ζ is a useful way to
categorize the prototype second order system and has its own name.

Version 1.2.2816, September 1, 2020

99 CHAPTER 4. FIRST AND SECOND ORDER SYSTEMS

Definition 4.2.2. The damping ratio of the prototype second order system is the parameter ζ in (4.7).
The natural undamped frequency of the prototype second order system is the parameter ωn in (4.7). The
prototype second order system is called:

(i) undamped if ζ = 0,

(ii) underdamped if 0 < ζ < 1,

(iii) critically damped if ζ = 1,

(iv) overdamped if ζ > 1.

The step responses for an underdamped, critically damped and over damped prototype second order system
are shown in Figure 4.6.

Figure 4.6: Representative step responses of the prototype second order system for different values of ζ and fixed ωn .

Exercise 4.7. Simulate and plot the step response of an undamped prototype second order system.
Table 4.2 lists key properties of the prototype second order system (4.7). In the rest of this section we study

Table 4.2: Properties of prototype second order system (4.7).

Name Poles Zeros Steady-state gain Step response

p
Underdamped −ζωn ± jωn 1 − ζ2 None K Oscillatory
Critically damped −ωn (repeated) None K Not oscillatory
p
Overdamped −ζωn ± ωn ζ 2 − 1 None K Not oscillatory

in some detail the properties of underdamped, critically damped and overdamped second order systems.

Version 1.2.2816, September 1, 2020

 4.2. Prototype second order system 100

4.2.1 Underdamped systems

Example 4.2.3. (Mass-Spring-Damper) Using the expression for the damping ratio from Example 4.2.1
we conclude
√ that the mass-spring-damper system is underdamped if and only if the damper constant b satisfies
0 < b < 2 kM . N

Exercise 4.8. Suppose that L = 160 µH and C = 10 µF in the RLC circuit from Example 4.2.2. Find the range of resistance
values so that the circuit is underdamped. p
The poles of an underdamped system, written in Cartesian form, are s = −ζωn ± jωn 1 − ζ 2.
Exercise 4.9. Verify that
p  p 
−ζωn ± jωn 1 − ζ 2 = ωn , ] −ζωn ± jωn 1 − ζ 2 = ± (π − arccos (ζ)) .

In light of Exercise 4.9, the polar form of the poles of an underdamped prototype second order system is

s = ωn e±j(π−arccos (ζ)) .

The angle of the poles is completely determined by the damping ratio ζ while the magnitude is determined by
the undamped natural frequency ωn , see Figure 4.7. We now analyse the frequency response of an underdamped

Figure 4.7: Pole location of an underdamped second order system. In this figure θ = arccos (ζ).

prototype second order system. If we assume that the steady-state gain is unity, i.e., K = 1, then we can write
the TF (4.7) as
1
G(s) =  2   .
s s
ωn + 2ζ ωn + 1

Recall from Section 3.8.3 that this is exactly the standard form for drawing the Bode plot of a system with
complex conjugate roots. The Bode plot for an underdamped system with K = ωn = 1 is shown in Figure 4.8
for various different values of ζ. We observe from Figure 4.8 that the system bandwidth is approximately ωn and

Version 1.2.2816, September 1, 2020

101 CHAPTER 4. FIRST AND SECOND ORDER SYSTEMS

40

20

0
Magnitude (dB)

−20

−40

−60

−80
0

−45
Phase (deg)

−90

−135

−180
−2 −1 0 1 2
10 10 10 10 10
Frequency (rad/s)

 √
Figure 4.8: Bode plot of the prototype second order system with K = ωn = 1 for ζ ∈ 0.05, 0.1, 0.3, 0.5, 1/ 2, 1 .

√
that, when 0 < ζ < 1/ 2, the magnitude plot has a “peak.” It can be shown that the value of the magnitude
plot at that peak is
1
20 log p
2ζ 1 − ζ 2
p
and that it occurs at ω = ωn 1 − ζ 2 .
Having identified the bandwidth of the underdamped system (ωBW ≈ ωn ) we turn to the impulse and
step responses to see how the damping ratio and undamped natural frequency affect them. We start with
the impulse response. The impulse response of the underdamped second order system is found, using partial
fraction expansions and doing some messy calculations, to be
ωn −ζωn t
 p
2

g(t) = K p e sin ωn 1 − ζ t , t ≥ 0. (4.9)
1 − ζ2

Observe that the decay rate of −ζωn t while the

p the impulse response is determined by the exponential part e
frequency of oscillation is ωn 1 − ζ 2 . Therefore the decay rate depends on the real part of the poles (see
Figure 4.7) while the oscillation frequency depends on the imaginary part of the poles (see Figure 4.7). Figure 4.9
shows the impulse response of various underdamped systems. In these figures, as the parameter ζ gets closer
and closer to zero (and the poles become purely imaginary), the response becomes more oscillatory.
To find the step response we can either integrate the impulse response or use partial fraction expansions on
Kωn2 1
Y (s) =
s2 + 2ζωn s + ωn2 s
to get !
1  p 
y(t) = K 1− p e−ζωn t sin ωn 1 − ζ 2 t + arccos (ζ) , t ≥ 0. (4.10)
1 − ζ2
Figure 4.10 shows the step response of various underdamped systems.
Notice that in the equations for both the impulse response (4.9) and the step response (4.10) that whenever
time t appears, it is multiplied by ωn . Therefore, for fixed ζ, the parameter ωn acts as a time-scale factor and

Version 1.2.2816, September 1, 2020

 4.2. Prototype second order system 102

1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0

−0.2 −0.2

−0.4 −0.4

−0.6 −0.6

−0.8 −0.8

−1 −1
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10

(a) K = ωn = 1. (b) K = 1, ωn = 5.
 √
Figure 4.9: Impulse response of an underdamped system with ζ ∈ 0.05, 0.1, 0.3.0.5, 1/ 2, 1 .

1.8

1.6

1.4
1.8

1.2
1.6

1.4
1

1.2

0.8
1

0.6
0.8

0.6 0.4

0.4

0.2
0.2

0
0
0 1 2 3 4 5 6 7 8 9 10
0 1 2 3 4 5 6 7 8 9 10

(a) K = ωn = 1. (b) K = 1, ωn = 5.
 √
Figure 4.10: Step response of an underdamped system with ζ ∈ 0.05, 0.1, 0.3.0.5, 1/ 2, 1 .

as ωn gets larger the response speeds up. Similarly, as ωn gets smaller, the response slows down. Therefore, we
once again see that the higher the system bandwidth, the faster the response.
These plots also show that as ζ gets closer to 1, equivalently as the complex conjugate poles get closer to
the real axis, the response become less oscillatory. The undamped natural frequency ωn is the frequency of
oscillation when ζ = 0 which provides some justification for its name.
Just as in the first order case, we see that the further to the left in C− that the poles are, the faster the
response. Also, the closer the poles are to the real axis (for fixed ωn ), the less oscillatory the response. We will
make much use of the above observations in analyzing and designing control systems.

4.2.2 Overdamped systems

The prototype second order system (4.7) with ζ > 1 has two distinct real poles located at
p
s = −ζωn ± ωn ζ 2 − 1. (4.11)

Version 1.2.2816, September 1, 2020

103 CHAPTER 4. FIRST AND SECOND ORDER SYSTEMS

By factoring the denominator of the TF (4.7) we can re-write the transfer function in the form

Kab
G(s) =
(s + a)(s + b)

where a and b are the roots (4.11).

Exercise 4.10. Show that the bandwidth of the overdamped system approximately equals ωBW = min {|a|, |b|}.
The impulse response of an overdamped system can be found using partial fraction expansions

g(t) = L −1 {G(s)}
( )
ab ab
−1 b−a a−b
= KL + (4.12)
s+a
s+b
ab  −at 
=K e − e−bt , t ≥ 0.
b−a
The time derivative of the impulse response is
d ab  −bt 
g(t) = K be − ae−at .
dt b−a
Setting this derivative equal to zero and solving for t we get

ln (a) − ln (b)
t̄ = .
a−b
This means that g(t̄) is an extreme value of the impulse response: the absolute value of the the impulse peaks
at that time. The step response can also be found using partial fraction expansions. If
 
Kab 1 K K a b
Y (s) = = + − .
(s + a)(s + b) s s b−a s+b s+a

then  
1  −bt
y(t) = K 1 − ae − be−at , t ≥ 0. (4.13)
b−a
The time derivative of the step response is
d ab  −at 
y(t) = K e − e−bt .
dt b−a
This shows that ẏ = 0 only when t = 0. In other words, there are no extreme values of y(t) for t ≥ 0 and the
step response does not have “peaks.” Figure 4.11 shows typical impulse and step responses of an overdamped
system.
Exercise 4.11. Find the pole locations of the system used to create the plots in Figure 4.11.

4.2.3 Critically damped systems

The prototype second order system (4.7) with ζ = 1 has two repeated real poles located at

s = −ζωn = −ωn . (4.14)

Therefore, in the critically damped case, we can re-write the system’s TF as

Kωn2
G(s) = .
(s + ωn )2

Version 1.2.2816, September 1, 2020

 4.3. General characteristics of a step response 104

0
0
Time (seconds) Time (seconds)

(a) Impulse response. (b) Step response.

Figure 4.11: Time response of an overdamped system with ζ = 2, ωn = 1.

Exercise 4.12. Show that the bandwidth of a critically damped system is approximately ωBW = |ωn |.
The unit step response can be found using the inverse LT and partial fraction expansions

y(t) = K 1 − (1 + ωn t)e−ωn t , t ≥ 0.


(4.15)

Using similar analysis as in the overdamped case, Section 4.2.2, you can show that the critically damped system
has no “peaks” in its step response. Its response is slower than a underdamped system’s response but it does
not “overshoot” its steady-state value during its transient period. Figure 4.12 shows the typical impulse and
step responses of an overdamped systems.

0
0
Time (seconds) Time (seconds)

(a) Impulse response. (b) Step response.

Figure 4.12: Time response of a critically damped system with ζ = 1, ωn = 1.

4.3 General characteristics of a step response

In this section we look at common measures used to quantify the quality of a step response. These measures can
be applied to any stable system, even if it isn’t first or second order. The key characteristics of a step response
are illustrated in Figure 4.13. While the quantities in Figure 4.13 can be numerically or experimentally found
for a system of any order, we will see that the prototype first and second order systems are useful because
we can obtain closed-form expressions for the various characteristics. These closed-form expressions motivate
many of our design equations.

Version 1.2.2816, September 1, 2020

105 CHAPTER 4. FIRST AND SECOND ORDER SYSTEMS

1.4

1.2

0.8

0.6

0.4

0.2

0
0 1 2 3 4 5 6

Figure 4.13: Measures of step response performance.

4.3.1 Overshoot
Informally, overshoot refers to the situation when the magnitude of a stable system’s step response gets larger
than the magnitude of its steady-state value. We have seen that stable underdamped systems always have
overshoot while first order systems, critically damped and overdamped systems never do.

Definition 4.3.1. The percentage overshoot of a stable LTI system with scalar TF G(s) and unit step
response y(t) = L −1 G(s) 1s is

kyk∞ − |G(0)|
%OS := .
|G(0)|

To derive an expression for the overshoot of an underdamped prototype second order system (4.7), we
differentiate the step response (4.10) and then solve for the smallest time t̄ > 0 such that ẏ(t̄) = 0. Then
kyk∞ = |y(t̄)| is the least upper bound of the step response (see Figure 4.13). Doing these calculations yields
the following formula
√−ζπ
%OS = e 1−ζ 2 (underdamped prototype second order system). (4.16)

Notice that the percentage overshoot only depends on the damping ratio ζ and not on ωn . We rearrange (4.16)
to obtain the percentage overshoot as a function of the damping ratio

Version 1.2.2816, September 1, 2020

 4.3. General characteristics of a step response 106

0.9

0.8

0.7

0.6

0.5

ln (%OS)
0.4
ζ = −q . (4.17)
0.3 π 2 + (ln (%OS))2
0.2

0.1

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Figure 4.14: Damping ratio as a function of %OS (4.17).

Observe that the more damping (larger ζ) the less overshoot. We can use this observation to help us impose
constraints on a system’s damping ratio as follows. Let %OSmax denote the maximum allowable overshoot.
Then using (4.17) the specification %OSmax is translated into a specification on the damping ratio

ln (%OSmax )
%OS ≤ %OSmax ⇐⇒ ζ ≥ − q .
π 2 + (ln (%OSmax ))2

Lastly, since ζ = cos (θ) where θ is the angle in Figure 4.7,

 
max
ln (%OS )
%OS ≤ %OSmax ⇐⇒ θ ≤ arccos − q  =: θmax .
π 2 + (ln (%OSmax ))2

Therefore, in order to meet an overshoot specification, the second order system must have all its poles in a
“cone” in the left half of the s-plane

{s ∈ C : | arg (s)| ≥ π − θmax } .

Example 4.3.1. Suppose that L = 160 µH and C = 10 µF in the RLC circuit from Example 4.2.2. Find the
range of resistor values R > 0 so that the system is (i) stable (ii) underdamped and (iii) its step response
satisfies %OS ≤ 0.05.
From Exercise 4.8 we know that (i) and (ii) are satisfied if R < 8 Ω. In this example %OSmax = 0.05 so that
(iii) is satisfied if
ln (0.05)
ζ ≥ −q = 0.69.
2 2
π + (ln (0.05))
Using the expression for ζ from Example 4.2.2 we obtain the constraint
r
R C R
ζ= = ≥ 0.69 ⇐⇒ R ≥ 5.52.
2 L 8
In summary, the design specifications are met if 5.52 ≤ R < 8. N

Exercise 4.13. Draw the region in the s-plane in which the poles of an under damped second order system must lie so that
%OS < 0.05.

Version 1.2.2816, September 1, 2020

107 CHAPTER 4. FIRST AND SECOND ORDER SYSTEMS

4.3.2 Settling time

Settling time refers to how long it takes for the step response to lie within 2% of its final value.

Definition 4.3.2.
The 2% settling time of a stable LTI system with scalar TF G(s) and unit step response
y(t) = L −1 G(s) 1s is the smallest Ts > 0 such that

|G(0) − y(t)|
for all t ≥ Ts ≤ 0.02.
|G(0)|

For an underdamped second order system, based on the expression for its step response (4.10), we can
crudely estimate the settling time by looking at the envelope e−ζωn t to get

e−ζωn t ≤ 0.02.
Solving for t we get the approximate inequality t ≥ 4/ζωn to obtain1
4
Ts ≈ (underdamped prototype second order system). (4.18)
ζωn
From this relationship we see that the larger the ωn , the faster the response. Note that Ts only depends on the
real part of the poles of G(s). This is consistent with our previous observations and motivates the following rule
of thumb for underdamped second order systems: the higher the bandwidth (larger ωn ) the faster the response
(smaller Ts ).
Suppose our design specification is that the settling time must be less than Tsmax . This means that
4
Ts ≤ Tsmax ⇔ ζωn ≥ .
Tsmax
Therefore, in order to meet a settling time specification, the second order system must have all its poles
sufficiently far to the left of the s-plane
 
4
s ∈ C : Re (s) ≤ − max .
Ts

Exercise 4.14. Find the settling time for the system from Example 4.3.1 when R = 0.4 Ω.
Exercise 4.15. Draw the region in the s-plane in which the poles of an under damped second order system must lie so that Ts < 3
seconds.

4.3.3 Time to peak

The peak time Tp is the amount of time it takes for the step response to reach its maximum value.

Definition 4.3.3. The peak
time or time to peak of a stable LTI system with scalar TF G(s) and unit step
response y(t) = L −1 G(s) 1s is the smallest Tp > 0 such that

kyk∞ = |y(Tp )|.

1 5.8
It can be shown that the 2% settling time for a critically damped system is approximately given by Ts ≈ ωn
. Unless explicitly
stated, we will use (4.18) for computing the settling time of second order systems.

Version 1.2.2816, September 1, 2020

 4.4. The effect of adding poles and zeros 108

To derive an expression for Tp for an underdamped prototype second order system (4.7), we (cf. Section 4.3.1)
differentiate the step response (4.10) and then solve for the smallest time Tp > 0 such that ẏ(Tp ) = 0. Doing
so yields
π
Tp = p (underdamped prototype second order system). (4.19)
ωn 1 − ζ 2
Once again we observe that the larger the bandwidth (ωn ), the smaller Tp is and the faster the response
becomes. Also note that Tp only depends on the imaginary part of the poles for a second order system.
Exercise 4.16. Draw the region in the s-plane in which the poles of an underdamped second order system must lie so that Tp < 3
seconds.

4.3.4 Rise time

The rise time Tr of a step response is the time it takes for it to go from 10% of its final value to 90% of its final
value.

Definition 4.3.4. Let G(s) be the scalar TF of a stable LTI system with unit step response y(t) = L −1 G(s) 1s .

Let t1 > 0 and t2 > 0 be the smallest times such that y(t1 ) = 0.1yss and y(t2 ) = 0.9yss . Then the rise time of
G(s) is
Tr = t2 − t1 .

The approximate relationship for a second order system is

2.16ζ + 0.6
Tr ≈ (0.3 < ζ < 0.8). (4.20)
ωn
Once again we observe that the larger the bandwidth (larger ωn ), the faster the response.

4.4 The effect of adding poles and zeros

We stated at the start of this chapter that second order systems are very useful in the analysis and design of
more general LTI systems, i.e., higher order systems and systems with zeros. In this section we start to justify
this statement by looking at what happens when additional poles and zeros are added to the prototype second
order system.

4.4.1 Adding a stable pole

We start with the prototype second order system (4.7) but now we multiply it with a stable first order system
with a steady-state gain of one, i.e.,

Kωn2 1
Ga (s) = , τ >0 (augmented plant).
s2 + 2ζωn s + ωn2 1 + τ s

The pole that we’ve added is located at s = −1/τ ∈ C− . As τ → +∞ the new pole approaches s = 0. As τ → 0
the pole approaches −∞. From the expression for Ga (s) we see that as the new pole approaches −∞ (τ → 0)
we recover2 the TF of the second order system (4.7).
2
In this case the effect of the additional pole is only noticeable when analyzing the high frequency response.

Version 1.2.2816, September 1, 2020

109 CHAPTER 4. FIRST AND SECOND ORDER SYSTEMS

To see the effect on the system’s step response when the pole approaches s = 0 (τ → ∞) response let’s
assume that K = 1 and re-write Ga (s) using an appropriate partial fraction expansion

a bs + c
Ga (s) = + 2
1 + τ s s + 2ζωn s + ωn2

where
τ 2 ωn2
 
a ζωn 1
a= , b=− , c= − 2 a.
1 − 2ζωn τ + ωn2 τ 2 τ τ τ ζωn
From this we see that as τ → ∞, we get b, c, → 0 and the augmented system’s response approaches a first order
response.

Example 4.4.1. Consider the underdamped second order system

8
G(s) = .
s2 + 2s + 4

Here ωn = 2 and ζ = 12 . Consider the augmented plant

1 8 1
Ga (s) = G(s) = 2 .
1 + τs s + 2s + 4 1 + τ s
Figure 4.15 shows the Bode plots and step response of this system for τ ∈ {0, 0.01, 0.1, 0.2, 1, 2, 10}.
50 2.5

0
Magnitude (dB)

−50 2

−100

−150 1.5

−200
0

1
−45

−90
Phase (deg)

−135

0.5

−180

−225

−270
−3 −2 −1 0 1 2 3 4 0
10 10 10 10 10 10 10 10 0 10 20 30 40 50 60
Frequency (rad/s) Time (seconds)

(a) Bode plot. (b) Step response.

Figure 4.15: Effect of adding a stable pole to a second order system (Example 4.4.1).

We summarize our observations as follows: As the additional pole gets closer to the origin (i.e., τ increases)
the bandwidth decreases, the phase is more negative and the step response is more sluggish. These effects
become less prominent as the extra pole moves further into C− (i.e., τ decreases).

4.4.2 Adding a left half-plane zero

We again start with the prototype second order system (4.7) but now we augment it with a minimum phase
zero
Kωn2
Ga (s) = 2 (τ s + 1) , τ > 0.
s + 2ζωn s + ωn2

Version 1.2.2816, September 1, 2020

 4.4. The effect of adding poles and zeros 110

The zero we’ve added is located at s = −1/τ . As τ → +∞ the zero approaches s = 0. As τ → 0 the zero
approaches −∞. To see the effect on the system response let’s assume that K = 1 and find the step response
 
−1 1
y(t) = L Ga (s)
s
ωn2
 
−1 1
=L (τ s + 1)
s2 + 2ζωn s + ωn2 s
2 ωn2
 
ωn 1 1
= L −1 + τ s
s2 + 2ζωn s + ωn2 s s2 + 2ζωn s + ωn2 s
2 ωn2
 
−1 ωn 1
=L +τ 2
s2 + 2ζωn s + ωn2 s s + 2ζωn s + ωn2
= step response of G(s) + τ × impulse response of G(s).
From this we conclude that as the extra zero approaches −∞ (i.e., τ → 0) then y(t) approaches the standard
second order step response. As the zero gets closer to s = 0 (τ → +∞) the impulse response term becomes
more noticeable in the step response.
Example 4.4.2. Consider, as in the previous example,
8
G(s) =
s2 + 2s + 4
and the augmented plant
8
Ga (s) = G(s)(τ s + 1) =
(τ s + 1).
+ 2s + 4 s2
Figure 4.16 shows the Bode plot and step response of this system for τ ∈ {0, 0.5, 1, 1.5, 2}. N
20 6

10

−10 5
Magnitude (dB)

−20

−30

−40
4
−50

−60

−70
3
−80
45

0
2

−45
Phase (deg)

−90
1

−135

−180
−2 −1 0 1 2 0
10 10 10 10 10 0 1 2 3 4 5 6 7
Frequency (rad/s) Time (seconds)

(a) Bode plot. (b) Step response.

Figure 4.16: Effect of adding a minimum phase zero to a second order system (Example 4.4.2).

We summarize our observations as follows: As the minimum phase zero moves closer to the origin, i.e., as τ
increases, the bandwidth increases, the phase is more positive and the response gets faster but more oscillatory.
As the zero moves closer to −∞, i.e., as τ decreases, the response approaches that of the prototype second order
system. Because of these relationships we will generally try to keep zeros away from the origin.

4.4.3 Adding a right half-plane zero

Zeros in the closed right half plane are called non-minimum phase (see Definition 3.8.2). Starting once again
with the prototype second order system (4.7) we augment it with a right half-plane zero
Kωn2
Ga (s) = (1 − τ s) , τ > 0.
s2 + 2ζωn s + ωn2

Version 1.2.2816, September 1, 2020

111 CHAPTER 4. FIRST AND SECOND ORDER SYSTEMS

The added zero is located at s = 1/τ . Generally, a non-minimum phase zero makes a system hard to control
(see Section 9.6.2). It slows the system down and causes the system to go in the “wrong direction” at first. To
see why, let’s assume that K = 1 and find the step response
 
−1 1
y(t) = L Ga (s)
s
ωn2
 
1
= L −1 (1 − τ s)
s2 + 2ζωn s + ωn2 s
2 ωn2
 
−1 ωn 1 1
=L − τs 2
s2 + 2ζωn s + ωn2 s s + 2ζωn s + ωn2 s
= step response of G(s) − τ × impulse response of G(s).

The negative sign in front of the impulse response in the above derivation is what causes the response to “go
the wrong way.”

Example 4.4.3. Consider again

8
G(s) =
s2 + 2s + 4
and the augmented system
8
Ga (s) = G(s)(1 − τ s) = (1 − τ s).
s2 + 2s + 4
Figure 4.17 shows the Bode plots and step response of this system for τ ∈ {0.1, 0.5, 1, 1.5, 2}. N
20 3

10

0
2

−10
Magnitude (dB)

−20
1
−30

−40

0
−50

−60
360
−1

270
−2
Phase (deg)

180
−3

90
−2 −1 0 1 2 3 −4
10 10 10 10 10 10 0 1 2 3 4 5 6 7
Frequency (rad/s) Time (seconds)

(a) Bode plot. (b) Step response.

Figure 4.17: Effect of adding a non-minimum phase zero to a second order system (Example 4.4.3).

We summarize this section as follows: As the non-minimum phase zero moves closer to the origin, i.e., as τ
increases, the bandwidth increases, the phase is more negative. Furthermore, the step response goes further in
the wrong direction near t = 0.

4.5 Dominant poles and zeros

The formulas we derived in Section 4.3 apply to underdamped second order systems. In general, it is very
difficult to obtain closed-form formulas to evaluate the performance of higher order systems. However, in many
cases, there are natural separations among the system poles and zeros. For example, some poles and zeros are
much closer to the imaginary axis than others as illustrated in Figure 4.18.

Version 1.2.2816, September 1, 2020

 4.5. Dominant poles and zeros 112

Figure 4.18: Pole-zero configuration with p1 , p̄1 as dominant poles.

In Section 4.4 we showed that poles and zeros “far to the left” do not have a significant impact on the
low frequency response3 of a system. Hence good low-order approximations can be found by appropriately
neglecting the less significant poles and zeros. Typically, we neglect poles and zeros that are at least 5 times
further away from the imaginary axis. The poles and zeros close to the imaginary axis are called dominant
poles and dominant zeros. The poles and zeros at least 5 times further away from the imaginary axis are
called non-dominant poles and non-dominant zeros.
Example 4.5.1. (Model Reduction)
s + 10
G(s) =
(s + 11)(s + 12)(s2 + 2s + 2)
s + 10 1
= 2
.
(s + 11)(s + 12) s + 2s + 2
| {z } | {z }
=:Gfast (s) =:Gslow (s)

The portion of the time response due to Gfast (s) reaches steady-state much faster than the portion of the time
response due to Gslow (s). This motivates us to approximate Gfast (s) by its steady-state gain
10
Gfast (0) =
132
and hence
10 1
G(s) ≈ Gfast (0)Gslow (s) = 2
=: Ĝ(s).
132 s + 2s + 2
Figure 4.19a shows the frequency response of G(s) and Ĝ(s). As expected, at low frequencies the approximation
is reasonably good. The step responses for the actual and approximated systems are shown in Figure 4.19b. N
When reducing the order of system model it is important to determine the frequency range in which the
approximation should be valid.
Example 4.5.2. Consider the transfer function
1 1



1 + 60 s 1+ 900 s

G(s) = 100 1

1

1 1

.
1 + 100 s 1 + 500 s 1+ 600 s 1+ 1000 s

We want to obtain a model that describes the process well in the frequency range ω ≤ 600. In this case
1
1 + 900 s
Gfast (s) = 1
1 + 1000 s
3
The Bode plot is affected at high frequencies where the gain is typically small.

Version 1.2.2816, September 1, 2020

113 CHAPTER 4. FIRST AND SECOND ORDER SYSTEMS
0

0.04
−20

−40

−60 0.035

−80
Magnitude (dB)

−100
0.03

−120

−140
0.025
−160

−180

−200 0.02
0

0.015

−90
Phase (deg)

0.01

−180

0.005

−270 0
10
−2 −1
10
0
10
1
10
2
10
3
10 0 1 2 3 4 5 6 7
Frequency (rad/sec) Time (seconds)

(a) Bode plot. (b) Step response.

Figure 4.19: Bode and step responses of G(s) (black) and Ĝ(s) (grey) from Example 4.5.1.

and the approximated transfer function is

1


1 + 60 s
Ĝ(s) = 100 1

1

1

.
1 + 100 s 1 + 500 s 1+ 600 s

4.6 Summary
The main purpose of this chapter is to build intuition about how pole locations affect the time-domain response
and bandwidth of LTI systems. Upon finishing this chapter you should know the following.
1. The form of the TF for prototype first and second order systems.

2. Understand how changing the time constant τ changes the time domain response and bandwidth of a
prototype first order system.

3. The relationship between damping ratio ζ, undamped natural frequency ωn and the pole locations of the
prototype second order system (see Figures 4.5, 4.7).

4. The qualitative differences in the step responses for under-, over- and critically damped second order
systems.

5. Perhaps most importantly, understand how changing the damping ratio and undamped natural frequency
changes the step response of an underdamped prototype second order system.

6. Section 4.3 is very important and serves as a basis for many design techniques. You should know the
common measures used to quantify the performance of a step response.

7. Compute the overshoot, rise time, settling and time to peak for an underdamped second order systems.
You should also be able to translate specifications on these parameters into allowable pole locations for
second order systems. You should be able to draw these allowable pole locations on the complex s-plane.

8. Qualitatively understand the effect that adding poles and zeros to a second order system has on its step
response. You should be able to relate this effect to the location of the additional pole/zero.

9. Know what is meant by “dominant poles.”

10. Find a reduced order model of a system when there are dominant poles.

Version 1.2.2816, September 1, 2020

Chapter 5

Feedback control theory

In this chapter we start to develop the basic theory and tools for feedback control analysis and design in the
frequency domain. “Analysis” means you already have a controller and you want to study how good it is;
“design” of course means you want to design a controller to meet certain specifications. The most fundamental
specification is stability. Typically, good performance requires high-gain controllers, yet typically the feedback
loop will become unstable if the gain is too high.

Contents
5.1 Closing the loop . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
5.2 Stability of feedback systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
5.3 The Routh-Hurwitz criterion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
5.4 Steady-state performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
5.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135

5.1 Closing the loop

We start with a series of examples to show how a typical control design might go. The first example deals with
modelling.

Example 5.1.1. (Pendulum on a Cart - modelling)

A favourite toy control problem is to get a cart to automatically balance a pendulum1 as shown in Figure 5.1.
We follow the modelling procedure from Section 2.3.1.

Figure 5.1: Pendulum on a cart.

1
The pendulum in this example is a mathematical pendulum as opposed to the physical pendulum from Example 2.3.3. The
difference is that in a mathematical pendulum the mass is assumed to be concentrated at the end and connected by a massless,
rigid rod of length L.

114
115 CHAPTER 5. FEEDBACK CONTROL THEORY

1. Our reference frame is such that down is the positive vertical direction and right is the positive horizontal
direction. Positive angular displacements correspond to clockwise rotations of the pendulum.

2. The coordinates are already suggested by the figure to be (q, θ).

3. The system has two mechanical components: the cart and the ball.

4. The free body diagrams are shown in Figure 5.2.

Figure 5.2: Free body diagram for the pendulum on a cart.

5. The position of the centre of mass of the cart is q. We ignore its vertical position since it has no motion
in that direction. The position of the centre of mass of the ball is q + L sin (θ) (horizontal direction) and
L − L cos (θ) (vertical direction). So when θ = 0 (upward vertical position), the ball is located at (q, 0).
When θ = π (downward vertical position) the ball is located at (q, 2L).

6. Newton’s law for the cart is

M1 q̈ = u − F1 sin θ.
Newton’s law for the ball in the horizontal direction is
d2
M2 (q + L sin θ) = F1 sin θ
dt2
and in the vertical direction is
d2
M2 (L − L cos θ) = M2 g − F1 cos θ.
dt2

These are three differential equations in the four signals q, θ, u, F1 . Use

d2 d2
sin θ = θ̈ cos θ − θ̇2 sin θ, cos θ = −θ̈ sin θ − θ̇2 cos θ
dt2 dt2
to get
M2 q̈ + M2 Lθ̈ cos θ − M2 Lθ̇2 sin θ = F1 sin θ
M2 Lθ̈ sin θ + M2 Lθ̇2 cos θ = M2 g − F1 cos θ
M1 q̈ = u − F1 sin θ.
We can eliminate F1 : Add the first and the third to get

(M1 + M2 )q̈ + M2 Lθ̈ cos θ − M2 Lθ̇2 sin θ = u;

multiply the first by cos θ, the second by sin θ, add, and cancel M2 to get

q̈ cos θ + Lθ̈ = g sin θ.

Version 1.2.2816, September 1, 2020

 5.1. Closing the loop 116

Solve the latter two equations for q̈ and θ̈:

u + M2 Lθ̇2 sin θ
    
M1 + M2 M2 L cos θ q̈
= .
cos θ L θ̈ g sin θ

Thus
u + M2 Lθ̇2 sin θ − M2 g sin θ cos θ
q̈ =
M1 + M2 sin2 θ
−u cos θ − M2 Lθ̇2 sin θ cos θ + (M1 + M2 )g sin θ
θ̈ = .
L(M1 + M2 sin2 θ)
The natural state is (positions and velocities) x = (x1 , x2 , x3 , x4 ) = (q, θ, q̇, θ̇) so that in terms of state variables
we have

ẋ1 = x3
ẋ2 = x4
u + M2 Lx24 sin x2 − M2 g sin x2 cos x2
ẋ3 =
M1 + M2 sin2 x2
−u cos x2 − M2 Lx24 sin x2 cos x2 + (M1 + M2 )g sin x2
ẋ4 = .
L(M1 + M2 sin2 x2 )
Again, these have the form
ẋ = f (x, u).
We might take the output to be the positions
   
q x1
y= = = h(x).
θ x2

The system is highly nonlinear; as you would expect, it can be approximated by a linear system for |θ| small
enough, say less than 5◦ . N
In order to apply the analysis and design tools of this course we need an LTI plant. In the next example we
linearize the pendulum on the cart.
Example 5.1.2. (Pendulum on a Cart - linearization) An equilibrium configuration (Definition 2.5.1)
x̄ = (x̄1 , x̄2 , x̄3 , x̄4 ) , ū satisfies f (x̄, ū) = 0, i.e.,
x̄3 = 0
x̄4 = 0
ū + M2 Lx̄24 sin x̄2 − M2 g sin x̄2 cos x̄2 = 0
−ū cos x̄2 − M2 Lx̄24 sin x̄2 cos x̄2 + (M1 + M2 )g sin x̄2 = 0.
Multiply the third equation by cos x̄2 and add to the fourth: We get in sequence

−M2 g sin x̄2 cos2 x̄2 + (M1 + M2 )g sin x̄2 = 0

sin x̄2 (M1 + M2 sin2 x̄2 ) = 0

sin x̄2 = 0
x̄2 = 0 or π.
Thus the equilibrium configurations are described by

x̄ = (arbitrary, 0 or π, 0, 0), ū = 0.

Version 1.2.2816, September 1, 2020

117 CHAPTER 5. FEEDBACK CONTROL THEORY

This makes good physical sense because the system is at rest when there are no applied forces; all the velocities
are zero; the pendulum is pointing either up or down; and the position of the cart doesn’t matter. We have to
choose x̄2 = 0 (pendulum up) or x̄2 = π (pendulum down). Let’s take x̄2 = 0. Then the Jacobians compute to

0 0 1 0
   
0
 0 0 0 1   0 
A= , B= .
 0 −MM1 g
2
0 0   1
M1

M1 +M2 g 1
0 M1 L 0 0 − LM1

Remark 5.1.1. Example 5.1.2 uses the general method of linearizing introduced in Section 2.5. In this partic-
ular example, there’s a faster way, which is to approximate sin θ ≈ θ, cos θ ≈ 1 in the original equations:

M1 q̈ = u − F1 θ

d2
M2 (q + Lθ) = F1 θ
dt2
0 = M2 g − F1 .
These equations are already linear and lead to the above matrices A and B. 

Now we close the loop on the cart-pendulum system.

Example 5.1.3. (Pendulum on a Cart - control design) Consider the linearized cart-pendulum from the
previous example. Take M1 = 1 kg, M2 = 2 kg, L = 1 m, g = 9.8 m/s2 . Then the state-space model is
   
0 0 1 0 0
 0 0 0 1   0 
ẋ = Ax + Bu, A=  0 −19.6 0 0 
, B =  1 .
 

0 29.4 0 0 −1

Let’s suppose we designate the cart position as the only output: y = x1 . Then
 
C= 1 0 0 0 .

The transfer function from u to y is (verify!)

Y (s) s2 − 9.8
= P (s) = 2 2 .
U (s) s (s − 29.4)

(a) Pole-zero map. (b) Block diagram.

Figure 5.3: Linearized pendulum on a cart.

Version 1.2.2816, September 1, 2020

 5.2. Stability of feedback systems 118

Figure 5.4: Control architecture for cart-pendulum example.

The poles and zeros of P (s) are shown in Figure 5.3a. Having three poles in Re (s) ≥ 0, the plant is quite
unstable. The right half-plane zero doesn’t contribute to the degree of instability, but, as we shall see, it does
make the plant quite difficult to control. The block diagram of the plant by itself is shown in Figure 5.3b.
Let us try to stabilize the plant by feeding back the cart position, y, comparing it to a reference r, and
setting the error r − y as the controller input as shown in Figure 5.4. Here C(s) is the transfer function of the
controller to be designed. One controller that does in fact stabilize is

10395s3 + 54126s2 − 13375s − 6687

C(s) = .
s4 + 32s3 + 477s2 − 5870s − 22170
The controller itself, C(s), is unstable, as is P (s). But when the controller and plant are connected in feedback,
the system is stable. If the pendulum starts to fall, the controller causes the cart to move, in the appropriate
direction, to make the pendulum tend to come vertical again. You’re invited to simulate the closed-loop system;
for example, let r be the input shown in Figure 5.5a. This reference signal corresponds to a command that

1.5

0.5
y(t)

−0.5

−1

−1.5
0 2 4 6 8 10 12
time (s)

(a) Reference signal (desired cart position). (b) Output signal (cart position).

Figure 5.5: Closed-loop response of pendulum on a cart (Example 5.1.3).

the cart move right 0.1 m for 5 seconds, then return to its original position. Figure 5.5b is a plot of the cart
position x1 versus t. The cart moves rather wildly as it tries to balance the pendulum—it’s not a good controller
design—but it does stabilize.
We mentioned that our controller C(s) is open-loop unstable. It can be proved (it’s beyond the scope of this
course) that every controller that stabilizes this P (s) is itself unstable (see the notion of strong stabilization
in, for instance, [Doyle et al., 1990]). N

5.2 Stability of feedback systems

Our objective in this section is to define what it means for the feedback system in Figure 5.6 to be stable. We
will give two definitions. The two definitions look quite distinct at first, but they are exactly the same under
mild assumptions.
Consider the setup in Figure 5.6. The signals and systems in this figure are:

Version 1.2.2816, September 1, 2020

119 CHAPTER 5. FEEDBACK CONTROL THEORY

Figure 5.6: Unity feedback system.

systems P (s), plant transfer function

C(s), controller transfer function

signals r, reference (or command) input

e, tracking error
d, disturbance
u, plant input
y, plant output.
The signals coming from the outside — r and d — are called exogenous inputs. The next example shows
the there is actually something to talk about in regards to stability of interconnected transfer functions and
illustrates the difficulties that can arise due to even simple interconnections.
Example 5.2.1. (Unstable Pole-Zero Cancellation) Consider the series connection shown in Figure 5.7.
Using the ideas of Section 2.9 we have

Figure 5.7: Cascade connection of Example 5.2.1.

Y (s) 1
=
R(s) s+1
which by Theorem 3.5.4 is BIBO stable. On these grounds we’d say that the system is BIBO stable, wash our
hands of the stability question and walk away. In doing so we’d be too hasty. To see why this is so, suppose
that the system admits some noise as in Figure 5.8.

Figure 5.8: Noise at the interface of two systems.

The transfer function from n to y is

Y (s) 1
= .
N (s) s−1
which by Theorem 3.5.4 is unstable. Thus any slight perturbation in the signal as it passes from the first system
to the second will potentially be dangerously magnified in the output. N

Version 1.2.2816, September 1, 2020

 5.2. Stability of feedback systems 120

Definition 5.2.1. The feedback system in Figure 5.6 is well-posed if 1 + C(s)P (s) is not identically equal to
zero. Otherwise, it is said to be ill-posed.

An ill-posed system is not meaningful, does not work properly, and should be avoided.

Example 5.2.2. (Ill-Posed) Consider the feedback system in Figure 5.6 with

P (s) = 1, C(s) = −1.

In this case 1 + C(s)P (s) ≡ 0 and the system is ill-posed. The dependent signals e, u and y cannot be uniquely
determined from the independent signals r and d. N

We make the following standing assumption throughout this section.

Assumption 5.2.2. The plant P (s) and controller C(s) are rational, the controller C(s) is proper, and the
plant P (s) is strictly proper. J

Exercise 5.1. Show that Assumption 5.2.2 implies that 1 + P C is not identically zero.
Assumption 5.2.2 is also ensures that the closed-loop system is causal as illustrated in the next example

Example 5.2.3. Consider the feedback system in Figure 5.6 with

1 + 2s + as2
C(s)P (s) =
s2 + s + 1
where a ∈ R is unspecified. If a 6= 0 then Assumption 5.2.2 is violated. The transfer function from r to y is
(see Section 2.9)
Y (s) C(s)P (s) 1 + 2s + as2
= = .
R(s) 1 + C(s)P (s) (1 + a)s2 + 3s + 2
If, and only if, a = −1 the TF is improper and hence non-causal2 . N

5.2.1 Internal stability

For this concept we set the reference input r(t) and the disturbance input d(t) to zero, i.e., r(t) = d(t) = 0. We
bring in state-space models for the plant and controller as in Figure 5.9. The closed-loop equations are

Figure 5.9: Internal stability.

ẋp = Ap xp + Bp u
ẋc = Ac xc + Bc e
u = Cc x c + D c e
e = −Cp xp .
2
And, incidentally, also unstable by Theorem 3.5.5.

Version 1.2.2816, September 1, 2020

121 CHAPTER 5. FEEDBACK CONTROL THEORY

Let the closed-loop state be xcl := (xp , xc ) so that (verify!)

  
Ap − B p D c C p B p C c xp
ẋcl = Acl xcl := .
−Bc Cp Ac xc

Definition 5.2.3. The feedback system in Figure 5.6 is internally stable if the state model ẋcl = Acl xcl is
asymptotically stable.

The concept of internal stability means that with no exogenous inputs applied (r = d = 0), the internal
states xp (t), xc (t) decay to zero for every initial state xp (0), xc (0). By Proposition 3.4.2 the feedback system is
internally stable if and only if all the eigenvalues of the matrix Acl have negative real part.

Example 5.2.4. Suppose the plant has state-space model

ẋp = xp + u
y = xp

so that Ap = 1, Bp = 1 and Cp = 1 and P (s) = 1/(s − 1). The controller is given by

u = 2e

so that Ac = 0, Bc = 0, Cc = 0, Dc = 2 and C(s) = 2. Therefore

Acl = Ap − Bp Cc Cp = −1.

The 1 × 1 matrix Acl has one eigenvalue at λ = −1. Therefore, the unstable plant 1/(s − 1) is internally
stabilized by unity feedback with a proportional (pure gain) controller 2. N

Example 5.2.5. We re-visit Example 5.2.1 and show that we cannot internally stabilize the plant P (s) =
1/(s − 1) by cancelling the unstable pole. Consider the controller C(s) = (s − 1)/(s + 1) and the system in
Figure 5.7. We saw in Example 5.2.1 that the transfer function from r to y is BIBO stable but that any noise at
the interface of the systems will potentially be dangerously magnified in the output. We now show that while
the instability is not apparent by only looking at input and output signals, it is apparent when we examine
internal stability.
Figure 5.10 shows state models for each transfer function. We have, with r = 0,

Figure 5.10: State-space model for Example 5.2.5.

ẋp = xp − 2xc
ẋc = −xc

and therefore     
ẋp 1 −2 xp
= .
ẋc 0 −1 xc
The system is not internally stable since the eigenvalues of this matrix are {1, −1}. N

Version 1.2.2816, September 1, 2020

 5.2. Stability of feedback systems 122

5.2.2 Input-output stability

We now turn the second way of thinking about stability of the feedback loop in Figure 5.6. To do so, we first
find all of the transfer functions from the independent signals r and d to the dependent signals e, u and y. We
use our systematic method of doing block diagram reduction. First, write the equations for the outputs of the
summing junctions:

E = R − PU
U = D + CE.

Assemble into a vector equation:     

1 P E R
= .
−C 1 U D
In view of our standing assumption (P strictly proper, C proper), the determinant of
 
1 P
−C 1

is not identically zero. Thus we can uniquely solve for E, U :

1 −P
 
   −1    
E 1 P R  1 + PC 1 + PC 
 R .
= =
U −C 1 D  C 1  D
1 + PC 1 + PC
The output is given by
PC P
Y = PU = R+ D.
1 + PC 1 + PC
We just derived the following closed-loop transfer functions:
1 C PC
R to E : , R to U : , R to Y : ,
1 + PC 1 + PC 1 + PC
−P 1 P
D to E : , D to U : , D to Y : .
1 + PC 1 + PC 1 + PC

Definition 5.2.4. The feedback system in Figure 5.6 is input-output stable provided e, u, and y are bounded
signals whenever r and d are bounded signals; briefly, the system from (r, d) to (e, u, y) is BIBO stable.

Input-output stability is equivalent to saying that the 6 transfer functions from (r, d) to (e, u, y) are BIBO
stable, in the sense that all poles are in Re (s) < 0. Since whenever r and e are bounded, so is y = r − e, it
suffices to look at the 4 transfer functions from (r, d) to (e, u), namely,
1 −P
 
   
E  1 + PC 1 + PC  R
=  D .

U  C 1
1 + PC 1 + PC
Remark 5.2.5. An alternative definition for input-output stability of an interconnected system is: a feedback
system is input-output stable if and only if every possible transfer function in the system (imagine introducing
noise everywhere) is BIBO stable. 

Version 1.2.2816, September 1, 2020

123 CHAPTER 5. FEEDBACK CONTROL THEORY

Example 5.2.6.
1 s−1
P (s) = , C(s) =
s2 −1 s+1
The four transfer functions are
(s + 1)2
 
s+1
   s2 + 2s + 2 (s − 1)(s2 + 2s + 2)   
E   R
=  D .

U 
 (s + 1)(s − 1) (s + 1)2 
s2 + 2s + 2 s2 + 2s + 2
Three of these are stable; the one from D to E is not. Consequently, the feedback system is not input-output
stable. This is in spite of the fact that a bounded r produces a bounded y. The problem here is that C cancels
an unstable pole of P . That isn’t allowed. N

Example 5.2.7.
1
P (s) = , C(s) = K.
s−1
The feedback system is input-output stable if and only if K > 1 (check). This again shows that a feedback
system can be stable even if the individual components are unstable. N

We now provide two ways to test input-output stability. Write

Np Nc
P = , C= .
Dp Dc

We assume the polynomials (Np , Dp ) are coprime, i.e., have no common factors, and (Nc , Dc ) are coprime too.

Definition 5.2.6. The characteristic polynomial of the feedback system in Figure 5.6 is π(s) := Np Nc +
Dp Dc .

The characteristic polynomial of a feedback system is the least common multiple of the denominators in the
four transfer functions from (r, d) to (e, u), i.e.,
−P
 1 
 
1+P C 1+P C
= 1 Dp Dc −Np Dc
 . (5.1)
C 1 Np Nc + Dp Dc Dp Nc Dp Dc
1+P C 1+P C

Example 5.2.8.
1 s−1
P (s) = , C(s) = .
s2 − 1 s+1
We have seen that with the plant and controller combination the system is not feedback stable because the trans-
fer function from d to e is unstable. Notice the unstable pole-zero cancellation. The characteristic polynomial
is
π(s) = s − 1 + (s2 − 1)(s + 1) = (s − 1)(s2 + 2s + 2).
This has a right half-plane root – precisely the root that the controller cancels. N

Theorem 5.2.7. The feedback system is input-output stable if and only if the characteristic polynomial has no
roots with Re(s) ≥ 0.

Version 1.2.2816, September 1, 2020

 5.2. Stability of feedback systems 124

Proof. (⇐ Sufficiency) If Np Nc + Dp Dc has no roots in Re(s) ≥ 0, then the four transfer functions on the
left-hand side of (5.1) have no poles in Re(s) ≥ 0, and hence they are stable.
(⇒ Necessity) Conversely, assume the feedback system is stable, that is, the four transfer functions on the
left-hand side of (5.1) are BIBO stable. To conclude that Np Nc + Dp Dc has no roots in Re(s) ≥ 0, we must
show that the polynomial Np Nc + Dp Dc does not have a common factor with all four numerators in (5.1),
namely, Dp Dc , Np Dc , Nc Dp . That is, we must show that the four polynomials

Np Nc + Dp Dc , Dp Dc , Np Dc , Nc Dp

are coprime. This part is left for you.

Exercise 5.2. Show that the polynomials Np Nc + Dp Dc , Dp Dc , Np Dc , Nc Dp do not have a common root. Hint: use the fact
that the pair (Np , Dp ) is coprime and the pair (Nc , Dc ) is coprime.

Definition 5.2.8. The plant P (s) and controller C(s) have a pole-zero cancellation at s = λ ∈ C if

Np (λ) = Dc (λ) = 0 (controller pole cancels a plant zero)

Dp (λ) = Nc (λ) = 0 (controller zero cancels a plant pole).

It is called an unstable pole-zero cancellation if Re (λ) ≥ 0.

Corollary 5.2.9. If there is an unstable pole-zero cancellation, then the feedback system is not input-output
stable.
+
Proof. If there is an unstable pole-zero cancellation at λ ∈ C , then

π(λ) = Np (λ)Nc (λ) + Dp (λ)Dc (λ) = 0 + 0.

+
Thus π has a root in C so that by Theorem 5.2.7 the system is not input-output stable.

Remark 5.2.10. In practice, it is not possible to exactly cancel a plant zero or pole because of modelling errors.
In Example 5.2.6 therefore the transfer function from r to y will also be unstable. However, it is important to
stress that even in the ideal case with a perfect pole-zero cancellation, we still get an unstable system. 

A second way to test feedback stability is as follows.

Theorem 5.2.11. The feedback system is input-output stable if, and only if,

1. The transfer function 1 + P C has no zeros in Re (s) ≥ 0, and

2. the product P C has no unstable pole-zero cancellations.

Exercise 5.3. Prove Theorem 5.2.11.

Example 5.2.9.
1 s−1
P (s) = , C(s) =
s2 − 1 s+1
Check that 1) holds but 2) does not. N

Version 1.2.2816, September 1, 2020

125 CHAPTER 5. FEEDBACK CONTROL THEORY

Figure 5.11: Non-unity feedback system.

Remark 5.2.12. Occasionally we will encounter non-unity feedback system as shown in Figure 5.11. This is
the case when we model the dynamics of the sensor that provides the feedback using a transfer function H(s).
The exogenous signal n models sensor noise. If we write
Np Nc Nh
P = , C= , H=
Dp Dc Dh

then you can check that the characteristic polynomial becomes

π(s) = Nc Np Nh + Dc Dp Dh .

In this case we have feedback stability if and only if all the roots of π(s) are in C− . 

Exercise 5.4. Find the 9 transfer functions from (r, d, n) to (v, u, y).

5.2.3 Internal stability and input-output stability

We have given two definitions for what it means for the system in Figure 5.6 to be stable. Internal stability
(Definition 5.2.3) dealt with the closed-loop state model whereas input-output stability (Definition 5.2.4) dealt
with the various transfer functions in the loop. It can be shown that the roots of the characteristic polynomial
π(s) are a subset of the eigenvalues of Acl , i.e., the roots of π(s) are a subset of the roots of det (sI − Acl ).
Hence
internal stability =⇒ input-output stability.
Usually the roots of π(s) are identical to the eigenvalues of Acl . In these cases, the two stability concepts are
equivalent. More precisely, if the pair (Ap , Bp ) is stabilizable and the pair (Ap , Cp ) is detectable then internal
stability is equivalent to input-output stability [Dullerud and Paganini, 2000]. The notions of stabilizability and
detectability are beyond the scope of this course and are covered in ECE488.
In this course internal stability and input-output stability of a feedback system are assumed to be equivalent
and hence when we talk about feedback stability there is no ambiguity. This is reasonable for most practical
applications.

5.3 The Routh-Hurwitz criterion

In practice, one checks input-output stability using numerical algorithms to compute the roots of a polynomial.
However, if some of the coefficients of the polynomial are not fixed, e.g., a controller gain, then we can’t solve
for the roots numerically. The Routh-Hurwitz criterion provides a test to check if the roots of π(s) are in C−
without actually finding them.

Version 1.2.2816, September 1, 2020

 5.3. The Routh-Hurwitz criterion 126

Consider an nth order polynomial with real coefficients π ∈ R[s], i.e.,

π(s) = sn + an−1 sn−1 + · · · + a1 s + a0 , ai ∈ R, n 6= 0. (5.2)

Definition 5.3.1. A polynomial π ∈ R[s] is Hurwitz if all its roots lie in C− .

The Routh-Hurwitz criterion is an algebraic test for a polynomial π(s) to be Hurwitz without having to
calculate its roots [Routh, 1877].

5.3.1 A necessary condition for a polynomial to be Hurwitz

We make the observation that a necessary, but not sufficient, condition for π(s) to be Hurwitz is that all the
coefficients ai have the same sign.

Proof. Assume that π ∈ R[s] is Hurwitz. Let {λ1 , . . . , λr } be the real roots of π and let {µ1 , µ̄1 , . . . , µs , µ̄s } be
the complex conjugate roots of π. Then n = r + 2s. Factor the polynomial as

π(s) = (s − λ1 ) · · · (s − λr )(s − µ1 )(s − µ̄1 ) · · · (s − µs )(s − µ̄s ).

For the real roots −λi > 0 since π is Hurwitz by assumption. For the complex conjugate roots we expand

(s − µi )(s − µ̄i ) = s2 − (µi + µ̄i )s + µi µ̄i = s2 − 2 Re (µi )s + |µi |2 .

Note that µi µ̄i = |µi |2 > 0 and −(µi + µ̄i ) = −2 Re (µi ) > 0 since π is Hurwitz by assumption. Therefore, when
we expand π(s), every coefficient will be have the same sign as the coefficient multiplying sn .

Example 5.3.1.
s4 + 3s3 − 2s2 + 5s + 6 (a bad root)
s3 + 4s + 6 (a bad root)
3 2
s + 5s + 9s + 1 (don’t know).
N

5.3.2 Routh’s algorithm

One of the most popular algorithms to determine whether or not a polynomial is Hurwitz is Routh’s algorithm.
We present it here without proof and refer the interested student to [Qiu and Zhou, 2010, Section 3.2] for a
proof. For the method of Routh, we construct an array involving the coefficients of the polynomial in question.
The array is constructed inductively, starting with the first two rows.
We again consider a polynomial π ∈ R[s] of degree n like (5.2). We construct Table 5.1. In Table 5.1 the
first two rows are populated using the coefficients of π. The third row is computed from the first two
an−1 an−2 − an−3 an−1 an−4 − an−5 an−1 an−6 − an−7
r2,0 = , r2,1 = , r2,2 = ,...
an−1 an−1 an−1
The remaining rows are computed recursively using the same pattern
  
ri−1,0 ri−2,j+1 − ri−2,0 ri−1,j+1 n−i
ri,j = , i ∈ {2, . . . , n} , j∈ 0, . . . , . (5.3)
ri−1,0 2
The construction terminates if we end up with a zero in the first column. There are at most n + 1 rows in the
final table.

Version 1.2.2816, September 1, 2020

127 CHAPTER 5. FEEDBACK CONTROL THEORY

Table 5.1: Routh array.

sn r0,0 = 1 r0,1 = an−2 r0,2 = an−4 r0,3 = an−6 ···

sn−1 r1,0 = an−1 r1,1 = an−3 r1,2 = an−5 r1,3 = an−7 ···
sn−2 r2,0 r2,1 r2,2 r2,3 ···
sn−3 r3,0 r3,1 r3,2 r3,3 ···
.. .. .. ..
. . . .
s2 rn−2,0 rn−2,1 0 ···
s1 rn−1,0 0 0 ···
s0 rn,0 0 0 ···

Remark 5.3.2. The definition of the coefficients (5.3) can be expressed using determinants
    
1 ri−2,0 ri−2,j+1 n−i
ri,j = − det , i ∈ {2, . . . , n} , j ∈ 0, . . . , .
ri−1,0 ri−1,0 ri−1,j+1 2

Theorem 5.3.3 (Routh-Hurwitz Stability Criterion). Consider a polynomial π ∈ R[s] given by (5.2) and its
associated Routh array in Table 5.1. The number of roots of π with real part greater than or equal to zero equals
the number of sign changes in the first column of the array.

Example 5.3.2.
π(s) = a2 s2 + a1 s + a0 , a2 6= 0.
s2 a2 a0 0
s1 a1 0 0
a1 a0
s0 a1 = a0 0 0
We conclude that all the roots of π are in C− if and only if a0 , a1 , a2 have the same sign. N

Example 5.3.3.
π(s) = 2s4 + s3 + 3s2 + 5s + 10.
s4 2 3 10 0
s3 1 5 0 0
3−10
s2 1 = −7 10 0 0
−35−10
s1 −7 = 45
7 0 0 0
s0 10 0 0 0
There are two sign changes in the first column so π(s) has two roots in C+ . N
The next example illustrates the utility of the Routh-Hurwitz criterion.
Example 5.3.4. Consider the unity feedback system in Figure 5.12 with
1
P (s) = , C(s) = Kp .
s4 + 6s3 + 11s2 + 6s
Find all Kp ∈ R such that the feedback system is stable. From Section 5.2.2 we know that the system is
feedback stable if and only if all the roots of π(s) are in C− . In this case

π(s) = Np Nc + Dp Dc = s4 + 6s3 + 11s2 + 6s + Kp .

Version 1.2.2816, September 1, 2020

 5.3. The Routh-Hurwitz criterion 128

Figure 5.12: System for Example 5.3.4.

We now apply the Routh-Hurwitz stability criterion

s4 1 11 Kp 0
s3 6 6 0 0
s2 10 Kp 0 0
3
s1 5 (10 − Kp ) 0 0 0
s0 Kp 0 0 0

We need all entries in the first column to be positive. From the s1 row we get the constraint Kp < 10. From
the s0 row we get the constraint Kp > 0. Therefore the system is feedback stable provided Kp ∈ (0, 10). N

Special cases
In using Routh’s array there are sometimes special cases that require extra steps to be taken. For example, we
see that in building the array, we cannot proceed when one of the elements in the first column is zero. In this
case we can immediately conclude that the polynomial is unstable.
Example 5.3.5. Consider π(s) = s5 + 3s4 + 2s3 + 6s2 + 3s + 3. We then have

s5 1 2 3
s4 3 6 3
s3 0 2
s2 r3,0
s1
s0

Since r2,0 = 0, the next element r3,0 cannot be computed. We can stop at this point and conclude that π(s) is
unstable since the elements in the first column aren’t all positive. N
The second special case we consider is when an entire row of Routh’s array is zero. If row sn−k is all zero
it means that the polynomial πa (s) = rk−1,0 sn−k+1 + rk−1,1 sn−k−1 + rk−1,2 sn−k−3 + rk−1,3 sn−k−5 + · · · divides
the original polynomial π. The polynomial πa is sometimes called an auxiliary polynomial. It consists of
only even or odd powers of s, where the coefficients come from the row above the row of zeros. By the result
from Section 5.3.1 we conclude that πa is not Hurwitz and therefore niether is π.
Example 5.3.6. Consider π(s) = s6 + 5s5 + 2s4 + 5s3 + 4s2 + 15s + 3. The associated Routh table is

s6 1 2 4 3
s5 5 5 15 0
s4 1 1 3
s3 0 0
s2
s1
s0 .

Thus πa (s) = s4 + s2 + 3 and π(s) = πa (s) s2 + 5s + 1 . The roots of πa (s) are s = ±0.785 ± j1.06.


N

Version 1.2.2816, September 1, 2020

129 CHAPTER 5. FEEDBACK CONTROL THEORY

5.4 Steady-state performance

For any control system, stability is mandatory; good performance is desirable. Two types of performance
measures are:

1. Transient performance: depends in a complicated way on the location of the closed-loop poles and zeros.
If the closed-loop system is dominated by a first or second order system, then we can use the results of
Chapter 4.

2. Steady-state performance: This refers to specifications like tracking reference signals and rejecting distur-
bance signals.

In this section we discuss steady-state performance.

5.4.1 Tracking reference signals

Cruise control in a car regulates the speed to a prescribed set point. What is the principle underlying its
operation? The answer lies in the final-value theorem (Theorem 3.6.1).

Example 5.4.1. Consider the unity feedback system in Figure 5.13 with

Figure 5.13: System for Example 5.4.1.

1 1
P (s) = , C(s) = .
s+1 s
Let r be a step input, r(t) = r0 1(t). Then we have
1
E(s) = R(s)
1 + P (s)C(s)
s(s + 1) r0
=
s2 + s + 1 s
s+1
= 2
r0 .
s +s+1
The FVT applies to E(s), and e(t) → 0 as t → ∞. Thus the feedback system provides asymptotic tracking of
step reference signals with zero steady-state error.
How it works: C(s) contains an internal model of R(s) (i.e., an integrator); closing the loop creates a zero
from R(s) to E(s) to exactly cancel the unstable pole of R(s). (This isn’t an illegal pole-zero cancellation.) N

Remark 5.4.1. Tracking error is always defined by e := r − y where r is the command or reference input and
y is the plant output. In a unity feedback control system, as in Figure 5.6, the tracking error is the input to
the controller. If the sensor TF in the feedback path is not 1, as in Figure 5.11, then the signal coming out of
the negative feedback summing junction is not the tracking error. 

Version 1.2.2816, September 1, 2020

 5.4. Steady-state performance 130

Step inputs
Let’s generalize the preceding example. Consider the unity feedback control system in Figure 5.6 and write

E(s) 1 Dp Dc
= = . (5.4)
R(s) 1 + CP Np Nc + Dp Dc

Suppose that R(s) = r0 /s, i.e., r(t) = r0 1(t) for some constant r0 ∈ R. If we assume that the controller C(s)
stabilizes the closed-loop system, then π(s) = Np Nc + Dp Dc has all its roots in C− , the TF (5.4) is BIBO stable,
and the final-value theorem applies. Application of the FVT to the tracking error yields
1 s r0 r0
ess = lim e(t) = lim sE(s) = lim s R(s) = lim = lim .
t→∞ s→0 s→0 1 + C(s)P (s) s→0 1 + C(s)P (s) s s→0 1 + C(s)P (s)

Therefore ess = 0 if and only if P (0)C(0) = ∞. This means that P (s)C(s) has at least one pole at s = 0, i.e.,
at least one integrator.
If P (s) does not have a pole at s = 0 and we want perfect step tracking, then we must choose
1
C(s) = C1 (s)
s
so that C(s)P (s) has an integrator. Then C1 (s) is designed to provide feedback stability. The integrator adds
more phase lag though and this can make the overall system harder to stabilize.

Ramp inputs
Now let’s try a non-constant reference input.

Example 5.4.2. This time take

2s + 1 1
P (s) = , C(s) =
s(s + 1) s
and take r to be a ramp, r(t) = r0 t, t ≥ 0 for some constant r0 ∈ R. Then R(s) = r0 /s2 and so (verify!)
s+1
E(s) = r0 .
s3 + s2 + 2s + 1
You can check using, say, the Routh-Hurwitz stability criterion, that the poles of E(s) are in Re (s) < 0 which
means that the FVT can be applied to get limt→0 e(t). Again e(t) → 0; perfect tracking of a ramp. Here C(s)
and P (s) together provide the internal model, a double integrator. N

Generalizing the preceding example, suppose that R(s) = r0 /s2 (ramp reference r(t) = r0 t, t ≥ 0, r0 ∈ R).
If we assume that the controller C(s) provides input-output stability then π(s) has all its roots in C− and the
TF (5.4) is BIBO stable. Let’s try and apply the final-value theorem

s r0 r0
ess = lim e(t) = lim sE(s) = lim 2
= .
t→∞ s→0 s→0 1 + C(s)P (s) s sC(s)P (s) s=0

We observe:

(i) If C(s)P (s) has no poles at s = 0 then we can’t even apply the FVT since sE(s) has bad poles. It
turns out that ess is unbounded in this case. This doesn’t contradict the assumption that C(s) provides
input-output stability because the input is unbounded.
r0
(ii) If C(s)P (s) has one pole at s = 0 then ess = C(0)P (0) is finite and non-zero.

(ii) If C(s)P (s) has two or more poles at s = 0 then ess equals zero.

Version 1.2.2816, September 1, 2020

131 CHAPTER 5. FEEDBACK CONTROL THEORY

Internal model principle

It is possible to generalize the above discussion to arbitrary reference signals.

Theorem 5.4.2 (Internal model principle). Assume that P (s) is strictly proper, C(s) is proper and the feedback
system is stable. If C(s)P (s) contains an internal model of the unstable part of R(s), then perfect asymptotic
tracking occurs.

The internal model principle of Theorem 5.4.2 can be understood as follows. Suppose that the Laplace
transform of the reference signal r(t) is R(s) = L {r(t)}. Write

Nr (s) Nr (s)
R(s) = = +
Dr (s) Dr (s)Dr− (s)

where the roots of the polynomial Dr− (s) are in C− (Re (s) < 0) while the roots of the polynomial Dr+ (s) are
+
in C (Re (s) ≥ 0). The internal model principle says that in order for the plant output y(t) to asymptotically
converge to r(t), the product C(s)P (s) should have the form

N (s)
C(s)P (s) =
D(s)Dr+ (s)

where N (s), D(s) ∈ R[s] are polynomials. That is, the product C(s)P (s) must have a copy of Dr+ (s) in its
denominator.

Example 5.4.3. Let

1
P (s) =.
s+1
We want the closed-loop system to be stable and to track the reference signal

r(t) = r0 sin (t).

Then
r0
R(s) = .
s2 +1
In this case Nr (s) = r0 and Dr+ (s) = s2 + 1, Dr− (s) = 1. Since the plant doesn’t contain a copy of Dr+ (s), the
internal model principle suggests that we should pick a controller of the form
1 1
C(s) = C1 (s) = C1 (s)
Dr+ (s) s2 +1

where C1 (s) ∈ R(s) is a TF yet to be determined. That is, we embed an internal model of the unstable part
of R(s) in C(s) and allow an extra factor C1 (s) to achieve feedback stability. You can check that C1 (s) = s
works. N

Remark 5.4.3. The internal model principle of Theorem 5.4.2 is only concerned with the unstable part of the
Laplace transform R(s) of the reference signal r(t). For example, if r(t) = e−t + sin (t) then

1 1 s2 + s + 2
R(s) = + 2 = .
s+1 s +1 (s + 1)(s2 + 1)

In this case Dr− (s) = s + 1 and Dr+ (s) = s2 + 1. The internal model needs to include the poles at s = ±j (roots
of Dr+ (s)) but not the pole at s = −1. This is because the part of r(t) corresponding to stable pole decays to
zero in steady-state and hence there is nothing to asymptotically track. 

To help understanding, it is useful to interpret the internal model principle in the frequency domain.

Version 1.2.2816, September 1, 2020

 5.4. Steady-state performance 132

20

-20

-40

Magnitude (dB)
-60

-80

-100

-120

-140

-160
135

90
Phase (deg)

45

-45
10 -2 10 -1 10 0 10 1
Frequency (rad/s)

Figure 5.14: Bode plot of E/R from Example 5.4.3.

Example 5.4.4. Returning to Example 5.4.3 we have that the TF from r to e is

E(s) 1 s3 + s2 + s + 1
= = 3 =: S(s).
R(s) 1 + C(s)P (s) s + s2 + 2s + 1

Figure 5.14 shows the Bode plot of the frequency response S(jω). Notice that the system heavily attenuates
signals at the frequency 100 = 1 rad/s. Indeed, the internal model introduces zeros at s = ±j in S which creates
a notch filter. Any reference signals at this frequency result in zero steady-state error as we saw using the FVT
in the previous example. N

Exercise 5.5. At what frequencies do you expect the Bode plot of E/R from Example 5.4.1 to heavily attenuate signals? Confirm
your response by plotting the Bode plot.
In summary, we have the following facts:

• If the closed-loop system is unstable, we can’t even talk about steady-state tracking performance since
there will likely not be a steady-state response.

• Even if the closed-loop system is stable, we have to be careful applying the FVT to compute the steady-
state tracking error because, depending on the reference signal, sE(s) may have poles in Re (s) ≥ 0.

• If the closed-loop system is stable and P (s)C(s) contains an internal model of the unstable part of R(s),
then perfect asymptotic tracking occurs.

• It does not matter if P (s) or C(s) provide the internal model. Perfect asymptotic tracking occurs when
the product C(s)P (s) has an internal model of the unstable part of R(s).

• If C(s)P (s) does not contain an internal model, then increasing the gain |C(jω)P (jω)| over the range of
frequencies that r(t) contains decreases the steady-state tracking error. This is consistent with the idea
that high-gain leads to good performance.

5.4.2 Steady-state disturbance rejection

A disturbance is an exogenous input over which we have no control, e.g., turbulence acting on a plane, load
torque on a car due to hills. Also signal noise, which in practice every signal in a control system has, can be

Version 1.2.2816, September 1, 2020

133 CHAPTER 5. FEEDBACK CONTROL THEORY

modelled as a disturbance. Consider a system with an input disturbance di , an output disturbance d0 and
sensor noise n(t) illustrated in Figure 5.15. We can once again use the final-value theorem to check for the

Figure 5.15: Feedback system with various disturbances and noise.

steady-state effect of certain types of disturbances and noise.

Example 5.4.5. Suppose that di = n = 0 in Figure 5.15 and that do (t) = 1(t). Find conditions on P (s) and
C(s) so that the steady-state effect of this disturbance on the output y is zero. The transfer function from do
to y is
Y (s) 1
= .
Do (s) 1 + C(s)P (s)H(s)
If we have feedback stability, then the above TF is stable and the steady-state effect of do is
1
lim y(t) = lim sY (s) = .
t→∞ s→0 1 + C(0)P (0)H(0)

Therefore the disturbance will have no steady-state effect on the output if and only if C(0)P (0)H(0) = ∞,
i.e., C(s)P (s)H(s) has at least one pole at the origin. We conclude that the output disturbance do (t) = 1(t)
has no effect on the steady-state output of this system as long as (i) the system is feedback stable and (ii)
C(s)P (s)H(s) has at least one pole at the origin. N
As seen in Figure 5.15, disturbances can appear at the input to a plant and/or the output to a plant. In this
section we adopt a unified description that encompasses both types of disturbances. Specifically, for a nominal
plant model P (s) with input U (s) and output Y (s), we’ll assume that a disturbance D(s) acts on the plant at
some intermediate point, i.e., we model the plant as follows:

Y (s) = P2 (s) (D(s) + P1 (s)U (s)) = P2 (s)D(s) + P2 (s)P1 (s)U (s). (5.5)

where P (s) = P2 (s)P1 (s). The generalized model of plant disturbances in a feedback loop are shown in
Figure 5.16. If P1 (s) = 1, then d(t) is an input disturbance. If P2 (s) = 1, then d(t) is an output disturbance.
Note that Figure 5.16 does not include measurement noise, i.e., it doesn’t include the signal n(t) from Figure 5.15.

Figure 5.16: Feedback system with generalized model of plant disturbances.

Version 1.2.2816, September 1, 2020

 5.4. Steady-state performance 134

Example 5.4.6. The generalized disturbance model allows us to handle more than just input and output
disturbances. Consider a plant with state-space model
     
0 1 0 0.3
ẋ(t) = x(t) + u(t) + d(t)
0 0 1 0
 
y(t) = 1 0 x(t).

Here u is the plant input while d is a disturbance. Find the transfer functions P1 (s) and P2 (s) so that the
input-output model of this plant has the structure (5.5).
The TF from U (s) to Y (s) is

 s −1 −1 0
      
Y (s)  1   s 1 0 1
= 1 0 = 2 1 0 = 2.
U (s) 0 s 1 s 0 s 1 s

The TF from D(s) to Y (s) is

 s −1 −1 0.3
      
Y (s)  1   s 1 0.3 0.3
= 1 0 = 2 1 0 = .
D(s) 0 s 0 s 0 s 0 s

Therefore
10 1 0.3
P1 (s) = , P2 (s) = .
3 s s
N

We are interested in understanding when a closed-loop system is able to asymptotically reject disturbances
in steady-state. Assume throughout that the controller C(s) provides input-output stability. The TF from
D(s) to Y (s) for the system in Figure 5.16 is

Y (s) P2 (s)
= .
D(s) 1 + C(s)P2 (s)P1 (s)

Suppose that D(s) = L {d(t)} is the Laplace transform of the disturbance signal and write

Nd (s) Nd (s)
D(s) = = +
Dd (s) Dd (s)Dd− (s)

where the roots of the polynomial Dd− (s) are in C− (Re (s) < 0) while the roots of the polynomial Dd+ (s) are
+
in C (Re (s) ≥ 0). We’d like to use the FVT to determine the steady-state effect of the disturbance on the
+
output so let’s determine conditions under which sY (s) has no poles in C . We have

P2 (s)
sY (s) = D(s)
1 + C(s)P2 (s)P1 (s)
P2 (s) Nd (s)
=s .
1 + C(s)P2 (s)P1 (s) Dd (s)Dd− (s)
+

Now write C(s) = Nc (s)/Dc (s), P1 (s) = Np1 (s)/Dp1 (s), P2 (s) = Np2 (s)/Dp2 (s) so that

Np2 (s)Dc (s)Dp1 (s) Nd (s)

sY (s) = s .
Dc (s)Dp1 (s)Dp2 (s) + Nc (s)Dp2 (s)Dp1 (s) Dd (s)Dd− (s)
+

Since the controller provides input-output stability, the characteristic polynomial π(s) = Dc (s)Dp1 (s)Dp2 (s) +
Nc (s)Dp2 (s)Dp1 (s) is Hurwitz. Therefore, sY (s) has no poles with Re (s) ≥ 0 if the product Np2 (s)Dc (s)Dp1 (s)
contains a copy of the polynomial Dd+ (s). In summary, we’ve shown that if C(s) provides input-output stability

Version 1.2.2816, September 1, 2020

135 CHAPTER 5. FEEDBACK CONTROL THEORY

and the product Np2 (s)Dc (s)Dp1 (s) contains a copy of the polynomial Dd+ (s), then the final-value theorem
applies and
Np2 (s)Dc (s)Dp1 (s) Nd (s)
lim y(t) = lim sY (s) = lim s = 0.
t→∞ s→0 s→0 Dc (s)Dp1 (s)Dp2 (s) + Nc (s)Dp2 (s)Dp1 (s) Dd (s)Dd− (s)
+

This suggests controllers of the form

1
C(s) = C1 (s)
Dd+ (s)
where the term 1/Dd+ (s) provides the internal model of the disturbance while the term C1 (s) is used to stabilize
the loop.
Remark 5.4.4. The derivations of this section show that in order to simultaneously reject input and output
disturbances, the controller C(s), not the product C(s)P (s), must contain an internal model of the disturbance.

Sometimes we can’t track and reject disturbances simultaneously as the next example illustrates.
Example 5.4.7. (Tracking steps while rejecting measurement error) Consider the feedback system in
Figure 5.17. Suppose that we want the closed-loop system to asymptotically track constant reference signals

Figure 5.17: Tracking steps while rejecting bias in measurements

r(t) = r0 1(t) while rejecting measurement error modelled as a constant bias n(t) = n0 1(t). Measurement noise
was not considered in our derivations of disturbance rejecting controllers so let’s apply first principles to see if
we can reject the error. The TF from N (s) = L {n(t)} to Y (s) is
Y (s) −C(s)P (s)
= .
N (s) 1 + C(s)P (s)
If N (s) = n0 /s and C(s) provides feedback stability, then by the FVT
−C(0)P (0)
lim y(t) = n0 .
t→∞ 1 + C(0)P (0)
This means that the steady-state error due to measurement error is zero only if C(s)P (s) has a zero at the
origin. However, if C(0)P (0) = 0, then y can never track a reference of the form r0 1(t) because the controller
can’t provide an internal model of R(s) without causing an unstable pole-zero cancellation. This shows that
the feedback system is not capable of rejecting measurement error of the form n0 1(t) if it is required to track
signals of the same form. This makes a lot of intuitive sense. If we wish to track a step and there is a step
measurement error, then the controller can never know the true value of the output and cannot take appropriate
actions. N

5.5 Summary
This chapter contains some of the most important ideas in this course. The following concepts should be
understood and gleamed after reading this chapter.

Version 1.2.2816, September 1, 2020

 5.5. Summary 136

1. We have to be careful when we start to interconnect LTI systems because stability of the interconnected
system does not follow from stability of the individual components.

2. We gave two definitions for what it means for the system in Figure 5.6 to be stable. Internal stability
(Definition 5.2.3) dealt with the closed-loop state model whereas input-output stability (Definition 5.2.4)
dealt with the various transfer functions in the loop.

3. You must know how to test for both internal stability and input-output stability.

4. You should understand why internal stability implies input-output stability.

5. You should know what the characteristic polynomial is for a feedback system and how to find it for unity
feedback systems and non-unity feedback systems.

6. Section 5.3 covered the Routh-Hurwitz criterion. You must be able to use these tests to check for stability.

7. You should understand the internal model principle (Theorem 5.4.2) and be able to determine the steady-
state tracking error and disturbance response for a stable feedback system using the final-value theorem.

Version 1.2.2816, September 1, 2020

Chapter 6

Root-locus method

The poles of a system tell us a lot about the behaviour of the system. They determine stability (directly)
and transient performance (in a complicated way). In this chapter we study the roots of polynomials when
the coefficients depend linearly on a parameter. In control systems, the parameter is typically the gain of a
feedback loop, and our interest is in choosing the gain so that the closed-loop system is internally stable.
The manner of studying such problems in this chapter is to understand how the roots move in the complex
plane as functions of the parameter. That is to say, we look at the locus of all roots of the polynomial as the
parameter varies, hence the name “root-locus.” The root-locus shows us, graphically, how the closed-loop poles
of a control system move around the s-plane as a single parameter is varied. The root-locus can be very easily
and quickly sketched by hand which makes it an efficient tool for analysis and, to some extent, design1 .

Contents
6.1 Basic root-locus construction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
6.2 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
6.3 Non-standard problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
6.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149

6.1 Basic root-locus construction

Consider a unity feedback system as shown in Figure 6.1. The characteristic polynomial of this system (see

Figure 6.1: Unity feedback system.

Definition 5.2.6) is π(s) = Np (s)Nc (s) + Dp (s)Dc (s). The closed-loop system is internally stable if and only if
all the roots of π are in C− , see Theorem 5.2.7.

1
The root-locus method was put forth by Walter Evans [Evans, 1948], [Evans, 1950].

137
 6.1. Basic root-locus construction 138

Example 6.1.1. To get a feeling for what a root-locus diagram looks we first consider a simple example in
which we can compute the closed-loop poles explicitly. Suppose that
1
C(s) = K, P (s) = .
s(s + 2)
2 +2s+K. Using the quadratic formula from high school the roots of the characteristic polynomial
Then π(s) = s√
are s = −1 ± 1 − K. From this equation we make the following observations.

• If K ≤ 0, the closed-loop system is unstable.

• If K = 0, the roots of π(s) equal the poles of P (s).

• If K ∈ (0, 1], the closed-loop system has two real poles in C− and is stable.

• If K ∈ (1, +∞), the closed-loop system has complex conjugate poles with real part −1 and hence it is
stable.

• As K → +∞ the imaginary part of the poles goes to ±∞.

From these observations we can sketch a plot that shows how the closed-loop poles vary as K goes from 0 to
+∞. The sketch is shown in Figure 6.2.

Figure 6.2: Locus of roots for s2 + 2s + K as K goes from 0 to +∞. The roots when K = 0 are denoted by the symbol ×.

This plot shows us that the system is stable for all K > 0 (infinite gain margin, see Chapter 8). The same
information could have been obtained using the Routh-Hurwitz criterion but this figure contains much more
information. We see that as K → +∞ the closed-loop system becomes less damped, the step response has more
overshoot and its frequency of oscillation increases. This suggests that for a good step response we shouldn’t
choose K too large for this system. N

We now show how to draw root-locus diagrams for higher order polynomials. Consider the system in
Figure 6.3 where we have pulled a gain K out of the controller.

Figure 6.3: System considered in basic root-locus.

Version 1.2.2816, September 1, 2020

 139 CHAPTER 6. ROOT-LOCUS METHOD

The characteristic polynomial of this system is π(s) = D(s)+KN (s) where N (s) := Np Nc and D(s) := Dp Dc .
The roots of the characteristic polynomial are of course

{s ∈ C : π(s) = 0} = {s ∈ C : D(s) + KN (s) = 0} . (6.1)

The root-locus is a drawing of how the roots of π(s) = D(s) + KN (s) vary as a function of K ∈ R. For
now we make the following assumption.
Assumption 6.1.1. Let D, N ∈ R[s] and let n := deg D(s) and m := deg N (s).
1. C(s)P (s) is proper, i.e., n ≥ m.

2. K ≥ 0.

3. N (s) and D(s) are monic, i.e., their leading coefficients are 1. J
In Section 6.3 we’ll see how to handle cases where items 1 and 2 in Assumption 6.1.1 don’t hold. As for item
3, that D should be taken as monic seems natural from (6.1). As for N , it can be made monic by redefining K
if necessary.
Remark 6.1.2. As with Bode and Nyquist plots, the root-locus of a polynomial can be generated using software.
The MATLAB graphical user interfaces started by the commands rltool or sisotool are very powerful. We
discuss how to do these plots by hand because, in most cases, it is more important to quickly determine the
trend of the root loci rather than the exact root-locus plot. 

Construction rules
Rule 1 The root-locus of D(s) + KN (s) is symmetric about the real axis.
This rule follows from the fact that π ∈ R[s] so that it’s complex roots always appears as conjugate pairs.

Rule 2 There are n branches of the root-locus.

This rule follows from the fact that, by Assumption 6.1.1, the polynomial D(s) + KN (s) has degree n.

Rule 3 The root-locus of D(s) + KN (s) is a continuous function of K.

This follows from the fact that the roots of a polynomial in R[s] are a continous function of its coefficients.

Rule 4 When K = 0 the roots of D(s) + KN (s) equal the roots of D(s).

Rule 5 As K → +∞ there are m branches of the root-locus that approach the roots of N (s).
This follows from re-writing D(s) + KN (s) = 0 in the form
1
D(s) + N (s) = 0.
K
As K → +∞, the s with finite magnitude for which this equation holds approach the roots of N .

Rule 6 The n − m branches of the root-locus that do not approach the zeros of N (s) asymptotically tend towards
infinity following asymptotes that originate from the point
P P
roots of D(s) − roots of N (s)
σ := (centroid of π). (6.2)
n−m
The angles of the asymptotes are φ1 , . . . , φn−m where
(2k − 1) π
φk := , k ∈ {1, . . . , n − m} . (6.3)
n−m
Figure 6.4 shows the asymptote patterns for different values of n − m.

Version 1.2.2816, September 1, 2020

 6.1. Basic root-locus construction 140

Figure 6.4: Asymptote patterns generated by Rule 6.

Rule 7 (“no–yes–no rule”) A point s0 on the real axis is part of the root-locus if, and only if, s0 is to the left of
an odd number of poles and zeros of P (s)C(s) = N (s)/D(s).
This rule follows from (6.1) and the angular constraint each point s0 on the root-locus must satisfy
 
1
]N (s0 ) − ](D(s0 )) = ] − =π (for K > 0). (6.4)
K

Rule 8 (angles of arrival and departure) As K varies from 0 to +∞ the angle that a branch of the root-locus
makes when leaving a root of D(s) (departure angle) and the angles of branches that arrive at roots of
N (s) (arrival angle) can be computed using (6.4).
To find these angles note that at each point on the root-locus we have

](s − z1 ) + · · · ](s − zm ) − ](s − p1 ) − · · · − ](s − pn ) = π (for K > 0)

where zi ∈ C are the roots of N (s) and pi ∈ C are the roots of D(s).
For example, let θpi denote the angle of departure from pi . Then θpi satisfies
m
X n
X
θpi = ](pi − zj ) − ](pi − pj ) − π.
j=1 j=1
j6=i

Similarly, the angle of arrival θzi at zi is

n
X m
X
θ zi = ](zi − pj ) − ](zi − zj ) + π.
j=1 j=1
j6=i

Version 1.2.2816, September 1, 2020

 141 CHAPTER 6. ROOT-LOCUS METHOD

Rule 9 Given a point s ∈ C on the root locus, the corresponding value of K is given by

D(s)
K=− .
N (s)

The value of K above is real if and only if s is on the root locus.

Rule 10 The values of K for which the root-locus crosses the imaginary axis can be computed using the Routh
array. See Examples 5.3.6 and 7.2.2.

Procedure for plotting a root-locus

Given a polynomial D(s) + KN (s) we can plot its root-locus for K ≥ 0 as follows.

1. Compute the roots {z1 , . . . , zm } of N (s) and place a at the location of each root in C.

2. Compute the roots {p1 , . . . , pn } of D(s) and place a × at the location of each root in C.

3. Use the “no–yes–no” rule (Rule 7) to fill in the portion of the real axis that are on the root-locus.

4. Compute the centroid σ using (6.2).

5. Draw the n − m asymptotes using (6.3).

6. Use Rule 8 to compute the departure and arrival angles. This is usually only needed for complex conjugate
pairs and, occasionally, repeated real roots.

7. If you are lucky, you can give a reasonable guess as to how the root-locus behaves.

The last step is in some sense the most crucial. It is possible that one can do the steps preceding it, and still
get the root-locus wrong. Some experience is involved in knowing how a “typical” root-locus diagram looks,
and then extrapolating this to a given example.

6.2 Examples

Example 6.2.1. Consider the system in Figure 6.3 with

 
1 1
P (s) = 2 , KC(s) = K 1 + (PI controller).
s + 2s + 5 0.25s

Draw the root-locus of the closed-loop system for K ≥ 0.

1. In this case N (s) = (s + 4) so m = 1 and z1 = −4.

2. We have D(s) = s(s2 + 2s + 5) so n = 3 and {p1 , p2 , p3 } = {0, −1 + j2, −1 − j2}.

3. By the no–yes–no rule only (−4, 0) ⊂ R is part of the root-locus.

4. The centroid is
0 + (−1 − j2) + (−1 + j2) − (−4)
σ= =1
3−1
5. The asymptotes are
π 3π π
φ1 = , φ2 = ∼− .
2 2 2

Version 1.2.2816, September 1, 2020

 6.2. Examples 142

6. The departure angle from p2 is given by

θp2 = ](−1 + j2 − (−4)) − ](−1 + j2 − 0) − ](−1 + j2 − (−1 − j2)) − π
= arctan2 (3, 2) − arctan2 (−1, 2) − arctan2 (0, 4) − π
= 33.7◦ − 116◦ − 90◦ − 180◦
= −352.9◦ ∼ 7.125◦ .
By Rule 1 this calculation also yields the departure angle of p3 . By Rules 4 and 5 we can easily deduce
the departure / arrival angles of real poles / zeros.
7. For the sake of completeness, we’ll compute the imaginary axis crossings though at this point we can
roughly plot the root-locus with the information we’ve already computed. The Routh table for π(s) =
s(s2 + 2s + 5) + K(s + 4) = s3 + 2s2 + (K + 5)s + 4K is
s3 1 K +5
s2 2 4K
1
s 5−K 0
s0 4K.

When K 1 2
√ = 5 we get a row of zeros in s . The auxiliary polynomial is πa (s) = 2s + 20 whose roots are
s = ±j 10. Thus the root-locus plot cross jR at these points when K = 5.

Figure 6.5: Root-locus for Example 6.2.1.

The resulting root-locus plot is shown in Figure 6.5. N

The previous example was done in great detail but most of the time a root-locus can be very quickly sketched.
This makes root-loci great for quick system analysis.
Example 6.2.2. Draw the root-locus for π(s) = D(s) + KN (s) where D(s) = s3 (s + 4), N (s) = (s + 1). Here
n = 4, m = 1. The centroid and asymptotes are
−4 − (−1) π π
σ= = −1, φ1 = , φ2 = π, φ3 = − .
4−1 3 3

Version 1.2.2816, September 1, 2020

143 CHAPTER 6. ROOT-LOCUS METHOD

With this information along with Rule 1 and the “no–yes–no” rule we can obtain the root-locus as shown in
Figure 6.6. N

Figure 6.6: Root-locus for Example 6.2.2.

Exercise 6.1. Re-create the root-locus in Figure 7.3 using the procedure outlined in this chapter.
Root-locus diagrams are useful for deciding on the structure of the controller as the next four examples
demonstrate.
Example 6.2.3. (Inverted pendulum on a cart and proportional control) Consider the inverted pen-
dulum on a cart from Section 5.1 where the angle of the pendulum is taken as the output
1
P (s) = .
M1 Ls2 − g(M1 + M2 )
Can a proportional controller stabilize this system in unity feedback? With C(s) = Kp the characteristic
polynomial of the system is π(s) = D(s) + KN (s) with
g Kp
D(s) = s2 − (M1 + M2 ), N (s) = 1, K= .
M1 L M1 L

Here n = 2, m = 0 and the root-locus is drawn in Figure 6.7. We conclude that a proportional controller
cannot render the closed-loop system internally stable. N

Example 6.2.4. (Inverted pendulum on a cart and proportional-derivative control) Continuing with
the inverted pendulum on a cart, determine if an (ideal) PD controller can render the feedback system stable.
With
C(s) = Kp (1 + Td s)
the characteristic polynomial of the system is π(s) = D(s) + KN (s) with
g 1 Kp Td
D(s) = s2 − (M1 + M2 ), N (s) = s + , K= .
M1 L Td M1 L
Our root-locus rules were derived for the case where there is only one undetermined parameter in π(s). In this
example we have two free parameters – the proportional gain Kp and the derivative time constant Td . We break
the root-locus into three cases determined by the location of the root of N (s).

Version 1.2.2816, September 1, 2020

 6.2. Examples 144

Figure 6.7: Root-locus for Example 6.2.3.

Case 1 (Td < 0) In this case the root of N (s) is in the right half-plane. This root is a zero of the closed-loop
system. Placing a zero in C+ means that, by Rules 5 and 7, we will not be able to stabilize the system.
Furthermore, right half-plane zeros limit the achievable performance of a closed-loop system (Section 4.4
and Section 9.6.2) so we avoid them if at all possible.
Case 2 (Td > (M1 L)/(g(M1 + M2 ))) In this case the root of N (s) is in C− between the left root of D(s) and
p

the imaginary axis. We have n = 2, m = 1. The root locus is shown in Figure 6.8a. In this case we see
that the PD controller can stabilize the closed-loop system but placing the zero between the poles limits
how fast we can make the response. The dominant pole can never move to the left of the zero no matter
how high we make the proportional gain Kp .

p p
(a) Td > (M1 L)/(g(M1 + M2 )). (b) 0 < Td < (M1 L)/(g(M1 + M2 )).

Figure 6.8: Root-loci for the PD controllers of Example 6.2.4.

p
Case 3 (0 < Td < (M1 L)/(g(M1 + M2 ))) In this case the root of N (s) is to the left of both roots of D(s). We
still have n = 2, m = 1. The root-locus is shown in Figure 6.8b. In this case we again see that the PD
controller can stabilize the closed-loop system. The root-locus shows that we can get both poles far to
left which means a high bandwidth and fast response. On the other hand, the control effort will be high.
N

Version 1.2.2816, September 1, 2020

 145 CHAPTER 6. ROOT-LOCUS METHOD

Exercise 6.2. Draw the root-locus for Case 1 of the previous example.

Example 6.2.5. (Inverted pendulum on a cart and proportional-integral control) Once again the
plant is the inverted pendulum on a cart. Determine if a PI controller can render the feedback system stable.
With  
1
C(s) = Kp 1 +
Ti s
the characteristic polynomial of the system is π(s) = D(s) + KN (s) with
 
2 g 1 Kp
D(s) = s s − (M1 + M2 ) , N (s) = s + , K= .
M1 L Ti M1 L

In this case n = 3 and m = 1. The choice of integral time constant Ti determines the location of the root of
N (s). We consider two cases.

+ M2 ))) In this case the zero (root of N (s)) is in C− between the pole at s = 0
p
Case 1 (Ti > (M1 L)/(g(M1p
and the pole at s = − g(M1 + M2 )/(M1 L). Since n − m = 2 there are two asymptotes. The centroid is
given by p p
− g(M1 + M2 )/(M1 L) + g(M1 + M2 )/(M1 L) + 0 − 1/Ti 1
σ= =
3−1 2Ti
which means it is to the right of imaginary axis and to left of the plant’s unstable pole. The root-locus is
shown in Figure 6.9a from which we conclude that this configuration for the PI controller cannot stabilize
the pendulum on a cart system.
p
Case 2 (0 < Ti < (M1 L)/(g(M1 + M2 ))) The rootp of N (s) is now to the left of both plant poles. The analysis
is the same as in Case 1 but now if Ti < 0.5 (M1 L)/(g(M1 + M2 )), then the centroid is to the right of
the plant’s positive pole. The root-locus is shown in Figure 6.9b

p p
(a) Ti > (M1 L)/(g(M1 + M2 )). (b) 0 < Ti < (M1 L)/(g(M1 + M2 )).

Figure 6.9: Root-loci for the PI controllers of Example 6.2.5.

We conclude that a PI controller will not stabilize this system. N

Exercise 6.3. Draw the root-locus for the previous example when Ti < 0.
Exercise 6.4. Re-create the root-locus in Figure 7.6 using the procedure outlined in this chapter.

Version 1.2.2816, September 1, 2020

 6.3. Non-standard problems 146

Example 6.2.6. (Inverted pendulum on a cart and proportional-integral-derivative control) For

our last example consider the pendulum on a cart controlled by an (ideal) PID controller
 
1
C(s) = Kp 1 + + Td s .
Ti s

the characteristic polynomial of the system is π(s) = D(s) + KN (s) with

 
2 g 1 1 Kp Td
D(s) = s s − (M1 + M2 ) , N (s) = s2 + s + , K= .
M1 L Td Td Ti Ti M1 L

The zero locations, i.e., roots of N (s), can be placed anywhere in the complex plane by appropriately choosing
the values of Ti , Td . By breaking things down into cases based on the zero locations, we can analyse the closed-
loop system in the same manner as the previous examples. It turns out that the PID controller can stabilize
the feedback system and it provides perfect step tracking to boot! This is hardly surprising given the discussion
in Section 7.3. The diagram of the root-locus for this system is left as an exercise. N

Exercise 6.5. Draw the root-locus for K ≥ 0 for Example 6.2.6 when the roots of N (s) are complex conjugates and in C− .
Exercise 6.6. Draw the root-locus for K ≥ 0 for Example 6.2.6 when the roots of N (s) are real, distinct and to the left of both
plant poles.

6.3 Non-standard problems

In this section we learn how to handle cases where Assumption 6.1.1 fails to hold as well as transfer functions
in the feedback path.

Non-unity feedback systems

Suppose that instead of system 6.3 we have the system shown in Figure 6.10.

Figure 6.10: A non-unity feedback system.

In this case the loop gain is L(s) = KC(s)P (s)H(s) and the closed-loop poles satisfy 1 + L(s) = 0. The
characteristic polynomial is given by π(s) = Dc Dp Dh + KNc Np Nh =: D(s) + KN (s). This polynomial has
the same form as before so the root-locus construction proceeds in the exact same way. The unity-feedback
situation becomes a special case in which Nh = Dh = 1.

Controller is not a linear function of its gain

Figures 6.3 and 6.10 implicitly require that the controller be a linear function of K which allows us to pull it
out of the controller block. This is not needed to draw a root locus. The requirement for drawing a root-locus
is that the polynomial’s coefficients be linear in the unknown parameter. That is, we only require that π(s)
have the form D(s) + KN (s). If K can’t be pulled out of the controller, then π may still have the desired form
except that, in general, N (s) 6= Nc Np and D(s) 6= Dc Dp .

Version 1.2.2816, September 1, 2020

147 CHAPTER 6. ROOT-LOCUS METHOD

Example 6.3.1. Consider the system in Figure 6.1 with

1
P (s) = , C(s) = 10 (1 + Td s) (ideal PD controller).
s(s + 2)

In this case C(s) cannot be written in the form KC1 (s). The characteristic polynomial of the system is

π(s) = s(s + 2) + 10 (1 + Td s) = (s2 + 2s + 10) + (10Td )s =: D(s) + KN (s).

Notice that the characteristic polynomial’s coefficients are linear in the unknown parameter K but that D(s) =
s2 + 2s + 10 6= Dp Dc = s(s + 2) and N (s) = s 6= Nc Np . The root-locus for this system is shown in Figure 6.11.
N

Figure 6.11: Root-locus for Example 6.3.1.

Improper loop gain

Consider the polynomial π(s) = D(s) + KN (s) and suppose that deg (D) < deg (N ). This situation can arise
when part 1 of Assumption 6.1.1 fails to hold as well as other situations. Observe that
1
D(s) + KN (s) = 0 ⇐⇒ N (s) + D(s) = 0.
K

If we define D̂(s) := N (s), N̂ (s) := D(s) and K̂ = 1/K then the polynomial π̂(s) = D̂(s) + K̂ N̂ (s) satisfies
Assumption 6.1.1 and its root-locus can be drawn for K̂ ≥ 0. In order to recover the root-locus of π from the
root-locus of π̂ we do the following.

1. Switch each to × in the root-locus of π̂.

2. Switch each × to in the root-locus of π̂.

3. Reverse the directional arrows.

Example 6.3.2. Consider the usual unity-feedback system in Figure 6.1 with
1 s+3
P (s) = , C(s) = .
s(s + 1) Ts + 1

Version 1.2.2816, September 1, 2020

 6.3. Non-standard problems 148

Then
π(s) = (s2 + 2s + 3) + T s2 (s + 1) =: D(s) + KN (s).
In this case deg (D) = 2 < deg (N ) = 3 so we use the approach described above. Define π̂ = D̂(s) + K̂ N̂ (s)
where
1
D̂(s) := s2 (s + 1), N̂ (s) := s2 + 2s + 3, K̂ = .
T
The root-locus of π̂ is shown in Figure 6.12a while the root-locus of π is shown in Figure 6.12b. N

(a) π̂(s). (b) π(s).

Figure 6.12: Root-loci for Example 6.3.2.

Negative gain
Assumption 6.1.1 asks that K ≥ 0 but it is possible to obtain a root-locus for K < 0. Such a root-locus shows
how the roots of π(s) = D(s) + KN (s) vary as K goes from 0 to −∞. The root-locus for K < 0 is called the
complementary root locus.
The analysis for K < 0 is very similar to the K > 0 case. The main difference is that, since ]− K1 = 0, the
various angle conditions are slightly modified. In particular, for K < 0, the rules presented in Section 6.1 are
modified as follows.

Rule 1 No change.

Rule 2 No change.

Rule 3 No change.

Rule 4 No change.

Rule 5 As K → −∞ there are m branches of the root that approach the roots of N (s).

Rule 6 The asymptote angle equation changes to

2kπ
φk := , k ∈ {1, . . . , n − m} . (6.5)
n−m

Rule 7 (“no–yes–no rule”) A point s0 on the real axis, is part of the root-locus if, and only if, s0 is to the left of
an even number of poles and zeros of P (s)C(s) = N (s)/D(s).

Version 1.2.2816, September 1, 2020

 149 CHAPTER 6. ROOT-LOCUS METHOD

Rule 8 (angles of arrival and departure) The angles of arrival and departure are calculated and interpreted in the
same way except that the angle criterion is

](s − z1 ) + · · · ](s − zm ) − ](s − p1 ) − · · · − ](s − pn ) = 0 (for K < 0)

where zi ∈ C are the roots of N (s) and pi ∈ C are the roots of D(s).

Rule 9 No change.

Rule 10 No change.

6.4 Summary
In this chapter we gave a method for plotting the root-locus of a polynomial by hand. These sketches are
useful for quickly understanding a system and how changing a gain affects performance. You should know the
following.

1. What exactly are we seeing when we look at a root-locus diagram?

2. How to draw a root-locus by hand.

3. Given a point on the root-locus, you should be able to compute the gain K for which the polynomial has
a root at the given point.

4. You should be able to do all the examples in Section 6.2 in full detail.

5. You should be able to draw the root-locus for any of the non-standard scenarios discussed in Section 6.3.

Version 1.2.2816, September 1, 2020

Chapter 7

PID control

In this chapter we briefly introduce Proportional-Integral-Derivative (P.I.D.) controllers. This particular control
structure has become almost universally used in industrial control1 . These controllers are popular because they
are simple and have proven to be robust in the control of many important applications. The simplicity of these
controllers is also their weakness since it limits the range of plants that they can control satisfactorily. The
references [Åström and Hägglund, 1995], [Goodwin et al., 2001], [Åström and Hägglund, 2006] contain much
more information for the interested student.

Contents
7.1 Classical PID controller . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
7.2 Empirical tuning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
7.3 Pole placement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
7.4 Integrator windup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
7.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164

7.1 Classical PID controller

The block diagrams of two closed-loop systems with ideal PID controllers are shown in Figure 7.1. The transfer

(a) Error feedback. (b) Two degrees-of-freedom.

Figure 7.1: Block diagram of closed-loop system with ideal PID controllers.
1
“Based on a survey of over eleven thousand controllers in the refining, chemicals and pulp and paper industries, 97% of regulatory
controllers utilize PID feedback” [Desborough and Miller, 2002], [Åström and Murray, 2019], [Samad, 2017].

150
151 CHAPTER 7. PID CONTROL

function of the controller in Figure 7.1a is

U (s) Ki
= C(s) = Kp + + Kd s. (7.1)
E(s) s
The corresponding equation in the time-domain is
Zt
de(t)
u(t) = Kp e(t) + Ki e(τ )dτ + Kd (7.2)
dt
0

where e(t) = r(t) − y(t) as usual. The constants Kp , Ki and Kd are called, respectively, the proportional,
integral and derivative gains.
In Figure 7.1b the integral action acts on the error, but proportional and derivative action act on the process
output y. The input output model is
 
Ki Ki 1
U (s) = E(s) − (Kp + Kd s) Y (s) = R(s) − Kp + Ki + Kd s Y (s). (7.3)
s s s
This setup is often preferred when the reference signal r is discontinuous, e.g., a sequence of step changes.
Differentiating a step function creates an impulse which we usually do not want in the control signal u.
Exercise 7.1. Let P (s) be real rational and write P = Np /Dp where (Np , Dp ) is coprime. Find the TF from R to Y for both
the systems in Figure 7.1. Show that the system in Figure 7.1a is input-output stable if and only if the system in Figure 7.1b is
input-output stable.
The PID controller (7.1) is sometimes expressed in the equivalent form
 
1
C(s) = Kp 1 + + Td s (7.4)
Ti s
where Ti is called the integral time constant or reset time while Td is the derivative time constant.
Exercise 7.2. Show that the PID controllers (7.1) and (7.4) are identical.
The PID controllers (7.1) and (7.3) are idealized because the resulting TFs are improper which makes them
impossible to implement without approximations2 . Furthermore, an ideal differentiator has the undesirable
property that it amplifies high frequency noise (see Figure 3.15). Therefore, in practice the derivative term is
“rolled off” and replaced by a low pass filtered version
s
Kd τd > 0. (7.5)
τd s + 1
The TF (7.5) approximates a derivative for low frequencies (ω < 1/τd ) and approaches the ideal differentiator
as τd → 0. Figure 7.2 shows the Bode plot for an ideal differentiator and the “rolled off” version. With the
approximated differentiator the PID controller (7.4) (equivalently (7.1)) becomes
 
1 Td s
C(s) = Kp 1 + + (7.6)
Ti s τd s + 1
while the two degree-of-freedom version (7.3) becomes
 
Ki 1 s
U (s) = R(s) − Kp + Ki + Kd Y (s). (7.7)
s s τd s + 1
Remark 7.1.1. In classical PID control design, τd 6= 0 was largely seen as a necessary evil, i.e., as a necessary
departure from a pure proportional, integral and derivative action. As a result almost all industrial PID
controllers set τd as a fixed fraction of Td , rather than viewing it as an independent design parameter in its own
right. A more modern approach to PID design is to treat the time constant of the derivative as an important
degree of freedom available to the designer. 
2
An exception is when the derivative term can be measured directly using sensors, e.g., a tachometer on a motor.

Version 1.2.2816, September 1, 2020

 7.1. Classical PID controller 152

Magnitude (dB)

90

60
Phase (deg)

30

Figure 7.2: Bode plot of an ideal differentiator (blue) and an implementable low pass filtered version (red).

Remark 7.1.2. There are various different ways of parametrizing a PID controller. The form (7.1) is useful
for analysis because it is linear in the controller gains and turning off the integral action corresponds to setting
Ki = 0. The form in (7.6) is called the standard form. In this course we usually stick to the standard
form (7.6) or its ideal version (7.4). However, when working with industrial PID controllers you should be
aware that different definitions of PID gains abound. 

Equation (7.1) is an idealization but it is a useful abstraction for understanding PID controllers. We now use
this idealized controller to develop some intuition about the closed-loop system in Figure 7.1a. In our discussion
we assume that the plant TF is given by
1
P (s) = . (7.8)
(s + 1)3
Of course, the behaviour of the closed-loop system in Figure 7.1a depends very much on the plant dynamics.
By focusing on a specific plant our intention is to simply develop intuition on the role that each component of
a PID controller plays.

7.1.1 Proportional control

Consider a pure proportional controller
C(s) = Kp ,
i.e., select Ki = Kd = 0 in (7.1). A proportional controller provides a contribution which depends on the
instantaneous value of the tracking error. A proportional controller can control any stable plant, but it provides
limited performance and generally gives nonzero steady-state tracking error. For the system in Figure 7.1a with
plant (7.8) the closed-loop TF from r to y is

C(s)P (s) Kp
= 3 2
.
1 + C(s)P (s) s + 3s + 3s + 1 + Kp

The Routh array for this system is given by

s3 1 3
s2 3 1 + Kp
1 8−Kp
s 3
s0 1 + Kp .

Version 1.2.2816, September 1, 2020

153 CHAPTER 7. PID CONTROL

Therefore the closed-loop system is internally stable if and only if Kp ∈ (−1, 8). Assuming internal stability
and applying the FVT, the steady-state tracking error for a unit step reference signal is
1 1
ess = = .
1 + Kp P (0) 1 + Kp

Thus we see that this system will always have steady-state error though the error gets smaller as Kp → 8. On
the other hand, as Kp → 8 the system is getting closer to becoming unstable. Figure 7.3 is called a root-locus
diagram3 . It shows how the roots of s3 + 3s2 + 3s + 1 + Kp vary in the s-plane as we increase Kp . From
1

0.8

0.6

0.4

0.2

-0.2

-0.4

-0.6

-0.8

-1
-2.5 -2 -1.5 -1 -0.5 0

Figure 7.3: Root-locus plot from Section 7.1.1 (proportional control).

Figure 7.3 and based on our discussion in Chapter 4, we observe that as Kp gets larger the complex conjugate
poles approach the imaginary axis and dominate the time response. As Kp gets larger, the dominant poles have
smaller real part (larger settling time), the damping ratio gets smaller (more overshoot) and the imaginary part
gets larger (increased frequency of oscillations). We conclude that while we can reduce the steady-state tracking
error by increasing the proportional gain, the transient performance gets worse and the system eventually goes
unstable. These predictions are verified in Figure 7.4 where we simulate the closed-loop system for different
values of Kp .
1.4 5

1.2 4

1 3

0.8 2

0.6 1

0.4 0

0.2 -1

0 -2
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Time (seconds) Time (seconds)

(a) Output y(t). (b) Control signal u(t).

Figure 7.4: Step response of system in Figure 7.1a with plant (7.8) and proportional control with Kp ∈ {1, 2, 5}.
3
See Chapter 6.

Version 1.2.2816, September 1, 2020

 7.1. Classical PID controller 154

To eliminate steady-state tracking error for step inputs the proportional term can be changed to

U (s) = Kp E(s) + Kf f R(s) (7.9)

where the gain Kf f is a feedforward or bias term. It is sometimes called reset in the PID literature. This
control architecture is shown in Figure 7.5 which justifies the term feedforward.

Figure 7.5: Proportional controller with feedforward.

Exercise 7.3. Find the TF from r to y for the system in Figure 7.5.
If we choose Kf f = 1/P (0) then the steady-state tracking error for step inputs will be zero so long as there
are no disturbances.
Exercise 7.4. Regarding the system in Figure 7.5, show that (a) input-output stability of the system in does not depend on Kf f
and (b) if the system is input-output stable and Kf f = 1/P (0) then ess = 0 for step inputs.
The problem with the solution for eliminating steady-state tracking error in Figure 7.5 is that (i) it requires
perfect knowledge of P (0) (ii) it only works if there are no disturbances and no noise in the loop and (iii) the
value of Kf f often needs to be re-calibrated. As we saw in Section 5.4, a far better solution is to use the internal
model principle to achieve perfect step tracking. This helps motivate the integral action in a PID controller.

7.1.2 Integral control

Consider a pure integral controller
Ki
C(s) =
,
s
i.e., select Kp = Kd = 0 in (7.1). The main function of integral action is to make sure that the system output
agrees with the setpoint in steady-state. The output of the intergral term is proportional to the accumulated
error. If Ki > 0 a small positive tracking error will always lead to an increasing control signal while a small
negative tracking error will always lead to a decreasing control signal.
For the system in Figure 7.1a with plant (7.8) the closed-loop TF from r to y is

C(s)P (s) Ki
= 4 .
1 + C(s)P (s) s + 3s + 3s2 + s + Ki
3

The Routh array for this system is given by

s4 1 3 Ki
s3 3 1 0
2 8
s 3 Ki
s1 1 − 98 Ki
s0 Ki .

Therefore the closed-loop system is internally stable if and only if Ki ∈ (0, 8/9). Thus we see that integral
action can make the system go unstable if the integral gain is too high. Figure 7.6 shows how the roots

Version 1.2.2816, September 1, 2020

155 CHAPTER 7. PID CONTROL
2.5

1.5

0.5

-0.5

-1

-1.5

-2

-2.5
-3 -2.5 -2 -1.5 -1 -0.5 0 0.5 1 1.5

Figure 7.6: Root-locus plot from Section 7.1.2 (integral control).

of s4 + 3s3 + 3s2 + s + Ki vary in the s-plane as we increase Ki . Again we see that as Ki gets large, two
complex conjugate poles get closer to the imaginary axis and dominante the time-domain response. Increasing
Ki results in longer settling time, more overshoot and increased frequency of oscillation. Figure 7.7 shows
simulation results for different values of Ki which match our expectations.
1.6 1.8

1.6
1.4

1.4
1.2

1.2

0.8

0.8

0.6

0.6

0.4
0.4

0.2
0.2

0 0
0 10 20 30 40 50 60 70 0 10 20 30 40 50 60 70
Time (seconds) Time (seconds)

(a) Output y(t). (b) Control signal u(t).

Figure 7.7: Step response of system in Figure 7.1a with plant (7.8) and integral control with Ki ∈ {0.1, 0.3, 0.5}.

7.1.3 Derivative control

Consider an ideal derivative controller
C(s) = Kd s,
i.e., select Kp = Ki = 0 in (7.1). This controller acts on the rate of change of the tracking error. It is
sometimes referred to as a predictive mode because of its dependence on the error trend. To see this consider
the time-domain expression of a PD controller
 
de
u(t) = Kp e(t) + Td .
dt

The term inside the bracket can be interpreted as a prediction of the tracking error at time t + Td by linear
interpolation as illustrated in Figure 7.8. The prediction horizon Td = Kd /Kp is the derivative time constant
of the controller.

Version 1.2.2816, September 1, 2020

 7.1. Classical PID controller 156

Figure 7.8: PD controller viewed as predictive action. Increasing Td increases the prediction horizon.

For the system in Figure 7.1a with plant (7.8) and an ideal derivative controller the closed-loop TF from r
to y is
C(s)P (s) Kd s
= 3 2
.
1 + C(s)P (s) s + 3s + (3 + Kd )s + 1
The Routh array for this system is given by

s3 1 3 + Kd
s2 3 1
8
s1 3 + Kd
s0 1.

Therefore the closed-loop system is internally stable if and only if Kd > −8/3. The fact that high gain derivative
control does not destabilize the system gives some justification to the intuitive notion that derivative control
has a stabilizing effect. Another interpretation of this intuitive notion is that derivative control acts like friction
in a mechanical system or resistance in an electrical system and tends to dampen oscillations and dissipate
energy resulting in a stable system.
A main draw back of derivative control is that it cannot provide zero steady-state tracking error. Indeed,
the simulations in Figure 7.9a show that in the case of plant (7.8), the steady state output is zero4 . Derivative
control can also lead to large control signals in response to high frequency tracking error as seen in Figure 7.9b.
This latter concern can be alleviated using the architecture from Figure 7.1b. The simulations in Figure 7.9
were done using a low pass filtered version of the differentiator (7.5) with τd = 0.01. Note that for the filtered
differentiator the closed-loop system can become unstable if Kd is too large.
Exercise 7.5. Let τd = 0.01 and use a Routh table to find the conditions on Kd so that the system in Figure 7.1a with plant (7.8)
and controller (7.5) is internally stable.
1
1000

0.9

800
0.8

0.7

600

0.6

0.5
400

0.4

200
0.3

0.2

0.1

0 -200
0 10 20 30 40 50 60 70 80 0 0.01 0.02 0.03 0.04 0.05 0.06
Time (seconds) Time (seconds)

(a) Output y(t). (b) Control signal u(t).

Figure 7.9: Step response of system in Figure 7.1a with plant (7.8) and derivative control with τd = 0.01 and Kd ∈
{2, 5, 10}.

4
This is not surprising since the TF from r to y has a zero at s = 0.

Version 1.2.2816, September 1, 2020

157 CHAPTER 7. PID CONTROL

7.2 Empirical tuning

Two classical methods for determining the parameters of PID controllers were presented by Ziegler and Nichols
in 1942 [Ziegler and Nichols, 1942]. The Ziegler-Nichols methods had a huge impact when they were introduced
in the 1940s. The rules are simple to use and give good initial conditions for manual tuning. The Ziegler-Nichols
methods are still widely used and often form the basis for tuning procedures by controller manufacturers and
process industry [Åström and Hägglund, 1995]. In both of these methods the criterion for optimization was the
minimization of
Z∞
|e(t)|dt
0

where e(t) is the tracking error due to a step input. A more modern approach to gain selection is to use a model
based technique like the one we discuss in Section 7.3.

7.2.1 Step response method

This method requires that the plant be open-loop stable.

Assumption 7.2.1. The plant P (s) in Figure 5.6 is open-loop stable. J

The tuning process works as follows.

1. Obtain the step response of the plant which by Assumption 7.2.1 and the FVT converges to a constant
value. The step response can be obtained either by simulating the plant model or, if we don’t have a model
to simulate, by running experiments on the physical system. Plot the step response as in Figure 7.10.

2. From the plot, extract the parameters σ and τ shown in Figure 7.10. Physically σ is the maximum slope
of the step response. The parameter τ is the intersection between the tangent and the coordinate axis.

Time (seconds)

Figure 7.10: Obtaining the parameters σ and τ for Ziegler-Nichols tuning of PID controllers.

3. Use the gain formulae in Table 7.1 to obtain a PID controller of the form 7.4.

Version 1.2.2816, September 1, 2020

 7.2. Empirical tuning 158

Table 7.1: Ziegler-Nichols PID gains using step response method.

Controller type Kp Ti Td
1
Proportional στ — —
9
Proportional-Integral 10στ 3τ —
6 τ
Proportional-Integral- 5στ 2τ 2
Derivative

Example 7.2.1. Consider the BIBO stable plant

1
P (s) =
(s + 1)3
and its step response in Figure 7.11a. From that figure we obtain σ ≈ 0.2705, τ ≈ 0.806. Using the gain formula
in Table 7.1 we obtain the PID controller
 
1
C(s) = 5.504 1 + 0.6203 + 0.403s .
s
The closed-loop step response with this controller is shown in Figure 7.11b. N
1 1.6

0.9
1.4

0.8

1.2

0.7

1
0.6

0.5 0.8

0.4
0.6

0.3

0.4

0.2

0.2
0.1

0 0
0 2 4 6 8 10 12 14 0 2 4 6 8 10 12 14 16 18 20
Time (seconds) Time (seconds)

(a) Open-loop step response. (b) Closed-loop step response.

Figure 7.11: Ziegler-Nichols tuning for Example 7.2.1 using step response method.

7.2.2 Frequency response method

Consider the TF from r to y for the system in Figure 5.6 when the controller is C(s) = Kp , i.e., a proportional
controller
Y (s) Kp P (s)
= . (7.10)
R(s) 1 + Kp P (s)
The method requires the following assumption.
Assumption 7.2.2. The closed-loop transfer function with proportional control (7.10) is BIBO stable for very
small Kp . Furthermore, as Kp increases there is a critical value Ku at which two of the closed-loop poles lie on
the imaginary axis while the rest remain in C− . J
Under Assumption 7.2.2, when Kp = Ku two of the closed-loop poles lie on the imaginary axis at, say,
s = ±jωu . With C(s) = Ku the steady-state closed-loop step response will oscillate with period Tu = 2π/ωu .
The parameters Ku and Tu are called, respectively, the ultimate gain and ultimate period. Once Ku and
Tu have been obtained the Ziegler-Nichols PID gains can be computed using the entries in Table 7.2.

Version 1.2.2816, September 1, 2020

159 CHAPTER 7. PID CONTROL

Table 7.2: Ziegler-Nichols PID gains using frequency response method.

Controller type Kp Ti Td

Proportional 0.5Ku — —
Proportional-Integral 0.4Ku 0.8Tu —
Proportional-Integral- 0.6Ku 0.5Tu 0.125Tu
Derivative

Example 7.2.2. We use the same plant as in Example 7.2.1. The closed-loop TF with proportional control
was found in Section 7.1.1. Figure 7.3 shows that Assumption 7.2.2 holds for this system. To find Ku and Tu
we can use the Routh table from Section 7.1.1. From that table we see that when the proportional gain equals
8, we get a row of zeros in the s1 row. The auxiliary polynomial (see Section 5.3.2) is

πa (s) = 3s2 + 9.
√ √
The roots of πa are s = ±j 3. From this we conclude that Ku = 8, ωu = 3 and Tu = 3.6276. The resulting
PID controller using the Ziegler-Nichols gains from Table 7.2 is
 
1
C(s) = 4.8 1 + 0.5513 + 0.4543s .
s
N

Exercise 7.6. Simulate the closed-loop step response from Example 7.2.2.

7.3 Pole placement

The tuning methods of Section 7.2 are a good starting point for obtaining PID gains but they use too little
information about the model of the plant. A more modern approach to PID gain selection is to choose the gains
so that the closed-loop poles are located at desired places in the s-plane. This method of controller design is
called pole placement. Pole placement is an extremely important concept in systems control and is covered
in great detail in ECE488 and ECE481. In this section we do not cover pole placement in general; our scope is
much more limited. Our objective is to use the ideas of pole placement to choose PID gains.
Proposition 7.3.1. Any controller C(s) ∈ R(s) of the form

g2 s2 + g1 s + g0
C(s) = (7.11)
s2 + f1 s
is identical to the PID controller (7.6) with

g1 f1 − g0 g1 f1 − g0 g0 − g1 f1 + g2 f12 1
Kp = , Ti = , Td = , τd = . (7.12)
f12 g0 f1 f1 (g1 f1 − g0 ) f1

Exercise 7.7. Prove Proposition 7.3.1.

Assumption 7.3.2. The plant P (s) ∈ R(s) is strictly proper and second order
b1 s + b0
P (s) = , b0 6= 0. (7.13)
s2 + a1 s + a0
J

Version 1.2.2816, September 1, 2020

 7.3. Pole placement 160

Remark 7.3.3. We cannot allow b0 = 0 in the plant (7.13) because then the plant has a zero at s = 0. The
PID controller (7.11) has a pole at s = 0 due its integral action and if connected to a plant (7.13) with b0 = 0
there results an unstable pole-zero cancellation. See Definition 5.2.8 and Corollary 5.2.9. 
If we connect the plant (7.13) in unity feedback configuration, as in Figure 5.6, with controller (7.11) then
the characteristic polynomial (see Definition 5.2.6) is given by
π(s) = (b1 s + b0 ) g2 s2 + g1 s + g0 + s2 + a1 s + a0 s2 + f1 s

= s4 + (a1 + f1 + b1 g2 ) s3 + (a0 + a1 f1 + b1 g1 + b0 g2 ) s2 + (a0 f1 + b1 g0 + b0 g1 ) s + b0 g0 .

Suppose that we would like the closed-loop poles, i.e., the roots of π(s), to be located at5 {λ1 , λ2 , λ3 , λ4 } ⊂ C− .
These desired pole locations can be chosen based on, for instance, desired characteristics of the step response
and the ideas of Section 4.3. The desired pole locations generate a desired characteristic polynomial
πdes (s) = (s − λ1 ) (s − λ2 ) (s − λ3 ) (s − λ4 ) =: s4 + α3 s3 + α2 s2 + α1 s + α0 ∈ R[s].
Equating coefficients between π and πdes yields the following linear equation in the unknown controller gains
    
1 b1 0 0 f1 α3 − a1
  g2  =  α2 − a0  .
 a1 b0 b1 0     

 a0 0 b0 (7.14)
b1   g1   α1 
0 0 0 b0 g0 α0
A few observations and facts about equation (7.14) and its derivation.
• If Np and Dp are coprime, then equation 7.14 has a unique solution (see Proposition 7.3.4).
• In deriving (7.14) we have treated τd as an independent design parameter as opposed to the approach
taken in classical control, see Remark 7.1.1.
• The plant (7.13) includes every rational strictly proper system of order two. When designing PID con-
trollers for higher order plants the ideas on model reduction in Section 4.5 can be useful.
Proposition 7.3.4. Equation (7.14) has a unique solution if and only if the numerator and denominator
of (7.13) are coprime.
Proof. Let M denote the 4 × 4 matrix in (7.14). Equation (7.14) has a unique solution
if and only if the deter-
2 2
minant of M is non-zero. The determinant of M is det (M ) = b0 b0 − b0 b1 a1 + b1 a0 . Under Assumption 7.3.2
b0 6= 0 so det (M ) = 0 if and only if b20 − b0 b1 a1 + b21 a0 = 0. We now proceed to prove the proposition.
(=⇒) Assume that the numerator and denominator of (7.13) are not coprime. Then b1 6= 0 because otherwise
the plant has no zeros and therefore its numerator and denominator are coprime. When b1 6= 0 the plant has a
zero at s = −b0 /b1 . By assumption this is also a pole of the plant

b0 2
 
b0
− − a1 + a0 = 0.
b1 b1
Simplifying this expression we obtain b20 − b0 b1 a1 + b21 a0 = 0 which shows that det (M ) = 0. We’ve shown
Not coprime =⇒ Solution to (7.14) not unique
which is logically equivalent to showing6
Coprime ⇐= Solution to (7.14) unique.
(⇐=) Conversely, suppose that det (M ) = 0. Then b20 − b0 b1 a1 + b21 a0 = 0 which, reversing the argument we
just made, implies that the plant’s numerator and denominator have a common root.
5
The desired pole locations must be symmetric about the real axis to ensure that the controller gains are real.
6
Formally, we have proven the contrapositive of the statement we sought to prove.

Version 1.2.2816, September 1, 2020

161 CHAPTER 7. PID CONTROL

Example 7.3.1. Design a PID controller for the plant

2
P (s) =
(s + 1)(s + 2)

which yields a closed-loop dominated by the roots of s2 +4s+9. We solve this problem using our pole placement
equations with
πdes (s) = (s2 + 4s + 9)(s + 4)2 , b1 = 0, b0 = 2, a1 = 3, a0 = 2.
Note that we’ve added the extra term (s + 4)2 to πdes so that its order matches the order of π. Also note that
the terms we’ve added are further to the left in C− so that s2 + 4s + 9 dominates the response. Equation 7.14
becomes     
1 0 0 0 f1 12 − 3
 3 2 0 0   g2   57 − 2 
 2 0 2 0   g1  =  136  .
    

0 0 0 2 g0 144
The unique solution is (f1 , g2 , g1 , g0 ) = (9, 14, 59, 72). Using Proposition 7.3.1 these gains correspond to a PID
controller (7.6) with
Kp = 5.6667, Ti = 0.7083, Td = 0.1634, τd = 0.1111.
N

Exercise 7.8. Design a PID controller for the plant from Example 7.2.1 using pole placement. Compare the step response of your
controller to that in Figure 7.11b.
The ideas presented so far can also be applied to plants of the form
K
P (s) = e−sT .
τs + 1
Many plants, especially in process industries, can be satisfactorily described by this model. This plant is not
rational because of the time-delay term e−sT . One way to obtain an approximate rational second order model
is to approximate the time delay by a first order all-pass system. In this case we have

K K − sT 2 +1
P (s) = e−sT ≈ sT
.
τs + 1 τs + 1 2 + 1

Our previously derived design equations (7.14) can now be used to design PID controllers for the approximated
plant.
If the plant has a first order model
K
P (s) =
τs + 1
and we are using the PID controller (7.11) then the closed-loop characteristic polynomial is

π(s) = τ s3 + (τ f1 + Kg2 + 1) s2 + (f1 + g1 ) s + g0 .

If our desired characteristic polynomial is

πdes (s) = s3 + α2 s2 + α1 s + α0 ∈ R[s]

then the design equations become

 
 f
τ K 0 0  1
  
τ α 2 − 1
 1 0 1 0   g2

 =  τ α1  . (7.15)
 g1 
0 0 0 1 τ α0
g0

Version 1.2.2816, September 1, 2020

 7.4. Integrator windup 162

Equation (7.15) does not have a unique solution. Letting α ∈ R be a free parameter, the solution is
     
f1 0 1
 g2   1 (τ α2 − 1)   τ 
 = K  + α  −K  .
 g1   τ α1   −1 
g0 τ α0 0
The resulting PID controller is
1 τ
(τ α2 − 1) s2 + τ α1 s + τ α0 − αs


K Ks +1
C(s) =
s2 + αs
where again α ∈ R is a free parameter.

7.3.1 When can PID control be used?

The discussion on pole placement shows that PID control is sufficient for processes where the dominant dy-
namics are second order. For such processes there are no benefits gained by using a more complex con-
troller. If the plant’s dynamics cannot be satisfactorily modelled by a strictly proper second order model,
then there are benefits to using a higher order controller that is not a PID. We refer the interested student
to [Åström and Hägglund, 1995, Example 3.3] for an example demonstrating this fact.

7.4 Integrator windup

Various modifications and improvements have been made to the classical PID control algorithm presented in
this chapter. Most of these modifictions are outside the scope of this introductory chapter. In this section we
discuss one important modification that is very useful in practice.
All actuators have limitations: a motor has a maximum speed, an operational amplifier has a maximum
voltage, a valve cannot be more than fully open or fully shut. For a control system with a wide range of
operating conditions it may happen that the control variable reaches the actuator limits. When this happens
the controller is no longer working in a linear regime: it can’t react appropriately because it cannot change
the control signal in one direction. If a controller with integral action is used, the error will continue to be
integrated. This means that the integral term will become very large or, colloquially, it “winds up.” It is then
required that the error has the opposite sign for a long period of time before things return to normal. The
result is that a controller with integral action may give large transients when the actuator saturates.
Example 7.4.1. Consider the system in Figure 7.12. Suppose that the plant and controller are given by

Figure 7.12: Unity feedback system with actuator limits.

1 Ki
P (s) = , C(s) = Kp + .
s s
The saturator block labelled satM in Figure 7.12 models the actuator limitations and its equation is
(
x if |x| ≤ M ,
satM (x) = (7.16)
sgn (x)M if |x| > M .

Version 1.2.2816, September 1, 2020

163 CHAPTER 7. PID CONTROL

For this example let M = 0.1, let the gains in the PI controller be Kp = Ki = 0.2 and let r(t) = 1(t). The
simulation results are shown in Figures 7.13 and 7.14.
2

1.8

1.6

1.4

1.2

0.8

0.6

0.4

0.2

0
0 10 20 30 40 50 60 70 80 90 100
time (s)

Figure 7.13: Plant output when there is integral windup.

0.8

0.1

0.6

0.4
0.05

0.2

-0.05 -0.2

-0.4

-0.1

-0.6

-0.15 -0.8
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
time (s) time (s)

(a) Control signal. (b) Output of integrator.

Figure 7.14: Control signal and integrator signal from Example 7.4.1.

Let’s go through these plots in detail to understand what is happening.

• At t = 0 the reference signal equals 1. The large initial tracking error is enough to cause the proportional
part of the control signal to exceed the actuator limitation.

• As a result, the control signal saturates at M = 0.1 just after t = 0. This is seen in Figure 7.14a. The
integral action starts integrating the tracking error and starts to increase. This is shown in Figure 7.14b.

• At t ≈ 10 the tracking error finally changes sign as y(t) finally exceeds 1. See Figure 7.13. As a result of
this sign change, we see in Figure 7.14b that at t ≈ 10 the integrator starts to decrease.

• Even though the integrator starts to decrease at t ≈ 10, it has accumulated a relatively large value over the
first 10 seconds, i.e., it has wound up. The integral part of the control signal dominates the proportional
part of the control signal. Hence the overall control signal in Figure 7.14a does not react and remains at
its upper bound.

• At t ≈ 18 the tracking error becomes sufficiently large (see Figure 7.13) so that the proportional part
of the control signal starts to dominate the accumulated integral. The control signal starts to decrease
(Figure 7.14a) but the tracking error has become so large that it again saturates, this time at its lower
bound, around t ≈ 20.

Version 1.2.2816, September 1, 2020

 7.5. Summary 164

• The above process begins to repeat except now the integrator value is negative. The magnitude of the
tracking error at t ≈ 20 is less than it was just after t = 0 so the integrator does not windup as much as
it did initially.

The wild transients in Figure 7.13 are a tell-tale sign that windup is happening. N

The core idea used to protect systems against the negative effects of windup is to turn the integrator off
whenever the input reaches a limit. This can either be done by a switch or by implementing the integrator in
such a way that it automatically turns off when the input reaches a limit. An example of the latter type of
solution, called back-calculation, is shown implemented on a PI controller in Figure 7.15. The system has an

Figure 7.15: Anti-windup based on back-calculation for a PI controller.

extra feedback path that changes the input to the integrator. The error signal es is the difference between the
actuator output u(t) and the controller output v(t). This signal is fed through a gain of 1/Tt into the integrator.
When there is no saturation es = 0 and the system acts like a standard PI controller. When the actuator
saturates, the signal es is fed back to the integrator in such a way that es goes toward zero. This implies that
the controller output is kept close to the saturation limit. The controller output will then change as soon as
the error changes sign and integral windup is avoided.
The rate at which the controller output is reset is governed by the feedback gain 1/Tt ; a small value of Tt
gives a short reset time. The parameter Tt is called the tracking time constant. It cannot be too small
because measurement noise can then cause an undesirable reset. The tracking √ time constant Tt should be larger
than Td and smaller than Ti in a PID controller. A rule of thumb is Tt = Ti Td [Åström and Hägglund, 1995].

Example 7.4.2. Returning √ to the

psystem from Example 7.4.1 we implement the back-calculation method shown
in Figure 7.15 using Tt = Ti = Kp /Ki = 1. The simulation results are shown in Figure 7.16 and 7.17. The
behaviour is very different from that in Example 7.4.1. In particular we see that the integrator output is kept
relatively small when actuator limits are hit. N

Exercise 7.9. Re-create the simulation results from Example 7.4.2.

7.5 Summary
The purpose of this chapter was to introduce you to the basic structure of one of the most common feedback
controllers used in industry. The key concepts to take away from this chapter are as follows.

• Know the ideal PID equation (7.1) (equivalently (7.4)).

Version 1.2.2816, September 1, 2020

165 CHAPTER 7. PID CONTROL
1.2

0.8

0.6

0.4

0.2

0
0 10 20 30 40 50 60 70 80 90 100
time (s)

Figure 7.16: Plant output with anti-windup.

0.02

0.015
0.1

0.01
0.05

0 0.005

-0.05 0

-0.1 -0.005

-0.15 -0.01
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
time (s) time (s)

(a) Control signal. (b) Output of integrator.

Figure 7.17: Control signal and integrator signal from Example 7.4.2.

• Understand the issues associated with implementing derivative action and the reasons that the low-pass
approximation (7.5) is used in practice. You should also know the frequency ranges in which (7.5) gives
a good approximation to a derivative.

• Understand why the two degree of freedom implementation scheme in Figure 7.1b is often better in
practice.

• Some intuition about what the “P”, “I” and “D” terms do to a step response. Understand the reason
why “D” is sometimes called a predictive mode.

• Be aware of the classical tuning methods of Ziegler-Nichols. You don’t need to know the specific formulae,
you just need to understand the basic concept of using experimental results to choose controller gains.

• Section 7.3 is very important. You should know how to design PID controllers using pole placement and
use this section to understand the power and the limitations of PID control.

• Know what integrator windup is and the core idea behind anti-windup solutions. Namely, to turn the
integrator off whenever the input reaches a limit.

Version 1.2.2816, September 1, 2020

Chapter 8

Frequency domain methods for stability analysis

In Chapter 5 we developed tests that determine when a feedback system is stable. In this chapter we introduce
a new test for stability of feedback systems called the Nyquist criterion [Nyquist, 1932]. The Nyquist criterion
tells us if a feedback system is stable but more importantly, it provides guidelines for modifying an unstable
system to make it stable. Nyquist’s results were an inspiration for many of the design methods that today are
called loop shaping. Nyquist’s results also allow us to quantify how robust the feedback loop is to modelling
uncertainty.

Contents
8.1 Cauchy’s principle of the argument . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
8.2 Nyquist stability criterion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
8.3 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
8.4 Stability margins . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
8.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186

8.1 Cauchy’s principle of the argument

The Nyquist criterion is a test for feedback stability. The criterion is based on the principle of the argument1
from complex analysis [Wunsch, 2005]. The principle of the argument involves two things: a curve in the
complex plane and a complex-valued function of a complex variable.
Consider a closed path (also called a closed curve or a contour) in the s-plane with no self-intersections and
with negative, i.e., clockwise (C.W.) orientation as shown in Figure 8.1. We name the path Γs .

Figure 8.1: Closed path in s-plane with no self-intersections and negative orientation.
1
Nyquist’s original analysis was fairly complicated. Only later was it pointed out that the result can in fact be easily obtained
from the principle of the argument from complex analysis [MacColl, 1945]. A more accessible version of Nyquist’s papers are found
in [Bellman and Kalaba, 1964], which also includes other interesting early papers on control. Nyquist’s paper is also reprinted in
an IEEE collection of seminal papers on control [Basar and Basar, 2001].

166
167 CHAPTER 8. FREQUENCY DOMAIN METHODS FOR STABILITY ANALYSIS

Now let G ∈ R(s) be a rational function of the complex variable s. For every point s in the complex plane,
G(s) is also a point in the complex plane. In other words, we can view the rational function G as a map C → C.
To make this clear and avoid clutter we draw two copies of the complex plane, s in one copy called the s-plane
and G in the other copy called the G-plane.
If the curve Γs does not pass through a pole of G(s), then as s moves around Γs from any starting point,
the point G(s) traces out a different closed curve ΓG , the image of Γs under the map G.
Example 8.1.1. For G(s) = s − 1 we could have :

Figure 8.2: Mapping the curve Γs to the curve ΓG when G(s) = s − 1. In this example each point s ∈ Γs is shifted to the
left by 1 unit.

Notice that ΓG is just Γs shifted to the left one unit. The angle of G(s) is ](s − 1). Since Γs encircles one
zero of G(s), the angle of G changes by −2π as s makes a circuit around Γs . Therefore ΓG encircles the origin
once CW. N

Example 8.1.2. We keep G(s) = s − 1 but change Γs :

Figure 8.3: Mapping the curve Γs to the curve ΓG when G(s) = s − 1. In this example, Γs does not encircle the zero of G.

Now Γs does not encircle any zeros of G. The net change in ](s − 1) is therefore zero after one trip around
Γs which is why ΓG has no encirclements of the origin. N

Example 8.1.3. Now consider

1
G(s) = .
s−1
The angle of G(s) equals the negative of the angle of s − 1
](G(s)) = ](1) − ](s − 1) = −](s − 1).

Version 1.2.2816, September 1, 2020

 8.1. Cauchy’s principle of the argument 168

Figure 8.4: Mapping the curve Γs to the curve ΓG when G(s) = (s − 1)−1 .

Since Γs encircles the point 1, i.e., a pole of G(s), the curve ΓG encircles the origin once in the counterclock-
wise (C.C.W.) direction. N
From these examples we observe that there is a relationship between the number of poles and zeros of G(s)
inside Γs and the number of times ΓG encircles the origin.
Theorem 8.1.1 (Principle of the argument). Suppose G(s) ∈ R(s) has no poles or zeros on Γs , but Γs encloses
n poles and m zeros of G(s). Then ΓG has exactly n − m counterclockwise encirclements of the origin.
Proof. Write G(s) in this way Q
(s − zi )
G(s) = K Qi
i (s − pi )
with K a real gain, {zi } the zeros and {pi } the poles. Then for every s on Γs
X X
]G(s) = ]K + ](s − zi ) − ](s − pi ).

If zi is enclosed by Γs , the net change after one trip around the curve Γs in the term ](s − zi ) is −2π;
otherwise the change is 0. If pi is enclosed by Γs , the net change after one trip around the curve Γs in the term
](s − pi ) is −2π; otherwise the change is 0. Hence the net change in ]G equals m(−2π) − n(−2π) which equals
(n − m)2π.

Remark 8.1.2. If we change the orientation of Γs so that it is positively oriented (CCW) then the above proof
yields m − n CCW encirclements. 

8.1.1 Nyquist contour

The special curve Γs we use is called the Nyquist contour shown in Figure 8.5. If we use a Nyquist contour
then the resulting curve ΓG is called the Nyquist plot of G(s). If G(s) has no poles or zeros on Γs , then the
Nyquist plot encircles the origin exactly n − m times CCW, where n equals the number of poles of G(s) in
Re (s) > 0 and m equals the number of zeros of G(s) in Re (s) > 0.
Theorem 8.1.3. Suppose G(s) has no poles or zeros on the Nyquist contour Γs . Let n and m equal the number
of, respectively, poles and zeros in C+ . Then ΓG has n − m CCW encirclements of the origin.
A few observations to wrap up this section.
• With respect to Theorem 8.1.3, G(s) has no poles on the Nyquist contour if, and only if, G(s) has no
poles on the imaginary axis and is proper.

Version 1.2.2816, September 1, 2020

169 CHAPTER 8. FREQUENCY DOMAIN METHODS FOR STABILITY ANALYSIS

Figure 8.5: Basic Nyquist contour.

• Similarly, G(s) has no zeros on the Nyquist contour if, and only if, G(s) has no zeros on the imaginary
axis and G(s) is not strictly proper.
• In our application, if G(s) does actually have poles on the imaginary axis, then we have to indent the
contour around them. You can either indent to the left or the right; we’ll always indent to the right as in
Figure 8.6.

Figure 8.6: Nyquist contour indenting around poles on jR.

• We indent around poles of G(s) on the imaginary axis because otherwise the function G(s) would be
undefined there; we don’t need to indent around the zeros of G(s).

8.2 Nyquist stability criterion

Now we apply the principle of the argument to the problem of feedback stability. The setup is shown in
Figure 8.7. Here K is a real gain and C(s), P (s) are rational transfer functions. The key observation is that if
the feedback system is to be input-output stable then the poles of the closed-loop transfer function
KC(s)P (s)
(8.1)
1 + KC(s)P (s)
must all lie in C− . The idea is that we work with the TF 1 + KC(s)P (s) to ascertain when the closed-loop
system is stable.

Version 1.2.2816, September 1, 2020

 8.2. Nyquist stability criterion 170

Figure 8.7: Unity feedback system.

Definition 8.2.1. The loop gain of the system in Figure 8.7 is the transfer function L(s) = KC(s)P (s).

We could also have a TF H(s) in the feedback path in which case L(s) = KC(s)P (s)H(s) but we’ll take
H(s) = 1 for simplicity.

Assumption 8.2.2. For the system in Figure 8.7

1. P (s) and C(s) are proper and C(s)P (s) is strictly proper.

2. The product C(s)P (s) has no unstable pole-zero cancellations.

3. The gain K is non-zero. J

The second part of Assumption 8.2.2 is needed because the Nyquist criterion doesn’t test for it, and such can-
cellations would make the feedback system not stable (see Corollary 5.2.9). The third part of Assumption 8.2.2
is needed only because we’re going to divide by K at some point.

Theorem 8.2.3. Let n denote the number of poles of P (s)C(s) in C+ . Construct the Nyquist plot of P (s)C(s),
indenting to the right around poles on the imaginary axis. Then the feedback system is stable if and only if the
Nyquist plot doesn’t pass through − K1 + j0 and encircles it exactly n times CCW.

Proof. The closed-loop transfer function from reference input to plant output is (8.1). Define G(s) := 1 + L(s).
Because we have assumed no unstable pole-zero cancellations, feedback stability is equivalent to the condition
that
Dc (s)Dp (s) + KNc (s)Np (s)
G(s) = 1 + L(s) = 1 + KC(s)P (s) =
Dc (s)Dp (s)
has no zeros with Re (s) ≥ 0 (Theorem 5.2.11). So we’re going to apply the principle of the argument to get a
test for G(s) to have no zeros in the closed right half complex plane.
Please note the logic so far: The closed-loop transfer function from R to Y is

KP (s)C(s)
.
G(s)

This should have no poles with Re (s) ≥ 0. So G(s) should have no zeros with Re (s) ≥ 0. So we need a test
for G(s) to have this property.
Note that G(s) and C(s)P (s) have the same poles with real part greater then or equal to zero, so G(s)
has precisely n there. Since Γs indents to the right around poles of G(s) on the imaginary axis and since G(s)
is proper, G(s) has no poles on Γs . Thus by Theorem 8.1.3, G(s) has no zeros in Re (s) ≥ 0 and hence the
feedback system is stable if, and only if, the Nyquist plot of G(s) doesn’t pass through 0 and encircles it exactly
n times CCW. Since C(s)P (s) = K1 G(s)− K1 , this latter condition is equivalent to: the Nyquist plot of P (s)C(s)
doesn’t pass through − K1 and encircles it exactly n times CCW.

Version 1.2.2816, September 1, 2020

171 CHAPTER 8. FREQUENCY DOMAIN METHODS FOR STABILITY ANALYSIS

Remark 8.2.4. If there had been a non-unity transfer function H(s) in the feedback path, we would have
taken G(s) = 1 + KP (s)C(s)H(s) in the proof. 

Remark 8.2.5. By Assumption 8.2.2 C(s)P (s) is strictly proper. As a result, see Definition 2.8.1, the infinite
part of the Nyquist counter Γs (labelled B in Figure 8.8) gets mapped to the origin. 

Figure 8.8: Relevant parts of Nyquist contour when plotting a Nyquist plot under Assumption 8.2.2.

Remark 8.2.6. Furthermore, see Remark 3.8.1, since L(s) ∈ R(s) we have that |L(jω)| = |L(−jω)| and
]L(jω) = −]L(−jω). Therefore the part of the Nyquist plot corresponding to the negative imaginary axis in
Γs (labelled C in Figure 8.8) is identical to the part corresponding to positive imaginary axis, except mirrored
along the real axis. This means we can get a complete Nyquist plot by plotting C(s)P (s) for s on that portion
of Γs in red in Figure 8.8. 

Remark 8.2.7. Together, Remarks 8.2.5, 8.2.6 show that under Assumption 8.2.2, the Nyquist plot is essentially
(aside from indentations caused by poles on the imaginary axis) a polar plot of the loop gain with s = jω. As
we saw in Section 3.8, the polar plot and the Bode plot of a TF are closely related. Thus we see the important
relationship between the Nyquist plot and the Bode plot of the loop gain. 

Procedure to apply the Nyquist criterion

1. Pick the Nyquist counter Γs as in Figure 8.6 indenting to the right around any poles of C(s)P (s) on jR.

2. Draw the image of Γs under the map C(s)P (s). If Assumption 8.2.2 holds then we only need to draw the
part corresponding the red portion of Γs in Figure 8.8, (see Remarks 8.2.5, 8.2.6).

3. Observe N , the number of counter clockwise encirclement of − K1 made by the Nyquist plot.

4. By the principle of the argument (Theorem 8.1.1)

N =n−m

where N is obtained from the figure in Step 3, n is the number of poles of C(s)P (s) in C+ (known) and
m is the number of closed-loop poles in C+ (unknown).

5. By Theorem 5.2.7 the feedback system is stable if and only if m = 0 which is equivalent, in light of
Theorem 8.2.3, to N = n.

Version 1.2.2816, September 1, 2020

 8.3. Examples 172

8.3 Examples
In this section you’ll learn how to draw Nyquist plots and how to apply the Nyquist criterion.

Example 8.3.1. Let

1
C(s)P (s) = .
(s + 1)2
In this case C(s)P (s) has no poles in C+ and no poles on the imaginary axis. Hence we use the basic Nyquist
counter shown in Figure 8.9. The left side of Figure 8.9 shows the resulting Nyquist plot.

Figure 8.9: Nyquist contour (left) and Nyquist plot (right) for Example 8.3.1.

We map Γs one segment at a time. For the segment A (solid red) we have s = jω and

1 1 − ω2 2ω
C(jω)P (jω) = 2
= 2 2 2
−j .
(1 − ω ) + j2ω (1 − ω ) + 4ω (1 − ω )2 + 4ω 2
2

To plot the image of segment A we make the following observations which allow us to plot by hand the red
portion of the Nyquist plot.

• On A, s = jω. When ω = 0, C(j0)P (j0) = 1 + j0. As ω starts to increase, the imaginary part of CP
becomes negative and never crosses zero again for any finite value of ω > 0.

• When ω = 1 the real part of CP equals zero, and hence switches from being positive to being negative.
In other words, the Nyquist plot of the segment A crosses the jR axis at C(j)P (j) = −j1/2. The real
part of CP never crosses zero again for any finite value of ω > 1.

• As ω → +∞ the real part of CP is negative and approaches zero. The imaginary part is also negative
while approaching zero.

Segment B of the Nyquist counter (dashed blue in left side of Figure 8.9) gets mapped to the origin since
CP is strictly proper, see Remark 8.2.5. The image of segment C (dashed orange in the Nyquist contour of
Figure 8.9) is a reflection of the image of section A. Figure 8.10 shows the Nyquist plot generated by MATLAB
for this system.
Now that we have the Nyquist plot, we are ready to apply the Nyquist stability criterion and determine for
what range of K the feedback system is stable. The transfer function CP has no poles inside Γs and therefore
n = 0. So the feedback system is stable if and only if the Nyquist plot encircles −1/K exactly 0 times CCW.
This means, does not encircle it. The number of CCW encirclements for different values of −1/K are shown in
Table 8.1
Thus the conditions for stability are −1/K < 0 or −1/K > 1; that is, K > 0 or −1 < K < 0; that is,
K > −1, K 6= 0. The condition K 6= 0 is ruled out by our initial assumption (which we made only because

Version 1.2.2816, September 1, 2020

173 CHAPTER 8. FREQUENCY DOMAIN METHODS FOR STABILITY ANALYSIS
0.8

0.6

0.4

0.2

Imaginary Axis
0

-0.2

-0.4

-0.6

-0.8
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
Real Axis

Figure 8.10: Nyquist plot generated by MATLAB for Example 8.3.1.

Table 8.1: Number of CCW for Example 8.3.1.

(−∞, 0) (0, 1) (1, +∞)

CCW encirclements 0 −1 0

we were going to divide by K). But now, at the end of the analysis, we can check directly that the feedback
system actually is stable for K = 0. So finally the condition for stability is K > −1. You can readily confirm
this by applying Routh-Hurwitz to the closed-loop characteristic polynomial, (s + 1)2 + K. N

Example 8.3.2. Now consider

s+1
C(s)P (s) =
s(s − 1)
for which
j−ω −2 1 − ω2
C(jω)P (jω) = = + j .
ω(1 − jω) 1 + ω2 ω(1 + ω 2 )
Since there is a pole at the origin we have to indent to the right around it. We use the Nyquist counter Γs
shown in Figure 8.11. The left side of Figure 8.11 shows the resulting Nyquist plot.

Figure 8.11: Nyquist contour (left) and Nyquist plot (right) for Example 8.3.2.

On segment A, s goes from jε, ε > 0 and small, to +j∞. When s = jε we have

j−ε −2 1 − ε2
C(jε)P (jε) = = 2
+j ≈ −2 + j∞.
ω(1 − jε) 1+ε ε(1 + ε2 )

Version 1.2.2816, September 1, 2020

 8.3. Examples 174

This point is shown on the Nyquist plot in Figure 8.11. As s progresses along the rest of segment A, we have
that the real part of C(jω)P (jω) remains negative while the imaginary part equals zero when s = j. Thus the
Nyquist plot cross the real axis at C(j)P (j) = −1 + j0. For ω > 1 both the real and imaginary parts remain
negative and approach zero as ω → +∞.
Since CP is strictly proper segment B gets mapped to the origin. The image of segment C is the same as
the image of segment A reflected about the real axis.
On segment D, s = εejθ where ε > 0 is small and θ increases from −π/2 to π/2. Therefore

εejθ + 1 1 1
C(εejθ )P (εejθ ) = ≈ − jθ = ej(π−θ) .
εejθ (εejθ − 1) εe ε

Thus the image of segment D is a semicircle of infinite radius whose argument goes from 3π/2 (equivalent to
−π/2) to π/2, i.e., CW. Figure 8.12 shows the Nyquist plot generated by MATLAB for this system.
10

2
Imaginary Axis

-2

-4

-6

-8

-10
-2 -1.5 -1 -0.5 0 0.5
Real Axis

Figure 8.12: Nyquist plot generated by MATLAB for Example 8.3.2.

Now to apply the Nyquist criterion. Since C(s)P (s) has one pole inside Γs , we have n = 1. Therefore we
need exactly 1 CCW encirclement of the point −1/K.

Table 8.2: Number of CCW for Example 8.3.2.

(−∞, −1) (−1, 0) (0, +∞)

CCW encirclements −1 1 0

Thus feedback stability holds if and only if −1 < 1/K < 0; equivalently, K > 1. N

Example 8.3.3. In this example take

1
C(s)P (s) =
(s + 1)(s2 + 1)

for which
1 1 − jω 1 ω
C(jω)P (jω) = 2
= 2 2
= 4
−j .
(jω + 1)(1 − ω ) (ω + 1)(1 − ω ) 1−ω 1 − ω4
You should be able to draw the Nyquist plot shown in Figure 8.13 based on the previous two examples. We
observe that on segment A of Γs , s = jω with ω ∈ [0, 1 − ε], ε > 0 and small. On segment B, s = j + εejθ with
ε > 0, small and θ ∈ [−π/2, π/2]. On segment C of Γs , s = jω with ω ∈ [1 + ε, +∞), ε > 0 and small. The
drawing yields the data in Table 8.3. Since n = 0, i.e., CP has no poles inside Γs , we conclude that the system
is feedback stable if and only if −1/K > 1, i.e., −1 < K < 0. N

Version 1.2.2816, September 1, 2020

175 CHAPTER 8. FREQUENCY DOMAIN METHODS FOR STABILITY ANALYSIS

Figure 8.13: Nyquist contour (left) and Nyquist plot (right) for Example 8.3.3.

Table 8.3: Number of CCW for Example 8.3.3.

(−∞, 0) (0, 1) (1, +∞)

CCW encirclements −2 −1 0

Remark 8.3.1. This section and Section 3.8.2 have dealt with graphical methods done by hand on paper.
Why did we bother? Why didn’t we just ask MATLAB to draw the Bode or Nyquist plots? The answer is
that by learning to draw the graphs, you acquire understanding. Additionally you should check that MATLAB
incorrectly draws the Nyquist plot in Example 8.3.3. MATLAB isn’t smart enough to indent the Γs -contour
around the poles on the imaginary axis. 

Example 8.3.4. (Capturing the essential features of the Nyquist plot from the Bode plot) Consider
the Bode plot in Figure 8.14 of some loop gain L(s). In Figure 8.14 the magnitude is in absolute units (not
dB), the phase is in degrees. The frequency is in rad/s.
We make the following observations:
• Since the magnitude plot is bounded from above, we conclude that the Nyquist contour doesn’t need
to indent around poles on the imaginary axis. Therefore we can use the basic Nyquist contour from
Figure 8.5 and the Nyquist plot is simply a polar plot of L(jω).

• Using the Bode plot, we see that at low frequencies (ω < 10−2 rad/s) the magnitude is |L(jω)| ≈ 5. The
phase at these low frequencies is ]L(jω) ≈ −π ∼ π and so L(j0) ≈ −5. Thus the Nyqusit plot starts at
−5 in the complex plane.

• Next we observe from the Bode plot that the phase of L(jω) is always between −π and −π/2. Therefore
the part of the Nyquist plot corresponding to moving up the imaginary axis is always in the third quadrant.

• The magnitude Bode plot crosses 1 at about ω = 6 rad/s and ]L(j6) ≈ −157◦ . Therefore the Nyquist
plot crosses the unit circle at an angle of −157◦ .

• The magnitude Bode plot increases a little at first and then goes to zero as ω → ∞. Therefore the Nyquist
plot moves aways from the origin before eventually going to zero.
From the above observations we can sketch the Nyquist plot in Figure 8.15. The Nyquist plot encircles −1
once in the CCW direction. Thus the feedback system is stable if, and only if, L has one pole in C+ .
N

Version 1.2.2816, September 1, 2020

 8.4. Stability margins 176

1
10

10 0

10 -1

10 -2

10 -3
-2 -1 0 1 2
10 10 10 10 10

-120

-130

-140

-150

-160

-170

-180
10 -2 10 -1 10 0 10 1 10 2
Frequency (rad/sec)

Figure 8.14: Bode plot.

Figure 8.15: Nyquist plot estimated from Bode plot.

8.4 Stability margins

If a feedback system is stable, how stable is it? In other words, how far is it from being unstable? This depends
entirely on our plant model, how we got it, and what uncertainty there is about the model. In the frequency
domain context, uncertainty is naturally measured in terms of magnitude and phase as functions of frequency.
In this section we look at this issue. The Nyquist plot is the most revealing, but we will also use the Bode plot.

8.4.1 Phase margin

Version 1.2.2816, September 1, 2020

177 CHAPTER 8. FREQUENCY DOMAIN METHODS FOR STABILITY ANALYSIS

Example 8.4.1. Consider the usual feedback loop as in Figure 8.7 and let its loop gain be given by
2
L(s) = .
(s + 1)2
The Nyquist plot of L is shown in Figure 8.16 where the image of the portion of the Nyquist contour corre-
sponding to negative frequencies is drawn with a red dashed line.
2

1.5

0.5

-0.5

-1

-1.5

-2
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2

Figure 8.16: Nyquist plot of Example 8.4.1 with phase margin indicated.

There are no encirclements of the critical point −1 + j0 so the feedback system is stable. A decrease in
the phase of L(s) will cause the blue portion of the Nyquist plot to rotate CW about the origin. If we rotate
it too far CW the Nyquist plot will eventually encircle −1 + j0 and we’ll lose stability. Hence, the maximum
amount of phase we can decrease by is exactly the angle indicated in Figure 8.16. This angle is called the phase
margin of the system and is related to the distance from the critical point −1 + j0 to the point where the
Nyquist plot crosses the unit circle.
The phase margin can also be found from the Bode plot of the loop gain because the two plots are closely
related (see Remark 8.2.7 and Example 8.3.4). The Bode plot of L(jω) is shown in Figure 8.17. The Nyquist
plot intersects the unit circle when |L(jω)| = 1. This corresponds to the magnitude plot crossing through
20 log |L(jω)| = 20 log |1| = 0 dB. Furthermore, the Nyquist plot intersects the negative real axis whenever2
](L(jω)) = π. This corresponds to the phase response crossing −180◦ . So to find the phase margin from a
Bode plot of L, we determine the frequency at which the magnitude is 0dB and then measure how far the phase
at that frequency is from −180, see Figure 8.17. N

Definition 8.4.1. Let L(s) ∈ R(s) be the loop gain TF for the feedback system in Figure 8.7.

(i) A gain crossover frequency, ωgc ∈ [0, ∞), for L(s) is a frequency at which |L(jωgc )| = 1.

2
Remember that we take the angles π and −π to be equivalent. See Appendix 3.A.

Version 1.2.2816, September 1, 2020

 8.4. Stability margins 178

10

-10

-20

Magnitude (dB) -30

-40

-50

-60

-70
0

-45
Phase (deg)

-90

-135

-180
10 -2 10 -1 10 0 10 1 10 2
Frequency (rad/s)

Figure 8.17: Bode plot of L(jω) for Example 8.4.1.

Let ωgc,1 , . . . , ωgc,` be the gain crossover frequencies for L(s) and assume these are ordered so that

](L(jωgc,1 )) < · · · < ](L(jωgc,` )).

(ii) The upper phase margin of L(s) is

(
180◦ + ](L(jωgc,1 )), if ](L(jωgc,1 )) ≤ 0
Φupm := (8.2)
undefined, else.

(iii) The lower phase margin of L(s) is

(
180◦ − ](L(jωgc,` )), if ](L(jωgc,` )) ≥ 0
Φlpm := (8.3)
undefined, else.

In simpler terms: (1) the upper phase margin is the amount of phase decrease required to reach the stability
limit and (2) the lower phase margin is the amount of phase increase required to reach the stability limit. The
upper and lower phase margins are easy to read off a Nyquist plot or Bode plot so there is usually no need to
compute them using the definition. See Figure 8.18. We adopt the convention that when we simply say phase
margin, we refer to the smaller of the upper and lower phase margins and denote it by Φpm .
Remark 8.4.2. In many texts one simply sees “phase margin” without reference to upper and lower. They
are usually referring to the upper phase margin which is more relevant for closed-loop stability when the loop
gain L(s) is minimum phase and has all its poles in C− . 

Example 8.4.2. Consider the system in Figure 8.7 with

−1
C(s) = 14, P (s) =
s2 + 4s + 15
The Nyquist plot of CP is shown in Figure 8.19a. The loop gain has no poles inside C+ and there are no

Version 1.2.2816, September 1, 2020

179 CHAPTER 8. FREQUENCY DOMAIN METHODS FOR STABILITY ANALYSIS

Figure 8.18: A Nyquist plot illustrating lower and upper phase margins.
1

0.8

0.6

10

0.4 0

-10

0.2
Magnitude (dB)

-20

-30
0
-40

-50
-0.2

-60
180

-0.4

135

-0.6
Phase (deg)

90

-0.8
45

-1
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1 0
10 -1 10 0 10 1 10 2
Frequency (rad/s)

(a) Nyquist plot. (b) Phase margin.

Figure 8.19: Phase margin from Nyquist plot in Example 8.4.2.

encirclement of the critical point −1 + j0 so the closed-loop system is stable. We observe that the blue (non-
dashed) portion of the Nyquist plot intersects the unit circle twice and that if we rotate in the CCW direction
(adding phase) by 27◦ we get an encirclement of −1 + j0 thereby losing closed-loop stability.
From the Bode plot in Figure 8.19 we determine that there are two gain crossover frequencies : ωgc,1 = 3.37
rad/s and ωgc,2 = 1.59 rad/s with (see Definition 8.4.1)
](L(jωgc,1 )) = 104.7◦ < ](L(jωgc,2 )) = 153◦ .
Thus by Definition 8.4.1 we conclude that lower phase margin is Φlpm = 27◦ and the upper phase margin is
undefined. MATLAB reports this as a negative phase margin of −27◦ . Conclusion: We need the Nyquist plot
for the correct interpretation of the phase margins. N

8.4.2 Gain margin

Example 8.4.3. Consider the usual feedback loop as in Figure 8.7 and let
1
C(s) = 2, P (s) = .
(s + 1)2 (0.1s + 1)

Version 1.2.2816, September 1, 2020

 8.4. Stability margins 180

The Nyquist plot of L is shown in Figure 8.20 where, again, the image of the portion of the Nyquist contour
corresponding to negative frequencies is drawn with a dashed line.
2

1.5

0.5

-0.5

-1

-1.5

-2
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2

Figure 8.20: Nyquist plot of Example 8.4.3 with gain margin indicated.

There are no encirclements of the critical point −1 + j0 so the feedback system is stable. The feedback
loop will remain stable for KC(s), where K ∈ R is a gain, so long as the point −1/K isn’t encircled. Since the
Nyquist plot crosses the negative real axis at −1/(12.1), this means we need
1 1
− <− .
K 12.1
Thus K can be increased from K = 1 up to K = 12.1 (21.67 dB) before the critical point is encircled resulting
in instability. This number, 21.67 dB, is called the gain margin. It is the distance from the critical point
−1 + j0 to the point where the Nyquist plot crosses the negative real axis.
The gain margin can be found from the Bode plot of the loop gain. The Bode plot of L(jω) is shown in
Figure 8.21. The Nyquist plot of L(jω) crosses the negative real axis whenever ](L(jω)) = π. Furthermore, the
magnitude of L(jω) equals 0db if |L(jω)| = 1. So to find the gain margin from a Bode plot of L, we determine
the frequency at which the phase is π and then measure how far the gain at that frequency is from 0 dB, see
Figure 8.21. N

Remark 8.4.3. The Bode plot in Figure 8.21 was created using the MATLAB command bode. By default,
this command “unwraps” the phase which is why the phase in Figure 8.21 goes below −180◦ . Figure 8.22 shows
the same Bode plot with phase wrapping on. Definitions 8.4.1 and 8.4.4 assume that phase wrapping is in effect.
The MATLAB commands used to create Figure 8.22 are shown below.

1 s = tf('s');
2 L = 2/(s+1)ˆ2/(0.1*s + 1);
3 opts = bodeoptions('cstprefs');
4 opts.PhaseWrapping = 'on';
5 bodeplot(L, opts);
6 grid on;

Version 1.2.2816, September 1, 2020

181 CHAPTER 8. FREQUENCY DOMAIN METHODS FOR STABILITY ANALYSIS

20

-20

-40

Magnitude (dB)
-60

-80

-100

-120

-140

-160
0

-45

-90
Phase (deg)

-135

-180

-225

-270
10 -2 10 -1 10 0 10 1 10 2 10 3
Frequency (rad/s)

Figure 8.21: Bode plot of L(jω) for Example 8.4.3.

20

-20

-40
Magnitude (dB)

-60

-80

-100

-120

-140

-160
180

135

90

45
Phase (deg)

-45

-90

-135

-180
10 -2 10 -1 10 0 10 1 10 2 10 3
Frequency (rad/s)

Figure 8.22: Bode plot of L(jω) for Example 8.4.3 with phase wrapping.

Definition 8.4.4. Let L(s) ∈ R(s) be the loop gain TF for the feedback system in Figure 8.7.
(i) A phase crossover frequency, ωpc ∈ [0, ∞), for a loop TF L(s) is a frequency at which ]L(jωpc ) = 180◦ .
Let ωpc,1 , . . . , ωpc,` be the phase crossover frequencies for L(s) and assume these are ordered so that
L(jωpc,1 ) < · · · < L(jωpc,` ).
(ii) The upper gain margin of L(s) is
(
1
, if there exists a k ∈ {1, . . . , ` − 1} such that L(jωpc,k ) < −1 < L(jωpc,k+1 ).
Kugm := |L(jωpc,k+1 )|
undefined, if L(jωpc,` ) < −1.
(8.4)

Version 1.2.2816, September 1, 2020

 8.4. Stability margins 182

(iii) The lower gain margin of L(s) is

(
1
, if there exists a k ∈ {1, . . . , ` − 1} such that L(jωpc,k ) < −1 < L(jωpc,k+1 ).
Klgm := |L(jωpc,k )|
undefined, if −1 < L(jωpc,1 ).
(8.5)

In simpler terms: (1) the upper gain margin is the smallest amount that the loop gain can be increased
before the closed-loop system goes unstable and (2) the lower gain margin is smallest amount that the loop gain
can be decreased before reaching the stability limit. As with phase margin, the upper and lower gain margins
are easy to read off a Nyquist plot or Bode plot so there is usually no need to compute them using the definition,
see Figure 8.23. We adopt the convention that when we simply say gain margin, we refer to the smaller of the
upper and lower phase margins and denote it by Kgm .

Figure 8.23: A Nyquist plot illustrating lower and upper gain margins.

Remark 8.4.5. In many texts one simply sees “gain margin” without reference to upper and lower. They are
usually referring to the upper gain margin which is more relevant for closed-loop stability when the loop gain
L(s) is minimum phase and has all its poles in C− . 

Example 8.4.4. Let

s+1
C(s) = 2, P (s) = .
s(s − 1)
The Nyquist plot of CP is shown in Figure 8.24a (see Example 8.3.2). The critical point is −1 + j0 and we
need 1 CCW encirclement to ensure the feedback system is stable. The phase margin on the Nyquist plot is
shown in Figure 8.24b
If KC(s) = 2K then the critical point is −1/K and K can be reduced from the value of 1 until −1/K = −2,
i.e., the smallest allowable K is K = 0.5 (6 dB). The Bode plot of this system generated with MATLAB is
shown in Figure 8.25.
Using Definition 8.4.4 we conclude that the lower gain margin is 0.5 and the upper gain margin is undefined.
MATLAB reports this as a negative gain margin of −6 dB. Conclusion: We need the Nyquist plot for the
correct interpretation of the stability margins. N

Version 1.2.2816, September 1, 2020

183 CHAPTER 8. FREQUENCY DOMAIN METHODS FOR STABILITY ANALYSIS

(a) Nyquist plot. (b) Phase margin.

Figure 8.24: Phase margin from Nyquist plot in Example 8.4.4.

60

50

40

30
Magnitude (dB)

20

10

-10

-20

-30

-40
180

135

90

45
Phase (deg)

-45

-90

-135

-180
10 -2 10 -1 10 0 10 1 10 2
Frequency (rad/s)

Figure 8.25: Bode plot of C(jω)P ((jω) for Example 8.4.4.

8.4.3 Stability margin

Even if both the gain and phase margin are reasonable, the system may not be robust as illustrated by the
following example.

Example 8.4.5. (Good gain and phase margin but poor robustness) Consider the loop gain

0.38(s2 + 0.1s + 0.55)

L(s) = .
s(s + 1)(s2 + 0.06s + 0.5)

The Nyquist and Bode plots for this system are shown in Figure 8.26.
From either plot we get the Φpm ≈ 70◦ and Kgm is infinite. These values indicate that the system is robust,
but the Nyquist curve is still close to the critical point, as shown in Figure 8.26a. This example illustrates the
limitation of gain and phase margins when it comes to quantifying robustness to modelling uncertainty. N

Version 1.2.2816, September 1, 2020

 8.4. Stability margins 184

1.5

40

0.5
20

Magnitude (dB)
-20

0
-40

-60

-80

-0.5 -100
-90

-120

Phase (deg)
-1

-150

-1.5
-1.5 -1 -0.5 0 0.5 1 1.5 -180
10 -2 10 -1 10 0 10 1 10 2
Frequency (rad/s)

(a) Nyquist plot. (b) Bode plot.

Figure 8.26: Nyquist and Bode plots for Example 8.4.5.

Let’s recap: The phase margin is related to the distance from the critical point to the Nyquist plot along
the unit circle; the gain margin is related to the distance from the critical point to the Nyquist plot along the
real axis. More generally, it makes sense to define the stability margin to be the distance from the critical
point to the closest point on the Nyquist plot.

Figure 8.27: Unity feedback system.

Consider the system in Figure 8.27 and define S(s) to be the TF from the reference input r to the tracking
error e
1
S(s) := . (8.6)
1 + C(s)P (s)
Assume that the feedback system is stable. Then the distance from −1 + j0 to the Nyquist plot of L(s) =
C(s)P (s) can be expressed as follows

min | − 1 − L(jω)| = min |1 + C(jω)P (jω)|

ω ω
h i−1
= max |S(jω)|
ω
= reciprocal of peak magnitude on Bode plot of S(jω).

In words: the point on the Nyquist plot which is closest to the point −1 + j0 is a distance equal to the reciprocal
of peak magnitude on Bode plot of S(jω). Thus, to increase the stability margin, one may wish to make the
function S(s) small.

Version 1.2.2816, September 1, 2020

185 CHAPTER 8. FREQUENCY DOMAIN METHODS FOR STABILITY ANALYSIS

Definition 8.4.6. Let L(s) ∈ R(s) be the loop gain TF for the feedback system in Figure 8.7 and let S(s) =
(1 + L(s))−1 . The stability margin, of the feedback system is
h i−1
sm := max |S(jω)| .
ω

Example 8.4.6. Let’s return to the system from Example 8.4.5. The Bode plot of the transfer function
1
S(s) =
1 + L(s)

is shown in Figure 8.28.

20

10

0
Magnitude (dB)

-10

-20

-30

-40
90

60
Phase (deg)

30

0
10 -2 10 -1 10 0 10 1
Frequency (rad/s)

Figure 8.28: Bode plot of S(s) = (1 + L(s))−1 for Example 8.4.6.

From the magnitude plot of S(jω) we get that the peak response is 11.3 dB which means that maxω |S(jω)| =
3.6728. The stability margin is therefore
h i−1 1
max |S(jω)| = = 0.2723
ω 3.6728
which is very low. N

There is a very nice geometric interpretation of the stability margin on the Nyquist plot. Suppose that, to
ensure good robustness, we have a desired stability margin of sm ≥ M . This means that we want the Nyquist
plot to be a distance greater than or equal to M from the critical point. In this case we want the Nyquist plot
to appear as in Figure 8.29.
When designing feedback systems, it will often be useful to define the robustness of the system using gain,
phase and stability margins. These numbers tell us how much the system can vary from our nominal model
and still be stable. Reasonable values of the margins are shown in Table 8.4.

Version 1.2.2816, September 1, 2020

 8.5. Summary 186

Figure 8.29: Geometric interpretation of a system with stability margin sm ≥ M .

Table 8.4: Reasonable values for stability margins.

Gain margin Phase margin Stability margin

6 − 13 dB 30◦ − 60◦ 0.5 − 0.8

8.5 Summary
This chapter is very important because the Nyquist plot and its interpretation lie at the heart of the design
methods discussed in the next chapter. After completing this chapter you should know the following.

1. You should know what it means for a closed curve in the complex plane to have CW (negative) orientation
and CCW (positive) orientation.

2. You should understand the basic ideas of Section 8.1 and especially the idea behind the construction
of Nyquist plots: A Nyquist plot is the image of a curve in the complex plane under a complex-valued
rational function.

3. Given a feedback system you should be able to find its loop gain TF L(s).

4. You must be able to draw and, even more importantly, interpret a Nyquist plot.

5. You should completely understand every example in Section 8.3.

6. Stability margins tell us how far a system is from being unstable. This depends entirely on how much mod-
elling uncertainty there is. For TFs, modelling uncertainty is naturally measured in terms of magnitude
and phase as functions of frequency.

7. Understanding the meanings of gain and phase margin, in terms of feedback stability, is essential. You
should be able to read gain and phase margins off Nyquist and Bode plots.

8. Know the definition of stability margin and be aware of the deficiency of using gain and phase margins
alone. You should know how to obtain the stability margin from a Bode plot of (8.6).

Version 1.2.2816, September 1, 2020

Chapter 9

Introduction to control design in the frequency domain

In this chapter we develop the basic technique of controller design in the frequency domain. Specifications are
in the form of bandwidth and stability margin. Of course our real interest is how the system will behave in
physical time (e.g. settling time, overshoot, steady-state tracking error) and in using the frequency domain we
are employing the duality between the two domains. The methods we use in this chapter fall under the broad
heading of loop shaping, as the objective is to shape the Nyquist plot to have desired properties.

Contents
9.1 Loop shaping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
9.2 Performance specifications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
9.3 Lag compensation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
9.4 Lead compensation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198
9.5 Lead-lag compensation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
9.6 Loop shaping theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
9.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211

9.1 Loop shaping

Consider a unity feedback system as shown in Figure 9.1.

Figure 9.1: Unity feedback system.

The design problem is this: Given P , the nominal plant transfer function, maybe some uncertainty bounds,
and some performance specifications design an implementable C. The performance specs would include, as
a bare minimum, stability of the feedback system. The simplest situation is where the performance can be
specified in terms of the transfer function
1
S(s) = . (9.1)
1 + C(s)P (s)
which is called the sensitivity function.

187
 9.1. Loop shaping 188

Definition 9.1.1. The sensitivity function of a feedback system with loop gain L(s) is S(s) = (1 + L(s))−1 .

Remark 9.1.2. The transfer function (9.1) is called the sensitivity function for the following reason. Let G(s)
denote the transfer function from r to y
C(s)P (s)
G(s) := .
1 + C(s)P (s)
Consider the relative perturbation in G due to a relative perturbation in P
∆G/G ∆G P
lim = lim
∆P →0 ∆P/P ∆P G
∆P →0
dT P
= ·
dP T 
d C(s)P (s) 1 + C(s)P (s)
= ·P ·
dP 1 + C(s)P (s) C(s)P (s)
= S(s).

So S is a measure of the sensitivity of the closed-loop transfer function to variations in the plant transfer
function. 
For us, S is important for two reasons: 1) S is the transfer function from the reference signal r to the
tracking error e. Thus we want |S(jω)| to be small over the range of frequencies of r (see Example 5.4.4). 2)
The peak magnitude of S is the reciprocal of the stability margin (see Definition 8.4.6). Thus a typical desired
magnitude plot of S(jω) is shown in Figure 9.2.

Figure 9.2: Typical desired shape for a sensitivity function.

Here ω1 is the maximum frequency of the reference signal r, ε is the maximum permitted relative tracking
error, ε < 1, and M is the maximum peak magnitude of |S|, M > 1. If |S| has this shape and the feedback
system is stable, then for the input r(t) = cos (ωt), ω ≤ ω1 , we have |e(t)| ≤ ε in steady state, and the stability
margin is at least 1/M . A typical value for M is 2 (see Table 8.4). In these terms, the design problem can be
stated as follows: Given P, M, ε, ω1 ; design C so that the feedback system is stable and |S| satisfies |S(jω)| ≤ ε
for ω ≤ ω1 and |S(jω)| ≤ M for all ω.
Example 9.1.1.
10
P (s) =
0.2s + 1

Version 1.2.2816, September 1, 2020

189 CHAPTER 9. INTRODUCTION TO CONTROL DESIGN IN THE FREQUENCY DOMAIN

This is a typical transfer function of a DC motor (cf. Example 2.8.5). Let’s take a PI controller:
Ki
. C(s) = Kp +
s
Then any M, ε, ω1 are achievable by suitable Kp , Ki . To see this, start with
1
S(s) = 10(Kp s+Ki )
1+ s(0.2s+1)
s(0.2s + 1)
= .
0.2s2 + (1 + 10Kp )s + 10Ki
We can simplify by reducing to two equal real poles:
5s(0.2s + 1)
S(s) = .
(s + K)2
Clearly, if we find a suitable K, then we can back-solve for Kp , Ki . Figure 9.3 shows the Bode plot of S(jω) for
K = 10.
0

-10

-20

-30
Magnitude (dB)

-40

-50

-60

-70

-80

-90
180

135
Phase (deg)

90

45

0
-1 0 1 2 3 4
10 10 10 10 10 10
Frequency (rad/s)

Figure 9.3: Bode plot of sensitivity function from Example 9.1.1.

Sketch the asymptotic Bode plot of S and confirm that any M > 1, ε < 1, and ω1 can be achieved. N
Exercise 9.1. Design a PI controller for the system in Example 9.1.1 for M = 2, ε = 0.01, and ω1 = 50 rad/s.
In practice it is common to combine interactively the shaping of S with a time-domain simulation. Now,
S(s) is a nonlinear function of C(s). So in fact it is easier to shape the loop transfer function L(s) = C(s)P (s)
instead of S(s) = (1 + L(s))−1 . Notice that if the loop gain is high, i.e., |L|  1, then
1
|S| ≈ .
|L|
A typical desired loop gain amplitude plot is shown in Figure 9.4. In shaping |L| instead of |S|, we don’t have
a direct handle on the stability margin, unfortunately. However, we can directly influence the gain and phase
margins, as we’ll see.

9.2 Performance specifications

A key element of the control design process is how we specify the desired performance of the system. Specifi-
cations are often given in terms of robustness to modelling uncertainty and responses to reference signals and
disturbances. They can be given in terms of both time and frequency responses. In this section we show how
to convert time domain specifications into frequency domain specifications. We will do so in the same spirit as
Chapter 4 by focusing on second order systems.

Version 1.2.2816, September 1, 2020

 9.2. Performance specifications 190

Figure 9.4: Typical desired shape for a loop gain. The loop gain is easier to work with for design than the sensitivity function
because it is linear in the controller. High gain at low frequencies provides good tracking and disturbance rejection. Fast
roll off to the gain crossover frequency gives good robustness. Low gain at high frequencies gives good rejection of high
frequency noise.

Relationship between damping ratio and phase margin

Consider the system in Figure 9.1 with
ωn2
L(s) = ∈ R(s). (9.2)
s(s + 2ζωn )
The transfer function from r to y is
Y (s) ωn2
= 2
R(s) s + 2ζωn s + ωn2
which is the prototype second order system from Section 4.2 with damping ratio ζ. To find the gain crossover
frequencies we set |L(jω)| = 1 and solve for ω to obtain
qp
ωgc = ωn 1 + 4ζ 4 − 2ζ 2 . (9.3)
Using Definition 8.4.1 (after some algebra) we get that the phase margin is
 
2ζ
Φpm = arctan  qp . (9.4)
4
1 + 4ζ − 2ζ 2

Observe that the phase margin only depends on the damping ratio ζ and not on ωn . Figure 9.5 shows a graph
of phase margin Φpm versus damping ratio ζ and also shows that it can be reasonably well approximated by a
line for small values of ζ. We conclude that, for the system (9.2) we have
If 0 ≤ ζ ≤ 0.7 then Φpm ≈ 100ζ (expressed in degrees). (9.5)
While the expression (9.5) was derived for the loop gain (9.2), we will use it as a general rule-of-thumb for
converting overshoot specifications into phase margin specifications.

Relationship between gain crossover frequencies and bandwidth

We saw in Chapter 4 that the speed of a system’s response is closely tied to its bandwidth. Hence performance
specifications like settling time and rise time can be cast as a specification on the closed-loop bandwidth ωBW .
It is therefore useful to approximate the closed-loop bandwidth from the loop gain transfer function L(s). If
L(0) is finite, then usually ωgc ≤ ωBW ≤ 2ωgc . For design purposes we employ the following useful rule-of-thumb
ωBW ≈ ωgc . (9.6)

Version 1.2.2816, September 1, 2020

191 CHAPTER 9. INTRODUCTION TO CONTROL DESIGN IN THE FREQUENCY DOMAIN
100

90

80

70

60

50

40

30

20

10

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Figure 9.5: Phase margin Φpm as a function ζ for (9.2) (black) and its linear approximation (grey).

Remark 9.2.1. In the next two sections we present two simple loop shaping controllers. These loop shaping
controllers work best for “nice plants” that (i) are stable or, at worst, have a single pole at s = 0 and (ii) have
only one gain crossover frequency ωgc and only one phase crossover frequency ωpc . 

9.3 Lag compensation

We separate the controller into two parts K and C1 (s) as shown in Figure 9.6. The controller is

Figure 9.6: Lag controller block diagram.

αT s + 1
C(s) = KC1 (s) = K , 0 < α < 1, T > 0, K > 0, (lag controller). (9.7)
Ts + 1
The lag controller has a steady-state gain of C(0) = KC1 (0) = K and its pole and a zero locations are shown in
Figure 9.7. The asymptotic Bode plot of C1 (s) is shown in Figure 9.8. A lag controller is like a low pass filter.

Figure 9.7: Pole-zero configuration of a lag controller: the pole is always closer to the imaginary axis than the zero.

Version 1.2.2816, September 1, 2020

 9.3. Lag compensation 192

Magnitude (dB)
Phase (deg)

Figure 9.8: Asymptotic Bode plot of a lag controller with K = 1.

Exercise 9.2. Prove that magnitude plot of C1 (s) levels off at 20 log |α| as indicated in Figure 9.8.
1
Exercise 9.3. Prove that ωm , the geometric mean of the pole and zero location, is given by ωm = T
√
α
.

Lag controllers are used for two distinct purposes:

1. To increase the phase margin. This is done indirectly by lowering the high frequency gain.

2. Boost low frequency gain to improve steady-state tracking and disturbance rejection without having too
much effect on gain margin, phase margin nor high frequency behaviour.
The loop shaping idea behind lag design is illustrated in Figure 9.9.
Magnitude (dB)
Phase (deg)

Frequency (rad/s)

Figure 9.9: The black plot shows the Bode plot of KP (jω) where K has been chosen to meet steady-state tracking and
disturbance rejection specifications. The red plot is the lag compensated system KC1 (jω)P (jω). The lag controller reduces
the high frequency gain without phase lag. This lowers the gain crossover frequency which, for “nice plants,” increases the
phase margin.

Example 9.3.1. (Increasing phase margin with a lag controller) The plant transfer function is
1
P (s) = .
s(s + 2)

Version 1.2.2816, September 1, 2020

193 CHAPTER 9. INTRODUCTION TO CONTROL DESIGN IN THE FREQUENCY DOMAIN

The specifications are:

(a) When r(t) is a unit ramp, the steady-state tracking error must be less than or equal to 0.05.

(b) A phase margin of 45◦ for adequate damping in transient response.

Step 1: Choose K to meet the tracking specification.
1
E(s) = R(s) (TF from e to r)
1 + C(s)P (s)
1 1
= αT s+1 1
1 + K T s+1 s(s+2) s2
(s + 2)(T s + 1) 1
= .
s(s + 2)(T s + 1) + K(αT s + 1) s
Assuming closed-loop stability, use the FVT
2
ess = lim e(t) = lim sE(s) = .
t→∞ s→0 K
So specification (a) is satisfied if K ≥ 40. We’ll pick K = 40 so that
40
KP (s) = .
s(s + 2)
Step 2: Draw the Bode plot of KP (s). It is shown in Figure 9.10a. From Figure 9.10a we find that the gain
50

40 4

30

20
3
Magnitude (dB)

10

0
2
-10

-20

-30 1

-40

-50
0
-90

-1

-120
Phase (deg)

-2

-150

-3

-180
10 -1 10 0 10 1 10 2 -4
-2.5 -2 -1.5 -1 -0.5 0
Frequency (rad/s)

(a) Bode plot of KP (jω). (b) Nyquist plot of KP (s).

Figure 9.10: Bode and Nyquist plot for Example 9.3.1 with only proportional control.

crossover frequency is ωgc = 6.17 rad/s and the phase margin is Φpm = 18◦ . Figure 9.10b shows the Nyquist
plot of KP (s). The system is closed-loop stable with K = 40 since there are no encirclements of −1 + j0 (not
shown in Figure 9.10b is the semi-circle of infinite radius on the right). However we see that a rotation of the
plot by −18◦ will result in an encirclement.
Step 3: We now design the lag compensator C1 . The idea is to reduce the high frequency gain and increase
Φpm while not changing the low frequency gain (thus preserving Specification (a) achieved in Step 1).
We want Φpm = 45◦ but we’ll actually aim for 50◦ , a bit more than 45◦ , to compensate for the fact that the
phase plot of a lag controller only approaches zero asymptotically as frequency increases.
From the Bode plot in Step 2 we find the frequency at which we have the correct phase margin. We obtain
◦
Φdes
pm = 50 = 180 + ](KP (jω)) when ω = 1.7 rad/s.

So we aim to make the gain crossover frequency 1.7 rad/s.

Version 1.2.2816, September 1, 2020

 9.3. Lag compensation 194

Again, from the Bode plot of KP (jω), we see that the gain at ω = 1.7 rad/s is 19 dB. Therefore we want to
reduce the gain by 19 dB at ω = 1.7 without changing the phase there. We have (see Figure 9.8 and Exercise 9.2)

20 log |α| = −19dB ⇔ α = 0.111.

We now pick T so that we don’t change the phase near ω = 1.7 by ensuring that
10
≤ 1.7.
αT
I’ll pick T = 52.7 (equality).
Step 4: We now plot the Bode plot of the lag compensated loop gain L(s) = KC1 (s)P (s). It is shown in
Figure 9.11. From Figure 9.11 we find that the phase margin is now Φpm = 44.6◦ . Close enough. Figure 9.12a
120

100

80

60
Magnitude (dB)

40

20

-20

-40

-60

-80
-90

-120
Phase (deg)

-150

-180
10 -3 10 -2 10 -1 10 0 10 1 10 2
Frequency (rad/s)

Figure 9.11: Bode plot of KC1 (jω)P (jω) (red) and KP (jω) (blue) for Example 9.3.1.

shows the Nyquist plot of the compensated system. We see that the curve is now much farther from the critical
point.
1.8

4 1.6

3 1.4

1.2
2

1
Amplitude

0.8

0.6

-1

0.4

-2

0.2

-3
0
0 2 4 6 8 10 12 14
Time (seconds)

-4
-2.5 -2 -1.5 -1 -0.5 0

(b) Closed-loop step response with proportional (blue) and lag

(a) Nyquist plot of KC1 (s)P (s). (red) control.

Figure 9.12: Nyquist plot and step response for Example 9.3.1 with final lag controller.

Step responses are shown in Figure 9.12b. The step response of KP/(1+KP ), that is, the plant compensated
only by the gain for Spec (a), is fast but oscillatory. The step response of P C/(1 + P C) is slower, typical of lag
compensation1 , but less oscillatory which was the goal of Specification (b). N
1
A lag controller reduces the gain crossover frequency which, using our rule-of-thumb from Section 9.2, reduces the closed-loop
bandwidth. Lowering the bandwidth, as we saw in Chapter 4, results in a slower response.

Version 1.2.2816, September 1, 2020

195 CHAPTER 9. INTRODUCTION TO CONTROL DESIGN IN THE FREQUENCY DOMAIN

The MATLAB code below was used to design the controller in the above example.

1 s = tf('s'); P = 1/s/(s+2);
2 K = 40;
3 [mag, phase, w] = bode(K*P);
4 Phi desired = 45; ∆ = 5;
5 [¬, i] = min(abs(phase + 180 − Phi desired − ∆)); % find freq. with desired phase
6 alpha = 1/mag(i);
7 T = 10/alpha/w(i);
8 C = K*(alpha*T*s + 1)/(T*s + 1);
9 [Gm,Pm,Wpc,Wgc] = margin(C*P); % check if the design meets the spec

Procedure for lag controller design

The given specifications are:

(a) Steady-state specification (determined by tracking or disturbance rejection requirement),

(b) Increase phase margin to be greater than or equal to Φdes

pm (determined by damping ratio or robustness
requirements).

1. Use FVT to fix the steady-state gain K of the controller.

2. Draw the Bode plot of KP (jω). Check Φpm .

3. If Φpm specification is met, we’re done (a proportional controller can do the job!). Otherwise, find ω ? such
that
dist (](KP (jω ? )), 180◦ ) = Φdes
pm + δ.

Here Φdes ◦
pm is the given specification and δ is a buffer to account for approximations, usually 5 .

4. Shift the gain down at the frequency ω ? from step 3 to get a new gain crossover frequency
1
α= .
K|P (jω)|

5. Ensure that the phase isn’t affected near the frequency ω ? from steps 3, 4
10
≤ ω?.
αT

6. Simulate the closed-loop system and check Bode plot of C(jω)P (jω) to make sure all specifications are
met.

The next example illustrates the second main use of lag controllers.

Example 9.3.2. (Boosting low frequency gain without changing high frequency gain) Consider the
plant with time-delay
1 + 0.48s
P (s) = 10.4e−0.05s .
s(s + 2)(1 + 0.13s)
The specifications are:

(a) When r(t) is a unit ramp, the steady-state tracking error must be less than or equal to 0.05.

(b) The phase margin and bandwidth should not be significantly changed.

Version 1.2.2816, September 1, 2020

 9.3. Lag compensation 196
40

20

Magnitude (dB)
-20

-40

-60

-80

-100
180

135

90

45
Phase (deg)

-45

-90

-135

-180
10 -1 10 0 10 1 10 2 10 3
Frequency (rad/s)

Figure 9.13: Bode plot of P (jω) for Example 9.3.2.

This is a typical lag compensation problem because we only want to increase the low frequency gain. The
Bode plot of P (jω) is shown in Figure 9.13. From the Bode plot we obtain that that the phase margin is
approximately 47◦ and that ωBW ≈ ωgc = 4.4 rad/s. We now have numbers to work with for the second
specification.
Step 1: Choose K to meet the tracking specification.
1
E(s) = R(s) (TF from e to r)
1 + C(s)P (s)
1 1
= αT s+1 1+0.48s
1 + K T s+1 10.4e−0.05s s(s+2)(1+0.13s) s2
(s + 2)(T s + 1)(1 + 0.13s) 1
= −0.05s
.
s(s + 2)(T s + 1)(1 + 0.13s) + 10.4Ke (αT s + 1)(1 + 0.48s) s
Assuming closed-loop stability, use the FVT
2
ess = lim e(t) = lim sE(s) = .
t→∞ s→0 10.4K
So specification (a) is satisfied if K ≥ 3.85. We’ll pick K = 4 so that
1 + 0.48s
KP (s) = 41.6e−0.05s .
s(s + 2)(1 + 0.13s)
Step 2: Draw the Bode plot of KP (s). Actually, there is no need in this case since Figure 9.13 has all the
needed data but remember to account for the increase in gain from our choice of K.
Step 3: The desired gain crossover frequency is ωgc = 4.4 rad/s and, since we’ve only added a gain of
K = 4, the gain of KP (jω) at ω = 4.4 rad/s is exactly 20 log |K|.
Step 4: We shift the gain down at ω = 4.4 to get
1 1
α= = = 0.25.
K|P (jω)| K
Step 5: Ensure the phase isn’t affected near ω = 4.4
10
≤ 4.4.
αT
I’ll take T = 9.1. Our overall lag controller is
0.25(9.1)s + 1 9.1s + 4
C(s) = KC1 (s) = 4 = .
9.1s + 1 9.1s + 1

Version 1.2.2816, September 1, 2020

197 CHAPTER 9. INTRODUCTION TO CONTROL DESIGN IN THE FREQUENCY DOMAIN
100

80

60

40

Magnitude (dB)
20

-20

-40

-60

-80

-100
180

135

90

45
Phase (deg)

-45

-90

-135

-180
10 -3 10 -2 10 -1 10 0 10 1 10 2 10 3
Frequency (rad/s)

Figure 9.14: Bode plot of KC1 (jω)P (jω) for Example 9.3.2.

Step 6: The Bode plot of the lag compensated loop gain L(s) = KC1 (s)P (s) is shown in Figure 9.14. From
Figure 9.14 we find that the phase margin is Φpm = 45◦ at ωgc = 4.38 rad/s. Close enough.
Exercise 9.4. Simulate the closed-loop step responses of P/(1 + P ) and CP/(1 + CP ). Verify that our lag compensated system
has a higher settling time.
Exercise 9.4 shows that the step response of the compensated system is slower. This is typical of a lag
compensator.
Exercise 9.5. Simulate the closed-loop response of the system to a ramp input and verify that our controller meets the tracking
specification.
Exercise 9.5 shows that our lag compensated system does a much better job tracking ramp signals. N

9.3.1 Proportional-Integral controllers

A lag controller is very closely related to the PI controller
 
1
C(s) = K 1 + .
Ti s

Expressing the PI controller in the form

 
K Ti s + 1
C(s) =
Ti s

it can be viewed as a special case of the lag controller in which the lag controller’s pole has been moved all the
way to the origin. The asymptotic Bode plot of a PI controller is shown in Figure 9.15. The high frequency
gain of a PI controller equals 20 log |K| and its high frequency phase equals zero2 . The PI controller’s gains can
be selected using the same procedure as a lag controller.

Example 9.3.3. Let’s re-do Example 9.3.1 using a PI controller. Once again the specifications are:

(a) When r(t) is a unit ramp, the steady-state tracking error must be less than or equal to 0.05.

(b) A phase margin of 45◦ for adequate damping in transient response.

Observe that the PI controller has integral action (a pole at s = 0) and so C(s)P (s) will have two poles at
the origin. Therefore ramps will be perfectly tracked in steady-state meaning that we only need to worry about
the phase margin specification.
2
In contrast to a lag controller whose high frequency gain is 20 log |α| and high frequency phase equals zero.

Version 1.2.2816, September 1, 2020

 9.4. Lead compensation 198

Magnitude (dB)
Phase (deg)

Figure 9.15: Asymptotic Bode plot of a PI controller.

Bode Diagram Bode Diagram
20 100

80

0
60

40
-20
Magnitude (dB)

Magnitude (dB)
20

-40 0

-20

-60
-40

-60
-80

-80

-100 -100
-90 -90

-120
Phase (deg)

Phase (deg)

-135

-150

-180 -180
10 -1 10 0 10 1 10 2 10 -3 10 -2 10 -1 10 0 10 1 10 2
Frequency (rad/s) Frequency (rad/s)

(a) Bode plot of P (jω). (b) Bode plot of C(jω)P (jω).

Figure 9.16: Uncompensated and PI compensated Bode plots for Example 9.3.3.

Step 1: We already meet the tracking specification so there is nothing to do in this step.
Step 2: Draw Bode plot of P (jω). See Figure 9.16a.
Steps 3 – 4: From Figure 9.16a we have that the phase margin is 76◦ at ωgc = 0.5 rad/s. We’re already
meeting the specification.
Step 5: Since the phase margin specification is already met, we select K and Ti to ensure that neither the
gain nor the phase is affected very much near the crossover frequency ωgc = 0.5 rad/s. Looking at the gain Bode
plot from Figure 9.15, we should pick K = 1 to ensure that the gain isn’t changed near the crossover frequency.
Next pick Ti to ensure the PI controller does not affect the phase near the crossover frequency. Looking at the
phase Bode plot from Figure 9.15, we should pick
10
≤ 0.5.
Ti
I’ll take Ti = 20. Our controller is C(s) = 1 + 0.05s−1 .
Step 6: The Bode plot of the loop gain L(s) = C(s)P (s) in Figure 9.16b meets all the specifications.
N

9.4 Lead compensation

We again separate the controller into two parts K and C1 (s) as shown in Figure 9.17. The controller is
αT s + 1
C(s) = KC1 (s) = K , α > 1, T > 0, K > 0, (lead controller). (9.8)
Ts + 1

Version 1.2.2816, September 1, 2020

199 CHAPTER 9. INTRODUCTION TO CONTROL DESIGN IN THE FREQUENCY DOMAIN

Figure 9.17: Lead controller block diagram.

The lead controller has a steady-state gain of C(0) = KC1 (0) = K and its pole and a zero locations are shown
in Figure 9.18.

Figure 9.18: Pole-zero configuration of a lead controller: the pole is always farther from the imaginary axis than the zero.

The asymptotic Bode plot of C1 (s) is shown in Figure 9.19. A lead controller is a high pass filter. Lead
Magnitude (dB)
Phase (deg)

Figure 9.19: Asymptotic Bode plot of a lead controller.

controllers are used for two distinct purposes.

1. To increase the phase margin Φpm by adding phase at the appropriate frequency while also meeting a
steady-state tracking requirement.
2. Increase phase margin while simultaneously increasing the closed-loop bandwidth to make the system
response faster.

Version 1.2.2816, September 1, 2020

 9.4. Lead compensation 200

The loop shaping idea behind lead design is illustrated in Figure 9.20. Before doing a design example we need

Magnitude (dB)
Phase (deg)

Frequency (rad/s)

Figure 9.20: The black plot shows the Bode plot of KP (jω) where K has been chosen to meet steady-state tracking and
disturbance rejection specifications. The red plot is the lead compensated system KC1 (jω)P (jω). The lead controller
increases the phase of the system at a designer specified frequency. An undesirable side effect is that it also increases the
gain at the specified frequency. This changes ωgc .

to derive a few design equations. We’ll need three formulas.

1 1
1. The frequency ωm (see Figure 9.19). This is the midpoint between αT and T on the logarithmically scaled
frequency axis. From Exercise 9.3 we have
1
ωm = √ . (9.9)
T α

2. The magnitude of C1 (jω) at ωm . This is the midpoint between 20 log |1| and 20 log |α|
1
log |C1 (jωm )| = (log |1| + log |α|)
2 √
= log ( α).
Therefore √
|C1 (jωm )| = α. (9.10)

3. The angle φmax : This is the angle of C1 (jωm ) (see Figure 9.19). Thus

φmax = ]C1 (jωm )

√
1 + αj
=] .
1 + √1α j

By the sine law (see Figure 9.21)

sin (φmax ) sin (θ)
√ 1 =
q .
α− α √
1+ α 1

But sin (θ) = √1 . Thus

1+α

√
 
1 1 α−1
sin (φmax ) = α− √ √ q = ,
α 1+α 1+ 1 α+1
α

Version 1.2.2816, September 1, 2020

201 CHAPTER 9. INTRODUCTION TO CONTROL DESIGN IN THE FREQUENCY DOMAIN

Figure 9.21: Law of sines.

and hence  
−1 α−1 1 + sin (φmax )
φmax = sin , α= . (9.11)
α+1 1 − sin (φmax )

Example 9.4.1. (Increasing phase margin with a lead controller) Let’s re-do Example 9.3.1 using a
lead controller.
1 (a) |ess | ≤ 0.05 for r(t) = t,
P (s) = Specs :
s(s + 2) (b) Φpm = 45◦ .
We start with a trick; express the lead controller (9.8) in the form

αT s + 1 K̂ αT s + 1
C(s) = K =: √ .
Ts + 1 α Ts + 1
√
The α term is ugly, but as we’ll see, it’s convenient for design purposes.
Step 1: (same as lag) Choose K̂ ≥ 40 so that K̂P (s) meets the tracking spec. Boost K̂ by 10bB (a guess)
√ 1
to account for the magnitude reduction that will result from the α term in our trick. I’ll pick K̂ = 40 · 10 2 .
Draw the Bode plot of K̂P (s) and find that Φpm = 10.2◦ at ωgc = 11.2 rad/s.
Step 2: We need at least 45 − 10.2 = 34.8◦ phase addition in order to meet the spec so we set φmax = 34.8◦ .
The lead controller will add the most phase at ωm so we set ωm = ωgc = 11.2 rad/s to get the maximum phase
addition at the phase crossover frequency.
Step 3: Now that we’ve selected numerical values for φmax and ωm , we can use our design equations. We
get
1 + sin (φmax ) 1
α= = 3.66, T = √ = 0.0467
1 − sin (φmax ) ωm α
and we determine that 1
K̂ 40 · 10 2
K=√ = √ = 66.13.
α 3.66
The lead controller is
αT s + 1
C(s) = K , α = 3.66, T = 0.0467, K = 66.13.
Ts + 1
Step 4: We now plot the Bode plot of the lead compensated loop gain L(s) = KC1 (s)P (s). It is shown in
Figure 9.22. From Figure 9.22 we find that the phase margin is now Φpm = 45◦ . N

The MATLAB code below was used to design the controller in the above example.

1 s = tf('s'); P = 1/s/(s+2);
2 hatK = 40*sqrt(10); % boost of 10dB
3 [Gm,Pm,Wpc,Wgc] = margin(hatK*P); % get margins of partially compensated system

Version 1.2.2816, September 1, 2020

 9.4. Lead compensation 202
100

80

60

40

20

Magnitude (dB)
0

-20

-40

-60

-80

-100
-90

Phase (deg)

-135

-180
10 -1 10 0 10 1 10 2 10 3
Frequency (rad/s)

Figure 9.22: Bode plot of KC1 (jω)P (jω) (red) and KP (jω) (blue) for Example 9.3.1.

4 phi max = 45 − Pm; Wm = Wcp;

5 alpha = (1+sind(phi max))/(1 − sind(phi max));
6 T = 1/Wm/sqrt(alpha);
7 K = hatK/sqrt(alpha);
8 C = K*(alpha*T*s + 1)/(T*s + 1);
9 [Gm,Pm,Wpc,Wgc] = margin(C*P); % check if the design meets the spec

Exercise 9.6. Simulate and compare the step responses of the lead and lag controllers from Examples 9.3.1 and 9.4.1. Generate
plots of the respective control signals. They both meet the specifications but the lead controller has a faster response (lead increases
ωgc which increases the bandwidth) at the expensive of using more control effort.
The next example illustrates the second main use of lead controllers.

Example 9.4.2. (Increasing phase margin and bandwidth with lead controller) Consider the plant
1
P (s) = 4e−0.05s
s(s + 2)

and the specifications

(a) Phase margin of at least 50◦ .

(b) Settling time of no more than 2 seconds.

We start by converting the second specification into a frequency domain constraint. Assuming the system
behaves like a second order underdamped system we can use the following approximations.
◦
Φdes
pm = 50 ⇒ ζ = 0.5,
4
Ts ≈ ≤ 2 ⇒ ζωn ≥ 2,
ζωn
⇒ ωn ≥ 4.

For an underdamped second order system ωBW ≈ ωn so we want ωgc ≈ ωBW = 4 rad/s. Of course at the end
we have to check that our controller satisfies the specifications for the actual plant and not just for a second
order approximation.
√
Step 1: Using the trick from the previous example, define K̂ = K α. Now choose the gain K̂ to meet the
bandwidth specification. The Bode plot of P (jω) is shown in Figure 9.23. From the Bode plot we see that we
must boost the gain by about 14dB which yields
14
K̂ = 10 20 = 5.

Version 1.2.2816, September 1, 2020

203 CHAPTER 9. INTRODUCTION TO CONTROL DESIGN IN THE FREQUENCY DOMAIN
40

20

-20

Magnitude (dB)
-40

-60

-80

-100

-120
180

135

90

45
Phase (deg)

-45

-90

-135

-180
10 -1 10 0 10 1 10 2 10 3
Frequency (rad/s)

Figure 9.23: Bode plot of P (jω) for Example 9.4.2.

Step 2: Draw Bode plot of K̂P (jω). Since we already have the Bode plot of P (jω), we have that the Bode
plot of K̂P (jω) is the same as that in Figure 9.23 with the magnitude plot shifted upwards by 14dB.
Steps 3 – 4: We want the crossover frequency to be 4 rad/s. From Figure 9.23 we have that the phase at
ω = 4 is about −165◦ . Therefore we need to add at least

φmax = 50◦ − 15◦ = 35◦ .

Step 5: For 35◦ the formula (9.11) gives

α−1
sin (35◦ ) = ⇒ α = 3.69.
α+1
√
This yields K = K̂/ α = 2.6.
Step 6: We want to add the phase at exactly 4 rad/s so using (9.9) we get
1
T = √ = 0.1301.
4 3.69
In summary, the controller is
αT s + 1
C(s) = K , α = 3.69, T = 0.13, K = 2.6.
Ts + 1
The Bode plot of the lead compensated loop gain L(jω) = C(jω)P (jω) is shown in Figure 9.24a. From
Figure 9.24a we find that the phase margin is Φpm = 47◦ at ωgc = 4.38 rad/s. Close enough. The step
response of the closed-loop system is shown in Figure 9.24b. The settling time is 1.4 seconds which meets the
specification. N

Procedure for lead controller design

The given specifications are:
(a) Increase phase margin to be at least equal to Φdes
pm (determined by damping ratio or robustness require-
ments).

(b) One of the following:

(i) Steady-state specification (determined by tracking or disturbance rejection requirement),

(ii) A desired closed-loop bandwidth (determined by time-domain specifications or noise rejection re-
quirements),

Version 1.2.2816, September 1, 2020

 9.4. Lead compensation 204
40 1.4

20

0 1.2
Magnitude (dB)

-20

1
-40

-60

0.8
-80

Amplitude
-100
180
0.6
135

90

45 0.4
Phase (deg)

-45

0.2
-90

-135

-180
0
10 -1 10 0 10 1 10 2 10 3 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8
Frequency (rad/s) Time (seconds)

(a) Bode plot of KC1 (jω)P (jω). (b) Step response of P C/(1 + P C).

Figure 9.24: Bode plot of loop gain and step response for Example 9.4.2.

√
1. Define K̂ := K α.
(i) Use FVT to select K̂ such that K̂P (s) meets the steady-state specification. Purposefully boost the
gain by 10dB to compensate for the distortion that comes from α.
(ii) Pick K̂ so that the gain crossover frequency of K̂P (jω) equals the desired bandwidth.
2. Draw the Bode plot of K̂P (jω).
3. Find ωgc and Φpm ; set ωm = ωgc .
4. Determine the amount of phase to add: φmax = Φdes
pm − Φpm .
√
5. Compute α using (9.11). Set K = K̂/ α.
6. Compute T using (9.9).
7. Simulate closed-loop system and check Bode plot of KC1 (jω)P (jω) to make sure all specifications are
met. If they aren’t met, you can return to Step 1 and use a different K̂.

9.4.1 Proportional-Derivative controllers

Recall that an ideal PD controller has the form
C(s) = Kp (1 + Td s) .
Its asymptotic Bode plot is shown in Figure 9.25.
As discussed in Chapter 7 this controller cannot be implemented exactly. Instead we low pass filter the
derivative term to obtain  
Td s
C(s) = Kp 1 + .
τd s + 1
Re-arranging this expression we have
 
(τd + Td )s + 1
C(s) = Kp .
τd s + 1
This is exactly the form of a lead controller with
Td
K = Kp , T = τd , α=1+ .
τd
Therefore we can use the loop shaping ideas of this section without modification to design PD controllers.

Version 1.2.2816, September 1, 2020

205 CHAPTER 9. INTRODUCTION TO CONTROL DESIGN IN THE FREQUENCY DOMAIN

Magnitude (dB)
Phase (deg)

Figure 9.25: Asymptotic Bode plot of an ideal PD controller.

Example 9.4.3. (Ideal PD controller) Let’s re-do Example 9.3.1 using an ideal PD controller.

1 (a) |ess | ≤ 0.05 for r(t) = t,

P (s) = Specs :
s(s + 2) (b) Φpm = 45◦ .

Steps 1 – 3: Again choose K = 40 to meet the tracking spec. Once again we draw the Bode plot of KP (s)
and find that Φpm = 18◦ at ωgc = 6 rad/s (see Figure 9.10a).
Step 4: As before, we need at least 27◦ degrees of phase addition.
Step 5 – 6: The phase contribution of an ideal PD controller at the frequency ω is

]C(jω) = ]Kp + ](1 + jωTd ) = ](1 + jωTd ) = arctan2 (1, ωTd ).

So pick Td using
1 1
Td = tan (27◦ + δ) = tan (27◦ + 5◦ ) = 0.1041.
ωgc ωgc
Here ωgc = 6 rad/s is the frequency from the earlier steps and the buffer δ = 5◦ is included because the PD
controller increases the gain at ωgc .
Step 7: The controller we’ve designed is

C(s) = 40 (1 + 0.1041s) .

The Bode plot of C(jω)P (jω) is shown in Figure 9.26a. The phase margin is 51.8◦ at ωgc = 6.9 rad /s. The
tracking error for a ramp reference is shown in Figure 9.26b. N

9.5 Lead-lag compensation

Consider the controller in Figure 9.27. The controller is
α1 T1 s + 1 α2 T2 s + 1
C(s) = KC1 (s)C2 (s) = K , α1 > 1, 0 < α2 < 1, T1 , T2 > 0, K > 0, (lead-lag controller).
T1 s + 1 T2 s + 1
(9.12)
The lead-lag controller is the product of a lead controller C1 (s) and a lag controller C2 (s). It has a steady-state
gain of C(0) = K and typical pole and a zero locations are shown in Figure 9.18.
Individually, both lead and lag controllers have weaknesses. Lag controllers reduce the closed-loop bandwidth
thereby producing a slower response. Lead controllers often produce very large control signals. This motivates
us to combine the controllers together into the lead-lag configuration to hopefully obtain a compromise.

Version 1.2.2816, September 1, 2020

 9.5. Lead-lag compensation 206
0.12
50

40

30

20 0.1
Magnitude (dB)

10

-10
0.08
-20

-30

-40

0.06
-50
-90

-95

-100

0.04
-105
Phase (deg)

-110

-115

-120
0.02

-125

-130

-135
10 -1 10 0 10 1 10 2 10 3
0
Frequency (rad/s) 0 1 2 3 4 5 6 7 8 9 10

(a) Bode plot of C(jω)P (jω). (b) Ramp tracking error e(t) = t − y(t).

Figure 9.26: Bode plot of loop gain and ramp response for Example 9.4.3.

Figure 9.27: Lead-lag controller block diagram.

Figure 9.28: Typical pole-zero configuration of a lead-lag controller.

Procedure for lead-lag controller design

The given specifications are:
(a) Steady-state specification (determined by tracking or disturbance rejection requirement),

(b) Increase phase margin to be greater than or equal to Φdes

pm (determined by damping ratio or robustness
requirements).
√
1. Define K̂ := K α1 . Use FVT to select K̂ such that K̂P (s) meets the steady-state specification. Pur-
posefully boost the gain by 10dB to compensate for the distortion that comes from α1 .

2. Draw the Bode plot of K̂P (jω).

3. If Φpm specification is met, we’re done (a proportional controller can do the job!). Otherwise, divide Φdes
pm
into two roughly equal parts Φdes
pm = Φ pm,1 + Φ pm,2 .

4. Design the lag controller C2 (s) to get Φpm = Φpm,2 .

Version 1.2.2816, September 1, 2020

207 CHAPTER 9. INTRODUCTION TO CONTROL DESIGN IN THE FREQUENCY DOMAIN

5. Draw the Bode plot of K̂C2 (jω)P (jω).

6. Design the lead controller C1 (s) to get Φpm = Φdes

pm for the partially compensated system.

7. Simulate the closed-loop system and check Bode plot of C(jω)P (jω) to make sure all specifications are
met.

Remark 9.5.1. The reason that the lag controller is designed first in the above procedure is that a lead
controller will tend to “flatten” the phase curve. This means that if we tried to then design a lag controller,
the crossover frequency will be very small and we’ll get a sluggish response. 

Example 9.5.1. Let’s design a lead-lag controller for the following:

1 (a) |ess | ≤ 0.01 for r(t) = t1(t),

P (s) = Specs :
s(s + 1)(s + 20) (b) Φpm = 45◦ .
√
Step 1: Define K̂ = K α1 . Assuming feedback stability, the FVT for r(t) = t1(t) gives
20
ess = lim sE(s) = .
s→0 K̂
1
To meet the tracking specification we need K̂ ≥ 200. Boost K̂ by 10bB (a guess). I’ll pick K̂ = 200 · 10 2 .
Steps 2–3: From the Bode plot of K̂P (jω) we find that Φpm = −5◦ at ωgc = 5.5 rad/s. Divide Φdes pm = 45
◦

into two Φpm,1 = Φpm,2 = 22.5◦ .

Step 4: We follow the standard lag control design procedure to design C2 (s) to achieve a phase margin of
Φpm,2 = 22.5◦ . This yields α2 = 0.088, T2 = 75.6 and
6.6s + 1
C2 (s) = .
75.6s + 1

Step 5: The phase margin of K̂C2 (s)P (s) is Φpm = 23.8◦ at ωgc = 1.5 rad/s.
Step 6: We follow the standard lead control design procedure to design C1 (s) to increase the phase margin
by φmax = Φdes ◦
pm − Φpm = 21.2 at ωm = 1.5 rad/s. This yields α1 = 2.13, T1 = 0.45, K = 433 and

0.961s + 1
C1 (s) = .
0.45s + 1
Step 7: The overall controller is given by

1207s2 + 1599s + 200

C(s) = KC1 (s)C2 (s) = .
9.796s2 + 28.94s + 1
and the phase margin of C(s)P (s) is Φpm = 45◦ as desired. N
The MATLAB code below was used to design the controller in Example 9.5.1.

1 Phi pm des = 45;

2 s = tf('s'); P = 1/s/(s+1)/(s+20);
3 hatK = 200*(10ˆ(1/2)); % boost of 10dB
4 [Gm, Pm, Wpc, Wgc] = margin(hatK*P);
5
6 %lag design
7 [mag, phase, w] = bode(hatK*P);
8 Phi pm 2 = Phi pm des/2; ∆ = 5;
9 [¬, i] = min(abs(phase + 180 − Phi pm 2 − ∆)); % find freq. with desired phase
10 alpha2 = 1/mag(i);

Version 1.2.2816, September 1, 2020

 9.6. Loop shaping theory 208

11 T2 = 10/alpha2/w(i);
12 C2 = (alpha2*T2*s + 1)/(T2*s + 1);
13 [Gm,Pm,Wpc,Wgc] = margin(hatK*C2*P); % check if the design meets the spec
14
15 % lead design
16 phi max = Phi pm des − Pm; Wm = Wgc;
17 alpha1 = (1+sind(phi max))/(1 − sind(phi max));
18 T1 = 1/Wm/sqrt(alpha1);
19 C1 = (alpha1*T1*s + 1)/(T1*s + 1);
20
21 % overall controller
22 K = hatK/sqrt(alpha1);
23 C = K*C1*C2;
24 [Gm,Pm,Wpc,Wgc] = margin(C*P); % check if the design meets the spec

9.6 Loop shaping theory

Although loop shaping gives us a great deal of flexibility in designing the closed-loop response of a system, there
are certain fundamental limits on what can be achieved. In this section we look at some theoretical facts that
we have to keep in mind while designing controllers via loop shaping.

9.6.1 Bode’s phase formula

It is a fundamental fact that if L = P C is stable and minimum phase and normalized so that L(0) > 0 (positive
steady-state gain), then the magnitude Bode plot uniquely determines the phase Bode plot.

Assumption 9.6.1. The loop gain L(s) = C(s)P (s) of the system in Figure 9.1 is stable, minimum phase and
L(0) > 0. J

The exact formula relating the magnitude to the phase is rather complicated, and is derived using Cauchy’s
integral theorem [Wunsch, 2005]. Let

ω0 = any frequency
ω ω
u = normalized frequency = ln ω0 , i.e., eu = ω0

M (u) = normalized log magnitude = ln |L(jω0 eu )| = ln |L(jω)|.

W (u) = weighting function = ln coth u2 .

Recall that
cosh (x) ex + e−x
coth (x) = = x .
sinh (x) e − e−x
The phase formula (in radians) is
Z∞
1 dM (u)
](L(jω0 )) = W (u)du (Bode’s gain/phase formula). (9.13)
π du
−∞

The key point of this is that it shows that ](L(jω0 )) can be expressed as an integral involving |L(jω)|. What
it tells us is that for systems that satisfy Assumption 9.6.1, if one wishes to specify a certain magnitude
characteristic for the frequency response (cf. Figure 9.4), the phase characteristic is completely determined.

Version 1.2.2816, September 1, 2020

209 CHAPTER 9. INTRODUCTION TO CONTROL DESIGN IN THE FREQUENCY DOMAIN

π2
It turns out we may approximate the weighting function as W (u) ≈ 2 δ(u). Then the phase formula (9.13)
gives
π dM (u)
](L(jω0 )) ≈ . (9.14)
2 du u=0
As an example, consider the situation where
c
L(jω) = near ω = ω0 .
ωn
Thus the slope of the magnitude plot is −n per decade near ω = ω0 . Then
c
|L(jω0 eu )| =
ω0n enu
c
⇒ M (u) = ln |L(jω0 eu )| = ln − nu
ω0n
dM (u)
⇒ = −n
du
π
⇒ ](L(jω0 )) = −n from (9.14).
2
Thus we arrive at the observation: If the slope of |L(jω)| near crossover is −n, then arg(L(jω)) at crossover is
π
approximately −n .
2
What we learn from this observation is that in transforming |P | to |P C| via, say, lag or lead compensation,
we should not attempt to roll off |P C| too sharply near gain crossover. If we do, ]P C will be too large near
crossover, resulting in a negative phase margin and hence instability.

9.6.2 The waterbed effect

This concerns the ability to achieve the specification on the sensitivity function S shown in Figure 9.29 (cf.
Section 9.1).

Figure 9.29: Typical desired shape for a sensitivity function.

Let us suppose M > 1 and ω1 > 0 are fixed. Can we make ε arbitrarily small? That is, can we get arbitrarily
good tracking over a finite frequency range, while maintaining a given stability margin (1/M )? Or is there a
positive lower bound for ε? The answer is that arbitrarily good performance in this sense is achievable if and
only if P (s) is minimum phase. Thus, non-minimum phase plants have bounds on achievable performance: As
|S(jω)| is pushed down on one frequency range, it pops up somewhere else, like a waterbed.
To understand the result, first note that we can always factor a non-minimum phase TF G(s) as

G(s) = Gmp (s)A(s)

where Gmp is the minimum phase part of G and A is the non-minimum phase part. We normalize A so that
|A(jω)| = 1. The transfer function A is called all-pass because it has gain 1 at all frequencies.

Version 1.2.2816, September 1, 2020

 9.6. Loop shaping theory 210

Example 9.6.1. Recall the TF from Example 3.8.11

(1 − s)
G(s) = .
(s + 10)(s + 1)
In this case we can write   
1 1−s
G(s) = Gmp (s)A(s) = .
s + 10 1+s
N

Example 9.6.2. Consider a time-delay system

G(s) = e−sT , T > 0.
Use an all-pass approximation of the delay as we did in Section 7.3
− sT
2 +1
2
−s
G(s) = e−sT ≈ sT
= T
2 .
2 +1 T +s
This system is non-minimum phase with Gmp (s) = 1. So a long time-delay T acts like a slow non-minimum
phase zero at s = 2/T . N
Here’s the result:
Theorem 9.6.2. Suppose P (s) has a zero at s = z with Re (z) > 0. Let A(s) denote the all-pass factor of S(s).
Then there are positive constants c1 , c2 , depending only on ω1 and z, such that
c1 log ε + c2 log M ≥ log |A(z)−1 | ≥ 0.

Example 9.6.3.
1−s
P (s) = , p > 0, p 6= 1.
(s + 1)(s − p)
The plant has a non-minimum phase zero at s = 1. If C(s) = Nc /Dc is a stabilizing controller, then
1 (s + 1)(s − p)Dc
S(s) = = .
1 + PC (s + 1)(s − p)Dc + (1 − s)Nc
So S has a non-minimum phase zero at s = p and therefore (s − p)/(s + p) is an all-pass factor of S. There may
be other all-pass factors, so what we can say is that A(s) has the form
s−p
A(s) = A1 (s),
s+p
where A1 (s) is some all-pass TF (may be 1). Evaluate at the RHP zero of the plant
1−p
|A(1)| = |A1 (1)|.
1+p
Now |A1 (1)| ≤ 1 (why?), so
1−p
|A(1)| ≤
1+p
and hence
1+p
|A(1)−1 | ≥ .
1−p
The theorem gives
1+p
c1 log ε + c2 log M ≥ log .
1−p
Thus, if M > 1 is fixed, log ε cannot be arbitrarily negative, and hence ε cannot be arbitrarily small. In fact
the situation is much worse if p ≈ 1, that is, if the RHP plant pole and zero are close. N

Version 1.2.2816, September 1, 2020

211 CHAPTER 9. INTRODUCTION TO CONTROL DESIGN IN THE FREQUENCY DOMAIN

The following quote is from Gunter Stein in the inaugural IEEE Bode Lecture at the 1989 IEEE Conference
on Decision and Control describing the waterbed effect:

Sensitivity improvements in one frequency range must be paid for with sensitivity deteriorations in
another frequency range, and the price is higher if the plant is open-loop unstable. This applies to
every controller, no matter how it was designed.

Stein called this, the principle of Conservation of Sensitivity Dirt. This image refers to the area under
the sensitivity function being analogous to a pile of dirt. If one shovels dirt away from some area (i.e. re-
duces sensitivity in a frequency band), then it piles up somewhere else (i.e. the sensitivity increases at other
frequencies).

9.7 Summary
In this chapter we introduced two basic loop shaping design methods. Control design is a vast subject area
with many other techniques not covered here. Key ideas to extract from this chapter are listed below.

1. Understand how adjusting the shape of the sensitivity function’s Bode plot can lead to good performance.

2. Understand Figures 9.2 and 9.4.

3. Understand why we shape the Bode plot of L(s) instead of S(s) and the advantages and disadvantages
of doing so.

4. Know what a lag controller is and what it is used for. Especially know Figures 9.7 and 9.8.

5. Know how to design a lag controller to meet specs on (i) stability (ii) steady-state tracking (iii) percentage
overshoot (converted into a Φpm spec).

6. Know what a lead controller is and what it is used for. Especially know Figures 9.18 and 9.19.

7. Know how to design a lead controller to meet specs on (i) stability (ii) steady-state tracking or closed-loop
bandwidth (iii) percentage overshoot (converted into a Φpm spec).

8. Be able to use the loop-shaping ideas of this chapter to design PI and PD controllers.

9. Understand the performance limitations discussed in Section 9.6 and that non-minimum phase systems
have inherent performance limitations which no controller can overcome.

10. What is the waterbed effect?

Version 1.2.2816, September 1, 2020

Epilogue

If this is the only course you take on control then it is hoped that you learned the following.

• The value of block diagrams.

• What it means for a system to be linear and the central role that linear systems play in engineering.

• The importance of mathematical models for systematic design.

• Why we use transfer functions and the frequency domain.

• What stability means.

• What feedback is and why it is useful.

• What makes a system easy or hard to control.

If you plan to take another course in control, it is hoped that you additionally learned

• How to model a system using differential equations.

• How to put a differential equation into state-space form.

• How to linearize a nonlinear state-space model.

• How to check if a system is stable using poles or eigenvalues.

• The Nyquist criterion and the meaning of stability margins.

• How to design a simple feedback controller.

The sequels to this course are ECE488 Multivariable Control Systems, which develops the state-space method
in continuous time, and ECE481 Digital Control Systems, which treats computer control. There is also ECE486
Robot Dynamics and Control. The subject of control science and engineering is vast and this course can take
you in many directions.

212
Bibliography

[Anderson and Ye, 2019] Anderson, B. and Ye, M. (2019). Recent advances in the modelling and analysis of
opinion dynamics on influence networks. International Journal of Automation and Computing, 16(2):129–149.
2

[Association, 2015] Association, N. H. T. S. (2015). Fatality analysis reporting system. http://www-fars.

nhtsa.dot.gov/Main/index.aspx. Retrieved March 25 2015. 1

[Åström and Hägglund, 1995] Åström, K. and Hägglund, T. (1995). PID Controllers. Instrument Society of
America, 2nd edition. 150, 157, 162, 164

[Åström and Hägglund, 2006] Åström, K. and Hägglund, T. (2006). Advanced PID Control. The Instrumenta-
tion, Systems, and Automation Society. 150

[Åström and Murray, 2019] Åström, K. and Murray, R. (2019). Feedback systems: An introduction for scientists
and engineers. Princeton University Press, 2 edition. 7, 8, 150

[Basar and Basar, 2001] Basar, T. and Basar, T. (2001). Control Theory: Twenty-Five Seminal Papers (1932-
1981). Wiley-IEEE Control Systems Society. 166

[Bellman and Kalaba, 1964] Bellman, R. and Kalaba, R. (1964). Selected Papers on Mathematical Trends in
Control Theory. Dover. 166

[Bequette, 2005] Bequette, B. W. (2005). A critical assessment of algorithms and challenges in the development
of a closed-loop artificial pancreas. Diabetes Technology & Therapeutics, 7(1):28–47. 2

[Cannon Jr., 2003] Cannon Jr., R. (2003). Dynamics of Physical Systems. Dover Publications. 18

[Chen, 1999] Chen, C. (1999). Linear System Theory and Design. Oxford University Press, New York, third
edition. 63

[Cobelli et al., 2009] Cobelli, C., Man, C. D., Sparacino, G., Magni, L., Nicolao, G. D., and Kovatchev, B.
(2009). Diabetes: Models, signals, and control. IEEE Reviews in Biomedical Engineering, 2:54–96. 2

[Desborough and Miller, 2002] Desborough, L. and Miller, R. (2002). Increasing customer value of industrial
control performance monitoring–honeywell’s experience. In Sixth International Conference on Chemical Pro-
cess Control, volume 98 of AIChE Symposium Series Number 326. 150

[Dorf and Bishop, 2011] Dorf, R. and Bishop, R. (2011). Modern Control Systems. Prentice Hall, 12th edition.
59

[Doyle et al., 1990] Doyle, J., Francis, B., and Tannenbaum, A. (1990). Feedback Control Theory. Macmillan
Publishing Co. 1, 118

[Dullerud and Paganini, 2000] Dullerud, G. and Paganini, F. (2000). A Course in Robust Control Theory.
Springer. 92, 125

[Evans, 1948] Evans, W. (1948). Graphical analysis of control systems. Transactions of the American Institute
of Electrical Engineers, 67(1):547–551. 137

213
 BIBLIOGRAPHY 214

[Evans, 1950] Evans, W. (1950). Control system synthesis by root locus method. Transactions of the American
Institute of Electrical Engineers, 69(1):66–69. 137

[Filieri et al., 2015] Filieri, A., Hoffmann, H., and Maggio, M. (2015). Automated multi-objective control for
self-adaptive software design. In Proceedings of the 2015 10th Joint Meeting on Foundations of Software
Engineering, pages 13–24. 2

[Goodwin et al., 2001] Goodwin, G., Graebe, S., and Salgado, M. (2001). Control System Design. Prentice
Hall. 13, 150

[Gopal, 1963] Gopal, M. (1963). Control Systems: Principles and Design. McGraw-Hill. 27

[Hellerstein et al., 2004] Hellerstein, J., Diao, Y., Parekh, S., and Tilbury, D. M. (2004). Feedback Control of
Computing Systems. Wiley-IEEE Press, New York, U.S.A. 2, 8

[Jia et al., 2015] Jia, P., MirTabatabaei, A., Friedkin, N., and Bullo, F. (2015). Opinion dynamics and the
evolution of social power in influence networks. SIAM Review, 57(3):367–397. 2

[Kailath, 1980] Kailath, T. (1980). Linear Systems. Prentice-Hall. 13

[Keshav, 2012] Keshav, S. (2012). Mathematical Foundations of Computer Networking. Addison-Wesley. 8

[Kim and Kumar, 2012] Kim, K. D. and Kumar, P. R. (2012). Cyber physical systems: A perspective at the
centennial. Proceedings of the IEEE, 100:1287–1308. 2

[Lessard et al., 2016] Lessard, L., Recht, B., and Packard, A. (2016). Analysis and Design of Optimization
Algorithms via Integral Quadratic Constraints. SIAM Journal of Optimization, 26(1):57–95. 8

[Lorenz, 1963] Lorenz, E. N. (1963). Deterministic nonperiodic flow. Journal of the Atmospheric Sciences,
20:130–141. 14

[Lundberg et al., 2007] Lundberg, K., Miller, H., and Trumper, D. (2007). Initial conditions, generalized func-
tions, and the Laplace transform troubles at the origin. IEEE Control Systems Magazine, 27(1):22–35. 59

[MacColl, 1945] MacColl, L. (1945). Fundamental theory of servomechanisms. Van Nostrand. 166

[Ny and Pappas, 2014] Ny, J. L. and Pappas, G. J. (2014). Differentially private filtering. IEEE Transactions
on Automatic Control, 59(2):341–354. 2

[Nyquist, 1932] Nyquist, H. (1932). Regeneration theory. Bell Laboratories Technical Journal, 11:126–147. 166

[Patikirikorala et al., 2012] Patikirikorala, T., Colman, A., Han, J., and Wang, L. (2012). A systematic survey
on the design of self-adaptive software systems using control engineering approaches. In 7th International
Symposium on Software Engineering for Adaptive and Self-Managing Systems, pages 33–42. 2

[Pothukuchi et al., 2018] Pothukuchi, R. P., Pothukuchi, S. Y., Voulgaris, P., and Torrellas, J. (2018). Yukta:
multilayer resource controllers to maximize efficiency. In Proceedings of the 45th Annual International Sym-
posium on Computer Architecture, pages 505–518. 2

[Qiu and Zhou, 2010] Qiu, L. and Zhou, K. (2010). Introduction to Feedback Control. Prentice-Hall, Toronto.
126

[Routh, 1877] Routh, E. (1877). A Treatise on the Stability of a Given State of Motion. Macmillan. 126

[Samad, 2017] Samad, T. (2017). A survey on industry impact and challenges thereof. IEEE Control Systems,
37(1):17–18. 150

Version 1.2.2816, September 1, 2020

215 BIBLIOGRAPHY

[Seierstad and Sydsaeter, 1986] Seierstad, A. and Sydsaeter, K. (1986). Optimal control theory with economic
applications. Elsevier. 2

[Smith, 2015] Smith, B. W. (2015). Human error as a cause of vehicle crashes. http://cyberlaw.stanford.
edu/blog/2013/12/human-error-cause-vehicle-crashes. Retrieved March 25 2015. 1

[Society, 2015] Society, I. C. S. (2015). Control systems are ubiquitous. http://ieeecss.org/general/

control-systems-are-ubiquitous. Retrieved March 25 2015. 1

[Wunsch, 2005] Wunsch, A. (2005). Complex variables with applications. Pearson, Toronto, 3rd edition. 166,
208

[Ziegler and Nichols, 1942] Ziegler, J. and Nichols, N. (1942). Optimum settings for automatic controllers.
ASME Transactions, 64(8). 157

Version 1.2.2816, September 1, 2020

Index

All-pass, 208 Equilibrium point, 37

Amplitude response, 75 Exogenous inputs, 119
Argument, 91
Principle, 91 Feedforward, 153
Arrival angle, 139 Final-value theorem, 73
Asymptotes, 85, 138 Fourier transform, 45
Asymptotic Bode plot, 85 Frequency response, 75
Asymptotic stability, 67
Gain crossover frequency, 176
Autonomous system, 14
Gain margin, 181
Auxiliary polynomial, 128
Heat equation, 15
Bandwidth, 81
Hurwitz, 125
Block diagram, 16
Bode plot, 80 Impulse response, 64
Bounded, 68 Input-output stable, 122
Bounded-input bounded-output stable, 68 Integral gain, 150
Break frequency, 84 Integral time constant, 150
Internal stability, 120
Cartesian form, 90
Cascade connection, 51 Jordan form, 63
Centroid, 138
Characteristic polynomial, 123 Lag controller, 190
Complementary root-locus, 147 Laplace transform, 40
Complex conjugate, 90 Lead controller, 198
Complex number, 90 Lead-lag controller, 204
Controller, 2 Lower gain margin, 181
Convolution, 44 Lower phase margin, 177
Coprime, 49
Matrix exponential, 61
DC motor, 3, 25 Minimum phase, 80
back emf, 26 Minimum phase system, 79
rotor, 26
Nilpotent, 61
stator, 26
Non-dominant poles, 112
Decade, 83
Non-dominant zeros, 112
Decibels, 80
Non-minimum phase, 80
Departure angle, 139
Non-minimum phase system, 79
Derivative gain, 150
Nyquist contour, 167
Derivative time constant, 150
Nyquist plot, 167
Difference Equation, 14
Disturbance, 2 Open-loop control, 3
Dominant poles, 112 Operational amplifiers, 24
Dominant zeros, 112
Peak time, 107
Equilibrium configuration, 37 Percentage overshoot, 105

216
217 INDEX

Phase crossover frequency, 180 Well-posed, 120

Phase curve, 66
Phase margin, 177
Phase plot, 66
Phase response, 75
Phasors, 91
Plant, 2
Polar form, 91
Polar plot, 76
Pole placement, 158
Pole-zero cancellation, 124
Proportional gain, 150
Prototype first order system, 93
Time constant, 96
Prototype second order system, 97
Critically damped, 99
Damping ratio, 99
Natural undamped frequency, 99
Overdamped, 99
Undamped, 99
Underdamped, 99

Region of convergence, 40
Reset, 153
Reset time, 150
Rise time, 108
root-locus, 138

Schematic diagram, 15
Settling time, 107
Signals, 15
Stability margin, 184
State of a system, 30
State vector, 29
Steady-state gain, 73
Summing junctions, 16
System, 2

Tracking time constant, 163

Transfer function, 47
improper, 48
proper, 48
rational, 48
strictly proper, 48

Ultimate gain, 157

Ultimate period, 157
Unstable, 68
Unstable pole-zero cancellation, 124
Upper gain margin, 180
Upper phase margin, 177

Version 1.2.2816, September 1, 2020