Part IB Paper 6: Information Engineering

LINEAR SYSTEMS AND CONTROL

Ioannis Lestas

HANDOUT 1

“Signals, systems and feedback”

Signals

+ input output
⌃ Plant

Systems

Feedback
Controller

(Filled-in and unfilled version of notes on Moodle)

Colour convention: Text in red in the filled handout is text that is animated
during the lectures, and also does not appear in the unfilled handout.

1
The Aims of the course are to:
Introduce and motivate the use of feedback control systems.
Introduce analysis techniques for linear systems which are used in control, signal
processing, communications, and other branches of engineering.
Introduce the specification, analysis and design of feedback control systems.
Extend the ideas and techniques learnt in the IA Mechanical
Vibrations course.
By the end of the course students should:
Be able to develop and interpret block diagrams and transfer functions for
simple systems.
Be able to relate the time response of a system to its transfer function and/or
its poles.
Understand the term ‘stability’, its definition, and its relation to the poles of a
system.
Understand the term ‘frequency response’ (or ‘harmonic response’), and its
relation to the transfer function of a system.

Be able to interpret Bode and Nyquist diagrams, and to sketch them for simple
systems.

Understand the purpose of, and operation of, feedback systems.

Understand the purpose of proportional, integral, and derivative controller

elements, and of velocity feedback.

Possess a basic knowledge of how controller elements may be implemented

using operational amplifiers, software, or mechanical devices.

Be able to apply Nyquist’s stability theorem, to predict closed-loop stability

from open-loop Nyquist or Bode diagrams.

Be able to assess the quality of a given feedback system, as regards stability

margins and attenuation of uncertainty, using open-loop Bode and Nyquist
diagrams.

2
What the course is really about

Understanding complex systems as an interconnection of simpler

subsystems.

Relating the behaviour of the interconnected system to the behaviour of

the subsystems.

(but we’ll only consider the feedback interconnection in detail)

Part 1 (1st 7 lectures or so)

Part 2 (last 7 lectures or so)

+
⌃

3
SYLLABUS
Section numbers
Course material book 1 book 2
Examples of feedback control systems. Use of block 1.1-1.13 1.1-1.3
diagrams. Differential equation models. Meaning of 2.2-2.3 2.1-2.5
‘Linear System’.
Review of Laplace transforms. Transfer functions. Poles 2.4-2.6 3.1-3.2
(characteristic roots) and zeros. Impulse and step
responses. Convolution integral. Block diagrams of
complex systems.
Definition of stability. Pole locations and stability. Pole 6.1 3.3-3.6
locations and transient characteristics. 5.6
Frequency response (harmonic response). Nyquist (polar) 8.1-8.3 6.1-6.3
and Bode diagrams.

Terminology of feedback systems. Use of feedback to 4.1-4.2 4.1

reduce sensitivity. Disturbances and steady-state errors 4.4-4.5 3.1.6
in feedback systems. Final value theorem.
Proportional, integral, and derivative control. Velocity 7.7 4.3
(rate) feedback. Implementation of controllers in various 12.6
technologies.
Nyquist’s stability theorem. Predicting closed-loop 9.1-9.3 6.3
stability from open-loop Nyquist and Bode plots.
Performance of feedback systems: Stability margins, 8.5 6.4,6.6
Speed of response, Sensitivity reduction. 9.4-9.6 6.9
12.5
References

1. Dorf,R.C, and Bishop,R.H, Modern Control Systems, 10th ed.,

(Addison-Wesley), 2005.

2. Franklin,G.F, Powell,J.D, and Emami-Naeini,A, Feedback Control of Dynamic

Systems, 5th ed., (Addison-Wesley), 2006.
Notes based on earlier version by Prof Glenn Vinnicombe.

4
Contents

1 Signals, systems and feedback 1

1.1 Examples of feedback systems . . . . . . . . . . . . . . . . . . 6
1.1.1 Ktesibios’ Float Valve regulator . . . . . . . . . . . . . 6
1.1.2 Watt’s Governor . . . . . . . . . . . . . . . . . . . . . 8
1.1.3 A Helicopter Flight Control System . . . . . . . . . . . 10
1.1.4 Internet congestion control (TCP) . . . . . . . . . . . . 11
1.1.5 The lac operon – E.Coli . . . . . . . . . . . . . . . . . 12
1.2 Block Diagrams . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.2.1 What goes in the blocks? . . . . . . . . . . . . . . . . 13
1.2.2 Signals and systems . . . . . . . . . . . . . . . . . . . 14
1.2.3 ODE models – A circuits example . . . . . . . . . . . . 15
1.2.4 Block diagrams and the control engineer . . . . . . . . 16

1.3 Linear Systems . . . . . . . . . . . . . . . . . . . . . . . . . . 17

1.3.1 What is a “linear system” . . . . . . . . . . . . . . . . 17
1.3.2 Linearization . . . . . . . . . . . . . . . . . . . . . . . 20
1.3.3 When can we use linear systems theory? . . . . . . . . 22
1.4 Laplace Transforms . . . . . . . . . . . . . . . . . . . . . . . 23
1.5 Key points . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

5
1.1 Examples of feedback systems The Ktesibios’ float
valve regulator is the
oldest control system
1.1.1 Ktesibios’ Float Valve regulator that has been reported
(Water-clock, Alexandria 250BC) in the literature. It
goes back to ancient
Supply Needs a Alexandria and it is a
qi feedback system that
constant flow regulates the outflow
rate at “E” rate q0 from a chamber
to a constant value
x Outflow (was used by ancient
E qo Egypticians as a water
clock!).
is a feedback control system.
In particular, if there
is a disturbance in the
supply rate qi , e.g. de-
creasing the water level
x, then the float in the
Block Diagram:
chamber would move
downwards, allowing
more water to flow in
outflow rate and thus recover x to
Net “qo ” its desired level (simi-
larly if the disturbance
Inflow Water increases x).
Rate Float level Orifice
⌃
“q” chamber “x” “E” A block diagram of the
+ float valve regulator is
illustrated on the left.
Each block has inputs
Supply and outputs which are
flow rate Float Supply Pressure signals (i.e. functions
& valve (Disturbance) of time). The block
“qi ” itself corresponds to a
system, i.e. a differen-
Signals have units (usually), are functions of time, and are represented by the tial equation the relates
the inputs and outputs.
connections:
e.g. Net inflow “q(t)” is measured in m3 /s It should be noted that
Water level “x(t)” is measured in m the block diagram is
not unique as each
block could be decom-
posed further into other
6 sublocks.

6
Systems have equations, and are represented by the blocks:
The equation on the
e.g. the Float chamber is described by left is associated with
Z t the system represented
1 by the "float chamber"
x(t) = q(⌧ ) d⌧ block. In particular, it
A 0 relates the output x(t)
cross-sectional area (water level) with the
input q(t) (rate of water
flow into chamber).
It should be noted that
an equivalent way to
write this equation is
1
ẋ(t) = q(t)
A

7
1.1.2 Watt’s Governor The Watt’s Governor
is a control system
that is a central
part of Watt’s steam
engine, an invention
that was at the core
of the industrial
revolution.

Its aim is to ensure

automatically that
there is a sufficient
flow of steam to the
engine as the load
applied to the engine
varies (see also next
page).

8
Watt’s Governor
The operation of the
Watt’s Governor is
based on a flyweight
mechanism that in-
crease/decreases the
opening of the steam
valve as the speed
of the engine de-
creases/increases. The
latter occurs due to
a change in the load
applied.

A block diagram repre-

sentation of the Watt’s
Governor is provided on
the left. As also men-
tioned in the previous
Is a feedback control system.
example described, each
Block diagram: block corresponds to a
system with inputs and
outputs that are sig-
Load torque nals.

Steam Engine Engine

pressure Engine torque + speed
& Engine
⌃
inertia
throttle
Net
torque
butterfly Pulley
angle
Fly-ball
Linkage
fly-ball Dynamics fly-ball
angle angular
velocity
Note: it would be wrong to label the input to the feedback system as
simply “steam” rather than “steam pressure”. Steam in itself is not a
quantity (although its pressure, temperature or flow rate is).

9
1.1.3 A Helicopter Flight Control System
Designing the flight
control system of a
helicopter is in general
a non-trivial problem
as heicopters are, so
called, open loop un-
stable systems, i.e. the
helicopter will crash if
its controllers are not
in operation.

As illustrated in the
Is a feedback control system
block diagram the flight
control policy will be
implemented by a com-
puter, hence an ADC is
needed to convert the
analogue measurements
to digital signals to be
Block Diagram:
processed by the flight
control computer. A
DAC will then convert
controls wind gusts measurements the digital outputs of
(inputs) (disturbances) (outputs) the computer to appro-
main rotor priate analogue signals
8 collective for the actuators that
> vertical accn 9
> determine the control
>
> >
>
>
> > inputs.
>
>
<
cyclic pitch Helicopter pitch rate > >
>
=
cyclic roll dynamics roll rate It should be noted that
>
> >
>
>
> >
> external disturbances
>
> tail rotor yaw rate >
>
>
: >
; can also affect the
collective dynamic behaviour of
pilot the helicopter.

actuators commands sensors

flight
DAC control ADC
computer

10

10
1.1.4 Internet congestion control (TCP) Internet congestion
control is a protocol
x1 that determines the
src 1 link A dst 1 rates with which
computers/devices
send data, based on
acks congestion signals
they receive from
link B the network.

Is a feedback control system

– in fact, the largest man made one in the world.
(in reality, of course, there are many source/destination pairs competing for
bandwidth over many links)
Note: This is NOT a block diagram
– it shows the flow of “stuff” (in this case packets) not information.

Files to be transferred across the Internet using the Transmission Control Protocol
(TCP) - eg a download from the web - are broken into packets of size typically around
1500bytes, with headers specifying the destination and the number of the packet
amongst other information. These packets are sent one by one into the network, with
the recipient sending acknowledgements back to the source whenever one is received.
Routers in the network typically operate a drop tail queue. If a packet is received when
the queue is full then it is simply discarded. Packet loss thus indicates congestion. If a
packet is received out of order, it is assumed that intervening packets have been lost.
The recipient sends a duplicate acknowledgement to signal this and the source lowers
its rate (in response to the congestion) and resends the lost packet(s). Whilst a steady
stream of successive acknowledgements is being received the source gradually increases
its sending rate. In normal operation sources are thus constantly increasing and
decreasing their rates in an attempt to make use of the available bandwidth.
Congestion (ie full queues and the resulting packet loss) can occurr anywhere in the
network - at the edges (eg your adsl modem, or at the exchange), in the core (eg a big
transatlantic link) or, very often, at peering points, which are the connections between
the networks that make up the Internet.

11

11
1.1.5 The lac operon – E.Coli (⇡ 130 million years Feedback control is
also ubiquitous in
BC!) biological processes.
A significant part
of the DNA has
a regulatory role,
controlling which
chemical reactions
take place and at
what rates. This is
crucial for the various
functionalities of the
cells.

(Only read this if you’re interested!) The diagram above illustrates 4 genes and some
control regions along the DNA of E.Coli. E.Coli’s favourite sugar is glucose, but it will
quite happily “eat” lactose if there’s no glucose around. If there is glucose around or if
there is no lactose around then there is no need to produce -galactosidase (the
enzyme which breaks down lactose, first into allolactose and then glucose) or the
permease (which transports lactose into the cell). In addition, when it is metabolising
lactose, it wants to regulate the amount of enzyme production to match the available
lactose. This is the control system which achieves this: The lacI gene codes for a
protein (the repressor) which binds to the operator (O) and stops the lacZ,Y and A
genes being transcribed (ie “read”). If there’s lactose in the cell, and at least some
-galactosidase, then there will also be allolactose (the inducer). In this case the
repressor binds with it instead, and falls off the DNA. In the absence of glucose, the
cAMP/CRP complex binds at the promoter (P), this encourages RNA polymerase to
bind and initiate transcription of lacZ,Y and A.

– for more details, see 3G1 next year . . .

12

12
1.2 Block Diagrams The figures on the
left illustrates various
1.2.1 What goes in the blocks? examples of systems
and the corresponding
Some of them act like “amplifiers” or “attenuators” equations that describe
1 them.
a= ⇥F
M ass In the first block the
F 1 a output is equal to the
input multiplied by a
Force Mass Acceleration constant.

“Gain” In the second block the

output is the integral of
But many are dynamic processes described by Ordinary Differential Equations the input.
(ODEs).
The third block corre-
sponds to an op-amp
circuit where the input
and output voltage are
related via an analogous
differential equation.
Z
a v
dt v̇ = a
Acceleration Velocity

ie Cdg
i
C
Vi
Vi
R
+
Vo
Vo Voltage
if V̇o = 1
RC Vi
Voltage
I i

(We shall (later) describe these by transfer functions.)

Note: By drawing this circuit as a block, we are implicitly assuming that any
current it draws has negligible effect on the preceding block and that the
following block draws insignificant current from it (i.e. that R is large and the
op-amp is close to ideal).

13

13
1.2.2 Signals and systems

Block diagrams represent the flow of information, not the flow of “stuff”.
(taking a numeric value as a function of time)
Blocks represent “systems ”, whose inputs and outputs are “signals ”.
(equations mapping inputs into outputs)
This is NOT a block diagram (in our sense)

This IS a block diagram

14

14
1.2.3 ODE models – A circuits example
A more involved system
r
i L Ldf described by an electri-
cal circuit is illustrated
on the left.
x C R y
The voltage difference x
T is the input of the sys-
i cdog tem and the voltage dif-
ference y is the output.
A differential equation
relating x and y is de-
di rived using the stan-
x y=L
dt dard differential equa-
tions satisfied by the
y
i = C ẏ + voltage and current of
R an inductor and capac-
=) ✓ ◆ itor, respectively (part
ẏ IA).
x y = L C ÿ +
R

which gives a 2nd-order linear Ordinary Differential Equation:

=)
L
LC ÿ + ẏ + y = x
R

x L y
LC ÿ + R ẏ + y = x

15

15
1.2.4 Block diagrams and the control engineer

For the control engineer:

Some blocks are given (fixed)
eg

Steam Engine Dynamics

Aircraft Dynamics
(the “plant”)
while other blocks are to be designed
e.g.
Geometry of fly-ball mechanism in Watt governor.
The program in an aircraft’s flight control computer.
(the “controller”)

16

16
1.3 Linear Systems A system is linear if it
satisfies the principle of
1.3.1 What is a “linear system” superposition, i.e. if in-
put u1 gives output y1 ,
Consider a “system” f mapping dynamic inputs u into outputs y and input u2 gives out-
put y2 , then the input
u(t) y(t) u1 + u2 will have as out-
f put y1 + y2 .

⇣ ⌘
y=f u
the “system” f is linear if superposition holds, that is, if
⇣ ⌘ ⇣ ⌘ ⇣ ⌘
f u1 + f u2 = f u1 + u2
| {z } | {z }
for any u1 and u2 .

In terms of block diagrams. If f is a linear system,

u1 y1 u1
f
+ +
=
⌃ ⌃ f
+ +
u2 y2 u2
f

17

17
 ⇣ ⌘ ⇣ ⌘
In particular, f 2u = 2 f u , eg The principle of super-
position is illustrated
1 by the diagram on
the left. Here the
Linear 0 input is multiplied by
0 1 2 3 4 2, hence for a linear
1 system the output is
also multiplied by the
0 same factor. If this is
0 1 2 3 4 not the case then the
system is non-linear.
Non-linear 1

0
0 1 2 3 4

The notions of causality

In addition, we shall also assume that all systems are: and time invariance
are also introduced
on the left. These
causal – the output at time T , y(T ), depends only on the input up to properties will be as-
time t (ie y(t), t  T is independent of u(t), t T ). sumed throughout the
response course and will not be
here is discussed further.

1 independent 1 In particular, causality

of input here is satisfied by all physi-
0 0 cal systems. It captures
0 1 2 3 4 0 1 2 3 4 the property that "the
time-invariant – the response of the system to a particular input doesn’t current output depends
depend on when that input is applied, ie if only on past inputs and
not on future inputs".
u(t) ! y(t), then u(t ⌧ ) ! y(t ⌧)
for any ⌧ . Time invariance can be
⌧ ⌧ violated in some ad-
u(t) u(t ⌧) y(t) y(t ⌧) vanced models (e.g. a
1 1 fault occurs on an air-
craft). These models
0 0 are though much more
0 1 2 3 4 0 1 2 3 4 difficult to analyse and
will not be covered in
this course.
18

18
Note that, if a system can be described by linear differential equations with
constant coefficients, and possibly delays, then it is necessarily linear (and
time invariant).
For example:
d2 x(t) du(t)
+ x(t T) = + 2u(t)
dt2 dt
describes a linear system, as if
d2 x1 (t) du1 (t)
+ x1 (t T) = + 2u1 (t)
dt2 dt
and
d2 x2 (t) du2 (t)
+ x2 (t T) = + 2u2 (t)
dt2 dt
then

d2
x1 (t) + x2 (t) + x1 (t T ) + x2 (t T)
dt2
d
= u (t) + u2 (t) + 2 u1 (t) + u2 (t)
dt 1

(because d
dt
u1 (t) + u2 (t) = du1
dt
+ du2
dt
)
which is just the superposition of solutions. If there are x2 terms or sin(x)
terms, for example, then this doesn’t work.
Almost all the linear systems considered in this course will be of this form,
although it is convenient to develop the theory in more generality.

19

19
1.3.2 Linearization

All real systems are actually nonlinear, but many of these behave
approximately linearly for small perturbations from equilibrium.
e.g. Pendulum:

F l cos ✓ + mlg sin ✓ = ml2 ✓¨ ✓ l

But, for small ✓

F
F l + mlg✓ ⇡ ml2 ✓¨ mg
or
l✓¨ + g✓ ⇡ F/m which is a linear ODE

General case
Suppose a system is described by an ODE of the form

ẋ = f (x, u)

where f is a smooth function. Assume that this system has an equilibrium at

(x0 , u0 ), by which we mean that

f (x0, u0 ) = 0.

where x0 and u0 are constants.

20

20
 Let x = x0 + x, u = u0 + u,
and use a Taylor series expansion to obtain:
0
ẋ0 + ˙x = f (x0 + x, u0 + u)
0 neglect
just a
@f @f
= f (x0, u0 ) + x+ u+ higher order
constant @x x ,u
| {z0 0}
@u x ,u
| {z0 0}
terms
A B
which results in the linear ODE

˙x = A x + B u

This is a simple example of a state-space model. This procedure can be

generalized to higher order systems with many inputs and outputs - see 3F2
next year.

21

21
1.3.3 When can we use linear systems theory?

Linearity is often desirable:

Hi-Fi audio system (non-linearities are called distortion).
Aircraft fly-by-wire system (for predictable response)
If we are going to design a controller to keep a system near equilibrium then
we can ensure that perturbations are small (and hence that behaviour is
approximately linear). This justifies the use of linear theory for the design!
– so linear systems theory is often very useful even when the underlying
systems are actually nonlinear

However: some systems are designed to behave

nonlinearly:
Switch or relay (because it is either on or off).
Automated air traffic control system.
(either have a collision or not) .
In such cases linear theory is of little use in itself.

22

22
1.4 Laplace Transforms

Laplace transforms are an extremely convenient tool for the analysis of linear,
time-invariant, causal systems. We shall now briefly review some pertinent facts that
you learnt at Part IA and introduce some new ideas.

DEFINITION:

Z 1
st
ȳ(s) = y(t)e dt
0

(provided the integral converges for sufficienty large and positive values of s.)

Note, a Laplace transform

is NOT a function of t

IS a function of s.

Various notations:

Z 1
st
L {y(t)} = Ly = ȳ(s) = y(t)e dt
0

Notation for the inverse transform:

1
y(t) = L ȳ(s)

EXAMPLES

Find ȳ(s) if y(t) = C (a constant )

Z 1  st 1
st e C
ȳ(s) = Ce dt = C = (taking Real(s) > 0 ).
0 s 0 s

Find ȳ(s) if y(t) = e at

Z " #1
1 (s+a)t
(s+a)t e 1
ȳ(s) = e dt = = (taking Real(s) > a ).
0 (s + a) s+a
0

23

23
Addition or Superposition property

If
y(t) = Ay1 (t) + By2 (t)
then
ȳ(s) = Aȳ1 (s) + B ȳ2 (s)

(A, B constants)
Proof:
Z 1
st
ȳ = (Ay1 + By2 )e dt
0
Z 1 Z 1
st st
= A y1 e dt + B y2 e dt
0 0

= Aȳ1 + B ȳ2

=) The operation of taking a Laplace transform is linear.

Transforms of derivatives
Z 1
dy st
Lẏ(t) = e dt
0 dt
Z 1
⇥ ⇤
st 1 st
= y(t)e 0
+s y(t)e dt
0

= sȳ y(0)

Z 1
d2 y st
Lÿ = e dt
0 dt2
 1 Z 1
dy st dy st
= e +s e dt
dt 0 0 dt

= ẏ(0) + s(sȳ y(0))

= s2 ȳ sy(0) ẏ(0)

24

24
Obvious pattern:

Ly = ȳ

Lẏ = sȳ y(0)

Lÿ = s2 ȳ sy(0) ẏ(0)

.. .. ..
. . .

dn y
L = sn ȳ sn 1 y(0) sn 2
ẏ(0)
dtn ⇣ n 1 ⌘
d y
... dtn 1
(0)

In particular, if y(0) = ẏ(0) = ÿ(0) = · · · = 0, then

Ly = ȳ

Lẏ = sȳ

Lÿ = s2 ȳ
.. .. ..
. . .

dn y
L = sn ȳ
dtn

differentiation (in the time domain) corresponds

to multiplication by s (in the s domain)

25

25
Laplace Transform of tn
tn
Define ȳn (s) = L .
n!

Z 1 n
t st
ȳn = e dt
0 n!
 1 Z 1
1 tn st 1 ntn 1 st
= e + e dt
s n! 0 s 0 n!
Z 1 n 1
1 t st
= e dt
s 0 (n 1)!

1
= ȳn 1,
s

(since for Real(s) > 0, and as t ! 1, then e st ! 0 faster than tn ! 1).

Thus we have
1
ȳ0 = L 1 =
s

1
ȳ1 = L t =
s2

t2 1
ȳ2 = L = 3
2 s

t3 1
ȳ3 = L = 4
3⇥2 s

tn 1
Similarly ȳn = L = n+1
n! s

integration (in the time domain) corresponds to

division by s (in the s domain)

26

26
Poles and Zeros

Suppose G(s) is a rational function of s, by which we mean

n(s)
G(s) =
d(s)
where n(s) and d(s) are polynomials in s.

Then the roots of n(s) are called the zeros of G(s)

and the roots of d(s) are called the poles of G(s)

Example:

4s2 8s 60
G(s) =
s3 + 2s2 + 2s
4(s + 3)(s 5)
=
s(s + 1 + j)(s + 1 j)

Zeros of G(s) are 3, +5.

Poles of G(s) are 1 j, 1 + j, 0

Imag(s)

X
Real(s)
X
3 5
X

X – denote poles
– denote zeros

27

27
 time functions s-plane pole positions

Im

2
Re

Time functions and pole positions for y(t) = t and y(t) = t2

Laplace Transforms of Sines and Cosines

y = ei!t = cos !t + i sin !t

1
ȳ = = L cos !t + iL sin !t
s i!
s + i!
=
s2 + ! 2

s
Equating reals : L cos !t =
s2 + !2
!
and similarly : L sin !t =
s2 + ! 2
poles at s = ±!i in both cases

NOTE: Results like this are tabulated in the Maths and Electrical Data Books.

28

28
Shift in s theorem

If Ly(t) = ȳ(s)

then Leat y(t) = ȳ(s a).

Proof:
Z 1
at (s a)t
Le y(t) = e y(t) dt
0
= ȳ(s a),

Example of use: poles at s = 1 + 10j, 1 10j

1 20 1 20
L = L
s2 + 2s + 101 (s + 1)2 + 100

t
= 2e sin 10t
1 10
because L = sin 10t
s2 + 100

time function s-plane pole positions

Im
10i

1 -1 Re
t

t
Time functions and pole positions for y(t) = 2e sin 10t

29

29
Initial and Final Value Theorems The Final Value The-
orem is often used in
If ȳ(s) = L y(t) then whenever the indicated limits exist we have control system design
Final Value Theorem: and will be discussed
extensively later
within the course.
lim y(t) = lim sȳ(s)
t!1 s!0
In particular, it
Initial Value Theorem: allows to quantify
the behaviour of a
lim y(t) = lim sȳ(s) signal as time tends
t!0+ s!1 to infinity, from its
Laplace transform.
Proofs omitted (as it’s a little tricky to prove these properly.)
It should be noted
However, for rational functions of s it is easy to demonstrate that these relationships that this theorem
hold: cannot be used for
signals that do not
Let a partial fraction of ȳ(s) be given as:
tend to a constant
b0 X bi
n X n value as t ! 1 (e.g.
ai t
ȳ(s) = + and so y(t) = b0 + bi e . sinusoids, or signals
s s + ai
i=1 i=1 that tend to infinity).

Hence
n
X
y(0) = b0 + bi and, provided ai > 0, y(1) = b0 .
i=1

On the other hand,

n
X sbi
sȳ(s) = b0 +
s + ai
i=1
hence
n
X
sȳ(s)|s=1 = b0 + bi and, provided ai 6= 0, sȳ(s)|s=0 = b0
i=1

which are the same expressions as above.

30

30
1.5 Key points

Feedback is used to reduce sensitivity.

We use block diagrams to represent feedback
interconnections.
Each block represents a “system”.
Each connection carries a “signal”.
We shall assume that systems are described by linear,
time-invariant and causal ODE’s.

We distinguish between causes (the input signals) and effects (the output
signals).

Large and complex systems can be constructed by connecting together

simpler sub-systems.

Laplace transforms are central to the study of linear, time-invariant

systems.

31

31
 Part IB Paper 6: Information Engineering
LINEAR SYSTEMS AND CONTROL
Ioannis Lestas

HANDOUT 2

“Impulse responses, step responses and transfer

functions.”

transfer function

ū(s) G(s) ȳ(s)= G(s)ū(s)

*
)
*
)

*
)

Z t
u(t) g(t) y(t)= u(⌧ )g(t ⌧ ) d⌧
0
= u(t) ⇤ g(t)
= g(t) ⇤ u(t)
Laplace transform
pair impulse response

1
Summary

The

impulse response,
step response and

transfer function
of a Linear, Time Invariant and causal (LTI) system each completely
characterize the input-output properties of that system.

Given the input to an LTI system, the output can be deterermined:

In the time domain: as the convolution of the impulse response and the
input.
In the Laplace domain: as the multiplication of the transfer function and
the Laplace transform of the input.

They are related as follows:

The step response is the integral of the impulse response.

The transfer function is the Laplace transform of the impulse response.

2
Contents

2 Impulse responses, step responses and transfer functions. 1

2.1 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.1.1 Definition of the impulse “function” . . . . . . . . . . . 4
2.1.2 Properties of the impulse “function” . . . . . . . . . . . 5
2.1.3 The Impulse and Step Responses . . . . . . . . . . . . 6
2.2 The convolution integral . . . . . . . . . . . . . . . . . . . . . 7
2.2.1 Direct derivation of the convolution integral . . . . . . 7
2.2.2 Alternative statements of the convolution integral . . . 8
2.3 The Transfer Function (for ODE systems) . . . . . . . . . . . . 9
2.3.1 Laplace transform of the convolution integral . . . . . . 10
2.4 The transfer function for any linear system . . . . . . . . . . . 11
2.5 Example: DC motor . . . . . . . . . . . . . . . . . . . . . . . 13

2.5.1 Impulse Response of the DC motor . . . . . . . . . . . 15

2.5.2 Step Response of the DC motor . . . . . . . . . . . . . 16
2.5.3 Deriving the step response from the impulse response . 17
2.6 Transforms of signals vs Transfer functions of systems . . . . . 17
2.7 Interconnections of LTI systems . . . . . . . . . . . . . . . . . 18
2.7.1 “Simplification” of block diagrams . . . . . . . . . . . . 19
2.8 More transfer function examples . . . . . . . . . . . . . . . . . 20
2.9 Key Points . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

3
2.1 Preliminaries The impulse function
is the limit of a square
or triangular pulse that
2.1.1 Definition of the impulse “function” has area equal to 1
and is zero everywhere
The impulse can be defined in many different ways, for example: apart within an interval
1 1 of width that tends
to 0.

As we will see later in

a) or b) as ! 0. this handout it is an im-
T t T t
portant concept in the
2 study of linear systems.

(area= 1 in each case.)

taking “limits”,
as ! 0

= (t T)

impulse at t = T
T t (impulse occurs
when argument= 0)

Note: The “convergence”, as ! 0, occurs in the sense that both functions

a) and b) have the same properties in the limit.

4
2.1.2 Properties of the impulse “function” The impulse function
(t) satisfies the prop-
Consider a continuous function f (t), and let h (t) denote the pulse erty
approximation to the impulse ( (a) on previous page): Z 1
8 f (t) (t T ) dt = f (T )
<1/ if /2 < t < /2 1
h (t) =
:0 otherwise This is shown on the left
by considering the inte-
f (t) gral above when (t T )
1 is replaced with a pulse
h (t T) h (t T ) of width
and area 1. Upper and
lower bounds to this in-
T t tegral are obtained and
it is then shown that
Let
when ! 0 the inte-
Z 1 Z T+ Z T+ gral tends to f (T ).
2 1 1 2
I = f (t)h (t T ) dt = f (t) dt = f (t) dt
1 T 2 T 2

Now,
1
I  = ⇥ ⇥ max f (t)
T /2tT + /2
1
⇥

T t
and also
1
I = ⇥ ⇥ min f (t)
1 T /2tT + /2
⇥

T t

So, it must be the case that lim !0 I = f (T ), that is

Z 1
lim f (t)h (t T ) dt =f (T )
!0 1

5
A similar argument leads to the same result for the triangular approximation
to the impulse ( defined above as (b) ).
Formally, the unit impulse is defined as any “function” (t) which has the
property Z 1
f (t) (t T ) dt = f (T )
1

(Strictly speaking, (t) is a distribution rather than a function.)

Laplace transform of (t):

Z 1
L (t)} = (t)e st dt= e s⇥0 = 1
0

2.1.3 The Impulse and Step Responses

Definition: The impulse response of a system is the output of the
system when the input is an impulse, (t), and all initial conditions are
zero.

Definition: The step response of a system is the output of the system

when the input is a step, H(t), and all initial conditions are zero.
where H(t) is the unit step function
8
<1 if t 0
H(t) =
:0 if t < 0

If you know the impulse response of a system, then the response of that
system to any input can be determined using convolution, as we shall now
show:

6
2.2 The convolution integral The slide on the left
shows that the output
y(t) of a linear system
2.2.1 Direct derivation of the convolution integral for a given input u(t),
is equal to the convo-
lution of the impulse
u(t) LTI system y(t) response g(t) of the
(impulse response system with u(t).
= g(t))
This is a very impor-
tant result as it shows
that the impulse re-
INPUT OUTPUT sponse of a linear sys-
tem fully characterizes
the system. This is in
the sense that it allows
impulse response
to deduce the output
(t) g(t) y(t) for any input u(t).

(t ⌧) g(t ⌧)

u(⌧ ) (t ⌧) u(⌧ )g(t ⌧)

Z 1 Z 1
u(⌧ ) (t ⌧ ) d⌧ u(⌧ )g(t ⌧ ) d⌧
1 1
| {z }
u(t)

Hence, the response to the input u(t) is given by:

Z 1
y(t) = u(⌧ )g(t ⌧ )d⌧
1

7
2.2.2 Alternative statements of the
convolution integral

The convolution integral

Z 1
y(t) = u(⌧ )g(t ⌧ )d⌧
1

is abbreviated as
y(t) = u(t) ⇤ g(t)
Let T = t ⌧ , so ⌧ = t T and d⌧ = dT
It follows that

Z 1
y(t) = u(t T )g(T ) dT
1

= g(t) ⇤ u(t)

NOTE: in either form of the convolution integral,

the arguments of u(·) and g(·) add up to t
It is shown on the left
If, in addition,
how the convolution in-
g(t) = 0 for t < 0 (CAUSALITY) tegral can be simplified.
In particular, we exploit
u(t) = 0 for t < 0 (standing assumption), then the fact that the im-
as g(t ⌧ ) = 0 for t ⌧ < 0 (i.e. ⌧ > t) pulse response g(t) is 0
Z t for t < 0 (due to causal-
ity). Furthermore, we
y(t) = u(⌧ )g(t ⌧ ) d⌧
0 will use throughout the
course the convention
as u(⌧ ) = 0 for ⌧ < 0 that u(t) = 0 for t <
and also 0. These two properties
imply that the convolu-
as u(t ⌧ ) = 0 for t ⌧ < 0 (i.e. ⌧ > t) tion integral only needs
Z t to be evaluated from 0
y(t) = u(t ⌧ )g(⌧ ) d⌧ to t (rather than from
0 1 to 1).
as g(⌧ ) = 0 for ⌧ < 0

8
2.3 The Transfer Function (for ODE systems)

As an alternative to convolution, the response of a linear system to arbitrary

inputs can be determined using Laplace transforms. This is clear when the
system is described as a linear ODE:
For example, if a linear system has input u and output y satisfying the ODE

d2 y dy du
+ ↵ + y = a + bu
dt2 dt dt

and if all initial conditions are zero, i.e. dy

dt t=0 = y(0) = u(0) = 0, then
taking Laplace transforms gives

s2 ȳ(s) + ↵sȳ(s) + ȳ(s) = a sū(s) + bū(s)

or ⇣ ⌘ ⇣ ⌘
s2 + ↵s + ȳ(s) = a s + b ū(s)

and so
as + b
ȳ(s) = ū(s)
s2 + ↵s +
| {z }
Transfer function
The function 2a s+b is called the transfer function from ū(s) (the input) to
s +↵s+
ȳ(s) (the output).
Clearly the same technique will work for higher order linear ordinary
differential equations (with constant coefficients). For such systems, the
transfer function can be regarded as a placeholder for the coefficients of the
differential equation.

9
2.3.1 Laplace transform of the convolution integral

We have seen that both convolution with the impulse reponse (in the time
domain) and multiplication by the transfer function (in the Laplace domain)
can both be used to determine the output of a linear system. What is the
relationship between these techniques?
Assume that g(t) = u(t) = 0 for t < 0,

⇣ ⌘ ✓Z 1 ◆
L g(t) ⇤ u(t) = L g(⌧ )u(t ⌧ ) d⌧
1
Z 1 Z 1
= e st g(⌧ )u(t ⌧ ) d⌧ dt
0 1
Z 1 Z 1
= e st g(⌧ )u(t ⌧ ) dt d⌧
1 0

Note that ⌧ is constant for the inner integration, and let

T = t ⌧, which implies t = T + ⌧ and dt = dT ,
giving (so e st = e sT e s⌧ )
⇣ ⌘ Z 1 Z 1
L g(t) ⇤ u(t) = 0 0 e s⌧ e sT g(⌧ )u(T ) dT d⌧
1 ⌧
Z 1 Z 1
s⌧
= e g(⌧ )d⌧ ⇥ e sT u(T )dT
0 0
⇣ ⌘ ⇣ ⌘
= L g(t) ⇥ L u(t) = ḡ(s) ū(s)

In words, a Laplace transform (easy when you get used to them) turns a
convolution (always hard!) into a multiplication (very easy).

10

10
2.4 The transfer function for any linear
system

It follows from the result of the previous section that, if

y(t) = g(t) ⇤ u(t)

is the response of an LTI system with impulse response g(t) to the input u(t),
then we can also write

ȳ(s) = G(s) ū(s)

where G(s) = Lg(t) is called the transfer function of the system.
It follows that all LTI systems have transfer functions (given by the Laplace
transform of their impulse response).

transfer function

ū(s) G(s) ȳ(s)= G(s)ū(s)

*
)
*
)

*
)

u(t) g(t) y(t)= g(t) ⇤ u(t)

impulse response

We shall use the notation x̄(s) to represent the Laplace transform of a signal
x(t), and uppercase characters to represent transfer functions (e.g. G(s)).

11

11
In general a system may have more than one input. In this case, the transfer
function from a particular input to a particular output is defined as the
Laplace transform of that output when an impulse is applied to the given
input, all other inputs are zero and all initial conditions are zero.
This is most easily seen in the Laplace domain: If an LTI system has an input
u and an output y then we can always write

ȳ(s) = G(s)ū(s) + other terms independent of u.

G(s) is then called the transfer function from ū(s) to ȳ(s). (Or, the transfer
function relating ȳ(s) and ū(s))
Here the “other terms” could be a result of non-zero initial conditions or of
other non-zero inputs (disturbances, for example).
NOTE: Remember that although the transfer function is
defined in terms of the impulse response, it is usually
most easily calculated directly from the system’s differential
equations.

12

12
2.5 Example: DC motor

The following example will illustrate the impulse and step responses and the
transfer function. We take the input to the motor to be the applied voltage
e(t), and the output to be the shaft angular velocity
˙
! := ✓(t).
R i L

J B
e(t)

K!(t)
⌧ ✓

1) ⌧ (t) = Ki(t) Linearized motor equation

2) ⌧ (t) B!(t) = J !(t)

˙ Newton

di(t)
3) e(t) = Ri(t) + L + K!(t) Kirchoff
dt
Find the effect of e(t) on !(t):
Take Laplace Transforms, assuming that !(0) = i(0) = 0:

1) ⌧¯(s) = K ī(s)
2) ⌧¯(s) B!¯ (s) = J s¯!(s) !(0)
⇣ ⌘
3) ē(s) = Rī(s) + L sī(s) i(0) + K !¯ (s)

13

13
We can now eliminate ī(s) and ⌧¯(s) to leave one equation relating ē(s) and
!
¯ (s):
First 1) and 2) give:
K ī(s) = (Js + B)¯
! (s)
and rearranging 3) gives:

ē(s) = (Ls + R)ī(s) + K !

¯ (s)

putting these together leads to

✓ ◆
(Js + B)
ē(s) = (Ls + R) +K !¯ (s)
K
or, in the standard form (outputs on the left)
K
!
¯ (s) = ē(s)
(Ls + R)(Js + B) + K 2
k
= ē(s)
(sT1 + 1)(sT2 + 1)
for suitable definitions of k, T1 and T2 (assuming real roots).

As illustrated in the
That is, output input example it is more
k convenient to find the
¯ (s) =
! ē(s)
(T1 s + 1)(T2 s + 1) output of a system
| {z } when a particular input
Transfer Function is applied by working in
We call G(s) = (sT +1)(sT k the transfer function from ē(s) to !
¯ (s) (or, the Laplace domain. In
1 2 +1) particular we find first
relating !
¯ (s) and ē(s)). the transfer function
The diagram relating the input and
output of interest.
ē(s) !
¯ (s)
G(s)
Then the Laplace trans-
form of the output can
be evaluated by multi-
represents this relationship between the input and the output:
plying the transfer func-
By using Laplace transforms, all LTI blocks can be tion with the Laplace
transform of the input.
treated as multiplication by a transfer function.

For the motor used in the lego demonstrations we have k ⇡ 2.2, T1 ⇡ 54ms
and T2 ⇡ 1ms.

14

14
2.5.1 Impulse Response of the DC motor
To find the impulse response, we let

e(t) = (t) unit impulse

which implies
ē(s) = 1
and put
!(0) = i(0) = 0.
So,
k
!
¯ (s) = ⇥1
(T2 s + 1)(T1 s + 1)
Split into partial fractions:

1 1 1 1
!¯ (s) = k ⇥ + ⇥
T1 sT2 + 1 T2 sT1 + 1
1 1
T2 T1

k 1 1
= +
T1 T2 s + 1/T2 s + 1/T1

Hence, 
k It should be noted that
!(t) = e
|
t/T1
{z } |
e {zt/T2} +
T1 T2 the response of the
Fast Transient Slow Transient system is dominated
e.g. T2 ⇡ 1ms T1 ⇡ 54 ms by the slow transient,
i.e. the term with the
50
slow transient largest time constant
k
40
Ti .
(T1 T2 )
total impulse response
30
Note also that the coef-
20 ficients multiplying the
initial value= 0 (= lims!0 s¯
! (s)) time t in the exponen-
! (rad/sec)

10
tial terms are equal to
0
the poles of the transfer
−10 function.
−20 fast transient
−30

k −40

(T1 T2 ) −50
0 0.05 0.1 0.15 0.2 0.25 0.3

t (sec)

15

15
2.5.2 Step Response of the DC motor
To find the step response, we let

e(t) = H(t)

which implies that

1
ē(s) =
s
and put
!(0) = i(0) = 0
(as for the impulse response.) So,

k 1
!
¯ (s) = ⇥
(T2 s + 1)(T1 s + 1) s
Split into partial fractions:
" #
1 T2 1 T1 1
!
¯ (s) = k + ⇥ ⇥
s T1 T2 s + T12 T1 T2 s + T11

Hence, As in the case of the


T2 T1 impulse response we see
!(t) = k H(t) + ⇥ e t/T2 ⇥ e t/T1
T T2 T T2 that the slow transient,
|1 {z } |1 {z }
i.e. the exponential
Fast Transient Slow Transient
term with the largest
2.5
step time constant Ti , deter-
k 2
mines how quickly the
output converges to its
1.5 total step response equilibrium value.
1
! (rad/sec)

0.5
kT2 fast transient
0
(T1 T2 )
−0.5
initial slope= 0
−1

−1.5
slow transient
kT1 −2

(T1 T2 ) −2.5
0 0.05 0.1 0.15 0.2 0.25 0.3

t (sec)

16

16
2.5.3 Deriving the step response from the
impulse response

For a system with impulse response g(t), the step response is given by
Z t Z t
y(t) = g(⌧ )H(t ⌧ ) d⌧ = g(⌧ ) d⌧
0 0

i.e.
step response = integral of impulse response

Check this on the example. Make sure you understand where the term H(t) in the
step response comes from.

2.6 Transforms of signals vs Transfer functions of It should be noted that

systems a function of s could
be either the Laplace
transform of a signal or
Mathematically there is no distinction, but in practice they have different
a transfer function.
interpretations
In particular, a trans-
G(s) Signals Systems
fer function corresponds
1
L G(s) ȳ(s) = G(s) · ū(s) to a system, and in the
m time domain it is repre-
sented by a differential
1 (t) y(t) = u(t)
Z t equation relating the in-
1
H(t) y(t) = u(⌧ )d⌧ (integrator) put and output. On
s
0 the other hand, a signal
1
s+a e at ẏ(t) + ay(t) = u(t) (lag)
Et Es in the time domain is a
E
function of time. These
!
sin(!t) ÿ(t) + ! 2 y(t) = !u(t) are therefore two differ-
s2 +! 2
ent interpretations. Ex-
amples of functions of
E E
sT
e (t T) y(t) = u(t T) (delay)
s and the correspond-
In each case, the “signal” is the impulse response of the “system”. ing signals and systems
in the time domain are
17 presented in the table
on the left.

17
2.7 Interconnections of LTI systems One of the main ad-
x̄(s) vantages of the use of
a) F (s) transfer functions is the
+ fact that they simplify
v̄(s) z̄(s) significantly the analy-
⌃ H(s)
sis when we have inter-
ȳ(s) + connections of systems.
G(s)
In particular, finding
the transfer function
represents the equations: of a system that is
an interconnection of
v̄(s) = F (s)x̄(s) + G(s)ȳ(s) other systems, becomes
a relatively simple
and
problem involving alge-
z̄(s) = H(s)v̄(s) braic operations. Such
examples are illustrated
⇣ ⌘
=) z̄(s) = H(s) F (s)x̄(s) + G(s)ȳ(s) on the left and in the
next page.
= H(s)F (s) x̄(s) + H(s)G(s) ȳ(s)
| {z } | {z }
transfer function transfer function
from x̄(s) to z̄(s) from ȳ(s) to z̄(s)

r̄(s) + ē(s) ȳ(s)

b) ⌃ G(s)

K(s)

represents the simultaneous equations:

ē(s) = r̄(s) K(s)ȳ(s)

and
ȳ(s) = G(s)ē(s)
=) ȳ(s) = G(s)r̄(s) G(s)K(s)ȳ(s)
=) (1 + G(s)K(s))ȳ(s) = G(s)r̄(s)
G(s)
=) ȳ(s) = r̄(s)
1 + G(s)K(s)

18

18
2.7.1 “Simplification” of block diagrams

Recognizing that blocks represent multiplications, and using the above formulae, it
is often easier to rearrange block diagrams to determine overall transfer functions,
e.g. from ū(s) to ȳ(s) below.
ū(s) + + ȳ(s)
⌃ G 1 (s) G 2 (s) ⌃ G 3 (s)
+
K1 (s)

+
ū(s) + G3 (s) ȳ(s)
⌃ G1 (s)G2 (s)
+ 1 + G3 (s)K1 (s)

+
ū(s) + G1 (s)G2 (s)G3 (s) ȳ(s)
⌃
1 + G3 (s)K1 (s)

G1 (s)G2 (s)G3 (s)

1 + G3 (s)K1 (s)
ȳ(s) = ū(s)
G1 (s)G2 (s)G3 (s)
1
1 + G3 (s)K1 (s)
or
G1 (s)G2 (s)G3 (s)
y(s) = ū(s)
1 + G3 (s)K1 (s) G1 (s)G2 (s)G3 (s)

19

19
2.8 More transfer function examples

To obtain the transfer function, in each case, we take all initial conditions to be zero.
The following three systems all have the same transfer
function (a 1st order lag )
1) Spring/damper system The first example is a
spring/damper system.
k The system has input
x and output y, which
are the displacements
at the two terminals of
the spring, respectively.
y(t) x(t)
The system equation re-
ẏ = k(x y) lating x and y is derived
by considering the dif-
=) ẏ + y = x
k ferential equation satis-
fied by the damper, and
=) sȳ(s) + ȳ(s) = x̄(s)
k Hooke’s law applied to
the spring.

✓ ◆
=) s + 1 ȳ(s) = x̄(s)
k
1
=) ȳ(s) = x̄(s) T = /k
Ts + 1

Step Response:
1 1 1 T
x̄(s) = =) ȳ(s) = =
s s(sT + 1) s sT + 1

y(t)

=) y(t) = H(t) e t/T

T t

20

20
2) RC network
The second example is
an RC electrical circuit
R
where the input x is the
supply voltage and the
x(t) C y(t) output y is the voltage
across the capacitor, as
shown in the diagram.
dy x y
i=C = The differential equa-
dt R
tion relating x and y
follows from the differ-
=) RC ẏ + y = x ential equation relating
the voltage across a ca-
pacitor with the current
=) (RCs + 1)ȳ(s) = x̄(s) through it (part IA).

1
=) ȳ(s) = x̄(s) T = RC
Ts + 1

3) Water Heater
The third example is
q(t) (J/s) a water heater system
(assuming perfect mixing – i.e. all where the input q is the
rate with which energy
the water in the tank is at the same
is supplied to the sys-
✓i M(kg) ✓o temperature, ✓o ) tem and the output ✓0
is the temperature of
ṁ specific the water in the tank.
(kg/s) heat capacity
The differential equa-
q = Mc ✓˙0 + ṁc(✓o ✓i ) tion relating these vari-
ables is a statement of
=) (Mcs + ṁc)✓¯o (s) = q̄(s) + ṁc✓¯i (s)
conservation of energy.
1 1 1 ¯ It quantifies the fact
=) ✓¯o (s) = q̄(s) + ✓ (s) T = M/ṁ that the energy sup-
ṁc T s + 1 Ts + 1 i
plied to the system is
input disturbance used to raise the tem-
(water flow rate ṁ is assumed to be constant) perature of the water al-
ready within the water
tank, and of the wa-
ter flowing into the tank
21 with constant rate ṁ.

21
2.9 Key Points
The step response is the integral (w.r.t time) of the impulse response.
The transfer function is the Laplace transform of the impulse response.

By using transfer functions, block diagrams represent simple algebraic

relationships (all blocks become multiplications).

22

22
Part IB Paper 6: Information Engineering
LINEAR SYSTEMS AND CONTROL
Ioannis Lestas

HANDOUT 3

“Stability and pole locations”

asymptotically marginally unstable

stable stable
Imag(s)

repeated
poles +
X

Unstable

X X X
Asymptotically Marginally Unstable
stable stable
Real(s)
X X X
Left half plane Right half plane

Imaginary axis

1
Summary

Stability, or the lack of it, is the most fundamental of system properties.

When designing a feedback system the most basic of requirements is that the
feedback system be stable.
There are different ways of defining stability. In this handout we shall:
Define the following notions:
Asymptotic stability
Marginal stability
Instability

Relate stability of a system to the poles of its transfer function

In addition, we shall:
Relate the transient response of a system to the poles of its transfer
function

Contents

3 Stability and pole locations 1

3.1 Asymptotic Stability . . . . . . . . . . . . . . . . . . . . . . . 3
3.2 Poles and the Impulse Response . . . . . . . . . . . . . . . . . 4
3.3 Asymptotic Stability and Pole Locations . . . . . . . . . . . . 9
3.4 Marginal Stability . . . . . . . . . . . . . . . . . . . . . . . . 12
3.5 Instability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.6 Stability Theorem . . . . . . . . . . . . . . . . . . . . . . . . 14
3.7 Poles and the Transient Response . . . . . . . . . . . . . . . . 17
3.8 Key Points . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

2
3.1 Asymptotic Stability
Definition: It should be noted
that this definition
An LTI system is asymptotically stable if its impulse response g(t) satisfies of asymptotic stabil-
the condition Z 1 yet Iggyuld e de ity guarantees that
|g(t)|dt < 1 g(t) ! 0 as t ! 1

Examples:
0
IlyAllactHdt as otherwise the inte-
gral in the definition
would tend to infinity.
E Uman 19aside
1. LCR circuit: g(t) = e t sin(3t + 2)
e t
|g(t)|
lying lil ill
Z 1 Z 1 t
t
|g(t)| dt  e dt = 1< 1
0 0

2. Delay line with lossy reflections:

1
X 1
g(t) = (t kT )
2k
k=0

t
Z 1 Z 1X
1 1
X
1 1
|g(t)| dt = (t kT ) dt= = 2< 1
0 0 2k 2k
k=0 k=0
E
integral
T ) asymptotically stable

of impulse
is I

3
3.2 Poles and the Impulse Response

Although stability is most easily defined in terms of the impulse response, it is

most easily determined (at least for systems with rational transfer functions,
the ones that come from ODE’s) in terms of pole locations. To understand
this, we first need to look at the relationship between the poles of a system
and solutions to its differential equation – in particular its impulse response.

Example: Consider the system with input u and output y related by the ODE
d2 y dy du
2
+↵ + y=a + bu.
dt dt dt
The Auxillary Equation for this ODE is
2
+↵ + =0

with Complementary Factor 1t 2t

yCF = Ae + Be
This decays to 0 as t ! 1 only when 1 < 0 and 2 < 0 (or, if the roots are complex,
when their real parts are negative).
In terms of transfer functions we have
as + b
ȳ(s) = ū(s)
s2
I
+ ↵s +
transfer function
The poles of the system’s transfer function are given by the roots of the denominator -
that is the solutions to
s2 + ↵s + =0
So, for a system described by an ODE, its poles are precisely the solutions to its Auxiliary
Equation.

4
4
Consider now a general LTI system described by an ODE, and consequently
having a rational transfer function G(s). That is, it can be written as the
ratio of two polynomials
n(s)
G(s) =
d(s)
(where the coefficients of d(s) come from the LHS of the underlying ODE and
the coefficients of n(s) come from the RHS).
We can factorize the denominator to give
n(s)
G(s) =
(s p1 )(s p2 ) · · · (s pn )

We will also assume that G(s) is proper, that is e.g. G(s) = s

(a differentiator)
deg[n(s)]  deg[d(s)]. is not proper

(This condition will always be satisfied for physically realizable systems. Moreover, any
system whose transfer function violates this condition is not asymptotically stable.)

In this case we can perform a partial fraction expansion to give

↵1 ↵2 ↵n
G(s) = + + ··· + +C
s p1 s p2 s pn

where ↵i = lims!pi (s pi )G(s) is called the residue at s = pi . (we are

assuming no repeated poles here, for simplicity of notation). Finally, by taking
inverse Laplace transforms, this means we can write the impulse response in
the form

g(t) = ↵1 ep1 t + ↵2 ep2 t + · · · + ↵n epn t + C (t)

Consider one of these terms, ept say. How it contributes to g(t) depends on
whether p is real or complex:

5
• If p is real: then ept is a real exponential, with time constant |1/p| .

p<0 p>0

• If p is complex then we need to consider <(↵ept ) (the imaginary part will

cancel out with the contribution from p⇤ , which will also be a pole† , since g(t)
must be real).
This will give either a damped or a growing sinusoid:

↵ ept ) = <(Aej ept ) = <(Ae t ej(!t+ ) )

<(|{z}
Aej
= Ae t cos(!t + )
(where we have put p = + j! again)
time constant = |1/ | frequency =!

2⇡/!

= <(p) < 0 >0

†
complex poles always appear in conjugate pairs since they are roots of a real polynomial

6
So each pair of complex poles contributes a term of the form

2Ae t cos(!t + )

where = <(p), ! = =(p)

Compare this with the impulse response of a second order system (see Mechanics
data book)
8
< = !n ⇣
!n ⇣t
Ce sin(!d t) =) p
:! = !d = !n 1 ⇣2

Clearly, the impulse response of any rational system can be regarded as a

combination of 1st and 2nd order terms. Furthermore, the contribution of the
second order terms can be understood in terms of the language of second order
systems, as the following very important figure makes clear :

We have assumed that no poles are repeated for this discussion. Repeated poles give
rise to terms of the form tm ept (or tm e t cos(!t + )), which have the same general
characteristics (as the exponential dominates the polynomial term).

=(s)
q
p= + j! ! = !n 1 ⇣2
X

!n sin 1 ⇣

cos 1 ⇣ <(s)
0
= !n ⇣

⇤
X p = j!
7

7
 This figure shows that, given the pole locations, in the complex plane, of a
second order system we can read off the natural frequency, the damping ratio
and also !n ⇣, the reciprocal of the time constant of the decay.
For a higher order system, we can read X =(s)
off the natural frequency and damping
X
ratio of each “mode” of the system
(each pair of complex poles). The
poles closest to the imaginary axis are
often called the dominant poles (their <(s)
2 X1 0
contribution dies away most slowly, and
so tends to dominate the response)

constant X
If damping ratio
X
then poles lie on lines

Im(s)
!n ⇣ = 1.25
2
⇣
=

⇣=0
1/
p 2

i
1

Sta ⇣=1 Re(s)

2 1 0 1 2
!n

gg
=
!n

0.
!n

1
1.
=
!n

0
1.
=

5
2.

2
0

This figure shows radial contours of constant damping ratio ⇣ and circles of
constant natural frequency !n as well as a vertical lines on which !n ⇣ is
constant.

8
3.3 Asymptotic Stability and Pole Locations

We will now show the following: The Theorem stated

Theorem: An LTI system with rational transfer function G(s) is on the left is an
important result in
asymptotically stable if, and only if, all poles of G(s) lie in the LHP
the study of linear
systems and in con-
=(s)
trol system design, as
asymptotic stability
is a property often
Left Half Plane Right Half Plane required in the design
LHP RHP of feedback systems.

In particular, the
<(s)
significance of the
<(s) < 0 <(s) > 0 Theorem is that it
allows to check the
Imaginary Axis asymptotic stability
<(s) = 0 of a linear system
without explicitly
evaluating its im-
pulse response, but
Proof: by simply checking
i) First we show that if all poles have a negative real part then the system is the position of the
poles of its transfer
asymptotically stable.
function.
For now, assume that the poles of G(s) are distinct
More precisely, the
i.e. that d(s) has no repeated roots system is asymptot-
(we shall remove this restriction later) ically stable if and
only if the real part of
then we can write the poles is negative.
n(s)
G(s) =
(s p1 )(s p2 ) · · · (s pn )
↵1 ↵2 ↵n
= ↵0 + + + ··· +
(s p1 ) (s p2 ) (s pn )
by partial fraction expansion, and so

g(t) = ↵0 (t) + ↵1 ep1 t + ↵2 ep2 t + · · · + ↵n epn t .

9
Now, let
k = <(pk ) and !k = =(pk )
so pk = k + j!k , for each k = 1 . . . n, then
|epk t | = |e( k +j!k )t | = |e k t ej!k t | = |e k t | |ej!k t | = e k t
| {z }
and so 1
|g(t)|  |↵0 | (t) + |↵1 |e 1 t + |↵2 |e 2 t + · · · + |↵n |e n t .

Now, 8
1

H
Z 1  1 >
<
1 , if <0
e t dt = e t =
0 0 >
:1, if 0
and furthermore, since every pole has k < 0, then
Z 1
↵ ↵ ↵n
|g(t)| dt  |↵0 | + 1 + 2 + · · · + <1
0 1 2 n
and consequently the system is asymptotically stable as required.

Repeated poles: If G(s) has repeated poles, i.e.

···
G(s) = ,
· · · (s p)l · · ·
where l denotes the multiplicity of the pole at s = p, then the partial
fraction expansion of G(s) will be of the form

1 2 l
G(s) = · · · + + + ··· + + ··· .
(s p) (s p)2 (s p)l
Hence, the impulse response g(t) will be of the form

l
g(t) = · · · + 1 ept + 2 tept + · · · + tl 1 ept + · · ·
(l 1)!
However, if p = + j! and < 0 (ie <(p) < 0), then
Z 1 Z 1
|tk 1 pt
e | dt = tk 1 e t dt < 1
0 0

for any k. Hence the conclusion remains valid.

10

10
ii) Now we show the converse, that if a system is asymptotically stable then
all poles have a negative real part. of jut
For all values of s for which <(s) 0, we have
so jw Esty

Ét t
Z 1 Z 1
|G(s)| = st
e g(t) dt  e st |g(t)| dt
Z 01 0

 |g(t)| dt ( since e st  1 for Re(s) 0 and t > 0)

0
= A< 1.
wind
since the system is asymptotically stable. This means that G(s) cannot have
any poles on the imaginary axis or in the right half of the complex plane. So
any poles it does have must have a negative real part, as required. .

So far, we have divided systems into two classes: those that are
asymptotically stable and those that are not. We shall now further classify the
systems that are not asymptotically stable into two classes: those that are
marginally (i.e. “almost”) stable and those that are unstable.

11

11
3.4 Marginal Stability

Definition: An LTI system is marginally stable if it is not

asymptotically stable, but there nevertheless exist numbers A, B < 1 such
that Z T
|g(t)|dt < A + BT for all T
0

Examples:

1. Integrator:
Z T
g(t) = H(t) = |g(t)| dt = T
T 0
G(s) = 1/s =) j!-axis pole at s=0

2. Undamped spring-mass system:

Z T Z T
g(t) = cos(3t) =) |g(t)| dt  1 dt = T
0 0
G(s) = s/(s2 + 9) =) j!-axis poles
at s = ±3j
3. Delay line with lossless reflections:
1
X Z T
g(t) = (t k), =) |g(t)| dt  T +1
k=0 0

4. Something which cannot arise as the impulse response of any ODE:

(g(t) ! 0, but system is not asymptotically stable)
Z T
1
g(t) = =) |g(t)| dt = log(T + 1)  T
t+1 0

12

12
3.5 Instability

Definition: A system is unstable if it is neither asymptotically stable nor It should be noted

marginally stable. that the impulse
response g(t) of an
unstable system will
Examples:
tend to infinity as
1. Inverted pendulum: t ! 1.

g(t) = e4t + e 4t This is also evident

1 in the examples illus-
G(s) = + · · · =) pole at s = 4 trated on the left.
s 4
2. Two integrators in series:

R R
g(t) = t
1
G(s) = =) double pole at s = 0
s2

3. Oscillation of badly designed control system:

g(t) = e0.01t sin(0.3t)

0.3
G(s) =
(s 0.01)2 + 0.32
=) poles at s = 0.01 ± 0.3j

Warning: Different people use different definitions of stability. In particular, systems which
we have defined to be marginally stable would be regarded as stable by some, and unstable
by others. For this reason we avoid using the term “stable” without qualification.

13

13
3.6 Stability Theorem

It should be clear from these examples that

if any of the poles of G(s) have a positive real part then the impulse
response will have a term that blows up exponentially (consider the
partial fraction expansion of G(s)).
Also, if G(s) has a repeated imaginary axis pole then the impulse
response will have a term that still blows up, although more slowly.
In either of these cases, the system is unstable.

Isolated poles on the imaginary axis, on the other hand, give rise to terms
in the impulse response which remain bounded (e.g. steps or sinusoids).

In this case the system is not asymptotically stable but is nevertheless

marginally stable (provided it has no RHP or repeated imaginary axis poles).

In fact, (for systems with proper rational transfer functions) it can be

shown that
Stability Theorem:

1. A system is asymptotically stable if all its poles have negative real parts.

2. A system is unstable if any pole has a positive real part, or if there are
any repeated poles on the imaginary axis.

3. A system is marginally stable if it has one or more distinct poles on the

imaginary axis, and any remaining poles have negative real parts.

Note: we proved part 1, and the converse statement that a system is not asymptotically
stable if any of its poles have a zero or positive real part, on page 6) The refinement of “not
asymptotically stable” into marginal stability and instability has only been illustrated by
examples. The proof of parts 2 and 3 is not difficult, but is messy (and so is omitted).

14

14
 X X
The examples on the
X X X X
left illustrate the Sta-
bility Theorem stated
X X
in the previous page,
asymptotically unstable i.e. how the location
stable of the poles determines
the stability properties
of a system.
X X X

X X X X
repeated
X j!-axis poles X X
unstable marginally
stable

Note: it’s the “worst” poles that determine the stability properties

15

15
 =(s)

<(s)
X

(s + 1.5)(s2 s + 1)
Poles/zeros for G(s) =
(s + 2)(s2 + 0.1s + 4)

Note: this is an asymptotically stable system.

pole at s = 2 pole at s = 0.05 + 2j

pole at s = 0.05 2j
5
zero at s = 0.5 + 0.87j
4

3
|G(s)|

1
2
zero
0
at s = 1.5 0
−3 −2
zero at s = 0.5 0.87j
−1 0 −2
1 2 =(s)
3
<(s)

16

16
3.7 Poles and the Transient Response

The term Transient Response refers to the initial part of the (time domain)
response of a system to a general input (before the “transients” have died
out). To a very large extent, these transients are a characteristic of the
system itself rather than the input.
If, for example a system with transfer function
n(s)
G(s) =
(s p1 )(s p2 ) · · · (s pn )
is given an input u(t), with Laplace transform ū(s), then the response is given
by

n(s)
ȳ(s) = G(s)ū(s) = ū(s)
(s p1 )(s p2 ) · · · (s pn )
1 2 n
= + + ··· + + other stuff
s p1 s p2 s pn

and so
y(t) = 1 ep1 t + 2 ep2 t + · · · + n epn t + other stuff
That is, the response y(t) contains the same terms as the impulse response
(although with different amplitudes) plus some extra terms due to particular
characteristics of the input.

17

17
3.8 Key Points
• The impulse response of an LTI system is a sum of terms due to each real
pole, or pair of complex poles.
• The system’s response to any input will also include these features.
The following figure shows a selection of pole locations, with their
corresponding contribution to the total response.
This again is an important figure.
Note:
The real part of the pole, , determines both stability and the time
constant, |1/ |.
The imaginary part of the pole, !, determines the damped natural
frequency (actual frequency of oscillation) in rad/sec.
The magnitude of the pole determines the natural frequency.
The argument of the pole determines the damping ratio.

Imag(s)
The figure on the left
illustrates how the
location of the poles of
the transfer function of
X X X X X a system (denoted by
X) affect its impulse re-
sponse (and hence also
its transient response
X X X X X when a general input is
applied as explained in
the previous page).
Real(s)
X X X X X In particular, as the
imaginary part of the
Left half plane Right half plane poles increases the
response becomes more
Imaginary axis oscillatory. Also, as the
real part of the poles
Pole locations and corresponding transient responses becomes more negative,
the response decays
more quickly to zero.

18 Note also that the ex-

istence of at least one
pole in the RHP implies
instability.

18
 Part IB Paper 6: Information Engineering
LINEAR SYSTEMS AND CONTROL
Ioannis Lestas

HANDOUT 4

“The Frequency Response G(j!)”

2⇡ arg G(j!)
x(t) ! y(t) !
1
|G(j!)|

t t

Asymptotically
stable LTI
system G(s)

x(t) = cos(!t) y(t) = |G(j!)| cos !t + arg G(j!)

+starting transient

1
Summary

If a pure sinusoid is input to an asymptotically stable LTI system, then the

output will also settle down, eventually, to a pure sinusoid. This steady-state
output will have the same frequency as the input but be at a different
amplitude and phase. The dependence of this amplitude and phase on the
frequency of the input is called the frequency response of the system.
In this handout we shall:
Show how the frequency response can be derived from the transfer
function.
(by substituting j! for s)
Study how the frequency response can be represented graphically. (using
the Bode diagram)

Contents

4 The Frequency Response G(j!) 1

4.1 What is the Frequency Response? . . . . . . . . . . . . . . . . 3
4.1.1 Derivation of gain and phase shift . . . . . . . . . . . . 6
4.1.2 Frequency response of an arbitrary linear system . . . . 8
4.2 Plotting the frequency response . . . . . . . . . . . . . . . . . 9
4.3 Sketching Bode Diagrams . . . . . . . . . . . . . . . . . . . . 12
4.3.1 Powers of s: (sT )k . . . . . . . . . . . . . . . . . . . 13
4.3.2 First Order Terms: (1 + sT ) . . . . . . . . . . . . . . . 14
4.3.3 Second order terms: (1 + 2⇣sT + s2 T 2 ) . . . . . . . . 16
4.3.4 Examples . . . . . . . . . . . . . . . . . . . . . . . . . 18
4.4 Key Points . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

2
4.1 What is the Frequency Response?
Our analysis in sec-
tion 4.1 will be as
2⇡
follows:
x(t) ! y(t) !
1 (i) We will first use
A
complex number ap-
proaches studied in
t t
part IA to illustrate
that for a system
with transfer function
Asymptotically
stable G(s) that has as in-
LTI system put a sinusoid, the
output is also a si-
nusoid with the same
frequency. Further-
x(t) = cos(!t) y(t) = A cos !t + more, the gain A,
and phase shift are
+starting transient
given by

A = |G(j!)|,
= 6 G(j!)
If the input to an asymptotically stable LTI system is a pure sinusoid then the (ii) We will then show
steady state output will also be a pure sinusoid, of the same frequency as the a more general and
input but at a different amplitude and phase. practically relevant
result. We will show
How do we find A, the gain, and , the phase shift ? that even when the
system is initially
at rest and then
a sinusoidal input
is applied, then as
t ! 1 the output
will be a sinusoid,
with gain and phase
shift given by the
same expressions.

(iii) This result will

be initially shown
for linear systems
described by differen-
tial equations. It will
then be generalised
to any linear system.

3
How does this relate to Part IA? Consider again a system with input u
and output y, If
d2 y dy du
+ ↵ + y = a + bu.
dt2 dt dt
then we can use the usual trick, letting

u = ej!t

so that
<(u(t)) = cos(!t).
We will find the response to the input cos(!t) by taking the real part of y, the
response to u = ej!t .
To find the solution y, we assume it takes the form

y(t) = Y ej!t (4.1)

for some complex number Y = |Y |ej arg Y , so that

<(y(t)) = <(|Y |ej arg Y ej!t ) = |Y | cos(!t + arg Y ).

Substituting (4.1) into the differential equation, and noting that

dy d2 y
= [j!]Y ej!t , = [j!]2 Y ej!t , etc
dt dt2
we obtain

Y [j!]2 ej!t + ↵Y [j!]ej!t + Y ej!t = a[j!]ej!t + bej!t

or
a[j!] + b
Y =
[j!]2 + ↵[j!] +
Note that the transfer function from ū(s) to ȳ(s) is given by
as + b
G(s) = 2
s + ↵s +
So, it would appear that Y = G(j!), suggesting:
Answer: If system has the transfer function G(s), then

A = |G(j!)|, = arg G(j!)

4
You’ve done this in Linear Circuits, Maths and Mechanical Vibrations in Part
IA, so I’m assuming that you are familiar with these arguments! The notation
above is as used in Maths and Mechanical Vibrations. In Linear Circuits ỹ was
used to represent a complex phasor, rather than the Y above.
Example:The Capacitor is described by the differential equation
dv
i=C
dt
so ī(s) = Csv̄(s) in the absence of initial conditions giving the transfer
function
v̄(s) 1
= .
ī(s) sC
1 ,
So, the frequency response of a capacitor (from current to voltage) is j!C
which equals its impedance, ṽ/ĩ. The advantage of transfer functions is that
they can be used to deduce the response to all possible inputs, whereas the
impedance can only be used for sinusoidal (ac) signals, providing the change
in their amplitude and phase.

5
4.1.1 Derivation of gain and phase shift

The previous argument is not the whole story though - we have assumed that
the input u = cos(!t) has been present since the beginning of time, and have
only shown that y = A cos(!t + ) is a possible response. What if the system
is at rest until t = 0, and then a sinusoid is applied? Take an asymptotically
stable system with input ū(s) output ȳ(s) and rational transfer function
n(s)
G(s) = d(s) , so
ȳ(s) = G(s)ū(s).
Let u(t) = ej!t , so ū(s) = s 1j! then, since G(s) can’t have a pole at
s = j!,
1 1 2 n 0
ȳ(s) = G(s) = + + ··· + +
s j! s p1 s p2 s pn s j!
How can we find 0 without calculating all the other terms?
Ans: Cover-up rule.

Multiply both sides by (s j!) to obtain:

(s j!) (s j!) n (s j!)
G(s) = 1 + 2 + ··· + + 0
s p1 s p2 s pn
=)
0 = G(j!)
=) y(t) = 1 ep1 t + 1 ep2 t + · · · + n epn t + G(j!)ej!t
| {z } | {z }
ytr (= CF) yss (= PI)
(with the obvious modifications if any poles are repeated).
Since all the pk , the poles of G(s), have a negative real part, then ytr (t) ! 0
as t ! 1 – leaving the steady-state response yss . The steady-state response
to the input u(t) = cos(!t) is given by the real part of this:
⇣ ⌘
j!t
<(yss (t)) = < e G(j!)
✓ ◆
= < |G(j!)|e j !t+arg G(j!)
as z = |z|ej arg z
= |G(j!)| cos !t + arg G(j!)
| {z } | {z }
gain A phase shift
6

6
Example:

Consider a system with transfer function

1
G(s) = 2
s + 0.1s + 2
and an input
x(t) = cos(0.5t)

(a sinusoid at 0.5rad/s, or 0.5/(2⇡) = 0.0796 Hz )

So, ! = 0.5 and
1 1
G(j!) = =
(0.5j)2 + 0.1(0.5j) + 2 0.25 + 0.05j + 2
1 1
= = 0.0286j
= 0.571e 0.0286j
1.75 + 0.05j 1.7507e
= 0.571 6 1.64
The following figures show the impulse response of this system, the input
x(t) = cos(0.5t), and the response to this input.
Impulse Response of
Note how the output initially G(s) = 1/(s2 + 0.1s + 2)
1

contains a component near g(t)

0.5

the resonant frequency of the

p 0
system !n = 2 (0.225 Hz) but
−0.5
that this quickly decays leaving only
−1
a sinusoid at the frequency of the 0 50 100 150 200
Input: x(t) = cos(0.5 t)
input ! = 0.5 (0.08 Hz) (and at an 1

amplitude of 0.571). 0.5

(The phase lag of 1.64 is too small 0

to be seen on this diagram) −0.5

−1
0 50 100 150 200
Response to the input x(t) = cos(0.5 t)
1
y(t) = g(t) ⇤ x(t)
0.571 0.5

−0.5

−1
0 50 100 150 200
(sec)
tCsec

7
4.1.2 Frequency response of an arbitrary linear The analysis in section
system 4.1.1 was for linear
systems described by
It’s not just systems described by ODEs which have a frequency response, any differential equations.
This is reflected in
asymptotically stable system, with impulse response g(t) and transfer function
the fact that G(s) was
G(s) = Lg(t), does. (See Q5 on Examples Paper 2 for an example). Consider assumed to be rational
the response of such a system to a sinusoidal input beginning suddenly at in s (polynomial di-
t = 0. vided by polynomial).

Let In this section it is

x(t) = <(ej!t ) = cos !t
proved that the same
then result holds (for the
✓Z t ◆ response when the
Z t
input is a sinusoid), for
y(t) = < ej!(t ⌧ ) g(⌧ ) d⌧ = < ej!(t ⌧ ) g(⌧ ) d⌧ any linear system. In
0 | {z } 0
x(t ⌧ ) particular, no assump-
✓Z 1 ◆ ✓Z 1 ◆ tion on the transfer
=< ej!(t ⌧ ) g(⌧ ) d⌧ < ej!(t ⌧ ) g(⌧ ) d⌧ function G(s) is made
0
| {z } | t
{z } and only use the fact
yss (t) ytr (t) that the output y(t)
is the convolution of
the input u(t) with the
Consider the second term: impulse response g(t).
Z 1 Z 1
|ytr (t)|  |ej!(t ⌧ ) | |g(⌧ )| d⌧ = |g(⌧ )| d⌧ It should be noted
t | {z } t
1 that we are considering
Z 1 asymptotically stable
but |g(⌧ )| d⌧ is finite, and hence limt!1 |ytr (t)| = 0. systems,R which implies
1
0 that 0 |g(⌧ )|d⌧ is
ytr (t) is called the transient part of the response , as it decays to zero, finite (from the defi-
eventually (and so is transitory in nature). nition of asymptotic
stability given in the
Now consider the first term: previous handout).
✓ Z 1 ◆ ⇣ ⌘
yss (t) = < e j!t e j!⌧ j!t
g(⌧ ) d⌧ = < e G(j!)
|0 {z }
G(s) s=j!
✓ ◆
j !t+6 G(j!)
= < |G(j!)|e = |G(j!)| cos !t + 6 G(j!)

yss (t) is called the steady-state part of the response, as it is the part that
remains when the transients have decayed.

8
4.2 Plotting the frequency response

The frequency response G(j!) is a complex-valued function of frequency !.

At each frequency !, the complex number G(j!) can be represented either in
terms of its real and imaginary parts, or in terms of its gain (magnitude) and
phase (argument).
There are a number of common ways of representing this information
graphically:
The Bode Diagram: Two separate graphs, one of |G(j!)| vs ! (on
log-log axes), and one of 6 G(j!) (lin axis) vs ! (log axis).
The Nyquist Diagram: One single parametric plot, of < G(j!)) against
= G(j!)) (on linear axes) as ! varies.
The Nichols Diagram: One single parametric plot, of |G(j!)| (log axis)
against 6 (G(j!)) (lin axis) as ! varies.

Example: G(s) = 1
2
s(s +0.2s+2)
Bode Gain Plot Nyquist Diagram
20 1

0
|G(j!)| (dB)

0
= G(j!)

−1

−20 −2

−3
−40
−4

−60 −5
−1 0 1 −5 −4 −3 −2 −1 0 1 2
10 10 10
! (rad/s) < G(j!)
Bode Phase Plot Nichols Chart
0 20
6 G(j!) (deg)

|G(j!)| (dB)

0
−90

−20

−180
−40

−270 −60
−1 0 1 −270 −225 −180 −135 −90 −45 0
10 10 10
! (rad/s) 6 G(j!) (deg)

9
Each has its use:
The Bode diagram is relatively straightforward to sketch to a high degree of
accuracy, is compact and gives an indication of the frequency ranges in which
different levels of performance are achieved.
The Nyquist diagram provides a rigorous way of determining the stability of a
feedback system.
The Nichols diagram combines some of the advantages of both of these
(although is not quite as good in either specific application) and is widely
used in industry. We shall not study the Nichols diagram in this course(!), but
the ideas behind its construction will be readily grasped once the two
fundamental diagrams that we do study are understood.

10

10
Consider G(j!) for ! = 2
1 1
G(j!) = =
!=2 2j((2j)2 + 0.2 ⇥ (2j) + 2) 2j( 2 + 0.4j)
Hence
1
|G(j2)| = p = 0.2451
2 22 + 0.42
and

arg G(j2) = arg 1 arg 2j arg( 2 + 0.4j)

=0 ⇡/2 2.9442
| {z } = 4.5150
⇡ atan(.4/2)

0 = 1
B C
B C
B C
B C
B C
B C
B C
B C
B
B because
X < C
C
B C
B C
B C
B C
B C
B C
B C
@ A

So,
G(j2) = 0.2451e 4.5150j = 0.0481 + 0.2404j
Also,
20log10 |G(j2)| = 12.2dB, 6 G(j!) = 258.7

The corresponding point has been marked with a cross on each of the previous
plots.

11

11
4.3 Sketching Bode Diagrams

The bode diagram of G(s) consists of two curves

1. Gain Plot: Gain |G(j!)| (log) vs freq ! (log)
2. Phase Plot: Phase 6 G(j!) (lin) vs ! (log)
It is straightforward to sketch, and gives a lot of insight.
Basic idea: Consider a transfer function written as a ratio of factorized
polynomials e.g.
a1 (s)a2 (s)
G(s) = .
b1 (s)b2 (s)
Clearly
log10 |G(j!)| = log10 |a1 (j!)| + log10 |a2 (j!)|
log10 |b1 (j!)| log10 |b2 (j!)|,
so we can compute the gain curve by simply adding and subtracting gains
corresponding to terms in the numerator and denominator. Similarly
6 G(j!) = 6 a1 (j!) + 6 a2 (j!) 6 b1 (j!) 6 b2 (j!)

and so the phase curve can be determined in an analogous fashion.

Since a polynomial can always be written as a product of terms of the type

K, sT, 1 + sT, 1 + 2⇣T s + s2 T 2 (for |⇣| < 1, i.e. complex roots)

it is suffices to be able to sketch Bode diagrams for these terms. The Bode plot of a
complex system is then obtained by adding the gains and phases of the individual
terms.
Note: Always rewrite the transfer function in terms of these building blocks before
starting to sketch a Bode diagram – you will find it much easier. For example, if the
transfer function has a term (s + a) first rewrite this as a ⇥ (1 + s/a) and then
collect together all the constants that have been pulled out. The transfer functions
in Question 6 on Examples Paper 2 are already given in the right form, but you will
need to rewrite the transfer function in Question 8.
500(s + 3s + 2) 500(s + 1)(s + 2)
Example: We would rewrite G(s) = 2
=
s(s + 5s + 100) s(s2 + 5s + 100)
as ✓ ◆
10 (1 + s)(1 + s/2)
G(s) = ⇥
s (1 + 0.05s + s2 /100)
and begin by considering each term individually.

12

12
 4.3.1 Powers of s: (sT )k

The simplest term in a transfer function is a power of s, which it is convenient

to write in the form
a(s) = (sT )k
where k > 0 if the term appears in the numerator and k < 0 if the term is in
the denominator. The magnitude and phase of the term are given by

log10 |a(j!)| = log10 (!T )k = klog10 (!T ), 6 G(j!) = 90k

The gain curve is thus a straight line with slope k decades/decade, or 20k
dB/decade, intersecting the 0dB line at !T = 1. The phase curve is a
constant at 90 ⇥ k. For T = 1, the case when k = 1 corresponds to a
differentiator, and has a slope 20dB/decade and phase 90 . The case when
k = 1 corresponds to an integrator and has a slope 20dB/decade and
phase 90 .

40dB (=100) The first figure on

the left illustrates
the Gain plot of
20dB (=10) a(s) = (sT )k for
k = 0, 1, 2, 2, 1.
Note that the slope is
20log10 |a(j!)| 0dB (=1) 20k dB/decade.

20dB (=.1) The second figure il-

lustrates the corre-
sponding Phase plot.
40dB (=.01) Note that 6 a(j!) =
0.01 0.1 1 10 100 90k
T T T T T

180
90
6 a(j!) 0
90
180
0.01 0.1 1 10 100
T T T T T

13

13
4.3.2 First Order Terms: (1 + sT ) For the Gain plot
the main features
Bode plot of G(s) = (1 + sT ) (for T > 0) illustrated are the
two asymptotes when
. . . replace s by j! to get G(j!) = (1 + j!T ) ! ! 0 (i.e. ! ⌧ 1/T )
and ! ! 1 (i.e.
! 1/T ). These
asymptotes intersect at
=) |G(j!)| = |1 + j!T |
| {z } the ! = 1/T .
p
1 + !2T 2 1 + j!T
6 G(j!)| = 6 (1 + j!T )
Note also that the low
j!T
| {z } frequency asymptote is
atan(!T ) horizontal and the high
frequency asymptote
Asymptotes: 1 has slope 20dB/decade.
! ! 0: (i.e. ! ⌧ 1/T )
The phase 6 G(j!)
20log10 |G(j!)| ! 20log10 1 = 0 increases from 0 to 90
6 G(j!) ! 6 1 = 0
as ! increases from 0 to
1 with the phase being
45 at ! = 1/T .

! ! 1: (i.e. ! 1/T ) See also the graphical

illustration in the next
20log10 |G(j!)| ! 20log10 |j!T | page.

= 20log10 ! 20log10 1/T

(which is a straight line with slope=20dB/decade and x-axis (i.e. 0db)

intercept at ! = 1/T )

6 G(j!) ! 6 j!T = 90

At ! = 1/T , we get

20log10 |G(j!)| = 20log10 |1 + j|

p
= 20log10 2 (3dB)
6 G(j!) = 6 (1 + j) = 45

14

14
Bode diagram of G(s) = (1 + sT ):

In both the Gain and

40dB (=100) Phase plots we illus-
high freq asymptote trate the two asymp-
( slope = 20 dB/decade) totes when ! ! 0 and
Gain

! ! 1, respectively.
20log10 |G(j!)| In the gain plot the
20dB (=10)
low freq asymptotes intersect at
asymptote ! = 1/T (follows from
the expressions derived
for the asymptotes in
0dB (=1) 3dB the previous page).
In the phase plot the
0.01 0.1 1 10 100
T T T T T phase is 45 at this
Frequency (rad/s) frequency.

10 high freq To facilitate the draw-

asymptote ing of the phase plot
T a straight line that ap-
90
6 G(j!) proximates the plot can
be drawn that passes
Phase (Degrees)

through the point with

! = 1/T , 6 G(j!) =
45 45 , it intersects the
low freq high frequency asymp-
asymptote tote at ! = 10/T ,
1 and the low frequency
asymptote at ! = 1/T
0 10T (see the phase diagram
0.01 0.1 1 10 100 on the left).
T T T T T
Frequency (rad/s)

try 3dB paint

15

15
4.3.3 Second order terms: (1 + 2⇣sT + s2 T 2) As in the case of first
order terms a main
1 feature we illustrate
Bode plot of G(s) = (for T > 0, 0  ⇣  1)
1 + 2⇣sT + s2 T 2 in the Bode plot of
second order terms,
1 in both the Gain plot
. . . replace s by j! to get G(j!) =
1 + 2⇣j!T !2T 2 and the Phase plot, are
the asymptotes when
=) 20log10 |G(j!)| = 20log10 1 ! 2 T 2 + 2⇣j!T ! ! 0 and ! ! 1,
⇣ ⌘ respectively.
6 G(j!) = 6 2 2
1 ! T + 2⇣j!T
As shown on the left
Asymptotes: the high frequency
asymptote has slope
! ! 0: (i.e. ! ⌧ 1/T ) 40dB/decade (in con-
trast to 20dB/decade
20log10 |G(j!)| ! 20log10 1 = 0 in the first order term
6 G(j!) ! 6 1=0 previously analysed).

Also the change in

phase as ! increases
from 0 to 1 is 180 ,
! ! 1: (i.e. ! 1/T ) with the phase being
90 at ! = 1/T .
20log10 |G(j!)| ! 20log10 | !2T 2|
Furthermore, the gain
= 40log10 !T at ! = 1/T depends on
= 40log10 1/T 40log10 ! the damping ratio ⇣,
and tends to infinity as
(which is a straight line with slope= 40dB/decade and x-axis (i.e. 0db) ⇣ ! 0.
intercept at ! = 1/T )
See also the graphical
illustration in the next
6 G(j!) ! 6 !2T 2 = 180 page.
At ! = 1/T , we get

20log10 |G(j!)| = 20log10 |2⇣j|

1
= 20log10
2⇣
6 G(j!) = 6 (2⇣j) = 90

16

16
 20dB The figures on the left
illustrate the Bode plot
8dB=2.5 (1/2⇣) of the second order term
⇣ = 0.2
1
G(s) =
0dB 1 + 2⇣sT + s2 T 2
6dB
Gain (dB)

for different values of ⇣.

⇣ =1
It should be noted that
20dB an overshoot is ob-
served in the Gain plot
for ⇣ < 1.

40dB

60dB0.01 0.1 1 10 100

T T T T T

⇣ = 0.2

−45

⇣=1
Phase (Degrees)

−90

−135

−180
0.01 0.1 1 10 100
T T T T T

17

17
4.3.4 Examples
5
Example 1: G(s) = (K = 5, 1/T = 0.1)
1 + 10s

20log10 |G(j!)| = 20log10 5 20log10 |1 + j!/0.1|

6 G(j!)| = 6 5 6 (1 + j!/0.1)

We now produce a sketch of the Bode diagram in two stages (this is for clarity
– all these constructions would normally appear on one pair of graphs ). First
we plot the asymptotes and approximations to the true curves for the
individual terms (The exact values are used for the plots here.)

20

20log10 |1 + 10j!|
Gain (dB)

20log10 |5|
0

−20
−3 −2 −1 0 1
10 10 10 10 10
Phase (Degrees)

0
6 1 + 10j! 6 5

−90
−3 −2 −1 0 1
10 10 10 10 10

Frequency (rad/s)

18

18
In the next pair of diagrams, the contributions from the individual terms (now
shown as dashed lines) have been added to give the Bode diagram of G(s)
(this has been done for both the asymptotes and the true gain and phase).

20

20log10 |G(j!)|
Gain (dB)

−20
−3 −2 −1 0 1
10 10 10 10 10
Phase (Degrees)

0
6 G(j!)

−90
−3 −2 −1 0 1
10 10 10 10 10

Frequency (rad/s)

19

19
The next example has both a pole and a zero.
It is an example of what is known as a “phase-lead compensator” (will be
studied again later in the course).

1 + 10s
Example 2: G(s) = 0.05
1+s

1 + 10j!
So, G(j!) = 0.05 ,
1 + j!

20log10 |G(j!)| = 20log10 0.05

+ 20log10 |1 + 10j!| 20log10 |1 + j!|

Note: 0.05 = 26dB

and
6 G(j!) = 6 0.05 +6 (1 + 10j!) 6 (1 + j!)
| {z }
0

20

20
First we draw the individual terms:

20
20log10 |1 + 10j!|
20log10 |1 + j!|
0
Gain (dB)

−20

20log10 |0.05|
−40
−3 −2 −1 0 1 2
10 10 10 10 10 10

90
6 1 + 10j!
Phase (Degrees)

6 0.05
0 6 1 + j!

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

21

21
and then we sum them (Note how the phase terms sum to produce a
T
maximum phase advance of only about 55 .)
20

0
Gain (dB)

−20 20log10 |G(j!)|

−40
−3 −2 −1 0 1 2
10 10 10 10 10 10

90
Phase (Degrees)

6 G(j!)
0

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

22

22
 1 + 2s 0.05
Example 3: G(s) = 0.05
100 = ⇥ (1 + 2s)
s s
20
20log10 |1 + 2j!|

0
Gain (dB)

0.05
20log10
j!
−20 = 20log10 0.05 20log10 !

−40
−3 −2 −1 0 1 2
10 10 10 10 10 10

90
Phase (Degrees)

6 1 + 2j!
0
0.05
6 = 6 j
j!
−90
−3 −2 −1 0 1 2
10 10 10 10 10 10
Frequency (rad/s)

23

23
100
T

20

0
Gain (dB)

20log10 |G(j!)|

−20

−40
−3 −2 −1 0 1 2
10 10 10 10 10 10

90
Phase (Degrees)

0
6 G(j!)

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

24

24
RHP poles and zeros

6 (1 j!T ) = 6 (1 + j!T )

20log10 |(1 j!T )| = + 20log10 |(1 + j!T )|

That is, if we have a term (1 sT ) instead of (1 + sT ) then the gain plot is

unchanged but the term’s contribution to the phase plot is reversed in sign (so
the contribution of a RHP zero to the overall phase diagram is the same as
that of a LHP pole at the same location).

Comments on Bode sketching

A alternative technique for Bode diagrams would be to ignore the asymptotes,

and just calculate and plot the true gain and phase over a grid of frequencies.
This is not recommended for a number of reasons. Firstly, a lot more points
are required to get the same accuracy (particularly when the diagram is to be
used for control system analysis and design, as only a small region is required
accurately in this case). Secondly, the structure of the diagram is then lost. A
practising control engineer will often prefer a good sketch, with the
asymptotes shown, to an accurate computer generated diagram – since this
gives a better idea of how things can be changed to improve the behaviour of
the controlled system.
When a question asks you to “draw” a Bode diagram (as in the questions on
Examples Paper 2) it’s really asking you to produce a drawing showing the
straight line asymptotes and approximations and a rough approximation to the
true gain and phase by rounding the corners appropriately.

25

25
4.4 Key Points
The frequency response is obtained from the transfer function by
replacing s with j!.

At each frequency !, G(j!) is a complex number whose magnitude gives

the gain of the system at that frequency and whose argument gives the
phase shift of the system at that frequency.
The gain and the phase shift are conveniently shown on the Bode
diagram.

26

26
 Part IB Paper 6: Information Engineering
LINEAR SYSTEMS AND CONTROL
Ioannis Lestas

HANDOUT 5

“An Introduction to Feedback Control Systems”

r̄(s) + ē(s) ȳ(s)

⌃ K(s) G(s)

z̄(s)
H(s)

z̄(s) = H(s)G(s)K(s) ē(s)

| {z }
L(s)
Return ratio
1
ē(s) = r̄(s)
1 + L(s)
| {z }
Closed-loop transfer function
relating ē(s) and r̄(s)
G(s)K(s)
ȳ(s) = G(s)K(s)ē(s) = r̄(s)
1 + L(s)
| {z }
Closed-loop transfer function
relating ȳ(s) and r̄(s)
1
Key Points
The Closed-Loop Transfer Functions are the actual transfer functions
which determine the behaviour of a feedback system. They relate signals
around the loop (such as the plant input and output) to external signals
injected into the loop (such as reference signals, disturbances and noise
signals).
It is possible to infer much about the behaviour of the feedback system
from consideration of the Return Ratio alone.
The aim of using feedback is for the plant output y(t) to follow the
reference signal r(t) in the presence of uncertainty. A persistent difference
between the reference signal and the plant output is called a steady state
error. Steady-state errors can be evaluated using the final value theorem.

Many simple control problems can be solved using combinations of

proportional, derivative and integral action:
Proportional action is the basic type of feedback control, but it can be
difficult to achieve good damping and small errors simultaneously.
Derivative action can often be used to improve damping of the closed-loop
system.
Integral action can often be used to reduce steady-state errors.

2
Contents
5 An Introduction to Feedback Control Systems 1
5.1 Open-Loop Control . . . . . . . . . . . . . . . . . . . . . . . 4
5.2 Closed-Loop Control (Feedback Control) . . . . . . . . . . . . 5
5.2.1 Derivation of the closed-loop transfer functions: . . . . 5
5.2.2 The Closed-Loop Characteristic Equation . . . . . . . . . 6
5.2.3 What if there are more than two blocks? . . . . . . . . 7
5.2.4 A note on the Return Ratio . . . . . . . . . . . . . . . 8
5.2.5 Sensitivity and Complementary Sensitivity . . . . . . . . 9
5.3 Summary of notation . . . . . . . . . . . . . . . . . . . . . . 10
5.4 The Final Value Theorem (revisited) . . . . . . . . . . . . . . 11
5.4.1 The “steady state” response – summary . . . . . . . . . 12
5.5 Some simple controller structures . . . . . . . . . . . . . . . . 13

5.5.1 Introduction – steady-state errors . . . . . . . . . . . . 13

5.5.2 Proportional Control . . . . . . . . . . . . . . . . . . . 14
5.5.3 Proportional + Derivative (PD) Control . . . . . . . . 17
5.5.4 Proportional + Integral (PI) Control . . . . . . . . . . 18
5.5.5 Proportional + Integral + Derivative (PID) Control . . 21

3
5.1 Open-Loop Control

Demanded By open-loop control

Output Controlled we refer to the case
(Reference) Controller “Plant” Output where we aim to control
K(s) G(s) the output of a system
r̄(s) ȳ(s) without feedback.

As explained on the
In principle, we could could choose a “desired” transfer function F (s) and use left this will not work
K(s) = F (s)/G(s) to obtain in practice since even
small disturbances and
F (s) model uncertainty will
ȳ(s) = G(s) r̄(s) = F (s)r̄(s) cause the output to de-
G(s)
viate from its desired
In practice, this will not work value (like trying to
drive a car with your
– because it requires an exact model of the plant and that there be no eyes closed!).
disturbances (i.e. no uncertainty).

Feedback is used to combat the effects of uncertainty

For example:
Unknown parameters
Unknown equations
Unknown disturbances

4
5.2 Closed-Loop Control (Feedback Control) The block diagram
on the left illustrates
Input Output a typical feedback
For Example:
disturbance disturbance control configuration.
The aim is for the
d¯i (s) d¯o (s) controlled output
Demanded Error ȳ(s) to follow a refer-
Output Control
Signal Controlled ence signal r̄(s).
(Reference) Signal Output
The plant G(s) is the
+ + system we would like
+ Controller + “Plant” +
⌃ K(s) ⌃ G(s) ⌃ to control, and K(s)
r̄(s) ē(s) ū(s) ȳ(s) is the controller that
needs to be designed.

Disturbances acting
at the output and
Figure 5.1 input of the plant
are also considered,
corresponding to
measurement and
actuation noise, re-
5.2.1 Derivation of the closed-loop transfer functions: spectively.

The transfer func-

⇥ ⇤ tions from d¯0 , d¯i , r̄, ē,
ȳ(s) = d¯o (s) + G(s) d¯i (s) + K(s)ē(s) respectively, to
the output ȳ are
ē(s) = r̄(s) ȳ(s) calculated. These
⇥ ⇤ are referred to as
=) ȳ(s) = d¯o (s) + G(s) d¯i (s) + K(s) r̄(s) ȳ(s) closed-loop transfer
✓ ◆ functions to empha-
=) 1 + G(s)K(s) ȳ(s) = d¯o (s) + G(s)d¯i (s) + G(s)K(s)r̄(s) size the fact that we
are in a closed-loop
configuration.
1 G(s)
=) ȳ(s) = d¯o (s) + d¯i (s)
1 + G(s)K(s) 1 + G(s)K(s)
G(s)K(s)
+ r̄(s)
1 + G(s)K(s)

5
Also:

ē(s) = r̄(s) ȳ(s)

1 G(s)
= d¯o (s) d¯ (s)
1 + G(s)K(s) 1 + G(s)K(s) i
✓ ◆
G(s)K(s)
+ 1 r̄(s)
1 + G(s)K(s)
| {z }
1
1 + G(s)K(s)

5.2.2 The Closed-Loop Characteristic Equation and

the Closed-Loop Poles

Note: All the C losed-Loop Transfer Functions of the previous The equation
section have the same denominator: 1 + G(s)K(s) = 0
1 + G(s)K(s)
is referred to as the
Closed-Loop Char-
The Closed-Loop Poles (ie the poles of the closed-loop system, or feedback acteristic Equation.
system) are the zeros of this denominator. Its solutions give the
For the feedback system of Figure 5.1, the Closed-Loop Poles are the roots of location of the poles in
the closed loop transfer
functions.
1 + G(s)K(s) = 0
As discussed in the pre-
vious handout, the lo-
Closed-Loop Characteristic Equation cation of the poles are
(for Fig 5.1) very important for the
behaviour of a system
as they determine its
stability properties, and
The closed-loop poles determine:
also its transient re-
The stability of the closed-loop system. sponse.

Characteristics of the closed-loop system’s

transient response.(e.g. speed of response,
presence of any resonances etc)

6
5.2.3 What if there are more than two blocks? The diagram on the
left illustrates a feed-
For Example: back configuration
where in addition
to K(s) and G(s),
Demanded there is an additional
d¯i (s) d¯o (s) Controlled
Output transfer function
Output
(Reference) H(s). The latter
could be part of the
+ Controller + + “Plant” + + implementation of
⌃ K(s) ⌃ G(s) ⌃
r̄(s) ē(s) ȳ(s) the control policy
or could include the
sensor dynamics.
Sensor
H(s)
Figure 5.2

The closed-loop
transfer functions
derived on the left
illustrate that the
We now have denominator has
G(s)K(s) a form analogous
ȳ(s) = r̄(s) to that when only
1 + H(s)G(s)K(s)
K(s) and G(s) were
1 G(s)
+ d¯o (s) + d¯ (s) present; i.e. the
1 + H(s)G(s)K(s) 1 + H(s)G(s)K(s) i closed loop poles are
the solutions to the
This time 1 + H(s)G(s)K(s) appears as the denominator of all the
equation
closed-loop transfer functions.
Let, 1 + L(s) = 0
L(s) = H(s)G(s)K(s) where L(s) is the
i.e. the product of all the terms around loop, not including the 1 at the product of all the
transfer functions
summing junction. L(s) is called the Return Ratio of the loop (and is also
round the loop (not
known as the Loop Transfer Function). including the 1
The Closed-Loop Characteristic Equation is then at the summing
junction).
1 + L(s) = 0
and the Closed-Loop Poles are the roots of this equation.

7
5.2.4 A note on the Return Ratio

0 0

+ + + + +
⌃ K(s) ⌃ G(s) ⌃
0 ē(s) ȳ(s)
b̄(s)

ā(s) H(s)

Figure 5.3
With the switch in the position shown (i.e. open), the loop is open. We then
have
ā(s) = H(s)G(s)K(s) ⇥ b̄(s) = H(s)G(s)K(s)b̄(s)

Formally, the Return Ratio of a loop is defined as 1 times the product of all
the terms around the loop. In this case

L(s) = 1⇥ H(s)G(s)K(s) = H(s)G(s)K(s)

Feedback control systems are often tested in this configuration as a final

check before “closing the loop” (i.e. flicking the switch to the closed position).

Note: In general, the block denoted here as H(s) could include filters and
other elements of the controller in addition to the sensor dynamics.
Furthermore, the block labelled K(s) could include actuator dynamics in
addition to the remainder of the designed dynamics of the controller.

8
5.2.5 Sensitivity and Complementary Sensitivity

The Sensitivity and Complementary Sensitivity are two particularly important

closed-loop transfer functions. The following figure depicts just one
configuration in which they appear.
d¯o (s)

r̄(s) + ē(s) ū(s) + + ȳ(s)

⌃ K(s) G(s) ⌃

The Sensitivity S(s)

and Complementary
Sensitivity T (s) are
Figure 5.4 defined, respectively,
✓ ◆ as
L(s) = G(s)K(s) 1
S(s) =
1 + L(s)
L(s)
T (s) =
1 + L(s)
G(s)K(s) 1 where L(s) is the
ȳ(s) = r̄(s) + d¯o (s)
1 + G(s)K(s) 1 + G(s)K(s) return ratio.
L(s) 1
= r̄(s) + d¯o (s) The block diagram
1 + L(s) 1 + L(s)
| {z } | {z } on the left illustrates
Complementary Sensitivity a configuration where
Sensitivity S(s) they appear as par-
T (s) ticular closed-loop
transfer functions.

Note: Note that the sum-

1 L(s) mation of S(s) and
S(s) + T (s) = + =1 T (s) is always equal
1 + L(s) 1 + L(s) to 1 irrespective of
the form of L(s).

9
5.3 Summary of notation
The system being controlled is often called the “plant”.
The control law is often called the “controller”; sometimes it is called the
“compensator” or “phase compensator”.
The “demand” signal is often called the “reference” signal or “command”,
or (in the process industries) the “set-point”.
The “Return Ratio”, the “Loop transfer function” always refer to the
transfer function of the opened loop, that is the product of all the transfer
functions appearing in a standard negative feedback loop (our L(s)).
Figure 5.1 has L(s) = G(s)K(s), Figure 5.2 has L(s) = H(s)G(s)K(s).

1
The “Sensitivity function” is the transfer function S(s) = . It
1 + L(s)
characterizes the sensitivity of a control system to disturbances appearing
at the output of the plant.
L(s)
The transfer function T (s) = is called the “Complementary
1 + L(s)
Sensitivity”. The name comes from the fact that S(s) + T (s) = 1. When
this appears as the transfer function from the demand to the controlled
output, as in Fig 5.4 it is often called simply the “Closed-loop transfer
function” (though this is ambiguous, as there are many closed-loop
transfer functions).

10

10
5.4 The Final Value Theorem (revisited)

Consider an asymptotically stable system with impulse response g(t) and

transfer function G(s), i.e.

g(t) *
) G(s)
|{z} | {z}
Impulse response Transfer Function
(assumed asymptotically stable)
Z t
Let y(t) = g(⌧ ) d⌧ denote the step response of this system and
0
G(s)
note that ȳ(s) = .
s

We now calculate the final value of this step response:

Z 1
lim y(t) = g(⌧ ) d⌧
t!1 0
Z 1
= exp( 0⌧ ) g(⌧ ) d⌧ = L g(t) s=0 = G(0)
0 | {z }
1
Hence,
Transfer Function
Final Value of Step Response ⌘
evaluated at s = 0

“Steady-State Gain” or “DC gain”

Note that the same result can be obtained by using the Final Value Theorem:
✓ ◆
for any y for which
lim y(t) = lim sȳ(s) both limits exist.
t!1 s!0
G(s)
= lim s· = G(0)
s!0 s

11

11
5.4.1 The “steady state” response – summary

The term “steady-state response” means two different things, depending on

the input.
Given an asymptotically stable system with transfer function G(s):

The steady-state response of the system to a step input that is equal to

U for t 0, is a constant G(0)U .

The steady-state response of the system to a sinusoidal input cos(!t) is

the sinusoid |G(j!)| cos !t + arg G(j!) .
These two statements are not entirely unrelated, of course: The steady-state
gain of a system to a step input, G(0), is the same as the frequency response
evaluated at ! = 0 (i.e. the DC gain).

12

12
5.5 Some simple controller structures

5.5.1 Introduction – steady-state errors

r̄ + ē ū ȳ
⌃ K(s) G(s)

Return Ratio: L(s) = G(s)K(s).

L(s) 1
CLTFs: ȳ(s) = r̄(s) and ē(s) = r̄(s)
1 + L(s) 1 + L(s)

Steady-state error:(for a step demand) If r(t) = H(t), then

L(s)
ȳ(s) = 1+L(s) ⇥ 1s and so

L(s) 1 L(0)
lim y(t) = s⇥ ⇥ =
t!1 1 + L(s) s s=0 1 + L(0)
and

1 1 1
lim e(t) = s⇥ ⇥ =
t!1 1 + L(s) s s=0 1 + L(0)
| {z }

Steady-state error
(using the final-value theorem.)

Note: These particular formulae only hold for this simple configuration –
where there is a unit step demand signal and no constant disturbances
(although the final value theorem can always be used).

13

13
5.5.2 Proportional Control Proportional Control
is the simplest con-
trol policy where the
K(s) = kp controller is just a
multiplication with a
r̄ + ē ū ȳ constant kp .
⌃ kp G(s)
As discussed on the left
and in the next page
there are tradeoffs asso-
ciated with the choice
of kp . A large kp leads
Typical result of increasing the gain kp , (for control systems where G(s) is a smaller steady-state
itself stable): error, but this is at the
expense of a more oscil-
Increased accuracy of control. “good” latory response (or even

}
Increased control action. possibly instability).

Reduced damping. “bad” It will be seen later

in this handout and
Possible loss of closed-loop stability for large kp .
also in handouts 6 and
7, that more advanced
Example: controllers can better
1 manage these tradeoffs.
G(s) =
(s + 1)2
(A critically damped 2nd order system)
1
kp
kp G(s) (s + 1)2
ȳ(s) = r̄(s) = r̄(s)
1 + kp G(s) 1
1 + kp
(s + 1)2
kp
= 2 r̄(s)
s + 2s + 1 + kp

So, !n2 = 1 + kp , 2⇣!n = 2 movement of A diagram that shows

closed-loop poles the closed-loop poles as
p 1 for increasing kp parameter kp changes is
=) !n = 1 + kp , ⇣ = p X
1 + kp called a root locus dia-
p gram.
Closed-loop poles at s = 1 ± j kp
”root locus diagram”

14

14
Steady-state errors using the final value theorem:
kp
ȳ(s) = 2 r̄(s)
s + 2s + 1 + kp
and
1 (s + 1)2
ē(s) = = 2 r̄(s).
1 + kp G(s) s + 2s + 1 + kp final value of
So, if r(t) = H(t), step response
kp kp
lim y(t) = 2 =
t!1 s + 2s + 1 + kp s=0 1 + kp
and
(s + 1)2 1
lim e(t) = 2 =
t!1 s + 2s + 1 + kp s=0 1 + kp
| {z }

1
Steady-state error
(Note: L(s) = kp =) L(0) = kp ⇥ 1 = kp )
(s + 1)2

Hence, in this example, increasing kp gives smaller steady-state errors, but a

more oscillatory transient response .

• However, by using more complex controllers it is usually possible to remove

steady state errors and increase damping at the same time:

To increase damping –
can often use derivative action (or velocity feedback).
To remove steady-state errors – can often use integral action.

15

15
 1
For reference, the step response: (i.e. response to r̄(s) = ) is given by
s
kp kp
(2 + s)
1 + kp 1 + kp
ȳ(s) = 2
+
s + 2s + 1 + kp s
so
!
kp p 1 p kp
y(t) = exp( t) cos( kp t) + p sin( kp t) +
1 + kp kp 1 + kp
r ⇣ ⌘
kp p kp
= 1+kp exp( t) cos( kp t ) +
1 + kp
| {z } | {z }
Transient Response Steady-state response
where = arctan p1
kp
But you don’t need to calculate this to draw the conclusions we have made.

16

16
5.5.3 Proportional + Derivative (PD) Control In this case, in addition
to the multiplication
K(s) = kp + kd s
of the input to the
controller with a con-
r̄ + + ū ȳ stant gain kp , we have
⌃ kp ⌃ G(s)
the addition of a term
+ where the input to the
controller is differenti-
ated and multiplied by
kd s a constant kd .

As with proportional
control there are trade-
Typical result of increasing the gain kd , (when G(s) is itself stable): offs associated with the
choice of the gain kd
Increased Damping. within the derivative
term. In particu-
Greater sensitivity to noise. lar, increasing kd makes
the response less oscil-
(It is usually better to measure the rate of change of the error directly if
latory, but this is at
possible – i.e. use velocity feedback) the expense of amplify-
ing noise.
1
Example: G(s) = , K(s) = kp + kd s
(s + 1)2

(kp + kd s) 1 2
K(s)G(s) (s+1)
ȳ(s) = r̄(s) = r̄(s)
1 + K(s)G(s) 1 + (kp + kd s) 1 2
(s+1)
kp + kd s
= 2 r̄(s)
s + (2 + kd )s + 1 + kp

So, !n2 = 1 + kp , 2c!n = 2 + kd =)

p 2 + kd Xkp = 4, kd = 0
!n = 1 + kp , c = p
2 1 + kp movement of
closed-loop poles
for increasing kd

17

17
5.5.4 Proportional + Integral (PI) Control

In the absence of disturbances, and for our simple configuration,

1
ē(s) = r̄(s)
1 + G(s)K(s)
Hence,
steady-state error, 1 1
(for step demand) = 1 + G(s)K(s) =
1 + G(0)K(0)
s=0

To remove the steady-state error, we need to make

K(0) = 1 (assuming G(0) 6= 0).
e.g
k
K(s) = kp + i
s

r̄ + + ū ȳ
⌃ kp ⌃ G(s)
+

ki /s

Example:
1
G(s) = , K(s) = kp + ki /s
(s + 1)2

18

18
 (kp + ki /s) 1 2
K(s)G(s) (s+1)
ȳ(s) = r̄(s) = r̄(s)
1 + K(s)G(s) 1 + (kp + ki /s) 1 2
(s+1)
kp s + ki
= r̄(s)
s(s + 1)2 + kp s + ki

1 1
ē(s) = r̄(s) = 1
r̄(s)
1 + K(s)G(s) 1 + (kp + ki /s)
(s+1)2
s(s + 1)2
= r̄(s)
s(s + 1)2 + kp s + ki

Hence, for r(t) = H(t),

kp s + ki
lim y(t) = = 1
t!1 s(s + 1)2 + kp s + ki s=0

and
s(s + 1)2
lim e(t) = = 0
t!1 s(s + 1)2 + kp s + ki s=0

=) no steady-state error

19

19
PI control – General Case
The presence of an inte-
In fact, integral action (if stabilizing) always results in zero steady-state error, grator in the controller,
in the presence of constant disturbances and demands, as we shall now show. ensures that there is no
steady-state error if the
Assume that the following system settles down to an equilibrium with closed-loop system is
asymptotically stable.
lim e(t) = A 6= 0, then:
t!1 di (t) do (t)
This follows by contra-
A
diction. In particular,
r(t)+ + u(t)+ + + + if the error e(t) tends
⌃ kp ⌃ ⌃ “G(s)” ⌃
e(t) to a constant value that
+ is non-zero, the output
of the integrator would
tend to infinity, which
“ki /s”
contradicts the asymp-
totic stability of the sys-
tem.
“H(s)”

=) Contradiction

(as system is not in equilibrium)

Hence, with PI control the only equilibrium possible has
lim e(t) = 0.
t!1

provided the closed-loop

That is, limt!1 e(t) = 0
system is asymptotically stable.

20

20
5.5.5 Proportional + Integral + Derivative (PID)
A PID controller can
Control combine the advan-
tages of derivative and
ki integral action. Ap-
K(s) = kp + + kd s
s propriate tuning of the
controller parameters is
though important.
ki /s
Systematic methods for
+ designing such dynamic
r̄ + + ū ȳ controllers will be dis-
⌃ kp ⌃ G(s) cussed towards the end
+ of the course.

kd s

Characteristic equation:

1 + G(s)(kp + kd s + ki /s) = 0
• can potentially combine the advantages of both derivative and integral
action:
but can be difficult to “tune”.
There are many empirical rules for tuning PID controllers (Ziegler-Nichols for
example) but to get any further we really need some more theory . . .

21

21
 Part IB Paper 6: Information Engineering
LINEAR SYSTEMS AND CONTROL
Ioannis Lestas

HANDOUT 6

“Feedback stability and the Nyquist diagram”

If L(s) is stable , then:

(either marginally or asymptotically)
= = =

1 < 1 < 1 <

L(j!) L(j!) L(j!)

L(s) L(s) L(s)

=) is =) is =) is
1 + L(s) 1 + L(s) 1 + L(s)
asymptotically marginally unstable
stable stable

That is, the closed-loop system is stable if the Nyquist diagram of the return
ratio doesn’t enclose the point “ 1”.

1
Summary
The Nyquist diagram of a feedback system is a plot of ⇣the frequency
⌘
response of the return ratio, with the imaginary part = L(j!) plotted
⇣ ⌘
against the real part < L(j!) on an Argand diagram (that is, like the
Bode diagram, it is a plot of an open-loop frequency response).

The Nyquist stability criterion states that, if

the open-loop system is asymptotically stable (i.e. the return ratio
L(s) has all its poles in the LHP) and
the Nyquist diagram of L(j!) does not enclose the point “ 1”,
then the closed-loop system will be asymptotically stable (i.e. the
closed-loop transfer function L(s)/ 1 + L(s) will have all its poles in the
LHP)

The real power of the Nyquist stability criterion is that it allows you to
determine of the stability of the closed-loop system from the behaviour of
the open-loop Nyquist diagram. This is important from a design point of
view, as it relatively easy to see how changing K(s) affects
L(s) = H(s)G(s)K(s), but difficult to see how changing K(s) affects
L(s)/(1 + L(s)) directly, for example.

In addition, the Nyquist diagram also allows more detailed information

about the behaviour of the closed-loop system to be inferred. For example
Gain and phase margins measure how close the Nyquist locus gets to
1 (and hence how close the closed loop system is to instability).

2
Contents

6 Feedback stability and the Nyquist diagram 1

6.1 The Nyquist Diagram . . . . . . . . . . . . . . . . . . . . . . 4
6.1.1 Sketching Nyquist diagrams . . . . . . . . . . . . . . . 7
6.2 Feedback stability . . . . . . . . . . . . . . . . . . . . . . . . 8
6.2.1 Significance of the point “ 1” . . . . . . . . . . . . . . 9
6.2.2 Example: . . . . . . . . . . . . . . . . . . . . . . . . . 9
6.3 Nyquist Stability Theorem (informal version) . . . . . . . . . . 12
6.4 Gain and Phase Margins . . . . . . . . . . . . . . . . . . . . . 13
6.4.1 Gain and phase margins from the Bode plot . . . . . . 14
6.5 Performance of feedback systems . . . . . . . . . . . . . . . . 15
6.5.1 Open and closed-loop frequency response . . . . . . . . 16
6.6 The Nyquist stability theorem (for asymptotically stable L(s)) 19

6.6.1 Notes on the Nyquist Stability Theorem: . . . . . . . . 20

3
6.1 The Nyquist Diagram

The Nyquist diagram of a system G(s) is a plot of the frequency response

The Nyquist diagram
G(j!) on an Argand diagram. of a transfer function
That is: it is a plot of =(G(j!)) vs <(G(j!)). G(s) is a plot of the fre-
quency response G(j!)
= G(j!) but with different quan-
tities on the x and y
axis relative to a Bode
6 G(j!1 ) plot. In particular, it
is a plot of G(j!) on
< G(j!) an Argand diagram, i.e.
a plot of the real vs
the imaginary part of
|G(j!1 )| G(j!).
The Nyquist locus
! = !1 As with the case
of Bode plots, on a
Nyquist diagram we
illustrate the behaviour
of G(j!) as ! ! 0
and as ! ! 1. It is
Examples:
= G(j!) convenient to consider
Integrator: the magnitude and
phase of G(j!) in these
1
G(s) = regimes in order to
s draw the diagram .
1 j
G(j!) = = !=1
j! ! < G(j!)
|G(j!)| = 1/!
6 G(j!) = 90
20dB
j
!=1
!
0dB0.1 1 10

!
0 0.1 1 10
90

4
 Time Delay: = G(j!)
The Nyquist diagram of
G(s) = e sT a system with transfer
G(j!) = e j!T function e sT , i.e. a de-
lay T , is a circle with
|G(j!)| = 1
centre the origin and ra-
arg G(j!) = !T < G(j!) dius 1. Note that the
(radians) phase of e j!T tends to
!=0 1 as ! ! 1.
20dB
! ⇡
0dB0.1 1 10
T T T !=
2T
!
0 0.1 1 10
T T T
90

A system with transfer

First-order lag: function 1/(1 + sT ) is
1 |G(j!)| 6 G(j!) referred to as a first-
G(s) = order lag. The Nyquist
1 + sT !=0 1 0
1 diagram of such a sys-
G(j!) = 1 1 tem is illustrated on the
1 + j!T != p 45 left. It can be shown
1 T 2
|G(j!)| = p ! =! 1 0 90 that this is part of a cir-
1 + !2T 2 cle centred at the point
6 G(j!) = arctan(!T ) = G(j!) 0.5 with radius 0.5.
20dB
!
0dB0.1 1 10
T T T !=1
< G(j!)
0.5
!
0 0.1 1 10 !=0
T T T 0.5j
90
! = 1/T

5
Time Delay with Lag and Integrator:

e sT1
G(s) =
s(1 + sT2 )

e j!T1
G(j!) =
j!(1 + j!T2 )

1 1
|G(j!)| = |e j!T1 | ⇥ ⇥
| {z } |j!| |1 + j!T2 |
1

6 G(j!) = 6 e j!T1 6 (j!) 6 (1 + j!T2 )

| {z } | {z }
!T1 90 When the transfer func-
Clearly, as ! ! 0 then |G(j!)| ! 1. But this is not enough information to tion has a pole at s = 0
then |G(j!)| ! 1
sketch the Nyquist diagram. Precisely how does |G(j!)| ! 1? To answer
as ! ! 0. In order
this, we use a Taylor series expansion around ! = 0. to determine in what
direction G(j!) tends
to 1 we do a Taylor ex-
e j!T1 ! 1 j!T1 and pansion around ! = 0
1/(j!T2 + 1) ! 1 j!T2 , and consider the real
and imaginary parts of
0 G(j!).
(1 j!T1 )(1 j!T2 ) 1
=) G(j!) ! = (T1 + T2 ) + j!T1 T2 .
j! j! In the example on the
left we see that as ! !
0 the real part is con-
stant and the imagi-
= G(j!) nary part tends to 1.
There is therefore a ver-
tical asymptote as illus-
trated in the diagram.
(T1 + T2 )
< G(j!)

6
Second-order lag:
1
G(s) =
(1 + sT1 )(1 + sT2 )
1
G(j!) =
(1 + j!T1 )(1 + j!T2 )
1
|G(j!)| = q q
1 + ! T1 1 + ! 2 T22
2 2 = G(j!)

6 G(j!) = arctan(!T1 ) arctan(!T2 )

!=1
< G(j!)

|G(j!)| 6 G(j!)

!=0
!=0 1 0
! =! 1 0 180

6.1.1 Sketching Nyquist diagrams

Unlike the Bode diagram, there are no detailed rules for sketching Nyquist
diagrams. It suffices to determine the asymptotic behaviour as ! ! 0 and
! ! 1 (using the techniques we have seen in the examples) and then
calculate a few points in between. Note that if G(0) is a finite and non-zero,
then the Nyquist locus will always start off by leaving the real axis at right
angles to it. 1 If G(0) is infinite, due to the presence of integrators, then

we must explicity find the first two terms of the Taylor series expansion of
G(j!) about ! = 0, as in the example with a time delay, a lag and an
integrator.
1
This is since G(j✏) = G(0) + j✏G0 (0) ✏2 G00 (0) · · · ⇡ G(0) + j✏G0 (0)

7
6.2 Feedback stability The remainder of this
handout will focus on
the Nyquist Stability
r̄(s)+ ē(s) ȳ(s) r̄(s)+ ē(s) ȳ(s)
⌃ K(s) G(s) ⌃ L(s) Theorem. This is a
⌘ very important result
in control engineering
that facilitates control
system design.

Its key feature is that

G(s)K(s) allows to deduce the
Closed-loop poles ⌘ poles of stability of the closed-
1 + G(s)K(s)
loop system from the
⌘ roots of 1 + G(s)K(s) = 0 frequency responce of
the return ratio L(j!),
i.e. from open-loop
properties.
It is difficult to see how K(s) should be chosen to ensure that all the
In the next few pages
closed-loop poles are all in the LHP. But . . .
we will discuss the
intuition behind the
Nyquist’s Stability Theorem allows us to deduce closed-loop properties: theorem, and state first
an informal version of
G(s)K(s) the theorem; i.e. a
the location of the poles of ,
1 + G(s)K(s) simpler statement that
holds for the systems
from open-loop properties we will come across in
this course. Towards
frequency response of the return ratio the end of the handout
L(j!) = G(j!)K(j!). we will state also the
full version of the theo-
The basic idea is as follows: Negative feedback is used to reduce the size of rem.
the error e(t) in the above figures. If y(t) is too large (i.e greater than r(t))
then e(t) is negative, which will tend to reduce y(t) (provided the signs of It should be noted that
we will not provide the
K(s) and G(s) have been chosen appropriately). However, for any real
full proof of the Nyquist
system the phase lag from the input to the output ( 6 L(j!)) will tend to Stability Theorem as
increase with frequency, eventually reaching 180 . When this happens, the this is quite involved
negative feedback is turned into positive feedback. If the gain |L(j!)| has not and is outside the scope
decreased to less than 1 by this frequency then instability of the closed-loop of this course.
system will result.

8
6.2.1 Significance of the point “ 1”

If the Nyquist locus passes through the point “ 1”, i.e. if

L(j!1 ) = 1 for some !1
then the closed-loop frequency response L(j!)/ 1 + L(j!) becomes infinite
at that frequency, ie
L(j!1 )/ 1 + L(j!1 ) ! 1 This is not a good thing!
In this case, if e(t) = cos(!1 t) then in steady-state we have
⇣ ⌘
y(t) = |L(j!1 )| cos !1 t + 6 L(j!1 )

= cos(!1 t + ⇡) = cos(!1 t).

However e(t) = r(t) y(t), which means that

r(t) = e(t) + y(t) = cos(!1 t) cos(!1 t)= 0.

That is, there is a sustained oscillation of the feedback system even when
there is no external input!

6.2.2 Example:
In the example on the
Let 1 left we will investi-
G(s) = 3 2
, K(s) = k,
s + s + 2s + 1 gate the position of
k the closed-loop poles,
=) L(s) = 3 . and hence the stability
s + s2 + 2s + 1
properties of the closed-
The closed-loop poles are the roots of
loop system, as param-
k eter k varies. In each
1+ 3 = 0 () s3 + s2 + 2s{z+ 1 + k = 0}
2
s + s + 2s + 1 | case we will also plot
CLCE the Nyquist diagram of
and the frequency response of the loop is: L(j!) and consider its
shape relative to the
k point 1.
L(j!) =
j( ! 3 + 2!) + ( ! 2 + 1)
p
At ! = 2, L(jw) is purely real. That is
p k
L( 2 j) = p p = k
j( 2 2 + 2 2) 2+1

9
 k=1
Nyquist diagram, k=1 Closed-loop step response
2 2

p L(j!) 1
L(
1 2 j)
0
Imag Axis

0 −1
0 10 20 30 40 50
L(s) 1
−1
L 1 ·
1 + L(s) s

−2
−2 −1 0 1 2
Real Axis

Closed-loop poles are at the roots of s3 + s2 + 2s + 2 = 0, i.e.

p
s= 0.0000 + 1.4142j, (because 1 + L( 2j) = 1 p
0.0000 1.4142j, and so 1 + L(s) = 0 at s = 2j)
1.0000
=)

Closed-loop poles

=) closed-loop system is marginally stable

10

10
 k = 0.8
Nyquist diagram, k=0.8 Closed-loop step response
2 2

1
1
0
Imag Axis

0 −1
0 10 20 30 40 50
Closed-loop is
−1 asymptotically stable

−2
−2 −1 0 1 2
Real Axis

Closed-loop poles are at the roots of s3 + s2 + 2s + 1.8 = 0, i.e.

s = 0.0349 + 1.3906j, 0.0349 1.3906j, 0.9302

k = 1.2
Nyquist diagram, k=1.2 Closed-loop step response
2 2

1
1
0
Imag Axis

0 −1
0 10 20 30 40 50
Closed-loop is
−1 unstable

−2
−2 −1 0 1 2
Real Axis

Closed-loop poles are at the roots of s3 + s2 + 2s + 2.2 = 0, i.e.

s = 0.0319 + 1.4377j, 0.0319 1.4377j, 1.0639

11

11
6.3 Nyquist Stability Theorem (informal
version)
We can now give an informal statement of Nyquist’s stability theorem:
“If a feedback system has an asymptotically stable return ratio L(s), then the
feedback system is also asymptotically stable if the Nyquist diagram of L(j!)
leaves the point 1 + j0 on its left”.
This is unambiguous in most cases, and usually still works if L(s) has poles at
the origin or is unstable.
For completeness, a full statement of this theorem will be given later.
Definition: We say that the feedback system (or closed-loop system) is
L(s)
asymptotically stable if the closed-loop transfer function is
1 + L(s)
L(s)
asymptotically stable, that is if all the poles of (i.e. the roots of
1 + L(s)
1 + L(s) = 0) lie in the LHP.

12

12
6.4 Gain and Phase Margins

L(j!) encirling or going through the 1 point is clearly bad, leading to the
closed-loop not being asymptotically stable. However, L(j!) coming close to
1 without encircling it is also undesirable, for two reasons:

It implies that a closed-loop pole will be close to the imaginary axis and
that the closed-loop system will be oscillatory.

If G(s) is the transfer function of an inaccurate model, then the “true”

Nyquist diagram might actually encircle 1.

Gain and phase margins are widely used measures of how close the return
ration L(j!) gets to 1.
The gain margin measures how much the gain of the return ratio can be
increased before the closed-loop system becomes unstable.
The phase margin measures how much phase lag can be added to the return
ratio before the closed-loop system becomes unstable.

The diagram on the left

illustrates how the gain
< margin and phase mar-
✓ gin of a feedback system
1 ↵
can be deduced from
the Nyquist diagram of
its return ratio L(s).
L(j!)

1
Gain Margin = Phase Margin = ✓
↵
In this example we have ✓ = 35 and ↵ = 0.75. Hence
Phase Margin = 35 and Gain Margin = 1/0.75 = 4/3.

13

13
6.4.1 Gain and phase margins from the Bode plot The diagram on the
|L(j!)| (dB) left illustrates how the
gain margin and phase
margin of a feedback
system can be deduced
0dB also from the Bode plot
Gain Margin
of its return ratio L(s).

10dB For the gain margin we

find first the frequency
6 L(j!)
where the phase of
0 L(j!) is 180 . We
log10 ! then find |L(j!)| at
this frequency.

For the phase margin

find first the frequency
Phase Margin where |L(j!)| = 1. We
180
then find the phase of
L(j!) at this frequency
Gain Margin = 20log10 4/3 = 2.5dB. Phase Margin = 35 (as before)
and how much it devi-
ates from 180 .
Hint: Given a Nyquist diagram of L(s) = kG(s) for k = 1, it is easy
to find gain and phase margins for k 6= 1 (just look at the “ 1/k” point
instead of “-1”). =

“ 1/k” 0.75 <

1 ✓ (=phase margin when k = 0.8)

L(j!)

1.25 = 5/3 (= 4.4dB), and

If k = 0.8, as here, then Gain Margin= 0.75
Phase Margin=80 .

14

14
6.5 Performance of feedback systems
The slide on the left
Good shows that a small mag-
“Small” 1
nitude in the frequency
feedback () sensitivity
properties 1 + L(j!) response of the sensitiv-
ity function
For 1) rejection of disturbances.
1
S(s) =
1 + L(s)
d̄(s) leads to good closed-
ȳ(s) loop properties.
L(s)
In particular, it leads
−
to a rejection of output
disturbances as the
1 sensitivity S(s) is the
Transfer function Transfer function
with f/b = ⇥ without f/b
closed-loop transfer
1 + L(s) function from d¯ to ȳ. It
Plus, 2) reducing the effects of uncertainty. also reduces the effect
of parameter uncer-
tainty (second point in
– if L(s) depends on an uncertain parameter (eg the slide).
1
L(s) = 2 ) then
s +2 s+1

d L
(1 + L) ⇥ dL (L) ⇥ dL . L
d 1+L = d d
L (1 + L)2 1+L
| 1+
{z L } d
L
relative change 1 d
=
in closed-loop
|1 +
{z L} L}
| {z
S relative change
in open-loop

Good design aims for sensitivity reduction over an appropriate range of

frequencies
1
Typically, by requiring that ⌧ 1 for ! < !1 where !1 here
1 + L(j!)
denotes the desired control bandwidth.

15

15
6.5.1 The relationship between open and closed-loop
frequency responses

Ultimately what we are always interested in are properties of the closed-loop

system, such as its frequency response and pole locations. The following plots
are representative of a typical feedback system, and correspond to a feedback
system with a Return Ratio of
2
L(s) =
s(1 + s)
As is typical, the feedback reduces the effect of disturbances at low
frequencies, up to !1 , as evident from the plot of Sensitivity S(j!). !1 is
defined here as the lowest frequency at which |S(j!)| = 1. The closed-loop
system will respond to reference inputs at frequencies up to around !2 , as
evident from the plot of the Complementary Sensitivity T (j!). !2 is defined
here as the highest frequency at which |T (j!)| = 1. Between these
frequencies both disturbances and reference signals are amplified.
The actual value of the frequencies !1 and !2 , and the size of these peaks,

can be determined directly from the open-loop frequency response.

1
A: |S(j!)| = |1+L(j!)| = 1 when |1 + L(j!)| = 1, which is when the
distance from the point 1 to the Nyquist locus equals 1 (this is the point
! = !1 overleaf).
|L(j!)|
B: |T (j!)| = |1+L(j!)| = 1 when |L(j!)| = |1 + L(j!)|, which is when the
distance from the point 1 to the Nyquist locus equals the distance from
the origin to the Nyquist locus (this is the point ! = !2 overleaf).
1
C: |S(j!)| = |1+L(j!)| is maximized when |1 + L(j!)| is minimized, that is
when the distance from the point 1 to the Nyquist locus is at a minimum.
|L(j!)|
D: The easiest way to find the maximum value of |T (j!)| = |1+L(j!)| is
probably to try a few points around where |1 + L(j!)| is minimized.

16

16
 2
Nyquist diagram of the return ratio L(s) = s(1+s)

=
! = !2 “1”
1 0.5 <

“1 + L(j!1 )” “L(j!1 )”
! = !1
L(j!)

1 L(j!)
! L(j!) 1+L(j!) 1+L(j!)
1 1 j j 1 j
1.732 .5 .289j 1.5 + .866j .5 .866j
1 L(j!)
Closed-loop frequency responses: S(j!) = 1+L(j!) , T (j!) = 1+L(j!)

dB |T (j!)| |S(j!)|
d¯o (s)
r̄(s) + ē(s) + + ȳ(s)
⌃ L(s) ⌃
!
0dB
!1 !2

17

17
Small gain and/or phase margins correspond to there being frequencies at
which L(j!) comes close to the 1. We now see that this also corresponds
to making |1 + L(j!)| small and hence there being resonant peaks in the
closed-loop transfer functions.
So,
Small gain and/or phase margins are
bad for robustness, and
bad for performance.

18

18
6.6 The Nyquist stability theorem (for
asymptotically stable L(s))
On page 12 we gave an informal statement of the Nyquist stability criterion.
The formal statement of the Nyquist stability theorem requires counting
encirclements of the point 1:
As before, we take L(s) to be the return ratio so that the closed-loop
characteristic equation is 1 + L(s) = 0.

We make the following simplifying assumption:

L(s) is asymptotically stable
This also guarantees that L(j!) is finite for all ! ( i.e. it has no j!-axis
poles), and that L(1) is finite (since L(s) must be proper – see Handout 4).

Under this condition the “full” Nyquist diagram of L(j!), for

1 < ! < +1, is a closed curve (since L(j1) = L( j1) = L(1)).

Note that, since ( j!) = (j!)⇤ , it follows that L( j!) = L(j!)⇤ . So the
section of the Nyquist locus for ! < 0 is the reflection in the real axis of the
section for ! > 0.

With this assumption we have:

The Nyquist Stability Theorem (for stable L(s))

Consider a feedback system with an asymptotically stable return ratio

L(s). In this case, the feedback system is asymptotically stable (i.e.
L(s)
1+L(s)
is asymptotically stable) if and only if the point 1 + j0 is
not encircled by the “full” Nyquist diagram of L(j!), for
1 < ! < +1.

19

19
6.6.1 Notes on the Nyquist Stability Theorem:

1. Encirclements must be ‘added algebraically’. If there is 1 clockwise and 1 The points mentioned
anticlockwise encirclement then they ‘add up’ to 0 encirclements. on the left will be illus-
trated via examples in
2. L(s) often has one or more poles at 0 (due to integrators in the plant or the next two pages.
the controller). The theorem still works, but one has to worry about what
happens to the graph of L(j!) at ! = 0 (as the locus is no longer a
closed curve). It can be shown that, if L(s) has n poles at the origin,
then the Nyquist locus should be completed by adding a large n ⇥ 180
arc, in a clockwise direction.

3. If L(s) is unstable, and has np unstable poles, then the theorem must be
modified as follows: “The feedback system is stable if and only if the ‘full’
Nyquist diagram makes np anticlockwise encirclements of the point
( 1 + 0j).”

4. (This is a repeat of the informal statement of the Nyquist stability

criterion from page 12.)
A potentially ambiguous statement of the theorem, but one which almost
always works, is: “The feedback system is stable if the Nyquist diagram of
L(j!) ‘leaves the point 1 + j0 on its left’ ”. This still works (usually) if
there are poles at the origin or even if L(s) is unstable.
This informal version of the theorem is adequate for the
examples which follow, for any others that you will encounter on
this course and indeed for any you are likely to encounter in
practice.

20

20
Examples: Formal application of Nyquist stability theorem.

= =
The solid black line
shows the Nyquist di-
agram for positive fre-
quencies and the green
1 < 1 < dashed line shows the
diagram for negative
frequencies. The arrow
illustrates the direction
of the plot as the fre-
quency increases.

no encirlements of 1 2 clockwise encirclements of 1

L(s) L(s)
=) is =) is
1 + L(s) 1 + L(s)
asymptotically unstable
stable

= = In the fourth example

there is a pole at the ori-
gin and hence the mag-
nitude in the Nyquist
diagram tends to 1 as
1 < 1 < ! ! 0. It should
be noted that here we
have applied the sec-
ond point in the previ-
ous page; i.e. an arc was
drawn in a clockwise
direction (dotted line)
1 clockwise + 1 anticlockwise no encirlements of 1 (note 2) to complete the Nyquist
encirclement of 1 L(s) diagram.
i.e. 0 net encirclements (note 1) =) is
1 + L(s)
L(s) asymptotically
=) is
1 + L(s) stable
asymptotically stable

21

21
Examples: Informal application of the Nyquist stability theorem (based on
The diagrams on the
note 4).
left illustrate that the
= = informal statement of
the Nyquist stability
criterion is also valid for
the examples that have
been considered.
1 < 1 <

-1 to left of locus -1 to right of locus

L(s) L(s)
=) is =) is
1 + L(s) 1 + L(s)
asymptotically stable unstable

= =

1 < 1 <

-1 to left of locus -1 to left of locus

L(s) L(s)
=) is =) is
1 + L(s) 1 + L(s)
asymptotically stable asymptotically stable

Hence: the informal application of the Nyquist stability criterion works for all
these cases.

22

22