INDEX

S. No Topic Page No
Week 1
1 Introduction to Systems and Control 1
2 Modelling of Systems 17
3 Elements of Modelling 27
4 Examples of Modelling 48
5 Solving Problems in Modelling of Systems. 73
6 Laplace Transforms 86
7 Inverse Laplace Transforms 109
8 Transfer Function of Modelling Block Diagram Representation 125
9 Solving Problems on Laplace Transforms and Transfer Functions 156
10 Block Diagram Reduction, Signal Flow Graphs 169
Week 2
11 Solving Problems on Block Diagram Reduction, Signal Flow Graphs 188
12 Time Response Analyzsis of systems 201
13 Time Response specifications 232
14 Solving Problems on Time Response Analyzsis ans specifications 256
Week 3
15 Stability 267
16 Routh Hurwitz Criterion 289
17 Routh Hurwitz Criterion T 1 300
18 Closed loop System and Stability 318
19 Root Locus Technique 344
Week 4
20 Root Locus Plots 356
21 Root Locus Plots(contd) 373
22 Root Locus Plots(contd.) 390
23 Root Locus Plots(contd..) 403
24 Introduction to Frequency Response 419
Week 5
25 Frequency Response Plots 439
26 Relative Stability 467
27 Bode plots 489
28 Basics of Control design Proportional, Integral and Derivative Actions 531
 Week 6
29 Basics of Control design Proportional, Integral and Derivative Actions 549
30 Problems on PID Controllers 563
31 Basics of Control design Proportional , Integral and Derivative Actions 578
32 Control design in time domain and discusses the lead compensator 607
Week 7
33 Improvement of the Transient Response using lead compensation 624
34 Design of control using lag compensators 642
35 The design of Lead-Lag compensators using root locus 653
36 Introduction design of control in frequency domain 667
Week 8
37 Design of Lead Compensator using Bode Plots 684
38 Design of Lag Compensators using Bode Plots 706
39 Design of Lead-Lag Compensators using Bode plots 721
40 Experimental Determination of Transfer Function 741
Week 9
41 Effect of Zeros on System Response 759
42 Navigation - Stories and Some Basics 782
43 Navigation - Dead Reckoning and Reference Frames 803
44 Inertial Sensors and Their Characteristics 832
45 Filter Design to Attentuate Inertial Sensor Noise 858
46 Complementary Filter 878
Week 10
47 Complementary Filter 1 906
48 Introduction to State Space Systems 919
49 Linearization of State Space Dynamics 960
50 Linearization of State Space Dynamics 1 984
51 State Space Canonical Forms 1013
Week 11
52 State Space Solution and Matrix Exponential 1026
53 Controllability and Pole Placement 1034
54 Controllable Decomposition and Observability 1055
 Control Engineering
Dr. Ramkrishna Pasumarthy
Department of Electrical Engineering
Indian Institute of Technology, Madras

Module – 01
Lecture – 01
Introduction to Systems and Control

Hello everybody. Welcome to this first lecture on Control Engineering. So, what we will do
is we will just keep it very short, a series of short lectures about 20-25 minutes each
sometimes even less. So, we will begin with a little introduction to Systems and Control, like
what kind of tools you might eventually use in the course and some very basics or some very
basic mathematical tools as well.

(Refer Slide Time: 00:53)

So, let us start by defining- what is a system? It could mean several things, but we will
restrict what makes sense for us in the course. So, by definition, it could be a collection of
some physical components, biological components, sometimes even abstract components;
which talk to each other to kind of perform an intended objective. From the control point of
view, we can say that a system when subject to a certain input also called excitation gives us
a certain output. We can also call it a response of a system and usually schematically we can
draw it as a diagram like this.

1
Systems could also be composed of several subsystems, and we will eventually see physical
examples of this. So, like one system could be connected this way SS1 you know it could be
connected this way or this way you know SS3 to SS4 and so on. So, the overall input will
give me some kind of overall response of the system.

(Refer Slide Time: 01:46)

So, what could be examples? some of them are like very obvious, take a case of a motor
where I just give an electrical input and I just see the output as mechanical energy in form of
torque. Air conditioner the input again is electrical energy; the output you can sense in the
terms of heat energy which changes the ambient temperature. The human body which is
infected with the virus, the input could be the drug prescription which you take; and output
could be the way the drug is distributed in the body, its effect on the body in killing the virus.

We will see all this, how this could actually be modeled mathematically. And another not
very surprising example is that of a car where the input could be some kind of a signal which
generates acceleration and the output which you could obviously see is the vehicle
displacement.

2
(Refer Slide Time: 02:37)

So, before we start: I’ll just introduce you to some notations in the course, the small letter t
would denote the time variable, input would usually be denoted by u(t), output would be
y (t), if there is certain delay in the time signal we will denote that as a δ, disturbance would
usually be denoted by w (t ) which is a time-varying disturbance and f denotes any function;
right which may be maps one variable to the other.

(Refer Slide Time: 03:08)

So, how do we classify systems? Based on the nature of systems we can have several
varieties or several classifications; some of them which we would have already learned in our

3
signal systems course or our circuits course and so on. So, the first one could be something
like a linear and non-linear, static or dynamic, systems which vary with time, and system
which does not vary with time as we call them as time-invariant systems. And lastly what we
will deal with is causal or non-causal systems.

(Refer Slide Time: 03:42)

So, what is a linear system? The system for which the output varies linearly with the input;
and we all know that where it just satisfies the principle of homogeneity and superposition.
For example, a resistor, so let’s see how this works.

(Refer Slide Time: 04:07)

4
So, if I just take a resistor right, so why do I call it a linear system. Let's say first, let’s see if
we can define what is a linear function. If I say f ( x) then a function is linear if
f ( x 1+ x 2)is f (x 1)+f ( x 2); this what we called as the superposition thing or a superposition
property. And something called the homogeneity if I just take f (ax ) where a is just the
constant, I just have a, af (x). Why do I call a resistor a linear system is essentially because; if
I take an R, I give it an input voltage V the output I see is I =V / R.

Now, what happens if I have two voltage sources V1 and V2, well of course with the
appropriate sign conventions and R, what is the current? So, the current I see here is of
course, V adds up like (V 1+V 2)/ R which essentially is V 1/ R+V 2/ R and this is the
response I1 which I would get only with voltage V1, and I2 the response which I would get
only with V2. So, this is like I could just take the individual voltages; take the response and
then combine them.

How does homogeneity work? So, instead of V if I say replace this by some constant a some
constant b and I look at the current would just be (aV 1+bV 2)/ R, that simply would be again
(aV 1)/ R+(bV 2)/ R and so on. So, this is like again the response with voltage V1 scaled by a
factor a added directly to the response or the current with the voltage V2 with a scaling b. So,
this is an example of a linear system.

(Refer Slide Time: 06:22)

5
We could also see well what if I have instead an inductor say voltage is L(di /dt) or in other
words the current across an inductor follows this one V / L. So, what would happen if I have
two voltage sources? So, my total voltage would be V 1+V 2 which goes across a resistor L
and that would generate some amount of current let's call this I total here di T / dt .

So, what is this di T / dt ? This di T / dt now can be written as (V 1/ L)+(V 2/ L); the same way as
we did earlier for a resistance. This again should be what the V 1/ L is essentially if I just take
the response with just V1 as the source similarly V 2/ L would be the response which I would
take with only V2 as a source and you say the add up, right just define the superposition thing.
And similarly, even the homogeneity would follow.

So, what does the non-linear system do? Well, non-linear system; well the response does not
vary linearly with the input. And of course, it does not satisfy homogeneity and
superposition. For example, a diode let us work out this as we did earlier for the same.

V
( )
So, iwas i e τ −1 . And you see if I take again V as a combination of two sources V1+V2, I
0

could never write i as some individual combinations. So now, the total iwould be

V 1 +V 2 V1 V2
(
i0 e τ
−1 ), and this is not equal to i (e
0
τ ) (
−1 +i0 e τ −1 ). So, superposition doesn't hold.
And similarly you can check with the homogeneity as well say- if I just put a here, it will not

V
[ ( )]
be equal to a i e τ −1 ; just kind of a straight forward to check.
0

6
(Refer Slide Time: 08:58)

The other kind of classification is between static and dynamic systems. So, for a static system
at any time the output depends only on the current value of the input. In other words, they are
also called memoryless systems, whereas for dynamic systems the output depends on the past
as well as the present inputs. And these are systems which are called systems with memory.
What does it mean? I can just work out another example.

(Refer Slide Time: 09:34)

So, let's again take the example of V=IR a linear resistor. So, if I apply a voltage V it results
in a certain current across my resistor; V, this is R and this is the output. And as soon as I say

7
switch to V =0, I just remove this guy. Then what happens? Well, I instantaneously go to 0;
which means the input is gone the output automatically is gone or it goes to 0.

Now, take an example where I have a V, an R and an inductance; this is my R and L if I write
down the dynamics it would simply be V =iR + L(di /dt ). And I say I just put this voltage
source, I let the system go on for a while, say for a few seconds. And later on what I do?
sorry, I just take out the voltage which means I am now looking at the circuit which looks
like this. I do this after some time, when I switch it off for a sometime; You would know that
the inductance accumulates some flux or its stores some energy. Or in other words, the initial
current when I just say I do this at t=0 is not 0, there will be some current. And I write down
the equations it will be iR + L(di / dt )=0. And let me say i at 0 is some number i(0). Then the

−R
current for all future times would be ( ( ) ). So, what you see is that, as soon as V goes to
i0 e
L
t

0, ifollow a certain pattern it does not go to 0 immediately, it will take some time to go 0
depending on what the value of R and L are and these are usually referred to as a time
constants of the system.

So, this is how we classify between what are called as static and dynamic systems. And the
way; the reason we say memory is that, well they store some energy and remember
something and that they don't instantaneously go to 0. Mathematically it would mean that
y (t)=f (u(t )), just at that time instance. Or like for example, which we saw of the linear
resistor. For the dynamics system y (t) would depend on u(t) its previous instances what was
like if we talk in terms of the stored energy, it would just mathematically look something like
this. If I want to say i at 1 second, it depends on all the previous instances starting from t=0.
For example, as we saw the inductor.

8
(Refer Slide Time: 12:35)

Other classification would be in terms of systems which vary with time and systems which do
not vary with time; as we call time-invariant and time-variant systems. And this should not be
confused by the dynamics of the system. Like for example, if in my earlier example here i
was varying with time, but this is also a time-invariant system. We are not really interested in
how i varies. So, what we mean by time-invariant is the following: that the output of the
system is independent of the time at which the input is applied.

So, if I take the example of a resistance I experiment say today for a certain V, I get a certain
value of I say- V is 1-volt, R is 1 ohm and my I obviously would be 1 ampere. So, if I do
this experiment say tomorrow the same time I should get the same reading. And those are
systems which are time-invariant, which means here we are concerned about the parameters
of the system- the value of R.

Now, what does the time-varying system do? Well, the output of the system depends on the
time at which the input is applied. If I apply u I get a certain response y, and if I do the same
experiment after a while the response might change. How does the response change? What
does it mean? It means, take an example of the mass of an aircraft: as the aircraft moves it
keeps burning on fuel and its mass inherently reduces. So, if the mass at take-off would be
different than the mass at landing, which means the parameters of the system are actually
varying. So, we are not classifying the time dependence in terms of the dynamics, but in

9
terms of system parameters. So here the resistance would be the parameter and in the case of
the aircraft, the mass would be the parameter.

(Refer Slide Time: 14:37)

The last kind of classification which we would be interested in the course is that of causal
versus non-causal, a very basic signal systems concept. A causal system is a one where the
output is depending only on the inputs which are received either in the present or in the past.
These are also called as non-anticipatory systems. Non-causal systems they depend also on
the future inputs. And the system anticipates the future inputs based on the past.

For example, well mathematically this would mean y (t) will depend only on the inputs it
received att , it is received at ( t −1 ) and all if at all any inputs at ( t −2 )and so on. Well, a
causal system could also do the reverse; it could also see what could happen at ( t +1 ) and get
its outputs based on that. Well, a basic causal system is all systems we see around you know
most physical systems are causal in nature a motor or a generator it doesn't; my AC does not
react to see what might happen tomorrow to give me an output today. Whereas, if I look at
the weather forecasting systems I would look at what could happen tomorrow to kind of
predict today, or a more general example which we are which we are used to when I prepare
for an exam I prepare keeping in mind what could be the questions in the exam these are non-
causal systems we keep in mind or we anticipate what could be the future inputs or if I am
playing a game or a cricket match. I decide, you know what kind of pitch I am playing; what

10
kind of bowlers the opponents would have. And then I prepare myself based on that. These
are essentially non-causal systems.

(Refer Slide Time: 16:20)

Now, what are the implications of this on control? Well, what is the simple control
mechanism? Well, a simple control mechanism could be a mechanism which directs the input
takes it through other systems and regulates their output. And a control system also alters the
response of a plant or a system as desired. Typically, it could look something like this. You
have a plant which is to be a controlled could be a motor, it could be a car, a missile or
whatever.

So, the controller gives it a certain control input and then we want to all this entire system the
plant plus the controller to perform the desired objective; at I will say well I want to say set
the speed of my car to a certain desired speed and I could measure that speed. So, this plant is
controlled via a control input which the controller generates based on what is the desired
reference.

11
(Refer Slide Time: 17:25)

So, what could, will this always work? Well, sometimes not you have disturbances, which are
essentially unwanted signals, which affect the output of the system. For example, people
entering and or leaving a room while an air conditioner is on, sometimes external
environments, so at morning 6'o clock my external conditions would be different then what
they would be at 12 or 6 in the evening that could be modeled as a disturbance. And my
controller should also be smart enough to eliminate the effects of these disturbances. And
therefore, we need what is called as feedback in control.

(Refer Slide Time: 17:52)

12
So, what does the feedback do? The feedback essentially it senses the output, it gives a signal
which can be compared to the reference. Say this is 24 degrees- if I talk of an air conditioner,
this is 27 degrees. This guy will send the signal back to the desired reference and say well
your output is not equal to the desired reference and this difference will tell the controller to
give a certain input which might give, which might set or eventually from 27 degrees I could
go to 24 degrees.

So, the feedback senses the output of the plant and gives a signal which can be compared to
the reference. And the controller action, the controller input is chosen or it changes based on
the feedback, for example, the desired is 24 and the actual is 21 degrees; the controller action
would be different when the output is 21, than it would be when the output is 27; so the
controller action changes based on the feedback. And feedback enables the control system in
exciting the desired performance from the plant even in the presence of a disturbance.

Say now, in (Refer Time: 19:16) this is 24 and I say well my temperature is also 24
everything is perfect. Then I say there is some disturbance some 10 people enter the room or
10 people leaves the room. Then if 10 people enter the room then you would expect the
temperature to go up to say 26, then this is being continuously monitored here continuously
being compared here. Then this guy will say oh hold on there is a difference in temperature
so the controller should perform an action to give certain input to the plant to get back the 26
to 24. So, this feedback also helps in eliminating the disturbance. And we will see the exact
analysis of how to eliminate the disturbance in the future lectures.

13
(Refer Slide Time: 19:54)

What could be examples that we have been; of course, motivating ourselves to the examples?
So, a typical open-loop air conditioner would look something like this. So, this room is the
plant to be controlled, the air conditioner is my controller, the control action could be on and
off switch of the air conditioner, and I will have a temperature setting nob or even as simple
as simple as a regulator on my fan.

So, when I look at a control system I should identify what is the process or the plant, in this
case, it is the room. What is the output? The output which I measure is the actual temperature
of the room. What is the desired input? If I want a certain temperature what could be the
desired input over here? Are there any subsystems involved in the middle and what is the
actuator? So, we will slowly quantify or even model each of these and see what these be in
the analysis and design of control systems.

14
(Refer Slide Time: 20:47)

So, this was in the open-loop case, now we could also or all the air conditions now we see are
close-loop in nature. We know that the air conditioner is it helps us in maintaining the desired
temperature. The plant, in this case, is my room, the control system or the controller is the air
conditioner, the reference signal is the desired temperature which I want the room go to, the
control input could come as in terms of an on-off switch, the output which I measure is the
output temperature, disturbance could be factors affecting the ambient temperature like
people entering or leaving the room. And then the feedback signal is the actual temperature is
compared with the desired temperature to perform the on-off control action.

(Refer Slide Time: 21:34)

15
Other examples could be well when I am actually driving a car I want to go from point a to
point b. So, this is the desired so I can fix a destination to say a point b.

Now, at each point, I see how far I am from point b, and then I have a control system which is
essentially my control action when I drive the car. The reference is the desired destination;
the control input is the steering mechanism what drives or steers my cars from point a to
point b. And the output which I measure or I compare myself to is the actual position. The
disturbance could be several in terms of traffic conditions, road conditions; you hit upslope or
a slope and so on. And this constitutes a close-loop control mechanism.

(Refer Slide Time: 22:20)

So, what we have done so far? We have looked at the classification of systems which would
be important from the control point of view. We have looked at a very basic notion of
feedback and its significance. So, what we will be doing in the coming lecture? We will look
at how to build models and the significance of models, types of mathematical models, and
what are the other what are the steps involved in modelling of systems. So, we will do that in
the second lecture.

Thank you.

16
 Control Engineering
Dr. Ramkrishna Pasumarthy
Department of Electrical Engineering
Indian Institute of Technology, Madras

Module – 01
Lecture – 02
Modelling of systems

In this module, we will talk about the Modelling of Systems, and various aspects of
modeling, what are the things we need to be careful of while modeling a system.

(Refer Slide Time: 00:32)

First, what is a model? By definition, we could say that it is an elemental or even

mathematical representation of a system. And this model which we derive will help us in
analyzing the system. And more in control terms it would help us more analyze the
input-output behaviors of the system. This model should be such that it captures the
dynamics of the system; what are these dynamics? This refers to the evolution of system
variables. For example, how the temperature behaves or how fast or how slow it goes to
the desired temperature as soon as the AC is switched on? How the speed of my car
changes before it reached or reaches the desired speed?

Similarly, if I look at an electrical circuit how a current or a voltage evolves. So, if I look
at a model in terms of very basic, why do I need a model of a system? And modeling is
something that we do in our everyday lives. For example, if I go to an interview and I am

17
being interviewed for the post of a control engineer how would I judge, if that person is
fit enough for my job or not, I would essentially ask him a set of questions. First I would
look at his background, his educational qualifications, experience in some sense, and
then I would question him with some of the basics of control possibly.

and over there it does not make sense for me to ask who is your favorite football player
or which IPL team do you support, because that answer would not help me judge if he is
a control engineer or not or a good control engineer or not. And the model essentially is I
will ask him a set of questions and based on his responses I should be able to judge if he
is fit enough for the job or not. That is an essential part of modeling.

(Refer Slide Time: 02:40)

So, in our course what we will look at is a couple of types of mathematical models. A
basic model would be a differential equation model; this the where the dynamics would
be represented in terms of some first order, second order differential equation sometimes
even higher-order differential equations. And these are essentially or usually also called
as a time-domain representation of the systems.

We also know that we could transform these differential equations via Laplace
transforms into linear equations. And can derive also the transfer function models. And
this also helps; this Laplace transform models also help us analyze the frequency domain
response of the system or the frequency response of the system.

18
Lastly, which we would like to investigate or something which helps us understands the
system behavior better in the control point of view are the state-space models. What is
the state? The state is a set of variables that describes how a particular quantity in the
system varies and these dynamics are represented as some set of a finite set of linear
differential equations. How does this look like?

(Refer Slide Time: 03:57)

Let us take a model or take an example of a very simple linear RLC circuit. So, if I just
asked you to write down the Kirchhoff's law what you will tell me is that [dv/dt]= [di/dt]
and so on. I could do the Laplace transform of this and I can write it as an input-output
representation as- a voltage being the input gives me a certain current as an output and
which is given the ratio of the output to the input is given by something called as a
transfer function.

This is just rewriting my previous equations. The other thing I could also do is to write
down; So, this is a second-order differential equation, and I could write it as a set of two
first-order differential equations which looks like this as a bit of matrix structure and this
is what we will call as a state-space model. And we will eventually learn each of these in
detail and see which model is fit for what, which model gives us what kind of
information and so on.

19
(Refer Slide Time: 05:02)

The measures of modeling could be classified as either analytical modeling or

experimental modeling. In analytical modeling, we look at systems as application of
some kind of basic laws; basic physical laws that are used to model the systems. And
their interconnections and then we combine all these physical systems to evolve at some
or to get to some mathematical model of the system.

Experimental modeling could be the selection of a mathematical relation which best

describes the observed input-output data of a system. This also called modeling by
synthesis. This is something which we would have done already in one of our electrical
engineering labs.

(Refer Slide Time: 05:54)

20
So, let say I want to I am given a resistor and I am given a variable voltage source; say
which could vary from say 0 to 100 volts. And I would ask you a question; can you
derive a model of this? So, what would I do? I would connect this, I would put a current
measuring a meter here, call the ammeter, I know what is the input voltage. So, I start
from say applying 1 volt, I analyze the behavior here it gives me a point here, I do the
repeat experiments for 2 volts it could give be me a point here, I repeat it for a 3 volt
could give me a point here, 4 volts and I keep on doing this when I generate a series of
observations.

And I just see if this is a linear resistor that this is R which defines the relationship
between V and I and based on this model or this experiments I could conclude that
V=IR. That's what this is graph tells me and this is something which we derive in our
basic circuit experiments, to when you observe the behavior of how the current changes
through a resistor when we apply a certain voltage. So, this is also one of the ways in
which we could model a resistor.

(Refer Slide Time: 07:28)

21
If I would look at the analytical model what could be the steps. First is, I should define
for myself what is the purpose of the model. Like as I said the about the example when I
am interviewing a candidate. Then I should define what are the boundaries: if I
interviewing him for a control engineering post I should not go beyond what is the job
specifications. Then I postulate a structure; the structure is what we will talk about
shortly.

Select the variables of interest; look at the mathematical description of each of those
elements applied relevant physical laws if any. Get some kind of a final form of the
mathematical model. Why did that to see if the behavior is correct or not, and modify the
model if necessary.

(Refer Slide Time: 08:17)

22
Let us go through these steps one by one. Purpose of the model: it should be first based
on the intended objective. For example, if I were to study the model of a transformer for
its losses and efficiency I would use a equivalent circuit of the transformer, whereas if I
were to study some kind of other properties of the system the magnetic behavior I would
look at something else; that the core behavior and the core saturation.

Next is to define the boundaries of the system. The system of interest is separated from
the rest of the world. For example, if I am looking at say improving my own cricket
team; I just set my boundaries as the set of players that I have, I do not really bother what
is happening with the other team. In some sense, I could do, but mostly not. This
boundary could either be real or imaginary. And this boundary says defined again with
some purpose on hand.

(Refer Slide Time: 09:16)

23
Then the third step is to postulate a certain structure. Most physical systems if you look
at what we have learned, what we learn in circuits, what we learn in some applied
mechanics or any of those courses is systems what do they do: the system components
either store energy, they dissipate energy or transform energy from one form to the other.
For example, storing elements could be an inductor, a mass element that stores kinetic
energy, dissipative elements are obvious resistors, and elements which transform energy
from one form to another would be either generators or motors and so on.

And for all these things we identify simple elements that characterize these operations of
the energy. And these basic elements would have some puts which we could connect
these systems to the outside world. And the overall system is now captured as an
interconnection of all these small subsystems. And this is referred visually to as the
physical modeling.

The next step would be to select the variables of interest. For example, how does my
current build-up through the inductor? And we assign variables to all these things which
are of interest. For example, the current, the voltage, a certain velocity profile and so on.

(Refer Slide Time: 10:44)

24
Now each of these elements would come with its mathematical representation: could
either be a linear relation, a static relation, or a dynamic relation. Once I have models of
all these individual elements I interconnect them and apply relevant physical laws. For
example, Newton’s laws or even the Kirchhoff’s voltage laws, current laws,
Conservation of energy laws, and things coming from Thermodynamics and so on.

(Refer Slide Time: 11:09)

And, once I do all these things I just get my final mathematical form. If there are
constraints that I could eliminate, I could eliminate and simplify the model as much as
possible I analyze and validate the model. If there are some inaccuracies in the model
behavior, then I repeat certain steps that could be necessary.

25
(Refer Slide Time: 11:30)

So, in this lecture what we have looked at is: why do we need a model, how to obtain a
certain model, what are the type of mathematical models, and several steps or the various
steps involved in the modeling of systems.

In the next lecture, we could classify our modeling process into a certain class of
physical systems and we will restrict ourselves to mechanical and electrical systems. We
will see analog between these systems, is there any relation between some kind of
mechanical models which we get and some kind of electrical modeling which we will
get. We will see what are the basic elements in the mechanical domain, in the electrical
domain to arrive at models and how what could we do with those models how do we
analyze them. So, those things will be a part of the next lecture.

Thank you.

26
 Control Engineering
Dr. Ramkrishna Pasumarthy
Department of Electrical Engineering
Indian Institute of Technology, Madras

Module – 01
Lecture – 03
Elements of Modelling

In today’s class, we will talk about some basic elements of modeling. And this is not
something which is very new, you might have learned some of these concepts in your basic
course on networks and circuits something coming from mechanics or even as early as your
high school physics. And we will see how all these basic elements which we will define can
be used to model physical systems and we also find out if there is any analogy between
different kinds of sub subsystems.

(Refer Slide Time: 00:54)

So, we start by classifying things into several types: it could either be an electrical system, a
mechanical, electronic, hydraulic, thermal, and several other systems. And each of these
models as we will see or each of these systems can be modeled in terms of certain very basic
elements. And then we will see if there is any analogy between you know, cross-domain
between electrical and mechanical systems.

27
(Refer Slide Time: 01:26)

And through this lecture, we will concentrate mainly on electrical and mechanical systems.
So, in electrical systems based on the type of source, we could classify systems either as
being voltage controlled or current controlled systems.

(Refer Slide Time: 01:44)

28
So, if I talk of a voltage-controlled system or a circuit my basic system variables are the
voltage and the charge. We will see why this is true why it is a voltage and the charge why
not voltage and directly the current or why not something else. And of course, the three basic
elements which we all know from circuits are a resistor, inductor, and the capacitor; and
throughout this class, we will restrict ourselves to linear elements.

Similarly, for a current sourced circuit or a system, my basic building blocks or the basic
system variable are the system current and the corresponding flux, and I'll also show you why
this is true. of course, the basic elements then remain the same the resistor, inductor, and
capacitor.

(Refer Slide Time: 02:29)

So, how are these things defined well we all know what a resistor is; it is an element which
resists the flow of current in a system, and we all know this ohm’s law V=IR, models simple
linear resistor. If I go to an inductor I know that the voltage across an inductor is given by
dϕ/dt, or in the case when the flux is linearly related to the current we will have V=L(dI/dt)
say if I just write down some basic laws which we learn in electromagnetic. So, ϕ is
proportional to the amount of current and in the linear case ɸ=LI, where L is an inductance
and then you have this linear relation between or this relation V=L(dI/dt) describing a linear
inductor.

29
Similarly, for the capacitance we know that the charge is proportional to the amount of
voltage and in case of linear elements q=CV and I have a relationship that the current which
is nothing, but the rate of charge is simply C(dV/dt).

(Refer Slide Time: 03:42)

So, what we will concentrate now is on these two things highlighted in the blue, how do we
apply relevant physical laws and get some final form of the mathematical model?

(Refer Slide Time: 03:54)

So, again coming from basic circuits we all know what are the nodal and the loop analysis,
and these are essentially coming from part six right. So, whenever we do nodal and the loop

30
analysis we apply in the context of either the Kirchhoff’s voltage laws or the current laws and
so on.

(Refer Slide Time: 04:23)

So, once we identify basic elements we interconnect them in a certain way which defines a
circuit, and then we write down on the relevant equations. So, the nodal analysis as we know
is based on the current laws and the loop or the mesh analysis we write down the voltage laws
and this is again not very alien to us, and the Kirchhoff’s Current Laws says at any node the
directed sum of currents flowing out of that node is equal to 0.

Similarly, the sum of voltages across the closed-loop is always 0 that is what the Kirchhoff’s
voltage law tells us right. So, if I will just very simply look at this little circuit kind of thing
here, I would know that i1+i2+i3 = 0 and if I just look at this as loop, I can say that V1+V2 = 0.
So, just look at the signs here I go from plus to a minus plus v2 is equal to 0 or in other
words, v1 is just minus v2.

(Refer Slide Time: 05:17)

31
Let us start with the simple example: We can start with very very basic examples which we
have already known and then try to build upon those things. So, I have a current source,
resistor, inductor and a capacitor which are all linear elements; and if I apply the Kirchhoff’s
Current Law at node 1 when the circuit is the single node that the input current which is I is
equal to the current which goes through the resistor through the inductor and through the
capacitor.

So, what is I through the resistor given V, that is simply I can I write it as V/R, across the
inductor it will just be (1/L) ∫Vdt. It's just this all these equations again come from the basic
laws which we have written earlier right V =L(di/dt), I can rewrite this as that the current is
equal to (1/L) ∫Vdt. So, we had earlier that V=L(di/dt) or (di/dt) =V/L and I simply becomes
(1/L) ∫Vdt. So, just rewriting things which we already know and finally, the current across or
current through the capacitor is C(dV/dt).

What else do I know? I already know that V is simply dϕ/dt which means a current now can
be written as C (d2ϕ/dt2) this term and across the resistor I=V/R, what is V? V=dϕ/dt. So, I
have (1/R) dϕ/dt and across the inductor that will simply be ϕ/L; and therefore, in the earlier
slide when I said; let us just revisit that slide for a while that the basic system elements are
the current and the flux- this is what I mean: So, I have current here and the flux and of
course, these are the basic elements which build my circuit the capacitor, an inductor here
and the resistance here right.

32
(Refer Slide Time: 07:35)

Now, if I do a loop or a mesh analysis you also called something like a series RLC circuit
with a voltage source again all elements are linear, I have a voltage source V then if I just
apply my voltage laws that with the total voltage V would be the voltage across the resistor
plus the voltage across the inductor plus the voltage across the capacitor where the current is
flowing in this direction.

Again I know these things; again then V the voltage across a resistor given I is simply IR,
and V across an inductor I already know from a previous slide that V=L (dI/dt) and voltage
across the capacitor can be written in terms of this (1/C) ∫I dt in the similar way as we have
written the current expression for an inductor. So, where does this current come from, well
current is simply the rate of charge or I= dq/dt, which means that V can be written as
L(d2I/dt2), just put this guy over here I(dq/dt). So, we will have L(d2q/dt2) and IR. So, again
I=dq/dt. So, I have R(dq/dt) + (q/C). Again here the basic things are the voltage and the
charge we will quickly go back and check if this is true or not. So, when I said that for the
voltage sourced system or a circuit the basic system variables that define my dynamics of the
equations are the voltage and the charge right. And we were somewhere in these steps, we
applied the relevant physical laws and got some final mathematical model.

33
We will see, again verify that. We started with the basic physical model or writing down the
basic conservation laws to arrive at this dynamics of the system. Similarly, over here a
voltage control circuit or a voltage source circuit; I start with my basic laws and arrive at this
differential equation which is the mathematical model of the system.

(Refer Slide Time: 09:37)

Now, I go to mechanical systems this is something which we learn much earlier than we do
an electrical circuit right. When we do things in high school physics. So, I classify motions
either translational motion or a linear motion as I called or rotational system or something
which has an angular motion about a fixed axis. So, what are the basic things here? If I talk of
translational motion my basic system variables would be a force which causes a certain
displacement; in the rotational motion or the angular motion, I have a torque that produces
some kind of angular displacement.

And what are the building elements I will have a mass which is essentially like a kinetic
energy element, I have a spring which is essentially like a potential energy element, and
damper which represents the losses or fiction in the system or we can also have an external
damper to the system. Rotational motion I would have the moment of inertia, I will have the
torsional spring and again the corresponding damping or the resistance element in rotational
motion.

34
(Refer Slide Time: 10:47)

So, what is mass? Mass is a property of an element which stores kinetic energy and what does
this mass do? So, when a force is acting on a body of mass m it causes a certain displacement
x, and what is Newton’s second law tell me, that the force is just the rate of change of
momentum right; this is the statement of the second law F= dP/dt and if I use the relation
between momentum and velocity. So, momentum is related to velocity as P=mV, where
velocity if I could write in terms of displacement I will have (dx/dt) and then I will have this
final relation that F=M ẍ or you usually called as or also referred to as force is ma or mass
times the acceleration.

Similarly, in case of the rotational motion, I have the moment of inertia the property of an
element that stores kinetic energy in the rotational motion, in such a way that when a torque
is acting on a body of inertia J causing a displacement θ then torque is J ӫ with the same
analogy over here. This is a direct one to one relationship between F=Mẍand T =J ӫ

35
(Refer Slide Time: 12:15)

So, the second element is the damper. Well, what is this guy? The damper is an element that
essentially generates a force acting opposite to the direction of motion, right it essentially
blocks your motion or resists the motion. What are the examples- natural friction which
happens when we walk or when we drive a car or dash spot? So, these are the two examples
and how are they modeled? So, if I have a force again causing a certain displacement x, the
damper is modelled as F is B which is a constant which is like the damping coefficient dx/dt,
a relationship linear relationship between force and velocity.

Similarly, in the rotation motion, I have the rotational damper or the friction as a linear
relationship between the torque and angular velocity which is ፅ written as T =D ፅ.

36
(Refer Slide Time: 13:15)

The linear spring, as said earlier linear spring is an element that stores potential energy in a
way that when the spring when you have a force causing a certain displacement, then the
spring as we know we will have a restoring force which will act on the opposite direction. So,
if this is my force here F and so, this is nothing, but the spring and then the force of the
spring would be somewhere here I will call it Fs right in such a way that F. So, if I write
down the conservation of the force here that is the sum of all forces is 0, I will have F+Fs=0
and what we learn in school is we model this Fs as -Kx right and then I put this here to I
substitute in F+Fs=0 resulting then in this equation.

A similar thing happens in the rotational domain also right that the torsional spring is a
property of an element it shows potential energy in the rotational motion, in similarly when I
have a torque that causes an angular displacement θ, then the torque and θ are related linearly
as T=Kθ.

(Refer Slide Time: 14:36)

37
Now, how do we do equivalent nodal analysis for mechanical systems? In the electrical
systems, we had direct laws right we have a circuit interconnected then you write the
revenant identify the loops or identify the nodes, write down the voltage laws write down the
current laws and you have set of beautiful equations, which describe for you the entire
dynamics of a system. Let see the analogous of those kinds of equations or those kinds of
conservation laws in the mechanical domain. So, we should first have our system structure
that would suit the nodal analysis and let us see with the help of an example.

(Refer Slide Time: 15:17)

38
So, I have a system composed of a spring, I have a mass element, I have another spring and I
have a damper or dashboard and this is my reference node and with some external or some
force over here.

Step 1: This would be to identify the number of nodes and these are identified usually by the
number of displacements. So, I will have one corresponding to this guy here and the other
corresponding to this guy here. So, these are my two nodes right node 1 and node 2, which
will give me these two displacements.

(Refer Slide Time: 15:49)

Step 2: So, once I identify these two displacements or these two nodes, I just also fix identify
these two nodes x1 and x2, I mark them here and then I have a reference node.

Step 3: we just go back to how these guys were connected between node 1 and node 2, and
then with the reference. I had the mechanical mass between node 2 and the reference, K2
between node 2 and the reference, B between the node 2 and the reference, between node 1
and node 2. I have the spring K and the force was kind of acting on the node 1. So, we just
draw some kind of a circuit diagram kind of a thing over there. So, I have a force x1 and x2
are my two nodes.

39
(Refer Slide Time: 16:27)

So, all these are the same this is also x2 this is also x2 this is my reference node. So, is this
one this guy also a reference node. So, I have the force, I have a spring K1 the mass element,
the second spring, and the damping element.

Step 4: So, once I do these connections as stated or as in the previous picture I have
something like a circuit diagram right.

Step 5: And the force F acts on this node 1.

(Refer Slide Time: 17:03)

40
Step 6: Then I apply Newton’s laws at node 1 that I have a force and I have a spring. So, F
would be K times a displacement and since I am dealing with two displacements x1 and x2;
I'll simply have F is K1(x1-x2).

Similarly, at the second node I say that sum of all the forces is 0, there is external force
coming from here, therefore I will have the force corresponding to the inertia element as Mẍ,
the force corresponding to K2 with displacement x2 would be K2 x2, here would be B x˙2 and
K1 the displacement or the force corresponding to the spring K1 would be, K1(x2-x1) and I
have these two balance laws for the for node 1 and node 2 which will give me the overall
mathematical model of the system.

(Refer Slide Time: 17:54)

So, in summary, what are the how do we identify the steps of a nodal analysis how do we
come about getting our models.

So, first is to identify the number of nodes which is equal to the number of displacements,
select a reference node and then connect all the basic elements, which are the mass elements
or the spring elements and so on according to their positions, similarly do for the spring and
the damper. If there is an external source that generates or which produces a force or a torque
also add that to your system. And then finally, apply Newton’s laws of motion to arrive at
your desired equations.

41
Now if you go back and you see that there are these kinds of equations look similar to what
we had even for electrical systems, and we then see where is their analogy because both
systems are written a set of conservation laws right.

(Refer Slide Time: 18:51)

If I look at the electrical systems I have the voltage conservation or the summation of currents
equal to 0 in the Kirchhoff’s Current Laws, here I also I have some kind of conservation laws
that summation of all forces is equal to 0. And based on these conservation laws we will just
try to investigate is there a good analogy between these two systems- Mechanical systems
and electrical systems.

So, let us investigate that one. So, that we could typically arrive at two kinds of analogies one
would be where you can say that the force or the input torque is equivalent to the voltage,
where we have the if we called the Force-Voltage analogy or F-V analogy, or we could also
have situations where the force is analogous to current or what we call as the F-I analogy.
How to establish this?

42
(Refer Slide Time: 19:44)

Let us start with the circuit right. So, again based on the previous steps which were listed I
can easily write down that the force that the system which is exerted on this mass of the
spring-damper is Mẍ +Bẋ + Kx which has a mass, a dissipative element and a spring with a
constant K similarly for rotational motion I can write down the equivalent of it.

(Refer Slide Time: 20:07)

43
So, let's see I have the equations for this guy and see how do the equations on the right-hand
side look like. Based on KVL we derived earlier also that L q̈ + R q̇ +(q/C) is the voltage
that is supplied. Now, this looks very similar right. So, I have a force as an input here, voltage
as the input, the force produced a certain displacement, here voltage resulted in some charge
and q̇ was the current and so on. So, therefore, I could say that force here is equivalent to the
voltage, the mass would, in this case, correspond to an inductive element, B the dashboard or
the damper would correspond to a resistance. K the spring would correspond to a capacitor
and the basic system variable the displacement now corresponds to the charge.

(Refer Slide Time: 21:09)

Let us go to the Force-Current analogy. So, instead of having a voltage sourced circuit, now I
have a current source circuit which has elements in parallel which I called as a parallel RLC
circuit. This thing remains the same F is M ẍ+ B ẋ +Kx and if I write the equations for the
circuit on the right-hand side, what I see that based on KVL I get, I =C ϕ̈+( ϕ̇/ R)+(ϕ/ L), the
sum of all the three currents through the resistance the inductance and the capacitor.

44
There is some analogy now force could be analogously seen as the current, the mass could
correspond to the capacitor, B would be the conductance or the inverse of the resistance, the
spring would correspond to the inductance like with the relation K goes to 1/L and finally the
basic system variable- the displacement here now corresponds to the flux here, and that is a
reason we call it like the force to the current analogy. So, in my previous slide, the mass
element corresponded to an inductor now it just corresponds to a capacitor and so on.

(Refer Slide Time: 22:29)

To summarize what we have done so far is to find or to establish an analogy between the
mechanical systems and the electrical systems, the force to voltage analogy or the force to
current analogy. Similar analogy would be even for the rotational domain where I start from
torque to the voltage or torque to the current and this table summarizes the analogy between
various elements between the mass and the inductor or the mass and the capacitor or the basic
system variable which has the displacement x here or a displacement θ to the charge or the
flux.

(Refer Slide Time: 23:08)

45
So, other kinds of basic elements which we learn are systems which just transform energy
from one system to another. In the electrical domain, we all know the transformer which
transmits electrical energy from one circuit to another through some electromagnetic
induction properties. Analogously in the mechanical domain, we have gears that are again, it
just transforms mechanical energy from one mechanical system to another via gears.

Here, well I can say; I can measure that the transformer as either a step-up or a step-down
transformer which would change the voltage level, here it could change the speed and the
direction of motion and they just look like this right. So, if I have a transformer with turns

N1 V 1 N1 I
ratio . is and the current would be the reverse of it; it will be 2 right it will be a step
N2 V 2 N2 I1
up or step down based on if N1 is greater than N2 or N2 is greater than N1.

Similarly, in the mechanical domain, I have the torques on both sides related by this relation
torque T1 is equal to T2 via the radius or inversely related to the angular velocity. So, these
are again equivalent elements in the electrical domain to the transformer and to equivalently a
gear in the mechanical domain.

46
(Refer Slide Time: 24:35)

So, what have we learned so far? We have classified physical systems and as specifically
concentrated on mechanical and electrical systems and their basic elements, we saw how by
using basic conservation laws where the nodal or the loop analysis helps us get the dynamics
of the system, we also established a beautiful analogy between electrical and mechanical
systems and towards the end also learned about energy transformation devices like this the
transformers and the gears.

So, what we will do next is to also learn some more physical examples right. So, starting
from cruise control of a car, we also look at some more details of the transformer of why
what we saw in the previous slide is not really fit for analysis, the simple pendulum and
something which is which might be a little new or call the predator-prey models which I will
explain to you while we reach there.

Thank you.

47
 Control Engineering
Dr. Ramkrishna Pasumarthy
Department of Electrical Engineering
Indian Institute of Technology, Madras

Module - 01
Lecture - 04
Example of Modelling

In this lecture, we will learn a few more examples about modelling. Again using all the
steps of modelling that were listed earlier; to see if you could still identify basic
elements, again we go with conservation laws or any of those physical laws that govern
the system. So, the first example that we will deal with is a very simple model of a cruise
control of a car.

(Refer Slide Time: 00:42)

So, cruise control refers to the control of a velocity of a car on a road, where the input
comes in the form of some kind of an accelerating force, the measured output is usually
the car velocity. And then I usually or we usually measure the position of the car how
much distance it has travelled; typically, from reference point.

So, how do we capture these dynamics? There is a certain force which generates a
displacement and in turn a velocity.

48
(Refer Slide Time: 01:14)

So, if we go back to the steps of modelling the first step would say well, what is the
purpose of the model? well the purpose of the model is to understand the velocity
behaviour of a car, and therefore to control the velocity. How do we define the
boundaries? Well, we look at the car as a system to be controlled and then it travels on a
highway or a road that forms a part of the environment.

So, what could be a typical structure here? So, there is again a force, displacement and
velocity. So, we could say that in this system the thermal energy which comes from the
engine structure is converted into kinetic energy or the velocity. Some part of the energy
is lost due to friction between the road and car. We will not take or take into
consideration for simplicity the thermal losses in this car.

So, the car would have a certain mass and then since we already said that there is some
energy loss, there should be some kind of a damping element introduced in the model,
and before we write down the relevant energy conservation or the force conservation
loss. And also to be a little simple, to begin with, we will neglect the rotational inertia of
the wheels.

49
(Refer Slide Time: 02:33)

So, once I do this I would like to select some basic variables of interest. So, the variables
of interest are the displacement x which again gets the velocity, all via this force F. So,
have this car modelled as a mass M and the friction between the car and the surface on
which it is moving the road is modelled as a damper with the B.

(Refer Slide Time: 03:01)

So, there are two elements here: one is the inertia element which is modelled as F=M , as

we did in the previous lecture. The frictional force would be related to the velocity via a
damping coefficient B. Are there any physical laws? Shall once we identify this structure

50
and the structure here I can directly write that the summation of all forces is 0, which
means something like this is my external force, the frictional force and the accelerating
force.

So, what is the frictional force? Well, this is B and the accelerating force is M and

substituting that is the velocity. I can write my equation as F-B = M or the final

mathematical model would look like + = ; a very simple model which captures the

cruise control in a car. This is a very simplified model, but still gives some basic
understanding of the dynamics of the system.

(Refer Slide Time: 04:20)

The next example, again we come back to the electrical domain is that of the
transformer. So, what is again the transformer? Just to recollect it's an electrical device
which transfers electrical energy from one circuit to the other and it works based on the
laws of electromagnetic induction introduced by Faraday. How does it look like? Well, it
has core something called a primary winding and a secondary winding. And an energy
transfer happens by inducing a voltage in the secondary winding via the core; and where
does the core get that energy? The core gets its energy through some electrical source in
the primary winding.

So, if I just quickly write a very small diagram of this could say that this is my core, so I
have this all this as my primary winding, this is my secondary winding, I have a voltage

51
source and then I will call this Vi which will generate some output voltage Vo here. And
then this is the magnetic core there will be some flux, and with all those electromagnetic
induction laws, there would be a voltage induced in the secondary and then you will have
certain N1 & N2. So, is that is again = . And then what is a
relation between the power here and the power here? Let's just write it very simply the
input power should be equal to the output power or V1I1 is V2I2.

(Refer Slide Time: 06:01)

Now, this is something which we do in our course. All textbooks have some problems
like this. Consider a problem where I was given a rating of a transformer goes from 200
to 400, a step-up transformer, frequency is given to be 50-hertz, single phase
transformer, some power factor, and if you calculate the efficiency of the transformer
operating with the load with some load at 100 percent load, at 80 percent load, 50
percent load, and 25 percent load.

So, how do I solve this problem? Do I take this transformer and say well, let me get a
voltage supply; let me get a load which is 100 percent and then do my experiment each
time for 100 percent, 80 percent, 50 percent, and 25 percent and get the answers? Or
there is any other easier way? The risk when I do this experiment is that, well I may not
be very careful or I could burn down the transformer and so on.

52
(Refer Slide Time: 07:07)

And therefore, what we equivalently see this problem is a problem translated to

something like this which is the equivalent circuit model of a transformer. Now, once I
read this question; solving it through this picture seems very tricky right; that I really
have to get this transformer down open all the wires, we know to get a big Crane because
this seems a very big transformer. But if I look at this transformer as something like this
then things get much easier.

Now, what is this process called? This process is what we call as the model of the
transformer for solving problems of this kind. And why do problems of this kind arise?

When we say that is or ; if this was true all the time then we

would never possibly need an equivalent circuit, because this thing usually does not; it is
not always true that V1 I1 =V2 I2 there is something going wrong here and that something
going wrong is what we will try to observe.

53
(Refer Slide Time: 08:17)

So, we will propose a model for analysing losses and in turn, we could compute what is
efficiency. So, in the first step, we identified what is the purpose, the second step is to
define the boundaries which is the transformer, possibly connected to a load, and then
the structure. Well, the structure of the transformer- we have the primary winding, we
have a secondary winding, we have voltages, currents and since I have winding would
have its own resistance, its own inductance, the core would come with its own properties
which might lose some energy and so on.

And just for a movement, we will assume that we will neglect the saturation of the core.
If I take into consideration all these other things of what in the textbooks would be called
as the core losses and the copper losses.

54
(Refer Slide Time: 09:10)

So, I then give a structure to that. So, I have a primary winding and a secondary winding,
then I call this; well, each of these guys would have some resistance let me call this R1
and R2. There would be some flux leakages in both the windings of primary and
secondary I will call them as L1 and L2. There are certain losses which happen in
addition to the losses in the winding also in the core, they are I will quantify them as R0
the resistors which quantifies the heat loss in the core and similarly the inductance of the
core.

(Refer Slide Time: 09:51)

55
So, once I have this, I have this beautiful looking circuit. I have a voltage source, a
resistance, L1-the inductance, R0, I have L0, I have given here the turns ratio L2, R2 and
so on. So, this is the physical model of the transformer. Now given a transformer, who
tells me what is R1, L1, what is R0? what is L0? what is L2? and so on. The load I would
know because I am just connecting this externally to the transformer.

(Refer Slide Time: 10:23)

Once, I have this equivalent circuit I would know that I could calculate the efficiency of
the transformer. And who gives me the values of those R-several resistances and
inductances? Well, we know that we could do something called an open-circuit test or a
short-circuit test. So, when I do the open-circuit test what do I do? In the open-circuit
there is no load that this guy just be open and that's why we call this an open-circuit test.
I apply a voltage here which is equal to the rated voltage. Then most of the current
consumed would just be here. And based on the measurements of the power and the
current and so on I could estimate what are these guys? What is R, what is L? Just by the
relation between the voltage, the current, and the power. That we call in our lab
experiment as the no-load test.

Similarly, to identify the resistance and the inductance of the winding I do something
called a short-circuit test, where the low voltage side is the short-circuit. So, this guy is
simply a short-circuit. And I keep on increasing the voltage from 0 to a value where the
current goes to the rated value. That is usually typically given in your problem statement

56
or even if you look at a physical transformer I will have it as the rated current; you know
the supplier would give me some.

So, once I have the rated current going through I this I1 and I2 then this becomes a little
negligible and all the parameter which I measure or all the losses would mainly be
accounted in I2R losses in the primary winding and I2R losses in the secondary winding.
And doing just these two experiments would give me all these values of R1, L1, R2, L2,
R0, and L0.

(Refer Slide Time: 12:17)

So, this is the process: Do the open-circuit test, the short-circuit test, identify the
parameters, so once I identify all these parameters the calculation just becomes now a
basic problem in circuits. So, I can just look at the circuit and all these problems translate
to; so this is given maybe possibly with a certain power factor or this just could be a
resistive load. So, I am just solving problems using a pen and a paper.

So, what have I done? I have looked at this guy the big transformer which is sitting at a
substation, I want to do experiments on this which will give me answers on how this
transformer would behave to different changes in the load; no loads, from full load with
different power factor, resistive load, inductive load, all these experiments I have
translated to simple experiments which I could just use a pen and a paper and possibly a
calculator. And that is the beauty of the model I can answer all the questions without
having to touch the transformer.

57
(Refer Slide Time: 13:19)

So, the next example is something which is very familiar to us something which we learn
in high school. The simple pendulum; well it's by construction it has a mass M, a little
string L, hinges to a pivot point P and we all know that if I start from this position given
θ; I just let it go it will just keep on oscillating back and forth around its equilibrium
point. Now how do we explain these dynamics or explain what this pendulum does as a
mathematical model?

(Refer Slide Time: 13:54)

58
Let us go back to high school physics for a while. The purpose of the model again is I
want to study the motion of the simple pendulum, how would it behave if I release it
from θ of 5° or θ of 10° or 90° and so on, and identifying the relevant boundaries of the
system. And what is the structure of the system? In this system, I know that there is some
kind of energy transformation happening in the system.

So, if I see here it goes all the way here till -θ and comes back. And at this point where it
stops and goes back, that is a point where we have complete change from the kinetic to
the potential energy: kinetic energy goes to 0, all the energy is the potential energy. This
is the point where all the energy is into the kinetic energy and back and forth. So, the
system just oscillates between points of maximum potential energy to maximum kinetic
energy and so on.

We will see how it looks mathematically. As the structure is similar to a rotational

mechanical system right you have rotational motion, possibly a torque or external torque
or the systems own torque because of its inertia and so on.

(Refer Slide Time: 15:07)

Neglect friction, I just draw something like the free body diagram which I am used to in
my twelfth standard physics. I have the rotational torque, there is some that could
possibly some frictional torque, a moment of inertia J and some angular velocity ω.

59
So, the variables of interest here are I would as in the car I was interested in how the
position and the velocity evolves. Here I would be interested in how the θ the angular
displacement and the angular velocity evolve with time based on these parameters of the
system- the M, the L, and g which is the acceleration due to gravity.

(Refer Slide Time: 15:49)

Again just if you recollect those basic building blocks in a mechanical system I had the
inertia element modelled as torque is related to the angle acceleration via J. And again if
I just recollect some of my high school physics I would know that the moment of inertia
is ML2. Now I also, in addition have gravity right in my system and the torque due to
gravity would be in this direction and can be easily computed to be Mg sin θ multiplied
by L.

So, this is my gravitational force Mg the force in this direction would be Mg sin θ and
the torque would be given by this form. And both these torques are acting in the opposite
direction. So, at the moment I don't consider friction over here, it's just a lossless system.
So, ML2 the torque this guy is equal to or is compensated by this guy -Tg. And

therefore, I have the dynamics written as this one: + = 0. So, this is just a

lossless pendulum. Once you leave it from this position θ it will keep on oscillating
forever.

So, as a simple exercise can try and find out an equivalent of these kind of behaviour in
an electrical circuit, where I just start from an initial condition and my systems are such

60
or the system component are such that the total energy is of course, conserved and it also
shows an oscillatory behaviour. Think about it try to do it may be in the next class to see
if there is an equivalent of simple pendulum behaviour in an electrical circuit.

(Refer Slide Time: 17:37)

Next, we move on to modelling friction. In addition, now I have the frictional torque Tf
again if you recollect the previous lecture it is modelled as the relation between the

torque and by a resistive element β. So, I will have one more equation in my or one

more term in my equation that Tr would be the negative of Tg and the negative of Tf
both acting in the opposite direction, and I just write down the relevant substitute this Tr,

Tg, and Tf over here to get a final mathematical model something in , , and sin θ.

So again, this is my kinetic energy element, this is my potential energy element, this is
my frictional element if you really want it to relate to mass-spring and the damper kind
of system. Think of another example of an electrical example of which exhibits
behaviour like this. Like a lossless behaviour combined with the dissipative element;
think of it.

61
(Refer Slide Time: 18:42)

So, something which we will do now is something outside what we have learned so far in
textbooks. So, I'll just briefly explain to you what we are trying to do here.

(Refer Slide Time: 18:56)

So, let me start with some scenario here say, where I have like a little pond here with lots
of water, little weeds here, little vegetation inside and I have some you know small fish
and this small fish will feed on this little vegetation inside the pond. So, this is a good
looking scenario where these guys, bunch of those small fish, another small fish here,
some hiding inside here and so on. So, this small fish feed on all the vegetation and then

62
they live happily, they multiply, they grow in population, and so on. So, how can I model
this population?

So, let me call this small fish as some s and I want to see how these guys evolve with
time. And if there are a bunch of them with some initial condition a nonzero initial
condition they have plenty of food to eat you will imagine that they will keep on growing
forever. So, this could be modelled as some constants say I will call this K1s; they will
keep on growing forever. This is a very simple model which captures growing on forever
phenomena.

Now, let us see there is another pond here where there are bigger fish; many of them, but
these fishes such that they do not feed on vegetation but they feed on some smaller fish.
And if there is no small fish here for them to feed on it is natural that all of them will die
out. There is no food the fish go on dying. And I will model this as -K2S. So, the first
equation tells me how the small fish will grow when they; of course, I assume that they
do not naturally die and they have lots of food to eat. So, if I write down this just be like
an exponential growth like so the s at any time t will be s(0)eK1t; this is a solution to this
equation and then if I just draw a graph between s and t it might just grow like very fast.

So, these guys what will happen? They do not have food to eat so they will die down
eventually and start with the initial condition they have nothing to feed the solution will
not look very good, we have the S(0)eK2t and with time they go smaller and smaller
asymptotically it goes to 0.

Now interesting thing happens when I take some of these guy’s small fish and put them
here. So, there is the small fish nice tiny little ones; I say one of them could be nemo
with one of its fin lost. So, what will happen now? So, these big fish have some small
fish to eat. So, what we would expect to happen in the second equation. So (dS/dt), these
guys had nothing to eat so they all were dying now have something to eat.

So, naturally, we would expect that their population will grow. So, I will write that this
big fish they come in contact with the small fish at some rates I will call this say from
K3. And you will see how the population increases that because they have food to eat
they can multiply and so on.

63
What happens to this guy? Will this guy still grow exponentially? What happens to
(ds/dt)? Here we assume that they never die, they have a lot of food to eat and therefore
they keep on multiplying. Now, this (ds/dt) which was earlier K1s will now see some
decrease, and the decrease is based on how many big fish come in contact with the small
fish and at what rate and we call this K1. So, this equation together with this equation;
this is small s and big S I hope you're able to distinguish between those let me just this is
my ds. So, these two equations will together in some way explain to me the behaviour of
how small fish and big fish, their population will change with time-based on maybe
certain initial conditions, availability of food and so on.

So, we are just written down this vaguely in terms of some mathematical equations. Just
made some assumptions and we assume that they grow exponentially or sometimes they
die when they have food and or when they small fish die when they come in contact with
big fish and so on. So, let us do this little systematically and see how it goes.

(Refer Slide Time: 24:53)

So, these models are typically referred to as the predator-prey models. So, what is this?
So, there are two species: one of them serves as a source of food for the other. So, the
bigger fish are the predators and the smaller fish are called the prey. And this prey they
feed on vegetation. So, how do we go about describing a model for this? So, earlier these
are based on certain assumptions and observations.

64
So, we make the following assumptions when we derive this model. And the prey birth
rate is propositional to the size of its own population. And the way the prey die is only
when they come in contact with the predator, and therefore their death rate would be
proportional to in some way to the size of the predator population. The third assumption
we make is that the predator's birth rate is proportional again to the size of both predator
and the prey population because, when they are in contact with each other they have food
to eat and then they could multiply. And the predator death rate is proportional again to
its own size.

A further assumption is that: the predators (Big fish) have nothing else to eat except the
small fish; or in other words it means that the predator species is totally depended only
on a single source the small fish for food supply. And the prey species are such that they
have unlimited food supply and nobody else eats them to expect the big fish; there is no
other thread to the prey apart from this specific predator which is the big fish.

(Refer Slide Time: 26:05)

So, let us say I want to define; these are my system variables like in a way the x(t) is the
population of the prey species at some instant of time t; y(t) the population of predator at
time t. So, based on how we derived the model previously (dx/dt) would be some
constant α greater than 0 which is the, first term will denote how the predator population
increases, then they are their own and they have a lot of food to eat. The second term

65
-βxy captures how their population decreases when they come in contact with the bigger
fish or the predator.

The next expression (dy/dt) = -γy which means the first term means that they will die in
the absence of food. The second term captures the increase in their population when they
come in contact with the prey or the smaller fish then how their population increase. And
all these parameters α, β, γ, and δ depend on how fast they die or you know how
frequently the predators and the prey they come in contact with each other; all these are
like greater than 0.

So, these equations are usually referred to in the literature as the Lotka-Volterra
equation.

(Refer Slide Time: 27:26)

So, again we just rewriting what we just said earlier: in the absence of predators what
happens to the prey, in absence of prey what happens to the predators and so on.

66
(Refer Slide Time: 27:40)

If I were to just look at the graph of what happens right. So, let me say that if I want to
just right down these things- if I say well what if the initial condition says that x=0, y=0;
well, this means nothing. What if say some x is say 15 and y is some 25, who will
increase first? Who will decrease first? Let us say, for example, if I start with say
something x is 100 and y is say 10000. So, what we would see initially is that there are
lots of big fish to eat the small fish, so you will see a drastic reduction first in the size of
x.

And then now these guys will have less food to eat then y will go down for a while. And
if y goes down these guys will increase and then they keep on increasing and decreasing
forever, unless there is a point where if there is some number x* and y = y* they start
from these numbers and remain that forever; which means that if x* is constant for all
times, y* is constant by all times which means dx/dt will go to 0 and also dy/dt will go to
0. Let us see if this is possible or not.

So, I have two equations here: αx- βxy going to 0 and -γy + δxy going to 0. So, the first
equation would mean xα- βy is 0, the second equation would mean y-γ+δx is 0. We are
not interested in this solution x = 0 and y = 0. Therefore, the solution would be y is α/β
and x let me call this is a * because I'm just looking at a solution when dx/dt goes to 0
and dy/dt goes to 0. So, this x* would be if I rearrange γ/δ and y* would be α/β. So, if

67
initial conditions are such that these two things hold then my population will just be the
same throughout; that's what this model takes place.

What happens for any other values? So, this point here the centre point is a x star y star
somewhere here. For every other thing, I just see that the population of x increases, then
because x increases y will also increase, but we will y increases x will see some decrease
and so on. Let us be some oscillatory kind of behaviour depending on where you start, if
I from here then I will just follow this curve here; if I start close to here I might just have
curves which look like here. If I start here I will just remain here. There is no other thing
because I cannot go here because x and y can never be equal to 0 and I don't even would
not start over here. So, these are fairly good conditions where I start at some point and
then I just keep on revisiting the same point over and over again.

This is just simulations for some certain values of α, β, γ, and δ. So, this is what we call
as a fix point or an equilibrium point when dx/dt goes to 0 and dy/dt goes 0, this is like
the equilibrium points and we will define this eventually more formally; equilibrium
points they are also called fix points. So, don't worry about this now, but we will define
this formally little later right.

So, this point is 1 comma (1/2); I'll may be share the code of how to draw this kind of
pictures; just for your own understanding and all this could be drawn quite easily in
Matlab. So, is there any drawback of the model?

(Refer Slide Time: 31:52)

68
When we look at the steps of the model we say no step five-step six-step seven says
come to the mathematical model. And then there is another step which says validate your
model; if I say validate my model will I know if I just go to a pond, I just own a pond for
a while: there is vegetation, I buy small fish and big fish and I say well, you know I do
this experiments will this hold true? Well, they may not because the assumptions we
made are really strong.

Now let us see if we need to make some more assumptions or relax some assumptions or
maybe there should be some stronger assumption to arrive at more realistic models. So,
the drawback of the previous model was that the prey population would grow unbounded
in the absence of a predator, which means they would never die by themselves which is
contradicting to nature. If you know there will be people, there will be small fish who
will die naturally.

And therefore, to make it realistic we replace the exponential growth term. So, earlier we
had dx/dt was αx which represented an exponential growth in population, now we will
also add some term which will take care of the death rate. Say some constant a multiplied
by x2. Therefore, if you look at this equation if there is no big fish y goes to 0, and in the
absence of y the prey population well it could stabilize at some value, it will not be that
really a large number of fish, that is not enough to eat, but will. So, the assumptions are
little unrealistic.

This is an assumption which will be closer to reality or which is sustainable in the

environment.

69
(Refer Slide Time: 33:42)

Similarly, if we say that the predator, we assumed earlier that it will, of course, die
naturally, but if what we also assume that if there are no smaller fish they will just die
because they have nothing to eat. That is also a very strong assumption, but if I say that
the predator has some other alternative source of food then the predator population then
they will not go to 0 due to starvation. So, mess is not the only source of the food for
hostel guys so they could go out and they could live you know much longer.

So, that I would account for say dy/dt this takes into consideration my increase in
population, because of the small fish, I modify this a little bit to take care of the account
of the natural death –by2 and this is the increase in population when I come into contact
with some other restaurant food, not the mess food. So, in this manner we could model
complex food chains, you can just look at some more assumptions something’s could be
based on observations, we could also model what if there are multiple species interacting
with each other. That is how the entire ecosystem of the nature is built right that one
species is depending on the other species for survival. The tigers are depending on deer’s
and deer’s are depending on trees and so on.

So, all these nice kinds of natural balance equations can be captured just by observations
and then writing down in terms of some of the simple differential equations.

70
(Refer Slide Time: 35:24)

So, what we have learnt so far is very general notion of systems; what could we do with
systems in terms of control? open-loop control, we had close-loop control, we had
classified system into two types of a physical system which we delt extensively were
electrical and mechanical systems. And we also see well of you know how to build
models or approximate all this real-world phenomenon in terms of some basic elements
from electrical and mechanical systems.

And even if I do not have these basic electric and mechanical systems I could still model
systems in terms of some kind of differential equations just by observation and looking
at some facts which I know that all living beings will die eventually; they eat food to
survive and so on.

71
(Refer Slide Time: 36:13)

So, next lecture what we will talk is: before that we will do some problems and after we
finish with those problems we will do something called as a linearization of non-linear
systems, more systems we counter are essentially non-linear, how do we linearize them
and then how does a linearization help us in the analysis. We will just look at why is
linearization required. And then we will recollect some tools which we learned in maths
or even in signals and system course; the Laplace transforms- its properties, inverse
Laplace transforms and why do we need Laplace transforms and what is its relevance
through the rest of the course.

Thank you.

72
 Control Engineering
Dr. Ramkrishna Pasumarthy
Department of Electrical Engineering
Indian Institute of Technology, Madras

Module - 01
Tutorial
Lecture - 05
Solving Problems in Modelling of Systems

Hello everybody. In today’s class, we will do a few problems related to the theory which we
have learned so far in terms of several steps of modelling. So, I start with the question which
I had asked you last time.

(Refer Slide Time: 00:36)

So, starting from a lossless pendulum for which the equations look like this right; Where θ

g
was the angular displacement and θ̈ + sin θ = 0, and this represented a simple harmonic
L
motion. Is there an electrical analog for such behavior or such dynamics?

Well, let's take this circuit with L and a C. So, if I write down the voltage loss I will have L

2
di 1 d q q
+ ∫ i dt = 0. Now, if I write it in terms of the charge I will have L 2 + = 0. So
dt C dt c
this, the solution to this will be periodic, in the same way as it will be to these equations.
Only difference here is that I am just using both of them are linear elements.

73
So, here the sinusoidal term comes because of the non-linearity of the element, of course, we
will discuss about the relation between this model and this model when we do linearization of
non-linear systems, but solutions to both of them will just be periodic right and when we had
the lossless or the pendulum with some frictional element we had an additional term.

g
So, my equations looked like θ̈ + sin θ + Bθ̇ = 0, and analogous here would just be an RLC
L
circuit. So, just add a resistance R. So, this is an L, R, and C and my equations would just be

2
d q q
L 2 + Rq̇+ = 0 and then the behavior again; it is the same this is again, I am just
dt c
ignoring the non-linear element here for a while, but the linear version of this would look
something like this and this is the analogous behavior of the simple pendulum in the
mechanical system case.

(Refer Slide Time: 03:23)

So, we will do some more examples related to this. So, let's take a system which looks like
this. So, all this is my reference points or reference node as I would call it, I have a mass
which I call M1 and which goes around this surface between the mass and the reference point
is a spring with constant K1, there is another mass M2 and between these two masses is a
damper with its value B12, a spring connecting these two masses with spring constants K12,
the friction of the surface is B1, the friction over this is B2 and there is an external force which

74
is being applied here called F, which results in a displacement here I will denote this as x2
and this displacement I will denote as x1.

And then my problem here or what I want to write is the dynamical equations governing the
system or the model of this system. So, if I go through the steps of the modeling, the first
thing to do is to identify the number of displacements, in this case, it is x1 and x2. So, the
number of displacements is 2; then what are the nodes? If I follow the step, I would choose x1
and x2 as my nodes together with a reference node.

(Refer Slide Time: 05:59)

So, if I just mark my nodes here, I call this x1 and let me call this x2 and maybe a reference
node here. Now if I look at each of my elements right the first element M1 is between x1 and
the reference node. So, I will just draw it this way. So, x1 and the reference node in between
is M1. Now look at even K1; K1 is also between x1 and the reference node. So, I could draw it
and similarly also with B1. So, B1 the damping element would show up here call this B1.

Now, between the nodes, x1 and x2 is a dashpot or a damping element and also a spring. So, I
would write this as a spring. So, between x1 and x2 a damping element and a spring with
coefficients B12 and K12; now I go to node number two there is a mass between node number
two and the reference similarly a damping element here so that would look something like
this. So, between this there is a mass M2, there is the frictional element denoted B2 and there
is a force which is being applied which looks like this, ok?

75
Now, once we have the system written in form of its nodal components, then it’s easy for me
to write down the dynamics or I just write down the conservation laws at node one: node one
where I have the mass, a spring, a damper or actually node x1. So, how will the equations
look like? All the forces corresponding to each of these elements this would be M1 x¨1, plus
this is modelled as K1 x1 the damper goes like B1 x˙1that takes care of these three elements. So,
for node two there are also these other 2 elements and these equations would be B12 ( x˙1 − x˙2)
+ K12 ( x1 −x2) is 0.

So, the sum of all the four sets here is 0; now allocate the second node or node 2 or denoted
by x2. We have an input force, force of this guy over here, force of the mass and these 2
elements right. So, this would just look like F is M2 x¨2 + B2 x˙2 correspondent to this element
and I have these two guys here B12 ( x˙2 − x˙1) + K12 ( x2 −x1). The sum of all the forces here is
equal to the external force which is being applied through F. So, these are the entire equations
which govern the dynamics of the system.

So, we also learnt how to write this in the Force-Voltage analogy and also the Force to
Current analogy. So, let’s just recollect how these analogy looks like. So, here we learnt the
analogy between electrical and mechanical systems via the Force-Voltage analogy, and also
the Force-Current analogy.

(Refer Slide Time: 11:20)

76
So, let’s draw this equivalent mechanical system in terms of its force say starting with voltage
analogy.

(Refer Slide Time: 11:35)

So, the Force to voltage analogy tells me that force is related to the voltage, mass with an

1
inductive element, friction with a resistor, spring with a capacitor with being the value. So,
C
if I just use these directly; My equations? let's write down the equations first that might help
us in drawing the picture a little better.

So, M becomes L1, the basic elements here are the displacement, the displacement will be
q1 1
mapped to the charge. So, I have L1q¨1+R1q˙1+ + R12(q˙1 −q˙2)+ ( q −q ). So, the mass is
C 12 1 2
C
1 1
replaced by this inductor B1 by R1, K1 by , you can look at the analogy table why this is
C C
1
then R12 replaced by or it represents B12 and represents K12 all this equivalent to 0.
C 12

The second equation would look like the force would be analogous to a voltage V=L2q¨2+R2q˙2

1
+ R12(q˙2 −q˙1)+ ( q −q ); I am sorry there will be a dot here as well because of the x, x˙1 − x˙2
C 12 2 1
.

77
Now, if I were to draw the equivalent circuit well I have the voltage source V which is an
analogy of force then I will have R2, L2 representing the mass M2 then these two elements B12
and K12 would look something like this I have an R12, C12 and then the mass M1 with L1 the
element B1 with R1 and the spring K1 is analogous to C1. So, if I write down treating this as a
circuit right and then write down the equation with possibly maybe this direction of the
current, I just get these two equations right and this is essentially the force to voltage analogy.

Similarly, we could also write down the force to current analogy; in the force to current
analogy my force is analogous to current, mass instead of an inductor now is analogous to a
capacitor, the frictions element is now a conductor, spring is now equivalent to an inductor,
and the displacement is now analogous to a flux element. So, let's see how the first equation
goes M1 x¨1+K1 x1+B1 x˙1+B12 ( x˙1 − x˙2) + K12 ( x1 −x2) going to 0. So, M1 from the table would
represent the capacitor C1. x is equivalent to the flux ϕ.

(Refer Slide Time: 16:00)

1
I will call it ϕ¨ 1plus the spring element is equivalent now to an inductor, I have ϕ , now with
L 1

1 1 1
a resistor I will have ϕ̇1+ (ϕ˙1− ϕ˙2)+ ( ϕ −ϕ ) all this is equal to 0.
R1 R 12 L12 1 2

Now, look at the dimensions of all this, all this have dimensions of current. So, I am just
writing down a current conservation law at a particular node. Earlier I had written down the
voltage conservation law. Similarly, in the second equation my source F is equivalent to a

78
current and again going back to the analogy table I can write down the equations in the

1 1 1
following way: C2ϕ¨2+ ϕ̇2+ (ϕ˙2− ϕ˙1)+ ( ϕ −ϕ ).
R2 R 12 L12 2 1

So, how would this look like in the circuit diagram? The mass, I will start with the mass will
be replaced by C1 then I have the spring with L1, B1 replaced with R1 then I have these 2
elements here. So, this is R12, L12 then I have the other three elements following right the
current source, I have M2 replaced with C2, and B2 replaced with R2, and this is my
equivalent circuit in the F-I analogy right.

So, we started with a system we had that in a form where we could write the equations with
node 1 and node 2 thus the summation of all the forces is 0 then we came to the Force-
Voltage analogy using the table and the Force-Current analogy right and we even could come
up with the equivalent circuits.

(Refer Slide Time: 19:35)

So, the second example is a motor driving a load through a gear train. So, how does it look
like? let me denote this block as my motor with this motor which is connected to a gear train,
and this gear train is then in turn connected to a load, ok? So, the motor produces a torque
Tm right. I'll also denote the moment of inertia here as J1, right the displacement as θ 1 again
in this direction right and a friction coefficient D1 over here.

79
Similarly, here if this is moving this way θ 2 will go in this direction right, imagine this as a
gear rotating this way and then this will force it to rotate the other way, and there will be a
certain load torque. Now this gear has a radius of r1 with the number of teeth as N1, here it is
r2 and N2. So, I can call this as the primary and secondary right analogous to what I call in the
transformer. Now I want to write down the dynamical equations for this or the model for this.

So, start with the motor shaft thing first; what are the elements here? well the motor generates
or gives me this torque Tm, now I have J1 θ̈ 1+ D1θ˙1. Now, this would be the dynamics if
there is no load or there is no gear train also right. So, the input power would be consumed by
the inertia here, the J1 and the frictional element; now what this load does is well it requires
some torque and how is this torque reflected here, it is reflected via this gear train to
something here. So, there will be additional torque which is experienced here or additional

load because of or additional torque because of the load let me denote this as Tl.

So, Tm now compensates for J, D or J1 D1 and also this Tl right I'll not call this Tl say I will

call this T1. This Tl is reflected here through the gear train as T1. So, I will have a T1 here.
Now at the load side, at the load side. So, I have a J2 here and a D2 similarly I have J1 and D1
here. So, here this T1 is experienced here as some T2 right I will tell you the relation between
T1 and T2 shortly. So, here I have the term corresponding to J2 as J2θ¨2, then I have the friction

element here θ˙2plus I have the load torque Tl. So, where does all this come from? So, I have
to compensate for the load, compensate for J2, compensate for D2 and this comes via T1
through the gear train and let me call this as T2.

Or let's see the entire process here right. So, I have Tm- the motor torque, it first compensates
for J1 and D1 and gives me a torque T1 here, this T1 is transferred via the gear train to the
secondary side called T2 and this T2 has to compensate for the load, compensate for D2 and
compensate for J2 right and each of this models which we had seen earlier right why is this J2
θ¨2 ? why is this D2θ˙2? and so on ok.

80
(Refer Slide Time: 25:31)

Now, what is the relation between T1, T2, r1, N1, and r2 ,N2? So, we will use the relation that
the linear distance travelled by both the gears is the same which means θ 1 r1 is θ 2 r2. Since the

θ2 r1 N1
linear distance travelled by both the gears is the same I have that is is . Further, I
θ1 r2 N2
assume that in the transfer of power between the primary side of the gear to the secondary

T 1 θ2 N 1
side there is no power loss, which means T1θ 1 is T2 θ 2. Now this means, that is is .
T 2 θ1 N 2

θ̈ 2
Now, the relationship here can also be written in the following way at , the angular
θ̈ 1

θ̇ 2 N1 N1
acceleration,
θ̇ 1
the angular velocity is
N2
. Now T1 can be written as
[ ]
N2
T2; now what is

T2? T2 is J2 θ¨2+ D2θ˙2+ Tl. So, the relationship between T1 and T2 is given via substituting for

T2 over here that is J2 θ¨2+ D2θ˙2+ Tl.

81
(Refer Slide Time: 28:22)

Now, substituting for this T1 in this equation gives me the following. That Tm is J1 θ¨1+ D1θ˙1

N1
+ Tl . So, I am substituting for T1 here that is
[ ]
N2
(J2 θ¨2+ D2θ˙2+ Tl). Now from this

expression, I can rewrite this expression again and then everything in terms of θ 1 right. I want
to write everything in θ 1 or θ¨1and θ˙1. So, this will look like the following J1θ¨1 let us do; let's
take this term; J2 θ¨2 is in terms of θ 1 would be J2 again from this expression what is θ¨2 ? θ¨2 is

N1 N
θ¨1. So, J2 θ¨2 would be J2 θ¨1 with 1 .
N2 N2

N1 N1
Now, here I have
N2 [ ]
, J2 θ¨2, therefore,
N2
J2 θ¨2 would simply become this guy. Now there

is this term corresponding to θ¨1 and also J2 θ¨2 can be written in terms of θ¨1. So, substituting

2
N1
for this guy over here I get the following. So, I have J1+J2
[ ] θ¨ . Similarly, I could do for
N2 1

the second term right I can write D2θ˙2 in terms of θ 1right. We are just using the relation that

82
 N1 N1 N N1
θ˙2 isθ˙1
[ ] N2
, which means
[ ] [ ]
N2
here 1 D2θ˙2 is D2 what is θ˙2 ? θ˙2is
N2
θ˙ . So, I have
N2 1

2
N
[ ]
θ˙1 1 .
N2

2
N1 N1
So, the second term would be D1 +
N2 [ ]
(D2θ˙1 ¿, plus now the remaining term is
N2
Tl . All

this is equal to Tm.

(Refer Slide Time: 32:37)

Now, I can call this J equivalent, I can call this entire guy as D equivalent, and write down

my overall equations, is that J or let’s call J1eq or D1eq.

83
(Refer Slide Time: 33:00)

N1
Let, J1eq θ¨1+ D1eqθ˙1+ [ ]T is Tm. So, this is the dynamics referred to the motor shaft and if
N2 l
you remember something from transformers this would also resemble the equivalent circuit
dynamics referred to the primary side.

Now, I could also do it referred to the load shaft, where I just write down the equations. So,
you will have J2eq θ¨2, where I just want to write down all equations now in terms of this one,

this equation in terms of notθ¨1, but now in terms of θ¨2, plus D2eq θ¨2 + Tl would remain as it is

N2
and Tm would get transformed via the gear ratio to
[ ]
N1
Tm. So, exactly the same

procedure which we did here in terms of writing in θ¨1 we just write in θ¨2 using this relations,
and the first 2 equations which govern the dynamics.

So, these are 2 ways of writing down the entire equations of motion of this one referred to
what we in the transformer terminology call as a primary side or the motor shaft in this case
or referred to the second risks side or as the load shaft in this case sorry.

Student: Sorry this will be θ˙2 thanks for correcting that. So, this will be J2eq θ¨2D2 will come

N2
always to the θ˙2 + Tl is [ ]
N1
Tm.

84
Thank you for your attention.

85
 Control Engineering
Dr. Ramkrishna Pasumarthy
Department of Electrical Engineering
Indian Institute of Technology, Madras

Module – 02
Lecture – 01
Laplace Transforms

Hello everybody, in this series of the lecture or in this module we will essentially revisit what
we had learned in our basics signals and systems course. So, in order to make the course a
little self-content, We'll quickly run through those concepts especially of Laplace transforms
and the several properties associated with it. We will take a side detour, because at the
moment we not do linearization of non-linear systems. So, we will do that slightly later when
we will deal with straight space.

(Refer Slide Time: 00:50)

So, starting with the basic definition of signals or basic classification of signals, what we
would learn in basic systems course would be to essentially differentiate between time
domain and frequency domain.

So, what does the time domain do? the time domain shows how a signal changes over time,
similarly the frequency domain graph shows how much of the signal you have within each
frequency. For example, even when you do things like filters like low-pass, high-pass, band-
pass, and all-pass filters and so on. So, in the time domain, the magnitude of the graph at each

86
instant is represented when you draw a graph, I will show you examples of this very short and
very briefly when we do this. when we learn of what this actually means? Similarly, in the
frequency domain, this signal is represented by sum of the sinusoids of different frequencies
with certain magnitude, and also in some cases a certain amount of phase shift.

(Refer Slide Time: 01:49)

So, what do this mean? So, let's take a very simple looking signal, signal x (t ) which is a

( π3 )+2 sin 2t. So, at t=0 if I compute the value,

composition or addition of two signals, sin t+

so this guy goes 0 ,I am looking at sin ( π3 ) and I get a certain value. At t=π / 6 the value
becomes 2.7 at t=π / 2 and so on and I can get this nice beautiful looking graph over here
right.

87
(Refer Slide Time: 02:20)

So, what happens in the frequency domain again? So, here the signal is again composed of

( π3 ) the graph of it looks like this and sin 2t looks like this.
two signals, sin t+

So, typically a signal when we represent, we write it or at least a sinusoidal signal as

A sin ( ωt +φ ) , with the amplitude, the frequency, and what you also called as the phase shift

π
right. So, here if you take these 2 signals t + , and 2 sin 2twell they have 2 frequency
3

components. What do I mean? now look at this signals, sin t+ ( π3 ) and if I compare with this
expression over here. So, here the amplitude is 1 right. So, the if I denote the amplitude as x

1
the amplitude would be 1, ω would be 1 and the frequency would, in turn be or the time
2π

1 π
period is 2 πtherefore, the frequency is and φ the phase shift would be right so this
2π 3

1 π
signal has ω as 1, frequency as , magnitude as 1, φ as .
2π 3

Now, look at the second signal 2 sin 2t; 2 sin 2t has. So, this is with the 2 here. So, ω would

1
be 2, frequency would therefore be or even its time period if you look at. So, this has twice
π

88
the time period as this guy, the magnitude is 2 and there is no phase shift. Now, if I were to
represent these two frequency components in a graph I look at the amplitude plot right. So,

this signal the first one: sin t+ ( π3 ) right exists only at the frequency of 21π ; at a frequency of
1 and with the magnitude of 1 right. So, this is its frequency. So, this is its frequency and
2π

1
the magnitude. So, look at the second signal 2 sin 2t which exists at a frequency of with the
π

1
magnitude of 2; in the graph it could be represented here and magnitude of 2. Now, what
π

about the phase plot? well the first signal sin t+ ( π3 )has a phase shift of π3 right. So, this
1 π
corresponds to frequency , I have a phase shift of and this guy does not have any phase
2π 3

1
shift. So, at , I'll just have 0.
π

So, this is the frequency domain representation of these two signals which are represented in

the time domain as a sin t+ ( π3 )+2 sin 2t. So, more on this will come later, but this is a nice
explanation of what is happening in the transition between time and frequency domain.

89
(Refer Slide Time: 05:33)

So, given function or a signal what we see here is we can actually switch between the time
and the frequency domain; and this can be easily done using some kind of a mathematical
transform right. And what is the transform? well Fourier already told us that any signal in the
time domain can be represented as summation of sinusoids of different frequencies, we have
this infinite series and what we call as the Fourier series and also resulting in the Fourier
transform and why are sinusoids so useful to us?

So, if I take a system which is LTI which means linear time invariant system and as the input,
I give some sine wave right, and then the output would be some sin wave may be of different
magnitude right and some phase shift. So, sinusoids are preferred because they do not change
shape when passing through an LTI system, there can only be an amplitude gain and phase
shift everything else remains the same. On the other hand, if I send square wave right the
output will not be a square wave and therefore, analysing sinusoids is important or is useful
because if I send the sine wave I’ll still have a sine wave over here. Only thing is if this is
sin ωt with a magnitude A , this guy could be with some other magnitude sin (ωt +φ), that
depends on what is sitting inside here, and this Á , φ depends on what is sitting inside this LTI
system, but the frequency remains the same this does not change; and of course, I can I could
analyse this signals based on the Fourier series that you know that any signal can be
represented as summation of sinusoids.

90
And therefore, I could even analyse these kinds of signals by making use of this
transformation at the sin wave if here transforms very beautifully to this side.

(Refer Slide Time: 07:47)

So, are there any advantages or disadvantages of writing things in the time or the frequency
domain; well each domain has or provides us with its own information that it depends on the
objective what kind of information we are looking for in the system. So, an analysis done in
one domain could be in sometimes advantageous over-analysis in other domain that we will
investigate slowly as the course progresses.

For example, a signal having characteristics that change with frequency can be easily
analysed in frequency domain compared to the time domain alright. So, here in this signal; in
this things if you see, so here the frequency domain is more or less not very informative right
that there is some kind of different things happening as the signal progresses over time, also
possibly here may be again depending on the objective we could either use the time domain
or the frequency domain analysis.

91
(Refer Slide Time: 08:44)

So, what is the relation between what we have learned earlier and then what we are trying to
do now right? So, we all know or we discussed this extensively of modelling a mass, spring,
damper system or a second-order system with M ẍ+ B ẋ +Kx = 0, and if I aim to solve this
equation it would have a e−σt term and some sinusoidal term right. May be a first course on
differential equations will tell you this, and this number σ , A , ω , φ will depend on this
parameter M, B, and K.

So, we are at the moment not interested in what is exactly the σ ? what is exactly the A or the
ω or the φ? We are just interested in the form of the solution and the form of the solution has
an exponent term which is well possibly decaying, in this case, it is decaying and then a
sinusoidal term right.

So, the solution to the equations has an exponential term and a sinusoidal term. The
exponential term is usually because of the damping term here. So, if B goes to 0 then I will
have a system which is M ẍ+ Kx = 0, which we know has a solution of the form A sin ωt ,
possibly with the phase shift or not. So, the solutions will just look like a periodic one right;
these ones. And this has a nice physical interpretation also right that if there is no damping in
the system then my system will just keep on oscillating and the energy will keep transforming
just between the mass element and the spring element here, and this once I have B, I have
some kind of a exponentially decaying term which the response would then look like

92
something like this right. That you know that this response that'll eventually go to 0. We'll do
this analysis little later right.

So, therefore the exponential term is due to the damper whereas, the sinusoidal term is due to
the interconnection between the mass and the spring, if there is no B they'll just keep on
exchanging energy, if there is a B the energy will slowly exponentially decay to 0 right; so
this how they look like. So, sin ( t )is just a nice periodic signal for all times t. e (−2 t)for
example, just goes down to 0 exponentially, and the combination of these two signals

e (− 2t ) sin t , it just looks something like this right these are also called as damped oscillations.

So far, what we observed is that solutions could have a combination of exponential terms and
sinusoidal terms or either just have exponential terms or also only a sinusoidal term as we
saw in this thing over here.

(Refer Slide Time: 11:43)

So, in the frequency domain transformation, the signals are decomposed into sinusoids which
are described by again an amplitude and phase at each frequency.

As we saw in the earlier examples and in addition to this now what we have to do is also to
account for instead of just sin (ωt +φ¿)¿ possibly with the magnitude A, we also have to
account for the exponential response which was e−σt in the previous slide right. So, can we
combine these two into a single idea or a single transformation that will help us analyse the
systems better. What we need is a new transformation, such that signals are decomposed into

93
both sinusoids and exponentials; why because our solutions essentially consisted of these two
components, the exponential component and some terms which were just sinusoids.

(Refer Slide Time: 12:47)

And that is where the Laplace transform comes into the picture; what does the Laplace
transformer do? The transform decomposes signals in the time domain into a domain of both
sine and exponential functions. So, here sine is not only sine, but it could also include cosine
terms and so on.

So, this domain is called the s domain or the Laplacian domain and invented by Simon
Laplace, and I think the name as the s comes from the Simon. So, s is essentially a complex
number right, s is usually σ + jω that we learn in our signal course is two dimensional; the
first dimension this correspondence to the frequency of the sinusoidal component, and
second, describes the exponent. So, if I just write down the signal of this form e power or a
component of this form e st . So, s is decomposed into the exponential term and the sinusoidal
term, and this expands as e σt ¿

94
(Refer Slide Time: 13:53)

So, by definition how do we find the Laplace transform? So, given a signal x (t ) right that
Laplace transform is given by L ( x ( t ) ), I call it as X ( s) because I go from the t domain to the

s domain usually computed as 0 to∞ , x (t )e−st dt . In most texts, you would see this as 0 minus
that would account to the question of what if there is an impulse at the origin or at when t=0;
we just start at t minus.

But for all purposes, we will stick to the notion of notation of just having the 0 here, right and
this L is the Laplace transform operator which transform the signal from the time domain to
the Laplacian domain via this expression. So now, the question would come can I solve this
integral? Does the Laplace transformer always exist? So, we'll try to answer those questions;
existence depends on the convergence of integral does this integral exists right, and in the
region in the splane for which this integral will exists is called the region of convergence, we
will see these things through the help of some examples.

95
(Refer Slide Time: 15:12)

So, let's take the very basic of the signals the impulse function right and I say, well find me
the Laplace transform of this, the impulse function right. So, which is usually denoted by the

δ ( t )and I just use the definition here ∫ x ( t ) e−st dt ; replacing x with δ am left with this
0

expression.

Now, I can just compute X ( s ) to be 1, where I just use the property of the impulse function
where the impulse is defined as 0 to ∞ ; δ ( t ) is 1 right and again most of the analysis here we
will do is again from 0, even though there exists something called the bilateral Laplace
transform which may which actually go from even the 0 replaced by -∞, but since throughout
this course we are interested only in causal signals or causal systems we restrict our analysis
to just this one right thing starting from zero. So, making use of this property that 0 to ∞or
even say this could also go from -∞to plus ∞ the δ ( t )equal to 1. I have X ( s)=1. So, the
Laplace transform of an impulse signal is just 1.

96
(Refer Slide Time: 16:41)

Now, let's take an exponential signal, Laplace transform of e−at . I substitute into the definition
of x, x ( t )here e−at e−st dt and I use the basic formula for integration and what I see is X ( s),

1
where x (t ) is e−at , the Laplace transform is given by ; now, is this enough? does this
S+ a
exists all the time.

Well now, let's take the case here where of here I am just evaluating these thing right, e−( S+ a ) t
right and I say well do this from 0 to ∞. Now what happens if this guy is positive right if S+a
is less than 0 then I will just have a term e? some number which is a positive number here
times t, and this ast goes to ∞ does not exist right it also goes to ∞, therefore, all these things

1
or the Laplace transform ofL(e¿¿−at )= ¿exists only in this region where S+a=0 right.
S+a

So, if I just draw it on the s plane spilt as σ and j ωright. So, the real part that σ , if I put, this
here- I call this −a. So, this guy will only exist for these reasons right. So, this σ >−a
is for all other σ this limit here will not exist; thus I will have something like loosely speaking

e and I want to avoid those things, and that is the notion of the region of convergence which
∞

we had talked about in the earlier slide right. The region in the s plane where the Laplace
transform exists is called the region of convergence and this is precisely the reason why we
need to define the region of convergence right just because of this one and this one.

97
(Refer Slide Time: 19:09)

So, let us quickly run through just computing Laplace transforms of few very basic signals if
x (t )=t , I can compute this just by standard integration by parts. So, there I just use this one
of integral from the udv formula as we famously call it, I just skip through these steps, but we

1
will see what it means. So, the Laplace transform of t is simply written as .
S2

n!
Similarly, if I have t 2or t n, I can just generalize this to an expression like this at L(t n) is .
s n+1

2! 2
See if I have to write Laplace transform of t 2that is 3 that will simply be 3 and so on.. right.
s s
So, this is kind of straight forward to compute.

98
(Refer Slide Time: 20:08)

Let's take sinusoids what happens to the Laplace transform of sin at? well I just put into the

a
formula, and what I get is the Laplace transform of sin at is , and I can, we will follow
S +a2
2

this little computational process here where I can write sin as something to do some of some
exponential signals. Similarly, I could even do for cosine signals the cos Laplace transform of

S
cos at is .
S +a2
2

(Refer Slide Time: 20:47)

99
Now, why are this important? or do these guys come with beautiful properties? Well we will
investigate them one by one. Sometimes it may happen that I may not be able to compute the
Laplace transform just by definition, because I need to solve some complicated integrals
right. So, to simplify this process we learn certain basic properties of Laplace transform, and
these basic properties have been derived again from some basic definitions; and by making
use of this basic definitions we will; the basic definitions plus these properties we can then go
on to get Laplace transforms more kind of more complicated looking signals.

(Refer Slide Time: 21:36)

So, let's investigate these properties slowly and one by one. First is the linearity property
right. So, if I take two signals x1 (t )with the Laplace transforms X 1 ( s ) , x2 (t ) with the Laplace
transform X 2 ( s ) and I define a new signal which is a combination of these two:
ax 1 (t )+b x2 (t ). So, the linearity condition or the linearity property says that the Laplace
transform of this signal is simply a right with its equivalent Laplace transform, plus b with its
equivalent Laplace transform i.e.,aX 1 (s)+b X 2 (s).

Let us take a very simple example. So, I have two signals e at as my x1 (t ), and say e bt as my
x2 (t ), and I want to find Laplace transform of a signal which looks like this e −e . So,
at bt

what is this? This is simply according to the property is L eat −L ebt , and that is simply

1 1
computed to be − . So, I am just individually computing and then adding up.
S−a S−b

100
The next property would be the time-shifting property, where if I have a signal x (t ) right now
I have t shifted by some number t 0, the Laplace transform under the time-shifting thing
would just be e−st X ( s ). So, this is kind of also straight forward to verify. Say if I have signals
0

that x (t )=t right which essentially looks like this and then I have another signal which is. So,
this x; let me call this x ( t −a ) =t−a , which is like this. So, the Laplace transform of this guy

1
I know is , now what is a Laplace transform of x ( t −a ) this would simply be, well I still
S2

1
have this X ( s) which is and in addition I have e−s with the amount of time t 0 which means
S2

1
the signals has shifted; so this guy a. So, this is e
−sa
. So, this is just how it looks likes.
S2

(Refer Slide Time: 24:31)

So, the third property is the time scaling property where I have a signal again x (t ) with an

1 S
equivalent L X ¿), I have x (at ) would just be X
|a| a ()
. So, let us see this in terms of say a

simple say sinusoidal signal.

1
So, from the earlier derived formulas let's say Laplace transform of sin t would be 2.
S +1
Now if I instead of t, I have to find what is the Laplace transform of sin atright. The time is

101
scaled with the factor of a like in one of our examples we had the signals 2 sin 2tright in the

1
second slide. So, what would this be? based on this property this would be , with the
|a|

1
S
Laplace transform of the same signal with s replaced by . So, I have this S 2 . Now I
a
a
+1 ()
a
write this down and expand this and what I have is now this guy would be , this is what
S +a2
2

we had exactly derived earlier.

Now the next property is time reversal is kind of straight forward, you just replace the −t
with the with the −s, we do not do any examples from this.

(Refer Slide Time: 26:11)

Now, the next important property is that of time differentiation. So, I have a signal x (t ) with

dx (t)
its equivalent LX (s). So, what the property tells me is a , the Laplace transform of this
dt
would be sX ( s)−x(0), where this is an initial condition of the signal. And similarly, I can
even do for higher order derivatives. So, let us see if you would like to do an example with

102
this. So, let's say I have a signal say, I start with sin t for which I know that the Laplace

1
transform is 2 .
S +1

d
Now Laplace transform of sin twith this formula assume that you know x ( 0 )=0 for
dt

d
simplicity, that sin t, the Laplace transform of this guy with this property would simply be
dt

1 d
s( X (s)) right. Now what is x ( s ) ? This is 2 and therefore, Laplace transform of sin t
S +1 dt
would be s multiplied by the Laplace transform of the signal itself and this is again nothing,
but the Laplace transform of cos t as we had verified earlier ok. Similarly, with the time

1
integration right so if I have a signal I integrate with 0 to t, this is just as replacing it by if
S
we just do the reverse of this example and that will be kind of obvious here.

(Refer Slide Time: 27:56)

Next would be frequency differentiation. So, if I take a signal x (t ) right and I multiply it. So,
x (t ) has an equivalent Laplace transform of X ( s), I multiply this by t this would amount to
just differentiating the Laplace transform of this original signal by s and adding a negative
sign. So, let's say I have to do this Laplace transform of t of some signal sin at.

103
 −d
So, this would by this property just be right this guy times the Laplace transform of only
ds

a
this guy right, X ( s) that is and this becomes now a straight forward thing for me for
S +a2
2

compute right and than just taking the formula then you know multiplying this by e−st and so

2 as
on. So, this just becomes 2 right. So, this is kind of a very useful property for me and
( S2 +a2 )
I can just send this to when I am just multiplying it with t n times.

Similarly, I have the property of frequency integration that if I have x (t ) with the equivalent

∞
1
Laplace transform X ( s)= x(t ) would just be this ∫ X ( u ) du.
t s

(Refer Slide Time: 29:30)

Frequency shifting right if I have again the signal, signal x (t ) with the equivalent Laplace
transform X ( s) I'll multiply this signal with e S t it would just be amount to substituting in the
0

original Laplace transform of the signal swith s−s0. So, let's say I want to do this something
like this Laplace transform of my original signal is sin at and I say I just multiply this to the
left by e at.

104
 a
So, this would simply be what is this? So, the Laplace transform of sin at is . Now
s + a2
2

multiplying this by e atwould just mean substituting s=s−s0 in this case s0 is a. So, this

a
would just be 2 2 and similarly if I have periodic function with some period p the
( s−a ) +a
Laplace transform is simply given by this formula, that's why I don't really need to go from 0
to ∞, but I can just make use of substitute.

We not derive each of this we may have learn this in our earlier courses but I am just doing a
bit of quick recap of these things. Now some other important properties; so first is called the
Initial Value Theorem.

(Refer Slide Time: 31:05)

So, what is the initial value theorem do? it relates again the signal to the in the time domain to

the one in the s domain as time approach is zero. So, lim

t →0
x (t) would be lim sX ( s). So, let us
s →∞

take this little example. So, the signal x (t ) goes as 3+4 cost, if I just do the limit 3 stays as it
is the cos(0) becomes 1. So, I have a four here. So, the limit goes to 7.

So, in the s domain I just use this formula that slim sX ( s) is lim ¿ and I multiply with s , and
→∞ s→ 0

then the Laplace transform of x (s). 3 is like a step of size 3 that will be the Laplace transform

3
would be , and 4 cost and we in our earlier slides had computed the Laplace transform for
s

105
cos t. So, I just do all the computations take the limit and I come at with this number 7. So,
only thing which we need to be careful here is that the Laplace transform should exist. I will
give you counter-example of this shortly.

(Refer Slide Time: 32:26)

Similarly, with the final values theorem what happens to my signal as time goes to infinity?

So, the final value theorem says that lim x (t) is equivalent to lim sX ( s ). So, let's take an
t →∞ s→ 0

example.

−2t
1−e
So, I have this example x (t ) is , I do this simple computation and I get the final value
2

1
that this signal convergence to as e−t goes to ∞ as tgoes to ∞. So, this term will disappear.
2

1 −2t 1
So, how does it look in the Laplacian domain? Well 1 becomes , e becomes , I apply
s s+2

1
the limits and I get . Again we have to make sure that the limit actually exists. So, let us
s+2
do something very quickly of a case when the limit may not exist all the time. Let us take the

1
case of a signal , s−3 and I say I just blindly apply this final value theorem at limit s going
s

−1
to 0, s times this would give me a value of minus now is this correct?
3

106
So, let me also do the time domain analysis like I do it over here. So, this signal can be

1 1 −1
written as − we have a s−3 this going away and minus here, and if I write on the
s s−3 3

1 1
equivalent time-domain representation of this. So, what I will have here is , and
3 s

1 3t
corresponds to 1 in the time domain like Laplace transform of 1 is −e . If I take the limit
s
of this signal as t → ∞, I see that this term actually blows up right. The limit does not exist
and therefore, I just do not blindly apply I should make sure that first the limit actually exists.
So, this final value theorem and also the initial value theorem is applicable only in the cases
where Laplace transform exits and it limit exists as s→ 0 and also the final value should exist
that is what we should careful of when we apply the final value theorem.

(Refer Slide Time: 35:18)

So, what we have done so for is had a quick recap of Laplace transforms basic properties and
how we use those properties to compute Laplace transform for some little complicated-
looking signals, and then the initial value theorem, the final value theorem and what we
should be careful of while applying those theorems. Next, we will discuss the inverse of
Laplace transform. So, if I can go from the time domain to the s domain can I actually come
back and another important property called the convolution which will be very important for
us in the control setting.

107
Thanks for your attention.

108
 Control Engineering
Dr. Ramkrishna Pasumarthy
Department of Electrical Engineering
Indian Institute of Technology, Madras

Module - 02
Lecture - 02
Inverse Laplace Transforms

In this lecture, we will continue with what the properties which we had learned with the
Laplace transform. Now we will do the inverse Laplace transform say if I can go from t to s,
can I come back and is the transformation unique; what are the properties that are preserved
in this transformation going from the s domain to back to the t domain? So, as I mentioned
earlier what we will deal with essentially are causal signals and therefore, the integration we
compute is essentially from 0 minus or 0 to ∞, in that case, the transformation or the inverse
Laplace transform tends to be unique.

(Refer Slide Time: 01:04)

So, I said earlier the inverse Laplace transform helps me to go from thes domain back to the
time domain, and there is a reason why we are actually doing this, which will be obvious few
slides from now. So, how do I compute this if given X ( s) the inverse Laplace transform is

1
computed by this formula L−1 of X ( s) is ∫ e st X ( s ) ds; this L−1 is called the inverse
2 πj
Laplace operator. Now, one could be tempted to believe that I write any expression in s and

109
there will exist an inverse Laplace transform while that's not true. Inverse Laplace transform,

2 2
S +a
for example, does not exist when X ( s) equal to s or even when X ( s) is for example,
s
or maybe this s may be replaced by a. These are special kind of transform which we will deal
with later.

But just to suggest that anything written as a function of s does not necessarily have an
equivalent time-domain representation.

(Refer Slide Time: 02:22)

So, similar to the Laplace transforms there are also some good properties of the inverse
Laplace transforms right, and then these properties, well they simplify the process of finding
inverse Laplace transforms for a complicated looking expressions right, and this is the one of
the usefulness of Laplace transform is also due to the fact that the inverse is most of the times
easy to determine and therefore, we can start from the time domain go to the s domain do the
computations and come back again to the time domain.

110
(Refer Slide Time: 02:56)

So, back to all these properties which we have learned earlier they also exist here. The first
thing is the linearity property. Say I have two signals in the Laplace domain denoted as
X 1 (s)¿ X 2 (s) with their equivalent inverses being x1 ( t )and x2 ( t ). Now if I have a combination
of these signals, that I have aX 1 (s)+bX 2 (s) the inverse of this entire sum would be a times
the inverse of X 1 ( s ) which is x1 (t ), plus b times the inverse of X 2 ( s ) which is x2 ( t ) i.e.,
ax 1 (t )+bx 2 (t). So, which means let's do a little example. I want to compute L−1 of a signal

1 1
which looks like this + .
S S+ a

So, what I do I just do because of the linearity property I compute L

−1
( 1S ) + L ( S+a
−1 1
) , and
I know that the L
−1
( 1S )is 1 and this L ( S+a1 ) would be e
−1 −at
. So, that's just a straight forward

property.

Next, I have the time-shifting property, this again is which we had in the transformation from
x (t ) to X ( s). So, suppose I have X ( s) with an inverse of x (t ) and now I replace swith s−s0,
then the equivalent time-domain signal would be of course, x (t ) would remain as it is and I
have an extra signal, e s t which is the shift in time here.
0

111
 S−a
So, let's take a signal which looks like this and I want to compute L−1 of 2 2 right. So,
( S−a ) +ω

S−a
here I have s shifted by this numbera; 2 2 . So, this could be written as. So, if I just
( S−a ) +ω

S
look at independently of this how x (t ) or X ( s) would look like. So, this is like right
S +ω2
2

and with s being transformed to s−a, this thing looks like this. So, this means that I can use
this property of time-shifting that my signal in the time-domain would now just look e at.

Now, I know the inverse of this guy this is my X ( s), the L−1 of this guy is simply cos ωt. So, I
used this very nice beautiful property over here, s−s0 would be the original signal plus or
multiplied by e s t , where s0 is the amount of shift here.
0

(Refer Slide Time: 06:14)

Similarly, with time scaling; so I start with the signal again X ( s) with an inverse x (t ) and I

1 t 1 t
multiplys by a then the inverse is just given by
a
x ( t ), the tnow replaced by i.e., x
a a a
. ()
So, let's again do some example and these are pretty simple looking examples L
−1
( 4 S2 S+1 )
2

2S
this is L
−1
( 2)
( 2 S ) +1
. This is almost like s replaced with 2 s. So, s replaced with 2 s if I just

112
 S
write down the signal without the time scaling it will look as 2 ; again the cosine square
S +1

1 1
kind of signal. So, this will simply be , what isa? a here is2, .
a 2

t
Now, this is the inverse of the original signal cos tnow with t replaced by . That's a simple
2
way of computing this. Time reversal is very straight forward X ( s)goes to x (t ) and therefore,
X (−s)goes to x (−t )no need of an for example, here.

(Refer Slide Time: 07:41)

Next is multiplication by s. So, take again a signal X ( s) with an inverse in time domain being
x (t ) , then s X (s) the inverse would be again I start with the signal x and I just differentiate
it; well assume that dx at the initial condition is 0. So, again a simple nice looking example.

aS a
Say take L
−1
( 2
S +a )
2 , now this is inverse Laplace of this signal ( )
s 2 2 , this is my sand
S +a

a
this is myX ( s)and this looks very familiar now right X ( s) being ( S +a2
2 ) what is the
equivalent x (t )? x (t ) here is simply sin at.

113
Now, going by this one, so s X (s) would simply be the differentiation of the original signal.

d d
So, the inverse Laplace transform of this guy would be right. So, this guy of the
dt dt
original signal, what is an original signal? this is sin at. Now, this can be computed to be
cos atwith a. Now you go take the Laplace of this and you back to here, what is inside this
bracket here. Similarly, division by s would give you the integral of it, X ( s) with the

X (s)
equivalent of x (t )would be. So, therefore, the inverse Laplace would be this one.
S

So, you see we observe that we are actually using these arrows both ways; that means, I can
go from here to here and also here to here. So, well I don't think we need an example, let's

1 1 1
quickly do this. L
−1
( )
S ( S+1 )
. I can write this as L−1 ; I have ; I have
S S+1
this looks like

1
this is my X ( s), this is . Now this can now be equivalently written as integral from 0 to t;
S
now the original function. So, this should be x. So, the original signal now look at this guy,
this is my original signal X ( s), the original signal would, therefore be e−t . So, I am just

∫ e−τ dτ , and that will just be ( 1−e−t ). Now you can just do the Laplace transform of this and
0

recover the original signal.

(Refer Slide Time: 10:59)

114
 d
Now frequency differentiation, so if I in my Laplace domain differentiate my signal that is
ds
simply here amounts to multiplying it with ( a−t ).

1 d
Let me do a simple example here, let's say I have a signal, I say it starts with X ( s) is , is
S ds

−1 d 1
S 2
. Now the inverse of this S 2is inverse of
ds S ()
. Now what is the original signal? X ( s)

1 1
going to x of that is right , that in the equivalent time domain would just be the unit step
S S

−1
or I'll say 1∙u(t ), this multiplied by −t. So this will simply be−t. So the inverse of is
x2

−1
sorry if the inverse of this signal is simply −t and similarly with a frequency integration.
x2

∞
So, I have X ( s) again going to x (t ), the ∫ X ( u ) du will again go back to a signal like this
s

again I'll not do an example for this, but its straightforward to verify once we know property
number 7, 8 is easy to verify.

(Refer Slide Time: 12:39)

The other important property is what we learn again in signals is the property of convolution,
and why do we do this? is essentially because in a control terminology or a control system,

115
you essentially what we saw in one of our first lectures is I have an input which goes through
a system could be something here and if I and then I want to measure the output y (t). Now,
this y (t) is in some sense, some kind of a combination of u(t) and h(t) right and in signal's
term, this would be it's a convolution of u ( t ) h(t), this is the convolution is defined by this
property and this is also commutative h(t) convolved with u(t). So, convolution is a
mathematical way of combining two signals, so here my input together with the system to get
my output of the system.

And this is how we will view things in control. So, it is an integral that measures the area
overlap of one signal as it is time-shifted over the other, we'll see a little graphical
representation of this. And why is this important for us? Because throughout this course it is
important for us to study the response of system response is y (t) the system is h(t) for a
given input signal right.

(Refer Slide Time: 14:12)

So, this is not simply multiplication that u(t) times h(t) will give me y (t), it's a little more
complicated than that. So, let's see that I have two functions a rectangular function and g(t) is
also a rectangular function, and we will see how the convolution of these two gives me a
triangular signal.

So let's, I think; I hope it will play again. So, this is some of the videos directly taken from
Wikipedia. So, thanks for Wikipedia to make life a little easier for us. So, you see this is one
rectangular signal convolving with this red guy, and you see that it just the yellow one or the

116
area. So, this is the area that's being computed here. So, this goes and then the area reaches a
term maximum here and then again the yellow line keeps or the yellow area keeps decreasing
and goes to zero. So, the convolution of a rectangular signal with another rectangular signal is
just a triangular wave; the black signal here.

(Refer Slide Time: 15:11)

So, this actually looks quite complicated and if you look at even the expression that I have to
actually compute you know x (τ ) y ( t−τ ) dτ and so on to get signals which are the output of
those, and this actually becomes too complicated sometimes in doing the integration. But
then when they are transformed into the Laplacian domain the convolution things become
very straight forward. So, what is the convolution theorem says? say if I have a signal in time
domain x (t ), and I want to see its convolution with another signal y (t) it just becomes a
simple multiplication of their individual Laplace transforms.

So, the convolution of x (t ) with y (t)is this: a very simple multiplication of X ( s) and Y (s).
So, therefore, if I now look at in terms of control what I had written earlier that I had a u(t)
here, I had h(t) and I had a y (t) here right. So, my x (t ) which earlier was h ( t ) convolved with
u(t), now I can write this simply has Y (s) is some H (s)U ( s) as simple as that. So,
convolution becomes a very simple multiplication or a product. And this actually helps now
to solve lots of problems. So, if I want to see how this signal looks in the time domain given
u(t) and h(t), I first go to the Laplacian domain I compute h , I compute u, I do the
multiplication and now y (t) can be easily computed as the L−1 H ( s ) U ( s), this tricks we know

117
how to compute right with several properties and therefore, studying convolution becomes a
very easy property just by a multiplication and the inverse Laplace transform.

(Refer Slide Time: 17:07)

So, as an example right; so I want to find the convolution of a signalt u(t ) with sin t u(t ), this
is like the unit step. So, x (t ) is sin t u(t ), and y (t) is t u(t ) if I just go by the formula x∗y

∞
going ∫ x ( τ ) y ( t−τ ) dτ. I just substitute for each of the signals, and since u(t) is the unit step
−∞

starting at t=0, I need to compute this kind of complicated looking integral right. Instead
what I could do is I can just go to the Laplace domain or the s domain where I compute the

1
Laplace of sin t u(t ) that is this guy plus this t sin t would be 2 , Laplace transform of t
S +1

1
would be .
S2

1
And I can write this as sum of these two Laplace transforms and I know that the inverse of
S2

1
is t and the inverse of 2 is sin t. So, the inverse of the convolution x∗y in the time
S +1
domain simply looks like this, and I don't have to go through this complex process of
computing this integral.

118
(Refer Slide Time: 18:26)

So, one advantage we saw so far is that, the Laplace transform easily helps us in computing
the convolution, where convolution is just a multiplication. The second advantage is in
solving Ordinary Differential Equations. So, solving an ODE in the sdomain is much simpler
because. So, let's see this little example. So, I have this example

2
d x dx
2 +2 −x(t), this is my differential equation maybe this is equal to 0 on the right side.
dt dt
So, now, what does this expression become? well this expression on the right side becomes
by using those formulae for differentiation ( S2 +2 S−1) X ( s ). well, this is under the
assumption well we just make it for simplicity that the initial conditions all are 0.

So, a differential equation here is now transformed to a linear equation, I know very well how
to solve linear equations, and this Laplace transform is applicable to lots of signals
continuous signals, piecewise continuous, periodic, step and impulse functions.

So this little question that this should actually be −x (t ). So, this x (t ) transforms via L to

2
d x
X ( s), thanks for pointing out. So, 2 transforms to S2 X (s) with an initial condition being 0,
dt

dx transforms to
2 2 S X ( s) and x ( t )simply transforms to X ( s ) . So, this should actually be a
dt
x (t ) here thanks for pointing that out.

119
So this transformation from a differential equation to a linear equation would make
computations very easy, and we also know so if I determine X ( s) from this like me equating
this to 0, and I can then find what is x(t) is simply the inverse of X ( s ) .

(Refer Slide Time: 20:49)

So, we will just see, how we could do that? So, let's consider a simple RLC circuit. So, I have
a voltage source, I have a resistor, inductor and a capacitor and let's say all initial conditions
are at 0 and V is simply a unit step.

And for computational purposes we will assume that R is 2, L is 1 and C is 1. So, first thing I
would know is well I will just apply a KVL to get these dynamics of the system right. So,

1
di 1
V ( t ) =RI ( t ) + L + ∫ Idt right. Now the Laplace transform V transforms to V (s), I
dt C 0
transforms to I ( s), the differentiation of I would multiply it by s with I ( s) the L remaining
as it is and with some initial conditions which are assumed to be 0 and we'll see that why this
guy also eventually will go to 0, because an inductor does not allow a rapid change in current

I (s)
and then the integral property right ∫ I dt would be with the C remaining as it is.
S

V ( s)
So, the current if I were to solve for the current. The current would be I ( s) is 1
R+sL+
sC
right just this differential equation now transforms to a very nice linear equation over here.

120
(Refer Slide Time: 22:18)

So, let's try to find out how the solutions look like this for these values. So, I have

1 1
R+sL + , I substitute values, u(t) is a unit step, v(t ) is a unit step therefore, V (s) is and
sC S

1
for values of R, L and C, I think this should correspond to R=2 right. R=2 would be and
S

1
all these expressions and I just write this to be I ( s) is 2 , and now I have this it in the
( S+1)
Laplacian form I can easily compute what is the time equivalent of that.

1
The time equivalent of that is an inverse of this guy, Laplacian form inverse of 2 and I
( S+1)
can use the properties of my inverse Laplace transforms to just get that the signal in the time
domain is just t e−t , which is the solution which I am looking for right. So, this differential
equation or the dynamics initially were differential equations here; I convert that into a linear
equation to get an expression for the current in the s domain, and I do the inverse Laplace to
get the expression in the time domain.

121
(Refer Slide Time: 23:47)

Now, I will do some other example with non-zero some initial conditions,
ÿ−2 ẏ−8 y=0, y (0)=3 and ẏ ( 0 )=6.

So, the first step would be to write this equation in the Laplace domain. So, this would be

s Y ( s ) −sy ( 0 )− ẏ ( 0 ) −2 ( sY ( s )− y (0) )−8Y ( s )=0. So, this would mean that Y (s) could be
2

3s
written just in term of an expression in slike this, 2 or just that the denominator
S −2 S−8

3s
could be written as a product of two of these terms . So, applying the rule of
( S−4 ) ( S+2 )
partial fractions which we would have learned while dealing with solutions of equations, I

2 1
can write Y (s) as the composition of two signals now, + .
S−4 S+2

122
(Refer Slide Time: 24:59)

1
And I can now easily do the inverse Laplace to find out that the 2 L
−1
( S−4 )+ L ( S +21 )
−1

gives me the following expression for y (t) right. So, again the process becomes very
straightforward I am just now solving for linear equations in the s domain, and then I am
applying the inverse Laplace transform to get the solutions to equations right.

(Refer Slide Time: 25:24)

So, to summarize what we have learned in this lecture is inverse Laplace transforms and its
properties, a very beautiful property of convolution where the convolution test translates now

123
into a very simple multiplication in the s domain and solving of ODE's using Laplace
transforms. In the coming lecture what we will look at is to write down the system or a given
set of equations as a transfer function and then explore its properties, we will try to do
stability analysis, its response to various kinds of signals and so on that will be in the coming
lectures.

Thank you.

124
 Control Engineering
Dr. Ramkrishna Pasumarthy
Department of Electrical Engineering
Indian Institute of Technology, Madras

Module – 02
Lecture – 03
Transfer Function Modelling
Block Diagram Representation

Hello everybody. In this module on module 2, lecture 3; we will talk about Transfer Function
Modelling and what we also call as the Block Diagram Representation. And I'll shortly tell
you what these things mean? And the advantages of the transfer function modelling and how
the block diagram representation simplifies for us the analysis of systems?

(Refer Slide Time: 00:45)

So, let's start with an example again this we have seen earlier, let's say I start with an RLC
circuit and given a certain initial voltage; I would ask you to compute what could be; so, if
this is the input I'll ask you to compute, what is the output voltage to the capacitor? Or even
what is I (t )? So, what would I do? I would start writing down the appropriate voltage loss

dI 1
that V i ( t ) =RI ( t ) + L + ∫ Idt , this is the voltage.
dt C

125
(Refer Slide Time: 01:28)

Now, the problem is to find V o for a given V i, so what does this mean? To find that the time
response or how V o changes with a given V i; we need to solve certain set of differential
equations; ordinary differential equations or even sometimes integro-differential equation;
you see the little you know integral term over here. So, what you also earlier saw is that when
I transform my differential equations into the Laplacian domain or the s-domain the
differential equations turn out to be simple algebraic equations which are then easier to solve
you know linear equations are usually easier to solve than differential equations.

So, we transform the model into the s-domain and then we would define what is called a
transfer function and this is a signal which will relate a given input to output which will be
computed again by what is called the transfer function of the system.

126
(Refer Slide Time: 02:37)

And this is, of course, a model which is applicable for all kinds of input signals. So, what is
the transfer function of a system; well the transfer function of an LTI or a Linear Time-
Invariant system, time-invariant again let's recall the definition that the parameters of the
system would not change with time for example, the mass or a spring constant or a resistance
and inductance all these will be invariant of time.

So, for a linear time-invariant system let me also recollect that what we will here focus on
systems which are just single input, single-output systems also called SISO systems. So, for
these kinds of LTI SISO systems, the transfer function is the ratio of the Laplace transform of
the output to the Laplace transform of the input assuming 0 initial conditions. So,
mathematically if U ( s) is a Laplace transform of the input signal or the input function Y (s)

Y (s)
is the Laplace transform of the output then the transfer function G(s) is .
U (s)

So, how does this again look like, so I have input u(t) and output y (t) and this is some
system setting here. Some I can even say some say ordinary differential equation, this system
is represented by; so I transfer it into the Laplacian domain, I can have a Y (s) and the linear
relation which describes, how a certain input results in a certain output is called the transfer
function. So, given this transfer function, I can simply say that given input signal U ( s), the
response of the system or the output is simply the multiplication of G(s) and U ( s), we call
that in the time domain; we would actually have to do a convolution here. Convolution-So,

127
let say if y (t) is my output signal let me call g(t) as a system and some convolution with the
input.

So, now we get rid of these things, so this g(t) is replaced by G(s) and this G(s) is called the
transfer function.

(Refer Slide Time: 05:20)

So, a simpler way or one of the ways of looking at transfer function is to look at the impulse
response of a system which means the response of the system when the input is an impulse.
So, earlier so what we had; we had Y (s) was G(s) times U ( s) and then if U ( s) is an
impulse, we know that the Laplace transform of an impulse is just 1. So, I am just left with
G(s), so let us keep it here for the moment.

So, what is an impulse by definition is it is an infinitesimally narrow and you know and
infinitely tall signal; you just you can you know in the speaking terms just says that it takes
value of 0 everywhere; here its 0, here its 0 and I accept that t = 0; it is like an infinitely long
signal such that the area is 1. So, how do we construct this impulse signal right it might sound
a little superficial here, but let's say I start with the little rectangle here and this rectangle has

1
a width of aand a height .
a

128
 1
So, now what is the area of this guy? it’s 1; area is a is 1 and if I call this signal as say the
a

rectangle r (t) then I have ∫ r(t ) would be 1; there is 0 everywhere else to the left of this
−∞

point and even the right of this point and this width is a; from here till here. So, you can call

a a
this as in the negative side, an in the positive side.
2 2

So, based on this how do I construct an impulse? So, let me take the limit of this signal as a

1
goes to 0; nothing changes right it's still, the area is still a and the limit of this signal a; if I
a
take the area of the signal, it would still be 1 and therefore, I have this signal like this that

∫ δ ( t )is 1. So, if the input to the system is the unit impulse, then the output is called the
−∞

impulse response and you see in that case; the output which I observe is equal to the transfer
function; at least mathematically I can see that. So that means, this transfer function is just the
impulse response for an LTI system again, you look at the initial conditions to be 0.

So let's do a little example of this, so I can take the case of simple pendulum, a point mass
say everything is normalized to 1 and say if I just measure θ this way and I say I just give an
impulse here; it may not be practically possible to give an impulse like this, but it might you
can just look at it as a signal like this it is a very small a. So, what would you expect I just
you know loosely speaking I just push it away so that so I gave it a little push?

So, it will just somehow follow its natural response, if it is a system which has no damping it
will keep on oscillating all the time. If it is a system which has damping, the oscillation will
die down; this is a natural response to the system. So, or even look at it in a different way say
the natural response of a human being, especially at 7 in the morning. So, the alarm clock is
the impulse response, so the alarm perturbs you; you just you know switch it off everything
happens in the flash time and then you go back to sleep and that is your natural transfer
function; is the sleeping condition and that is the impulse response.

So, but if there is somebody trying to wake you up for a very long time then your response
might be different, we will see how those responses would be but, the impulse is just the
natural response of the system.

129
(Refer Slide Time: 09:41)

Now, the question is given a system; how do I find the transfer function? Well the first step is
to write down the equations. So, where do the equations come from? This again come from
my voltage laws KVL; KCL and maybe even the Newton’s laws all those things which we
have been dealing till now.

So, first write down the model equations of a system; second, identify the system input and
output variables and we will see that you know the choice of output can be different or it is
not always unique. Now, take the Laplace transform of the model equations which are
essentially differential equations and assume 0 initial conditions and then find the ratio of the
Laplace transform of the output to the Laplace transform of the input; I can always do this
because it is a linear equation.

130
(Refer Slide Time: 10:35)

So, again let us come back to this example which we started with, so I have V i as the input R,

Vo
L and C the circuit elements generate certain current I (t ) and I want to find what is ? So,
Vi

dI 1
this is easy to find out V i ( t ) =I ( t ) + L + ∫ Idtand this is my V o, this guy here.
dt C

(Refer Slide Time: 11:06)

Now, what is the first thing is to identify the input and output variables; input is the voltage
which I apply and output I identify as the voltage across the capacitor. So, my equation which

131
was an integral differential equation, if I just take the Laplace transform and use the
appropriate properties that differentiation becomes loosely speaking a multiplication with s,
integration becomes a division by s and I end up with an equation like this.

So, now step 2; identify input, identify output, do the Laplace transform and the last step

Vo Vo
would be just take this ratio . So, I could I always do this right just take this guy,
Vi Vi

1 1
divided by V i; I have this one
sC (
; I ( s) R+ sL+
Cs )
and I just end up with this transfer

function. Now I could also do it, you may ask a question what if I just write down here; what
if I ( s) or even say i(t )is the output; this is another choice of output, I will say instead of
measuring the voltage here as the output, I just measure what is the current? I could always
do that.

So, how does the Laplace transform look like; it is again we have kind of straight forward to

1
[
compute from here. So, I say V ( s )= R+sL+
sC ]
I ( s ). Now the transfer function G(s); just to

distinguish between this put a I here and there is an output here, how dimensionally

I ( s)
measuring the current as the output is the ratio of the output ; which from this
V (s)

1
expression can be simply computed to be something like this; 1 . Again these are
R+sL+
sC
transfer functions for the same system, but here I am measuring V o as the output and here I’m
measuring the current as the output.

132
(Refer Slide Time: 13:39)

Now, if I just start with some relation between input, so usually we will throughout this
course denote u(t) as the input signal and y (t) as the output signal. So, if I am given a
differential equation in Y (s) and U ( s) like this with again 0 initial conditions; how do I find
the transfer function well again I just do the same steps, I take the Laplace transform of the
left-hand side, so Y (s). So, I am just differentiating it three times, it will be s3 ,differentiating
twice will give me an s2, differentiating once will give me asand so on. On the right-hand

dU
side, I have +u, which will just transform to 10 s+1. So, once I have this; I can easily
dt

Y (s)
write down what is .
U (s)

133
(Refer Slide Time: 14:32)

So, given this transfer function what are its properties; well the first says that the transfer
function of a system is independent of the magnitude and the nature of the input; which
means it is essentially system-level property, it's just based on what is the system at hand; not
what is the input signal?

For example, if I take a system let's say any system with their transfer function G(s) and I
measure the certain output with the certain input U ( s). So, if this U ( s) is an say impulse, I
will have a certain output it could be whatever; the G(s) would be the same as when this
could be a step input with the corresponding output, it doesn't really depend on the choice of
the input, it is just a property of the system by itself.

The nature of the pendulum will not change if the input is a ramp or the input is an impulse or
even a step; the transfer function with the one which transfers in a way a given input signal to
an output signal is a property of the system itself. So, once I have G(s); I can easily study
what could be the transfer function for different kinds of inputs. For example,
Y ( s )=G(s)U (s), so this is fixed.

So, give me any signal U ( s); I can always study what is the response? if I were to do the time
domain, I will just go find the inverse Laplace transform for which I know lots of tricks by
now and so once I have the transfer function, I can study the response for various kinds of
inputs to see how my system would behave. Well now, is the transfer function unique? so
give me a transfer function can I say what kind of a system is this? can I say is it car, is it a

134
bus, is it an aircraft or whatever. Well, the answer is no, the transfer function does not
provide any information concerning the physical structure of the system; which means that
two different physical systems can have the same transfer function.

So, let me just take an example of the mass-spring-damper system, which we have studied in

1
our earlier lectures. It had a transfer function which was like 2 ; if I just take a
Ms + Bs+ K
simple case of M =B=K =1, I have this transfer function and the RLC circuit which we was
you in a previous slides, had a transfer function of this sort which for R, L and C being equal
to 1; again translates back to this one.

So, these are the same transfer functions for two different systems; therefore, if I just say well

1
2
, it could mean more than one system. Physically, I could realize with some
s + s+1
mechanical system or even as an electrical system; however, what the transfer function does
is; say if my input is the voltage and output I know is the current; let me write in the
Laplacian domain, then this G(s); will have the appropriate units. So, it could be the
impedance or the reverse of the impedance and so on, it depends on how we look at the

V (s)
transfer say if I just write it in; so, V (s) would be the output I ( s), would be . So, the
G(s)
G(s) here would capture the units of the inverse of the impedance similarly over here also.

So, the transfer function actually it captures the appropriate units to transfer the input signal
to the output signal, if it is a force under velocity it will have an appropriate unit over here
and so on.

135
(Refer Slide Time: 18:38)

So, in general, a transfer function would look of this form where Y (s) could be any
polynomial of order m in s, U ( s) the Laplace of the input could be written as any polynomial
of ordern which comes in the denominator. I could also factor that as some number K ' , I'll tell
you shortly what this is and do a series of what we call as zeros and what we call as poles. So,
here I call n as the order of the system, m would be the number of zeros; n would be the
number of poles.

So, there is something called also the DC gain here; which is essentially what happens when I
just supply a DC input. So, this DC gain is just obtained by substituting s with 0 and in this

bm
case, it would just become . If you are wondering where this comes from, it is just an
an
application of the final value theorem to the output Y (s)=G(s)U (s). So, when U ( s) is a step
and I apply the final value theorem s going to 0, sG( s), what is the G(s)? For a step input it's

1.
s

So, these guys go away so what I am left with is G(0), so this is we'll referred to this as the

bm
DC gain of the system computed as . So, this z’s are called the zeros of the system; I'll
an
shortly define them formally, these are called the poles of the system and if I write my system

136
in this form; my DC gain then would be G(0), that would be let's call this as K . So, the K
here would be this number K ' over or multiplied by the product of all the zeros with a minus
over the product of all the poles.

So we will see shortly, when this K ' is important to us, so one important thing to remember
here is that the number of poles is always greater than or equal to the number of zeros
because if the system has more number of zeros. So, if the system actually becomes non-
causal, if you have a number of zeros more than the number of poles. For example, if I have a

2
s +1
system and this is a non-causal system and this is also not physically realizable by any
s+1
known components.

So, and you could also look at it in a way that if I try to compute the inverse Laplace

2
s +1
transform of this guy; ; it will not exist. Similarly, if I just take even simpler ones and
s+1
therefore, if in the future classes, when we are designing a controller its transfer function; if
you know analytically say well I got a beautiful controller which again has a structure like
this; this is incorrect.

This means this never be physically realizable, so what should be important in all through our
problems is that the number of poles should always be greater than or equal to the number of
zeros, if not the system becomes non-causal and it will never be physically realizable. So, we
will keep this in mind and recollect this again when we do a control design and we will see
how to actually overcome this problem.

137
(Refer Slide Time: 22:52)

So, by definition what are poles? well these are the roots of the denominator polynomial of
the transfer function; how does this equation looks like just take this denominator polynomial
and just equate this to 0; it will give me set of n solutions and those are the poles of the
system and what happens to the system; when s which is the value of poles when s is equal to
some P i; G(s) close up to ∞ at a very simple example could be something like this when I

s+1
have .
( s+2) ( s+3 )

So, here the poles are -2 and -3, so if s equal to -2 this value just goes to ∞ . Similarly, zeros
are the roots of the numerator polynomial said the numerator to be equal to 0 and what I then
get are the zeros of the system. So, what happens when I evaluate the transfer function at
zeros, so this will always be 0 just substitute; so, in this case, the zeros are at -1 and if I
substitute -1 for s here, might the value of transfer function becomes 0; because this goes to
0. So, these poles and zeros together with the system gain K ; characterize the entire input-
output response of the system.

So, what do I need or whatever have I learned so far is, given a transfer function I can write
my system as a set of poles, a set of zeros together with a possible system gain and these
three will tell me the exact behavior of a system given a certain input.

138
(Refer Slide Time: 24:45)

6 s+12
So, let's see the system right given by and I want to find; what are the sources
s +3 s2 +7 s+5
3

of system gain? what are the poles and the zeros? So, the DC gain just said s equal to 0 and I

12
can easily compute this, the DC gain of the system is . What are the zeros? Set the
5

−12
numerator polynomial to 0; 6 s+12=0; then this will give me s equal to =−2.
6

So, the zero would be a −2; so sorry this a little error here, so if I compute the zeros would
turn out to be at −2. Similarly, I take the denominator polynomial equated to 0 and I get s
equal to these three roots. So, at s=−1 ,−1+2 j ,−1−2 j, so what do we observe from here?
of course, the gain is will always be a real number, the poles and zeros they could either be a
real number like a −2 here, −1 here and if they are complex, they will always be in
conjugate pairs.

For example, there will never just exist a pole at −1+2 j; if −1+2 j exist, there will always
be a −1−2 j or in general the poles the complex poles will always be of this form;
a ± jband this is because all the coefficients of my transfer function are always a real value.

139
(Refer Slide Time: 26:20)

The second thing is; well if I have a system with different components; How do I visualize it?
Is there a nicer way to represent that system and how to understand the flow transformation
of signals in the complex system or given a series of interconnected systems; how do I find
its overall transfer function? We will try to understand that phenomenon of it and this is what
we will usually call as the block diagram representation.

(Refer Slide Time: 26:51)

What is a block diagram representation by definition? It is just a shorthand pictorial

representation of a system which depicts each of the systems sub-component, how signals

140
flow within components or from one sub-component to the other and this block diagram will
give me an easier representation of this kind of systems. There could be multiple systems
connected to each other and interacting with each other and we'll see also how having this
kind of representation, will give us some easier methods to find the overall transfer function
of the system; given that I know transfer function of individual components or the individual
subsystems.

(Refer Slide Time: 27:36)

So, various simple building blocks for block diagrams will be of four components once I will
have G(s); which is the transfer function, as we have defined so far; I will just have arrows to
represent the direction in which the signals flow.

141
(Refer Slide Time: 27:52)

Then I will have summing points say I have a signal U 1 ( s ) , a signal U 2 ( s ) ; there is a plus and
a minus here; the Y (s) would be computed as U 1 ( s ) −U 2 ( s ); if there is a plus here it will be
simply U 1 ( s ) +U 2 ( s ) ; take-off points well these are just like you are just adding an additional
wire over here or something like this is a U ( s) here; if I just put a wire here and take the
signal would still remain the same; since U ( s) here and the U ( s) also here, it just represents
the branching of a signal.

(Refer Slide Time: 28:23)

142
So, let us try to draw a block diagram of the simple circuit which we had started with; so, I
have a V , I have a R and a C over here; the L we just omit for simplicity. So, I can just say
that V i ( s )is my U ( s);V o (s) is my output signal and this has a very simple block diagram

1
representation of this format. So, we can just compute this transfer function to be ;
1+ sRC
this is a very simple representation of the system.

(Refer Slide Time: 28:55)

So, well what happens if I have systems which are in this form; this is U ( s) which generates
a certain Y (s) via G 1 (s); this output serves as an input to G 2 (s); G 2 (s) will have a certain
output which will serve as an input to G 3 (s) and so on. In this case, the transfer function is

Y (s)
simply this one is G 1 (s)G 2 (s)G 3 ( s). Let's see this, this actually is a straight forward to
U (s)
compute, so let me call this as Y 1 (s).

Now let me call this as U 2 (s), let me call this as Y 2 (s) and U 3 (s) and so on. So, for G 1 (s),

Y 1 (s) Y 2 (s)
=G 1 ( s ), similarly for G 2 (s); this will be the ratio of . So, there are; so
U ( s) U 2 (s)
Y 1=U 2, Y 2=U 3 and so on. So, Y 2 ( s )=G 2 ( s ) U 2 ( s ), now what I know is this is G 2 ( s ) . Y 1 ( s ).

143
 Y 2 (s)
Now, what is this Y 1 ( s )? Y 1 ( s )=G 1 ( s ) U 1 ( s ) ; so, therefore, if I were to compute what is
U ( s)
that will simply be G 1 ( s ) G 2 ( s ).

Similarly, I could even do the third cascade, so I will get the overall

Y (s)
=G 1 ( s ) G 2 ( s ) G 3 ( s ). So, here is what I call as the series connection; series
U (s)
interconnection in which case the transfer functions just to multiple.

(Refer Slide Time: 31:07)

Similarly, I can also have components interconnected in parallel and then what would be the

Y (s )
overall ? So, let's say that this U 1 ( s ) via G 1 ( s ) generates an output of Y 1 ( s ); here I have
U ( s)
Y 2 ( s )and here Y 3 ( s ); by the definition of the summing block Y (s) would be
Y 1 ( s ) +Y 2 ( s ) +Y 3 ( s ); what is Y 1 ( s )? Y 1 ( s ) comes via G 1 ( s ) and U 1 ( s ) . So, I will have

G 1 ( s ) U ( s ) +G 2 ( s ) U ( s ) +G 3 ( s ) U (s) and resulting in this expression that the entire G(s) is

Y (s)
=G 1 ( s)+G 2 (s)+G3 (s).
U (s)

144
(Refer Slide Time: 32:14)

I could also have things connected in what I call as a negative feedback loop, so I have a
signal here it goes via G(s), it is and this Y (s) is fed back via H (s) again over here. So, how
does the transfer function look in this case, so in the feedback form, so I want to find out

Y (s)
what is ? So, let's start from here where does this Y ( s )come from? Y (s) comes via G(s)
U (s)
and E(s) is written as Y ( s )=G(s) E( s); this is nice. Where does E ( s )come from? E ( s )comes
as U ( s ) −Y ( s ) H ( s); this is little summing block here with a plus and a minus. So, E(s) is
plus of this signal minus of this signal, so this I write here U ( s ) −Y ( s ) H ( s).

And I just do a little simplification, so I just say Y ( s )=G ( s ) U ( s )−G ( s ) H ( s ) Y (s), I eliminate

Y (s) G(s)
or I just get this guy to this side and I just get that = and this is as we call
U (s) 1+G ( s ) H (s)
as the negative feedback loop and we will loop and we will use this very extensively
throughout this course, so it might help you just to memorize this thing as a formula that if I

Y (s)
have a G(s) here, H (s) here, or the overall transfer function or I could even call as the
U (s)
overall transfer function between Y (s) and U ( s); given this individual transfer functions
here, the forward transfer function block has G(s), the feedback one has H (s) and this is the

Y (s) G(s)
overall transfer function from = .
U (s) 1+G ( s ) H (s)

145
(Refer Slide Time: 34:08)

Now, is this always true? if I say that you know I have G 1 (s), I have a G 2 (s) that the overall
thing is always G 1 (s)G 2 (s) and when I do this; I just everything is in s. So, I just sometimes
omit this, but let's assume that this is all in the Laplacian domain, is this always true? so let's
first talk of with an example, it's in this slide.

(Refer Slide Time: 34:35)

So, I could possibly you know if I just be a little lazy, I could just look at this system as you
know I have this one RC circuit connected to another RC circuit, how did the transfer
function of this RC circuit look like? let's go back few slides. So, this looks like something

146
 1 1
like this , so the transfer function of the first guy would be , the second guy
1+ sRC 1+ sRC

1
would be the same .
1+ sRC

Now, the question to be asked is; Is this the overall transfer function? let's see if that is the
answer.

(Refer Slide Time: 35:43)

Let me just derive these transfer functions starting from the four steps which I had listed out
earlier, I just write down the individual equations with the K V L here, a K V L here and then

V0
just compute (V output by V input). The first loop says
Vi

1
V i ( s )=R I 1 ( s ) + ( I (s)−I 2 (s) ), similarly I could write down the expression for V o and V o
sC 1
(s) is explicitly written down is in this form.

147
(Refer Slide Time: 36:15)

Now, I do all the math and the transfer function obtained by just eliminating I 1 ( s ) and I 2 ( s )
would be simply like this.

(Refer Slide Time: 36:22)

And this guy is not equal to the product of the individual transfer functions, the overall
transfer function of this system is not equal to the product of the transfer function of the two
individual RLC circuits even though they seem as if in cascade; What is going wrong here?
So, there is a catch here is that whenever I want to do an interconnection like this, the

148
assumption is that while deriving the transfer function; there is no loading that there is no
power drawn at the output of the system.

So, this assumption must be satisfied all the time while deriving the transfer function in this
form. So, if one component is acting as a load on the other component; the transfer function
you know it just cannot be determined individually. So, we should actually look at both
components together, they should look at in combination. So, what is happening here? if you
see this circuit, the overall transfer function is not equal to the product of the individual
transfer functions. This is because again the assumption of no loading fails; how does it fail?
The second RC circuit, it draws energy from this component here; this one, it draws energy
from; this is where my output is measured for the first; I write from here till here.

So, this circuit draws energy from the first one and hence it individual transfer function is not
valid when it is in cascade. So, this is what we have to be careful when we apply the formula
of two transfer functions in cascade, the overall transfer function being the product of it. We
will possibly see little more examples on this; for the moment we will just conclude by saying
that this is not always true. Of course, and they are ways to make this which we will postpone
to a little later, but before that let's try to do what I would call as a block diagram
representation of a transfer function.

(Refer Slide Time: 38:29)

A very simple example; I take DC motor which is also called as separately excited DC motor,
the field is supplied by an external source; I have the armature, certain armature resistance, a

149
certain armature inductance and this we know produces a back emf, this back emf results in a
torque; which produces a certain displacement θ; if I consider J and D to be the moment of
inertia and frictional component of the motor. So, in this sub-systems; what are my
components? So, if I denote R a as the armature resistance, La as armature inductance, I a is the
current of the armature.

(Refer Slide Time: 39:08)

So, if I just write down the equations for this; the first one, so I have E a as the applied
voltage, this is R a; this is La and this will have a certain back emf. So, the first of my
equations if I were to write down would be E a equal to and if I just call this I a ; from these
notations here this is I a ; R a I'll write everything in the Laplacian Domain R a + s La + E b or
E a −E b =I a(R a + s La).

So, what happens next? Once this back emf is generated, I will have a certain torque resulting
in a displacement θ and of course, taking care of the moment of inertia of the motor or
possibly even the load and D would be the combined frictional component of the motor and
the load. So, there are two things here; one is the back emf constant and the other one is the
motor torque constant.

150
(Refer Slide Time: 41:03)

Let's first look at how this block diagram is generated; the first thing in the previous slide I
had this thing that E a −E b is the armature current R a + s La. So, if I were to write this as a
block, so the first block would say that E a generates a certain armature current here. So, as
this E a as the input possibly or which results in certain armature current.

So, how will that look like, so it is only trying to achieve you know this bigger block diagram
here. So, I start by saying E a ; there is a negative sign here, So, E a −E b if I would write this in

E a−E b 1
a better way =I a. So, E a −E bgoes via a block to give me I a ; everything is
Ra + s La Ra + s La
again in the Laplacian domain in this direction.

So, what do we know in a motor that the torque is proportional to the field current and the
armature current? The basics of a DC motor teaches this way when here my field is constant
the separately excited machine. So, if you go see this picture here this is always constant E f is
constant. So, the torque as a result of this would have a constant effect, therefore, what is
variable here is just the armature current as a load varies the armature current would vary and
therefore, the torque I can re-write as T is some torque constant times the armature current.

So, I will have K T and the T m here ok, now what does this torque do? well this torque, so we
had seen this relation for rotational motion between torque and the angular velocity and what

151
does it go through? This torque goes through the J; the moment of inertia of the motor and
the load and the friction. So, the torque is related to the T m; goes via this thing via J s2 + Ds .

So, the relation between the torque and the angular velocity goes through this, it generates a θ̇

1
and this θ̇; I can integrate this, an integral component would look like and I get a θ(s).
s
Now, how to close this loop, so I have that the back emf is proportional to the speed of the
motor. So, here we define K b as the back emf constant, so if you recollect the basics of DC
motors, I will have K b θ̇. So, this θ̇(s) via K b completes my entire loop here. So, again let's go
through the process, so with an arrow in this direction. So, my input voltage first generates an
armature current.

Student: (Refer Time: 45:06).

I was about to say that, so there is a question of what is this θ̇(s), he said we will eliminate
differential equations, but then there is a θ̇(s). So, it is just an abuse of notation θ̇(s)actually
means sθ( s). So, I am just abusing the notation here a little bit, but this actually means this
one, so that we'll keep that in mind. So, I just wrote this just to make a little obvious that we
are dealing with the angular velocity and we are used to this differentiation.

So, θ̇ ( s )essentially would mean sθ( s), so again back to this. So, we start with input voltage
which generates a certain armature current and this armature current results in a torque;
torque results in a certain angular velocity and this angular velocity, I can integrate to get the
displacement and I also know that the to close the loop that E b is related to θ̇(s) via this
expression and then I have this entire closed loop system.

152
(Refer Slide Time: 46:11)

So, I can rewrite; how do I get the transfer function for this, so let's for simplicity say I just
ignore L for the moment. So, I will have E a ; a negative feedback loop and so I'll have a

1
single block here; all this would simply be multiplied. So, I have a , I have K T , I have
Ra + s La

1 1
; I have a θ̇ ( s ) then , I have a K b here; this guy goes from here till here and this. So
Js + D s
now, how do I write the overall transfer function as the ratio of the output displacement to s?

So, I'll just recall what properties I learnt about the block diagram earlier as, so here I have

G(s) . So, I go back to my block here, so let me call this G(s); I call this
G(s) H (s)=
1+G ( s ) H (s)

G(s)
H (s). So, my overall transfer block here would look like and then this would be
1+G ( s ) H (s)

1
multiplied by and I can just write down the final expression that this will look like
s
straightforward computation which can be done very easily.

KT
θ( s) Ra ; this is we can just do these calculations from here to here.
= 2
E(s) Js + s ( D+ K T K b ) R a

153
(Refer Slide Time: 48:31)

So, what have we done so far? well we had defined transfer functions and its properties we
had seen how defining a transfer function helps in block diagram representation of complex
systems and also saw some special cases where two systems in cascade do not direct their
Laplace transforms, do not directly multiply or the transfer functions do not directly multiply
in case there is loading effect on the transfer functions.

(Refer Slide Time: 49:00)

So, what we will next look at is, if I have a system which has lots of summing points lots of
you know negative feedbacks; how do I simplify those block diagram of a complex system

154
and then to find the overall transfer function. We'll also then see how we can actually use this
transfer function to see how my systems would behave to different kinds of inputs.

Thank you.

155
 Control Engineering
Dr. Ramkrishna Pasumarthy
Department of Electrical Engineering
Indian Institute of Technology, Madras

Module - 02
Tutorial
Lecture - 09
Solving Problems on
Laplace Transforms and Transfer Functions

During this hour or so, we'll just do some problems related to Laplace transforms, inverse
Laplace transforms and also some simple examples on calculating the transfer functions.

(Refer Slide Time: 00:34)

So, let's start by finding Laplace transform of the signal which looks like this say in the time

2
sin t
domain is +t e sin ( at ) , and during the computation of this we will use lots of
−at
t
properties of Laplace transforms which we learnt earlier. So, this is a combination of two
signals f 1 ( t ) and f 2 ( t )right and using the linearity property, the Laplace transform of f ¿) can
be individually computed as the L(f 1 ( t ) )+L(f 2 ( t ) ).

So, let's start with this thing first. Finding of Laplace transform of sin 2 t . So, do not worry
where the t has disappeared I will tell you magic of how to take care of that t. So, this would

156
 1−cos 2t
be sin 2 t can be equally written as , here I can further use the linearity property this
2

would be L ( 12 )-L ( cos22 t ); now I just recollect the formulas. So, this would be 21s − 12 ( s s+4 )
2

2
sin t
, right my original signal is of the form .
t

(Refer Slide Time: 03:17)

2
sin t
I need to find the Laplace transform of . So, far I have found what is the Laplace
t
transform of sin 2 t ? Is there any property which helps me in finding the Laplace transform
when I just divide a signal by t? So, just recollect the frequency integration property.

So, what does the frequency integration property tell me is that if I have a signal in time

∞
1
t[ ]
domain x (t ), L x ( t ) =∫ X ( u ) du.
s

157
(Refer Slide Time: 04:12)

2
sin t
Therefore, the L ( ) t
given that the L ( sin 2 t )is this, I just used the frequency integration

∞
1
formula this would become ∫ 1 − u du.
(
2 s u u2 +4 )
1
So, I just. So, what would this be? ¿(u2 +4)] and everything going from s to ∞? So, this
2

2 2
1 s +4 sin t
simplified would look something like this
4
ln
( )
s
2 . So, this is the L
t( ).

So next, we need to compute the Laplace transform of the second signal this guy

L [t e sin at ]. Now I use equivalently the frequency differentiation formula. Now I just use
−at

the frequency differentiation formula here, what does the formula say that

−d
L[t x ( t ) ]= X ( s ) right here I have t multiplied by this signal. So, this would be
ds

−d
ds
[ L ( e−at sin at ) ]. Now, next I use the frequency shifting formula right, but the time by e−at .

158
 −d
So, this would be [ L ( sin at ) ]. So, this would be now I know the Laplace transform of
ds

d a
sin at ,− [ L ( sin at ) ]. So, L ( sin at ) = 2 2.
ds ( s+ a ) +a

(Refer Slide Time: 08:21)

−d a 2 a ( s+ a )
And that [
ds (s+a) +a
2 2 would be][( s+a)2 +a2 ]
2 right and therefore, the Laplace transform

2
sin t −at
of our signal f (t ) is a combination of +e sin ( at )would be is this one, the first guy is
t

2
1 s +4
this one over here, ln
4 s
2 ( )
plus. So, this will be s+a here if I differentiate the

2 a ( s+ a )
denominator I have 2 a ( s+a ) , thus 2 . So, this illustrates that you are using the
[( s+a)2 +a2 ]
frequency integration, the frequency differentiation, frequency shifting and of course, the
linearity property.

So, next if I have a signal let’s see f, f (t ) looks something like this right it is 1, for t between
0 and 2 its minus 1 for t between 0 and 4

10 ≤t <2
f ( t )= {−12 ≤t < 4 }

159
Together with the property that f (t +4)=f (t ), it is like a periodic signal with the period of 4.

(Refer Slide Time: 10:10)

So, how can I plot this signal? So, this is f has a value of 1 fromt going from 0 to 2. So, it
looks something like this right and at 2 it takes a value of -1. So, this until it reaches 4 and
keeps on doing this. So, this I have 6 here and 8 and so on. Now, this is a periodic function.

P
1
So, what is the Laplace transform for a periodic function L [ f (t) ]= ∫ e−st f ( t ) dt right
1−e−Ps 0
this guy P is the time period, which in our case is 4.

2 4
So, this would be simply now translates to
1
1−e
−4 s [∫0
e
−st

2
]
dt +∫ −e−st dt . Then I just keep all

the computations this is a very straight forward integration which you learn very early in your

−2s
1−e
high school. So, this would simply be −2 s , this quite straight forward right once you
s ( 1+e )
just plug in this formula these are easy to compute.

160
(Refer Slide Time: 13:20)

So, next we look at solving non-homogeneous differential equations or also called first
differential equations. So, just looks like this, say ẏor say y ' + y=sin t , with the initial
condition y (0)=1. Now, I take the Laplace transform on both sides. So, I have

1
[ SY ( s ) −Y ( 0 ) +Y (s)] in the right hand side, I have sin t the Laplace transform of sin t is 2 .
S +1
So, I substitute ( S+1 ) Y ( s ) , y (0)=1.

1
So, is 2
+1 now just take this guy over here, answer should give me
S +1

2
s +2 right just re-arranging terms. So, I can rewrite this as
Y ( s )=
( s+1 ) ( s 2 +1 )

3 1 1 1 1 s
+ 2 − 2 and this is just by applying the method of partial fractions of solving
2 s+1 2 s +1 2 s +1
this kind of equations. Now y (t) would simply be the L−1 Y ( s). So, the first

1 1 s
L−1 [ ]
s+1
=e−t, while this inverse is a sinusoidal signal sin t and the last one is 2
2 s +1
is cos t

as simple as. So, just this step here is a crucial one of identifying these numbers here and to
write this in terms of fractions.

161
(Refer Slide Time: 16:21)

So next we will look at deriving transfer functions from given simple circuit models. So, let's
say I have a circuit with the input voltage V i, R here. So, I will call this as R 1 the capacitor C 1
, L1, this would be C 2 and this is R 2. So, I need to find what is; So, all initial conditions are 0
this is the assumptions I will make right, the first is to find the transfer function. So, this is

V C ( s)
capacitor C 1 let me call this voltage as V C , find the transfer function 1
.
1
V i (s)

So, let's step by step follow what we had learned earlier, how to determine my transfer
function? So, first is well, write down the equations right; or the differential equations which
model my system. So, let me call this current as i1 as I would do it in any circuit analysis and
i 2.

162
(Refer Slide Time: 18:27)

So, the first equation for loop 1 would look like this right that

1 d i1 .
V 1=i 1 R 1 +
c
∫ ( i1−i2 ) dt +L1
dt

Student: (Refer Time: 19:06).

So, there is also another way of doing this, now since we are used to writing things in linear
equations I'll just write this entire circuit in Laplace representation right. So, I will just try to
draw this here, this a capacitor here a V i, this guy the inductance here capacitance R 2, this is
C 2 this is my C 1 , L1and R 1. So, in the Laplacian domain I could write this as sL1 capacitor I

1 1
could rewrite as , this guy would equivalently become right and this would be I 1 ( s )
sC 1 sC 2
the current in this loop, the current in this loop would be I 2 ( s ) . So, let's go back here and if I
were to write these equations in the Laplace domain.

1
So, that will be I 1 ( s ) R 1. So, it will be same; this would be ( I (s)−I 2 (s) ) +s L1 I 1 (s) right.
C1 S 1
So, I can even write it directly from here right. So, I have V 1 (s)=I 1 (s) R1, I can write I don't

1
even need to bother about writing the integral equation now, I have ;I can just look upon
s C1

163
 1
this now as a linear element, the current flowing through this is I 1 ( s ) −I 2 ( s ) plus the
s C1
current here I 1 ( s ) s L1right. So, I can just write down the equations without even going
through this step, if I can transform from this circuit to something like this right and for the
loop 2

So, loop 2 I have V C ( s ) equal to. So, I just use this circuit now this is my V C ( s ) here V C ( s )
1 1 1

1
then we do the convention like this, V C ( s ) is current through here is I2 ( s) + R I ( s ).
1
s C2 2 2
Now I have equation 1, I have equation 2. So, now, what do I need to find? I need to find out

V C ( s)
1
. So, just take this one right, the voltage across this capacitor V C ( s ) is also, well what
V 1 (s) 1

is this? right this voltage if I look at this equation here this simply this one right here this one

1
or this one; this is [ I ( s ) −I 2 ( s )]we call this as equation number 3. So, what do I need to
C1 s 1
do right? So, I need to eliminate I have three equations this, this, and this and I need to
eliminate I 1 ( s ) ∧I 2 ( s )right. So, from 2 what I have that I 2 ( s ) can be written as sorry I can just

1
write this equivalently as I 2 ( s ) +R .
s C2 2

s C2
So, I 2 ( s ) would be V C ( s ) divided by this entire thing. So, this would be times this
1
1+ s R 2 C 2

one, V C ( s ) . So, all this will be s. So, now, I eliminated this from here. So, I know I 2 ( s ) in
1

terms of V C ( s ) . So, we have eliminated I 2 ( s ) from equation 2 and from equation 3 now what
1

we have is I 1 ( s ) =V C C 1 s, this goes here plus I 2 ( s ) . So, this is V C C 1 s plus what is I 2 ( s ) ? this
1 1

s C2
entire thing V ( s ). So, I will call this 4. So, in my equation 4 I eliminated I 1 ( s ) . So,
1+ s R 2 C 2 C 1

I 1 ( s ) could now be written in terms of V C ( s ) , I 2 ( s ) can also can be written in terms of V C ( s )

1 1

from this expression here and. So, I can mix or this guys here to get my final expression in
the following form.

164
(Refer Slide Time: 25:44)

V C ( s)
1
is, I'll skip the computations but there should be straight forward to do and
V i (s)

1+ S R 2 C 2
2
.Yes all this is in the denominator this
( s R1 C2 + s LC 2 )+ ( 1+ S C 1 R 1 + S2 LC 1 ) ( 1+S R 2 C 2 )
multiplied here right. So, what have I done right again. So, first is write down the individual
loop equations and try to eliminate each of the variables, to get an expression which will give

VC (s)
us the ratio of 1
. So, similarly I could also compute the transfer function from V 1 ( s )or
V 1 (s)

V i (s) to V C ( s ) right; nothing changes so all this.

2

So, instead of this is my V C ( s ) the voltage across my capacitor C 2 , I can equivalently then
2

1
find my transfer function as this one right. So, equivalently what is V C ( s ) ? V C ( s ) is .
2 2
s C 2 I 2 (s)

V C ( s)
So, you could do the remaining thing as an exercise to find out what is 2
given that
V i (s)

V C ( s ) can be written as an expression like this right it's just the same steps that we will
2

follow.

165
(Refer Slide Time: 27:50)

So the last problem for today; so again to in the last problem for today we are required to find

X (s)
the transfer function . So, what does this electromechanical system do right. So, I have a
E (s)
E which generates a current here, this current results in a back emf over here and there also
force which this system exerts on this one right. So, what are the constitutive relations here

dx
right? So, the back emf e b=K 1 and the force is proportional to the current K 2 i right the
dt
force which is experienced by this mechanical system over here. So, we are required to find
out what is the transfer function starting from this input e resulting in an output here. So,
these are two subsystems right. So, one is this electrical component, and one is the
mechanical component. So, let's analyse each of them individually.

So, the mechanical component for M here which possibly this experiences the force in this
direction like this. So, the K here there is a K here and there is a D here right. So, well I can
directly write down the equations for this because I already know this right; so
M ẍ+ D ẋ+2 Kxfor the spring. So, you have two spring that will be 2 K x, D is the is the
damping and K are the spring constants, this guy would be equal to the force and what is the
force? the coupling between this mechanical system and these electrical systems comes this
via current i; so this F i. So, I can write down this as ( s2 M +sD +2 K ) X ( s ) is sorry I just made a
mistake here this is K 2 i so this would be K2I(s), that's with this part here over here.

166
(Refer Slide Time: 32:09)

Now, the electrical part right I can write down this as equivalently in the s domain directly.

1
So, I have a , a R here, sL and I have a E and I have some I just draw block here called the
Cs
e bor the back emf and this is the current I which I'm interested it, and let me call this current

1
over here as I 1 right. So, the first equation would sayE=R I 1 ( s ) + ( I −I ). And the second
Cs 1

1 dx
expression would be ( I −I 1 ) + sLI +e b=0. So, what is this e b? e b is K 1 . So, I could
Cs dt
simply write this as K 1 sX ( s )is 0. So, I have this equation relating X and I, I have these 2
equations relating the currents and the input voltage.

So, I can just make use of these three equations to eliminate I 1 and I and ultimately find the

X (s)
transfer function right. And I will leave the steps for you, but they should be very
E (s)
straightforward right. So, take this equation you can eliminate one of the current variables
from here, one of the other current variables from here and here you can directly eliminate
possibly write I directly as something on X 2 substitute that I over here. Here you could
eliminate I 1 you already know Iin terms of X , do all the computation and thus get the
equivalent expression I will just use the computations again for you as an exercise.

167
So, these were some simple examples in you know of course, of solving Laplace transforms,
solving ordinary differential equations, and starting from system configuration how to arrive
at the equivalent transfer function again via all you know everything in the Laplacian domain.
So, for future purposes what would be easier is to equivalently draw this circuits in this form
right then you just have to treat them just as here impedances and then right the equations
directly as linear equations, you might have done this in some of your circuits course circuits
or networks course. So, just a little re-collect of that would help you write on the equations in
a much compact and much easier way.

Thank you.

168
 Control Engineering
Dr. Ramkrishna Pasumarthy
Department of Electrical Engineering
Indian Institute of Technology, Madras

Module - 03
Lecture - 01
Block Diagram Reduction Signal Flow Graphs

Hi guys. Today we will learn simplification techniques for block diagram. So, if you
have a system with a lot of interconnected components connected in series or parallel or
feedback interconnection and so on, are there any methods where we could simplify and
get the overall transfer function? We will also see, what is also known as signal flow
graph method for determining the transfer functions, these are very small tricks which
you need to learn and I will quickly run you through those tricks before we do some
examples.

(Refer Slide Time: 00:55)

So, what is a block diagram? So, it's a shorthand representation of the system where you
have each subsystem or each component represented by its transfer function. And the
interconnection is captured as the flow of signals from one system to the other. And well
what were the components? How did we draw the block diagrams? So, the blocks
represent the components, equivalently represented by a transfer function, arrows
indicate the direction of which the signal flow, summing points to show-do signals add

169
up or do they know it is a 1 plus and a minus and so on, and take off points indicate how
we branch of signals.

(Refer Slide Time: 01:41)

So, what is this reduction business then? So, we are aiming to reduce or simplify the
block diagram in order to get the overall transfer functions. We do not need to really
write down equations all the time and then eliminate each variable one by one, that was
what we did in yesterday’s examples right then we have the block diagram
representation, but then if it is a complex form, we just have to simplify it you know we
need to get the overall transfer function.

So this simplification is done by some very basic rules of block diagram algebra. And
these are very simple algebraic methods when where we just reduce our variables or
reduce the blocks and so on. Just very basic things like addition and subtraction and so
on.

(Refer Slide Time: 02:30)

170
So, we learned this thing yesterday. So, if I have two systems in cascade. So, if the two
systems are in cascade then I think the overall transfer function is just G 1 G 2, and the
output Y is simply U G 1 G 2 or the transfer function is just this guy G 1 G 2. This we learned
yesterday and it is just kind of a very straight forward proof. So, this is one of the basic
building blocks which we will use.

(Refer Slide Time: 02:58)

So, similarly if I have things in parallel then I can just start with this know, this diagram
which looks like this is a block diagram and I know that you know these two signals this
guy here is U G 1, this guy here is U G 2. And then they just add up according to this thing.
So, this is a +¿ here and a −¿ here. So, it will be U G 1and sorry this should be a minus

171
here right if this a −¿here. So, U G 1is what comes from the top and then U G 2 comes
from the bottom with the −¿ sign it will be U G 2.

So, it is very straight forward to do and then the equivalent representation can be like this
U G 1. So, if this is a −¿ here. So, it will just be G 1−G 2, and similarly here the G 1−G 2; so
apologies for this little thing here. So, if it is a +¿everything will be +¿and then things
will be fine. So, this the guy on the left looks a little you know things with a bigger
component, and the guy on the right is something which is simplification of the block
diagram. And we will again use these very basic building blocks.

(Refer Slide Time: 04:11)

So, we also studied yesterday about the feedback interconnection right. So yesterday we
started with feedback interconnection this is my U my output signal Y if I have a negative

G
feedback. So, this is a minus here then my overall transfer function would be
1+GH
right. And then Y would be U times this entire thing, this is with the negative feedback.
So, I can still have plus sign here also right, a plus here; this would be called as a positive

G
feedback. In which case, my overall transfer function of Y would be U ( 1−GH ) right.
So, the lazy way to remember this is a negative feedback here you will end up with a

172
plus and if it is a positive feedback you will end up with a minus here. And equivalently I
can write these entire things as just one component here as just one single transfer
function.

(Refer Slide Time: 05:28)

So, if it is negative feedback I have + and if it is a positive I have a −¿right. So, now,
some more things which we could do with these block diagrams. So, say I have a
summing point here, you can call this say 1, and then I have a summing point 2 and if I
want to interchange them. So, in a way that I want to put a 1 here, and I want to put a 2
here, what should not change is the output signal right, if I just change them arbitrarily
then what is coming out here may change, and I don't want to do that; I want to preserve
the entire input-output structure of the system. So, what should come out here would still
be A ± B ±Cdepending on what is the sign of summing block here.

So, I start with A still goes in, but now it will go to the summing block 2 and then what is
the input to the summing block 2 this C over here and then well the resulting signal here
would be A ±C. That will go all the way here, and then I have an input here B, this is C,
this is A and I can just interchange them this is my summing block 2. So, this is my
summing block 1 right; this is very kind of straight forward.

(Refer Slide Time: 06:45)

173
This is a little, not tricky, but again you know it is kind of easy to understand and we will
know why we are doing this kind of tricks is, we will want to get things either in the
feedback form like here or things in which are interconnected in parallel, or which are
interconnected in series and these three formulas I know very well and these are very you
know easy to derive and even easy to handle.

So, all these tricks I will use to just simplify into one of these three basic configurations
right. So, this is my summing point and then I have a block. So, what I aim to do here is I
aim to move the summing point somewhere here, such that this guy moves backwards
somewhere here. So, what should happen here is that the signal should still be the same.
So, what does not change is the way U 1 is given to me; it is given to me from outside.
So, I just cannot change anything here and then U 2 comes to me from here. And what is
measured at the output at this take of point if I call this the output, that would still be the
same; this will still be G ( U 1 ± U 2 ). And I want to move the summing point ahead which
means the summing point should come somewhere here and then there should be a G
here.

So, let's try to do this in a just blindly replace it and see what happens. So, this guy goes
here U 2 goes into the summing point and I have this I have a +¿ and ±. So, the signal
here is U 1 G . The signal coming from here is U 2. what is the overall signal now? that is
U 1 G+U 2. This is not what I want right, what I want is a signal something like + or −¿;
what I want is G U 1 +GU 2. So, what do I do? So, what I see here is at U 2should be

174
multiplied by a block with G. And if this is G this guy is the G U 2 GU 1 and then I
preserve my original signal that is what is exactly happening here.

So, U 1goes to G, I have a U 1 G here; U 2goes via Gand I have your signalG ( U 1 ± U 2 ),
which means I have a signal here the inputs are the same signals and the outputs are also
the same.

(Refer Slide Time: 09:23)

Now, if I want to do the reverse, I want to move this summing point somewhere here.
again we just follow the same structure right. U 1s given to me as it is U 2 comes from
outside. So, this is also the same and I want to even preserve what is the output
U 1 G ±U 2, and then I want to move G to the right. So, let us say I have a G here and
then I move the summing point here, and then I connect this with a + sign and say just I
just directly connect this U 2 here without doing any change and this way.

So, what am I left with? So, from here I have. So, my signal with this connection would
like the following; this is U 1. So, here I will have U 1 ±U 2and what comes out here is
G ( U 1 ± U 2 ), but you said what I want? no I want something like this I want U 1 G ±U 2. So,
I have here U 2 is multiplied by G, I do not want that I want U 2 just to be by itself,

1
therefore I add a block here which is just now the signal looks little better right. So,
G

175
 U2
here this U 1 ± this will go via Gto give me this guy, that's exactly what's happening
G
here.

(Refer Slide Time: 11:06)

Now, this is kind of straight forward right. So, if I have this take-off point I just want to
put it here. So, again I have U 1 going in and what should come out is also U 1but now I
want this guy to go after the G right. So, this is U 1 G , where this is still U 1 G , but what I
want to measure is U 1. So, I just divide it by G, so this kind of straight forward to do this
one.
(Refer Slide Time: 11:40)

176
The reverse I don't think I need to do the proof. So, if I want to move this guy over here,
again so, the input is U 1the output here is U 1 G , and I still want to you know just take
off something the signal coming from this take-off point still to be U 1 G ; if I get this guy
here I just have to multiply it by G.

So, these are some of the very basic tricks which we will use and we will do some
examples later in the lecture.

(Refer Slide Time: 12:06)

So the other, so sometimes it may be very cumbersome to do this kind of exercise right
that you know then it might get a little messy and you might forget which is to be moved
what, you know and so on. So, is there any simpler representation and the answer is
provided by the signal flow graph. And this is an alternate way or one to represent the
block diagram or the physical system which I write the set of physical equations, I
interconnect them and then have an equivalent block diagram representation. It's a
simpler representation of the relationship between the signals between the system
variables. And I don't really need to do any reduction; I just kind of memorize a small
formula. We'll not derive that formula, but we just try to use that formula, formula
developed by Mason and therefore, it is called the Mason ' s gain formula. So, how does
the signal flow graph look like?

(Refer Slide Time: 13:02)

177
So, I start with a block diagram right. I could also start directly from the systems
equations, I can even skip the block diagram part of it I can just start with system
equations which are essentially linear equations, if I do it in a Laplacian domain right the
equations in the Laplacian domain and of course, as said earlier all our analysis for the
Laplace base things was based on linear system this, is also for linear systems and we'll
also do a bit of this during the state space analysis which would be towards the end of the
course, but for the move time we'll just concentrate on the first three points here.

(Refer Slide Time: 13:36)

So, what are the components? So, I have a node which represents the variable which is
equal to the sum of all the incoming things at the node. As so, this you can also relate to
circuits right, if I have a node and then if I want to compute the currents, and you just

178
take the sum of all the elements which you know enter into that and then outgoing
signals. So, in addition to that, well the outgoing signals wouldn’t affect the variable at
the node. For example, if signal here if I have something coming here right. So, this
would be say a signal a, a signal b. Well, what is going out will not really affect right?
So, this will be kind of a+b thisc will be a+b+c and so on. So, what is the computation
here? will just depend on what is the incoming of the signal; the outgoing signals do not
affect the variable at the node.

This is, we will also see this from some kind of this KCL kind of stuff right. A branch,
well the signal travels along a branch from one node to another. So, if this is a node and
well this is signal traveling here, this is node just say n1. So, this is noden2and this guy
will be called a branch. And each branch could possibly have a gain right if I talk in
terms of circuit elements, there will be some elements sitting over here right some gain
let's call this gain k right. And then if a signal is passing from here to here the signal
passing through a branch gets multiplied by its gain this is not very surprising to see.

(Refer Slide Time: 15:04)

So, if I come back to my original feedback configuration. So, I have an input signal, I
have an output signal here Y this transfer function G , H and I know how to derive this
transfer function by now. So, what could be the nodes here? well, I can say something
starts from here this is a node which just has one branch going out of it here. So, this we
denote it as the error signal. So, Ehas U coming in from here and something coming
from here right here in this case it is YH, depending if the sign is a + or a −¿. So, I add
another node. So, I can go from here till here I have a H , and this + or −¿ I just represent

179
it with a −¿ sign if it is a −¿1. I could also keep this process and this call this entire
branch directly has −H nothing changes, I just multiply these two. They also like if you
look at a transfer function H multiplied by this again −1, it will just be −H and from E
to this node Y , I have a gain G.

And then of course, these are very straight forward computations right. So, what is
happening at B=YH which is what we said here, that the signal passing through a branch
gets multiplied by it's gain, that is exactly what is happening here. And then So, these are
the same things right; B=HY , this Y =¿and so on and Eis so E has a combination of
signals one coming from U with a gain of one this guy Bcoming with a gain of −¿or you
can also talk about this directly depending on what kind of analysis you do.

(Refer Slide Time: 16:54)

Some basic terminology what is an input node, the input node is the one which has only
outgoing branches. For example, here the U is the input node it just has outgoing
branches; Now output node, So, the output node is the one which has only incoming
branches and if this condition is not met then we can just draw an additional branch with
unit gain. So, what does this mean in the context of this standard feedback system. Let’s
take so, here I really do not satisfy this definition directly, the node with only incoming
branches and when this condition is not met what do I do is I just take this guy and I put
a gain of 1 put an arrow here this will still be Y . Now this will satisfy the definition now.
Because it just has branches which are just incoming into it and I think no other branch
here satisfies that criteria.

180
You could also maybe pull out something from here right this is nobody stops you from
doing this also this also you could call as the output node. So, the path the definition is
not surprising it's just the traversal of the connected branches in the direction of the
branch such that, no node is traversed more than once. Here a path would simply be if I
want to go from U tillY just go from U to E, from E to Y and so on. This is not allowed I
cannot go from U to E, E to Y again come back via this path and go. So, each node
should be traversed only once that's what we will be careful about.

(Refer Slide Time: 18:38)

A forward path well is a path from the input node to the output node and then a loop is a
path which originates and terminates at the same node. So, this is a forward path, this is a
loop starts from E and ends atE this, this outside guy is also a loop, and then non-
touching loops could be the loops which have no common nodes. So, in the previous
example, there are no non-touching loops and then the path gain would be just a product
of branch gain encountered in traversing a path or even a loop.

So, for example, if I traverse from U till Y , the gain is one times Gtimes one, it will just
be G, but if I traverse around this path or along this path the gain would be G × H ×−1, it
will be −GHor this one G ×−1also −H .

(Refer Slide Time: 19:40)

181
So, now we are ready to learn the Mason's gain formula, what is the purpose? The
purpose is again to find the overall gain of the signal flow graph. And why do we find
the gain and this gain we will see is just a transfer function of the system right. So, what
are the things which we need to know? So, the formula looks kind of simple it is the
N
pk ∆ k . So, what do these mean,
TF=∑ TF is of course the transfer function, p k is the
k=1 ∆
gain of the k th forward path. There could be more than one forward path, we will shortly
see with the help of an example.

(Refer Slide Time: 20:29)

182
N is the number of forward paths, ∆ is the determinant of the graph, and we will again
define what this ∆ means? we will just go to the definition of the ∆. So, delta is by
definition which is also called the determinant of the graph is just ,
∆=1−∑ of all the individualloop gain.So, let's keep that example in mind in that
example. So, instead of going I'll just re-draw it again over here. I'll just draw the
Mason's circuit; I have a G, this way I have a −H , I am talking of negative feedback, I
have U and I have Y and this goes in this direction right.

So, the determinant is 1−∑ of all the individual loop gain. How many gains here? I have
∆=1−(−GH ), there is just one loop right, this is ∆=1+GH . It does not stop here, it
goes further and says ∑ of gain products of all combinations of two non−touching loop ,
well there are no 2 loops here this guy will go away
minus the ∑ of products of all combination of 3non−touching loops, of course, there are
no second guy is absent the third guy will also be absent and then this formula just goes
by right.

So, this is how we determine the ∆, now ∆ k, ∆ k is the value of ∆ for the part of the graph
not-touching the k thforward path. What does that mean? In this case there are number of
forward paths is 1, just this is the forward path. Now if I remove this right. So, this value
of ∆; ∆will again be computed this way. So, just give me a graph I can just look at
1 −¿minus loops and the non-touching loops, 3 non-touching loops and so on. And these
ones if I remove this node, this node, this node nothing remains right, therefore, the ∆ k
here is just 1.

N
pk ∆k
Now, let us see the transfer function by the formula; ∑ and N=1. So, I will just
k=1 ∆

p1 ∆ 1
have . what is this p? p is the gain of the k th forward path. There is only one
∆

G ∆1
forward path what is it is gain from here till here that is , this is the overall delta;
∆
∆=1+GH and this is exactly what we derived earlier for a negative transfer function. If
it is a + here, this will just become a−¿. Now let us see a little one more example which
will help us understand these concepts of loops and forward paths and so on.

(Refer Slide Time: 23:20)

183
So, from U till Y how many forward paths I have? well I can goes this here,here,.. and
here. The forward path, let me call this as p1, p1 would be G 1 or the gain of the forward
path would be 1× G1 × G2 × G3 × G 4 ×1, which is G 1 G 2 G3 G 4. Now is there any other
forward path where I can reach from U till Y , well of course, I can go here I can jump all
the way where G 5 come to E and go to Y . So, that is the second forward path, p2 with the
gain of G 5. How many individual loops are there? well this one, So, loop one which has a
gain L1 is −H 1this is also a loop, I start from C, I go to Dand I come back again to C
here I just start from B and again come back toB. So, like a self-loop; here L2this is G 3
times −H 2 that is −G 3 H 2.

And look at the third loop there is also a loop here. So, this is D to E and again back to
D ; the gain with G 4 times −H 3that is −G 4 H 3 ; are there any other loops no right. So, I
cannot do this way, just this three loops and then here are there any non-touching loops.
So, this loop is touching this loop this is not a non-touching loop. So, this is one loop
which does not touch this loop for example, again the same loop does not even touch this
loop. So, what are the non-touching loops that is L 1 if I call this loop as L 1, I call this
loop as L2, I call this loop as L3. So, L1 and L2 are non-touching and their gains would be
−H 1times this guy that will be H 1 H 2 G 3,

Similarly, L1and L 3 are not touching each other, is the total gain is H 1 G 4 H 3 . Then the
minus gets canceled out. So, we have two forward paths, three loops and two non-
touching loops.

184
(Refer Slide Time: 26:41)

Now I want to write down the entire transfer function for this. So the formula read, let's
N 2
pk ∆k p ∆
recollect the formula, this is ∑ . So, here my N=2. So, TF=∑ k k . Now the
k=1 ∆ k=1 ∆
first thing is to find ∆. So, let us see what the definition says the definition of
∆=¿1 −¿ sum of all individual loop gains, let's do this first, this is 1−¿sum of all
individual loop gains; what were the loop gains? H 1 this and this, 1−¿ minus the
individual loop gains would be −H 1 −G3 H 2−G 4 H 3.

So, what does the next step say + sum of gain products of all combinations of two non-
touching loops, gain products if I have two loops L 1 and L2 which are non-touching? So,
I will just take the product gain of each of this L1 and L2. So, this is this one,
H 1 H 2 G 3 + H 1 H 3 G 4. So, I have plus here, so I start with this guy H 1 H 2 G 3 plus this times
this, that is H 1 H 3 G 4 and so what does it say plus the sum of gain products of
combinations of 3 non-touching loops, well, there are no 3 non-touching loops here right;
there are no 3 loops here. So, there are 3 loops here, but this loop is touching this loop.
So, there are no 3 non-touching loops. So, this is my ∆, i.e.,
∆=1−(−H 1 −G 3 H 2−G 4 H 3 ) +(H 1 H 2 G 3 + H 1 H 3 G 4 )

Now, next I want to find out what is p 1 and it is corresponding ∆ 1 . p 1 as we found earlier
is just G 1 G 2 G3 G 4. What is the corresponding ∆ 1 . ∆ 1 again by definition is the value of ∆
for the part of the graph not touching the K thforward path. So, how does the graph look
like? So, this is the first forward path and all the nodes go away right this is nothing left

185
right. So, ∆ 1 would simply be 1. Now look at p2, p2again they have this path, this here
and this one; that is G 5. Now ∆ 2; So, ∆ 2 looks how does it look like? So, this node is
gone this node is gone, and this node is gone this node is gone. So, I am left with B to C,
D, and I have this one with −H 2 G 3 and self-loop here −H 1.

So, this how my graph looks like without the nodes U, A, E, and Y because they touch
the second path. Now I again apply the definition of ∆ for this guy. So, that will be 1 −¿
again let's recollect the definition Sum of individual loop gains, how many loops are
there this is one loop, and this is one loop 1−H 1 −H 2 G 3 . These are all ∆ 2and then plus I
have to look at non-touching loops. So, this loop is not touching this loop. So, I have
around one pair of non-touching loops plus overall gain that is H 1 H 2 G 3,i.e.,
∆ 2=1−(−H 1 −H 2 G 3 ) +(H 1 H 2 G 3 )

p1 .1+ p2 ∆ 2
So now, I am pretty much ready with my formula, the TF= . This guy is the
∆
overall ∆ right, this is p1, this is my p2, and this is the overall ∆ 2. I just plug all these
guys here, to get my transfer function as simple as that. Just that you have to be alert in
finding are there any non-touching loops and so on.

(Refer Slide Time: 31:53)

So, that like ends the lecture for today, where we have just learned block diagram
reduction techniques, construction of the signal flow graph and the Mason’s gain formula
to derive the transfer function. What do we next focus on? before we do some problems,
or after we just finish some problems in the block diagrams in signal flow graph is to

186
look at time-domain analysis of the system. I will not go to details of what will be there,
but that will be explained in the next class. Thank you.

187
 Control Engineering
Dr. Ramkrishna Pasumarthy
Department of Electrical Engineering
Indian Institute of Technology, Madras

Module - 03
Lecture –02
Tutorial - 01
Solving Problems on Block Diagram Reduction and Signal Flow Graphs

Hello everyone. In this tutorial session, we will try to solve some problems in relation to
block diagrams and signal flow graphs that we have studied in the previous classes. So,
to begin with, we will try to draw a simple control system with the plant and a feedback
loop.

(Refer Slide Time: 00:39)

So, this is how the block diagram looks like. So, this is the input U which goes with the
positive sign into and this is the control system with transfer function C and then we are
adding a disturbance signal along with a control signal, both we take as positive sign.
And then we have the plant with the transfer function G and we also have a feedback
loop with transfer function H and this is the output and this node that for the sake of
simplicity we are just ignoring the Laplace thing, but everything is in Laplace domain.
So, these are in fact, C (s), G(s) and H (s) which we are representing as C, H and G for
the sake of simplicity. So, what we will try to do here is; we will try to find the transfer
function. So, here transfer function there are actually 2 input signals, that we have to

188
notice, one is the normal input signal that you know, and there is another disturbance
signal which you do not know. So, we can find the transfer function both with respect to

Y (s)
both these inputs. So, that is we can find out what is and then we can also find out
U (s)

Y (s) .
D(s)
So, not only finding the transfer function what we will do is we will try to show that Y (s)
is actually a response due to the input and the disturbance, when they act individually
while the other is completely 0. So, while doing this what we do is while finding out this
transfer function, we will make D=0. And while finding out this transfer function we
will make U =0. And then finally, we will try to find out the overall transfer function
using superposition.
So as these are parts, I will call them Y 1 and Y 2. So, this is Y 1 is the response when U

Y1
alone is acting and Y 2 is the response when D alone is acting now. So, to find out we
U
make this 0 and so our block diagram changes.
Y1
So, this part goes off completely and sorry we make D=0. So, this part goes off
U
completely and what I will be remaining with is something like this. So, we are taking
negative feedback. So, this is Y 1. Now, this is a very simple feedback loop and we know

Y1
the transfer function of it directly. So, is nothing but the forward loop transfer
U

CG
function the feedback transfer function. Now what we will do is for the sake of
1+CGH

CG
simplicity we will even take H=1. Then our transfer function becomes . So, this
1+CG
is 1. Now we will find the other transfer function by making this 0, when we make that 0
we will get a diagram like this.
(Refer Slide Time: 05:02)

189
I’ll put this as C because it is a negative feedback, I will take it as −H . So, C and −H go
into the same loop because this completely vanishes and now this is Y 2. So, again this is

Y2 G
a feedback loop and you can take the transfer function to be = right. Now as
D 1+GCH

G
we took H to be 1, we can just write it as ; this is my second transfer function.
1+GC

Now, my overall response Y can be actually written as Y 1 +Y 2, by the principle of

UCG DG
superposition. So, if I take Y 1 +Y 2which will be. So, Y 1= , Y 2= ; therefore
1+CG 1+CG

UCG DG UCG+ DG
Y 1 +Y 2= + . So, this is nothing but . So, that is the overall
1+CG 1+CG 1+CG
response of the system. Now, if we even do the same thing normally if you take the
complete system and try to find the overall response that is what you will be getting. So,
this is a kind of verification of the superposition principle using this block diagram.

Now what we will do is we will take up a slightly complex block diagram of a system,
and then try to find its transfer function. And then we will convert the block diagram into
a signal flow graph, and then we will again find the transfer function. So, we will have a
comparison of how difficult it is to find the transfer function using the block diagram
directly and then using the signal flow graph using the Mason gain formula. So, this is
the block diagram that we will take.

(Refer Slide Time: 07:55)

190
So, as you can see this is a pretty much-complicated block diagram. And it has about 1,
2, 3, 4, 5, 6, 7, and 8 blocks. Each block has it is own transfer function, and you can see
there are three feedback loops and one parallel combination as well. Now what we will
do is, we will simplify this block diagram using the rules of block diagram algebra and
so how do we begin with? The first thing you can observe is this is a feedback loop
which we can easily simplify. This is a parallel and this is also another feedback loop,
but there is a small complication here.

So, this point and this point are slightly off positions, for us to easily apply the feedback
formula that we know. So, what we can do is, we can shift this here. And once we shift
that we can easily find out or simplify these blocks and find out their overall transfer
function. So, once we transfer that block this is what we get. So, I only draw this part of
the block diagram and see how it looks like once we shift that.

So, once I shift that point comes here, but for shifting that there is something which we
need to compensate. If you look at here what is the signal that is coming out? Whatever
this is, say if I call this as some the signal at this point as A. So, the signal here will be
A G 3. Now if I shift this from here to here, I will be getting only A ; I will be losing out
on this G 3. So, what I'll do is, I will put it here H 2 G3. So, whatever I lost this gain, by
shifting this point here I will be compensating it by including it in the feedback transfer
function. So, I this transfer function is now H 2 G3. And everything else remains the same.
So, now, we have a simple kind of a diagram, this point here and this thing. So, what we
will do is? we will apply the rules that we already know we will find out the transfer
function of this block, this block and this block.

191
(Refer Slide Time: 13:29)

G4
So, this will be . And this will be rest of the summation G 3 +G 5.
G4 H 1

(Refer Slide Time: 13:51)

G2
And this will be . Once I do this, I can redraw my whole block diagram in a
1+G 2 H 2 G 3
much simpler fashion, which is looking like this.

(Refer Slide Time: 14:15)

192
This G 2 will be H 2 and G 3, because this G 3 has got multiplied here, when we shifted this
point.

So, this is how the simplified block diagram looks like. So, we just have G 1 coming here

G2
and this reduction going into and this parallel block going as G 3 +G 5, and
1+G 2 G 3 H 2

G4
this feedback loop going as . So, now, we have 2 more simplifications to do.
1+G 4 H 1
One is the cascade of all these 4 blocks and then applying the feedback law right.

(Refer Slide Time: 16:09)

193
So, once I apply the cascade thing will be something like this

G2 G4
G1
( 1+G 2 G3 H 2
()G 3 +G5 )
(
1+G 4 H 1 )
. So, this will be in a single block now, I can just

add a unity feedback to this. So, now, I have a very simple feedback loop. And I can
apply the feedback loop formula and get the transfer function. So, the overall transfer
function will be like this. So, I am doing all the simplification and then giving you the
final answer. You can do the simplification yourself and check.

So, this will be the overall transfer function of the system. So, that is what we did was we
started with a very complicated block diagram, and then we use the rules that we already
studied in the previous class, and try to simplify the whole block diagram, and finally
arrived at the transfer function.

(Refer Slide Time: 19:23)

Now, what we will do is? we will draw this block diagram in terms of its signal flow
graph, and derive the same thing and see if both are equal or not. So, this is how the
signal flow graph looks like. So, while converting a block diagram to signal flow graph.
One should be very careful to ensure that some branches are not missed out like this. So,
when you go from G 3to G 5, this is one take-off point followed by a summer, and so you
cannot take it to be one node.

So, every take-off point and every summer should have it is own node, and because there
is no gain over this, we will take a gain of 1. So, similarly when you write negative
feedback, you should ensure that you put the minus sign here, because there is no

194
summer kind of thing in the signal flow graph the negative sign goes into the gain, and
this we have discussed in the class that you should add a branch to ensure that your input
and output are as per the defined; as per their definitions.

So, now this is how the signal flow graph looks like, and now what we will do is we will
apply Mason’s gain formula and try to derive the transfer function.

(Refer Slide Time: 22:56)

So, to begin with, what are the forward paths? One clear forward path is U to Y , and
when we go, we have G 1 G 2 G3 G 4 and other forward paths G 1 G 2 G5 G 4 and Y . And so
there are no other forward paths, this graph has only 2 forward paths. What about the
loops? one loop is this. So −G 4 H 1 , and another loop is G 2 G 3−H 2 . And another loop is
G 1 G 2 G3 G 4 with unity feedback. And you should be careful that there is also another
loop with G 5, that is G 1 G 2 G5 G 4 with negative feedback. So, generally, people miss out
on this. So, we should be careful here. So now, we have the forward paths and the loops
now we will go to the non-touching loops.

So, out of 4 loops if you look at these 2 loops these are not touching each other, they do
not have any common nodes or branches. So, among the non-touching loops, we have
L1 L2 and G 2 G 3 G 4 H 1 H 2. And if we observe there are no other non-touching loops
because this loop and this loop are touching, this and this are touching and even with G 5
these 3 are touching each other. So, there are no other non-touching loops, and if we
even go 3 non-touching loops there is nothing. There are no 3 loops which are not

195
touching. So, this is where we stop. Now we all have all the required components to
substitute in the Mason’s gain formula.

So, we looked at the Mason’s gain formula giving us the transfer function as given by
N
pk ∆ k . So now, we know both p k and before going to we need what is ? So, we
∑ pk ∆k ∆
k=1 ∆
will define what is ∆? ∆ is as per the definition it is 1−the ∑ of individual loops. So, I
will take a + sign because all of them have a −¿ and I can simply add them up. So, these
are all the individual loops plus the products of the non-touching loops. So, we have only
one set of non-touching loop, I will just add it here.

So, this is delta. So, what about delta 1? So, if I take the first forward path there is no
loop that is actually not touching that forward path. So, my ∆ 1 is simply 1. And even if I
take the second forward path, this case is the same. So, my ∆ 2 also equals to 1. Now I can
simply substitute everything into the Mason’s gain formula, and what I will be getting is
the same transfer function. So, you can simply observe thus this delta is nothing but the
denominator which we got here. And if you just add up the forward path it is nothing but
the numerator that we got; because ∆ k is 1 in both the cases. So, this is how you got the
transfer function in both cases.

And the most important point to note here is when you have a signal flow graph; it was
actually very easy to derive the transfer function. In this case, you have to do lots of
computations, lots of simplifications and before you could arrive this at this transfer
function, but here it is very simple, and even if they are more complicated block
diagrams its always observed that signal flow graph does very well in finding out the
transfer function. So, that is the advantage here. And now for the last problem of this
tutorial, we will try to find out a signal flow graph from a network directly. So, we were
told that apart from finding out from the block diagram we can also find out the signal
flow graph from the network equations and that is what we will try to do in the next
problem.

(Refer Slide Time: 29:52)

196
So, I will take a network with some RLC components. So, this is the network that I'm
taking, I have an input source V 1, two resistors R 1 and R 2 and one capacitor C and one
inductorL; our output variable is we defined it as the voltage across the resistor R 2. So,
here I will take the branch currents to be i1 and this branch current to be i2.

So, this will be simply i1 −i2. So, we will be dealing with these variables. So, first what
we will do is we will write down the equations based on Kirchhoff’s law that we know
and so the voltage here also I am defining it as V 1, this will be anyway V 0. what is my
V i (s)? I am directly writing everything in Laplace domain because now we are
comfortable with writing down network equations in s domain directly. So, my V i (s) is
nothing but R 1 I 1 ( s ) +V 1 (s). So, this equals to this plus this. And my V 1 (s), I can write it

1
as ( I (s)−I 2 ( s)) and V 0 (s) can be written as R2 I 2 ( s). So, we have these 3 equations.
sC 1

(Refer Slide Time: 33:08)

197
Now, while drawing the signal flow graph, we have to go from V i to V 0. So, in between
as we said the nodes are actually the system variables. So, we include all these variables
as nodes i1 , i2 . So, the easiest way to do it is always put it in the order in which they
appear in the series circuit i1 , V 1 , i 2∧V 0.

So, first I'll put I 1 , V 1 , I 2 ¸∧V 0. Again everything is in s domain, just using the variables

directly without representing s. So, how do we connect these nodes using a signal flow

graph as we said we are using the equations to find out what will be the signal flow

graph. So, I have certain equations which relate these variables; now I need to find the

right gains which will connect these things. So, firstly what is the connection between V i

and I 1.

So, V i is nothing but R 1 I 1. So, because I will be writing I 1 in terms of V i, I need to

V i−V 1
reconvert this equation first, what is I 1 ( s ) it is nothing but right. So, my I 1 is
R1

1
depending on V i and V 1. So, what is the factor with which it is depending on V i is .
R1

1 −1
So, my gain on this will be and with a , it is depending on V 1. So, I will write it as
R1 R1
feedback, because this is appearing after I 1, I will put it as feedback. So now, what is the

198
 ( I ¿ ¿1−I 2 )
relation between I 1 and V 1. So, we know that V 1 is ¿. So, V 1 is depending on
sC

1 −1 −1
I 1 with a factor and it is depending on I 2 with a factor of . So, I will put .
sC sC sC

Now, what is the relation of I 2 and V 1? So, we already know this relation. What we will
do is we can take another relation as well, which will actually give us a nicer this thing.
We will take V 1 to be Ls I 2 ( s ) +V 0. This is fine right, I can write because this is current i2
and this is an inductor, I can write it as Ls I 2 ( s ) +V 0. So, why did I write it in this fashion?
Because I have to observe that in the nodes I took I 2 is appearing between V 1 and V 0. So,
it will be easy to draw the signal flow graph if I have I 2 in terms of V 1 and V 0 right. So,

1
what is I 2 in terms of V 1 and V 0. . So, I 2 is depending on V 1 with a
Ls(V 1 ( s )−V 0 ( s ) )

1 −1
factor . I can simply connect this and it is depending on V 0 with a factor .
Ls Ls

Now, finally, what is V 0? V 0 is simply R 2 I 2. I can simply write R 2 here. So, my V 0 will
be simply R 2 I 2. Now if you look at this signal flow graph. This will exactly represent all
the equations that we have written here. And so there is nothing that is missing the
complete network is actually coming into this signal flow graph.

So, now, what we will do is we will try to find out the transfer function. Earlier we used
to apply all this thing and do the elimination and find out the transfer function now
because we have a signal flow graph we can apply Mason’s gain formula, and find out
what is the transfer function. So, as you can see there is only one forward path, here. And

R2
its gain is 2 . So, this is my forward path gain. Then what about the loops? I have
s + R 1 LC

−1 −1
3 loops here right. One is this loop, I have another loop which is 2 and the
s R1 C s LC

−R 2
third loop is right.
sL

So, we have 3 loops and one forward path. So, what are the non-touching loops? So,
these 2 are touching, these 2 are touching, but this and this are not touching; So, I can

199
 R2
write L1 L3, which is 2 . So, now, simply we apply the Mason’s gain formula to get
s R 1 LC
the overall transfer function, which is; before the Mason’s gain we also need to find out
what is ∆? I will write down what is ∆ here. 1 minus because everything has a minus I

1 1 R R s
will put a plus 1+ ( + 2
s R 1 C s LC sL )(
+ 2 + 2 2
s R 1 LC )
. So, this is my ∆. Now once we have

and what about ∆ 1 ? ∆ 1 should be 1 because there is no loop which is actually not touching
the forward path. So, when you go from here to here every loop is being touched. So, ∆ 1
is 1.

(Refer Slide Time: 41:58)

p1 ∆ 1
So, what will be the overall transfer function is simply , which is nothing but this is
∆
p1 divided by this whole data. So, that will give you; you can just simplify it and find out
the overall transfer function.

So, instead of doing all the analysis and trying to eliminate everything and finding out

V0
what is , we can find simply draw the signal flow graph and directly get the transfer
Vi
function. So, that is it for the tutorial today.

Thank you.

200
 Control Engineering
Dr. Ramkrishna Pasumarthy
Department of Electrical Engineering
Indian Institute of Technology, Madras

Module - 03
Lecture - 03
Time Response Analysis of Systems

Hello everybody. So, in today’s thing, we will start a new topic called time response
analysis of systems. So, so far what we have learnt is given a system how would you
model that? model that using basic conservation laws from voltage laws to current laws,
laws of force and so on. And from that, we could derive models based on transfer
functions if we have systems which are interconnected transfer functions then we had
techniques of how to get the overall transfer function via block diagram reduction or the
signal flow graph techniques. So, what we will learn today is once I have a model. So,
what could I do with the model or how does the model behave?

(Refer Slide Time: 01:23)

So, the first thing is after I obtain the model, how to know the performance of a control
system for any input signal? And next is how to design a system which meets some
desired response? it could just be some natural response or even to fit a control
requirement.

So, how do we visualize this? So, given a system or given a model. So, let me say when
we are subject to the driver’s license test right. So, the driving inspector doesn't really

201
test us for all possible conditions on the road, which could include high speed on
freeways, it could include heavy traffic on the city lanes, it could involve chaotic traffic
with buffalos running around, kids playing cricket on streets. We don't get tested for all
those things right. So, instead what the driving test does is under standard Indian
conditions. So, we will just have a little pre-defined track which you have to drive to
pass your learner's test. In that track you have several things right, you have to go
straight, you have to take a sharp turn, you have to make a figure of 8, you have to take a
reverse, you have to park and several other things; these are the standard things which
we will encounter while we drive on roads.

In a similar way, if I am being interviewed for software engineer right. So, I go as a

model; my model could be my B-Tech degree with a CGPA of 8.65 with several lists of
courses. Now I want to hire you as a C programmer right. So, I will not give you a real-
time project to work on immediately right or I would also want to test how would you
react or how would your performance be on a certain project or a certain deliverable.
Then I would ask you some basic questions on sorting algorithms, it could be things
related to pointers, linked lists, some functions, interconnection with a database and so
on. And based on all these things I will get an estimate of how your performance could
be while handling a real-time project.

Now, this could differ from person to person right. Either you could be a fast
programmer right, you could write a program or code very fast; you could some other
person could take a longer time, but run write a code which is more efficient. So, all
these I would get to know with some basic standard test signals. Similarly, what will
happen is when we design a control system or we analyze a control system or a model
we will subject it to some standard test signals because we would not be able to emulate
a real-time control input kind of signal. So, these test signals will tell me how my system
will react under changing conditions. The conditions for a system could be well I have
kind of a shock which could be an impulse, I could have a time-varying signal which
could be modeled as a ramp signal; I could have just a sudden disturbance which could
be tested with the step signal and so on.

So, throughout this we will see how my system would behave, the system could be of
different nature right. So, how my system would behave based on certain standard test
signals. And this will broadly constitute the entire time domain analysis and design
specifications.

202
(Refer Slide Time: 04:58)

So, the time domain analysis refers to the analysis of the system performance in time,
that is how my system variables particularly the output will vary with time right. So, how
long does it take for a signal to reach a certain value which I want like the speed of a
car? You say whenever you see this car related programs in this news channels usually
they come on Sunday morning, you say oh this you know Audi A 4 or A 6 they go from
0 to 100 in 5 seconds, somebody goes in 6 seconds, my own little car would take about
30 or 40 seconds right, sometimes it possibly would never reach.

So, how; what describes the evolution of the system variables? So, in standard terms,
there could be 2 kinds of response, right. The natural response and forced response also
classified as the transient response and the steady-state response. So, I will shortly define
what these things mean, but in both cases, the complete response is given by the
combination of both the natural response and the forced response as well as a transient
and the steady-state response.

(Refer Slide Time: 06:13)

203
So, what is the natural response? It is also is referred to as a 0 input response, which
means the system's response to a certain initial conditions with all external forces set to
0. So, if I take a simple RLC circuit, this would be the response of the circuit with initial
conditions that inductor could have some initial current or a capacitor could have certain
voltage, with all independent voltage and current sources set to 0 which means there will
be no external sources.

Say, for example, I have a capacitor and I have an R, a C with some initial charge I
would call q0 , the natural response would be that the charge if I plot with time, or
possibly even the current that entire charge gets dissipated through heat energy in this
resistors. That is the natural response, I can then add further elements like an inductor
with initial charge and so on. The Forced response is the system's response to external
forces again assuming that all initial conditions are 0. So, again if I look at this circuit
and I add a voltage source here, together with the resistance it will have a different kind
of response, we will shortlist what these kinds of systems mean? and how these kinds of
responses would look like?

So, here there is this external voltage source, here there is no external voltage source
right. So that's the little distinction between the natural response and the forced response
it is also called the zero state response.

(Refer Slide Time: 07:54)

204
Similarly, I have two classifications in terms of the transient response. So, all this
analysis when I make this for the statements I would be referring to systems which are
stable right. And I will define formally what is stability a little later, but for the
understanding of this particular module, it is sufficient to know the following definition
that stability means. So, if I am at an equilibrium position take my pendulum is just at
this equilibrium position, if I perturb it slightly or I start from this initial condition it will
eventually come back to this, its natural state or the equilibrium position.

So, that so all systems which are subject to some external perturbation which then retains
their original configuration would be called stable systems. That definition is enough for
now right. You can think of it again as simple as the example of a pendulum right if just
somebody comes and just pushes this, it just goes back to the equilibrium position. Same
like an alarm clock which wakes you up in the morning right, so your natural
equilibrium position is the sleeping position, and if you have your alarm clock waking
you up, that's a little disturbance or a little shock and then you switch off the alarm put it
to snooze, and then go back to sleep which means your stable equilibrium position which
is sleeping position is always stable.

So, we will talk of; know those kinds of systems? So, in this context, the transient
response is the response that goes to 0 or this response vanishes as the time tends to be
large. So, we will shortly see an example with this, it could also be written or it could be
tied to any events that affects the equilibrium. Right, as I said in the case of this
pendulum right, that the response it just vanishes as time goes to a very large value.
Steady-state response well this is the response of the system after all the transients have

205
died down right or as time goes to infinity, again we will see this with the help of an
example. So, the total response of any system would have two components, the transient
response, and the steady-state response. A steady-state response could also be just the
zero thing itself ok?

(Refer Slide Time: 10:18)

Now, as I said earlier that we will or based on this nature of responses, we will subject
our system or the models which we have or even the control system to some standard test
signals right. So, this standard test signals would be a unit impulse, which could be just a
shock, a unit step which could be just a sudden change in the system. For example, if I
have 10 people in my room with my air conditioner set to 24 degrees, 10 more people
add up to the room and just stay there that is a sudden change right, it is like a step
disturbance; my occupancy is going from 10 people all of sudden to 20 people. Or a
ramp or a constant velocity where you know one person is being added every minute or
parabolic right which has like kind of units of or you can visualize it as something
moving with a constant acceleration. And of course, you can have periodic signals like
for example, sinusoidal signals. So, for the movement, we will do analyze: impulse, step,
ramp, and parabolic signals and postpone the sinusoidal analysis when we will do the
frequency domain analysis. Again these inputs are chosen because they capture many of
the possible variations that can happen in a real-time scenario right. As the driving
example or even the C programming example right; and then your design specifications
could also be based on the response of my signal to a certain input.

(Refer Slide Time: 11:51)

206
So, just to recall how these signals look like, well a unit impulse as we saw in our earlier
∞
lectures, is just signal for which I integrate from ∫ δ ( t ) =1,the area goes to 1 and it just
−∞

looks something like this right. Similarly, a step is a signal that switches to 1, so it's just
at t equal to 0 plus, I am at 1 and I then stay at 1 and for tall t <0 , I am just at 0.
(Refer Slide Time: 12:26)

1
So, the Laplace transform we know of an impulse is 1 and of a step, I know it is . The
s
ramp is something which is just you know with constantly increasing with time at maybe
a slope of 1 or it could also have a different slope depending on what is the value of this

207
 A
A, and then the Laplace transform of this would be Ok. similarly, to a parabola. So,
s2
2
At
we have and then the response increases actually faster than what it was here.
2

2
At
So, one thing which you could notice is I take this parabola and I differentiate it,
2
what I get is At ok. Similarly, I differentiate; So, this is a parabola I differentiate it to get
a ramp, that is we can see it here, now I differentiate a ramp and I get a step right, ok.
Now just keep this in mind for a while that I differentiate a parabolic signal to get a
ramp, I differentiate a ramp and I just end up with a step, this is a step of size A. Now
we'll see does the response also have certain behavior like this right. We will just see if
that is true or not.

(Refer Slide Time: 13:51)

So, the simplest of all systems are systems with only one pole and are called the first
system. There could be some simpler systems, where I just go from here to here this is
just a number 1 and I am not worried about anything right. So, those are not of interest;
there are some dynamics here right in terms of τs.

So, if I just do the standard feedback interconnection I compute the transfer function, I

Y (s) 1
just get that the transfer function from R(s) to Y (s) is = and usually R(s) is it
R(s) τs+1
could be the reference signal which my system is subject to and I will see how my

208
system behaves for different reference signals right. The τ is a system constant and it
characterizes the speed of the response of a system to an input and we will see this with
the help of an example. So, let's just start with dealing with first-order transfer functions

1
which look like this .
τs+1

(Refer Slide Time: 14:46)

So, if I just take the impulse response, now we will use all the tricks which we learnt in
you know Laplace transforms, solving of differential equations and so on. So, my input
or the reference is just 1, Y (s) which is the transfer function G times R looks like this.

1 −t / τ
y (t) I just take the inverse Laplace of this guy and I get this guy. So, e , whereτ is
τ
the time constant of the system. So, if I just look at the plots quickly right. So, if I just
choose a particular value τ =1, then you know my response just dies this way; for τ =2, I
have something like this ok. So, you can see that with changing time constants your
response also changes, for example, the blue line which has a smaller time constant
approaches the horizontal line faster than the black line which has a higher time constant.
Ok.

How does it look in terms of the steady-state or and the transient response? So, this guy

1 −t / τ is called the transient term. If you see that you know as to how do, we define the
e
τ
transient term? that these guys tend to vanish as time becomes very large right. So, as
time tgoes to infinity this guy goes to 0. And of course, the steady-state term here is just

209
0 right, as I said earlier the steady-state could term that the Y eventually so, what the
term or the value where which the transient term settles down could actually be the
steady-state value which is true in this case.

(Refer Slide Time: 16:24)

Now I go slightly further right, I look at the step response or more simply a unit step I
can change the magnitude, but nothing much changes in the analysis.

So, Y (s) is again the transfer function is just a first-order transfer function the reference
is a step. So, my overall output would look like this ok. Then I do some tricks here and
−t
get the inverse Laplace transform as 1−e τ , that is something different here than what

we had learnt earlier. If I just plot this for 2 values say at τ =1, I start at 0, and I go and I
end up at 1 asymptotically or eventually. I take τ =2, I still end up at Y =1, but a little
slower than before right. So, in this case, there are like 2 terms. So, this is the term which
vanishes as at infinity right, the transient term and this is the steady-state term; at steady-
state what remains is only this guy 1 and at the steady-state, if I just write that down or
lim y ( t )=1.
t →∞

So, I could also compute it directly from here right just to see how the steady-state looks

1
like, you can just use the final value theorem, which goes lim sY ( s )is lim s ,
s→ 0 s→ 0 s(τs+1)
these guys vanish and I am just left with 1 that also is an agreement with what I get here.
And the system is stable so, I can always use the final value theorem. So, I have here I

210
have a transient term and I can see very explicitly and the steady-state term which is
actually non-zero, in the previous case the entire response was just going to 0.

(Refer Slide Time: 18:33)

Now, something good happens here right. So, look at this signal right. So, what I have is
at y (t) is 1−e−t / τ , and say t=0, y (0)=0 and that's what even the plot says. And at t
equal to the value of the time constant(τ ) I have y (t) is 1−e−τ / τ . So, I am just looking at
1−e is like .632 or 63.2% of the final value right. So, what I could observe here is, ok
−1

let’s say what happens even at t=2τ I reach. So, my y (t) would be 86.5% of the final
value, I do it for a 3 τ. So, y (t) at this for thist would be above 95% and at 4 times the
time constant i.e., 4 τ, I have y (t) is like 98.2% of the steady-state value which is 1
right. So, I can just plot this here at t=1, I can just for the blue line this will work nicely
here. So, I will just be somewhere this is like .632.

So, few things to observe here, so, one is let's see what decides this, if who goes fast and

d
who goes slow. So, let's try to differentiate this signal right. So, I have y (t) that is; so,
dt

1 −t / τ d 1
I have e , now at t=0. So, y (t) is . So, the initial velocity or the initial speed
τ dt T

1
here is of a slope . And then just goes here and just meets this point here. So, if I keep
T

1
on going at the speed of , I will hit this point exactly at t=τright. At this point, this is
T

211
my time constant, similarly if I go. So, here with the speed, it will be at t=2, that I am
getting this value right like this, this and this one you can compute it right, for different
time constants this is a general formula for that and of course, the slope keeps on
decreasing, decreasing and so on.

So, why is this important? and why can’t I, you know maintain the speed all the time and
just go to 1 in .6 seconds than just you know going this way? Well, this phenomenon
essentially is because of the dynamics of the system which gives you a response like this
or these curves could be explained by what is in here. So, this we could this something
what we also experience in our daily lives right. So, if I take an elevator from say floor
number 1 to floor number 6 you see that the initial velocity is a little higher and then as it
goes up and up you don't have the same velocity all the time until you reach the sixth
floor and then abruptly come to a stop; you go keep on going up and up and the velocity
keeps on going down, and then you just stop at the sixth floor very gradually. Same even
when I am driving a car from point a to point b right, I build up velocity and then I just
don't go and hit that point in the stop abruptly right, I just stop gradually.

So, this is a phenomenon that could be explained you know in that kind of fashion right.
And then this if you look at this the slope also decreases monotonically as t goes to
infinity and until it reaches the slope of 0. Ok. Now, what about these guys here? how do
how would I say that I have actually reached some kind of a steady-state value right. I
don't have to wait till infinity. So, how do we define is if I say of what I will slowly
define a term now call this settling time, or when I would say that my system has actually
reached the steady-state is when I am within 2% of the final value. So, if I reach .98 like
for example, after 4-time constants then I say that my system has actually reached a
steady state. This will be very useful for analysis purpose right this definition of settling
time.

(Refer Slide Time: 23:51)

212
So, after this well, you can look at the unit ramp response. So, even before I do this right.
So, let's let me do something else also here. So, here my reference was a unit step if I just
look at the standard configuration of the plant and controller. So, I had nothing, if I just

1
remember correctly, it will be ; I had a unit step here which was myr (t) this. So, y (t)
τs
plus-minus; so, if I call this say the error signal right, E(s); the difference between the
reference and the output. So, let me define something right. So, the error
e (t )=r(t )− y(t ). So, what is r (t)? it’s just a unit step 1, and what is y (t)? y (t) is
1−e
−t / τ
. So, this is simply e−t / τ, again this is all times greater than or equal to say this is
0 plus.

Now, what happens to the error at infinity right, as tgoes to infinity this is just this is my
term right that it goes to 0 means at infinity. So, if e ∞ is 0 from this expression I can say
that r (∞)is y (∞). And so this is exactly what is happening here right, at infinity both
signals become the same that my system I can say is asymptotically tracking a unit step.
It’s no matter what the value of the time constant is; this guy does a little slower, but it's
here. You could have even slower guys which could do here, but they would eventually
again go to go to the value 1. So, this system I would say again I am slowly defining
these terms, but we will elaborate these all a little later, has a steady-state error of 0.
Right I can just call this I am running out of space here, but let me write it here.

So, how do we define this? as a difference between the reference and the output and how
does it go at infinity. So, ok now, what happens when the input is a ramp or when I'm

213
 1
actually passing my system to track a ramp? R(s) and I do all these things and I
τs+1
see do the inverse Laplace. So, a response to a ramp would look like this t−τ + τ e−t / τ ok.
Again how do we classify or how do we differentiate between the transient response and
the steady-state response? Well the transient response, well the one which goes to 0 for
very large time. So, this guy and at steady-state, this guy still remains. So, it need not be
a; So, earlier we saw in the first example that the steady-state was 0, the second example
when I was tracking a step, the steady-state was 1, and now you see it is actually
depending on time also.

We will see what these things actually mean. So, let's see first what is the error? right,
does at t=∞ are my responses just meeting the original signal or not, again I just draw it

1
for the sake of completeness here, that here I have r (t), I have plus and a minus and I
τs
have Y (s). so Sorry, y (t) or whatever notation you want to follow; ok better Y (s)
because this iss, this also bes capital R(s). So, at infinity does my output track the ramp?
So, that depends on what is the error signal here E(s). So, what is E(s), or e(t)? e (t ) is
r (t) which is my t− y(t ). y (t) is the entire thing, t−τ + τ e−t / τ, Ok so what I'm left with
−t / τ
is τ −τ e . Now if I do these things e (∞) would be, well this nothing will change
here τ +e∞ , this guy at infinity will go to 0. So, nothing will change here; τ +0 is τ .

Now, what does this mean? So, this is my r (t), this is my reference signal right. So, as t
goes to infinity, I am not touching this guy, but I am actually just at a distance of τ ,
which is 1 in this case right. Here at a distance of 2 right. So, there is some kind of a
steady-state error here, which is just governed by this number τ . The smaller this number
smaller the steady-state error. So, what I conclude is that well a first-order system, if it
has a ramp it will have some steady-state error ok. How do I encounter the steady-state
error? let's say I can take this example of a pendulum right? And I start it with some non-
initial condition and I expect it to go to θ=0 if this is θ, but there could be something in
the system which stops it going, maybe it just stops at θ equal to say .1°right. So, that is
the idea of this steady-state error.

So, it does not really reach the final the desired value, but some number very close to that
there is an error, but there is a way to determine that error, how small it is? how large it
is? I can directly quantify by just looking at the dynamics of the system.

(Refer Slide Time: 30:04)

214
 2
t
Now finally, we will see another test signal it’s the parabolic input. Well, r ( t ) = , the
2

1
Laplace transform is , and I do all these things and I get signal which is
s3
2
t
−τt+ τ −τ e . Again this has a steady-state, this transient term which we will just
2 2 −t / τ
2
die down as time goes to infinity and a steady-state term; well this actually looks
possibly quite worse than the additional one it has a t 2 term now right this is no longer
constant.

What does the error do here right? So, can my system track a unit parabola? So, the error

2 2
t t
e (t ) is my reference signal that is
2 ( 2
)
− −τt +τ . I would exclude this exponential term
2
because I know it will not contribute to the steady-state error right. It'll go to 0 anyways;
So, this will be τt plus or with a minus τ 2. So, e (∞) is actually a little strange here right
e (t ) when t goes to∞ this guy also goes to ∞ . So, the error will be very large. So, if you
see that the distance between these 2 this is my actual reference signal and this is my
y (t) as I go to infinity, this distance will keep on increasing right, it will keep on
increasing increasing and then if you keep plotting this forever, you will see that you
know the distance actually becomes unbounded, similarly with here. With the different
time constant of τ =2 right.

215
Before this right, so in one of my, you know earlier slides, what I had shown was that
these things right. I take a parabola; I differentiate it I get a ramp. I take a ramp I
differentiate and I get a step. Now let's see if the responses also follow some kind of a
pattern like that. So, this is my response here. So, this is this, this entire thing this, this

d
guy. So, just differentiate this with time. So, y ( t ), in this case, would be t−τ, and then
dt
here you have plus τ e−t /τ tau will be here, something like this ok. Now, just remember
this t−τ + τ e−t / τ, and this looks something like this right. Exactly this one right; this is

dy
just the differentiation of for a unit parabola. So, I differentiate the response I get,
dt
now the response to a unit ramp. So, the differentiation holds until now.

dy dy
Now, let me differentiate this again, . when the input is a ramp is this will be 1, this
dt dt
will go to 0, this should be a −e −t / τ ok. Now ok, this e this looks 1−e−t / τ. We'll go
backslides, and ok how this is the same right this one and I can keep doing this right. So,
you see that if the signals are preserved under differentiation, in this way that I
differentiate a parabola to get a ramp the responses also follow the same pattern. If I
know the response to a parabolic input, how do I find the response to a ramp? I just
differentiate this again from here to here one more differentiation, here to here one more
differentiation ok.

ok good.

(Refer Slide Time: 34:25)

216
Now, we'll go to something slightly complicated, but still not very complicated, but very
useful. So, second-order systems, these are the systems which have 2 poles. We have
done lots of examples on this say a typical RLC circuit would have a transfer function
like this if you remember vaguely this was the transfer function between the voltage of
the capacitor and the input voltage. Similarly, for the mass-spring-damper system where
I was taking the input was a force output was a position or a velocity one of those. And
then, in general, I can write transfer function as, well a second-order polynomial in the
denominator and some number in the numerator and else slowly we will learn why these
are equal? what are these numbers? and so on.

So, the response here will be a little more tricky than what we saw earlier ok. So, how do
I derive the system? let's start with an open-loop thing which looks like this. I have unity
feedback and I go to a closed-loop transfer function which looks like this ok. Some
terminology which will shortly be clear is that ω nis what I call the natural frequency of
the system. 𝛇 is a system damping ratio.

(Refer Slide Time: 35:25)

Let's just remember it with these names at the moment and we will see slowly what they
could be.

(Refer Slide Time: 35:49)

217
So, just some properties of it, the damping ratio is a dimensionless quantity and it
describes some kind of damped oscillation during the transient response of a system. So,
we will leave for the next lecture or the lecture after that. That is, it is the ratio of the
actual damping to critical damping; no need to understand this statement at the moment.

But we will recollect this when we go and then you know try to justify what this means.
Natural frequency is the frequency, at which the system tends to oscillate in absence of a
damping force, this will be clear shortly. So, again together with ω dwhich is the damped
frequency of the system ok. So, we will again come back to this as we learn it through
some examples.

(Refer Slide Time: 36:36)

218
So, the response is essentially now depending on the damping ratio 𝛇. Some
classifications before we understand them physically is when ζ >1, I call my system to be
overdamped. ζ =1 it is a critically damped,0<ζ <1 under-damped, ζ =0 undamped
systems. You might have encountered these words somewhere earlier in some of the
courses. If not do not worry we will try to define all those things and of course, we are
not really interested in negative damping at the moment, because again all the things we
are doing are restricted to stable systems.
(Refer Slide Time: 37:24)

But you know a system with a negative damping ratio could resemble the behavior of an
unstable system, which we are not interested in at the moment ok. what is damping right?
So, damping is something which kind of reduces or prevents oscillations in the systems
right. So I could have systems which are overdamped in a way that transients
exponentially decay without any oscillations, I would have critically damped where you
know the systems decay to the steady-state value with or in the shortest possible time. I
could have oscillations which actually go to 0, and I could have undamped systems,
systems which just keep oscillating at its natural frequency without any decay in
amplitude. Ok.

(Refer Slide Time: 38:04)

219
So, a typical response the red color here would be an undamped system. So, let me see if
I could explain you with a simple pendulum example right. So, I am here this is my θ
equal to 0 position, and say I have an initial condition starting at say θ equal to some 15 °
right, some other θ and I am starting here if the system does not have; it is not subject to
any sort of friction either here or in the air. So, what we will be, we expect is it just keeps
oscillating right. So, we just go from θ equal to 15 ° if this is a plus, if this is a minus 15,
it will go to plus 15 and just keep on doing this right. So, this will correspond to the
behavior when 𝛇 is 0. And I will explain to you this mathematically very shortly. There
could be cases when I could have a situation where this guy goes from 15, maybe it will
go to a 14.9 here come back and it will keep on you know oscillating go back and forth,
back and forth until it reaches this point. So, these are called underdamped systems right.
When ζ <1. There could be a situation where I start from θ=15, I go and I just slowly
settle here, I don't go to the other side I just go and I just stop here ok.

There could be say, like this is sometimes in 1 second. There could be situations, where I
start from here again and I go very slowly, say it I am going in 3 seconds; there could be
even worst situations. So, I am not going you know I am not crossing this line, I am not
crossing this line, the Lakshmanrekha here, so you can call it. There could also be
situations where I start at θ is again 15 or minus 15 depending on the convention, that I
am going very slow, very slow, very slow and I just stop here, this could take me 10
seconds ok. So, few kinds of responses ζ =0, I keep on the oscillating right. For some
value of ζ <1, I am oscillating, but my oscillations are dying down, dying out before I
reach this steady state position, there could be a situation where I just go from here till

220
here in 1 second and just stop. I don't go beyond θ=0 it could take 3 seconds it could
take 5 seconds.

So, the fastest value of and these things will vary with 𝛇 right; ζ =0, ζ <1, and then all
this behavior would be ζ ≥1. So, the fastest where I can reach here and stay in the
critically damped response. Which is like the blue line here and for damping anything
larger than that I will just keep taking a longer time I could have a response also here
right some something like here which could be 𝛇 still greater than 1 right. So, these are
typically the four kinds of responses which we will encounter in designing a system. If I
take, for example, I want to design a lift or an elevator and I want to go from say this is
level 1 to say level 2; I am lazy, and I do not want to walk and I take this elevator and I
go up I don’t want my system to go to 2, 2 point you know halfway to the third floor,
come back here and then do this and I do not want it to do this right.

Now, that will be a very unpleasant right. Even if you know, as you see on TV even if
Deepika Padukone is stuck with music it'll still be very unpleasant. And therefore, what
we design these systems are to be critically damped or like slightly overdamped right that
I just go here and I just slowly settle down here right. So, again depending on my
requirement I would just want to you know have my system to be either be
underdamped, critically damped or even sometimes overdamped.

(Refer Slide Time: 42:22)

We'll characterize all this very shortly. So, how does my system behave again with the
standard test signals? So, here I'll have to be careful that there is another thing just not

221
just the test signal, but something depends also on the value of 𝛇 right. So, earlier my
formulas were very easy to remember, but here you know this 𝛇 adds more complexity
2
ωn
to me right or more it presents me with more test cases. So, Y ( s )= . So,
s2 +2ζ ωn s+ωn 2
ζ =0 would be this guy would go away. Y (s) is this one and I can do the inverse Laplace
and I can just say I just have a sinusoid of some frequency ω n ok, let's just highlight this
very nicely ω n ok. How does the response look like? well, the response is just a sinusoid,
I say what is its amplitude? well, the amplitude is this one, could now what I am
interested in is the frequency ok. ζ =0 and if there is no damping it will oscillate at ω n
and therefore, I call this as the natural frequency as defined earlier.

So, here so, did we define that yeah yes the system natural frequency: The angular
frequency at which the system tends to oscillate in the absence of damping force. That is
exactly what we did now, there is no damping force, 𝛇 disappears and then my naturally
my system just oscillates ok.

(Refer Slide Time: 44:09)

So, this is the natural frequency when there is no damping. So, let's make things a little
complicated. So, 𝛇 between 0 and 1 ok so, before I do this I'll just do some very short
calculation here. So, look at the denominator I have s2 +2ζ ωn s+ωn 2 ok. So, just you say
and if I look at you know how do I find the roots of this? This roots would be so,

−ζ ωn ± ( ζ ωn ) −ωn2 right. So, this is how roots will look like, you'll divide by 2 and
2
√
you know we saw the standard formula, or this is −ζ ωn ± ωn √ ζ 2−1 . Now, when

222
0<ζ <1 this will become a complex number, and therefore, the roots will look something
like this. So, here and I call this guy. So, if this is complex I will just write it here
−ζ ωn ± j ωn √ 1−ζ 2 .

And this guy I will call as ω d ok, and this is how these roots now expand right,
s+ζ ωn− j ω d and s+ζ ωn + j ωd exact what is here ok.

Now, I just want to see the natural response or the impulse response and I see I just do all
the Laplace transforms, inverse Laplace transforms which I learnt earlier. I see that y (t)
is some number here ω n blah blah blah, then I have an exponentially decaying term and
this guy I say exponentially decaying because all these guys are positive and with the
time you know it just falls in value and goes to 0. It still now earlier here I had sin ωn t
whereas, here now I still have a sin curve or a sin wave, but with a different frequency
that is ω d right and therefore, sin ωd t and I called this as the damped frequency. The
system damped frequency, angular frequency at which the system tends to oscillate in
the presence of a damping force and while 0<ζ <1 and this is exactly the formula we use
there; this ω d is not because of any magic that I defined it just because it just sits in
beautifully here. In the term of sin ωd t right, this is why I call this ω d t , as the damped
natural frequency this is ω n √ 1−ζ 2.

So, this is the damped frequency. And of course, I have this exponentially decaying term
and if you see the response where it will keep oscillating and eventually go to 0. And this
is the response of the pendulum which we are talked about when 0<ζ <1 right and this is
the sin here says that there will actually be oscillations and I will just go to whatever this
number here is come back and of course, these numbers depend on what are the specific
values of 𝛇 and ω n right.

(Refer Slide Time: 47:54)

223
Now, another case could be when ζ =1. So, when ζ =1 my expression for Y just looks
like this s2 +2ω n s+ω n2, and the inverse Laplace would just be like this right. And if you
see the sinusoidal term has disappeared which means there will not be any oscillations;
however, you will see that as time goes to infinity I just come back to 0, this is again for
a particular value of ζ =1 and ω n, we'll later investigate what happens for larger values of
𝛇 and so on.

So, this is; so, here for a system which we actually called now as a critically damped
system. The transients in the system decay to a steady state without any oscillations and
in the shortest time possible as we have defined in the case of the pendulum this is the
shortest time and this is the critically damped case. Ok.

(Refer Slide Time: 48:49)

224
Now I could go overdamped system right this ζ >1, in which case this roots will just be
real. So, ζ <1 I had complex roots, but I will just have real roots. So what happens here?
2
ωn
So, , I can just write it as a composition of 2 real roots like this for which
s2 +2 ζ ωn s+ωn2
the response looks like 2 exponentially decaying signals right 𝛇 ω n blah blah blah and 𝛇
ω n blah blah blah. And again this response you could see there is no oscillating term
here, there is no sinusoidal term. So, you just go here and then just come back possibly
when 𝛇 equal to you know 3, you might have even much slower response right it could
take even much larger time like this and so on.

So, we saw 3 cases now right first is the or actually 4 cases. The undamped case where I
have the system just oscillates at the natural frequency. Then I had well I had
oscillations, but they just died down. They died down because of this exponentially
decaying term here right. And this will oscillate at now a new frequency which is my 𝛇
defines, again you can just put ζ =0 and say it will again go back to the damped natural
frequency, and then a critically damped system and what we call as an overdamped
system.

(Refer Slide Time: 50:26)

225
Now, this was all about the impulse response. Now how will these guys behave when

1
there is a step response? So, step is in my R(s)= , then Y (s) would be something like
s
this right, I just have s over here and blah blah ok. Case of ζ =0, an undamped system
well nothing should change in principle as what was observed from the case when I had
just an impulse response.

So, Y (s) I just do all the math and I see well y (t) is 1−cos ( ωn t ) ok. So, this system well
is just1−cos ( ωn t ), it will just be like this right. And you can ask me, sir, what is transient
and steady-state for this, this guy will never reach the steady-state right. Whereas these
guys here these guys will have you know this kind of transient terms which go to 0, even
the previous guy would have a transient term which goes to 0, initially the steady-state is
again coming back to 0 correct, here also you will have the transient term which actually
is going to 0 this entire term could be viewed as a transient term and you know the
steady-state term is again 0, here well there is no nothing like that right there is nothing
which is actually vanishing with time.

So, there is no possibly we cannot know nicely classify what is a transient and what is a
steady-state its transient behavior is actually it is steady-state behavior and it will never
convert to any fixed value like a 0 or a 1 as in the previous case. So, the same thing
happens when I have a unit step response, I am still doing this one right none of the
terms actually disappears as time goes to infinity, all terms are just intact this is again
ζ =0, ω n=5and the reference just being a step.

226
(Refer Slide Time: 52:16)

Now, we'll do all the same cases which we did earlier, I will just skip all the math here,
but we'll see how my final expression looks like.

So, for 𝛇 equal to this guy I will start with this and then I will have this one, I will, of
course, have complex roots and all this will be taken care of here. So, my response has 1
minus some exponentially decaying term again. You have some oscillating terms here
which could be further simplified to look something like this.

(Refer Slide Time: 52:46)

227
 −ζ ωn t
e
That is 1− 2
sin ( ωd t +θ ),all these exponentially decaying terms and this one or I
√ ( 1−ζ )
could call this as my transient response. And this as my steady-state response right and,
but θ is we'll know at the moment we'll just say that well the θ=cos−1 ζ . Of course, we
are not really worried about why what this means, but all the math would tell me it is just
this one. Ok.

So, how does the response look like y (t) goes? So, I want to track this number 1 and I
go up and overshoot and then I will go below 1 and then I eventually go to this number
ok. Which I can directly say here now I do not really need to compute what the steady-
state error just by looking at this thing I can say that the steady-state error at infinity is
just 0. We'll do more analysis on this, but just as a passing observation I am just
introducing you to these terms at the moment right and of course, we can have different
behaviors for different 𝛇 right as 𝛇 is. So, intuitively you can see that as 𝛇 goes to 0 you
know is very 𝛇 is very close to 0 say 0.01 then you could actually expect a response
which is like this, when 𝛇 is very close to 0 you could expect a response like this, but
which is very slowly decaying right. So, it might actually take a very long time to decay.

So, if I just draw something here, here and I will see these oscillations for a very long
time and but they will slowly die down right. And of course, as 𝛇 goes to very close to
one then you will see that the oscillations are also minimal again this ω d guy just sits
here right, the damped frequency of the system. Now, this is again because of this guy
the 𝛇 right. A justification to why we introduced these terms specifically and called it the
damped frequency of the system because this is the frequency of the signal when there is
in the presence of damping.

(Refer Slide Time: 54:57)

228
So, next what happens when I am critically damped, I will just skip all this math and just
see that well the sinusoidal terms disappear, and I have just this one.
−ωn t −ω n t
y ( t ) =1−e −ωn t e right. And I can just compute this response either with the final
value theorem here or just put t=∞ and I can just compute the error; the error is again 0.
So, this is like it goes from 0 to y in the shortest possible time and just stays there.

(Refer Slide Time: 55:26)

The next one is the case of an overdamped system. Again we'll have two exponentially
decaying terms, which will just be like this right, it will go a little slower than the
previous one, and you could just plot you know things 𝛇 goes to 3 and possibly it will

229
have a much slower response 𝛇 equal to 4 and much slower response and the limiting
case you will say well there is a very large 𝛇 then you just refuse to move.

(Refer Slide Time: 55:30)

So, these are also literature referred to as very sluggish systems, which are very slow to
respond right. So, also, in this case, you can just you know naturally see that the steady-
state error goes to 0. So, what we have we have learnt so far, today is what do we do with
the models which we obtained from our previous classes.

(Refer Slide Time: 56:08)

So, once I have the models I need to subject them to some real-time conditions which I
possibly would not be able to emulate when I am just having a system at hand. So I'll just

230
give it certain test signals which will be closer to what I can see in reality. And therefore,
I have a series of signals in my case I have the impulse, the step, the ramp, and the
parabolic kind of signal, and we just saw very; saw how we could define what we call as
the settling time? What are the transients? what are the steady-state? And in the, we did it
for the first order, second order case was a little more interesting because the damping
actually played a role in the dynamics of the system. If I have oscillations if I have
sustained oscillations, if I have decaying oscillations or if I have no oscillations at all
right. So, that was a summary of what we did today, and we'll do a little more analysis or
we'll elaborate more on this in the next lecture, when we do more analysis on this time
domain specification.

Thank you.

231
 Control Engineering
Dr. Ramkrishna Pasumarthy
Department of Electrical Engineering
Indian Institute of Technology, Madras

Module – 03
Lecture – 04
Time Response Specifications

Okay. So, what we will look at today is something called the time response specifications. So,
last time what we had done was given a first-order or a second-order system.

(Refer Slide Time: 00:32)

We just saw how it responds to different kinds of signals and in the first-order system, we saw
there was something called a steady-state error we will formulize that in a better way today.
Second-order systems we saw how the response of the system varies with changing in the
damping term; where, where we saw for an underdamped system your response is just
continuous oscillations still you have critically damped systems to overdamped systems and
so on.
So, what we will see now is given some design specifications, what is the impact of the time
response? So, loosely speaking if I say that I would want to go from say you know some
signals say y (t); I want it to take some value of say y=1, somewhere here. So, ideally what
if I would want to design a system I would want this guy to go instantaneously to this point I

232
just say t=0+¿ ¿, that I switch on my air conditioner and immediately my temperature goes
to the desired 24 degrees or 21 degrees ok.
Sometimes might be difficult to do it instantaneously, but it could be possible to do it a little
you know something like this right or something like this right. So, not exactly like this, but
we could you know at this time or this time, but what I would expect is if I, you know want to
run from this point till this point. And do very fast then I just cannot stop here right I have to
go a little further right before then I again come back and I run a little faster then I go here,
here, here and so on right.

However, if I wouldn't want these things to happen I would have to run a little slowly. So
that, I just go here and stop right. So, these are the kind of behavior we will analyze is to do
with how fast can my system respond or how fast can it reach the given the desired value, but
if it runs very fast are there any overshoots and if there is any overshoot is there a limit on the
overshoot or is, can I really say how say what is the value of y at this point given this guy to
be 1. Or if you know when does it really reach the desired value right and so on that's what
we'll try to you know quantify in this lecture. So, and we'll restrict ourselves to the step
response of the system and the analysis becomes much easier and it's also you know quite
intuitive too.

So, in general all these specifications of how fast I go, how fast I settle down are a part of the
design specifications, if someone gives me a problem would say, okay design me something
which reaches the steady-state value in so and so time that it should not overshoot beyond a
certain limit say this could be just .1 and I would say I would not want to overshoot 1.1 right,
for example, and these indices answer the following question. So, based on this are humans
what are we interested in: Is how fast this system moves to a given input? Is the response
oscillatory? and how oscillatory is the response? which is again we know that it is depending
on the damping that under damped systems will have an oscillatory response and how fast
does it reach the final value?

(Refer Slide Time: 04:09)

233
So, here we'll restrict ourselves to underdamped systems right. And we'll focus on the step
response or the unit step response. So, to quantify this we will define for ourselves few things
right, first is the delay time: the time required for the response to reach 50 % of the final value
at the first instance. And I could calculate this based on the formula which I had derived
earlier, I will come to this we’ll just, first define what these are and then go to see how these
numbers are achieved.

So, the delay time is a time required for the response to reach 50 %of the final value at the
first instance, you could have that in an oscillatory response or it could reach the 50 %you
know several times. The rise time is usually defined as a response to rise from 10 %to 90 %of
the final value for overdamped systems and since we are here looking at underdamped
systems we are we will look at, what time does it reach the peak the desired value the first
time or from 0to 100 %of final value at the first instance. For example, if I just were to plot it.
So, this till this and I say this is my desired value these things right. So, here well this is what
is called the rise time right. So, to where it reaches the value of one the first time. This is
where you get this is value here, of 1 here and so on it keeps on you know oscillating, until it
settles down to its value final value and the delay time is this time right. So, this is at t d where
my response takes the value of 0.5. So, this is my time, this is my response y (t).

234
Similarly, the peak time is the time required for the response to reach the peak value. So,
somewhere here. So, this is my y peak time. Then I will also quantify if I reach a peak if I am
going from the desired value to some other value, how much is this overshoot?

(Refer Slide Time: 06:23)

So, this is quantified or called the peak overshoot and this is just computed as the percentage
of overshoot like. So, this how much I deviate from the steady-state value at the first instance.
So y (t p ) which is what I overshoot here, this one, this is the value of y (t p ) minus, the steady-
state value and then you just calculate that it in the percentage of deviation from the steady-
state value.

The settling time. So, we had vaguely defined this last time when we were dealing with first-
order systems, and we'll recall that all over again the settling time was defined as the time
required for the response to reach and stay within a specified tolerance band. So usually I
would say well I am this system has reached some kind of a steady-state value, if it is using
2 % of its final value, sometimes even 5 % depending on my tolerance levels. So, if I in this
case, where the steady-state value is 1, I would say the settling time is when it reaches 2.98.

So, again all these things; so settling time we said we thought or we had derived the settling
time concept earlier for first-order system right. And this concept the peak time and the peak
overshoot are not defined for overdamped and critically damped systems. That is because, ok
let's start with the response. So, this is the overdamped response, the critically damped
response would be something like this right, it will never cross this value of 1, or the desired

235
value. This is a critically damped system. So, this is the overdamped system. This is
overdamped, this is critically damped. So, t pand M pare typically defined only for
underdamped systems. And that is what we'll focus most of our analysis on. Ok.

(Refer Slide Time: 08:11)

So, this what, what it looks like. So, a typical response of a first-order system this we derived
last time it looks like; this is like has some kind of damped oscillations.

So, again I have the delay time, the rise time, the peak time, and the settling time right. So, the
first time when it reaches this guy. So, this could be within 2, 2 % could be 0.98 or even 1.02,
but once with it is here it always remains within that 2 % value right, it will never go down
you know below 0.98 because the oscillations are continuously being damped. So this is these
are the values which we'll focus on t p, t r , M pand all this settling time.

Now, the idea would be can I quantify this based on my system parameters? right.

(Refer Slide Time: 09:00)

236
So, start with the second-order underdamped system with transfer function which we had
derived which we had said last time was has a typical transfer function if I would call it G(s)

2
ωn
was . And the response of this system and we'll do this 𝛇 is less than 1 and of
S 2 +2ζ ωn S+ωn 2
course, greater than 0 i.e., 0<ζ <1.

So, the rise time, so this system subject to a step input had a solution like this and we derived
this last time at with t no not replaced by t r right. So, this t r is the reason we have t r is. So, the
rise time is a time taken by the step response to go from 0 ¿100 %. So, this is the 100 % this is
the solution of the equation and I want to calculate the time t at which it reaches this value of
1 and therefore, I replace this t by t r in the solution to that equation. So, I do this, 1 and 1 go
away. So, I am I am left with sin ωd this is the damped natural frequency if you remember
was related to the underdamped frequency in this way okay, this was ω d. So, when this goes

π −θ
to 0 for the first time. So, which means ω d t r +θ=π or t r is just this one, and what was θ?
ωd
So, θis written here θ=cos−1 ζ from the damping coefficient right. And then this relation
between ω d and ω n, the damped natural frequency with the natural frequency of the system
ok.

(Refer Slide Time: 11:03)

237
Similarly, I can calculate the peak time is the time taken by the step response to reach its peak
value. Ok. Now, how do I find what is the time? right. So, if I find here I am looking at this
value right. So, this is a function if I just. So, y is a function of t. So, it starts here, it goes
here and again comes down right. So, this is the maximum value that the function can take.
And from calculus, I know that the maxima can be computed by taking the derivative of y.

dy
And then setting this to 0, What this really means is I take , I set it to 0 and I compute the
dt
time at which this happens right, so, the time over here.

So, I do all these computations. So, I take the derivative of this. So, this 1 disappears, and I
have is this one 0 is all these guys. And then I solve this step by step to get this value right
ω d t p is this one, and therefore, again I am just interested in the first instance right. So, this is

dy dy
one peak and then would go to 0 here, could go to 0 and at several points right here and
dt dt
so on, but I am interested only the first time and this is the value which I would be interested
in right.

(Refer Slide Time: 12:37)

238
 π
So, this t p would be simply . So, how do I express this overshoot as a percentage? right.
ωd
So, I calculate the peak value and then the steady-state value is 1, the desired steady-state
value and here, in this case, it will just be M p would be t p −1 right. So, how much I overshoot
from here this, this number from here till here ok. Ok, so I can just substitute y (t p ) here. And

π
I just y (t p ) would be just I know the value of t p, that this is . I plug it into this solution ok,
ωd
this the solution over here and I just see how much it deviates from the steady-state value of
1.

So, y ( t p )substitute for t p in the solution of y (t) and this with all these computations would be
something like this. So, M p ise power blah blah blah and then this term gets canceled because
of cos θ, if you see here is defined this way right, θ is cos' ζ or cos θ=ζ and then I just use
sin θ+ cos θ=1and this guy just disappear. So, the peak overshoot in percentage is e to the
2 2

power you know this guy. Ok, some properties of this one right. So, if I say well what are the
limiting cases here? right. So, we go from say 𝛇 equal to 0, when there is no damping what
we would expect is that M pwould be the highest.

So, if my steady-state value is 1, then M p here would be,ok I substitute for 𝛇 is 0; 0 and so
M p would be 1, which means my peak overshoot would go from this is a value is 0 to 1, this
is the desired value it will go all the way till 2 and then come back and just keep oscillating

239
like this. This is the standard response which we saw, it is almost it’s like a hundred percent
overshoot.

Now, if 𝛇 keeps on increasing; I see that my overshoot keeps on decreasing right, so if I just
were to plot the peak overshoot versus 𝛇, this 𝛇 going from 0 to 1. I see that this will go from
1 and all the way till 0 right. This is a little important right. So, one of the design parameters
here could be if I were to restrict my peak overshoot then my problem would be to have; to
design my system with the appropriate 𝛇.

(Refer Slide Time: 15:36)

So, similarly, for settling time the settling time is the time required for the response to reach
and stay within say at 2 % band of it is final value. So, how do we compute this? So, let's say
some pictures here that I am here and my response is again something like this right. And this
is my steady-state value, let's say this is 1. Okay, we'll have to go a little further up here.

So, if I look at this, this response here let's say it is actually bounded from below by a curve
like this and bounded from above by another curve like this ok. Now, what are these curves?
−ζ ωn t
e
these are simply these exponential curves here the guy on the top would be 1+ . Ok so,
√ 1−ζ 2
this, this r would just go away here; this will just be a standard t. So, this is the curve on the

240
 −ζ ω n t
e
top, and this is the curve on the bottom; this is 1− and then so this value here would
√ 1−ζ 2
1 1
be 1− 2 , and the value here would be 1+ . So, this is all good right. So, if I say
√ 1−ζ √ 1−ζ 2
well when does this curve inside, this solid line reaches within 98 % and stays there; well the
answer could be that whenever this curves you know this, this guy when whenever this reach
the within the 98 % and then stay there because these are enveloped by these curves right this
the guy on the solid line is just enveloped by these 2 curves.

So, I can just now look at the response of this system right. So, I just say let's take the bottom
−ζ ω n t
e
one here, for example, 1− . And this, if you remember this looks like the response of a
√ 1−ζ 2
first-order system right, where we had the other day that the response was 1−e−t / τ right, and
if you just look at that response we had some things defined like this. This is my 1, this will
be my response y (t) and this, this kind of looks similar right.

1
So, this is a time constant of t and time constant . So, this guy reaches t at the first time
ζωn t
constant, it reaches the value of 63% and then at t equal to 2 times the it reaches some value
of 0.8 right and then at t=3 t it reaches 95% and at t equal to 4 times the time constant it
reaches 98.2% right. Somewhere here. Right, this we had done this last time. This 2t, this is
3 t, this is 4 t. Ok. Now, if this guy reaches settling time at 4 times the time constant, this guy
would also be you know reaching in the same time right because this is just enveloped by
these; by the curves there right. So, now, my settling time can be defined directly in terms of
the time constant right. This is for the 2% tolerance band and settling time t s and then t s
would be 3 times the settling time when I am at the 5% tolerance band.

(Refer Slide Time: 20:04)

241
So, how do I write this? So t s , so if I again look at my curve; the time constant there was say I
−ζ ωn t
−e 1
just take the bottom curve 1 . So, this has a time constant of , and therefore, my
√ 1−ζ 2 ζωn

4
settling time would just be 4 over, 4 times the time constant that'll be or the settling time
ζωn

3
this is for the 2% criterion and this would be for at the 5% criteria.
ζωn

So, what does this mean that the settling time is inversely proportional to 𝛇 and ω n again let's
say the limiting case, let us say the limiting case of 𝛇 being 0 my settling time would be
infinity, this is not surprising because if my settling time is infinity or when 𝛇 is 0, I am just
having like an oscillatory response right. My system will never go anywhere within the 2% or
even the 5% or even 50%.

So, settling time would be infinity. So, as 𝛇 increases the settling time tends to reduce right.
So, based on this formula here 𝛇 increases and then the settling time reduces. So, I don't want
to draw a graph for that at the moment, but we'll just understand that the increase in 𝛇 will
cause a reduction in the settling time. Ok now, what is; is that sufficient to increase 𝛇? So, my
peak overshoot is also some function of 𝛇. In such a way that if I have a smaller 𝛇, I have a

242
smaller peak time or then even a smaller rise time right, if I want a faster response my 𝛇 is
smaller, but then my overshoot will be larger right.

So, usually in design specifications, my peak overshoot would be specified, and my 𝛇 would
be calculated based on the peak overshoot. And therefore, if I were to look at settling time it
just depends on the natural frequency of the system.

So, to write the settling time I can play around with this guy ω n right the ω n also over here.
Because 𝛇 is already defined by the performance specifications of the peak overshoot. So,
when this is fixed, I am may not be able to change it here again because you know the smaller
𝛇 would maybe increase in a bigger overshoot which may not be very desirable. So, I can just
play around with this ω n. That is only we'll see now what how we could smartly design these
things. So, that I have not a big overshoot, but also a smaller settling time and so on.

(Refer Slide Time: 23:18)

So, ideally what we would like is to have a faster response and a smaller settling time, ok? So,
what could be the application of damped or this several kinds of systems right? So, let's start
with overdamped systems, now this is push-button, water tap shut off valves. So, just before
we started this recording, my student was telling me a nice analogy of this right. So,
originally if you look at the trains, the taps in the train we would have normal taps which
people would just forget to turn off and then we could waste water, then, later on, they had

243
these things, you know I don't know if I could draw a picture of this right. So, it is like this
and then you will have a tap and then the water flows through here right, and when you keep
on pressing this and then once you release this the water will go will not come anymore right.

So, what they do now is that you just press it. So, the system will you know give us water for
some time and eventually the valve will close. Well, that is like the example of an
underdamped system, like the emergence in technology is can be seen you know when you
travel by train right. I don't know if the trains still have this kind of smart things or not, but
this is in more sophisticated thing you have this sensor you just put the hand and then you just
actuate this and then water just you know the valve just close of closes off gradually. The
same thing you can see in this automatic door closes right, they don't really be they cannot
really have a fast response and then bang on to the wall or to the stopper right. So, these are
automatic door closers are also, typically overdamped systems. Critically damped systems
you would want them to have elevator mechanisms or even gun mechanisms right. So, you
want the elevator to have a faster response, but you don't want it to overshoot. So, we talked
of this last time as well and of course, underdamped systems all string instruments would do
that because you know if you pluck a string, say I have a guitar I pluck the string and usually
it has this nice wave kind of formation right, but if you pluck the string from here it goes here
and again comes back here it may not generate the sound which you want.

Similarly, all electrical or mechanical measuring instruments, right. So, you take a Volt-meter
if it were to measure via a scale like this. So, if this is a 0 position it will typically go to the
desired value and then come back and these are usually underdamped systems right all this is
moving coil or moving measuring instruments.

(Refer Slide Time: 25:41)

244
So once we do these things we are also interested in does my output follow the reference all
the time.

So, we saw last time when we were doing the response for a first-order system, in some cases
we saw that there is a steady-state error which means the reference is not tracked directly by
the output, but it is with some error which was defined by the time constant right. Ok so, let's
try to formulise that a little more and then we'll revisit the original examples which we had.
So, the steady-state errors. So, a typical feedback loop looks like this, you have the reference
signal, the output, the feedback loop, and that the difference between the reference and the
output is termed as the error signal. And this error I would typically want to go to 0. Now
given these characteristics of the system, the G(s) and this 1, it will also be H (s) nothing
would change in the analysis. So, given this G(s) can I find what is the error? right;
depending on the nature of the signal here ok.

So, what is the steady-state error? It is an error between the actual output and the desired
output as time goes to infinity right. So, we know from the final value theorem if I call this

signal as e (t ) that e ss is lim

S→0
sE (s) in the Laplacian domain. Now, this E is R−Y , and Y I

G(s)
know is, so, this transfer function here is . So, this Y given this R. So, I can write this
1+G(s)

−GR R
as R and I get this one . So, this is my error signal.
1+G 1+G

245
So, in this thing the steady-state error, the lim sE (s) would be lim sR (s) . Now, we'll make
S→0 S→0 1+G( s)

use of this little thing here and it will give us lots of information of the system subject to
several inputs.

(Refer Slide Time: 27:56)

So, let us start with a step input right. So, this will be fixed all the time. So, for the step input

1 sR (s) 1
R(s) is , and what is the steady-state error? e ss is lim and this R(s) is right. Ok
s s→ 0 1+G( S) s

1 1
so, this s and term has disappeared and I am left with this. So, this limit would be
s 1+ K p

right. So, this K p would be lim

s→ 0
G ( s ) and this limit will always exist, because we are looking at

causal systems right. So, well with this limit will always exist; we don't really care about if
this might block to infinity or not. And this also we with this like this Eigen will always exist.

So, the K p is computed as lim

s→ 0
G ( s ) is called the position error constant right. So, I am just

1
tracking a position, a fixed position . Now, similarly, if I go to an input which is ramp,
s
where I am tracking a ramp. So, this steady-state error is again I do all the computations. One
of the s disappears and I am just left with this and this guy goes to 0, let's compute that. So, I

246
 1 1 1
have lim s
s→ 0 s (
2
1+G(s) )
. So, this would be lim
s→ 0 (
s+sG (s) )
, 1 of the s's would disappear and I

have this. So, now this would go to 0 thiss. And what I am left with is lim
s→ 0 ( sG(1 s) ) and I call
this the K v . So, where K v is lim
s→ 0
sG (s)and I call it the velocity error constant.

So, if there is, if I have a feedback loop, I will just draw it here again. So 1, reference, plus-
minus, Y (s) and so on right. So, if the input or the reference is a step, then there will always
be an error. So, the response, so Y (s), if I just take an underdamped whatever kind of system
and if I just look at the steady-state curve. So, let's say this is my desired value, my actual
value will just be somewhere here right. So, this is the steady-state error.

So, we might think of if this is a position error, we might sometimes think that the velocity
error, is actually that error in the velocity, but that's not true, it just a velocity this is an error
in it is not the error in the velocity, but error in position due a ramp input or a due to due to
velocity input, for example, I am you know maybe chasing a car here right, is say a car C 1 and
then this is my reference car and then I am just I just want to meet it right and at C 2 this is my
car.

So, as time goes to infinity. So, we are always at say a distance of say 1 meter from this right.
So, this is just error in the position, we are not having an error in the velocity even though we
are just we could just be moving at constant velocities right. So, this could be some velocity
V 1, this will also be the same velocity V 1, but I will just have an error in the position. So,
that's it should not be confused with error in the velocity. Similarly, if I were to say well give
me a temperature of 24 degrees I just go to 23 the steady-state error here is 1 right. So, I am
just tracking a position right I just want to go to 24.

But however, if I say keep on increasing my temperature profile with time in this way. So, say
this is it increases let's say 1 times t ok. This is my t and this is my temperature, the reference
temperature right, with time it just keeps on increasing. So, if I say that the velocity is you
know there is a velocity error I will start from here and I will just be here. So, if this has to
increase constantly at t, well I will always be behind by say 1°. So, at say at some t=50, I
might just be at 49 at t=100, I might just be at 99. So, it just a little difference in the position
not really the rate of my car here or the rate of change of temperature here.

(Refer Slide Time: 32:50)

247
 1
Similarly, if I have a parabolic input for which the Laplace transform looks like I can do all
s3

1
the math and I say that the steady-state error is , with K a defined in this way. And this guy
Ka
is called the acceleration error constant and I just you know this has an extra s over here
right. So, the error constants, these 3 things and these are the standard test signals which we
also used in the earlier lecture to study the response of systems.

So, the error constants K p, K v and K a they describe the ability of a system to reduce or
eliminate steady-state errors. For example, if I will come to that right. So, and then these
values depend mostly on the type of the system. While this type is important and as a type
becomes higher more the steady-state errors are eliminated. We'll just define this type in the
next slide.

(Refer Slide Time: 33:48)

248
(Refer Slide Time: 33:53)

So, I will come back to this slide a little later, but just the type of a system is defined here by
the number of poles at the origin. So if I have a system in the pole zero, configuration G(s)
could be typically some gain with a set of zeros with the set of poles and additionally some
set of poles at the origin.

249
So, if there are n poles, I will call it a type n system, if n is 0, I will call it a type 0 system and
so on right. And we'll see how does this type actually now influences the way I compute the
steady-state error. So, that is a type of system. So, the steady-state error is a measure of how
accurate my system is right, it does it really track my reference accurately or is there. So, the
accuracy should. So, in an ideal scenario, I would always want it to track it to its desired
value right.

However, there could be situations where it may not be possible to do so; in that case, I would
like the error to be as small as possible. And that is a very important performance parameter
right I don't want the desired temperature of 24 and end up at 12 or 31 for example, right. So,
this steady-state error now depends on 2 factors. And it is the nature of the signal which I am
tracking and also what is the nature of the physical system or the type of the system right.
And of course, all these things we'll only calculate for closed-loop systems. Of course, I will
calculate for closed-loop systems only based on the open-loop information I don't really need
to see what is happening within the loop and so on.

(Refer Slide Time: 35:37)

So, we'll see via how we'll do that. So, start for a type 0 system this G(s) is again K ' set of
zeros and there are no poles at origin right. So, n in the previous slide, this guy would be 0.
So, I'll just have these poles.

250
 1
So, how does it respond to a step signal well, for a step signal I have this guy, . So, that
1+ K p
there will always be some constant, some constant error right at the steady-state right. It can
track position signal, for example, this is my position signal; this guy can track well possibly I
would take it as the response, but it could be something like this and maybe it will go, go, go
and then there will be some error right. So, there will be some steady-state error here when
you're just tracking this step.

Now, this system can never track a velocity signal. Why? because the steady-state error is

1
computed in this way lim , this is infinity. For steady-state it will never be able to track
s→ 0 sG( s)
a signal which is like this and the error will keep on increasing right. So, this, this guy as we
keep on moving further the error will keep on increasing. Similarly, here it will also never be
able to track a parabolic input. So, the acceleration error will also be infinity.

So, to summarize a type 0 system has a constant position error and infinite velocity and
acceleration errors all at a steady-state. So, all these are related to this steady-state behavior.
At the moment, I am not worried about what is a peak overshoot? what is a rise time and so
on? I am only interested in what or how my system behaves at the steady-state?

(Refer Slide Time: 37:31)

251
 1
Type 1 system, well I just do all the computations again I have 1 my lim and since
s→ 0 1+G( s)
this has a pole at 0 this guy will go to infinity right the K p.

Let's again go back to see how we had defined K p. K p was defined at lim
s→ 0
G( s). So this guy

since s has a pole at the origin this will go to infinity and the error will go to 0, similarly the

1
velocity error would be lim . So, this s and this s would cancel out and I'll have a
s→ 0 sG( s)
constant here. And the acceleration error will go to infinity. So, if I take a type 1 system and I
ask it to follow a step input, then it will at steady-state there will be no error it will actually
follow this input very nicely.

However, if it takes the same system I ask it to follow a ramp it might have some steady-state
error right, I just go their response could be something like this and eventually go to a
constant error here like a straight line. Similarly, acceleration it will never be able to track a
signal which is parabolic. So, it will have a zero position error, a constant velocity error, and
infinite acceleration error at steady state.

(Refer Slide Time: 39:05)

252
Now, type 2 systems right. So, nothing will change here, the position error will still be 0

1 1
because I have this s in the denominator I put lim , G(s) goes to infinity and goes
s→ 0 1+G( s) ∞
to 0. Similarly, the velocity error. So, I do sG( s), one s still remains in my denominator. So,

1
this will still be and the velocity error goes to 0; however, my acceleration error constant
∞
now has some number here it was infinity in the previous 2 cases. So, I could summarize by
saying that if I take a type 2 system, I ask you to track position or a step it will track it
perfectly right. There will be no steady-state error, if I ask you to track a ramp it will also do
it perfectly. If I ask you to track a parabolic signal it will do it, but there will actually be a
small error determined by this acceleration error constant.

Now, we see that you know if I say give a little you know problem saying take a type 0
system and make the position error to be 0. What would I do? I'll just say is this look similar
to all this theory now and I say see just put an integrator right or add a pole at the origin. I can
always do this is a system which has more poles than zeros I can possibly always realize this,
and I can just put it here right. So, this is; can this be done? can this not be done? or can this
be done with some being a little careful? we'll see all those things, but these are the design
specifications which would be given to us and then you know we can see that these errors can
be quantified just based on the type of signals which we are tracking and the type of the
system.

So, having an integrator in my system or having a pole at the origin gives a little more hope to
me in designing the system.

(Refer Slide Time: 40:55)

253
So, in this entire module what we started with block diagram representation, signal flow
graphs and then reducing the complexity of finding the transfer function by reducing the
complexity of the block diagram or simply using the signal flow graph formula to find the
transfer function, then we also saw the time response of first and second-order systems, with
standard inputs and we also justified why we were using the standard test inputs because in
real-time we can see all these signals the real-time signals could be a combination of all these
test signals. And then we actually quantified our performance in terms of the steady-state
error in terms of how much do I overshoot what is a rise time and so on and we had explicit
formulas for those right.

(Refer Slide Time: 41:40)

254
And in the next lecture, we'll and so far what we assumed that the system is stable and we had
a very crude version of the definition of stability saying if it starts with 0 initial conditions it
should come back to it is original configuration. So, we'll have notions of define these notions
formally of stable and unstable systems. And see how can we define these things directly by
the transfer function looking at the poles and zeros, just by looking at the poles and zeros can
is say this system is stable or not. Or another notion is where you know what is called as the
bounded input- bounded output stability and then given a transfer function of fairly you know
the complex level, which has lots of poles and zeros. I will just teach you how we use the
Routh Hurwitz criterion for the stability of to determine the systems as stable or not.

Thank you.

255
 Control Engineering
Dr. Ramkrishna Pasumarthy
Department of Electrical Engineering
Indian Institute of Technology, Madras

Module - 03
Tutorial 2
Lecture - 14
Solving Problems on Time Response Analysis and Specifications

So, we'll just do some problems related to the time response and its specifications. So, I'll
start with a fairly simple looking problem.

(Refer Slide Time: 00:29)

So, it is to find the step response of the system. So, I have a reference which is a unit

25
step. there is a 1 here and this is my Y (s). So, there could be two ways to do this
s (s+4)
right, one is right applying the formula for y (t) and if we remember this essentially was
based on ω nand 𝛇. So okay, first let's see how the closed-loop looks like the closed-loop

2
Y (s) 25 ωn
transfer function here = 2 and this is of the form 2 right.
R(s) S +4 S+25 S +2ζ ωn S+ωn 2
Therefore, I can easily compute ω n=5 rad / sec and 2 ζ ω n=4 and therefore, 𝛇 can be
easily computed to be as 0.4 and this will be radians per second.

256
Ok, right. And then we can clearly see that this is an underdamped response or an
underdamped system and I know that the response just like this

−ζ ωn t
e
sin ( ωd t+cos ζ ) which I had this I had termed this as θ in my earlier
−1
y ( t ) =1− 2
√ 1−ζ
classes right. And then once I know so I know the ζ , I know the ω n, I could compute

ω d =ωn √ 1−ζ 2 which turns out to be 4.58 rad /sec and I do all the computations and I just
end up with this expression
1−0.44 ¿, that I'll just skip the intermediate step. But, it should be easy to compute it this
way.

(Refer Slide Time: 04:35)

25 1
Another way to do it, that I know Y (s)= 2 . . And I could just write this
S + 4 S+25 S
down in this way. And so this would be 1 the inverse Laplace transform y (t) would be

−2t
e
sin ( √ 21t) okay and this is equal to this. Because I
−1 −2t
L ( Y ( s ) )=1−e cos ( √ 21t ) −2
√ 21
can compute directly by the formula or using the inverse Laplace transform right based
on whichever is comfortable.

257
(Refer Slide Time: 06:20)

25
So, the next problem I am given G ( s )= , H ( s ) =1 and I have a negative feedback.
s(s+5)
Now, here I have to find out what is ω n, 𝛇, ω d, rise time, percentage overshoot, and

Y (S) G(S)
settling time? So, the transfer function of = that would be
R(S) 1+G ( S ) H (S)

2
25 ωn
2 . So, I just compare this with the general form of 2 gives me
S +5 S+25 S +2ζ ωn S+ωn 2
ω n=5 & 2 ζ ω n=5 and therefore,ζ =0.5 okay.

(Refer Slide Time: 09:06)

258
So I know ω n, I know 𝛇, I know ω d =ωn √ 1−ζ 2 and that will be 4.33 rad /sec and I have

π −θ −1
π −cos ζ
the rise time(t r ) computed as and θ=cos−1 ζ , therefore and this is
ωd ωd

π
0.48 sec. The peak time(t p), we computed this as . So, ω d is this, so over here and that
ωd
will just be 0.725 sec okay, or and then the percentage overshoot was dependently
2
directly depending on the damping that was given by e−πζ / √ 1−ζ .100to get the percentage.
And I could compute this as 16.36 % and finally, the settling time for the 2 % tolerance
criterion is 1.6 sec and quite straight forward right. Once you write down the close loop
and compute what is ω n and 𝛇 then everything else is for fits into the formula right.

(Refer Slide Time: 10:02)

259
 10 4
The third problem is well I am given a system like this , , a plus, a minus this,
S+2 S+8
this is 1, Y (S). In the problem is to find the steady-state error for a unit step input okay,
if I had done this three or four lectures ago, I would have done you know used to

Y (S) 40
calculate first the transfer function = , where 40 is G ¿) over
R(S) ( S+2 ) ( S+8 )+ 40

40
1+G( s) H ( s) that’s computed as shown above, which is 2 . So, two weeks
S +10 S+56

40
Y (S)
ago I would just do this right. So, S and then I would do, this is my
2 R(S)
S +10 S+56
transfer function and then I would compute what is the final value, I am not really
interested in finding the entire signal in with time, but I'm only interested in finding this
error, steady-state error.

So, what is the final value of Y (s)? So, from the final value theorem, a final value says

lim SY (S) for a step input this would be lim SY (S). So, Y (S), sorry this would just go
S→0 S→0

40 40 40
right. So, Y (S)would be 2 . So, I have 2 and this would be
S +10 S+56 S +10 S+56 56

260
which is 0.714 and the steady-state error is the actual signal r (t) or the actual reference
minus the actual output. So, is 1−0.714 is 0.286, 0.286.

But since, these are a slightly longer process, but now I know directly how to calculate
the steady-state error given the type of the system right. So, I have if I look at this, this
was called the G(S) it is a type 0 system and I am tracking a unit step and I know that
there will be some position error, of course, this tells me already that. So, how would I
compute that?

(Refer Slide Time: 14:29)

1
So, E ss was and this is for a type 0 system, with a step input and then
1+ K p

K p =lim G( S)right that was what is G(S)? So, is guy 40, so G(S) is this one right. So,
S →0

10 .4
G(S) and I'm just looking at the open-loop transform function, is .
(S+2)(S+8)

So, lim G( S) was 40 =2.5. So, this is my K p. So, the steady-state error is 1 is
S→0 16 1+ K p

1 and that can be easily computed to be

0.286 . So, the advantage here is I don't
1+2.5
really need to compute the closed-loop transfer function. I can just look at the loop

261
transfer function or the open-loop transfer function and then compute these things
directly. So, again this is the lead to the same thing not surprising.

(Refer Slide Time: 16:09)

Okay so, in the last problem. So, I am given a system which looks like this R(S), E(S), I
have given K , I have a certain disturbance and then I have a plant which looks like this.

1 S+2
, where h is any number or any integer value. I have right this.
S h
(S+1)(S2 +10 S+40)
So, this is my entire plant as it looks and negative feedback, as usual, a plus here a minus
here which is a1 here. So, this is my D(S) this is my K and so on and this is my Y (s).
So, in the first part, I would want to find hsuch that for suitable values of this gain K ,
additionally let me assume that the disturbance is identically equal to 0. So, the h should
be such that the steady-state error for step reference or step input is 0 and the steady-state
error again for a ramp reference is a non-zero constant, which means we are just
interested in finding the type of the system.

Okay so, let's recollect quickly. So, if I have a type 0 system and the input being a step

1
my E ss = this was again for h=0. So, this step input is what I am looking at that
1+ K p
the steady-state error for a step reference is 0. Now when h=0I would have some steady-

262
state error this is not 0. So, I should at least haveh to be 1,h to be 1. So, that I have a type
one system and then a step input, then E ss goes to 0 that we know. Now is that the
answer? well if just this was the statement I could say that h could be 1 or even moreh
could also be 2, h could be 2 and this statement would still be satisfied, that the steady-
state error would be 0 for type 1, type 2, type 3 and so on. However, the second statement
says that in addition to the steady-state error for a step input to be 0 it should also have a
non-zero constant error for a ramp input or for a velocity input.

So, what was the nature of these things? So, if I have a type 1 system and the input is a
ramp. So, I know that a type 0 system will not be able to track a ramp and the error is
infinity and h=0 is ruled out anyways right. So, I'll start with h=1. Type 1 system, a

1
ramp input, I know that there is a constant error given by I think or something and
Kv
we know how to how to compute that K v . So, the answer would be that h well it, of
course, cannot be 0 it can be 1 or greater because of the second statement. And the third
statement restricts it only to be 1, because if I have a type 2 system then the error is not a
non-zero constant, it will go to 0. Right at type 2 if I have a type 2 system and the ramp
input my E ss goes to 0 right. So, this is not allowed. So, I'm just a type 1 system and
therefore, the answer is h should be equal to 1.

So in the second part, we will do something a little more interesting.

(Refer Slide Time: 22:07)

263
So, let me call this block as G ' (S) and the second part what I'm I want to find a K such
that when this disturbance signald (t) is a step and the reference r (t)=0 and I want to

find the tlim

→∞
y ( t )=0.01 right. So, I want to find a gain K such that when there is no

reference and a disturbance which is a step that y (t) should go to 0.01. Now after this for
this value ofK , for this value of K find the steady-state error for unit ramp and d (t)=0
right okay. So, let's do the first part until here. Find a K which under a step disturbance
and no input will give me an output at a steady state at 0.01. So, this my

'
Y (S) G (S) and we saw how to derive this transfer function in the previous
¿R ( S )=0 = '
D(S) 1+ K G (S)
tutorial.

So, now, what I would want is D(S) is a step right. So,

1 (S+2)
Y ( S )= . 2 right this entire thing here expands to this one,
S s ( S+1 ) ( S +10 S+ 40 ) + K (S+2)

straight forward computation. Now y ∞ is again I apply the final value theorem lim
S→0
SY (S)

2
and I do all the math and I get this is right the 2 and everything else will disappear
2K
and this y ∞ should be equal to 0.01 right. So, the question was to find a gain K which for

264
D being a unit step results in a steady-state y (0.01). So, I just solve this and I get at the
value of K=100.

Now in the second part, assume there is no disturbance I take this value of K=100. I will
also, this was the first part so I am not even interested in what is the value of hat the
moment, for this so this 100 was computed. So, we did h=1 right, we keep the value of h
from the previous one and then compute the value of the gainK such that the output is
0.01 and thatK was computed to be 100. So, I know what is the h now, I also know what
is the K .

(Refer Slide Time: 26:39)

So, for these 2 values of h and K , I want to find the steady-state error for a unit ramp and
assuming d (t)=0. So, this is again a type 1 system and I have an input which is a unit
ramp and I know that the error is definitely not equal to 0 and its some constant value.

1
So, let's compute what is that constant value. So, the steady-state error is and K v is
Kv

the velocity error constant. And what was K v ? K v =lim

S →0
SG (S)getting into the math

lim SG (S). So, I'll take this entire G now, now this should be S . 100(S+2)
2 S
S→0 S( S+1)(S +10 S+ 40)
times K is 100 in the numerator I have S plus 2 denominator h equal to 1. So, I have one
pole at the origin okay. So, I can do all the computations. So, this guy goes away and in

265
 200
the numerator, I am left with and this is my K v right, this is 5 and E ss therefore, is
40

1
=0.2. That's just straight forward using the formulas, the only trick is to how to get
Kv
this transfer function from D to Y putting R=0, finding the appropriate value of h and
then finding K , again by using the final value theorem and then using just the formulas
for steady-state error given that the system is type 1 and the input is a unit ramp. Yeah.

Student: That is just Y (S)

This is just Y (S)

Y (S)
Oh sorry, D(S) will go away. Or I'll just do this one right this is is so there will be
D(S)
no Dhere and that D will be taken care of here and all this we assume here that that the
R=0 right.

Thank you.

266
 Control Engineering
Dr. Ramkrishna Pasumarthy
Department of Electrical Engineering
Indian Institute of Technology, Madras

Module – 04
Lecture – 01
Stability

Hello everybody. So, far in the lectures or in the previous module what we saw was a
typical response of a first-order system to various test inputs. We also justified why we
had those typical test inputs like an impulse, a step, a ramp and so on. We also saw the
notion of steady-state error while were just introducing the response of first-order
systems. We even looked at the response of second-order systems and we saw there are
lots of other things to take care of while designing a second-order system like the peak
overshoot, like the rise time and then you had the notions of settling time and so on. And
these were inherently related to the system's damping and also the natural frequency.

Later on, we characterized steady-state errors of the system, and we saw how the addition
of an integrator or a pole at the origin helps in achieving a good steady-state behavior
right. Only in terms of error right, not necessarily with reducing the settling time. And in
all these analyses what we had assumed was that the system was inherently stable. And
for that analysis purpose, we had just defined stability as a property where if the system
is subject to some initial condition, it will regain its equilibrium position or its original
configuration.

So, what we will do in this lecture is to first formalize the notion of stability of a system.
Because before we even try to go for an analysis of steady-state or even the peak
overshoot, what is important before any design is carried out on the system is that the
system should be stable.

(Refer Slide Time: 02:04)

267
Now, what is this stability? So, this is a very important property before we design a
control system right. So, let's see a couple of videos to just get a little idea of what this
stability business is all about. And of course, so these are videos from these websites. So,
there is a guy driving and oops. So, nothing is left for him right and you have this aircraft
which. So, by the way, this is, not an animation video this is something which could
happen in the real-time right, you have this big turbulence and then your aircraft behaves
in a very abnormal way. So, these are not just the two definitions or these are not just the
two pictures which define instability right.

Okay so, how is this characterized by the system? So, by the system, we had so far said
that the entire course would be on linear time-invariant systems. Typically, with a single
input and single output and what we had as an information of the system was essentially
the transfer function. Now we will see how the transfer function gives us some
information on the stability of the system.

(Refer Slide Time: 03:21)

268
So, the first definition would be, a system is stable if the system comes back to its
equilibrium state, sometimes even called the zero state when the system is subject to an
initial condition. And again the simple example which we saw so far, is the simple
pendulum right with certain damping. So, if the pendulum is at the equilibrium position it
will always remain at the equilibrium position.

However, if I start at some position say which is somewhere here say, theta equal to
some say 20° or whatever then I just. So, this is my initial θ. I just leave it, it will just
oscillate and then come back to its position which is θ=0 right. So, this is stable
behavior when subject to an initial condition of θ=20, the system regains its equilibrium
configuration the same thing happens when θ is say somewhere here or here or here and
so on.

Now, what is the meaning of the unstable or what is the definition of unstable? The
system is unstable if the output diverges without bound from its equilibrium state when
the system is subject to an initial condition. So, let's do the reverse of the pendulum, right
it's here. So, I have this guy. So, this is my pendulum, not a very good drawing, but as
long as it's in the upward position it will always remain in the upward position right and
that's what physics tells us, once you are in the equilibrium you are always in the
equilibrium position.

Now, I say I do a very small perturbation or subject it to initial conditions say theta equal
to 1 degree, for example, right then what would you expect is that the system does not go

269
back to its original configuration, it's just open-loop system there is no control nothing.
So, what it does it just goes down right. And then it's energy increases quite a bit right.
So, most systems it does not really go out of bound because of this physical limitation
here, Right. But mathematically we'll say well the system is kind of going out of bound.

So, this system with what we call as the inverted pendulum system is unstable in a way
there. So, if this is starting with this notion θ=180 ° this is also an equilibrium position,
but this is unstable behavior. And the system is marginally stable if it tends to oscillate
about its equilibrium state, when subject to an initial condition. Or again consider this
simple pendulum without damping. Now, say θ=20 ° right. So, what will you expect it
will go to say −20 here and then come back again just keep on oscillating like forever.

So, here the output or the system does not really go out of bound. It just remains within
this nice bound of θ=20. If I start at θ=40 it will just keep oscillating between +40 and
−40 and so on. So, these are systems which are called marginally stable.

(Refer Slide Time: 06:24)

So, stability also depends on the nature of input and how my system behaves to that
certain input. So, what we in input-output notion stability is defined also as small
changes in the input should not lead to very large changes in output. Small changes in
input could lead to small changes in output that is acceptable.

270
For example, here if I small change in initial condition could be well I am going from 20
to 40, but then my output doesn't really go like you know the behavior doesn't go to like
say 400° or something. Even here if I start from an initial condition of22 well the output
doesn't change really much right. It's still very much within the desired behavior right.
So, how do these bounded signals look like well? So, this is one the exponentially
decaying signal is an example of a bounded signal. So, this is kind of bounded in values.
So, you all can also have a constant input this guy is also bounded. A sinusoid is also a
bounded signal right. So, it doesn't really go outside and then unwanted signals are one
which just increases very large with time.

(Refer Slide Time: 07:26)

Let's say these guys just keep on increasing, similarly these oscillations right, they just
the amplitude just goes on increasing with time and these are essentially unbounded
signals. So, so they keep on attaining very large values for large values of the time.

(Refer Slide Time: 07:49)

271
So, in the case of input-output what we call as bounded input bounded output stability or
in short BIBO stability, a system is stable if for every bounded input signal, the response
should be bounded right say for example, like have a bounded input here the response
should just be bounded.

The response goes like something like this then it is unbounded. So, how will we define
unstable systems in input-output settings? So, if I find any one signal which is bounded
for which the output goes unstable, this or output goes unbounded the system is unstable.
There could be say 25 inputs for which the system behaves like this, but for the 26th
bounded input if it goes if the output goes unbounded that is the system I would still call
unstable. So, the system is unstable if, for any bounded input signal, the response goes
unbounded.

272
(Refer Slide Time: 08:42)

So, this is how it looks. So, I have bounded input I pass it through an unstable system and
I get a response which is unbounded.
So, there is also a notion of exponential stability in linear systems. So, what does this
mean that in the absence of input or if the input is 0, if the system response
asymptotically decays to 0. Again within an envelope then the system is called
exponentially stable. For example, a response of say e−t is an exponential stable system.
So, we will not go into details of this, but just to make a side note that if I take a linear
time-invariant system, and if I say that the system is exponentially stable, and then it's
also BIBO stable.
So, just the definitions of stability what have we, how do we how did we define stability?
First is that the system, if it is subject to a non-zero initial condition will eventually go
back to its original configuration or the equilibrium position. The second notion in terms
of input and output was if the system is subject to small changes in input it will result in
small changes in output. Or more formally if the input is bounded the output should be
bounded and if the input is bounded, output goes unbounded I call it an unstable system.

273
(Refer Slide Time: 10:04)

So, let's see what that means in terms of what we have learnt so far in terms of the
transfer functions or in terms of the block diagrams. So, let's take the standard you know
I have a plant system here, I was certain input or a certain reference generating a certain
output C (s) right and then the output we know in the Laplacian domain is given by
C (s)=G(s) R(s). Or in the time domain, we use the convolution property that

c (t )=∫ g ( τ ) r ( t −τ ) dτ .
0

(Refer Slide Time: 10:36)

274
What do I know here, is that the input or the reference signal is always bounded right?
So, this is a kind of necessary right. So, the input is bounded and then the magnitude of
the output. So, so I could just say I just say. So, this c (t ) if I just look at the absolute

value of it, c (t )=∫ g ( τ ) r ( t −τ ) dτ and I just take the absolute value to the inside I get this
0

∞
one, ∫|g ( τ ) r (t −τ )|dτ.
0

Now, this r I know is bounded and let me assume that it's bounded by some number M
and this absolute value of this output c (t ) then takes this form. Now, I ask the question
r (t) is bounded when is c (t ) bounded? Well, the answer has got to do with this guy now,
because I already know thatr is bounded by M . So the answer of when does a bounded
input result in a bounded output a lies in here in this expression here right. So, the output

∞
is bounded when this guy is also finite or when ∫|g ( τ )|dτ < ∞ or its boundary.
0

So, and this guy is essentially the impulse response of the plant right. So, the system is
BIBO stable if and only if the impulse response is absolutely integrable that the limit of
this integral exists or it's a finite value otherwise it goes unstable.

275
(Refer Slide Time: 12:16)

Now, in terms of the transfer function we know that the transfer function is represented
as a ratio of two polynomials in s, you can also write this in terms of the pole zero-
configuration right. So, how did we define the poles? The poles of the system are the
roots of the denominator polynomial, for which now D s I equate D s =0 and then find out
what are the roots.

Similarly, for the zeroes, I equate N s =0 and find out the values of h for which this
equation holds true right. So, this system here has ‘n’ poles and ‘m’ zeroes. And of
course, the number of poles will always be greater than or equal to the number of zeroes
is what we will follow throughout.

(Refer Slide Time: 13:08)

276
Let's see with the help of an example, I have a system which is like this. So, I just look at
the numerator equate it to 0 and I just get two zeroes right, Z=−3and Z=−4. And there
are three poles one at the origin (i.e., p1=0 ) and I have a complex conjugate pair
p2, 3=−1 ± j 2.

(Refer Slide Time: 13:30)

277
So, if I look at go to thes plane, with σ sis typically like this guy σ ± jω right. So, it has
real values and imaginary values. So, a pole at the origin would be. So, this guy entirely
goes to 0. So, I am here. So, the pole is usually denoted by a cross. So, I have a pole here
complex conjugate poles at −1+ j 2, −1− j 2and two zeroes one at −3, one at −4 and
zeros are this is typical. So, this is the notion for 0.

(Refer Slide Time: 14:07)

Now, we will see what is what to do with stability? What has these locations here 1 2 3 4
5 guys, do they give us any information on stability right. So, that is what we will try to

278
investigate; so again the roots. So, the denominator polynomial D S =0 is called the
characteristic equation of the system.

So, we'll do a little bit more of this later, but at the moment you can just assume that I'm
just interested in the solutions of the denominator equation, D S =0 . And then the roots
are called the poles and if I assume for simplicity that the poles of the system are distinct.
I can rewrite this from my Laplacian theory as like this side, so ( s−s1 )( s−s2 ) and so on.
And these poles can be real, imaginary or complex just that they are not repeated we will
come to the repeated case a little later.

(Refer Slide Time: 15:06)

First: let's take some partial fraction expansion of like this. And say well I am just
interested in this guy say one guy, say the pole is real, say Sk =a then the pole lies in the
left half-plane for a=0 right. So, it will be somewhere here, here. On this negative line,
now the inverse Laplace of this guy we know is A k eat u ( t ). And therefore, when a<0,
the pole lies somewhere here and denoted by a cross. And then the impulse response I
know is something like this right; it just lies down to 0, after well asymptotically it goes
0 and this actually a bounded output response. And hence the system I would call is
stable for a<0.

279
 A
So, my first observation I could make is if I have a transfer function of the form ,
s+ a
and alies in the left half-plane. Then my impulse response of the system is bounded and
therefore, by the first definition I could say that the system is stable.

(Refer Slide Time: 16:31)

Now, similarly when a>0, when a is greater than 0 my pole is sitting over here. And the
impulse response is some constant A k eat , a>0 therefore, I would expect this guy to keep
on increasing right. So, this is an unbounded response. If t goes very large this guy
assumes very large values. And therefore, I can say that the system if the pole is sitting
on the right side then my system is unstable right second observation.

Now, third observation what if my pole is just sitting here. When a=0 the pole lies at the
origin. So, this guy put a cross here. And the impulse response is constant with time
right. So, this the output does not increase with time, but; however, if I just integrate for

∞
this case g(t), ∫ g(τ )dτ then I see that the area. So, g this is essentially the area, that this
0

area is not bounded right the area is not bounded, but my signal doesn't really increase
like it increases here right. So, this system the simple integrator or with the pole at the
origin is not absolutely integrable and we can just call this as some kind of a marginally
stable system.

280
So, now we saw three things what about pole on the left, a pole on the right, and the pole
at the origin and the okay, nothing to do with zeroes so far.

(Refer Slide Time: 18:09)

What happens if my pole is now complex conjugate? Well and say there could be two
possibilities that the poles could. Say, this is my σ , this is jω the poles could either be
here or be here. So, if a<0 then these are my poles right just for a<0. And the poles are
on the left half-plane not just on the left of the real line, but on the left half-plane this
entire thing. Now if I just do all these partial fraction thing then I get that the response of
the system is given by A e at cos(bt+ φ).

Now, this is a general expression no matter what the sign of a is. Now for a<0 the pole
lies the poles lie here. And the impulse response since this is e−at is decaying. It decays
exponentially to 0 indicating that is also absolutely integrable.

281
(Refer Slide Time: 19:13)

So, it will be something like this right, this is integrable just see the area is finite. The
response is unbounded. And therefore, this system is BIBO stable. And of course, it's
also stable by both the definitions for a<0, for a>0 the poles are somewhere here on the
right half-plane and the impulse response increases without any bound.

So, it just does this. So, you see that the oscillations keep on increasing in amplitude right
and this is an unbounded behavior.

(Refer Slide Time: 19:47)

282
So, what if a=0 again right if the pole sk is imaginary and say I just have here, so one
+ jω and − jω. Then the response is simply of the form A cos(bt + φ). So, the response I
know just the response of a sinusoid, it's just like this right it just has sustained
oscillations, and therefore I call this system to also be marginally stable right. It neither
increases unbounded in value nor do the oscillations died down to 0 as in the previous
two cases. So, this response or this oscillatory behavior is what I would call as
marginally stable systems.

So, for what we have seen is if there is a distinct pole like s=a, if a<0 it is stable, a>0
it is unstable, a=0 has some kind of a tricky response that the output grows unbounded,
but the system is not absolutely integrable, we could call this as a marginally stable
system. If I have complex poles I am just interested in what does the real part do. If the
real part is less than 0 then the system is stable, the real part is greater than 0 we see an
unbounded response, and if the real part is 0 which means my poles are just on the
imaginary axis then I just see an oscillatory response.

Refer Slide Time: 21:14)

283
Let's see what if now we have repeated roots, say I have say a couple of roots ata, at if I
just have to draw it here and this is my −a or +a right here. So, my response would be
some constanta, plus some other constant times t eat and they again this depends on what
is the value of a. If a<0 then the impulse response is stable and if a>0 then the impulse
response goes unstable and that could be computed quite easily. Similarly, if there are
repeated complex poles say there are a couple of guys sitting here and a couple of guys at
the same location here, then again depending on the sign ofa my system is stable or
unstable. So, this region is again the stable region when a<0 and a>0 is always unstable.

So, the same thing holds right if the poles or the real part of the poles are less than0, I am
still strictly less than0 I am still stable. And if they are strictly greater than 0 then my
system is unstable. Now earlier we saw something about the root at the origin something

1
like or when the pole is at the origin then I saw that my response was just some straight
S
line right. It was bounded in value, but not integrable what if I have two of this. If I have
two of this my response would look something like this right. It will just keep on
increasing with time and that's because of this kind of thing right,
A k + A k+1 t +…, and if I just plot it with time this is my response it just goes unbounded
with time.

Now, let's take these guys further apart and say well I had a jω, I had a − jω and I saw
earlier that this response is just oscillatory. Now tricky things will happen when again I

284
have multiple things here, and multiple things here. Then my response is A 1, okay this is
the response when there is only one pair of poles which are imaginary. And if a second
guy comes and then I have this guy A2 t cos(bt+ φ2)and then t 2 and so on. So, these
responses will actually keep on increasing like this way and this also is an unbounded
response.

So, these two guys as long as a<0, a<0 has a stable behavior. Whereas, these two guys
have a strange kind of behavior right. If there is one pole at the origin, then I am in some
kind of a marginally stable behavior. If there are two poles at the origin, then I go just
unbounded, similarly for 3 4 and so on. Also here if there is one pair of the complex pole
on the imaginary axis, then I have some kind of a marginally stable or an oscillatory
behavior. Whereas, this guy makes it unstable if there is more than one pair of poles on
the imaginary axis.

(Refer Slide Time: 24:16)

So, just to summarize, what is the relationship of the stability with poles of the transfer
function. If all roots of the characteristic equation have negative strictly negative real
parts, then their impulse response is bounded and eventually decreases to 0. This also
means that this integral which I was trying to compute is finite and therefore, the system

285
is BIBO stable. If any root of the characteristic equation has a positive real part it could
just be 1 it could be 500 roots of the system even if one is on the right half-plane, then
my system is unstable as g t is unbounded and this guy this integral is not defined.

Thus, the system will become an unstable system lastly if next if the characteristic
equation has repeated roots on the imaginary axis right. So, again the same thing it could
result in an unstable system. So, lastly if one or more non-repeated roots of the
characteristic equation are on the imaginary axis then g(t) is bounded, but the integral
would not be bounded, so that's as we saw here this integral here is unbounded this one
root at.

However, if the input signal has a common pole on the imaginary axis, then c (t )
becomes unbounded or the output becomes unbounded. And there are also cases where
in the absence of any common pole is output is bounded and has sustained oscillations
like these two cases, here I have sustained oscillations or here is just it's a straight line
and this kind of systems are typically referred to as a marginally stable system.

So, we have categorized stable, unstable and marginally stable systems. So, on the left of
the plane it's pretty clear on the right of the plane it's pretty clear. On the imaginary axis
if there is just one pair of the complex pole then there is some kind of marginal stability,
but if there are repeated roots, then I go to an unstable level.

(Refer Slide Time: 26:21)

286
To just pictorially say what I have said. So, this is a stable region left of plane unstable
and this is the imaginary thing right, where it could just be marginally stable again
depending on how many roots or how many repeated roots are there.

(Refer Slide Time: 26:35)

So, what have we learnt so far is to define the concept of stability first from a system
response to a non-zero initial condition. We also talk defined bounded signals and
studied or defined what is called the bounded input bounded output stability, in which
case we said if the input is bounded and the output is bounded then the system is stable,
or if there is any one bounded input which results in an unbounded output the system is
unstable.

However, what we need to be careful is what if the input is unbounded and the output is
also unbounded, can I say something about the stability well the answer is not is kind of
no. Because if I am tracking a ramp as we saw in some of our previous examples when
we are doing the steady-state error, my input is unbounded right, it just t which just
increases with time and the output is also expected to increase due to time because I am
tracking this signal back, this is an unbounded input and an unbounded output.

So, there I just cannot categorize saying that input is unbounded output is unbounded and
therefore, the system is unstable no we have to be a little more careful with that. So, what
we did next was to relate stability to poles and zeroes, but typically to the poles and we

287
said depending on the location of the poles, I can judge or I can say something about the
stability or not of the system.

Now, next we will see is if I'm given polynomial of say some order 10 or 15, how do I
compute its roots, or how do I at least say where the roots lie. I don't at some point of
time I need to explicitly compute where the roots are, as long as I know that they are
somewhere in the left half I am happy. Or just say there is one guy in the right half-plane
I don't really know where he is, need to know where he is sitting right even if he's at plus
1 or plus 100 it's still unstable.

So, we will just try to learn some tricks call the Routh Hurwitz criterion which will tell us
about the stability of a given transfer function.

Thank you.

288
Control Engineering
Dr. Ramkrishna Pasumarthy
Department of Electrical Engineering
Indian Institute of Technology, Madras
Module – 04
Lecture – 02
Routh–Hurwitz Criterion

Hey guys. In the last lecture, what we had seen was to define properly the notion of stability and
relate it to the transfer function or how does transfer function have or does transfer function have
any information on the stability? And we characterize by saying that if all my poles are strictly
on the left half of the S plane then my systems are stable right. And then we also had some
special characterizations when there are repeated complex conjugate roots, how the system goes
to unstable behavior. And even when you just have a single pair of complex conjugate poles
where the system was defined to be marginally stable right, and all these involved computing
roots of some higher-order polynomial could be 6, 7, 8, depending on how many poles there
exist in the system.
So, today we will try to find an algebraic way to see if there are better or maybe more efficient
tests of finding out the stability of the system. We may not be directly interested in the exact
location of the poles, but an answer to whether the system is stable or not would be enough for
the initial analysis purpose. So, what we will do today is to just look at a polynomial of a fairly
higher-order for which we cannot just compute roots by hand. What are the methods which were
which evolved you know historically over time?
So, this method is typically referred to as the Routh-Hurwitz criterion.

So, we are again interested in the transfer function analysis where I have on the numerator a
polynomial of m and n, and then n is usually typically greater than or equal to m right. And then
the characteristic equation which essentially what happens in the denominator when that is
equated to 0, this is referred to as the characteristic equation of the system, no matter if your
system is open-loop; if you are doing an open-loop analysis or even a closed-loop analysis. We
will formalize this characteristic equation shortly, but at the moment you are just interested in the

289
denominator of a given transfer function for which we are interested to perform this stability
analysis.
So, the definition of stability leads us to a test which says or which asks suppose this was the
question are there any roots of the system on the right half-plane? if the answer is yes then my
system is unstable. Okay so, let’s do some steps here. So, q of s is a polynomial here which I can
write it in this way as the product of. So, this will have some roots let me call those roots are p 1
to p n and these roots are the poles of my system. So, if I expand this expression in terms of the
poles, the expression would look something like this right again in decreasing powers of n until
decreasing powers of s until you reach n equal to 0, just to simplify this.

So, what happens when there are three poles right when this expression q of s, q of s let’s even
assume that a 3 would be equal to 1, then I have s minus p 1 s minus p 2 s minus p 3. And this
would expand to the following you have s cube and s square, so you will have sum of all the
roots. That’s what it says is p 1 plus p 2 plus p 3 plus the term in s which is p 1 p 2 p 2 p 3 p 3 p
1 minus the product of all these, like here n equal to 3, so I’ll have a minus p 1 p 2 p 3 is 0. Okay
so, stability says that all the poles should be on the left half-plane. So, let’s assume just say that
you know to solve p 1, p 2, p 3 are less than 0. Let’s just for simplicity assume they are real even
though if they are complex nothing much would change also.
So, if p 1, p 2, p 3 are less than 0 then this entire number would also be negative and the
coefficient of s square there needs to be positive, right so here. Similarly, if p 1 and p 2 both are
negative p 2 p 3 are negative p 3 p 1 is negative all this would; the products would be positive
numbers. Similarly, if p 1 is negative, p 2 is negative, p 3 is negative, their product would be a
negative number and this minus would make it positive. So, what does the first observations say
right, if I have a polynomial like this and if I want to observe or if I want to say something about
the location of poles in the stable region, the coefficients of this should be non-negative, that’s
one of the necessary conditions. It may or may not be sufficient that we’ll see a little later. And
all the coefficients they should well be; they should have the same sign possibly it could be a
minus, minus, minus, minus all the time that could that’s also allowed because you just equate to
0 and then all the minuses will become pluses.
Okay so, the first observation just by an expansion like this and keeping in in mind that stability
would mean the poles are on the left half-plane would lead us to say well, the necessary

290
condition is that all the coefficients are non-negative right. It could be a 0. What if in this thing,
if this is true all the coefficients would be positive right. Okay so, this follows just from the
general conditions of this polynomial this is a general expansion when you have it for degree n.

So, some straightforward computation here would show that this ratio or like this sum of poles
from p 1 to p n could be expressed as a ratio of a n minus 1 and a n, and similarly the products
and all. So, based on again these observations here we could also say that all the ratios should be
non-negative and of the same sign for the system to be stable and I’m just rewriting those
conditions in terms of polynomials and just dividing throughout by a n. And since our a n we
assumed to be 1 for simplicity I think this would be very straight forward to find out okay, but
these are not sufficient conditions these are necessary right.
For example, I take a system which follows this rule over here right the s cube plus s square plus
2 s plus 8 equal to 0. Well, this satisfies all these conditions that the coefficients are all positive.
Now, does it guarantee stability we’ll just you know compute the roots and the roots turn out to
be minus 2, this is a stable guy, and you have 0.5 plus minus j now this guy sits in the right half-
plane okay. So, therefore, the system is unstable even though all the coefficients are positive. So,
this the test that these things all the coefficients are non-negative is just a necessary condition, it
will not guarantee things right. However, if instead of a plus here I have a condition like say s
square s cube minus s square plus 2 s plus 8 equal to 0; it violates the condition that all this guy
should be non-negative this is definitely unstable, this is because I am violating the necessary
condition.

And then to have the roots computed or at least to check where the roots lie in the complex plane
Routh and Hurwitz, I will tell you a story of this a little later. They came up with a condition,
which is which was both necessary and sufficient for the stability of linear time-invariant
systems. So, how does this go we’ll not do the proof of it, but we’ll just learn the technique. So,
the method has two steps, first is I just generate something called a Routh array and interpret the
array in terms of locations of poles in the s-plane. So, we will read this statement once we see
how the array goes.

291
So, I have this characteristic equation. And I just write it into an array, I start with s power n, I
take the coefficient here, I put a power n and then I take s minus 1, I took this coefficient, I take
this coefficient put it here and then I do all the coefficients of the of n, n minus 2, n minus 4 and
so on until here, and then n minus 1, n minus 3 until I reach, until I reach the end of the
polynomial here. So, these two are given to me. How about the remaining rows? So, given s n, s
n minus 1, s n minus 2, so this number is b n minus 1 computed in the following way, that I take
a n minus 1 times a n minus 2 minus of this guy divided by a n minus 1 and just a simple
formula.
Similarly, I compute b n minus 2 as again I start with a n minus 1, I go to the next row here a n
minus 4, again a n, a n minus 5 divided by this guy. So, this is I think you could just look at this
as a matrix and see the numerator is just the negative of the determinant of that matrix right and I
just divide it by this coefficient here.

And then keep on doing this going to in this recursively. Then I can have, how do I compute the
third thing c n minus 1? I start from here I take b n minus 1, I go to this guy a n minus 3 subtract
the product of with the product of a n minus 1 b n minus 2 and divide by this guy. Similarly, I do
for c n minus 2, I go from here till here subtract this guy and divide again by b n minus 1 and so
on. So, okay so I do all these steps and what does this last statement tell me? The last statement
tells me that the Routh-Hurwitz criterion states that the number of roots of the characteristic
equation with positive real parts is equal to the number of sign changes in the first column of the
Routh array. So, I have this thing here is the first column of the Routh array.
So, a n, a n minus 1 to go to b n minus 1 and c n minus 1, if there is a sign change in any of these
numbers. So, by construction, a n and a n minus 1 would be of the same sign or they would right
with they typically both could be positive. So, if b n minus 1 becomes negative right then there is
a sign change and we’ll see that with the help of an example and then and if this is a plus say- for
example, and this is plus, this is a minus and this again becomes a plus. So, you see well there is
no sign change here, there is one sign change and two sign changes. So, this sign changes will
tell you that there are or they will tell you how many poles are there on the right half-plane right
and that’s what again; we’ll read the statement again.
That the number of roots of the characteristic equation with positive real parts is equal to the
number of sign changes in the first column. It will not tell you the exact locations, but it will tell

292
you a first test is my system stable or not or all the roots of the characteristic equation are they on
the left half-plane or not or if they are on the right half-plane how many are on the right half-
plane right and so on.

So, you can just keep on computing this, this array until you reach s power 0 and this is you just
look at this column. Hence, check if there are any sign changes. If there are no sign changes, then
my system is stable. If there are sign changes then there are roots on the characteristic root roots
of the characteristic equation on the right half-plane. How many are there that will depend again
on the number of sign changes.

Okay so, just to repeat, so for the system to be stable, it is sufficient that all elements of the first
column are positive. Again this is we’ll see why it is just a sufficient condition. If this condition
is not met when the way if there is a sign change then the number of roots with the positive real
part is equal to the number of sign changes. So, if you just look at it this way I will go from a n to
a n minus 1, I ask my question is there is sign change I go from a n minus 1 b n minus 1, and I
keep on asking this question is there a sign change? is there a sign change? And when the answer
is no then I can conclude that the system is stable. And if there are sign changes, if the answer is
yes even once then the system is unstable. And the number of times I change the sign that equals
the number of roots on the right half-plane.

Okay so, just as an example, so I start with a fairly simple looking polynomial s cube 4 s square
9 s and 10, everybody is of the same sign. So, the necessary condition is met right. So, let me go
about drawing the Routh table. So, I start with the s cube which is the highest order of the
polynomials, so I just have 1 and 1, then I go to s square and this is a 4 here, then s n minus 1
would be I am looking at s power 1, it is 9 here, and then I apply in it 10 here.
Then I just compute, how do I do this? So 4 times 9 minus 1 times 10 divided by 4. So, I get a
number here 26 times 4 and then a 10 here just by looking at the formulas for b n minus 1, c n
minus 1 in the previous slide. Okay so, what question do I ask? I just look at this column, and I
ask is there a sign change from 1 to 4? the answer is no; from 4 to 26 by 4, again answer is no;
26 by 4 to 10 again the answer is no. So, all there is no sign change, no sign change, no sign
change and therefore, the system is stable.

293
Now, in the case of unstable thing, so you have s cube and s square 3 s and minus 5; so this again
by inspection is unstable. Okay now, let’s just verify what the Routh guy tells you. So, I just do s
cube and so 1 and 3 then s square have 1 and minus 5, then I do all the thing and then I just find a
sign change here right, from 8 to minus 5. And therefore, the system is unstable you could even
compute the roots exactly and say that. So, we have one this is the unstable guy, and then okay
these are two stable poles, there is a one sign change. So, one pole on the right half-plane right,
this is what this last statement tells you.

Now, let’s do some other examples. So, I have s 5 plus s 4 2 s cube plus 2 s square 3 s plus 15 is
0. And my necessary condition here is met. So, again I start constructing the Routh array. So, s 5
is 1 s cube is 2 s power 1 is 3, and s 4 is 1 s square is 2 and s power 0 is 15. Okay so, something
strange happens here right. If I look at this matrix well the determinant is 0, so something strange
happens here and these are like like the special cases right, where there is a 0 in the first column.
What do I do? do I just leave it here and say well there is no answer to this problem, no, there is
a way out of this. If the first element of any row is 0 then just replace this by a small positive
number and call it epsilon. And this epsilon, the value of epsilon is allowed to approach 0; it is
like a very, very small number and then we compute the Routh table. Now, this now with this
epsilon being a small positive number and then interpret the table for the stability okay, how do
we do this?

So, the way we compute is instead of the zero I substitute an epsilon which is considered to be
greater than 0. Then this element is computed as twice epsilon plus 12 over epsilon and I have a
15 here and then the s power 1 term turns out to be this again doing this, these computations.
And now closely look at this term right, as epsilon becomes very small you could even take the
limit as epsilon goes to 0, this guy becomes a negative number right, I just have 144 by 12. So,
this and then the last number becomes 15. So, I just keep traversing here, I say well there is no
sign change from 1 to 1, no sign change from 1 to epsilon as epsilon is assumed to be positive
since epsilon is assumed to be the positive this is also a positive number. Then I go here this guy
becomes negative and again the last guy is positive.

294
So, I see there is a sign change and there is another sign change. So, since there is one negative
element here, the system is definitely unstable. And then how many sign changes are there, this
is a plus to minus, one sign change, and again a minus to plus. So, the system is unstable.
Further, this table or this row tells me that there are two poles of the system in the right half-
plane corresponding to two sign changes.

So, well let’s see another case. So, have s cube 5 s square 6 s and 30 and then I put this in the
table I pluck pluck these numbers say in 1, 5, 6 and 30. Okay so, there is some strange happening
here right. So, the entire row is now going to 0 right. What do I do now? Do I really substitute
epsilon and epsilon again and do the analysis, well the answer is no. Because you’ll still again
get encounter 0 if you do that right. So, what do we do here? we will be a little careful, that if an
entire row is 0, we form something called an auxiliary polynomial with the row immediately
above this guy. So, this is s 1, I find I form a polynomial with this guy s square which has entries
5 and 30. And what I do is then this polynomial is differentiated with s and I will have then, so if
this is a polynomial of order 2 then I will have something in order 1 here. The rows of zeros of
this guy is in replace with the coefficients of this differentiated term of these differentiated
auxiliary polynomials with respect to s.

Okay, let’s see how this works. So, I have a 0 here, the polynomial just above this is 5 s square
plus 30, I differentiate with s and I get this is 10 s plus 0. I take this 10, I plug in here I take this
0, and I plug in here.

So, it’ll so my new or my new table would now look like this right and then Routh array. So, I
differentiate I take 10 and 0 and then I proceed with the rest of the computations. So, this guy
will just be a 30. And then I again do the sign change test, is there are sign change? The answer
is no, no, no. And therefore, I can conclude that the system is stable because all elements of the
first column are positive right. Okay there is something; something little more here right that
since this guy is entire row is going to 0, there is some there is an existence of a pair of roots
which are on the imaginary axis and that is simply given by solutions of this equation 5 s square
plus 30 equal to 0.

295
You solve this; this will give you the exact location of the roots on the imaginary axis. So,
whenever an entire row goes to 0, then you detect a pair of roots on the imaginary axis.
Now, this is is the conjugate pair.

So, what if there are more of these auxiliary polynomials. So, the auxiliary polynomial is
essentially it has the highest power is of course, and even number actually it could have s 4 s
square and so on or maybe say s 6 or s 8, so this is the highest power. And then you could have
all the other powers could be s square, s power 0, s power 4, s power 2 and so on right, it’ll be all
just have even powers s 6, s 4 and so on. In this case, there are what we see is there are roots that
are symmetric about the origin and symmetry could occur under several conditions.
Symmetry could be like here at so from here we have a minus 1 you could have a plus1 here they
could be symmetric just and lie just on the imaginary axis plus j omega and minus j omega there
could be the something like this also right, you have plus 1 plus j omega minus 1 minus j omega
you have sorry this is a minus 1 plus j omega this, this is the symmetry from here till here. So,
this is a minus 1 this will become a plus 1 if this is plus 1 plus j omega this will be minus plus 1
minus j omega, and here this is symmetry about this imaginary axis. So, if this is a plus 1 the real
part will become minus 1 and so on. And the roots of the auxiliary equation are also the roots of
the characteristic equation.
This again appears only when you know when some of these kinds of things happen.

So, let’s do another polynomial here, I have s 5 plus 2 s 4 plus 2 s cube 4 s square s 2 all being
equal to 0. Okay so, I construct my Routh table and I take s 5. So, I have 1 then s 3 the 2 goes
here, 1 goes here for s 4, 2 sits in here, 4 goes here and this goes here okay. So, whenever you
have things like this you could just even replace this, just divide this by its common factor. So,
this could even be 1, 2 and 1. So, you can divide by 2 divide by 2 divide by 2, it will still be the
same analysis.
And you see that the rows are the same right, and therefore this entire row of s cube goes to 0
okay. What do I do now? I first construct the auxiliary polynomial that is in s 4. So, you have s 4
plus 2 s square plus 1 equal to 0. And even will be the same, even if you write 2 s 4 plus 4 s
square plus 2 equal to 0, I just take the 2 factor out. And then I differentiate d A by d s would

296
give me 4 s cube plus 4 s plus 0. Now, I go here right, I just substitute these entries 4 and 4 over
here.

So, this would be 4, 4. And since 4 is common I just get rid of the 4 and I can just write
equivalently 1 and a 1 okay. Now, I compute s square and what I get is again a 1 and a 1 and not
surprisingly because this is 1 1 this is also 1 1 there is another 0 entry here. Now, I construct
further one more auxiliary polynomial. So, I keep on doing at each time, I encounter a row which
has entries which are completely 0, and then I take this guy. The second polynomial I call it as A
prime and s square plus 1 equal to 0, I differentiate I get 2 and then I get an 0 here. So, I just do
this.

Now, there are no sign changes, can say well this system is stable? Well, from my stability
theory we had learnt something right, that if let’s just draw the complex plane here I have a
sigma j omega everything here is stable. So, if I have a root where this is plus j minus j, this I
called as marginally stable. If I had something like this, say plus 2 j minus 2 j this was also stable
marginally even though marginally, but if I had repeated roots like this plus j, plus j, minus j,
minus j then there was we derived that this roots actually lead the system to instability okay.
Now, where is this? So, I know that there are now four roots on the imaginary axis, two coming
from, because of this row and two more coming because of this row.
So, let me calculate the roots. So, s square minus 1 equal to 0 would give me roots as plus minus
j. And I even further here if I calculate the roots that will just give me another plus minus j. So, I
have four roots which are plus j, plus j, minus j and minus j and these are repeated roots on the
imaginary axis. And we concluded in the stability analysis class that when I have repeated roots
on the imaginary axis then the system goes to being unstable. And therefore, we conclude that
this system is unstable. So, these are the special cases which we need to be to be careful of while
we are computing stability. So, we just need to remember all those cases which we analyzed
during the stability class.

Okay, a bit of, a bit of history here. So, where did all this start from? So, this had roots the
Routh-Hurwitz stability criterion was developed somewhere in the 19th century when Maxwell
now the famous guy of the Maxwell’s equations they became interested in the stability of

297
dynamical systems. And this Maxwell was motivated by this analysis of centrifugal governors
which was essentially invented by James Watt for his steam engine. So, Maxwell was a first to
make some initial guesses around this thing he was first to provide a theoretical analysis of
feedback systems using some linear system, so linearized differential equations.
So, this was how the governors look like and this is a picture which we got from Wikipedia. So,
what did Maxwell say in his paper? So, this is from his paper in 1868 called on governors which
appeared in the proceedings of the Royal Society you can still download it. So, what Maxwell
said that the condition for stability of the governor. So, this guy here is mathematically
equivalent to the condition that all the possible roots they still didn’t define this because they
didn’t have a notion of right half-plane, left half-plane, you know stable roots, unstable roots,
marginally stable roots and so on. Thus all the possible roots so instead of that the stability was
equivalent to the condition that all the possible roots and all the possible parts of the impossible
roots of a certain equation shall be negative.
So, this is a little misleading because they did not have any formal terminology to define it, so
essentially he meant that all roots must be negative. So, possible roots were all the real roots they
shall be negative and then you could have impossible roots which were imaginary and you say
the possible parts of the imaginary roots were the real parts. And this possible parts of the
imaginary roots, so of the impossible roots say for an example, minus 1 plus j omega was
regarded as an impossible root, but this was the possible part all these should be negative. And
then Maxwell could do it for a third-order equation. So, he said that I have not been able to
completely determine these conditions for equations of some higher degree than 3 and then he
posed this problem to a set of mathematicians.

And then, so this analysis by Maxwell lead to the subject for the Adam’s prize in 1877 with
Maxwell was being as one of the examiners. And then the problem which was thrown at
mathematicians was to find a general criterion for dynamic stability of systems right and then
later on Edward Routh in his essay called A treatise on the stability of a given state of motion
presented this condition which we learned just now from the Routh table.
So, parallelly and not surprisingly from Russia there were also people working in similar areas
and then the results again of a steam engine with a governor these were used by some guys to
design water turbine governors. And then Stodola with his collaborator at ETH, in Switzerland

298
independently also arrived at the stability criterion and these two actually evolved independently
right using some different methods. And to honor this parallelly evolved result. So, the stability
criterion is now what we know as the Routh-Hurwitz stability criterion. And this is a little
interesting story of how this stability criterion evolved.

So, what we and these are the papers which you could actually refer to see the history of all these
analyses of stability of differential equations. So, what we learnt so far is the Routh-Hurwitz
criterion some special cases of what happens when there is a 0 in one of the entry, what happens
if there is a 0 in the entire row then did the analysis of the auxiliary equation.

So, next week we will look at more, so at the moment we just looked this analysis only in terms
of roots of a certain polynomial. So, next time we will do give it a little more control flavor
analyze this in terms of feedback. Then we will do things related to sensitivity. We will also
slightly introduce the concept of relative stability; what relative stability means? And we will see
if we could solve some of those relative stability criteria or relative stability problems using the
Routh-Hurwitz criterion.
Thank you.

299
 Control Engineering
Dr. Ramkrishna Pasumarthy
Department of Electrical Engineering
Indian Institute of Technology, Madras

Module – 04
Lecture - 17
Tutorial
Routh Hurwitz Criterion T 1

Hello everyone. So, today we are going to do a tutorial session on Routh-Hurwitz stability
criteria. So, we are going to do some problems and based on the theories, we studied till
now like how to determine stability, how to find out unstable systems in using Routh
Hurwitz criteria. We will also see special cases one as zero as the first element and also
zero in the entire row. Now so, we start with the basic problem how to determine the
stability of a system.

(Refer Slide Time: 00:53)

So, determine the stability of the system whose characteristic equation is s to the power
4 plus 2 s cube plus 8 s square plus 4 s plus 3 is equal to 0. Also, comment on the location
of the roots of the system the poles of the system. So, this is our characteristic equation.
Now, for the characteristic equation, the Routh array will be so we take first s to the
power 4 then s to the power 3. Now, s to the power 4 - first 1, alternatively 8 and then 3;
and for s to the power cube, s cube, so 2 and then 4 and then 0; so as we saw that this

300
term comes from multiplying these 2 minus multiplying these 2 divided by this factor.
(Refer Slide Time: 03:18)

So, we can do this way that 1 by 2 divided by I just take this is the determinant; so 1 by 1,
2, and 8, 4. So, what we are doing basically this is a n, this is a n minus 2, this is a n minus
1, this is a n minus 3. So, what we are doing is a n into a n minus 3 minus a n minus 1 into
a n minus 2 divided by a n minus 1 minus. So, basically, the minus will come here and this
will be plus, so that is what we will do. So, here we get minus half 4 minus 16. So, here it
will be 6.

(Refer Slide Time: 04:58)

So, similarly, if we do 1, 2, and 3, 0, we will get 3. So, here you get 3.

301
(Refer Slide Time: 05:12)

Now, s to the power 1; so similar way 6 into 4 minus 3 into 2 that means, and s to the
power 0, will be just 3. So, we see that there is no sign change. So, just let’s write the first
column only.

(Refer Slide Time: 05:44)

So, the first column is 1, 2, 6, 3, and 3. So, the signs are positive. So, all are positive. So,
there is no sign change. So, we can say the system is stable, there sorry all the roots lie in
the left-hand plane hence the system is stable.
(Refer Slide Time: 07:29)

302
So, the second problem we are going to do on how to find out if the system is unstable
or not. So, the second problem will be, so using Routh-Hurwitz. So, this is RH means Routh
Hurwitz, find the stability of the system with characteristic equation s to the power 5. So,
it will be 3 s cube.

(Refer Slide Time: 08:50)

So, Routh array for the given system is first s to the power 5, then s to the power 4, s to
the power 3, s to the power 1. So, 1, alternatively 3, then 16 and here comes 1, 9, 10. So,
this point will be, so this is easy 3 minus 9, so it will be minus 6, 1 into 16 minus 10. So,
this will be 6. Okay now, this is 10; 10 into 6, 60 minus, so 12 and this is 10.

303
(Refer Slide Time: 10:18)

So, now the first column is 1, 1, minus 6, 10, 12, and 10. So, positive, positive, negative,
so there are two sign changes from here and from here. So, we learn that number of sign
change is equal to number of roots in the right-hand plane. So, number of therefore is 2.
And the system is, and the system is unstable.

(Refer Slide Time: 12:49)

Now, we solve a problem if the first element of the row is 0. So, okay let’s consider the
characteristic equation of a system s to the power 5 plus s to the power 4 okay. Now, the
Routh array is, so first will be 1, then 2, then 3, here will be 1, 2, and 5. So, 2 minus 2 this

304
will be 0; and 3 minus 5, this will be minus 2. So, we see that this term is becoming 0. So,
what we learnt is we will replace it by a positive number, small positive number epsilon
here. So, the 0 is replaced by a smaller positive number epsilon now. So, this will be
epsilon. Now, so the next term becomes 2 epsilon plus 2 by epsilon, and here we get 5.
So, calculating we find this will be 5 epsilon to the power 2 plus 4 epsilon plus 4 by 2
epsilon plus 2, and s to the power 0 it will be 5 okay.

(Refer Slide Time: 15:33)

So, what we do is now we consider these terms, we check the signs considering that
epsilon is a small positive number tending to 0, but it is positive then 2 epsilon plus 2 by
epsilon, this is a positive quantity. Now, you can write it as 2 plus 2 by epsilon. So, this is
greater than 0 and epsilon is positive. So, this is also greater than 0. Now, the other term
this one, limit epsilon tends to 0 minus of 5 is equal to minus 2 that is negative. Now, we
consider the first column first one is 1, so positive; second is positive, third one is epsilon
we assume that epsilon is a small positive quantity, so this is positive; this is also we found
positive. Now, this quantity we found that this is negative, so this is negative; and finally,
5 so positive. So, there are two sign changes here and here. So, number of roots in the
right-hand plane is equal to 2; and the system is unstable.
(Refer Slide Time: 18:32)

305
So, let’s consider one example where the entire row is 0. So, how we use the auxiliary
polynomial and find out the roots and determine the stability of the system, okay. So, let
the characteristic polynomial equation be given by s to the power 6 plus 3 s to the power
5 plus 5 s to the power 4 plus 9 s cube plus 8 s square plus 6 s plus 4 is equal to 0. So,
construct the Routh array s to the power 6, s to the power 5, s to the power 4, s to the
power 3, s to the power 2, and then s to the power 0. Now, first comes 1, then 5, then 8,
and then 4. And here is first element is 3, then 9, and then 6. So, here will be, first element
will be 2 - 15 minus 9 by 3 are 2, and then 6 and last term is 4. So, now, for s to the power
3 all the terms are 0, 18 minus 18 – 0, 12 minus 12 - 0. So, I get here, so 0.

Refer Slide Time: 20:35)

306
So, what we learnt is we consider auxiliary polynomial A s, and we go to the immediately
upper row, so which is for s to the power 4, so here it will be 2 s to the power 4 plus 6 s
square plus 4. And the auxiliary equation s to the power 4 plus 3 s square plus 2 is equal
to 0. So, we take common; we take common 2, so divide all other terms by 2 and we get
this polynomial equation. So, what we now do is, we find the derivative of the auxiliary
equation that auxiliary polynomial that is d A s with respect to d s that is 4 s to the power
3 plus 6 s plus 0. So, the coefficients where will be replaced by 4 and 6. So, 4, 6 and 0; for
s to the power 2, the terms will be 3 and here will be 4. Now, so 18 minus 16, so it will be
2 by 3 and here it will be 4.

So, just consider the first column. So, there are no sign changes. So, all the terms are
positive, positive, positive, positive. So, there is no sign change that means, there are no
roots in the right-hand plane, but we can’t, for determine the stability of the system we
need to find out the roots of the auxiliary equation, because we learnt that the roots of
the auxiliary equation is also the roots of the characteristic equation. So, the roots will
determine stability.
(Refer Slide Time: 23:53)

307
So, here we can see that s square plus 1 into s square plus 2 equal to 0. So, for s square
plus 1 equal to 0, we get s is equal to plus minus j into 1; and for s to the power 2 plus 2.
So there are 4 roots on the imaginary axis. So, if we construct the, we just draw the s
plane and this is j omega this is sigma. So, and there is one at plus j, one at minus j and
one as minus j 2, one as plus j 2, square root.

What we learnt in the first lecture of module 4 that if there are non-repeated roots on
the imaginary axis, then the system will be marginally stable. So, here we find that there
are 4 roots on the imaginary axis, but none are repeated. So, the system is marginally
stable, so done;

Now, some problem on feedback. Now, we do some problems on determining the gain
of the feedback systems what we learnt in lecture 3 of module 4. So, how we can we’ll
see how we can use Routh-Hurwitz criteria and apply it in feedback systems to determine
the gain. So, quickly we will do.
(Refer Slide Time: 26:57)

308
So, this is problem number 5. So, let the characteristic equation of a closed-loop feedback
system be s to the power 4 plus 2 s cube plus 4 plus k s square. So, first we construct the
Routh array s to the power 4. So, first term will be 1, second term is 4 plus k and third
term is 25. Now, here first term will be 2 and second term will be 9. So, here we get is k
minus half and here 25.

Now, just after calculating we will find that this term is 18 k minus 109 by 2 k minus 1,
and this term will be 25. Now, if the if we want the system to be stable, so let’s find out
what will be the range. So, what will be the range of k for the system to be stable? So, for
stability, what we know is that all the terms should be positive.
(Refer Slide Time: 29:34)

309
So, k minus half, first condition is k minus half should be positive that means, k is greater
than half. And second condition is, so this term will also be positive. So, 18 k minus 109
by 2 k minus 1 is greater than 0 that implies that k is greater than 109 by 18. Now,
definitely, this value 109 by 18 is greater than half. So, the system will be stable when the
value of k is greater than this. So, system is stable for k greater than 109 by 18.

So, so similar problem if we just give a parameter to the system, and we tune it and find
out how the parameter what will be the range of the parameter such that such that the
system will be stable or marginally stable or unstable. So, we consider another example
like that.
(Refer Slide Time: 32:02)

So, characteristic equation s to the power 3 plus k s square plus k minus 3 s plus 4 is equal
to 0. So, we have to find out what will be the range of k for the system to be stable and
marginally stable. So, Routh array s cube, so first term is 1, then k minus 3 and this is k,
this is 4. So, this term will be k square minus 3 k minus 4 by k, and this term will be 4. So,
for stability the conditions will be, for the system to be stable.

(Refer Slide Time: 34:11)

310
So, all the elements of the first column should be of the same sign. So, it should be
positive here. So, k first condition is k is positive and second condition is k square minus
3 k minus 4 by k is greater than 0. So, since k is positive, it implies that k square minus 3
k minus 4 is greater than 0 that is k minus 4 into k plus 1 is greater than 0. So, here we
find that k is greater than 4 because k cannot be greater than minus 1 because that
condition it lies between minus 1 and 0, so it does not hold first condition. So, k will be
greater than 4 if the system is to be stable. So, part a is done.

(Refer Slide Time: 35:46)

Now, part b; so we now start we do part b. So, for marginally stable, this term will be 0
that k square minus 3 k minus 4 is equal to 0 that is when k is equal to 4. Now, the auxiliary

311
polynomial A s is equal to 4 s square plus 4, because k is equal to k is equal to 4 and d A
by d s 8 s plus 0. So, replace k by 4, replace this by 8, and this will be 4. Now, we find that
the all the sign elements of the first column are of the same sign.

So, there are no roots in the right-hand plane and the roots of the auxiliary equation s
square plus 1 taking 4 common equal to 0, s is equal to plus minus j 1. So, there are two
roots on the imaginary axis at plus j 1 and at minus j 1, and hence the system is marginally
stable with oscillating at the frequency of omega is equal to 1 radian per second. So, k
will be 4.
(Refer Slide Time: 38:05)

Problem number 7: so we find the number of roots which are in the left of the axis, s is
equal to minus 1. Okay, so, how to do that? so we have to find out how many roots are
there in the left of s is equal to minus 1. So, we consider a dummy variable s n such that
s n is equal to s plus 1. Now, if we replace the, if we right the characteristic equation in
terms of s n, so it will be s n minus 1 whole cube plus 4 s n minus 1 whole square plus 6
s n minus 1 plus 4 is equal to 0. So, if you calculate you will find that the characteristic
equation in terms of s n will be s n to the power 3 plus s n to the power 2 plus s n plus 1
is equal to 0.

(Refer Slide Time: 40:42)

312
Now, Routh array for this system, Routh array for the new equation is s n to the power
3, s n to the power 2, s n to the power 1, here 0, so 1, 1, 1, 1. So, the entire row is going
to be 0. So, we consider the auxiliary equation A s is equal to s to the power 2 plus 1 is
equal to 0. So, d A s by d s is equal to 2 s plus 0. So, this is 2 and this will be 1. So, all the
elements of the first column are positive. So, there are no roots on the right-hand side
of s n, it is no roots, s is equal to minus 1.

Now, for the auxiliary equation if we find the roots, so we will get sorry, so s n square
plus 1 is equal to 0 that is s n is equal to plus minus j. So, what we get is that there are
two imaginary two roots on s n is equal to; s n so imaginary axis of the shifted plane s n.
So, s n plane is plus this is j omega, sigma here. So, there will be two roots.

(Refer Slide Time: 42:45)

313
Now, if we see the entire picture in the s plane that is not in s n plane, but in s plane what
we are going to see, find that. So, this is our original sigma, this is our j omega. So, we
have shifted the axis to minus 1, and there are two roots at minus j, plus minus 1 or better
to write minus 1 minus j, minus 1 plus j. So, there are two roots here. So, this kind of axis
shift is very useful in phase systems where there are number of imaginary sorry complex
roots. So, more than one complex pair of roots are there. So, if we are drawing to
determine the relative stability of the system based on how the roots effect on the
damping of the system, so this axis shift method using Routh-Hurwitz criteria will be
helpful.
(Refer Slide Time: 44:40)

And the final problem for today’s tutorial. So, final problem for today’s tutorial. Now, let’s
consider the same problem we did in the module 3 for time response. So, this is the error
E s. So, in that problem for time response what we found is if the steady-state error is
going to be 0, but not for ramp input. So, each will be 1. So, we can modify the loop gain
of the system as G s is equal to k into s plus 2 by s, say h is equal to 1 s plus 1, s square
plus 10 s plus 41.

(Refer Slide Time: 46:53)

314
And the characteristic equation will come from 1 plus G s. So, the numerator of this 1 plus
G s it is so the numerator of this term, so this part. So, this is the characteristic equation
of the system.

(Refer Slide Time: 47:46)

And so, you have to find out what will be the range of k, for what values of k the system
is stable. So, the numerator will be the characteristic equation.

Student: You can do that?

Yeah.

(Refer Slide Time: 48:23)

315
So, the characteristic equation is the numerator of the part of 1 plus G s that is CE is s to
the power 4 plus 11 s cube plus 51 s square plus 41 plus k s plus 2 k is equal to 0. Okay
so, we form the Routh array. So, the first term will be 1, second 51 and finally, 2 k and
here it will come eleven and 41 plus k. So, this term will be 520 minus k by 11 and this is
2 k, s to the power 1 will be minus k square plus 237 k plus 21320. So, we have to find the
values of k for which the system is stable. So, there are three terms related to k in the
first column.

(Refer Slide Time: 50:20)

316
So, the first condition will be 520 minus k by 11 greater than 0 that implies k is less than
520. Second condition is minus k square plus 237 k plus 21320 by 520 minus k will be
greater than 0. So, from condition one this term is greater than 0. So, we have to make
this term greater than 0. So, let’s consider the quadratic equation k square minus 237 k
plus minus 21320 is equal to 0. So, the roots of this equation plus 237 plus minus just
calculate this you will get 141449 divided by 2. So, if this term is going to be greater than
0, so what we need is k should be k should lie between plus 237 plus root over 141449
by 2 and 237 minus root over 141449 by 2 what we find that. So, this is almost equal to
367. So, the upper limit of k now reduces to from 520 to 367.

(Refer Slide Time: 53:14)

Now, the third part third condition comes from, I write it here third condition is k is
greater than 0. So, this part is negative. So, this part will be replaced by 0. So, the range
of k is now 0 greater than less than k less than 367. So, within this range, the system will
be, the entire closed loop feedback system will be stable that is it.

Thank you.

317
 Control Engineering
Dr. Ramkrishna Pasumarthy
Department of Electrical Engineering
Indian Institute of Technology, Madras

Module – 04
Lecture – 03
Closed-Loop System and Stability

Hello everybody, today we will talk a bit more details about the closed-loop system and
stability. So, we had studied so far stability from the BIBO point of view or the bounded
input bounded output point of view and also stability in terms of the equilibrium points.
And then we characterize those as the location of poles of my open or closed-loop system
depending on wherever you are. So, what we will do today is to also look a little more
details about feedback, is feedback really helpful all the time or there are some drawbacks,
we’ll skip the drawbacks part or we will you know do not go too much into the details.

(Refer Slide Time: 01:15)

And we’ll just see how feedback helps us in a lot of ways and also define something called
the relative stability.

Okay so. So, open-loop systems are heavily prone to disturbances right. So, the output is
affected by the change in system parameters, system parameters could be like I’m talking
of a mechanical system- a mass could change or any other thing could change in the
system. It could also change because of external environmental conditions like

318
temperature, sometimes disturbances act into the system and I would I really wouldn’t
know how to, how to bring my system back to the original state for example, if I’m driving
a car and my pedal is at a constant position, if I hit a ramp you know or an upslope that’s
a disturbance right. And if I keep the same pedal position and I keep going through the
ramp, I will not be able to maintain the desired speed and after a while, I might even lose
the speed and I might just go to zero speed right. So, there is no automatic correction if I
set my pedal speed to its open-loop value and just don’t change it when the disturbance
comes, right.

(Refer Slide Time: 02:21)

In closed-loop systems, the output is actually measured and compared with a reference
signal and which generates an error signal by which I update my control law and through
the control law, I adjust my actuator which actually, which gives the actual control input
to the system right. So, typical things that could happen are you have the effect of
disturbance and you also have measurement noise right. So, my measuring devices may
not be very accurate and you know they might come with a bit of noise and so, how do we
deal with those these things? the disturbance signal here and the measurement noise. So,
we want to keep or design our controller or keep in mind such that I reject disturbances
and I also should reject the measurement noise. Okay.

319
(Refer Slide Time: 03:16)

So, let’s start with an example right. So, this is a model for automatic control of speed for
an automobile on a straight run. So, I have a reference speed, I have the measured speed
and this generates an error signal which will decide the position of my throttle, then this is
the throttle angle. There could be disturbances in terms of the road slope, it could either be
a positive slope or a negative slope; no disturbance would just mean no slope. And this is
a very simple model of my car, right. So, the car which now looks very big with lots of
components actually you know looks very small here right, it’s just a first-order differential
equation or a system with just one pole and I just get the desired speed, right.

So, this is like the standard measuring devices like the tachometer and all and this will
convert it to be appropriate speed.

320
(Refer Slide Time: 04:06)

So, the engine for simplicity is modeled as a first-order system with a certain time constant
and a gain of 𝑲 with a typical value of 1.5. The input is a throttle angle measured in degrees
and the output is the desired speed right and the speed feedback is obtained by a tachometer
right somewhere here. So, this comes as an; in the feedback loop and this is used to obtain
the error signal for changing the throttle angle right and for desired purposes we say that
throttle angle has a time constant much less than lesser than the system time constant and
therefore, we could ignore it right and it’s just modeled as a small gain 𝑲𝟏 .

So, the error signal via 𝑲𝟏 gives me the control input and some typical value we could
assume it in this case to be 𝟓𝟎 the disturbance signal appears on the automobile runs up or
on a downslope right, it is expressed in terms of percentage of the road slope and I have a

321
𝟗𝟎° would actually mean that you actually hitting a wall and all and just going vertically
upwards. So, and this signal is converted into equivalent throttle angle and then you have
this, this 𝑲𝒈 which takes care of the units and you have this like 𝑲𝒈 is 𝟏𝟎𝟎° per unit
percentage of the slope.

So, why are we doing this? So, let’s see if what happens if I just do not measure this guy
right, I have a reference and I just have the output right and there was no, no, feedback and
also take the case of feedback.

(Refer Slide Time: 05:48)

So, given the first condition or the first problem that my reference or not reference my
desired speed is 60 kilometers per hour, that is in terms of the picture that would be 𝑽(𝒔).
So, 𝑽(𝒔) is 60 kilometers or say 𝒗(𝒕) more nicely and then we have to find the 𝑹 or the
reference signal with the system being first open-loop and second being closed-loop and
we’ll also see how the steady-state error changes with this, right.

So, let’s say in the open-loop case. So, at steady-state, steady-state speed is 60 and how
𝒌 𝒌 𝒌 𝒌
will my transfer function look like. So, it will be 𝝉 𝟏+𝟏. So, this output is 𝝉 𝟏+𝟏 and some
𝒔 𝒔

𝑹(𝒔), okay this will be a step signal right I have to determine the size of the step. So, let

322
 𝑹
me just call it some number say and I do the final value theorem and I just get that 𝟔𝟎 =
𝒔

𝑹 𝒌𝟏 𝒌 right. And then 𝒌𝟏 was this guy 50, 𝑲 is 1.5 and I do these computations and I get
that 𝑹 is 0.8 this is kilometers per hour.

Now, let’s see what is the steady-state error right. So, also if I look at in the Laplacian
domain right 𝑬𝒔𝒔 , 𝑬 steady-state the Laplacian is 𝑹(𝒔) −𝒀(𝒔). Okay so, what we want to
do is compare the errors; that is 𝑹(𝒔) − 𝒀(𝒔). So, 𝒀(𝒔) here or whatever is this is, 𝒀 here
𝒌 𝒌
is the velocity. So, 𝑽(𝒔) would be equal to 𝒀(𝒔). So, 𝒀(𝒔) is again take this 𝝉 𝟏+𝟏, this is
𝒔

𝒌𝟏 𝒌 𝒌𝟏 𝒌
∙ 𝑹(𝒔). So, the error is 𝑹(𝒔) [𝟏 − ] and at steady-state, sorry. This error just
𝝉𝒔 +𝟏 𝝉𝒔 +𝟏

becomes 𝑬𝒔𝒔 , I do the final value theorem as you know 𝑬𝒔𝒔 = 𝐥𝐢𝐦 𝒔 𝒆𝒔𝒔 (𝒔) that will just
𝒔→𝟎

be. So, 𝑹(𝒔) is the step of some size 𝟏 − 𝒌𝟏 𝒌 i.e., 𝑬𝒔𝒔 = 𝐥𝐢𝐦 𝒔 𝒆𝒔𝒔 (𝒔) = 𝑹(𝒔)[𝟏 − 𝒌𝟏 𝒌 ];
𝒔→𝟎

this is my open-loop error ok.

So, let’s do the closed-loop case. So, in closed-loop the reference is still 60 is 𝑹 can you,
I’ll directly write the steady-state expression right, I just take care, you know to get rid of
all the transients. 𝟏 + 𝒌𝟏 𝒌. So, I am just finding this 1 right. I am finding 𝑹 in the closed-
loop case when the desired speed is 60. So, I do all the computations and I get 𝑹 as 60.8
right the reference is again a speed.

So, now what is the error here, is again I look at the signal here right; that is again
𝒌 𝒌 𝟏
𝑹(𝒔) − 𝑹(𝒔) 𝟏+𝒌𝟏 𝒌. So, this is 𝑹(𝒔) and this times 𝟏+𝒌 𝒌. So, just for simplicity assume
𝟏 𝟏

that 𝑹 is just 1. So, in the open-loop case if I just say what is the steady-state error this
would be for the values of 𝒌𝟏 and 𝒌 which were 𝟓𝟎 and 𝟏. 𝟓 this would be −𝟕𝟒 and the
𝟏
and the steady-state error here would be 𝟏+𝟕𝟓 and this is a significantly smaller than here

right. So, I have a big steady-state error over here and I have a fairly smaller steady-state
error over here right. So, so you see how the feedback or closed-loop reduces the steady-
state error.

(Refer Slide Time: 12:11)

323
In the next case, let’s do some transient analysis right. So, earlier we have done the steady-
state analysis. So, we want to find at what time does the vehicle reach 𝟗𝟎% of its steady-
state value for the open and closed-loop system. So, let’s just do first for the open-loop
−𝒕⁄
case okay. So, in the time domain I could write this directly, 𝒗(𝒕) = 𝟔𝟎(𝟏 − 𝒆 𝟐𝟎 ), so
this is like straight forward to write from the expression. So, 90% of this value would be
54, is 60 and I want to find out the time, let me call this 𝒕𝟏 at which this occurs. So, I have
an equation here where there is one unknown. So, I can easily compute 𝒕𝟏 to be 46
seconds. Okay.

Now, let’s see what happens in the closed-loop case, so, in the closed-loop case. So, let me
𝒌𝟏 𝒌
just first write 𝑽(𝒔) = 𝝉 and this is multiplied by the 𝑹(𝒔). So, I just do some
𝒔 +𝟏+𝒌𝟏 𝒌

manipulations and computations. So, in this case, my 𝑹(𝒔) for the closed-loop it comes
from what I had computed over here. So, 60.8; 60.8 and then it’s Laplacian; in the Laplace
domain, it’ll be 60.8, whereas in the previous case my 𝑹 for the open-loop when I use I
will use this 𝟎. 𝟖 𝒌𝒎/𝒉𝒓 right, so, here when I do it here. So, in the open-loop case my
𝑹 = 𝟎. 𝟖 𝒌𝒎/𝒉𝒓. So, just be careful of that.

So, now I just write down the final expression for 𝒗(𝒕), I’ll skip a bit of computations. So,
−𝒕⁄
what I have 𝟔𝟎(𝟏 − 𝒆 𝟎.𝟐𝟔𝟑 ) and I just do nothing I just substitute these values and then
compute the inverse Laplace transform and we by now know how to do these things. So,
−𝒕𝟐⁄
again this 90% would be 60, I have 𝟏 − 𝒆 𝟎.𝟐𝟔𝟑 and I compute 𝒕𝟐 that is actually turns

324
out to just say 0.6 seconds right. So, this is the open-loop and this is the closed-loop and
you see that this is 𝟕𝟕 approximately times faster right. So, if I just have a closed-loop my
steady-state error is much smaller and also my response of the system is pretty fast. So,
this will motivate us towards you know analyzing what are the effects of feedback on the
system.

First, we saw well it helps in the steady-state response, it also helps in the transient
response okay. Is there anything else associated with feedback?

(Refer Slide Time: 15:50)

Now, in the initial block diagram, we mentioned a few other things in addition to the
reference and the output. So, one was the noise sorry the disturbance and second was the

325
noise, okay. So, in the presence of disturbance and the noise I can compute the overall
output of system or overall response of the system as the sum of response just to the
reference signal and while computing this I set 𝑫(𝒔) and 𝑵(𝒔) equal to 0, then I compute
the response by setting 𝑹(𝒔) and 𝑵(𝒔) to be equal to 0, this will give the response to the
disturbance signal and lastly for the noise I set the remaining two inputs to 0 and I just use
the superposition. So, this is the total response to the reference signal, to the disturbance
signal and the noise.

So, let me just define the loop gain as 𝑮𝒄 (𝒔) 𝑮(𝒔). So, if you have just forgotten where
these 𝑮𝒄 (𝒔) and 𝑮(𝒔) comes from well these are just these guys, right. So, 𝑮𝒄 (𝒔) is my
controller, this is the plant, this is 𝑯(𝒔) is what happens in the feedback, I have reference
signal here, the disturbance, and the noise okay. So, the error now becomes, of course, a
function of what is the reference and something to do with the controller transfer function
and the plant transfer function. Similarly, with the disturbance and also with the noise. So,
I can just compute this directly and these are I think in one of the tutorial classes we saw
how to compute the effect of disturbance and then take the total response of the system as
the sum of the response to the disturbance and combine it to the to the response of the
input right, just the superposition.

(Refer Slide Time: 17:44)

326
Okay so, when we started this lecture we said well there could be some parameter
variations right and then now we’ll see what is the effect of parameter variations on the
system. So, assume there is no disturbance and noise we’ll come to this very shortly. So,
say my transfer function or the plant changes from 𝑮(𝒔) to ∆𝑮(𝒔) and then the error signal
changes accordingly. So, I am not I am just doing something simple here. So, 𝑬(𝒔) which
𝟏
is 𝟏+𝑳(𝒔) 𝑹(𝒔), 𝑳 = 𝑮𝒄 (𝒔) 𝑮(𝒔) I just substitute instead of 𝑮(𝒔)=𝑮(𝒔) + ∆ 𝑮(𝒔) . So, this

is how my error looks like right.

Now the change in the tracking error can be computed, you can just say what is ∆ 𝑬(𝒔)?
𝟏
you take this guy here and then 𝑬(𝒔), I just substitute this guy 𝑬(𝒔) = 𝟏+𝑳(𝒔) 𝑹(𝒔) and I

get this expression right. So, this is if again if ∆ 𝑮(𝒔) = 𝟎 this goes to 0, that’s the trivial
case to check this. Okay, the change in the tracking error is something like this 𝑮𝒄 (𝒔) and
∆ 𝑮(𝒔) and you have lots of terms of ∆ 𝑮(𝒔) influencing this particular number.

(Refer Slide Time: 19:13)

327
So, if I just take an open-loop right. So, I have okay then this; for convenience, I will just
draw it here and I hope I will use a write notation. So, this was 𝑹(𝒔) and this was 𝑪(𝒔),
now the error is usually defined as 𝑹(𝒔) − 𝑪(𝒔). So, in this case, this is 𝑹, I’ll just omit
the 𝒔 for the moment, 𝑪 is 𝑮𝒄 𝑮 𝑹 okay. Now 𝑬 + ∆𝑬 = 𝑹[𝟏 − 𝑮𝒄 (𝑮 + ∆𝑮)]. So, this is
how my error looks like. And therefore, ∆𝑬 would be this guy minus the original error,
original error was what? That was 𝑹[𝟏 − 𝑮𝒄 𝑮], all in the 𝒔 domain. Okay, if I just
compare these things. So, I just subtract this from this what I am left with is
∆𝑬(𝒔) = −𝑮𝒄 (𝒔) ∆𝑮(𝒔). So, this is again in the open-loop.

Now, let me compared to what happens in the closed-loop. So, I have the error signal is of
course. So, I do some approximation that 𝑮𝒄 (𝒔) 𝑮(𝒔) is much larger than the very small
change and I can get an approximate expression like that, this is the change in error,
𝑮𝒄 (𝒔)∆𝑮(𝒔)
∆𝑬(𝒔) = 𝟐 𝑹(𝒔). So, if you compare this with this you see that the tracking error
(𝟏+𝑳(𝒔))

𝟐
is reduced by a factor of 𝟏 + 𝑳(𝒔) or the square, (𝟏 + 𝑳(𝒔)) and for large values of 𝑳(𝒔),
I can just again ignore this 𝟏 and just write 𝑳(𝒔) and I just get this expression
𝟏 ∆𝑮(𝒔)
∆𝑬(𝒔) = − 𝑳(𝒔) 𝑹(𝒔), this is not difficult to obtain right. So, the way you do this is
𝑮(𝒔)
−𝑮𝒄 ∆𝑮
you have ( 𝑹(𝒔)) and this is what remains right, you put a ∆𝑮 and a 𝑮 and right
𝑳(𝒔)𝑮𝒄 𝑮

and then okay. So, what does this mean? that large values of 𝑳(𝒔) result in smaller changes
in the tracking error okay. Is there something which I could change in 𝑳(𝒔) well. So,
𝑳(𝒔) if we just recall was 𝑮𝒄 (𝒔) 𝑮(𝒔).

So, there could be well you know change the plant, where I really don’t want to do that
right because the plant is already I have, there is something which I’m given to control I
just cannot change the subject or the plant. So, I can just manipulate my 𝑮𝒄 (𝒔), in such a
way that I it results in small smaller changes in tracking error. So, this is about how
feedback changes the error or helps in reduction of error.

(Refer Slide Time: 22:52)

328
Now, so, what we saw was essentially this ∆𝑬(𝒔) related to small changes in the system
parameter. Now let’s try to quantify that a little more right. So, we call the system
sensitivity. So, we’ll define this new term as the ratio of change in the transfer function to
the change in the process transfer or some small parameter. So, I have the 𝑻 as the transfer
function. So, let’s just understand from here. So, I want to check how my overall transfer
function changes with small changes in 𝑮 right. So, this is what I call as the sensitivity
change this is 𝑺, how does it, how sensitive is a transfer function 𝑻 to changes in 𝑮 i.e, 𝑺𝑻𝑮 .
So, what is 𝑻 here? this is my closed-loop transfer function ok.

𝝏𝑻 𝑮
So, the sensitivity I define. So, from here I just I assume 𝝏𝑮 𝑻 okay, and this is how it looks

like, the sensitivity of my transfer function with respect to 𝑮 or changes in 𝑮 is simply this
𝟏
one 𝟏+𝑮 .
𝒄 (𝒔)𝑮(𝒔)

(Refer Slide Time: 24:05)

329
I will go here right, okay again I just start with I am just looking at the sensitivity with
respect to disturbance right, in which case I set the reference and the noise to be 0 and I
𝑮(𝒔)
can compute the tracking error right. So, this is 𝑬(𝒔) = − 𝟏+𝑳(𝒔) 𝑫(𝒔). So, if I were just to
𝟏
write the sensitivity in terms of 𝑳(𝒔) this would just be 𝟏+𝑳(𝒔) where 𝑳(𝒔) is again what I

call as the loop gain L(s), and just to recall again 𝑳(𝒔) = 𝑮𝒄 (𝒔)𝑮(𝒔).

𝟏
So, this is my error with respect to disturbance and it takes this term right. So, 𝟏+𝑳(𝒔) is the

sensitivity; sensitivity of the transfer function which changes in −𝑺𝑻𝑮 (𝒔)𝑮(𝒔)𝑫(𝒔) right.
Okay, if 𝑮(𝒔) is fixed which is usually the case then with increase in the loop gain, so I
want to reduce the error right, typically my error should be 0 if there is disturbance there
should be no error right. So, how do I well, but then you know if I have a disturbance I
would expect that there is some error. So, if there is some error I would want to minimize
the error right. So, if 𝑮(𝒔) is fixed then with increase in loop gain this guy the effect of
disturbance can be decreased right. Okay.

So, something which I will briefly introduce now is okay, is something related to the
frequency we’ll talk a little more of this when we talk of frequency domain. So,
disturbance signals are typically signals of low frequency right, it may be like a step or
even if there is oscillatory behavior in in the disturbance they will just be at a very low
frequency, in the car example- my disturbances are either I hit an uphill or a downhill or

330
something like, it’ll just be the constant thing without any disturbance right. So,
disturbance signals are mostly introduced or these usually occur at low frequency. Okay.

So, my 𝑳(𝒔) is, okay if I just look at how it varies with the frequency I can just see the
magnitude of |𝑳(𝒋𝝎)| at frequencies right, so, at low frequencies. So, I want. So, what
does it say that these expressions tell me that with an increase in loop gain the effect of
disturbance decreases, now I do not have to increase it like forever right? So, 𝑳(𝒔) I know
depends on frequencies and my disturbance signals are associated with certain frequencies
or certain lower frequencies. So, I want this 𝑳(𝒔) to be increased at low frequencies. So,
this should be bigger for lower frequencies. So, this essentially means that I increased the
controller gain at low frequency right or in another way I am just. So, if this increases my
sensitivity decreases. So, I just I ideally I would want I don’t want to be sensitive to
disturbance right, whatever happens, I just should reject it.

But typically that doesn’t happen right. So, we can only decrease sensitivity. So, when I
when I try to increase this at low frequency my sensitivity decreases right at low
frequencies I don’t really care what happens at high frequencies at the moment because
there are no disturbance signals which come at high frequencies.

(Refer Slide Time: 27:37)

Now, what about this noise right. We so far, have not talked about it, but then usually my
measurement comes with some noise right.

331
The tracking error in terms of the measurement is something like this right again is from
𝑳(𝒔)
the original expressions which we obtain. So, this guy is also known as the
𝟏+𝑳(𝒔)

complementary sensitive sensitivity function. So, some people denote it as 𝑻 some people
denote it as 𝑪 and since we have already used the notation 𝑻 and 𝑪 we’ll not use, we’ll just
say this may be just for; I can just call this 𝑺′ the complimentary sensitivity function okay.
And this noise, so small loop-gain if this 𝑳(𝒔) is very small it leads to a good noise
attenuation right because this the effect of noise would be reduced and typically noise
signals are signals which occur at a very higher frequency or which get into the system
these signals are typically high-frequency signals.

So, for the purpose of noise attenuation the loop gain is turned to be of low values at very
high frequency right previously what we had that I want the loop gain to be higher at lower
frequencies that will eliminate my disturbance for the purpose of noise attenuation the loop
gain is tuned to have lower values at high frequencies right. So, when I do this the
controller gain is increased at low frequencies for disturbance rejection and decreased at
high frequencies for noise attenuation, right. So, it’s like some kind of a complementary
effect over there right. So, these are the two ways which you know based on what is the
|𝑳(𝒋𝝎)|, I can deal with disturbance and noise together or reduce the effect of disturbance
and the effect of measurement noise.

We’ll do more of this when we do frequency response like you know what it actually
mean- what is low frequency? what is higher frequency? and so, on and how do I actually
go about computing these things? that we’ll keep till a little later when we do frequency
response analysis. Okay, and something which we’ll use throughout from now on is what
is called the characteristic equation, right.

(Refer Slide Time: 29:48)

332
So, what is the characteristic equation? Well, it just, I take the transfer function and I look
at what is in the denominator. So, I know how to compute this right, given 𝑮(𝒔) in the
forward loop, 𝑯(𝒔) in the negative loop, the overall transfer function 𝑻(𝒔) from 𝑹(𝒔) to
𝑮(𝒔)
𝑪(𝒔) is 𝑻(𝒔) = 𝟏+𝑮(𝒔)𝑯(𝒔) okay.

(Refer Slide Time: 30:18)

So, the characteristic equation is when I equate the denominator of this guy to 0 right.
𝟏 + 𝑮(𝒔)𝑯(𝒔) = 𝟎 okay. And so, if 𝑯(𝒔) = 𝟏 then I am kind of something very simple
right, in 𝟏 + 𝑮(𝒔) = 𝟎. Okay.

333
So, now this is a system with non-unity feedback. Now, is it possible that I do some magic
and I convert this to a system which looks like this let me call this 𝑮′(𝒔), and then a plus,
a minus with this to be 1 this guy still being 𝑹(𝒔) and this guy still being 𝑪(𝒔) right. So, I
just want to convert this non-unity feedback system to a unity feedback system. Can I do
this? Well, the answer is yes. So, what I do is I just equate. So, how will this transfer
𝑮′ (𝒔)
function look like? That is simply 𝟏+𝑮′ (𝒔) and I find out what is 𝑮′ (𝒔) in terms of 𝑮(𝒔) and

𝑯(𝒔). So, a non-unity feedback system can be transformed into a unity feedback system
by just using this transformation this should be very straightforward to compute because
we are just dealing with very easy looking linear equations over here.

𝑮(𝒔)
So, 𝑮′ (𝒔) would be this one 𝟏−𝑮(𝒔)+𝑮(𝒔)𝑯(𝒔), 𝑯′ (𝒔)would be 1 right, and while I do this the

characteristic equation remains the same. So, if I just look at 𝟏 + 𝑮′ (𝒔)𝑯′ (𝒔) this is
𝟏 + 𝑮′ (𝒔) and I’ll just get this; this again simple computations will take you from
𝟏 + 𝑮′ (𝒔)𝑯′ (𝒔) to 𝟏 + 𝑮(𝒔)𝑯(𝒔) right this is nothing special happening over here yeah.

(Refer Slide Time: 32:22)

So, let’s revisit stability for a while. So, I start with the; that the system that looks like this
right. So, this is I know that it’s stable because I have, sorry this is unstable because there
is a pole on the right-hand side. So, anything on the right-hand side unstable right. So, can
I do something with this? Right, can I do something so as to make this system stable? Well,
I just say I just take the negative feedback 𝑯(𝒔) = 𝟑 and I take the closed-loop transfer

334
function set the characteristic equation to 0 and the poles of the closed-loop system are
𝟏 √𝟑
now minus − 𝟐 ± 𝒋 right.
𝟐

So, these are complex conjugate poles, but which are again in the LHP right, in the left
half plane right so this guy; so, all poles are to the left-half plane and therefore, the closed-
loop system is stable right. So, I start with an open-loop system which is unstable, I can
stabilize it with just some feedback right. So, this is one advantage which I get when I do
feedback right I can make unstable systems stable. Is it possible always that give me any
unstable system can I make it stable via feedback well that may not be always possible,
but in several cases it is possible.

(Refer Slide Time: 33:48)

So, what does what are the advantages? The advantages are I can see that I can have
improved stability, I can have decreased sensitivity to parameters, I can deal with
disturbances, I can deal with noise right and we could also increase bandwidth. We’ll see
this when we do the frequency response of the system. But then my overall gain of the
system reduces right because I have some multiplicative factor over there right, so; so here,
so okay I go here and this kind of gives some attenuation which may not always be
desirable. So, that’s one of the drawbacks and of course, I would have increased the cost
to deal with and sometimes a bit of complexity too, but of course, you know the advantages
are a lot more compared to the disadvantages.

335
(Refer Slide Time: 34:40)

So, sometime now we have seen you know how the system could be sensitive to change
in parameters right ah. So, we’ll see does that have any effect on stability? and that is
called relative stability. So, at the moment we just talked about- is the systems stable or
unstable? Right, nothing more right. So, now, given a system can be defined as some kind
of relative stability? Relative stability means can a small change in the parameter of the
system make it unstable right and then that we’ll define it as the; there will also be some
measures of stability that if my system parameters are between so and so, I could still be
stable right.

So, for example, if I am making, cooking food and you know I just download something
from some Khaana dot com or you know all those websites and I say they say one
tablespoon of salt, but if I just put one tablespoon plus 1 percent or minus 1 percent it does
not really kill the taste right. But if I put two tablespoons that might be awful, if I put 0
tablespoon that may still not be very good so, but then there is some amount of relative
stability right. You can go slightly higher or slightly lower and still maintain stability right
stability in terms of the taste, but if I am making
Tandoori chicken and it says put on the oven for twenty minutes, but if I say 22 minutes
then I might you know not really get something very edible right. So, that that, that little
thing could be unstable over there, okay 20 minutes 1 second could be ok. But if I do 20
minutes plus 2 more minutes that could you know give me something very strange right.
So, so these are things which you know changes in little parameters sometimes affect

336
stability sometimes wouldn’t affect stability. So, if I am looking that I am, I know I am
playing a cricket match and my team is chasing a score of something in some 20 overs and
at the end of 15 overs I am you know I still need 40 runs with 5 wickets in hand I could
still say well I can still win the game right. Because my required rate is not too much, but
all of sudden if I lose 2 wickets right then that could be recipe for disaster I could go from
a winning zone directly to a losing zone right.

So, these are things where you know, you are stable, stable, stable and you can all of
sudden become unstable, and that is exactly what is the concept of relative stability. So,
we don’t really deal with food or sport here we deal with physical systems. So, what is it
deal mean in our case right.

(Refer Slide Time: 37:44)

So, we look at our stability in terms of the location of the poles right well typically complex
poles or real poles or whatever right. So, so what I said is in terms of systems if I say well
I am here, here. So, these are my poles and I say I have a system which typically would
look like this K, a G, a H, the plus, a minus, this is my R, this is my C or the output and
say I had design my system I work very hard and I say well my system is stable it has poles
at −0.1. Okay.

So, I am given a fantastic design problem and I just say this my poles are at 𝟎. 𝟏 ± 𝒋
whatever say 𝒋 𝟐, right I work very hard and I submit this as my thesis and I say my
controller works because my poles are here, right. So, these are my poles and two days

337
later you know my supervisor comes and checks it, you know some something would have
gone wrong there right you know my mass would have been a little different than earlier
or you know some rats might have chewed on to something where my mass slightly
decreases and I when, when I run the system again. So, my maybe you know these guys
just jump here right or maybe I did it in air-conditioned environment the power goes off
and I am presenting this to the examiners or my supervisor in a different environment
where the parameters might change and say oh what I see the response oh. So, it goes
unstable right.

So, my poles could just be instead of minus 0.1 it would just be 𝟎. 𝟏 ± 𝒋𝟐 this could happen
right. So, and we’ll see there are lots of cases where this will happen and therefore, when
I design things I will say well I don’t just, don’t want stability, I want thing like more
stable right. It’s, it’s like this right somewhere. So, so I want my poles to be somewhere
here not just left of 0, but say at −𝟏 or further left of −𝟏. So, I want to be to have a little
buffer for myself to do these things right.

So, these are things which we can talk about you know in of relative stability, now do we
have any tools where I can say where we can actually quantify relative stability? We have
learnt the Routh table right. So, let’s see if I could say something about relative stability
may not be very useful in terms of exact analysis, but some preliminary analysis I could
do about relative stability right. So, I have a closed-loop characteristic equation where I
just you know factorize that, them in terms of n poles right. So, these are direct relation
over here.

(Refer Slide Time: 40:38)

338
So, the relative stability can be determined by using Routh Hurwitz criterion by just
shifting the axis right.

(Refer Slide Time: 40:53)

So, if I say start with this guy you know say consider a system, a third-order system and I
say is it stable? I do the Routh thing and I say everything is positive, positive, positive, no
sign change and therefore, my system is stable.

(Refer Slide Time: 41:21)

339
But this does not help, because you know if my temperature changes slightly or some little
thing goes wrong here and there, then I go to unstable, therefore, I will pose a slightly
different problem to be a little more comfortable and say well 𝑹 the plane. So, are all the
poles over here too, are they at left of −𝟏 until now I was only interested in left of 0, but
if I have at −𝟏 I have a little margin here right. So, if the temperature increases which
shifts the pole may be from −𝟏 to 0.8 I am still stable right and therefore, I want to push
them you know as slightly further to the left.

(Refer Slide Time: 41:51)

340
So, the way we check this is like this. So, let me define a new variable called 𝒔𝟏 which is
𝒔 + 𝟏. So, my 𝒔 now will become 𝒔𝟏 − 𝟏 and I rewrite. So, this is my new coordinate
frame where the origin is (−𝟏, 𝟎) . So, if I take a characteristic equation which is
𝒔 + 𝟐 = 𝟎 which has a pole at −𝟐 how will this pole look in like in the new coordinates.
So, my 𝒔 is (𝒔𝟏 − 𝟏) + 𝟐 = 𝟎 this means I have 𝒔𝟏 + 𝟏 = 𝟎 this is okay right. So, in my
new coordinates sorry in my old coordinates I was at a distance of −𝟐 from the origin if
this is the new origin at −𝟏, I should be at a distance of −𝟏 from the new origin right. So,
that is what this guy tells us.

So, simply put if I were to check for a general polynomial here. So, are the roots of this
equation to the left of −𝟏? I just substitute this guy right. So, in the new coordinates 𝒔𝟏 I
just show this substitution my characteristic equation looks like this now I do the entire
Routh table thing for this one right. So, this is a 1 4 1 4. So, this row will go to 0 and I’ll
have auxiliary polynomial and I’ll just solve for the auxiliary, you know you know
differentiate this and then I do all these things. So, first is there is no sign change and
therefore, there are no roots on the shifted plane there are no roots which are on the right-
hand side of −𝟏.

However, there are two roots here. So, we can say this is in some sense called marginally
stable with respect to this axis, of course, in the original coordinates it will not have a
sustained oscillatory behavior that we should be little careful of. So, I am just saying well
it is not left of −𝟏, where they’re a couple of guys who are just sitting on −𝟏 marginally
stable with respect to this new shifted coordinate that is −𝟏 and 0.

(Refer Slide Time: 44:15)

341
So, what we have done so far, is we started with the concept of stability we defined some
criterion for BIBO stability, we also found out how we given a transfer function how I
computed stability, given then higher-order polynomial we had an algebraic method to
compute are there any unstable poles, we looked at closed-loop system, advantages of
feedback and some analysis on relative stability some very basic analysis on how stable
my system is. So, what we’ll do next is to focus on these kinds of things of relative stability,
and see if there are more sophisticated analytical tools which is called the root locus
method. We’ll see how my pole locations respond which changes to certain things in the
systems.

(Refer Slide Time: 45:08)

342
So, these are essentially you know conditions divide no derived by Evan and called the
root locus techniques, we’ll study some construction rules of this graph called the roots
locus technique. Okay.

Thank you.

343
 Control Engineering
Dr. Ramkrishna Pasumarthy
Department of Electrical Engineering
Indian Institute of Technology, Madras

Module - 05
Lecture - 01
Root Locus Technique

In the previous class, we concluded with the concept of relative stability; That if I want
my poles to be not in the left-half plane, but left of -1, can I do that? or does my Routh-
Hurwitz criterion give me an answer to that? The answer was yes, you know I could find
out if the roots are to the left of -1, slightly to the right or on the axis itself, but that may
not necessarily be enough when I am handling design problems right. So, what we will do
today is to learn a technique called the root locus, which will tell me a little more about
relative stability right. So, what is this root locus all about?

(Refer Slide Time: 00:59)

So, when I design a system. So, I want the system to be such that it meets the desired
specifications, and if we go to our earlier lectures the specifications were usually either in
the transient response or in the steady-state response. So, in the transient response, I
wanted to speed up the response, but then I had a constraint of the overshoot. So, I wanted
to speed up in such a way that I don’t overshoot too much and I also wanted the system to
have well very small or no steady-state error at all, but before I do this I always wanted
that the system should first be stable, if it is stable then I can look at, at performance right.

344
So, until now what we did? we just followed some analytical approach while we do while
we were doing the Routh-Hurwitz criterion.

So, it does not provide even though it gives me some information of relative stability it
does not give me too much information; I may need something more than just ask, I just
ask the question do the roots lie to the left of -1 or left of -2 and so on right? And it does
not really help in choosing parameters in that way right? it does not help in choosing a
control where you know I really want a little more fine-tuning of the system parameters.

(Refer Slide Time: 02:16)

Okay so, if I, if I look at a typical closed-loop system like this right and I just ask a question
what happens if I change this gain? It could be from all values possible values from
−∞ to 0 and to +∞, what happens to the system well the Routh-Hurwitz doesn’t really
give me an exact answer to this question, neither does it tell me any information about the
settling time or the rise time, overshoot or even the steady-state error. So, for those things
I need something little more sophisticated or little more, more in detail right. Okay. So,
let’s see are there any techniques like that right. So, as again so what we will be interested
in is the characteristic equation, the solutions of which will give me the location of poles
of the closed-loop system. So, we would like to know how the poles of the closed-loop
system vary by changing the gain 𝑲.

(Refer Slide Time: 03:24)

345
So, this thing or this the locus of the migration of the roots of the characteristic equation
as the gain 𝑲 varies is called the root locus techniques introduced by W. R. Evans, you
know in his paper in 1948. So, what is this guy? this is, this root locus is a technique, it’s
a graphical technique for sketching the locus of the roots of the characteristic equation as
a certain design parameter.

Here, it’s this gain 𝑲. It could be something else also, we’ll see how we deal with those
when this design parameter is varied and these are the first two, two papers where this root
locus was first introduced. Okay so let’s, let’s not get into that at the moment let’s start
with something, something very small right.

(Refer Slide Time: 04:12)

346
So, I start with a second-order system if I were just to draw it I have in my forward transfer
𝟏
function , I have 𝑯(𝒔) = 𝟏, a +, a - here again all these usual things of 𝑹 and 𝑪 as
𝒔(𝒔+𝒂)

one right. So, the characteristic equation of the system 𝟏 + 𝑮𝑯 = 𝟎 gives me something
like this okay. Now, the roots okay so, I think I should also put a gain 𝑲 here right, another
block. So, this will give me the characteristic equation in this way. So, the roots of this
−𝒂±√𝒂𝟐 −𝟒𝑲
equation, second-order equation I have two roots, .
𝟐

So, let’s analyze what this, how these roots change and what are the two parameters here?
𝒂 changes and 𝑲 changes. So, for positive values of 𝒂 right? when 𝒂 is greater than 0; −𝒂
is always less than 0. Additionally, if 𝑲 is also greater than 0 system is always stable right
and because of this guy right, and the roots of the system always lie in the left half. Okay,
the relative stability now depends on the location of the roots right.

So, then I could see that by varying 𝑲 which corresponds to my 𝝎𝒏 in the second-order
analysis this guy corresponds to 𝟐𝜻𝝎𝒏 , you know dealing with the damping coefficient
also right. So, the desired transient, the transient response can be obtained by just changing
this may be if 𝒂 is fixed. So, given a transfer function, 𝒂 could be fixed. So, what happens?
So, given 𝒂 there could be few cases because the roots of the characteristic equation
depend on what is the sign of this guy.

(Refer Slide Time: 06:34)

347
 𝒂𝟐
So, for case one when 𝟎 ≤ 𝑲 ≤ , the roots are distinct right I can just see very easily.
𝟒

When 𝑲 = 𝟎 the roots are at 𝒔𝟏 = 𝟎 and 𝒔𝟐 = −𝒂. So, let’s plot that for. So, okay it just
be plotted here. So, I start with 𝑲 = 𝟎, I am here that is 0 and I am here this is −𝒂 in the
𝒂𝟐
s plane 𝝈 and 𝒋𝝎 right; for 𝟎 ≤ 𝑲 ≤ . So, this guy would be real in that case. So, the
𝟒
roots would be distinct. So, they will be somewhere I know that that there will be distinct
they will just be real, just some somewhere on this in this axis could be here and wherever,
but I know that a stability is always maintained because 𝒂 and 𝑲 are positive you don’t
even need to look at the Routh table right here maybe you can just trivially see that the
Routh condition is satisfied.

And therefore, when we were doing the second-order system analysis for given 𝛇 and 𝝎𝒏
which was always positive we never worried about stability, because stability was very
obvious in that case. You can now see that the Routh criterion satisfies this very trivially
𝒂𝟐
and therefore, we were really never worried about stability. For 𝑲 = , this guy
𝟒
𝒂
disappears and I have equal roots, 𝒔𝟏 = 𝒔𝟐 = − 𝟐,. So, let me just say that this 𝑹 when
𝒂𝟐 𝒂 𝒂𝟐
𝑲 = 𝟎, and when 𝑲 = , well this is at − 𝟐 occurs at 𝑲 = . So, when 𝑲 changes from
𝟒 𝟒
𝒂𝟐
0 till at least it reaches , what I see that this guy starts moving here and this guy starts
𝟒
𝒂𝟐
moving here until they both meet at this point right at 𝒂𝟐 , at 𝑲 = and this point which
𝟒
𝒂
is − 𝟐.

𝒂𝟐
Now, what happens after this? So, 𝟎 ≤ 𝑲 ≤ . Well, I just say that these are somewhere
𝟒
here, well I knew that they were somewhere on the real line, but now I know that even
𝒂𝟐 𝒂𝟐
𝑲= you know they meet here. So, this guy travels here; what happens for 𝑲 ≥ ?
𝟒 𝟒
Well, the roots are complex and conjugate and given by this number.

𝒂𝟐
So, what you see that after 𝑲 ≥ , 𝒂 is given to us right. So, this guy remains fixed. So,
𝟒

𝒂 √𝒂𝟐 −𝟒𝑲
all the real parts will be at − 𝟐 and the complex part will follow this guy ±𝒋 which
𝟐
𝒂
means well I am just. So, there will be all the real parts are − 𝟐 here and the complex parts
𝒂𝟐
will keep on changing. So, when 𝑲 = , I am here, and then I see that I am just moving
𝟒
around this, these lines, these lines and these lines.

348
 𝒂𝟐 𝒂𝟐
So, this is for 𝑲 ≥ similarly here, 𝑲 ≥ and we know that if there is a root here, there
𝟒 𝟒
always be a root here right because of the nature of the system right. So, the real part
remains constant and the imaginary part varies as the gain 𝑲 varies, thus the roots move
along a vertical line right here 1 1. So, they go this way and this way. So, what do we
observe right?

So, as gain 𝑲 for 𝑲 = 𝟎, I’m just at the open-loop poles okay, as 𝑲 increases both these
guys meet up here and they again go away right to, it’ll just keep going till infinity. So,
something nice is happening here right. So, what information I used? I just used this guy
𝟏
𝑮 is right, I’m not doing anything with the transfer with the characteristic equation
𝒔(𝒔+𝒂)

now, and I’m just interested in how this 𝑲 𝑮 and then now this will be the closed-loop
okay.

I have not drawn this one, this is. So, I am just taking this information right this is this guy.
Okay so, what I observe for 𝑲 = 𝟎, I am at the open-loop poles and as 𝑲 increases these
guys tend to move and they direct they decide to move they decide to meet here and then
again separate.

(Refer Slide Time: 11:47)

So, this is how I just play a very small video of this right. So, this is how the roots will go;
Okay something you saw there right. So, if I give. So, again I’ll just play this again. 𝑲
increases, so this guy will go and then for lowering values of 𝑲 and then go small and then

349
again come back to their position. So, this is how the roots will move. So, this guy will go
here and move this guy direction this guy will go here and move this direction right.

(Refer Slide Time: 12:17)

So, what did we know? what did we see here that? So, so the root locus indicates the
manner in which the closed-loop zeros and poles move so that the response meets some
specification, just my specifications are usually in terms of 𝛇 and 𝝎𝒏 and these
specifications usually translate to locations of poles. So, this could be some 𝛇 and 𝝎𝒏
𝒂𝟐
which would be maybe satisfied by 𝑲 = right then I would know right.
𝟒

So, I exactly know where the roots are maybe somewhere here and somewhere here right
and then I would actually, exactly know what should be the value of 𝑲 such that my
closed-loop pole satisfies some, some desired specifications or some desired performance
criteria okay. So, I would know how to modify the gain 𝑲 such that my performance is
met and performance is usually mapped to some pole locations okay.

And I am just doing this only with the knowledge of this guy right, I started with 𝒔 the
open-loop poles 𝒔 at 0, 𝒔 at −𝒂 and I am just now increasing the gain 𝑲, we’ll see how,
how we’ll formalize this a little later right. So, just by looking at these two, I start with the
open-loop poles and these, these are all the closed-loop things now right? As 𝑲 increases
I am already in the closed-loop right; then how the closed-loop poles are modified starting
from the open-loop poles. So, I just need to know how the open-loop system looks like
and I can just do a good study of the closed-loop system. So, I don’t really need to compute
the closed-loop poles and then analyze things and so on.

350
(Refer Slide Time: 14:09)

So, how are poles related to? pole locations related to the performance? Well, we analyze
this while we were doing first-order systems. So, just say if for different values of poles at
𝒔𝟏 I will have a faster response 𝒔𝟐 slightly slower, and the slowest will be for 𝒔𝟑 right. So,
this is like how the how the location of poles influences the performance this we did while
we were doing the analysis of first-order systems okay.

Now, in general, so the first order system was very easy right. So, all the poles are on the
real line; then based on their location I would determine how are they fast? are they slow?
The slowest would be here and the fastest would be very you know further away from the
origin.

351
(Refer Slide Time: 14:56)

However, for second-order systems there were little more details right, we started well we
had this un-damped system where we had the maximum overshoot, the settling time was
infinity, then we had the case of underdamped systems where we had quantified or we
defined underdamped, nature in terms of 𝛇 right and then we had also when the system
was un-damped the frequency of oscillation was 𝝎𝒏 , for 𝝈 = 𝟏 the system was critically
damped and then we also had the notion of over damped system where this 𝛇; So, it was
greater than 1.

(Refer Slide Time: 15:33)

So, if I just look at, you know the poles I typically well I am just looking at a solution to
this equation, right? 𝒔𝟐 + 𝟐𝜻𝝎𝒏 𝒔 + 𝝎𝒏 𝟐 okay. So, if I just look at the underdamped case

352
my roots were −𝜻𝝎𝒏 ± 𝒋𝝎𝒏 √𝟏 − 𝜻𝟐 , this is for 𝛇 less than 1, if we remember it properly
and this guy I said this is 𝝎𝒅 . So, this is the real part here. So, if I just take a complex
conjugate pair right 𝑷𝟏 and 𝑷𝟐 the real part, this distance would be this will be at −𝛇𝝎𝒏 ,
and then this guy would be 𝝎𝒅 this distance. 𝝎𝒅 = 𝝎𝒏 √𝟏 − 𝜻𝟐 right this is just like. So,
this is a right angle triangle and I could say that this 𝝎𝒏 is just you know square root of
this guy plus this guy right from the Pythagoras theorem; Okay, now something else right.

So, this is what we also call as the 𝝎𝒅 right. Okay now, look at this guy 𝜽 right and if I say
𝜻𝝎𝒏
how does 𝐜𝐨𝐬 𝜽 look like. 𝐜𝐨𝐬 𝜽 is okay, this guy which is just 𝛇 right. So, from this
𝝎𝒏

𝐜𝐨𝐬 𝜽 =𝛇 and therefore, 𝜽 is 𝐜𝐨𝐬 −𝟏 𝜻. Okay so, let’s look at this right. So, if I am at this
point 𝜽 remains the same whereas, 𝝎𝒏 changes right if I go here, if 𝜽 is the same then 𝛇
is the same. Okay.

Now if I go from here till here, this distance which is 𝝎𝒏 changes, but 𝐜𝐨𝐬 𝜽 still remains
the same right slightly here also if I go from here till here, what happens? my 𝛇 remains
the same, because 𝜽 is the same, right? if I’m moving along this line right and then 𝝎𝒏
changes right for some given 𝛇. So, all these lines are lines which I call as constant 𝛇 lines.

So, I can okay go here, right? If I just take any arbitrary line here right for which a 𝜽 would
be fixed, if I move along this line right what happens is my θ is fixed, whereas 𝝎𝒏 changes
okay. So, this behavior could be seen from 𝜻 = 𝟎, I am here I keep on increasing 𝛇 and
there are all these, all these lines are the lines for which the 𝛇 remains constant. So, at this
point, you will have different 𝝎𝒏 and this point you’ll have different 𝝎𝒏 , but the 𝛇 here
is the same similarly on the other side. Okay? So, these are all constant 𝛇 lines or we just
say how the locus of the poles looks like for constant 𝛇.

Now, let me keep 𝝎𝒏 fixed and keep moving right and I, I just want to vary θ. So, what
will I do? Well, 𝝎𝒏 is fixed; So, 𝝎𝒏 is fixed and I just move around some semicircles, I
don’t want to do a circle because I don’t want to go into the unstable region. So, if you
look at, you know this guy over here 𝝎𝒏 is just same, 𝝎𝒏 is same, but each of these will
have different 𝛇 right this will have a different angle, this would have a different angle and
this would have a different angle.

353
So, if I if I keep on moving 𝝎𝒏 then I just see I just move, I’m just moving along the circle
and this guy also you know since all poles exist in complex conjugate pairs. So, this guy
will also move just according to how the 𝝎𝒏 varies. So, you just have some semicircles
here. So, there is a pole here, there is also pole here and this will correspond to certain 𝛇,
if I go here this pole will move here and this pole will keep on will move here right the
complex conjugate, but the 𝛇 will move.

So, these are semicircles where 𝝎𝒏 is constant and 𝛇 is changing, these are lines where 𝛇
is constant and 𝝎𝒏 is changing. Right? Okay, what is the importance of this? When I say
well design something such that 𝛇 is some number and 𝝎𝒏 is also some number. Say, 𝛇 is
say 𝟒𝟓° and 𝝎𝒏 = 𝟏 okay. So, I have this constant 𝝎𝒏 circles and then I, I just draw a
line of say 𝛇 or also the θ being 𝟒𝟓°. So, I just draw a line from here till here and if say
this is 𝝎𝒏 = 𝟏 then my pole should be here. Okay?

So if I have, again I will repeat; if I want a design specification where 𝛇 should be, say
some number which corresponds to the angle of 45 degrees, given a given θ, I can always
compute 𝛇 right and then add a certain 𝝎𝒏 let’s say this is 1. So, I just find these two lines.
So, this is my 𝛇 right and then this is my 𝝎𝒏 = 𝟏 and the point where they intersect is
where I want my closed-loop poles to lie and this I know how you know from the previous
example that I could vary this closed-loop poles just by not varying the gain. Now I just
investigate is there any gain which will do me this? Now, how do I test right?
So, that is exactly what we will study in the root locus case. Okay?

354
(Refer Slide Time: 21:32)

So, this was just, just a little example right, and now we’ll see how we can generalize this
right? Do the roots? So, what we saw that the roots were, were traveling towards each
other they meet and then they go away do they always meet and if they meet do they again
have to go, do they have to go to the imaginary axis? or they could possibly just stay over
𝒔𝟏
there? There was no zero in our example right it was just , what happens if there are
𝒔(𝒔+𝒂)

zeros in the system.

So, all these things we will try to analyze a little more, this will lead us to some conditions
on the root locus right. So, all these rules are what we will do in our next lecture.

Thank you.

355
 Control Engineering
Dr. Ramkrishna Pasumarthy
Department of Electrical Engineering
Indian Institute of Technology, Madras

Module - 05
Lecture - 02
Root Locus Plots

In the previous lecture, what we had seen was given a second-order system in the closed-
loop, we saw how the roots moved as the gain increased or the gain was increased from 0
to infinity, and we saw some strange behavior that the roots were moving towards each
other, they meet and then they go to towards infinity and so on. So, all this was; the
motivation for this was to try and understand the concepts which we had not, which we
were not very clear in the Routh stability criterion with respect to relative stability.

So, we’ll today formalize those ideas into a set of rules and see for given a general system,
how would the poles and zeros move; especially the poles move as the system gain is
varied from 0 to infinity.

(Refer Slide Time: 01:15)

And this is called the Root Locus Plots. So, how does it go? So, let’s consider a typical
closed-loop system where I have the reference signal, a certain output, a gain 𝑲, 𝑮(𝒔) and
a 𝑯(𝒔) 𝑮(𝒔) in the forward loop, 𝑯(𝒔) in the backward in the feedback loop plus and minus

356
and so on and I know that for this set of system the overall transfer function, 𝑻(𝒔) is simply
𝑲𝑮(𝒔)
and we by now, know how to derive this.
𝟏+𝑲𝑮(𝒔)𝑯(𝒔)

So, this is it again just ratio of some polynomial 𝒑(𝒔) in the numerator and 𝒒(𝒔) in the
denominator, and we’ll see how the gain varies as sorry how the system behaves as the
gain 𝑲 is varied from 0 to infinity, we could also do for 𝑲 being negative right. So, not
minus infinity, but that is a little straightforward extension once we know how it behaves
when 𝑲 goes from 0 to infinity.

(Refer Slide Time: 02:18)

So, where do we begin with well the characteristic equation which we had defined last
time is simply the denominator going to zero right? The 𝒒(𝒔) = 𝟏 + 𝑲𝑮(𝒔)𝑯(𝒔) = 𝟎.

So, I just do some manipulations 𝑲𝑮(𝒔)𝑯(𝒔) the 𝟏 goes here I call 𝑲𝑮(𝒔)𝑯(𝒔) as some,
𝑷(𝒔) and it is −𝟏 and what is 𝒔? the argument of 𝑷 is a complex number right, 𝝈 + 𝒋𝝎
right when we do; draw things in the complex plane. So, 𝒔 = 𝝈 + 𝒋𝝎, 𝑷(𝒔) is also then a
complex number. Since 𝒔 is a complex variable 𝑷(𝒔). So, this −𝟏 can also be then written
as a complex number right; so with this guy. So, I can split this complex number into its
magnitude, its phase and correspondingly this guy can also be split into its own magnitude
and phase and we’ll see what this exactly looks like right in terms of the magnitude of
|𝑷(𝒔)| and the ∠𝑷(𝒔).

357
So, what does it mean? right. So, this from here to here that the magnitude of this number
−𝟏 + 𝒋𝟎 should be the magnitude of P, this |𝑷(𝒔)| = 𝟏 and then the angle should be the
angle of this. So, what is this angle? So, this angle is 𝟏𝟖𝟎° or in a way odd multiples of
𝟏𝟖𝟎°, ±(𝟐𝒒 + 𝟏)𝟏𝟖𝟎°. So, what does this mean? right. So, again I am looking for the roots
of the characteristic equation as 𝑲 varies. So, how do I test if a particular point 𝒔𝟎 is a
solution to this equation? I am just finding solutions to this equation.

So, this could be seen just by satisfying the angle criterion and the magnitude criterion
right and more we’ll slowly build up a set of rules based on just these two things right, and
slowly you’ll realize that what we do in the rest of the course or most of the course which
is pending from now on is just analyze complex numbers nothing much.

(Refer Slide Time: 04:36)

So, a system with the loop transfer function of the form 𝑲𝑮(𝒔)𝑯(𝒔) is again can be written
as 𝑲, a set of zeros, set of poles, 𝒏 ≥ 𝒎. So, this gave us what we called as 𝑷 in the
previous slide right. So, 𝑲𝑮(𝒔)𝑯(𝒔) was 𝑷 and we need to find the existence of a point or
conditions of when does a point lie on the root locus? Well, the answer is whenever it
satisfies these two conditions this we’ll call this as the magnitude criterion and this the
second condition as the angle criterion.

So, how does this look like? So, if I just take this guy here and try to compute its magnitude,
this will be 𝑲|𝑮(𝒔)𝑯(𝒔)| should be 𝑲 and then okay 𝑲 is a positive number so I don’t
really draw the magnitude sign here and then the magnitude of all the zeros divided by

358
magnitude of all the poles right and this should be such that the overall magnitude is 1. So,
𝑲 with the product of all magnitude; magnitude of all zeros divided by the product of the
magnitude of all the poles you know 𝒔 + 𝒑𝒊 this guy should be equal to 1 right and
similarly with the angle. So, this is a complex number or a phasor as you could even call
it.

So, the contribution of the angle of a constant is zero there will be some angle
contributions. So, these are in the numerator. So, this will be summation of all the angles
of the zeros we’ll slowly see how to compute this, minus since these guys are in the
denominator minus the summation of all angles of this poles and what should this satisfy
to be on the root locus this should satisfy this angle criterion. So, the summation of all
angles of zeros minus summation of all angles of poles should be ±(𝟐𝒒 + 𝟏)𝟏𝟖𝟎° or like
odd multiples of 𝟏𝟖𝟎°.

So, these are the two conditions which we will verify. So, does the point lie on the root
locus? The answer is yes if it satisfies these two conditions. So, what was the root locus
just to recall what we have done last time? So, we had started with an open-loop pole at 0
and at −𝒂, and what we saw that as the gain increased these two guys meet at some point
𝒂
that is − 𝟐 and then they just go to infinity. So, this movement of poles as the gain 𝑲
increases from 0 to infinity is the root locus and you can see and we’ll see slowly of why
what this plot means in terms of this and this criterion.

(Refer Slide Time: 07:30)

359
Okay so, what’s a criterion? Every point 𝒔 in the complex plane that satisfies this equation,
lies on the root locus of the system with the loop transfer function given this right. So, in
this, I am just looking at the open-loop transfer function right. So, how does then the
magnitude criterion simply becomes? So, from here 𝑲 is. So, this guy will go to the
numerator; the product of all magnitude of all the poles and then the zeroes right. So, this
should satisfy something like this, 𝑲 should be such that these two things should be
satisfied for my any root to lie on the root locus. So, let’s see what this means right? So,
in terms of an example.

(Refer Slide Time: 08:19)

𝑲(𝒔+𝟐)
So, I have . So, give me. So, let me just draw something right. So, I just mark my
𝒔(𝒔+𝟏)

poles at 0 and −1 and there is a 0. Now, I say, say give me any arbitrary point here right
this guy. Now, I see if the gain increases will this you know. So, so as the gain keeps on
increasing will this point lie on the root locus? The root locus could be any kind of curves
on this complex plane right, they would also go to the unstable region, what we’ll see that
a little later. Okay so, for any point 𝑠0 to be on the root locus it should satisfy the angle
criterion right. So, I’ll have well this guy, this guy, and this guy and I just measure the
angles this way.

So, this is how it measures the angle of. So, this is zero at −2, it’s a pole at −1, pole at 0,
zero is in the numerator. So, the angle of this guy minus the angle of 𝒔 of this guy minus
the angle of this guy, all this should be, you know should satisfy this equality. Similarly,

360
𝑲 should satisfy this guy right so this one the magnitude criterion. So, summation of this
magnitude 𝒔𝟎 + 𝟏 and 𝒔𝟎 + 𝟐 right; so, this is what this picture says.

(Refer Slide Time: 10:05)

So, if I just were to draw it. So, this A would be the distance of this pole from my desired
or the point which I’m testing if it is on the root locus this is A, similarly for −1 it is B
and C. So, this guy. So this, if these numbers satisfy this equality then the point is on the
root locus together with satisfying this conditions. So, I will. So, then these are the
corresponding angles.

(Refer Slide Time: 10:38)

361
Okay so, the first rule right. So, when whenever we draw a root locus, well, what are we
interested in? So, let’s just recall the earlier example we did right. So, I had, I had 0 a pole
at −𝑎, and the root locus is and also this guy moves this side and then they go to infinity.
The first thing is the root locus is symmetrical about the real axis and the number of
branches is equal to the order of the characteristic polynomial, or the number of poles of
the open-loop transfer function or the number of poles of 𝑮(𝒔)𝑯(𝒔). Okay so, what does
it mean?

So, first if I, so these are this is this the root locus is such that it will always give me the
location of the poles for different values of gain 𝑲. So, if there is a pole here, there is
always a pole here because all the roots exist in complex conjugate pairs. So, you take any
equation or any polynomial, 𝒔𝟔 , 𝒔𝟖 you see that all complex poles end up or they are in
common in they, they come across as complex conjugate pairs there just be no pole here
right and therefore, we can say that if there is a pole here, they already be pole here, if
there is a pole here well, it does not matter there is a pole here, there will be a pole here
and so on.

So therefore, the root locus, these roots could be either they are real, they could be
imaginary or complex conjugate right, but then if I just look at this thing if there is a pole
here, there is a pole here, there is a pole here, there is a pole here and so on. And therefore,
I can say that well, whatever happens, it will always be symmetric along the real axis. So,
the first thing is that the root locus is symmetric about the real axis or in other words if I
just take the positive half here, and I put a mirror here this guy will just be the mirror image
of this guy right and okay, how many branches will the root locus have the root locus will
have number of branches that is equal to the number of poles right. So, here I have two
branches right? One branch comes from here and goes here, another branch comes from
here and goes here.

Now, what decides where that why is this guy going here? why don’t they do here and
then come back to their own position? or why don’t they do this and then go this way?
Why only perpendicular? why not this way? or why not just keep circling all these things?
which we’ll say shortly.

362
(Refer Slide Time: 13:03)

The second rule says that all branches of the root locus they start at the open-loop poles
when 𝑲 is 0, and end at either open-loop zeros or infinity as 𝑲 goes to infinity. So, like
𝑲
what we saw last time right we had a transfer function of the form 𝒔(𝒔+𝒂) and we had two

poles here, a pole at 0, and a pole at −𝑎 and these two guys well, they started moving one
guy to the right, other to the left they meet and then they diverge; so the number of
branches. So, see that two branches meet and then they go to infinity. Now, why do these
two go to infinity? Well, the rule says at the number of branches terminating at infinity is
equal to the difference between the number of poles and the number of zeroes of 𝑮,
𝑮(𝒔) 𝑯(𝒔).

So, here there are two poles and no zeros. So, both of them will disappear to infinity, okay,
now how to generalize this right? So, our characteristic equation is of this form
𝟏 + 𝑲𝑮(𝒔) 𝑯(𝒔) written as you know the product of zeros and poles it looks like this. So,
what am I looking at now? I am looking at, what happens when or what are the solutions
when 𝑲 = 𝟎 solutions means values of 𝒔. So, I take this equation I substitute 𝑲 = 𝟎. So,
this term goes away. So, I am left with this guy. Right?

363
(Refer Slide Time: 14:41)

So, I am. So, what are the solutions? What are the values of 𝒔 here, such that this
expression goes to 0? Well, these are precisely the poles of the open-loop system. Thus
for all the open-loop poles are a part of the root locus, and the root locus exactly starts
from those poles and these are again the open-loop poles.

Now, where do they go? What about the zeros? I can rewrite this equation, this get divide
by 𝑲 and I get something like this right now the equation can be rewritten in this form.
Now, what am I looking at? I start from 𝑲 = 𝟎 and I ask another question, now what are
the solutions for again values of 𝒔 when 𝑲 goes to infinity? When 𝑲 goes to infinity, I
know that this guy disappears; if this guy disappears, what I’m looking at is an expression
which is like this.

364
(Refer Slide Time: 15:46)

Now, what are the solutions to these guys right? The solutions of these guys are the open-
loop zeroes. So, at 𝑲 = 𝟎, we start at the poles and 𝑲 equal to infinity we end well that is
this final point which we look at, we end at zeroes. Okay?

So, what I can say that also the zeroes lie on the root locus and these are the terminating
points. So, I, a pole at 𝑲 = 𝟎, I start from the pole and I end at the 0, okay but what we
know is that the number of poles is usually greater than the number of zeros right that
could also be equal. So, but these are 𝒎 number of poles are 𝒏. So, what I could say is
that only 𝒎 poles terminate at zeros right? Okay, what happens to the remaining 𝒏 − 𝒎
poles.

365
(Refer Slide Time: 17:06)

So, if I say- well this remaining 𝒏 − 𝒎 poles they go to infinity. So, let’s see- okay, what
this could mean? Again, I’m looking at solutions of the characteristic equation which is
𝑲 ∏𝒎
𝒊=𝟏(𝒔+𝒛𝒊 )
𝟏+ ∏𝒏
= 𝟎. Okay? So, let’s do a little example and I say well I take an example
𝒊=𝟏(𝒔+𝒑𝒊 )

like. So, I say 𝑮(𝒔)𝑯(𝒔) = 𝟎 and I want to see okay what, how does this look like? This is
𝟏 𝟏
(𝒔 + 𝟐)(𝒔 + 𝟑) + (𝒔 + 𝟏) = 𝟎. So, I want to see. So, this is exactly what we had right?
𝑲 𝑲
the product of the poles and zeros this guy should go equal to 0. So, for sure one solution
is 𝒔 = −𝟏 as 𝑲 goes to infinity, but this guy disappears and I’m just looking at this
solution.

Okay. Now, are there any other solutions to this equation? So, let’s rewrite write it in a
𝟏 |𝒔+𝟏|
slightly different way. So, I have = |𝒔+𝟐 ||𝒔+𝟑|, if I just write the magnitude part, and I
𝑲
am interested again in what happens as 𝑲 is going to infinity? one solution again as 𝑲
goes to infinity this guy goes to zero, so 𝒔 = −𝟏. So, what we could always what we can
always or also observe is that well something similar happens even when 𝒔 which is the
magnitude goes to infinity and we have some whatever angle here. Right?

Another solution could be just, you can just substitute it here and you take the limit and
𝟏
you see that as 𝒔 goes to infinity this guy goes to 0 right? This is to and check all the
𝒔𝟐
stuff. So, another solution is 𝒔 going to infinity right? Okay, what happens then at 𝒔 equal

366
to infinity? The value of the transfer function or the open-loop transfer function 𝑮(𝒔)𝑯(𝒔)
you know then when 𝒔 is infinity of course, with 𝑲 also being infinity this guy goes to 0.
Now, these are what I call as zeroes at infinity because even at this value at infinity the
value of the transfer function goes to 0, right? Because there is a definition right, 𝑮 at some
zero is 0 right? Now, take this example right?

𝟏
So, if I just take the other, the other day, we were doing the example of , 𝒔 + 𝒂 right?
𝒔
Similar, thing holds here and both the roots are actually now going to what we call as again
zeros at infinity. So, the thing is, the first. So, here there are two poles and one zero, one
pole will end up at a zero which is −𝟏 and the other pole will go to the zero at infinity.

(Refer Slide Time: 21:16)

Now, the third rule will say that, okay take a point on the real line say for example, this
point right here. Sorry, say this point and say does this point lie on the root locus? Well,
let’s see if we try and find an answer to that. A point on the real axis lies on the root locus
if the sum of the poles and zeros on the real axis to the right is an odd number.

So, this sum is actually just the number of poles plus the number of zeros, it’s not the
values of the poles right? Let’s just say, for example, I am looking here, from here to the
right, I just say well there are two poles and one zero, three poles and one zero, four poles
and one zero, four poles and two zeros and so on and I take the total numbers. I have here
I have three poles and zeros, four poles and zeros as well right. So, we just look at the
number of poles and zeros, not the values of it. So, why is this true? So, let’s. So, again

367
the root locus should satisfy the angle criterion. So, let me take a point here and I call this
point as s not.

Now, what is the angle? Okay, sorry there is already s not is mentioned here. So, this is s
not right. So, what happens at s not if I measure the angle well from this pole contributes
an angle 𝟏𝟖𝟎°, this pole contributes an angle 𝟏𝟖𝟎°, this pole contributes an angle of 𝟏𝟖𝟎°.

So, this is a zero. So, it will contribute an angle of +𝟏𝟖𝟎°, right? So, all these angles will;
this will be the contribution, what about these guys? If I just measure the angle from the
right so, they will just be 𝟎° they don’t contribute to the angle; so 𝟏𝟖𝟎°, 𝟏𝟖𝟎°, 𝟏𝟖𝟎°. So,
we’ll see if this point satisfies the angle criterion or not.

(Refer Slide Time: 23:19)

So, why don’t we what about conjugate complex conjugate poles? Well, the sum of the
angles is always 0. So, as you can see from here right. So, this angle sorry this angle here
plus this angle will always go to 𝟎° or 𝟑𝟔𝟎°, the angle contribution by every poles and
zeros on the real axis to the right of the test point is 𝟏𝟖𝟎°, these are to the right to the right,
to the right, it’s 𝟏𝟖𝟎°, and then to the left of the test point is always 𝟎°.

These guys don’t contribute 𝒑𝟑 , 𝒛𝟐 , 𝒑𝟒 they do not contribute. Similarly, with and then
you can see the complex conjugate guys here, the overall angle goes to 𝟎°. Okay so, what
is the total angle contribution? Right, the total angle contribution is angle of all the poles
and the angle of all the zeros. Well, so here 𝒑𝟏 contributes 𝟏𝟖𝟎°, 𝒑𝟐 𝟏𝟖𝟎°, 𝒛𝟏 𝟏𝟖𝟎°. If I
use the appropriate signs, −𝟏𝟖𝟎° for the poles at 𝒑𝟏 , −𝟏𝟖𝟎° for the pole at 𝒑𝟐 , and all

368
these guys are 0, the first zero here contributes an angle of 𝟏𝟖𝟎° with the plus sign and this
is right and this guy is an odd multiple of 𝟏𝟖𝟎° and therefore, 𝒔𝟎 is a point on the root locus
right.

So, therefore, this justifies our thing saying that the point on the real axis lies on the root
locus if the total number of poles and zeros to the right of that point is an odd number.

(Refer Slide Time: 24:58)

Similarly, I could even do for some other point say 𝒔𝟏 which is now. So, earlier 𝒔𝟎 was
between 𝒑𝟑 and 𝒛𝟏 , now I take this point right and I call it 𝒔𝟏 , and I just check is 𝒔𝟏 on the
root locus right and I just look at the angle criterion again. So, again I have 𝒑𝟏 contributing
an angle 𝟏𝟖𝟎° in the negative direction, the pole 𝒑𝟐 again −𝟏𝟖𝟎° . I have a zero that
contributes +𝟏𝟖𝟎°, 𝒑𝟑 contributes −𝟏𝟖𝟎°. Contribution of 𝒑𝟒 and 𝒛𝟐 will be 𝟎° and the
complex conjugate guys will anyways be −𝟎° . Okay? So, let’s do this:
−𝟏𝟖𝟎° − 𝟏𝟖𝟎° − 𝟏𝟖𝟎° − 𝟎° + 𝟏𝟖𝟎° + 𝟎° + 𝟎° = −𝟑𝟔𝟎° right? and 𝟑𝟔𝟎° it does not satisfy
this criteria that it should be an odd multiple. Right?

So, 𝟑𝟔𝟎° is an even multiple. So, if you look at the angle criterion what was the angle
criterion? it should be ±(𝟐𝒒 + 𝟏)𝟏𝟖𝟎°, it should be the ∠𝑮𝑯. So, this is an even multiple
therefore, 𝒔𝟏 is not a point on the root locus.

369
(Refer Slide Time: 26:13)

Okay so, let’s end with an example. So, if I take a system with poles at. So, poles at −𝟐,
−𝟒, a pole at the origin and these two zeros right. So, what’s the first observation is that
the number of poles is three right, and therefore the root locus will have three branches.
Okay, now where will these branches go? that depends on what are the locations of the
zeros right. So, take these two zeros and say well there are two zeros sorry and two poles
sorry two zeros and three poles. So, two of those three poles will go to the zeros and
remaining one will go to infinity.

(Refer Slide Time: 27:00)

370
So, let’s quickly draw the pole zero diagram. So, I have a pole at 0, I have a pole at minus
−2, I have a pole at −4.

So, let me just take for example, say does this point lie on the root locus? I see, well there
is only one pole to the right. So, of course, this is on the root locus, this is also on the root
locus, this entire thing will be on the root locus right. So, this pole will start here and then
move towards this zero, and we know that well, the poles terminate at zero. So, this is
when 𝑲 = 𝟎, this happens when 𝑲 = ∞. Now, look at what happens to this guy, does this
guy move towards the right or towards the left? we’ll again see the same condition number
three. Now, is this point on the root locus? Well, the number of poles plus zeros is 2 it is
an even number even is not allowed. So, this entire region here is not on the root locus.

So therefore, this guy will; Okay now let’s check this region, is this region on the root
locus? Okay, look to the right there is 1, 2 and 3 the total number of poles and zeros is an
odd number and therefore, this entire region is on the root locus. So, this guy will slowly
move this side and then terminate at zero again 𝑲 = 𝟎 till 𝑲 = ∞ okay? Now, what about
this guy? can this guy move towards the right? Well, just see what happens in this region?
I have 1, 2, 3 and 4 poles, which is 4 poles. 1, 2, 3 and 4 are the number of, total number
of poles and zeros, which is an even number and that is not allowed. So, this entire region
is also out of the picture.

Now go here right to the left is this, does it satisfy the condition number 3? Okay say 1, 2,
3, 4, 5 are the total number of poles and zeros of course, this is on the root locus, this is on
the root locus, and this entire this real axis to the left is on the root locus right. Okay so,
what does the second condition say? that when 𝑲 = 𝟎, I start at the poles and end up at
zero when 𝑲 = ∞. simply this happens here there are, in this case, there are 3 poles and 2
zeros. So, 2 poles go to zero, and by the rule 2: one pole goes to infinity or the zero at
infinity and you see that entire real line from here, from to the left of −4 till infinity is on
the root locus. So, this is a third guy that’ll, which will go to infinity.

371
(Refer Slide Time: 29:37)

And this entire this is this will tell me the entire behavior of the system as the gain 𝑲 is
varied from 0 and 0 to infinity. So, these are the first three rules in construction. So, in the
next lecture, we’ll see some other rules related to these things.

Thank you.

372
 Control Engineering
Dr. Ramakrishna Pasumarthy
Department of Electrical Engineering
Indian Institute of Technology, Madras

Module - 05
Lecture - 03
Root Locus Plots

In the previous lecture we saw first three steps about constructing a root locus. First step
was a little easy to understand, that it is actually symmetric about the real axis, then we
also characterized where do the root locus; where does the root locus start when 𝑲𝑲 = 𝟎𝟎 and
what happens as 𝑲𝑲 goes to infinity, and then we concluded that they start at the poles and
end up at zeros and then what if there are more number of poles than zeroes the remaining
poles go to, as what we defined as zeros at infinity. We also easily computed to see if what
areas of the real line are on the root locus. So, that was just a small test based on how many
poles and zeros lie to the right of that particular region of the real line.

(Refer Slide Time: 01:20)

So, continuing on that. The further, the next step here is how do the root locus branches
go to infinity. For example, in the motivating example we saw, that we had this guy a pole
here: 𝟎𝟎 and −𝒂𝒂; that they meet here and they go to infinity. Now, what decides this? why
don’t they go something like this right or why don’t they go something like this or why
don’t they just instead of going on a straight line just you know, follow a lazy path and

373
you know do a little more adventure and go this way. So, who decides how they go or in
this example it was little easier because it was just like a solution of a second order or a
quadratic equation. So, what if there are more poles, who meet at this point and can they
go to infinity in the way they want or there is someone; something that decides that right.
So, the statement says that the 𝒏𝒏 − 𝒎𝒎 roots. So, we are only interested now on the roots
or on the poles which go to infinity as 𝑲𝑲 goes to infinity.

So, the 𝒏𝒏 minus roots that proceed to infinity, they do so along asymptotes with this, this
angle right and this, I’ll show you shortly how this is computed right. So, consider a test
point at infinity right. So, let’s do a little, let’s erase this and okay. I say; okay let’s say I
have my 𝝈𝝈 and 𝒋𝒋𝒋𝒋 axis here and say this is my infinity here. So, there is this guy 𝒔𝒔𝟎𝟎 at
infinity. So, and then there would be you know open loop poles, zeroes and whatever. So,
just say there are these 2 poles right? and I say; I just look at how the angle approaches it,
goes here and then keeps going here and then it reaches this guy.

So, this guy is still very far. So, I just do something right go here and then meet this guy
here. So, this test point which is at infinity. So, what are these angles right? Those are
angles made by the line joining the test point and the open loop poles or zeroes, they are
almost equal to each other right. So, if I really look at 2 points from very far away they’ll
seem just very close to each other right? this is like the infinity and let’s say both of them
kind of make the same angle 𝝋𝝋𝑨𝑨 and 𝝋𝝋𝑨𝑨 .

So, how many numbers of angles are. So, there would be 𝒏𝒏 − 𝒎𝒎 number of angles right,
that would be the difference between the number of poles and the number of zeros. So, the
net angle made by all open loop poles and zeros to the test point would just be well this
guy right, 𝒏𝒏 − 𝒎𝒎 with a negative sign because I’m dealing with poles times 𝝋𝝋𝑨𝑨 right. So,
there are 2 things here right. So, this will be 𝝋𝝋𝑨𝑨 and 𝝋𝝋𝑨𝑨 , 2 times of that right. I think we
are only interested in 𝒏𝒏 − 𝒎𝒎 guys.

So what should these angles do? So, all these guys they should satisfy the angle criterion.
So, what is the contribution at infinity to this guy? So, the total angle contribution at 𝒔𝒔𝟎𝟎
must satisfy the angle criterion. What is the angle criterion? It says that the angle should
be odd multiples of 𝟏𝟏𝟏𝟏𝟏𝟏°. What is the angle contribution? That’s, this guy −(𝒏𝒏 − 𝒎𝒎)𝝋𝝋𝑨𝑨 °
±(𝟐𝟐𝟐𝟐+𝟏𝟏)𝟏𝟏𝟏𝟏𝟏𝟏°
is this guy, and this means I can just compute 𝝋𝝋𝑨𝑨 = (𝒏𝒏−𝒎𝒎)
. So, this angle is now

computed just by this term and this is not surprising right. So, let’s do that the same

374
example again right, I start with 𝟎𝟎, I start with −𝒂𝒂; and let’s say well, I know that two;
these 2 guys go to infinity from some point. So, what is 𝝋𝝋𝑨𝑨 there? 𝝋𝝋𝑨𝑨 is say 𝟏𝟏𝟏𝟏𝟏𝟏° over
±𝟏𝟏𝟏𝟏𝟏𝟏°
what was 𝒏𝒏 − 𝒎𝒎 there it was 𝟐𝟐; with a plus minus, i.e., 𝝋𝝋𝑨𝑨 = . So, what I have is +𝟗𝟗𝟗𝟗°
𝟐𝟐

and −𝟗𝟗𝟗𝟗° that’s what exactly happens right.

So, this pole it starts here and goes to infinity at an angle +𝟗𝟗𝟗𝟗°, and here the other guy this
guy starts from here and goes here at an angle −𝟗𝟗𝟗𝟗°. So, that’s what this guy tells us right,
that the 𝒏𝒏 − 𝒎𝒎 branches 2, in this case. That go that proceed to infinity, do so along
asymptotes with angles which are depending on the number 𝒏𝒏 − 𝒎𝒎 right. If there are more
of these guys the angle will change accordingly, if there is only one say for example, if
𝟏𝟏
this is say, I’m just looking at say for example, right it is my 𝑮𝑮𝑮𝑮. So, I can just draw
𝒔𝒔+𝟏𝟏
the root locus. So, this is going to infinity. So, the angle would just be −𝟏𝟏𝟏𝟏𝟏𝟏° or +𝟏𝟏𝟏𝟏𝟏𝟏°
both are the same right? either this or this, it’ll just keep on going this side.

(Refer Slide Time: 06:47)

Okay now, so these asymptotes, well they’ll somewhere you know they’ll have some
starting point right from on the plane on the complex plane. So, the centroid or the point
of intersection of the asymptotes with the real axis is computed in this way, I guess sum
of all the real part of the poles minus sum of all the real parts of the zeros divided by
number of poles minus number of zeros, this will again be 𝒏𝒏 − 𝒎𝒎; 𝒏𝒏 − 𝒎𝒎. So, where does
this come from? Okay so, what was the open loop transfer function? That’s 𝑲𝑲𝑲𝑲(𝒔𝒔)𝑯𝑯(𝒔𝒔).

375
So, what I had was a set of zeros in the numerator set of poles with m being less than or
equal to 𝒏𝒏.

Now, I just you know do some, do some high school algebra just to say that, well this I
can write it as decreasing powers of 𝒎𝒎, 𝒔𝒔𝒎𝒎 , 𝒔𝒔(𝒎𝒎−𝟏𝟏) , with the coefficients being the sum of
all these guys until I reach the last point 𝒔𝒔𝟎𝟎 will just be; sorry here it’ll be, yeah it’ll be the
sum of all these guys till I reach here, where it’ll just be the product of all these guys.
Okay, similarly the denominators I start with 𝒔𝒔𝒏𝒏 , the 𝒔𝒔𝒏𝒏−𝟏𝟏 will be the product of all these
guys and sorry the sorry the sum of all these guys and the 𝒔𝒔𝟎𝟎 guy will just be the product
of all these guys. So, I’m just dividing you know one guy with the other.

So, how does the characteristic equation look like? Characteristic equation, I’m interested
again the solution of the characteristic equation 𝟏𝟏 + 𝑲𝑲𝑲𝑲(𝒔𝒔)𝑯𝑯(𝒔𝒔) = 𝟎𝟎 is what I’m aiming
to solve right.

(Refer Slide Time: 08:34)

Okay so, 𝟏𝟏 + 𝑲𝑲 looks like this I just add a 𝟏𝟏 here and go here right. So, I just divide right.
I just get this guy down here and then see how it looks like. So, how; what will be now
the, now how the denominator now looks like. So, it’ll have start with powers of 𝒏𝒏 − 𝒎𝒎,
it’ll have powers of 𝒏𝒏 − 𝒎𝒎 − 𝟏𝟏 and so on.

So, what will be the coefficients of 𝒔𝒔𝒏𝒏−𝒎𝒎−𝟏𝟏, that will just be the coefficients of this guys,
the sum of the poles minus the sum of all the zeros and then you just have this series. Okay,

376
and now where is my; I’m looking again. So, my axis 𝒔𝒔 is here and my test point is at
infinity here, 𝒔𝒔𝟎𝟎 this is my infinity. So, if the test point is very far which means I’m looking
at very large values of 𝒔𝒔, I can just do my analysis just with these 2 terms and then I can,
the contribution of the remaining terms would be very small, then really looking at really
large values of 𝒔𝒔.

So, my characteristic equation can equivalently be written as 𝟏𝟏 + 𝑲𝑲𝑲𝑲(𝒔𝒔)𝑯𝑯(𝒔𝒔) is just these

𝑲𝑲
guys: 𝟏𝟏 + = 𝟎𝟎 , right. So, I just have 𝟏𝟏 + 𝑲𝑲, 𝒔𝒔𝒏𝒏−𝒎𝒎 with a
𝒔𝒔𝒏𝒏−𝒎𝒎 +�∑𝒏𝒏 𝒎𝒎
𝒊𝒊=𝟏𝟏 𝒑𝒑𝒊𝒊 −∑𝒊𝒊=𝟏𝟏 𝒛𝒛𝒊𝒊 �𝒔𝒔
𝒏𝒏−𝒎𝒎−𝟏𝟏

coefficient of 𝟏𝟏, and 𝒔𝒔𝒏𝒏−𝒎𝒎−𝟏𝟏 with the coefficients that’s decided by the nature of poles and
the nature of zeros. Now, okay let’s do something else now right. So, let me start with a
transfer function which has 𝒏𝒏 − 𝒎𝒎 repeated roots at some point 𝝈𝝈𝑨𝑨 and it has no zeros,
how will that look like? Well, the 𝑲𝑲𝑲𝑲(𝒔𝒔) like just say can assume 𝑯𝑯(𝒔𝒔) = 𝟏𝟏 or just add
𝑲𝑲
𝑯𝑯(𝑺𝑺) doesn’t really matter. So, 𝑲𝑲𝑲𝑲(𝒔𝒔) would be something like this okay.
(𝒔𝒔+𝝈𝝈𝑨𝑨 )𝒏𝒏−𝒎𝒎

(Refer Slide Time: 10:38)

Now, again the characteristic equation I just add 𝟏𝟏 and I get this one. So, this characteristic
equation, how will this root locus look like? Well, this root locus has no zeros and 𝒏𝒏 − 𝒎𝒎
poles okay? So, this root locus of this open loop system here with 𝑯𝑯(𝒔𝒔) = 𝟏𝟏 will have
𝒏𝒏 − 𝒎𝒎 branches, all of them go to infinity. They start at 𝝈𝝈𝑨𝑨 . So, that’s what it tells us right;
so what happens when? So, this is my 𝝈𝝈𝑨𝑨 and then there are say, 𝒏𝒏 of these roots, all 𝒏𝒏
will go to infinity; so some here, so this here, where whichever way they want to go. Okay,

377
that’s decided again by the nature of these things right? So, I’m looking at the equation
like 𝟏𝟏 + 𝑲𝑲 over this guy, 𝒏𝒏 − 𝒎𝒎 and all these , 𝒏𝒏 − 𝒎𝒎 branches they go to infinity. Now,
let me just expand this a little bit right. So, this is like s power a; (𝒔𝒔 + 𝒂𝒂)𝒏𝒏 kind of thing.

So, we do the binomial expansion and in the denominator I have again powers of 𝒔𝒔,
𝒏𝒏 − 𝒎𝒎, and then now I’m just summing this way right. So, if I just substitute this, this
formula directly summation of all the; all the poles, how many poles are there? 𝒏𝒏 − 𝒎𝒎,
where are the poles? at 𝝈𝝈𝑨𝑨 . So therefore, the second term would simply be (𝝈𝝈𝑨𝑨 ) 𝒏𝒏−𝒎𝒎 these
are the summation of all the, values of all the poles 𝒏𝒏 − 𝒎𝒎 − 𝟏𝟏 and so on. Again, for large
values of 𝒔𝒔, I just stop here and say well this is a good approximation at largest values.
Therefore, the straight line roots; the straight line root locus branches of the transfer
function in this equation are asymptotes of the transfer function with centroid, this one
right? So, everybody starts here and then they go to infinity, we’ll say how they go. So,
this is just a very vague picture. So, don’t really think that I’m drawing something very
correct here, just to give some vague idea about what’s happening right.

So, this straight line root branch; root locus branches of the transfer function which has 𝝈𝝈𝑨𝑨
poles; 𝒏𝒏 − 𝒎𝒎 poles at 𝝈𝝈𝑨𝑨 and then the negative side, they are asymptotes of the transfer
function with centroid 𝝈𝝈𝑨𝑨 , all the all the asymptotes start at 𝝈𝝈𝑨𝑨 . So, what is the relation
between this here right? So, we started with this one right we wanted to find out where;
what is the centroid point of the asymptotes, where how do, where do the asymptotes
intersect if at all? right. So, compare this guy, right. Again so what did I assume here is I
assumed here that of course, I started with 𝒎𝒎 poles sorry; 𝒎𝒎 zeros, 𝒏𝒏 poles I approximated
my characteristic equation I just did this right, so this was gone. So, this in a way the
denominator of 5.3.11 and these 5.3.8 are the same.

378
(Refer Slide Time: 13:58)

Okay, if I just equate this you know what I would get is something like this n minus m
sigma A is this guy. Now, what is 𝝈𝝈𝑨𝑨 in the general case? Because what I showed earlier,
then we were even looking at the angles, is that if I start from here this is my test point 𝒔𝒔𝟎𝟎
and these are 2 poles 𝒑𝒑𝟏𝟏 and 𝒑𝒑𝟐𝟐 if I look from very far it might just look at; it might just
seem as if these 2 poles are on the same location, because angle contributions are almost
the same. So, I can take this analogy and I can just say well I can equivalently view this as
a system where I have 2 poles −𝝈𝝈𝑨𝑨 , I’m not changing a location of the poles, I’m just
seeing how they look like from infinity. So, how if there are 2 guys placed here at 𝝈𝝈𝑨𝑨 and
𝝈𝝈𝑨𝑨 in the left half plane from infinity it’ll look. So, these guys 𝒑𝒑𝟏𝟏 and 𝒑𝒑𝟐𝟐 will exactly look
like this, that are actually close to each other. And therefore, this centroid point of the
asymptotes can be calculated this way, just by equating this expression here to this guy
because I know that from infinity they, it looks as if all the roots are together.

Because they contribute the same kind of angle, correct. So, what is this guy? So, the
centroid point. So, I sum the values of all the real parts of the poles, real parts of the zeros
and 𝒏𝒏 − 𝒎𝒎, right? And this is how I calculate the centroid point.

379
(Refer Slide Time: 15:35)

Okay so, let’s see how it looks like you know in terms of a small example. So, take
𝑲𝑲
𝑮𝑮(𝒔𝒔) = , there are 3 poles for the system, okay again 𝑯𝑯(𝒔𝒔) is assumed to be 𝟏𝟏. So,
𝒔𝒔(𝒔𝒔𝟐𝟐 +𝟒𝟒)

wherever 𝑮𝑮(𝒔𝒔) is given you can safely assume that 𝑯𝑯(𝒔𝒔) is 𝟏𝟏 and if we also we all; we
always know how to transfer; transform a non-unity feedback system to a unity feedback
system. So, we can do the same analysis.

So, coming back to this, so given the transfer function of this form I have 3 poles and 3
branches of the root locus. So, these 3 branches let me just approximately draw something
here right. So, I have a pole here at origin, 𝒔𝒔𝟐𝟐 + 𝟒𝟒 = 𝟎𝟎. So, I’ll have a complex conjugate
pole at sorry, at 𝒋𝒋𝒋𝒋𝒋𝒋 and −𝒋𝒋𝒋𝒋𝒋𝒋 or 𝟐𝟐𝟐𝟐𝟐𝟐; however, you want to write this. So this, when 𝑲𝑲 =
𝟎𝟎 I’m just here right, 𝑲𝑲 = 𝟎𝟎 just at the poles, 𝑲𝑲 = 𝟎𝟎 at the poles, 𝑲𝑲 = 𝟎𝟎 my root locus looks
like this right, it satisfies the thing that I’m symmetric about the real axis and all of them
go to infinity because there are no zeros.

Now, how do they go to infinity? Now, first the rule number 1 told me that it is symmetric.
Rule number 2 told me that all will go to the; to the infinity. Rule number 3 told me about
well, is a point like this on the root locus? You know is a point is zero which is on the real
line on the root locus, then I compute well to see if I’m here, I just look at here and see
how many poles are there, well there are 1, 2 and 3.

380
There are no zeros, this is odd number and therefore, this point will be a root locus, this
point will also be a root locus, any point on this real line if I look towards my right I’ll
have 3 poles which is an odd number. So, this entire line will be on root locus, all the entire
negative real axis. So, what about the other 2 guys? So, when we look at how do they go
to, well this guy goes to infinity in this way right, because entire real line is on the root
line. Now, how do the other guys go? So, I look at the, the way they look; they go to
infinity. What kind of angles they make?

(𝟐𝟐𝟐𝟐+𝟏𝟏)
So, the angle here is computed as again 𝟏𝟏𝟏𝟏𝟏𝟏°. So, I get 𝟔𝟔𝟔𝟔°, 𝟏𝟏𝟏𝟏𝟏𝟏° and 𝟑𝟑𝟑𝟑𝟑𝟑°. 𝟔𝟔𝟔𝟔° is
𝟑𝟑
this guy, then I have 𝟏𝟏𝟏𝟏𝟏𝟏°, and I’ll have 𝟑𝟑𝟑𝟑𝟑𝟑° here. 𝟑𝟑𝟑𝟑𝟑𝟑° I can also write as −𝟔𝟔𝟔𝟔°. So, this
is, this is okay here right? Because, at 𝟏𝟏𝟏𝟏𝟏𝟏° already this guy is branching out, at 𝟏𝟏𝟏𝟏𝟏𝟏°. So,
that’s a good validation and then these guys they will. So, these are how my asymptotes
go. So, these guys will meet this line at 𝟔𝟔𝟔𝟔°, at infinity, and this guy will meet this line at
infinity the guy at −𝟐𝟐𝟐𝟐𝟐𝟐.

So, at infinity not that they just come here and then they go along right; so they keep
moving, moving, moving and they meet this line asymptotically, this guy makes this line
asymptotically right. So, this is about the angles.

(Refer Slide Time: 19:04)

381
Okay so, next where do these, these asymptotes originate? and I just look at this this
formula here, about computing the sum of the real parts of the poles, real parts of the zeros
with 𝒏𝒏 − 𝒎𝒎.

𝟎𝟎−𝟎𝟎
So, this is and this 𝟎𝟎. So, these asymptotes actually start here and just go here. So,
𝟑𝟑
they meet exactly at this point. Okay so now, how the root locus will look like? Well, this
guy will start moving little towards this direction, we’ll see what direction exactly it moves
in a little, in a little later, but I can fairly assume it moves something here, it doesn’t, it
never touches this line right? It just goes here and then it only touches it at infinity. This
guy starts moving this way just a mirror image of this, and then goes to infinity this way.

So, we have found out which direction these guys go or at least I know that now this, I
may not exactly know this location, this direction now, but I know that at infinity it meets
this line and therefore, it has to move somewhere vaguely in this direction right? And we’ll
define this a little later, similarly with this guy. So, we know how; in what way or what
decides how these roots branch out to infinity.

(Refer Slide Time: 20:14)

Okay so, again coming back to the first example: so I start at 𝟎𝟎, I start at −𝒂𝒂 and I know
𝒂𝒂
here that this guy moves here, this guy moves here, they both meet at − and then go to
𝟐𝟐
infinity. At this point, where they meet and then they again go away right; they don’t really
stay together, just come here come together just say hello and then you know, go in your

382
 𝒂𝒂
own path right. So now, who decides why; where this is? Why − ? Why not something
𝟐𝟐
here, why not something here? Again this, I know because I could just compute from the
computations of my quadratic equation, but what in general right? I just I may not be able
to again compute all the roots and then you know, it might be a little little cumbersome.

So, this was good enough for understanding. So, who decides this point and does that point
exist right? That also is the thing. So, if I’m just giving you a characteristic equation which
has a polynomial of order 𝒔𝒔𝟗𝟗 + (𝟏𝟏𝟏𝟏𝒔𝒔𝟖𝟖 + ⋯ ) and lots of things. I should first know that does
a breakaway point excess or not right? So, who tells me that? Again everything is from the
characteristic equation.

Now let’s assume. So, at this point what happens? That there are multiple roots right, the
root here, the root is 𝟎𝟎, the root is −𝒂𝒂 or the poles and then they come here, here both the
𝒂𝒂 𝒂𝒂
poles have the same value − and − . So, when this happens my characteristic equation
𝟐𝟐 𝟐𝟐
𝟏𝟏 + 𝑲𝑲𝑲𝑲(𝒔𝒔)𝑯𝑯(𝒔𝒔), well I will have here the root at 𝒔𝒔𝟎𝟎 with some multiplicity 𝒓𝒓 and times
𝑿𝑿(𝒔𝒔). Again, this 𝑿𝑿(𝒔𝒔) does not contain this factor and we will be interested in situations
where 𝒓𝒓 is at least 𝟐𝟐 right. Now, I differentiate this guy, I differentiate this on the left hand
𝒅𝒅
side I have of this guy, on the right hand side I differentiate again with respect to 𝒔𝒔,
𝒅𝒅𝒅𝒅

differentiating with the first guy I’ll have 𝒓𝒓(𝒔𝒔 + 𝒔𝒔𝟎𝟎 )𝒓𝒓−𝟏𝟏 𝑿𝑿(𝒔𝒔), then again same guy here and
𝑿𝑿′ (𝒔𝒔), where 𝑿𝑿′ (𝒔𝒔) is again the derivative of 𝑿𝑿(𝒔𝒔) with respect to 𝒔𝒔. Let’s leave this here
for a while and then come back to this again.

So, in the pole-zero form I can write 1 plus G, 𝟏𝟏 + 𝑲𝑲𝑲𝑲(𝒔𝒔)𝑯𝑯(𝒔𝒔) as 𝟏𝟏 + 𝑲𝑲 some polynomial
in the numerator and some polynomial in the denominator. Now, let’s see what this guy
has to do.

383
(Refer Slide Time: 22:54)

So, I can differentiate this guy and what I have is, is this thing; here exactly I was just
doing the same thing right. I’m differentiating this and I’m differentiating this again here.
𝒅𝒅𝑯𝑯
So, in terms of 𝑷𝑷 and 𝑸𝑸, the condition that d H or = 𝟎𝟎 looks like this, just the numerator
𝒅𝒅𝒅𝒅
being equal to 𝟎𝟎.

Now, look at this guy here. So, I know okay of course, that at 𝒔𝒔𝟎𝟎 ; if this is 𝒔𝒔𝟎𝟎 is a root the
characteristic equation is obviously 𝟎𝟎. Now look at the derivative right, a derivative what
is the value of derivative at 𝒔𝒔 = 𝒔𝒔𝟎𝟎 ? Let 𝒔𝒔 = 𝒔𝒔𝟎𝟎 well this, this guy goes to 𝟎𝟎, this guy also
𝒅𝒅 𝒅𝒅
goes to 𝟎𝟎; so this. And this, in this expression [𝑲𝑲 𝑮𝑮(𝒔𝒔)𝑯𝑯(𝑺𝑺)] at 𝒔𝒔 = 𝟎𝟎 is 𝟎𝟎; sorry
𝒅𝒅𝒅𝒅 𝒅𝒅𝒅𝒅
𝒔𝒔 = 𝒔𝒔𝟎𝟎 , where 𝒔𝒔𝟎𝟎 comes with a multiplicity of 𝒓𝒓. So, that’s what this means, so the roots
of this equation are the roots of this guy. Now, I can equivalently you know derived my
𝑲𝑲𝑲𝑲(𝒔𝒔)
characteristic equation something like this. So, 𝟏𝟏 + 𝑲𝑲 𝑮𝑮(𝒔𝒔)𝑯𝑯(𝑺𝑺) or 𝟏𝟏 + 𝑸𝑸(𝒔𝒔)
, I can just
𝑸𝑸(𝑺𝑺)
write it as 𝑲𝑲 = − 𝑷𝑷(𝒔𝒔). ok? There is a reason why you are doing this now. So, first is well

I’m just looking at how to detect if there are multiple roots and that is when
𝒅𝒅
[𝟏𝟏 + 𝑲𝑲 𝑮𝑮(𝒔𝒔)𝑯𝑯(𝑺𝑺)] goes to 𝟎𝟎 that has a multiple root right, at that point. Right, then that
𝒅𝒅𝒅𝒅
point here is.

384
So, therefore, I can just say that well the solutions are of this form right. Now, let me do
𝒅𝒅𝒅𝒅
this another something else. So, 𝑲𝑲 is again now a function of 𝒔𝒔, looks like this. And
𝒅𝒅𝒅𝒅
𝒅𝒅𝒅𝒅 𝒅𝒅𝒅𝒅
when does go to 𝟎𝟎? goes to 𝟎𝟎, when the numerator of this guy goes to 𝟎𝟎, i.e.,
𝒅𝒅𝒅𝒅 𝒅𝒅𝒅𝒅
𝑷𝑷(𝒔𝒔)𝑸𝑸′ (𝒔𝒔) − 𝑸𝑸(𝒔𝒔)𝑷𝑷′ (𝒔𝒔). Now look at these two, these two are the same sorry this
expression and this expression are the same right; now if the equivalent the condition here.
𝒅𝒅𝒅𝒅
So, when does 𝒓𝒓 have a multiplicity at least 𝟐𝟐? Sorry if 𝒓𝒓 was just 𝟏𝟏, would not be 𝟎𝟎
𝒅𝒅𝒅𝒅
right? It would be something else because this guy would go away.

But if 𝒓𝒓 is 𝟐𝟐, this guy would go to 𝟎𝟎, this guy would also go to 𝟎𝟎. So, if 𝒔𝒔𝟎𝟎 occurs twice it
𝒅𝒅
means [𝟏𝟏 + 𝑲𝑲 𝑮𝑮(𝒔𝒔)𝑯𝑯(𝑺𝑺)] = 𝟎𝟎. So, if 𝒔𝒔𝟎𝟎 occurs twice, that meant that
𝒅𝒅𝒅𝒅
𝒅𝒅 𝒅𝒅
[𝟏𝟏 + 𝑲𝑲 𝑮𝑮(𝒔𝒔)𝑯𝑯(𝑺𝑺)] =
𝒅𝒅𝒅𝒅
𝑮𝑮(𝒔𝒔)𝑯𝑯(𝑺𝑺) = 𝟎𝟎 and you can have the 𝑲𝑲 or not. Now, this being
𝒅𝒅𝒅𝒅
𝟎𝟎, tells me that there is a multiple root, this being 𝟎𝟎 also means that this is 𝟎𝟎.

Now, compare this condition and compare this condition here, they both are the same and
𝒅𝒅𝒅𝒅
therefore, I can check; I can just equivalently say = 𝟎𝟎, this is equivalent to
𝒅𝒅𝒅𝒅
𝒅𝒅
[𝟏𝟏 + 𝑲𝑲 𝑮𝑮(𝒔𝒔)𝑯𝑯(𝑺𝑺)] = 𝟎𝟎 and for obvious reasons this is for me equal to; easy to compute
𝒅𝒅𝒅𝒅
because I know 𝑸𝑸(𝒔𝒔) and 𝑷𝑷(𝒔𝒔) and this might be because, I may have to you know 𝑮𝑮 and
𝑯𝑯 might have numerators, denominators and may have to do a lot of other stuff, but then
this is easier to compute. So, this condition tells me that, well there is a breakaway point
right. The breakaway points are the points at which multiple roots of the characteristic
equation occur, the solution of this will be the exact breakaway point, you can compute
this, I thing here it would be very trivially satisfied.

So let’s again, so see compute for this example and check right, you know in this way.
𝒂𝒂 𝑲𝑲
Why this is − ? So, what was the characteristic equation? that was 𝟏𝟏 + 𝒔𝒔(𝒔𝒔+𝒂𝒂) = 𝟎𝟎 or
𝟐𝟐
𝒅𝒅𝒅𝒅
𝑲𝑲 = −𝒔𝒔(𝒔𝒔 + 𝒂𝒂) okay. would be −(𝟐𝟐𝟐𝟐 + 𝒂𝒂) and if I equate this to 𝟎𝟎, I get the solution
𝒅𝒅𝒅𝒅
𝒂𝒂
as 𝒔𝒔 = − 𝟐𝟐 right? And this is the breakaway point and of course, in this case it also turns

out to be the centroid point and it can happen in many cases. So, this is like okay, you can
𝒂𝒂
see why this is using this explanation over here.
𝟐𝟐

385
(Refer Slide Time: 28:46)

Okay so, let’s do a; sketch a root locus for this guy. So, I have again a pole at −𝟐𝟐, −𝟑𝟑, −𝟒𝟒;
zero at −𝟏𝟏 - 3 poles and 1 zero. Again the number of branches in the root locus would be
the number of poles, that is 3 and these; of these 3 branches one will go to the zero at −𝟏𝟏
and the other 2 will disappear to infinity. Again starting from 𝑲𝑲 = 𝟎𝟎 they will start at the
poles and then go to infinity. So, let’s first draw the basic conditions. So, at 𝑲𝑲 = 𝟎𝟎 my root
locus will be at this poles which is – 𝟐𝟐, −𝟑𝟑 and −𝟒𝟒.

So, rule number 3 which part of the root locus on the real line or the real axis will be on
the root locus. Let’s start from here, well there is there is nothing here. So, there would not
be anything here. So, and then let’s look at this area between −𝟏𝟏 and −𝟐𝟐, if I look to the
right there is 𝟏𝟏 zero. So, I’m good here right, there is an odd number this one. Okay now,
look at the area here between −𝟐𝟐 and −𝟑𝟑; how many guys are sitting out here one pole,
one zero that’s an even number I don’t like it. So, I just go here between −𝟑𝟑 and −𝟒𝟒, I just
look to the right. So, there is 1, 2 and 3 oh its an odd number.

So, I like it right. It’s here, and then what happens that after −𝟒𝟒? I just look to the right I
have 1, 2, 3, 4. Okay so, this area is also not good for me. So, one thing for sure I know
that 𝑲𝑲 = 𝟎𝟎 points are these guys, 𝑲𝑲 = 𝟎𝟎, 𝑲𝑲 = 𝟎𝟎, 𝑲𝑲 = 𝟎𝟎, this guy will go quickly and then
sit here right. Now, these guys have they; have to find their way to infinity, they both try
to run away this way right, but it is not allowed because these points are not on the root

386
locus. So, they have to somewhere just be around this one and then decide how to go to
infinity.

(Refer Slide Time: 31:10)

Now, let’s see how; who tells these guys where to go to infinity. So, how do they go to
infinity? There are 2 guys who are going to infinity and they proceed to infinity along
(𝟐𝟐𝟐𝟐+𝟏𝟏)
asymptotes with angles again given by 𝝋𝝋𝑨𝑨 = 𝒏𝒏−𝒎𝒎
𝟏𝟏𝟏𝟏𝟏𝟏°; at 𝒏𝒏 − 𝒎𝒎 was 𝟐𝟐. So, you will

have 𝟗𝟗𝟗𝟗° and 𝟐𝟐𝟐𝟐𝟐𝟐°.

So, I know at least that. So, these 2 guys, let me just draw it for. So, I have a zero here this
is at −𝟏𝟏, −𝟐𝟐 sorry; these are all poles, pole, a pole, and a pole. So, whatever they do in
between they will go to infinity at these angles 𝟗𝟗𝟗𝟗° and 𝟐𝟐𝟐𝟐𝟐𝟐°. So, how is the centroid point
given? The centroid point where of this is again you compute the sum of the poles of the
real parts and zeros and you get this one. Now, what I also know is that, okay now this is
a centroid point right. So, asymptotes will be like this at −𝟒𝟒. So, these guys will move and
then they will meet these guys at infinity and where do these lines at infinity come from?
They come from −𝟒𝟒. So, I know that one now right. So, I’ll just remove this. So, I didn’t
know which direction they go to infinity, but now I know, they go from −𝟒𝟒.

Now the sec, the next thing is okay, these guys well, where do they meet? does this guy
wait for this guy to come all the way here and then they just move away or they meet
somewhere in the middle or then go to; then move apart. So, that’s we’ll find. So, there is

387
a breakaway point right? Is breakaway point −𝟒𝟒 or −𝟑𝟑 or whatever, let’s try to find out
𝒅𝒅𝒅𝒅
and that is given by solution of = 𝟎𝟎.
𝒅𝒅𝒅𝒅

(Refer Slide Time: 33:13)

𝒅𝒅𝒅𝒅
So, I just do all the math here = 𝟎𝟎, would mean well it gives me 3 roots: −𝟑𝟑. 𝟓𝟓𝟓𝟓, −𝟐𝟐. 𝟑𝟑𝟑𝟑,
𝒅𝒅𝒅𝒅
and −𝟎𝟎. 𝟏𝟏𝟏𝟏𝟏𝟏𝟏𝟏. Now, look at these points: −𝟐𝟐. 𝟑𝟑𝟑𝟑 that these regions are not on the root
locus, this region is not on the root locus, this region is not on the root locus, this region is
not on the root locus. So, only point where they are allowed to meet is −𝟑𝟑. 𝟓𝟓𝟓𝟓. So, that’s
the point here.

So, what happens is well this guy, this guy goes here, these 2 guys from the asymptotes
and centered computation I know that they will meet these 2 lines at infinity. So, they will
just start from here, so this guy goes starts here and goes to infinity. This guy starts here
and goes to infinity along this line. So, that’s a very complex. So, which means again,
which; how do they move? why not they do something like this or you know something
like this, you know more.

388
(Refer Slide Time: 34:12)

So, that that will see a little later. Well, at the moment that seems interested now little
towards the asymptotic behaviour of this. So, we’ll to this; we will end this lecture over
here and then we will see a little more details and then conclude the root locus by looking
at the last 3 rules.

Thank you.

389
 Control Engineering
Dr. Ramakrishna Pasumarthy
Department of Electrical Engineering
Indian Institute of Technology, Madras

Module - 05
Lecture - 04
Root Locus Plots

So, continuing again on the root locus thing. So, we just left with a couple of rules.

(Refer Slide Time: 00:28)

So, the first one was angle of departure from an open loop pole, and why is this important to
compute. So, in one of our earlier examples, where we had a pole at origin, then we had
something at j 2 ω , − j2 ω , one guy was going to infinity through the real line and then
we had asymptotes this way and these guys were going to like this right? like this to infinity,
this guy was going like this to infinity.

So, who decides right of how this guy goes? why doesn’t it take a steeper angle and go like
this? or why does not it go, you know wander here and then again come back and say oh this
guys here, and not like this. We’ll say who decides or what is there any anything that will tell
us which direction I should be moving right? So, this is decided by something called the
angle of departure or which direction I should go. So, the angle of departure the formula says
the following it φ p =± (2 q +1 ) 180 ° +φ from an open loop pole. Let’s see I am just looking
at one pole under consideration is again it should satisfy the angle criterion right. So, all the

390
time we are just looking at this one at 1+ KG ( s ) H ( s )=0 . So, all this criterion, all we do is
just try to satisfy this criterion right? Nothing else or just that you know the magnitude
criterion and the angle criterion, as simple as that. So, the angle of departure is given by again
this angle criterion. So, little modification of that plus φ . Okay, what is; what is φ ? Say
I am interested in some pole and this φ is the net angle contribution to this pole by all
other open loop poles and zeros.

Now this is for pole right, I go away from a pole and then I end up at a zero. So, I arrive at a
zero. So, which direction do I arrive? Do I arrive along the axis? or do I just do some circular
behavior or whatever, how do I arrive? That’s also important right. So, the angle of arrival is
again defined similarly. So, let’s try to understand what this means and say draw a little
picture here. So, let’s say I have 6 poles and one zero, and I am interested of you know how
does p5 move? So, let’s get a little scenario here, Okay, let’s erase this. So, this is my
plane and say well I have what we call this p1 , p2 , p3 , p4 and say I have a p5
and p6 be here right. p1 , p2 , p3 , p4 , p5 and p6 and then there is a zero
right.

So, let’s say that the zero is sitting somewhere here. Now, I want to see which direction this
guy should start moving, should it start moving vertically downwards or somewhere here or
somewhere here, for any arbitrary scenario. And let’s say I just take a point very close by here
this one and I call that point as s 0 . Now, for this s0 to lie on the root locus it must for
sure satisfy the angle criterion okay. Now and this so which this is very close by right. So,
let’s see. So, I just draw these angles: this will be θ4 , so θ3 , I have θ2 , I have θ1
and the angle here would be this guy right; so some joining this line from here. So, that would
be some θ6 , and then say a angle of this of the zero, I just call this θ .

And then this guy is somewhere here right. So, the angle here would be. So, this guy is
actually moving in this direction right. This is the angle of θ P , the pole which I am
interested in. So here, there are φ ’s and θ ’s, but does not matter right, you can just use
whatever. So, for s 0 to lie on the pole, so what should happen? At all, the sum of all these
angles should again be multiples of 180 ° . So, what is the; how is angle criteria look like,
the angle of the zero is just θ , and then I have angle of all the poles that is θ1 , θ2 ,
θ3 , θ4 , I have this little θ p , and θ6 , this again ±180 ° (2 q+ 1) blah blah blah.
i.e.,

391
 θ−(θ1 +θ2 +θ 3+θ 4 +θ p +θ 6)=±180 ° (2 q+1)

Now, I am interested in this angle right, this guy this little angle, this will; it should tell me
where I will go. Therefore, I just you know substitute this, and this, this θp or even φp
whatever is simply ±(2 q+1)180 ° , plus all the other remaining angles they’ll just go here
' '
and I just call the sum, some θ . So, where this guy or this θ here or I just call this
φ , we should be consistent with what is the notation over there, for all the others you
guys. So, this φ is the net on angle contribution to this pole by all other open loop poles
and zeros. So, that’s how I compute this θ p .

(Refer Slide Time: 06:28)

So, this is this how it would look like. So, this φ is the angle of the zeros minus all the
angle of the poles and this is how I get it, and it’s a kind of straightforward to look at it right
and again because you know s 0 is very close to the departure point or the pole this will be
a departure point and as s0 tends to the pole p5 , well that particular angle contribution
will come zero and then you know this and then the other poles or the angle contribution by
the other poles would satisfy the angle criterion.

Where I am not moving I am just at that pole and the the limiting condition as s0 goes to
s 5 , but s0 being very close to s5 how do I compute the angle? Well, that is just by
this φ I wrote here this guy. Similarly, I could even do for a zero right? its exactly the
same procedure what I would do for a zero, take this zero and you can just compute take a

392
point just to the left or right or in a top or bottom of the zero and then the angle could be
computed kind of quite easily.

(Refer Slide Time: 07:37)

So, we do an example: its case the root locus for a unity feedback system with open loop
transfer function again. We have a complex conjugate poles, open loop, this is zero at −1 ,
n=2 , m=1 , number of branches is 2 this one and there is a 1 zero. So, one pole will go
to zero and other will go to infinity and the 2 branches originate from here when K=0 ,
and m=1 , one of the branches will go to this guy, another will go to infinity. So, let’s see
how the plot looks like.

So, I’ll just plot you a rough sketch of it before. So, this is a good space here. So, I have a
zero at −1 and my poles are here, −2+ j3 and −2− j 3 , a pole here and a pole
here. So, I just want to see you know which direction this move, but before that well let me
do the first rules right, which point is on the real axis? I know that when I start from K=0
I’m here, K=0 I’m here, K=0 I’m here, now just look at this, there is nothing here
right. So, this point would not this region would not be on the root locus. There will be
nothing here.

Then look at the point here right. So, from here I go to the left of −1 , there is just one
zero-odd number, then I go again here, there is again an odd number; 1, 2 and 3. Sometimes,
we can even neglect this complex poles for counting because they always are in conjugate
pairs. So, there is even before this guy, there will always be an even after this guy. If it’s an

393
odd there will be an odd. So, at least I know that the entire this guy is on the real axis you. So,
sorry is on the on the root locus correct? I will not even draw the arrow at the moment just to
be a little safe. So, this entire line is on the root locus starting from the left of −1 .

(Refer Slide Time: 09:50)

So, again I do all the basic computations that well there is one guy which goes to infinity.
What is the angle in which that guy goes to infinity? Well, that angle is 180 ° and is just
very straightforward to compare, n−m=1 . So, this φ will be 180 ° , well and these

dK
guys will do they really meet at some point and then go away, I can just find by
ds
¿0 and the solutions would be, well this one −4.16 and 2.16 . I don’t really; I am
not really interested in this guy right this is on, this point does not really lie on the root locus.
I am not ignoring this because it is an unstable region right, I am just ignoring because the
root locus tells me that I just cannot go here, because I cannot be on any of this line because
there are no poles or zeros to the right of this. So, I am looking at −4.16 . So, this is this
real part here is −2 and say this is this point is −4.16 .

394
(Refer Slide Time: 10:53)

So, I just go now to compute the angle of departure, departure from this side and departure
from this side right and then just this formula φp if I am looking at this guy would be
±180 ° plus the contribution of the other two guys, and if I am computing φ for this
guy it would be this φ would be the angle contribution of these other two guys.

(Refer Slide Time: 11:26)

I can just compute; this is kind of very straightforward to check that φ for this guy would
be 198.43° this way. And φ for this guy −2− j 3 would be 161.57 ° . So, I just

395
know that, I am just from here sorry I just move in this direction, here I just move in this
direction, I also know where I am meeting right at −4.16 .

So, the locus will start from here, then will come here, then they come here and therefore, the
directions would be something; maybe something like this. This guy moves here, this guy
moves here, this guy goes here and this guy goes here right. So, now, I know why is it not
like something like this or any other thing why is it not like this right. So, this I just this angle
tells me why how it behaves.

(Refer Slide Time: 12:11)

So, this is something not very surprising this rule, how do I know if my root locus intersects
with the imaginary axis, well when I hit the imaginary axis say I have a root locus which is
like this: let’s say I have three guys here and I say well this guy is, I’m just doing an
approximate plot right. So, this guy there could be something like this right there could be
possibilities of this thing and these guys are going here. So, these are all K=0 , K=0 ,
K=0 , and I see that after some values they are actually going this side, they are going in
the unstable region. Can I really compute this value when it becomes unstable, there was
another example where my poles were here, here and here and then the root locus, well one
guy was going this side and other, well other two guys went here. So, for any value of
K >0 it was always unstable right. That was not difficult to compute, now how do I handle
these conditions.

396
Well, I know something all the Routh Hurwitz criterion right, the intersection of the root
locus with the imaginary axis can easily be determined by the Routh criterion just to see how
can. So, we had done some problems where we were computing stability with respect to a
parameter. If K was between this number and that number, the system is stable. If K is
greater than this number system is stable and so on. So, and then the open loop gain K at
any point on the root locus is just given by this, I think this we had even discussed in one of
our very earlier slides, what given a point K how do I find the sorry; given a point on the
root locus how do I find the gain K at that point? well this is just simply like, there is
nothing special here.

Of course, now if you could do and on Matlab you just move the cursor it will actually tell
you the values, but where does it come from? that comes from something here.

(Refer Slide Time: 14:09)

K
So, just to do another example: so we have G(s)= 2 , I have a pole at the
s (s +8 s +32)
origin and 2 other poles which are complex conjugate −4 ± j 4 . And another pole at the
origin. So, I start from here. So, I have a pole at the origin and I have two poles here. So, this
is −4+ j 4 , −4− j 4 and this is the origin, in the σ and jω axis. So, there are
three branches; all of three, all of these things will go to infinity, because there are no zeros;
so which points of the real line lie on the root locus. So, this is fairly obvious because at any
point if I go to the left of zero, I look at the right there always be odd number of poles. So,
one until I reach here and after that there will be 3 poles.

397
(Refer Slide Time: 15:04)

In the usual process of finding how the root locus or how the roots proceed to infinity, they
will be at angles of asymptotes at 60 ° , 180 ° and 300 ° with the standard formula
and then the centroid point of these asymptotes would be −2.66 : so somewhere over here.

dK
So, this is my σ A . So, the breakaway points, well I just look at ¿ 0 , I find K .
ds

(Refer Slide Time: 15:45)

dK
And then solve for ¿ 0 , which gives me the breakaway roots at this point. And I
ds
could actually see from here that these points are very unlikely to be on the root locus right.

398
 −8 ± j 4 √2
So, look at this . So, I am at somewhere here, somewhere here right, then what
3
do my asymptotes suggest is that well I am at 60 ° , 180 ° and 300 ° . So, which is at
an angles like over here; so this is where.

So, I it shows that I should actually be moving towards the asymptotes, so somewhere in this
direction; somewhere in this direction right. So, there is no way that I could reach these
points based on the location of the poles, the centroid point, and the asymptotes. So, for sure

dK
that even though ¿ 0 gives me a solution, these points will not lie on the root locus.
ds
You could also test this by the angle and the magnitude criterion. Now I am looking at the
angle of departure at these guys. So, angle of departure would be computed by you know so,
this angle and this angle.

So, that should be straightforward to compute. So, this angle would be the angle at
−4+ j 4 would be −45 ° , and −4− j 4 would be +45 ° . Now since these are
moving in this direction and the asymptotes are towards the right of the imaginary axis, there
could be chances of these guys meeting the imaginary axis. So, let’s see if that is true. So, I

K
write down the characteristic equation and I have 1+ 2 and I draw the Routh
s (s + 8 s+ 32)
table.

(Refer Slide Time: 17:40)

399
 s3 1 32

s3 8 K

s
3
256−K
8

s3 K

8∗32−K
So, 1 , 32 , 8 , K , this would be, . So, what happens when say, when
8

256−K
does this guy go to 0 ? this guy will go to 0 when =0 or for a value of gain
8
K=256 , I have this entire row equal to 0 . Now let’s see how the auxiliary equation
2 2
looks like, I have 8 s + K=0 or 8 s +256=0 which gives me s=± j √ 32 and
this is the value of ω , where it will cross the imaginary axis.

So, this will be at somewhere here, √ 32 in a positive direction, √ 32 in the negative

direction right. So, this gives me a lot of information now that, if there are no breakaway
points the root locus will actually cross the imaginary axis and it would actually do it this
way.

(Refer Slide Time: 18:44)

400
So, the angle of departure was −45 ° . So, it will initially tend to move this way and this is
√ 32 in the j−axis and then it will just asymptotes this way.

Similarly, this guy has angle of departure of +45 ° . So, it will move this direction and it
will cross the imaginary axis at −√ 32 and I will just do this one right. So, we saw few

dK
possibilities here right that there are no breakaway points even though gives me a
ds
solution, second is to explicitly compute the angle of departure right and also to see exactly
where the root locus meets my imaginary axis and this where the Routh table was very useful
for me.

So, if I now want to have a problem where I said: well, what are the values of gain K for
which the system is stable, I’ll say well for values of gain K between 0 and 256
my system is stable and for all other values it might go to the unstable region. So, I have a bit
better understanding now of the relative stability in terms of the gain K .

(Refer Slide Time: 20:11)

So, that’s about it and this guy is kind of straight forward right, you just keep on moving to
the left until it reaches infinity right very straightforward root for this guy. So, to summarize
module 5, we have learnt construction root; the root locus techniques, we also knew now
does the system go to instability.

401
(Refer Slide Time: 20:23)

And next when we meet we will do lots of things on the frequency response of the system.
Now we were interested in what happens when time goes from 0 to infinity or a gain goes
from 0 to infinity. We’ll see what happens with changing frequency, so that will do in the
in the next module.

Thank you.

402
 Control Engineering
Dr. Ramakrishna Pasumarthy
Department of Electrical Engineering
Indian Institute of Technology, Madras

Module - 05
Tutorial
Lecture - 23
Root Locus Plots

Hey guys. In this, but will just quickly run through a few root locus plots; I will just run you
through three or four problems just to tell you how to use those steps

(Refer Slide Time: 00:31)

So this is the first problem: G(s) is, all through again I assume that everything is a unity
feedback system so don’t really worry about what should happen to the H and then I just
get the K in here as well. So, let us try plotting this. So, I have m=1 , n=3 . So, the
n=3 , so I will have 3 branches right? So, two starting from the origin here, I have a pole
at −9 and I have a zero at – 1 . Okay, this is not drawn to scale; so that should be okay.
And this is again the σ and jω just for one; these are the two roots at the origin. Okay.

So, the first thing is there the three poles 1, 0; the root locus will have 3 branches, it has 1
zero, so one of these poles will go to a zero and two of them will go to infinity. We need to
find out how these two guys go to infinity. So, first is we need to; step four we need to find
out what is σA ? or the centroid point of the asymptotes so that is just like −9 and the

403
zeroes are here −(−1 ) ; n−m=2 , so this will be −4 . So, somewhere here is my
centroid point again not drawn to scale is somewhere here.

And then what are the angles of these asymptotes? Angles of asymptotes we compute them as

(2 q+1)180 °
(2 q+1)180 ° over n−m=2 (i.e., ), so that will again give you
2
± 90° ; so somewhere in the direction, vertically above and vertically downwards.

Now, other thing which we need to look at is, do the roots cross the imaginary axis? So, I just
write down the characteristic equation of 1+G(s )=0 and this gives me that
3 2
s +9 s + Ks+ K=0 . We’ll see when does this actually meet the imaginary axis, while I
quickly draw the Routh table. So, I have

s
3 1 K

s
2
9 K

s
1
8
9
K

s0

3
So, s I have a 1 and K , I have a 9 , I have a K and what comes here is

9 K −K 8
. So, that would be K . So, for what values of K this goes to you
9 9
know 0 ; okay, K=0 ; this guy goes to 0 , but if I substitute K=0 over here I see
that well there are no real; there is no other point apart from the origin that it meets that the
root locus actually meets this jω axis. So, the roots, do the roots cross the imagery axis?
The answer is no; it is just obvious from this table here.

dK
Now, the breakaway points: so what do we do, we just take =¿ 0 we equate it to
ds
0 . I will not do this, I think that should be easy to compute so will have s=0 , s=−3
and s=−3 . So, one, of course, you can see here, they’re already a breakaway point

404
because there is multiple roots here and then we have two guys sitting here at −3 . So, this
is how it looks like.

Now, we have to decide how do these guys move from here, which direction they go do both
of them go here, both of them go here or they just go around no wander here for a while and
come back, I know that they never go to the right half plane because of this Routh criterion.
So, they just have to be somewhere in the left half plane. And how do they go away? Well,

± 180
they go away from here at an angle , where r here is a number of the roots. So,
r
there will be ±180 , so the initial moving point will be sorry it will be like the tangent
would be here ± 90° . And then this these two guys go here and then they have to arrive
here, arrive here so some at this point.

Look at this guy so this guy has: ok so even before that I think what we forgot is to see what
point of or which part of the real line is on the root locus. So, I start from here what about the
area between 0 and −1 , I just look at my right I have two poles, but this number
should be odd. So, this area is ruled out I go from −1 to −9 I just look at the right I
have 1, 2 and 3. So, this entire part is on the root locus: the part from −1 till −9 . So, I
know that these two guys will move away from each other and they will arrive at this point
−3 .

Now where does this guy go? This guy moves to the right, right because it is on the root locus
and this guy also moves and it also arrives at this point because you have two times −3
coming in. Now, these guys go away at ±180 . And, again now look at how. So, there are
three roots coming in and three roots will go away. So, I just use the angle criterion I have

180
(2 q+ 1) . So, that will give me 60 and then 180 and then 300 . So, this is
3
60 , this is 180 and this is 300 or even −60 , this is 60 , −60 .

So, here the roots will arrive here this way, this way after they come here they just go away to
infinity like this along the asymptotes which are given by this and ± 90° . So, I just try to
join the dots here. So, I will have a root locus which looks like this from here, from here and
so on. So, all the steps are clear right.

405
(Refer Slide Time: 07:35)

So, this is how it goes. So, I start: so this is a just do not draw a further to the left because
there is nothing happening there. So, there are two roots, here there you go, they meet at -
−3 which is the break in and break away point. So, we just roots break in here and they
also go away from here and then the three guys which come in 1, 2, 3 and then they go away
here, here, and here. One pole is going to the 0 here; the remaining two poles are going to
infinity.

So, that’s all we are just following all this steps as before.

(Refer Slide Time: 08:16)

406
Okay so, the second guy: so here I have 2 and 1: 3-zeros and I have 1, 2, 3, 4, 5: 5-poles. So,
5 poles would mean 5 branches. So, let us just locate these poles and zeros. So, I have a zero
at −1 and then this is the zero at 0.1+ j 6 , another zero at −0.1− j 6 . Now these
two will tell me that there is a pole, again I am not drawing to scale, but just to make this a
little more visible 0.1+ j 6 .6 and here is will be −0.1− j 6.6 and somewhere here is my
origin; sorry, this is my origin and these are my poles. So, the distance is 0.1 .

And then there additionally there are two poles here at the origin, so we have 1, 2, 3, 4 and
one guy is sitting here at −12 ; so 5− poles and 3−zeros . So, the root locus will
have 5 branches, 3 branches will go to the zeros: 1 zero here, 1 zero here, and 1 zero here.
And then there would be two of them going to infinity, we will find out how they would go to
infinity.

The first question we would like to ask is well which part of the real line lies on the root
locus. So, I look at the region between 0 and −1 , I don’t really look at these guys
because they don’t really change the even or odd numbers. So, if it’s even before and after
them, if it’s even before them it’ll be even after them and so on. So, I don’t really need to
count this. So, I start from the origin here between 0 and −1 , I just see two guys to the
right. So, this is not allowed. Then I come here and I say −1 , I just see between −1
and – 12 , I see three guys to the right so this entire region is on my root locus. So, one
thing I know that this guy will move in this direction.

So, then I can just look at it and see well maybe this is the pole goes to 0 so maybe just
goes this way right; that is a lazy way to look at it from this front from here and this one to
here this one is going to infinity and these two guys I do not know just here they just do not
know what to do and then you go to infinity from here that may not be true, because there are
some other rules which we need to follow as well. So, I just remove this in, the not about this.

So let us see the; well, let’s see the angle of the asymptotes. Angle of the asymptotes well
you could easily compute them to be ± 90° like two guys which are going to infinity. The
centroid of the asymptotes σA that would be again of all the poles is −12 minus; I can
I just take all the real parts if you remember how we derived these things; minus the zero here
which is −1 and then the real parts of the poles 0.1−0.1 everything divided by
n−m that is 5−3 and this will give me a centroid point of 12

407
 −12−0.1−0.1−(−1−0.1−0.1) −11
(i.e., ) and . So, somewhere here 5.5 ; this is
5−3 2
my σ A −5.5 .

The next thing I would like to do is, well the angle of departures. Departure from let me say
this pole −0.1+ j 6.6 . So, I need to look at the angles which these two guys make here,
right? and then I need to look at the angles which these guys make. So, let me call this zeros
as, let me call this as Ψ1 the angle made by this zero, let me call Ψ2 by this zero, and
then Ψ 3 by this zero. And then I have five poles right; so, let me call this φ2 , φ3 ,
φ 4 this is the zeros and φ5 .

So, the angles of departure and I just need to compute first what is φ1 ? and then φ1 ; so
is all these angles right? I’ll just say angle by all the zeros minus all this φ2 , because there

is no angle contribution from φ1 to itself, φ2 +φ 3+ φ4 + φ5 . And then to just get the angle
of departure, what I do is I just add an angle of −180 ° here
(ie., ( Ψ 1 +Ψ 2 +Ψ 3 )-( φ2 +φ 3+ φ4 + φ5−180 ° )). Right, okay.

So, let us compute each of these angles. So, I am just looking at the zero starting from Ψ1 .
And then, from here till here if I measure the angle that will just be a +90 ; so I have a
90 plus, what is Ψ 2 ? Let’s do Ψ 3 , that’s easier right, Ψ 3 would also be 90
and then this angle would be the angle over here. So, that is the inverse tangent of this
distance that is 6.6 divided by the horizontal distance that is 0.1 or 1−0.1 . So, that
is 0.9 ; so this one. And then minus we have the angles from phi 2 that is very very nearby
so I could as well I approximate this by 90 degrees. So, φ2 would be 90 , φ3 would
also be 90 , φ4 would also be 90 , and then I have this guy that is tan −1 of from
this zero again so that is 6.6 by the horizontal distance 11.9 ; and all of this −180,
so that will give me an angle of sorry; of 142.6 .

(
i.e., φ1= 90+ 90+ tan
−1
( 6.6
0.9 ) (
) − 90+90+ 90+tan ( 6.6
11.9 ))
−1
−180=142.6

So now, I know that from here I am actually going this direction. Similarly, I could also
compute what is the angle of arrival here: the angle of arrival here would just again depend
on well, this angle. Again all these all the other contributions. And that we could easily

408
compute again just by doing all the same things to be similarly about −(142) , so it is just
arriving from here. And then angle of −142 approximately right? because there will be
some little change because of this 0.6 number.

So, I know that this guy is not directly falling on to this directly here which is taking a little
detour and coming here and doing this. Similarly, this guy also will do the same right this is
also should be easy to compute this and this. Now, what happens to these guys; so this guy I
know goes and sits up here. So, we need to compute what are the breakaway points. So, break

dK
away points could be computed by just looking again at =¿ 0 . So, again it is a
ds
little complex expression here, but it will give us some points and let’s see how the final plot
looks like.

So, what we would expect is of course, there will be one breakaway point here, these guys
should meet at some point and then go to infinity along asymptotes given by this one.

(Refer Slide Time: 17:54)

So the plot would say well there is a breakaway point here, there is a break in point here. And
then let us look at each of these roots. This is easy to find out this guy goes here, this guy
goes here, these two poles well this guy goes here, this guy goes here, this guy travels here
and then they come here and then they just go away to infinity this is a point −5.5 . So, I
am just avoiding computing this one because it will just be a little too ugly to write down
these expressions, but we could compute.

409
So, the idea here is to just that the complexity here is not to blindly say that this guy will just
directly follow on to this guy and this guy you just directly fall out of this guy, right. So, they
will actually follow a different path like this and like this and then the rest all is just a

dK
computational procedure you have to and then equal to 0 and then you get this
ds
plot which looks like this, right. This is a little ugly here, but then we have to draw it by hand
so it looks a little distorted. But this is more or less how the pot actually looks like.

(Refer Slide Time: 19:07)

The other example: so again I have all and again a unity feedback system let’s see where my
poles and zeros are I start by drawing this diagram here. So, I have again a pole at origin, a
pole at −2 and then I have a pole here at −1+ j2 , −1− j 2 . So, m=0 , n=4 .
So, which part of the root locus lies on the real axis? There is one pole at the origin, so I look
here from the right. So, then this entire line lies on the root locus; this entire thing. So, this
guys, okay let’s just mark as these guys will travel this direction. So, what is the next thing
over here?

Then you have to compute the angles of asymptotes. Angles of asymptotes, again there are
four of them. So, we’ll have 180/4 then you keep adding 45 ° , then you have 135 °
and you have −45 ° and −135 ° . And then you are also looking at the center of the
asymptotes σA is you can just check that this is a –1 right? I am skipping all these

410
computations. So, if this is your center of the asymptotes right and then they go these ways;
this way and this way. Okay, this is 45 ° , this is −45 ° and so on; 135 ° , −135 °

Okay, so let’s see where this guy goes angle of departure from −1+ j2 . So, what are the
angles I am interested in? I am interested in this angle, I am interested in this angle, and I am
interested in this angle. So, that φ angle of departure is just the negative with all our poles,

2
it’ll just be negative.
−1
for the first guy, then you will have −tan −1 ( 21 ) , this guy at

90 ° , and then I add 180 ° from the formula.

φdep =−tan −1 ( −12 )−tan ( 21 )−90+180=−90 °

−1

And I get all this, we’ll just add up to −90° . So, this guy will just vanish but still because
it’s you can also see the symmetry or so it just vanish.

So, one thing I know is from here I am moving here. Similarly, I can compute here that I just
go here and then I just going this way. So, if I am moving this way I just keep moving this
way, and I see that well something like this happens right that the entire line between
−1+ j2 and −1− j 2 is on the root locus. This is the angle of departure for −1+ j2 ;
here I can compute easily right? It’ll just be −1 from −1− j 2 it would just be
+90 ° .

So, there is something interesting happening right. So, everybody is meeting here, well that is
what it seems like right and then what do they do after they meet here. Well, now let me

dK
compute the breakaway points again by setting =¿ 0 . So, this means I’ll have of
ds
course, one breakaway point at −1 and I will also have breakaway points at −1 ± j 1.22
. So, this is 2, this is 1, so I am having breakaway points even here; something more
interesting happening. This is 2, this is 1 here and I am having at 1.22 . So, guys come here
they even go away. So, this guy comes here and then maybe then it says it seems that this guy
will go and these two poles will meet here at this guy goes here from here till here and they
meet here this guy goes from here till here and then this guy they meet here. So, something
strange is expected to happen here.

411
Now, look at even the asymptotes right: they look as if this might actually come here and then
cross the imaginary axis. So, does the root locus cross the imaginary axis? Well, I just look at
the characteristic equation and then I am looking at the Routh criterion. So, my characteristic
equation turns out to be 4 ok. So, just computing characteristic equation again from
1+G(s )=0 , I have a characteristic equation which looks like this
4 3 2
s + 4 s +9 s +10 s + K=0 .

And I do again the Routh table; and what it gives me is that, well there is a gain K at a
number 16.25 I will skip all the details. By now you should be experts in computing these
guys. And also at an ω 0=1.58 . So, they cross the imaginary axis at this point, who is
running approximately again not to scale this is 1.58 with a minus sign and this is 1.58 .

So, now I know that it seems two of the roots will cross the imaginary axis. So, let us again
try to analyze what is this; this guy is going here and well it is a breakaway point. So, he is
just waiting for somebody else to come and meet him here so that he could just say hello and
go by. Similarly, this guy, this guy starts here there is again a breakaway point. So, these are
complex breakaway points. So, these guys come here and this guy also comes here they meet
and I say well we have to go away because it is a breakaway point. So, this guy comes here,
he goes and then they meet here. After this well they decide where should they go they should
go asymptotically along this line right; this is this very this guy will go, this guy will go this
way right. Again at angles 45 ° , −45 ° and so on.

So, to the asymptotes, again this will be like this and this will be like this; sorry just messing
up the scale here, but the point where it meets the imaginary axis is at this omega naught
which we computed from the Routh table this is 1.58 .

412
(Refer Slide Time: 26:49)

So, I will just show you how the plot-turn outlooks likes. So, this guy comes here goes here,
this guy goes over here, this guy goes over here, this is, this point is 1.58 , this point is
−1.58 , again this guy here, here, this, this, and this. So, just we just using all those steps
to just see the entire plot right? And then see this intersection with the imaginary axis. So, I
know that after some point if I keep increasing the gain the system might actually go unstable
it is not might, it will actually go unstable after my gain increases to beyond 16.25 , right?

You could just do this on MATLAB; I will shortly post a little tutorial for yourself to see how
to use MATLAB to get these basic plots. Other plots like the step response on how to give a
command for the transfer function, find the poles, and zeroes and so on. So, I will just post a
little little a tutorial on that. I will not do it in the class over here, but that tutorial would be
self-sufficient for you to learn by yourself.

413
(Refer Slide Time: 28:07)

So, far things were nice. If I give you a block diagram which looks like this and I say for the
system given plot the root locus as this guy K1 is varied from 0 to ∞ ,with this
parameter β=2 , and then this should be 5 here; repeat with β=5 . Not only that, then I
keep K1 fixed right? and then I vary β from 0 to ∞ . So, how do I handle this? So
far, everything was given to me that a G( s) was K in the numerator, some zeroes here
and then denominator. So, how will I do this?

So, first is I need to get the characteristic equation. Now given this guy do I know anything
which I learned in the past; that would directly give me the characteristic equation. So, for

1
∆∑ K K
that I will do the signal flow graph analysis, the transfer function had TF= P ∆ or

something, right? I don’t really need to do the entire transfer function; what I just need to
know is what is in the denominator of the transfer function, the characteristic equation what
was it 1+G(s ) H (s) if I had a K outside this was 0 this is just a denominator of the
closed loop transfer function.

414
(Refer Slide Time: 29:57)

So, first let’s draw the signal flow graph of this. So, the signal flow graph would look

2 K1 1
something like this. So, 1 , 5 and , then I had and then I had ,
s+10 s+ λ s

2 K1
with a 1 and a Y . So, this is we have R( s) , 5, , ; sorry I called it a
s+10 s+ β
λ here. So, let’s erase that and put a β . So, what happens after all this? So, I have after
this guy, I have a loop of −1 , from here till here I have a gain of 0.1 , this is plus or

K1 1
minus, okay it is plus. Then later, between and I had a 0.2 here, and here I
s+ β s
have 1 . And here I have 1 . And, now these are all the directions of the arrows.

So, I am just interested in finding the overall ∆ . So not really the transfer functions,
because I know that I am only interested in the denominator which I call the characteristic
equation. Why is that? Because, I am just interested in how my poles or how the poles of my
closed loop system vary as I change a parameter from 0 to ∞ , which is K 1 in this case.

So, the ∆ here is 1−( L1+ L2 + L3 ) , well you look at all the individual loops there are
three loops here right: one loop this arrow is here, one loop is this one, second loop is this
one, and the third loop is the outermost loop. So, the ∆ turns out to be 1 plus; I just look at

1
this one right, so I have 1+ plus; I look at the other guys right. So, I have 10 times
s+10

415
 2 K1
K 1 over 0.2 sorry, there is something wrong. So, and this take the 5 and
( s+10)(s+ β)
0.2 just becomes 1; so that just goes away from here.

And then the last term: so I go the entire distance here I have 5 and 2, 10 times K1 over
again. So, I have s in the denominator s , s +10 , s + β . So, this is

1 2 K1 10 K 1
1+ + + =0 , this is my characteristic equation. So, if I
s+10 (s+10)(s+ β ) s ( s+10)( s+ β )
want to draw the root locus in terms as with K 1 as a parameter, I need to get the equations

in the form 1+ K 1 times something in the numerator polynomial a( s) , something in the

denominator polynomial b( s) equal to 0. So, it should look something like this

a( s)
1+ K 1 =0 .
b( s)

Okay, now can I rewrite this equation that way; I just do some manipulations and what I get

s ( s+ β ) +2 K 1 s+10 K 1
is. So, first this will become 1+ =0 ; which I could also write as, this
s(s+ 10)( s+ β)
is equal to s ( s+ 10 ) ( s+ β ) + s ( s+ β ) +2 K 1 s+10 K 1 . i.e.,

s ( s+ β ) +2 K 1 s+10 K 1
1+ =0=s ( s+ 10 )( s+ β ) + s ( s+ β ) +2 K 1 s+10 K 1
s(s+ 10)( s+ β)

So, this will further simplify into s ( s+ β )( s+11 )+ 2 K 1 ( s +5 )=0 , which I could also

2 K 1 (s +5)
equivalently write as 1+ =0 . Okay
s (s + β)( s+11)

So now I have, well I can just do this right so I have assuming β is always greater than 0.
I have a zero at −5 , I have a pole here, and a pole at −11 . And then depending on
wherever the β is, I will have the location of the other pole. I’ll just do the simple one for
myself so β=5 . So, I will have a pole also here. So, this guy will just go away, so there
will just be nothing here. So, I will just be left to the transfer function which looks like

416
 2 K1
1+2 K 1 , these guys will cancel out s ( s+11) (i.e., 1+ ). And, I easily know
S (S +11)
how to draw the root locus for this right; I am not going to enter into the details of that.

Now, when K is fixed and β is varying; I just again look at this entire expression and
now I want to write it in the terms of a transfer function of where 1, I have 1+ β ; some

'
β a (s)
other polynomial a' ( s) divided by some other polynomial b' ( s) (i.e., 1+ ).
b ' (s)
So, again I can write it down in a fairly straightforward way again right, so I just start from
here; from here. So, I just skip a bit of the computation, but I will just

β s( s+1)
1+ 2
=0 , where K is a fixed value now according to the problem
s ( s +11 ) +2 K (s+5)
statement. Well, you can just say any just K 1=1 . And now I can look at what happens to
the root locus when now β varies from 0 to ∞ , keeping K in to fixed, this being
keeping K fixed. So, the idea is to just write down this characteristic equation in this form

β a' ( s)
1+ =0 . I will skip these steps because now again the looking at the steps of the
b ' ( s)
root locus should be kind of very straightforward now. So, those steps sorry its it a β ;
β and β . Let me do this again.

Okay, we’ll now, see how to draw the root locus with β as a parameter or in other words I

'
β a (s)
want to write the characteristic equation as 1+ =0 . So, I just go from here, I just
b ' ( s)

βs ( s+11)
skip the computations here so I just have 1+ 2
=0 . So, for a fixed K
s ( s +11 ) +2 K (s+5)
I know what will be the poles here I know what are the open loop zeros and I could just say
as β going from 0 to ∞ . I will skip the steps because these are just the routine step.

Just to give you an idea of how to formulate a problem in a way that I could use the root
locus techniques directly. It was not very obvious from here because so far all the problems
were very nice. That I was just given G and H and the K and I just had to do. If it
is a little complicated there is nothing really to worry I can just look at my signal flow graphs
and then just compute the ∆ , which is a denominator I don’t even really need to look at

417
what the numerator is like because I am just interested in the characteristic equation which is
essentially the denominator; and then the steps will follow.

So I just stop here, and I will leave these things for you as an exercise to do.

Thank you.

418
 Control Engineering
Dr. Ramkrishna Pasumarthy
Department of Electrical Engineering
Indian Institute of Technology, Madras

Module - 06
Lecture - 01
Introduction to Frequency Response

Hi guys, so far what we have done is classified as Time Response Analysis. So, we started by
defining the transfer function, we saw how the step response of the system looks like, and we
characterized things like the overshoot settling time, we also characterized various errors like
the position error, the velocity error, acceleration error and so on.

Later on, we used some algebraic methods to compute if the system is stable or not, whether
or not certain poles or all of the poles lie to the left half of say some point of −1 because
if I reduce stability, then we had the root locus criterion where we had a graphical technique
to see how the roots of the closed-loop system change as the gain K was increased from 0
to ∞ . So, we have not done any design problems so far, but what we learnt through the
root locus technique is that I can guess or I can predict accurately the behavior of the closed-
loop by just looking at the open-loop behavior. So, when I do the root locus, I just need to
know what is G and H , right, and then, I can just tell how my closed-loop system
would behave based on changes in k from 0 to ∞ . I also saw examples where after a
while we keep on increasing k , the system becomes unstable, and so on.

So, what we will do today is to look at the other part of the analysis which is the frequency
domain analysis to see what happens to the system or how the system responds to inputs of
different frequencies. So, let us do a little bit of introduction before we actually go to more
deeper analysis.

419
(Refer Slide Time: 02:15)

So, I start with as usual systems which are linear time-invariant, right. So, the frequency
response of a system is the steady-state. Now, this is important. The steady-state response of
the system towards sinusoidal input signal and I will tell you why we are only interested in
the steady-state response. So, the reason why we study these kind of sinusoidal outputs is if I
give pass a signal through LTI system, if this signal is sinusoidal, the output is also
sinusoidal, of course, what will change is its magnitude and there will also be a bit of phase
shift depending on what is sitting inside the system, right. So, we will exactly characterize
how the B will change based on what is sitting inside here, and similarly, how the φ
will change is again based on what is sitting inside here, ok.

420
(Refer Slide Time: 03:24)

So, why do we need to do all this analysis? So, most of this analysis or what we call also as
the frequency response is independent in a way of the amplitude and phase of the input test
signal, right. So, we will see why that is true, and when I go to design a certain system. So,
the design and parameter adjustment of the closed-loop can be carried out equally or rather
easily in frequency domain than in the time domain. So far we haven’t looked at design
problems, but this will be obvious when we start handling design problems in the next, in the
coming lectures.

We also saw what the effect of noise and disturbance in terms of the frequency, right. So,
when we were talking about sensitivity and the complementary sensitivity function, we
classified the disturbances as low-frequency signals and noise as high-frequency signals and
now, we will see how we can actually kill the effect of noise in a certain setting, right. So, the
effect of noise and disturbance parameters can be easily visualized, and then, of course, there
is always a correlation between the time domain and frequency domain specifications which
we will derive shortly.

So, given you know certain performance specifications in the time domain, I can actually
even translate it into frequency domain and another thing what I could also do is if I would
not know how the transfer function of a system looks like, I can just give it certain signals of
varying frequency and based on those frequency response, I can easily estimate what is the
transfer function of the system is. You may also say well why not just look at the step

421
response of the system and get the transfer function because the step response included ζ ,
it included ω n .

I could measure the peak overshoot, I could measure the rise time, the settling time, bunch of
other things, and based on these things, I could get the transfer function. Then, what is the
problem? Well, one drawback over there is that we just analyzed second-order systems, right
and sometimes it might be a little cumbersome to even compute those things. What if the
system is overdamped, right? What if the system is how do I find out if it is actually critically
damped or it is not, there could be some difficulties, right in just looking at the step response
and looking at and then estimating the transfer function based on that the frequency response
analysis or just looking how the frequency, how the system magnitude and phase I am using,
these terms without defining them, but it will be clear in the next couple of slides, how the
magnitude and phase of a system changes with different frequencies and I just do a plot of
that and I can get a very good estimate of how the transfer function looks like. Not only that I
could even look at what could be the error constants, right.

(Refer Slide Time: 06:49)

So, we will see that slowly as we progress through the remaining of this module, as a simple
example let us start with LTI system, where I have usual things at the system, some transfer
function G( s) , I have a input signal which is sinusoid of this form and I will just be
interested in how my y (t) changes. So, the steady-state response of a stable, well this guy
is important, right; if the system is unstable. Then, I don’t really need to do any kind of

422
analysis. The steady-state response of a stable LTI system to a sinusoid does not depend on
initial conditions and therefore, we can just assume that all the initial conditions are 0 and
we will also see why this is true. If I just take a signal of this form, I know that R ( s )=¿

Aω
2 2 . This was just the Laplace transform of this guy and the output now becomes
s +ω

Aω
C ( s )=¿ R ( s ) , ok.
s + ω2
2

(Refer Slide Time: 07:54)

So, what I know I can just do all the partial fraction expansion and all those things which I do
to get how the time response of the system or how the solution of C( s) looks like in time
domain. I get something like this, right. So, why do I call this as a transient response?
−t
Essentially because these guys will consider term will have terms like k1 e and
−2t
depending on where the locations of the poles are it could be k2 e and so on. There
−2 t
could also if there are repeated roots, there could also be terms like t k 3e , but I know
that this term goes to 0 as time goes to ∞ and therefore, I am just looking at the
steady-state response because as t goes to ∞ , this entire thing goes to 0 and
therefore, I also said earlier that I am interested in stable systems.

If the system is unstable, it would mean there is a pole on the right half-plane and that could
6t
be something like k4e and you know that this guy actually blows up to ∞ as time

423
goes to ∞ , right. Therefore, there would be nothing left to analyze, right. So, these guys
go away. So, the output contains of course the transient terms and the steady-state
components. Now, at steady-state, these guys die down and what is remaining is the effect of

k k¿
only these terms + , right and these are essentially due to the input, right this
s+ jω s− jω
guy, ok.

Now, what is this response at steady-state when the input is a sinusoid? If you look at the
earlier analysis in the step response, even though we had oscillations in the system. Those
oscillations were dying down, right say if I just take an undamped system, we had response
which was like this for the step response and see there were still oscillations, but they were
actually going down right with time. So, there you had terms like e−t cos ωt and so on,
right. So, even though whatever frequencies we had, the oscillations just died out and
whatever remains after these things go away is this guy.

(Refer Slide Time: 10:34)

So, I just look at how the response looks like C ss and at the end, I am calling as ss

k k¿
because of steady-state + and I just use the simple tricks from what I learnt
s+ jω s− jω
in Laplace transform of how to find these constants k and k ¿ . They turn out to be
something like this, right. k would be this guy and the k¿ would be this guy. I don’t
really need to teach you how to compute these. These are extremely straightforward things,

424
right and then, I take the inverse Laplace transform. Well, these are also kind of easy to
compute, and what I get as the result is I do all the computations and what I get you know
this is important, right.

(Refer Slide Time: 11:14)

So, the response is B sin ( ωt + φ ) , where the input what was how did our input look like.
The input looked like A sin ωt . So, A sin ωt now gets transformed to a signal
B sin ( ωt + φ ) . What does this B , where does this B come from? Well, this B is
simply B= A|G( jω)| , right the system which you are passing the sinusoid through.
Similarly, φ , I could just even compute it as just the angle of this guy, right. φ is just
the angle of G( jω) given this is G . I know G is a complex number and I could
easily find out what the angle is and what is also important to notice here is that well I used
the word the sinusoidal transfer function, right which is essentially the substituting s= jω
because I am only interested now in the steady-state things.

Only the frequency components are remaining, though I can just ignore the real component,
right. I’m just looking at these guys s= jω and therefore, I just take a transfer function, I
substitute s= jω , call it the sinusoidal transfer function and that will help me in the entire
frequency domain analysis, right. So, this is little important and this is little simple trick also.
Here you could see that right all the other guys just disappear and what I am left is just jω
terms, just these two poles are remaining right and this is why I call them as the sinusoidal
transfer function because, after all these things, this is what is left, right. So, that is even

425
obvious through the entire steps which we do, right. So, we just compute k =G(− jω) that
is just when you get the partial fraction substitution k would be ok.

So, k would be just I think let’s do some computations here substitute s=− jω and
then, you get this thing, right. Similarly, this guy you substitute s= jω and then, compute
k¿ and then, you see all these are jω , jω , jω , right. Nothing to do with the real
j
term, right, this is actually a result of a systematic computation and I know that e of
−j
something, e would actually give me a sinusoidal signal. This is again some basic
complex analysis, right. So, what is the conclusion or the first observation here is I pass a
signal A sin ωt through a transfer function G( s) . What I see in the output is that I have
a different magnitude, I still have a sinusoid with a little phase shift and the magnitude is the
magnitude of original signal minus, sorry multiplied by the magnitude of G ( jω ) the
transfer function. Now, I also know why I only take this jω , and then, of course, the φ
is result of something like here, right. Let’s do a very simple example to see what this
actually means.

(Refer Slide Time: 14:52)

So, let say I have G(s) , I have r (t)/R (s) whatever we call it and then, I am just
interested in how C( s) looks like or the c (t) , right all in steady-state. So, let me

k
assume that G(s) is something like some constant over . So, this is what is
TS +1
called as the transfer function in the time constant form. I am not really writing it in the pole-

426
zero form which is just be that you know s plus something. I’m just dividing that and this
is called the time constant form and it is little useful to do this in the frequency domain
analysis, right. That will actually be obvious when we do things like the Nyquist or the bode
plot.

So, I just first write down the sinusoidal transfer function for obvious reasons. So, this will be

k
the G ( jω )=¿ and let me assume that the input is just say a sinusoidal transfer
jTω+ 1
function A sin ωt . So, what this entire analysis here tells you is that the output or the
steady-state output which I call as c ss (t) is simply computed by the |G ( jω )| . What is

k
the magnitude here? Well, that is straightforward to compute that is . Similarly,
√1+ω2 T 2
−1
the ∠G ( jω )=−tan (Tω) . This is just the complex numbers and we know how to
compute those angles.

So, c ss (t ) would be A multiplied with the magnitude of G ( jω ) . What is the

k
magnitude of G ( jω ) ? That is and sin ( ωt +φ ) , this is the
√1+ω 2 T 2
φ=−tan−1 (Tω) . So, this is actually straight
k forward now, right.
c ss (t )=¿ A sin ( ωt−tan−1 (Tω))
√1+ω T 2 2

Just look at what G is. Compute its magnitude, compute its angle and I can easily see
what is the output signal. I don’t each time, I really wouldn’t need to compute the inverse
Laplace transform and then, do all that, substitute s=− jω , + jω nothing, right, how
the magnitude changes? Magnitude just changes with the magnitude of G ( jω ) . How does
the phase change? Phase just changes with the angle of G ( jω ) . So, the reason of doing the
entire previous steps was just to get this expression in a reasonably straight forward manner,
ok.

427
(Refer Slide Time: 18:36)

G(s)
For single loop systems, well we just have the transfer function and then, I
1+G ( s ) H (s )
am interested in this sinusoidal transfer function. So, I just substitute s= jω and then, I

|G( jω)|
can just compute the . Similarly, with the angle, would be the angle of
|1+G ( jω ) H ( jω)|
the numerator guy that is the ∠ M ( jω)=∠G( jω)−∠ (1+G ( jω ) H ( jω)) , right. So, things
get a little easier to compute, right. This has become the just as addition or multiplication or
additions or subtractions, ok.

428
(Refer Slide Time: 19:25)

So, the idea is to check now how the magnitude changes, with changes in ω . Is there,
some strange thing happening over there which might give us some information. Similarly,
what happens to the angle as the frequency is increased from 0 to ∞ , right we have all
kinds of frequency signals available, right. So, there are few terms here which we will slowly
derive.

One is called the resonant peak. It is in some way you can say well this is not directly related
to, but you could when I was doing the time response you know as time was going from 0
to ∞ , I defined what was or at what time did I reach the peak overshoot M p as I called
it, right and then, I just did some, put d of the magnitude by d (t) equal to 0 and I get
some time where it actually reaches the peak value and I could even exactly compute what
was the peak value something depending on zeta e power minus pi zeta some formula we
had, right. So, similarly here does the magnitude reach a peak at some frequency and this
frequency, if it exists, is called the resonant peak, right. The system with a large resonant peak
will exhibit a large overshoot. We will see that, right. So, in general, this guy gives indication
of relative stability of closed-loop system. We will just keep this in mind for the future thing.
There will also be a corresponding to the resonant peak, the resonant frequency, the
frequency at which the output becomes maximum.

So, this guy becomes maximum at certain frequency and that frequency is what we call as the
ω r . Earlier we had defined it as a peak overshoot Mp and Tp and the time where

429
actually I reached the peak overshoot, then something called the bandwidth which I think is
the term which we use very often, but formally it means the bandwidth is the frequency at
which the magnitude drops to 70% of its original value or its 0 frequency value. This
scaling we will define little later, ok.

(Refer Slide Time: 21:54)

So, this is how a typical response would look like depending on ω r . So, this is on the
horizontal axis. I have increasing frequency. This is on I will also tell why we use this in in
the log scale. Some books will also give you this in the normalized frequency form, where

ω
u=¿ ( )
ωn
and this is plotted as u goes from 0 till ∞ , so that we will see that

a little later, right. So, typically I will also tell you why this is 1 and u=0 or ω=0 .
I start this way and I could reach a peak and then, I go down this peak is Mr as defined
earlier. We called this the resonant peak and the frequency at which this occurs is called the
resonant frequency.

Now, you see this starts from 1. It goes up and then, falls off to certain value and this is what
I call as the cutoff rate, right or the bandwidth, right. This frequency from point from 0 till
this is called the bandwidth because for other frequencies, you see my signal gets attenuated.
Why I say this is because again I have input signal some A sin ωT . Here you have
G( s) and you have the output some c (t) and you see c (t) is what it is A times
this magnitude and after this, you see that actually the signal is getting attenuated too fast

430
with changing frequencies, right. So, until this, I can get see some good output. After this, I
may not really be interested because my signals just become very small right.

(Refer Slide Time: 23:53)

1
This is essentially what I call as the cutoff frequency which is where magnitude is
√2
times of its peak value. This is not really the maximum value, but the 0 frequency value. I
will just change this in the slide, right, and then, the cutoff rate is the rate of change of slope
at the cutoff frequency, right this slope of this line. So, again we start with a typical
configuration of the closed-loop and I have a certain reference signal, then I have G( s)
given by this guy, then I have a unity feedback, I have certain outputs C( s) plus-minus
and so on.

431
(Refer Slide Time: 24:26)

So, the loop, closed-loop transfer function which we derived even when we were doing the
second-order system was this guy and I just substitute s= jω to get the sinusoidal transfer
function and this is how it looks like. So, let’s see what information does this guy give us. So,

ω 2n ω2n
M ( jω )=¿ 2
= 2 2 , ok.
( jω ) + 2ζ ω n ( jω ) +ω2n −ω + j2 ζ ω n ω +ωn

(Refer Slide Time: 25:21)

432
 ω
Now, let me define a quantity u which is a normalized frequency as . You may even
ωn
avoid using this, but I think the computations get a little easier. So, I can just write this guy in

1
terms of u as 2 . So, how does the magnitude of M look like? This is
( 1−u ) + j 2 ζu
what I am interested in calculating the magnitude and the phase. The magnitude will decide
how my original signal will get transformed, what will be the frequency of the result, sorry
the amplitude of the resultant signal, and the angle will decide how much is the phase shift.
So, this M ( jω) is, what’s the magnitude.

So, in terms of this guy, the real part and the imaginary part, |M ( jω)|=¿

1
. Similarly, the angle of M or even there is a closed-loop. So,
√( 1−u ) +( 2 ζu)
2 2 2

2ζu
∠ M ( jω) which I call as φ=−tan −1
( 1−u ) 2 . Again u is the normalized

ω
frequency which is and ω n is the natural frequency of the system. So, let me take
ωn
my reference earlier we had a unit step. So, let’s keep the unity thing here again
R=1∗sin ωT and see how this guy looks like.

1
So, c (t )=
√ ( 1−u ) + (2 ζu )
2 2 2 (
sin ωT −tan−1
2 ζu
( 1−u 2)) . Now, let’s see what happens as we

change the frequency. Well, I don’t really want to do all the values from 0 to ∞ , but
just look at some typical values first is when u=0 , just put u=0 here, in that case,
M =1 and what is phi? at u=0 ; φ=0 , right. So, when the input frequency is 0 ,
well the output is just you know it is just the same thing, right. Nothing comes out. What just
comes out is 1∗sin ωT , okay.

Now, second is when u=1 which means ω=ω n , the natural frequency of the system.

1
What happens to M at u=1 ? m is, so this guy goes away. I have M= and φ ,
2ζ

π −π
when u=1 , this guy becomes very large, and tan −1 ∞ , I know is . So, φ=
2 2

433
or with a negative sign. Third thing for very large values of frequency when u=∞ , the
magnitude goes to 0 and what happens to the phase φ as u becomes very large, the
phase becomes −π , right.

So, I know well M starts at certain value of 1 , that is for 0 frequency signals, then
at u=1 , where ω=ω n . The natural frequency I have some other value and as u
goes to ∞ , my magnitude goes to 0 and φ=−π , then we can ask ourselves a
question, ok.

(Refer Slide Time: 30:43)

Based on all these three values, u=0 , u=1 , u=∞ , when does M have a peak if
at all? So, what do I do? So, what is changing here? The frequency is changing. So, I just

dM
look at =0 to get its peak. So, this I just do the computation. I just keep here
du

ur =√ 1−2 ζ 2 or equivalently ω r=ω n √1−2ζ 2 . So, if I substitute this ur in the

1
expression for M and φ , what do I get? That M r=¿ and similarly,
2 ζ √ 1−ζ 2

√ 1−2 ζ 2
φr =−tan
−1
( ζ ) , right. So, what do we observe from here? Let’s see what happens

434
 1
when ζ is between these values, right. You see right from this expression; ζ= . I see
√2
that this goes to 0 , right. So, I put an equality here, right.

1
So, when 0<ζ ≤ , the resonant frequency ω r ≤ ω n , right and then, the resonant peak
√2
1
M p for these values. Look at these 0<ζ ≤ . So, what was the resonant peak? The
√2
1
resonant peak by the formula was M p= because this is for ζ >0 . Therefore, I am
2ζ
not looking at the value of 1 and I am also not looking at the value of very large
frequencies, right. So, this is the thing and therefore, for these guys the peak, the resonant

peak will always be greater than 1 . So, the second case when ζ>
√ 1
2
, I will ask myself

dM 1
well is =0 for any of these? right. So, it turns out well when ζ> which
du √2
means this number will not exist, right.

dM
So, there is no ur , ω r for which =0 or its peak will only be when u=0 . So,
du
let’s try to just plot these two things, right. I am just plotting over u with M . When
u=0 , M =1 . So, if I look at these things, I get a plot which looks like this. Whereas

1 1
So, for the first case when 0<ζ ≤ and then, for all other values when ζ> , I just
√2 √2
1
get something like this, right. This is for ζ > . There will be no peak, ok. So, now all the
√2
bandwidth and all I can define. So, this is my 1 . So, this is say 0.707 , this guy will be
my bandwidth, this guy will be ωr or where u=1 and so on, right. This is the
explanation to why we actually get a picture which looks like this, right.

So, then in this case, well again there is no resonant peak. So, M decreases monotonically
as u increases, then you see that well just look at these things over here M r . What is
this M r ? Well, this Mr is again depending on ζ . So, the resonant frequency has
some, sorry Mr has some direct relation with the damping factor, right and similarly, this

435
guy ω r=ω n tells me something about the natural frequency of the system given a
particular value of damping, given ζ =0.1 , then well ωr has something to do with
ω n and if you remember these were essentially the two things which were defining all the
time domain specifications ζ and ω n , right and therefore, there is a nice correlation
between what was observed in the time domain and what is now observed in the frequency
domain.

One more observation is that I will just write down here control systems are usually low pass
filters, you see right here. So, when u=0 , M =1 all the time, right and you see well if

1
my ζ is less, it actually allows this guy and if ζ> , then it still allows some low-
√2
frequency signals to pass through. So, all control systems, most of them are usually low pass
filters especially because of this and then, what we observe also from here and then, little
more about the bandwidth, right. What is bandwidth? When the magnitude reaches 0.707

1
or , so this entire thing from here till here is the bandwidth. So, if I look at, I know the
√2
1
formula for M right the magnitude and so, M =¿ . Then, this
√( 1−u ) +( 2 ζ u )
b
2
b
2

actually gives me that this little approximation gives me that ub=−1.19 ζ +1.85 , ok.

So, the bandwidth is also related to the damping factor of the system, right. So, there is also
lots of good relation between the system parameters. What were the system parameters when
we were dealing in? What were the system parameters when we were analyzing the time
response? It was essentially ζ and ω n and these two guys characterize lots of things of
the system. Similarly, here give me ζ I can find out what is the bandwidth, give me ζ I
can find out the resonant peak, and so on. I can determine the bandwidth, resonant peak, I can
even find out what is the ω r , right.

436
(Refer Slide Time: 40:37)

So, just to summarize we just looked at this is the resonant peak, this is the resonant
frequency, the bandwidth. Again equivalently we computed. This way I will skip this
computation, but this is again not really difficult to find out just by looking at formula like
this, ok.

(Refer Slide Time: 40:56)

So, we just introduced ourselves to the concept of frequency response, we listed out few of
those advantages, we just listed out, but we will slowly build upon explaining or justifying
those advantages. We define the concept of frequency response, the sinusoidal transfer

437
function, we also defined few frequency domain specifications with respect to second-order
systems. In the remaining of this lecture, we will or the lecture 2, we will look at what are
called the frequency response plots. Typically, we will look at Nyquist plot and define what
are the stability margins, right. Relative stability is what we were talking of for a long time
and we can see how in the frequency response analysis we can actually characterize this
relative stability with the help of what we call as a gain margin and the phase margin.

We will also see some important tool called the bode plot. You may have heard of bode plot,
maybe while doing circuit analysis or networks and systems, but if you have not, don’t really
worry. We will also look at the bode plot and we will see how these two actually give us a lot
of information about the system, the system stability, gain margin and phase margin, and how
we could formulate our design problems as appropriate bode plots of the closed-loop system.

Thank you.

438
 Control Engineering
Dr. Ramkrishna Pasumarthy
Department of Electrical Engineering
Indian Institute of Technology, Madras

Module - 06
Lecture - 02
Frequency Response Plots

Hello everybody. So, to continue on our discussions within the frequency domain
specification, so what we had seen last time were few very basic things. That okay, why this
analysis is important for us. The first part is that well I can see how my system behaves at
different frequency signals, right. This will also help me as we saw in terms of sensitivity of
my system, sensitivity to disturbances, sensitivity to noise and so on and these signals occur
at different frequencies.

We also saw that if the input to a linear time-invariant system is a sinusoidal signal, the
output is also a sinusoidal signal. The only difference would be that the magnitude would
change and the change in the magnitude would depend on what is sitting inside my transfer
function and also, the phase shift which is observed and that phase shift is again depending
upon the transfer function or the angle of the transfer function. Then, we characterized a few
things which we called the resonant peak. We called, you know we had characterized the
bandwidth and so on and all this were very beautifully again related to the system
specifications like the damping, coefficient, the natural frequency and so on. So, we still
didn’t worry about stability while doing that, because we were just dealing with second-order
systems and the kind of systems which we were dealing with in second-order, we always
know that it is stable, right.

When we write ω n2 + s2 +2 ζs ωn , we know it is always stable. So, what do we do if I have a

system which has more than 2 poles, if 5 poles or 6 poles.

439
(Refer Slide Time: 02:06)

So, today we will try to answer a few of those questions with the help of some frequency
response plots. So, these plots are what are called as polar plots. So, what is a polar plot? The
polar plot of a sinusoidal transfer function, again this is important and we know why in the
steady-state analysis in the frequency domain, we just replace s with jω because after
all the transients die out, only the sinusoidal components remain and we had a little derivation
also of that last time.

So, we are actually plotting the magnitude versus the phase angle as the frequency is varied
from zero to infinity, right. This polar plot is also known as the Nyquist plot. The only thing
we have to be careful in the Nyquist plot as we will see shortly is that well, we also deal with
negative frequencies and I will tell you why that is important. So, why do we need this polar
plot? So, the advantage is that it depends, it depicts very nicely the frequency response
characteristic over the entire frequency range just in one single plot. Just one graph will tell
me what happens at zero frequency till infinity, but then I really wouldn’t know the
contributions of each of the individual factors of the open-loop transfer function. I just look at
everything as one single block and just one single graph, nevertheless, there are still lots of
nice things which we could do, ok.

440
(Refer Slide Time: 03:40)

So, starting with some very basic transfer functions, so an integral factor right which has a

1
transfer function , the polar plot I just write down the sinusoidal transfer function which
s

1
is . It is just the negative imaginary axis. Why is this true? When I just say ω=0 ,
jω
well I have a magnitude which is well at minus infinity and as ω tends to infinity, my

1
magnitude goes to 0 and since it is just . My angle is always over here, right. I am
jω
always on the imaginary axis.

Similarly, I just take the derivative factor of just a pole with just s is a transfer function.
The sinusoidal transfer function becomes just jω and this is just the positive imaginary
axis as you would see, right. So, when ω is 0, the magnitude is 0 and it keeps on
increasing. As ω goes to infinity, you may ask sir this is not a good transfer function
because the number of poles are less or there are actually no poles. There is just one zero, not
physically realizable. All that is okay, and I am just giving you a little indication of how these
things look like in the frequency domain. Nothing about saying that this is a transfer function.
Therefore, I don’t even call it transfer function. I just say it is a derivative factor, right
M ( j ω) . When ω is 0, I have a magnitude of 0 and I just keep on increasing the
frequency, ok.

441
(Refer Slide Time: 05:28)

Actually, this ω is equal to 0 should be somewhere here. I will just correct. So, something
else this looks a little more useful right because I can realize this system and what this system
is, it is just a simple RC filter. I can just draw a little circuit diagram for it just for us to
make a little sense. So, I have a R element. We have done this while we were deriving
basic transfer functions. So, I have R , I have C . The input voltage here and the output
voltage here. So, I will not derive a transfer function, but I will assume that you know this,
right. So, I will just directly write down the sinusoidal transfer function, not even the overall
transfer function.

1
So, this M ( jω )=¿ , where T is the time constant, T =RC . Now, again
1+ jωT
I am interested in how this guy, this is a complex number, right. It will have a magnitude and
a phase or a phasor. How will this guy move or change as the frequencies go from 0 to

1
infinity? So, this guy I can write this in terms of the magnitude as and the angle
√1+ω 2 T 2
would be the ∠ −tan −1 ωT . This is called the magnitude M and this is the angle φ ,
right. So, let’s observe what happens; first is when ω is 0 right? This is just the DC signal,
when ω is 0, the magnitude M , I just omit the mod symbol. It should be obvious the
magnitude is 1 and the ∠ φ=0 , right. Just put ω=0 and 0 , it is obvious, ok.

442
So, at ω=0 , I just have a phasor which is 1 ∠ 0 , magnitude 1 angle 0 . My
horizontal axis is my reference. Therefore, this will be this one, right. So, this will be the
entire thing when ω=0 . Now, let’s do the other extreme. What happens as ω goes to
infinity? As ω goes to infinity, what would happen as ω goes to infinity? What you see

−π
is that the magnitude goes to 0 and the ∠ φ=−90 , right. The −tan−1 ∞= . So, I
2
have a phasor which looks like 0 with an ∠ −90 . So, it will be just some tangential,
right. This one, this is 0 ∠ −90 , ok.

1
So, the third thing is what happens when ω= , in that case, you can compute the
T

1
magnitude M to be just and the ∠ φ=−45 . So, we are looking at the phasor
√2
1 1
which I want to plot ∠ φ=−45 ° , that is this one and at ω= , the magnitude is
√2 T

1
and the angle is −45 degrees. So, if I just keep on you know computing for other
√2
values, I just see that, this plot looks like a very nice semi-circle from ω=0 and here I
increase all the plot travels, this direction, this direction; ω=∞ and it stops here. So, I will
tell you why we are actually doing these plots, but we just do some other interesting plots
before we even try to understand why we are doing this, ok.

(Refer Slide Time: 09:51)

443
So, this guy just had a pole to the previous transfer function. So, what do I get here is, well let

1
me see how I can analyze this. So, I have g, sorry M ( jω )=¿ , right. I just
jω(1+ jωT )
do the rationalization which I learn in school. So, I just multiply and divide this by

(1− jωT )
. Similarly, for this guy, I just multiply the numerator by j and the
(1− jωT )
denominator also by j . That is how I get rid of j , the denominator. So, while doing this,

−T
I just get this one, right. If I split up the real and the imaginary parts, I have 2 2 as
1+ ω T

1
my real part and the imaginary part j , ok.
ω(1+ω 2 T 2)

I just do the computations and I just say well what happens first when ω is 0. So, when
ω is 0, the magnitude becomes −T and well, this guy goes up to infinity, right. So, I
have some complex number which looks like this – T − j ∞ . So, what I am interested in
how can I write this as a phasor, right? How can I write this as a phasor in terms of its
magnitude and the angle? So, let’s look at it carefully, right. So, magnitude if I have a
complex number a+ jb , the magnitude is √ a2 +b2 . So, look at this closely. So,
magnitude is actually ∞ and the angle is tan
−1
of the complex part over the real part.

−1 ∞
So, tan . So, that would just be −90° , ok.
t

Next thing when ω approaches infinity, when ω approaches infinity, well this guy
becomes 0 . Let’s keep this – 0− j 0 . So, this looks like a phasor of 0 and an angle
of ∠−180° . So, at ω=0 , so somewhere here, right. So, I just cannot plot that
because I will never reach that at ω=0 . What happens is, well I have a magnitude in the
real part. So, I can just split into the real part and the imaginary part, right. Just given this,
just looking at this complex number over here, so the real part has a value of T , sorry this

1
should be just T instead of . This is just T . Sorry about it.
T

So, the real part goes or it touches this line −T asymptotically. So, this is the phasor
– T − j ∞ . So, at ∞ , I am at this one. I just asymptotically touch this line right the real
part is T and the imaginary part goes to infinity as ω equal to; as ω goes to ∞ I

444
then come back to the origin here at somewhat here right, 0 and −0 at an ∠−180°
, right. This is how the plot would look like for increasing ω . So, these are just two things.
So, it is to understand from here, right. We just write down this explicitly as the real part and
the imaginary part and compute these guys, ok.

(Refer Slide Time: 14:17)

So, the next thing what happens if I have two poles, ( 1+ jω T 1 ) , ( 1+ jω T 2 ) and just be
careful when you do these polar plots is that your transfer function should always be in the
time constant form and not like S +1 , but something like ( jω T 1+1 ) . That should be
very easy to derive, right? We are not going into the derivations of those things. Okay so, my

1
M ( jω ) . . So, this guy has a magnitude. It is just the magnitude of
( 1+ jωT 1 )( 1+ jωT 2 )

1
this times the magnitude of this, this is angle would be φ is
√1+ω T √ 1+ω T
2
1
2 2
2
2

−1 −1
−tan ω T 1 −tan ω T 2 . Now, again I do all, I could do the rationalization and all I
can write it down in the real and the imaginary parts and it would look something like this

(1−ω 2 T 1 T 2)
( −
jω(T 1+ T 2 )
(1+ω T 1 )(1+ω T 2 ) (1+ω T 12)(1+ ω T 22 )
2 2 2 2 ) .

445
So, let’s do all the steps again. First what happens when ω is 0? When ω is 0, I just get
that the magnitude is 1 and the angle φ is 0 or I’m just looking at a phasor 1∠0 ° ,
right. This one, right at ω is equal to 0 and just here, just this point; just this point, not
the; not the horizontal axis right. I’m just interested in this point. Therefore, this, this line is
not entire blue color; this one. Okay, second what happens when ω tends to ∞ , well
when ω tends to ∞ what I see is that well this, naturally goes to 0 and the angle
φ is well −90° for this guy, −90° . For this guy, that would be −180 ° , ok.

Now, I will ask myself a few questions which we could have answered even earlier, but well
for some reason, we did not because they were not important. So, the first thing is does the
polar plot touch or intersect the jω axis, right? This is a valid question to ask. So, just say
there is some point here which it touches at this point. The real part goes to 0 . So, look at
this expression, right. So, does the real part go to 0 for any value of the frequency, you
2
just equate this real part to 0 . So, what you get is 1−ω T 1 T 2 is 0 or ω is

1 1
. So, this is how you can substitute ω . ω is over here and what
√T 1 T 2 √T 1 T 2
1
you get is that the magnitude or the magnitude of |M ( jω)|ω = is simply
√T 1 T 2

√T 1 T 2 . This could just be easily obtained by substituting this in the magnitude expression
T 1 +T 2
for magnitude that is marked here, right.

1
So, this is ω is and this magnitude here is this much and then, again as ω
√T 1 T 2
goes to ∞ , you again reach the origin, you arrive at you know if you use the root locus
terminology, arrive at an angle of −180 ° to this point. So, again the same steps. Look at
the magnitude, look at the phase, just compute simply what happens at ω is 0 , ω is
infinity. Just check if the polar plot touches the jω axis. You could also check you know
why don’t we, you may also ask why don’t we see the intersection for the real axis
somewhere here or here. That you could just check by computing or equating the imaginary
part to 0 and you will only get the solution that it only touches this guy at ω=∞ and
ω=0 that you can just check from this text.

446
(Refer Slide Time: 19:49)

Now, the next thing is, so what if I add a polar j origin to the previous transfer function.
So, how does this go? So, this is just the previous guy to which I add jω . Let’s do some
simple computations before when ω=0 . I can easily see that the magnitude M goes to
∞ . When ω is 0, this guy does not contribute to the angle. The only contribution
comes from this guy that is as φ=−90 which is similar to what we had in the second
example, right. The magnitude becomes ∞ and φ=−90 , right. So, this one here also
sees that this number which was earlier T1 is now T 1 +T 2 . So, here it was just T .
Now, it has just become sum of these two, T 1 +T 2 , ok.

Then, what happens to this plot as ω goes to ∞ ? At ω goes to ∞ , the magnitude

goes to 0 . What is the angle contribution as ω goes to ∞ ? Well, it will just be φ
would be −90° , right because this guy has a −90° , sorry it will be
+90 ° ,−90 ° ,−90° ,−90 ° . It will be −270 ° and therefore, you see everything is ∞
, right. tan ∞ ,
−1
tan ∞ ,
−1 −1
tan ∞ because all with the negative sign because these
are poles and you have −270 ° or even +90 ° . So, this is what happens here, right. I can
also see now does this intersect the real axis, ok.

So, this can be obtained by setting the imaginary part of M ( jω) to 0. So, I will just keep
those expressions, right. So, we can easily just do this, right. So, to this magnitude, what you

1 1
do? You just add . So, you can just see, right. So, if I add , this guy will become
jω jω

447
my imaginary part now because of this j , let’s you know quickly do this how does the
imaginary part look like. So, this imaginary part is just the real part of the previous guy

2
1−ω 2 T 1 T 2
(1+ω T 12)(1+ω T 22)
and I just do
1
jω
i.e.,
1
( 1−ω T 1 T 2
jω (1+ ω T 12 )(1+ω T 22) ) . This will also

be similar to writing. So, I just get rid of j from the denominator. I just have j in the
numerator with a minus sign. Again when does this go to 0 , this goes to 0 when

1
ω= , ok.
√T 1 T 2

(Refer Slide Time: 23:59)

Now, substitute this ω in the real part to get the value. I just leave that simple computation
2
to you, ok. So, based on these things, we can just keep on adding one more pole here jω .
I possibly add 0 here and then, I just keep on doing a bunch of things, but we will just
stop here, otherwise, we will have to spend a few years doing all those plots. So, let’s see
what do these plots mean, right. What do these plots mean and then, can I infer something
more just than saying what happens when ω=0 , what happens when ω=∞ , what is
the angle? what is the magnitude? and so on, ok.

So, Nyquist gives us a nice answer. So, before we do that, let’s just do some stuff. What do
we know? Now, we know this for a long time now that stability depends on the roots of the
characteristic equation that the roots should be or the poles should be on the left half

448
 s− plane is what the transfer function stability criterion tells me. Now, this Nyquist
stability criterion, it determines the stability of a closed-loop system from its open-loop
frequency characteristic and also its open-loop poles, we will all today, what we will do is
just to see what this statement means, ok.

So, again what does this do? It relates the open-loop frequency response. So, again all the
analysis, this is important because I just look at the open-loop transfer function GH and I
can tell you what the closed-loop behavior is, right. So, that is what Nyquist will tell us very
shortly, right. The Nyquist plot relates the open-loop transfer function, open-loop frequency
response to the number of poles and zeroes that lie in the right half s-plane of the
characteristic equation. Now, we will see how we arrive at this. Just consider a function
q (s) which is again in the usual terms you can say it is a transfer function which has m
number of zeroes, n number of poles and usually these n is greater than or equal to
m , ok.

So, what is S here? S is a complex variable real part: σ ; imaginary part: ω , right.
If this is a complex variable, the function q (s) is also a complex variable. Let me call that
u+ jv in the new axis. So, I start with the σ + jω axis and I go to σ + jv axis via this
map q , right. So, q takes s , it computes all these things and gives me q (s) . No,
not complicated, don’t get worried, right. So, what does this mean is that for every point s
in the s− plane .

(Refer Slide Time: 26:44)

449
So, if I take let me just see here, right. So, this is my s-plane. So, for every point here s ,
let me call this s 1 , there might be a corresponding point q (s 1) here, not there might be
there will be, right.

For all the points at which q (s) is analytic, where I can actually compute this, right. So,
we will not go into the definitions of analytic functions, but we just say that well I can
compute it almost at all points except when q is analytic, right. So, for s 1 , there will

exist a q (s 1) . Similarly, I take s 2 here, there will exist q (s 2) here and so on. Whatever
points here, here, here and at some corresponding points here, here and here, right. So, the
function q (s) , this maps the s− plane . So, this is my s− plane and this is via
q (s) . This function q maps the s− plane which equals σ + jω to the
q (s)−plane with the coordinates, I can call u and j v , right. This is what this
particular expression tells us, right. This is nothing surprising and nothing new, ok.

So, we can just say well you know we avoid points like s=β 1 , right. That time q (s)
may not exist. So, we avoid all those points or what we call as singular points, ok.

(Refer Slide Time: 28:21)

That is what the contour in the s− plane which does not pass through any singular point,
there corresponds, there exists a corresponding contour in the q (s)−plane . The contour is
like this. It is entire map which I take here, right. I just take this little closed-loop here and I
say something else happens here. Well, I just may have to make sure that I don’t hit a point,

450
where s=β 1 or s=β 2 or I don’t hit s at the poles and because at that point q (s)
is not defined, but for all other points, it is defined, ok.

We will not necessarily be interested in the shape of the contour, right. We will just say there
is a contour here which maps to a contour. On the other side, our aim here is to detect if there
are any poles of the closed-loop system in the right half-plane and when I detect this, I have
you know there is direct implication on stability. So, I do a trick where say well, I ask myself
a question. Does there exist a pole on the right half-plane? I do a graphical technique. If the
answer is yes. The system is unstable. If the answer is no. Well, system is stable. So, let’s just
ask few questions what does it mean when the q (s)−plane encircles the origin, right. So,
this is little important here, right. So, why I am not really interested in the shape? It could be
this one or it could just be something I am just interested in this origin. Does the contour on
the q (s)−plane contain the origin? Not necessarily here. Here it could be whatever I will
tell you in some other stuff related to this, ok.

Now, let start by considering the case when the s− plane contour encloses only one of the
zeroes of q (s) . So, what does that mean? So, q (s) has m zeroes and let say that
this guy encloses one zero. So, let me draw zero here and I call it as α 1 . So, all the other
poles and zeroes are outside the contour. Now, given this complex number, I can compute the
magnitude of q (s) as blah blah blah and I can also compute the angle, right the angles of
the zeros minus the angles of the poles.

(Refer Slide Time: 30:48)

451
So, let’s see here. So, this is the contour n the s− plane which encircles a zero, ok? So, let’s
see as the point s follows the path in the clockwise direction. Now, let see I am here and I
just take a clockwise direction. I come all the way and I just end up here now with respect to
this, I just say looking at this particular phasor, right. What does this do? This actually has to
rotate all the way and come back which is it does a net angle or it generates a net angle of
−2 π because the positive angle we measured as the anticlockwise. Okay now, what does
this do? This angle from the pole at β 1 , well it goes here, this phasor goes here, here, here,
here and again comes back here, right. So, it just goes this and just comes back this, right. It
just goes this and comes back this, whereas the other way it just starts here and does the
complete revolution like this. So, the pole just as this and this comes back over. So this, how
much angle does this generate? Well, the angle is 0° and similarly, this guy again just
goes, again and again, comes back to same position. It does not really take a revolution or a
rotation similarly with α 2 , so α1 generates or the phasor s−α 1 . So, if I just was to

mark this, this guy would be the phasor s−α 1 . The phasor s−α 1 , first you can write it
here. It generates an angle of −2 π while all other phasors generate 0 ° net angles, ok.

Carefully we will go to the next statement. What does this mean that q (s) phasor also
should undergo a net phase change of 2 π ? So, it is something like this, right. If I take a
q (s) here, I know that q maps the s− plane to the q (s)−plane . So, if s
undergoes a phase shift of 2 π , so this also should undergo the phase shift of 2π just
from the expression of q (s) and the tip of the q (s) phasor which is this guy, it
describes a contour about the origin. What happens when s=α 1 ? This is the origin. So,
this guy maps to the origin here, ok.

So, again I have here two zeroes: α 1 , α 2 ; two poles: β 1 , β 2 . I just know like this, pick

this guy α1 and I just generate a contour along with this in the clockwise direction. So, I
go here I see that this guy also generates a contour, right a close curve again in the clockwise
direction, right and then, I can just measure this phasor and also the angle that is just from the
alpha ones and beta ones and beta twos. Not only that if I am encircling a zero here, zero; I
will encircle the origin here, that is because when s=α 1 , my map what does q (s) do
when s=α 1 , it maps with the origin. It is a very simple way of understanding this. There is
actually much-complicated proof, some complicated, I am sorry some complex analysis
which we will not go into, but we will just learn what complex analysis is telling us, we just
believe what people from complex analysis told us, but we will just use it in our own way,

452
right. We will just be a little smart and translate in our own language and say well, this
information which they proved is correct and will actually help me in my stability analysis.

(Refer Slide Time: 34:53)

So, what happens when my contour encircles or encloses two zeroes. So, let me just draw
this, say this is zero here, there is a zero here and I just got on this. So, this is my s− plane
with σ and jω . So, what will happen here is that I will encircle the origin twice, once
like this and once like this, something like this again. The shape is not really important. What
we are interested is in the encirclements of the origin, right. So, you have a u and jv ,
ok?

Similarly, if there are n zeroes, I will encircle the origin n times. Again n this is clockwise
direction, this is also the clockwise direction. Lastly, if the contour does not enclose any pole
or zero as I said earlier, so if I just here I will just be here, right. I will not trouble everything
if I am just here, I will just be here. Nothing happens. It will never encircle the origin. That is
what the q map that tells me, ok?

453
(Refer Slide Time: 36:01)

Now, what is the encirclement of a pole? So, let’s say again I and do the same thing just draw
it for here. So, let’s say I have a pole here and this was my β 1 . When I had a pole here,
earlier this was β 2 , I had a zero somewhere here and a zero somewhere here, α1 , α2
, ok.

Now, again let’s say I am just having a contour, where I do this, right this guy and I start from
here. So, this phasor is my s−β 1 . So, what happens again let’s see all the angles, right. If I
am starting from here, this is my point s and say there is a phasor from here, there is a
phasor from here and a phasor from here. Now, how does the phasor α1 change? If I just
prove with this pen, the phasor α1 and the clockwise, it goes here, traverses all the way
and just is here, right again which means it generates an angle of 0 ° , right. So, again just
look at here, it goes here and again comes back here.

Similarly, β2 contributes an angle 0° , α2 also contributes to an angle 0 ° . What

is β 1 doing? β 1 is in this position. It just goes again all the way, this way, right. Again
the phasor s−β 1 , I will just show this here, it generates or it traverses an angle of −2 π
as s traverses along close path. Therefore, what will happen here is that q (s) will now
have an angle change of 2π say somewhere here and this is +2 π because s−β 1
occurs in the denominator angle always becomes minus when I am talking of the angle of a
pole and this means, again I just encounter or encircle 1 , I encircle the origin once in the
counter clockwise direction.

454
Now, again we don’t really see why exactly we are encircling the origin. It is the same
arguments as before, right. So, what we will just try to understand is the following that if in
the s− plane , I encircle zero in the clockwise direction, the corresponding contour in the
q ( s ) −plane encircles the origin once in the clockwise direction. If there are more than
one zeroes, say 5 zeroes, I will encircle 5 times. If I am encircling a pole in the clockwise
direction correspondingly in the q (s)−plane , I will encircle the origin once now in the
counter-clockwise direction because I am looking at the pole which contributes an angle of
negative. It goes to the denominator, ok.

(Refer Slide Time: 39:20)

So, therefore, I can just say something more now if there are P poles and Z zeroes of
q (s) encircled by the s− plane , so let’s say you know I just draw this and I say I am
not really interested in where these poles are. There they could be stable, they could be
unstable on the imaginary axis at the origin, whatever. I am not talking about stability at all. I
am just saying well say there is a pole here, a pole here, a zero here, a zero here and say one
more pole here, right and I say I just want to have some fun and I say I just enclose this all
the time. Now, I just you know say in the s− plane , I encircle all of them. So, this is
σ , jω in the s− plane , ok.

Then, the corresponding q (s)−plane contour must encircle the origin Z times in the
clockwise direction. In this case, it should be two times in the clockwise direction, P
times in the counterclockwise direction, number of poles; two zero two times in the

455
clockwise direction, three poles three times in the counterclockwise direction and the net the
encirclement is P−Z in the counterclockwise directions. That way if this number is say
well, I will always say this number will always be greater than 1 possibly or even if it is
less than 1 , it does not really matter because I am just its −1 times counterclockwise
means 1 time in the clockwise direction. So, the sign will take care of themselves, ok.

So, we now know about encirclements, the mapping from the s− plane to the
q (s)−plane . What happens if I encircle poles and zeroes in the s− plane ? How does
it reflect on the q (s)−plane ? That I know now what is the relation to stability, the
standard pole-zero form, I know that my open-loop transfer function is G ( s ) H ( s) has
again a set of zeroes, a set of poles, right and my characteristic equation q ( s ) =1+ K , again
this side of zeroes and this side of poles and I can write this as an equation of the form
s+z' till z n ' . This number will no longer be m , but will be n because this will

( s+ p1 )(s + p n)+ k ( s+ z1 ) + …(s+ z m )

go here say this is again their open-loop poles. So, this
( s+ p1 ) ( s+ p2 ) … … .( s+ pn )
will also be s power n and this will also be s power n . So, this will have n
zeroes and n poles, ok.

(Refer Slide Time: 42:14)

The zeroes of q (s) are the roots of the characteristic equation. So, this is a transfer
function, with a set of zeroes and a set of poles, q (s) going to 0 would give me that all

456
these (s + z 1 ')( s+ z 2 ') ……. ( s + z m ' ) should go to 0 and therefore, the zeroes of
q(s) nothing to do with the plant, nothing to do with GH now, we are just doing with
1+k , this one and this one or 1+GH , only the characteristic equation. And the poles of
q (s) are the same as the system poles, ok.

Now, when is the system stable? For the system to be stable, the roots of the characteristic
equation, what is the root of, what is the characteristic equation? 1+k or

( s + z 1 ) …(s + z m )
1+k =0 or this guy is 0, right. So, this guy is equal to 0. So, the roots of
( s+ p 1 ) …(s + pm )
this equation must lie in the left half of the s− plane . That is what we know from the
earlier lectures, ok.

Now, these roots of the characteristic equation turn out to be the zeroes of this q (s) . So, if
zeroes are confusing, you just leave it. You can just say the roots of the characteristic
equation, nothing else even if I write root in italics. You can also just not read this if the term
zero, why am I interested in zero is confusing. You don’t read that. Just say well the roots of
the characteristic equation, that is all whether it happens that these guys just are in the
numerator, right and then, the numerator terms you always call them zeroes, ok.

Then, the question arises what if the open-loop system has pole in the right half-plane, right
which means that the open-loop system is unstable say this p1 is at plus 1 does not
matter, right. This p1 even if it is in plus 1, I can or I may have all these z 1 till zm
to be on the left half. The expression for q (s) , therefore, such as that even if the open-loop
system is unstable. The close loop system maybe stable because these are the open-loop poles
and these are the closed-loop poles. We will see an example also. So, meaning to say that
even if the open-loop system is unstable, I can make it stable. This is not what Nyquist will
teach us, how to do that Nyquist will just tell us though is the close loop system stable or not.

Now, what we need to investigate are there any roots of q (s) in the right half s− plane
, right or any of these numerator guys z 1 ' . Till zn ' is there, any one sitting in the right
half-plane is z1' , z2 ' , zn ' is anyone sitting on the right half-plane. That is what we
will try to investigate and that is what leads us to define the Nyquist stability criterion.

457
(Refer Slide Time: 45:51)

So, now the encirclements will come into the picture. Now, the polar plots will come into the
picture and so on, ok.

So, what do I need to investigate? I need to investigate the presence of any right half 0. This
is again the right half zero of q (s) , right. The zeroes of q (s) are my close loop poles.
Just you write down that and I will just try to understand what it means and I am not really
talking of zeroes here. I am still talking of the closed-loop poles which happen to be the
zeroes of the denominator transfer function, 1+GH or 1+kGH which is q . Now,
choose a contour that completely encircles or encloses the right half s− plane , right and
this is called as the Nyquist contour. Again, it is directed clockwise and comprises of an
infinite line segment j ω , c 1 which is infinity and then, an arc of again infinite radius right,
ok.

What happens along c 1 ? Along c1 what is happening is s is just j ω . This is c1,

this line, s is all jω with ω varying from - ∞ to + ∞ . So, this one right. I am
starting from - ∞ all the way till + ∞ , the entire imaginary axis and along c 2 . What is
c2 ? s=R e jθ . Just the phasor here. Again with θ , what does θ do? Theta goes
from here till here or here till here, right? whatever whichever if you measure from here to

π −π
here, this is at θ varies from to , right. So, this means I’m actually I am
2 2
taking care of the entire right half-plane including the imaginary axis, ok.

458
(Refer Slide Time: 48:20)

Now, we will ask some questions. First is, why do we need to define a contour? such a
contour I just have to start from 0, go till infinity in you know + jω going to ∞ and
jθ
– jω going to ∞ . Not only that, I also have to travel Re with R going to ∞ .
Why do I need to do this? Well, such a contour encloses all the right half-plane, all the right
half poles and zeroes of my characteristic equation. Now, suppose this starts everything will
now start to make sense of what we did so far. Suppose there are z zeroes and p poles
of q (s) in the right half-plane, again we just write out z zeroes. So, zeroes of q (s) ,

where ( s+ z '1 )( s+ z '2 ) … ..(s+ z n ' ) and then, you have poles (s + p2 ) . These are all open-
'
loop poles. In other words, all these z 1 till zn ' are the closed-loop poles, ok.

So, as s moves along the Nyquist contour in the s− plane , this is a Nyquist contour
which we define as the entire right half-plane. Now, as s moves along the contour in the
s− plane , a close contour is traversed in the q (s)−plane which encircles the origin.
So, let’s say, sorry this is my Nyquist contour. So, it has zeroes here, poles here, few more
zeroes, few more poles. Whatever now if I map this contour, what I wanted to go from the
s− plane , so this is in the s− plane , the Nyquist contour and I want to see what
happens in the q (s)−plane . This q (s)−plane will encircle the origin, right. That is
what I learnt from these previous slides here, right these things the encirclement of the origin,
ok.

459
So, as s moves along the Nyquist contour in the s− plane , the closed contour is
traversed in the q (s)−plane which encloses the origin and equal to ( p−z ) times in
the counter clockwise direction. That is what we also observed somewhere here, the
encirclement of the origin is ( p−z ) times in the counterclockwise direction. Now, what
has this to do with stability? I am still talking of so many things. I talked of polar plots, I
talked of ω from 0 to ∞ , I talked of the encirclement of poles and zeroes and then,
correspondingly I said something happening in the clockwise, something happening in the
clockwise direction. Now, we will see this actually make sense, right? Now, how do we test
stability with these encirclements of origin?

Now, I will again ask myself a question when is this system stable? Well, for the system to be
stable, there should be no zeroes of q (s) , this guy in the right half s− plane because
this zeroes of q (s) are the roots of the characteristic equation and these are also the poles
of the closed-loop system. Now, this system for this system to be stable, z should be equal
to 0. There should be no zeroes on the right half-plane. Now, this is met if, sorry this is met if
the number of encirclements is p . It is not z ; it is n . z is anyways 0. So, if the
number of encirclements in the counterclockwise direction and this is n the number of
encirclements in the counterclockwise direction is equal to the number of poles of q (s) in
the right half-plane because I am just my contour is just right half-plane.

So, there might be p1 sitting here, p1 , p2 , p3 , p4 and so on. See there are four of these
guys sitting here. Then, this n should be equal to 4 if the system or for the system should
be stable, there should be no zeroes. All these z 1 till z n should be on the left half-plane.
So, none of them should sit on the right hand side. So, z should be equal to 0. Now, if
z is equal to 0 and I measure the number of counterclockwise encirclements, what should
that match? That should match the number of poles from p1 till pn on the right half-
plane, ok.

460
(Refer Slide Time: 53:24)

This means that for the closed-loop system to be stable, the number of counterclockwise
encirclements of the origin of the q (s)−plane should equal the number of right half
s− plane . That is what I just explained right through these things, ok.

These are open-loop unstable poles. So, I don’t really worry now is the open-loop is some
unstable, open-loop is unstable. I cannot do anything. So, let’s close the books and go home.
No, there is still hope. If N=P , then my closed-loop system is stable, right and if open-
loop system is a nice guy, if it is stable, then close loop system is stable if and only if
N=P=0 . So, if all these guys are well behaved, then there is no P which is on the
right half-plane which means this number P here, this number P here will go to 0. So,
N=P=0 .

So, my stability now translates to the number of times I encircle the origin in the
counterclockwise direction, right. For example, if all these poles are stable and say N=1
or −1 to say N=−1 , then my system is unstable because well there are no poles or all
P’s are equal to 0 and whenever z> 1 or ≥1 , then I am unstable, right. I always want
z to be 0. This is a necessary condition. Now, N=P=0 which means no encirclements
of the origin. So, in this entire region here, I don’t encircle zeroes which I should never
encircle. I should never encounter zeroes and I also never encounter poles here. So, maybe
my contour here when there are no poles and zeroes might just be here and therefore, I am
happy, ok.

461
So, well you might say you know sir this is actually too complicated because I start with an
expression which already looks complicated. There are m poles, sorry m zeroes n
poles. I have to compute all this, count these numbers that might be cumbersome. However,
this is actually given to me, right. The plant is given to me, the feedback loop H is also
given to me and what I promised you or what I told you is that just by looking at the open-
loop transfer function G and H , I can tell you the closed-loop behavior, but no, well I
actually made you compute all these z1' till zn ' to see where they are and then, I said
something to do with you know z− p , p−z encirclements and so on, but we will just
simplify things a little bit. This is just to again get to the point which I promised you that just
by the open-loop behavior, I can tell the close loop system stability, ok. So, this is my close
loop characteristic equation 1+GH and I can just rewrite this as the q (s)−1 .

So, this is just my q (s) I am just rewriting this as q (s)−1 . Now, this is a complex
number just say I am looking at 0+ j0 and I am just doing −1 . So, what happens is
that the real part shifts with −1 ; ie ., 0+ j 0−1=−1+ j0 . Similarly, say I have 1+ j 2 .
So, I am just doing a −1 . So, what happens is, I will just you know have
1−1+ j2=0+ j 2 . Similarly, if I have just any complex number right, σ + jω , I do a
−1 . I will just get a ( σ −1+ jω ), right. The imaginary thing always remains the same
and the real part is shifted by 1 , right. If this is the σ , it just becomes σ −1 , ok.

If the contour of GH , this is the same as the contour of 1+GH just that I am just
shifting the real part to −1+ j0 . I will show you a plot of this and therefore, if I say well
have I encircled the origin of 1+GH , this is actually equivalent to saying have I encircled
or asking a similar question have I encircled the point −1+ j0 of G H . So, I am
looking, I am comparing encirclements of region of 1+GH to this is −1+ j0 of
GH . So, I will just show you a plot, right. So, this is 1+GH , ok.

462
(Refer Slide Time: 58:48)

Now, I just shift it by −1+ j0 . This is 1+GH . I just do −1 . So, this just shifts here
which means the origin which I am interested in the encirclement of the origin, now I just do
a −1 all of a sudden. So, therefore, this origin I was interested in the encirclement of the
origin of 1+GH or the contour of 1+GH , now I am just doing 1+GH −1 . So, the
origin also shifts here, sorry I am really sorry. So, the plot shifts by −1 in magnitude. So,
you will see all this distance would be −1 and then, from here I shift here, ok.

So, the encirclement of origin of 1+GH is equal to the encirclement of the point
−1+ j0 of GH . Now, this is easier to check because I know G , I know H ,I
don’t need to compute z1 ' all the way till z n ' . So, now, all the results of n= p−z of
the origin will transfer to again N=P−Z not of the origin, but now of the point
−1+ j0 . This will be my point of interest. So, let’s read the statement. If the contour
GH of the open-loop transfer function corresponding to the Nyquist contour, if it encircles
the point −1+ j0 in the counterclockwise directions as many number of times as the right
half poles of GH , the close loop remains stable, everything remains same. It is just that
the encirclement of origin now just translates to the encirclement of −1+ j0 , everything
you just go to the previous slides, replace the origin with −1+ j0 and you will have this
thing, ok.

So, if the open-loop system is stable, again I am just rewriting all those things which I said.
Then, the closed-loop system is stable if the contour does not encircle the point −1+ j0 .

463
Earlier I said if the open-loop was stable, then the closed-loop system or the closed-loop
contour of 1 plus j g h should not encounter, they should not encircle the origin. Now, you say
well if the open-loop system is stable, the contour of g h should not encircle the point
−1+ j0 . Just origin control c and then, this control v , right or control f and control
v , find and replace. So, now, this is our point of interest −1+ j0 .

To summarize all these things, we will just, so this is the statement of the Nyquist stability
criteria, right and I am just interested in the point −1+ j0 , ok.

(Refer Slide Time: 01:01:54)

Let’s summarize with the help of example, right. Consider a feedback system whose open-
loop transfer function is this guy. Now, this is similar of a polar plot which we had in one of
the earlier slides. So, let me slowly go there and this one, right. So, just what I am interested
little more in doing is there I plotted as ω was from 0 to ∞ . Now, I plot a little more,
right. I am also interested in −∞ because if I just plot ω from 0 to ∞ , I am just
looking at this one, this contour, but I am interested in the entire right half-plane. So, I just
the same thing, right. I just look at what happens when ω is - ∞ , nothing changes. I just
have to compute again the angle and the phase, the angle and the magnitude or the magnitude
and the phase.

So, for this example, well the infinite semi-circular arc, it gets mapped to the origin the plot
of GH . So, this is my origin, somewhere here is the point −1 . It is not even in the
graph. So, I am nowhere near to −1+ j0 . So, what do I conclude is that the plot of GH

464
again there is some distortion here. I will just take care of this. So, this is the plot of GH ,
right and then, it does not encircle −1+ j0 for any positive values of kt 1 and t2 ,
sorry for the distortion I will take care of the slide and therefore, N must be P−Z
open-loop system is stable. Therefore, P is 0, Z should be 0 for stability. Therefore,
N should be 0 that of the point −1+ j0 , but if you look here, I am really far away from
−1+ j0 . Therefore, this system is stable.

(Refer Slide Time: 63:49)

S+ 2
Now, look at this example . This open-loop system is unstable. So, this is
( S+1)(S−1)
some interesting case. Now, I just draw the Nyquist plot, I just look at what happens at
ω=0 , ω=−∞ , ω=+∞ and I just plot this, right. So, I start from ω=0 and from
here and then, I reach here. So, I start from −∞ , I go all the way here and I go all the way
here and I just plot this and I say where am I. So, this is my origin here. I am at a distance of
2 and this 1 . So, here my P is 1, Z should as always be 0. Therefore, N
which is P−Z should be 1 in the counterclockwise direction, ok.

Let’s just check this. So, if I just do move in the counterclockwise direction, am I encircling
the point −1+ j0 . So, this is the point −1+ j0 . I am encircling it one time and
therefore, the close loop system is stable. I start with an open-loop unstable system and
Nyquist tells me that it is actually closed-loop stable, right with 1+GH as the feedback
and I am closed-loop stable, right. It does not tell me still how can I make a close loop system
stable, can I do it always. None of those questions are answered. Just that well I can make or I

465
can start from open-loop unstable system. Nyquist will tell me well is it close loop, stable or
not, right just via these plots. So, the polar plots started from 0 to ∞ , here I am even going
from 0 to −∞ . That is the only little change, but nothing changes in the analysis, ok.

(Refer Slide Time: 65:46)

So, what we learnt so far is polar plots and we just started to define for ourselves the notion
of stability in the frequency domain leading to the Nyquist stability criterion, did a couple of
examples and next, we will do some special cases of the Nyquist criterion. What if I hit a
singular point, what if there is a pole on the imaginary axis things like that and then, we will
also define. So, now, everything was to do with −1+ j0 , how far I am from
−1+ j0 . So, we will define relative stability in the frequency domain and introduce the
concepts of gain margin and phase margin, ok.

Thank you.

466
 Control Engineering
Dr. Ramkrishna Pasumarthy
Department of Electrical Engineering
Indian Institute of Technology, Madras

Module - 06
Lecture - 03
Relative Stability

Welcome, to this new lecture again we will continue today with the Nyquist plots. So, what
we saw yesterday was or in the previous lecture was how to analyze stability in the frequency
domain. So, we answered several questions based on the Nyquist plot, the first question was
what Nyquist would tell me is how many zeroes are or how many poles of the characteristic
equation lie on the right half-plane. So, that would be determined by the number of
encirclements of the point −1+ j0 and we also see saw how we shift from the origin to
−1+ j0 and how −1+ j0 analyzing this point will just help us getting the closed-loop
behavior by just looking at the open-loop transfer function right and so what we will do today
is to see or just handle some special cases right.

So, what we said when we were defining the conformal maps from the s plane to the
qs plane is that this map should not pass through any single point or through any poles,
but that is well it is mostly possible, but sometimes you could see that there might be a pole
on the imaginary axis and lets for simplicity say I just start with the pole at the origin ok.

(Refer Slide Time: 01:34)

467
So, the question we will ask is how do we handle these conformal maps or the Nyquist
contour when the open-loop poles are on the jω axis.

So, if we use the Nyquist criterion directly, the s plane contour would pass through a
singular point of 1+G(s ) H (s) this would just blow up. So, this could be similar in a way
to see you know when we were doing the Routh Hurwitz table of when you get an ε ,
when you get a 0 right, in one of your columns one of your entries becomes zero. So,
what do you do, you just substitute with a very small number called ε , you assume it to be
greater than 0 and take the limit as ε goes to 0 and then do the rest of the
computations. So, we will do very very similar here it is a little tricky here than what we did
last time. So, let us see how we will do this. So, we need to make modifications, so as to
bypass these points in the jω axis and therefore, what we will do is if there is a pole here I
just take a detour here, and this detour I will just say well I will just take a little semi-circle of
radius ε .

jθ
So, this is how the semi-circle looks like εe and take the limit as ε→0 right, so if I
just write down a little more formally. So, what we would have is take the semi-circle indent
around say there is a pole at the origin. So, if I say this guy there is a pole at the origin, I just
jθ
draw a little semi-circle and S=lim ε e . So, as theta goes from −90 to 90 and then
ε→0

we’ll just do this.

(Refer Slide Time: 04:20)

468
And then put this back here and see how the plot goes like. So, to just get a little
understanding we will start with the help of an example right we have feedback system I have
a pole at the origin and another pole. So, sinusoidal transfer function we just substitute s
with jω and we know already why we know why we do this right, why we just substitute
s with jω for the sinusoidal transfer function ok.

So, let us start by dealing with the pole at the origin right the standard thing that I learnt in the
last class will not work because there is a singularity when I substitute S=0 . So, I take

this little thing limit. So, this guy limit S or lim ¿ this guy is the value of S . Now
ε→0

jθ
we will see how this value of S defined by this little semicircle given by this thing εe
, ε →0 and θ going from −90 to 90 . So, how this little region will map to the
qs plane, I take

lim ¿ K lim ¿ K − jθ
jθ jθ
=¿ e
ε→0 ε e (Tε e +1) ε→0 ε

K
Okay so, this guy →∞ , as ε→0 and what does θ do? θ varies now from or
ε

+π −π
this is −θ right, −θ varies from +90 to −90 or to . So, I just
2 2
first want to see how this little region here or I will draw it a little better, this region; I still
jθ
can’t get the line straight this little region. So, this is εe . So, how will this map into the
qs plane right and s , σ , jω ; u , jv ; this is how we defined it last time
right? So, what happens? how does this guy map here? So, what I have is as ε → 0 , the
magnitude goes to ∞ and the angle goes from +90 to −90 . So, I will just have arc
like this right something like this is of radius infinity and it will travel in this direction here,
here, here and here. So, this is just to do with the pole at the origin and the rest of the Nyquist
plot follows the same procedure as before ok.

469
(Refer Slide Time: 07:54)

So, if we remember the polar plot of this guy it just looks something like this right. So, the

1
polar plot of was something like this, it will asymptote here this guy is T ,
jω( jωT + 1)
ω=0 , I start from ∞ . So, I start from somewhere here, I go here and all the way, I
come back to ω=+∞ right and if I want to do the Nyquist you know, I just take the even
the other frequency right something like this right. So, this is how the plot will look like, this
is what more MATLAB will tell you right when you draw these plots and this we derived last
+¿¿
time, so how do I really look at it right. So, I start at ω=0 or let’s call it 0 right.

+¿
When ω=+∞ I will be here, when ω=−∞ I will be here, ω=0¿ I am going here,

−¿
ω=0¿ for negative frequency, I will just be here right.

+¿
Now, what happens at omega equal to 0 right. At omega equal to 0, no sorry ω=0¿ , I’ll go
−¿
to ω=0¿ and then I have to really complete this, this thing right; where, so this thing will
be the region of infinite radius. So, just, this region is what I have to map now right this is
what when ε → 0 this little region right this also I have to map it here and therefore,

470
 −¿
So, the plot is complete by starting from here I started ω=0¿ all the way come to

+¿
ω=0¿ and then go to ω=+∞ . Minus infinity till 0 plus, 0 sorry I will say it again

−¿¿ +¿¿
ω=−∞ , 0 , 0 , and 0 and ω=+∞ . So, if we draw on MATLAB it will

give you this line, these 2 lines like this one and this is the polar plots it would not really give
you this line because it would not be able to plot, but this is like this is just the infinite radius
from here drawing this very approximately, but you know we could look at it this way. Now
the question, is the system stable? Well, what was stability? I just looked at encirclements of
the −1+ j0 point right, where is the −1+ j0 it is somewhere here right. This is the
point of interest. Now, what was the number of encirclements was N=P−Z , P=0 ,
Z =0 therefore, the number of encirclements should be 0 right. So, since the number
of encirclements is 0 we can say that the system is stable ok.

(Refer Slide Time: 11:15)

Now, what if there are 2 of these poles at the origin. So, just do again the same kind of

K
analysis. So, I am just dealing with this term alright. So, I have lim 2 j2θ . So,
ε→0 ε e (ε e jθ +1)

K −j 2 θ
this I am just looking this like this. So, the K in the numerator; lim e right. So,
ε→0 ε2
of course, the magnitude goes again goes to ∞ , and as θ varies from −90 to

471
 +90 in the s plane, the qs plane it varies from angle of 180 ° → 0 ° →−180 ° ;
this is what the, then the plot would actually look like.

So, what MATLAB will give you is this, this kind of blue line here right. So, I just go here
this I terminate here at omega equal to minus infinity and this; sorry I terminate here at
+¿
ω=+∞ , I start from here at ω=−∞ and up here at ω=0¿ . So, what happens, later
on, is again and just looking at this of infinite radius going from angle −180 ° to
−¿¿
+180 ° . So, the plot is completed in this way −∞ , 0 ; I go here, go here and do all
the way, okay. So, there is some kind of an encirclement it is little tricky, but it is you can just
understand it how to complete these lines. So, what MATLAB would show me is these 2
+¿ −¿
things right, starting from ω=0¿ , ω=∞ , ω=−∞ and ω=0¿ , and these guys
have to meet somewhere right and they will meet again via this expression. So, it will go all
the way till infinity with this +180 ° to −180 ° and then come back here ok.

So, what about the stability here? Right, so this is how the Nyquist plot looks like. Now can
we infer stability of closed-loop system? So, again we look at the number of encirclement,
right so, what I am interested here is in the again the point −1+ j0 and if I see here that
the closed-loop encircles the point −1+ j0 this is our point of interest. so, if I look at it
where it actually encircles it two times, so the total number of encirclements is two. Now,
how many open-loop poles are in the right half-plane, well that answer is P=0 and if
the system were to be stable, what we need is N=0 . Now, N≠0 , N=2 therefore
the closed-loop system is unstable.

So, we could also look at it, if I look at it from the root locus, Right. So, If I just plot a quick
root locus for this, so I have two poles here and one pole somewhere here so well on this side
for all values of K >0 my root locus would look something like this and this we are
plotting, well you just say some non-zero value of K my poles always will be here and
these are the two unstable poles which is also captured by Nyquist plot and therefore the
closed-loop system here is unstable. Ok.

So, this is all about special cases of Nyquist we will not really go into plotting complex plots,
that’s not the aim of the course of you know given a you know something of s4 and s6
and so on; that is not important right. So, we just try to understand what this guy tries to tell

472
us. So, there is something also called as relative stability right. So, most of the times I will
say, well I am interested in stability; the system is just about stable.

So, then the next question I will ask is how stable it is right. If it is stable will it remain there
for all time or will a certain change in parameter push it to the verge of instability, certain
change in gain will push it to the verge of instability? So, I want a reasonably good, good
measure of how stable my system is ok.

(Refer Slide Time: 15:44)

So, what we know after we say, well the first requirement is that the system should be stable,
and next is to ensure that it is adequately relatively stable, now how does one measure
relative stability sorry. So, look at these 2 things this brownish kind of color, I don’t know
what it is called. So, this is a set of poles here open-loop poles and these are also a set of
poles, right. Now I go, so this is my s plane- σ , jω these are the poles or also
called the dominant poles, I will tell you later why I use the word dominant poles, but let’s
say the system just has 2 poles and I plot the Nyquist for this right. So, this guy blue guy goes
here and this brownish guy goes here.

Now, here I look at stability in terms of the distance from the imaginary axis right. So, if I am
closer here get closer closer until here, and there could be a point where I just tip-off to the
right side it is here and here. Even if am just very slightly to the right I am still you know
unstable even if I lose by one run in a cricket match or 100 runs, I still lose right. So, it is the
same thing here. So, how does the Nyquist thing now look like here I just cannot directly

473
measure relative stability, I can say well this is further from the origin. So, it will take you
know it will still be safe if there are some change in parameters or things like that it may just
come slightly to the right or maybe even move to the left and so on.

But I can say that this looks little more hopeful the blue line than the brownish one, this
paired poles. Now let’s come back here right. So, what is the point of interest right? So, as
long as P=0 say the open-loop system is stable, I just want Z =0 right, and if Z =0
Nyquist tells me that the system is stable, now who decides this. So, I just. So, this is my
guard right who says well you cross this line you trespass and you become unstable. So, the
idea would be key well I just stay as far as possible from this line this number −1+ j0
right.

So, look at the blue line here it starts here and goes here the brownish line is still closer here
right this is this guy inside is the −1+ j0 thing right and I want to stay away from here,
say something changes and then you know if I could just redraw plot over here like this. So,
this is my point −1+ j0 and if I do, I know something like this then I am gone right?
Because I am actually encircling the point −1+ j0 I didn’t really draw a smart picture
here, but something like this. So, I want to you could see this one if something keeps
changing you may just stick to over to the to up, and then the what happens when I am here
that is poles on the imaginary axis I am just here. So, the Nyquist here I will just draw it here,
it will just be on the horizontal line this and this depending on the ω increasing or
decreasing right ok.

So, based on this can I quantify something as a measure of relative stability. Again the idea is
to stay as far as possible from the point −1+ j0 or further to the left, of course, you
cannot really say that I will have poles at plus-minus 1 million plus-minus j whatever
something like say j 10,000 this is also not practically feasible right. So, we have to make
some reasonable assumptions there or so get some reasonable estimate of things ok.

474
(Refer Slide Time: 19:52)

Now, what is then the beauty of the Nyquist plot? So, Nyquist plot indicates not only whether
the system is stable, but also the degree of stability.

So, we will also see how it will give some information if the system stability can be improved
right and then since the deciding factor is this guy −1+ j0, and this entire stability
criterion is depending on the encirclement of the point −1+ j0 . So, what does the plot tell
us, either the polar or the Nyquist that as the plot gets closer and closer to −1+ j0 the
system is on the verge of instability, okay. Closer it is to −1+ j0 larger is also the
maximum overshoot in the step response and so, is the settling time, closer I am to the
−1+ j0 which means closer I am to the origin here which means the damping is very
small and things like that now ok.

475
(Refer Slide Time: 21:02)

Now, okay so let’s do it a little graphically, so let’s say this is a I’m not really, I am not
interested really in you know in in the Nyquist in the exact transfer function, but say
something looks like this right this is the Nyquist plot this guy or even the polar plot and this
is the point −1+ j0 and I just draw a circle around here and say this is the distance from
here till here is −1 , the distance from here till here is a . So, what is this distance? This
distance is when the imaginary part goes to 0 ? So, let me call this
Img [ G ( jω ) H ( jω)¿ω=ω ] =0
2
or in other words, you can even say that the angle would be
−180 ° . So, this is the thing here right at ω2 ok.

So, what could happen that if this a is getting closer and closer here, then I can you know
I am on the verge of instability, say I have something like this here and say I have something
changes; I increase the gain and I go here. So, once I go here I am actually encircling this
point right. So, I will just do some detour like this and I can say that the system is going to
the verge of instability here is this thing. So, these are not important. So, there is something
which this a can tell us the distance or how far this guy is from this point, not only that
now look at this circle, and say well it intersects this guy here right at frequency ω1 the

|G ( jω ) H ( jω)|¿ ω=ω =1
1
. Now keeping this, this constant and if I keep on moving along
this line which means I am adding some kind of angle to the system right I am just moving
from here till here then if I add say some angle of φ in this direction, then I hit the point
−1+ j0 . So, Once I hit the point well then I am on the verge of instability.

476
So, there are 2 things the distance how far is this a from here and second also how far is
this angle from being −180 ° right. So, these 2 will quantify this a and φ will give
us some measures of relative stability. So, far I just said just be further from the origin in the
s plane, when I would talk in terms of poles, that the poles they are on the imaginary axis
the system is just marginally stable slightly to the left well it is stable, but it is on the verge of
instability right it is not very. So, now, I can actually quantify these things ok.

(Refer Slide Time: 24:28)

So, let’s say these things a little more formally now. So, the Nyquist crosses the imaginary
axis at a frequency ω=ω 2 with an intercept of a right, here ω2 and this intercept is
called a . And then the unit circle intersects the locus of G( jω) H ( jω) with the
frequency ω=ω 1 here right and I just draw this unit circle because this is the point of
my interest right −1+ j0 . So, the phasor of this makes an angle of φ in the negative
real axis, again the usual measurement is we measure positively in the counter-clockwise
direction. So, this angle φ from here till here and then well this is now kind of known to
us right and we know why this is true, that as locus approaches −1+ j0 , the relative
stability reduces and this happens in 2 cases right.

So, this happens when the value of a this should be a little bit highlighted; because it is the
when the value of a approaches unity or φ approaches 0 or you know. So, for this
angle, if I measure from here the ∠ G ( jω )=−180° . So, a approaching 1 would be

477
bad news for me, in the similar way φ approaching 0 will also be bad news for me. So,
therefore, we use these guys a and φ as a measure of relative stability.

(Refer Slide Time: 26:07)

So, we will use these 2 quantities a and φ to define the measures, now we will name
each of them right. So, this guy related to a , we will call as the gain margin and this guy
related to φ we will call as the phase margin the names will already tell us what they
mean.

(Refer Slide Time: 26:36)

478
But let’s understand that in little more formally in little more details. Gain margin what is this
guy? This is the factor by which the system gain can be increased to drive it to the verge of
instability. So, slowly we will try to understand. So, at ω=ω 2 this guy omega equal to,
what is the phase angle? The phase angle is −180 ° and the |GH|=a , now I want to
say well. So, say GH now I will not use this ω2 . So, |GH|=a , I can add some gain

right, I can have some gain. So, it will be K 1|GH|=K 1 a . I say well at what point does it

1
reach 1 ? how much can I increase the gain factor well that will be K 1= right then,
a
|GH|=1 right. So, which means that if the gain of the system is increased by a factor of

1
, this is K 1 I am just multiplying with gain right. So, don’t confuse with addition here
a
right.

So, somewhere okay, what is the gain margin? They’ll say oh this is 1 , oh how far I am, I
am just 1−a , you are not 1−a because the gain comes as a multiplicative factor right,
so here this guy, that if I add more gain it will show up here. So, this multiplied by the actual
magnitude a when does that goes to 1 ? That goes to 1 when this additional guy

1 1
here K 1= . So, if the gain of the system is increased by a factor of then the
a a
magnitude of G1 becomes a this is the original magnitude, multiplied by the new

1
magnitude right, K1 the additional gain which we put and this is 1 ; this means
a
that the plot of GH okay now again I am just you know not being a little lazy not writing
the entire thing of G( jω)H ( jω) , but that should be obvious. Which means that the plot of
GH will pass through this point −1+ j0 , driving the system to the verge of instability
ok.

So, therefore, now how do we define this? what is this gain margin in terms of a number

1
right? So, this gain margin may be defined as the reciprocal right, reciprocal of what?
a
reciprocal of the gain at the frequency at which the phase angle becomes −180 ° right. So,
that is this one. So, this frequency we defined as ω2 right, at ω2 :

∠ G ( j ω2 ) H ( j ω2 ) =−180 ° right. So, at this ω2 , I calculate the gain. What is the gain at

479
this ω2 ? This, the gain at this ω2 where the angle is −180 ° is a . Now, the gain
margin is defined as the reciprocal of this a , what is this a ? The gain at this frequency
which is ω2 at which the phase angle becomes −180 ° and the frequency at which this
happens is called the phase crossover frequency. ω2 is the phase crossover frequency, the

1
gain margin here is ok.
a

So, by what factor should the gain be increased such that the system goes to the verge of

1
instability? that is , a again is defined as the value of |GH|¿ ω . what is ω2 ?
a 2

ω2 is the frequency at which the angle becomes −180 ° , similarly the phase margin ok.

(Refer Slide Time: 30:31)

So, we could just you know say oh now I understand gain margin. So, phase margin should
still be the same right some similar definition, you will say oh additional amount of phase-lag
needed to bring this system to the verge of instability, well this is vaguely true but we need to
be little more careful in formulating the statement or getting the exact value. So, we ‘ll just
see what are those steps. So, first is well we are interested let’s revisit the picture again for a
while.

So, what are we interested again is here right this intersection here, this φ this guy I call
this ω1 and what is this ω1 ? The point at which you know the magnitude becomes one

480
here this guy. So, the frequency at which |G ( jω ) H ( jω )|=1 is called the gain crossover
frequency this is ω1 right, the magnitude is 1 how did you find this? Just by the
intersection of the plot with the unit circle. So, this is the frequency at which the plot
intersects the unit circle, of course, this unit circle is centered at the origin for obvious
reasons because we are just interested in −1+ j0 , I just want to pull it along that angle to
that point −1+ j0 and at this frequency what is the phase angle? At this frequency, the
phase angle if I just measure is 180 ° or −180 ° , either wise 180 ° and this φ ok.

Now, if an additional phase lag of φ is introduced in the system, right if I just push it, this
guy little up with an angle φ the phase becomes −180 ° and while I am doing this
along that unit circle the magnitude remains unity right, the magnitude remains one and once
I reach that point where the angle is −180 ° , what will be GH then? The GH or the
plot of GH will pass through −1+ j0 , once I have added that extra angle φ and just
see this one right, I am just pulling this ω1 here along the circle right. I am adding some
1° , 2° and so on, I add 5 ° I end up here right, what is the magnitude here? –1 .
What is angle? 180 ° or −180 ° right. So, I am just adding this extra amount of angle
φ . So, now, this additional phase lag φ is called the phase margin now where do, how
do I compute this φ well.

(Refer Slide Time: 33:30)

So, now formal definition is the phase margin is defined as the additional phase lag at the
gain crossover frequency and this is the frequency ω1 , where the magnitude of

481
 |GH|=1 . So, what we started off saying that well I need to add some phase now where do
I add this phase? This phase additional phase I add at the gain crossover frequency again the
standard terms of measurement will be in the counterclockwise direction from the negative
axis and so on. For stable systems phase margin is always positive that is maybe even true for
the gain margin. So, if I just want to formally write down what is φ from that which you
think I had earlier here this guy φ is the angle of ∠G ( jω ) H ( jω ) ¿ω=ω +180 ° ok.
1

(Refer Slide Time: 35:15)

Now, let’s see how this system looks like. So, now, I have a system of the form well, I have a
pole at the origin and then jω T 1 +1 , and jω T 2 +1 . So, I can write this as, or I split this
into the real and the imaginary part, okay I’ll skip this computation I will just write down the

final value
−K ( T 1+T 2 )− jK ( ω1 ) (1−ω T T )
2
1 2
okay, just check this computation for
2 2 2 4 2 2
1+ω ( T +T ) +ω T T
1 2 1 2

yourself also. So, if I look at again that plot which I had shown and right and if I say I am
interested in finding the gain margin, so at this ω2 which I defined earlier right. So, the
imaginary part should go to 0 ; imaginary part goes to 0 when this term goes to 0, this

1
term goes to 0 , when ω=¿ ok.
√T 1 T 2

482
Now, at this, so this is what? this is ω2 right, at this frequency
|GH|¿ 1 =¿
ω=
√T 1 T 2

−K T 1 T 2
, If I just put the magnitudes this will be the plus. So, this will go away. So,
T 1+T 2
now, for stability what should what do I want if this is the point −1 . So, this magnitude

K T 1T 2
which is now this 1 right ¿ 1 . If I put less than or equal to then there is
T 1+T 2
some kind of marginal stability and therefore, this system is stable as long as the gain

T 1 +T 2
K <¿ . See it’s quite straightforward.
T 1T 2

So, I start. So, my thing is well I want to compute what the values of gain K is the closed-
loop system stable well. So, I just see use this criterion right. So, this how far is this guy from
the point −1+ j0 how do I compute this? I just put the imaginary part of this guy to 0
and I compute the frequency, right this frequency is ω2 or what we call as the phase
crossover frequency, what is the magnitude at this phase crossover frequency? It is

K T 1T 2 T 1 +T 2
and the gain margin is this one right. So, and what we have defined
T 1+T 2 T 1T 2

1 K T 1T 2
earlier? This was and what is a here? a is . So, we just get the inverse
a T 1+T 2
of that.

483
(Refer Slide Time: 38:18)

Okay just to see some plots here. So, I start. So, this is the same transfer function with T1
and T1 being equal to 1 and I keep on increasing the gain K right. So, this is at
some point here, I reach the point −1 , and this point I can now exactly compute right. So,

T 1 +T 2
K=¿ . So, at the gain K=2 . I just reach this point these are all K for
T 1T 2
1 , 1.5 , 1.75 , and so on. So, as gain K increases my root loc sorry my Nyquist
shifts to this side and if it is shifting to this side I am reaching closer and closer to −1 to
this. So, this was my ω2 there, ω2 , ω2 and now I am here. So, this is we will just
show you how the plot changes with varying gain how to quantify that well that is this one
right this guy right.

K T 1T 2
So, what was this magnitude here , what happens when K=2 , when
T 1+T 2

2∗1
T 1 ¿ T 2 =1 this guy becomes =1 . So, that is what is exactly happening here at
2
K=2 , I can compute analytically and also check by the plots ok.

484
(Refer Slide Time: 40:28)

So, this thing we had check, seen last time also right. So, I had a system with an unstable
open-loop pole at s=1 , I had a zero and this one. So, well the system has an open-loop
pole at s=1 and the plot. So, this is the plot −1+ j0 , and I see that the plot encircles
−1+ j0 once in the counter-clockwise direction right and P=1 . So, what should N
be? N=P . P=1 and N=1 therefore; the closed-loop system is stable ok.

Now, what is the gain margin? Gain margin is we had said well the reciprocal of a , how
did we define a ? a was defined like this. So, this was a this was the point

1
−1+ j0 , . Now which means I am looking at the intersection with the real axis now
a
see where this intersects with real axis. So, it intersects here; what is the magnitude? well, it
is 2 , now is the gain.

485
(Refer Slide Time: 42:02)

1
So, is the gain margin , well we will see what that means, right? So, if I start with gain
2
of K=1 , sorry I am somewhere here right this is the original plot; this plot right s +2 ,
s +1 , s−1 with K=1 . Now oh sorry this for K=1 I am fine right. So, this is
K=1 , N=1, or N=P where P=1 system is stable.

Because I am encircling the point −1+ j0 and I come down, come down, come down here
I say well K=0.5 , K=0.5 well, I am just here, just at the point −1+ j0 .
K=0.4 what happens? The K=0.4 what is the number of encirclements of −1+ j0
if I take this direction, well there is no encirclement of −1+ j0 right this one. So,
N=0 therefore, system is unstable; Okay, what is happening here? Now let’s compare the
root locus right, the root locus will have a pole at 1 another pole at.

So, I have this is another pole at −1 , I have a zero at 2 . So, one pole goes to the origin
other goes. So, this is starting from K=0 , I reach some point here right say K=0.5
and when am I stable. So, under what conditions is conditions on K is this system stable
usually let’s then do something else right, say I have a system with a pole at −1 , pole at
−2 , pole at −3 and there are no zeroes and ok, I will just do all the rules of the root
locus. So, one guy will go to infinity this way, these guys will meet here right. So, they do
this way right. So, there is some value of gain K say some number say K ’ right this is
K=0 right. So, what are the values of gain K for which the system is stable, I will say

486
for gain 0≤K< K' right if this is the value of the gain K when it intersects
the imaginary axis ok.

What can I say over here? Well, the system is stable, what happens for K=0 ? K=0 I
am unstable, K=0.1 I am still unstable, until I reach K=0.5 here right. Where I am
just at the point −1+ j0 here. So, this system is stable for K >0.5 correct. So, I start
with an open-loop unstable system I have to reduce. So, the system is unstable until I have a
certain gain. So, therefore, what we call these systems are as a system with a negative gain
margin right. So, the earlier problems were formulated in such a way that how much can I
increase the gain K , before it becomes unstable now it’s the reverse. I am starting
K=0 is already unstable, K=0.5 is already unstable.

But there comes a threshold where after a certain limit I pull the poles back to the stable
region and unstable for all other gains. So, that is the concept of a negative gain margin that
first I have to make a system stable starting from an unstable open-loop configuration, I
increase the gain K and I push it to the stable region. Now, why is it negative we will
quantify that a little later while we do the bode plot analysis in terms of decibels. So, we will
stop here with the Nyquist thing, again getting or you know really spending time to get
beautiful plots by just writing the imaginary and the real parts may not be very helpful, but
this is just we have to keep in mind for our basic understanding right.

Now, we will try to do an equivalent version of the plots called the bode plots and there the
quantification will be much more clear, the plots would be much easier for example, here I
really wouldn’t know how does this cross here there are also plots where you may say
encounter things like this right like I am doing this and this and this and so on and something
like this much more complicated than what your driving exam would ask you to do it just
asks you to make one 8 figure, here it will be several combinations of those. So, we are not
really going to get expertise on drawing these plots MATLAB will do it for us or there are
several other tools also right. So, once we have the basic understanding of the concept of the
stability and the relative stability in the frequency domain via Nyquist.

487
(Refer Slide Time: 47:05)

We will go in the next class to see the construction of bode plots. You might have seen this
somewhere in your networks course if not then don’t worry we will do it all over again.

So, I will just start doing this on the blackboard first right, just to slowly explain you the
construction and then we will go to the MATLAB and then I will show some slides on that
then what are the equivalent of you know how do I get gain and phase margin using bode
plots and then lastly some interesting thing we will also see is if I am just given a bode plot.
Can I reconstruct back the transfer function of the system and that is also kind of pretty
exciting, and that will give you insight of I just start with a black box I just do not know what
is in the system I gave several frequency signals, I construct the bode plot and I could kind of
get an understanding what is sitting inside the system right? So, all these things we will do in
the next class.

Thank you.

488
 Control Engineering
Dr. Ramkrishna Pasumarthy
Department of Electrical Engineering
Indian Institute of Technology, Madras

Module - 06
Lecture – 04
Bode Plots

Hi guys. So, we are almost half-way through the course. So, today we’ll compute to
complete module 6, and Let’s see, Let’s see how it goes. So, I will start with what is
called the bode plots. So, last time we plotted Nyquist and that gave us a lot of
information. It gave us if we start with an open-loop system which is unstable can the
close-loop system be stable. It also gave us accurately or with some good measures or
some good quantification of relative stability. And we defined two different terms called
the gain margin in a way how much the system gain can be increased before it reaches
the verge of instability. And otherwise phase margin, how much phase can be added to
the system before it reaches the verge of instability. We also saw open-loop system which
was unstable and then we had to actually decrease the gain to get back the system into a
stable configuration. So, what we’ll do today is to look at a little easier way of doing this
right. So, it is called the bode plot it is also called the logarithmic plot. So, I guess
somewhere in circuit you might have encountered this.

(Refer Slide Time: 01:27)

489
So, it consists of two graphs again we start with a sinusoidal transfer function. So, I first
plot the logarithm of the magnitude and, so this magnitude it will be worst of the
frequency and also the phase angle against the frequency. So, some basic nomenclature
or some basic notations will define. So, the standard way to represent this is in terms of
the log to the base 10 and multiply it by 20 ( 20 log 10 ). Not really important of why
this is 20 and why not logarithm to the base n we’ll not worry about that right. So, the
advantage when I do this and we’ll be which will be fairly obvious in the next few
minutes is that multiplications of magnitudes can simply be converted to additions. So,
we’ll see how these things actually translate to just or plotting the frequency response in
the logarithmic scale just translates to just plotting straight lines, like y=mx +c . And
if I have two lines I know how the addition of those two looks like right. So, this kind of
lots of simplification whereas, if when I was in the Nyquist plot I had to be careful of lots
of things, right. It wasn't really clear what happens at ω=0 sometimes. So, there are
poles on the imaginary axis and then, here are the bunch of things that there could be
multiple intersections on the real line, how do I detect those all those were fairly tricky
and then it required lots of computation by hand which could sometimes go wrong right.

(Refer Slide Time: 03:15)

So, We’ll see an easier method of doing that. So, what are the advantages of these? Again
that low frequency characteristics, I can really see very well, right. And in most practical
systems low frequency characteristics are very important in terms of disturbances, right.
Disturbances always occur at low frequencies or their characteristic is always low

490
frequency characteristic. The experimental determination of a transfer function is simple.
So, I just plot how the system gain and the phase changes and I in a way that I just give
the system lots of signals of varying frequencies, of course you know sinusoidal signals
and I just I see what the output is. So, earlier I said that I will do this experimental
determination of a transfer function as a part of this lecture. I’ll do a slight detour right?
So, I will do this when in the module where we discuss non-minimum phase systems.
And it will be little more you know appropriate over there because So far we haven’t
really talked of the classification of minimum and non-minimum phase. I will not even
make the distinction of that because what we do from now on we know we’ll try to avoid
that and keep that separately for an elaborate discussion a little later, right.

So, what we’ll also see interesting is that it is very easy to get this static error constant
through the bode plots, right. Which was not very obvious in the Nyquist diagrams, right.
And what We’ll see here these are just fairly easy to plot than Nyquist diagrams. So, I
will leave the screen for a while and I will just go to the go to the blackboard to do some
initial things and then I will come back to the screen again. So, there is a reason I am
using the board because I think I can draw a little more freely I just wouldn’t directly
want to show you the Matlab plots because that may not be very obvious of what I am
doing right.

(Refer Slide Time: 05:14)

491
So, what do what would we do in bode plots? Okay so, we are given a transfer function
in the sinusoidal form G( jω) and the standard procedure is to plot 20 log of the
magnitude of jω versus the frequency( ω ), right. And the frequency will be on a
log scale. We’ll see what is also the importance of doing this on a log scale ok.

So, this magnitude is usually in this term referred to as decibels . Of course, ω is

in the standard things. Now if I look at a transfer function. So, what are the elements of a
sinusoidal transfer function G( jω) ? Okay, first thing there will be gain K , right?
Then there could be integral or derivative factors, an integral factor would mean a pole at
the origin given by ( jω )± 1 and a derivative factor would be a 0 at the origin would
which should mean like +ω . If I could write it respectively will be minus plus. Minus
for the integral this way and if there are multiple poles it will just be ( jω )± n . Then
there are first-order factors again both for the pole and the zero. And they would look
like this 1+ j ωT and remember I want to write the transfer function always in the time
constant form. It will be −1 for a pole and +1 for a zero and then lastly I will
have a quadratic factors and this quadratic factor represents complex poles. How would

2 ±1
they look like? Just pick
[ ( ) ( )]
1+2 ζ
jω
ωn
+
jω
ωn
. This will also be

±1 depending on if the poles are complex, depending on if you are dealing with poles

ω
or if you are dealing with zeros. And earlier we had one ease of notation define
ωn

¿ u and we’ll sometimes use this notation for u , or for ω=ω n , ok?

492
(Refer Slide Time: 09:16)

Let’s start with a basic of all these things. Let’s say my transfer function just has the gain
term K . So, so what will we usually should we do? right. We’ll say I take 20 log
of the magnitude. So, 20 log of the magnitude which is simply K . This will
represent the magnitude curve. So, if I would just plot it for increasing frequencies. So, I
would just see that it is a horizontal line for all frequencies, right. So this horizontal line
of magnitude K for all frequencies, right. And the phase contribution φ will
always be equal to 0. Say the imaginary part is 0. Another thing to just look at. So, if I
have K , so how will the plot of 10 K look like? The first phase will still be 0. So,
let me just 20 log ⁡( 10 K ) . So, this is 20 log (10)+20 log K . log 10=1 right. So,
this would be just 20+20 log K you know. So, I’ll have a line somewhere here.

So, this is all in decibels, right. Similarly, I can have 20 , 40 . So, it will just be
again we are just adding constants right. So, the plot becomes quite straight forward.

493
(Refer Slide Time: 11:50)

And lastly well this is 20 log K=−20 log ⁡ ( K1 ) , Okay? So, this is the basic one, right.

How do we plot the gain K ? It just a straight line horizontal line verse when plotted
versus the frequency. Constant for all frequencies, not surprising there is no frequency
term here there is no jω term in this sinusoidal transfer function. And the phase angle
is always 0 ok.

(Refer Slide Time: 12:47)

494
 1
So, let me take the integral factor . This is good for us because it helped in the
jω
steady-state analysis or we tried adding an integrator always helps in reducing the
steady-state error as we saw. So, how will the magnitude look like? So, this is looking at

20 log | jω1 | right? So, this is I expand this it will be 20 log 1−20 log ω ; this guy is

0. So, what is left is −20 log ω ok?

So, in this thing. So, what we’ll do is on the ω scale, how is the frequency defined?
So, this is ω and I use this on a log scale right. So, this is essentially log ω . So, if I
just you know plot it on the x1 , x 2 . So, this is x 2−x 1 . So, it will be like the unit
in which we say sometimes: 1-centimetre, 1-meter, 1-millimetre and so on. It’s always
drawn to a scale. So, here we’ll not do frequency from 1, 2, 3, 4, 5, 6 but we’ll do in
something called decades well, I will explain you what it is in decades. So, the frequency
is in decades. And there is a good advantage of why we do in decades. So, a decade is a
band from ω1 , some frequency ω1 till 10 ω1 ok. Now, what does a unit change
in not ω , we are looking at log ω , log ω mean? Again if we draw in the
standard x− y plane if we talk of distances and time. So, one-unit change would
mean 1-second, 1-minute, that’s what we write here, right. Time in seconds, distance in
meters and so on.

So, what does a unit change in log ω mean? A unit change in log ω . So, if
it is this by the time I will say this is t1 , t2 and then t 2 −t 1 is 1-second, right. And if
here if it is said some other thing distance, right. x1 , x2 if I say this is a velocity in
meters per second x 2−x 1 you say 1-meter per second, right. This was the unit. So,
which we learn in high school how to draw the graphs, now what does the unit in log
(ω) here mean? So, this is log ( ω 2 )−log ( ω 1 )=1 , a unit change. Now this is also

ω2
written as log =1 . Which means that ω2 =10 ω1 . So, which also means that
ω1
here, if I say well this is t3 , t 4 and so on. The distance between t 2 and t1 which
is 1-second is also same between t4 and t 3 , right? All the graph is drawn to that
scale. So, here the distance ω=1 to ω=10 is the same as from say, ω=4 to

495
 10
ω=40 . You can do this, right? log , right? You will get a, you’ll say log
1

40 10 40
10; here also log right? So, this is a log is 1, log is also 1
4 1 4
right?

So, the distance between 1 to 10 is the same as the distance between 4 to 40, 40 to 400
all in multiples of 10; 400 to 4000 also, right. Now, will come back to this one right. So,
We’ll be interested in measuring the frequency decades. So, what is here is in a decade.

(Refer Slide Time: 18:29)

So, the frequencies usually will be on the log to the base 10 ( log 10 ) as multiples of
1 2
10 : 0.1 , 1 , 10 , 10 , 10 and so on, right. And the distance between
the ω2 and ω1 this is unit distance, unit distance, unit distance all measured in the
log scale this way, right. And this is what we need to be careful of this is the only
important thing here. Now once we do this everything now translates to just writing
down equations of a straight line.

496
(Refer Slide Time: 19:12)

So, just say this is my y−axis magnitude of G( jω) . So, I am plotting the
magnitude of G( jω) in decibels versus the horizontal line versus the horizontal axis
that is log ω . And this is you can see is sorry, this of the form y=−mx , right.
Magnitude y , x is log ω ; this is a straight line right.

So, if I were to just plot this, from I will not plot at ω=0 because the log value is
not defined, but any value slightly if, apart from 0 we could plot. So, at ω=0.1 the
magnitude is −20 log 0.1 , this will be 20 . Similarly, at ω=1 . So, this is again
all in decibels at ω=1 , the magnitude is −20 log1=0 . ω=1 or say ω=1 0
the magnitude is −20 log 10 this would be, −20 . And similarly you could do for
ω=100 and so on. So, when we compare with this guy, we say well we are interested
in the slope so give me the slope and I could plot y=mx , right. This is a −m .

So, a negative slope. So, here you say unit change in y . So, what is m ? alright,

y2 − y1
this is like 11 standard geometry , right. Unit change ∈ y over
x2 −x1
unit change∈ x , Okay? So, here we’ll see right? what is the unit change? Unit change
here is from 0.1 to 1 decade, right. And, then here I’m measuring the y−axis
in decibels. So, let’s try to plot this, right. Say let me say this is my −40 decibels, this
is −20 , 0 we’ll draw it a slightly better; −40,−20 , this is 0, this is +20 .
Now from point at ω=0.1 the magnitude is 20 . So, I am here and at ω=1 the

497
magnitude is 0 on this here it will be a line here. Now, at ω=1 0; I am here, right
and so on. So, I just draw a line which looks like this ok.

Now, what is a slope, right? See how much the y−axis changes, from here till here
when I am measured from the frequency of 0.1 till 1 it changes 20 in the
negative direction, again 1 till 10 it goes from 0 to −20 again negative
20 right. So, this has a slope of −20 . −20 what? What is in the horizontal
2
axis? From 0.1 to 1 or 1 to 10 or 10 to 100 there will be 10 ,
3
10 . So, this falls at a rate of −20 per decade; −20 decibels per decade. So,
this is how we’ll measure the slope here again it is very similar to what is happening
here, right. y 2− y 1 that is 20 decibels; x 2−x 1 , right. That is a decade, right.
And the advantage of doing this is that I will have a 103 , 104 and so on., right? And
frequencies are usually from very low frequencies of 1 Hz till 103 , 104 . And if
I were to plot that on a graph sheet in the market, I would need to plot a simple transfer
function a graph sheet as big as a football stadium, right. And I wouldn't want to do that
right.

What instead I do, I can just do that in a log scale, right. In the log scale my frequencies
well from 0 , 0.1 till 104 , 106 now I can really get very big frequencies in a
small now A4 size this much of paper, right. Whereas, I cannot do it on a normal graph
sheet. That is another advantage why we do this in the log scale right. So, the magnitude
of this looks like, like this it has a negative slope of −20 decibels per decade. Okay,
what about the phase? φ is what? I’m just looking at the magnitude of this I am sorry,

1
the angle of this − jω . So, the phase the angle or just say the angle or phase of
jω
that is φ is always −90° . I draw the phase here, if I say this is my 0° line; the
phase will always be constant at −90° , okay?

So, the integral factor will have a magnitude plot at −20 decibels per decade and a
phase plot which is −90° . Now how will the derivative factor look like? Let me use a
different colour. So, the derivative factor this is jω . Then I just write down these

1
steps again, So 20 log . So, I am not looking at 20 log , but I am looking at
jω
20 log jω which is simply 20 log ω is the magnitude, right. Now how will the plot

498
of this look like? Let’s say at ω=0.1 , I am looking at −20 log . So, it will not be
minus anymore this is a plus here this one 0.1 i.e., 20 log 0.1 . So, this would be
now +20 . Now sorry, 20 log 0.1 , 0.1 is – 1 . So, this will be −20 .
Similarly, at ω=1 the magnitude of 20 log 1=0 . ω=10 the magnitude of
20 log 10 is this becomes +20 and so on ok.

Now, Let’s start plotting this, at ω=0.1 , I am at −20 . And at ω=1 this one
this line, or this colour ω=1 , I am at 0, ω=10 , I am at +20 ok.

(Refer Slide Time: 27:42)

1
So, this is the plot of jω , the white line is a plot of , right. Just a mirror image
jω
which is not surprising. And of course, what is a phase of sorry, phase of jωφ is
always +90 ° right. So, I will be all the time here just see this the phase of jω at

1
+90 ° and white guy is the phase of at −90° all the time, right. Stay
jω
straightforward, right. You just need to remember something like this and then at the
slope. So, this slope of the purple guy is +20 decibels per decade, right. Because I go
from here till here I increase 20 again, I increase 20 and so on.

So, each decade my magnitude goes up by 20 decibels in a straight line why is it

straight line because of this guy. Something slightly, what if, okay I’ll erase this I won’t
make it anymore mess. What if this is 2 ? Ok.

499
(Refer Slide Time: 29:19)

So, I am looking at 20 log

| |1
( jω )2
. Now this is −20 log|( jω )2| or

−20∗2 log jω or say just ω if I just look at the magnitude. So, this 2 just
comes and sits here, right. Therefore, how will the straight line look like? It will still be a
straight line, what is its slope? −20∗2 that is −40 right? So, it will be something
like this −40 decibels per decade. And what will be the phase? Phase well, this is

1
phase of =¿ −90° I have another of the same guy this will be −180 ° all
jω
integrals. Just write straightforward I will not really now draw this scale and what is this
and all that is a kind of now easy to compute. What happens to this guy? 20

log |( jω)2| this is 20∗2 log ω , right. Look at this number now this will be
something like this, 40 decibels per decade and an angle and an angle φ of
90 ° +90 ° is 180 ° , okay? So now, I can just say well generalise this too; here it
should be 2 , right. For this analysis I can generalise this to n here and n here,
right? So, what will change here is minus 20∗n , right. Here also plus 20∗n . So,
the slope would be minus −20∗n decibel per decade and here it would be +20
times decibel per decay right?

So, the angle would be −(n) time’s 90 ° the angle here would be +( n) time’s
90 ° , straight forward a little exercise, right. I without, so if I have a transfer function

500
 10
now of say, say , how will this go like? Well I am just looking at the magnitude
jω
20 log . So, this will be 20 log 10−20 log jω . So, it is just being shifted by this
number right. So, I will not draw this, right. You just tell that y=mx +c kind of thing
there is a slope still here this is y on the horizontal axis this is will my c is the
intercept −mx+c , right. Very, very straight forward. One observation we could make
here is that the plot or at least the magnitude plot of the 0 is just a mirror image of the
pole right. So, there it was going at −20 and +20 , right. We’ll see that still
continues or not. We’ll see now not these things, but what was the third kind of thing,
third kind of thing was first order factors.

(Refer Slide Time: 34:05)

1
First-order factors are this . Now how the magnitude looks like? 20 log
jωT +1

| jωT1 +1| , right? The magnitude of this thing. So, what is the magnitude? Well, you
can just get this 20 log I get this upstairs that would just be −20 log √ 1+ω 2 T 2 ,
right all again in decibels. Now, it is a little more interesting it is no longer like
y=mx +c kind of form. So, little more complicated, but we’ll try to simplify this.
Let’s try to draw this looks like, this is my ω this is my magnitude of G( j ω) in
decibels somewhere here would be the phase right angles. So, let’s see what happens at

501
 1
low frequencies, low frequencies would mean ω is very small(i.e., ω≪ ), okay?
T
ω is very small this term can be ignored a bit and you can say that −20 log of this
~
guy is approximately −20 log1=0 (i.e., 20 log √1+ ω2 T 2 −20 log 1=0 ).
So, what can I say is that for low frequencies if this is my 0 dB line and I will just be
here, I will use a colour chalk for this one, here. Until what point?

Let’s see when this changes? At high frequencies, high frequencies ω is significantly
large as compared to T . So, −20 log √ 1+ω 2 T 2=−20 log ωT . So, this one goes
away and then ω2 T 2 , the square root will just be ωT . So, how does this look like?
This is −20 log ω−20 logT . So, Let’s see at one particular interesting number

1
here would be at, not here no here, this I write this here at ω= in this the
T
magnitude, right. Sorry like 20 log sorry 20 log of this entire guy 20 log ωT
with a, what sign here? −20 log ωT would be −20 log1 and that will be 0. So,
you see at some point of time or some frequencies are somewhere here which is

1
ω= this guy goes to 0. And after that well I am just moving downwards at this
T
frequency −20 logω . And how does the product just look like? Well, I know it looks
like this, right. This is again at a slope of −20 decibels per decade and this is also
well I can say where it just moves here these two guys meet here and this goes to.

So, well I can I am just making some approximations here I am just starting at very low
frequency then I say I am just at the 0 dB line. I keep on continuing, continuing and at
a higher frequency range I am just doing this going down at −20 decibels per decade.
And this is you can say that the plot can be approximated by two lines right. So, this is
one asymptote this is the second asymptote. And these two meet at this point. What is

1 1
this point? ω= and this ω= is the corner frequency. Now, we can see that
T T

1
this way that for frequencies, for frequencies between 0 to ω to my plot is
T

1
approximated by the 0 dB line, again for ω= infinity it is can stop at −20
T
decibels per decade. So, these are two approximations we get. Now what is the actual

502
 1
plot? The actual plot may have some errors, right. We just may at this ω= it may
T
not be a zero, but something somewhere close to 0. So, we’ll see what that is ok.

(Refer Slide Time: 41:06)

So, now how does the actual plot look like? We’ll come to the phasor a little later. We’ll
see what does, we’ll settle this business first. So, the error is usually the actual minus the
approximate. So, what is the actual? The actual is well I can write this, the square root
can go here this is −10 log ( 1+ ω2 T 2 ) , this is actual thing. The approximate was what?
10 log ω . So, the minus and the minus become plus approximate also had the −10

1
(i.e., −10 log ( 1+ ω2 T 2 ) +10 log ωT ) right? So, what happens at ω= ? This
T
becomes −10 log 2 . So, this will be ωT here. So, from here, right. The higher
frequency things. So, −10 log 2 this will be 0. This would be −3 decibels no per

503
 1
decade right. So, at ω= , I am just having an error of −3 decibels. So, I am
T
somewhere here where this length is −3 dB or similarly I could ask myself what
happens at other points.

(Refer Slide Time: 43:10)

1
So, the same guy at ω=
2T
would look like well ( 14 )+ 10 log ( 12 )
−10 log 1+ okay.

1
So, this you could compute to be −1 dB right. So, ω= . So, if I just add this is
2T

1 1
somewhere here let’s say somewhere here is . So, here the error is 1 dB .
T 2T
Similarly, I can compute for ω=2T the error is again −1 dB at here. So, this is
say ω=2T . So, the actual plot would look something like this. If I just use this purple
colour at the low frequency it is still starts here and at some point it will start will go it
will go and this will go down and then meet this guy asymptotically, right.

Student: sir you explained 20 log ω , T is approximate.

Here?

Student: That one yeah.

504
This should be 20 square root disappears. So, the again let’s recall the approximate
1
thing that was −20 log ( ω2 T 2 ) 2 that is −20 log ωT , sorry about that this will again

be 20 , right. Thanks for pointing that, just be I am sorry about this the approximate
ones. This will still have 20 , this will go 10 because the square root just comes
here. So, this is the approximate plot and the exact plot would look something like this.
And you see at that the maximum error is just at the corner frequencies. So, if were to
just you know vaguely plot how the error changes with frequency right. So, Let’s say this

1
is this is 0 error. So, this will be some plot like this, right. And these are in this
T
case this is −3 dB ok.

So, what about the phase here? So, you can keep doing this for you know ω=3 T and
so on and just compute the error. That is again for a straightforward procedure ok.

(Refer Slide Time: 46:30)

So, we’ll now come back to the phase angle. Phase angle well, this is computed as the

1
−1
tan ωT first one here. So, at φ and at ω=0 , we’ll have φ=0 ; ω=
T
and φ=−45 ° . So, it can be a minus here and as ω goes to infinity and the φ
and the phase goes to −90° . So, if I were to just plot it here. So, let’s say this is the
0° line see somewhere here is −45 ° somewhere here is −90° right. So, the
plot would look something like this at ω=0 and asymptotically here ω=45 °

505
 1
sorry, ω= and at 45 ° . So, the phase would look something like this. Again for
T
exact values I could compute substitute for ω and T and this stuff.

1
Right, so this was first order factors of the form , I could do this exactly in the
jωT +1
same way for a 0 or first order factor jωT + 1 in the numerator. So, nothing would
change just that is −20 will now become +20 everything will be the same just
that in this case φ would be 0° , φ would be 45 ° , φ would be 90 °
well, these 3 cases right. So, this is +90 ° , 45 ° then the plot would have, I’ll use
the another chalk and so on right. Something like this and the magnitude plot would just
be again the same, right. If I just draw the asymptotic magnitude plot using the same
colour. So, I am assuming both are the same corner frequencies it will just be like this
right. So, it will have a +20 decibels per decade and the exact plot would look
something like this while computing the errors, right. This and then asymptotes here ok.

So, it is a very straightforward computation all these processes would go to my −s

would go to plus and so on. And the way you compute for low frequency it will still be
exciting the same for higher frequencies the approximation would still be the same for
low frequencies I can just get rid of. So, low I can just get rid of this term for higher
frequencies I can just get rid of this term and it will just be like a mirror image, right.
Similarly, even for this kind it gets down this guy goes up I will pause here for a while.
And then we’ll continue to look at these guy’s quadratic factors these are all little
interesting, right. To plot the errors. So, we’ll deal with this quadratic factors.

506
(Refer Slide Time: 50:00)

So, we’ll say they are of the form 1 over this is my G( jω) is

1
2
jω jω right. So, again as you, I will just do it first for the pole and the
1+2 ζ
ωn( )( )
+
ωn
zero is just again a mirror image ok.

So, how does a magnitude look like? So, I am just looking at 20 log of this entire guy,

| |
1
20 log 2
right? jω jω this is a, so real part this is
1+2 ζ
ωn
+
ωn ( )( )

[ √( )]
2 2 2
−20 log 1− ( )) (
ω
ωn
+ 2ζ
ω
ωn
. Again as usual we’ll first look at how the

ω
asymptotic plots look like. So, here we’ll see well what happens when which we
ωn

ωn
defined as u earlier is very small, right. If . So, this is sorry, this should be a
ω
log here. So, for lower frequencies So, the magnitude

507
 20 log |G ( jω )|=−20 log it remains 1 this will be 0 . So, well nothing but,
nothing but changes here. So, for low frequencies I am still at the 0 dB line here ok.

(Refer Slide Time: 52:29)

ω
So, what is the other approximation? The other approximation comes when ≫
ωn

1 . In that case my magnitude, this becomes −20 log . So, all these guys will go

2
ω
away all these guys will go away. So, here just remains is just this guy, ( )
ωn
. So, we

2
ω
ignore this guy, we ignore this guy. So, we’ll have ( )
ωn
and this square root. So, this

will this square root and this 2 will cancel. So, I am having something like this. So, this

ω
would be −40 log ( )
ωn
or −40 log ω−40 log ω n correct? So, again at what point

will these two asymptotes meet this is again a straight line of slope, now −40
decibels per decade and some intercept here, right. This can be computed because the
ω1 is already given to you, right. And when ω=¿ ω n , when these two frequencies
match then this sorry, this is −40 this will be a plus here and because it just jumps to
the numerator, right. And then when ω=ω n , the magnitude goes to 0. So, I am here I
say, I’ll just be 0, 0, 0 until this point, what is this point? At this point, I will use the

508
coloured chalk, this guy this till here this is ω is ω n , and after this point I go down
at a slope of −40 decibels per decade right.

So, again I can just compute the errors as it is. So, what is the actual one? The actual
magnitude is, if I just take the square root here it is −10 log of 1 minus, let me again
call this term as u I said earlier, (Refer Time: 54:59) So,
2 2 2 2
−10 log ( 1−u ) +4 ζ u +10 log1 . This is for lower frequencies. Then well again for
higher frequencies the error would again be the same right. So,
2
−10 log ( ( 1−u2 ) + 4 ζ 2 u2 )+10 log 1 is the actual value plus, I have this 40 this is
higher frequency approximation, right. 40 log u and this guy. Now what we see here
is that is no longer just computed by substituting a value of ωn is given to us right.
So, this is this comes from. So, ω they are substituting I just cannot get a close some
expression. So, what we see here it is depending on of course, ωn and also for
different; it will be different for different values of ζ . We’ll not really spend time in
computing the simple I just show you how the plots would look like. Just before that how
would the angle look like?

(Refer Slide Time: 56:22)

[ ] ( ωω )
2ζ
−1 n
The phase angle φ = tan 2 ok.
1−
( ωω )
n

509
So, when ω=0 ; φ=0 , when ω=ω n ; φ=90 ° and when ω → ∞ ; the φ which
is to be −90° . The φ would go to −180 ° right. So, similar thing will happen
when I have this guy in the numerator, will just be this one all other errors is will be the
same right. So, we’ll go back to the screen and see how these things would look like
right.

(Refer Slide Time: 57:27)

So we saw that the error was depending on � for a given ω n . And these are my
asymptotes right. So, just the 0 dB line till here and then this guy is somewhere here,
somewhere here. And you see that the error is maximum when � is very low I am just
plotting for 0, 8, 0.1, 0.2, 0.3, 0.4 is be 0.9 and then as this guy increases when you see
the error kind of decreases, right. Similarly, with the angle right. So, ζ =0.1 , I have
this angle and ζ =0.9 my phase plot looks something like this I just plotted
this directly from Matlab it should be straightforward to compute these guys given ζ
and ω n , but this to avoid all that computational process I am just showing you this
one I could not even draw this accurately on the board ok.

510
(Refer Slide Time: 58:34)

So, now given all these things how would we in general plot a bode diagram? given any
general transfer function. So, the first thing is we rewrite the sinusoidal transfer function
in the time constant form. We identify the corner frequencies associated with each other
transfer function corner frequencies were the one where the asymptotes met, what is
sample here? well this was the corner frequency ω=1 . Then you draw the asymptotic
magnitude plot like these two straight lines. Then what we do is we determine the errors
for lower frequencies for higher frequencies, how do we determine the error? The actual
minus the approximated value again for both lower frequencies and higher frequencies.
And we draw the phase plot well this is kind of straightforward you just look at the what
is the formula for φ and then you add all these phases to get the overall phase plot ok.

511
(Refer Slide Time: 59:40)

Okay, let’s just try to do a few examples on how to how to plot the bode, how to plot the
bode diagram. So, the first thing is to write down this transfer function as a sinusoidal
transfer function or in terms of the time constants.

So, I just rewrite the final expressions that will look like 7.5 , I have the quadratic

( jω3 +1)
7.5
term which should look like this . So, how many terms
jω ( +1 )[ + + 1]
2
jω ( jω ) jω
2 2 2
here we have the gain 7.5 and then we have this guy pole at the origin, first order

1 jω
terms as to +1 and then we have the other first order term in the form of a
jω 2

zero
jω
3
+1 and then quadratic term, [ ( jω )2 jω
2
+ +1
2 ] . So, well I will call this

7.5 the constant as g , I will call this guy g1 , g2 , g 3 , and g4 . So, well let’s
identify what are the corner frequencies there is no corner frequency associated with this
guy, there is no corner frequency associated with this guy, well this guy has a corner
frequency of well 2 then 3 and this guy has √ 2 . So, let’s try to plot this
individually right. So, the first guy the magnitude would 20 log 7.5 right.

512
So, it will be a straight line, right. At that in a proper magnitude and then I cannot read

1
out scale by telling just then would be something like this, right. With a slope of
jω
−20 decibels per decade. What about g2 ? g2 is the first order term it is a pole.
So, I will have in the axis. So, this goes straight for a while and then go down. This is
like again a −20 decibels per decade and the corner frequency here is 2 . Then for

jω
the for this guy +1 . So, I go till 3 and I go up at +20 decibels per decade
3
this is ω=3 . And then the last term well, I don’t have space let’s just draw it here; So
it will start at ω equal to at the 0 dB line right. So, this is at 0 dB line; for this
guy this is 0 dB line for this guy and after a frequency of √ 2 it will go at −40
decibels per decade ok.

So, let’s see right. So, now what how will the overall plot look like? The overall plot
would be the combination of all these 1 2 3 4 and 5 lines. So, well this is like this just
addition of two lines. So, so I start with this looks close to −20 decibels per decade
until this is a first corner frequency this is, so this line is just a combination of these two,
right. And after I did the first corner frequency my slope decreases further. So, this will
be −40 until I reach 2, at 2 I have a pole. So, I will go further down to −60 and
after this is at √ 2 , this is 2 and at so this will be 20 ; say – 40−20 , this would
be – 60 ; it should be −60−20 is −80 and then I hit a 0 right. So, that will
add −20 . So, I will have −80+20=−60 is a slope right? So, this it looks look
something like this and I can actually compute what are the errors by the formula for the
errors which we used ok.

513
(Refer Slide Time: 64:18)

So, just to summarise. So, this is how the first guy would look like the constant then the

1
green line would be the term, right. The red line would be the first order factor
jω
which corresponds to the pole. The greenish line I do not know what this curve is called,
but this would be the 0; the first order term which appears in the numerator and then the
purple line would just be a second order term. I add all these lines and I get a plot which
actually looks like this. So, well can we get some more information about what how does
the close-loop system look like. So, what I do on Matlab is, Matlab actually shows me
the stability margins. So, we quickly see what is happening here and will postpone the
complexity of it to a little later. But what it shows me is that the closed loop system is it
is stable well the answer is no similarly here, right. If I look at how did I do define gain
margin, the gain margin was I look at the point at which you know my phase becomes
180 ° and I look at how far it is from the point −1+ j0 in the Nyquist thing. Here
it translates to decibels right. So, I am just looking at this line and I am also looking at
this line for the phase margin where say I hit the −1 point, right. Or the unit where
my magnitude becomes unity. Magnitude becoming unity in the log scale would be
0 dB line right.

I do all these things, as well the closed loop is actually unstable, what does it mean?
Right. So, I have an, if I just look at. So, let’s just try drawing the root locus for this. I’ll
just directly go to Matlab and I will just draw it.

514
(Refer Slide Time: 66:05).

I hope I just remember the commands properly g , I will call as a transfer function be
the poles. So, I will have 10 and 30. And in the denominator I will have fourth orders one
4 3 2 1 4 3 2
as s ,3s , s , s , and 0 (i.e., s +3 s + 4 s +4 s ) I will just write the earlier
transfer function in this way and this is what I get. So, if I just look at the root locus of
this guy g ok.

(Refer Slide Time: 67:02)

So far, so if I just put my cursor here it will tell you what is the gain right. So, gain is 0
and if I look at well I have the system looks like the system here has a gain of 10

515
right. So, if I see, if I go till here. Well, I am already in the unstable region, right. For all
gains greater than this guy a 0.14 in absolute terms, I’m stable. So, all gains from
0 till 0.1 and if I am at the gain of the system which is 10 , I will become
unstable right? So, that’s what at gain at 10 my poles are on the, right half plane and
my system becomes unstable. And that is what this guy is trying to tell us here, right.
Where it says at the closed loop, the system is actually unstable ok.

Not to worry much about this we’ll keep on revisiting this several times, until we
understand this exactly what it means right. So, We’ll kind of postpone this where we’ll
try to understand at this gain my closed loop system is unstable, what is the closed loop
system? The characteristic equation of 1+GS=0 , right. Ok.

(Refer Slide Time: 68:16)

So, now we’ll go to this example, right. Which will look a little bit familiar if I say this is

(s+ 2)
. So, this is we had plotted the Nyquist for this, right. Which will
(s+1)(s−1)
actually did something like this; this is 0 , this was −2 . I was encircling the point
−1+ j0 ones and then the system was stable, right. We said open-loop system has a
you know this is serious open-loop unstable, but I know that the closed loop is stable. So,

how will the bode plot look for this guy? So, I will just write this down as 2 ( jω2 + 1)

516
and again jω+1 and jω−1 (i.e.,
2 ( jω2 +1) ). And the cross over
( jω+1)( jω−1)

ω1
frequencies would be 1, 1 and 2 okay? From this, ok.
T

Now, a few interesting things here. So, let’s see well the plot of two would be just the

jω
straight line, let me this is a plot of two would be straight line the plot of +1
2
would be something like this, they meet at 2 and then they do this. Now look at these

1 1
guys right So, , . So, these guys will do something interesting. So, the
( jω+1) ( jω−1)
magnitude plot for the first guy say I will take the stable guy first and it will just go to
the ω=1 sorry, this is not drawn to scale, but you know and it goes here, right at
−20 decibel per decade. Now, the second guy again; so, this is plot for the unstable
guy this goes till here to the ω=1 and still be −20 decibels per decade. This is
because the magnitude is the same for this guy. So, what is the magnitude for the first

1 1
guy that is that is same here also magnitude is, the magnitude is
√ ω 2+ 1 √ ω 2+ 1
also ok.

So, just by looking at the magnitude plot I may not be able to say if the system is stable
or not. Because this will correspond to as +1 , this also corresponds to ±1 right.
So, what will change is the phase right. So, this will have different phases and I will
show you the exact plots now how they look like right.

517
(Refer Slide Time: 71:07)

So, this red guy is the So, this red guy would be you know the zero at −2+ K , right.
And then the green guy So, these plots actually over lap and this will be both the stable
and the unstable pole, right. And then now look at the phase right. So, we see that you
know if I were if you just remember how the first order terms look like. So, the angles
started from 0° it went to −45 ° at the cross over frequency and then here.
Similarly, with this guy and this is how the overall plot actually looks like again.

So, a couple of questions which we can try to try to answer here. So, can just by looking
at the bode plot can I say if this is if it corresponds to a stable or an unstable pole? Well,
just by looking at the magnitude plot; well at least I can say that this is a pole, right.
Because this zero will go up. Now what was the how did the magnitude behave for a

1
term like this one? Sorry, the phase , I start with at ω=0 , I start from
jωT +1

1
φ=0 then ω= , I start at then I go at −45 ° and as ω → ∞ , φ=−90 ° right.
T
So, that is happening in the green line, but the blue line does something strange, right.
This starts at 180 ° and ends up at −90° . So, this is one thing which you can
distinguish just by looking at the magnitude plus the phase plots of if it corresponds to a
stable pole or not. So, how does the overall system now look like? Overall system well
the transfer function looks something like this and this ok.

518
So, what was a gain margin well the closed loop system is stable, right? If you just look
at 1 plus you know this thing being equal to 0 the closed loop system is stable, but
it has a negative gain margin, right? −6 So, which essentially means if I draw the
root locus. So, I have a pole here, a pole here and a zero here; this is +1 , this is
−1 , this is – 2 , this is 0 here. So, for some values of So, at K=0 , I am
unstable well, not I am not unstable the system is unstable. Then I keep on moving until
at K=0.5 , I reach this point. So, my system is stable after this thing. So, the gain
margin would be 20 log 0.5 , right. And this will turn out to be negative or 6
decibels right. So, −6 dB is the thing. And if you remember last time, I said these are
systems which have negative gain margin, right. Even the Nyquist thing right. So, look at
how the Nyquist looks like. So, this was – 1 , this was −2 and for a system I am
just really messing up this, but that’s okay. For a system when we started the Nyquist

1
arguments we say if this was an a , was the gain margin right. So, if you
a

1
increase the gain by a factor of the system was stable. Here what I am doing? I am
a

1 1
decreasing it by a factor of , because this you know in the negative direction
a 2
and therefore, I have this negative gain margin – 6 .02, this I think the Matlab will tell
you directly.

Of course lots of things again we’ll come on this well we are dealing with stability and
also minimum and non-minimum phase system, at the moment you can just understand
this from what is happening through, what we learnt through the Nyquist that I really
have to decrease my gain for this for K decrease my gain. Because the open-loop
gain is one and also for the root locus it is obvious, right. That I start from an unstable
system and I have to make it to a stable system. And the close-loop is stable here unlike
the previous case. And that is what you know Nyquist tells us. So, before this I will just
do try to do one more example right. So, I will just let me say I am looking at. So, so far
we have just been doing, you know a negative transfer function. So, in the negative gain

1
margins. So, let’s say I take a system and it is still not ok.
( s+1)(s +2)( s+ 3)

519
(Refer Slide Time: 76:08)

1
So, G ( S )= . And I could see which is also plot and check at this is
( s+1)(s +2)(s +3)
−1,−2,−3 . And, the root locus would look this guy will go here these two guys will
'
meet up here. And they will go here at this is K=0 and this is K=K that is we’ll
find out.

1
jω
Now, Let’s try to write this in the sinusoidal form. So, this is
(
( jω+ 1 )
2
+1 )( jω3 + 1) .

1
And what will come here is 6 (i.e., jω
6 ( jω+1 )
2 (
+1 )( jω3 + 1) ) So, I have a thing of

1
and then I have cross over frequencies as 1, 2 and 3. So, Let’s draw the root locus
6
and the Nyquist both for this I will check ok.

520
(Refer Slide Time: 77:17)

So, I go here I will call g1 as my transfer function. So, on the numerator I will have
3 2
1 and then the denominator polynomial would be s +6 s +11 s +6 now perfect,

1
i.e., 3 2 . So, I would want to draw the root locus. So, I will have root
s +6 s +11 s+6
locus of g1 and let’s see how it looks like ok.

(Refer Slide Time: 78:01)

So, this is almost like what I have drawn. So, you see at a gain of around 60 or
slightly around right. So, I think if we are around gain of 60 , my system is on the

521
verge of instability. So, that is I will say some 0 to 60 the system is stable
otherwise the system is unstable.

(Refer Slide Time: 78:23)

Let me say what the bode says, bode of the same guy g1 ok.

(Refer Slide Time: 78:33)

So, this is how the bode diagram looks like and if I ask the system to show all it is
stability margins on the grid then this is something nice right.

522
So, you have gain margin of you know this is a positive 35.6 dB which will actually
be 20 log 60 . And then similarly it will have well let me phase margin well it is stable
for all figures. How do we compute this? Gain margin again we look at the phase cross
over frequency at 180 ° , right at 180 ° some somewhere here. So, this is a gain
margin. The phase margin well I have never actually touched this zero line. So, there is
no question of I computing the phase margin here, but this is system which is which
starts with a stable configuration when K=0 it stays stable until K=60 in
absolute terms or 35.6 dB and later on it just goes on to the verge of instability, right.
There are also systems with Let’s say an infinite gain margin.

(Refer Slide Time: 79:40)

Let’s say a standard second order system say g2 is tf it is like this one, this is a very
simple second, order system and this is what the first example we had done when we are
even doing root locus 1, 2 and 1 and this looks good and let me just first draw the root
locus of g2 ok.

You see the root locus is exactly what we have drawn right. So, this is for all gains K ,
it is stable.

523
(Refer Slide Time: 80:20)

Then you look at how the bode of this looks like. Now I will go to the grid and I see how
the stability margin look like, right? So, you will have a phase margin of 180 ° and
then there is actually the this actually never reaches well the 180 ° which is at infinity.
So, you will just have the gain margin to be infinite, you see it never reaches 180 °
right. So, these are systems which have infinite gain margin and a phase margin of
−180 ° means kind of very good for us right. So, back to the slides. Now, to just
before we conclude what we’ll try to do is can I actually find out the error constants from
the bode plot? ok.

(Refer Slide Time: 81:20)

524
So, let’s say first we look at the position error constant K p , right. How was a typical
transfer function like? Well, if I write it in the sinusoidal form? I have K , I have all
the set of zeros, say I will call this T some number m over some poles at the
origin and I will have ( T 1 s+1 ) … . ( T n s+1 ) (i.e., G ( s )=¿

K ( T a s +1 ) ….. (T m s+1)
N ). So, when I can I just write this in the sinusoidal transfer
s ( T 1 s +1 ) … . ( T n s +1 )
function just by substituting s with jω .

So, let’s say n=0 and this is the only case where I am interested in the position error
constant right. So, when s=0 , well how will my bode plot look like? So, my bode
plot initially will have a certain constant magnitude. Because this guy doesn’t appear
right. So, this is just one. So, it will just be K times this or things whichever is this.
So, all these guys at the lower frequencies are at the 0 dB line what else changes
K right. So, this will be something like this and then they will go whatever if there is
a here and then there is a zero and again a pole and so on. Right. So, this is the
−20 dB per decade and this is the first corner frequency. So, what is this magnitude?
The magnitude here is so, how do we define a K ? The position error constant that
was a limit s going to 0, G( s) . Now, same thing if I say, what happens when
limit? ω going to 0, G( jω) is well that is simply K , right. This guy goes away
I just put ω=0 this is what I am left with.

Now, this thing is again at ω=0 right So, 20 log K p and this K is essentially
K p , right? So then I write it in the time constant form. So, how do I find K p ? Well
I just see where does this intersect this, this what is this constant? And then this is
actually 20 log K p , when I once I know 20 log K p ; I can find out what is Kp .
Then we look at the velocity error constant.

525
(Refer Slide Time: 83:59)

And this has typically will we are interested only in a type−I system. So, which
means the air for lower frequencies the plot would look something like this. Or I can say
a plot would look something like this and then well maybe after a corner frequency I just
go down and you know do whatever. So, for lower frequencies G( jω) because the
system is type−I will look something like this. These are again for very small ω s
ok.

So, 20 log of the magnitude of | jωK | . So, let’s say that well there is some point.
So, I just extend this line, right. Till where I do not really worry about what happens at
the corner frequencies. And, say what is this value at ω=1 ? Well, at ω=1 the
value is 20 log K v . So, I take this say ω=1 here I extend this line and this value
would be 20 log K v . So, this is my magnitude in decibels and this is frequency now
well something else has happened right. So, I am just going down and somewhere I am
in intersecting the somewhere here, right. The 0 dB line. 0 dB line essentially

K
means at the magnitude at some say ω=ω 1 is 1 let me call this frequency has
jω
ω1 right. So, I just keep going down and this one in the decibels turns out be 0 dB
. I am just looking at 20 log 1 that would be a 0 dB line. So, I will ask myself what

K
is the frequency at which the magnitude of goes to 1, well that turns out to be
jω

526
 K=ω1 , right. And, what is this K ? This K is essentially my K v , right. All

these are K v S where I can say what is Kv ? lim sG ( s ) or similarly,

s →0

ω → 0( jω)(G( jω)) .

So, if I take a type−1 system this will be 1. So, whatever remains after everything is
substituted to 0 which is K and this K is my Kv right. So, G( jω) for lower
frequencies and this is Kv , Kv and. So, the frequency ω1 at which this line
intersects the 0 dB line this ω1 is then my velocity of a constant. So, I can do it
two ways see what is a magnitude at ω=1 , right. And that will give something
20 log K v ; given 20 log K v I can always find out what is Kv or other ways I
can just extend this line until I get the 0 dB line. In absolute terms what does this

K
mean? 0 dB means a magnitude of at ω1 and this K is essentially
jω
Kv because of this guy this is K . And this is from the definition of Kv this is
K . So, Kv can be found out by the intersection of the initial line with a 0 dB
line. And the frequency at which this happens is my Kv then this should be
straightforward to visualise.

(Refer Slide Time: 87:52)

We look at the acceleration error constant, right. In this. So, so We’ll have n=2 or a
type−2 system right. So, here So this initial thing will be −20 decibels per

527
decade line. And here well, what happens for low frequencies? I am now looking at
G( jω) is K and this Ka again turns out to be the acceleration error constant,

right. By ( jω)2 how do I compute K a=lim s2 G (s) , right. And then due to the
s→ 0

jω and whatever K is left is actually the K a ok.

Again these are for lower frequencies. So, if I just plot I will have a −40 decibels per
decade line and then maybe something else happens something else happens and so on.

ω
j¿
¿
¿
So, I just extend this line. So, I am interested in log; Ka , what happens to this
¿
¿
20 log ¿

ω
j¿
¿
guy at ω=1 ? This guy at ω=1 is simply
| )
Ka
¿ω=1
=20 log K a
. So, Let’s say this

¿
20 log ¿
is at ω=1 , right. And their intersection is 20 log K a . Other way to also compute is
to see at what point of time this touches this axis. So, I am looking at magnitude of

ωa
ωa j¿
j¿ ¿
¿
¿ ¿ 1 sorry, or equivalently in decibels Ka is, 0 dB line, right. It is a
¿
Ka ¿
¿
¿ ¿
20 log ¿
0 dB magnitude of one in absolute terms corresponds to this guy. So, what will this
tell me is that ω a , so let me just call this at ω a , right. At ω=ω a , so

528
 ω a=√ K a ,square root of the acceleration constant. So, I just keep going this till I reach
the point ω equal to till I reach the 0 dB line and the frequency at which this
happens is the square root of the acceleration error constant ok.

So, again these just occur only for type−2 systems, Kv occurs only for
type−1 system and K p occurs only for type−0 systems. And we know why
this is true right. So, we saw we saw several advantages now of the bode plots, right.
Much easier to compute than the Nyquist, right. We just have to deal just by adding
straight lines, right. And well a bit of we could visualise how the gain margin and phase
margins look like even though we’ll really extensively discuss that a little later. What we
also knew towards the end is that given a bode plot, I could actually compute what are
the K p−¿ the position error constant, K v −¿ velocity error constant and K a−¿
the acceleration error constant? right. So, that is what we learnt today.

(Refer Slide Time: 91:38)

So, what we’ll postpone the experimental determination of transfer functions via of bode
plot when we finish learning what are the minimum phase systems I will not really even
attempt to define that at the moment, but let’s assume that there is something excess or
just to please you I would say well minimum phase system are the ones for which all the
poles and zeros are to the left half plane. For example, this system was a non-minimum
phase system because there is a pole on the right. So, that everything is into the left is a
minimum phase system. So, just juts remember it that way for the moment, right. Ok.

529
So, next class what we’ll do is to slowly understand the concepts of designing a
controller, right. And what are the basic elements? These are well what they call as a
proportional integral and a derivative control. So, we’ll see how to smartly use these
individual elements when to use this is a proportional controller enough for sometimes I
need to add an integral control, whatever a derivative control can I actually realise a
derivative control, if I cannot realise the derivative control what are the other methods I
will do all these ways slowly try to build up the theory behind it, right. Ok.

Thank you.

530
 Control Engineering
Dr. Ramkrishna Pasumarthy
Department of Electrical Engineering
Indian Institute of Technology, Madras

Module - 07
Lecture – 01
Part – 01

Basic of control design – Proportional, Integral, Derivative actions

So, in this module, we will start to familiarize ourselves with some basics of design. Now so
far what we have done is modeling and lots of analysis of the system, analysis in terms of
stability, starting from the Routh criterion, the root locus gave us lots of other information
about how my closed-loop systems change as I vary the gain from 0 to infinity. There’s also a
parallel of what to do you know for gains from 0 to minus infinity, so straight forward as I
said, we’ll keep that for a little later that’s not important at the moment. We also saw analysis
in the frequency domain, we characterized relative stability in terms of the gain and the phase
margins. So, what we will do now is to slowly learn concepts of design.

Now, what are typical design problems? So, while we were doing time response analysis or
even the frequency response analysis, we had some basic parameters or basic things which
we observed like the rise time, the settling time, the peak overshoot, and we saw how all
these were related to the natural frequency and the damping. Similarly, the frequency domain
concepts like the bandwidth, the resonant frequency, were also in a way related to these
system parameters. Now any design problem would be based on setting these values like the
peak overshoot or the rise time or the settling time, the steady-state error, to some desirable
numbers. Now, that is what we will focus on how do we go about doing this. So, there are
three basic elements in this control design. So, we start with a proportional controller, the
integral, and then the last would be the derivative actions.

531
(Refer Slide Time: 02:26)

So, let’s slowly go through these things right. So, we started defining in our earlier lectures,
the closed-loop system of the form; of this form where you know I had a plant to be
controlled via this controller, I have a reference signal R( s) which I need to track here,
then there was disturbances in the system, there was noise and so on right. So, this feedback
configuration we saw is capable or it does improve the stability, it also in some sense helped
in meeting the performance specifications, we will elaborate on this a little more shortly, we
also saw how feedback helps decreasing the sensitivity to certain parameter variations of the
system, disturbance ejection was also very obvious while we were doing the feedback control
and of course, attenuating of the measurement noise and so on right.

So, of course, what was the defining factor even when we looked at the error constants was
you know this is the signal E( s) right.

532
(Refer Slide Time: 03:30)

Okay so, a typical first-order system with some time constant T looks something like this,

1
G1 ( s ) =¿ right and then it’s response was supposed to be like this right,
1+T s
depending on the time constant how fast or how slow it was and it also had inherently some
steady-state error. So, this u(t) , this is okay let me call this y (t) or c (t) depending
on whatever terminology or notation we are using. Okay, the second-order system was of this
form I had the natural frequency ω n2 , I had the damping ratio � and so on right. And for a
second-order system we defined some basic properties like the rise time again, it was
depending on � and ω n , the delay time was also depending on � and ω n , similarly, the
peak overshoot was only depending on the � and settling time again both � and ω n .

533
(Refer Slide Time: 04:34)

Okay so, typically any performance matrix is defined in terms of these 4 or 5 things which we
defined here- rise time, delay time, settling time, overshoot, and sometimes also the steady-
state error. Okay, now if you see that you know all these were defined only for second-order
systems and typically you may say well all; not all systems we come across are second-order
systems, it might be; we actually saw third-order systems, we also did examples where we
had a system with seven poles and so on. So, how do we deal, how do we translate these
situations or these design parameters or performance matrix into cases, where we have say
several poles in the system. Okay so, let’s say I have a configuration like this right; I have
these two poles here, I have these two conjugate complex poles here, a pole here and a zero
here right.

So, this is something which has like 1, 2, 3, 4, 5 poles and a zero right. So, say well what is
the contribution of the poles and zeros to the response of the system. So, you can say that
these two poles right, I call them as the dominant poles of the system because these as time
increases the response of this now it dominates more than these 3 or 4 guys over here right.
So, we’ll see; you know actually I’ll show you some plots, and then also we’ll do some
analysis on this right. So, let us first start by analyzing the system where there is an extra pole
or zero along with this dominant pole, and I think over the next few slides I think this concept
of dominant pole should be more or less clear ok.

534
(Refer Slide Time: 06:37)

So, we start with a typical second-order closed-loop system and we know that this is always
stable, so we don’t need to worry about stability here, and just for you know computational
purpose or for simplicity I say well, I set the damping ratio to 0.5 and ω n=1 . Now, I
will start by investigating what is the effect of adding a zero. Say, where do I add the zero,
let’s say I add a zero at s=−a , again in the left half-plane. Well, I just do a little
manipulation here to maintain the DC gain at 1 nothing really changes, but just for things
to look a little neat, I add you know divide this s , so the way I get this is as follows. So, if
I just use the zero as s +a , my DC gain would be a right, and I just want to make it

s+ a
1 . So, I’ll just do for which the DC gain is 1 ok.
a

So, the response would now; well, how will my transfer function now look like C1 (s) is
this guy the zero at −a , in such a way that the DC gain is 1 divided by this guy and I
will get something like this right. So, I have the original thing. So, when ζ =0.5 , ω n=1

1
; this guy just becomes 2 , okay. So, now, look at something interesting here right.
s + s+1
So, this is my original transfer function which is already here and then what happens to the

s
transfer function if I add this +1 . Well, I still have this guy in the denominator and a
a
numerator at s and we know what is the idea of this s right, in the Laplacian domain; it
has just to do with differentiation, loosely speaking multiplying by s would mean

535
differentiating that particular signal right and therefore, if I look; write it in terms of in the

1
time domain c 1 (t)=c( t)+ ć (t ) .
a

(Refer Slide Time: 09:05)

Okay now, what happens to the response? Well, this just for simplicity take the step response.

1
So, Y ( s )=C 1( s) right over you know C1 (s) with a step, unit step has a Laplace
s

1 C (s) s C ( s)
of . So, this is + right. So, where does this come from? There is
s s a s
something here right. So, this is if I were just to write it properly. So, this would be in this

s
C1 ( s )=C ( s )+ C ( s) okay, and that’s what happens here, okay. So, this is
a

s
Y 1 ( s )+ Y 1 (s ) right.
a

So, I’ll just define the output corresponding to the original plant C1 ( s ) as Y 1 ( s ) . So,
now, I go back to the inverse and do the inverse of the Laplace go back to the time domain

1
and what I have is y ( t )= y 1 ( t )+ ý 1 (t) the original response without a zero. Okay so,
a
when, so I just denote ý 1 (t) as y 2 ( t ) , and the total response is y ( t )= y 1 ( t )+ y 2 ( t ) .

So, let’s again; let’s just you know remember this for a while, y1 ( t ) is the original

536
response without the zero, y2 ( t ) is just the response of the zero, of adding the zero and
then y (t ) is the total response.

(Refer Slide Time: 10:37)

Okay so, if I just take this system and I say I add a zero at s=−0.5 , okay. So, what was
my original response I’m calling this as y 1 (t) . So, this is the signal in the black right this

is y 1 ( t ) , now I add a derivative term and what does sorry, I add a zero and what the zero

1
does is adds this extra term into the response y 2 (t) , what is y 2 (t) ?
a

1
y 2 ( t )= y´1 (t ) , okay. So, the total response y= y 1 ( t ) + y (t) , is just y this is red line
a 2
here, okay. So, this, the black line or the black curve was original response, I add a zero at
s=−a and then I get this one again by maintaining the zero, maintaining the DC gain to
be 1 , okay. So, this is at s=−0.5 . So, where were my original poles of the system? the

( −1 ± j √3 )
original poles of the system were at . So, I have something here. So, this is my
2

−1 ± √3
−1 , I’ll have , somewhere here and here right. So, these two are the poles
2 2
of the system and I’m just adding a zero over here and this is again σ and the j ω axis
ok.

537
Now, I say I add a zero I remove this guy and I add a zero over here, at s=−1 and I
remove this guy right, this guy no longer exists. So, what I have is the second-order system
plus a pole at sorry; plus, the zero at −1 . Okay so, again we follow the same thing and see
what happens to the response, the response in the black remains as it is, the scales are
different, and therefore, this looks a little bigger right ok.

1
Now, y 2 looks like this, which is ý 1 , and the overall response now is something
a
like this, okay the red line. Now I do, what do I do; I just remove this zero over here and I say
I add a zero much further away right. So, somewhere here and I say this is a zero at −8
okay, now what happens well y 1 just still remains the same, y 2 well, the blue line does
something like this, and then I have the total response is something like this, okay. Now again
let’s compare these three plots, right. So, as the zero is closer to the origin you see here
difference is quite big here right between the original response in the black curve and the red
and the response in the red which is the effect of adding a zero, at s=−0.5 . If I add a zero
at s=−1 , well there is still some effect but the curves get much closer. Add a zero further
to the left, right? at very far; at s=−8 , say that the response is more or less the same in a
way I can just say well, I just don’t add the zero right and nothing changes ok.

So, apart from these distance in the curves here, what else do we observe? what is the
difference between this black curve, and the red curve here? Well, one thing is obvious that
this has an increased overshoot okay. Now here, so the steady-state value is one. So, I’m my
overshoot is like almost between 60% and 70% okay, this is at zero at −0.5 . Now here the
pole is; the zero is slightly to the further to the left. So, the overshoot still increases, but not
by too much okay. Well, here well it’s more or less the same. Well, first is an increased
overshoot depending on the location of the zero, but what happens here right? There is no too
much change in the settling time and it’s much easier; they more or less raise to 98% of the
original value in you know, fairly the same time right, all 3 guys including you know the
original one and then well what is good here is that I have a faster response in a way that I
reach the peak time much faster here, here I was taking about two and half seconds, here I
reach in less than one second.

Similarly, here also the response is faster right now I reach this time you know when;
obviously it’s like two and half seconds here, I reach about little over one second and so on
right. So, my response is faster and as I keep shifting or pulling my zero further to the left and

538
I keep doing this, what I observe is that as a goes further to the left and furthest then the
contribution to the zero or the contribution of the zero to the system response decreases, and
we get back the original behaviour. Say, this is −8 0 , I don’t really need to plot this curves
also, it would just be, almost exactly be the same right. So, there is some effect of adding the
zero and it is based on where we add the zero. So, we want them somewhere close to this,
these guys to see the maximum effect, of course, there are also drawbacks here right, about
that in a way that the overshoot actually increases, but nothing changes in the settling time.

(Refer Slide Time: 16:31)

Okay now, what if I add a zero in the right half-plane right? Again we do the same thing that
to maintain the DC gain at 1 , we just divide the entire thing by a . So, my transfer

s 1
function of the system is now 1−¿ is again looks fairly the same, I have 2
a s + s+1

−1 1 1
; I just have a minus here, or in other words c 1 ( t )=c ( t )−¿ ć (t ) ,
a s +s +1
2
a
is at a plus earlier.

Similarly, the response here also translates in the same way, that the total response, is the
response y1 and then, in addition, we have this response due to the zero that is placed at
s=a .

539
(Refer Slide Time: 17:24)

Okay so, what is happening here? Well, again the black curve is as same as in the previous
case, the blue curve is the effect of adding a zero on right half-plane, now the red curve which
is the combination of these two looks something strange now right. So, initially, if you look at
this plot you might see that, well what I what do I want to reach? I want to reach the point
plus 1 right? In, well with less number of overshoots and you know settling time should be
fairly small and also no steady-state error.

But what I’m doing is, initially if I just switch this on, and I see that well I’m actually going
away from zero which might mean, well you know I just want to stop it over here because I
think well the system is doing something very weird right, but then if I just let it go for a little
while, it reaches you know the minimum here and then it actually starts going up. Going up
and then does, you know something same as the previous one, that there is some increased
overshoot, of course, you know there is nothing much in the response is a little slower, the
overshoot has increased and of course, the settling time doesn’t do much things ok.

So, what is the first observation is that well, so far we have only analyzed overshoots right?
How much, so I start from 0 my desired value is 1 , how much do I overshoot before
settling down at 1 right. So, that is over here right this is this was the overshoot, now in
addition to that we also experience something called an undershoot, I go down from −0.5
till just almost −0.8 and then come back here right and then this undershoot what it does,

540
is that it actually increases my delay time, the rise time, and the peak time and so on right.
Which means that the, my response is quite sluggish now.

So, when poles and zeros are in the right half-plane, these are called non-minimum phase
poles and zeros. We defined this last time and we’ll still stick to that definition that if all the
poles and zeros are to the left then it’s a minimum phase system, and if some guy sit to the
right it’s a non-minimum phase system. Okay, we will deal with this when we talk about
design in frequency domain or maybe some special module on just dealing with right half
zeros.

So, at the moment we’ll just restrict ourselves to this definition that will be enough for our
analysis. Just that we just remember that if I have a right half zero, I experience an
undershoot and this will always happen and I’ll show you a little proof for that a little later.

(Refer Slide Time: 20:20)

Okay so, here what we will see is the effect of adding a pole to a standard step response of a

1 1
second-order system. So, well there’s the standard response is this one, . Now I
s + s+1 s
2

add a pole right, I add a pole at different locations a . So, first I see that if I just look at
y 1 , this is a natural response or the response without adding a pole, if I add a pole very
close to the origin and you see there is a significant change in the response also good in a way
because you know I kill the overshoots and other stuff. Similarly, if I had a pole slightly

541
further may be like little 5 times to the left and I see well the responses almost look the same
right there is a very little margin here and here and so on.

Now, if I go; keep going further I see that the red line actually touches the black line right.
And that is kind, not surprising right. So, as a→∞ the contribution of the pole, it
decreases and as you know as this pole a keeps on going to ∞ there is no effect at all.
So, whenever we do in a design process where we have to add a pole such that it does not
have too much effect on the transient response, we take it to be 5 times the real part of the
dominant pole as in this, in the middle case when we have a pole at s=−2.5 . So, we’ll
keep this in mind while we were, while we will explicitly talk about design problems when
we are looking at you know improving the transient performance or the steady-state
performance and so on.

(Refer Slide Time: 22:01)

Okay so, this leads to the definition of our characterisation of the dominant poles of the
system. So, the dominant poles are the closed-loop poles that have dominant contribution on
the transient response of the system, and it is quite possible that this occur in conjugate pairs,
they could also just be sitting here right on top of each other, say at s=−1 , and higher-
order systems are generally adjusted such that there exists a pair of dominant complex
conjugate poles, which in that case, we would just restrict our analysis to these two poles and
we could do the entire analysis and design right.

542
But for that, well it is desirable to have the real parts of other poles and zeros at least 5 times
further away than the real parts of the dominant poles. So, if this is −1 , this guy should be
at least at −5 and further away. This is at −1 and there is a say a dominant not really
dominant, but say some other pole pair at this one, then I cannot really call these, the blue
guys as the dominant poles. So, these are like typically because the response of these guys
over here they die down faster, the response of these guys is even like little slower than this,
but still, fast enough then this also dies down fast and what is remaining is just this one right,
and the limiting case is when the poles are here right, the response never dies down right. So,
this if this plus this is the case then these are the dominant poles because this will still go to
die down to 0 after may be a fairly large time, but this will remain as it is.

So, if I; you know look at this in terms of a classroom, the guys who are on the, you know
sitting on the last bench furthest away from me go to sleep much faster, than these guys, than
these guys and possibly you know the dominant guys are the ones who are on the first row or
the second row right. So, it’s just like the attention span also right. So, these guys loose
attention faster, these guys little slower, well these guys possibly are a little more attentive
than these other guys. So, if these guys are here well they still, you know go to 0, but a little
after; possibly after 45 minutes of the class or something. Okay and this is this guy is me here
right the teacher on the imaginary axis, he cannot go to sleep right he will always be talking
talking talking.

(Refer Slide Time: 24:37)

543
Okay so, now we’ll see what is effect of adding poles to the open-loop transfer function. So,
2 2
far we were just doing the closed-loop one right, s +2 ζωn s+ ωn right. Now we’ll see,
what is the effect of adding poles to the open-loop transfer function well, we will do this in
terms of the root locus. Now the step response sometimes may not give us too much
information, but we know root locus actually is a little easier to interpret and things are much
straight forward some, in some sense to see from the root locus. So, we will switch to the root
locus plot right.

1
So, let’s say I just have a G ( s )=¿ it’s just a pole here, how the root locus look
s+ 2
like as the gain K increases it just goes further and further to the left. It’s stable all the
time. Okay, now to this guy, I add another pole at s +4 . So, this is my original pole, I add
another pole here and what I know from the first class of root lecture; root locus is that well
this guy will move towards the right, this guy will move towards the left, they meet each
other and then they go their own ways ok.

Now, then I say well I’ll just do some more experiment, I’ll just add s +8 . So, I have s
at −2 , −4 , and −8 . So, I have 3 guys; all three would go to infinity because there
are no zeros, this guy will choose the lazy path go here, these guys have to move; this guy has
to move to the right, this guy has to move to the left because something tells me that a point
on the root locus lies on the real axis only if the number of poles plus zeros to the right is an
odd number, okay. Therefore, this is not allowed to be on the root locus. This guy, well this
entire line is allowed to be on the root locus, so he just takes the easy path and just goes runs
away. Now, these two guys meet here, I calculate the angles of the asymptote that will be
60 ° , then this would be 18 0 ° and a −60 ° .

So, these guys will go this way, okay. So, what is happening right? So, first well the root
locus was, which were just going to infinity in the first case is now pulled now a little bit to
the right, right? So the root locus exists just in this domain, it was earlier existing in this
entire domain here right and I add one more pole and it shifts it further to the right, in such a
way that it can also go to the verge of instability after I attain a particular value of the gain
K right. So, these are very simple things right there is nothing to actually remember as a
formula or anything like that, but just that we are just plotting various things just kind of just
doing some initial experiments to then you know evolve towards a general theory. Okay?

544
(Refer Slide Time: 27:46)

Now, what happens if I add a zero now to the open-loop transfer function? Okay so, let me
say I start with this open-loop configuration. So, I have a pole at the origin, a pole at −2 ,
a pole at −5 and of course, you know if I just plot the root locus this guy will go away to
infinity, these two guys will meet up each other and then they again go to infinity, and after a
particular value of K at the Routh Hurwitz will tell me that I’m going towards the
unstable region. Okay so, let me just add a zero just slightly towards the left of −5 right.
So, this was −5 and this is a zero.

So, this guy which was now running away is now trapped by this zero here, not only this
right. So, this guy also does something nicer, he not only holds up this guy here, he also does
things in a way that the root locus now is well more or less shifted to the negative region
right, there is no point of instability here because my asymptotes are just the imaginary axis,
okay. Now, I add pole somewhere not really you know very far at −7 , but somewhere
between these two poles right this and this. So, well this guy again it’ll just move to the right
and stop here.

And now I see that the root locus has improved much because if I were to deal with this gain
here, you see that I’m close to the verge of instability right this and this right. So, this this this
kind of adding a zero it makes my system stable because it is pulling this blue and the red
lines to the left, but it is; well it may not be too good in terms of relative stability. Now, I add
a zero at −3 and I see well it is actually fair, fairly stable for all values of gain K right,

545
and if I add it further add −1 then I get a configuration which looks like this right. Okay
so, what does the observation tell us? The observation tells us that the addition of a zero to an
open-loop transfer function it pulls the root locus to the left, right. And the effect of zero is
prominent when it is close to the imaginary axis, okay. So, in these plots, we just start with a
third-order system and then study several effects right ok.

(Refer Slide Time: 30:13)

Now, based on these observations we’ll just start with the basic element of control. One is
the, first one is the proportional control is it just means what happens if I just keep on
adjusting the gain, does adjusting the gain achieve my control objectives or not right. You
either increase the gain, decrease the gain and so on and we’ll see under what conditions just
a proportional controller or just by adjusting the gain K we meet our performance
specifications; we’ll also see if there are any drawbacks right just by adjusting the gain even
the system is stable does it always help us achieve good results. Okay, so that’s what we will
do here right. So, again I have an R( s) , the G( s) is the plant, and my controller is just
a proportional controller, okay. So, as usual, this is my reference signal this is my output there
is a certain error and then and then the controller is supposed to adjust according to itself
based on the error and so on, okay.

So, the first question is when will a proportional controller work. So, the steps we usually do
here are the following, let’s say I’m given that well that my peak overshoot should not be
more than 20%, I’m also given a specification that the settling time should be well say 5

546
seconds or less correct and maybe some other things on Tr and so on. So, what this will
translate to is some � and some ω n , okay. Now, I just keep on moving my gain from zero
and see at any point of the root locus, if these specifications are met, these are my dominant
poles, if it is met at this point and this point then I say yes I can just adjust the gain and then
that value of gain K would sit in the controller and do my objectives.

But if, this � and ωn translate to these pair of dominant poles, then no matter whatever I
do with my gain, I’m never here right. So, when does this work well, of course, we have the
freedom to vary the gain K , which means we are restricted to move on the root locus and
therefore, this works when the poles of the root locus meet this specification right. So, I just
translate the given specifications into � and ω n , these � and ω n translate to appropriate
locations on the root locus or all or on the complex plane, if by adjusting the gain I could
arrive at those complex planes, then okay I did, I just say this is my controller with that value
of gain K , if not well we need to something else right.

(Refer Slide Time: 33:33)

Okay, so what else does the proportional control do? right? So, again just depends on the
controller adjusts according to the error and so on, what happens in the bode plot right. In the
bode plot, the gain K would just shift the magnitude plot, this doesn’t contribute to the
phase, the phase remains the same; the magnitude could go like up and down depending on
whatever is the value of gain K , okay. Again so this is the case of an unstable system

547
could be easily seen here and we could see that I know by adjusting the gain I could actually
push the bode plot or the system to the verge of stability.

548
 Control Engineering
Dr. Ramkrishna Pasumarthy
Department of Electrical Engineering
Indian Institute of Technology, Madras

Module - 07
Lecture – 02
Part - 02
Basics of control design – Proportional, Integral, Derivative actions

(Refer Slide Time: 00:22)

So, the proportional control action now right. So, let me start with a first-order system, a

K
proportional controller with the closed-loop transfer function is , okay. What is
Ts+1+ K
the error? If the reference is a step function right so, the steady-state error by doing all these

1
calculations we had done this before also is right, the first observation is that the
K +1
steady-state error is non-zero; I would ideally want it to be 0, but it is not zero in this case.

Now, what does this proportional control do? Well, the first observation is that the control or
the proportional controller it results in improving the time constant, which means the
response is faster on the flip side there is a steady-state error right. Now what we could do is
or just by observing this say that well my steady-state error can be made smaller and smaller
by increasing the value of the gain K , but this may not always be practically possible or
even desirable because I might amplify unwanted signals and so on right.

549
In some cases of higher-order plants which I’ll show you in an example increasing the gain
could also have the tendency to destabilize the system right for example, if I were to look at
this root locus for I just keep on increasing the gain, but I just notice that I am now on the
verge of instability right and that is not desirable ok.

So, the proportional controller it improves the response of the system, it also helps improve
or lessen the steady-state error, but only up to some factor right I just cannot get it to 0 ok.

(Refer Slide Time: 02:27)

Now let’s see what a pure integral control action do? and this actually go very well with the

1
terms, right? So, integral control action is well I have a K , I divide by s ; has an
s
integral action right? The Laplace transform of an integral of a signal just was translated to

1
of the original signal. Okay so, the output is something like this. So, now, what is the
s
error? The error is well something nice happened here right. So, if I just go through all these
computations the steady-state error becomes 0, this we also can see in when we did the error
constant analysis, right? Type−I system if I am tracking a step the steady-state error is
always 0.

So, with addition of the integrator as a controller, well the steady-state error is 0, but the
system order has increased from 1 to 2 and if we keep on dealing with higher and higher-
order systems increasing a pole or adding an additional pole at the origin which is essentially

550
an integrator might lead the system to the verge of instability. So, two things to observe here
right first is proportional control, it improves my transient response it also helps in reducing
the steady-state error, but doesn’t eliminate it completely. Integral action well, I add an
integrator to the system which means I add a pole at the origin it helps in reducing the steady-
state error to 0, in this particular example which we are talking about or in this particular case
of what is now the type−I system.

(Refer Slide Time: 04:23)

Okay, higher-order systems it may not always be desirable right that I just keep on buying
integrators and dumping into the system, it may not work and we’ll see why it may not work.
So, if I take this example here well, I need to see well how the system looks like now right.
So, so total G(s) , if I call this as the open-loop transfer function; G ( s )=¿

K
2 ok.
s (Js+ b)

Now, okay let me just say that K=1 , J =1 , b=1 and I see well when is the closed-
loop system stable, well I can actually do the Routh analysis, but I’ll do the root locus right
because now we are fairly familiar with doing that. Okay so, where are the poles? Well, there
is a pole here and a pole here, right? I’ll write s=0 , in the σ and jω axis or the
complex plane, there is another pole at s=−1 , now I think we now remember the rules of
the root locus by heart. So, this guy will go here and we see that this guy is actually go this
way, we also did this particular example while you were doing the root locus plots and you

551
see that just adding an integrator here doesn’t help me much because the system is now
becoming unstable.

(Refer Slide Time: 05:57)

Okay so, we have to be careful in in that way, okay now what is the effect of disturbances
right? When, when we use these kind of controllers. So, again let me take this kind of plant
here which is a rotational element with a moment of inertia J , and some frictional
coefficient b . Now let’s say there is some disturbance D( s) and I’ll see what is the
effect of this disturbance on the response of my system and can I get rid of those disturbances
by adjusting the value of K ok.

552
(Refer Slide Time: 06:38)

So, let’s say well I am just interested at the moment in a step disturbance. So, I want to see

Y (s)
what is , means the effect of the disturbance on the output Y ( s) , I can just do this
D (s)
by setting R ( s )=0 . We learned this, how to derive this transfer function much earlier in
the course. So, so now, what happens is well E( s) because of D(s) just becomes the
negative of Y ( s) , right. So, I am just. So, Y ( s) goes here there is no R( s) . So,
E( s) is simply the negative of Y ( s) .

So, I can also see what is the effect of the disturbance on the error signal? This is straight
forward to see because what we do all; when we do this, we just compute this with
R ( s )=0 , R ( s )=0 there is nothing here, Y ( s) enters here. what is E( s) ? E( s)
is 0 minus the signal what is coming here. So, E( s) is the negative of Y ( s) ok. Okay
so, what is the steady-state error because of this disturbance? Well, the steady-state error, I

−T
just do all the formulas, use the final value theorem. And, I get that this is . Now I can
K
take care of this steady-state error again by increasing the value of gain K , okay? Now,
what is the effect of this gain K , I am keeping on increasing gain K .

Now, where does the gain sit in my characteristic equation, the gain is sitting here. So, K
2 2
now has some direct relation to ω n , right? s +2 ζ ωn s +ω n right, if I just normalize it

553
properly. Okay so, if K increases the system becomes more and more oscillatory right and
for a higher-order plants as we saw earlier it can also go to the verge of instability ok.

(Refer Slide Time: 08:43)

So, next, we see what is the effect of integral action, and we have this integral action in-
addition to the proportional action and therefore, we call this a proportional plus integral

Y (s)
control action. Okay so, again , I could easily compute it to be this way, and I see
D (s)
that you know earlier we had stated that increase or introducing an integral action increases
the order of the system. So, now, I have three poles, okay. With, R ( s )=0 the effect of the
disturbance on the error becomes something like this ok.

554
(Refer Slide Time: 09:23)

So, what are things which I have in my control? from the control that is Ti and I have a
K here. Okay now, I might be tempted directly to use the steady-state error formula from
the final value theorem to see what is a steady-state error to step disturbance. Okay so, in
order to compute what is the effect of the disturbance on the steady-state error, while we are
using a proportional plus integral control, we might be tempted to use the final value theorem
directly, but only thing which we need to be careful is that when we use the final value
theorem we must first ensure that the system is stable, which means I need to choose the
K and Ti such that the overall closed-loop system is stable or in other words the roots
of the characteristic equation have negative real parts, J and b are given to me. So, I
am; I can just play around with K and T i .

Now, under this assumption that the system is stable or that K and Ti are chosen such
that the system is stable, the steady-state error I just use the formula for the final value
theorem is 0. So, the observation here is that proportional plus integral control action helped
us in eliminating the steady-state error, now this is a steady-state error due to the disturbance.
I’ll just make it a little more explicit here, which means the effect of disturbance can be
nullified in the system. Now, you may ask a question why did I not choose only an integral

555
 K
1 1
control action? Why do I need (
K 1+
Ti s ) , why not only this guy
Tis
or 1 ; well
Tis
the answer may be found in the previous exercise.

So, you know just try to write down things, and then well the answer is very obvious there
right. So, to conclude the, what we can say is that to eliminate the effect of disturbance, we
need both the proportional and the integral terms. The proportional term was not enough as
we saw earlier; adding an integral term alone gives us problems which we found in the
previous exercise. So, the proportional term here ensures the stability while the integral terms
eliminates the steady-state error, and that we could compute from here. Okay now, let’s look
at a derivate control action.

(Refer Slide Time: 12:10)

So, I just have my plant as a rotational element with a moment of inertia J . So, I’ll; let me
say I just start with a proportional control action. So, why do I even need to use a derivative

K
action? So, I just say well put a K here, the closed-loop transfer function is , if
J s2 + K
I plot the root locus it will tell me that the poles are always on the imaginary axis as the value
of the gain K increases from 0 to infinity right.

So, my response of the system will be purely oscillatory right and these oscillations wouldn’t
die down. Now I would like these oscillations to be eliminated right and we’ll see how these

556
oscillations can be eliminated and when I say that the oscillations are to be eliminated it
means I must introduce some damping into the system. So, this system here does not have
2
any damping term, J s +K there is no term corresponding to the coefficient s right,
2 ζ ωn s right.

(Refer Slide Time: 13:29)

So, the oscillations can be damped if we incorporate a derivative term in the controller, and
incorporation of a derivative term leads to introducing the damping term in the system. I’ll
shortly show you how that looks like. Okay so, again take a proportional plus derivative
controller, where the controller now looks like K (1+T d s) , and the output from the

controller Y c (s ) is all of these guys. So, I have y c ( t )=Ke ( t ) + K T d é (t) right, we saw
this while we were discussing the effect of adding a zero right? this is actually like adding a
zero to the system right.

So, you have the error and the derivative of the error also. So, the output now is which was
earlier in the case of the proportional control only depending on the error is now also
proportional to the rate of the error change, é (t) . So, what does this derivative term do?
This derivative term anticipates a large overshoot with a é , how fast it is moving right?
é (t) , tells me how fast my error is moving. So, it anticipates a large overshoot based on
the rate, and when the rate of the change of the error is too high it takes a corrective action
right. So, this é tells me how fast my error is growing right and this effect can be noticed
in the improved damping in the system.

557
(Refer Slide Time: 14:55)

Okay so, what is the effect of adding a derivative or a proportional plus derivative term to the

1
plant which did not have any original damping. So, the output of the system is now
Js 2
well, this K this entire thing in the numerator, now look at the characteristic equation or
the denominator J s 2+ K T d s + K .

(Refer Slide Time: 15:45)

Now, this is just a second-order system and we know that for positive values of J , K
and T d the system is always stable. Now, what did the derivative control do? it just added

558
this extra term here which was absent earlier right, it added a damping term into my
characteristic equation and this damping term is due to the derivative control action. Okay,
now let’s combine all these things; we started with proportional, we started with proportional
plus integral control and now we, later on, saw proportional plus a derivative control. Now,
let’s combine all these things together and see what is the effect of each of these things right.
Okay so, let us look at the effect of P plus D plus I on a second-order unstable plant. So, let

1
me say that I am starting with a plant which looks like this, G( s)=¿ 2 . And
( s −b )
for b>0 , the system is always unstable because my poles are ± √b .

Okay so, the first thing right. So, in an earlier slide, I said proportional control helps in you
know improving the stability of the system, let’s verify if it’s true or not. So, I add a Kp

here and with this K p , I am just calling this with the you know subscript p , because
I’m just dealing with a proportional control action. With this proportional control, the closed-

Kp
loop transfer function is now this one, . Okay now, is the system stable? Well,
s 2−b+K p
at best it could be marginally stable right because there is no damping term here right, but at
least starting from this unstable system I could at least get it to a marginally stable
configuration by appropriate choice of the value of Kp .

(Refer Slide Time: 17:19)

559
Okay, now let us add a derivative term, the derivative term well what we saw will add some
kind of a damping to the system. So, the proportional control got an unstable system to
become marginally stable, adding a derivative term will actually make it what we call also the
asymptotically stable system right.

Y (s)
So, now the response has an additional damping term in the system, and for Kp
R (s)
and Kd ; Kd>0 and K p >b , the system is always stable right and therefore, I can
arbitrarily place the closed-loop poles or given any � and ω n ; I can design the system for

any � and ω n , this is what I meant by pole placement, right. Because the poles are decided
on the values of this, I can place this wherever I want right. Based on appropriate choices of
Kd and K p . Okay so, this K p >b is also necessary here right when this K p >b
then I have marginal stability.

(Refer Slide Time: 18:27)

Okay now, what happens to the steady-state error? Well the steady-state error I just look at
the again the formula and I get that well it is I would like it to be 1 because I am tracking this
reference signal well this is not 1 right, therefore, there is a steady-state error in the system.
Now, we’ll see what is the effect of adding an integral term, in one of the earlier slides we
said that adding an integral term eliminates the steady-state error.

Now, let’s add an integral term to this guy. So, I have a Kp the proportional term, the
derivative term, and now I have an integral term. So, the response through this and I compute

560
the steady-state error. So, lim y ( t) this is one right. So, what I’m. So, my R( s) here is
t→∞

a unit step, but I am not tracking a unit step here right. So, my steady-state value is K p−b .
So, I am not computing the steady-state error here, but I am just computing the final value,
the final value should be 1. Because I’m tracking a step signal. It is not 1, but depending on
K p and b . So, again appropriate choices of values of K p we saw earlier also right, I
can keep on increasing the K p to be very large value, but that may not help me all the time
right. Now adding an integral controller gives me the output as 1, which is what I’m tracking
right I’m tracking a unit signal, sorry. I’m tracking a unit step right. So, this adding of an
integral term eliminates my steady-state error or it means that I can actually track a unit step
right.

So, the conclusion is that we have perfect tracking of this unit step, moreover, we also know
that having an integral control action rejects constant disturbance inputs. So, we individually
saw what are the effects of adding a P term the proportional controller, a derivative term,
useful in adding damping to the system or improving the transient performance, and an
integral controller which is useful in improving the steady-state performance in terms of the
steady-state error.

(Refer Slide Time: 20:47)

So, just to summarise what we have learnt in terms of the proportional derivative and integral
control, is well, we could start with an unstable system, second-order and we could stabilize it
right. We could also place the poles arbitrarily or in a way given any choice of � and ω n or

561
any performance specifications, I could choose my gains Kp and Kd appropriately so
that I achieve those performance objectives the system has perfect tracking because of the
integral control and what integral control also does it rejects constant disturbance inputs right.

(Refer Slide Time: 21:27)

So, to summarise we defined what are dominant poles and zeros, we by just you know using
MATLAB plots we saw what were the effect of adding poles and zeros to the system right
that was for second-order systems, we also saw the same thing when I add poles and zeros to
an open-loop system, what is a consequence on the closed-loop system and we defined three
basic control actions proportional, integral and derivative control action. So, in the next
lecture, we will see some issues in implementing PID controllers.

So, this stands for proportional integrative and derivative controllers and we will see slowly
define what are lead and lag compensators, what could be the performance specifications in
the time and frequency domains, and the correlation between these two specifications if any
ok.

Thank you.

562
 Control Engineering
Dr. Ramkrishna Pasumarthy
Department of Electrical Engineering
Indian Institute of Technology, Madras

Module - 07
Lecture – 03
Tutorial

(Refer Slide Time: 00:18)

So last time we had learnt the basics of PID controller or what was essentially like a
proportional component, an integral component and a derivative component right. So, before
we go into further analysis let’s do some problems to build up a case for each of these
components is are they useful are they not always useful is it sometimes not very you know
wise enough widely add a component. So, that it reduces my steady state error and so on. So,
we will slowly build that up with some, with some motivating problems and then go back to
analysis of that. So, I start with a simple looking example, which is like a pendulum with no
damping and I will tell you why this represents a pendulum with no damping and then you
have a controller and we expect this controller to do a set of things for us right.

So, you have a reference input and then you have the output, the disturbance and the standard
feedback configuration right. So, this little block here, if I do the block reduction; this block

1
would just be a series block of the form 2 , right and this has the imaginary roots
(s +K )

563
 ± √K with a you know with a j right and that response is oscillatory, and the response
is always oscillatory and therefore, I call it’s a system with no damping.

And these are the dynamics of a simple pendulum right we did in one of our very first
lectures. So, first question which we would like to answer is what conditions must D(s)
satisfy so that the system can track a ramp reference. So, what should be are there
specifications on the steady state error. Well the specification is my tracks are ramp reference
with constant steady state error ok. So, first thing we would like to do is to find out transfer
function right, I will not spend time of this on this because we know how to do this now. So,

D( s)
this would be 2 .
s + D ( s) + K

Y (s)
So, this should be and similarly Y (s) or the effect of disturbance on the output is
R (s)

1
another transfer function of the form 2 . Now what does the problem require
s + D ( s) + K
me to do? Well, it requires me to find what must; what are the conditions on D(s) , if not
the exact nature of D( s) , so that the system tracks the ramp with a constant steady state
error. This is very similar to what we did while finding out the error constants right at type
one system can track a velocity ramp signal with certain steady state error and so on where a
little table for that right ok.

So, first thing is, what is the error? The error is the reference minus the actual output. So, in

2
s +K 1
this case the error would simply turn out to be R( s) ok.
2
s + D ( s) + K s2

So, what do; we do need that the steady state error which is lim s E(s) . E( s) is the
s →0

[ ]
2
s +K 1
entire guy here, that is 2 2 , this thing should be a constant which means
s + D (s )+ K s
well. So, this some usual cancellations.

So, this means that limit. So, here I put a s=0 , there is constant here anyways. So, if I just
2
look at this term here, the lim s(s + D ( s ) + K ) should be a constant, now sorry lim ¿
s →0 s →0

564
here. So, the first guy goes to 0, third guy well it should also go to 0, because I cannot change

K , now let us see what it is. So, this means that lim sD ( s )=constant . So, what should
s →0

be the nature of D(s) such that sD ( s ) is a constant, well D(s) should be such that it

D ' ( s) lim ¿
should be some right, in this case is this cancel out and then I just take the
s s →0

this will be some constant ok.

So, the condition on D(s) such that these things are satisfied is that D(s) must have a
pole at the origin right. So, this guy must have a pole at the origin ok.

(Refer Slide Time: 07:19)

This answers the first part; now in the second part for the D( s) which we obtained in part
one; which has a pole at the origin, now for that D(s) that stabilises the system, right.
Find the class of disturbance w (t) or W (s) that the system can reject with 0 steady
state error. So, I need to find well for this D(s) which had a pole at the origin, I also
assume that this it stabilises the system what are the class of the disturbance signal that I can

Y (s)
reject. So, I will use this guy here right. So, is this guy, ok.
W (s )

So, what I would want is lim s

s →0 [ 2
1
s + D (s )+K ] , so this is my Y ( s) right or Y ss if

I would call, right. At this steady state there should be no effect of the disturbance on the

565
output which means the W (s) should not affect this Y ss , at steady state and that is only
because of W ( s) , I am not really considering what is happening with R( s) at the

moment. So, this would mean that, this is lim s

s →0 [ 2
1
] 1
. L , let me assume that
s +D (s )+K s
W is of some form s L right; could be a step, could be a ramp or parabolic kind of thing
and so on. So, I want this to be 0 . Now, when is this 0 ? This happens if and only if
L−1
lim s D ( s )=∞ , right. So, why do I say this? Well, look at the denominator here.
s →0

L−1
So, this would cancel out, this will be s , this L−1 would go here, that term would
L−1
become 0 , Ks would also become 0 , when I substitute s=0 . So, for this to
go to infinity, this guy should sorry this; for this effect to go to 0 there should be infinity

in the denominator and that is possible when lim s L−1 D ( s )=∞ , and this happens when
s →0

L=1 because of; let’s see what happens when L=1 . So, when L=1 this guy is

'
1−1 D (s )
s , what is my D (s ) ? D ( s ) is ok.
s

(Refer Slide Time: 11:35)

So, this guy just goes away and lim ¿ , I know this D ( s ) has a pole at the origin and this
s →0

guy goes to infinity right. So, in the third part: show that although a PI controller satisfies
conditions of part one of the problem, it will not yield a stable closed loop system. So,

566
 KI
D ( s ) is a PI controller which means it looks something like this K p +¿ or I can
s

s K p+ K I
just write it as right. So, this satisfies the condition of the problem one that
s
D ( s ) should have a pole at the origin, and if D ( s ) has a pole at the origin then I. So, the
condition one the part one of the of the problem said find the conditions on D ( s ) such that
the error, when the reference is a ramp is a constant, this is satisfied right? D (s ) satisfies
the conditions of part a ok.

Now, with this D (s ) let us find out what is Y ( s ) or the transfer function of the system.

D (s ) KI
Well this becomes 2 , what is my D ( s ) ? My D ( s )=K p +¿ here
s + D ( s) + K s

2 KI
and again I have the denominator s + K p +¿ + K . So, this would; okay I just do all
s

the computations K p s+ K I , here I will have s 3 +s ( K p + K ) + K I . Now is this system

stable? How do I find out? I find out by the rotors of the characteristic equation or the poles
of the system.

Now, look at this guy right. So, s 3 +s ( K p + K ) + K I =0 , and I do not even need to see the
Routh table right to say that this is always unstable right for whatever values of K , you
may just want to do it by the Routh table, but you see that there is no s 2 term here, s2
term is missing and as long as that is missing you will never have a closed loop system that is
stable, as a little exercise you can just convince yourself by drawing the Routh table.

So, the conclusion here is that although D (s ) satisfies the conditions of problem number
one with a PI controller or which has a pole at the origin, it leads to a closed loop system
which is unstable and therefore, even though a PI controller has the effect of reducing the
steady state error, if this was not true in problem number one this would never be able to
track a ramp right. So, with by putting a pole at the origin I see that well I track the ramp, but
with some error right, but then the closed loop turns to be unstable.

So, these are little things which we need to be careful while we design PI controllers or use
them blindly. So, the conclusion is even though this theoretically satisfies it will be
meaningless because my system is unstable. So, this saying that D (s) has a pole at the

567
origin means nothing because the system is unstable. For any of these analysis you need the
closed loop system first to be stable it is not stable therefore, this is not true; it could be true

s+ 1
for some other D (s ) which has a pole here and it could be some that could be
s+ 2
true, that could lead to a stable D (s) right and here I am not really talking of a stabilising
D (s ) right, I am just saying condition one. The second part of the problem assume
D (s ) to be stabilising, if D ( s ) was not stabilising there I do not need to really care
about disturbances also right, only assumption in the second part was D (s ) has a pole at
the origin and it has a stabilising affect.

KI
So, you see that just adding a K p +¿ or a PI control does not have a stabilising
s
affect. In this case and it could turn out that’s in any higher order cases this could be true. So,
we need to be careful by before looking at the problem statement of improving the steady
state performance through a PI controller is always comes with a bit of warning, that you
should look at the closed loop system to be stable ok.

(Refer Slide Time: 17:44)

Okay, next is, now part 3 told me that PI was not a very wise decision or a wise choice. Now
can I use a PID control? So, a general structure of the PID control would be ok.

568
(Refer Slide Time: 18:03)

You have Kp proportional term, the integral term and the derivative term, in which case

Y (s)
the closed loop transfer function becomes, ok I will just skip the computations I just
R (s)

K D s 2+ K p s+ K I
write down the end result, . So, this gives me a little bit of hope
s 3+ K D s2 + ( K p + K ) s+ K I

because there the s 2 ≠ 0 and D ( s ) still has a pole at the origin ok.

So, the part one is satisfied, now let’s see what happens with this one. So, I can just draw the
Routh table I have s
3
terms as 1, K p+ K ; s
2
terms KD and K I ; then I could

1
K D ( K P + K )−K I 0
compute the s terms accordingly as and a 0 here, and s
KD
would simply be K I . Now this, this is positive, this is positive, this is positive, and I can

choose my KP and KD such that this is true. So, just find out. So, this is table for
K D >0 and K I > 0 right, of course, K P is always greater than 0 ok.

So, this is very straight forward to find out. So, KI>0 , K D >0 ; if these two are true,
you plug it in here, and you will find this also to be greater than 0 right. So, this the
additional condition we need to impose is K D (K P+ K)> K I , right for the system to be
stable right. The second one would mean K D >0 , K I > 0 , and this is third one. So, as

569
stated earlier, this guy has an effect on the steady state performance of the system and the
derivative component has an effect on the transient performance of the system right. So, this
is a very basic illustration of when to use PI, when to use PID. Again not a very general
answer it is a just special case of it right of why we should be careful, why while we are using
PI controllers we will see a little more different types of problems.

(Refer Slide Time: 21:44)

In the second problem, well let’s say it is a very simple model of automobile speed control
using a PI control, when there is disturbance also right. So, and we will see what all we could
do with this right. So, first I would say part a we will do is, with zero reference velocity input
which means this guy V c =0 , find the transfer function relating the output speed; let me
call this V , relating the output speed V to the wind disturbance W here well. I can

V ( s)
it is not really a big deal assuming V to be 0 , I can easily compute this as
W (s )

ms
2 ok. So, this is straight forward I will not go into the details of doing
s + m K 3 s+ m K 1 K 2
it.

570
(Refer Slide Time: 23:19)

Okay, in the second part, what is a steady state response of V if w is a unit ramp again still

keeping V c =0 . So, well typical formula V ss =lim s V (s) , what is V (s) ? This is
s→0

ms 1
V ss =lim s 2 2 ok.
s→0 s +m K 3 s+m K 1 K 2 s

1
So, this will turn out to be. So, this, this well this will go away I am just left with .
K 1 K2
Third part what is the type of this system? May be, with respect to some reference signals.

[ ]
K1 K2 m
s(s+ m K 3 )
So, we look at in terms of E=V c −V =1−¿ and usual; the way we
K1 K 2m
1+
s (s +m K 3)

compute the steady state error is through the final value theorem lim s E(s) . So, here the
s →0

position error, okay I’ll skip the computations lim G ( s )=∞ . So, the steady state error with
s →0

K1 K2 K3
a step input is 0 right. Similarly, K V = or e ss (ramp)= right, which
K3 K 1 K2
means just a type one system that could be easily computed.

571
(Refer Slide Time: 27:06)

So, in the third and the last problem we deal with again a very simple model of a satellite
altitude control. So, with J , being the moment of inertia. So, we have a certain reference
altitude and then you have the actual θ right ok.

So, and then as usual W is my disturbance, disturbance torque; in the first part with a
proportional controller, we’ll slowly see the effects of P, PI, PD, PID and see what they
mean with respect to reference tracking and even disturbance rejection. So, first thing is to
start with the proportional controller, which means my D(s) is simple a constant KP .
can I stabilise this system right. So, can I stabilise. So, let’s say so for some of this purposes I
would sometimes assume that this H r , H y =1 even though nothing will change even if
they are non-unity. So, when D( s) is just KP then my characteristic equation is

D KP
1+¿ 2 assuming this to be true; which means 1+¿ ¿0 or
Js J s2

s=± j
√ KP
J
,yes with the; since this will be in the imaginary axis as KP and J are

always greater than 0 ok.

So, this system will always have poles no matter whatever I do with the KP the poles will
just lie on the imaginary axis right and therefore, this system is not stable; at best it can be

572
called as marginally stable. So, with just a proportional controller I cannot stabilise the
system this is a first thing ok.

(Refer Slide Time: 30:35)

What could I do now with a PD controller, in that case well the transfer function

θ(s) ( K p+ K D s )
= 2 , now you look at. So, earlier we just had roots which were
θ r (s) J s +K P + K D s
2
imaginary this had this term J s +K P now have this additional term K D , which
introduces damping to a system and this is a second order system with all constants being
greater than 0 , I know that this system is actually stable now, and by appropriate choice of
KD and K P ; I can play around with the peak overshoot or the settling time the rise
time and so on ok.

So, first thing with a proportional controller, I could not stabilise the system I could just move
the roots of the system along the imaginary axis, I put a derivative term, I add damping to the
system right, in the first case we just D( s) being the proportional controller; I did not
have this damping term, this term corresponds to 2 ζ ωn which means ζ =0 if
K D =0 . There is no damping so, the system will always be oscillatory the second case I
introduce a damping and stabilise the system.

573
 1
So, if my reference is , what is the steady state error? So, for this I will just use some
s
constants to compute these values, that J =10 since it is in the moment of inertia thing.
So, it is being Newton-meter 2
second /¿ radians then H y =H r =1 . So, when I just used

1
that θr =1 or θr ( s )=1 , then when H r= H y the steady state error is 0 , I will
s
leave that as an exercise to you because you can see lots of things happening here right, you
have well if I just consider this I have two poles at the origin ok.

(Refer Slide Time: 33:35)

So, therefore, I can easily track a step signal same case what is steady state error with

θ(s)
disturbance inputs well for that I need to find out what is a transfer function, or the
W (s )
effect of W on θ , then I could compute again this straight forward computation

1 1
2 right and then well if the disturbance is then θss and steady
J s +( K P + K D s ) H y s

1 1
state value of θ where W = is θss =−¿ .
s KPHy

So, this kind of controller does not eliminate disturbances of this form it will get more
2 3
complex when you have s and s and so on. So, what do we observe from here right.
So, again go back to the plant, the plant has two poles at the origin and these two poles were

574
ok to eliminate the steady state error when I was tracking a step reference, but these guys are
not useful enough to reject disturbances, that is seen from here and therefore, a pole at the
origin helps in the reference tracking, but if I were to reduce the effect of disturbances I may
need an integral component here or a pole at the origin for the controller ok.

(Refer Slide Time: 35:41)

So, in the next part we will just see what happens when I use a PI kind of controller. So, what
does this controller do. It is again not very different than what we did earlier right. So, in this
3
with PI control the characteristic equation becomes J s + K P s + K I =0 this is again
assuming this to be true. Now, this is unstable right? I don’t really need to analyse this any
further. So, this with a PI controller I know it’s unstable, now it does not really help me to
compute the steady state error anymore because whatever I do the plant the system is
unstable therefore, just by doing a PI controller and doing a steady state analysis is really
meaningless right.

So, first we always have to check for stability before we do any other analysis even though

the math of that you know of e ss =lim s E (s) will give me a number, but this was valid
s→0

only if this is a valid transfer function, or E( s) is a stable signal if this is unstable then
nothing can be done right. So, this we even did while we were discussing about the final
value theorem right the system must be stable so, as to apply the final value theorem right, so
that we need to be careful of all the time.

575
So, similarly we could also do the effect of you know disturbances right. So, the effect of this
I will leave as an exercise, effect of disturbance with a PI control. Just check this will the;
well I’ll just use the word closed-loop even though it is not the actual closed-loop. The
closed-loop with respect to W stable now with a PI control, right. Just try to solve this it
should be once you know this should be very obvious. So, the last part now is to check again
what happens with a PID control.

(Refer Slide Time: 38:08)

KI
So, PID control again I will have K P + K D s+¿ , I skip all the all the steps and you
s
3 2
know you just have the transfer function of the form J s + K D s + K P s+ K I =0 . To similar
to what we had in one of our earlier examples right and now I can really find out conditions
under which the close loop system is stable right.

Now, again, so again I just leave this for you as an exercise. So, analyse the steady state error
for different inputs may be for step and just for a ramp right. So, I will just give this as an
exercise because it is just a very manual thing. So, what we have learnt right just to
summarise is, we need to be careful when we are adding an integral control. It might seem a
little misleading that it will improve my steady state performance, but it might lead to an
unstable closed-loop system right and then of course, when I see this PID well I just have to
then we should choose the constants KP , K I , and K D such that the closed-loop
system is stable, this can come from the Routh’s table right this we are now very well familiar

576
of how to analyse stability of this right. So, what we will do next is to take this as bit of
background and analyse these kind of things a little more can I directly do this practically are
there any difficulties right.

So, can we relate to what we learnt in terms of the root locus plots, can we relate these things
to what we learn in terms of the Nyquist or the bode plots, can do these guys talk anything
about relate to stability at the moment we just haven’t talked anything about relative stability
right. So, all those things we will analyse once we have understood what adding of these
controllers mean, are they good, are they bad, are they sometimes good, and sometimes bad
ok.

Thank you.

577
 Control Engineering
Dr. Ramkrishna Pasumarthy
Department of Electrical Engineering
Indian Institute of Technology, Madras

Module - 07
Lecture – 04
Basics of control design – Proportional, Integral and Derivative actions

So, in the last lectures, in the last couple of lectures what we saw were designing of basic
controller or compensating elements. The basic elements were a proportional element, a
derivative element, an integral element, or a combination of proportional and a derivative or a
proportional-integral control or a combination of all of these three these elements. And
through some problems, we saw well when can we add a PI, when can we not add a PI, what
are the effects of adding just a proportional controller, what are the effects of adding just a
derivative controller and so on and those problems gave us little idea of what happens
actually when we try to design controllers, to do it blindly without worrying about stability or
not is what PI controller asks us to be careful of right. The derivative control we saw had it is
in the transient response of the system in case when a proportional control is not good enough
to stabilise my system. So, what we will do here is to do a little further more analysis of those
components.

(Refer Slide Time: 01:23)

578
So, what we did so far is well just to sum up. So, we had a proportional element, a derivative
element, and an integral element, the standard thing is just to be; the objective being to track
a certain reference at the output. So, well proportional controller its well, what we saw was it
very very simply simple to implement, but it may not be able to meet some specifications
sometimes it is not even good enough to stabilise the system.

Stability, so when I look at proportional and the derivative control we saw that it actually had
a very good transient response, it could stabilize the systems and then based on the value of
the Kd which added damping into the into a simple second-order system it provides more
control over where I should place my poles of the closed-loop systems. Integral control well,
it had well perfect tracking of constant reference inputs because it has a zero, adds a zero at
the origin and I know that type−1 systems if the reference signal is a step then I could
track it perfectly. And it also had good effect of rejection of constant disturbance inputs where
we saw that if the plant had some zeros at sorry some poles at the origin they were not good
enough to neglect or reduce or reject the effect of disturbance, in that case, we actually
needed to have a controller which had a pole at the origin.

So, for second-order systems all these proportional derivative and integral control they add to
stability. I can also look at where to place the poles, placing the poles I know has lot of effect
on the relative stability of the system, perfect tracking of constant reference inputs and also I
could reject constant disturbances with the help of an integral control.

(Refer Slide Time: 03:27)

579
So, what we will do here is to just get little more understanding of this integral and derivative
actions and what are the practical issues while I implement PID controllers and this would
lead us to some kind of approximation of this PID components called the lead and the lag
compensation right. And then, of course, we will also look at these performance
specifications both in the time and the frequency domain and see how they actually like,
connect to each other.

(Refer Slide Time: 04:01)

So, let’s start with a very basic first-order system with a proportional controller, what do I
usually want to do is to track a reference at the output. So, as long as a> 0 , I know that
this system is stable and it will have a steady-state value of y=0 , right some response like
this. And through a feedback controller, I just want to see if I could try to shift this steady-
state value to 1 right, for example, if I am just; something is lying on the table, if I were to
lift it I will just need some external force right can I just lift it to some 1 meter above the
ground right. So, here as we were through a feedback controller can I try to shift the steady-
state position to one.

So, let’s try to understand does a proportional controller help me achieve this. So, what is the
structure? So I have the reference, I have the output, the error goes via Kp and the
U p ( s )=K p E( s) and this goes as a input to the controller and then generates an output. So,
typically what I would like is e (t) →0 as t→∞ right or this steady-state error should
be zero or which means that this y (t) should track this reference asymptomatically right.

580
(Refer Slide Time: 05:32)

So, let just try to understand this with the help of the block diagram here right. Okay so, let’s
just see with this help of this block diagram just what do I want e (t) →0 as t→∞ or
y (t) to track r (t ) as t → ∞ .Okay, let’s see how this propagates. So, once
e (t) →0 then I will have that this is 0. 5 times 0 , this is also 0 . So, this will be
0 okay. Now, what is this actually is - this actually is computed to be this one which is
actually 0 , now if this is 0 , I go back here and I compare this with this thing, I see well
the error is actually not 0 , right?

So, there is actually a contradiction here I cannot have the steady-state error to be 0 and in
this setting, it could be something else. I am not really worried about that something else. I
am just worried about keeping my steady-state error to 0 or y tracking r
asymptotically right which means that if this was assumed to be true or if this was feasible it
would mean that this actual input to the plant would be 0.5 times 0 and the controller
applies no input right, its natural state is 0 without applying any input I cannot get any
output here right. So, the controller applies no input to the system to maintain the system at
the new steady-state which contradicts, this is contradiction right because if I were to hold
this at some non-zero steady-state the controller must supply some input right the controller
must apply some non-zero input to the plant to maintain the output at some non-zero value or
a non-zero equilibrium if I may call so. Therefore, in this case, the steady-state error cannot
be 0 , we are only worried is the steady-state error 0 , when I have a first-order plant
with a proportional controller in the closed-loop is it tracking a step? well the answer is no.

581
We also know these answers via the time domain analysis, we also know the answers via the
error constants which we had found out.

(Refer Slide Time: 07:37)

Now, let’s say well I am not really worried about having a steady-state error of 0 , but
some number. So, if I take this plant controller configuration then I can compute that the error
is 0.1 right. So, this is again follows the first-order system analysis which were learned in
the earlier course I think module 4 or something like that.

Okay, now let us see how is this consistent. So, there is a steady-state error of 0.1 this
error gets multiplied by 9 and input signal here as 0.9 , and then the output because of
this input will be 0.9 at steady-state. So, I am holding this guy at 0.9 , but to hold this
guy at 0.9 , I am actually having a non-zero input right which is actually also 0.9 right.
So, I can with some adjustment of this guy reduce it reduce the steady-state error, but of
course, for the steady-state error go to 0 this has to go to infinity, and again it will lead to
some other contradictions. The steady-state will error will never go to 0 , but we can
reduce it to some small number depending on the choice of the K p right.

So, the proportional controller it reduces the steady-state error but does not eliminate it
completely. Of course, there are practical issues of noise amplification and things like that
when I choose Kp to be very large it may not be practically feasible sometimes and could
come up with some other issues ok.

582
(Refer Slide Time: 09:04)

So, can I add an integral control? Well, integral control would mean that just add an
integrator here this would increase the type of this system to 1 and I know by the error
constant analysis that type−1 system if I have a step here I will track it perfectly right.

So, what does it mean again in terms of this signals? So, if I compute the error signal then the
error is something like this and it goes to 0 asymptotically. Now, the input has some value
like this right, it has some input value even at steady-state, and based on this input value the

1
output is tracking the reference asymptotically. So, U p this guy is now e( t) or in if I
s
talk in terms of the time domain it is an integral of the error right, this signal you integrate it
and you get U p , earlier it was just getting multiplied if it was 0 it was 0.5 times
0 , if it was 5 , it’s 0.5 times 5 and so on right. So, the controller input is now the
integral of the error signal right. And when I do this what I also observed is the controller
output is non-zero see it is all always non-zero. So, there are some points here, the error goes
to 0 , error goes to 0 and whatever, but the control input is non-zero even though the
error has gone to 0 .

Similarly, I could also show we also did this in some problems earlier that when there is a
constant disturbance which could possibly come from here the error due to disturbance
accumulates and it is used to reject the effect of disturbance. So, in this case, the output of the
controller is the integral of the error so the controller output which is the input to the plant

583
right this guy is non-zero even though the error has gone to 0 , you can see this here right,
here the error has gone to 0 , the input to the plant is non-zero right and then as it goes to
0 there is some steady-state value for the input right. Similar analysis I could also do for
the disturbance rejection as we have seen in some problems earlier. So, is that all is life now
easy that just add an integrator and then steady-state errors are gone well it also comes with
one set of rules.

(Refer Slide Time: 11:31)

So, what are the important points to observe when I have plant with integrator? When the
plant already has an integrator the integral control may not be necessary for 0 steady-state
error that we have also seen in problems earlier that I had a second-order system it was
tracking a step signal perfectly asymptotically right.

So, and we also saw that the integrator in the plant does not help in disturbances rejection
right, and therefore, for this kind of scenarios I would need to have an integral control via the
compensator or the controller.

Now second thing, is called the integral wind up phenomena and this is very very practical
issue right. So, first is well these guys are not always stable, right. I have a pole at the origin
right, and if I just it could at best be marginally stable right. And then not only that when
there is a limit on the actuation capability let’s say my actuation signal is some kind of force
right, to which I cannot just apply infinite force right, if I just buy an equipment which could
give me some force it is bounded by above and below by some values right? That phenomena

584
is usually called as the actuator saturation right. So, in this case, well this is my controller
input. So, the input to the controller from the error it goes via this and this Up is now has
to go through some functions like this because my control action or my actuator has a limit
here which does not allow the input force or the input voltage to go beyond a certain value.

So, what happens here is that I may need if I just you know roughly draw a graph I may need
an input which is somewhere here or an input which is somewhere here to drive the error to
0 or to meet some performance specifications and so on, but well my thing, my controller
via this actuator could just give me you know may be 4 volts even though I need 10
volts right. So, in this case, what happens is that the integral, so there will, of course, an error.
So, I need 10 volts, but all I am getting is 4 volts the error will still keep on
accumulating accumulating accumulating and then the output could go to very large values
right. So, it does not really work in cases where I have some kind of an of limitation in my
actuator capabilities.

So, that also I need to be careful of right? Okay, we will not really discuss on how these kind
of situations will be handled in practice right, but this is good to know that it is, it just that the
integral control action even though it looks extremely beautiful, it looks very simple, just

1
, it eliminates lots of steady-state errors, it rejects disturbances, but it comes with two;
s
you know two of these, this not really drawbacks, but something which we should be careful
of while we are deciding a system.

585
(Refer Slide Time: 14:39)

Also what we could observe is that an integral control leads to sluggish closed-loop response
for a first-order system. What does this mean? Let’s just take well normal proportional

1
controller and a plant as , the root locus will tell me that as K increases this guy
s+ 3
just runs away to infinity. Now, if I add an integral control which means I am adding a pole at
the origin then my root locus is like somewhere here right here and here. So, I am actually
having my dominant say for a; here for a larger value of K and somewhere here if I just
put the same value of K here, I might just end up somewhere here. So, this means that for
that value of K where you know my pole in without the integral action was say at −12
or −10 or somewhere in between this response is faster than just having a poles or this
complex poles at −1.5 ± j something right which means the response has becomes
sluggish which means my poles or what I would now define as the dominant poles are now
closer to the origin and that might. So, here this system might be oscillatory here they are not
oscillatory ok.

So, and then if I look at higher-order plants they may lead to instability as we had seen in the
problems right. So, we just add some cases where I just had J s 2+ s+ 1 and you get some
constants here, right? And no matter what the values of the constants were this system is
always unstable right.

586
(Refer Slide Time: 16:41)

Okay now, what about the original plus derivative control action? The input to the plant now
comes out where it looks ( K p + K d s ) E (s) is K p e ( t )+ K d é (t) , where é (t) the
derivative of the error right. So, what we had seen earlier in the first lecture of this module
was that a derivative control not only acts on the error but also on the rate of change of the
error which improves the relative stability right. So, we are adding a zero to the open-loop
transfer function in the left half plane and this zero has the effect of pulling the root locus to
the left, right? Away from the imaginary axis in the stable region of the more away I am from
the imaginary axis the more I can say that I am more, more and more relatively stable right.

And not only that right the derivative term is in some sense it has some anticipatory effect or
some anticipatory behaviour it predicts how fast the error is changing and the and therefore, it
could take corrective actions much better than if I wouldn’t have this é term.

587
(Refer Slide Time: 17:54)

So, what is the effect of adding this proportional and derivative terms right? We have seen
this again in earlier problems, but it is its good for us to remind us every now and then. So,
this js
2
if it just had a Kp without a K d , I know that it is not stable then the roots
will always be on the imaginary axis, as soon as I add a K d , I have I introduce a damping

term and by proper choice of this K d , I could choose this system is underdamped, over-
damped, critically damped or I can play around with the response of the system right.

Now Kd influences only the s1 term, Kp influences s0 term right, but then well
to have control over the closed-loop pole locations I need two things right. So, the standard

2
ωn
thing was I had 2 2 and this ωn together with ζ characterize the
s +2 ζωn s +ω n
complete behaviour of the closed-loop system right. So, therefore, here both these terms
K p and Kd are necessary to have complete control over the closed-loop poles.

588
(Refer Slide Time: 19:11)

So, in the previous slides, we had said well the derivatives term is in some sense anticipatory
or predictable it can predict what is going to happen in the future and these systems we called
as non-causal systems moreover while defining transfer functions we said. Or we defined a
proper transfer function in such a way that the number of poles should be equal to the number
of zeros or the number of poles should be greater than the number of zeros where the number
of zeros were greater than the number of poles as in this case there is one zero and no pole
this were improper transfer functions and I could not realise or I would not be able to realise
in practice right or I can just cannot just go to the market or say amazon.com and say well
I’m; I was looking for a derivative control that will not happen right.

And amplification of noise when I am looking at a derivative component here and say I have
first you know like where you have frequency noise, the effect of the derivative could be very
bad in terms of the amplification of the noise right. So, the derivative term acts on the rate of
the error right, and therefore, the output is highly susceptible to noise if it is changing or a
higher frequency and this noises can cancel can come either through the measurement sensors
or through the reference signal itself. So, just for practical purposes sometimes it is good for
us to add the derivative control in the feedback loop if I have no noise sensor right. So, I can
then affect, so I can reduce the effect of noise which are evaluating through this right. So, in
this case, the noisy reference signal gets filtered before or it gets filtered in the forward path
before encountering a differentiate it does not the noisy signal does not really affect or
encounter the differentiator right, but of course, still, I have to take care of the noise which is

589
coming from y (s) that is a bit of a trade-off here, but just some issues which we have to
be careful of while dealing with these things in practice.

(Refer Slide Time: 21:24)

So, the point is now well I have this derivative controller I design nice properties, and as I
also told you that cannot buy it in the market, well so, what do I do? Well, can I make an
approximation, so to address both issues, I’ll tell you why I am addressing actually both of
these issues: amplification of noise and I can also take care of this non-causality to address
both these issues, I have an approximation of a PD controller now right. So, in such a way
that as so I just if you see here I am just taking Kd s which is the actual derivative term

s
and I am adding a pole here and I do this just to maintain the DC gain. So, T d (s) is
p
causal I have one pole, one zero. Now and as p→ ∞ , T d ( s)→ K d s and we recover the
original derivative controller again which is non-causal ok.

1
So, what does this guy do? s
+1
, this acts as a low pass filter right as we saw in that the
p
frequency response well it goes something like this right and you have some kind of
bandwidth right a cut off frequency as we had called. So, this multiplicative term acts as a
low pass filter and eliminates high-frequency content before passing through the
differentiator right and usually, noise is a high-frequency signal so there. So, if I just add an

590
additional pole here that will take care of the high-frequency signals right and of course,
again as p→ ∞ we have an low pass filter and we get back a pure differentiator.

So, I have done two things here: I have made this guy to be causal or a realisable transfer
function still comes with its own set of you know it still not done and what we also saw is
that I can place the poles further away from the origin. So, in one of the earlier lectures, we
saw what is the effect of adding a pole to a transfer function what is the effect of adding a
zero to transfer function, if I add poles to further to the left then the effect becomes small and
small right. So, I can smartly choose the location of poles such that the effect is like a
derivative controller, it is causal and it also eliminates this the issues related to amplification
of noise right.

(Refer Slide Time: 23:55)

Now, this brings us to some kind of approximation of this PID controllers which are in text
referred to as lead and lag compensation.

591
(Refer Slide Time: 24:07)

Well just start with a lag compensation which essentially is an approximation of a

proportional plus integral controller, a PI controller well I know provides perfect tracking for
constant reference inputs and helps in rejecting constant external disturbances we said this
repeatedly will keep saying this over and over right. Of course, this is not by itself stable it
has a pole at the origin and so, therefore, well it is not a stable system and we also saw we
have this integrator wind up property and also in some cases we had seen for higher-order
plants that it can actually destabilise the closed-loop system. I add a integrator and I do the
math it says steady-state error is 0 , but say hold on you know the system is actually
unstable you cannot apply the final value theorem blindly.

So, now let’s say, let me say take a condition where small steady-state errors are acceptable
right. In that case, a perfect tracking may not be necessary right? If I look at the cricket ball, I
do this and it goes to, no I do this right it goes to the DRS it’s okay right even if the you know
ball is hitting the stumps and if the margin of error is say 0.01 millimetre I am okay with
that right it will still be the umpires call 3 reds or 3 greens and all those drama which adds up
while we are watching cricket these days, in that case, a some small steady-state errors are
acceptable right. Now these things leads us to the following approximation of a PI controller.

+K I KI+K p s
So, we start with the ideal one which is Kp = , but I just add not a pole
s s
at the origin, but very close to the origin, could be as much as 0.01 with p being very
close to 0 . Now, the design procedure we will see how close to zero it should be right. So,

592
and further well this now looks like a gain K, a zero, and a pole; in such a way that the pole is
very close to the origin and some zero right, this we will see how to design this right where

KI
K=K p and z= right. So, the approximation of PI which is a lag compensator what
Kp
does it do - instead of having a pole at the origin and see this was the pole at the origin, a zero
on the left half plane. Now we have a pole and zero both in the left half plane right and it is a
stable controller right z< 0 , p<0 it is always stable and this approximation is called a
lag compensator.

Now, we may ask ourselves from when you know actually, you are actually you know
playing around too much with a system right you are teaching us something and, but then you
are approximating now we will see is the effect of these approximation disastrous or is this
good enough.

(Refer Slide Time: 27:02)

So, again we will start with the first-order plant, I have this PI controller and the static
position error you know all these I do the error as say well because of this integral action the
steady-state error goes to 0. Now consider the same plant with a lag compensator this was the
ideal one right, which we said has very good properties in addition to some not very good

s+ z
properties too, now if I do this K and I pass it through my plant I do all the error
s+ p

593
 1
analysis and what I get is that the steady-state error is z this p is very small
1+ K
p
0.001 or 0.01 and say this you know so based on these choices I can make this error as
close to 0 as possible right. So, this is not equal to 0 , but this is well as close to 0
as possible. So, that would be the objective right.

The steady-state error will not be 0 with a lag compensator right. However, what is in my
control, I can have K , I have z , and I have p and I can make this as large as
possible to achieve as smaller errors as possible. Now, what is the effect of adding the lag
compensator on the root locus?

(Refer Slide Time: 28:22)

So, let me start with a plant well it just kind of fairly good looking I have two poles and there
is no zero, so everybody goes to infinity, and just by adding a proportional controller well I
see you know this the red line right. So, the poles are always at an, this complex conjugate
poles and then I have a certain steady-state error sorry this I could compute by hand.

Now if add a lag compensator, first let’s see what is the ideal case right the proportional plus
integral controller. I do all these things and get a steady-state error to be 0 which
theoretically I learn; I have been learning since a very long time now. Now if I instead of this,
I do something like this; I still have s +0.15 and I have a pole very very close to the origin,
that’s like 0.015 right and then the dominant closed-loop poles are at this location and

594
some somewhere here right. And then the steady-state error instead of 0 is now 0.038
which is you know a kind of fairly acceptable right and if you look at the root locus it does
not really change much. So, the qualitative behaviour of the system with respect to gain has
not changed much if not little it will change via very small margin, but we still be in a very
safe zone right.

So, I will start with 0 and 0.05 this is right. So, what these plots tell us right, so I have
a plot here with the P and a PI. So, and all the time when we’re talking of PI, we were talking
of perfect tracking right disturbance rejection. So, all these are steady-state behaviour these
are all important add the steady-state, now while I am playing around with a steady-state
behaviour or I am trying to improve the steady-state behaviour does it have any bad influence
on the transient behaviour. So, the first plot which is just adding a proportional controller and
comparing it with a PI controller you see that this section is fairly close enough. So, there is
not too much of compromise on my transient behaviour it will be very small right. However,
now this is the ideal PI controller now I say well this is actually not very good and I need to
actually have go to something called as lag compensator which is an approximation of a PI
controller and here; I say well they are more or less they are very much the same right, this
show so much seem that with this plot I cannot really find out any difference if I zoom in like
a hundred times possibly something would happen here right.

So, first thing is the PI controller has a good effect on the steady-state performance without
compromising too much on the transient behaviour and the PI is close to the lag compensator
and therefore, I can conclude that the lag compensator if I implemented this way it improves
my steady-state performance without compromising too much on the transient behaviour.
That’s what that we will keep in mind whenever we are doing a steady-state error design for a
lag compensator design.

595
(Refer Slide Time: 31:34)

Now, having found a causal approximation for a derivative controller and we would again
like to do something similar as we did for the PI controller right. So, under a similar

K (s + z)
approximation, we will again , where again K here I can find what are the
( s+ p)
constants, z and so on. Now again this is causal, this is stable and again this is less
susceptible to high-frequency noise right because it is not a pure derivative term right. So,
both the pole and zero are in the left half plane with a pole further away from the origin than
the zero. Now, we will really see why this is true right. So, here it’s so, even though both the

K (s + z)
structures were the same , here I also had the same structure right it really
(s+ p)
depends then on the choice of the poles and zeros to decide which one is the lead
compensator which one is the lag compensator and I will shortly tell you that.

596
(Refer Slide Time: 32:47)

But before that let’s again analyse what is the effect of adding a lead compensator on the root
locus. So, what we saw earlier was adding a derivative component significantly improved my
transient response without possibly much changing in the steady-state response.

1
So, start with for which I saw that all the time that my poles, how much ever I change
s2
the gain they are on the imaginary axis. Now adding a proportional derivative control which
is I am just adding zero, it just has the effect, it just pulls the root locus to the left-right. So,
the poles for a certain gain which were here, sorry which were at this location have now
becomes somewhere here right approximately right. So, through this PD control and a smart
choice of K and z , I can place my poles anywhere on this circle or the circle could be
different for different kind of stuff right.

Now, look at this thing right. So, this is desirable right that the root locus experiences a
significant shift to the left, and shift to the left is required because I am interested in
improving the transient performance. Now, this PD controller how does it compare to when I
just look at a lead compensator which is an approximation, well say I have a, say a zero at
−1 here and a pole significantly further right, now when I do this you will only see that
there is not much difference right the green and the blue are fairly close enough to each other
and again with some little adjustments here I can get my closed-loop here we to perform in a
certain manner right.

597
So, what we conclude here is that the ideal PD controller it affects or it improves the transient
performance via this kind of a behaviour that it pulls the root locus to the left it also helps in
increasing the relative stability, but this approximation also does more or less same kind of a
job and this is realisable, this possibly is not this is never realisable right and you can say well
then what does this pole do right. So, the only significant difference here is I have a s+z ,
but I have a pole. So, the reason I had a pole very further away is that the effect of pole, as it
goes further and further to the left it reduces right as we have seen in the first lecture.

(Refer Slide Time: 35:38)

So, I will add a pole to make it a proper transfer function, but I will keep it far enough in such
a way that it does not really affect my behaviour of the system right. So, both compensators
now have this structure s+z it has a pole, it has a zero, it has a gain right. So, when
z> p then it is a lag compensator. If I were just to write down in a very; you know very
nice way. So, the lag compensator should have an integral action I just cannot have a pole at
the origin. So, I place it somewhere very close by, and then the zero would be somewhere
here right. So, that the zero is to the left of the pole for a lag compensation, what did we see
for the lead compensation is that if I add a zero here, it pulls the root locus to the left which is
nice; which is I want right. It improves the transient response. But to make it realisable I have
to add a pole. So, I add a pole further to the left which means the effect of this on the
transients will be as minimal as possible right and then of course, when we will see why these
are called lead and lag right. So, I will skip this discussion when we actually draw the bode
plots of this compensators.

598
(Refer Slide Time: 36:49)

So, having had an approximation for both PD and PI, well will there exist what about PID
controllers right? So, this problem is solved by using lead and lag compensators in cascade to
each other. So, as to approximate a PID control action right.

(Refer Slide Time: 37:11)

So, after all these design approximations the lead and the lag, and the PID being a cascade of
a lead and a lag how do we relate this to performance specifications. So, while we started our
time response analysis we said that the performance was based on the overshoot it was based
on the rise time, settling time, steady-state error and so on, then we also had an equivalent

599
frequency domain analysis where we talked of resonance frequency, we talked of bandwidth,
and how are all these related to each other and how are all these related to the design
specifications and how could we do this via a lead and a lag. So, before we do that we will
just compare the performance specifications both in the time and the frequency domain and
see if they are actually related to each other. So, so what we have done so far is from the
description and specification of the closed-loop in terms of the overshoot as I said the rise
time, settling time, steady-state error and these are very nice to visualise right.

(Refer Slide Time: 37:54)

And this actually came across as finding all these, all these were like second-order system
analysis right. So, all these were nice when we say there are several poles in the system, but if
there is dominant second-order pole maybe all other guys lying further to the left and there
will some zeros here, zeros here. So, this is what detected my behaviour of the closed-loop
system right. So, all those analyses could be just made based on this dominant closed-loop
system right.

So, however, we did not really talk of; talk much of relative stability noise rejection and this
were nicer to understand in the frequency domain in terms of what we saw when I actually
add an integral component or a pole in my what was said the lead compensator it had a good
effect; two good effects - one was it made that derivative controller look a little more
realisable it also had the effect of attenuation of higher frequency signals which are
essentially the noise signals right. So, and these are in some sense better to understand in the

600
frequency domain and therefore, we see sometimes these parameters or the design
specifications are well, to loosely speaking simpler than the time domain. Simpler not in
terms of that the problems which I gave in frequency domain, they are simpler than the time
domain. It will actually turn out to be the reverse that you know this might need a little more
you know trial and error or some approximations than the design data domain. So, simplicity
comes from this from one or two parameters, I can actually get lots of information right.

So, before we do that we just find out is there a correlation between these parameters such as
the rise time, overshoot to it’s; to the frequency domain parameters which were the resonant
peak, the resonant frequency, bandwidth, the gain margin, phase margin, and so on. Again all
over analysis for obvious reasons of the dominant pole behaviour was or we tried to find this
correlation for second-order systems right this correlation holds very nicely when the closed-
loop systems has a pair of dominant poles right. And for higher-order systems where there is
possibly no concept of this domination poles, we may have to do some better simulations to
achieve the results, but for our analysis purpose we just restrict ourselves to the dominant
poles and that also gives us lots of rich information. So, starting with a second-order system I
know all these formulas now, well possibly by heart right. So, peak overshoot and the settling
time, the rise time and I have this error constants Kp , Kv , K a , and so on right.

(Refer Slide Time: 40:40)

601
(Refer Slide Time: 40:54)

And at the same time I also have if I have a frequency response like this, this is Bode diagram
of a typical second-order transfer function, I will have a resonant peak. So, this is the
resonant peak occurring at some resonant frequency, and these formulas we derived earlier. It
will also have a bandwidth right, this one of this is an ugly looking expression, but I will just
quickly tell you how to do it by hand.

(Refer Slide Time: 41:30)

602
 2
ωn
So, I start with a second-order system which is 2 2 , then I do the sinusoidal
s +2 ζ ω n s+ ωn

2
ωn
version of it, the T ( j ω )= 2 2 . So, when I define the bandwidth I look at
−ω +ω n + j 2 ζω ωn
this thing right. So, at which the magnitude is about 0.7 times the magnitude at a here. So,

1
the magnitude of this guy |T ( jω )|=¿ , where u I had defined
√( 1−u ) +4 ζ u
2 2 2 2

ω
earlier it also to be right. So, first is how do you define the bandwidth or even the cut
ωn
off frequency right. So, this is the frequency at which the magnitude of the response now in
terms of the frequency response, not the time response is 3 db or below right and the
bandwidth isn’t defined as this 0<ω <ω b , right.

So, how do we find this? we just looked at; take the transfer function and evaluate it to

1
in the absolute scale or in the log scale we say 20 log |T ( jω )|=−3 and I will just
√2
keep the computations because it is just kind of the business very manual right. You can just
do it by yourself. And then I can say that the bandwidth is related to ζ and ω n via this
you know kind of ugly looking expressions square root inside the square root and so on.

So, I will not ask you to differentiate this or anything like that, but we will just try to not even
remember this formula only to show that there exists a relationship between the bandwidth,
which is essentially a frequency domain concept which relates directly to ωn and ζ
which came from the time domain specifications right. For a system to track inputs quickly it
must have a large bandwidth this again comes from here right what is bandwidth is
depending upon ωn and ζ right. And however, sometimes it will be it is not desirable
from the noise perspective right because if I have a higher bandwidth as I saw, when I were
just having a pure derivative controller it was an all-pass filter and it had the drawback of not
dealing very well with high-frequency noise signals right. So in practice, we will actually
look at trade-offs between good and bad right.

603
(Refer Slide Time: 44:59)

Now, the Mr or the resonant peak it depends only on ζ . So, somewhere here and then
we also have this have the formula for the resonant frequency. So, in this case, well it has an
exact correlation with M p , if I just draw M r . So, look at these two formulas right they

are kind of nice. So, Mr is directly depending on sorry Mp the peak overshoot in the
time domain specifications is directly depending on ζ nothing else. Similarly, the Mr

also has a relation. So, if I find an M r here with gives me a value, I could correspondingly
find a value for M p right.

Now can I do it all the time well let us just see that right resonant peak only exists for these
frequencies right and therefore, whenever I look at the correlation between Mp and just
from here till here from points point till the it is this .707 right, if you just because of this
guy over here? So, for these guys well I can just see that you know there is a direct
correlation between Mp and Mr of course, this is not defined for ζ =0 right. And
then this is really too much of concern not really because the peak overshoot, for ζ being

1
greater than is very small right. So, for a given ζ either the bandwidth or the
√2
resonant frequency can be used to specify ω n via these guys.

Given this guy you know I can compute these things right and therefore, the bandwidth and
the resonance frequency influence the rise time in the step response and of course, I can just
you know based on the formula for the rise time I can compute, when will how is this related

604
to the bandwidth or ω r ; larger the bandwidth smaller is the rise time and so on could just
be computed by just looking at this expression here right. So, this is my bandwidth and this
one.

So, the aim is not to introduce you to complex formulas which are difficult to derive and even
worse to by heart, but just to show that there actually exists a good relation between the time
and the frequency domain things and in your ends an exam whenever you take or if you take
this I will give you all the formulas which are required for you. So, you don’t really need to
need to memorise formulas, you just need to know which how is approximately each of the
formula looks like.

(Refer Slide Time: 47:38)

So, what are what can we conclude or based on all these things right. That exact correlation
holds for second-order system between the time and the frequency domain specifications, for
higher-order systems if there is a pair of dominant poles then we can still use this correlation
right and as said earlier well the frequency domain control design gives us little more
information than in the time domain I will still very carefully use the word simpler right. And
of course, the specification in the frequency domain after design are translated back to time
domain to verify if the time domain requirements are met or not. And again for higher-order
systems, we may need to do designs through extensive simulations, but then these are not
very important words at the moment right. So, basic understanding of the second-order

605
system will give us good amount of intuition while we do this an analysis for higher-order
systems just based on some lots of experiments and simulations.

(Refer Slide Time: 48:45)

So, what we have done so far is to have a qualitative analysis of both the integral and
derivative components. We have also saw what were the practical issues of realisability, well
we saw accurate saturation and so on, and we had some good approximation of this
components called lead and lag compensators. And we saw also how the performance
specifications in time and frequency domain are nicely interchangeable.

So, next, we will see how to implement these or how to design these lead-lag and lag
compensators. Especially find the value of K , find the value of z , find the value of
p . So, this were the important things right when we were realising those lead and lag
compensators. We will do that using the time domain or essentially the root locus technique
and also the Bode plot technique which is essentially a frequency domain technique. So, and
that we will try and use lots of MATLAB, we will plot lots of things for ourselves, we will
ask lots of questions before we come to a conclusion and hopefully, that would be a little
more intuitive while using MATLAB.

Thank you.

606
 Control Engineering
Dr. Ramkrishna Pasumarthy
Department of Electrical Engineering
Indian Institute of Technology, Madras

Module – 08
Design using the Root Locus
Lecture – 01
Introduction to design in the time domain

Hello everybody. So, in this module we will focus on explicit design methods, what we saw
previously was several controller components; like there was a proportional controller, there
was a derivative controller, there was an integral action, and we had analyzed when under
what situations do we need a derivative component, under what situation do we need to add
an integral component. What are the effects of adding zeros to the system, what are the
effects of adding extra poles to the system like for example, in terms of the root locus what
we had seen was adding a zero has the effect of pulling the root locus further to the left, and
the more the root locus is further from the imaginary axis the better the transient response
right?

So, to speak and there are also methods where if I add an integrator or a pole at the origin I
could derive that the steady-state errors go to 0 , but we need to be careful because what
we saw with a help of few examples is just that, just blindly adding integrators might even
lead the closed loop system to be unstable. So, we’ll formalize those things a little more.

(Refer Slide Time: 01:34)

607
So, what we studied earlier was the entire analysis based on dominant closed-loop poles. Say
for example, if I have 10 poles in my system all the analysis could be based on the poles
closest to the origin because that is what really decides the behavior of the system in the long
term because the other poles which are to the left, their response actually dies down pretty
fast. So, for example, if I just take a system which has two poles here some guys sitting here,
here, here and so on. So, the pair of poles will decide how my system behaves more or less
can be approximated by just looking at these two poles; because this transients die-out die-out
pretty quickly because we are looking at something like exponential of some number Kt
(i.e., −Kt
e ); bigger the K faster they decay, smaller the K slower they decay. So,
this actually corresponds to smaller K , and if you see that if they are on the imaginary
axis then these guys never decay. So, the constant oscillation process ok.

So, we also looked at issues in the implementation of the nominal derivative or PID
controllers right. So, in a way that if I just take a P and D controller, we saw that it cannot be
realized physically as it is a non-causal system. So, this will lead us to have slight
modification of these components in terms of physical realizability like for example, I can
never have a perfect integrator right. So, this kind of systems are very difficult to realize in
real world. So, an approximation could be to classify PD controllers as lead compensators,
and PI as lag compensators. Compensators also means controller. So, whenever I say
compensator it also means controller and vice versa right and okay. So, given a system how
are the performance criterion specified, how is the performance criterion specified. It’s
usually specified in terms of some kind of peak overshoots, some kind of settling time, and so
on and this loosely translate with some you know you know the formulas these are
specifications of � and ω n , and I can construct the dominant poles based on this right.

And we see that these are, how these are related both in the time domain and the frequency
domain like there is a direct one to one relationship between the phase margin and the
damping coefficient. So, here we will see explicitly how we construct this lead and lag
compensators, and the tool we will use for our analysis is the root locus technique. So, we
will use make use of the root locus plots, which we learn to design lead and lag compensators
to meet a certain performance specification.

608
(Refer Slide Time: 04:29)

Okay so, in a typical closed-loop system looks like this. So, I’m given a plant G( s) and
suppose G( s) does not behave the way I want, well the constraint is that I may not be able
to alter the plant or it might just be expensive. In some cases, it may also happen that a
proportional controller may not achieve specifications, in that case, we need to add. So, these
are like; like static controllers right just, I’m just increasing the gain.

If increasing a gain does not help me I need to add a dynamic compensator right which has
some poles and zeros right to overcome the shortcomings of the plant, in meeting the desired
specification as simple as tracking a reference signal. So, when I say I’m adding a
compensator, it essentially means that I will have a system with possibly some gain, some
zeros, so on and some poles, and so on. Again, I’ll make sure that the number of poles is
greater than or equal to the number of zeros to maintain the causality condition. So, we had
earlier discussed the effects of adding poles and zeros to the open-loop transfer function right.
So, now, how do I make use of the root locus to achieve this a desired a performance
specification, that’s what we will, we will analyze and all this analyze, analysis will be in
terms of the dominant pole pair.

609
(Refer Slide Time: 06:12)

Okay. So, when I talk of dominant pole pair it is easy to specify this transient performance
through a standard second-order system right. So, how does the second-order system look

ω n2
like? I have the transfer function as and we know what ζ means, ζ
s 2+2 ζ ω n s+ ωn2
is a damping coefficient, which decides the system is underdamped, overdamped, critically
damped and so on; ω n is the natural frequency of the system.

So, a certain step response to the system leads us to these expressions, at the peak overshoot
is directly related to a ζ and the settling time depends on ζ and ω n . Similarly, there
are formulas for the rise time, the peak time, and so on ok.

So, now given M p , given t s , these guys would result in a certain ζ , a certain ω n ,
through these formulas and how would they be related to the dominant poles. So, given a
ζ and ωn can I actually construct dominant poles? Mainly, just by this formula right.
So, how; just the poles are the solutions of this system right. For a typically for an
underdamped case. So, when I do all this design procedure and if I just design my system
based on these numbers here, I must ensure that this design procedure which I follow in terms
of the second-order system these are actually the dominant poles, that these are the poles
which are closest to the jω axis.

610
(Refer Slide Time: 08:00)

Okay so, so first question is there are any dominant poles if I look at it graphically. So, given
a ζ and given a ω n , I can look at this thing, right? So, these are lines which are
constant ζ lines right, and how is this angle θ determined? θ=cos−1 ζ . And these
are lines which have constant frequency, for different frequencies, there has to be circles here
or here and so on and we had seen this in one of our earlier lectures. So, if I’m given a certain
range of ζ and ω n , say my performance specifications if they are say let the peak
overshoot be less than 0.1 or 10 that will give little range of values right, it is not
strictly that it should be 0.1 . similarly, for ω n in terms of the settling time. So, if I have
given some ranges that the ζ should be between this number and this number or the ωn
should between this number, I will have a range here right I can choose any pole within this
blue region right.

So, if I look at specifications in terms of M p and t s this translates to ζ being within

a certain bound right. So, these are like the ζ lines. So, any ζ within this will give me
Mp less than c, c1 right, and this, we can just you can go anywhere all the way till
infinity on left-hand side.

Similarly, if I look at the settling time which is again depending on ζ and ω n ; these are
the lines right, this is ζ , ω n . and then anything here would be this specification. How
both of these are met is through the area here. Say, choose a pole anywhere here right
somewhere here or here it would be these specifications ok.

611
(Refer Slide Time: 10:04)

Okay so, the transient performance specifications are usually in terms of ζ and ω n , and
the steady-state performance specifications are usually in terms of this steady-state error,
depending on if I’m tracking a step or I’m tracking a ramp or a parabolic input. And based on
these specifications and or these definitions and the type of the system we had characterized
several error constants. So, what we concluded from that analysis is that if it is a type−0
system, which means G will not or G(s)Gc (s) will not have a pole at the origin, then it
will be able to track a step input, but with some error and we know how to compute this
K p right, with this simple formula.

Similarly, a type−0 system will never be able to track a ramp signal and parabolic signal
and other higher-order signals. The system is of type−1 which means G(s)Gc ( s) has
a pole at the origin, then it can track a step input with 0 steady-state error. It can track a
ramp with some steady-state error and it will never be able to track a parabolic signal and
similarly for type−2 systems, it is the steady-state error for a step in step reference is
0 , steady-state error for a ramp input is 0 , but for a parabolic input it some infinite
number right.

612
(Refer Slide Time: 11:37)

So, this analysis is what we did much earlier in the course. Okay so, let’s start with an
example to motivate how we are going, how we will build our case to design a controller.

1
So, let’s start with a simple plant and I interconnect it to a controller in the
s (s +1)
standard feedback setting, and the transient specifications are that the peak overshoot should
be at best 10 or at most 10 and the settling time should be less than 5 seconds
right. And the steady-state specification or the; is in terms of the velocity error constant and it
should be less than 10 . So, let’s first translate this into the locations of closed-loop poles
in terms of the dominant poles. Okay so, M p=10 ; I put it into the formula. t s=5
seconds, I put it in the formula and I get this number as ζ =0.59 , and ω n=1.35 .

With these two numbers, I get the location of the dominant closed-loop poles at
−0.8± j 1.0914 . So, the first thing to do here would be to see, look at this in terms of the
closed-loop specifications in terms of the dominant poles. So, first what I would do is see
what is the root locus of the system right. So, this is kind of very straight forward to plot, we

1
have drawn much more complicated plots and this is a point , this is my pole, this is my
2
pole. So, this guy goes here, this guy goes here and then they we will split this way ok.

613
Now, I want my compensator to be such that the closed-loop poles or the dominant poles
should lie on the root locus, this is what the system is if I just look at well. So, this is the root
locus of G(s) , I can just say well can I just add a controller K , proportional controller,
and check if my requirements are met. How do I do this? I just move along the root locus and
see if for any value of K ; for any value of K am I reaching this point or are these
points on the root locus for some value of K . Okay, now does there exist a K ? right.
So, if there exist a K then my Gc (s ) would just be that K right, just a proportional
controller would satisfy that. So, here I can easily see right. So, this line is −0.5 line: this
is −1 , this is −0.5 . And my dominant poles are somewhere here.

So, what I see here is no matter how much I increase or decrease the gain I will never be able
to reach this points. So, the thing is that or the or the concept here is that this point or this
dominant poles should lie on the root locus of the closed-loop system. That is, that is how we
will make use of the root locus plot to design controllers. So, more on this will commence
shortly.

(Refer Slide Time: 15:10)

So, what we should check is that does the root locus of the system pass through the desired
pole locations? the answer is no. Another way to check is also the angle criterion right if I just
do not know how to plot; I do not have Matlab with me I will just check what is the angle
right. So, I’m looking at the given dominant pole, and I will always denote this as Sd .
Okay, now does this lie on the root locus of the closed-loop system? And I know that a good

614
way to check this is at the angle of the compensator or the controller plus the plant should
satisfy this condition ±180 blah blah blah. So, if my controller is just a K right and
then G , I know what is, at which point I’m taking the angle, it’s at say this guy G(Sd ) ,
now this turns out to be −205.85 ° , it’s not 180 ° . Therefore, there will never exist a
gain K such that KG( S d ) is 180 ° . Okay so, just by adjusting the gain we are not
able to meet the desired specifications right.

Okay, another way to find out is just to look at the intersection of the constant ζ and the
constant ω n line, and that intersection turns out to be this one right, just to find graphically
the dominant or the or the desired closed-loop dominant pole pair. And it is easy to check that
these are not on the root locus right? But, what we would need is this points the one in the
reds here, to pass through the root locus of the closed-loop system which is Gc (s )G(s) .
Okay so, which essentially means I want to pull the root locus to the left, and if I remember
what I had learned in module 7 is add a zero; add a zero so that, it will be to the left. Now,
how much; where to place the zero is the first question. Say okay, add a zero; do I really add
this blindly anywhere. So, this is the question which we will try to answer.

(Refer Slide Time: 18:15)

Okay so, before I know where the zero will be placed, I’ll just recollect what I observed
earlier in terms of the effects of adding a zero to a system right. So, okay I just have the

615
 1
standard system where this is my given plant has a root locus like this right.
s (s +2)( s+ 5)
So, and for certain gains, it goes to instability.

Okay, let’s say I add a zero at s=−7 somewhere here, and I see that the root locus is
pulled slightly to the left that okay. I’m actually ensuring stability here for certain values of
K or for certain values greater than this number my system is becoming unstable, here my
system is stable right for all values of gain K or at worse it is marginally stable. However,
this is a system which has a very lower relative stability, that slight changes in system
parameters can actually push this to the right and lead to the verge of instability. Okay so, I’ll
add a zero further to the right, say at s=−3 . Let’s see okay, some you know nicer
behavior here right, at the this is a system which has a bigger stability margin than this one.
Hence finally if I add a zero at s=−1 you see where it has something more, something
much better in terms of the relative stability. Not only that if I say that the transient response
of this system would be much poor compared to this system, and this root locus is to the,
further to the left of this root locus. So, this will have a better transient response.

So the effects of adding a zero is to pull the root locus to the left, to the left here, further to
the left, further to the left. Now, the thing is how much to pull? how much to pull is answered
in a way that you pull it just enough. So, that this S d , sorry a dominant pole lies on the

root locus of G times G( s) or this Gc (s )G(s) .

(Refer Slide Time: 20:21)

616
Okay so, the first thing which we would like to see is, how much do I pull? Well, we pull just
enough such that it meets the angle criterion. At the angle of my compensator evaluated at
S d , angle of my plant evaluated at Sd should be −180 ° okay? Now, what is my
plant? my plant has a pole at the origin and a pole at −1 , okay? So, now I compute the
angle contribution to S d through each of these poles; a pole at 0 , pole at −1 , okay?

S
And I get a certain number. So, the difference. So, what should this (¿¿ d ) do such that the
Gc ¿

S
total is 180 ° , that it should make the required adjustments in the angle of (¿¿ d ) such
G¿

S
that the total is 180 ° . So, let me call this angle as φ right. So, (¿¿ d )=−180° , if the
φ+∠ G ¿

S
angle of (¿¿ d ) by itself is −180 ° , I don’t need to do anything here and the
G¿
compensator need not add an extra angle right.

So, in this example, it turns out that the φ should be 25.8 ° if you just these are very
simple computations. So, this compensator or controller must contribute an extra angle of this
one, and therefore, I call it a lead compensator it adds an angle to the root locus. Now let’s,
let me just assume without really worrying about if I can realize this compensator or not, let
me just do some analysis before I really have a constructive way to, to determine how the
lead compensator looks like. So, let’s say my compensator just is a proportional plus
derivative controller with the same constant right. Now let’s, let me assume that I just put my
zero at some arbitrary location Zc and this is what I have to find out, okay? Now the total
∠ Gc G=−180 ° . So, I just use this little triangle here right. So, this φ is given to me,
φ is this one. Now θ is also known to me which comes from −1
cos ζ , I also know
what is ω n ok.

Now, I know several things here: I know θ , I know ω n , I know this guy, now can I
calculate what is my Z c ? where I just use a very simple trigonometric rule right. So, sin of

617
 ωn Zc ωn
the Z c , over all these numbers would be (i.e., = ) and
sin ( φ ) sin ( π −θ−φ ) sin ( φ )
just this computation. So, θ is known to me, φ is known to me, ω n is known, φ
is known I just get the location of Zc which is 3.0518 somewhere here right. So,
choosing a zero at; now I’m not really worried about what is this K because this K
does not really contribute to the angle right or it does not contribute at all. So, if I place my
Zc at this value then my contribution of ∠ Gc G=−180 ° .

(Refer Slide Time: 23:43)

Okay so, now how does Gc G look like? Gc G looks like this K , the zero which I
added because of the compensator and the poles of the plant. Now what is this K and I
just simply cannot choose any K , this K is then chosen by the magnitude criterion.
Such that the overall magnitude of K times the zero and the poles should be 1 , okay.
and this K can be computed by again looking at what is the gain contribution at Sd .
So, all these are evaluated at s=S d , this is also evaluated at s=S d . So, again these are
just simple computations.

Okay now, let’s see what is happening with the root locus, well the root locus it has a pole
here, a zero at −1 , sorry a pole at 0 , a pole at −1 and the new zero which is the
compensator this is my controller. And I see that the root locus now which was earlier, which
was something like right? It was going up here, down here, has now been pull to the left in
such a way that the closed-loop root locus actually points passes through this point.

618
(Refer Slide Time: 25:02)

Okay so, how does now the plot look like or the response look like? Well, this is my

Gc ( s ) G(s)
compensated system with a closed-loop transfer function , we know how to
1+Gc ( s ) G (s)
derive this by now. The compensated system has a peak overshoot if I compute of 11 and
the settling time of 4.08 second which are more or less what we wanted right?

So, the system has a zero here and this along with the dominant poles. Okay, now let’s see
how close are these two plots first all my specifications are in terms of dominant second-
order this is little redundant I’m writing second order, but say this for completion, dominant
second-order poles right. So, what I designed in terms of ζ , ω n , and Zc and the
appropriate K , were only; was in terms of just taking the response of a standard second-
order system. Okay, now in addition I add the compensators. So, I have in reality this is not
true right. So, my dominant pole analysis is based on this kind of transfer function,

2
ωn
2 2 .
s +2 ζ ω n s+ ωn

Now in addition I have a zero here right. So, I have the plot we’ll consider; will be the plot
will include the dominant second-order poles for sure plus the effect of zero. Now well what
is the effect of zero well there is some not very good thing happening here right if I say if I
just draw a plot of the dominant second-order system which is without considering the effect

619
of zero this is how my design goes right all my specifications are translated to a second-order
system. This is how I would want the look like, like in the black line.

But however, the compensated system has a zero, zero at this number 3.05 , and so on. So,
the compensated system now has the responses like this at the blue line; and therefore, you
see, well there is slight increase in overshoot than what I actually wanted. So, what I design is
for this, we design for this dominant poles thinking that the closed-loop system would
actually have only this kind of dominant poles and the other guys wouldn’t matter much;
however, the actual response has a zero. So, what we, so this is not the completion of the
design process because there is some gap here right, it does not really ensure complete
dominance of the desired pole locations.

So, we would want to select the zero such that these two guys come as close as possible right
and this will be clear in a sec. So, this is still not a constructive procedure to design a lead
compensator, it is just by following steps of what we had learnt earlier. That if I want my
transient performance to improve, I need to pull the root locus to the left, I can pull the root
locus to the left by just adding a zero. Where do I add a zero? it is given by this angle and
sorry the angle criterion and the magnitude criterion. Now, once I add a zero is it really
dominant well you see here, it’s actually close to dominance, but just that no there is a little
error right.

So, the actual system this has 11 . Sorry, this has 10 overshoot, but this guy has 11 .
So, we will try to see how we minimize this gap right. That also we should be careful of.

620
(Refer Slide Time: 29:19)

Okay so, now, we will see. So, the response so far is okay, we looked at dominants second-
order systems, which should say this is, for example, consider a system like this, with a zero
somewhere here, now we will see. So, what we wanted we wanted this plots in blue and black
to be as close as possible. Now are they related in any way to the location of zeros if they are
closer to the origin is it good for me or if they are further away from the origin are they good
for me? So, this is the kind of a response I’m looking at. Okay so, let’s start with this system
and say I’m adding a zero at s=−0.5 ok.

So, this y 1 is the step response of this guy, okay, and y 2 is just adding a derivative right
ý 1 and y is the total response right, this y is a total response now, I’m adding a zero
at −0.5 . And, okay so, now, I add a zero at s=−1 and I see that the red and the black
'
lines actually are coming closer to each other right this. So, the red is the response of G ,
and the black is the response of G and I want to place the zero in such a way that these
two are closer to each other, and if I have pulled further back at s=−8 , then actually see
that well they are, they are very close to each other. So, depending on the location of the
zeros my red and black curves come closer or go further away ok.

Now, another effect of adding this zero is that well there are no; there is no significant change
in the settling time. In addition, to say, say that well this actually; the response here is faster
right? This is slower to respond. The black line is slower, I pull the root locus to the left and I
see the faster response. But faster response will also lead to increase in overshoot right and in

621
the limiting case as a→∞ the contribution of the zero decreases and we get back the
original system behavior, but we really do not want a to be so large enough that it doesn’t
really matter right.

(Refer Slide Time: 31:58).

So, we need to come to a nice trade of what is a good a to select. Okay so, once I know the
effect of adding a zero there is some drawback here that, okay can I just go to the market and
buy a controller which looks like this? Well, the answer is no we had seen earlier because this
is a non-causal system right ok.

So, somebody I think in the message forum also asked, you know what is inverse Laplace
transform of s and where the answer is that it does not exist. So, therefore, to make it
realizable we add a compensating pole in such a way that the effect of adding this pole is just
for construction purposes just for practical purposes, but this guy should just keep quite
should not really try to interfere too much in the process.

So, now, this compensator, so this again the angle of Gc (s ) and angle of G(s) should
always be −180 ° , all those condition should be satisfied. So, if I say that now earlier only
the zero had to contribute certain angle now I have a zero and a pole. So, these two together
should contribute an angle φ right. So, this is my Sd , μ1 is the contribution of; of
contribution of Zc , μ2 is the contribution of Pc . So, this Gc (s ) will now have
angle contribution because of the zero and because of the pole and the total thing should
satisfy this one, okay? Now, this should more should always contribute a positive angle right

622
which means the angle this guy should always be greater than 0 ° okay? So, earlier we see
in the, in the example φ=28 ° , this always contributes a positive angle. And, for it to
contribute the positive angle, the pole should be to the left of zero, I cannot that have a zero
here and I have a pole here, no this will not work for this reason ok

Now, this first conclusion here before we go further is that I just cannot have only a Zc . I
need to place a pole, where do I need to place the pole is to the left of this zero why because
of this condition here positive angle condition right.

(Refer Slide Time: 34:46)

Now, we’ll slowly see okay where should this guy sit and where should this guy sit right?
And, that we will do in the in the next lecture right we’ll, so far what we have learnt is the
effect of adding a zero helps me in transient response or improving the transient response of
the system I cannot realize zero by itself. So, I add a pole to make it physically realizable now
we will try to ask a question where to place this pole’s and zero’s given a certain performance
metric so that we will do in the next lecture.

Thank you.

623
 Control Engineering
Dr. Ramkrishna Pasumarthy
Department of Electrical Engineering
Indian Institute of Technology, Madras

Module - 08
Design using the Root Locus
Lecture - 02
Improvement of the transient response using Lead Compensation

Hello everybody. So, in this module, we will continue with the design of the lead
compensation and see what are the techniques or how do we place the poles and the zeros.
Earlier we had a little formula to place the zero, but we saw that zero by itself is not
realizable. So, we need to add a compensating pole.

(Refer Slide Time: 00:41)

Okay so, let’s start slowly with this analysis, right? So, what we concluded from the previous
lecture is that well this is the compensating zero and I should have a pole for practical reasons
to realize this. And, we also knew why the pole should be to the left of zero. So, what we will
now try to answer is where should the zero be? and how far left should the pole be from the
zero? and that’ll actually lead to some nice exiting observations.

624
(Refer Slide Time: 01:17)

Okay so, before we proceed we’ll see how we can realize or construct the lead compensator
with help of some basic circuit elements. So, it turns out that if I take resistance R1 , R2
, C1 , I apply a input voltage and measure the output voltage across the resistor R2

which are connected in this way. So, Vo - the output voltage by Vi which is; has the
construction of a zero and a pole with these numbers, and these numbers can be computed

R2
directly right. So, where α= and we’ll clearly say this is less than 1 and
R 1+ R 2
τ =R1 C 1 . And this α < 1 ensures that the pole is to the left of zero, okay. So, this is
how we will construct a lead compensator. And this guy you know this guy we will have
some angle contribution, some gain contribution and so on right because there is some
dynamic element here in terms of C1 .

625
(Refer Slide Time: 02:22)

Okay so, now how does this, this look like right? So, if I look at the sinusoidal transfer
function of the lead compensator, it looks something like this. I have α here, I have 1 over
j; 1+ jωτ and 1+ jωατ , just rewriting this in the in the sinusoidal form. So, this guy
goes here, this guy goes here and so on. Okay so, when α<1 for low-frequencies the
compensator provides a gain which is less than 0 db . So, if I look at the bode plot here and
if I were just to draw with the regular α my plot would start somewhere, from not from
0 this is the 0 db line; my plot would start somewhere from here, depending on the
values. Okay, this is not really desirable because a control system is essentially a low pass
filter and I really do not want the low-frequency signals to get attenuated, right? So, these
low-frequency things I should preserve.

1
And therefore, what I do is I just add a gain of and call V i (t) as the output of the
α
controller where my compensator now becomes this one and the α thing is gone now. So,
when I do this my bode plot now starts at the 0 db line or an absolute gain of 1 and it
increases this way. Okay, I should also be careful that I do not, so first I would not want to
attenuate the low-frequency signals and at the same time, I also wouldn’t want to amplify the
higher frequency signals. So, a trade-off would be to select this α smartly enough such
that this is preserved; whereas, this length here is not big enough. So, I have to maximize this
α in such a way that the high-frequency signal attenuation is as minimal as possible.

626
And just look at the phase portrait, this will; we were always talking of a phase lead, right?
So, this is the phase lead that that it provides, so this guy it is adding an angle right, this thing
here okay. We’ll do little bit of more of where to this angle and all this when we do the
frequency design compensation, but at the movement it is we’ll just concentrate on this thing
here, this thing here, and what to do with this α ? Okay so, it is desirable to have α as
large as possible so that this length is as small as possible which means the higher frequency
signals are not amplified beyond the reasonable limits.

(Refer Slide Time: 04:57)

Okay now, we come back to design. So, a warning before we do this design procedure is that
if I follow the procedure it is not guaranteed that I will get the exact solution. I will be very
close to the exact solution after which I may; I may have to do a little bit of trial and error,
just to maybe just shift the pole slightly to the right or the zero to slightly left and so on. Okay
so, this is how the construction looks like. So, given a certain performance specification in
terms of the peak overshoot and the settling time; so again just to just to remind ourselves.
So, these are all transient response specifications right, I am improving only the transient
response. What will happen to the steady-state is something which we’ll see parallel as we go
through these steps.

Step (i) is again find what are the dominant poles given M p and t s , okay. Now, step (ii)
is in designing a controller is plot the root locus and see if there is a gain K which passes

627
through this S d . If it passes through the Sd is and then its perfect right, I don’t really
need to do much, but usually it does not as we saw in the previous example, right.

(Refer Slide Time: 06:18)

So, we will again start with the same example and see how we go about constructing a
compensator which looks like this. Okay so, this is what we had in the example in the last
lecture where which; where M p=10 and settling time, t s=5 seconds translates to
closed-loop pole locations at this numbers.

(Refer Slide Time: 06:44)

628
So, the angle deficiency is obvious by this computation, and we know that; we now conclude
that the root locus should be pulled to the left and so on.

(Refer Slide Time: 06:55)

Okay now, earlier construction just focused on this zero right. So, we didn’t have a p .
Okay now what changes when we have this p here, right? So, first, let me just say that
this is at some arbitrary location zc and pc is some arbitrary location. What I know is
just the φ , this I have already computed from here right that the compensator should
provide an additional angle of 28 ° with this formula it’s also done in the previous lecture.
So, I know this guy this is 28 ° something, this θ is cos−1 of the desired damping.

So, the location of zc now depends on where is; on you know how; where is this γ ?
For example, this φ when I say that the location, that this is may not be unique, if I just
choose a zc here and I just draw line here, okay. I can just add some number φ , the
same number, right. If this is at 25 ° , I can add 25 ° and I will say that this is also a
pc right, which also gives the same φ . There could be several possibilities right if this

is my Sd , zc , pc with φ could even be, well this is z c , much further pc

and so on, this is also φ . This is my zero; this is my pole.

So well, now what is a good location? Because there are infinite possibilities of zc and
pc which gives me this lead? Now, let’s see; let’s draw a line from here till here, the ωn
line, and call this angle γ right, we’ll see that the locations now depend on how much is

629
the angle γ , right? So, for a given γ , right so this I can use this φ , this θ to write
down this formulas with the ω n . I’ll not really derive this formulas, but given this
configuration: this angle is φ , this is γ , this is θ and an ωn , zc is related to
γ , θ and ω n via this this formula; pc is related again to ω n , γ , θ , φ ,
and this formula. So, what is; this is unknown because my problem is where to find, where to
place z c ? where to place pc ? On the right-hand side, well ωn is known, θ is
known, φ is known, what is unknown is this γ .

So, the location of poles and zeros depends on this γ . So, γ is the deciding factor.

(Refer Slide Time: 10:04)

Okay? Now, let’s just arbitrarily select few γ s and check. Okay so, let’s see what is
important when we choose this γ . First and the most important is the dominance
condition, second is the gain K , third are the error constants and four is the α right,
because here we wanted α in such a way that this length here from here till here was as
small as possible. Okay? So, let’s just do again some trial and error, start with γ 1=15 ° ,
take this γ 1=15 ° ; I put in these formulas and I get appropriate locations of zeros right and
the pole.

630
(Refer Slide Time: 10:46)

I say that the pole is fairly to the left of zero so that condition is trivially satisfied. The gain
K=1.4 , the velocity error constant is K v =0.5 , α is small enough 0 .4 23 and the
settling time is t s=9.23 seconds. Okay so, with these numbers, what we see is, okay the
gain is 1.4 , the velocity error constant has decreased from 1 to 0.5 , which means
the steady-state error increases, α has some number 0 .4 23 which is less than 1 ,
which is again trivially satisfied. Settling time is pretty poor right? What I want is a settling
time of less than 5 seconds.

Okay, why does this happen? Say well even though my lead compensator is successful in
giving me this lead or this ensures that ∠Gc G=−180 ° . Why are these things really bad?
The errors have increased, settling time is bad. This is something to do with the dominance
condition, right? That the dominance condition is not satisfied, which is obvious from this
plots, right? So, this is what we would like to decide, we would like our closed-loop system
to behave, but actually, it behaves like this, of course, it’s a little strange behavior here that
there is no overshoot, but there is still a little, it goes down and then it’ll result in something
like this.

The Mp criterion is met, but we are not really happy because that’s not just the objective,
your settling time is fairly large and the errors have increased. So, γ =15 ° is a wrong
choice even though this is satisfied. Okay so, this plus this is not the sufficient condition.
Okay so, why is this happening? Why is the dominance condition failing? Because, very

631
close to the closed-loop poles is a zero, right? and that is here right. So, if you look at; so this
guy, so the; how did we define the dominant poles? that they should be here such that well
there is nobody sitting over here and the other guys should be like further away from some
limit. But here when there is this zero here which loses or which helps losing the dominance,
right?

So, these guys are no longer dominant. And therefore, even though the root locus passes
through S d , now this is very important. The root locus is actually passing through Sd ,
but the settling time is not guaranteed, because these are not the dominant poles anymore,
because there this guy sitting over here, right. And therefore, this combination does not work
even though the root locus is passing through this point because the nominal design what we
do is just for a second-order approximation this one, okay?

(Refer Slide Time: 14:05)

Okay, let’s choose another number γ =50.2 ° , right. γ =50.2 ° then I get some with the
same formulas; I get some location of the zero, some location of the pole, the pole being to
the left of zero. And, it’s kind of quite good rather that the velocity error does not change
much. α is good and well the peak overshoot is more or less which I wanted it was 10
was the desired; now I have 11 , settling time is fairly good enough. And, you see that well
some kind of dominance is ensured because both this blue and the black lines are very close
to each other. So, this works well right, I say this is actually something nice for me, right?

632
(Refer Slide Time: 14:49)

Okay now, I become a little greedy and say that from 15 ° to 5 0.2° , I had this I could
get this blue and the black lines close to each other. Now, let me get a little, little more
greedy, and say well I will choose the 6 5 ° , okay. Again the same φ and all would be
ensured, I choose γ 1=6 5 ° and then zero comes at −1.4 , pole comes at −2.3 and
some so, this is certain gain, the velocity error constant has changed again from 1 to
1.3 , and the α has now decreased from 0.62 to 0.6 , right?

Okay so, now let us look at the numbers well the peak overshoot is not really good right it’s
13 ; it actually slightly increased. Settling time is still okay, right it’s 4.27 seconds. And
you see that the dominance is slowly has started, you know the system has the or these poles
have slowly started to lose the dominance because well the distance between the red and
sorry the blue and the black curves is increasing. Okay so, so why is this happening with the
increase in γ ? Okay so, now, let’s look at what is happening with the location of the poles
and zeros right. So, in the; Okay, in the previous case the value of gain K at 1.89 ,
which results in the closed-loop poles to be here, here and here; you see that the distance
between the pole and the zero here for this value of gain K is very minimal in such a way
that they actually cancel out each other.

So, the closed-loop system will just have like two poles, one pole here and one pole here.
That is like dominance is ensured by these two being very closed to each other. Ideally, we
would want these to cancel, you know these two guys to get cancelled. Whether that’ll

633
possibly happen only in the limiting case, when K is very large. So, for this value of gain
K , these are my dominant poles and they’re very close to each other. So, they will almost
nullify the effect of each other. Whereas, in this case, I see that there is some distance
between these two, that this don’t cancel out each other; and they in some way affect the
performance or you know they contribute to the plot here to the step response. And therefore,
we see that as a distance between these two again increases, the distance between these plots
also increases.

So, this is a little the nice thing happening here and nice observation that for this γ the
effect of these pole-zero pair is the minimal.

(Refer Slide Time: 17:48)

Okay now, is there really a method of finding right, what is a good choice of γ ? I just
cannot do a trial and error all the time, right. So, and if we look at the relation between γ
and α first case I had γ =15 ° and α =0.4 ; γ =50 ° , α =0. 62 ; γ =65 ° ,
α starts decreasing. So, there is one γ which maximizes the α and this is the γ
which I’m interested in right. And maximizing α is also good for me in terms of the
attenuation, okay. Because, if you see this is 4.45 db here, this is 4.06 , this is 7.47 .
So, this is I want; as I said earlier, I want to keep this length to the minimum. So, a good γ
is the one which maximizes α right, and not only that this also provides us or helps me
satisfies the dominance condition, right? In such a way that the transient specifications are
met, okay.

634
What is the effect on the steady-state error? We’ll not really worry about that, but some
adjustments can be made to accommodate steady-state requirements. So, here I am at prime; I
would primarily be interested in just improving the transient response without having too
much effect on the steady-state as we said, see how we saw here right, the effect on the
steady-state is not really too much there is some effect but we could possibly try and keep this
to a minimum.

Okay so, now the γ was the variable, you know so in this expression for zc - γ was
unknown, pc - γ was unknown, right. And α is related to zc and pc , right?
from the construction of the compensator. Let’s go back to the slide and say, well how is
zc and pc related via α ; here right this is z c , this is pc , right? And

z c 1/τ zc
= , 1/ τ and on the denominator I have 1/ ατ , so this is =α . So, how
p c 1/ατ pc
to maximize this?

zc
So, α which is , α is now like you know like as some non-linear function of
pc
γ , then I can just use standard calculus, what calculus teaches me? that the max is that, the

dα
maxima occurs when ¿ 0 . I do all the computations and I get a γ , these γ
dγ
maximizes the α . Now, the unknown when I was computing the location of the zeros and
the poles were the γ . Now, I get the γ in terms of θ and φ , these two are known
to me. So now, I know where to place these poles and zeros right given a γ .

635
(Refer Slide Time: 21:11)

Okay so, I just do the computations as before, I get that; start from θ I found, I find what
is θ from cos ζ , I get how much is the gain, and then
−1
γ turned out with that
formula to be 50.2° . Exactly as what we; what we had guessed in the second guess of
ours, which has the good α =0.62 , okay. And okay, this plus are exactly the same as what
we saw right at the poles and zeros are very close to each other. So, they nullify each other’s
effect and this and these two guys then ensure the dominance of the closed-loop. So, this is
exactly what we had before.

So, what we have answered here so far is how to place the zc and pc . First is that the
∠ Gc =φ . Now, next is, is there a unique location? well, the answer is no. The zc and
pc were depending on the γ and we chose a γ in such a way that it maximizes α
right. So that is, method one and it gives us some beautiful results here, okay.

636
(Refer Slide Time: 22:35)

Now, if I’m really lazy to you know, go through all this calculus and you know these are
another method of doing this well it’s a nice graphical method that I know the location of
S d , right. So this, this is Sd here right and then okay I know θ and ω n , I just
draw a line to the left here, right. And what I do is I bisect this entire angle, this entire angle
right. I draw a bisector to this angle which is this one right. So, this is the bisector right and
then I know φ right, φ is got, is derived by easiest computations. So, what I do is I go

φ φ
to this side right draw a line this is z c , I go to this side draw a line and this is
2 2
pc , this is exactly the same as what we have done before. All that analysis which we
talked in terms of maxima and minima would lead to the same thing if I would just do this
bisector thing. So, there are two ways of understanding this right. So, this γ which I
obtained by this geometric method actually maximizes α as we saw earlier okay.

637
(Refer Slide Time: 23:46)

Now, method two is there another way right, if I really do not want to compute γ , I forgot
the calculator to the exam and things like that; well what could I do. So, what we had seen
here is the following right. So, look at again these distances, this and this. And things were
depending on how these were located from each other right. So, from this observation, I can
say that in cases where an open-loop pole is in the region below the desired pole locations
right. So, this not really far-right this is 0.8 and this is 1 , in cases where the open-loop
pole −1 is in the region below the desired locations in the close vicinity of it, then I just
place a zero just slightly to the left of that pole, right. So, here I have a pole at −1 and I
place the zero just say at you know 1.07 .

And once I know this zero, I can easily calculate where is the pole, just by; based on the φ
or the lead angle which it should contribute right. So, this is also provides desirable value of
α while guarantying the dominance condition. In our condition, this is kind of easily if
you see here this is kind of met easily, right. So, this zero is placed slightly to the left of this
−1 , I would not want; you can you may even ask why not instead of 1.07 place it at
0.93 , if I place it at 0.93 . So, this pole according to the root locus condition will move
towards the right and I do not want anybody to move to the right. So, to ensure dominance of
these two poles here, so I want them to go to the left, but this is a very restricted condition
only in this case; in the case where an open-loop pole is placed just slightly, no just in the
vicinity, just below this dominant closed-loop pole okay.

638
(Refer Slide Time: 26:00)

So okay, so now, I found out what is the z c ? and what is pc ? Now, the unknown now
is; what is the closed-loop gain K . This I can easily find by the root locus magnitude
criterion right, use the magnitude criterion to find the gain K at which the root locus
passes through the desired closed-loop poles. So, we could see also; what is the effect on the
steady-state specifications, even though that is not really the objective of this controller, right.
And if the error constants are close to the specifications we can just do some minor
adjustments, right. So, in our example 1, where the locations of zero was at 1.07 , pole at
1.7 , resulted in a gain K=1.89 this can be computed by the magnitude criteria
criterion. And, this is very straight forward and we had also seen the formula for this
K v =1.1 per second, right.

The steady-state requirement here is not met right then can we actually do a minor adjustment
of the lead compensator. Well, this can only provide a minimal you know some very small
adjustment, but that can somehow even alter with your peak overshoot and all. So, in those
cases where we want to alter both the transient and the steady-state, we would do, also look at
what we would call a lag compensator, which we will study this lag compensator in the next
lecture, right.

639
(Refer Slide Time: 27:37)

So, before we finalize a design we just check with simulations to verify if all the criterion are
met or not, right. So, where did we start with, we started with specifications in such a way
that Mp was less than 10 , settling time was less than 5 second and I do all the
design, choose the best α it gives me 11 of M p and 4.38 seconds of the settling
time. And these are kind of within acceptable limits.

(Refer Slide Time: 28:07)

So, I could just summarize the design process into the following steps. First is find the
dominant poles; second is check the gain condition, just go through the root locus and check

640
is there any value of gain K such that the dominant poles lie on the root locus; if no then I
consider a lead compensator and then calculate where to place the poles and the zeros. Next
step is to calculate the gain K right, and then see is there too much effect on the steady-
state error constants; if there are, then I would go further I would go for a further design
procedure which would also include a lag compensator and I would verify all this design to
by appropriate simulations to see all, to see if all these specifications are met or not; if not
then I could just slightly push my pole or zero slightly to right or left to see that the
appropriate specifications are met.

(Refer Slide Time: 29:07)

So, in the next lecture, we will see how to meet specifications on the steady-state error or the
error constants via a lag compensator.

Thank you.

641
 Control Engineering
Dr. Ramkrishna Pasumarthy
Department of Electrical Engineering
Indian Institute of Technology, Madras

Module - 08
Lecture - 03
Design using the Root Locus

Okay so, continuing with our design methods using root locus, what we will see in this
particular lecture is design criterion to improve the steady state performance.

(Refer Slide Time: 00:26)

And, steady state performance essentially means reducing the steady state errors either
getting them to 0 or having them below some predefined values. This is done with
something which we call a lag compensation. Lag compensation is an approximation of
an integral control. We saw, how adding an integrator increases the type of the system.
Increasing the type of the systems, helps me track signals in a better way right. So, we
had this little relation between, what is the relation between a type−2 system and the
acceleration error constant, the velocity error constants and so on, right.

So, this lecture we will focus on the lag compensation which is an approximation of a PI
controller.

642
(Refer Slide Time: 01:20)

So, let’s start with an example. So, given a plant; certain transient specifications which
translate to ζ =0.5 77 , ω n=2.7 . These are my transient specifications and the
steady state specification says that my Kv should be greater than 10 right. Okay
so, first I draw the root locus and I see that at a certain value of gain K which is
32.4 . My root locus passes through the dominant poles and root locus passes through
the poles which are defined by this ζ and ω n . So, I don’t, I don’t really to worry
about the transient specifications because my root locus actually passes through those
points by just mere adjustment of the gain.

Okay, Let’s now worry about the steady state error. The error constant K v =1.8 . And I
want to; to make this larger than 10 . So, the question here is, can we design a
compensator in such a way that it has K v >10 , but not really alter this? I don’t want
this thing to be messed up right? I want that to design a controller such a way that this
steady state specification is made by not disturbing too much of what is happening here
in terms of the transient specifications.

643
(Refer Slide Time: 02:42)

Okay, let’s see how we can do that right. So, consider again a standard closed-loop
system right. So, I have the error, a gain K , a plant G( s) , the open-loop transfer
function is given by this now, where G( s) has a set of poles sorry, a sets; G( s) has
set of zeroes, a pole at the origin and set of other poles.

So, the transient specifications require the dominant closed-loop poles to be at some
s d , right. Which are completed by these 2 numbers? The first thing is assuming there
is a K >0 , such that the root locus passes through s d . We compute the gain
K in the following way. Again, this is just the magnitude criterion right
G(s) H (s)=1 or in this case K G(s) would have its magnitude to be equal to 1
.

644
(Refer Slide Time: 03:40)

∏ zi
i=1
Okay, then we look at the steady state error right, K v =lim n of this entire
s→ 0
∏
j=1
pj

number. And I am deliberately looking at a type−1 system right. Because a

type−0 system cannot track ramp input whereas, for type−2 system the steady
state error is 0 , K v =0 .

So, if Kv is some finite number, I am essentially looking at a type−1 system.

Again, we can refer to this table. Okay so, how would this look like? Well, the lag
compensator provides this facility. So, the first thing I would, if I look at the ideal you
know integral compensator, I would say that I should place a pole at the origin. Which
means well, which would eliminate the steady state error completely. But, what I told
you is that an ideal integrator cannot be realized. So, what you will say; well just put it
slightly to the left, this can be realized. And, that would be the first thing right. In the
lead compensator we just said, okay just add a zero, because it pulls my root locus to the
left. And, we see that it has its own set of troubles.

Now, if I just add a pole slightly to the left say instead of zero, let’s say I added 0.05 .
Now, does this meet the steady state conditions possibly? But what this will do, is it will
contribute certain angle right. So, this G or Gc will contribute certain angle to the
plant and then this condition would be violated. That’s easy to check. So, we want to add

645
this pole here, in such a way that it does not contribute to the angle. Or the angle
contribution is as minimal as possible and therefore, I start with this pole and I add a zero
slightly to the left. Again say, why not to the right? Well, I don’t really want, So I want
again, the dominance condition to be ensured. So, that the pole moves to the left and by
construction also it is kind of obvious.

So, the lag compensator by construction looks like this, in such a way that the β now
which is analogous to the α earlier. This β> 1 and z and p must be placed
so close to each other, that the angle contribution is less than 5 ° . I cannot make it
0° because, if I make it 0° the pole would sit on the zero they would cancel out
each other and there would be no compensation right. So, it is desirable to have G ,
the angle of Gc to be as small as possible to have the least effect on the transient
behavior. So, that is what we have to be careful in this, in this design procedure. So, here
in this case, this is the compensating pole, the zero should be to the left of the pole right.
Contrast to what was happening earlier at the zero was to the right of the pole.

(Refer Slide Time: 06:45)

Okay, first let’s see what is the effect of the closed-loop gain K while we are at the
lag compensator. Earlier, we had computed the gain to be something like this right. K
this in such a way that sd was passing through the root locus. Now, how much does
the gain change when I add a lag compensator? Say ^
K which is the total gain when I
add a lag compensator is the original gain K plus this extra factor. Now, the reason

646
why we place these 2 close to each other or also that these the ^
K should be as close to
K as possible right, and this magnitudes if you see, they are kind of almost equal,
because these are very close to each other. So, ^
K for ^
K should to be equal to
K this z ; s d + z g the magnitude of this should be the same as this one.

So, this is also what is desired, in such a way that we really do not disturb the transient
requirements.

(Refer Slide Time: 07:52)

Now, what is the error constant for the compensated system? The velocity error constant

^ Gc ( s ) G(s) zg
is lim K . And this turns out to be this is Kv and this K v , is
s →0 pg
the Kv of the uncompensated system, this guy. Kv of the uncompensated system.
Now, when I were to design this compensator, I am now worried about what is the
location of this poles and zeros. And the unknown, one of the factors that was deciding
was this β , β is an unknown and τ is the other unknown. Kv of the
compensated system is related to the Kv of the uncompensated system, via these 2

zg
numbers z g and pg . And, based on this construction of the compensator =β
pg
ok.

647
Now, I know how to find β right. This is known to me right; this is from the given
specification. So, I can find out what is β . So, for the appropriate design we must
have β as the ratio of the desired error constant to the error constant of the
compensated system which helps me in finding the location of zg and pg
appropriately.

(Refer Slide Time: 09:40)

So, how are these guys placed? So, first thing is to obtain a large range of factors, I
would have them as close to the origin as possible, say at 0.1 or 0. 001 and so on.
So, the factors can be increased by just you know. So, if I am at 0.1 or 0. 01 and I
want to have a factor of 10 , I can easily go to 0.1 right and so on. So, the zero
must be made as small as possible and pg must then be fixed, so I first select the zero.

Once I set the zero, since I know β I can fix what is my pg right. So, what must
also be ensured is that the angle contribution is as minimal as possible right, less than
5 ° . Unlike the lead compensator, there is no constructive way here neither a graphical
way of choosing z c and pc right. There we had a beautiful way of finding the γ
which maximizes the α and we could explicitly compute the formulas. So, here what
we will do is, we’ll just look at a graphical method without much proofs of what is
happening? But what this method will ensure is that the angle contribution is less than
5° .

648
So, what do I do is I take this constant � line or the line joining the origin to s d . From
this line, I draw another line which is at an angle less than 10 ° , should be 7° or
8° really you can do a little modification once we have the first iteration of design.
So, once I have this say, let; I fix this angle at 9 ° , I just draw another line and this
will give me the location of the zero. Once, I know the location of zero I can easily find,
what is the location of the pole right. Now, with this formula because, I know β and
β is derived by comparing the desired error constant to the error constant of the
uncompensated system.

(Refer Slide Time: 12:10)

So, again back to the example: alright, so we have; we knew that just by adjusting a gain
to 32.4 the transient specifications are met. Now, I’m interested in this steady state
specification right. The Kv of the open-loop or the uncompensated system is 1.8
and this is not satisfied it’s quite obvious.

649
(Refer Slide Time: 12:32)

^
Kv
Now, this is the desired error constant. So, β= =5.4 and β is always greater
Kv
than 1 right. So, now, just by doing the method as suggested earlier if I put my zero at
0.1 , the pole location turns out to be at 0. 108 and the lag; form of the lag
compensator is something like this. I will not really have a gain K because I don’t
really want to it to do. So, the gain contribution of this entire thing should be as close to
1 as possible right. And what is the angle contribution? Well its just 0. 89° is less
than 5 ° .

So, the lag compensator does not really alter the transient performance too much. As we
see here right. So, the uncompensated system the red one which was, which was the root
locus of the un-compensated system with just the gain adjustment, is the red line and we
see the blue line is just slightly to the right. This is not very surprising because I’m
adding a pole at the origin and this has the effect or pole very close to the origin this has
the effect of pulling the root locus slightly to the right.

But we say, what we saw here is we are improving the steady state performance by
having just very little effect on the transients. And we could you know maybe choose
alter this slightly to 2.09 and check the effect is minimal. It will never have, it will
never be a condition that the red and the blue lines match completely.

650
(Refer Slide Time: 14:19)

So, what we would see from this plot is that the original gain of 32.4 would not work,
because it does not really lie on the root locus. The root locus has moved slightly to the
left. So, what is the new gain K that we would choose? One method is to choose, is
to fix up the constants � line right and then see the intersection of the constant � line with
the new root locus which is slightly shifted to the right and then compute the new gain
right. And that would turn out to be it is 32.65 . So, that is what we would do here
right and if, so a constant � line would ensure that the peak overshoot does not change at
all. But if we were, are more interested in the settling time then I would look at the
intersection with the constant ω n line and then choose the gain appropriately.

It depends on what kind of specifications I am looking at. So, this slight gain adjustment
it’s just done by the root locus here by looking either at the constant � lines when my
transient, when my peak overshoot is more important or the constant ωn line when
my steady state or the settling time is more important.

651
(Refer Slide Time: 15:31)

So, this is just like the basic design procedure for a lag compensator. So, we have seen
how lag compensators helps in improve the steady state response in terms of the steady
state error, in the next lecture we will see given a certain system where neither the
transient specifications are met, nor the steady state performance requirements are met.
In that case how we will compensate the system, where we would need combination of
both the lead and the lag compensator. And we’ll see how to go about the design
procedure in that case.

Thank you.

652
 Control Engineering
Dr. Ramkrishna Pasumarthy
Department of Electrical Engineering
Indian Institute of Technology, Madras

Module - 08
Lecture - 04
Design using the Root Locus

Hello everybody. In this last lecture of module 8, we will still continue with the design of
root or Design using Root Locus methods.

(Refer Slide Time: 00:30)

Now, by looking at improving both the transient and the steady state behavior. Which
means we have to design both the lead and the lag compensators. So, how do we go
about doing it? So, far what we have learnt is that the lead compensator is suitable where
the or in scenarios where the transient response is unsatisfactory.

653
(Refer Slide Time: 00:43)

And we say that this is essentially that you are adding a zero, which will pull the root
locus to the left and then you add a compensating pole so that you could realize it
physically right. So, once the transient response is satisfactory or in problems or in
specifications where we don’t have to worry much about the transient response and we
have to improve the steady state response or improve the steady state errors we use a lag
compensator, right.

And in this and while we do this to design procedure. So, while designing the lead
compensator we would like the effect on the steady state to be minimal or maybe
sometimes improve. Whereas, when we design the lag compensator we would want to
design it in such a way the other system transients are not affected. And therefore, we
look at keeping the angle to less than 5 ° . And we also saw the effect of adding a lag
compensator it’s just is that it just shifts the root locus slightly to the right. In case of
specifications where neither the transients specifications are met nor the steady state is
satisfactory we would go for designing both lead and lag compensators.

654
(Refer Slide Time: 02:10)

So, how would it look like? Well so, if I look at the construction, well you have a general
gain of the compensator, we’ll have a lead component, a lag component here in such a
way that α < 1 , this ensures that the pole is further to the left of the zero, β> 1
ensures that the pole is to the right of the zero.

So, in a standard procedure is whenever we have to design a lead and the lag together we
first design the lead compensator and then a lag compensator. So, first we try to meet the
transient specifications and then go to the steady state compensation. So, if I just look at
okay, what do I need to know here? Well the unknowns are τ1 , τ2 , α , β , and
of course, they can be computed directly from the root locus once this specifications are
made.

655
(Refer Slide Time: 03:09)

Okay so, if I were to look at how does the realization of a lead-lag compensator look. So,
first case (i) where if you look at here. So, α<1 , β> 1 . So, I can always choose a
combination such that α β=1 .

So, this will ensure that α<1 and β> 1 . So, this circuit here it realizes a lead, a
lead-lag compensator just with passive elements, right. You know all these guys here are
passive elements. However, you have little less freedom right. Because, once you chose
the α here, the β here is fixed. Or vice-versa so once you fix your β you don’t
have too much control over the α . So, first we design a lead compensator which fixes
parameters τ 1 and α .

So, with α fixed I can prove or what we even saw earlier, was at the lag compensator

1
that this guy can improve the steady state error by a factor of . Now we have this
α
α which comes as a result of designing the lead compensator is not or does not
provide the satisfactory steady state response then I would go for the design separately.

656
(Refer Slide Time: 04:23)

In the next case, where I would want to choose my α and β separately in such a
way that α β ≠1 . Then okay so, first let us look at how the; how the series, how the
cascade of these 2 compensators looks like. To maintain this transfer function, I would
have to ensure that there is no loading at the output of system 1 right. So, if you see; if I
just remove this amplifier and I just plug this in here, this may not be; this will not be the
overall transfer function. As we saw in one of our very earlier lectures when we were
deriving transfer functions right. So, what I put here is a, is a amplifier of gain 1 in
such a way that the overall transfer function now is just the transfer function of this
multiplied by transfer function of the lag compensator here.

So, this amplifier acts as a buffer between the lead and the lag and helps in designing the
lead and the lag separately. So, here I can choose τ1 and α independently of τ 2
and β right. And I still have to satisfy this one and α should always be less than
1 and β should always be greater than 1 .

657
(Refer Slide Time: 05:35)

Okay so, we’ll start with a problem directly right. Now, we know the theory of what are
the steps involved when I want to design a lead compensator and what are the steps
involved when I want to derive a lag compensator. So, let’s say I am given a plant which

10
looks like this, G ( s )= . And the desired specifications are the damping
s (s +1.2)( s+ 6)
coefficient being ζ =0.5 and ω n=2 . And this corresponds to a peak overshoot of
16.3 .

So, I can I know the formulas to compute this and the settling time of 4 seconds(
t s=4 sec ). In addition, I also want or I desire that the error constant ( K v ) should
be 50 . So, based on these 2 specifications I can, I can compute what are the desired
closed-loop poles or where should be the dominant closed-loop poles, right? That is

s d =−ζ ωn ± j ωn √ 1−ζ 2=−1 ± j1.732 . that turns out to be these numbers right,
somewhere here and here.

So, as a first test to see; what is an appropriate compensator I just plot the root locus of
this guy. And just see if for any gain K , I will check if there is any gain K such
that the s d lies on the root locus right. So, if I just move along this lines well, I would
see that the s d does not lie on the root locus right. And therefore, just a minimum, just
a mere gain adjustment would not suffice.

658
(Refer Slide Time: 07:29)

So, the root locus does not pass through the desired closed-loop. When I am just
considering a proportional controller or I’m just adjusting the gain. So, what we conclude
from the previous slide by just using a proportional controller is neither the transient
specifications are made nor are the steady state specifications. So, we need to go for
designing both the lead and a lag compensator. So, let us start with a simple case when
α β=1 .

So, first I would do is to design a lead compensator. So, what are the steps right. So, if
my root locus of the closed-loop has to pass through these desired poles, then the angle
contribution at that point which is, which I call sd of the plant and the compensator
should be −180 ° . So, now, I know how to compute this G(sd ) right.

659
(Refer Slide Time: 08:29)

So, it’s a straight forward procedure; so what I find from all this computations is, the
angle that the compensator should contribute is 42.5 ° , right? Very straight forward
computations as we did when we were just designing a lead compensator. Now, the
second thing while we are designing is, now I know that my compensator should provide
an angle of 42.5 ° .

Now we saw last time, there are several of this pole-zero pairs that could achieve this
one. And what we saw is the best one, is the one which is computed by this one, right. If
you remember from the diagram what was γ right. So, once I know the γ I can
compute the locations of my poles and zeros, exactly the same procedure right you just
have to, just write down the formulas on the left hand side and then start filling the table
on the right hand side.

So the method to choose or to find the poles and zeros was something like this right. And
this gives me a compensator which is essentially a lead compensator with zero at
−1.2 and a pole at −3.15 and all these conditions are satisfied that the pole
should be to the left of zero. And the gain at which this occurs on the root locus is at
K=2.915 and this gain also includes this factor here right. So, 10 K is what we

'
' K
call as K right, and therefore the gain is ¿ 2.915 . Okay, now what is the
10

660
 1
impact of this on the steady state error? Well, if I look at substituting α= and
β
computing the steady state error, I get K v =1.623 .

Okay so, let’s first look at, what is how does the compensated system’s step response
look like? So, the black one is 1 , when I just look at these 2 dominant poles, these 2
right? And then I want the compensated system which is essentially the dominant poles
plus the compensator to be as close to the desired as possible. So, this is what is desired,
the black one, right in terms of the dominant pole analysis.

And this is actually, fairly we are close to each other. And this is the maximum, this is the
closest we could get as we saw earlier in the analysis based on the computation of this
γ . When you see that, well in the compensated system my M p=17 which is close
by to what I desire. And settling time is also like fairly what I wanted around 4
seconds.

(Refer Slide Time: 11:04)

So, now just to do a little bit of analysis on the contribution of these extra poles and why
we have this blue line and why do not the blue and the black line match each other?
Right. So I have, the systems the poles of the closed-loop system at −7.05 , −1.31
and then −1 ± √ 3 . And then the transfer function of the closed-loop system looks
like this. So, look at the dominance condition right. So, this pole is very far from the
dominant poles right.

661
And then there is a zero close by −1.266 and this is closer to the pole at −1.31 .
So, if you remember last time that we actually would, it was desired that there should be
a cancellation, as far as possible between the zero of the closed-loop system and one of
the poles here right. And, see these are placed fairly close to each other and therefore,
this minimizes the effect or the negative effect on the dominance of the remaining 2
poles right ok.

So, what we see in this when we do the partial fraction expansion is well there are, there
is a pole at −1.31 , there is a pole at −7.05 , this we now is much further away
from the desired or from the dominant pole. So, this effect is minimal whereas, the effect
of this pole is getting multiplied by a very small number on the numerator and this
essentially takes care, tells me that this pole-zero combination is almost like facing a
facing a cancellation and then you are just left with this part of the transfer function
which essentially deals with the dominant poles ok.

(Refer Slide Time: 12:47)

So, the design of the lead compensator ensures that the transient response is satisfactory.
Okay now, let’s see what is the effect on this steady state only by this lead compensator.
So, if I compute the velocity error constant it turns out to be that it is just 1.623 which
is must less than 50 , this is just with the lead compensator.

662
 1
Now, let me design a lag compensator with the β being fixed by α as β= . In
α
that case if I compute what is my steady state error constant the new one with the lag
compensator it just turns out to be this guy, 4.057 . And this actually comes in the
design procedure somewhere over here right. With α fixed the lag compensator can

1
improve the steady state error by a factor . And that’s exactly what is happening
α

1
here right . Well, this is still not good enough for me because, what I want is Kv
α
to go to 50 .

Therefore, we need to design a lag compensator independently of α right. Or in other

words we choose the lead-lag compensator which looks like this. So, we have the lead
now separately and the lag separately in such a way that α and β are chosen
independently, but we are still satisfying α being less than 1 and β being
greater than 1 . So, what is this β which I want to design? This is the desired, the
β is the ratio of the desired Kv which is 50 and the original Kv which I
have.

So, what is the Kv which I have now, it is whatever is left after I design the lead

50
compensator. So, I can now get the β to be the ratio of and that turns out to be
1.6
around 30.8 .

663
(Refer Slide Time: 14:41)

So, with this design; well now I have to fix where are the poles and zeros of the lag

1 1
compensator. So, let’s simply choose =0.1 and this gives me the location .
τ2 β τ2
And, what we also have to make sure is that the angle contribution of this lag
compensator is less than 5 ° . And that is ensured by this one that that the angle
contribution at sd which is the dominant poles, pole locations is 2.4 which is very
satisfactory. Therefore, the effect of lag compensator on the transients is very small. So,
my overall lead-lag compensator looks like this. So, this is the gain, this is the lead
component and this is the lag component. And let just compare it with the help of some
plots.

So, this is what I really want to track right the unit ramp. So, just by designing a lead
compensator I see that, well you know the red line, that there is a significant amount of
steady state error. And by designing the lead-lag compensation I see well, that the steady
state error actually reduces significantly right. This is, it is obtained by the looking at the
blue line.

Similarly, what is the effect of designing the lead-lag on the root locus? So, first when I
do my time domain specifications well the lead compensation the root locus looks
something like this and you know, I see that the root locus actually passes through the
desired poles. Now, if I do a lead-lag compensation I see that well its more or less

664
preserved by just which is not very unexpected that the root locus shifts slightly to the
right.

(Refer Slide Time: 16:23)

So, in this lecture we had seen both the passive and the active circuit realization of a
lead-lag compensator, where this active element came in terms of this amplifier with a
gain 1 over here right. And we had seen also how to design a lead-lag compensation
based on the root locus method. So, we end this module here and so we just did one
problem in each case, but more or less the problems will follow the same procedures.

There are some special cases which we need to be careful of, but we’ll post some extra
nodes based on that and we could possibly discuss them on at a later stage if you have
any more difficulties. And as a part of the assignment in this module we will give you a
problem or at least a set of problems where you need to start from modeling of the plant
you need to use matlab several times to check, you actually would be made to solve for
design problem with the help of matlab. And we will give you as many instructions as
you require.

And I hope it actually goes well when we when we, we learn a little more when we
actually solve a problem by ourselves. So, that will be about the assignment of this
module. And the next module we will look at the frequency domain compensations,
where we will see; what are the specifications, transient and steady state in terms of
frequency domain parameters, like the gain margin or the phase margin bandwidth and

665
several others. And similarly as we made use of the root locus in the time domain
analysis we will make use of the bode plot in while designing compensators in the
frequency domain. That’s coming up in the next lecture.

Thank you.

666
 Control Engineering
Dr. Ramkrishna Pasumarthy
Department of Electrical Engineering
Indian Institute of Technology, Madras

Module - 09
Design using Bode plots
Lecture - 01
Introduction to design in the frequency domain

Hello. In this module, we will learn how to use bode plots for meeting our design
purposes. So, we will call it as design in the frequency domain. So, so far what we have
done in the previous module was to learn how to design compensators and our main
analysis tool there was the root locus plot. So, we had the issue of addressing transient
performance specifications via a lead compensator, addressing steady-state performance
specifications via a lag compensator, and in cases where neither of these are met in the
open-loop, we go to design a lead-lag compensator.

And, so here we’ll see it from the slightly different perspective: from the perspective of
the frequency domain. So, as we saw that the lead compensator was used to improve the
transient response, lag was for the steady-state and in case we want to improve wanted to
improve both, we use a lead-lag compensator.

(Refer Slide Time: 01:24)

667
So, here the tool which we used was the root locus. And now in the frequency domain
design techniques, we will make use of the bode plots to verify our design or even for the
analysis purpose.

(Refer Slide Time: 01:39)

Okay so, let’s quickly again recollect, what was the root locus based approach to design?
So, the transient requirements were given to me in terms of the peak overshoot, in terms
of percentage and I had the settling time. And these 2 specifications were translated to
the location of the dominant poles. And the root locus was used in such a way that I
would want the root locus of the compensated system to pass through the dominant poles
alright.

And then in the final design, we ensure that the dominance conditions are met very
closely. We may not really get an exact overlap between the dominant poles and the
compensated plant, but we know techniques of how to get them as close to each other as
possible. Okay so, what will we do, or what will be our philosophy in the frequency
domain techniques? So, in this case, the transient requirements are specified in terms of
the phase margin, the gain crossover frequency, sometimes the resonant peak, and even
quantities like the bandwidth.

So, what we saw while we were analyzing the frequency domain plots or the frequency
domain response was that there is some relation between the specifications in the
frequency domain and the specifications in the time domain. We’ll again recollect those.

668
So, in this case, the design would involve reshaping of the bode plot to meet the
specifications. So, as analogous to what was happening in the time domain that we
wanted the root locus to go through or pass through the dominant pole location; similar
specifications will have to be met even in the frequency domain via the bode plot and
we’ll see how we do that.

This is also advantageous to me in more than one ways, right? In cases where the model
of the system is not available. And, and in cases where I could get or I could construct
the open-loop frequency response, this method is extremely helpful. So, in a way, it
means that I could actually construct the transfer function with by just seeing the
response of the system for various frequency signals. I could, once I could construct the
transfer function I will have a model of it and from the model, I can go to design the
frequency domain compensator. So, this is the part which we will do in the next module
which will be titled, how to? Should we title experimental determination of the bode
plots?

So, given a frequency response or given a bode plot, how do I obtain the transfer
function? That will we will do in the next module, but at the moment it’s safe to assume
that if I do not have a model, I can actually construct a model by using the frequency
response techniques.

(Refer Slide Time: 04:40)

669
Okay so again, just a very quick recap. So, a standard second-order plant in closed-loop
looks like this. Where we have the damping coefficient, the natural frequency, a typical
step response for an underdamped case looks something like this, where the overshoot
was calculated as a function of the ζ , settling time dependent on both ζ and ω n
and similarly, we had other quantities like the rise time.

And in addition to these 3 specifications, we also had the steady-state specification in

terms of the position error, the velocity error, and the acceleration error or the error
constants so to speak.

(Refer Slide Time: 05:19)

Okay, the closed-loop characteristics have few components. So, first is, so if you see this
is the response to different frequencies. So I will start from here, we’ll have a peak and
they again go down right. So, this we earlier termed it as the resonant peak and the
frequency at which this occurs is the resonant frequency and we computed that this was
directly depending on the damping factor. And the resonant frequency was depending on
the damping factor and also the natural frequency.

Similarly, for the bandwidth right, so bandwidth was computed in the following way that
it’s depending on ω n and also the ζ . Okay so, what is observation from these three
things? That systems which have large bandwidth or bandwidth close to ωn which
means the ζ would be closer to 0 and smaller the ζ would mean you will have a

670
smaller rise time and so on. Of course, you will have to trade-off in terms of the
overshoots.

However, I would from the filtering perspective, my noise signals appear as high-
frequency signals. And I would want to filter them out as much as possible, therefore,
just by merely increasing the bandwidth or designing my system to have a very large
bandwidth is sometimes not desirable. And therefore, we need to have a tradeoff between
the bandwidth and the rise time or even the overshoot, okay.

(Refer Slide Time: 07:06)

Okay, now a little bit more on those specifications. So, M r or the response of the freq
or the resonant peak here depends only on ζ . Therefore, and if you look at even the
formula for Mp in the time domain, it also dependent only on ζ . And therefore,
we can say there is actually a very nice correlation between Mr and Mp right,
therefore given a specification on M r , I can directly you know, have translated into a

specification on M p or you know specification on ζ , right.

So, in the same way as peak overshoot, the resonant frequency is also an indicator of the
relative stability of the system. Okay so, the expression for the resonant peak tells me
that it exists only for these values of ζ and therefore, the correlation between Mp

671
 1
and Mr exists only for ζ between 0 and and it’s given by the plots like this
√2
here right. So, at ζ =0.02 I have 0.707 , I have this number that Mp goes to
closer to 1 right, which you can get from this expression.

Okay so, what happens for ζ which is larger than this value? Well, for ζ larger

1
than square root, ζ >¿ the peak overshoot is very insignificant right if you just
√2
compute it from the formula for the M p . And therefore, I don’t really need to worry
about what happens for larger values of ζ because the overshoot is very minimal,
okay. Now for a given ζ right, either ω or the bandwidth or the resonant
frequency can be used to specify ωn right, which is apparent from these 3 formulas,
okay. And therefore, you can say that the bandwidth and the resonant frequency they
influence the rise time in the step response.

Again just by these 3 formulas and I am relating it to the formulas in the time domain
here right; the rise time has ζ and ω n , okay. And therefore, we can conclude that
larger the bandwidth or ω r , which we also argued here, larger the bandwidth or the
resonant frequency smaller is the rise time.

(Refer Slide Time: 09:38)

Okay now, further, how do we translate performance specifications from the time domain
to the frequency domain? Or are there any direct correlations? So, let’s see from with the

672
help of the plots here. So, the expression for the gain crossover frequency is depending
on ω n and ζ .

And similarly, the phase margin has a direct relation with ζ given by this formula,
right. And moreover a standard second-order system, this system is always stable, right?
It has an infinite gain margin and the phase margin is always a positive constant right.

(Refer Slide Time: 10:25)

So, has it, that is obvious now that the phase margin depends only on ζ , at this one.
And defining the ζ has a direct correlation with the peak overshoot in the time
domain. Therefore, specifications on the peak overshoot directly translate to
specifications on the face margin of the system right, through this expression. Okay now,
again larger the damping ratio larger is the phase margin, and therefore, the system is
relatively more stable. Larger the damping ratio more the poles shift to the left in the root
locus or in the time domain, and therefore, we can say the system is more stable, right.

Okay now if we look at the plot of ζ versus γ the phase margin which is
essentially a graph of this function. What we see is that for lower values of ζ between

γ
0 and 0.5, I can get this linear relationship that ζ =¿ and this is a very close
100
approximation if we just look at this plot. Okay, this actually means, that if I design a,
want a phase margin of 45 ° it approximately means that I am looking at a damping
ratio of 0.45 , right.

673
And for a given phase margin I can always relate it to the gain crossover frequency and
the bandwidth, again we just go through the expressions before. And therefore, ω gc or
the gain crossover frequency at which this one decides the rise time in the step response,
we’ll again come to come to these things when we analyze the low-frequency regions,
the middle frequency regions, and the high-frequency regions. So, at the moment it is
important for us to observe that if ζ is given right here.

(Refer Slide Time: 12:46)

Then ω gc decides the rise time in the step response right because again this is because
of this nice correlation over here. Okay now, what are the exact specifications, right? So,
when the objective is to improve the relative stability of the system, right? Then we will
specify the requirements directly in terms of the gain margin and the phase margin, okay.

Now, how are they related to transients? Well, the damping ratio was related to the phase
margin via this expression and then a given ζ and ω n the gain crossover frequency
was related via this one. Okay?

674
(Refer Slide Time: 13:23)

Now, we go now to the steady-state specifications. And the steady-state specifications

are usually in terms of the error constants; could be either the position error constant, the
velocity error constant, or the acceleration error constant. Okay so, we had in our; in the
lectures of bode plots learnt how do we get these error constants directly via the bode
plot, we’ll just try to try to recollect those again.

So, given this bode plot what is K p ? And Kp will exist only for a type−0
system. Which means that I’m actually having a transfer function of this form. So, if I

take the limit of ω ; lim G ( jω) , I get that the intersection here is 20 log K p in
ω →0

decibels or K p if I just look at in terms of absolute numbers.

675
(Refer Slide Time: 14:21)

Similarly, I look at a type−1 system to compute the velocity error constant, right. So,

type−1 , I have a pole at the origin and I know that this formula here lim sG(s)
s →0

will give me the value of K v . Now how do I compute the value of Kv from the
bode plot? Well if I look at very low frequencies which means ω is very less than 1 or
much less than 1, I just ignore these terms.

So, for low frequencies, this expression would hold that G( jω) can be approximated
simply as Kv because of this thing by jω . And the Kv can be obtained by
looking at the intersection of the initial slope of this line, of this guy with the 0 db line

Kv
or at which point the magnitude of which is essentially this guy, the bode plot for
ω
low frequencies becomes 1. And the frequency at which this initial line intersects the
0 dB line is actually my K v , we have done this before.

676
(Refer Slide Time: 15:31)

Similarly, with the acceleration error constant; I’m essentially now looking at a

type−2 system and I know that Ka is given by this formula of lim s2 G( s) .

s →0

Again, I look at very low frequencies; ω much less than 1. In which case I can

Ka
approximate my transfer function G ( jω )=¿ , right. And I just keep doing all
( jω )2
the same exercise, I look at the intersection with the 0 dB line or in which or in the

absolute terms | |
Ka
ω2
which is the magnitude of this becomes 1 well, it becomes 1

here, exactly at the point where ω=√ K a right, which is from this formula.

So, what is K a ? Well, it is, it can just be computed by this formula, it’s a square of the
frequency at which my initial line intersects the 0 dB line.

(Refer Slide Time: 16:32)

677
Okay now, let’s talk more in terms of now this the performance specifications in the
frequency domain. So, the relative stability requirements are in terms of the phase and
the gain margin. And steady-state are usually again in the same terms of Kp , Kv
and Ka right. Similar to the time-domain again we recollect this, in terms of
overshoot, the rise time or the settling time. These are now translated via appropriate
formulas or relations to specifications on the gain margin and the gain crossover
frequency.

And steady-state specifications remain the same right. And what we have to be careful
here is that the correlations between the time and frequency domain parameters used in
this lecture are exact for a second-order system right. So, we started our analysis all the
Mp , M r ; these were formulas derived only for second-order system. And in the
time domain, we translated this specification in terms of the dominant poles right, where
we wanted the exact plots and the plots with the compensator and the dominant response
to be as close to each other as possible. Right?

678
(Refer Slide Time: 18:07)

So, same thing we could do even in the frequency domain. Okay now, depending on, so
so far we saw relations between you know; how are the phase margin and the gain
margin related to the time domain specifications in terms of ζ , ω n , the bandwidth,
the resonant frequency, and so on. Now, based on what we had argued so far, if I look at
the low-frequency region, again look at these plots, the low-frequency region here, the
low-frequency regions here, now ω being less than 1. Or the low-frequency regions
here they were deciding if I call them as region 1 they were deciding the steady-state
response right. So, the region 1 will decide how my system behaves at the steady-state
right, what is steady-state specifications either in terms of K p or Kv or Ka ?

Now, look at this region here, in terms of the gain crossover frequency. And if we go
back a few slides, the gain crossover frequency was related to the transient response of
the system. So, this frequency range here which I call the region 2 will decide the
transient behavior of my closed-loop system or also the relative stability. Region 3 which
is much ahead or much after the region 2 or much after the gain crossover frequency, this
indicates in some sense the complexity of the system. What are the extra poles and zeros
in the system?

And at this region what we must be careful is that the gain at this, in this region should
be attenuated to reduce the effect of noise. So, we, the design specifications can be
categorized essentially into 2 regions, region 1 the low-frequency taking care of the

679
steady-state requirements, region 2 taking care of the transient response which is
essentially the region slightly to the left and slightly to the right of the gain crossover
frequency. And only a specification or only thing which we need to be careful in region 3
is that the gain must be attenuated right, in in in order to reduce the effect of noise on the
system.

(Refer Slide Time: 20:31)

Okay so, how do these things look by construction? So, in the time domain, I know that
the lead compensator had a zero, and a pole placed such that alpha was equal to 1, and
similarly, here I can write down the sinusoidal transfer function. The lead compensator
goes exactly the same way. And similarly, for the lead-lag compensator, I’m just writing
down the sinusoidal transfer functions here. Here, I just call the gain K' to be just
αK and here to be βK and so on.

680
(Refer Slide Time: 21:10)

Okay so, what, how do we go about doing this? Now we have a nice correlation between
the time domain specifications and the frequency domain specifications, you also have
nice explanation of the low-frequency region corresponding to the steady-state
requirements, then the mid-frequency range around the gain crossover frequency
corresponding to the transient specifications right. So, the first thing, as usual, we would
do is to see in terms of the gain adjustment, just by adjusting the gain, do I get the
appropriate gain margin or the phase margin. And by the root locus plots I actually sorry,
by the bode plots I actually know how to compute the gain and the phase margin.

And once I design or look for an appropriate gain K and if it does not exist, just by
merely adjusting the gain if my closed-loop specifications are not met, then I would use
an appropriate compensator could be again in terms of a lead compensator and the lag
compensator.

681
(Refer Slide Time: 22:23)

But as usual, the very simple first step is to is to just look at a proportional controller or
just look at adjusting the gain, right. And then if they are not met then we look at
appropriate regions for the compensation.

(Refer Slide Time: 22:37)

So, in this lecture what we saw is how the time domain and the frequency domain
specifications are related to each other, the various regions how they are categorized in
terms of frequency? And in the next lecture, we will see exact construction of the lead
compensator using bode plots.

682
Thank you.

683
 Control Engineering
Dr. Ramkrishna Pasumarthy
Department of Electrical Engineering
Indian Institute of Technology, Madras

Module - 09
Design using Bode plots
Lecture - 02
Design of Lead compensators using Bode plots

So, continuing on our discussions with respect to designing lead compensators in the
frequency domain making use of the bode plots.

(Refer Slide Time: 00:27)

So, let’s just recall what we had in the last lecture. So, specifications; the performance
specifications were usually in terms of the gain, and the phase margins this were like what
determine the transient performance together with the error constants in terms of the position,
the velocity, or the error the sorry the acceleration error constant. So, what happens in the
frequency domain? So, we saw that the lower frequency region decides my steady-state
performance in terms of a certain Kp , K v , or K a . In the region 2 where the gain
margin and phase margin kind of concepts come they determine the relative stability and the
transient performance, and the higher frequency region determines how we are handling
noise is it amplified or not and so on.

684
(Refer Slide Time: 01:30)

So, here we would be interested typically in region 1 and region 2. And we will see how we
go about looking at a design problem. So, just look at the sinusoidal transfer function and
then we plot the response of Gc ( jω ) G( jω) . So if Gc ( jω )=K ' G' c ( jω ) with a gain
coming out explicitly then Gc ( jω ) G( jω) can be written down explicitly like this right.
And then this guy can actually go into what we also can say as the plant dynamics. So, what
we did in the root locus when we had a problem where we wanted to compensate for the
transient performance and also ensure a certain steady-state error or study certain bounds of
the steady-state error, we first design the transient controller which is essentially the lead
compensator. And then went onto design the steady-state compensation which was a lag
compensator.

So, here we do slightly different approaches. The first step would be to meet the steady-state
specifications and it will be obvious in the next few slides why we need to meet the steady-
state requirements first. So, in order to meet the steady-state requirements, we look at the
frequency response of this thing. And then we choose the gain K' such that the steady-
state specifications are met. And once the steady-state specifications are met we choose an
appropriate controller which we denote here as G' c to achieve the desired gain margins
and the phase margins. So, this is how my transient performance specifications are stated to
be, they are given to me in terms of the phase margin and the gain margin and we know that
there actually exists a one to one relation between the phase margin and the things like the
damping coefficient and all.

685
So, the first step in any control design is to just start with a proportional controller as we did
even in the root locus is there. So, we mapped that given transient performance specifications
to a set of dominant poles. And if the dominant poles were lying on the open-loop root locus,
we just compute that gain at that location and say well I can just design a proportional
controller of that particular gain to ensure my behavior in terms of the transient performance.
So, similar things we also do here right, we begin first with a proportional controller.

(Refer Slide Time: 03:57)

So, well just to make things a little easier, let’s start with the help of an example. And I am

4
given G(s) the plant has with specifications of this form, that the overshoot
s (s +1)
should be less than 15 and the settling time of less than 5 seconds which translate to
appropriate � and ω n while we do the formulas for this. Which in turn translate to a phase
margin of 53 ° and a certain gain crossover frequency of 1.19 . And in addition, I am
given that the steady-state specifications which I should meet is 40 sec−1 .

Now, just look at the open-loop bode plot right. So, here the open-loop things would tell me
that my phase margin is 28 ° and I know how to find out the velocity error constant from
the bode plot, I just extend this line and find the frequency at which it intersects the 0 db
line and that frequency is actually my K v . So, this is what we did when we were
discussing how to find out error constants just by looking at the bode plot. So, the
specifications of the open-loop system has the following characteristic K v =4 , and the

686
phase margin is 28 ° . And we would like to have it something like this that K v =40
with a phase margin of 53 ° .

(Refer Slide Time: 05:31)

So, let just start with the proportional controller where Gc ( s ) =K . How does what changes
in the magnitude and the phase? Well, the log magnitude changes according to this thing
20 log |KG ( jω)| . So, you see that this gain K just adds up this magnitude to my open-
loop frequency response or open-loop bode plot, this is my plant. The phase angle remains
the same. So, with a proportional controller, the bode magnitude plot shifts by this number
20 log K the phase plot remains the same, but this is well the phase plot remains the same,
but something strange happens here.

687
(Refer Slide Time: 06:17)

Let’s say I start with this plant and I say I just keep on increasing my gain to see what
happens. So, I keep on increasing the gain and I see that what happens is with an increase in
gain, this bode plot shifts, the magnitude plot shifts up. And when it shifts up, what we see is
that this frequency which we call as the gain crossover frequency this keeps on shifting to the
right. So, even though the phase plot remains the same, the phase margin changes because the
phase margin is defined now at the new gain crossover frequency. So, if I just keep plotting
some of this and I see that my phase margin actually keeps on decreasing as I arbitrarily keep
on increasing the gain and that is essentially because this changes this frequency changes
right.

So, this is corresponding to this thing I was keep going here; even though I see that an
improvement in my steady-state is visible. So, because this frequency at which the initial
slope intersects the 0 db line will tell me what is my K v . So, I really wanted it to be as
further as possible. So, for a gain of K=10 , I have K v =40 , but the phase margin is
very, is highly unacceptable it’s just 9.04 ° when the original system had a phase margin
of 28 ° and the desired is 53 ° .

688
(Refer Slide Time: 08:00)

Now, what we observe is that well I just, when I just do a proportional controller, well I have
a big trade-off between the steady-state and the transient specification. So, based on this plot
here, what I see is that to achieve a desired steady-state performance or if Kv should be
large this actually means that the gain in region 1 should be as high as possible to achieve low
steady-state errors, so if you see here right. This keeps on increasing, and this keeps on
increasing, my Kv goes on increasing, and therefore, my steady-state error decreases.
Now, the region 1 thing is taken care of. Now, I want to do something over here in such a
way that my phase margin requirement is also met.

(Refer Slide Time: 08:54)

689
And this is where we come to this point of designing a lead compensator using bode plots.

(Refer Slide Time: 09:02)

So, the transfer function of the lead compensator is exactly the same as we did in the time
domain, where I have a gain, zero, a pole. Where pole was significantly to the left of the zero.
I could write it in time constants in this way. And this is my sinusoidal transfer function. And
as usual α< 1 . So, I am looking at designing Gc ( jω )=K ' G' c ( jω) where this is this
guy is a just the G' c and K' just take care of the gain factor here that is αK . So,
first, we design this K ' , the gain which ensures a desired steady-state performance and
then I go about finding what are these numbers here, the unknowns are the ω , unknown is
the capital T , and unknown is the α .

690
(Refer Slide Time: 09:53)

So, just by making α < 1 , well I know that the pole is to the left of the zero on my complex
plane. So, a typical bode diagram the asymptotic plot is given in the black, it would look
something like this. And when you see that the phase actually has a nice shape here. So, if I
go back and compare to what was happening earlier, you see that the low-frequency region
right, so here, so this is now this is my region of interest right. So, nothing much should
happen over here. So, what I see is at the low-frequency region the addition to the gain is
very small or almost 0, similarly with a phase.

So, I am not really interfering with the low-frequency region; however, here I see that well
the phase margin should increase which means this plot should be pushed to the right, so
sorry to further a little up to ensure a bigger phase margin. And this curve here ensures that I
push the phase margin or push the phase plot to slightly upwards so that I have the desired
phase margin. So now, the lead compensator is essentially seen as the high pass filter, right.
You can see for higher frequencies you have a higher gain and so on. So and unlike the
proportional controller, the lead compensator raises the bode magnitude plot at only high
frequencies not at the low frequencies right so nothing is happening at the low frequencies.

So, I am ensuring something in the mid-frequency range without actually disturbing my low-
frequency range. Now, if you see how this plot goes, there is a point where the phase attains a
maximum value and it again goes back close to 0.

691
(Refer Slide Time: 11:55)

Now, there is a frequency at which this compensator provides a maximum phase lead right,
and this frequency what we will use in the design process. Now, what is this maximum phase
lag and what are we trying to identify here, we are trying to identify these two corner
frequencies and this α . So, once we know these two control frequencies, we know the
α then our design process is complete.

Now, the angle which the compensator provides is given by this one. So, I have a
tan −1 ( ωT ) coming from the numerator and the −tan −1 (α ωT ) . And then you see that
this angle φ which is angle contribution here, it varies with frequency and it has a
maximum at a particular frequency. This is like a nice looking continuous function. So, I can
just find the maximum of it. I can say what if I ask myself a question what is the frequency at

dφ
which φc attains a maximum value and calculus tells me that just do this thing
dω
¿ 0 , which leads to me to this expression.

1
And I say the ω at which the maximum phase occurs is just this one ω=¿ .
√α T
Now, this frequency is a geometric mean of the two cut off frequencies as you see this is one

1 1
cutoff frequency at and this is a cutoff frequency of . And this is just a
T αT

692
geometric mean of both of them right. So, the frequency at which the maximum peak occurs
is a geometric mean of these two cutoff frequencies.

(Refer Slide Time: 13:40)

And then we conclude that the maximum phase lead occurs at this particular frequency. Now,
at this frequency what is how does φc look like. Well, in this expression, now I know what
is this ω right, where maximum φc occurs that is this one. So, I substitute this
expression into this, into the computation of the angle, and I do all these things. And I see that
the maximum phase lead is related to α via this expression, just a simple computation,
right. Then first, I find out what is the frequency at which the maximum phase lead occurs I
substitute it to into this expression and I get what is the maximum value of φc .

So, two questions: what frequency? Well, it occurs at this frequency the maximum phase
lead. What is the value at maximum phase lead? Well, that value of that angle is written down
in terms of α with this, this is little steps here. So, given α , I can find a φc ¿ω m
or
vice versa. So, if we’re given how much phase lead I need to add to the system or how much
should be the extra angle contribution I can find out what is α . So, in this expression, now

I find out α , let say φc ¿ω =sin

m
−1
( 1−α
1+α )
and this occurs at this frequency.

693
Now, what is happening to the magnitude at this frequency? so somewhere here, in this

1
magnitude is at |G 'c ( jω)|ω=ω =20 log this also can be computed pretty easily that I do
m
√α
not really need to derive expressions. So, we already know via bode plots of how we do this,
and I am just substituting in this expression these values which I compute right. So, substitute

1
ω=¿ in this expression find out what the magnitude is and I just get this one,
√α T
1
20 log is the magnitude at this frequency. And what is this frequency this is a
√α
frequency at which the maximum phase lead occurs due to the compensator.

(Refer Slide Time: 15:57)

4
Okay now, let us come back to a design problem. So, I have G ( s )=¿ and then
s (s +1)
I translate all this again to specifications to such a way that I want γ or the phase margin
−1
to be 53 ° together with K v =40 sec . So, the open-loop characteristic is at γ =28 °
and K v =4 as we have seen even earlier.

694
(Refer Slide Time: 16:23)

So, first is well can I design, can I at least first take care of my steady-state requirements.
Well, how is the steady-state requirement? So, Kv the velocity error constant is s times
the entire transfer function. So, entire transfers function. So, entire transfer function consists
'
of this K ' G' c ( s ) G( s) . So, what I know is that if I look at this expression lim G c ( s )=1 .
s →0

Why is that obvious? That is obvious because just substitute ω=0 , here you always get
1 . So, this really does not contribute anything to that. So, that’s also indication that
nothing really is happening with this compensator for this steady-state error, it doesn’t really
contribute anything either positive or negative.

So, I am just left with this expressions K v =K ' lim sG( s) . What is required K v =40 .
s →0

And I know that lim sG ( s )=4 , this is given to me right from here lim sG ( s )=4 and
s →0 s →0

' 40
then so I can just find out the new gain K =10 , ; Or this is the additional gain that
4
my compensator should provide with K ' =10 , I can get K v =40 . Now, I just simply add
this compensator and I just plot KG(s) . So, this is my controller the proportional
controller which gives me a steady-state error constant of 40.

And what happens to the bode plot, well this is fine. This initial slope intersects at
40 rad / sec and this is the value of my K v , I look at this gain crossover frequency. And

695
I find something terrible has happened here is that the gain. The phase margin has reduced to
9 ° right, with this gain K . So, I want this guy here the phase margin to be like 50 ° .
So, if I just say well, I want to put a controller here such a way that you know, say if I do
something like this then the phase will increase to give me a desired phase margin. This is
essentially what I want to do, I want to add some angle here; add an angle such that the new
γ =53 ° . Now how will we do this? So, here the gain crossover frequency is roughly about
6.28 right, you can just see from this plot here.

(Refer Slide Time: 19:12)

So, I have to perform this operation in such a way that the effect here is minimum, right. so
that is what we even said that. So, now, given that the steady-state performance is
satisfactory, I must choose the compensator that does not disturb this low-frequency content.
And then if I look at a lead compensator as I said earlier that it is a high pass filter and
provides no attenuation to low-frequency signals no gain, no attenuation also. Now, we also
required that the additional phase margin at the gain crossover frequency to meet the transient
specifications, and then this lead compensator provides me that additional thing over here.
So, this actually should not be 9 ° , but this should be somewhere like close to 50 ° .
Now, this is what we will achieve with the help of a lead compensator.

696
(Refer Slide Time: 19:58)

Now, let us just do some experiments right to say well now I want to design a compensator,
which looks like this. The only information I have is α< 1 and I know the K’ which
gives me a K v =40 . So, we want to select α and T in such a way that I should meet
the desired specification. So, I just play around with different values of α . So, for different
values of α , the bode plot of the compensator looks something like this.

So, that is an angle here and then the frequency at which the maximum angle occurs keeps on
shifting to the right, and you have the magnitude plot going this way. And you see the low-
frequency behavior more or less remains the same, there is nothing happening in this region.
So, we are not really playing around or we are not really disturbing what is happening to the
Kv .

697
(Refer Slide Time: 20:58)

Now, we require an additional phase angle at the gain crossover frequency. So, this is how the
phase margin is defined right, the amount of phase of 180 minus the formula. Now, where do
we compute this at the gain crossover frequency? Now, at this gain crossover frequency, the
angle should be 53 ° . Now, I already have an angle of 9.04 ° now how much should the
lead compensator give me, well I have through this computation that the lead compensator
should give me an additional angle of 44 ° at this point.

So, this is this angle plus 44 ° then I am fine right, I am; my transient specifications are
met, my steady-state specifications are met and I am happy. So, how do we go about this? So,
here well this is my gain crossover frequency 6.28 and I add this much of angle at this
frequency, that’s what I would naturally do, right. So, that this is my gain crossover
frequency, and what I need is angle the overall angle here to be like 53 ° .

698
(Refer Slide Time: 22:10)

Now, let us just do this right arbitrarily. Now, what I know, how do I compute this thing? well
I know that φm is related to α via this formula. So, φm is the additional amount of
phase my controller or the compensators will should provide. So, for this φm , it turns out

that the α =0.179 . Now, this φm the maximum phase lead, I know that it should occur
at the gain crossover frequency right here, I want this to be 53 ° over here at this
frequency. Now, this frequency will be my ω m .

Now, once I know ω m , I know α , I can find out where what is the T ; because this
ω m occurs at the geometric mean of these two cutoff frequencies. So, now, I know α ,I
know what is the ω m , so I can compute what is T , based on the formulas which we
derived. Now, I know K ’ , I know ω , I know; so I don’t need to know ω right, I
know the ω m , based on this ωm I compute what is α and what is T . And this is
how my compensator would look like. Now, let us verify if this is actually helping me get the
desired phase margin or not.

699
(Refer Slide Time: 23:41)

Now, the bode plot; I would now plot; see the plot of not just G(s) , but it will be K’ ,

with this K ’ gives me a K v =40 sec −1 . Now, I have G and I also have Gc . This is
my Gc , G is given to me as a plant. Now, I just draw these things and I find out well
what is the phase margin, the phase margin actually turns out. So, I should look at the phase
margin, not of G , but now of Gc ( s ) G( s) ; Okay, Gc ( s ) G( s) . This is my, this is
where I will measure the phase margin, this is the gain crossover frequency of the
compensated system. I go down and I check well it is not 53 ° but it’s actually 43° .
What has gone wrong here?

So, if I just compare the bode plot of the uncompensated system, what I see is that there is a
change in the magnitude plot in such a way that my crossover frequency is no longer at
6.28 , but it will somewhere here. And we design our compensator such that the maximum
phase lead occurs at this frequency, but that does not really satisfy the closed-loop conditions
because this point has shifted to the right. And therefore, we should have γ at this new
crossover frequency to be 53 ° not here; this is not useful to me.

So, just by looking at this crossover frequency and designing a compensator does not help me
much because the gain crossover frequency of the compensated system shifts to the right that
is obvious because you see there is a little bit of a magnitude change also in the
characteristics of the compensator, not only the phase changes but also the magnitude

700
changes. Therefore, the natural tendency is that the gain crossover frequency will shift to the
right because we actually adding up some magnitude.

(Refer Slide Time: 25:59)

Now, how to fix this problem? Because of the positive magnitude contribution of the lead
compensator the gain crossover frequency naturally shifts to the right. And the maximum
phase contribution from the lead compensator it no longer occurs at the new gain crossover
frequency it occurs still at the all the gain crossover frequency, but we want the 53 ° to be
at the new gain cross over frequency. So, add the new gain crossover frequency it is 43° ,
but I really want 53 ° .

Now, well how do we know? So, now, I know already that the bode plot shifts to the right.
So, can I appropriately find a new gain crossover frequency? Well, one technique is so
instead of 43° this should have been 53 ° . So, you can I actually convert this 53 °
which I was designing to be here to be some 63° for example. So, this is I push it a little
further up this will naturally go up right, so that is one technique. The new gain crossover
frequency the phase contribution should be greater than by 10 .

So, what I should then do is I should shift this by some extra 10 ° which means that even
though this is going down here this can be 53 ° and the φm here can be much higher.
So, I just add up 10 ° here, so that I compensate here. Now, if I design for φm =54 ° , I
can find the appropriate α again from the same formula nothing changes here; instead of
4 4° now just to compensate for this shift I’m adding some 10 here, okay?

701
(Refer Slide Time: 27:52)

So, with this new α , I can find well now we must fix the T . Now, how do we fix the
T , this is the question. Now, look at what happens at ω m . At this ω m , if I go back to
what my bode plot does, we just take a much simpler one the first one here. At this ωm ,
there is a certain substitute magnitude here. So, this is an extra magnitude contribution right
to my closed-loop bode plot and therefore, it shifts a little to a right and therefore, I select
ω m , at ωm I do not know this ω m , but I know what is the contribution of the
magnitude.

So, based on what we had derived earlier this one. I don’t really need to know ω m , but I

know that at a particular ωm where the maximum phase lead occurs. The contribution of

1 1
the magnitude plot is 20 log . Now, in this example that turns out to be 20 log
√α √α
1
with this new α turns out to be 20 log =9.799 .
√α

702
(Refer Slide Time: 29:16)

Now, what do I do? If we choose ωm to be the frequency at which the magnitude of the
gain compensated system is the negative of this, then ωm will be the new gain crossover
frequency. Let us again go back to our plot and see what this means. Now, this is the old gain
crossover frequency. This doesn’t help me right. Now, what should be the condition at new
gain crossover frequency, call it ω c . At this for the new frequency should be such that G;

20 log |K ' G( jω)G ' c ( jω)|=0 , which means that now I know what is this at the new

1
crossover frequency. And I can just this compute it by 20 log . So, it should be such
√α
1
that. So, if this is my 20 log . So, whatever this is 9.7 here.
√α

Then this is the new frequency which I will choose because at this frequency right, here the
magnitude of G by itself is −9.7 this entire guy K ’ G( jω) . And what I am adding
to this to this −9.7 , I am adding now a +9.7 which comes from the compensator so
that the overall contribution become zero. Now let’s read this statement again. Choose a
ωm to be the frequency at which the magnitude of the gain compensated system that is
K ’ G ( jω )=−9.7 right, then ω m will be a new crossover frequency.

Now, from the bode plot I can find out what this is right, this frequency is found out to be
11.2 rad / sec . Now, once I know the ω m , I could compute easily what is the T .

703
Once I know the T , I can get the structure for the compensator. And what this gives me is
well this is kind of very nice here.

(Refer Slide Time: 31:44)

So, this gives me a phase margin of 59 ° just put this is exactly what I want right, and well
the gain crossover frequency here is 11 .2 . So, this is how we just do it a little bit by trial
and error to find out to compensate further shift in the bode plot to the right. So, there was
when we added a lead compensator to the root locus it was pulling it to the right, here the
magnitude plots shifts slightly to the right. And therefore, I need to find out what is a new
gain crossover frequency. This new gain crossover frequency is just a frequency at which the
gain compensated system has this much of magnitude contribution.

So, this is how it’s done right. So, with the help of a very simple example; other example
would also follow the same nature of it right. So, this is nothing really complicated about
taking a system which has four poles and five you know five thee zeros and things like that.

704
(Refer Slide Time: 32:50)

So, this was a very simple and yet powerful illustration of how we design lead compensators
in the frequency domain. So, the idea was just ok, so then the thing was first I need to find
out what are the locations of the poles and zeros. First unlike the root locus, I here first
compensate for the steady-state behavior and then come to the transient behavior. So, I again
as usual compute what is a phase lead, I compute the new gain crossover frequency and I
have now techniques how to compute the new gain crossover frequency and based on that I
can find out what is my α and T which will give me the location of the poles and the
zeros.

So, similarly next time, we will see how to design a lag compensators and eventually lead
into lead and lag compensators using bode plots.

Thank you.

705
 Control Engineering
Dr. Ramkrishna Pasumarthy
Department of Electrical Engineering
Indian Institute of Technology, Madras

Module - 09
Design using Bode plots
Lecture - 03
Design of Lag compensators using Bode plots

Hey guys. So, in this module, we’ll do another smaller lecture, where we will learn how
to design lag compensators via bode plots. So, just to recall where we started off all this.

(Refer Slide Time: 00:35)

So, in the frequency domain, we classified the entire frequency range into 3 regions. As
the low-frequency, the middle-frequency, and the high-frequency range. And what we
observed was, or what we were desired that the gain in region 1 should be as high as
possible, so as to achieve low steady-state errors. Region 2 must-have satisfactory gain
and phase margins. And if there are conflicting requirements we need sophisticated
controllers and so on. And then there are the higher frequency region takes care of
attenuation of unwanted noise signals, which usually occur at higher frequencies ok.

706
(Refer Slide Time: 01:16)

So, how does this lag compensator look like? Well, by construction well I have a gain
K , I have a zero and a pole placed in this manner. Exactly the same as we did while
we were doing compensation via root locus right, in the time domain.

(Refer Slide Time: 01:35)

The value of β is chosen such that it is greater than 1, which means that the zero is to
the right of the pole, again as we did in root locus. Now in the frequency domain, I can
write down the sinusoidal transfer function in the following forms. So, I have; I just
rearrange these β s and write the sinusoidal transfer function. So, I have βK

707
 1+ jωT
and similarly, the denominator with a β . And I call this number
1+ j βωT
' '
K =β K and then these terms remain. As usual, β> 1 and K =β K .

So, the compensator now I split into 2 terms. One is purely to do with the gain K’
' '
here. And then G c , where Gc is just this transfer function without this gain K ,
which we already accounted for over here. So, what is important to note is that this
K’ also contains this design parameter β ; however, when we design it will not
really be explicitly as mentioned, but we need to keep in mind that we need to design
K’ in such a way that it also accounts for β . So, this K and this K ’ are
different in such a way that they are related via β .

So, some books may say that, once they end up computing the controller they say desire;

K’
you know divide the original gain K by and so on, but once we just do it in
β
this steps we will avoid that complication right. I just design K’ and I say. And then
the compensator, the dynamic part of it will just be this one. So, we’ll really eliminated
again you know dividing or multiplying by certain number. So, in our process, we first
design the gain K ’ , and then we go towards designing this compensator with the
dynamics part ok.

(Refer Slide Time: 03:31)

708
So, how does it look like when I do the bode plots for a typical value say, β=10 my
bode plot looks something like this. So, I’m not really accounting for the gain here, right.
I’m just looking at this part. And therefore, my magnitude starts from the 0 dB line.
And then it goes down and just like this. Similarly, with the phase lag right, this is
essentially why it is actually called a lag compensator.

So, the first observations here is that it is a low pass filter right. So, the gain is 1 or
0 dB for low frequencies and then it keeps attenuating the higher frequencies, okay.
So, first, it’s a low pass filter because it provides attenuation at higher frequencies. And
we will see how this actually results in providing a sufficient phase margin.

And as we see here, this phase lag characteristic is of little consequence as the pole and
zero are close to each other and then, even if you look at in the root locus design in the
lag compensator, we wanted the angle contribution of the lag compensator to be very
small. And therefore, we said that while designing I should make sure that the angle
contribution of the lag compensator is less than 5 ° . Okay, many textbooks say it is of
no consequence, but that’s not completely true and we will see why that is important.
And I say and I’ll also tell you why I use this word little. It has some effect, but a small
effect.

(Refer Slide Time: 05:04)

Okay now, I will teach you this thing with the help of an example. So, that the steps are
clear right. If I just write down the steps and then do the example later it may not really

709
match, but we will motivate ourselves and learn the procedure with the help of an
example. Okay so, I’m given this open-loop transfer function. And my closed-loop
design specifications are the phase margin should be 40 ° , the gain margin should be
10 dB , together with a steady-state specification of 5 ok.

So, I just look at what the plant by itself means, well it has a phase margin of 32.6 ° .
And the gain margin of 9 ° . So, gain margin is fairly okay, phase margin is a little out
of what I have is 32.6 ° and what I want is 40 ° , Kv is really bad; I’m having 1,
but I want to change it to 5.

(Refer Slide Time: 06:05)

So, the first step while we design in frequency domain is to adjust the gain. So, let us
again look at the standard structure of the compensator, of the lag compensator in this
case where Gc ( jω) , I separate that into the gain part and then the dynamics part. So,
first is to adjust the gain K ’ to meet the steady-state requirements. Now how does the
steady-state error change with lag compensation? Well, I look at K v , so this is my
formula to find out K v . And this Kv is K’ with this this K ’sG (s ) . This
K ’=Kβ . Okay so, this is actually Kβ .

So, now while I do this thing, so what I find is that lim G'c ( s )=1 , because of this
s →0

construction I put zero, zero here and what I’m left with is 1. So, on the left-hand side, I

have K v , now this guy is K ’ lim sG ( s ) . Now, this is, what is this?
s →0

710
 uc
lim sG ( s )=K v of the uncompensated system right. Let me just call this uc ok. So
s →0

this is Kv of the uncompensated system in general.

Now, here it turns out that this is 1, okay and the desired K v =5 , and therefore, K’
which is the ratio of the desired Kv to the original Kv turns out to be 5 . Okay,
now again let’s see through the construction also right. So, this is my compensator design
here. Gc , K and then the K also includes this β . So, this is K’ and so
on. Now when we do the design later, I will tell you what is exactly the relation between
you know, K’ because we are actually not yet calculating the value of β , but we
are only calculating the value of K ’ , which is actually βK ok.

So, here what I have is, this is the Kv desired over Kv of the uncompensated
system is K ’ , and this is 5 and this 5 is actually equal to the K β , okay.
Now with this K ’ , well I just plot the bode again, and I see that something strange
has happened. Well, K’ is actually 5 that’s perfectly alright. But well the phase
margin is a −13 ° , and the closed-loop is unstable. And I really wouldn’t want this
right. So, a steady-state error improvement means nothing if the system is unstable.

Okay now, what I would like to do and you see also these frequencies are typically not
very high frequency; they’re like 1.8 , you know 1.8 . Even the cross over
frequency is one you know this is; so at this, this cross over frequency is around one
point 1.5 or something, okay. So, what I would want to do is well, you might think
why not just use a lead compensator over here and push it upwards right. So, that my I’ll
have the desired phase margin. Well, first is this frequency ranges which we are talking
about are very if you see is low-frequency ranges, and therefore this kind of thing may
not necessarily work.

711
(Refer Slide Time: 10:13)

Okay so, how would I do this? Okay so, the idea is to; so there are the other way of
looking at this is also, that if I look at a point say somewhere here okay, at this point
where the angle is −140 ° , okay. Which means if, at a certain frequency right, I can
actually map that frequency to somewhere here, may be between 0.1 and 1 and I’ll
actually compute what the frequency is. So, if I pick up this frequency I go here and I
make this frequency at which the angle is −140 ° , this will correspond to a gain
margin of 40 , okay. If I pick this frequency go here and I say make this the new gain
cross over frequency ω gc , then I’m fine right. Ok

So, all I need to do is to identify this frequency and pull this magnitude to zero here, such
that this will be the new gain crossover frequency. Now that is what is happening in a in
this lag compensator right. I’m actually having a negative number in the magnitude right.
So, if I superimpose this over here it could have an effect of pulling down the overall
magnitude, okay. So, that’s what we will do.

So, the idea is to choose a new gain cross over frequency where the phase angle of the
uncompensated system is −140 ° . This uncompensated is I’m also looking at
uncompensated system with the gain adjusted one. Gain adjusted in such a way that I
have my K v =5 right. I will talk everything after I adjust the gain. Okay now, this

712
frequency turns out to be 0.631 like here right. So, somewhere here this is about
0.631 ok.

Now, I’ll find out what is the magnitude at this frequency of the uncompensated system,
which is essentially I’m looking at interested in this magnitude, right this one. Okay now,
we at this frequency, we need to bring the magnitude down to 0 dB , okay. At this new
gain cross over frequency. This means that the lag compensator must provide an
attenuation of −16 dB . By the construction of the lag compensator, so the β can

1
now be designed in such a way that, well the magnitude 20 log =−16 , right. That is
β
well this magnitude right.

This is the magnitude which I should compensate via a lag compensator. Okay so how do
we go about doing this? So, if we look at the structure of the lag compensator, in lower
frequencies you see that you know it’s just for a ω very small of this order of say
10−4 or even less. The gain is approximately about 0 db line. And if I keep on

1
going to the higher frequencies I see that the gain is can be approximated by , okay
β
that can be seen from here also and if I just look at this magnitude this would be

1
20 log which I equate to −16 in my problem right, so this is what was required
β

1
so 20 log =−16 in such a way now that at this frequency, I want the total gain
β
adjusted plant plus the lag compensator to be 0 db in such a way that this frequency is
now my new crossover frequency. Okay now, why does this work right? So, it’s just an
approximation, you just take a, for a larger values of ω this approximation works,
even if you look at this particular example where β=10 you just see that this
magnitude is about −20 decibel at even the frequencies starting from 1 hertz and

1
that’s the reason we equate this number −16 to 20 log , to get the new gain
β
crossover frequency such that the magnitude here becomes zero.

Now, it turns out; well if I just do this that β=6.31 . Okay now, if I choose the first
corner frequency to be 0.1 rad /sec or T =10 . So, if I look at the construction,

713
what are my unknowns? Unknown is the β and unknown is the T . If I contrast
with the lead compensator there is no real constructive procedure to exactly find out
what this T and β are. I can explicitly find out what is β that’s much easier for
me. T , I will just choose to be closer to the original, that’s all I know that these guys
should be closer to the origin, even from the root locus. So, I select T =0.1 , and then
the desired compensator will now be something like this, okay. And I see that the new
phase margin is well where am I? I’m here at this new gain crossover frequency, I’m at
32.1 .

(Refer Slide Time: 15:55)

And, okay this is the gain margin, gain margin is okay I don’t really see how much is it,
but it’s like close to 10 to 11 that’s fine. Okay, but what I what do I really want is
this to be 40 ° , not 32° . So, where am I going wrong in the design procedure?
Now, this error is due to the lag angle introduced by the compensator. If I go back to the
construction you see that, well there is some bad effect going on over here. So, maybe
for a simple construction say I’m looking at a frequency of 0.6 , and you see there is
some 8 ° to 10 ° which is lost because of this one. It’s not significant, but still, it is
small, but it’s of consequence. Therefore, I say it’s of little consequences it’s not of no
consequence. So, somehow I have to compensate for this.

714
(Refer Slide Time: 16:47)

So, in order to compensate for this what I do is to add an angle 5° to 12° degrees
to the phase margin. Based on some trial and error in such a way that I don’t really look
at this number now. But I aim at a slightly bigger number, so not −140 ° , but say
somewhere around −15 0 ° sorry, −13 0 ° so that my phase margin is 50° ,
okay. Just to compensate for the loss of angle because of the lag compensation, because
of the angle criterion of the line or the or the phase nature of the lag compensation. So, I
just say instead of 40 , I choose it to be 52 . And this 52 I’ll find out at which
frequency it occurs somewhere here and it turns out that this 52 occurs at a frequency
of 0.465 . And at that number, at that frequency 0.465 what is the magnitude?
That’s exactly the same procedure. I found that what is at one here and then I go here
right.

So something is error, ok whatever. Okay so, here I see that the magnitude is 20 dB it
was 16 earlier now it is 20 . So, I need a different β than earlier. So, β
would be 10 and I just now just again choose the first corner frequency to be 0.1
'
and I have this G c . And let’s see how the closed-loop response is like. Well, the
closed-loop response has a good gain margin. So, somewhere over here and then, now
the phase margin is quite satisfactory it’s 41.6 , and of course, the closed-loop system
is stable right.

715
So, the design procedure is straight forward, I don’t now really need to write down the
rules for you. Well, the rules are pretty straight forward right. So, you start with adjusting
the gain first, and you see well what is happening. So, I want the signal to be attenuated,
such that I have a new gain crossover frequency, which is lower than the original one.
I’m just you know shifting this gain crossover frequency to the left so that I will have a
better phase margin, right? And of course, this is just now design procedures, I add an
extra angle between 5 and 12 , to compensate for the lag introduced because of the
phase nature of the lag compensator.

(Refer Slide Time: 19:18)

Okay now, if I compare the time response plots I see a decrease in the overshoot right.
And this uncompensated essentially means that I’m not even looking at the gain adjusted
system; I’m just looking at the original plant. So, in this in the left-hand side, I’m just
looking at the original plant right.

Okay so, if I again go back to things, well I had said that well with K ' =5 , my a
system turns out to be unstable. So, here let me look at the root locus plots; well what
happens to the root locus plots. If I look at the uncompensated system like this, now is
the gain adjusted system. So, this, these colors here and these colors here are they have
no relations, these are just 2 different plots.

So, if I look at the blue plot which is the gain adjusted system at a gain of 5 , this is
'
what we found outright, that K =5 , I see that my root locus shifts to the right, and

716
therefore, my system is unstable. Now I introduce a lag compensator and with the same
gain, I’m now here right. My system is now stable and I see the effect of lag
compensator right, it just shifts the root locus slightly to the right okay, but then the
things are taken care of here, right that I’m actually ending up with a stable system, okay.

(Refer Slide Time: 20:53)

Okay, so to summarize at lower frequencies, the lag compensator provides an additional

gain. So, when I’m look saying of this additional gain I mean that I am looking at the
'
plot of, this is my entire compensator right. K times this thing right. This is the entire
compensator, I just split it into the gain part and the dynamic compensation part right.
So, if I look at the entire plot this will just go up with 20 log K depending on again
what is the value of β for example.

Okay so this is, how it goes and at higher frequencies, it has a smaller gain. And
therefore, the lag compensator acts like a low pass filter right. In such a way that it
reduces the gain at the high-frequency range and improving the phase margin, right.

z
That’s what the example also told us. Okay, and the ratio is chosen to meet the
p
steady-state specifications, right. So again the steady-state specifications were through
this gain and you see that this gain was equal to Kβ .

717
Now, what, where is this β coming from? This β is coming from the construction

zc
or even from here that is the ratio of the zero to the pole, right if I call this this is
pc
β . Okay so, these are not disassociated from each other right, this β actually has
to do something with here also. But just that I make my design procedure easy to look at
it as at the gain independently and this part independently, but this also has to do with
β , right; this directly sitting in over here. Okay.

And of course, the lag compensator reduces the bandwidth of the system because the
gain crossover frequency now shifts to the left. And we always claim that the poles and
zeros should be placed close to the origin ok.

(Refer Slide Time: 23:04)

So, how close to the origin? So, and what happens when I place poles and zeros close to
the origin or even much closer to the origin, is that the lag compensator it creates an
additional closed-loop pole-zero in the same region. So, it creates an additional closed-
loop pole in the same region as the zero and the pole of the lag compensator. And which
and this pole, if it is closer and closer to zero, will create some additional dynamics, for
which the settling time might be higher as we go close and close to the origin.

Okay, what does that mean right? So, without going too much into the details, say if I,
say just for trial basis just select another value of T right, to be 20 with this
actually we’ll place the poles and zeros a little closer to the origin, in that case, well I

718
think my gain margin and phase margin are good, but now look at this right. So, if I’m
away from the origin I see that here I’m settling down little faster right. So, this is what I
had with earlier with T =10 right. And I say I call this away from the origin,
relatively away and this plot is for T =2 0 .

And you see that the settling time is actually higher if you just zoom in here you could
actually see right, there is a bit of a difference in the settling time. And that is due to the
additional pole which is created because of the lag compensation, you could just check
for the closed-loop poles in Matlab and that’ll actually give you the exact locations of the
poles. That’s not really important, what is important is to just understand these
phenomena, how close do we go to the origin.

Now, closer to the origin also means that I’m actually looking at a PI kind of controller
right. So, this lag is an approximation of a PI controller. And one drawback of this PI
controller was that it could lead to instability, as we have seen right. In in in the module
number 7 and even in 8. That this could actually lead to a closed-loop unstable system.

(Refer Slide Time: 25:27)

So, these are the things which we need to be careful of, apart from that the process is
very simple. So, next lecture what we will see is to design lead-lag compensators and this
is a little philosophically different than when we do in the root locus or in the time
domain, where an obvious interpretation was if I were to improve the steady-state
performance, I design a lag compensator, if I were to improve the transient performance,

719
I go to a lead if my specifications are a combination of both then I directly go for a lead-
lag compensator.

Over here it’s a little different because we have additional things introduced into the
system like the bandwidth. And we’ll see how we make use of lead and lag compensators
to achieve little more objectives than what we had in the time domain, and see if it is
straight forward or what are the pros steps involved with that.

Thank you.

720
 Control Engineering
Dr. Ramkrishna Pasumarthy
Department of Electrical Engineering
Indian Institute of Technology, Madras

Module - 09
Design using Bode plots
Lecture - 04
Design of Lead-Lag compensators using Bode plots

In this last lecture of design with using frequency-domain methods or in particular the bode
plot, we will look at designing both lead and lag compensators.

(Refer Slide Time: 00:30)

So, just to again recap what we did earlier, was you have this region 1 which was the low-
frequency region, where we want the gain to be as high as possible so that we have low
steady-state errors. In the region 2, it should have some satisfactory gain and phase margin.
So, this corresponds to the transient region, the steady-state region, and the high-frequency
region.

721
(Refer Slide Time: 00:57)

Okay so, what is this lead-lag compensator all about?

(Refer Slide Time: 01:03)

So, if we remember what we had done while we were doing the design with root locus
methods. We had a component which was essentially a lead component and the lag
component and of course, taking care of all the lead and lag components and β should be
1 for something, β> 1 and an α and β here are 1 . If we directly map to what
we had done in terms of that one, αβ =1 . So, I have a lead and the lag components and
essentially I’m now looking at a designing this compensator with a gain K
'
and Gc is

722
again split into the gain component and then the dynamic component. So, something like this.
'
So, as usual in the design process, we design the gain K first and then go about designing
this one.

(Refer Slide Time: 01:58)

So, let’s say how the frequency response or the frequency characteristic of this lead-lag
compensator looks like. So, if I just draw the bode, so what I see is well I have the lag
component here and then the lead component; I could also see it with the magnitude plot. So,
this response to my lag where you’ll have attenuation at higher, at these frequencies for
example, and you’ll also have this to be the lead component.

So, what we observe here is, of course, these are my corner frequencies 1, 2, 3 and 4, and this
is this will these four corner frequencies will decide how my design of the compensated look
like or how the poles and zeros of my compensator are placed, and if I look at the phase plot,
there is a certain frequency over here at which the angle goes to 0 . Now, what is this
angle? this angle, the frequency at which the phase angle is 0 , if I call this as ω1 , the
frequency at which the phase angle is 0 is this is one.

Now, how do we compute this, now what are this T1 and T 2 ? This T1 and T2

essentially come from here, the T1 corresponds to this part of the compensator, T2
corresponds to this part of the compensator. So, how do we do this? So, I just quickly run you

1 1 β 1
∠ Gc ( jω )=∠ jω+ + ∠ jω+ 723
−∠ jω+ −∠ jω+
T1 T2 T1 βT 2
 1
through the proof. So, you have the ∠ Gc ( jω) is so you have the ∠ jω+ . Similarly,
T1

1
you have the ∠ jω+ , now the angle of the pole. So, the negative sign
T2

β 1
−∠ jω+ −∠ jω+ . i.e.,
T1 βT 2

ωT 1
This could also be written as the (
tan −1 ωT 1 +tan −1 ωT 2− tan −1
β
+ tan−1 ωT 2 β ) . So,
there I will just split these two quantities this guy and this one, and I just use the simple

tan a+ tan b
formula which we learn in trigonometry. So, tan ( a+ b )=¿ .
1−tan a tan b

So, if I take the tangent of this entire guy, what will I will be left with is I have

ωT 1
ωT 1 + ωT 2 + ωT 2 β
. Similarly, on this side, I will have β . Now, what happens
1−ωT 1 ωT 2
1−ωT 1 ωT 2

1
when this is true, when ω1 or at a frequency over ω1 = , I substitute it over here
√T 1T 2
and I see that the denominator tends to 0, denominator tends to 0 here also. So, what I will be
left with is ok, so the tan of this entire first term is ∞ , if I take the tan
−1
what I get

π
is that whatever is remaining inside has an angle of similarly here. So, the tangent of
2
these two terms added up would be ∞ .

π π
Therefore, whatever terms in the bracket is . So, I have a . And this happens only
2 2

1
at these frequencies. I am looking at the situation where ω1 = . And at this
√ T 1T 2

724
 π π
frequency, I get if I substitute it here, I get a − , and therefore, the angle contribution
2 2

∠ Gc ( jω ) ¿ =0
the ω1=
1 .
√ T1 T 2

(Refer Slide Time: 07:24)

So, just a little summary of what we learnt so far. So, if I look at lead compensation well as
the name suggests it provides a phase lead which means, it improves the stability margins. It
also has a higher gain cross crossover frequency which means it has a higher bandwidth. And
we know what is the relation now between bandwidth and the damping ratio that suggests
that higher the bandwidth which would mean lesser settling time. And of course, it needs an
additional gain increasing gain to offset the attenuation, the α< 1 , therefore, we have to
put some kind of a gain there it so as to offset this attenuation.

So, the lag compensator it reduces the system gain at higher frequencies in such a way that it
does not affect the system gain at lower frequencies. So, only the higher frequencies are
attenuated. Reduced bandwidth obviously mean slower response and it helps in steady-state
accuracy right as we see in also in the time domain. Because of this property of reducing the
system gain at higher frequencies, it helps attenuating the higher frequency noise and another
drawback was well this pole-zero combination near the origin. So, the compensator poles and
zeros in the lag case are placed very close to the origin and this will lead small some kind of
variation in the transient response in terms of the settling times, it will have a long tail with
small amplitude in the transient response.

725
(Refer Slide Time: 09:03)

So, now if both fast responses and good static accuracy are desired, then you go for a lead-lag
compensation when increase in the low-frequency gain would mean improving steady-state
accuracy. And of course, the bandwidth and stability margins can also be increased when we
do the lead and the lag design together. So, let’s start with an example to see what I am trying
to say here that I am given a transfer function, which looks like this and it is desired that
−1
K v =30 sec . If I write down the units, it is per-second, phase margin
≥50 ° , and the bandwidth of certain number of 12rad /sec .

So, what I will do here is, I will try to solve this in MATLAB; I just plotted up everything for
you. And we will see how the design process goes. Many times while designing I had
mentioned that, well the design is not a one-step process; you may have to do it in over and
over until you reach closer to the solution. So, in the first iteration, you may be close to the
desired results, but not then; in the may be in the secondary iteration you may tune your
parameters which essentially in the β , may be you can choose a different α and β
as we did in the time domain analysis to just tweak it a little bit. So, that you get the desired
response. So, I will just quickly run you through that process of going about that.

So, before that right so we are given this K v , now K v =30 sec−1 . So, what is the Kv

? in my case Kv in my case is the standard formula lim sG( s) that becomes K

s →0

here. So, for K v =30 , this guy K should also be 30. So, I am looking at a gain

726
compensated transfer function or the plant also right, the gain compensated plant which now
looks like say 30 and the rest all you may say, jω(0.1 jω+1)(0.2 jω+1) . Now, let’s go to
MATLAB and try to see what this gain means in terms of stability margins.

(Refer Slide Time: 11:40)

So, I have g=30 and here I will have 0.02 , 0.3 , 1 and 0 right, a transfer

30
function, 3 2 , this is the transfer function for this guy.
0.02 s +0.3 s + s

(Refer Slide Time: 12:24)

727
So, if I do the bode plot of this to see, what does the gain compensation do to me. So, I have
plot here. Okay so, I have the grid, the characteristic all stability margins. Okay so, I have
phase margin of −17 and negative gain margin and it shows that the closed-loop system is
unstable, something not very good has happened here. So, and also if I look at what is the
bandwidth, so then at what time am I reaching −3 . So, look at this go here, here. So,
roughly about eleven some something about eleven point what you say which would say
about 11.3 rad / sec . So, let’s just note that down.

So, the gain compensated system has, well, this is first this is unstable; has a negative phase
margin and negative gain margin, and also it has a bandwidth of and a roughly about
11 rad /sec . Ok, so this bandwidth is what I measure looking at the open-loop bode plot
with a gain K=30 , so specifications are usually in terms of the closed-loop bandwidth
and in the previous lectures we saw how closed-loop was in a way related to � with some
kind of complicated formula, ok so this bandwidth of 11.8 radians per second is for the open-
loop, how does the closed-loop bandwidth look like? Well typically ah the closed-loop
bandwidth is if I say denote this as the closed-loop, it is usually greater than the bandwidth
for the open-loop, so if this is 11 here the closed-loop bandwidth would be roughly about
13 . So, we can check this with the help of the closed-loop bode plot or if we were to do
the design through the design process ah check the bandwidth we would we may have to use
complicated plots like the Nichol’s chart but we will try to avoid those. So, what we will here
do is I mean time and again we’ll measure the open-loop bandwidth and typically the closed-
loop would be about 20 higher than the open-loop ok, there is no reason for me to say
20 but I will just make sure that my design for the open-loop is around maybe 9 or 10 so
that the closed-loop is close to 12, right. So, this we may have to check, you may take the
bode plot for the closed-loop system and see that the bandwidth for the closed-loop system is
actually 13 rad /sec . However, this is not very useful to us because I’m just here dealing
with an unstable system. Now, let’s just try what happens we just using a lag compensator. I
will just write down the steps here. Let me just take arbitrarily any ok let’s start with say a
lead compensator.

728
(Refer Slide Time: 15:52)

1+ jω 0.25
A lead compensator, So, let me say it looks of this form , I take a very
1+ jω∗0.025
liberal α which gives a very big phase lead of say 0.1 . So, now let’s see what happens
with this, with this lead compensation. So, there is no basis for me to choose this, but just to
see a basic analysis what it is.

(Refer Slide Time: 16:35)

729
So, this is my gC 1 , I will call it is as a transfer function. So, it has one sorry 0.25 , 1 ;
0.025 , and 1 . So, this is how my compensator looks like. So, now, if I do the bode plot
of G sorry, g with gC 1 which is my lead compensator, well I get something like this.

(Refer Slide Time: 17:24)

Okay I see, I have some improvement in the phase margin, good thing is that the closed-loop
system is stable it has a phase margin of 10 . And gain margin of about 3.5 db, and not
bad because at least I could get the system close to stability and the bandwidth if I go back
bandwidth of close to 20 ok, roughly about 20 . So, let’s just note this down. So, just
by using a lead compensator what I achieve is well the closed-loop system is stable.

Now, what is the phase margin, phase margin is about what was that phase margin was about
10 ° and bandwidth was fairly large 20 rad /sec . So, what is good, well this is good,
this is these are this is a good thing, but it is far from desired, desired was more than 50 .
so this is just about designing with the lead compensator. Let me just do with the help of a lag
compensator. And just take some standard, this one some standard compensator. Let’s say it

1+ jω10
could be like . So, let’s see what MATLAB says because of this.
1+ jω100

730
(Refer Slide Time: 19:30)

So, let me call this gC 2 , so I have 10 jω+1 ; 100 and 1. Now, if I do the bode plot of
this plant with the lag compensator.

(Refer Slide Time: 19:56)

And it looks something like this, there it is. And I am looking at the stability margins. So, I
see that my phase margin is quite good now, it is 42 a good enough gain margin. Now, let’s
say what is happening to the bandwidth, when I am at −3 db is roughly about 3 point;
about 3.4 rad / sec . So, with this my phase margin has shown some improvement. Well, the
closed-loop system is stable which is good. My phase margin has become almost like more

731
than 46 ° then I have the bandwidth is 3.3 something like this, it goes to
3.4 rad / sec , okay.

So let’s now compare these two scenarios so if I look at the lead compensator the good thing
was well the closed-loop system was stable, phase margin is 10 , earlier with the gain
compensated system of K=30 , my system was unstable now it is at least stable, the
bandwidth is 20 so this again I’m measuring for the open-loop and if I just measure for
the closed-loop; the closed-loop bandwidth is about 26 rad /sec and this thing is too large
right, and then the design specification in the previous slide here was in terms of the closed-
loop specification so this is too large and therefore it will be sensitive to noise, okay. Now,
this is about the lead part of it, now what does the lag compensator do? Well, good thing is
the system is again stable, phase margin is close to the desired, it’s 42 , 46.2 and what
we want is about 50 now, the bandwidth is is 3.34 this again for the open-loop, the
closed-loop would be about 4 rad /sec or even less anything is like far from desired right,
so here I’m not really computing this but I will just use the open-loop bandwidth as an
indication of what is happening in in in the closed-loop system so comparing these two I see
that well I have stability in both cases phase margin is good but the bandwidth is too large on
the right-hand side for the lag compensator system is stable which is good phase margin is
close to desired but the bandwidth is too small so you see there is a little trade-off here right
if I increase the phase margin the bandwidth goes down and if I want to increase the
bandwidth the phase margin goes down so I may have to go for the lead plus lag compensator
under such scenarios.

732
(Refer Slide Time: 23:04)

So, as a rule of thumb with may not be very proper justifications, we first design the lag
compensator. So, partially or partial compensation is via a lag compensator. And if you get
the system to behaves slightly better, then we can use the lead to overcome the deficiencies
where the lag compensator; say here, the bandwidth increases in phase margins are ok kind of
thing then if it is like; say the desired is 50 , this guy gives me 30 . I can you know
compensate this with a lead and also the bandwidth can in can be increased because the lead
naturally has this tendency to increase the bandwidth of the system. So, how do we do this?
So, we will so sorry first thing we need to choose is; what is the new crossover frequency. So,
I will again go back to my plots.

733
(Refer Slide Time: 24:30)

So, I go back to my plot here, which is just the gain compensated system with K=30 and
I choose just by trial and error some partial compensation. Let me say I choose this frequency
here where the phase is −145 which means the phase margin is about 35 ° . So, when
will this become the phase margin when this frequency 3.52 is actually the gain crossover
frequency. So, I go here. So, roughly about this point, I want the magnitude to be 0 . What
is the magnitude here? it’s about 16 db .

So, this is the lag compensator should provide a negative gain here of −16.5 such that this
is the new gain crossover frequency. I am just assuming some partial thing you can also do,
go to 150 and check, you can also go to one −140 and check, so no hard and fast rule
here. So, what I need to do is compensate for this magnitude via the lag compensator, or in
other words, I need to first design the β . Ok.

So, how is the β designed? Well, I am looking at now 20 log β=16.4 in decibels this
means that log β=0.82 and which essentially means β=6.6 , and let’s assume that it is
7 . Now, now I need to assume or find out what or assume something for the corner
frequencies. Let me just start by assuming T 1 =1 and with β=7 ; T 1 =1 , my lag
compensator if I call it Gc lag would look something like this. And I have yes, sorry

1+ jω
. Now, let’s see what happens when I just plug in this compensator into my plant, I
1+7 jω
will call this as my gclag . Okay so, something like this.

734
(Refer Slide Time: 27:34)

And then I look at the bode plot of the lag compensated system. So, g is the gain adjusted
and then I have gclag . So, what is happening with this now? So, I have now a phase
margin of up to 22° , it is fine and I have a gain margin of 8.2 and well let’s check
what is the bandwidth, the bandwidth is it’s like very small say about 4 4 point, say close to
4.3 , ok. Now, I have this lag compensated system which now has a phase margin of
22.5 and a bandwidth of about 4 four and half.

Now, what I know now is that I could use the lead compensator to provide an additional lead
here; and with the lead compensator, my bode plot also shifts to the right. So, I could improve
the phase margin and also improve the bandwidth. So, first what I did is to partially
compensate with the help of a lack compensator and I will just do the lead part now ok. So,
how we will go about in the lead part? The step two would be design the lead part because
this compensator had given me a phase margin of 22 together with a bandwidth of
foremost about 4.3 rad / sec . Ok.

So, now the lead part; so what is that? β is fixed and αβ =1 . So, I know what is α

( )
1−
−1 7
now? So, the φm , if I write it in terms of this β , would be φm =sin . This is,
1
1+
7
comes out to be 48.6 ° . Now, what will change is there should be a another new crossover

735
frequency, ok. So, this new crossover frequency, now we have to adjust this into the lag
compensated system; lag compensate about this one. So, these Gc where G∗Gc lag
where G was a gain at this is just a system with gain K=30 .

So, once I know φm , I also know α , I can compute α or in this case; I know β ,

1
where α= , I can compute the new crossover frequency, where the magnitude of the lag
β
compensated system take this value −10 log β . So, this is β=7 now; so
−10 log 7=−8.45 and this 8.45 occurs at ω m , I will show you this one in the bode
plot is about 6.15 .

Now, once I get this ω m , so I know α or β ; I know ω m , I know φm , I can

compute T 2 =¿ √ β = √7 =0.43 .. So, if you remember the formula, it was like

ω m 6.15

1
=ωm √ α , let’s just go back to the slides of the lead compensator you’ll get to know
αT
this.

So, I am looking at this magnitude this where the lag compensated system has a magnitude of
−8.45 that will be the new crossover frequency. So, that is turns out to be 6.15 as we
will see from the bode plot, bode plot of the lag compensated system.

(Refer Slide Time: 32:25)

736
So, that is it −8.45 is roughly around here right 6.15 right, this is the frequency here
right. So, now, based on these things I can write down how my lead part of the compensator
would look like.

(Refer Slide Time: 32:43)

So, the lead compensator has now the form based on ω m=6.15 , β=7 , T 2 =0.43 .
This T2 is essentially that the time the T the time constant of or the numerator part of
the lead compensator that which we had here. So, the unknowns in my design were T1 ,
β , and T 2 . So, now I am actually going to find out what is T2 and by doing all this
process, I get T 2 =0.43 , which gives my Glead , if I call it Gc lead to be of the form

0.43 jω+1
. Let us see how the overall compensation now looks like. So, I have
0.0614 jω+1

30 1+ jω
G=¿ , I had the Gc lag which looked like and I
jω(0.2 jω+1)(0.1 jω+1) 1+7 jω
have the lead.

737
(Refer Slide Time: 34:36)

So, let’s see how my bode plot has change when I plug in this lead compensator. So, my
gclead is 0.43 and a 1 ; 0.0614 , and a 1 , something like this. So, my overall
compensated system with the lead and the lag compensator I have gclag and I have a
gclead . Ok.

(Refer Slide Time: 35:25)

So, what are the characteristics now? First, ok, phase margin is 48.1 . Let’s look at how is
the bandwidth doing, well bandwidth not doing too well, it is roughly about 8 rad /sec . So,
with this, the phase margin is 48 but the desired was 50 was to be greater than or equal

738
to 50 and the bandwidth is about 8 rad /sec . Ok, again this bandwidth is for the open-
loop case the closed-loop bandwidth would roughly be about 11 rad /sec and we could
check this by the bode plot of the closed-loop system and the desired for us was 12 . Now,
see the first the steps involved. So, what did we do is with the lag compensator, we had some
partial compensation where we at had the closed-loop the gain compensated system with the
lag compensated to be stable, and a decent enough phase margin. We still far away from what
is desired and as some amount of bandwidth.

Now, I know that given this configuration I can now use a lead compensator to increase the
bandwidth and also to push up the phase margin. So, here the design involved choosing an
appropriate β then I had to choose T 1 and I was also taking this special case where this

1
α= , so given an α , I had to choose a β which satisfied this relation so and if you
β
look at the design what does the design give me well I’m close to the desired phase margin of
48 , I am close to the desired bandwidth which is close to 12 , here I’m at 11 ; so
what I could do is I could design the lead and lag component separately, this essentially
means choose an α which is not really depending on β or just fix the α what we
had earlier you can choose a different value of β to be 9 or 10 or whatever and then you
can just come back to this so what is required, so now we are close to the to meet the closed-
loop specifications so what are required here are just little bit of tweaking of these values of
α and β so that we get to the exact values not as close but really greater than 50 on the
bandwidth about 12rad /sec . so I will not run you through that process I will leave that to
you as an exercise to play around with different values of α or even designing the lead
compensator completely separately, the aim of this lecture was to run you through that
process of the design, well how to possibly use MATLAB appropriately or how to read the
MATLAB plots and so on.

739
(Refer Slide Time: 38:45)

So, to summarize what we learnt today was to design compensators in the frequency domain
or across this module. In the 10th module, we will do something, more something little
different. So, these are not necessarily a part of a regular control curriculum, but these are
important and interesting and also useful things to know. The first is experimental
determination of a transfer function.

So, given a system for which I have no physical model; physical model essentially means I
cannot realize them in terms of any circuit components or mechanical components. Can I just
look at its frequency response generator a bode plot and get the transfer function we will do
the later part of it? Given bode plot can I generate a transfer function and that leads to a very,
very nice concept what we should be careful of is that of non-minimum phase systems.

So, these are essentially the two things which we will deal in module 10, might be a smaller
module in terms of the length of the lectures than compared to the other models, but still very
informative and useful enough.

Thank you.

740
 Control Engineering
Dr. Ramkrishna Pasumarthy
Department of Electrical Engineering
Indian Institute of Technology, Madras

Module - 10
Lecture - 01
Experimental Determination of Transfer Function

In this module we will learn a couple of things. So, we will start with the experimental
determination of a transfer function. So, what we had learnt earlier in terms of modelling was,
you have a system you first build a physical model of the system. Physical models would
mean that you realize it with some basic components, if it’s from electrical domain you will
have resistors, capacitors, inductors. From mechanical domain, you will have equivalent
elements in terms of mass, spring, dampers. You could also model electromagnetic coupling
and so on.

And once you have the physical model, what you would do is write down the physical laws
like the KCL, KVL, the Newton’s laws, and arrive at a mathematical model. So, another way
to do, is what if I do not have any information on the system. I may not be able to realize it
with physical components and therefore, I may not be able to directly write down the transfer
function in terms of or first by writing down the KCL and KVL kind of things.

(Refer Slide Time: 01:26)

741
So, can I do something else, can I start with determining the transfer function with the help of
experiments. So, we need to be careful while we do that. So, and that leads first to define the
concept of minimum and non-minimum phase systems. So, I will just briefly run you through
these things before I tell you why we need these things. So, systems which have neither poles
nor zeroes in the right half s plane are minimum phase systems. So, far we characterized
poles being on the right hand side and identified them with essentially unstable systems. We
never talked about zeroes either on the left or on the right, all we knew while we were
drawing the root locus was that the poles go to the zeroes and so on.

So, now what are these non-minimum phase systems? So, these are systems which have some
amount; some number of zeroes, finite number of zeroes in the right half s plane. So, how
do I identify when is a system a minimum phase or a non-minimum phase. So, there could be
several systems with the same magnitude characteristic for example, if I and we also revisit

s+ 1
these example. So, if I have a transfer function which is like and I compare it with a
s+ 2

s−1
transfer function . If I plot down the magnitude characteristics both of these systems
s+ 2
will have the same magnitude characteristic ok.

Now, how do I identify which is. So, by just by looking at the magnitude plot, I may not
know if it is a minimum phase or a non-minimum phase. So, what happens in case of non-
minimum phase systems is that, the range in the phase angle of the minimum phase transfer
function is minimum along among all such combinations right, all such systems which have
the same magnitude characteristics. So, you have same magnitude characteristic here and
here, how do I identify which one is the minimum phase, you just look at the range in the
phase angle of this guy, and the range in the phase angle of this guy. And you take that one
which has the minimum range right. The range in the phase angle of the minimum phase
transfer function is the minimum along all such systems right.

So, I’ll show you some plots also shortly. So, while the range of any other thing would be.
So, if this one is the minimum any other thing would be the non-minimum characteristic. And
now for a minimum phase system the transfer function can be uniquely determined by the
magnitude curve alone. So, when I; so once I know this I can do something nice right, if I just
give you a magnitude characteristic, if this is my frequency, this is my magnitude; magnitude
G( j ω) it looks something like this.

742
Then I know that this essentially what I do is a bode a plot, and with the corner frequencies
the gains at say ω=1 , I can actually write down the transfer function here. So, for that I
need to first make sure that I know that my phase would be of minimum phase. So, for a
minimum phase system the transfer function can be uniquely determined from the magnitude
curve alone. So, I will assume that whenever I am doing this experimental determination that
my system would be minimum phase. So, more on this we will keep coming.

(Refer Slide Time: 05:33)

So, what are the other ways how of distinguishing between minimum and non-minimum
phase. So, for a minimum phase system the magnitude and phase characteristics are uniquely
related. So, just take any minimum phase like this. So, given the magnitude plot there will be
a direct one to one relation with the phase plot. So, if the magnitude curve is specified over
entire frequency range, then I could also find the phase angle curve. So, non-minimum phase
systems, what do they or why are they not really good for us, and I will spend an entire
lecture on analyzing these three bullet points which are coming up.

So, non-minimum phase systems are slow in response because of the faulty behaviour at the
start of response. So, if we remember while we were in one of the earlier modules I think in
number 7, we were looking at the role of zeroes in speeding up the response. There was also
one plot where we talked about a zero being on the right hand side, where we experienced
and an under shoot. So, I will spend time on that in the next lecture. For the moment we will
concentrate on determining experimentally the transfer function of a system.

743
So, a common example we will again come to this examples is that of a transport lag systems.
So, when can non-minimum phase situations arise practically? Well, if the system includes
non-minimum phase elements right. So, I think we will do this; revise these examples or
revisit these examples in detail, or when internally there is a minor loop that is unstable.

(Refer Slide Time: 07:19)

Okay, now what are further techniques just look at the transfer function or the phase
characteristic and the magnitude characteristics, is there any way can we can identify between
minimum and non-minimum phase systems. So, for a minimum phase system as the
frequency tends to ∞ right the phase angle becomes −90° ( q−p) . Well, this is just a
difference in the number of the poles and the zeroes right? Where p and q are the
degrees of the numerator and the denominator, or just essentially number of poles minus
number of zeroes.

For a non-minimum phase system this will not be to be true, that the phase angle as ω
tends to ∞ will be different from this number. However, in either case right, it could be a
minimum phase or a non-minimum phase; the slope would be equal to this guy. So, the slope
at ω=∞ or as ω tends to ∞ would be equal to −20 or some multiple of −20
, sorry would be equal to −20 (q− p) . So, the multiple will come over and if the slope of
the log magnitude curve is equal to −20 decibel per decade, and the phase angle is again
−90° ( q−p) , then the phase system is minimum phase else it is non-minimum phase.

744
So, what uniquely determines is if it is minimum phase or non-minimum phase is this number
or the phase angle −90° (q−p) ok.

(Refer Slide Time: 08:54)

1+ s T 1
So, let’s start with an example right, so I consider a transfer function G1 ( s ) = both
1+ s T 2
of them are greater than 0 (i.e., T 1 , T 2 >0 ), and let’s assume for simplicity that
T 2 >T 1 or in other words this zero lies to the left of the pole, in the left half plane. So, it is
a stable system and also minimum phase as per the definition which we just had. So, well
both poles and the zeroes lie in the left half plane. So, the transfer function is a minimum
phase transfer function.

Now, look at the magnitude and let’s compare with what we had said over here in this 4 bullet
points. So, the magnitude and the phase can be determined by this way. So, |G1 ( jω)|
would be the magnitude of this guy. And the magnitude so, say something like this

√
2
1+(ω T 1 )
|G1 ( jω)|= 2
. Similarly, with the angle right; the angle contribution of this
1+(ω T 2 )
minus the angle contribution of this one. So, as ω tends to ∞ in this case, we see that
the magnitude goes to the 0 dB line or in absolute terms it will just be 1 . And the angle
goes to 0 ° . So, what should be from here?

745
So, the magnitude for a minimum phase system as ω tends to ∞ should be
−20 (q− p) , what is q in this example, what is p ? Both are equal to 1 . So, what
is −20 (q− p) ? That would just be 0 ; q− p=0 therefore, you see as ω goes to
∞ the magnitude goes to 0 . What happens to the angle? Well, as the angle as ω
increases and goes to ∞ , we can see that the angle goes to 0 ° . Now is this consistent
with the definition here? Well, the angle should become −9 0° (q− p) , q− p=0 .
Therefore, the angle contribution should be 0 ° . So, based on these observations here or
this remarks, we can notice that the magnitude and the phase satisfies the properties of a
minimum phase system. Both the magnitude criterion and the phase criterion both are 0
because q and p both are equal to 1 .

(Refer Slide Time: 11:19)

Now, as a contrast let’s take another example, where I have a zero in the right half plane, the

1−s T 1
system is still stable by the way. So, G 2 ( s )= , T1 , T2 again greater than 0
1+ s T 2
(i.e., T 1 , T 2 >0 ), with T 2 >T 1 , the zero lies in the right half plane. So, the transfer
function as per our definition is a non-minimum phase transfer function. So, let’s do all the

comparison of the magnitude and the phase again, so |G2 ( jω )| . So, this magnitude is like

this |G2 ( jω )|=

√ 1+(ω T 1 )2
1+(ω T 2 )2
, and it is exactly the same as in the previous case. So, the

746
magnitude just remains the same and therefore, even as ω goes to ∞ the magnitude
will go to 0 dB nothing will change here right.

So, this does not give me any information if my system is a minimum phase or a non-
minimum phase. Okay let’s go to the angle, well if I do this one and my angle now is

ω( T 2 +T 1)
∠ G2 ( jω )=−tan−1
( 1−ω 2 T 1 T 2 ) . So, a little different right, we have a T 2 +T 1 here

instead of a T 2 −T 1 . So, now, if I compute the angle as ω goes to ∞ , I can easily

check that the angle goes to −180 ° and this is not equal to. So, this angle is not equal to
−9 0° (q− p) . q− p in our case is 0 therefore, this angle if it is a minimum phase
system should also be 0 , and since this does not hold we say that the system is a non-
minimum phase from these two conditions.

So, sorry this condition, this is satisfied trivially, but this condition is not satisfied. So, to
conclude our observations both G1 (s ) and G2 (s ) have the same magnitude criterion,
but they have different phase angle as ω goes to ∞ . G2 has −180 ° , G1 has
0° . G1 satisfies this criterion −9 0° (q− p) whereas, G2 does not and therefore,
G2 is the non-minimum phase and because of this definition of the zero being on the right
hand side, it satisfies these four conditions.

(Refer Slide Time: 13:48)

747
So if I were to just plot this right for some values of T1 and T 2 , you see that this is
G1 and this starts from 0° goes back to 0° , G2 will start here and then go to
−180 ° . Again so, given these two how do I identify which one is minimum phase? Well I
go here because if I say I just do not know what is the number of or what are the number of
poles and zeroes. So, I take the phase plots and I say the range in phase angle of the minimum
phase transfer function is the minimum among all subsystems. So, the range in the phase
angles here it goes from 0° to say −45 ° , and here it keeps on going down. So, this is
the minimum phase system and this is the non-minimum phase system, because of the range
also of the phase angles.

So, I need not always uniquely, if I do not know what is the number of q ’s and p ’ s, I
can make use of the other definition.

(Refer Slide Time: 14:44)

Ok. So, now, this brings us to the question which we started with can we experimentally
determine the transfer function, and by experimentally determining I say well can I subject
my system to different frequencies and look at the magnitude plot ok.

So, as we saw earlier the first step in analysis and design was to derive a mathematical model.
The mathematical model which we dealt or which we used so far very extensively was the
transfer function. Sometimes it may not be possible analytically or even I may not even have
any information of what are the physical components sitting in the system. Therefore, I

748
employ a technique where I can just plot the frequency response of the system, and see if I
could get to the transfer function.

So, I measure the amplitude over a large range of frequencies and this and because of this I
can easily plot the bode. So, bode essentially was that right. I am looking at how my
magnitude changes, which changes in frequency. And the transfer function can be determined
by the asymptotic approximations, and I will show you a couple of examples how we do this
and of course, I can just keep on adjusting the corner frequencies until I get as close to the
real model as possible; and we also know how to uniquely determine what were the errors
associated with the corner frequencies for each of the cases I have a zero, a pole, complex
conjugate poles and zeroes and so on.

(Refer Slide Time: 16:28)

So first we’ll just recollect bode plot. So, this is just recollecting bode plots in a reverse way
right. So, first I will just draw a asymptotes of the of all the bode plots and well all this
asymptotes must be of multiples of ± 20 dB/decade depending on if it’s a pole, if it’s a
zero, if there are multiple poles, multiple zeroes and so on. So, if the slope changes by
−20 dB/decade , then I know that there exists a factor resembling a pole like this. Just
remember what we did; recall what we did in while we were constructing the bode plots, and
I’m just doing the reverse of that.

And similarly if there is a slope of +20 dB/ decade , then I’m actually looking at a zero.
Now, this is again I assume that everything is minimum phase and therefore, I could do this.

749
And similarly, if the magnitude changes by −4 0 dB /decade then there exists a factor
something like this, these are complex conjugate poles and similarly for +4 0 dB/decade it
would be complex conjugate zeroes. And then the un-damped natural frequency here is the
corner frequency and the damping ratio can be found out by measuring the amount of
resonant peak near ω2 . So, we had plotted the bode for different values of � at the corner
frequencies.

So, based on those observations which I had earlier I could get an estimate of what is the
damping ratio.

(Refer Slide Time: 18:01)

Okay, the gain can be determined from the low frequency region right and we also saw how
the low frequency region also helps me, what is low frequency region helps me estimate the
error constants Kp , Kv , Ka and these are directly related to the low frequency gain
right. So, the low frequency behaviour I can find out what is the gain of the system. And well
these are all we know so, we know these from earlier right at low frequencies.

So, these derivations we did when we were looking at computing the errors constants from
the bode plots. So, at low frequencies I am just interested in this kind of things; there could,
where β is like the type of the system. If you just, if you’re just a type−0 system
G ( jω )=K for a very low frequencies or in the log scale, 20 log |G( jω)|=20 log K
right.

750
And then in this case the low frequency’s behaviour is just a horizontal line at
−20 log K dB /decade . So, it’ll just be a straight line of −20 log K and just be a
constant line, across all frequencies if I just keep on drawing this. So, this is a low frequency
behaviour when I have a type−0 system.

(Refer Slide Time: 19:28)

Similarly, for a type−1 system: So, G ( jω ) can be approximated for very low

K
frequencies as and I see that the behaviour is something like this, which indicates
( jω )
that the low frequency has a slope of −20 dB/decade and where it intersects the 0 dB
line this was my constant or this constant was also equal to the velocity error constant.

Similarly, for type−2 systems: I will have a slope of −4 0 dB /decade , and the
intersection with the 0 dB line will tell me the gain right. With this, we have derived these
formulas right when we were looking again at how to determine Kv and Ka directly
from the bode plots.

751
(Refer Slide Time: 20:27)

Okay so, just as an example. So, how do we do this? So, I just do a frequency testing of a
signal I subject frequency testing of a system, I subject my system to several frequencies and
I see a plot like this, ok.

So, and I say now; well if I say that I do not know anything of the system, I just give you this
plot just the magnitude plot, forget even the phase for the moment. If I just have the
magnitude plot, can I find out what is the transfer function?

(Refer Slide Time: 21:00)

752
The way we do this the way we identify, now the transfer function is to first construct an
asymptotic plot of the magnitude plot of the bode. I said earlier, I am not too worried about
what happens in the phase plot, because I’m essentially dealing with minimum phase system
in which case I can purely determine my transfer function based on the asymptotic plot.

Okay so, what happens first in the low frequency region? Well, if I know that if it were just a
straight line then the system would be of type−0 , now there is some initial slope. So, the
system is no longer of type−0 , it could be type−1 , type−2 , type−3 and so on.
So, let’s find out what is the slope of this one. So, I start here at a frequency of 0.1 , I reach
0
a decade here at 10 this is 1. So, let’s measure the magnitude. So, here I am roughly
0
about 38 dB , and I go down and then where is a point, where I am at a 10 . So, I’m
somewhere here; I go here and say I’m roughly about 18 .

So, in a decade from 0.1 to 1 , my slope has decreased by 20 decibels. So, this is
20 decibels per decade right which means I have a system which is now of type−1
okay. Now, the second thing what we need to identify are the corner frequency. So, I start as a
slope at 20 dB/decade here and I see that my slope changes from −20 to 0 dB here
right; so at this corner frequency. So, we will identify what this is, so I start with this is
−20 , this is 0 dB/decade and I keep going and I see that again at this corner
frequency, I am at +2 0 dB/decade and I go here again at somewhere around this
frequency again it becomes 0 dB per decade, then at this frequency I am at −20 and
here I am at −40 ok.

1
So, let’s see. So, here this is a type−1 system. So, I will definitely have a kind of
s
thing or a pole at the origin. So, I also need to identify what is the gain K . So, we will
come back to that. So start from here, so here what is this corner frequency? If I go here this
occurs roughly at a corner frequency of 1.7 okay, now next I go further. So, this is now I
start at −20 , I go, I increase a slope by, I encounter a quantity which increases a slope by
+20 . So, there is a zero here, similarly at this frequency which is about 4 rad /sec , I
can just measure it through these lines here.

4 rad /sec , I encounter another zero, because slope increases from 0 to 20 , at this
frequency which is roughly about 8 , the slope decreases from 20 to 0 . So, I have a
pole, again at this frequency I go which is roughly about 17 , again there is a pole and so

753
on until there is the slope decreases from 0 to −20 , −20 to −4 0 here. So, also
here there is a pole at a frequency of about 42 rad /sec . Now the next thing to find out is
what is the gain K ? So, the way we find out the gain is it is a value at which the initial
slope intersects the 0 dB line. So, here that’s we, we also saw this when we learnt how to
compute the velocity error constant through the bode plot.

So, this is 8 . So, my constant K which sits in the numerator, where the initial low
frequency region is of this form, so, the K is simply found to be 8 by the intersection
of the initial line with the 0 dB line. Now things are straight forward right.

(Refer Slide Time: 25:41)

So, I just can write down my transfer function of the form G ( jω ) is now I have the gain,
this is the low frequency behaviour I am just writing down this sinusoidal transfer function,
then what happens after this. I encounter a zero at 1.7 . So, this would show up in this way

( 1.7jω +1) , after that there is another zero at a frequency of 4 rad/sec.

So, this will be 1.7 sorry, ( jω4 +1) , and then after 4 , I am at a corner frequency of 8

right. So, where the slope now decreases from +20 to 0 therefore, I encounter a pole

8 jω jω
G ( jω )=
(
jω 1.7
+1
4 )(
+1 )
jω jω jω
(
8
+1
17 )(
+1
42
+1 )( )
754
here ( jω8 +1) , similarly I have another pole at 17 , and another pole at ( 42jω +1)
i.e.,

At 42 , where the slope now becomes −40 dB/dec . So, what is the simple process that
we follow here right? So, we start with first by experiments, these experiments lead to a
frequency response which is the bode plot. From this bode plot I get I can construct the
asymptotic plot and finally, from the asymptotic plot I can find out what is the transfer
function of the system.

(Refer Slide Time: 27:34)

Again the key is the assumption of minimum phase. So, let’s do another example where I am
directly given the asymptotic bode plot. So, things are much easier for me here. So, how do I
go about doing this? So, first is well I have this low frequency behavior, which I can say that;
well, my numerator or this is like 2 0 dB/dec and therefore, there is a some like a zero at
the origin and then there is. So, what is this corner frequency? So, at ω=3 , I am at 0
and I pass one decade to be 20 . So, this is what is a decade? You multiple by 10 , this is
30 rad /sec .

Now, similarly now I go here and there is another corner frequency here. So, at this point I
want to find what is this corner frequency, I am at 20 , I go to 0 at 50 0 and this all

755
happens in a decade. And again what is the decade? So, here I divide by 1 0 , this is at
50 rad / sec . So, now, I have something like a zero at the origin, then one pole here and
another pole here. Now something strange is happening here at ω=3 my magnitude is
0 , at ω=1 , where I don’t really know what this is, ok.

So, now these are just straight lines at, which go at 2 0 dB/dec if I look at in in in the log
scale. So, what I would know, if I just called this as a ω2 , ω1 . So, what I know is that
log ω 2 , I will just write it here; now compare with the y−axis right? So, what is on the
y−axis ? So, I have a zero let me call this some number y , this magnitude here 0 .

y2 − y1
So, is the slope right, that’s what we learn in the coordinate geometry. So, here on
x2 −x1
this axis I have log 3−log 1=20 .

Ok. So, now, from this I have y=−20 log 3 , this means y=−9.54 decibels. Now I
should find; what is the value of gain K . So, I compute it in the following way. So, at
ω=1 , I have 20 log K=−9.54 and this gives me a value of K=0.33 okay. So, now,
I just want to write down my transfer function ok.

(Refer Slide Time: 30:50)

So, G ( jω ) , I know well it has a zero starting from the origin, a gain at 0.33 . what
happens after this? So, from here, here till here I have a pole now right.

756
My slope decreases from 20 dB/dec to 0, it’s like a −20 from here right. So, there will
be a pole, what is the corner frequency corresponding to that pole? That will be
30 rad /sec . I will write this down here, sorry this has to be jω ; this will be 0.33 jω

jω jω
, +1 and here this is a corner frequency of 50 . So, this would be +1 .
30 50
0.33 jω
G ( jω )=
jω jω
(
30
+1 )(
50
+1 )
So, this is the overall transfer function of my system and we can re-check this by just plotting
the bode of this. So, we get to this asymptotic plot right. So, so what we have seen right?

So, we have seen two examples where we start with a pole at the origin, we start with the
zero at the origin and so on right. So, this is now a very general procedure you could apply
this to several other systems, it is just a matter of manual exercise picking the asymptotes
carefully and accurately, and you just arrived at the appropriate transfer function ok.

(Refer Slide Time: 32:24)

So, what we have learnt here is well we started with the aim of experimentally determining
the transfer function, but the catch was that we had to distinguish between minimum and non-
minimum phase, and we said that well for the minimum phase there is a unique relation
between the magnitude plot and the transfer function and also the phase right. So, we learnt
about these things, and then we plotted or we given a minimum phase transfer function, we

757
learnt how to get the transfer function from a bode plot. So, I know that the system is
minimum phase.

So, we will next class we will see a little more on this effect of zeroes. Zeroes; the effect of
zeroes to the left half plane, we did quite extensively in module 7, even in module 8 while we
were designing what we called as the lead compensators here. Here we will see some strange
behaviours of zeroes being on the right half plane or a non-minimum phase zeroes. So, that
will be coming up shortly.

Thank you.

758
 Control Engineering
Dr. Ramkrishna Pasumarthy
Department of Electrical Engineering
Indian Institute of Technology, Madras

Module - 10
Lecture - 02
Effect of Zeros on System Response

Hello everybody. In this particular lecture we will consider effect of zeros on the system
response. And I will not really worry now about; well, let me with an abuse of
terminology call it and call it a stable zero, we would be mostly interested in zeros on the
right half plane. What about the zeros on the left half plane? We have done several
analyses on that in module 7 and even through other design process procedure in module
8 and 9. So, we know; what is the effect now of zero on the left half plane how it helps
improving the transient response and so on.

So, so far we have not really discussed about in detail about zeros. We talked about zeros
only when we were facing some problems in the design or in the performance
specifications. All we were interested even while drawing the root locus was how my
poles change as the gain varies from 0 to ∞ . Similarly, even while we were doing
the bode plot, you never really worried much about what is the role of zeros and what are
the implications of having different kinds of zeros.

(Refer Slide Time: 01:36)

759
So, in this I just start with some basic things right. So, what is a mathematical zero well,
so if you back to history a little bit well there was no symbol for 0 , right? If you
remember what we learn in roman numerals in school, there was no concept of zero
there, you had X , V then L and the C , I do not even know, why we are
forced to learn those things even though we had a much better number system.

So, and there were few other things other civilizations where they had some kind of
symbol for 0 and first day of each Mayan month was zero. But significant advances
happened in through Indian mathematics. So, the 0 invented here went through the
Arabic world and then to the west through Europe and so on. So, if you typically ask
European who invented 0, they would actually think it is actually an Arabic number
system because that is where they imported it from. But it is this is the biggest
contribution of Indian mathematics ok.

(Refer Slide Time: 02:43)

So, we will not discuss about the Indian mathematics here, what we will look at is the
role of zeros in our context; in the context of control systems. So, so far in our analysis
even when we started with poles or when we started with the analyzing stability
everything was depending on poles: �, ω and everything and the characteristic
equation everything came in the denominator. And for a system of the form now this is;
these are what as referred to as state space equations.

760
This will come up in detail in module 12, but for the moment assume that we have our
system in a differential equation form of this way, where x́ are my states. A is the
system matrix, B is the input matrix, y is an output vector and C is an is the
output matrix. So, typically what this would be mean is well, I have a x as an n -
T
dimensional state vector, it could be [ x1 … … . xn ] , A is an n x n matrix.
Similarly, B if there is just one input, would just be an n x 1 matrix. C if
there is only one output. So, one input one output means u and y are simply
scalars. And C would then be 1 x n matrix in some cases there could be D
also here and ok.

So, how do I go from here to here? The way I go from here till here is, So I just take the
Laplace of this SX ( s ) =A X ( s )+ B U ( s) . And then Y ( s )=C X (s ) . And then I keep
eliminating things and what I find is I will skip the steps here, but they should be very
straight forward to find out. So, this would be C X ( s) comes from here, this would be
C [ SI −A−1 ] B . So, so you can just eliminate X (s) from here. So, this will be

[ SI − A ]−1 X ( s )=B U ( s) and then you could do the manipulations. So, this is what we
call as the system in the state space form and we had discussed this very briefly in one of
our first lectures of course, we have not analyzed much on what is the nature of A ,
B and C that will come in detail later.

But from here to here is this is the very straight forward procedure, you just take the

Y (s)
Laplace transforms and take the ratio of . And if you have a D here, the
U (s )
D will well, very plainly sit here. Okay? So, you are eliminating X from this
equation and just substituting for here sorry, this is just some error here. So, this would
+1
be, this would not be an inverse this would be plus here [ SI − A ] X ( s )=B U ( s) and
−1
when I do it on the other side X ( s )= [ SI−A ] B U ( s) . And then it is a little straight
forward. So, this is a sorry I just made a mistake here, this should be a plus 1. Okay, now
this is also what is, this is our transfer function. And if you look at it carefully what goes
in the denominator is this [ SI − A ]−1 right.

So, you have this the way you compute the inverse of a matrix is you just take the
determinant and you take the adjoint and so on. So, the determinant or the denominator

761
 −1
here will only depend on what is [ SI − A ] , I am not talking anything about C ,
B or even D at the moment. The poles determine whether a system is stable or
not, in addition to the decay rate and oscillation frequencies of the initial condition
response right.

So, it is it talks about stability, it also talks about the transients, the decay rate, the
frequency of the oscillation in terms of ω d , decay rate is determined by �. And what
is important to note here is while I do this analysis is that the poles do not depend on
B or C . Therefore, instead of this having this if I just say I want to analyze
x́= A x the autonomous system, the characteristic equation will just give me the poles
[ SI − A ] =0 it is same thing right. So, I don’t really worry about what is B , C
and even D for that matter.

(Refer Slide Time: 07:42)

So, the message from this slide is that the poles only depend on the A matrix okay,
just because of this guy. And they have nothing to do with B , C or even D .
Now zeros are determined by the system matrix A and as well as the input B and
the output C . Now what is this, this inputs suggestion? What is output suggestion?
The input matrix or the way the input enters to the system is via placement of a certain
actuators, actuating elements. In similarly y are the measurements, which we will or
the y ’s are the quantities which we will measure physically. And therefore, y

762
comes through some sensors and y ’s will determine or the C will determine
where my sensors are placed.

So, these zeros are determined by the system matrix A as well as the input matrix
B and output matrix C . Practically what do they mean, practically the zeros
depend on the physical placements of the sensors and actuators relative to the underlying
dynamics. Now this is what distinguishes a control system from a standard dynamical
systems emanating from physics. So, the concept of zero is what distinguishes control
theory from dynamical systems theory and of course, throughout this.

So, most of the analysis we will do is for strictly proper transfer functions. Let’s see this

with some very, very simple example. So, I have [

A= 0 1
−1 0 ] ; say, B= 0
1[] ;

C=[ 0 1 ] , and say

−1
D=0 . And, then I do the math C [ SI − A ] B , D is 0.

s
So, what I get as a transfer function G ( s )= 2 . Now I keep all this same, so A
s +1
does not changes, B does not change, C does not change, D does and say, I

s 2+ s+ 1
put D=1 then what I get as a transfer function G ( s )= . You see nothing
s 2+ 1
changes in the denominator the poles are the same.

What changes from here till here is the location of the zeros, there is one more zero

added here and you can similarly check you know may be []
B= 1
0
, C=[ 1 0 ] and

you will get different numerators for each of them. So, the zeros are determined by not
only A , but also B , C and then also in many cases D ok.

763
(Refer Slide Time: 10:20)

So, what do exactly these zeros do, right? So, a zero of a transfer function is the root of
the numerator polynomial of the transfer function and it is, it can be a real or a complex
number or a just be a; the number 0 also.

Then when the transfer function is asymptotically stable; so in this analysis we will only
be interested in transfer functions which are asymptotically stable which means all the
poles are to the left half plane and each zero then has the specific effect on the
asymptotic response of the transfer function for certain inputs. And this is what we will
do throughout this lecture. So, let’s just re-visit the slides which we had last time ok.

764
(Refer Slide Time: 11:14)

So, what we will in this lecture typically be interested is in non-minimum zeros, it is

zeros in the right half plane. So, what we discussed last time about all these things how
do I distinguish between minimum phase and non minimum phase via the bode plot. We
also concluded that if the system is minimum phase, the magnitude response will
uniquely determine the transfer function. And well okay, so there is things like we said,
and we’ll actually detail this a little more that non minimum phase systems are slow in
response, because of the faulty behavior at this at the start of the response.

Common example is transport lag. So, let’s say I have very, very simple example that I
have a plant here G , I have a controller C and well may be some reference signal
R( s) then, as usual some output y (s) here, there is some input to the plant, I say
well there is a little delay between the time the control signal is sent and so, then the
control sends the signal to the actuator and the actuator generates a control signal.

And this is modeled in transfer function as e power minus s tau, assuming tau is the
delay. And this actually comes from the Laplace transform of a delay signal
F ( t−τ )=e−sτ F (s) , this is a straight forward derivation we will not go in to the details
of that. But why do I say that this system has, is; has or it will be faulty behavior at the
start of the response?

−sτ
Now, what is the contribution of this delay here? That in the numerator it adds up e
and then you have a C ; G( s) and C( s) , (i.e., e−sτ G ( s ) C(s) ). And if I take a

765
small approximation of this, this would just be e
−sτ
can be approximated at 1−sτ
and then you have ( 1−sτ ) G ( s ) C (s) . Now what this and this is not a very
elaborate proof? But just a very small evidence of what happens when I when I do this.
So, when I take a when I have a transportation delay or transportation lag. So, this is a
system. So, what is the nature of this. Now, this introduces a zero in the right half plane.
And this will result in a system which is now non minimum phase, ok.

So, we will see what is the role of this zero on the right half plane. How is it related to a
transported lag? Well, it is related via this relation. And what is the faulty behavior here
that we are talk talking of, right. So, we will talk of these two points also right. So, non
minimum phase situation arises when the system included non minimum system has a
non minimum phase element or a minor loop inside the system which is internally
unstable, ok.

(Refer Slide Time: 14:31)

So, let us first talk of internal stability and what does it mean? Mathematically a zero can
cancel a pole when there are two transfer functions in cascade. Let’s say loosely speaking

1
say I have and there is a cascade s +1 . So, mathematically they will cancel
s+1
each other and this is 1 . Now, we’ll not worry about how can I realize this and so on.
That’s not the worry at the moment.

766
Similarly, when I have unstable things this is s−1 , this is s−1 , they will also
cancel each other and I will have 1 right. So, first is, is this a stable realization? So, if
I say I have a unstable plant, I want to track a reference say some R( s) , it’ll be capital
R and the input here. So, this is s−1 and then this is the output. Okay? Then each
time I could say that, well I have a unstable plant here, unstable pole, cancel it with
unstable zero and I will get a output which is exactly the reference, I want to track. This
is straight forward, right. Just looking at it mathematically, but is it practically feasible?
Well, it may not be allowed practically because if this changes by even a very small
number say, by 1 million for example. Then, well there is some instability here, there
is no cancellation. Moreover, what if the initial conditions are not zero? Then you can
never do this because the signal from here till here is unstable, right. And we will talk of
this, when we will talk of unbounded internal signals. So, even if there is no discrepancy
between these numbers there could be an unbounded internal signal ok.

(Refer Slide Time: 16:23)

What does that mean? So, if I take a configuration like this, right? Say I have a standard
reference, I have the output, the error, a unstable pole here and I have a zero on the right,
and I have a s +1 here. So overall, if I just look at the cancellation. Right? So, the
system might look stable, not only that how does the error go? So, what we are worried
about in the system is usually the error; and all the analysis in the transients or the studies
sorry, all the steady state analysis we were doing in terms of this error signal ok.

767
Now, this error signal or the transfer functions from r to e looks something like
this, which looks; which looks nice, right? This was, this is actually a stable, okay?
However, if I look at the transfer function from r to u well, I have something
strange, I have s−1 here. Okay? Now this is unstable and this is not allowed right.
So, there is some amount of hidden instability in the system, so the signal from here to
here. So, the u which goes to this plant if I may say so, is actually unstable right. So,
this u would be unbounded and therefore, I cannot do much about this. This system
is, system which has a instability hidden within the system right

(Refer Slide Time: 17:43)

So, now what do we do, right? In these cases, well therefore to determine stability it is
necessary to determine the stability of all the transfer functions within the system. So,
even when there is a perfect cancellation we should not just blindly, say the unstable pole
is cancelled by a zero on the right half plane so the system is; is stable, No. So, even with
the arguments in the previous slide we just cannot do this. A non-minimum phase
controller zero cannot be used to come to cancel an unstable plant pole. And in the same
way an unstable controller pole cannot be used to cancel a non-minimum phase zero,
either ways right?

768
(Refer Slide Time: 18:36)

So it’s, this is we have to be very careful of this of this cancellations. Okay.

Second thing, what have; what happens with this zeros? Again we are just talking of
zeros in the RH plane or right half plane. Now, say I just look at say a very simple

s−1
transfer function of the form say , and I just want to plot the root locus of this.
s+ 2
So, I have pole here and at −2 and zero here at +1 . So, what happens as the gain
K increases, well I go here and as K goes, gain K goes to ∞ , I am in the
unstable region in the s plane; of the σ and jω axis.

So, these non-minimum phase zeros also limit the closed loop performance, well how do
I do this? Well, my root locus plot tells me that as the loop gain is increased the poles
move towards the zeros, right. And then once I cross this line here I am already in the
unstable region. And therefore, whenever I have a zero in the right half plane my gain
margin would be limited. And if the gain margin is limited it would imply that the system
has lesser robust stability properties.

1
Contrast this with this case, right. Where we had a transfer function of , a pole
s (s +1)
here, a pole here, it was stable for all K ’s no matter what you do. So, this is, this is
the effect also of a right half zero where it in; where it introduces some additional

769
performance limitations. Now is that all, that right half zero does, there are some more
interesting things which right half zero does ok.

(Refer Slide Time: 20:21)

So, while doing the rest of the analysis we will not do proofs, some of the proofs are still
open problems, but we’ll just try to understand the several phenomena associated with it.
And try to relate with what we had already studied in the earlier lectures ok.

So, the first effect is what is called as a blocking effect of a zero. So, let us take a simple
example, here where the number 0 this is the zero of the transfer function which means
G ( 0 )=0 . Then the steady state response of it is also 0 . And which means the DC
gain is also 0 . So, the steady state is a DC gain. So, for example, if I take this plant
here and I say well what is this response? Well, the response initially goes up and it just
goes back to 0 . This is when I have G ( 0 )=0 ; however, what am I tracking? I am
tracking is this is a step input right. So, this is a step response of a system which has a
zero at this number 0 ok.

770
(Refer Slide Time: 21:42)

Next what happens when I have a sinusoidal input? Say input of the form, input in such a
way that the frequencies of the input is sinusoidal with some frequency ω . And then
if the imaginary number jω=0 , of course, − jω will be also a 0 , if these two
ω ’s are the same, so in this right? So, I have roots at ± jω or just ± j 1 . So, if I

1
have an input of the form sin t , which has a Laplace transform of 2 then,
s +1
typically what we would assume or what we had learnt earlier that if the input is a
sinusoidal signal the output for a linear time invariant system will also be a sinusoidal
signal.

What will change? Well, the magnitude will change. Frequency will remain the same and
there might be a bit of phase shift depending on what are the components sitting in my
plant or in my process. But that is not always true, now this some; something strange will
happen here, is if I subject this signal, this plant to a sinusoidal thing which means I am

1
essentially multiplying it with 2 . So, this guys goes away and what remains is just
s +1
the impulse response of this signal which is stable, right. I just recover back or I just start
from 0 reach a peak and then go back to 0 . So, the output is not a sinusoid for this
case at least. And the condition is what? that this ω must be equal to this ω . If this
2
is s +2 and something else might happen, you might actually see a sinusoidal signal
at the output.

771
(Refer Slide Time: 23:47)

So, this is also. So, these two are called blocking effect of a zeros. Now something even
stranger, when we talked about stability, So one of the first ways. So, how was stability
defined? Stability was defined in a way that if my plant is or if my system is subject to an
initial condition, will it come back to its rest state. From physics it will be: will it come
back to its point of minimum potential energy, right? So, that’s how we defined stability.

Another notion of stability was bounded inputs leading to bounded outputs. And this
analysis of bounded inputs leading to bounded outputs helps us to characterize stability
in terms of the location of poles of the system. But we never talked about what happens
when the input is unbounded if it gets me an unbounded output is the system stable?
Well, I do not know the answer, what I know is if the input is bounded and if that results
in an unbounded output the system is unstable. Now what if the input is unbounded,
right? So, something strange happens here. So, I take a system with transfer function of

s−1
this form 2 , right. It has again a right half zero and I subject it to an
s +2 s+1
unbounded signal e t , unbounded input signal ok.

So, what would you; we typically expect? we would expect that the response of the
system would also be you unbounded. Now, the number 1 is the zero of the system
which means G (1 )=0 . Then the response of the system surprisingly is not only
bounded, but it also converges to 0 , right. You can just see with the plot. So, which is

772
again start from 0 , e
−t
is the input and I just come back to 0 So, this is one
strange thing. So, so what we had left out while discussing about stabilities what about
unbounded inputs. Well unbounded inputs could also result in bounded outputs in under
certain conditions here. Where the system is naturally stable because the poles of the
system are always on the left half plane ok

(Refer Slide Time: 26:09)

Now, some more strange phenomena: so after we discussed about what we called as the
blocking effects of zeros, that we see something strange happening to what we had learnt
in the earlier 9 modules, right? If a input is a sinusoid, output is also a sinusoid. Well, no
no no there is something strange happening here. Input is unbounded, output might also
be unbounded, no, something strange is happening here that output actually is bounded,
right. Ok.

Now, there is something called initial undershoot right. So, we’ll analyze these, some of
these examples here. So, let me take a step response to a signal which looks like this and
again it has an unstable zero, if I may call it so or a zero in the right half plane. So, what
you see here is starting from 0 , if I were to track a step I am actually moving in the
correct direction. I move in the correct direction, but here something happens and I just
go back to 0 . So, after sometime passes the system reverses and moves in the wrong
direction right. So, if this was a plan say, I may be I hire a new CEO in my company,
which was not doing very good until yesterday. There is a big positive sentiment in the

773
market the stock prices go up after few months, I realize that this was not a good
decision to have this new CEO, nothing changes in my profits or even my turnover and I
see well that the response actually goes down right.

So, this could be encountered in several situations and inappropriate plan, right. Even
sometimes you elect a new government, right. Which promises a lot during the election
campaign, the market sentiment is high everybody is very happy, stock prices go up, but
after few months it is all the same story again and again it goes back to 0 ok.

(Refer Slide Time: 28:12)

Now some other things could also happen. Here what is happening I take a decision and
the initial trends look very nice, but then after a while I have made I realized that I made
a mistake by hiring this new CEO or you know even hiring or even voting for new
government because of the trend that follows later on. Now, here something else happens
right. So, if I say, again I take the same transfer function with a negative sign here. And
what I observe here is, okay if I just start here and I just wait for a while and see what is
happening here, I am actually tracking this step properly okay, there is no error nothing.
But what is happening initially is I am actually going down. I go down and then I start
rising after a sometime. For example, if you look at the demonetization, the initial trends
suggested that the productivity in the market was going down, some analysis; analysts
suggest that, maybe the government would say it oh it actually it went from 0 to 1
instantly.

774
But we are not we are not looking at what the politicians say, but just the initial trends
would suggest that because of demonetization there was a cash crunch and we are
somewhere here now, right. At this, at this stage as we are recording this lecture, but
what we hope or the government hopes is that we will go somewhere here eventually we
will start rising, possibly rise to higher levels, but; however, the opposition things that
we’ll just go down and down until the next election, when the opposition gets selected
and then the upward trend will start moving. So, we’ll not do the analysis of what the
politics or the economics say, but decision sometimes leads to the movement in the
wrong direction and eventually it might just pick up.

So, a response of this departs in the non asymptotic direction, what is the asymptotic
direction? that I want to track this number 1 . And this results in a initial error growth.
So, the error, well if I just started t=0 plus, it will just be about zero, but the error is
actually growing, but after while you see that its decreasing. So, the step responses of an
asymptotically stable strictly transfer functions. Strictly proper at this is a little
important, for proper transfer functions some something else might happen and we will
not worry about that at the moment. So, the observation is that the step response of an
asymptotically stable strictly proper transfer function exhibits and initial undershoot, if
and only if it has a odd number of positive zeros.

Again we will not really do the proof of this one, but if there are odd numbers of positive
zeros we expect a trend like this.

775
(Refer Slide Time: 31:00)

Okay now, what happens when I have an even number of zeros? So, I have s−1 and
s−1 . The response initially moves into the correct direction. And so say, after a while
and then I actually goes down, and again I am actually picking up, right. So; however,
the system reverses here it goes negative and again reverses here. So, this the step
response of a system with multiple positive zeros can exhibit multiple reverses in the
direction.

Again I will not do the proof of this, right. It just; we will just observe few things what
are happening and these are things which would not be done in standard control syllabus
nor in the gate or even under undergrad texts, but these are actually little bit of fun things
to know. What are the things nice things or even the bad things that could happen with
non minimum phase zeros ok.

776
(Refer Slide Time: 31:59)

So, what we see in these plots is that well there is a 0 crossing, ideally what you
would expect? Well, if it is a say under damped or critically damped system the response
goes like this, if it is a under damped sys, So if it is over damped or critically damped
this is this is the case, if it is under damped then I expect some oscillatory behavior,
overshoot and so on. Here, I experience something else right, I start here and I go here, I
again cross the 0 once, I come here, I again cross the 0 once more. Here, I am
crossing the 0 once I go down, I cross this 0 here ok.

So, when do these 0 crossings happen, and how many of them happen? And is why
what is the evidence that is actually happens? So, if an asymptotically stable transfer
function possesses at least one positive zero, then the step response undergoes at least
one 0 crossing. So, if there is at least one zero sitting on the right half plane there will
be at least one time where the plot will cross 0 or when the response will cross 0 .
So, how do we verify this?

Let’s say the number z denotes a positive zero of the asymptotically stable transfer
function. This is important, right unstable systems will be; there is nothing we could
analyze. The Laplace transforms of the output for and if I am just looking at a step is

777
 G(s )
Y ( s )= . And what happens at a s=z ? Which is the positive zero, well I have
s

G( z)
Y ( s )= , G ( z ) =0 it follows that Y ( z ) =0 , and what is Y ( z ) ?
z

Now, this Y (s ) is the Laplace transform coming from this formula. This is simple

formula ∫ e−st f ( t ) dt , right. I am just substituting z for s , and what I see here
0

is this ∫ y ( t ) e−z t dt , the value of the exponent is positive all the time, right. This guy
0

−z t
e and what could only go to 0 is y (t) and therefore, we say that there is
actually a 0 crossing, when this happens that Y (s ) or G(z) going to 0 and
actually Y ( z ) going to 0 actually means something like this.

(Refer Slide Time: 34:28)

−z t
And because e is always positive, Y (z ) should go to 0 sometime. That’s
what this argument says, since e−z t is positive for all times it follows that y ( t ) must
cross 0 somewhere in this interval, right. Now this is important here, right. Note that,
this only depends on the positive number z , it does not depend on either the poles or
the remaining of zeros, just this guy is what is causing it to cross 0 .

778
So, at least one 0 crossing occurs if the transfer function is proper and at least and has
at least one zero on the right half plane. Now, if there are more then you will have
multiple reversal of directions in such a way that the step response must possess at least
two direction reversals if it has a non-zero even number of positive zeros. So, I am just
writing down this statement which I observed here. I have two 0 , two direction
reversals and two times I cross 0 , one is here and one is here. And the number of
0 crossings is equal to the number of positive zeros. This is only for a strictly proper
transfer functions. And these are also most, many of these are just by observations and I
think this statement is still not been proved mathematically ok.

(Refer Slide Time: 35:49)

Now, 0 crossings are they only with positive zeros, one may ask what happens when
my roots are imaginary or the zeros on the right half plane are in or complex conjugate.
Well, it still has two 0 crossings right. So, things also extend to the case when we
have poles which are not only positive, but all sorry, right half zeros which are not only
positive, but also complex conjugate again these are just a little examples, ok.

779
(Refer Slide Time: 36:23)

So what we, what we have done. So, far is in this module. So, which is possibly shortest
of all modules is to analyze the role of zeros from a different perspective. Not just
looking at a lead compensator or a derivative control action. We are seeing what happens
when the zeros are on the, right. There is there are several other facets of this particular
topic.

But we just restrict to just what we need just the basic information of what we need and
why were we interested in this? Because we were interested in doing the experimental
determination of transfer functions via the bode plots via the magnitude of the bode plot,
magnitude plot of the bode thing. So, in that case there was a unique correspondence
between the transfer function and the bode plot if and only if the system was minimum
phase.

So, what we saw here was a minimum phase systems and then the definition of non
minimum phase and how the concept of minimum phase systems was crucial in the
experimental determination of transfer functions. And we saw the effect of right half
zeros on the system response things like the unstable pole zero cancellations, blocking
effects and so on. So, all this material from this, from I just referred directly to this paper.
It is a very beautiful paper to read if you if you have time we can there are several other
examples of the inverted pendulum some examples related to the to even server systems

780
and so on. And also good examples of you know of day to day life examples where when
you say I am riding a bicycle and I want to turn to the left.

A typical answer if you ask anybody how would you turns your bicycle to the left he
would think you know it is it is stupid question to ask. That you just going, going and
then you just do turn left all of a sudden, but that’s not true right. So, what you do is you
first turn a bit initially to the right create an angle So that then you turn to the left. Just
observed this when you are next time on a bicycle right. So, these are several phenomena
which are not very obvious to us, but there if you go a little bit detailed into detailed
observation you observe the effects of what we actually saw in this lecture in terms of
plots ok.

So, in the next module what we would do is I mean most people are interested in
learning some practical applications or applications could be several, right. From process
industry a very popular application is a robotics, and people say where all robotics is just
buy a robot you know that a microcontroller program it and do. So, we will spend an
entire module on some, some applications. So, a colleague of mine Dr. Vishwanath will
be handling that module. He is very experienced in; he has a rich experience in industry
he was experienced he was involved in the development of the LCA which is now called
the tejas. He worked with honey well for a very long time. So, he has good experience in
a in practical aspects of control and he will teach you few things about inertial
measurements and sensors and so on.

So, I hope that will be a little more interesting again, this will not be a part of the
standard control curriculum for either gate or even while crediting a normal course in the
university. But never the less we thought it will be little more interesting to add a bit of
practical aspects. We cannot add all the possible application but at least one application
where you can directly see what you are learning and how to apply that in real life.

And I’ll see you back again during the last lecture of in the module 12 on state space
analysis.

Thank you.

781
 Control Engineering
Dr. Viswanath Talasila
Department of Telecommunication Engineering
Ramaiah Institute of Technology, Bengaluru

Module – 11
Lecture – 01
Navigation – stories and some basics

Hi my name is Viswanath Talasila, I’m an associate professor at Ramaiah Institute of

Technology. And I am offering this course together with Professor Ramkrishna
Pasumarthy of IIT Madras. We are going to now move to module to 11 the first lecture of
this. We’re going to be looking at some applications. Applications of some of the theory
which you have learnt in the previous lectures and the specific application which we are
going to focus on is a navigation.

To be more specific we are going to be looking at the notions of how transfer functions,
filter design, you know so what you have done in bode plots and those kind of things,
low pass, and the high pass filter. So, we are going to be looking at how they are relevant
in the context of navigation. We could take a lot of different applications, there is a
specific reason why I have taken navigation, and it is as follows. So, in most control
courses we see that the traditional application areas are for electric circuits like power
systems or mechanical systems which are robotics and so on. But navigation is another
very big application area, and there is quite a lot of interest in control systems also
modeling which is very relevant and required in navigation. It is important for all
students to remember that when you talk of control systems, controls does not exist
independent of a model right. So, which is why any controls course the first one or two
units will be predominantly about the model of the system, only then we go for analyzing
say how to design controllers and so on.

In fact, the first 3 units will always be about modeling and analysis of systems. So, we
will look today at how we can model the problem of navigation. And we will see how to
bring in two specific concepts of control system which is the filter design and of course,
modeling. The another reason why navigation is very important is many young engineers
when they go to industry or higher research. So, when your looking at autonomous
vehicles. You are looking at say the flight control of an airplane or a missile. The story of
navigation becomes very important. And some of you would have already heard of these

782
filters called the Kalman filters, maybe if you are lucky you all have even heard of
particle filters and so on. So, these filters are a very standard and they’re very important
in the field of navigation, specifically, the Kalman filter is a classical feedback control
design written in a slightly probabilistic or stochastic manner. So, while Kalman filter is
one of the standard techniques we use for navigation, what I will be presenting in the
series of lectures along with some experimental demonstrations is a complementary
filter. And a complementary filter is a very simple filter it is a combination of low pass
and a and a high pass filter. And we take the signals coming out of these 2 filters and we
fuse them in order to get more meaningful answer to our navigation problem right.

But the Kalman filter does is something far more complex. And it turns out for lot of
applications we do not really need the Kalman filter which is a feedback control design.
It is enough to do just you can think of this as pre-filtering we can just filter our signals
and still get very meaningful answers to our navigation problem. So, this is a
combination of modeling and control. So, let’s go ahead and if you look at the slide over
here you will see that I have used these words stories. And the reason for this is ash it is
always nice to know how certain fields have evolved right. So, be it control system be
signal processing or whatever. And navigation is actually very fascinating topic. So, the
story of navigation is as old as a story of mankind or even animals and all species. And
we will see with a few stories how navigation has been accomplished over the past say
thousands of years. And we will see that the basic concepts of navigation have remained
very similar over the last 5000 years. But change, of course, is a technology for example;
GPS was not there more than few decades back right.

So, technologies have changed, but we will see that the story of navigation the basic
concepts have remained very similar ok.

783
(Refer Slide Time: 05:23)

So, first, let’s start with some stories and the content is so, first, we will see what is
navigation, we will see various types of navigation. So, we look at some things some
nice stories from history which includes some sad stories as well. How animals navigate,
then we will look at one specific aspect of navigation in a little bit of detail called dead
reckoning. We all have actually studied dead reckoning slightly before in the concept of
trigonometry, we will see that we will conclude this lecture with studying reference
frames. So, the basic idea of a reference frame is when I describe the motion of an object
it is with respect to my own reference frame.

So, if I see, for example, this duster moving from this point to this point, this motion is
with respect to my reference frame. So, the really bad way of explaining this is for me
this is a motion from right, to left whereas, for you, this would be a motion from left to
right. That is a very simplistic way of explaining, but if you remember that we all live in
a 3D world. So, we describe every point in space with a XYZ say coordinate system, and
my coordinate system will not be the same as your coordinate system right. And it is very
important when we do navigation to be very clear about which reference frames we are
going to use.

784
(Refer Slide Time: 06:52)

So, we will conclude with the notion of reference frames. Okay so, what is navigation?
Well, there is many definitions and all definitions are equally true, the basic idea is that it
is a technique or the art of moving from one geographical area to another right. So, you
can move from Chennai to Bangalore you could move from say place x to place y and so
on. Now when you want to navigate or move from place x to place y, the first and most
important notion in navigation is to know one’s own position right. Say you want to
move from this point to this point right. Through whichever route you want to take
whichever trajectory you want to take, you have to know where you are otherwise it is
literally impossible to know where you are actually going to land upright. So, we will see
some examples of calculating our own position, and using our known position we will
see how we calculate the path to a certain object. Then we will look at very briefly
maybe in just one slide or So, other navigation sensors very advance navigations sensors
some of which are based on GPS, some of which are based on Wi-Fi localization and
camera-based localization. So, localization basically means knowing where I am locating
my position.

So, we will do this very briefly just to give you an indication of how advanced field of
navigation has become. Our primary focus will be on inertial sensors; we will see that a
little bit later. So, point number 5 is really crucial and it basically says while the nature of
sensing GPS sensors cameras Wi-Fi’s and so on while they have changed even
computing has changed over thousands of years, 3 basic concept of navigation have

785
always remained and still being used in exactly the same way. The first is know where
your current position is, we already talked about that. The second is using information
about your current position and velocity we will see that velocity is not the only
requirement, but for now, let’s assume velocity.

So, using the information about a current position and velocity you can actually compute
what will be your future position. And finally, you always will make use of external
landmarks to make sure that you actually don’t get into I mean what I call hopelessly lost
this is actually technical problem called drift. Before we go ahead let me give a small
indication of how we normally navigate if I need to go if I need to move from this
position where I am sitting right now to if I need to leave this particular room I have a
fair estimate of the map of this area right. So, I know where the walls are I know where
the chairs are I have a rough estimate of the distance which it takes for me to go from
this point to exiting the door.

If I close my eyes and I start walking, and all of you should really try this experiment at
home, if I close my eyes and start walking we have a reasonable idea as long as you do
not fall down we have a reasonable idea of how much distance we are covering right. For
example, if I close my eyes and walk from this point to the end of the room I won’t say
that I walk one kilometre, I won’t say I walk one meter. I will be in the range of maybe 5
meters 10 meters or something. So, I have a fair idea of how much distance I have
covered in which direction I have gone alright. So, that is a little bit the problem of
navigation. If I open my eyes now I can see external landmarks, I can see there are chairs
over here I can see these gaps in the corridor all these additional. So, called landmarks
they help me to navigate far better.

So, in navigation we always at least the advance navigation we always use 2 kinds of
techniques. So, the first one would be called it is actually called dead reckoning, and it
basically is a means to estimate my future position based on my current information. And
the second thing is the use of external landmarks. And for those of you who read this
subject in a little bit more advanced way, you will see that GPS is actually one of these
beautiful concepts where we use external landmarks which are satellites and space to
actually help us to localize where we are, and these inertial sensors are the ones which
are going to be used to estimate my future positions and so on ok.

786
So, the concluding point of this slide is the last line in read that the technology for
navigation has changed dramatically, but the overall concepts are still very similar alright

(Refer Slide Time: 12:07)

So, what are the various types of navigation? Obviously, we had say coastal navigation
you could also include land base navigation right. In the ancient days so, people would
move from say from Calicut all the way to Indonesia. So, that is entirely the coastal
navigation the sea routes. Then people may have wanted to move from, let’s say 2
different kingdoms in ancient India and there will be the land base navigation. We will be
focusing more on a little bit on coastal navigation in our lecture. So, coastal navigation
basically is navigating alongside a coast.

Alongside a coast means you usually are able to visually observe or see the coast, and we
will see why that’s very important. And the second type of navigation is dead reckoning;
we will go into lot of detail about that. The third type of navigation is celestial navigation
where we use stars, planets, moon so on, to actually help us to navigate and we will
again go into a little bit of example to see this. We will look at the concept of dead
reckoning with electronic navigation in a lot of detail, by specifically using what are
called inertial sensors. And inertial sensors many of you are aware of at least one of
them, the famous accelerometer which you get in your phone when you rotate the phone
from portrait mode to landscape mode right, the screen also rotates.

787
And we are also going to look at another sensor called the gyroscope. So, first, we look
at coastal navigation we look a little bit at dead reckoning, how humans do it, how
animals do it, and then we look at celestial navigation. And finally, the rest of the whole
lecture series will be on electronic navigation for dead reckoning.

(Refer Slide Time: 14:09)

So, what is coastal navigation? It basically loosely speaking if you talk of ancient times,
coastal navigation was when you need to move from one point on the coast to another
point may be hundreds of kilometres away you would use specific natural features. So,
for example, you would use maybe some mountains which are along the coast. You
would use certain species of animal fishes and other kind of marine species which you
see along the coast, you would actually use tides and currents. So, along the coats, there
would be certain regions which are very strong currents and other regions which have
really very mild currents right. So, we you would use all this information to be able to
say that I am at this particular position. At a different position my tides would be
different, the species of animals would be different, I may not have a mountain at all and
so on. Or you could use lighthouses light vessels and all these other features.

So, if you look at this particular map over here it is called a nautical fishing chart and all
the, all the pink shaded areas maybe it looks like magenta all the pink shaded areas are
where fishes are available in plenty. So, a person who is actually navigating along this
coast say in this particular route. If the person actually is able to see that there are lots of

788
specific types of fish in this particular area, and they will actually have a fair
approximation of the location. They will know that they are somewhere within this
vicinity whereas, if they come in this region and they see a different species of fish over
here they will know that there could be somewhere in this viscidity. They will certainly
not be here right because those species of fish don’t exist over here. So, these are the
nautical fishing charts. Like that you, of course, have the nautical charts for the
coastlines which depict the mountains so on and so forth. So, that is coastal navigation
and this is one of the predominant ways that ancients would actually navigate along the
Indian coast or other countries as well.

(Refer Slide Time: 16:46)

Let us see one very interesting story of coastal navigation from history. And this focuses
on the Emperor Alexander’s great march to India, where he fought some battles, and then
eventually he had to go back. And we will see what actually happened. So, there is a
place called Taxila over here, and Alexander after he finished his battles he wanted to
march say further into India, but his generals they were actually very tired there was
almost going to be a rebellion amongst the soldiers. So, they convinced Alexander that
let’s go home enough war is done. So, a decision was taken that from Taxila they would
actually have to go back to Babylon where the empire was. So, let’s say Taxila was
somewhere over here, and Babylon which is modern-day Baghdad in Iraq was
somewhere over here.

789
And the idea was that Alexander and his army would actually navigate from Taxila to
Babylon or modern-day Baghdad. And based on the surveys which his people had done
they believed. So, if you look at this yellow line over here this is the it is the Indus river.
So, it is much more clearly visible as these blue lines in this particular chart right. So,
this is the Indus River and what they believed is if they sailed down the Indus River and
they reached this coast. So, this particular coast they wrongly believed that this coast was
the same as the upper reaches of the Nile, which is I would have said yes which is over
here. And it is must closer to Baghdad and is much easier to get from Taxila.

And the reason why they believed that this area over here was basically the upper
reaches of the Nile River near Egypt was because the local flora and fauna the species of
animals and other things fishes actually matched that of these two places. So, the kind of
fishes you would get here were very similar to the fishes you would get there, the kind of
trees which would grow were very similar in both cases. So, what Alexander did then
they marched all the way down to this coast and what do you get you do not get the
Mediterranean Sea over here right. You get the Arabian Sea when you come over here,
having discovered the mistake now because this was a completely wrong coastal
navigation, they now had to march all the way across the desert where thousands and
lakhs may be lakhs of soldiers died before they reach Baghdad.

So, this was a great story in coastal navigation which almost ensured the Alexander’s
army was wiped out, and actually, he never recovered from that and he eventually died.
And the So, the story here is that coastal navigation which is built on using certain
features like say flora the mountains along the coast and so on and so forth, can be highly
erroneous. And we will see that this is a typical feature of most navigation techniques
right. So, that is about coastal navigation.

790
(Refer Slide Time: 20:27)

Now let’s move to dead reckoning dead. Reckoning is a very simple and very easy to
understand. So, of course, coastal navigation is not always you may not be at the coast or
you may not have interesting landmarks to observe, even on land and then navigation
becomes a little bit tricky. That is why we use the concept of dead reckoning, which is
the technique of computing your current position based on your past known position and
the speed and direction in which you have gone. And the problem in 2D, let us look at
the problem in 2D it is scales to 3D in a very similar way.

The problem in 2D basically says the following, if you start at this particular location and
you moved 45 ° , 45 ° with respect to something, but let us say in this case it is
respect to the x-axis. If you start at this position and move 45 ° with respect to the x-
axis, and you moved at a speed of say one meter per second, alright in this direction.
Where would you be after 10 seconds ? Where would you be after 10 seconds ?
Now, this is very easy to calculate from trigonometry right. So, if you call this initial
position as x old and y old we know from trigonometry what these two coordinates will
be right. I encourage you to try this before you move on to the next slide.

791
(Refer Slide Time: 21:59)

And it basically follows from the basic trigonometric relations of this one.

Where d is basically the distance which is traveled. So, this is basically dead reckoning
in 2D. Note that we are not talking about accelerations velocities and other kind of thing
measured by sensors, we actually know at what speed we are traveling at like how the
ancients would do, have an estimate of the velocity you are traveling at and based on that
you actually calculate your current position. Although this is done in 2D exactly the same
concept extends to 3D as well ok.

(Refer Slide Time: 22:36)

792
So, that’s dead reckoning. Now just a small example of how we would do dead
reckoning in the ocean. So, this is in marine navigation, these charts which is shown over
here is basically one way of computing or plotting. So, I would actually say plotting your
course or plotting your path right. So, they normally use a word course in navigation.
And the basic story here it goes as follows. Your ship has started at 9 am in the morning,
and it is moving at a speed of 10 knots, knots is a marine way of computing speed. So, it
has a certain relation to meters per second or kilometres per hour.

So, your ship is moving at 10 knots at an angle of 90 ° , at a heading of 90 ° and

this is a magnetic heading. So, with respect to your compass which looks at magnetic
north, you are moving at an angle of 90 ° with respect to magnetic north. Then what
the navigator does is to plot? What would be the expected position of the ship alright at
10 am? And in this particular example what has happened is at 10 am the ship captain
decided to maintain the same speed and to move at the same what course of 90 ° with
respect to navigation. And then the navigator waited till 10:30 when the captain decided
to make a change in the path or the course of the ship to 60° moving in the same
speed. Then what the navigator does is to plot what is the expected position at 11
o’clock. So, apart from the first one, all these are expected positions, where you expect
your ship to be. In the same way, they would plot all the positions.

Now we will see in the next lecture that if this is the computed position based entirely on
dead reckoning the actual position, now I am drawing this roughly this would not be the
real case it may be far worse than what I am drawing. The actual position may actually
be something like this. And this would be the actual position where the ship actually
lands up. And the reason for this massive difference in the estimated position versus the
actual position is because of the condition called drift, which we will go through in a
while ok.

So, that is how you do marine navigation dead reckoning, you are not using any coastal
landmarks or any such thing just only dead reckoning. How do animals do it or even
humans? So, if you look at this particular desert ant it is Saharan desert ant and it uses a
technique called proprioception it is a biological term which basically means that I have
an idea of how much I am actually moving. It is a technique of proprioception and it is
able to count it is leg movements to estimate the distance traveled. So, it is able to count
the leg movements right. To know how much distance, it has traveled, notice that we still

793
are not talking about how this guy gets the course or the direction of movement, I will
come to that in the next slide. So, animals, humans, all species they actually have a
means of estimating how much they have actually traveled, that is a basic notion here
and it is based on dead reckoning.

(Refer Slide Time: 26:27)

So, let’s look at celestial navigation and we will see how these ants are actually able to
navigate accurately with celestial navigation. So, celestial navigation is the art of
navigation by using celestial bodies, stars, moon, and is really how it was done in the
ancient days. The cool thing about celestial navigation is that while it is not very
accurate. You can always be sure reasonably be sure about your position. And we will
see why that is so. So, you will never be completely off. So, you will never have errors
the way that you had in your dead reckoning, which leads you to really I mean
completely erroneous position then where you really are.

794
(Refer Slide Time: 27:17)

Celestial navigation avoids that problem. In fact, we will see one example of this shortly.
And the example is something like this. So, let’s say you are in a ship in the middle of a
ocean. And this is your current position. You do not what a current position is by the way.
So, I am telling you are over here, but you are the ship captain and you don’t know
where you are. All you can see assume it is night time; all you can see are the stars in the
sky. So, you are somewhere in the middle of an ocean all you can see are the stars in the
sky the question is where exactly on earth are you. So, how you, how would you solve
this problem? So, the way it was done was actually this. So, let us say you are at this
unknown position in the middle of ocean or somewhere.

795
(Refer Slide Time: 28:03)

And you see a star over here. I do not know, how to draw a star I am imaging it is
something like this. So, you see a star over there right. So, this is all your beautiful ocean
over here it is a moonlit night maybe you are listening to some nice Lata Mangeshkar
songs, but you have absolutely no idea where you are.

Now, how would you actually compute your position? So, to be more precise, let’s say
you have seen this star at 10 pm at night. Now you have a book I believe it is called a
star map and the location on earth based on where the stars are. It basically says if you
are able to see this particular star at 10 o’clock, then if the star is perfectly above your
head. So, if I draw perpendicular line from the star and it falls on earth over here. If the
star is perfectly above your head at 10 pm you know this location more specifically you
know the coordinates of the location. It is usually given in latitude, longitude, altitude,
but we will skip that for now. So, you know the actual location where you are if the star
were perfectly above your head. The star is not above your head, you are somewhere
over here right. And you are seeing the star from this position. Now let us say that you
have a means of measuring the angle at which you are able to observe the star from here
let us say it is some 55 ° . I am standing here I am seeing the star is 55 ° with
respect to me. I have this fancy book which says that at 10 o’clock if I were here I would
see the star exactly above my head.

796
Now it turns out with very simple trigonometry you can actually compute what is the
distance from your unknown position to the actual position if you if the star were exactly
above your head. This is a very simple trigonometry, and it basically assumes that you
know the distance from your position to the star. And the ancients had really interesting
methods of calculating this star again based on standard trigonometry.

So, if you know this distance alright, and you can calculate this you can assume that you
are somewhere over here; however, life is not that simple because what would actually
happen is all it is saying is that with respect to XYZ with I have drawn over here, I can
be anywhere at a particular angle of 55 ° right. It can be here it can also be here
55 ° and so on. I can be anywhere on a circle right. And you do not know exactly
where you are. So, actually, it is a complete circle I cannot draw it here because it looks
weird. But otherwise, it is a complete circle and anywhere on this circle, you will see this
particular star at 55 degrees. So, you know approximately where you are on a circle, but
you still need to know where exactly you are. And the way they would do that is to use
another star. So, you would see another star and see I can now draw stars properly, you
would use another star and you would again do exactly the same. So, you are in this
unknown position I have a certain angle reference to that position and it says that maybe
I am at some angle at one 110 ° or something like that right. I really do not know
exactly, but say someone. And let’s say you have seen this other star again at 10 pm at
the same time as you have seen this star.

We again know from this from the star map chart of us, that if this star were visible at 10
pm you would have to be somewhere over here. At this new x’ , y’ , z ’ . So, I
now know two specific positions if the star were above my head XYZ other is x’ ,
y’ , z ’ . You saw with XYZ I am able to calculate this distance, and I know that I
am somewhere on this great arc on the great circle. With this new position I can again do
the same, and what I would get is a different arc. So, maybe it is something like this,
which means that anywhere on this circle I will see this star at that one 110 ° whatever
you have mentioned right. Anywhere on this circle. Now we see that these two circles are
intersecting at this position right. And that is your actual position. For those of you who
are little bit more curious, you may have either guessed by now or you may want to
know that you still cannot get an exact position fix by doing this because this area over
here is really large with twos with two circles you take the intersection of these two

797
circles you will actually get a very large area. And you could be anywhere in this area
and get exactly the same information.

So, the even better way to do it is now look at more stars in the sky. 3 stars, 4 stars, 5
stars then I get all these circles and the intersection of these circles will become a much
smaller region right. And that much smaller region will actually be where you really are.
So, this is how you would do celestial navigation. It is basically the problem of
triangulation, nothing more than that. And triangulation we see is a technique that is
being I mean followed even in these days. So, right so that is we have done dead
reckoning we have just done celestial navigation.

(Refer Slide Time: 34:17)

Now, let’s see what happens when we combine dead reckoning with celestial navigation.
Know that celestial navigation it is not always easy to get an exact say distance
measurement. Because as we have seen in the previous slide we get a fairly reasonably
accurate position of where we are, but the position is not accurate to within few meters
something like that like. You would say that if I am able to see the star from this position
and after doing all these calculations which we have done over here, you would say I am
approximately in this area. And that area could be as large as a few meters. And why that
is not good? It is not good if you are an ant or an animal which is searching for food
right. And then it needs to come back to it is house.

798
So, let’s say there is this ant over here, and by the way, this is also true of autonomous
vehicle navigation. So, you are over here, you do all this hungama, you are looking all
over the place and you find some food over here. Now when you come back you need to
know exactly where you need to come back. If your position calculation is erroneous
even by a few meters, the ant instead of going to this nest may land up over here, where I
do not know there may be some predator which eats ants. I am not sure. So, it is not
really a very advisable thing to use only celestial navigation. We know that dead
reckoning is very inaccurate. When you combine dead reckoning with celestial
navigation we actually get a far more accurate picture. And the idea is very simple. Dead
reckoning is accurate for small time scales, specifically, this means if I am navigating for
a very short amount of time like say a minute or two dead reckoning is very accurate.
Celestial navigation is celestial navigation has bounded errors for any time scale. So,
even if I navigate for 1 hour, 10 hours, 100 hours, 200 hours, one year, I know that my
error will always be bounded. We have seen that in the previous slide if I am able to see
the correct stars or any such thing I will always know reasonably well where I am. It will
not be perfect, but it will be reasonably well.

So, when you combine dead reckoning with celestial navigation, you can get the
accuracy at the small time scales, and you will always make sure that you don’t have
unbounded error which is a common feature of dead reckoning through drift for any time
scale. And in fact, the desert ants actually use these kind of things. So, what this guy does
over here it uses these proprioceptors which counts the leg movements, it knows how
much distance it has traveled, it uses the concept of optic flow to get the feel of a
velocity. So, you take this object again. So, let us say that I move this object like this
right. I have an approximate feel of the velocity of this object right. I know that it moves
from here to here in about a second. The same thing if the object is static and I am
moving across like I am an ant which is actually moving across this or I am robotic
autonomous vehicle moving across this, I have exactly the same perception of
movement. And I know how fast I am moving with respect to an object. So, that is a
concept of optic flow. So, it gives velocity whereas, proprioceptor gives distance. So,
now you have a lot of information with you. And then this desert ant it actually uses the
position of the sun for the relative angular information right.

799
So, this is the celestial navigation which is done. So, it combines. So, that it always
knows approximately where it is it never gets really badly lost. And the dead reckoning
ensures it is always accurate in the small time scales. So, the fusion of these two will
make sure that the ant always lands up at home alright. So, that is a combination of dead
reckoning and celestial navigation.

(Refer Slide Time: 38:41)

I couldn’t resist putting this slide, because I think it is really fascinating and it shows
how we ancient Indians actually did some amazing navigation. So, Indian navigation is
around 3000 BC old or even older. They had lot of trade links with Mesopotamia which
is current Iraq, and also Kuwait, that’s around 2900 BC really 5000 years ago. And the
world’s first dock for ships Lothal was built around 2400 BC. And one of the earliest
navies which India had was during the Maurya time, around 300 BC the Chola empire
lots of us would have heard, had it is influence extending all they across these regions
and of course, you could not have done that without extraordinarily good marine time
navigation.

In fact, if you look at this picture over here, this is in a museum somewhere in Calicut or
somewhere I am really sorry, I am not able to recall that. They actually got the wreckage
of a Chola era boat, and they reconstructed based on the drawings that are available from
the Chola empire. How the boats would have looked back in those days and by those
days I mean, around say around 1000 AD. So, this is actually a boat which was from that

800
era is so beautiful. And of course, fortunately, or unfortunately as history has taught we
had the visit from Vasco Da Gama 1498, and things changed after that. And certainly, we
have not done any exciting navigation at least marine navigation beyond that. Until the
last few years, when the Indian navy now has started to go, to Antarctica and other really
neat places. So, it is a brief history of the great adventures of Indian navigation, there are
some externally good books on this it is fascinating and beautiful and you should
actually read this to get a feel of what navigation was in those days.

(Refer Slide Time: 40:55)

Ok so, done with history done with stories now let’s move onto some other cool stuff.
So, we have seen coastal navigation, dead reckoning, we have seen celestial navigation.
Now the more modern navigation techniques are highly based on electronics right. And
the one of the most classical examples which all of you would have used or really aware
of is a GPS based navigation. So, this gives you your position on earth in latitude,
longitude, and altitude. So, GPS works basically by how the receiver in your phone or
your GPS receiver look is looking at satellites in space.

The other thing is inertial navigation we’ll be discussing that lot in these lectures. And
inertial navigation uses accelerometers and gyroscopes. The really accurate type of
navigation on earth is by a combination of INS and GPS. So, it is basically a fusion it is
called INS GPS system inertial nav GPS based system. And here again, as we have
talked about in the dead reckoning and celestial navigation, the inertial navigation system

801
is very accurate over small time scales, over small time scales. And GPS has bounded
error, it is not perfect as many of you would have observed if you are trying to navigate
with GPS you actually get lost sometimes like couple of streets away from where you are
from where you should be, but it is a bounded error.

So, if you are in Chennai it would not show Bangalore and vice versa. Whereas, if you
navigate with a INS for say a few hours and you are still in Chennai it would actually
show Bangalore or something else. So, these are the standard navigation sensors which
are used. Recent research has focused on the use of cameras, laser-based ranging sensors
and radars to also do navigation, there is also Wi-Fi-based navigation you would have
possibly not used Wi-Fi, but you are you would have used your cell tower-based
localization, when sorry you would have used your cell tower-based localization when
you are booking your ola or uber vehicles.

So, when actually doing the booking it shows where your current position is on the map
and that is by the cell tower-based localization. Wi-Fi localization is a similar approach
based on Wi-Fi signals. So, for those of you who are really interested in looking at this
you have you need to look at autonomous navigation and if you give any of these
keywords say LIDAR’s or cameras it actually shows up a lot of really interesting
literature; however, the focus of our lecture is going to be only on inertial navigation.

802
 Control Engineering
Dr. Viswanath Talasila
Department of Telecommunication Engineering
Ramaiah Institute of Technology, Bengaluru

Module - 11
Lecture - 01
Part – 02
Navigation – Dead Reckoning and Reference Frames

Hello and welcome to module 11 lecture 1 part 2 of navigation. It is a warm day in

Chennai. So, I hope I do not sweat too much when talking now alright. So, in the
previous lecture we saw some basics of navigation. So, we discussed some really
interesting story of Alexander’s march, how desert ants search for food? How to do
celestial navigation? and so on.

We briefly discussed dead reckoning, celestial navigation, coastal navigation and

something about GPS and other things although that was very brief. What we are going
to do today is to dig a little bit deeper into the basic the first cut equations of dead
reckoning, and we will see why reference frames are very important. In the next lecture, I
will actually show you the kind of errors that are induced because of dead reckoning
alright and those errors primarily because of sensor characteristics very good.

(Refer Slide Time: 01:32)

So, let’s start inertial navigation. So, the reason why it is called inertial navigation is
because we use sensors which respond to Newton’s laws. So, you consider an

803
accelerometer or gyroscope and these objects have certain inertia like all physical
objects, and these objects they maintain their velocity either translational velocity for
acceleration or angular velocity for the gyroscope. They maintain their acceleration and
velocity unless they are disturbed by an external force that is a Newton’s inertial law
right.

So, this sensor when they make the measurements, when they disturbed by an external
force this is why they are called inertial sensors, alright. So, what do they actually
measure well they measure two things. One is the translational motion which is say the
motion of this object as it is going in a particular line right. So, it could rather go straight,
it could go left, right, up, down alright. It could also measure the rotational motion and
the rotational motion would basically be I have this aircraft, and I have an angular
motion with respect to an axis, I have this rotation and I have the third rotation we will
see a little bit more about this. So, since we live in a 3D world we are always going to
look at 3D translator motion and 3D rotational motion alright. Now there are various
types of inertial sensors the two sensors we will deal within in this and the next series of
lectures are the accelerometer and the gyroscope. We could also be doing the
magnetometer, but due to lack of time we will focus on just these two.

So, what does an accelerometer measure? Well loosely speaking it measures the
acceleration, and the units of that is meters per second square alright and accelerometer is
a very interesting object. So, whenever we talk of accelerometer or gyroscopes from this
time onwards, we will be referring to what are called as three axis sensors right. And
three axes basically mean that because we live in a three dimensional world, we have
motion along an axis we can call it X axis, we have a motion along perpendicular to the
X axis called the Y axis. So, that is in a plain, and then perpendicular to the plain we
have the Z axis. In the literature they commonly do not call it as X, Y and Z, they are
specific notations for this we will see this a little bit later, but for now we really need to
know that motion takes place in three dimensions, and each of these three axis sensors
they measure the motion along each of these axis alright. It could be a gyroscope it could
be an accelerometer as well.

So, I have also included some links for an accelerometer and for a gyroscope both are
analogue devices, but there are plenty of other options as well there is InvenSense and

804
there are Honeywell products and so on. This is just an example for you to see what these
products are about. A gyroscope measures the angular velocity we will see this a little bit
later what I specifically mean by this. So, the gyroscope is used to measure the rotational
behavior of an object, the accelerometer is used to measure the translational behavior of
an object alright.

Historically, accelerometers and gyroscopes would come as two independent products,

for may be the last decade or two decades’ manufacturers have integrated these
accelerometers and gyroscopes into a single device called the IMU. IMU is basically an
Inertial Measurement Unit. So, you these small chips which we shall show you may be in
a couple of lectures with some experiments which one of my research associates Vinay
Sridhar will be presenting, we will show you the basic IMU which are the sensors
mounted in that and examples of these ok.

So, we are going to be working predominantly with IMUs which have an integrated
accelerometer and a gyroscope on board good.

(Refer Slide Time: 06:18)

So, let’s look at little bit of dead reckoning in the very basic sense. So, this slide is
basically on dead reckoning alright. So, as I have said the general principle of navigation
using inertial sensors, is based entirely on dead reckoning. Now let’s look at the example
of accelerometers in this slide, in the next slide we will go to gyroscope. So, what does

805
an accelerometer measure? We saw that it measures the acceleration. So, here I have
written.

So, let’s say you have a three axis accelerometer, we consider a single axis let’s say the X
axis, you could also consider Y or Z and the measurement along that X axis will simply
we can denote it as, we can denote the measurement along the X axis as simply ac c x ,
and along the Y which I am not mentioning here would be ac c y along the Z axis
ac c z alright. So, let’s say we have made a measurement along the X axis, and now we
want to see how to compute position based on this. By the basic laws of physics, we
know that the integral of acceleration is nothing, but velocity right, inverse of that is
derivative velocity is nothing, but acceleration. So obviously, if you have if you are
integrating over the X axis, the velocity which you compute will be also along the X
axis.

Again from basic physics we know that the integration of velocity is nothing, but
position alright the position of the object. And as before if you are integration along X
the position computed will also be along X. So, that is basically what this expression
over here says right, which is basically the same as saying if you double integrate the
acceleration what you are going to end up with is a position X right we will do this. So
now that we have seen the basic say calculus of how we compute position from
acceleration, all of you have done this in high school. Now remember that our sensors
are usually connected to micro controllers right or to computers or to any electronic
device. Being computers and micro controllers, they can take only indiscrete values of
the actual data. So, if your real let’s say this is time axis, this is the acceleration along
let’s say they Y-axis right.

(Refer Slide Time: 09:18)

806
So, let’s say the vehicle was accelerating in the particular manner, when you actually try
to convert this or connect this sensor to a micro controller, you need to pass it through an
analogue to digital converter right, and depending on the resolution of the analogue to
digital converter, you actually get a 16 bit, 10 bit, 8 bits so on and so forthright.

So, the accuracy increases with the increase in resolution. So, with whatever accuracy
with whatever resolution we have taken, we will be able to recover some parts of this
data right. So, the complete curve is a true acceleration, what your sensor is able to
measure are these points it cannot measure anything in between unless you increase the
sampling rate of the sensor and unless you increase the resolution analogue to the digital
converter alright. So, now, in the discrete world or in discrete times, this is the
continuous-time. So, in the discrete-time which as all of you know nothing, but samples
how will this look like the measurements. So, I call it ac c y actually measured or
measured and then discretized and sampled it will basically look like this right that is
going to be your measurement.

Now, this is the actual data which is going to come into your microcontroller. So, I would
call this as a sample number 1, sample number 2, 3, 4, 5 and so on. Now based on these
measurements which we are taking the acceleration measurements our objective is to
compute our position right that is a whole point of dead reckoning, based on whatever
you know in the past you want to predict your future. So, based on the current
acceleration, I would like to know what happens next or a different way of putting based

807
on the acceleration which I have measured now I want to know where exactly I am. So,
you have to double integrate the acceleration.

(Refer Slide Time: 11:55)

How did we do this in the discrete world? Because you are now inside a computer your
data has come inside your microcontroller or your computer. So, how we actually

dv
compute this? Well, let’s look at the basic equations first right. So, we see that
dt
where V is a velocity is nothing, but your acceleration. So, let’s just take it along the X-

808
 V x (t+ ∆ t)−V (t)
axis this is nothing, but the limit is equal to ac c x ,we know this
∆t
from basic calculus that the derivative is expressed as a limit ∆ t → 0 ,blah blah blah.
So, on now of course, in the discrete world we do not have the delta t really tending to 0,
it is really small you know it could be 0.01, may be 0.0001 and so on, but it is never
perfectly 0. So, let’s fix the ∆ t to a certain value we will just call it as ∆ t , it could
be any number depending on how fast you are able to sample your signal alright. So, we
can remove the limit for now because it never really goes to 0 and you are left with this
basic expression in the discrete time. How do I if I expand this expression? I will
basically get v (t +∆ t )=v ( t)+( ac c x (t )× ∆ t) . This is your difference equation which
does the following; it uses the current information of the measured acceleration note that
this is a function of time, it uses the current velocity which is known from the previous
calculation which we have done and it calculates the next velocity. So, it predicts what
the next position the next velocity is going to be. Now I have expressed this slightly
differently in this equation over here. So, instead of t I have taken k and ok, so,
let’s see how this looks like. So, if I write this equation over here this will basically look
as follows. So, V x (k) note that in my case I took (t+ ∆ t) , in this case, I am taking
k because I am doing a (k −∆ k ) in the other place. So,
V x ( k ) =V x ( k−∆ t ) + ac c x ( k)× ∆ t alright. This is the same as this equation except that
I have used a slightly different notation over here and here.

So, we will not be using this notation so much this is the kind of notation which you will
see when you write your code, and I will show you this a little bit later. For our
discussion now we will use this notation we will come back to this when I explain the
code to you. So, this is how you compute the acceleration the velocity from the given
acceleration alright. Now how do we compute the position in exactly the same way that
we have looked at the computation of the velocity equation? So, now, that we know what
is a current velocity which we have computed by using the acceleration measurement, we
are just gaping to repeat the same process for computing the position and that again

dx
follows the same rule that is nothing.
dt

(Refer Slide Time: 15:29)

809
But Vx the velocity, this basically would say if I write this in the notation of k , it
basically say x (k ) plus or let’s do it. So, x (k ) is nothing, but x (k−∆ t) the
previous computed position plus the current velocity which I have computed from
integrating the acceleration before times the ∆ t ok.

So, this is the previously computed position, this is the current computed velocity; let me
recall the previous equation of how we computed the velocity. So, there we had
v (k)=v ( k−∆ t )+ac c x (k )∗∆ t . So, in that case again as we have seen now say. So,
this was the previously computed velocity right this is the current measured acceleration
by the accelerometer. So, this would be step one this would be step two. So, you first
compute the velocity based on the previously known velocity and the current
acceleration then having the value of this velocity now you plug this into this equation
for computing the position.

So, the current position is now equal to the previous computed position plus the current
computed velocity times ∆ t and that is basically this expression over here. So, this
will just be x (k )=x (k −∆ t )+Vx (k )×∆ t , that is wrong. We are going to take the
velocity plus velocity which is computed at time k ∆t right. So, these are my dead
reckoning equations, will basically follow from the mathematical principle that if I
double integrate the acceleration I get the position. To do that in a computer you need to
first integrate it once to get velocity integrate that the second time to get position. So,
these are the equations we are going to be using in our microcontroller programming
which I will show you tomorrow to compute the actual position.

810
(Refer Slide Time: 18:35)

So, this is the translation position because we using the accelerometer data, what happens
when we use the gyroscope data if you want to visualize this just think of this as let’s say
there is a disk over here, and it is spinning along this axis. So, this disk will have some
inertia, it will have some mass and so on and you apply a torque on this disk and you
rotate it right.

Say, for example, I have connected a motor over here and I am continuously rotating the
disk, and here if I place a gyroscope at this location it is going to measure the angular
rotation rate in the units of degrees per second. Now that is what the angular rotation rate
is. Now what I want to compute is the angular position. At every instant of time, I want
to know where this disk is. I want to know whether I have rotated 5 ° or 10 ° and
so on. Given the measurement is angular velocity.

So, of course, by basic physics, we know that integration of angular velocity is nothing,

d θx
but angular position and we use exactly the same story as before since along the
dt
X-axis, of course, is ω x . I will rewrite this in time in the discrete-time domain which
we get after sampling the angular velocity data from the gyroscope. So, this will
basically become θ x at the sample number k is the previous computed angular
displacement. So, k −∆ t plus the current measured angular velocity, which comes in
from my gyroscope times ∆ t and that is exactly what we have over here ok.

811
So, this is actually should be delta t it is more a computer notation which I have used
over there. So, this will come to later when we looking at the code, we will consider this
equation for now. So, we see again this is a first-order difference equation because we
only had a first-order difference equation over here alright. So, in both the cases for the
accelerometer and the gyroscope we compute the difference equations in order to get
the angular position and the angular and the translatory position alright. So, that is done
and now, of course, we have a slightly different issue, which is basically that each sensor
as I have said before is a three-axis sensor.

(Refer Slide Time: 21:29)

So, this specifically means is that let’s consider an accelerometer. So, again I will explain
this very briefly before and let’s say that you have a vehicle it is a car or something like

812
that it is a pretty ugly car, but is good enough for our story. So, let’s say you have a car
and it is moving, it is its moving on a terrain right it is not a street road it is moving on a
terrain like an off-road vehicle and in the Himalayas or somewhere, and let’s say we have
this axis which we denote with X, Y and Z. Now if it is moving on only one in only one
direction in the only axis say it is a perfectly flat road.

So, let’s say it is moving only along the X-direction, the three-axis accelerometer which
can be visualized like this. So, this will be one axis the other axis and the third one each
one is orthogonal or perpendicular to each other. So, this three-axis accelerometer will
measure the acceleration of the car only in the direction in which it is moving right. So, it
will not measure this direction or the acceleration measurement will be 0 in this
direction, it will also be 0 in this direction alright.

It measures only in the forward in which it is moving similarly if it is just moving left
and right say a plane or helicopter moving left and right, along this axis over here right.
So, along this axis, it actually gives me a non-zero acieration, but along the forward axis
it will show 0 along the along the vertical axis it will show 0 right. If you move in a
combination of two different directions, both the X and the Y axis measurement of the
acceleration will show non zero values. If it moves in 3D all three axes will show non
zero values. So, that is what an accelerometer measures, a gyroscope in a similar way
again have.

So, I can call this as acc x along X, acc y along Y, acc z along Z or gyroscope
will measure the angular rotation rate along with respect to an axis I will give a brief of
that now. So, let’s say that I have my three axis sensor, we will call this the one pointing
say pointing towards you as a forward axis, the one pointing to the right as one axis and
the vertical axis completes the orthogonal system or the perpendicular axis system. So, I
have three axes all perpendicular to each other. Now let’s say I take the system and I
rotate it about this axis alright how would I rotate it about the forward axis? I keep the
forward axis fixed and I rotate it right. So, you can see the motion over here.

So, this motion is an angular rotation about the forward axis, what about the rotation
about saying the right side axis the X the Y axis. So, I keep this one fixed and I rotate it
this way alright. So, this is our angular rotation about this axis finally, I keep this fixed
and I rotate it over here, this is another rotation about this axis. So, we have three

813
rotations and the rotations are measured with respect to each of the axis. So, it would
look like this. So, I have a rotation about the X axis or I can have a rotation about the Y
axis or rotation about the Z axis or you can have a combination of rotations.

So, it need not just be a rotation like this or rotations like this, it can be a rotation like this
right and that is a combination of all three rotations. When you take all of these then
basically what you will get are these three equations in the computer in the discrete time
case right. So, this is the computed position along the X direction assuming you have
computed the velocity along X, you have the position along Y position along Z and
similarly we have the angular position along X, angular position Y, angular position
along Z after we made the measurements along X, Y and Z angular velocity
measurements. For those of you who either watch lots of movies or play some nice video
games like flat simulators and those or read some crazy novels, you may immediately
recognize that these three angular positions are very commonly known as the roll the
pitch and the yaw alright we will see this in the next slide.

(Refer Slide Time: 26:33)

So, just small terminology the word attitude in navigation it doesn’t refer a good attitude
or bad attitude, it refers to these three angles we are talking about. The roll angles the

814
pitch angle and the yaw, and note that each of these angles is a rotation about one
specific axis. So, for example, the roll angle is usually called as a rotation about the
forward axis. So, I keep the forward axis fixed and I roll my vehicle right. If I keep the
side axis fixed and I pitch up and down, you know.

So, the pilot who is sitting in the plane they pull on the you can call it a joystick they pull
on the joystick and the plane pitches up you push it down it pitches down alright the final
one is a rotation about the Z axis, this is actually the heading or the yaw. So, you keep it
over here and you ask the plane to move to change the direction to the right. So, it rotates
to the right left it rotates to the left about this axis. So, that is a roll pitch and yaw and
you can see this small illustration over here. So, you mount your sensor at this place the
gyroscope and this is the forward axis, this is the vertical down axis as it is called, and
this is the right axis which completes the orthogonal axis system. So, whenever you talk
of roll you typically talk of the aircraft rolling with respect to the forward axis, as you
can see my forward axis is not moving. When you talk of pitch this axis doesn’t move.
So, it is just a rotation like this and the same for the yaw. This doesn’t apply only to
aircraft it applies to cars as you can see over because of cars especially off road cars
when they move on uneven terrains you have a roll, pitch and yaw.

As many of you would have experienced if you take a car ride through let’s say
Bangalore’s amazing roads filled with potholes, you often you know you go into these
kinds of roll, pitch and yaw. And of course, this applies to all movement you can talk of
human movement right. So, as I am standing here with respect to you, I may be doing a
yaw, I may be doing a pitch up, my hand may be doing a combination of angular
rotations and so on. Every one of this can be measured by the use of a gyroscope fine
now attitude ok.

(Refer Slide Time: 29:51)

815
So, the roll pitch yaw attitude is good, but sometimes attitude is not good and it is always
problem with attitude well. Let’s see what is a problem with attitude over here as it is not
just attitude, I need to be more precise attitude as well as a translational position or
translation movement you can call it. So, let’s say there one of you cool people are lying
on a beach you know wearing some dark shades and they are just looking up at the sky,
and you see a nice aeroplane flying across over here alright. And you know there is a
friend of yours who is sitting in another part of the town or the city, and you want to call
up this guy and say, hey! look man, I saw this amazing the LCA Tejas aircraft you know
India’s own fighter combat plane, brilliant stuff.

Now, this guy is going to ask. So, the guy who is sitting in the town he is going to ask his
friend dude, where did you actually see this plane. Now, this guy is going to say I saw the
plane let’s say approximately 2 kms above me. This guy has no clue what it means by
2 kms above him. He doesn’t know where he is right. So, the guy in the town does not
know where the guy on the beach is. He doesn’t know what 2 kms above him means
especially if this guy is sitting on top of a mountain, and this guy is on the beach at
almost at the sea level.

So, he has no clear idea of where the plane actually is with respect to him. The person on
the beach knows where the aircraft is, but the person here based on the information
which the beach dude has given him he is unable to configure or figure out where exactly
the aircraft is. Is a further problem, you have the aircraft shown over here now, of course,

816
we need the aircraft flight control system needs to be able to compute it is own position.
So, it places this gyroscopes accelerometers and all these things in the aircraft. Now,
where does we place it let’s say for example, it has placed at let’s use a different color
white let’s say it has placed at over here. So, this box is actually a very small box is
placed over here, and this box is actually going to the measurements of angular rotations
as well as the accelerations the IMU.

Now, what is a problem with this as you can see the aircraft axis is shown over here with
a roll the pitch and the yaw, and aircraft axis is typically defined with respect to the
center of mass of the aircraft or the center of gravity of the aircraft. The IMU, in this
case, is placed at a slightly different location and it is not aligned with the aircraft axis.
So, if the aircraft axis is like this, the IMU axis could be like this right because it depends
on how I placed my aircraft. So, if this is an aircraft very bad example I agree, if this is
an aircraft let’s say this is an IMU. if I place it perfectly aligned with an axis of the
aircraft that is good, but because of various reasons, I have placed it like this and the axis
are completely miss-aligned. So, in that case, the axis of the IMU will be like this that
would be the forward axis this would be the pitch axis if you want to call it, and that
would be the vertical axis. And you can see that the two axes are not aligned at all, which
means the IMU which measures that aircraft acceleration and angular velocity, it is
measuring in it is own frame of reference right the three axes of the accelerometer. With
respect to these three axes it is telling the aircraft is rotating about certain degrees per
second, but the aircraft is in a different axis representation.

So, the aircraft is in this axis representation the IMU is in a different axis representation
alright. So, you essentially have now a bigger problem; a guy the pilot in the aircraft if
she or he looks at the data coming from the IMU. So, let’s say the data coming from the
IMU is displaced on the screen of the pilot and it is saying that look man let’s say your
pitch angle, is currently going like this.

And the pilot is not really sure if this is correct or not because this pitch angle is
measured by the sensor, and it is not the same as the pitch I actual pitch angle of the
aircraft because it is in the slightly different position. For all you know, if this is on my
left hand this is aircraft axis the right hand is the sensor axis, the way the designers
mount the sensor on the plane it could be completely 90 ° shifted. So, the pitch axis

817
becomes the roll axis becomes the pitch axis. So, when the pilot is looking at the data
from the pitch over here, it is actually the roll axis of the aircraft and the pilot has to
make a decision based on that it can be catastrophic.

So, this is a problem with reference frames the problem in attitude, the problem in
transitional movement, every object which measures movement, measures it in it is own
frame of reference, it is own axis or coordinate system. The actual system the plane or a
robot or a human walking is experiencing motion in it is own reference system. So,
typically you have the system which could be a plane or a fancy robot a human walking
and so on, this has it is own frame of reference it is own coordinate system this is called
the body reference frame, a body reference frame.

A sensor which is mounted on the body to measure the motion of the body or you can
actually say there is a camera over there and it is capturing my movement, the camera is
recording my movement or the sensor on my body is recording my movement in it is
own reference frame called the sensor frame, and then you have an observer like you or
me standing outside and just you know watching this cool plane flying above, then you
have an observer frame. Now with three different frames how are you going to say the
aircraft is exactly there or the robot is exactly over here, you know autonomous robot
which are wandering all over the place. Like this terminator movies, how would you say
that the terminator is right over here. You can only do that if there is a standard reference
frame and this is actually a very important problem in robotics in ship and aircraft
navigation, human movement everything. Every movement must be expressed with
respect to what standard reference frame it cannot be expressed with respect to his
reference frame or her reference frame, it should be with respect to a standard reference
frame which everyone understands.

However, measurements are done in the sensor frame, the body is moving in it is own
frame, the observer has his own frame. So, what we will need to do is whatever
measurements are done in the sensor frame we need to translate it or we need to
transform it into the standard reference frame. Whatever the body is experiencing
transform that into a standard reference frame, then you know exactly what is happening
to the motion of the object alright.

(Refer Slide Time: 38:26)

818
So, this is the problem which I have been talking for the last two three minutes, and the
way that people address this problem in navigation is, when a sensor measures you know
the acceleration and angular velocity and all three axes. We transform the measurements
usually first into the body frame into the aircraft or to the robot. So, we know exactly
what is a measurement of the aircraft movement, we do not care what this sensor is
experiencing really we care what the aircraft is experiencing specifically the movements
right. So, we transform the sensor measurements into the coordinate system of the
aircraft or the robot or the human being who is walking, once you are in the body frame
of course, you are body frame is different from his body frame right it is the body
coordinate system is different from his body coordinate system. So, you need to now
transform it into the standard reference frame. So, for each of these transformations you
have something called a rotation matrix. So, apply the rotation matrix and you actually
go from one reference frame to another reference frame. We will see now one simple
example of how to actually do the rough the rotation matrix then we will go to the
complex 3D case. So, let’s see how we can do this reference frame transformation, we
will start with a very single simple example.

Let’s frame the question properly. So, the question would actually be given a motion
which is measured in one reference frame, let’s say it is the sensor reference frame ok.

(Refer Slide Time: 40:02)

819
How would you represent this motion in another reference frame, let’s say the body
reference frame alright? So, that is the basic question which we are asking, we will
actually pose the question as follows. So, given a motion which is measured in one
reference frame, how would we represent it in another reference frame? We always have
to keep track of two quantities one is the translation the other is the rotation. We will
focus primarily on the rotation over here, I will briefly explain how to take of the
translation afterwards. So, we have two reference frames as you can see here the sensor
reference frame and the body reference frame, you need to represent one in terms of the
other, let’s take a very simple 2D example, and in this 2D example I will assume I have
got a vehicle and the center of gravity of the vehicle is over here, with an axis as follows.
So, I will call this x body axis( x b ) and y body axis( y b ).

Now what I have done is to mount IMU at exactly at the centre of gravity of the vehicle,
now usually this will not be the case and if it is not the case you need to take care of the
translation part as well. If you mount it exactly at the centre of gravity of the vehicle, you
only have a rotational problem which is what we will deal with here. So, let’s say we
mount the sensor like this and the sensor origin coincides with the origin of the body
coordinate system. Except that the sensor is misaligned with respect to the body
coordinate system. So, the sensor coordinate system will be like this.

I will call this as or let me rewrite that a little bit, I will call this as X sensor( x s ) and
this is Y sensor( y s ¿ . So, I can see that the two origins are perfectly aligned, there is a

820
difference in the rotation between the two sensors; and it is in intuitively very clear
already by now to most of you all you need to do is to merely rotate the sensor reference
frame by certain angle to match with the body reference frame alright, and that is exactly
what the 2D example is all about. So, let’s say that in the sensor reference frame the
x s and y s , I have said with respect to the origin, of course, we can call it (0,0) ,
it doesn’t really matter it can be any (X , Y) because two origins are perfectly aligned.

Now with respect to the origin let’s say I have measured a certain motion which is over
here and this could be acceleration, it could be angular velocity, whatever it is depending
on whether I am using an accelerometer or a gyroscope over here any one of the sensors
you can assume. And you see that any point in this is a measurement with coordinates
and it can be called as x 's and y 's . What is x 's and y 's ? With respect to the
origin it is simply a vector right basic high school physics? So, this is a measurement
'
which is done, I need to now take this vector which is the sensor measurement x s and

y 's , and I need to transform this into the body reference frame. Because I want to
know in my body reference frame what is the measurement, which I am seeing I do not
care what is in the sensor reference frame I want to know in my body reference frame.

So, given this how do I transform it into the body reference frame, very simple. So, we
have the two this is the sensor frame, and then you have the body reference frame right
x b and y b this is x s , y s . Now you need to know only one important quantity in
the 2D example only one quantity, which is what is the angular displacement between the
two reference frames that is all I mean it is intuitively clear to all of you by now what it
is that that we need. So, if we know what is this angular difference let’s just call it θ. If
you are able to measure this θ we know exactly what we need to do well; let’s say our
gyroscope is actually able to measure this θ or we know from a third sensor or we have
manually measured it ourselves, that the two axis have slightly deviated with respect to θ
and you can actually measure this right.

So, you can actually put some sensors over here and you can say that there is an angular
displacement between these two reference frames. Once you know that you just going to
' '
do the following. So, given x s and y s X sensor 1 and Ysensor 1, I want to express
' '
this in my body coordinates right. I want to know what is x b and y b ? X body 1 ,

821
 Y body 1 . And the way you will do it is very simple you would actually do it as follows.
' '
So, x b=x s cos θ X body 1 would be nothing, but x sensor one times cos (θ) ;

y 'b= y 's sin θ Y body 1 would be nothing, but y sensor one times sin(θ) and how

would you express this as a matrix? very simple. So, I would just put [ cos θ
0
0
sinθ ]
that’s it.

So, when you do this, now I have a measurement in sensor frame, now I am able to
express the measurement the sensor measurement in the body frame brilliant, simple
straight forward in the 2D case. The 3D case becomes a little bit tricky, but it is almost
the same logic, how we do that well as follows, in a 3D case note that you can have a
rotation about all three axes.
(Refer Slide Time: 47:23)

So, if you look at the previous slide, we basically said that there was a rotation of θ°
about the X axis, which means even the Y axis will also be θ° rotated with respect to the
other Y axis alright. So, what I am basically saying is if this is θ° over here by the
property of orthogonality or perpendicularity, this will also be θ° exactly the same
because both axes are orthogonal for both the reference frames. Not true for 3 axis
system, it is little bit more involved and the reason is little bit more involved is as
follows, 2 reference frames let’s take the forward axis this one over here, and let’s say

822
that you have a slight yaw of the second reference frame, yaw is a rotation about the
vertical axis right.

So, the above the vertical axis the second reference frame is slightly yawed by some
angle γ . Now, because it’s 3D this is not the only case which we can have, about this
axis over here, which I am showing with my left finger, about this axis I have a slight
pitch with respect to this guy over here. And now I still have a third axis left the forward
axis, the roll axis I have a slight displacement with that. So, I have to now consider three
angles, because my axis can be misaligned with respect to this axis in three different
ways, I can have a yaw displacement I can have a pitch displacement and a roll
displacement. All three conditions have to be considered it is not difficult it is fairly
straight forward, you just need to use the correct math.

Again we take the same example of the sensor frame and the body frame, but please
remember you can use this for any two frames. So, the assumptions I am making here is
that the sensor frame and the body frame have the same origins. So, in that case, we have
this origin over here the two reference frames, because we have three possible angular
displacements between the two reference frames, we are going to consider each angular
displacement independently.

So, first in this particular example I am taking we will assume that the sensor frame is
rotated by an angle γ right over here, with respect to the body frame with respect to
the Z axis or the body frame. Turns out if you do the math properly, because when we are
doing a rotation about the Z axis notice that the bottom one is fixed. So, when you rotate
about the Z axis, the third component which is the Z axis component will all be zeros
except the Z axis itself because the Z axis is fixed, it is not moving it is the other two
axes which are moving right.

The other two axes by standard trigonometry relationships will follow this relationship,
you do not need to worry about the derivations of this maybe those of you who go for
higher studies, you can study this later not too relevant for now it is important to know,
that if I want to go from the sensor frame to a frame called b1 which is body frame 1
where the Z axis are now aligned. So, remember I had 3 misalignments, a misalignment
about the Z axis, misalignment about the Y axis that is a pitch axis and about the roll let
me correct these one by one. So, first let me correct the Z axis. So, I do a rotation about

823
the Z axis this one over here. So, that rotation I do first the second rotation which I will
show in the next slide will be about the pitch axis. So, I will do like this, and the rotation
will be about the roll axis. So, I am going to dot this step by step.

So, about the Z axis, if you blindly go ahead and use this expression this is called the
rotation matrix, and this rotation matrix helps you in this example to align sorry about
that, it helps you to align the two Z axis of the two reference frames that is all what it
does. So, which means that if I have a vector my measurement xs ys zs in the sensor
frame, this I can use this rotation matrix and bring it to the body frame where the Z axis
of the two reference frames are aligned we still have to take care of the other two
reference frames. So, this is step 1; step 2 let’s say that your body frame was now rotated
by some θ about the X axis, in the previous case it was γ about the Y axis now θ
about the X axis. In this case again since it is about the X axis, X axis will always remain
all terms will be 0 and X axis itself will be fixed. So, it will always be 1, and we look at
the rotations in the other two components and it has a very similar structure as the
previous one. So, this is again my rotation matrix and notice what it is doing in the
b1
previous example we had this rotation matrix with the notation Rs which means it is
a rotation or a transformation from the sensor frame s to the sensor frame b1 where
the Z axis are aligned.

b
In this slide we have Rb , this is a rotation from b1 reference frame to b2
2

reference frame where Z axis is aligned as before and now we are aligning the X axis as
well. So, now, two axes are aligned and you can see that over here right both the axis Z
and the X are aligned after you apply the rotation and again this is a standard expression.
So, the measurement in body frame 1 where the Z axis were aligned, is now transformed
to a measurement in body frame 2 where X and Z are both aligned, that you can see in
this slide after the previous two transformations the Z and the X are now aligned with
respect to reference frames; and what is left out is the one axis.

So, let’s say you have the angular displacement by β with respect to the new Y axis,
do exactly the same thing as we done before. So, we have had from the sensor frame to
body frame 1, after this we have done body frame 1 to body frame 2 in this case we align
the Z axis. In this case we have aligned the Z axis and the X axis and now finally, we
look at the rotation between the two Y axis of say body frame to the final correct body

824
frame b2 to b, and now all three axis will get aligned X axis, Z axis and the Y axis,
and you can see that over in the next slide.

So, when you rotate this by the appropriate angle all three axis will be perfectly aligned
with each other right. So, like. So, and this is a rotation to go from b2 to b. The second
body axis to the final body axis, and interestingly all you need to do in order to do this
very complex sounding mathematical description is one very simple operation.

(Refer Slide Time: 56:04)

Take the three rotation matrices and multiply them in this specific order; on the right
hand side from the sensor frame to body frame 1 that is one rotation matrix, multiply that
with body frame 1 to body frame 2 the second rotation matrix and finally, multiply that
result with the body frame 2 to the actual body frame the third rotation matrix. You
actually do the multiplication and you get this horrible looking expression, which you do
not need to worry about considering you are an under graduate student, you can actually
use this as it is. And this follows from very basic trigonometry rules and really nothing
more than that.

So, the interested student can contact me on the forum NPTEL forum or you can look up
in Google to find out how you actually derive these expressions. Now this means. So,
this formula what it actually means is this, it means that any measurement I do in the
sensor frame. So, I can call it xs ys zs any measurement done in the sensor frame, if I

825
multiply it with this rotation matrix, I will exactly get what the actual measurement is in
the body frame. Now the body and the sensor frame are aligned, they know exactly what
they are both talking about like where the vehicle actually is, when you just multiply
with this which is this matrix alright.

Now, note that I have talked only about rotation, I have not talked about translation. Well
translation is slightly different it is even simpler. So, let’s say that you have two rotation
frames like this, and a second rotation frame over here origins are not aligned, that is a
basic difference between the example considered here and the example I am now
considering well how to handle this really simple. So, this example we will write down
over here and it is really straight forward how we are going to do this is as follows. So,
again we have the measurement in the sense of range z s , the first thing we are going
to do is to align the rotation axis. So, by multiplying with our matrix which we have
already computed, after doing that we shift, I will actually draw this over here. This is
one sensor axis and this was the other sensor axis the body axis let’s say. Now what we
are going to do when we apply Rbs , this axis over here it gets shifted displaced rather
angularly displaced to this axis.

So, now, the angles at which the two reference frames are being considered in are exactly
the same, what is missing is I need a translation from here to here, I need to go I need to
shift the origin from this place to that place. Now let’s not worry how that is actually
calculated, let’s say someone actually tells you by measurement. So, you have your car
and this is the center of gravity and you have placed your. So, it is in some x, y, z body
position origin, and you have placed your IMU over here, you can physically just
measure it right.

So, you know that this is some x 1 , y 1 , z 1 where you have placed your sensor and you

know the difference between x and x 1 , you know the difference between y and y1 ,
z and z 1 subtract all those differences and all you need to do is to add that small
difference. So, I will call it x difference in the translation position, y difference in the
translation position, Z difference in the translation position. If you add this to the rotated
values you are going to get perfectly xb , yb , zb .

826
So, first take care of the rotation between the two axes, then align the origins done
essentially done three dimensional inertial navigations. So, let me try to summarize this
entire story which we have been talking about which may seem a little bit complex when
you encounter it at the first time, if you go through this two or three times it actually
fairly simple all we are doing is just trigonometry in 3D. So, let me just try to summarize
it as neatly as possible. So, the problem was as follows.

(Refer Slide Time: 61:47)

So, given two reference frames let’s call this as a body frame, and another reference
frame is the sensor reference frame. Now we know that a measurement is always done in
the senor reference frame, the sensor measures values with it is own three axis reference
frames, which needs to be converted to the body measurement frame. Then of course, we
need to convert the body measurement frame to the standard reference frame, but will
come to that later.

So, how do we actually do this well as we have seen there are two steeps, the first is
align the axis rotations align the axis of the two reference frames that is. So, if you align
the axis of the two reference frames even though they are widely separated from each
other, as you can see this is angularly different from this one let’s first align them
together alright.

827
Now to align them you only need that one Rbs matrix you know that big three cross
three matrices which you had. You only need three angles as you saw there was a γ , θ
and what was the last one a β . So, if you have three angles, you can actually align the
reference frames together angularly align them. How do you compute these γ , θ and
β there are various techniques, for our case let’s actually supposedly we do it
manually? So, as I said before you have a vehicle with the center of gravity over here,
and the sensor mounted over here I mean that’s a axis like this all we need to do is to find
out what is angular displacement between these two things alright it is a bunch of ways
of doing it, but we can just literally do it manually take I mean if you are really lazy just
take a protractor and try to align them and that is that will be reasonably accurate alright.
If you want very high degree of accuracy there are really fancy techniques, but let’s not
worry about that.

So, all you had to do was to compute the Rsb matrix, and then multiply the vector the
sensor measurement that you do with a Rsb matrix, then you get the measurement in
the body coordinates with axis aligned. So, this is the axis angularly aligned body
coordinates it will still not align in the translation way for that all you need to do is to
just compute this guy. So, if you know the x 1 , y 1 , b 1 of the body and you know the
x 2 , y 2 ,b 2 of the sensor.

So, again very easy to do the easiest way to do it without actually breaking your head too
much is to assume the body axis origin is 0 0 0, just assume it 0 0 0 no worries. You
compute the displacement along the x direction of the body to see whether sensor is. So,
if the body is over here let’s say this is the origin alright origin of the sensor. You
basically just go along the x direction and you measure and you get the displacement
delta x with respect to the sensor the x displacement, same thing does with y.

So, keep going until you see what is that y and the same thing with z then all you need to
do is to add the ∆ x , ∆ y , ∆ z . This is the displacement the translation displacement
between the origins of the axis. And once you do all that you get a completely aligned
reference frames and whatever you measure in the sensor axis now you know it is
exactly the same as what you actually have in the body axis that is a basic principle you
are free to contact me if you want specific code for this if you want more explanation
literature about all this you are free to contact us on the forum.

828
(Refer Slide Time: 66:50)

Ok, well this is fairly simple there are fortunately or unfortunately, there are many too
many reference frames with which are actually in use. So, while we like to say that the
there is one standard reference frame it really depends on what your purpose is. If you
are on earth one of the standard reference frame is called the earth centered inertial
reference frame, where the origin of the reference frame is the center of mass of the earth
and one axis points towards north other axis point towards the vernal equinox, and the
third one just completes the right angle triangle along the equator.

So, this is one reference frame. So, anywhere you are on earth right can I see India here
no I cannot see in the map, but anywhere you are on earth you can define your position
with respect to these three axis, and actually this is what will give you the standard
latitude longitude and altitude measurements. But this is if a vehicle is on earth, but for a
satellite, you have a different reference frame altogether and one of them is called the
orbital coordinate system. If you are inside a room like the one that I have it really makes
no point to worry about orbital coordinate system and for very specific applications, you
may not even care about earth-centred reference earth-centred inertial frame, you may
just literally worry about this room axis.

How do you define the room axis? I can just take the corner of the room, I go down
where all the walls are meeting at one corner and I can define that to be the origin. So,
reference frames are very important and it is up to you to decide what reference frame

829
you want, you can choose the reference frame of this room. Assume your robot is
navigating within this room it is only job is to worry about what it is doing in this
particular room. A person outside who wants to monitor the robot is also interested in
what the robot is doing in the room we do not really care if this room is in Chennai or
Bangalore or Mumbai or wherever right. It is enough for us to know that in this room the
robot is doing it is job properly, for that you just need a room reference frame I can take
the axis over here and then define it.

If you are going from actually I am planning a trip from Chennai to Mahabalipuram
tomorrow, because you know I am visiting Chennai from Bangalore then my room
reference frame is of is of no use, my car reference frame is of no use I really need an
earth-centered inertial reference frame. I need a GPS kind of reference frame the latitude
longitude altitude that is more relevant for me. So, what you use how you use is up to
your task, in all cases the way that you would compute the way that you would express a
motion in one reference frame in another reference frame is exactly by the story which
was described. This was the basic story of say navigation where we discussed different
navigation techniques.

(Refer Slide Time: 70:18)

We looked at very briefly what is dead reckoning which was basically the integration
process, and we represented the integration process with a differential equation which is
going to be implemented in a micro controller and finally, we concluded with reference
frames.

830
In the next lecture what I will be talking about is one of the most important or the most
important problem in dead reckoning it is a problem called drift and we will see that drift
arises because of certain noise characteristics of sensors, and very briefly I will introduce
one or two techniques of how we can remove this noise in order to get a better
computation or estimation of position. And we will finally, conclude with a very nice
detailed experimental example right here my research associate Vinay Sridhar will
actually be taking care of that.

Thank you very much.

831
 Control Engineering
Dr. Viswanath Talasila
Department of Telecommunication Engineering
Ramaiah Institute of Technology, Bengaluru

Module – 11
Lecture – 02
Inertial Sensors and their characteristics

Hi there and welcome to the next series of lectures in control systems. So, in this lecture, we
going to be talking about little bit more detail about inertial sensors and their characterisation.
This is usually a topic that students in instrumentation engineering take up, but this is
becoming far more relevant for people across various disciplines because the measurement
and analysis of movement; right is now spanning across multiple fields, while initially this
was treated as a purely instrumentation issue which was mostly concerned with aircrafts and
ship like navigation, ever since movement began started to become of great interest in
robotics, in human movement analysis for medical purposes and so on.

The overall usage of these inertial sensors has become far more critical, right. So, now, we
have people from sports who are worried about movement analytics, people from even dance
and theatre who were worried about movement and so on and so forth. So, it is not just core
engineering, it is also lot of allied subjects as well, which is why we are going to be looking
at some characterization of inertial sensors, how they affect the computation of position and
attitude, attitude means roll, pitch, and yaw angles and why they are important.

832
(Refer Slide Time: 01:52)

So, the story so far, we have seen various types of navigational techniques and we
specifically said that we are going to be focusing primarily on dead reckoning, right and in
the previous lecture, we discussed how to numerically integrate; we discussed how to
numerically integrate, the measurements which are coming in from the accelerometer and the
gyroscope in order to compute angular position and angular and translatory position. So, just
to refresh your memory, to compute the angular position we did something like the angle at
time k . At the sample, k is the previously computed angle at the previous time sample
th
plus the current measurement of the angular rate at the k sample times ∆ t , right.
θ ( k ) =θ ( k−∆ t ) +ω (k )∆ t

So, this was the numerical integration equation we used in order to compute angles and we
did a similar thing for positions as well. What we are going to do today is to look at
accelerometer we already seen, how to compute velocity, position, we specifically going to be
looking at noise characterization. More specifically, we are going to be looking at a problem
of drift which really comes about mainly in the gyroscope thing, it is actually not part of
accelerometer, the reason I put this here is, it is very common mistake that people think that
drift is a problem with IMUs, it is not. It is a problem which comes only with gyroscopes, not
with accelerometer. So, you should remember that.

Similarly, we going to look at the gyroscope, various noise characterization of that and then
what kind of conditions it has; noise characterization and then we are finally, going to

833
conclude with a very brief description of what kind of filtering techniques we may use in
order to reduce navigational errors, right, good. So let’s go ahead and quite a few of my
students have contributed to various works in navigation that are been involved over the past
few years, we have done projects with hospitals and autonomous vehicles and these people
have actually contributed a lot and specifically Vinay Sridhar, he is going to be showing some
experimental results over the next few lectures, alright.

(Refer Slide Time: 04:37)

So, in this lecture; well, this is what I have basically described while you need to remember
one thing, there are some extraordinarily important details of sensor characterization which
we will be brushing aside really quickly, without going too much into the details of that’s
actually very important. If you want to have extraordinarily good navigational precision,
accuracy, you need to really worry about a lot of details, but the objective for second or third-
year undergrad student is really to just have enough information to develop your own
navigational system by the end of this course, alright, that is really the objective of this set of
lectures.

834
(Refer Slide Time: 05:28)

So, general introduction; right, there are various types of inertial sensors, and these basically
dependent on how you actually fabricate and design them. So, you have mechanical optical
MEMS and so on, we are going to be exclusively focusing on the use of MEM sensors in
these series of lectures, I will explain this while later, the main disadvantage of mechanical
sensors is they contain moving parts, right. So, there is a significant amount of wear and tear,
fairly accurate, but they suffer from this problem optical sensors are extraordinarily accurate,
but they are very very expensive MEM sensors are the kind of sensors that you know, you
have in your phones, when you do the portrait to the landscape mode, you know, it shifts, you
also have these sensors in various drones and UAVs, small robots which you keep playing
around and so on. So, MEM sensors are really cheap. In fact, as I have written here you can
buy simple MEMS 3 axis accelerometer for just a couple of 100 rupees and yeah and the
really important reason why they are so useful is that they are highly versatile, I can actually
just take a MEM sensor of the market and directly interface it to my accelerometer and start
to work with that to do that with optical sensors mechanical sensors require a lot of additional
electronics.

The problem with MEM sensors is that they are significantly corrupted by noise as compared
to the optical or the mechanical sensors and we will see a few of these noise
characterizations, but to make sure that you still continue to use these MEM sensors, there are
really nice algorithms or primarily you can actually call them as filters or filtering techniques

835
which had been developed to compensate for this noise and 2 of the most commonly used
techniques one is the complementary filter.

(Refer Slide Time: 07:50)

Which we will be discussing in the next series of lectures and other one is a Kalman filter,
right well. So, let’s go ahead, effect of sensor noise on dead reckoning computations. So, we
have seen the first-order difference equation for computing the angle angular displacement
once you measure the angular rate, right. So, this is the first-order differential equation, this is
actually what gets implemented in your computers. Now it is important to remember that
every sensor, not just inertial sensors every sensor which is actually going to use is, the
measurements are highly corrupted by noise. So, you can actually see this very often when
you try to take you know, your selfies if you do 10, slightly poor lighting, you get lot of
grainy images, if even if you do it in a extremely good lighting conditions and you try to
zoom into the camera, into the image, you start to get lot of blurring effect and grainy images
and so on this is true for all sensors.

You measure the temperature of this room by putting a very accurate digital thermometer
costing you know 100s and 1000s of rupees, but still not going to get the perfect temperature
because always going to be noise and of course, the same thing happens for inertial sensors as
well. So, every measurement you make is actually a combination of two things, one is the
actually dynamics which is happening. So, for example, you are measuring temperature, it is
actually temperature in the room plus the noise. So, this can be written for a gyroscope as

836
follows. So, the measured angular velocity or the measured angular rate is basically the true
angular velocity plus and I am going to write this with a circle, I will explain why plus noise
now.

So, this is actually what you measure every time and you are never going to directly get the
true angular velocity its always corrupted by noise the meaning of this circle over the plus
sign is some noise are additive in nature. So, you actually simply add it up some noise can
actually be multiplicative in nature; there are various kinds of noise as well. So, what we are
going to consider here is primarily the additive noise. So, the angular velocity measured by a
gyroscope is a true angular velocity plus noise that is a basic kind of equations we are going
to be dealing with or behaviour.

(Refer Slide Time: 10:45)

So, let’s see one of the consequences of this. So, let’s suppose and we look at a very specific
example of one noise parameter which is a constant bias, what is a meaning of a bias let’s
again take the example of the thermometer, let’s say the temperature in this room, you know
god or somebody has told us it’s exactly 23.1℃ , exactly no other decimal points
23.1℃ .

The thermometer is going to show always 23.12 ℃ . So, the exact temperature in the room
was 23.1℃ , god or someone has actually told you this, your thermometer shows
23.12 ℃ , this is by the thermometer, let’s go to another room, a different room where the
temperature in the new room is 43℃ and your thermometer will show 4 3. 02 ℃ . Go

837
to another place, there will always be an offset from the actual true temperature. In our
particular example by 0.02℃ , this constant offset; this constant offset is called a bias is
there in every sensor. So, let’s see, and actually, it turns out that this is one of the most
dominant noise effects which you find in all sensors.

So, let’s see now in the next 2-3 slides, what is the effect of this bias on our computations.
Now, remember, I am going to first start off with an example of compute in the angular
position by using the angular velocity measurements. So, we look at what happens to the
angular position. The angles when bias is present, let’s take us even more specific example in
a simulation, I will do a practical experiment, later on in a simulation, let’s say the bias error
or the bias is 10 deg/ sec , it means every second I have a bias of
−3
0 . 001 ℃ .
Remember, I am measuring the angular velocity. So, every second, there is this factor added.
So, this factor is added every second as basically what you have over here.

Now, let’s take a even more specific example, let’s say I have this laser pointer in my hand
and think of it as a pendulum, right which oscillates like this. Now if it goes say from this
vertical position to the horizontal one, we say that there is a 90 ° change, if it takes
1 second to go from the vertical position to the horizontal, we say that the angular velocity
is 90 ° deg/ sec , if it takes 5 seconds to go from this horizontal position to the vertical

90
position, we would say the angular velocity is ° deg / sec , right. So, that’s the rate
5
or the velocity at at at which it moves. Now let’s forget about motion, let’s do something far
simpler, let just keep it as is, it is perfectly stationary, it is not moving.

So, the measurement which you get, we would expect the measurement to be 0 ° deg/sec ,
right. It is not moving, it is perfectly stationary; however, every measurement is corrupted by
this bias value of 0.001 , what would happen if I include this bias value in my numerical
integration, let’s see, what would have happened. So, to understand that we need two
important, we need an important thing which is the amount or rather the sampling rate. This
is a gyroscope a typical MEMS gyroscope, you can sample it anywhere from 50 hertz , all
the way I imagined to around 300 hertz or something you can even do it higher, but the
accuracies are not so good.

What this means is that my gyroscope is throwing out let’s in a suppose we 100 hertz , just
for the sake of discussion, it means my gyroscope is throwing out 100 samples of data every

838
second, in 1 second, I get 100 samples of data from the gyroscope. So, if I take 0 to 1 second,
then 2 seconds over here, I will have 100 samples of data, what do we do? So, this is
basically the ω the angular velocity, what do we do with angular velocity; we integrate it.
So, we basically say that θ=θ ( k−∆ t ) +ω ∆ t , what happens when you are integrating,
remember, you are integrating the data at 100 hertz .

So, you are getting 100 samples per second, each sample is being integrated. So, let’s do this
following example, let’s suppose the initial value of the gyro which is perfectly addressed
θ , the angle θ ( 0 )=0° , this is the initial condition now. So, this is known, now I want to
compute θ at the next sample. So, in this case, k =0 , now let’s see what happens when
k =1 . So, that is θ (1) is nothing, but the previous angle which I had; the previous
angular position which I had which was θ ( 0) plus ω (1) , the current angular velocity
which is measured times ∆t whatever is the time interval, you are taking this ∆ t , in

1
our case will be equal to hertz because we are sampling at that rate.
100
θ ( 1 )=θ ( 0 ) + ω(1) ∆ t

So, we are going to look at this expression now. So, was θ ( 0 )=0° , what is ω (1) ? If
it’s a gyroscope made by god, it will show that its ω measured is exactly equal to 0, it is
not made by god. So, you always have a small bias; well let’s see what θ ( 1 ) becomes. So,
θ(1)=0.001 , this is at k =1 , what happens at k =2 , θ(2) is basically the
previous angle which I have measured plus the current gyroscope value times ∆ t . Let’s
for the moment ignore the effect of ∆t over here, we will just assume it to be 1 without
any loss of generality, you will get the same kind of results when you take the correct ∆ t .

So, we will wrongly assume it to be equal to 1, but you get exactly the same kind of
behaviour even if you take the correct delta t. So, θ ( 2 )=θ ( 1 ) which is 0.001 and again
I have a bias, the true measurement is 0 , but I have a bias, again I add 0.001 , alright
and this is nothing, but 0. 002 . Let’s do it one last time. So, θ ( 3 )=θ ( 2 ) + ω(3)∆ t add it
up again. So, you get 0.002 , the previous value plus 0. 001 and this becomes 0. 003 ,
what is happening? Every single time, I am integrating the gyroscope measurement, the
angular velocity, the actual angle computation, the angular displacement is drifting away, it
should be exactly at 0 , right because my measurement is actually 0 , there should not

839
be any angular velocity measured, but every single time we are integrating noise, the bias
value and its drifting away see 0. 001 , 0.002 , 0. 003 and so on.

You simulate this; I have done this for 16000 samples over here. So, k =1: 16000 in my
code, I will be sharing the code with you later and you do this for 16000 samples and you will
see that the drift actually goes linearly like this, it’s a perfect linear curve whereas, the true
value of the angle. So, this is the angular position in degrees, it should have been 0 for all
time and what is the computed angle, it drifts away like this. This is always going to happen
and a better quality sensor will reduce the amount of drift; maybe it will be somewhere over
here, really expensive sensor may keep it somewhere over here, otherwise, in every single
case, you are never ever going to get a perfectly computed angle with a gyroscope thing, this
is the phenomena of drift and drift here basically means you are drifting away from your true
angular position well that is true for the gyroscope.

(Refer Slide Time: 20:58)

Now, what happens for the accelerometer; it is you’ll see, it is even worse and why. Okay so,
2
let’s assume the same kind of noise profile. So, the bias is −3
10 m/s , again, we will
wrongly just assume ∆ t=1 , but that is not really important, you need to take care of it in
your actual code to make sure that ∆t is actually 1 over the sample rate( f s ) that
you should take care of that. In any case, you are going to get exactly the same behaviour as
we will show in the actual experiments, alright.

840
 2
So, let’s take the bias error for an accelerometer again as 0.001 m/s because that’s what
an accelerometer measures. Let’s again assume this the same condition, the laser pointer is
perfectly stationary. So, what would I measure 0 , right? So, we would expect this to be
0 always; however, there this bias which comes in. So, this is what you are actually going
to measure at every instant of time. So, for all t you are going to get 0.001 and we will
now see that in order to compute the angular position from the angular velocity, we had to do
a single integration for an accelerometer. To compute the translatory position, we need to a
double integration; you can already imagine; what is going to happen.

So, you are going to from your accelerometer measured value, you are going to compute a
velocity by a single integration form the computed velocity, I will call it velocity computed
from the computed velocity integrate it a second time and you get your position, okay which
means you are; if you go to the previous slide, in every single integration the value of this
bias; the 0.001 was added one time in each integration process, right. So, here we added it
once here, we added it the second time, here we added that the third time. You know, what is
going to happen here, well it’s going to add this bias twice because you are double integrating
the accelerometer.

So, if we know from the previous slide that if you plot the velocity as a function of time
which was obtained by a single integration with a bias of 0.001 , we going to get
something like this, a straight line this is the drift when the true velocity should have been
0 because even the acceleration is 0 . Now, remember that we are double integrating the
acceleration. So, it means that we take this linear drift over here integrate it one more time,
and what you are going to get is actually a Parabola single integration results in the linear
drift, a double integration results in a parabolic drift that’s what you are going to see over
here. So, this is the simulation which I’ve done in a MATLAB, I will happily share the code
with you and these are the 2 equations for one is for the velocity, one is for the position and
you see that in the first plot over here, you see in the first plot over here, right.

841
(Refer Slide Time: 24:34)

So, what you see is this curve is the velocity curve which has which you got by single
integrating the accelerometer data, this curve is actually a position which you got by once
again integrating the velocity and you can see the error between the two because of the
double integration has grown so much, right. Let’s just run the simulation for a sufficiently
long time, again all of this is synthetic data I have created this in MATLAB software, this is
not true data the real experiment will be conducted a little bit later in the next lecture.

So, let’s run this for 16000 samples, which is approximately 2 minutes and 30 seconds where
the sensor was stationary, perfectly stationary and you can now see after 2 minutes 30
seconds, my position which was supposed to be where 0 has now gone to almost
1.1 meters . So, it means the laser pointer which was supposed to be here, has gone even
farther away from that, more than 1.1 meters away, right, in just 2 minutes and 30 seconds
while it is really not moving that is the effect of drift with a double integration. So, this is
actually the problem with dead reckoning, you will always get drift in dead reckoning if you
remember the desert ant example which I gave before the poor guys is actually trying to use
his legs, the ant computes the number of movements of the legs in order to calculate how
much distance it has travelled.

Now let’s say that he is counting you know 1, 2, 3, 4, 5, 6 and it by mistake, it forgets to
count a few of the leg movements that error now starts to add up because there is an error
right a small error, it may have forgotten 2 leg movements, that gets integrated every single

842
time like and it is going to produce this kind of a drift. So, it may have walked 1 0 meters
and it thinks it has walked maybe just 5 meters or 50 meters , it does not know and it’s
true for all navigational processes. So, drift is bad it’s even worse in double integration. So,
now, before we head with techniques of how to remove this drift, let’s see in some detail what
are the characteristics of these sensors?

Let’s now let’s look at the gyroscope, first. The gyroscope is an interesting sensor, let’s say
you have a spinning ball or a gear as what you can see in your slide over here.

(Refer Slide Time: 27:37)

Now, remember the gyroscope is a 3 axis sensor, you can also have; you can also have a
single axis gyroscope, but we will be dealing only with 3 axis here. So, let’s say, you have
single-axis gyroscope which is mounted like this and you have a ball which is spinning;
spinning in a clockwise direction. Let’s fix the gyroscope on the ball. So, the ball is spinning
and gyroscope is; obviously, spins along with the ball with respect to the vertical axis. So,
this picks up the rotation rate of the spinning ball, right. So, there is a small illustration over
here.

So, these are the gears which are moving and you put a gyroscope in that and the gyroscope
measures the rotation rate with respect to its vertical axis and it will say maybe it is going at
10 ° /sec or whatever the actually velocity is; right. So, that is what you see. Now the cool
thing about the gyros is the following where is this, yes, the cool thing about the gyros is that
where if you have a sufficiently sensitive gyroscope, you can even pick up the rotation rate of

843
earth, right that is really amazing. So, you are on earth, earth is rotating all over you, I have
absolutely no, I mean I cannot sense the rotation rate of earth, I mean I am sitting everything
looks static to me, you place a sufficiently sensitive gyroscope and it can actually measure the
rotation rate of earth. The problem with the rotation rate of earth is that it is very slow, it’s
15 ° /hour , you calculate this per second, it’s a really small number and what is
15 ° /hour ? well, let’s see. So, that is 15 ° , every 3600 seconds , right.

15
So, what is the rotation rate for every second, well it is =¿ 0.004 ° /sec . So, this
3600
is the rotation rate of earth and most gyros are not sensitive enough to pick up this rotation
rate which is a good thing because every time you pick up such rotation rates you need to
now compensate for this, right. So, it acts like a bias, it is not the true motion of your vehicle,
it is not the true rotation motion of a vehicle, there is also on extra component of earth
rotation rate, luckily our MEMS gyroscopes are not so accurate and they do not really pick up
this thing and one the reasons why they do not pick up is that our noise component.

So, the actual bias of the sensors itself is usually much larger than the rotation rate of earth
and because of that your noise dominates the earth rotation signal and you can never actually
see it in our in these relatively low-cost MEMS gyroscopes, the other thing is if you and I
think you need to do this thought process yourself or thought experiment placing the gyro at
the equator pointing exactly at north will give you the rotation rate of earth place. The same
gyroscope, sorry, place, the same gyroscope exactly at the poles north pole or the south pole
you will not measure the rotation rate of earth I really want you to think of that experiment,
good.

844
(Refer Slide Time: 31:41)

So, that is what the MEMS gyroscope does, and well what are the errors which we can
encounter in a gyroscope. So, far we have seen only the constant bias error and it has a
constant offset, but there are lots of other errors. So, one which we saw was the constant bias
then we have something called the biased stability or the bias instability both mean relatively
similar things expressed differently and what this basically means is that we always assume
our bias is constant, but that is not really true, you turn on your gyroscope or MEMS
gyroscope and you calculate you keep it in the perfectly stationary condition.

So, let’s say, we are looking at one particular axis it can be any axis as a function of time in a
perfectly stationary condition. So, it may have maybe some noise like that right this is the
0 value, it may have some slight deviation from 0 over a function of time and when
you take this entire data and compute the mean of it over say about 5 minutes you will get a
particular value. So, we will call it noise 1( η1 ) you turn off the gyroscope and maybe after
couple of minutes turn it back on again and then what you are going to get is a slightly
different noise signal again it will around 0 , but it will be slightly different from the
previous one, okay.

And now when you compute the mean, it will be a different value slightly different, but it will
be different you can repeat the experiment again if you like if you are sufficiently interested
or bored in life and you compute the mean the third time, you are going to get a third, keep
repeating the experiment every single time turn it off, turn it on wait for 5 minutes collect the

845
data calculate the mean every time you are going to get a different answer which means your
bias is not really stable it is not a fixed constant bias. So, that is another problem the reason
why fixed constant bias would be amazing is because you can actually before the start of your
experiment navigation robotic navigation whatever you can just turn on the gyroscope wait
for 5 minutes calculate the mean of that is your offset right.

Now, every time in the future, if it is a same offset; what I can do is the following. So, W
measured, W meas=W true + ηbias , right, plus the noise and let’s assume only bias a fixed
constant, if I am able to model this one, if I am able to calculate this from all my previous
experiments, all I need to do is to subtract from the W measurement, my bias constant offset
and I will really get my true value this is actually a standard procedure which is called as
calibration and we do this all the time with gyroscopes or even with accelerometers, well it
would be nice except for the fact that your bias is also time-varying your bias changes with
time, it even has what is called the turn on turn off values you turn it on it changed turn it off,
it changes, you take the same gyroscope and you run an experiment for 1 hour, for the first 5
minutes, it will be one bias value may be, the next 10 minutes will be a slightly different bias
value, after another half hour maybe another slightly different bias value. So, bias is time-
varying there is also a change from turn on to the next turn on condition, then we will see
other parameters called the scale factor misalignment and white noise these are the 3 things
which we are going to see in the little bit of detail, right.

(Refer Slide Time: 35:45)

846
So, the fixed bias, we have already seen, we have also talked about bias stability. So, while
we assume the bias is perfectly constant, it can actually be changing as a function of time like
it will not go too much away from the actually value, but it could be changing, this is the bias
stability and it is a function of time, this is the assumed constant offset bias, this is not always
true. Now, what is the bias measured in typically in degrees per hour because degrees per
second is a very small number? So, typically they measure it in degrees per hour, the bias
stability is measured in degree per hour, per hour because it is a variation of your bias, yes.

(Refer Slide Time: 36:35)

Then you come to actually a very interesting noise parameter called the angle random walk
and basically, what happens is every measurement of your gyroscope as we have seen we
have the true value plus a corruption the noise( η ), the noise is a combination of your bias
¿
your fixed bias( ηbias ), let’s call it f bias, then you also have the time-varying bias(
ηtime varying bias ) and you also have white noise( ηw hite noise ) for the interested students, white
noise has some very nice statistical properties its 0 mean process for those who are
hearing this for the first time, I encourage you to just go in Google and read a little bit white
noise, you may have seen it in your television screen when the signal goes off and you get
that kind of sound, right. So, that’s basically a white noise.

So, the interesting thing with white noise is that it corrupts your measurements as we have
seen over here and the way in which it corrupts, it is a fairly complicated mathematical way
of deriving it, but basically, it corrupts it in such a way that the error grows as a function of

847
square root of time. So, you actually say degree per square root hour whereas, for the
constant bias, it was degree per square root hour, sorry, it was degree per hour, right, whereas,
for the angle, random walk is degree per square root hour.

So, the basic if you look at time and you look at the computed angle( θ ) from the measured
gyroscope, this would be the plot for the constant bias and now I would like you to think on
your own; what would be the plot if θ varies. So, in this case, θ is varying linearly
with time, right, it is varying linearly with time, I want you to think and plot and these will
also be set of assignments later on for you how would theta look if θ varies as a function
of the √ t , it is a very simple MATLAB code which you can write, okay. Then you have
something called the scale factor.

The scale factor basically says the following if the input the real angular velocity is let’s say
2° /sec in an ideal world the sensor which measures this should also say that the output of
the sensor is also 2° /sec , what happens is in reality, you have a scale factor. So, the input
is multiplied with some scaling parameter. So, and this is typically expressed as percentages.
So, if you say that you have a 5 scaling error and the input is 1° /sec the actual
measurement will show it to be 1.05 this is not the bias it is a scale factor, alright.

(Refer Slide Time: 39:59)

And then you have misalignment and cross-axis sensitivity misalignment basically is the
following we always talk about 3 axis measurements, right. So, the forward axis, the vertical
axis, and the axis which completes the right angle coordinate system. Now these axis system

848
has to be perfectly 90 ° offset from each other otherwise what is going to happen is the
measurement; let’s say you are rotating about this axis. So, the pitch angle; so, you are
actually doing the pitch rotation over here. Now if the axis on are perfectly aligned, there will
be a 0 measurement of angular velocity along this axis, right which I am holding, you will
only see a angular velocity in the other 2 axis; however, because of manufacturing tolerances
and errors and manufacturing these are not perfectly orthogonal to each other there is a slight
deviation. So, when I hold this and I rotate I do a pitch up and a pitch down even this axis
starts to measure a non-zero angular velocity.

So, to see this in a little bit more detail we would ideally want these to be perfectly
orthogonal. So, this should be perfectly at 90 ° , let’s take the 2 axis case the same story
applies for 3 axis. Now instead suppose you manufacture it in such a way that this axis was
slightly off and this axis stays well almost the same. Now when you do a rotation about this
axis alright you would expect. So, we’ll call this X , we’ll call this Y ; X' , Y' ,
this would be X and this would be Y when you rotate. Let’s actually rotate about the
other axis. So, when you rotate about say this axis you which means that you hold this
constant and you rotate about that; right, it means the angular velocity along x( ω X ) must
be equal to 0 and the angular velocity along y( ωY ) must be greater than 0 ; however,
because this axis is slightly misaligned and they are not perfectly orthogonal. So, this may be
like 89.5 ° or something you will find that ωX ≠ 0 it’s slightly larger than 0 and that is
not true, right. So, this is again a noise.

So, all these noise terms, they come together and they basically corrupt your measurement
true plus noise and noise as you can see we have seen about 5 different types of noise terms,
each one of them affects the overall computation in a slightly different way the most
dominant one is a bias that is a largest value and for most applications, it is generally enough
to compensate for the bias, we do not need to really worry about say this misalignment or
angle random walks scale factor and all these things; if you can really compensate for this
bias which is not too difficult, you can really solve many applications which do not require
very high precision, fine, a similar story holds for accelerometer except that every noise term
in accelerometer gets integrated 2 times. So, we need to be a little bit more careful.

849
(Refer Slide Time: 43:55)

Let’s see what an accelerometer actually measures, we know an accelerometer measures

acceleration right 3 axis accelerometer I am posing a question for all of you 3 axis
accelerometer it’s perfectly stationary, let’s say the I am inside a vehicle and I am and I am
holding a 3 axis accelerometer, it’s not moving the vehicle is perfectly stationary will this
measure anything apart from noise, it will always have some noise. Let’s ignore noise say the
god has made this accelerometer its absolutely no noise does it measure any non-zero
component well the picture on the right this beautiful colour picture which you see over here,
these 2 are basically the earth’s gravitational field every time you place an accelerometer at a
perfectly levelled position. So, the vertical axis is pointing directly towards the earth’s
vertical axis, what you are actually going to measure is the non-zero quantity along the
2
vertical axis and that non-zero quantity is basically 9.8 m/s your gravitational force or the
gravitational acceleration.

So, even if you are perfectly stationary place an accelerometer anywhere on earth you are
going to measure earth’s gravity, this would be fairly okay if earth’s gravity were constant.
2
So, everywhere on earth, if earth’s gravity is 9.8 m/s I know that I can compensate for
that unwanted measurement because I do not really want to measure earth’s gravity I want to
measure how fast my vehicle is moving not my vehicle motion plus earth’s gravity, I don’t
2
want to measure that. So, if it were perfectly 9.8 m/s , life would have actually be fairly
simple because, from every measurement of my accelerometer, I would simply subtract
2
9.8 m/s and then I will actually get say a much better measurement with just a noise

850
terms being available and your this one compensated the problem with that is that earth’s
gravitational field is irregular and you can see this picture.

And surprisingly even I didn’t know this it’s very fascinating to see this, India actually has
extremely well relatively speaking India has fairly low gravitational field over here. So, all
the colours in dark are where the gravitational field is much are large and the blue colours.
So, the bright red colour is where the field is very large and the blue colour is where the
gravitational field is really low and interestingly India and literally where I am sitting right
now. In Chennai, the gravitational field is really low. So, it will be less than our expected
2
9.8 m/s .

So, now, this is another problem right you need to compensate for this, but you cannot do it
directly unless you have this earth’s gravitational field data with you, right. So, really
professional navigators they always they compute the latitude, longitude, altitude of where
you are on earth alright and then they subtract out the corresponding gravitational
acceleration from that particular place, right, but otherwise basically if you ignore all of this
an accelerometer basically measures the acceleration and it will always measures the earth’s
acceleration as well, here also what we discussed before with a gyroscope the misalignment
becomes very important. So, if the axis were perfectly aligned. So, this vertical axis, let me
re-draw that bit nicely.

So, let’s say you have a vertical axis and the other 2 axis over here. So, this vertical axis
2
would measure 9.8 m/s , if it were perfectly vertical, it’s never the case because of
manufacturing defects. So, maybe it is something like this now this would be the real vertical
axis. Now because all the 3 axis are now slightly misaligned whereas, they should have
2 2
0 m/s , the other axis should also have been 0 m/ s . Now each of the axis will measure
2
a quantity greater than 0 m/s and of course, the vertical one will be the largest you know
depending on how much misalignment has happened it may be 9.75 m/s 2 or whatever, this
is the problem of misalignment and all the axis start to get corrupted by values which they
should not be measuring. So, the same thing with the accelerometers as well you have a fixed
bias you are doing a double integration and the drift grows really large and we seen that it
goes linearly with single integration and the position it grows as a quadratic term, that
parabolic shape.

851
(Refer Slide Time: 49:04)

The bias stability is similar to the angle random walk the behaviour is very simple and the
3/2
drift when you do the single integration it grows as a power of t , you do a double
5/2
integration it grows as t whereas the other one was the linear and quadratic this grows
at these rates. We will not really go into these 2 details except to say that the bandwidth of a
sensor and not just accelerometer even for gyroscope is extremely important if you have let’s
take my arm let’s say the arm is moving at the rate of one at a rate of 1 hertz right it means
I am able to do this motion in a 1 hertz cycle which means it takes me 1 second to do this
motion 1 second, 2 seconds, 3 seconds and so on.

If this is the motion which I always expect there is absolutely no need for me to use a very
high bandwidth gyroscope or accelerometer right there is no need for me to use like a
1 000hertz bandwidth accelerometer because I don’t do anything with all the remaining
frequencies it really just the 1 hertz component which I am interested in if I am a missile
and I am moving extremely fast, right, I do not know how fast these missile actually change
their parts, but you could really be looking at fairly high frequency of say 50 hertz or
something like that if you have these what beam these cantilever beams which are attached to
walls and you tap on them and you have these vibratory motion these can go really very high
frequencies and for that, you would use a different accelerometer as well, right.

852
(Refer Slide Time: 51:15)

So, the bandwidth is an important property finally, while I will not spend too much time on
these, it is important for everyone to realize this, the input range basically what I have just
been talking to you the input range is very important for high-speed dynamic motion, you
would require a high bandwidth gyroscope or accelerometer for a low speed or low dynamics
motion, you would require a low bandwidth accelerometer or a auto gyroscope and the range
actually determines that accuracy, of course, is important the more accurate your sensor is the
more expensive it is, but the more better it is resolution and sensitivity again are important.

Sensitivity basically says it is defined as a ratio of the change in the input to the actual
measured value, right, sorry, is defined as the change in the input to the change in the output
signal. So, if you take let’s say your sensor over here, let’s say, we give it an acceleration or
we give it angular velocity for a ∆ change in the acceleration means for a very small
change in the acceleration, if there is a corresponding change in the voltage or in the response
over here that is very good it means that we can measure really small acceleration.

However, if it takes a large input to have some amount of data coming in over here then it
means our accelerometer or gyroscope are not sensitive enough alright that is a basic meaning
we will ignore this non-linearity you can look up datasheets we expect behaviours of sensors
to be fairly linear around certain operating regions in real life as we have seen in the previous
lectures they can be slightly non-linear not as much as I have drawn, but slightly non-linear.
So, all these effects start to become important when you really start to worry about accuracy,

853
if you are comfortable that your application should have reasonable accuracy is what you
should define for most home automation or even most robotics application.

I mean a MEM sensor is good enough and even more specifically you do not need to really
compute angle random walk you do not need to worry about misalignment and so on and. So,
forth unless you want your robot which travels for say 2 kilometres to know exactly
where it is with very high degree of precision if you are not too worried about that focus on
the constant bias remove that through a calibration process which we will be showing you in
the experiments and we are pretty much done.

(Refer Slide Time: 54:11)

So, this is accelerometer datasheet from analog electronics and there is quite some interesting
information which they provide in the datasheet. So, you can see that the bias values fairly
high it is about 150 milligrams ; 150 mg ; the g basically
2
1 g=9.8 m/s . This is actually the acceleration, I am experiencing right now without any
motion, if I add an extra motion its 9.8 added to that extra acceleration which my body
2
will be providing at stationary positions, it is 9.8 m/s and as you can; that the bias is
150 mg that is fairly large which is why it is important to compensate for that; yes. So, we
are almost towards the end of this lecture and what we have seen is that MEM sensor are
fairly cheap for a few 100 rupees; you can buy accelerometers and gyroscope. In fact, I
encourage you to go ahead and buy these sensors.

854
(Refer Slide Time: 55:09)

Because two of these sensors cost the same as two ridiculously overpriced cappuccinos or I
do not know what you guys have in these fancy coffee shops, right. So, a complete IMU with
a accelerometer, gyroscope, and this something called a magnetometer is like a compass
complete IMU costs less than 2 cappuccinos, well, there are various techniques which we will
be seeing in the next lecture, on how to compensate for these noise problems. Finally, if you
are really desperate and you want incredibly good accuracy maybe, you are an engineer
working in aircrafts or in missile developments or even in autonomous land vehicles or air
vehicles where precision, accuracy is extraordinarily important for whatever reason that you
have you would really try to go in for optical sensors. They are of course, very expensive,
you may need to sell your house or car to buy one of these sensors, but that’s how it goes the
more expensive the better the accuracy.

855
(Refer Slide Time: 56:29)

The final slide is applications of inertial sensors, I will not spend too much time on this
except to point out that as the bias stability drops it is as the bias stability keeps reducing as
the scale factor stability keeps reducing meaning that the scale factor is less and less of a
problem, you can do far more complex navigation problems you can solve far more complex
navigation problems. So, you can look at missiles and you can look at somebody navigations
and so on if you have poor bias stability and poor scale factor stability you are only looking at
the consumer electronics like the accelerometer which are present in your cell phones same
with the gyroscope the poor bias stabilities will mean you are looking at consumer-type the
really low noise profile will be your submarine navigation missile navigation and so on and
these two graphs are taken from this excellent reference which I showed over here.

So, the point of this entire discussion was that we tried to look at accelerometer and
gyroscope say noise characterization where there are plenty of noise parameters which must
be considered for really serious navigations for most consumer type of navigation it is enough
to focus primarily on the bias, the constant bias and where these noise characteristics affect
your navigation is in the phenomena of drift solve drift reduce drift as much as you can focus
on what is the focus on constant bias for all consumer-type applications not for medical not
for military and all and for say medical and other kind of say military kind of applications
you would need to look at all the different noise profiles this becomes very important.

856
So, thank you very much. In the next set of slides, we will actually move on to filtering to
see, how we can reduce navigation errors specifically we will be looking at example of
complementary filter we will conclude that with a set of experiments.

Thank you very much.

857
 Control Engineering
Dr. Viswanath Talasila
Department of Telecommunication Engineering
Ramaiah Institute of Technology, Bengaluru

Module – 11
Lecture – 02
Part – 02
Filter Design to attenuate inertial sensor noise

Hello, and welcome to the next series of lectures in control engineering. So, we are going to
be looking at module-11: Lecture 2-part 2, and it is a continuation of the study we of the
discussion we had on the inertial sensors.

(Refer Slide Time: 00:40)

So, if you recall in the last lecture, I had discussed the various effects of a sensor noise
specifically: accelerometer sensor noise, and gyroscope sensor noise on deposited on the
position computation. And whenever I say position it could mean either translation or it could
mean the rotation position, the angle rotation position right alright.

So, we saw a couple of things one was that there are various types of noise, there is a constant
bias, the drifting bias angle random walks so on, and so forth. For the kind of applications,
we typically deal with, in consumer electronics and even sometimes to basic medical
applications, it is enough for us to deal with the constant bias. So, if you are able to estimate

858
the constant bias and then remove that, you will get actually a very good position estimation
ok.

So, the way that we are going to do it in this lecture, which will be immediately followed by a
discussion on the actual implementation of this in both MATLAB and on a microcontroller
alright. So, that will be done by my research associate Vinay Sridhar. We are going to start
with the discussion of a high pass filter, and we will see that we pass the gyroscope data
which is the angular velocity through the high pass filter. We will then go to a low pass filter
and basically, we will be passing the accelerometer data through the low pass filter alright.
And we will then see how to combine the high pass filtered angle output with a low pass filter
angle output, in order to get a much more accurate position computation and again when I say
position it could be either that x, y, z the transition position or the angular position the roll,
pitch, and yaw.

So, let’s go ahead and one last thing you should be very careful to remember this, we are
going to be looking at computing only two angular positions today. And the first will be the
pitch, the second will be the roll. We will not be computing the yaw. There is a specific
reason for this, which and it is basically that the accelerometer is a sensor which for various
reasons, which I’ll not discuss here; it is unable to compute the yaw angle which is the
rotation about the vertical axis right it is able to compute only the pitch or the roll, it cannot
compute the yaw. So, and we know that we can compute the yaw from the gyroscope because
it can measure the angular rotation, angular velocity about the vertical axis, but we know that
we cannot blindly use the gyroscope output because of drift right. So, whenever you integrate
the gyroscope output we saw that before right.

So, if you take the vertical axis as a z-axis, and then you integrate the angular velocity which
you get about the z-axis, you know that you are integrating two quantities. The true value
about the z-axis plus some noise, and you know that this leads to drift. And what we are
trying to claim in this lecture is if you combine the gyroscope with an accelerometer by using
these filters, you can actually reduce drift. But because we cannot compute the yaw angle
using the accelerometer, we will not compute the yaw angle directly with the gyroscope.
There is actually another way of doing it which is you combine the gyroscope with a
magnetometer and a magnetometer basically gives you the yaw angle or the heading angle.

859
So, this can again be done in the same complementary filter technique that I am going to
explain today, except that instead of the accelerometer we use a magnetometer alright. And
there will be basically to combine the gyroscope and magnetometer in order to get the yaw
angle. So, yaw is obtained from these guys, for today's lecture we are going to be focusing
only on combining the gyroscope and the accelerometer data in order to compute the pitch
and the roll ok let’s see how we are going to do that.

(Refer Slide Time: 05:48)

So, before we go ahead what we need to do is to first look at what is a basic first order high
pass filter and then we look at a low pass filter and then a complimentary filter. Now, you
may be surprised that we are trying to do a real application over here and we are still just
using first order filters, well it turns out that in most practical applications in the industry as
far as possible people try to prefer to use only first order, and maybe if necessary second
order filters typically we don’t like to go much higher than that at least in the area of
navigation ok.

So, what is the first order high pass filter? A physical representation of that could be
something as simple as this. So, you have a RC filter basically. So, the capacitor followed
by a resistor R . So, if I then measure the voltage across this over here that will be, I’ll just
call it as V out and we apply a voltage at the input terminals we call it V ¿ . Now this
functions like a high pass filter and you already know that a high pass filter is nothing but
what is called as a lead compensator. If you don’t know this that is not an issue, but for those

860
of you who are going to go into design of control systems the term lead compensator
essentially means a high pass filter and you will see later that a lag compensator is a low pass
filter anyway.

So, this is the physical realization of the high pass filter. So, the first thing we will try to do is
to derive the transfer function of this filter. So, transfer function of the high pass filter, some
of you would have already seen this with the basic derivation of electric circuits and transfer
functions. So, the way that we are going to do is as follows. So, let’s look at V out , the
output voltage across the resistor terminals it is nothing but, let’s say a current I is
flowing clockwise in the circuit it is IR ok.

So, let’s keep this aside alright. So, this IR . Now we know the basic; one of the basic
constitutive relationships between the charge and the voltage drop across the capacitor, which
is basically Q=CV . Now, if we take a look at our capacitor over here, the voltage drops
across the capacitor. So, let me just write this down. So, the voltage drop across my capacitor
C is basically V ¿ −V out ok.

So, this then implies that Q is nothing but, C( V ¿−V out ) , this I’ll call it a step number
2. Now, we go to step number 3, where we know the basic relationship between the charge in

dQ
the capacitor and the current flowing through the capacitor which is that I= and then
dt
we substitute this expression for Q over here back into this derivative term and you

d d
basically get C (V ¿ −V out ) . So, this can be written as C (V −V out ) . So, that is the
dt dt ¿
expression for the current through the capacitor.

Now, substitute into 1 what are we going to substitute, the expression for the current into 1.
So, substitute 3 into 1, this would give V out which is nothing, but

d d V ¿ d V out
V out =C (V −V out ) R
dt ¿
alright. So, this is RC ( dt
−
dt ) that’s basically the

expression that you are going to get. So, if I now write all of this in a Laplace transform
domain. So, transform; I’ll call this as this entire equation as equation number 4. So,
transform 4 into a Laplace domain and what do we get? Well, we get V out (s) , I know that

V out ( s )=RCs V ¿ ( s)−RCs V out (s)

861
 d
on the right hand side you would get RCsV ¿ (s ) that is because the term in the
dt
Laplace domain it is s alright, −RCs V out (s) this is the expression we get.

So, take this term on to the left hand side and take the V out (s) term common, and you will
basically get V out (s)(1+ RCs) on the left hand side is equal to RCsV ¿ (s ) . Now we
know the definition of the transfer function it is a ratio of the output response to the input

V out (s )
signal. So, we get G ( s )= alright and that will be nothing, but the transfer function
V ¿ (s)
of a first order high pass filter. So, this is the cool little guy we have been trying to derive all
along ok. V out (s ) RCs
G ( s )= =
V ¿ (s) 1+RCs

So, it’s a transfer function of a first-order high pass filter now for those of you who are
already comfortable in MATLAB, what you can go ahead and do is to well let me erase make
some space over here. So, that I can write. Okay, that’s enough let me not spoil this anymore.
So, for those of you who are comfortable with MATLAB by now what you can do is to take
different values of R and C and you can use the command tf in MATLAB in order
to define your system the high pass filter.

So, we can actually call it system or you can call it a filter if you like that is a better word. So,
we can call it hpf filter is equal to transfer function of the numerator comma denominator
[i.e., hpf filter=tf (num, den) ]. In this and the way you would write it down in MATLAB
is basically numerator is. So, you have a s term over here. So, it would be RC and
there is no constant term. So, that is a 0 that is a vector right and the denominator term
would be you have a constant term 1 and you have RC over here ok.

So, you can define a transfer function in MATLAB for different values of R and C and
then you can go ahead and try to look at you know the bode the, bode plot of your high pass
filter. So, the bode power plot hpf of the high pass filter. So, this is a command in
MATLAB [i.e., bode (hpf filter ) ]. So, when you actually plot the bode plot of the high
pass filter let me erase make some more space over here, right you are actually going to get

862
you should if you do it carefully, you are going to get a bode magnitude plot we will not
worry about the phase over here too much, where this is in a log scale the frequency on the x-
axis and then you have the magnitude of G( s) in decibels. And this being a high pass
filter you would actually expect to see a magnitude plot of this type alright where this is the
cutoff frequency f c .

All of you know how we can calculate the cutoff frequency. So, we have the high pass
frequency response plot over here and you see the cutoff frequency is shown over here
actually it should be slightly over here around the 3 dB point. So, let me correct that.
Okay, let me draw this again; that is a better plot. So, this is a cutoff frequency, and the cutoff

1
frequency as all of you know you can compute with the expression . So, given the
2 πRC
values of R and C you know now what is the specific cutoff frequency which you can
actually use alright.

So, what have we done in this slide we have derived the transfer function of a high pass filter
we have seen what are the command we need to use to plot the bode plot to first of all define
the transfer function in MATLAB, then plot the bode the frequency response in MATLAB
and you will actually get something like this the magnitude part you can see a similar one for
the phase. Now with this you need to remember that this is a continuous time filter right of
course, what we would like to actually do is to use this in discrete time on a computer or in a
microcontroller.

So, we will actually need to go for discrete time version of this continuous time filter, that is
what we’ll actually need to do and this is what I am going to do in the next slide. So, we have
the same high pass filter. So, we had the capacitor and the resistor right and for this we got
the basic expression for the output across the capacitor, as a function of the other variables.
Now if you look at each of these variables let’s just plot the input V ¿ as a function of time
and also the output V out as a function of time. So, let’s say V¿ is some signal like that
and ok.

So, we have the output filtered version something like this, now this is a continuous time
what you actually do is to pass this through A/D converter, and you take it into your
microcontroller or into your computer or something like that, and when you do that you
actually only have just discrete samples right. So, you may have this, this, this, so on

863
depending on how fast your sampling rate is right, and the same for the output as well. So,
you will not have the continuous signal, you will just have discrete parts of that and so on for
all the other points as well. So, how would we represent that well one way to do it is to
actually use the notation over here? So, the input signal V¿ you can see here the input
signal V¿ right. So, it is a collection of points. So, I’ll call the first point as x 1 , I can
call the second point as x2 , x 3 so on and so forthright.

So, that is what I have done over here. So, this n which you have will depend on the
sampling rate and for how much time you have actually collected the data. So, it can be a
really large number if you collect data for a few minutes. So, this will be essentially the
discrete version of the continuous time signal. So, these collection of points is a discrete
version of this continuous time signal. In the same way, we measure the output we again
collect all the discrete samples over here, and we call it as y (1) , y (2) , so on till
y (n) . Now when you do this we need to now basically plug these two expressions back

d
into this equation, and we know that in the difference time would basically be
dt
converted to a difference equation and you have seen this difference equation come about in
when we computed the angular position if we had a measurement as angular velocity right or
we computed the basic velocity when we had the acceleration measurement with us. So, we
did a single integration in both cases.

So, that is basically your difference equation. So, how would this look if I, actually write

dV¿
down the difference equation? So, the . So, remember that V¿ was nothing, but
dt
x (1) , x (2) , so on till x (n) depending on how much time you have collected your

dV¿
data. So, would be nothing, but x (i) , the current sample minus the previous
dt
sample: x (i−1) , divided by the amount of time, ∆t . Difference between these two
samples and the time difference
d V ¿ x (iwill
) −xbe essentially your sampling rate which we are looking
(i−1)
=
at ok. dt ∆t

864
 dV¿
So, that is how we write and we know that we had written V out as y (1) ,
dt

d V out
y (2) , or till y (n) and in exactly the same way would be nothing, but
dt
y (i) , the current sample which we are measuring minus the previous sample: x (i−1) ,
divided by the time difference between the two samples: ∆ t and that is basically what we
have over here. Now if I take this expressions, remember we are multiplying this with RC
as we see here and you do very simple algebraic manipulations and the derivation is given
over here you will end up with this final expression. So, what does the final expression look
like let me just write this down again. So, the output at the current time is a function of some
constants.

So, this is actually a time constant let’s call as RC , a function of the of these constants
times the current input minus the previous input plus again we have these constants over here
and the previous output, which we measured last time alright. So, what we do is to replace
this node with a different notation. So, we will typically call RC as τ , the time

RC
constant and we will call because this is a high pass filter. So, we will have
RC + ∆ t

τ
which in our case is nothing, but and this entire term we are going to denote by
τ +∆ t
α HPF ok.

So, it basically means the current output after the filtering right, the current output is
y (i ) =α HPF [ x ( i )−x (i−1) ] +α HPF y (i−1) . This is the expression that we are going to be
using in the discrete time implementation and we are actually going to show an example of
this in MATLAB as well as in the microcontroller code. So, this is what we are going to do
with the high pass filter, we have seen how to compute the frequency response we have seen
how to compute the cutoff frequency by ok.

So, if I go back we have seen that the cutoff frequency let’s use a different color. So, the

1
cutoff frequency was over here . Now we know that τ is basically RC right.
2 πRC
Which is nothing, but the time constant. And we will see in the next lecture how to
specifically to choose this value of τ based on the design constraints that we have to

865
consider which may depend on how fast your motion is, what kind of filtering you want to
use and so on. So, this will be done in the next lecture by Vinay Sridhar alright very good. So,
that is a first order high pass filter and just to recall the gyroscope angular measurement data
after computing the angle is going to be passed through this high pass filter, we will see that.

(Refer Slide Time: 25:00)

Now, let’s look at a low pass filter. So, how would a physical realization of a low pass filter
look like well you would have a resistor followed by the capacitor over there, and then we of
course, would write V out a signal over here this is your R and this is your C . Now
why don’t we apply KVL and see what we are going to get. So, let’s say you have a current
over here and if I apply KVL you would actually get −V ¿ because I am going in a
clockwise direction plus IR plus the voltage drop across the capacitor which is the same
as V out is equal to 0 this is by KVL.

−V ¿ + IR+V out =0

Now, the current in this loop is the same as a current which is passing through the capacitor
and what is the expression of the current which is passing through the capacitor,

d V out
I =C right that is the expression number 2 and we substitute now 2 in 1. So, this
dt

d V out
gives us −V ¿ +CR +V out =0 . Transform this to the Laplace domain and remember
dt

866
 d
that in the Laplace domain is basically s . So, when we transform this to the
dt
Laplace domain you are going to get −V ¿ ( s )+ CRs V out ( s ) +V out (s)=0 and now all you
need to do is to simplify this expression.

So, we take V ¿ (s ) on the right hand side and take the V out (s) term common over here.
So, this would imply V out (s) into I am sorry they should be C over here, because

d V out
I =C . So, in this expression when I replaced I had forgotten to include the C
dt
anyway. So, the C comes over here, you will also have a C over here alright good. So,
now, you take V out (s) common and you would basically have (1+ RCs) , On the left
hand side. On the right hand side, you will just have V ¿ (s ) this, of course, gives us the

1
transfer function which is the output response to the input signal, which is , this is
1+ RCs
the transfer function of the low pass filter. And as before if you try to compute the frequency
response of this and you look at the magnitude plot, you will get a response like this. So, this
is ω in a log scale this is the magnitude in decibels and this would be your cutoff

1
frequency, once again it is equal to very similar to the high pass filter derivation
2 πRC
so, good.

So, that is the low pass filter and now let’s. So, as we discussed before all of this is a
continuous time implementation, and what we will actually need is to have. So, we need a
discrete time version of this filter, because we want to put all of this into a microcontroller or
a computer.

867
(Refer Slide Time: 29:14)

So, how are you going, we going to get it discrete time version? Okay so, discrete time low
pass filter and to be very specific this is a first order low pass filter. So, we are going to use
this basic expression of the continuous time filter to compute our discrete time filter ok.

So, let’s recall that, −V ¿ ( s )+ CRsV out ( s ) +V out (s)=0 . So, we change the colors. So, we
had; was this KVL what we used? yes indeed. So, this was. So, we had the expression from
the Kirchhoff’s voltage law and it was derived as −V ¿ ( s ) plus RC let’s do this again. So,
with this expression from the Kirchhoff’s voltage law which was

d V out
−V ¿ (t)+ R C +V out =0 . Let’s try to write down the discrete time equivalent of this
dt
one in a similar way as we did before. So, let me choose blue. So, we have V¿ described
by a sequence of points x1 , x 2 , so on till x n right.

Similarly, we had V out described by a sequence of points y1 , y 2 , so on till yn

alright. Why don’t we substitute this in this and we will see what this looks like,
remembering that the derivative of the continuous time will become a difference equation in
the discrete time? So, let me choose a nicer color now let’s choose funny green. So, what will
happen first let me take V ¿ onto the right hand side, you will only be left with these terms

d
on the left hand side. So, RC obviously, constants and the now becomes a
dt

868

y (i ) − y (i−1)
difference equation, which will basically be let’s not say V out let’s use the term y . So,
we will actually say y (i) which is the current sample, that is y (i−1) the previous
sample divided by ∆ t , plus you have V out over here. So, it will just be again y (i)
the current sample and we had taken V ¿ (t ) on to the right hand side. So, the V ¿ (t ) will
simply be x (i) .

I hope you have noticed I made a small mistake over here which alright. So, this is. So, this
was i and this was i. So, let’s see how we actually compute this relationship. So, we multiply
everywhere with ∆t and we will get RC y ( i )−RC ( i−1 )+∆ ty ( i )=∆ tx (i) . So, from this
not that one let’s take the other one. So, from this term and this term, let me take y (i )
common out. So, I get y (i ) ( RC+ ∆ t) , and let me take this term over here on to the right
hand side. So, this will be nothing, but ∆ tx ( i ) + RCy(i−1) .
y (i ) ( RC+∆ t )=∆ tx ( i ) + RCy (i−1)

Now you can see that I can divide with RC +∆ t , will go on to the right hand side and then
eventually I’ll have my expression for y (i ) . So, this will be nothing, but

∆t RC
y (i ) = x ( i )+ y (i−1) . Now as we did before let me simplify this
RC + ∆ t RC +∆ t

∆t ∆t
expression by denoting α LPF = alright. So, α LPF = , I’ll also call this as
RC +∆ t RC +∆ t

∆t RC
the α LPF . In that case 1−α LPF=1− which is the same as ok.
RC + ∆ t RC + ∆ t

So, let me use this expression and this expression in this equation and we get the final
discrete time realization of a first order low pass filters. So, that’s
y (i ) =α LPF x ( i ) +(1−α LPF ) y ( i−1) . So, we started off with a continuous time version of the
filter over here and through these steps we have now arrived at the discrete time version of
the filter and again RC is nothing, but a time constant ∆t is the time difference
between two between two samples. So, depending on what data you have you will
appropriately choose the RC and we will see this in an example by Vinay Sridhar in the
next lecture very good.

869
So, now that we have a high pass filter and we have a low pass filter, what is the idea of this?
why are we doing all this? Well the idea of this is that we want to create something called a
complementary filter, and the complementary filter takes the best of both worlds it takes the
best that the accelerometer data can give us and it takes the best that the gyroscope data can
give us ok let’s see what I mean by that is as follows.

(Refer Slide Time: 36:31)

Using the accelerometer data and we will go more into detail in the next lecture by Vinay
Sridhar, we can actually compute two angles we can compute the roll and we can compute
the pitch. Using the gyroscope data and after we numerically integrate the angular rotation
which we get over here. So, there is a numerical integration, we can also compute the roll and
the pitch. So, this is from the accel and this is from the gyro ok.

Now, we have seen a specific problem with the gyroscope, that because of numerical
integration I get the problem of drift. So, if I just use the gyroscope alone in order to compute
these angles, eventually if my real angle is supposed to be some 30 ° , my actual angle may
just keep drifting away forever and you can have an unbounded error which is not good. So,
why don’t we use just an accelerometer; the problem with an accelerometer is that it is
susceptible or you can say corrupted, it is susceptible to very small changes which is basically
I am saying high frequency components, it is susceptible to very small changes in the
environment and it picks up everything it is it is a very sensitive sensor. So, if I place an

870
accelerometer on my hand right just suppose this pin is an accelerometer, and I am doing
these motions. So, the accelerometer will pick it up I do want to track this.

Now, while the pen is on my hand even if I just tap very lightly very very lightly, the
accelerometer is sensitive enough to pick up those kind of vibrations as well, and those start
to come into the equations and you are going to get erroneous answers. Furthermore,
accelerometer data as explained just now is susceptible to high frequency noise. So, if you
keep an accelerometer at a perfectly stable condition and we would expect the data to be
something like this, along one axis of the accelerometer because of this high frequency noise
and vibrations in the atmosphere all around you actually would be getting data, like this is
high frequency data and that is not good. It does not drift, but it does not give you good
accuracy. What you can do to remove this high frequency drift this high frequency noise is to
actually pass all of this through a low pass filter right and then you can actually get a much
cleaner response, and what we will do with the gyroscope is to pass it through a high pass
filter primarily for the reason that a gyroscope is very accurate in computing the high speed
dynamics it is not so good when your dynamics are really low speed.

So, and then when you combine the gyroscope with the accelerometer, the accelerometer
ensures that the gyroscope does not keep drifting away because it tries to bring it back
because accelerometer does not drift. So, the basic construction which Vinay will be talking
in more detail in the next lecture is that we have the accelerometer data in all three axes. So,
that is ax , ay and az all of this going into let’s say an angle computation block we
will see exactly how to do that, and this will be then passed through a low pass filter. Then
we have the gyroscope data, which will again you will take ωx , ωy , ωz the angular
velocities do a numerical integration and pass it to an angle come well and that actually gives
you the angles directly there is no angle computation block as such.

So, this gives you the angles directly after the numerical integration with drift and you pass it
through a high pass filter. And what you then do is to combine both the outputs in order to get
a complementary filtered output, and how does this look like? It is a little interesting. So, let
me erase this stuff over here. So, we know what the frequency response of an accelerometer
would of a low pass filter looks like right. So, if the low pass if the frequency response of the
low pass filter looks like this where this being the cutoff frequency f c , the frequency

871
response of a high pass filter would actually be like this and let’s assume that we are clever
enough to choose exactly the same cutoff frequency ok.

So, this is we are looking at the magnitude plot, what this tries to behave is like an all pass
filter. So, if you pass in exactly the same signal through this filter we will just call some
signal as q , you pass q through a low pass filter you pass the same to a high pass filter
both with the same cutoff frequencies, look and then you add them up the combination of this
behaves like an all pass filter. We won’t go more into details of this except to point out at this
stage, that we are looking at filters with gains or magnitudes of 1. So, that we don’t amplify
or attenuate these signals ok.

So, given that we will see in the next lecture by Vinay Sridhar, how the use of this
complementary filter the combination of a low pass and high pass, how the use of this will
actually lead to much more accurate say computation of the angles, and we will also see that
if you have really high speed motion I want to trust my gyroscopes more. If I have low speed
motion I want to trust my accelerometer more right. So, if you look at it in the frequency
response context if you have high frequency I want this part of the frequency response curve
to be activated right because if I have high speed dynamics, I really want this to be activated.
If you have low speed dynamics meaning my bandwidth of the system is low I don’t really
want to use this region, it is pointless it starts to bring in noise and other problems.

So, in that case I want to activate this part of the frequency response, how we would do that?
While we would do that by writing the complimentary filter by using a special parameter
called α CPF and you multiply this with the angle which is coming in from the

accelerometer ( ie. , α CPF θ acc ), and you multiply exactly ( 1−α CPF ) with the angle
coming in from the gyroscope ( i. e .,( 1−α CPF )θ gyro ) alright and you would. So, again it is
CPF. So, you would choose your α CPF to be such that if my dynamics are really fast
dynamics I would like to trust this one more. So, for fast dynamics I would like to trust the
gyroscope and how would I tell the equation to trust the gyroscope, well you simply put
α CPF to be fairly low almost close to 0 for really fast dynamics. For slow dynamics, I
like to trust my accelerometer more.

So, you let α CPF almost get close to 1 yeah of course, in reality, you would never use
0 and 1 because you remember the problem with drift that the gyroscope has. So, you

872
would always want to make sure your accelerometer angles are always coming in to prevent
that drift. So, it will not be really 0 ’s or 1 ’s, it will be small numbers greater than 0
large numbers, but less than 1 and Vinay Sridhar, in the next lecture will explain this in
more detail with the couple of nice experiments as well ok.

So, to conclude this talk we had started out by recalling some of the effects of sensor noise on
position computation translation and rotation position computation, we looked at two specific
implementations the discrete time implementation of the high pass filter and the low pass
filter, and we saw very briefly we will go more into detail into the next lecture, how the use
of a complimentary filter can give us far more accurate results. What I want you to remember
is that the MATLAB code and the microcontroller code that we have developed we will be
actually sharing it with all of you on the forum. So, you are welcome to experiment with the
codes that we are going to give you, the data that we are going to give you and you can
actually design your own navigation filter ok.

So, we saw in the previous 2-3 slides, we basically wrote down what was that the discrete
time realization or implementation of the high pass and the low pass filters right and the low
pass filters.

(Refer Slide Time: 47:04)

Now, let’s see once again why we would like to pass the gyroscope measurements through a
high pass filter and accelerometer measurements through a low pass filter, and then we will
see the role of the complementary filter ok.

873
So, first let’s take the accelerometer. If you actually look at the accelerometer data which we
will also be providing to you on the forum, even when you keep the accelerometer in a
perfectly stationary condition and we expect to see let’s take along any particular axis, we can
take along the x axis as a function of time. So, if it is stationary you would expect to see
around the 0 value, may be some minor variation like this right.

So, there is almost no variation and that is actually what we expect to see when the
accelerometer is perfectly stationary. So, this is basically the noise primarily because of bias
right. So, instead of being at 0 , it gets offset to a certain value, which we can remove by
calibration this is what we would expect to see. Now when you mount this on any stationary
surface and let’s say you are walking around that surface or you tap on your table or you
shake your table very gently, any motion can be easily picked up by a accelerometer because
it is very sensitive. So, instead of seeing the slight variation which is basically an offset that
the bias offset with some and the noise parameters, the accelerometer is capable of picking up
even minor vibrations in the environment and you may actually see the stationary
accelerometer data to be something like this right.

So, this is the kind of actual measurements which you get from the accelerometer and this
well it has many different terminologies one of it we can say that it is called as jitter. So, these
are the high frequency components which any minor vibrations in the environment the high
frequency vibrations are easily picked up by the accelerometer and then and then measured
and then you see them over there. Now because of the nature of these vibrations which are
typically high frequency if you pass this through a low pass filter and if you do a really good
job at that we would actually expect to see back of a nice little accelerometer data, but the
high frequency components are removed, of course, this is in an ideal condition it is never
this perfect, you may actually get something like this.

But that is still far better than what you are seeing in the previous plot. So, that is for the
accelerometer and that is a reason why we pass the accelerometer data through a low pass
filter. Let’s look at the gyroscope data. So, if you keep your gyroscope in a perfectly
stationary condition, again you would expect remember this is angular rotation over there, it
is this time you would again expect to see some fairly mild signal like that and it is not
exactly at 0 because you have offset or what is also called as a bias. Now unlike in the
accelerometer, the offset is present in all the axis.

874
So, when you compute the angles through the accelerometer, we would use trigonometric

ax
expressions of the type tan −1
(√ 2
a y +a z
2
) , and what actually happens then is you would

have a small bias measurement in each one of them alright, but because you are dividing one
by us with the other bias terms. So, this is the total bias contribution is quite small. So,
although this is not the true angle it is still not too bad you are slightly erroneous, but it is not
as if you are going to drift away completely from the true angle. In the case of the gyro as we
have already seen before this is not true, and the reason for that is basically because we are
integrating the signal and when you integrate the signal you also integrate this bias over there.
So, what you would actually get is if the true angle is supposed to be somewhere like this.

Let’s say 0 ° , I’ll call this 0 ° , your bias is actually going the integrating the angular
rotation rate is going to produce a bias of that nature alright because it is a fixed bias you
have a linear drift over there. So, how to solve this problem well what you can do is to simply
pass this through a high pass filter and what happens when you pass this signal through a
high-pass filter, it would try to pass all of the gyroscope data except the bias which is a
constant offset right. And a constant dc signal through a high pass filter is simply attenuated
off right and when you do that. So, then you would not have too much of a drift your drift
will be significantly reduced of course, depending on how will you choose your high pass
filter coefficients.

So, accelerometer data is passed through a low pass filter and the reason for that is to remove
jitter or the high frequency noise. The gyroscope data is passed through a high pass filter of
course, after the integration to remove drift that is a basic idea of what we do with the high
pass and the low pass filter. What do we do with the complementary filter, while we do
something more interesting. So, now, let’s look at the complementary filter let’s say that we
have high speed motion high speed dynamics say you are in a plane and then it is going at
extreme speeds your rotation speeds or low speed motion right.

So, the aircraft is moving very gently, now because of the nature of the two sensors which we
have, it turns out the gyroscope can actually capture the high speed motion very well. In fact,
typical gyroscopes have fairly high bandwidth like what and it is in the order of a few
thousand degrees per second whereas, accelerometers have very low bandwidth and it is

875
meant to capture really the low speed dynamics. So, we would like to capture the low speed
dynamics with the accelerometer and the high speed dynamics with the gyroscope.

So, how would you do this? So, we know that from the gyroscope we get an angle after doing
the high pass filtering. So, I’ll just call it as theta high pass filtered angle, θ H PF and this
has come. Out of the gyroscope out of the accelerometer, we compute the angle and then we
pass it through the low pass filter. So, we get another angle over here I have called θ LPF .
Now how do we combine these two angles in such a way that we can capture both high speed
dynamics, we can capture low speed dynamics or we can capture any dynamics in between
because there is no guarantee your dynamics is always high speed or always low speed. And
the way you would do that is by the use of the complementary filter, and it is a very simple
equation.

So, the output coming out of the complimentary filter the angular position is nothing, but a
coefficient which we will see how to calculate in later lectures by Vinay Sridhar, the
coefficient
θCPFof
=αthe
CPF complementary
θ L PF +(1−α CPF ) θfilter
HPF this is multiplied with the angle coming out of the

accelerometer plus (1−α CPF ) times the angle coming out of the high pass filter which is
the gyroscope.

How does this equation solve our problems? Well let’s say that we have really high speed
dynamics, I would like to trust my gyroscopes more because it can actually it has very high
bandwidth it can capture these high speed dynamics far better. If I want to trust my
gyroscopes more then I would like to actually use this far more than using the angle from the
low pass filter from the accelerometer how would I do that well I’ll simply set the coefficient
to be small; not 0, but small maybe ok.

If I have predominantly low speed dynamics, then I would like to trust the angle coming out
of the accelerometer more than the angle coming out of the gyro. Why is that we know that
gyro integration is affected by bias, even though you have done a high pass filter of the gyro
there is always a little bit of residual bias left and if the dynamics are really towards the low
speed dynamics we are forced to choose the cutoff frequency of the high pass filter actually
very close to the left axis. So, let’s draw that over here, if you have very high speed dynamics

876
a 100 Hz and above or 20 Hz and above, I can choose my high pass filter to have a
cutoff frequency. let’s say more than 20 Hz ,

But if I very low speed dynamics and I want to pass it through my gyroscope, I’ll need a cut
off frequency much closer like this right and the problem with coming closer and closer to the
lower frequencies is that I am actually going to pass in my bias that is going to get added up
in the integration process and you are going to have the problem with drift. So, whenever you
have low speed dynamics we really don’t like to use the gyroscope we don’t trust the
gyroscope data, in that case, what you would do? You would use the coefficient to be as large
as possible of course, what do I mean by small what do I mean by large well the basic idea is
that α CPF of the complementary filter it lies between 0 and 1 .

So, let’s say we are absolutely sure that our vehicle will experience very high speed
dynamics, in that case, I would like to trust my gyro more right. So, I would actually set
α CPF to be almost close to 0 , maybe even 0 . If I am always going to have very low
speed dynamics 1 Hz , 2 Hz kind of dynamics, I would start to prefer to trust my
accelerometer more right in that case I’ll set α CPF to be very close to 1 . So, that the
gyroscope data is not being used. So, the interpretation of a complimentary filter is basically
we choose the coefficient in such a way that we trust either the gyroscope measurements or
we trust the accelerometer measurements; this is true of course, for extreme readings for any
readings in between let’s say that you have medium speed dynamics.

So, neither too fast not too slow then maybe we trust both the sensors equally well in which
case I would choose α CPF to be maybe some way close to 0.5 . So, that is the basic idea
of the complimentary filter and we will see in the next lectures how Vinay Sridhar is going to
actually use this complimentary filter to actually compute angles on a real in a real
experiment.

877
 Control Engineering
Dr. Viswanath Talasila
Department of Telecommunication Engineering
Ramaiah Institute of Technology, Bengaluru

Module – 11
Lecture – 03
Complementary Filter

Hello, I’m going to teach you module 11, lecture 3 which is the complementary filter. The
complementary filter is coming from an aspect where you use your high pass filter and the
low pass filter as told by Dr. Viswanath. So, I will be continuing on that on the
implementation of the complementary filter. We have already seen the aspects of jitter of
drift, jitter in the accelerometer, drift in the gyroscope. So, we see how we can implement the
complementary filter to reduce the drift and the jitter.

(Refer Slide Time: 00:54)

So, what is sensor fusion? Sensor fusion is a technique where we use multiple inputs we fuse
the data to get a particular variable out of it. So, for example, over here if you see that there is
an IMU there is a GPS and there is a camera as Professor Viswanath has already spoken to
you about celestial navigation with dead reckoning that technique where you fuse the data
from the dead reckoning sensors and the celestial navigation.

878
So, we use something some block called the sensor fusion block where we add these 2
sensors to get a particular output, here I’m going to show you the fusion between the
accelerometer and the gyroscope.

(Refer Slide Time: 01:49)

Why do we need sense of fusion? As we can see over here, the angle which is computed here
can be used directly from the accelerometer by passing this angle information which uses the
accelerometer values to calculate the angle? So, this is the formula where we calculate the
angle from the accelerometer. It is atan2( Accel x ) , I’ll tell you the explanation

atan 2( Accel x ) divided by a √ ( Acce l ∗Acce l ) + √( Accel ∗Accel )

y y z z . So, if we consider
an accelerometer to be this duster, we have the x-axis which is facing towards us, the y-axis
which is facing parallelly to us, and the z-axis which is facing downwards from us. So, if this
is the accelerometer that we are considering x is always the forward axis, the y is the parallel
axis, and z is the perpendicular axis to the body of the accelerometer.

So, if we want to find out an angle, for example, the pitch angle which is supposed to be like
that a movement which happens upwards in this directions we consider Accel x to be the
forward direction axis, the Accel y which being this axis, and the z-axis which is
perpendicular to the axis of the accelerometer. So, we get meters per second square and g
values from the accelerometer g basically is the g force. So, if I consider an
accelerometer to be stationary like this in air, the x-axis which is the parallel to the ground
and the y-axis which is parallel to the ground does not measure any acceleration, but the z-

879
 2
axis which is measuring the gravity is measuring 9.8 m/s . So, we call that to be g and
we normalize the value of 9.8 m/s 2 into a g value.

So, you get ±1 g depending on the orientation of the sensor. So, the accelerometer value
that we get as g is passed through the angle computation block. So, the angle computation
in this block when we get the g value perform the certain function which is mentioned
over here this angle computation is then passed. So, we get a value of θacc . So, this θacc
is passed through the low pass filter which is removing all the high-frequency noise as we
know that the accelerometer has a lot of high-frequency noise, we can remove it by designing
a customized low pass filter for this particular function where this low pass filter reduces the
jitter from the accelerometer data. So, the value we get over here is θ LP . So, this is the
angle that we get from the accelerometer.

(Refer Slide Time: 04:54)

Similarly, we can do the same from the gyroscope. So, from the gyroscope, we can compute
the angle. The gyroscope basically gives you value in degrees per second. So, which is your
angular velocity? This value from the gyroscope which is degrees per second is passed
through the numerical integration which is shown as a function over here this is the
MATLAB function where we pass the gyroscope in through the numerical integration by
Gyro y∗(time(i)−time(i−1)) and add it with the previous angle of the gyro. So, this is the
function of a numerical integration. So, after we compute the numerical integration of the
gyroscope value we get an angle which is θ gyr .

880
So, as Professor Viswanath has already mentioned that the gyroscope values have after
integration give a lot of drift. So, the drift is compensated by passing this angle through the
high pass filter. So, the high pass filter basically cuts of low-frequency noises which is the
drift and gives you a value θ HP value which is again the angle. So, now that we know that
we can compute the angles from the accelerometer and the gyroscope, we tried to fuse these 2
sensors which is your accelerometer and your gyroscope to compute the angle for better
performance.

(Refer Slide Time: 06:34)

2
So, now we can see here that the accelerometer giving values in m/s and gyroscope
giving values in deg /s , the accelerometer is passed through the angle computation and the
low pass filter, the gyroscope is passed through the numerical integration and the high pass
filter, both of these which is the θ LP and the θ HP are shown previously is then added
and then we get the angle.

881
(Refer Slide Time: 07:17)

So, why do we use a low pass filter for the angle calculated from the accelerometer? So, now,
that we know that angle can be calculated from the accelerometer and this has to be passed
through the low pass filter, I will tell you why the low pass filtering is required, if we
considered this to be the accelerometer, we keep it stationary on the ground, we can get data
which is very similar to this. So, we get a very stationary data which is around 83 ° . So,
the data we receive over here is at a mean of 83 ° , but it has a lot of jitter. So, it varies
from around 82 ° to around 85 degrees, 84 ° . So, we need to pass it through low pass
filter to remove the high-frequency components. So, the red graph which is over here is the
high-frequency component jitter that we get when we calculate the angle from the
accelerometer, we pass the high-frequency component through a low pass filter with certain
cut off frequency.

So, this cut off frequency can be calculated in such a way that the jitter which we see in the
red graph can be removed and we calculate in this particular experiment to be 3 hertz . So,
this frequency that the cut off frequency that we calculate, gives us a response which is in the
yellow form, so, yellow form is the filtered out data and the red is the raw data. So, now, we
can see that when the accelerometer is kept stationary on the ground we get a better curve
which is centered at 83 ° .

882
(Refer Slide Time: 09:05)

So, now I will be showing you why we use the high pass filter. So, the gyroscope data that we
have, which is denoted by ω is in deg /s . So, once this ω is passing through the
numerical integration, we get the value of angle which is θ gyr .

So, this value is then pass through the high pass filter and now we get θ HP . So, after

getting the value of θ HP , we see the experiments which is conducted using the L3G4200D
which is a 3 axis gyroscope, if we consider this to be our gyroscope and keep it stationary on
the ground we see that we get a drift. So, this drift is, basically the drift where if we need to

∫ ω +η . So, the noise also gets integrated when we trying to integrate the gyroscope value

to find out θ gyr , and hence this causes are drift. So, this drift can be seen in this particular
experiment where we use a 3-axis gyroscope and the blue line is the angle from the raw
gyroscope and the black line which is above over here is the angle pass through the high pass
filter.

883
(Refer Slide Time: 10:49)

So, now we see the response of the high pass and the low pass filter, the high pass filter
which is on the left side has a response which is like this where the fc is the cut off
frequency of the high pass filter and you have a pass band from fc to ∞ . So, now, we
have we look at the low pass filter, where the low pass filter has a frequency response which
is similar to this there is a decay after the fc which is your cut off for the low pass filter.
So, if I consider frequency for example, as 3 hertz . So, after 3 hertz there is a 20 dB
slope decay in this response. So, the pass band is basically from 0 to 3 hertz . So, now,
what is the complementary filter? The complementary filter is the combination of the high
pass filter and the low pass filter.

So, let us consider the angle θTrue , this is measured from both your sensors as we know
from the accelerometer and from the gyroscope. So, this angle which is the true angle is
measured from your accelerometer and the gyroscope using their computations the angle
which is θacc is passed through the low pass filter and the angle that is θ gyr is passed
through the high pass filter. So, we have seen already, why we pass the accelerometer angle
and the gyroscope angle through the respective high pass and the low pass filters, this angle
which is low passed and high passed can be called as θ LP and θ HP is passed through a
combination of the high pass filter and the low pass filter.

So, I will show you the combination of the low pass filter by taking the cut off frequency as
3 hertz . So, if I considered 3 hertz cut off frequency, I have a response of my low pass

884
filter which is like this, and the response of my high pass filter which is like this. So,
essentially if I pass the θ LP and the θ HP through the complementary filter block, I will
get an output which is having a response with a magnitude of 1 . So, the y-axis is the
magnitude and the response is always one. So, at the output, we get θTrue . So, for example,

if my θTrue is a lower frequency signal we can see that the low pass filter actually weighs
more in this complementary filter. So, if my signal comes in this region.

My low pass frequency is, the low pass filter basically dominates over the high pass filter to
give the magnitude as 1 and if I get a higher dynamic frequency which is in this region,
my high pass filter takes over and the output from the high pass filter ways more in the
magnitude of 1 to measure the θTrue . So, now, I will be taking you through the design
steps of the complementary filter.

(Refer Slide Time: 14:41)

So, we basically know that the complementary filter algorithm helps to get us better accurate
angles and accurate computations. So, I will be taking you through the design steps of the
complementary filter. So, just to revise back the block diagram of the complementary filter.

885
(Refer Slide Time: 15:01)

This is the complementary filter where the accelerometer which is present over here it gives
2
you a value in m/s , the gyroscope which is present over here it gives you values in
deg /s .

So, right now that the accelerometer is giving acceleration data, the gyroscope is giving me
angular velocity in deg /s , we pass the accelerometer data through the low pass filter, here
is the low pass filter where we’ve already discussed what low pass filter does and then we
passed through the angle computation block where we get θacc of your accelerometer.
Now we already know that from the gyroscope, we can also get the angle by numerically
integrating the values of the gyroscope, we get the angle θ gyr over here and we get
θ gyr HP
. So, these 2 angles that we have that is your θacc and θ gyr HP
has to be added
together in the complementary filter block with the respective weights and we compute the
angles from that.

So, going to the design steps of how we design the complementary filter. So, we know that
the gyroscope and the accelerometer are the sensors which we are using. So, it is very
essential to see you the datasheets of these sensors because the sensors are the main part in
the acquisition of the data. So, here once we see the datasheets of the accelerometer and the
gyroscope, we get the values of the bias and also we get the random noise which is present
while we are doing data acquisition. So, then we carefully choose the cut off frequency, the
cut off frequency fc which is mentioned over here is to be chosen based on basically 3

886
parameters. If you see the note below, the cut off frequencies are chosen depending on the
frequency of movement that is how fast your actual movement is happening, and the noise
that is the jitter present in the accelerometer are shown. Previously, there is a lot of jitter
present in the accelerometer and then also the frequency of the drift which is present in the
angle computation of when we compute the angles from the gyroscope.

So, now, that we have chosen the cut off frequency, we know the formula for a cut off
frequency and the time constant. So, the cut off frequency over here has the formula which is

1
like this f c =¿ . Now depending on what cut off frequency, we choose we can
2 πRC
find the value of τ . So, τ =R∗C that is if you reframe the first equation which is above

1
R∗C=¿ . So, now, we have the values of fc and we have the value of τ .
2πfc
So, next, we see; how we put these values of your τ and you're cut off frequency to design
a low pass filter.

(Refer Slide Time: 18:23)

So, as you see on the left side, there is a circuit which is the low pass filter you have the
V ¿ over here, we have the R and the C circuit. So, this is the passive low pass filter.

So, now, we know that from previous lecture that if V ¿ is considered as x and V out
is considered as y , we get this equation of the low pass filter where yi is equal to

y i=x i ( RC∆+∆t t )+ y ( RCRC+ ∆ t )

887
i−1
 ∆t
xi which is your input into ; ∆t is basically your sampling rate. So, ∆t
RC + ∆ t
is your sampling rate and we know the value of RC which is equal to τ again plus
∆t plus y i−1 . So, this y i−1 is the output of the previous function. So, y i−1 into

RC
.
RC + ∆ t

So, now, with this equation, we can get a simplified equation of the low pass filter which is in
the corner below here, y i=α lpf x i +(1−α lpf ) yi −1 .

So, this is the simplified equation of a low pass filter where α lpf is given by the equation

α lpf =¿ ( RC∆+∆t t ) . So, with this we get from the previous slide we have seen that we

get f c , we get the value of τ .

(Refer Slide Time: 20:25)

And now we have the value of the α lpf , let’s move on to designing the high pass filter. So,
the high pass filter circuit is shown over here where V¿ is here, your capacitor, and your
resistor and this is a passive simple first-order high pass filter. So, previously, we have seen
the equations of this high pass filter to be

y i= y i−1 ( RCRC+∆ t )+( x −x )( RCRC+ ∆ t )

i i−1

888
So, this equation can also be simplified and written as
y i=( α h pf ) y i−1+(α h pf )( xi −xi−1 ) . So, here we know that the input again is considered as
x and the output is considered as y . So, with this equation, we get the filter coefficient

that is ( RC+RC∆ t )
α h pf = . So, now, for the high pass filter, we have the cut off frequency,

we have the value of the time constant and we have the value of the α h pf . So, now, we
have seen that for the low pass filter and the high pass filter, we have the parameters which
are stated here, we have to combine these 2 filters together to make something called a
complementary filter.

(Refer Slide Time: 22:15)

So, I will be showing you MATLAB implementation of the low pass filter. So, if you see over
here, if you remember from the block diagram the accelerometer values which is coming as
2
m/s is passed through the low pass filter and we get the value of LPF over here. So,
we know that we were using 3 axis sensor which is x, y, and z. So, hence we get LPF x

−¿ y
which is your low pass filter of the x-axis, which is the low pass filter of your y-
LP F ¿
axis, and LPF z which is the low pass filter of your z-axis. So, we pass this through the
equation which is shown here, I will be giving you the MATLAB implementation the
complete file which includes the low pass filter application, which also includes the high pass
filter applications and also the complementary filter imp implementation.

889
So, now we see that after we pass the values that we get from the accelerometer through the
low pass filter with the equation which is shown over here, we can see that the values are
more stable, there is less amount of jitter present in your accelerometer if you see the graph
2
on the right side the red graph is basically a raw acceleration value which is again in m/ s
2
or you can consider it to be g , where g=9.8 m/ s .

So, your raw acceleration value which is the graph which is in red is more jittery compared to
the Accel passed through the low pass filter which is in black. So, we see over here that once
we pass it through the low pass filter the mean of this whole line is very close to 9.8 meters
0.98 g , but otherwise, when we do not pass it through the low pass filter as we see in the
red graph it varies from 1 to 0.96 . So, we see that we get more accurate data when we
passed through the low pass filter.

So, this is the data when the IMU which is present, which is your sensor is kept stationery
now we will see an implementation where your IMU is actually moving.

(Refer Slide Time: 24:48)

So, we here again we have the same color combinations where the red graph is your raw
acceleration value consider this to be your accelerometer, we have an accelerometer which is
kept on my hand and I move the accelerometer from 0 ° to 9 0 ° and back to 0 ° . So,
this is the graph that I do get when I move the IMU which is my accelerometer over here
from 0° to around 80° for example, and the red graph, if you see, has a lot of jitter.
So, we pass it through the low pass filter where the Accel values are passed through the low

890
pass filter and is marked in black. So, we see that the black line is a more smoother line
compared to the red line.

(Refer Slide Time: 25:39)

So, now we will see the MATLAB implementation of the high pass filter. So, we know that
when we calculate angles from your gyroscope, where the Angle gyro is the computed
angle. So, here we have Angle gyro which we then pass it through the high pass filter. So,
we know the equation of the high pass filter which is previously shown. So, we pass the
Angle gyro through the high pass filter where the equation is shown over here as well and we
get HPF . So, the value of HPF . So, this HPF value is the value which we are
getting after we pass the angle computed from the gyro through the high pass filter, here on
the right side we can see a graph which shows the component of drift. So, if you see over
here the IMU is kept stationary on the ground you see that this is the stationary data we
always want the angle to be constant.

So, we keep it at around 0 ° because of this integration there is a buildup of noise and then
we see that the angle which we are computing this; the y-axis is the angle which we are
computing from the gyroscope is always drifting. So, at around 8000 samples we are getting
around 9 0 ° of drift. So, we see that there is an angle which is not stationary, but the noise
which is building up causing the angle to be drifted. So, now, when we pass this angle
through the high pass filter it does not allow changes as spoken previously and gives you a
line which is the black line.

891
So, this black line is the line where when we have a stationary value it when it pass through
the high pass filter we get almost a straight line. So, so from this we get to know that we need
to remove the component of the drift in your angle computed from the gyroscope and your
component of jitter from the accelerometer. So, this can always be done only by the good
selection of your fc which is your cutoff frequencies. So, if your frequencies are not
chosen properly, your angle which may be that your angle which you are calculating is
drifting away or your jitter which is the component present, which is the component of noise
present in your accelerometer does not get removed properly.

(Refer Slide Time: 28:17)

So, here is another MATLAB implementation of the high pass filter. So, on the y-axis, we
have angles again which is computed from your gyroscope and your x-axis is your data
samples. So, we see over here again we conduct an experiment where the IMU is present on
the hand we moved from 0° to 80° and back, and we continuously repeated for 7
times. So, we see that there is a graph which is coming which is your blue line the raw angle
which is calculated from your gyroscope. So, we see that the blue line over here has a lot of
drift if you can see I’m doing an experiment where I’m going from 0° to 8 0 ° and
coming back to 0 ° .

So, this experiment which I’m doing has a lot of drift where for example, after 9000 samples,
I’m getting it to be around 7 0 ° . So, which is this value of around 70° is not correct
because I have stopped the experiment at 0 ° . So, hence I pass this through the high pass

892
filter which is your black line again over here and I get a data where at 9000 samples I’m
almost at 0 ° . So, you will say that there is an issue with this graph where for example, the
data which we are computing in the blue line which is your angle computation from the
gyroscope has a bigger magnitude than the magnitude present over here in the black line.

So, your magnitude present over here is larger than the magnitude which is present in your
high pass filter. So, now, I will explain why we get this value I will just draw a filter
response, the frequency response of this high pass filter taking the cutoff of f c . So, in your
blue graph, we have some data which is being removed because the high pass filter is
designed in such a way that the fc is also removing some part of the signal. So, for
example, the signal which we are getting is around this range. So, anything which is below
the f c is being rejected and only the signals which is on the right side of the f c is being
passed hence we see that there is a dropped in the magnitude in this graph as compare to this
graph.

So, we need this is very important when we chose the cut off frequencies of your high pass
filter we do not want the signal also to be removed we just want the drift to be removed over
here. So, what do we do? So, what do we do over here?

(Refer Slide Time: 31:56)

So, now we know that this is my fc which is chosen, in this graph I want to chose an
fc which is much lower which come somewhere to the left towards 0 ° . So, that my

893
filter response looks like this and the signal which is present in this area is also passed
through. So, I’ll be showing you an implementation where we have actually changed the
f c we have made it a lower f c .

(Refer Slide Time: 32:40)

So, that we get better values of the signal. So, on the right side if we see over here, on the
right side if we see over here the magnitudes of the blue line and the black line is very similar
only that the component of drift is being removed. So, here we have chosen an appropriate
f c such that we do not reduce the magnitude of my response.

(Refer Slide Time: 33:06)

894
We have already seen from the previous examples as to how to calculate the angle this is the
MATLAB implementation of how we calculate the angle from your IMU, from the
accelerometer we calculate the angles in this manner. So, we take this equation and we get the
value of the Accel. So, this Accel angle is got from this equation which is above. From the
gyroscope we pass it through the numeric integration, we get the Angle gyro and we pass it

through the Angle G yr o HP to get the angle in the output.

So, here is how we calculate the angle from the gyroscope this is also shown previously. So,
this is a reputation of this and now we pass this through the block of the high pass filter to get
Angle G yr o HP which is this complementary filter is a filter in which we combine these 2
angles that is your Angle Acc and Angle G yr HP with certain weights to get an output
response which comprises of the true signal.

(Refer Slide Time: 34:36)

So, the complementary filter as we know follows this equation where the Angle from
your complementary filter is equal to αcf which is your filter coefficient of the

1−α
complementary filter into the angle that we get from your accelerometer plus (¿¿ c f )
¿
which is again your filter coefficient of the complementary filter into the angle we compute
Anglethrough the high pass filter,
from the gyroscope after passing
Angle
Angle= ( α cf ) ¿accelerometer ) +(1−α cf )(¿ gyroscope)
(

895
So, now, we need to choose the value of α c f . So, how do we chose the value of αcf ?
So, if you can say that the frequency of motion of, for example, you are trying to measure a
certain movement, for example, I’m just rotating moving my hands from 0 degree to 90
degrees again and back. So, in you are you actually want to see what is the frequency of
motion over here, and from the frequency of motion, we decide to give weightage to either
the accelerometer or give weightage either to the gyroscope.

So, we see that we are moving the IMU from 0° to 9 0 ° and back to 0 ° , we see the
frequency of movement of this hand which is measured by your sensor and if the frequency is
lower, we give more preference to our accelerometer. If your frequency of motion is higher,
we give more preference to the gyroscope here α c f is in between 0 to 1 where if we
give the value of α c f =0 , the angle from the gyroscope is dominant or if we give
α c f =1 , angle computed from your accelerometer is dominant.

So, we need to choose a value of αcf depending on how fast or how slow we are moving.
So, after we chose the filter coefficients of the complementary filter we see that your low pass
filter will give a response which is similar to this and you can say this is my fc and my
high pass filter will give a response which is similar to this. So, now, that we are getting a
response like this with the same f c , we know that we will get a response of magnitude of
1 when we pass the signal through the complementary filter.

(Refer Slide Time: 37:54)

896
So, now, we will be discussing a small problem on this design of the complementary filter,
here consider an induction motor which is rotating at 1100 rpm ,
1100 rpm=6600 deg /sec . So, 1100 rpm=6600 deg /sec which is approximately equal
to 18 hertz . So, we know that the equation for calculating from rpm to hertz , we

1
get 1rpm= hertz .
60

1
So, if we get 1100 rpm , then ∗1100 , we approximately get 18 hertz . So, we
60
considered the value of 18 hertz which is your frequency of your message signal and we
strap on and IMU to calculate the angle of this motor shaft. So, we have an accelerometer and
gyroscope which is present on the motor shaft which is computing the angles in the motor
shaft. So, we know that in this problem we have already been given the value of fc which
is f c =20 hertz .

So, here we need to design a low pass filter and a high pass filter to and combine them
together to get a complementary filter, we are also given the value which is an assumption
that C=1 μ F and your sampling rate which is f s=50 hertz . So, we know the equation

1 1
of fs and ∆ t , where ∆ t=¿ . So, now fs , ¿ 0.02 . So, we’re here
fs 50
if I have to just mark my parameter that I have I get f c =20 hertz , your value of
∆ t=0.02 and your value of C=1 μ F .

897
(Refer Slide Time: 40:34)

So, now we see how to design the low pass filter. So, we know that your value of
f c =20 hertz . So, we need to design a high pass filter and a low pass filter for the same cut
off frequency and we also we need to find the value of τ which is your time constant
where τ =RC , where your sampling rate is 50 hertz , we know the equation of τ

1 1
where τ =RC=¿ = , we get the value of τ =0.0079579 .
2 π f c 2 π∗20

So, now we know the value of τ =0.0079579 , after getting the value of τ , we calculate

τ 0.0079579
the value of R , where τ =RC ; R=¿ ¿ . So, your value of
c 1∗10−6
R=7.95 kilo−ohms , which is given over here. So, from this problem, we have found out a
few parameters which is your f c , which is your ∆ t , your C , your τ , and R
with these parameters.

898
(Refer Slide Time: 42:35)

We start to design the filter, on the left side if you see it is the filter circuit of the low pass
filter and on the right side we see that it is a filter circuit of the high pass filter.

So, we know the values of R=7.95 kilo−ohms , I mark it on the right side as well
7.95 kilo−ohms and C=1 μ F and C=1 μ F . So, we actually completed the first
question where we have designed the low pass filter circuit and we also design the high pass
filter circuit, now we need the values of α . So, now, on the left-hand side, we want to find
the value of α lpf where we can a put it into the equation. So, the value of α lpf if you

remember the equation is equal to ( RC∆+∆t t ) , if I have to rewrite again

0.02
( 0.00795+0.02 ) , this is your value of α lpf , where you get α lpf =0.7155 .

So, this value of α lpf you can put it into the equation of your low pass filter where y is
your output, and x is your input. So, from this, we have actually got the equation of your
low pass filter. Now we move on to the high pass filter where the equation for calculating the
α h pf which we want to put it into the filter equation the equation, the formula for

α hpf = ( RCRC+ ∆ t )=( 0.00795+

0.00795
0.02 )
. So, if we calculate the value of α h pf we get the

899
value 0.2844 , we put this equation into the equation of the high pass filter where y is
your output and x is your input.

So, now we have design a basically 2 circuits which is your passive low pass filter circuit and
your passive high pass filter circuit next.

(Refer Slide Time: 45:38)

We have also calculated the value of α l pf and the value of α h pf and also put it into the
equations of your low pass filter and the high pass filter. Now we move on to finding out the
transfer function of these 2 filters. So, now, we know the value of R is equal to point, sorry,
we know the value of R=7.95 kilo−ohms , and value of the capacitor 1 μ F .

Similarly, over here, 1μ F and 7.95 kilo−ohms . So, we already have studied Professor
Viswanath has taught you about the transfer function of these filters. So, we know the transfer

1
function of the low pass filter which is given over here H ( s)lpf =¿ . So, we
1+ RCs
know the value of RC , where τ =RC=0.00795 . So, we get the transfer function of

1
your low pass filter to be H (s)lpf =¿ .
1+ ( 0.00795 ) s

So, now we know the transfer function of the high pass filter where H (s)h pf =¿

RCs
, where again RC=τ=0.00795 . So, we get the value of H (s)h pf =¿
1+ RCs

900
 ( 0.00795 ) s
. So, this is the transfer function of your high pass filter and on the left side
1+ ( 0.00795 ) s
is transfer function of your low pass filter.

So, now we have come from the cut off frequency, we have calculated your R and C
values and your τ values and we’ve also calculated your α values, put it into the
equation of the low pass filter and the high pass filter and we’ve come to get the transfer
function of the low pass filter and the high pass filter. After we have the transfer functions of
the high pass filter and the low pass filter, we now find out what is the corner frequency cause
with the corner frequency we can get the bode plot. So, we can draw the bode plot by using
the corner frequency which is ω c which is in terms of radians per second.

(Refer Slide Time: 48:20)

So, now that we got the equation, I will again input the values of R and C .
R=7.95 kilo−ohms and C=1 μ F , 1μ F and 7.95 kilo−ohms . So, these are the

1
2 equations where we calculate the corner frequency which is ωc , ω c =¿ is
RC

1
equal to which is ω c =¿ when we do this computation we get
0.00795
ω c =125.786 rad /sec which is the same for your high pass filter as well.

So, your ω c for your high pass filter and your ω c for your low pass filter are the same,
where both ω c =125.786 rad /sec . So, now, with transfer function, there is in the

901
implementation of the MATLAB code that we will be giving it you, we’ve also shown the
bode plot. So, with the transfer function, you can actually draw the bode plot or in a freehand
drawing, you can draw the bode plot of this by in this manner. So, 125.786 is your
3 dB margin which comes over here. So, over here you get the value of ωc in
rad /sec .

So, we know the corner frequency of this bode plot where ω c =125.786 rad /sec . So, here
are the bode plots of the low pass filter and your high pass filter, it follows the curve as
shown over here and it follows the curve as shown over here. So, with this, we have actually
designed the low pass filter and the high pass filter and also got your bode plots. So, just to
revise, just to summarize on what steps we have taken. Steps for design can be written like
this. So, we choose the values of f c , then calculate the τ , where τ =RC .

Next, we get the values of R and C and input into the circuit respectively for the low
pass filter and the high pass filter after we get the circuit, we calculate the values of α l pf

and the value of α h pf . So, these are the filter coefficients, we put it into the equations of
low pass filter and high pass filter. After we get the equations of the high pass filter and the
low pass filter we get the transfer functions of these 2 filters, once we get the transfer
functions of these 2 filters we get the bode plots of these 2 filter.

So, we’ve actually completed the design of the low pass filter and the high pass filter now we
will see how to combine these 2 filters to get the complementary filter.

902
(Refer Slide Time: 52:26)

If you remember the block diagram, we had Accel values which is m/s 2 , your gyro values
which are getting in deg / s . So, this Accel value is passed through the low pass filter and
you get the value of Accel LP , from the value of Accel LP , we pass it through the angle

computation block which is shown previously and we get the value of θ LP .

So, this is the θ LP which is the angle got from the Accel value pass through the low pass
filter and gone through the computational block for calculating the angle. Now we go to the
gyroscope, where the gyroscope is passed through the numeric integration and we get the
angle θ gyr , this gyroscope θ gyr is passed through the high pass filter and we get the
value of θ H P .

Now, that we have the value of θH P and θ LP we combine these together with weights
of α cf and 1−α cf to get θcf which is your complementary filter. So, here is a
complete design of the complementary filter, you can actually implement it if you have an
accelerometer which is interfaced and a gyroscope which is interfaced; you can rig up a
circuit which is similar to this on the top, where you have low pass filter with a resistor and
capacitor you passing the input from the input end and you get the output from the output
end, you can pass this block through an angle computation block which is shown previously
to get θ LP and you use the value of α cf to pass it into the complementary filter.

903
So, from the gyroscope, you can actually do something like a numeric integration you will
get the value of θ gyr , you can take this value and rig up a circuit for the high pass filter
where your capacitor is 1μ F and your resistor is 7.95 kilo−ohms approximately
7.95 kilo−ohms and you will get the value of θH P you can use this θ H P which is

multiplied by this weight which is 1−α cf and also put it into the complementary filter. So,
this is how you can implement a complementary filter in real-time.

So, now we have shown the implementation of the complementary filter and how to actually
design it from end to end from choosing your cut off frequency and coming to your bode
plots and also integrating these two filters which is your low pass filter and high pass filter to
get a complementary filter. In the forum, we will be having some reference material as some
of you all would like.

(Refer Slide Time: 55:44)

We are going to give you an Arduino code, which interfaces an IMU which is basically
considering GY80, this is the sensor that we have chosen, this GY80 consists of an
accelerometer, a gyroscope, and a magnetometer we’ll be giving you the codes where the
GY80 will be computing accel values and your gyroscope values.

So, this accel values and the gyroscope values will be printed on serial monitor you can
actually interface it to MATLAB or something like that, where in our MATLAB code we will
also have an option where this serial data that we get from an actual sensor which is your
GY80 will be taken into MATLAB which will read the data from these accel or also read a

904
data from the CSV file, the choice will be given where 2 different MATLAB codes for either
reading the data or directly receiving the data taking the data from a CSV file will be given.
We’ll be able to plot the raw data which is your accel values, your gyroscope values. We’ll be
showing you the filtering techniques which is your high pass filter and the low pass filter and
also plotting it and we will be showing you in another file which is the complementary filter
equation for implementation of the complementary filter and also plotting the angles which
we get from the complementary filter.

I hope you guys have enjoyed this course. This was a practical implementation of the
complementary filter as I said the reference material will be given and if you require any help
please feel free to ping me.

Thank you.

905
 Control Engineering
Dr. Viswanath Talasila
Department of Telecommunication Engineering
Ramaiah Institute of Technology, Bengaluru

Module - 11
Lecture - 03
Part 2
Complementary Filter

Welcome to the lecture on control engineering. So, I will be taking you on to the
applications of the complementary filter.

(Refer Slide Time: 00:26)

So, the complementary filter is a sensor fusion algorithm where we calculate the angle
from the accelerometer and then after the angle computation block from the accelerometer,
we get the angle underscore accelerometer (𝑨𝒏𝒈𝒍𝒆_𝑨𝒄𝒄), we pass it through the low pass
filter and bring it into the combination block; from the gyroscope we also get the angle by
numeric integration. The formula for numeric integration is shown below this angle from
the gyroscope is passed through the high pass filter and it comes to the combination of the
low pass filter of your angle from your accelerometer and your high pass filtered angle
from your gyroscope.

So, as discussed previously the complementary filter’s equation is as follows,

𝜽𝒄𝒇 (𝒕) = 𝜶𝒄𝒇 ∗ (𝜽𝒉𝒑𝒇 ) + (𝟏 − 𝜶𝒄𝒇 ) ∗ (𝜽𝒍𝒑𝒇 ) where,

906
𝜶𝒄𝒇 − filter coefficient of your complementary filter.

𝜽𝒉𝒑𝒇 − angle which comes from the gyroscope.

𝜽𝒍𝒑𝒇 − angle calculated from the accelerometer.

So, these 2 are combined depending on the dynamics of the moment that we have. The
complementary filter equation is this, where 𝜶𝒄𝒇 is your filter coefficient for the
complementary filter, 𝜽𝒄𝒇 is the angle computed from the complementary filter, the 𝜽𝒉𝒑𝒇
is the high passed value of angle from your gyroscope and 𝜽𝒍𝒑𝒇 is the low pass value of
the angle from your accelerometer.

So, I will be showing you the MATLAB implementations of why we use the low pass
filter and the high pass filter and how do we combine these 2 angles from the accelerometer
and the gyroscope in the complementary filter.

(Refer Slide Time: 02:35)

Here are the 3 codes that we have for the actual implementation of the high pass filter for
the gyroscope angle and the low pass filter from your accelerometer angle and the
complementary filter block which combines the high pass filter and the low pass filter
angles in one particular computational block. So, I will show you first how we have the
drift in a particular stationary position and how we correct the drift in this particular
experiment.

907
So, here we have a csv file which is in the name 𝑠𝑡𝑎𝑡𝑖𝑜𝑛𝑎𝑟𝑦. 𝑐𝑠𝑣, I will show you the
excel file for that. So, this is the excel file where we have collected data from an IMU that
we have. So, this IMU consists of an accelerometer and a gyroscope where the
accelerometer is the ADXL345, and the gyroscope which is L3G4200D.

(Refer Slide Time: 03:39)

So, here we can see that how we save the data as, column number A which is 𝑎𝑐𝑐𝑒𝑙 𝑥. So,
this is the axis of the accelerometer which is pointing forward. So, if we consider this to
be the accelerometer and I keep it stationary, the axis which is pointing forward is the x-
axis, the axis which is perpendicular in this manner is the y-axis and the axis which goes
through the device is the z-axis.

So, axis-x of the accelerometer is in column A, axis-y of the accelerometer is in column

B, axis-z of the accelerometer is in column C. The same holds true for the gyroscope when
kept stationary. So, the axis-x of the gyroscope is in column D, the axis-y of gyroscope is
in column E and the axis-z of the gyroscope is in column F. So, I go back to the MATLAB
code and I read these values into a particular variable. So, from the file, I get it into a
variable which is called 𝑑𝑎𝑡𝑎. So, my 𝑎𝑐𝑐𝑒𝑙 𝑥 is the first column as we know. So, we call
it from the first column, 𝑎𝑐𝑐𝑒𝑙 𝑦 is from the second column, accelerometer z is the third
column, so we call it respectively. The gyroscope data which is the x-axis gyroscope data
is in column number 4, the gyroscope y data is in column number 5, and the gyroscope z
data is in column number 6.

908
So, this is how we read the data from the csv file, after reading the data from the csv file
we calculate the Mean Bias Value as Professor Viswanath has already spoken to you about
bias, we will be calculating the bias of these values that we have in an experiment where
we have collected the data. So, we take around 200 samples we calculate the sum of these
200 samples and divided by 200. So, we get the mean bias which is in a stationary
condition of the x, y, and z axis of the gyroscope, this mean gyroscope value has to be
subtracted from each of the axis, now we consider only the y-axis. So, I’m subtracting the
gyroscope mean value of y into the gyroscope y value.

So, the bias is a basically removed. So, here are some experimental results of the bias
where x-axis is measured as 0.04, the y-axis bias is measured as 2.85 and the z-axis bias
is measured as 1.029. So, here this is an experimental set up where we have measured
these bias which may not be true all the time. So, if you want to calculate the bias we have
to keep the IMU stationary on the ground and calculate the bias and remove the bias
initially from the gyroscope value.

So, next, we would say; what is the data rate at which we are getting the data from the
accelerometer, the gyroscope. So, the data rate at which we are getting is around 200
samples per second it’s in the code where we actually call the data in at this sample rate.
So, we get 200 samples per second. So, 𝑻𝒔 which is also 𝐷𝑒𝑙𝑡𝑎 𝑡 is 𝟏⁄𝟐𝟎𝟎. So, this is the
delta where 𝐷𝑒𝑙𝑡𝑎 𝑡 where we need to compute the 𝜶 for the high pass filter and the low
pass filters. So, here specifically for the high pass filter.

So, as we know that the sensor is kept stable, on stationary, on the ground, on the table.
So, we consider the first value of the angle of the y-axis to be 𝟎 and we have a for loop
where we integrate the gyroscope y value as we said we consider the y-axis you can
consider the z-axis or the x-axis, but here, in this case, we are considering the y-axis. So,
we keep the y-axis stationary or the IMU completely stationary and we calculate the angle
from the y-axis. So, 𝒂𝒏𝒈𝒚 is a variable where we get the value of the angle from the
gyroscope.

909
(Refer Slide Time: 08:22)

So, if I plot this value of the gyroscope. So, here is the graph of the angle measured by the
gyroscope. So, the y-axis is measuring the angle, the x-axis is measuring the samples. So,
we can see that over a period of time which is around 8000 samples my value of drift, my
value of the angle which is when kept stationary has moved from 0 to around 2.25. So, we
see that there is a lot of drift in this experiment that we have conducted. So, how do we
remove the drift of the angle computed from the gyroscope is by passing through the high
pass filter.

So, I will be telling you how to pass it through the high pass filter, as we have seen how to
calculate the filter coefficients previously in the lectures given by Professor Viswanath, I
will be going through the same again. So, 𝒇𝒄 is the cut-off frequency, where
𝟏 𝟏
𝒇𝒄 = . Now we calculate for 𝑹𝑪, which is your time constant 𝒕𝒂𝒖 = 𝟐𝝅𝒇 from this,
𝟐𝝅𝑹𝑪 𝒄

we calculate your 𝒕𝒂𝒖 value as 𝒕𝒂𝒖 = 𝑹𝑪 and when you want to calculate the filter
coefficient 𝒂𝒍𝒑𝒉𝒑 = 𝒕𝒂𝒖⁄(𝒕𝒂𝒖 + 𝑻𝒔 ).

So, the filter coefficient that we have calculated is for 𝟑 hertz. Your cut-off frequency of
this filter is at 𝟑 hertz and we get the alpha value of 0.8641. So, we pass this; this the angle
computed from your gyroscope through the high pass filter using this filter coefficient. So,
now, that we know that the alpha which for the high pass filter is 0.8641 we put it in the
high pass equation which is defined previously, which is explained to you by Professor
Viswanath.

910
So, we get the value of the high pass signal as 𝒂𝒏𝒈𝒈𝒚𝒓𝒚𝒉𝒑. So, if I have to plot this value
of 𝒂𝒏𝒈𝒈𝒚𝒓𝒚𝒉𝒑 with, I would get a graph which looks similar to this, yeah. Now you can
see that your blue line as seen previously is your angle from your gyroscope which is
drifting of even when kept stationary and when we pass it through the high pass filter we
get a very smooth line which is almost at 𝟎 all the time. So, here is how we correct the
drift from the angle of the gyroscope by passing it through the high pass filter.

Now, I will be taking you through how to reduce or how to reduce the jitter from your
accelerometer by passing it through the low pass filter.

(Refer Slide Time: 11:46)

Yeah. So, all these codes are available in your forum, if you have any doubts with the
codes you can always contact me. So, over here we are getting the data of the same
stationary value; we know that your accelerometer x data is in column number 1,
accelerometer y is in column number 2, accelerometer z is in column number 3. Here we
do not consider your gyroscope, but for computational sake, I have given the values over
here. As we know, from the previous experiment for the high pass filter 𝑻𝒔 is your 𝐷𝑒𝑙𝑡𝑎 𝑡
and my sampling rate is around 𝟐𝟎𝟎 hertz.

So, I get 200 samples per second, where 𝐷𝑒𝑙𝑡𝑎 𝑡 is 𝟏⁄𝟐𝟎𝟎, this has to be checked properly
while your conducting the experiment because every time you will not get a standard data
rate unless you have defined your data rates well. So, please check this before you use this
same code. Now we calculate the angle from your accelerometer. So, the angle from your

911
accelerometer is calculated using this formula. So, we have already shown to you as how
what the formula for the accelerometer angle computation is. So, this is how we implement
it in MATLAB.

So, once we have got this data I want to design a low pass filter for this accelerometer data.
So, as this, as shown before I will go through the cut-off frequencies and how we calculate
your filter coefficients from the cut-off frequencies; so, 𝒇𝒄 is your cut-off frequency, where
𝟏
the formula of 𝒇𝒄 = ; 𝑹𝑪 is also your 𝒕𝒂𝒖 which is your time constant which can be
𝟐𝝅𝑹𝑪
𝟏
said as . So, now, your alpha for a low pass filter is 𝒂𝒍𝒑𝒍𝒑 = 𝑻𝒔 ⁄(𝑻𝒔 + 𝒕𝒂𝒖).
𝟐𝝅𝒇𝒄

Once we substitute the values of 𝒕𝒂𝒖 and 𝐷𝑒𝑙𝑡𝑎 𝑡 or 𝑻𝒔 in the equation, we get the value
of alpha for the low pass filter again over here the alpha for a cut-off of 𝟑 hertz is around
0.0909. So, we consider this 𝟑 hertz as for the high pass filter we have also consider the
𝟑 hertz as the cut-off frequency, we consider the 𝟑 hertz cut-off frequency even for your
low pass filter.

So, we pass the angle computed previously from your accelerometer data through the low
pass filter. So, this is your equation which is being implemented in MATLAB the equation
is previously shown. So, you can implement this equation for your low pass filter in this
manner, once we pass the accelerometer data through the low pass filter we can see how
we remove the jitter from the angle computed from your accelerometer data.

(Refer Slide Time: 15:04)

912
So, here is the graph where when kept stationary the accelerometer is not moving. So, we
need to get an angle which is ideally a straight line, but as we see, that the angle calculated
from your raw acceleration value which is in terms of red has a lot of jitter and the angle
computed from your accelerometer after passing through the low pass filter is in yellow.
So, we can see there is a considerable amount of jitter which is being removed when passed
through the low pass filter.

So, now, I will be showing you on how we combine the low pass filter and a high pass
filter to get the complementary filter. Now I will be showing you how the complementary
filter block is designed, where we combine the angle calculated from the gyroscope and
the angle calculated from the accelerometer in the previous experiments I have shown.
Over here the basic steps are very similar. So, we have the data, we read from the stationary
file and then we assign it to variables which are a accx accelerometer x first column,
accelerometer y the second column, accelerometer z which is the third column, the fourth
is the gyroscope x, fifth is the gyroscope y, and sixth is the gyroscope z.

So, we find the mean which is the bias and we remove the bias in this way and we know
that the 𝑻𝒔 which is the 𝐷𝑒𝑙𝑡𝑎 𝑡 is 𝟏⁄𝟐𝟎𝟎. So, now, we need to calculate the angles from
each of the data; with each of the sensor, which is the accelerometer and is gyroscope. So,
we find out we keep the initial condition from the accelerometer which is this formula atan
2 accel x divided by square root of y square plus z square and the gyroscope is also initially
given the same angle which is accel x divided by root of tan inverse of accel x divided by
square root of y square plus z square.

i.e., 𝑎𝑡𝑎𝑛2 (𝑎𝑐𝑐𝑥(1), 𝑠𝑞𝑟𝑡 ((𝑎𝑐𝑐𝑦(1) ∗ 𝑎𝑐𝑐𝑦(1)) + (𝑎𝑐𝑐𝑧(1) ∗ 𝑎𝑐𝑐𝑧(1)))) ∗ (180/𝑝𝑖)

So, now, here we calculate the angle from the gyroscope in this way and we calculate the
angle from the accelerometer in this way, now we have the complementary filter which is
a combination of your low pass filter with your high pass filter using a particular filter
coefficient here we know that we have a stationary position. So, we need to trust the
accelerometer more than the gyroscope because the gyroscope has drift. So, we choose our
filter coefficients of our complementary filter in this manner where we choose 1 minus
alpha is equal to 0.98 and alpha is equal to 0.02. So, now, we can plot this signal.

913
(Refer Slide Time: 18:34)

So, we get it in this manner as we have seen previously the blue line is the angle computed
from your gyroscope and the green line which is superimposed by the red line over here is
your angle computed from your accelerometer and the red line is the actual complementary
filter angle.

Now, this is done without filtering without passing each of the angles through the high
pass filter or the low pass filter. So, let’s see on how the dynamics change when we pass
the angles calculated from the gyroscope and the accelerometer through the high pass and
the low pass filters.

914
(Refer Slide Time: 19:25)

As previously calculated our filter coefficients for the low pass filter and the high pass
filter is as follows, the alpha for the high pass filter is previously calculated as 0.8641 and
for the low pass is 0.0909, we have seen that in the previous experiment as to how to
calculate the alpha values, we initialize the values which are required for the computation
of these angels.

Now, we compute the angles of the accelerometer using this formula, after computing the
angle we pass this angle which is 𝒂𝒏𝒈𝒂𝒄𝒄𝒍𝟏 into your low pass filter where your
accelerometer data of the x, y and the z is passed through the low pass filter, next we
compute the gyroscope angle in this way. So, 𝒂𝒏𝒈𝒈𝒚𝒓 is added with the gyroscope y value
and multiplied with 𝐷𝑒𝑙𝑡𝑎 𝑡 where the numerical integration is the gyroscope into 𝐷𝑒𝑙𝑡𝑎 𝑡
which gives you angular position.

So, now we pass this gyroscope value through the high pass filter, the high pass filter of
the gyroscope, the angle calculated from the gyroscope is this equation. So, the equation
is already previously mentioned where we pass the angle from the gyroscope through the
high pass filter, now we see that this to have to be combined using the complementary
filter. So, the filter coefficients from your complementary filter is alpha is equal to 0.02
and the (𝟏 − 𝜶𝒄𝒇 ) = 𝟎. 𝟗𝟖.

So, now, we can see the difference in the graphs that we get with pre-filtering and without
pre-filtering from the high pass and the low pass filters yeah on the left side we see a plot

915
which is the angle calculated from your gyro which is the blue graph, the angle calculated
from your accelerometer is superimposed by your complementary filter which is the red
graph, but the green graph is your angle computed from your accelerometer, the red graph
over here is the complementary filter angle.

So, here we do not pass the angle from your gyroscope or the angle from your
accelerometer through the high pass or the low pass filters respectively, we just pass it we
just combine these two angles directly to get the complementary filter, on the right side
we have the angle from your accelerometer which is in green, the angle from your
gyroscope which is in blue, and the angle from your complementary filter which is in red
again, here we pass the angle computed from your gyroscope and your accelerometer
through the high pass and the low pass filter and we combine these two to get this
complementary filter value. I will tell you why the gyroscope value is towards 𝟎, that is
because we see over here that the gyroscope is basically stationary.

So, now the movement or what the angular velocity which is present in the gyroscope is
very low. So, we pass it, when we compute the angle it is basically a stationary angle with
lower frequency dynamics as we see that the angles computed from the gyroscope when
passed through the high pass filter the lower frequency which is below the cut off
frequency is actually filter out an attenuated hence we get the value which is close to 𝟎°
because all the frequencies which are present in stationary is actually attenuated by the
high pass filter and we get a 𝟎° angle. So, now, we can see that in a stationary condition
where the sensor is not moving we can trust the accelerometer more than the gyroscope.

So, now, we have, we have seen the complementary filter where we combine the angles
computed from the accelerometer and the gyroscope which is passed through the low pass
filter and the high pass filter respectively in a stationary position, now I will show you an
experiment where we have actually taken the gyroscope which is strapped on to the hand
and we move the gyroscope and the accelerometer which is a sensor which is over here in
𝟎° to 𝟗𝟎° manner and bring it back to 𝟎°. So, it goes from 𝟎° to 𝟗𝟎° and back all the way
to 𝟎°. So, here is a file where which is called 𝑚𝑜𝑣𝑒𝑚𝑒𝑛𝑡. 𝑠𝑒𝑟𝑖𝑎𝑙, I won’t run through that
code because it is the same code, I will just show you the graphs.

916
(Refer Slide Time: 24:55)

So, the graph on the left which is your figure 1 is where we find the angle from your
accelerometer and the angle from your gyroscope without passing through the high pass
and the low pass filter and combining these two angles together in a complementary filter.
So, here we can see that the blue line is the angle computed from your gyroscope, the green
line again which is superimposed by the red line is your accelerometer angle and we see
that the red line is your complementary filter angle, here we can see the value of drift,
where we can see that the angle computed from your gyroscope is drifting in a way where
we get peaks from 𝟎° to around 𝟔𝟎°, but at sample number 7000 we actually start the angle
value from 𝟔𝟎° to around 𝟏𝟐𝟎°.

So, here we can see the gradient in which there is the drift. So, now, again on the right
side, we have the angle computed from your gyroscope which is the blue line, the angle
computed from your accel which is your green line and the angle computed from your
complementary filter which the red line here we can see that. So, once we compute the
angles and pass it through the high pass and the low pass filter respectively, we get this
kind of a response from the complementary filter which is your red axis after computing
the angles we see that we are trusting the accelerometer more than the gyroscope, but we
can play around with the alpha value of the complementary filter to trust either the
accelerometer or the gyroscope we would trust the gyroscope more in a fast dynamic
movement.

917
So, an example to show you how we can trust the gyroscope more on the left side we can
see that till sample number 3000, we have a slow speed motion where we do not move
very rapidly, we have a value where we compute in a very slow manner, on the right half
where we start from around 6500 to 8000, we have fast dynamics. So, here if I zoom in
into the plot, you can see that the frequencies above 𝟑 hertz also are captured from what
is compared previously there are the dynamics of the fast motions which are also captured.
So, here we can actually compute the complementary filter by trusting the gyroscope more.

So, now that we see the higher dynamics of the gyroscope where the gyroscope can be
trusted more, I request you to play around with the alpha of the complementary filter where
if you trust the gyroscope more where the 𝜶𝒄𝒇 is given a higher weightage towards the
gyroscope, the complementary filter will actually follow the gyroscope. So, I request you
to please try by changing the alpha for the complementary filter if you have any doubts
you can contact me on the forum all these codes are posted over there you can get in touch
with me.

Thank you, all the best.

918
 Control Engineering
Dr. Viswanath Talasila
Department of Telecommunication Engineering
Ramaiah Institute of Technology, Bengaluru

Module - 12
Lecture – 01
Introduction to State-space Systems

Ramkrishna has already covered quite a bit of details in classical control. So, he has covered
Laplace transform, the transfer function modeling, stability analysis in root locus, bode, so on
and so forth. What I would like to do in these series of lectures is to go to the other design of
controller; other option of control systems which we call as modern control systems and
specifically in this lecture, I will be talking about state-space, alright.

So, we will see what is state-space control or the state-space systems. We will see some
modeling aspects of state-space systems, I will present a few examples in this and then we
will go ahead with other aspects as well. So, first of all, we have already seen what Professor
Ramkrishna has already taught about the transfer function modeling approach and the
question naturally arises; why do we want the state-space approach, alright. So, before we go
to the state-space let’s see why we need a different approach as compared to the Laplace
transforms. So, on the slide, I have got 2 boxes over here, on the left I have got the good
aspects of the transfer function approach, on the right I have got the drawbacks.

(Refer Slide Time: 01:30)

919
So, a transfer function is actually very useful and it’s actually one of the most dominant forms
of modeling in the control industry. So, you go for process control, aircraft any kind of
systems, ships, so on and so forth. The dominant way of modeling a system and then
controlling it is always from the viewpoint of a transfer function, alright. So, the basic
concept of a transfer function is that it is some kind of a generalization of the Fourier
transform, alright. So, once we see that within the Fourier transforms, we know that we are
essentially trying to have a signal processing viewpoint of a system. So, you have a system
and you give inputs of varying say frequency and the output is measured with respect to these
inputs of varying frequency, right.

So, this is more a signal processing viewpoint and more specifically, this is called the input-
output view and because we’re looking at signals with frequencies a natural aspect that arises
is a notion of bandwidth, alright and we know that bandwidth is very very important for
control systems, if your bandwidth of the system is fairly low, you cannot control the high-
speed dynamics, alright or you cannot analyze the high-speed dynamics and vice versa if the
bandwidth is too high you may start to bring in lots of noise and other aspects, alright. So, the
bandwidth is of course, dependent on the specific task and the problem which you are looking
at, but the bandwidth is again a very specific aspect of the signal processing viewpoint and
it’s a fundamental aspect in control systems.

The other nice aspect of control system of the transfer function approach is that we have these
algebraic properties of transfer functions. So, I can take multiple transfer functions, I can put
them in different say I can cascade them in parallel series, I can take the feedback loops, feed-
forward loop, so on and so forth. So, the idea is that I can actually combine different transfer
functions by using their algebraic properties, right. So, we have seen that you can just take 2
transfer functions in series and you can just combine them by the product of them or I can
take a transfer function I can do a feedback with another transfer function you know so on
and so forth and all these properties I’m able to combine these different blocks primarily
because of their algebraic nature, alright.

So, it allows us to do a lot of very useful things by combining different blocks from different
subsystems and we can build more and more complex systems, right and of course, as I said
before, this is very dominant in the industry you can take any advanced industry in the world
today transfer function modeling is a single most important aspect when you come to the
control of any system, alright. So, that is a good stuff.

920
Let’s see; what are some of the drawbacks of all the transfer function approach the first and
most important is they can be used only for linear systems, alright. So, we will go a little bit
about what we mean by linearity and so on in the next lecture, but for now, I’m assuming all
of us have a fair idea of what is linearity. So, transfer function approaches cannot be easily
used when we are looking at non-linear systems it can be done, but it is a little bit tricky.

Secondly, the other important thing is it is usually limited to single-input, single-output

systems right. So, I have got a system like this say a transfer function over here, I have got an
input and I have got an output right. So, it’s basically limited to these kind of systems. Now,
in general, we see that a lot of systems are of the multi-input, multi-output type. So, I have
got different inputs and for each input in the combination of a particular process in the plant I
have got different outputs right. So, y1 , y 2 , so on till y n ; all the way to Gn (s) .

So, I have got inputs r1 , r 2 , so until r n (s) . Now we know from the lectures that
Ramkrishna has already explained, we know how to handle this system, right, we also know
how to handle this system, we know how to handle every individual system. In this particular
say highly complex system which is an interconnection of many different systems, what we
still haven’t seen is when 2 systems are interconnected, alright. So, I will actually a give a
specific example, and then it is going to become clear.

So, let’s take I have got system 1 over here, system 2 over here, right, we know how to
analyze the stability properties, we know how to compute performance criteria, as all these
kinds of things for each of these individual systems. You take any complex system a chemical
process plant, aircraft, automated autonomous vehicles, so on and so forth; you never have
single systems, there are multiple subsystems and these subsystems are actually coupled to
each other, alright. So, you may have the engine control system and then you have; say, for
example, the automatic flight control system or in cars, you may have the ABS system or
advanced braking systems which we have.

Now, these are coupled systems. So, when we say coupled it means there is some kind of a
relationship between these 2 systems right and these are very tricky to handle when you have
multiple such subsystems and the number of such subsystems in a in advanced vehicle or a
chemical process plant can go to 100s, if not 1000s, right. So, for each one of this, we need to
keep on say modeling first in this, in the simple single input, single output approach, then we

921
need to start to look at what are the coupling effects between these 2 systems and the
coupling can actually be non-linear in many cases, right.

So, this becomes extremely cumbersome when you have multiple-input multiple-output
systems. Now, this doesn’t mean that people are not doing this. So, in the aircraft industry;
for example, you have extremely high order systems and there is really very complex amount
of say dynamics and controls which is required and it is all done in the classical approach,
right. So, you have beautiful transfer functions you have these couplings between the say
each of the transfer function loops which we’re looking at, and you can model all of this, the
problem with the reason why we had to do this in the good old days was primarily because
computing power was very limited, alright and we will see in a couple of slides why this
change in the computing power allowed us to advance in the notion of state-space and other
modern control say techniques.

So, the reason was because we had limited computing, everyone preferred to look at
individual loops, design the controllers for individual loops and then worry about the
coupling between individual loops, alright and then design another controller to take care of
the coupling those kind of things. So, this is; obviously, very tricky and it’s it can lead to
really messy designs and it is very difficult to really capture all the coupling effects. So on
and so forth. So, that’s another drawback.

The final drawback which is actually very important is that the internal behavior of a system
is not known remember that we just said that a transfer function is typically an input-output
view of a system, right. So, you give a signal over here you measure the output of the signal
and that’s all you are really concerned about. So, the transfer function then would be; just
give me a minute. So, the transfer function then would be something like you know the ratio
of the output to whatever is the input which is applied, right, what is missing is the internal
dynamics of the system and it can be a very complex system, right, but we have absolutely no
clue what is happening within the system, all we are concerned about in the transfer function
approach is simply the response to a given input that’s about it, alright.

922
(Refer Slide Time: 10:35)

So, these are the drawbacks, and let’s see how the introduction of this new concept of a state-
space system solves many of these issues, alright, but first of all, we need to see; what is the
state-space system. So, a state-space system or a state-space representation of a system is
basically a mathematical model that connects the inputs, the outputs and the internal variables
of the system called the states, right, it connects all of these 3 different parameters by first-
order differential equations, alright. So, as we drew in the previous slide.

So, as we drew in the previous slide say for example I have got the system, I have got
multiple inputs, I have got multiple outputs right just write it as r1 , r2 , r 3 , I have
got 2 outputs, but it doesn’t matter, I can have 10 inputs, hundred outputs, yeah so on and so
forth and then, there are multiple subsystems within this, right each of these subsystems is
important because it shows how the input is mapped onto the output and we’ll see a couple of
examples of this later. So, state-space representation of this system what it does is to basically
be able to take care of the inputs which I have all the inputs take care of all the outputs which
we have and all the internal variables which are actually hidden if you look at it purely from
an input-output point of view, alright.

So, and of course, the relationship between these 3 guys are essentially just first-order
differential equations and we will see why it is first-order differential equations in a couple of
slides. So, a classical example which I gave before which I briefly mentioned before was the
chemical processes. So, I can give a very short example of that one over here and there the

923
idea was that let me yeah. So, there the idea is that let’s say you have this chemical process
plant, there is a chemical process plant and you have multiple inputs, I have just drawn 3
inputs over here, but you can have more and you have a couple of outputs right although in a
real chemical process plant you would have at least 20 to different or 20 to 30 different inputs
and 100s of measurements which are the outputs basically.

So, now what could be the input? So, one input could be the steam, the volume of the steam
which is coming in from the boilers right the other input could be the pressure which the
compressors in the chemical process plant are applying to the particular liquid which is inside
this inside the chemical process. So, this is because of the compressors right and then maybe
we are having additional inputs like maintaining the temperature and so on, right. So, these
are 3 different inputs which are being given to your plant and maybe we want to maintain the
purity of a particular chemical that we are trying to manufacture say for example we want
99.9% pure ethylene right for example, ethylene is something that is used to make plastics
the plastics that you see over here and maybe you also want to measure the pressure and the
temperature and the volume because we don’t want these plants to go into a catastrophic
failure and blow up right.

So, these are the measurements over here all these are the measurements and these are the
inputs to the system and internally you have a lot of very complex dynamics going on right
and the state-space essentially manages to capture all these dynamics in one single
framework and we will see a couple of examples of this later. So, that is one very good thing
if you have to do this with the transfer function approach as I mentioned before you need to
model each path separately right then you need to worry about how these things are coupled
with each other and this becomes very complicated we will see in the state-space approach
why this is really very simple.

The second nice aspect of state-space systems is that we can actually isolate individual
subsystems. So, the keyword is this one we can actually isolate individual subsystems and or
say the components and we can actually say that these specific subsystems. So, to speak or
either controllable or uncontrollable we can say that observable or unobservable of course
when we say components we actually mean states we do not mean literally the component
itself, but the representation of the component, right. So, we are looking at a system where I
can model inputs, internal variables, outputs, I can even tell which specific internal variables

924
the states are controllable, uncontrollable, observable, and unobservable, so on and so forth,
alright.

So, we will see all this a little bit later and of course, the other important point is that we will
see that state-space is the linear version of state-space systems is entirely dominated by linear
algebra and there are excellent techniques and really good software tools for us to do analysis
and design, alright. So, to recall to summarize this slide, the basic idea is that a state-space
representation of a system is not is a mathematical model which captures your inputs, outputs
through these internal variables called the states and all of this relationship are basically first-
order differential equations.

(Refer Slide Time: 17:02)

So, let’s look at the general procedure to obtain a state-space system, I will derive one or 2
examples a little bit later, the basic idea is that we have seen in the transfer function approach
whenever you give in a physical system we write down the differential equations which
govern the dynamics right. So, for a simple mass-spring-damper, we have the second-order
dynamics for the RLC circuit, we have a second-order dynamics so on and so forth, right and
these dynamics are basically differential equations, right.

So, what we are going to actually do in state-space modeling is to basically take the
differential equation model of your physical system and then convert this Nth order
differential equation N could be 2, N could be 10, so on and so forth and we convert this Nth

925
order differential equation which essentially determines the dynamics of your system into N
first-order differential equations, alright that is the basic approach to state-space and so, how
many states we choose is basically one of the easiest ways to determine that is, its basically
the order of the differential equation which used to model the system with. So, if you have,
for example, an RLC circuit, it turns out the number of states we need are only 2 because the
differential equation which governs the dynamics is a second-order differential equation.
There is, of course, a different way to also decide how many states we need and the idea is
essentially this.

For an electrical system, we essentially say the numbers of states are basically equal to the
number of energy storing elements. So, take a RLC circuit you have the inductor and the
capacitor which are the energy storing elements and so, you actually have 2 over here, right if
you have 10 inductors and 10 capacitors you have a 20th order differential equation that is
basically the story and then you have 20 states, alright.

(Refer Slide Time: 19:09)

So, let’s actually go to a simple example and we will start off with the ubiquitous mass spring
damper system. So, we have as you can see here a mass m with the spring constant
attached to this wall on the left k . So, you have a mass m , it’s connected to this wall
over here by the spring constant k and then I have got a damper over here with the
damping ratio of b and I have applied an input force towards the mass in this direction
and so we are going to see what the dynamics are. Now we know by the basic force balance

926
equation. So, the applied force the total amount of applied force is distributed among each of
these 3 components.

So, by Newton’s laws, we know the component of force along the mass is basically mass
times acceleration, where x is the displacement or the position of the mass. So, we have
the component of force acting on the mass, then we have a component of the force which is
acting on the damping or rather the damping resists this component of the force and it’s

dx
basically a function of the velocity that is and of course, the damping ratio b and
dt
then the final component is this is a, the spring constant over here. So, the component of force
acting on the spring and the elastic force which it produces.

So, these are 3 components now we know from standard the transfer function approaches
which Ramkrishna has already taught the basic way to decide or to write down the transfer
function of this one is to convert this into the Laplace domain, right. So, we apply the Laplace
transform on either side. So, F goes to F( s) , then you have m s2 , s2 comes

2
d
because of the 2 , then you get the X (s) the displacement in the Laplace domain, the
dt
damping ratio is a constant, so it stays as it is, and the velocity gets a s term again
X (s) finally, plus you get a kX (s) , alright.

Now, if you remember the basic definition of a Laplace of a transfer function, it was nothing,
but the ratio of the measured output to the applied input, in this case, the applied force, alright

1
and there is nothing, but 2 and then we know that when we correlate this with
ms +bs+ k
the standard expression for the characteristic equation, we know that we can derive properties
of damping, the natural frequency. We can talk about stability, we can talk about
performance, so on and so forth; all the standard things, alright.

927
(Refer Slide Time: 22:14)

Now, let us see how we do this in the state-space approach. So, I will actually be doing this
derivation. So, the final state-space model is actually going to look like this and you can see
that we actually have and you can see that we actually have these 2 variables over here x́ 1 ,
x́ 2 these are actually the states of the system and we will see what these other terms
actually mean. So, let’s see how we can actually say derive this. So, if we recall the basic,

2
d x dx
derive the basic second-order differential equation. So, we had m +b
dt
2
dt

+kx , right.

Now, what I’m going to do is to introduce a new set of variables. So, let’s say a new choice of
variables and remember what I said that because we have a second-order differential
equation, the number of states will actually be equal to 2 . So, let us introduce 2 new
variables, we just call them x 1 you can call them whichever you like, but x 1 is a good
choice and x 2 , alright and specifically I’m going to assign x 1=x which is the basic
displacement and x2 to be the derivative of the displacement there is nothing, but the
velocity.

Now, with these new choice of variables let us see how I can actually write down the
dynamics and I can derive the; what I’m still calling is a state-space model which we’ve not
formally defined, but we will come to that later. So, let us see the relationship between x1

928
 d x1
and x2 and we will see from there. So, first, we see that is nothing, but x2 ,
dt

dx
right because if I take the derivative of x1 I get which is nothing, but x2 over
dt

d x2 d2 x
here, right and is nothing, but , right.
dt dt 2

2
d x
Now, this , we can rewrite it by using our second order differential equation and by
dt 2
dividing with this factor m everywhere, because I’m taking m onto the left-hand side.

F b dx k d2 x
So, this will simply become − − x . So, I have just rewritten my in this
m m dt m dt 2
particular form, alright. Now what I’m going to do, if you look at the first equation over here
you see that I have expressed x1 in terms of x 2 , what I would like to do now is to

express x 2 also in terms of x 1 and x 2 if possible.

d2 x
So, let’s see how to do that. So, I will actually do this here. So, we have this and you
dt 2

F
can see that I have got this term that’s obviously constants or the independent input,
m
they do not depend on the states, alright. So, that stays as it is and then I have got this

b dx dx
. Now if you look at what is this it’s nothing, but x 2 , right? So, I will just
m dt dt

b dx
write this as x , x 2 is nothing, but .
m 2 dt

k
Similarly, the last term I have over here is which is a constant it does not depend on the
m
state. So, that will stay as it is and then I have the term x (t) , x (t) is nothing, but
x 1 . So, this simply becomes x 1 . So, this is my equation 2, this is my equation 1, if I

dx
combine equation 1 and 2; I get exactly the state-space system and what you see is a
dt

d x2
is related to the state x 2 whereas, is related to both x 1 and x 2 , but of course,
dt

929
this is specific to this, specific example we have considered this need not be true in general;
however, the first row in the 2 X 2 states matrix will always be 0 and 1 and we will
see the structure repeating even for a Nth order state-space system and the inputs are always
treated separately.

So, you can think these are the input this is if you look carefully these are the actual dynamics
of the system. So, here is where you have your accelerations velocities and so on and these
are the states. Now let us say for this particular system, I also want to measure the output
because without measurement you actually cannot do control analysis any such thing. So, let
us say, I want to measure this as the output, now what are the 2 states which I have? As we
have seen the two states are x 1 and x2 , x 1 is nothing, but the position; x 2 is the
velocity now which you want to measure is entirely up to you depending on what problem
you are you are looking at let’s say I want to measure the position maybe I have got a say
position sensor in one case, in other case if you want to measure the velocity I may have an
odometer or any or one of these things.

So, if you want to measure the position I call my measurement as y and remember what is
accessible to me is x1 and x 2 , assume I can measure only the position or I want to
measure only the position. So, I write this in the form of this, alright. So, this is my
measurement equation, this is my measurement equation and this is my actual state-space
matrix as well as the inputs, alright. Now let’s actually go ahead, we will come back to this in
a slightly more complex example, we will also look at the basic theory of how to write a Nth
order differential equation in form of in the form of state-space systems, but the basic idea is
always this you start off with the basic differential equation, you make a new choice of
variables and the number of such variables are; is basically equal to the number of is equal to
the order of a differential equation, alright. So, in this case, I had second-order differential
equation.

So, I’m taking 2 variables; these variable I have call the states, once you make a choice of
these 2 states the idea is to relate the two the states with each other, alright and then you
essentially start to get these dynamics, alright and finally, of course, for as I said for analysis
for control all these things we need to make specific measurements. So, then you introduce a
measurement matrix as well and you get this structure.

930
(Refer Slide Time: 29:49)

Now we will actually see the more formal version of this. So, so this was a mass-spring-
damper; let’s actually see a very simple simulation of this one. So, let us say, I take mass,
m=1000 Kg , the spring constant, k =2000 N m
−1
and a damping ratio,
b=1000 N /ms
−1
.

Now, remember my A matrix, remember this particular matrix over here

[ 0 1
−k /m −b/m ] . So, it’s actually called the A matrix and we will see why it is called

the A matrix a little bit later. So, when I substitute the parameters k , m and b in

this matrix over here, essentially what you are going to get is this one; [ 0 1
−2 −1 ] and

then if you look at these inputs over here, I have a 1/ m for the moment let us assume F is
equal to a unit step force.

So, F=1 and 1/m which just becomes [ ]0

0.001
. So, this is the A , this is the

B and we will see what these things are later and if I do a simulation of this, I will run
through the code a little bit later, we do a simulation of this, we see a step response of the
following type and you essentially see that this is a damped stable response and with a very
small overshoot over here and we have seen from the transfer function modeling of a mass-
spring-damper that when you have this kind of a response, we can sort of guess; what kind of

931
poles that the corresponding transfer function would have. So, for this kind of response, it is
very easy to see that the poles would be like this.

So, you have a complex pair of poles and it’s in the left half-plane, right which is why you
have the stable and the damped response. Now, of course, we are in this so called state-space
version where the notions of poles don’t really hold, but we, of course, want to relate how the
system response comes with the notion of poles as well in the context of state-space systems.
So, one of the ways to do that is you have this matrix A over here. So, we see this is a
square matrix, and one of the natural questions that come in we see a square matrix is what
are the Eigenvalues of A .

So, let’s compute the Eigenvalues of A and it just follows the standard procedure you
compute the | A−SI | and when you do this, you see that the Eigenvalues are essentially

1
this and for the same system if you look at the transfer function which is 2 and
ms +bs+ k
you compute the poles of this transfer function, you will get exactly the same values as the
Eigenvalue. So, you get −0.5± j 1.3229 , alright.

(Refer Slide Time: 31:52)

So, in this particular case, we are seeing that the poles of the system are exactly equal to the
Eigenvalues of A . We will see this a little bit more formally why this is true, but this is
very interesting because it shows that this new modeling approach in the state-space system

932
at least with respect to say it the dynamics of this kind exactly corresponds to what we see
from the transfer function modeling approach. So, once we are done with this let us we have
talked about states of a system, right. So, for the mass-spring-damper, we have said that the
state one of the states is a position, other state is a velocity and we have done all these
derivations over here, we have this model, we have a simulation as well so on and so forth,
but we formally not actually defined what the state of a system is.

(Refer Slide Time: 33:37)

So, the definition here is it is not a very polished definition, but this actually conveys a
correct picture, a variable is called the state of a system because the future output of the
system depends entirely on the current state value and whatever input we are going to give,
alright. So, you have a system I modeled it by the so called states of the system and the
output of the system the future output the system depends entirely on the current state and
whatever input I give in the future.

So, I do not need to worry about what happened in the past, I will need to worry about the
current state and in other words, it also means the future output, in other words, it also means
the future output depends on the past input only through the state and not through any other
parameter, of course, assuming you have done a good job at the modeling and you have not
left out any variables or you do not have un-modeled dynamics. So, as long as you take care
of that the future output depends only on the past input through the state, alright.

933
So, if I put these 2 things together in effect this means that the state summarizes the effects of
all the past inputs on your system on the future outputs. So, the state actually acts like a
memory of the system. So, take, for example, simple, so, cute where I have a capacitor and
I’m applying a voltage across this capacitor, right. So, I look at the voltage drop across the
capacitor and I’m interested in the dynamics. So, we know that the current through the
capacitor is basically modeled like this right.

Now, what I’m saying is if this Vc which is a voltage drop across the capacitor if that is
the so called state of the system, this particular variable I’m saying that at time t=0 or
any specific time you fix, we can just say it to be equal to 0 at a time t=0 if you have
a certain value of the state at any time in the future V c (t) will depend on whatever
information you have in the past. So, the current value of the state and whatever input I’m
going to give afterwards. So, for example, at t=1 as long as I know; what is the value of
the state at time t=0 and I know the kind of input, I’m going to give at time t=1 , I
can always calculate V c (t) , I don’t need to know what happened one hour before, I don’t
need to know what happened one day before as long as I have the total information of the
state as of now, alright.

(Refer Slide Time: 36:54)

So, that is essentially it's loosely called as a memory of the system, alright now we have
talked about what is a state we have done a little bit example with the mass-spring-damper we
have seen that we have the so called input structure over here inputs over here, we have the

934
measurements over here and we have the so, called state dynamics as I call them let’s see this
as a formal definition. So, we have seen the, an example with the mass-spring-damper where
we have looked at the all the inputs which are given to the system, we have looked at the
measured outputs we have seen how to write down the basic equation for that one, we have
looked at what I have called as a dynamics of the system, though I have not really explained
that. We will come to that and of course, I have talked about the states, now what is the
general form of a state-space model and specifically a linear state-space model.

So, specifically, a linear time-invariant state-space model and it’s written in the following
way. So, x́ ( t )= Ax ( t )+ Bu(t ) , y ( t )=C x ( t ) + Du (t) remember all the variables are
functions of time x and u and we are assuming that this is a time-invariant system. So,
the matrices A , B , C , and D are actually constant for now in general, this need
not be true. So, let’s look at this particular equation the variable x is actually called as the
state of the system and we had a definition of that is basically the memory, it captures what
all has happened before. The variable y is the measured output and the measure output is
really your decision as well as it depends on the constraints of the system, there are some
states which we can measure some states which we cannot measure and in general you may
not even be able to measure states, alright.

But for this particular lecture, we will assume that y are the measured states, alright and
u over here is the input which we give to the system and u need not be a single input it
can be multiple inputs we are looking at state-space so multiple inputs are permitted. Now we
have these parameters A , B , C , and D ; A is basically the, there are different
words for this I can call it as a dynamics matrix or the state matrix or a much better word
would be the interconnection matrix, I will explain what this is. So, the interconnection
matrix basically tells how different states of my system are actually connected with each
other or related with each other, or I can say coupled with each other. So, if you looked at the
previous example the state x1 which was a position, is connected to the velocity through
this parameter over here right.

So, x1 is connected to x2 through the parameter 1 and x́ 2 is connected to the

states of the system through these 2 parameters, these are the interconnection parameters as
we call them and they define the dynamics; they define how different states are connected to
each other and so and interconnection structures are very important although we won’t be

935
dealing with that over here, but for us, integration structures here are basically the elements
or the values which come up when we connect states, alright then you have the matrix B .
Now, okay before I go ahead, please note the dimensions of the matrices are very important.

So, let’s say, we are looking at the Nth order state-space system; Nth order state-space system
and remember that you get a Nth order state-space system typically from a Nth order
differential equation, alright. So, if you have a Nth order state-space system the matrix a will
be of dimension n cross n, it cannot be more it is a square matrix of dimension n X n now
let and now let’s look at the matrix B . So, B is a matrix as you can see from here B
is a matrix which takes the inputs which are applied to the system and relates them to the
state. So, it basically tells how my inputs are actually connected to my states in the case of the
mass-spring-damper example which we took before the input matrix over here essentially
says that whatever force have applied does not directly go to state number 1 because I have
got a 0 in that entry over here, but this force directly goes to state number 2 which is the
velocity by the way through the factor 1/m . So, what B captures basically is how the
inputs actually affect the states.

So, we actually call B as the input matrix, next we have this matrix C and C can be
called as the output matrix, we have already seen in the case of the mass-spring-damper, we
had an example of when we wanted to measure only the position. So, we had an output
matrix like this where x1 and x2 is the state vector, and this particular row over here
was actually the C . So, so C is the output matrix. we are not going to be say dealing
with the matrix D over here D is actually called a feed through matrix the feed-
forward matrix and we will not consider that over here.

So, let me just recall this slide. So, we have this the general form of the linear time-invariant
state-space model A , B , C , D are constant matrices all the others are functions
of time. The A is the interconnection matrix, it tells how different states are coupled to
each other, B is the input matrix it tells how my inputs affect states which input affects
which state and with what magnitude, C is the output matrix it depends what I want to
measure and D is a feed-forward matrix. So, A is nXn matrix the interconnection
structure it is always equal to the number of states which we have, B is the input matrix.

Let’s assume we have m inputs; m≤ n . So, B will actually be nXm matrix and
let’s look at the output matrix C , let’s say we are measuring p states in which case,

936
 C will be a pXn matrix, now in general; we could actually measure all the states if
you have access to all the states in that case p=n . So, we have an n X n measurement
output matrix finally, we have the matrix D the feed-forward you know matrix which is
basically p X m . So, these are the 4 matrices which we are going to consider we will be
ignoring D for the rest of this lecture we are going to look at basically A , B and
C , alright.

So, this is a general form of the linear time-invariant system. Now we have always said that
we assuming we have a Nth order say differential equation how do we go from that to the Nth
order state-space system.

(Refer Slide Time: 44:59)

And let’s look at the general derivation over here. So, consider an Nth order differential
equation which actually models a linear time-invariant system. So, we have the differential
equation over here. So, u is the input, y is the output in this particular case. So, we
have these coefficients here which come about when you are modeling the system.

Now, what we are going to do here is as we have done in the mass-spring-damper case where
we had the second-order differential equation and we introduced 2 new variables x1 and
x 2 what we are going to do here is to introduce n new variables because we have a Nth
order differential equation we would expect to have n states. So, the way you would do
that is fairly straightforward. So, let’s say I have got a; I introduced a new variable called

937
 x1 , x2 all the way to x n , alright and again we have n because the order of the
differential equation is n , alright.

So, x1 I will relate it to the first variable over here y , alright x 2 I’m actually going

dy dy
to call it as and so on; now note that yeah. So, so I have x 1= y , x 2= and
dt dt

2
d y
x 3 in a similar way x 3= 2 , I can go on all the way to n and basically what I
dt

d n−1 y
would get is . So, this is exactly the same as a mass-spring-damper example where
d t n−1

dx
we had written x 1=x and x 2= exactly the same as that except I have got n states
dt
over here.

So, I have got the slightly longer looking expression, but is exactly the same, alright. Now as
we did in the case of the mass-spring-damper, we will now look at the derivatives of these
states and we will see how these states are actually coupled with each other. So, if I take

d x1 dy dy
we see that its nothing, but and as you can see is nothing, but x2 .
dt dt dt

d x2 d2 y d2 y
So, I get x2 over here, similarly, is nothing, but and is nothing
dt dt2 dt2
but x3 , x 3 I can go all the way here.

d xn
So, I can go will be equal to well this would depend on all the states which we have
dt

d xn
over here, alright. So, let’s see how to do that. So, is nothing, but if you look at this
dt

n
d y
particular expression over here is nothing, but n , alright and how do we derive this
dt
particular thing we take a look at our basic differential equation over here where I have

n
d y
n the Nth order derivative of y and we rewrite this expression and the way you
dt

938
rewrite that as you have done for the mass-spring example before would simply be u(t)
and take all these terms over here on to the right-hand side.

So, that becomes u ( t ) −an−1 then what do I do with this expression over here, I replace it
with x n , alright this particular term is nothing, but x n . So, it becomes
−an−1 x n−an−2 x n−1−… … … and you can go on like this and then you have this expression

dy dy
over here a1 I take it on the right-hand side. So, I get −a1 ; is nothing,
dt dt
but x2 , −a1 x 2 and then finally, this expression a0 y . So, take it on the right-hand
side, I get a0 x 1 , alright.

So, basically what I have done is exactly the same as a mass-spring-damper where I have
written all the states in terms of the other states. So, we essentially have now the coupling
between the states the interconnection matrix and when I put all of this in a matrix form we
want to know how that looks like in the case of the mass-spring-damper if you remember we

had x́ 1 , x́ 2 it was nothing, but

[ 0 1
−k / m −b/m ] , alright times []
x1
x2
.

Now, we are doing exactly the same thing except that we want to see how this looks on the
Nth-order system and it looks exactly like this.

(Refer Slide Time: 50:17)

939
So, you see that these first set of rows, okay the n−1 rows which I’m actually showing
you over here the n−1 rows are basically all the velocities which are related to the
previous one. So, x́ 1 would be nothing but x2 , x́ 2 would be nothing, but x 3 so on
and so forth same as the mass-spring-damper case, we have done. The final expression the

d xn
it basically comes from this guy over here.
dt

So, this expression I have rewritten as a last row of this matrix, and of course, we have an
independent variable here the u(t) , it is not a state variable it is an independent one. So,
we take that outside. So, this is basically my B u (t) this is my A , x (t) and of
course, this is x́ (t) . So, x́ ( t )= Ax ( t )+ Bu(t ) there is a standard expression for a Nth
order differential equation. So, this is for a Nth-order system and of course, as before we
want to measure the states now I can again decide which particular state I want. So, I’m just
calling ⃗x as a state vector which is nothing, but x1 , x2 all the way to xn and in
this particular case, again I’m measuring the first state, I could also measure the third state for
example, in which case I would have a matrix like this.

So, if I measure the third state, I will just have a matrix like this ⃗x or I can even have a
combination of measurements. So, I can measure state 2 and maybe state 3 and 4. So, I would
have something like this that’s really up to you how you can measure depends on the sensors
you have the access to these states so on and so forth. So, this is the basic state-space model
of a Nth order differential equation and I have just shown you how we can derive this one and
the logic for this basically follows the same process which we did for the simple two cross
two mass spring damper nothing more and nothing less, alright.

940
(Refer Slide Time: 52:54)

Now, let us work out a simple example. So, we have this. So, we have already done the
second-order state-space model for a mass-spring-damper we have seen the Nth-order
derivation for to derive a Nth order state-space system let us look at this simple third-order
differential equation and see how to did how to derive the state-space model, alright. So, what
we are going to do is to do what we have just done for the previous case since it’s a third-
order differential equation, we are going to define 3 new variables x1 , x2 and x3 ,
alright and as we have done before we relate the first variable to y , the second variable to

dy d2 y
and the third one to , alright.
dt dt2

Let me write this here and the first thing which we always try to do is to find out the
relationship between these states essentially to compute the interconnection matrix. So, for

dy
that, we take the derivative of the first state and we find that it’s is nothing, but and
dt

dy
is basically x 2 . So, the derivative of the first state is x 2 , the derivative of the
dt

d2 y d2 y
second state so x́ 2 is nothing, but and is x3 .
dt2 dt2

941
 3 3
d y d y
Finally, we have x́ 3 which is 3 and 3 we can actually derive from this
dt dt
expression, alright. So, what we do, should take all these terms onto the right-hand side and

d3 y
then divide with the factor 2 that is going to give and so, let us actually do that. So,
dt3

3
d y
we get is remember I’m dividing all the terms with the factor of 2.
dt3

10 −4 d 2 y d2 y
So, it is u(t)=5u (t) , this term is . So, −2 and then a
2 2 d t2 dt2

−6 dy dy −8
becomes −3 and finally, y becomes −4 y and what we do now
2 dt dt 2

dy
is to replace each of these y ’s or the with appropriate state. So, we know that y
dt

2
dy d y
, as we have defined here is x1 , is nothing, but x 2 and is x3 .
dt dt
2

So, this equation is nothing, but 5 u (t )−2 x 3−¿ 3 x2 −¿ 4 x1 that is a final the last
row of the state-space model and that is exactly what we get over here and the first 2 rows in
this case as we have seen will always be 0 1 0 0 0 1 and now we have the inputs in the final
matrix over here in this form and of course, again for no particular reason I’m considering,
that I measuring the first state that need not be true, I can measure the second state or I can
measure a combination of states any such thing that is really up to you.

942
(Refer Slide Time: 56:28)

So, let us take another example we will use a parallel RLC circuit and we will try to derive
the state-space model. So, actually, I have shown the final derivation over here. So, let’s see
what the actual circuit looks like. So, let us suppose we have a current source Is and we
have connected this parallel with the resistor branch, an inductor branch, and a capacitor
branch, and let’s say we measure the voltage across the capacitor and call it Vc or V out
as you like. So, this is C, this is L and this R and I will further denote the current through the
resistor being as I R , through the inductor being IL and through the capacitor being
Ic and now what we want to do is to write down the state-space model for this one and
you can see that I have already considered the 2 states as being the current through the
inductor and the voltage across the capacitor and again the reason for that is because this
particular RLC circuit has 2 energy storing elements.

One is the inductor, one is a capacitor hence you will have 2 states and not more than that.
So, how are we going to write the equations for this one? So, first what we are going to do is
to focus on Vc and then on I L . So, the way that we would do it is if you look at the
current through the inductor we know that the voltage drop across the inductor is given by the

d IL
constitutive relationship which is basically L is nothing, but V the voltage
dt
drop across that one and specifically which we I can write it as Vc itself because the
voltage drop across the capacitor over here is the same as the voltage drop across the inductor

943
which is in parallel with the capacitor. So, this is your equation number one when I take the

1
term L on to the right-hand side. So, that is Vc .
L

So, this is actually the first state equation to write the second state equation we will employ
Kirchhoff’s current law at this particular node by observing that the expression for the current
through the capacitor is nothing, but the total current which is coming in minus the current
which is going through the resistance branch minus the current which is going through the
inductance branch. Now we see that this we have I L in this expression.

So, I do not want to touch IL because we want the interconnection matrix between the
voltage and across the capacitor and the current through the inductor to contain both terms
what I will do is to replace the other terms if possible Is of course, will stay as it is
because it is an independent source, for IR that is the current flowing through this
resistance branch we know that it is nothing, but if you consider the voltage drop across this
to again be V c because all these branches are in parallel.

Vc
So, IR is nothing, but , IL of course, stays as it is, what about Ic the current
R

dV c
through the capacitor as you know is given by C . So, this is equation number 2 and if
dt
you plug these 2 equations into a matrix form, let me erase this one for one minute. So, if you
plug these 2 equations in a matrix form you are actually going to get this expression over

d IL
here. So, let me show how we can do that. So, that’s the derivative of the first state,
dt

dVc
derivative of the second state and please see here this term C the capacitance I
dt
will remove it from here and divide it on the right-hand side everywhere.

1 V −1
So, that is I s− c I and now we write down the A matrix and then the
C CR C L
state vector here is nothing, but just IL and Vc and I just have to fill in the elements of

d IL 1
the matrix. So, is only a function of Vc with the and the other 2 and for
dt L

944
 dVc
, you simply just fill these 2 terms over here. So, this is your A matrix. It is the
dt
same as this expression over here, this is A , this is your B matrix. This is the input,
alright.

So, this is how the final state-space model will look like let’s say that you want to make a
measurement of both the current flowing through the inductor as well as the voltage drop
across a capacitor, in that case, your y=Cx will have the following form will be [1 1]
and you will have I L and V c assuming you want to make both measurements, alright.

(Refer Slide Time: 62:18)

So, let’s look at the next example which is again electric circuit. So, we have two sources
Vi and Is and we have got basically 2 resistors, one inductor, and 2 capacitors right.
So, let’s see what is a state equation of this one and we are going to solve this again in the
same way as a previous example we see that we have three energy storing elements there are
2 capacitors and one inductor, alright and the voltage through capacitor one is written as
V c 1 , the voltage to the second capacitor is Vc2 those are the two states for the
capacitors and the current through the inductor is written as IL .

So, the states for this particular system will be V c 1 , V c 2 and I L and so let us actually
try to write down the dynamics of each of these variables the, the time derivative of this one.
So, let’s start off and we can analyze this circuit in two ways in two steps actually. So, the

945
first step is to basically write down the nodal analysis at these two particular nodes. So,
basically, we are going to write the KCL at these 2, and in the second step, we are going to
apply a KVL in this particular loop, alright.

So, first, let’s look at node one and we will write the kickoff current law at node one, and the
way, we can do this is we are going to assume all the currents coming in are positive all the
currents leaving that particular node are negative. So, they sum up to 0 that is a basic idea.

V c1−V i
So, at node one, we have by Kirchhoff’s current law. So, that’s the current which
4
is over here and then we have the current which is leaving this particular branch here and that
current is going to be given as V c 1 .

So, this current through this I’m going to write as by using this constitutive relationship

dV 1 dVc1
I=C . So, it’s basically going to be C=0.25 that is that’s a current
dt 4 dt
in this particular branch, right and a current in this particular branch which is going out is is

dVc1
just I L this is equal to 0 . So, if I rewrite this expression, I will actually get is
dt
nothing, but −4 I L that’s a term on the right-hand side −V c1 +V i , alright. So, this is
equation number 1. Now, this is the KCL applied at node 1, if I apply KCL at node 2.

I’m actually good at this particular node, I have the current IL which is coming in and I
have this current which is going out there is a current in this particular branch going out and
this Is is also coming in. So, I have I L+ I s the 2 incoming currents, I need to know

dVc
your minus the 2 outgoing currents. So, the first one will be again C , again I
dt

1
apply this constitutive relationship over here. So, I will basically get 0.5 which is
2

dVc2
that’s my C , that is a current in this branch over here and the current in this
dt

V c2
branch is nothing, but the current in this in this branch is simply , alright.
1

946
 dVc dVc2
So, if I now rewrite this in terms of my . So, I get is basically, I multiply
dt dt
with 2 everywhere, if you take all the terms on the right-hand side and then multiply with
the negative sign. So, you would actually get −2 V c 2+ 2 I L +2 I s right. So, this would be
my equation number 2. Now if I look at the loop equation, that’s this particular loop over here
with the ground at the negative potential these 2 as being positive defined over here.

So, by applying KVL, I would actually get V C 1 it is −V c1 because of the polarity plus
the voltage drop across the inductor which by the constitutive relation which I have written

dI d IL
over here V =L . So, it becomes −V c1 +2 +V c 2=0 , V c 2 is a voltage drop
dt dt
across this capacitor, right. So, plus Vc2 is equal to 0. If we rewrite this in terms of the

d IL
derivative expressions the states basically. So, I get remember I need to divide with 2,
dt
when I take it on the other side.

So, it becomes 0.5 V c 1−0.5 V c 2 . So, I had 3 states which I have defined. I now have the 3
first-order differential equations for each of these states. So, I simply put them together, I
have not yet talked about measurement, we will see the output, we will see that later. So,

dVc1 dVc2 d IL
when I put all of this together we get , , and . So, this is the state
dt dt dt
vector x́ , this will be related by the A matrix times the states itself. So, Vc1 , Vc2 ,

and I L and now this relationship is basically the ones which we have derived over here.

So, let’s look at equation number 1 on the right-hand side and we just plug in the terms which
you have computed. So, it is −4 I L . So, I get a −4 and a −V c1 , so –1 ; Vc2

is not there, so it’s 0 , we will see what to do about the Vi a little bit later. Now for

dVc2
, we see that we have a −2 V c 2 and we have a plus 2I L and Vc1 does not
dt

d IL
come over there, it is a 0 finally, for , it depends entirely on V c1 and Vc2
dt
and it's basically 0.5 and −0.5 , this is 0 .

947
Now, we still have to plug in the inputs because we have a Vi and Is over here. So,
what we are going to do with that is we have 2 inputs Vi and Is and what we would
need is a 3 X 2 matrix over here because this is a 3X3 this will be a 3X2 matrix,

dVc
alright and we see that from equation 1, it depends on V i . So, I have a 1 over
dt

dVc2
here, it does not depend on Is so, this is 0 . So, that’s equation number 2 that
dt
depends on I s and you have the constant being as 2 .

So, this is 2 , this one 0 , I L does not depend on d s, I L does not depend on Is
as you can see over here. So, this is all 0 s. So, this is the basic state-space model for this
electrical network. This is the A matrix. This is the B matrix. This is u(t) and this
is x (t) , right and the measurements again are up to you; you may want to measure the
current flowing through this inductor in which case my measurement matrix will be [0 01]
, my C matrix will be [ 0 0 1 ] or you may want to measure the current across in other 2
capacitors. So, the C matrix will change appropriately, alright.

(Refer Slide Time: 70:44)

So, the entire derivation; the final state-space matrixes are actually given over here.

948
(Refer Slide Time: 70:52)

Let us look at one more problem which is on a couple mass-spring-damper. So, we have 2
masses and they are coupled by the spring, with the spring constant K , mass 2( M 2 ) is
pulled in this direction by a force F and mass 1( M 1 ) is connected to the wall with this
damping with a damper D , alright and mass 1( M 1 ) it has a displacement x 1 , mass
2( M 2 ) has a displacement x2 and we are just going to assume that there is no friction
between the mass and the ground over here, alright.

(Refer Slide Time: 71:29)

949
So, there is no friction between these two, but you can easily incorporate that, that’s not a
problem. So, if you look at the; so, the basic way that we always write the differential
equations or the dynamics for a mass-spring-damper is always look at the force balance
equations, alright. So, the applied force will have to be distributed across the individual
components which are there. So, if you look at M 2 first. So, I’m applying a force F , it
resolves into two components one is an acceleration component on M2 .

So, that is mass times acceleration over here and then the spring which is coupled between
M2 and M1 and you have already seen in the previous lectures; how to write these
dynamics. So, it is basically K (x 2−x 1) that’s the displacement between the two things
between the 2 masses. So, this is the second-order differential equation for M 2 . In a
similar way, we write the equation for M 1 , note that the force F ; it does not directly
act on M 1 , it acts actually through M 2 . So, term on the right-hand side is 0 and
because we have these internal forces for M1 experiences an acceleration which is

d2 x1 d x1
and then you have the damping which acts in terms of its velocity and K
dt2 dt
is again couple between M 1 and M 2 . So, we have K ( x 1−x 2) , alright.

So, as we did before, we are going to define the basic we want to write down the state-space
model of the system. So, the first thing we need to do is to identify the states, right. So, the
way that we are going to do as we did for the original mass-spring-damper system. So, we
identify x1 is equal to basically the position x1 itself. So, x 1 ( t )=x 1 itself then we
have a second state which I can call as x 2 , but I can also call it as V 1 , alright, it’s the

d x1
velocity. So, I call it as , these 2 states are for mass M1 .
dt

Similarly, I have another state x 2 which is the displacement of M 2 . So, I’m calling it x
2 itself and sorry this is x 3=x 2 and then I have a state x4 which I’m going to call as

d x2
basically V 2 the velocity of M 2 . So, this is nothing, but .
dt

Now, this notation is a little bit confusing. So, what we are going to do is to simplify this a
little bit we will remove this x2 over here. So, we have x1 and V1 as the states of

950
the first mass then we will call this as x as x2 and V 2 , right otherwise everything
else stays exactly the same. So, these are the states for mass M 2 . Now all you need to do

d x1
since we know these already by the relation is nothing, but V 1 . So, one state is
dt

d x2
taken care of the second state is also taken care of by this expression and what we are
dt
left with is to actually to compute the time derivatives of these 2 states, and then you will get
the interconnection matrix.

(Refer Slide Time: 74:58)

dV1 dV2
So, if you go through the steps, you will actually see that and are
dt dt
essentially derived by using these expressions over here, right. So, we saw that V1 is if

2
d x1 dV1 d x1
you take V 1=¿ and now I’m looking at . So, it’s nothing, but 2
dt dt dt
over here and I can rewrite this expression as we have done before and then derive this these

dV1
terms over here for . In a similar way, we can do you can derive the expressions for
dt

dV2
, alright.
dt

951
So, we get these terms over here, put everything together in exactly the same way as we have
done before, remember that the states where x1 , V 1 , x 2 , V 2 and then you get this
A matrix over here and the B matrix, alright exactly the same way as we have done
before.

(Refer Slide Time: 75:59)

So, this example for the coupled mass-spring-damper; so, we have seen you know in a
previous example for the simple mass-spring-damper that the Eigenvalues of the A matrix
are exactly the same as the poles of the corresponding transfer function, right. Now let us see
if this property holds in general.

But before we do that; we need to actually establish the relationship between the state-space
and the transfer functions or how? So, given a transfer function model; how do you compute
a state-space model or vice versa given a state-space model, how do you compute the transfer
function model, right. So, this is a very simple derivation to show that. So, remember our
basic equation was x́= Ax+ Bu and ý=C x + D u , right, this is a general form of the
basic state-space system.

dx d
Now, when I convert this into the Laplace domain; so, , is basically the s
dt dt
term in the Laplace domain. So, you get s X ( s) and then you get A X (s) , remember
that A is a constant matrix. So, it just stays as it is as is for B , C and D . So, this
is the first term of the Laplace expression, right that’s what I have written over here. Now you

952
take this term on to the left-hand side and you subtract and you are going to get these terms.
So, ( s I − A ) X ( s )=BU (s ) , right, it is confusing, I can just write this down. So,
sX (s )− AX(s)=BU ( s) take the X (s) term common out and all you are going to get is
( s I − A ) X ( s )=BU (s ) .

Now, so, this is we can call this as equation 1. Now we have a second equation
y=Cx+ Du , write this in the Laplace domain. So, this is the Laplace equivalent of
Y ( s)=CX ( s)+ DU ( s) that is this over here. Now what we are going to do is to take this
X (s) which we have computed over here. X ( s )=( s I − A )−1 BU ( s) , right. So, I can

write this 1 implies X ( s )=( sI − A )−1 BU (s) , right. So, I take this X ( s) and I substitute
this, there is a small correction to be made in the slide, this is not U of x, they should be
U ( s) , alright.

Let’s continue. So, we have that X ( s) is equal to this expression, we substitute this back
into this term over here and then we derive this one, alright. So,
−1
Y ( s )=CX ( s )+ DU ( s )=C ( sI− A ) BU ( s )+ DU ( s) . Now we know that the transfer
function is defined as G( s) is equal to the output by the input. So, take the U terms
common out of these 2 terms, alright and then and then divide with U ( s) and you are
going to get this. So, this is the basic relationship between the transfer function and the state-
space model in this particular case, we are assuming A , B , C and D matrices
are given to you and the goal is to see how we can go and derive a transfer function and this
is the basic expression for that.

953
(Refer Slide Time: 79:41)

So, let’s actually derive a; solve a very simple problem I won’t go through the details the all
the details are there in the slide. So, we have x́= Ax+ Bu and then y=Cx . So, this is
the A matrix, this is the B matrix and this is the C matrix and the objective here is
to actually compute the transfer function given the state-space model and we know that it’s
given by this expression which we have just derived before, right. So, all we need to do is to
plug in all these terms and so, for s I minus.

(Refer Slide Time: 80:17)

954
So, let me write down the expression again here equal to C ( sI− A )−1 BU ( s ) + DU (s ) ,
alright. So, we need to first compute ( sI −A ) and compute the inverse of that. So,
( sI −A ) is basically this over here. So, that is basically sI is basically this sI ,
−A matrix was just this one, alright. So, when you actually subtract this, we actually get
this ( sI −A ) , then you compute the inverse of that. So, ( sI −A )−1 is basically the adjoint

adj ( sI− A )
of , if you remember the basic expression of inverse.
|sI −A|

Let us take our matrix, let me just give a different notation I will call it as matrix M . So,

adj ( M )
the matrix M is actually computed by taking the , alright. So, that’s exactly
|M|
what we have done here adjoint basically means these two terms are swapped and the sign of
these two terms are reversed that is exactly what is happened and we compute the
determinant here, alright. So, this is ( sI −A )−1 . Now to ( sI −A )−1 ; we need to multiply
with we need to pre multiply with C and post multiply with B , alright.

(Refer Slide Time: 81:43)

So, when you do that. So, this is the expression over here. So, you pre multiply with C
and post multiply with B and when you do that remember in our particular example, we
had D over here is equal to 0 . So, the D term does not exist. So, this doesn’t matter
the D term here really doesn’t matter for us, we can ignore that and what we get is

955
 −1
C ( sI− A ) BU ( s ) , you do all the multiplication and you are going to get this transfer
function, right. So, this is a simple example to show given a state-space model how to derive
the transfer function.

Now, it’s important to remember that you may have made a mistake in the derivation; the
simple way to double-check is to compute the poles of this transfer function right. So, there is

−b ± √ b2−4 ac
a quadratic equation. So, you just use and you compute the poles in this
2a
particular case, I think it will be −2 and −3 , these are the poles and to double-check
that these poles are actually correct and that we have obtained these poles from the correct
state-space expression, what you can actually do is to compute the Eigenvalues of the matrix
A and I will explain why that is true we already saw a simple example at the mass-spring-
damper that this was exactly the same. So, if you compute the Eigenvalues of the matrix A
, it turns out you will also get −2 and −3 , alright.

(Refer Slide Time: 83:30)

Now, let’s see why this is actually true. So, the Eigenvalues of A are and the relationship
to the poles of the transfer function and it turns out as you can conclude in this last line the
roots of the characteristic equation of a transfer function which are the poles are exactly the
same as the Eigenvalues of the A matrix of the equivalent state-space model. So, let’s the
derivation is fairly simple. So, we have our expression which is over here. So,
Y ( s )=C ( sI− A )−1 BU ( s ) +DU ( s) , alright. Now let me just rewrite this remembering that

956
the inverse of a matrix is the adjoint by the determinant. So, that’s what I have written over

adj ( sI− A )
here. So, this becomes C , alright, you multiply with BU ( s ) and add the
|sI− A|
DU ( s ) over here.

So, this entire thing is equal to Y ( s ) . Now just do a simplification over here. So, this

C [ adj ( sI − A ) ] BU ( s )+|( sI− A)| DU (s)

becomes , alright. Now you see in the numerator
|sI −A|
you again have this U term as constant in terms, of course, a function of s . So, take
U constant or take U as a common factor. So, and then if you divide by U on the

Y ( s)
left-hand side you would actually get on the left-hand side and that is equal to this
U ( s)
expression over here, alright.

Y ( s)
So, this was basically when I take U as a common factor in the numerator and
U ( s)
then and I bring it onto the left-hand side. Now, this transfer function you see is equal to this
one with the denominator being sI −A ; the |sI− A| ; how do you compute the poles of
a transfer function? We basically take a look at the denominator and we actually look at the
roots of the denominator right we see the roots of the denominator are basically the
|sI − A| . It is exactly the same, alright.

So, computing the Eigenvalues of ( sI −A ) is exactly the same as computing the poles of
your transfer function. So, that’s why in the previous example, we were actually able to
prove; I mean show numerically that the Eigen the poles of this transfer function which are
−2 and −3 exactly equal to the Eigenvalues which is also −2 and −3 and this
is true in general.

957
(Refer Slide Time: 86:24)

So, what we have learnt in this lecture was that we introduced the concept of a state and a
state-space system and we saw why the state-space systems are very important. They can
actually model non-linear systems. They can take care of multi-input, multi-output system, so
on and so forth which the transfer function approach could not easily do or could not do in
some cases.

We essentially introduced the notion of the state we saw the basic form of a linear time-
invariant state-space system which was the x́= Ax+ Bu and y=Cx+ Du right and then
we did a few examples for the state-space modeling and then we derive the relationship
between the state-space and a transfer function crucially, we proved that the poles of a
transfer function are exactly the same as the Eigenvalues of the A matrix of the state-
space system.

So, this establishes that the dynamics we are looking at in the state-space and the dynamics
we looking at to the transfer function only for linear systems are exactly the same. So, what
I’m going to do in the next lecture is to look at since we talking since we are saying that state-
space can handle linear or even non-linear systems, we look at what are the basic differences
between linear and non-linear systems. We will look at why it is important to linearize a non-
linear system and there’s really good reasons for doing so, because handling non-linear
systems is extremely difficult and we will actually look at the technique of the Taylor series
to linearize a non-linear system and finally, we will come conclude with 2 really nice

958
examples one is a predator-prey, we’ll write the non-linear dynamics and show how we can
linearize them and finally, we look at the Van Der Pol oscillator, okay.

Thank you.

959
 Control Engineering
Dr. Viswanath Talasila
Department of Telecommunication Engineering
Ramaiah Institute of Technology, Bengaluru

Module - 12
Lecture – 02
Linearization of state-space dynamics

Hello and welcome to part 2 of state-space modeling. So, today we will be talking about
linearization of state-space systems. So, in part 1 we saw that there are certain benefits of
moving from the transfer function to the state-space domain. Primarily we are dealing
with non-linear systems we could deal with multi-input multi-output systems and so on.
We saw that in the examples of the electric circuit where you had two sources a voltage
source and the current source you had multiple inputs and multiple outputs.

So, today what we are going to see is that practically all the systems we consider in the
world or for engineering are all completely non-linear right. And yet there is a good
reason for us to not work directly with non-linear systems. We actually would like to
work with the linearized version of non-linear systems. There are a couple of reasons for
doing this I will explain that. So, we will see why we need the linearization procedure
and I will explain one specific way of doing a linearization which is by using the Taylor
series and the so called Jacobian matrix and we will end with two fun examples yeah.

(Refer Slide Time: 01:40)

960
So, that is the basic structure. So, we will see what is linearity then we go to Taylor series
and the Jacobian, and then we will conclude good. So, our physical systems linear or
non-linear they are all non-linear. There is no physical system which is completely linear,
but before we even go ahead with seeing what do I mean by a physical system is non-
linear and then how to handle this. It is important for us to understand the meaning of
linear or non-linear. And the reason why I would like to stress a little bit more on this
even though you have done this in electric circuits or what’s called network analysis
because most students are not comfortable with the distinction ok.

So, we remember it as a definition that they should be a superposition property, but it is

not too clear what it actually means. So, let’s see a very simple illustrative example of
linear and non-linear.

(Refer Slide Time: 02:35)

2
So, we consider 2 functions, the first one is y=3 x and the second one is y=x .
And the question I’m trying to pose is which of the above two functions is not linear
alright. So, if I just plot this particular function let’s say I the x-axis is, of course, x
and then I plot the output over here y . So, if I look at y=3 x the curve will
actually be something like this is at the origin 0 ok.

So, when x=0 you have y=0 and then it goes linearly like this. If you look at
2
y=x it is a function like this, alright? And now the question is which of these two is
linear or which one is non-linear alright. So, the way that we have studied this in electric

961
circuit is or even in math, to answer this we actually invoke the property of
superposition. I will explain superposition in a slightly different way.

(Refer Slide Time: 03:41)

Alright So, let’s see which of these two functions is non-linear. And the way we are
going to do this is to assume that each of these systems each of these mathematical
functions are actually systems alright. So, let’s consider the first one, we have a system
with the gain of 3 . There is another system which is like, this I give in inputs for both
the systems.

Let us in this particular case take only two inputs and I have an output over here. In the
first case, I’m looking at y=3 x alright. In the second instance, I’m looking at an
2
expression y=x . And so, this is actually a squaring system over here. In both the
cases what I’m assuming in this specific example is that I have an input x 1 , I have an
input x 2 . In the other case also I have x 1 and x 2 . And for the sake of simplicity,
but this is not required at all, but for the sake of simplicity let’s just assume that
x 1=x 2 . And we are going to assume that this y=3 x or y=x
2
are basically
models of some physical system. And let’s look at these two tables on the left over here.
So, for y=3 x and remember that we are assuming x 1=x 2 in both the cases. So,
y=3 x I take x 1 and x 2 . And I’m going to define x=x 1+ x 2 alright.

962
So, this is the basic equation which I’m going to use. So, when x=x 1+ x 2 and I
consider x 1 and x 2 with these values, x then becomes the sum of each of those
right. So, 1 becomes 1+1=2 , 2 is then 2+2=4 and so on. Now let’s look
at the function y=3 x , I can look at it in two separate ways I can either look at it as
y=3 ( x 1 + x 2 ) because remember x=x 1+ x 2 or I can treat it in this way
y=3 x 1 +3 x 2 alright. And let’s see if both produce an identical answer. And we see in
this particular case they actually produce exactly the same answer ok.

So, this is exactly what we mean by a linear function or a linear system. And I will
explain intuitively what this means a little bit later. Now let’s look at the second function
alright.

(Refer Slide Time: 06:27)

Which is y=x 2 and again there are two ways of writing this y=( x 1 + x 2 )2 . I can

also assume the signals are being added in the following way and remember x1 and
x2 are signals which are input to my system. We all know this is high school math
that these two expressions are not equal to each other right. And we can actually observe
this over here.

So, what is the interpretation of this? So, let us see what is the interpretation of this
assuming that these two are models of two different physical systems. So, what it means
is as follows. So, let us take the first one which is 3 x .

963
(Refer Slide Time: 07:14)

So, I have two inputs it is not necessary to just have 2, I can have 10, 100 whatever you
like and this can be written in two ways right. And we have just seen this it is either
3 ( x 1+ x 2 ) or it can also be written as 3 x1 +3 x 2 .

So, what is the interpretation of this? If x1 and x2 are two signals right. As
functions of time, all I’m saying is that I can add up the two signals x 1 and x 2 and
I can pass them into this your system called 3 , or I can pass each of these signals into
the system 3 and add up the result later. How would that look like? That would look
as follows.

964
(Refer Slide Time: 08:04)

So, I can take x 1 , I can take x 2 , I can add these two up in a summation block. So, I
get x 1+ x 2 over here and remember both are signals and I can pass them into a block
all 3 , I get y . The other option is I can pass each of them into a block called 3
which is a model of the system. And I can add them up over here in order to get y .
And both these systems are exactly the same. And if this is the case, this is actually what
we mean by superposition, actually a part of superposition, not the full thing, but this is
what we mean by superposition. I can pass my signals through the model first and then
look at the output or I can pass each of the signals through the model and sum up the
response afterwards. And you are going to get exactly the same answer and we have seen
this in linear electric circuits where we apply the principle of superposition to actually
compute system response to different sources. What happens if it’s non-linear? Well, if
it’s non-linear we have just. So, we take the 2 signals again, x1 , x 2 and we pass so,
we add them up ok.

965
(Refer Slide Time: 09:19)

Okay so, we take two signals x1 , x 2 add them up you get x 1+ x 2 . And then you
pass it through this block called the squaring block and you get your y alright. Now,
this is not equal to taking each signal squaring it up and then adding them. This y
over here is not equal to this y . And this basically violates it violates the principle of
superposition. So, it violates the principle of superposition and we can see that in the first
2
case y is nothing but ( x 1+ x 2 ) and the second case what is it? It’s ( x 21 + x 22) . And

these 2 are not the same. There is a factor, a residual factor of 2 x 1 x2 which is left out,
right.

So, again if you violate superposition it essentially means that I cannot mix my signals
pass it through the system and say that it is the same as passing each signal through the
system individually and then mixing it up, okay. That is not true for a non-linear system.
So, good, now we know what is linear, what is non-linear in a general sense, we have
understood superposition this applies to all physical example electrical, mechanical and
so on.

966
(Refer Slide Time: 11:10)

Examples of non-linear systems, I will just put 3 over here there are thousands of
examples of non-linear systems.

So, there are your diode behavior it is actually one of the classic examples. So, you have
this kind of a response over here right. So, in this region it’s linear in this region it is
absolutely not it changes completely right. Even if you look at a non-linear mechanical
spring and there is no such thing as a linear mechanical spring, all springs in life are
really non-linear. So, if you look at the non-linear spring again you have this kind of an
effect. So, all these are non-linear examples and there are thousands of such. So, while
the claim is that all systems are actually non-linear what happens in practice many times
we see systems which behave as if they are almost linear if we ignore certain very small
effects. So, for example, I can have a spring which is like this which looks almost linear
everywhere. And my range of operation of this particular spring depending on the
application I’m looking at maybe it is only in this range right. And I do not need to worry
about what happens over here or over here because my operation actually happens in this
particular range.

In this case, it is enough for me to just use a linear model right. And I do not need to
worry about the non-linear effects and this is actually what happens in practical
engineering. As far as possible we try to use systems in their linear operating range
alright. It’s not always easy to do so, and we will see a procedure of how to linearize a

967
non-linear system in the so called linear operating ranges alright. And that is essentially
the topic of today’s talk.

Now, why we want to do this is primarily because of this reason. The analysis and
control of non-linear systems is very painful. It can get extremely mathematical, the
computational effort is very high and so on. On the counter, if you look at the linear
approaches you have excellent.

(Refer Slide Time: 13:30)

Simulation design tools, you have simulation packages so on and so forth. Primarily
relying on linear algebraic approaches alright. So, the only option we have because all
systems are non-linear the only approach we have is to linearize the non-linear system.
And we will explain how we are going to do that. We need 3 concepts the first is the
concept of equilibrium point, the second is the concept of the Taylor series, and the
related notion of the Jacobian. Using these 3 we will show you one specific way of
linearizing a non-linear system.

So now that we know that this is what we are going to do linearize the non-linear system.

968
(Refer Slide Time: 14:12)

So, first, let’s look at the general expression for the non-linear system. So, if you recall
linear state-space system was x́= Ax+ Bu , y=Cx+ Du , let’s ignore the inputs and
outputs for the minute. So, what we will have is basically a system of the type x́= Ax
alright. And notice that A is independent of x , that is the crucial thing A may
be a function of time, A maybe constants whatever it is, but it is independent of x .
In the non-linear system, you have the classical way of denoting a non-linear system
dynamic as x́=f ( x) . And f ( x , t) is specifically a non-linear function of the states
x (t) . Of course, in the linear case f ( x )= A .

Now, I can write my dynamics as simple example would be say 3 x2 alright. Or I

could write x́=sin( x )cos( x ) , these are all examples of non-linear system and we
essentially are just calling these as f ( x ) . So, that is a general notation for a non-linear
system.

969
(Refer Slide Time: 15:27)

Now, it is very important to understand the concept of an equilibrium point. And the
definition of an equilibrium point is as follows. So, given a non-linear system
x́=f ( x) as we have seen before a point or x or you can call it as x 0 . Some
people also use x¿ as another notation. A point x0 is said to be an equilibrium
point of this non-linear system x́=f ( x) , if and only if at that particular value of x0
the dynamics is equal to 0 .

So, this basically means that x́=0 when evaluated at x 0 . So, that’s the notion of an

equilibrium point. It basically says at that specific point x in x 0 in this particular

case your dynamics are 0 . Let’s look at one simple example of the so called scalar
case remember that when you talk of states you have state vectors. So, in general, we
would talk about x as actually being a vector of states right. For this particular slide
let’s ignore that. What we are going to really look here is just a single state alright.

So, let’s look at this example of the RC circuit. And we can write based on the
Kirchhoff’s voltage law. This relationship over here, where the voltage across a capacitor
plus the voltage drop across the resistor it basically sums up to 0 alright. So, when I

dV V
rewrite this a expression I get + ¿ 0 . Now, what is the equilibrium point
dt RC
for this particular system? That is a question. And the and the procedure to compute the

970
equilibrium point is really straightforward, we know that the definition says the
derivative must go to 0 at that particular point ok.

dV −V
Now, what we are going to do is to write = alright. To compute the
dt RC
equilibrium point I need to set this to 0 . Now, this is equal to 0 only when
V =0 right. Because no one will take R or C equal to 0 and construct a circuit like
that. So, R and C will be non-zero values and this equation holds true only when the
voltage across the capacitor is equal to 0 alright. So, this is an equilibrium point now
it’s actually an interesting equilibrium point and we will come to one of this notion a
little bit later in a more complex case.

So, first of all, remember we computed and said that when the voltage is equal to 0

dV V
for this model of the system + ¿ 0 , this is an equilibrium point because the
dt RC
dynamics is actually equal to 0 . Now the solution of this simple first-order
−t
differential equation is basically this one. So, V 0 e RC , where RC is your time
consumed basically. So now, the question is we know that when V =0 ? alright when
you start the circuit simulation with the initial condition, such that V (0)=0 alright. In
this case, for all future time, this answer will be 0 alright. It will never be a non-zero
value, it will always be 0 .

Now, what happens when we started a different initial condition say, for example,
−t
V (0)=5 V alright. We see here that V ( t )=5 e RC and for any reasonable say

combination of R and C as t → ∞ actually it doesn’t even need that much time, but as
t tends to a sufficiently large value V (t)→ 0 . So, what this means is that when my
initial condition is 0 , my the dynamics of the system will always stay at 0 . It
never goes to any other finite value. When I started any other initial condition my
dynamics will always converge to 0 alright. So, if I actually draw x and t as a
function of time, actually v sorry about that.

So, if I take v as a function of time if you start over here at the origin initial condition
is equal to 0 , you will never have dynamics going anywhere else it will always stay
there. You can start anywhere else over here or this can be an initial condition, this can

971
be an initial condition, this, this whatever it be, you will always converge to this value
0 . And it turns out that this is a specific case of what is called a stable equilibrium
point and more specifically a globally stable equilibrium point.

Now, we will not go too much into these definitions, but when you look at design
analysis of control system these concepts become very clear. So, let me summarize this
slide, basically saying that a particular point of your solution we are looking at v over
here, one of those values of v is said to be an equilibrium point if the dynamics is
always equal to 0 at that particular point. And we also have a simple procedure to
calculate what that equilibrium point is which is basically you equate the derivative to
0 and you solve for that f ( x) and that’s what you get ok.

(Refer Slide Time: 21:36)

So, this is basically what I said in the previous slide, that if you start off your dynamics
let’s say if you start a dynamics at x 0 , where x0 is your equilibrium point the
dynamics will always remain at that particular value for all future time t . It also
means that if I start at any other initial condition and accidentally or through design or
whatever the system manages to reach that particular state from that time onwards the
system will be essentially stuck at that state forever alright. So, that’s a basic notion of an
equilibrium point. And of course, there are various kinds of equilibria there are saddle
point, centers, all kinds of things we will encounter one or 2 of them in this particular
lecture otherwise it’s actually a fascinating study ok.

972
Now, we have been talking about equilibrium points, we have been talking about
linearization. Now there is a natural question which comes, why do we want to linearize
about an equilibrium point why not somewhere else? The answer to that is fairly one is a
very practical reason. And it’s basically that it’s very easy to compute the solution of
differential equations at the equilibrium point. All you need to do as we have seen in the
previous slide just equate the derivative to 0 and just solve it. Generally, it is fairly
straightforward as opposed to computing the general solution of the differential equation
itself which can be very complex.

The second and more relevant reason more important reason is that we have theorems
which actually guarantee that if you linearize a system about an, in the neighborhood of
this equilibrium point right. So, if you linearize the system at this equilibrium point we
are guaranteed that the behavior of the system at this equilibrium point is very similar to
the behavior of the system anywhere else around the equilibrium point in a small
neighborhood right. So, in the previous case, we had the RC example. And we saw that
the value of 0 is an equilibrium point you start anywhere close to 0 right. Say
0.00001 or whatever it be in a very small region around that equilibrium point, your
behavior will be very similar to what it is at the equilibrium point fine.

(Refer Slide Time: 24:18)

So, that’s basically why we look at linearizing around an equilibrium point. So, let’s
actually first look at a slightly graphical sketch of how we would approach the problem

973
of linearization, and in the next slide, we will actually see the so called Taylor series and
a more formal way of doing the linearization, okay; alright. Let’s say that we have this
parabolic like function in the blue line which we are seeing over here. And we want to
actually fit another function to this one alright. And or rather we want to approximate
this function and there is various ways of approximating such functions. So, in the first
case, I can simply have that at a particular, ok.

So, let me call this as x and let me call this as f (x) . x is an independent
variable and f ( x) is a function. So, as x varies we have a change in f (x) also.
Let’s take a particular value of x , say that we take x=0 , at this point. And then we
compute what is this f (0) we’ll get some value let’s say it looks to be about say
almost 1. Now, this is one approximation to this curve right. So, you have this curve and
one particular approximation to this curve says that I have I know this particular value I
don’t know the rest of the curve, but I know that one particular value well that’s a good
start.

Now, what can we do once we know this one particular value, well we can actually
compute the slope right? This looks like a nice curve and one of the nice things we can
do is just compute the slope over here alright. The slope. So, if this function is f ( x)

∂f
the slope over here would simply be alright. Now, this is another approximation.
∂x
So, the way I would actually write this is let’s take this particular value in x we will
call it as a . So, x=a . So, at the first instant I got the value of f when
x=a , we’ll call that as f ( a ) right.

Now, to that, I’m going to add the slope over here which is a straight line right. So, I will

∂f
add . And essentially I get this kind of an, this is an approximation. So, all I’m
∂x
saying is given my nice parabola I now have an approximation to the parabola which is a
straight line clearly this is not true right. I mean you still have all this deviation over here
which needs to be taken care of. So, what we could do is now to fit this in a slightly
different way. So, you had your parabola again and you computed this point, you
computed a straight line why don’t we actually try to fit a second-order say polynomial
around this particular point right.

974
Let’s say that we are able to fit something like this. This is another approximation. So,
what I’m going to do now is to add a second-order term over here. And when I add up all
these terms together the final fit to this particular curve will be something like this,
almost what we want. So, this is a very good approximation of this particular parabola.
Now in general you may not have just parabolas you may have say fairly high order
curves.

(Refer Slide Time: 28:14)

In which case I may need to take first is my constant value, then I need to take a straight
line I may need to take a second-order polynomial you need to take a third-order
polynomial so on and so forth. All these curves which you are trying to fit to this
particular polynomial are basically approximating this polynomial right.

So, this actually goes by the name of the Taylor series. And the basic idea of this slide is
that linearization is basically an approximation of a non-linear function. How do how we
do it formally is by the Taylor series expansion, okay. And so, we have already seen what
this thing is right. So, if I go back to the previous slide you can see the animation over
here it’s a second-order, it’s a third-order, fourth-order, and so on right. I can show this
again. So, there you go first order, second order, third order, and so on right. So, these are
different fits to the particular function.

975
(Refer Slide Time: 29:19)

And the way we compute those fits to the function is by using the notions of derivatives
we have already seen the first order derivative where we took the slope of the line, that’s
a slope of the line over here right. Slope of the curve sorry.

So, if we add up all these terms of the Taylor series essentially you are actually going to
have a perfect fit to your parabola alright. And that’s basically approximating the non-
linear function. And of course, where we approximate is important, we have already seen
that it is better to approximate at the equilibrium point because it is easy to compute the
solutions and other reasons right. So, we always try to approximate at the equilibrium
''
point. So, if you look at these terms over here the f ' ( a ) , f (a) , and so on, these are
nothing but the partial derivatives with respect to the independent variable, in this case,
''
we have only one independent variable, evaluated at the equilibrium point right. f
will be the second-order partial derivative evaluated at the equilibrium point and so on
right.

So, all the derivatives we are considering are with respect to the independent variable
x , in this case, one, but in general, it will be a multidimensional independent variable
right. Because remember that we have the state vector which was x1 , x 2 So on till
x n . So, let’s see how that works out.

976
(Refer Slide Time: 30:46)

So, in the previous case, we had a function of the type, for example, it could have been
2
x́=3 x right a scalar dynamics which you can call. Now what we are going to look at
is functions of 2 variables in the next slide we will go to n variables and see how we
can generalize this.

So, we look at functions of two variables this can actually look as follows. So,
x́=f ( x 1 , x 2) right. And the way I can write this down as one specific example it could
2
be something like 3 x1 x 2 , x́ 2=x 1 x 2 something like that. Now this would be my
f 1 (x 1 , x 2 ) this would be nothing but f 2( x 1 , x 2 ) and that’s basically this expression
expanded in 2 equations ok.

So, that’s basically what we are saying if you have a state vector with 2 states and the
dynamics of state one depends on state 2 and vice versa this is captured in the 2-
dimensional function f ( x1 , x2 ) . And that’s exactly what we have done over here. So,
this is a 2-dimensional state-space system and we have 2 functions f 1,f 2 remember in
the previous case when we had the dynamics like this x and f (x) right. Where
2
this parabolic thing says some whatever x whatever it is so when you had this kind
of an equation we actually considered at the equilibrium point in this case equilibrium
point is over here. So, we considered at the equilibrium point the derivative of f with
respect to x right.

977
So, basically, we are trying to understand how the dynamics of the system changes with
respect to the independent variable. In this case, we had only one independent variable.
In our example, in the current example, we now have 2 independent variables right. So,
if we try to look at it visually it will actually look something like this.

(Refer Slide Time: 33:02)

So, I have got a x 1 , one of the independent variables the other independent variable
is x 2 right. Then you have your function itself f (x) . So, for a change in x 1 and

a change in x 2 , I will have a particular value of f (x) right.

So, typically this would be a surface or something like that right. So now, because
f ( x ) which is basically x1 and x 2 , because f changes with respect to x1

and x 2 , I need to compute 2 derivatives. So, it one derivative will be respect to x1 ,

the change of the dynamics in the x 1 direction the other derivative will be the change

of dynamics in the x2 direction. Now we need 2 derivatives unlike in the scalar case
where we just use a single derivative alright ok.

978
(Refer Slide Time: 33:55)

So, how will that look like? Well, if you just write the Taylor series and you ignore the
higher derivative terms, it can call it the higher-order terms ok.

So, we take only the first-order terms over here for the first equation x́ 1 or the first
state in our case we actually have these 2 expressions over here right. The derivative with
respect to x 1 , the derivative with respect to x 2 . These are again I’m repeating this

simply capturing the change or the variation of your dynamics your system in the x1
direction. This one is capturing the variation of the dynamics in the x2 direction,
which is why we always have 2 derivatives.

Similarly, f 2 also varies as a function of x 1 and x 2 . So, again you have these 2
derivatives. If we put these derivatives in a matrix form you have this kind of a structure
over here. It is a square matrix again and essentially you are going to just take all the
partial derivatives and then put it over here. Now, this is a special kind of a matrix called
the Jacobian, and the Jacobian basically is derived from the Taylor series right. The first-
order terms of the Taylor series we are ignoring the higher-order terms because in
general first-order terms are usually good enough approximations to the non-linear
functions in general not always. Now, we know how to handle second-order dynamics,
how would we handle nth order state-space system well based on what we have just
seen, for a second-order state-space system if we compute the derivative with respect to

979
 x 1 , and the derivative with respect x 2 , for a nth order you will just keep on going
until you compute the derivative respect to x n alright.

Now, note that when I say f it is not just f it is f 1 and so on I will have f2 .
So, the derivative of f2 with respect to x 1 . So, until derivative of f2 with
respect to xn and so on. Put all of this in a matrix form and this will essentially
generalize to this kind of a structure.

(Refer Slide Time: 36:15)

So, this is your the classical the Jacobian matrix. And this is where we stop one part of
the lecture and what we have done so far, is the following.

So, we have a non-linear system or we have the dynamics of the non-linear system. We
have then computed the equilibrium points, we have defined it and just said how we can
compute the equilibrium points of the dynamic system. And finally, we have seen that to
approximate the non-linear function we actually use the Taylor series. And as far as
possible we typically use only the first-order derivatives of the Taylor series right. So, we
use a Taylor series the first-order derivatives and then we compute the Jacobian ok.

So, this is what we have done so far, now the question is how do we go from the
Jacobian to the linearized state-space system, how do we get the linearized state-space
system once we have the Jacobian. And by the way, I will be deriving one or 2 examples
for the Jacobian. So, this will become clear in a couple of slides.

980
(Refer Slide Time: 37:29)

So, how we go from the Jacobian to the linearized state-space system is as follows. So,
again all of this is at a high level when you look at the example it becomes much more
clear. So, first, we have computed the equilibrium point, which we have which you
talked about. And then we compute the Jacobian. So, I would actually include another
point over here I would say compute the Jacobian right.

So, first, we compute the equilibrium points of the system then we write down the
Jacobian the matrix basically. And the equilibrium points which we have, we substitute
the equilibrium points into the Jacobian and we will see an example of how to do that.
When you substitute the equilibrium points into the Jacobian what you are going to get is
a linear A matrix. And that’s it it’s as straightforward as that. So, the procedure is you
write down the non-linear state-space equations, compute the Jacobian, compute the
equilibrium points, substitute the equilibrium points into the Jacobian and this gives you
the A matrix. A point which we will come back to maybe in the next slide which is
very important is that the computed linear A matrix or the system matrix A is
valid only at the equilibrium point and in a small neighborhood around the equilibrium
point it is not valid anywhere else. We will see that very shortly ok.

981
(Refer Slide Time: 38:57)

So, if you rewrite this a little bit this. So, we have the basic non-linear dynamics then we
compute the equilibrium point, the Jacobian and you get x́= Ax and this is the
linearized model of x́=f ( x) alright. So, there we go now yeah. So, one example is if
you take the Tejas our Indian light combat aircraft or for any aircraft for the matter of
fact and aircraft the general non-linear dynamics are very complex right. And you really
can’t design simple controllers for these really non-linear dynamics. So, what engineers
have done over the past fifty years or so. They have design what are called as linear
models at each operating point of the aircraft flight path say, for example, just one
particular example if we. So, this is the terrain earth and the aircraft's trajectory basically
takes it along this path ok.

So, we have our wonderful Tejas which is actually flying around this path. Now at each
and every point on the flight path, multiple changes happen. So, what are these changes
let’s actually write some of them down? So, your wind, your conditions change,
atmospheric conditions change, the pressure conditions change of course when wind
changes even turbulence and other factors also change. So, your pressure changes that
these are the external parameters, the internal parameters are that the fuel is reduced a
little bit your plane may be having some extra drag because of the kind of wind
conditions which are affecting it and so on right. Each of these changes, every single
change actually changes the model of the aircraft dramatically.

982
So, these are actually called operating points right. So, as we go from one operating point
to another operating point the model of the system is actually different. So, what people
have actually done is to take each of these operating points and linearize the system
around this operating point. So, to get A matrix take another operating point and then
linearize it over there and so on and so forth. So now, you have a bunch of system
matrices A1 to An along the flight path of the aircraft for aircraft. For each of
these system matrices each of these A matrices they actually design a controller. So,
corresponding to each matrix you now have a controller right. Now you can call this
K 1 , I can call this K 2 . So, on until Kn and this is in flight control literature this
is actually called as gain scheduling, without getting deviated from our topic. The idea of
this example is that this is a highly non-linear problem right this flight path which
encounters. So, many changes in wind, pressure, fuel, drag, lift, so on and so forth is a
very highly Non-linear system.

And the only way we can handle it at each operating point of this flight path we actually
linearize the system right. So, then you get this set of linear models and for each model
we actually design controllers. So, this is how it is classically done in aircraft industry
and chemical process plants and so on.

983
 Control Engineering
Dr. Viswanath Talasila
Department of Telecommunication Engineering
Ramaiah Institute of Technology, Bengaluru

Module - 12
Lecture - 02
Part 2
Linearization of state-space dynamics

Hello and welcome to the continuation of lecture 2 of module 12, which is on the
linearization of state-space dynamics.

(Refer Slide Time: 00:30)

So, if you remember what we have done in the previous lecture is part one of lecture 2. So,
basically what we did was to consider a non-linear system. So, we model the non-linear
system right, then we computed the equilibrium points and we computed the Jacobian which
is you considered all the first derivatives of the Taylor series and in general it is a nXn
matrix depending on the number of states that you have. So, you compute the Jacobian
matrix, and then in order to compute the A matrix, you basically evaluate the Jacobian,
alright at an equilibrium point right.

So, this was the basic theory which we saw in the previous class. Now before I go ahead with
two very interesting examples for today I just want to tell, make a small note about the
equilibrium points itself. Now, in the non-linear case which is what we are dealing with in

984
general you do not have a single equilibrium point, you can have multiple equilibria. So, if
you consider a non-linear dynamical system it can have multiple equilibrium points I will
give a small example to illustrate that let’s say I have got a non-linear system of the following
type alright.

Now, if you actually graph this. So, you have x-axis and you have the cos ( x) on the
y-axis, x of course, will be in degrees or radians; however, you like, and so, cos ( 0 )=1 . So,
you will actually get a graph like this right and so on, now if you recall the definition of an
equilibrium point it is basically where the derivative goes to 0 and equilibrium point is a
point in your space of the independent variable. So, it could be any one of these points where
the derivative goes to 0 .

Now, you see that for this simple example of the cos (x) . So, if you look at the cos ( x)
and you try to compute the equilibria, you see that you have multiple locations the angles that
is where the derivative of the dynamics actually goes to 0 . So, this would be one place,
you have this, you have this, so on and so forth. So, for this, you actually have an infinite
number of equilibria depending on as long as x keeps on repeating. So, for a non-linear
system, it is not necessary that you have a single equilibrium point you can have multiple
equilibria alright.

So, with that in mind let us actually go to two interesting examples for today.

(Refer Slide Time: 03:27)

985
The first example is a predator-prey model and before going ahead with the mathematics of
this the reason why we introduce you know tigers, deers, and other kind of things in a lecture
in control systems or in control engineering is actually twofold. One is that the beauty of
control systems is that it allows us to model in principle any physical system, and the
physical system need not be a chemical process, it need not be just a electrical system, it can
even be an ecological problem as what I am going to present today, it can even be a problem
in economics right and so on that’s one thing.

The second is that when we look at the predator-prey model today, it is important to
remember that the kind of dynamics which I will be talking about are also applicable when
we model certain behavior in economics, in social behavior and so on. So, with that in mind,
let’s go ahead with the predator-prey model let’s take the specific case of tigers and deers. So,
these are two pictures which I have taken of the net and they show a tiger and the deer in the
Bandipur forest and the male Sambar deer is native to the Bandipur forest, and one of the
reasons why people do a lot of modeling for these tigers or predators in forest is because the
tiger is basically a symbol of the well being of the ecosystem.

The tiger depends on the prey, the prey depends on say trees and grass and other kind of
things, they depend on other subspecies. So, if the tiger is actually conserved you are
essentially conserving the entire ecosystem. So, that’s why the predator-prey models were
historically developed. So, what is the predator-prey model? First let’s look at the prey ok?

(Refer Slide Time: 05:33)

986
So, let H (t) represent the number of deers and it’s a function of time. Now in this
particular context of the predator-prey model t is not in seconds, it is typically in months
or in years ok, so that we need to keep that in mind.

So, if H (t ) is a number of deers in the absence of predators, we are going to make an

assumption, that the deer population keeps on increasing which is an assumption I am making
in my modeling and that essentially follows this particular equation. It basically says that the
rate of change of the population of the deers as a function of time and time being in months
or years is a positive rate of change where r h >0 , and h is basically the population of
the deers, r h >1 to be more specific right and this particular parameter rh H , rh
essentially is the reproduction rate of the deers or the growth rate of that population alright.

So, as long as r h >1 , we have the population of deers going on increasing which is
basically this what this equation talks about over here.

(Refer Slide Time: 06:54)

So, we know that this cannot be really true in general, that the population of the prey keeps
on increasing because what this particular equation seems to be telling us is that as time goes
to infinity or a very large time like maybe 5 years, 10 years, and so on, the population of the
deers which was say for example, 50 in the Bandipur forest, according to this equation would
keep on increasing right which is not true because your forest ecosystem cannot support such
a large population of deers right, and the way that we would model this constraint is to say

987
that, let K be the maximum population of deers which this particular jungle or the
ecosystem can actually support. And the number of deers cannot go more than K and the
way you would model that is by this expression over here and what this expression basically
tells us. So, this part is the same as what we have seen before the growth rate or the

H
reproduction rates of the deers, then have a 1− . H is generally less than K , in
K
the case where K is very large.

In the example where K is very large let’s say even approaching infinity, this term goes to

H
0 . So, 1− becomes 1−0=1 , and basically says exactly what we talked about
K
before that the growth rate of the deers is constant and it keeps on increasing and it can reach
a very large number possibly infinite. In general, H ≪ K , okay because the forest cannot
really support such large population of deers, there is not enough food basically. And as long
as H≪K this expression over here is less than 1 alright and. So, that basically means
that I do have a constraint on the total number of population of deers which my jungle can
support.

However, for the remainder of this stock without any loss in generality, for the remainder of
this stock, I will assume that K is very large meaning I have this hypothetical jungle
which can support as many deers as possible, it is not a correct assumption in reality, but we
will go with that for now to just show the modeling capability of state-space systems. So, this

dH
slide basically concludes with the equation which we are going to consider as =r h
dt
H . So, this is the dynamics of the prey in the absence of the predators ok.

988
(Refer Slide Time: 09:43)

Now, in the presence of predators you see that this growth rate of the deers is now
compensated by interaction say parameter between the deers and the tigers where L(t ) is
the number of tigers in the jungle, H (t ) of course as we saw before was the number of
deers, and the parameter a is the interaction parameter between the tigers and the deers.
So, we can see that now this equation basically says, yes I do have a growth rate in the deers
because of reproduction and there is plenty of available food for them, but then I have a
decay rate over here this negative term, this is an interaction when the tiger basically hunts
the deer and the population of the deers go down alright.

dH
So, this is the. So, this is our prey model. So, =r h H−aLH and it’s actually an
dt
interesting exercise I would like you to try this, this parameter over here the product of the
number of tigers and the number of deers is your non-linear term, this is a perfectly linear
term rh is a positive constant, H of course, is the population of deers, this is the non-
linear term over here I would like you to prove this on your own. So, that is the prey model,
now what about the predator.

989
(Refer Slide Time: 11:17)

Now, the assumption that we make for the prey model is that there is always plenty of food
available which is basically grass and other leaves which the deers eat, but for the predator, it
needs the prey right now let’s say that the predator is for whatever reason it is unable to
actually kill the prey and then have food, in which case you are actually going to see that the
predators are going to starve to death. And let’s assume this may not be always true, let’s
assume that the rate at which they starve to death is an exponential factor, okay an
exponential decay, in which case we get this expression. So, when L(t) is a number of
tigers in the jungle. So, the rate of change of the population of the tigers is this negative
growth rate over here ok.

So, this is in the absence of prey

990
(Refer Slide Time: 12:11)

In the presence of prey, the population of the other tigers will obviously recover, right. Now,
they get food and they can they are healthy they can reproduce and so on. So, every time I
have an interaction between the tigers and the deers, the population of the tigers over a period
of time: months, years will be a positive growth rate, this is good for the tigers, not so good
for the deers, but it is very good for the tigers. So now, this is the equation the dynamics for
the population of the predator, and again as in the previous case I would like you to see and
try to prove on your own that this product term between 2 variables which are actually the
states between 2 variables L(t ) and H ( t ) is actually a non-linear term, this is actually
what makes the dynamics non-linear, okay very good.

991
(Refer Slide Time: 13:12)

So, now we have 2 equations and it turns out these equations are coupled right. So, you have

dL
H which is a function of both H and L and then which is a function of both
dt
H and L . So, if you just rewrite this in a matrix form we essentially have this
expression over here, okay. This again your classic state-space thing and this is not the A
matrix as we call it in the linear case. So, this is basically x́=f ( x) , where my state vector
x is basically H (t) and L(t ) alright, and this is the f (⃗x ) over here, okay very
good.

So, what we have done so far now is to basically capture and to actually write down the non-
linear dynamics of the predator-prey problem right. It’s a first step, now the second step
would be to compute the equilibrium points and we go to the Jacobian and so on. So, let’s see
how we do that.

992
(Refer Slide Time: 14:16)

Before we do that there is an important modeling assumption which me should make. So, if
you see in this slide we’ve assumed, that the interaction parameter between the tigers and
deers is the same now this is not true in general and it’s not true for an interesting reason
because every time an interaction takes place between the tiger and deer. Typically, the deers
usually escape most hunts are not successful; however, when an interaction actually does take
place meaning the tiger manages to make a kill, the deer population will definitely reduce
immediately by at least of by at least one depending on how many the tigers kill, but this
doesn’t mean the tiger population is going to increase immediately right.

So, the population of deers decreases because one of them was killed, but the population of
tigers will not increase immediately it takes a much longer time. So, this interaction
parameter a it cannot be the same for both. And, more specifically in our case the
interaction parameter for the deers will always be greater than the interaction parameter for
the tigers, okay furthermore the growth rate of the deers is much larger than the death rate of
the tigers again it’s an assumption and these assumptions are usually proven to be valid in
actual field studies. So, we take this basic model and with the assumptions which we have
placed over here, we get the new model which is exactly it looks the same we made certain
assumptions on the parameters right.

So, we basically assume that a1 >a 2 , where a1 is the decay rate for the deers when the
interaction happens and rh and rl are the respective parameters for each of them. So,

993
now, we take this set of coupled differential equation and we will compute the equilibrium
points.

(Refer Slide Time: 16:17)

Now the and then, of course, linearize it and we go ahead with that. So, the way you couple.
So, the way you compute the equilibrium points we have seen before for a system of the type
x́=f ( x) , and equilibrium point x0 is such that when you evaluate the function
f ( x 0 ) =0 which is basically saying that x́=0 when evaluated at x 0 alright.

So, we are going to the same here, we are going to set each of these parameters to be equal to
0 and we will see what is the solution of these 2 equations and that is of course, fairly
simple because in this case is just algebraic expressions over here.

994
(Refer Slide Time: 17:08)

dH
So, we said ¿ 0 , we will see that this basically means r h H−a 1 LH =0 . So, this
dt
basically gives you let’s just call it equation number one, we will come to the other one later

dL
and we do the same thing where ¿0 and this basically would mean that
dt
−r l L+a 2 LH =0 .

We have 2 equations and 2 unknowns, the unknowns are L and H and remember these
are functions of time. So, we have 2 equations and 2 unknowns and the unknowns are L
and H . The trivial solution of this guy being equal to 0 and this equation being equal
to 0 , the trivial solution is L(t)=0 and H (t )=0 . So, this is one of the is an
equilibrium point, and remember in the previous example which we gave we had a single I
mean a one-dimensional system. So, the equilibrium point was just one single scalar value
now the equilibrium point x 0=( 0,0) alright. That is one equilibrium point and this is the
trivial equilibrium point this is not very interesting, but because it basically says that you start
off with 0 tiger, 0 deers in the jungle and of course, what will happen to your
dynamics, they will always remain at 0 this is not it, this is not interesting that’s what.

Now, what about the other equilibrium point if we solve these 2 expression these 2 algebraic
expressions.

995
(Refer Slide Time: 18:50)

rl rh
You are going to get this as your equilibrium point and and so, I will just call
a2 a1
this as step one. So, first, we had the non-linear dynamics steps 0, now we computed the
equilibrium points we are going to use this one and not the other one and now let’s see for
step three which is to compute the Jacobian and how do we compute the Jacobian? We use

∂f1 ∂f 1 ∂f2 ∂f2

the expression which we had before which was , , and . Where
∂H ∂L ∂H ∂L
f 1 =r h H −a1 LH and f 2=−r l L+a 2 LH . So, with these 2 functions if I actually compute
the Jacobian this is the expression, which I am going to get alright. So, this is a step 2.

Now, in order to derive the linearized system, we know that we are going to essentially
substitute the values of H and L into the Jacobian over here. So, when you do that. So,
this is going to become basically this expression over here.

rl rh
So, into the Jacobian, if we substitute, sorry and which is the equilibrium point
a2 a1
which we had over here into this if right. So, if we substitute this into the Jacobian you are
essentially going to get this linearized matrix. So, this is actually the A and with this, we
actually get your x́= Ax and of course, every time we see the linear matrix A , and you
can see that it’s linear we do not have any coupling parameter in the previous example of the
non-linear system we saw that we had a coupling parameter a , L , H .

996
So, this was a non-linear term now those terms have disappeared over here when you
evaluate the Jacobian at the equilibrium point. So, now, we have x́= Ax . So, this entire
Jacobian at the equilibrium point is a and we would always first thing you would always
do is to compute the Eigenvalues, and when you compute the Eigenvalues of this matrix
A in this particular case you see that the Eigenvalues lie on the imaginary axis, and when
you actually do the simulation of this which I will show you in a few minutes you will see
that these Eigenvalues they produce a very interesting behavior that the equilibrium point
which we have considered, that’s this one over here the equilibrium point actually behaves in
a very unique way which is called as a stable center ok.

So, how would that look? We saw in one of the previous examples in the previous lecture that
an equilibrium point one of the definitions is that, if you start off at an equilibrium point your
dynamics always remains in that equilibrium point for all time in the future, okay. Stable
center, has a very unique understanding or notion of what the dynamics would mean, and it
basically would say that if this is an equilibrium point. So, when I evaluate these 2 values

rl rh
right and I get this particular point over here. So, this would be my and , and if
a2 a1
you start at the equilibrium point as by the definition and we’ll also see in this in the
simulation shortly the dynamics will tend to stay at this point only it doesn’t go anywhere
else.

Now, because it is what is called as a stable center, it has a following really interesting
behavior. If you start at any other initial condition, the final behavior of the dynamics when
you plot H (t) Vs L(t) , this behavior will always be a closed orbit or a closed response
if you want to call it around the equilibrium point. Start at any other initial condition and it
will go around that equilibrium point, start over here it will go around that equilibrium point.
If you start somewhere over here it would actually tend into this part of the axis and then
again create a new orbit and so on.

So, the concept of a stable center is that you have an equilibrium point as before, if you start
your dynamics initial condition at the equilibrium point the dynamics will always stay there.
You start anywhere else around in a neighborhood of the equilibrium point or anywhere else
in your phase space, that is a space of your state variables anywhere else it will be a closed
orbit about that equilibrium point. And we will see very quickly that these closed orbits

997
basically mean that the dynamics are basically oscillating with respect to each other. So, this
orbit basically means that when the population of the deers is increasing, say assume I am at
this particular point, and the population of the tigers is actually reducing at the same time you
see the population of the deers is increasing, and it keeps increasing until the population or
tigers have completely come down, at this particular point the population the tiger starts
going up and the deer population starts decreasing.

So, if you actually plot this response you will actually see that if the tiger population was
going like this, the deer population would follow this, but with a slight delay. So, every time
the tiger population goes up the deer population would come down, and when the tiger
populations coming down the deer population would actually go up. So, you will see this
kind of a response; that is basically an outcome of how these closed orbits are there.

(Refer Slide Time: 25:52)

So now, we talked of the center, it turns out the other equilibrium point which we did not
consider which was a trivial one where did that go, yes. So, where L(t)=0 & H (t)=0
, it has a specific mathematical property that is called the saddle, we will not really go into
those things, we are more interested in this equilibrium point which actually has some
physical meaning ok.

998
So, now that we have our linear system of this form, let’s actually and we know that the
Eigenvalues are unlike this and we expect to see a stable center, let’s actually go ahead and
simulate this in MATLAB and we will see what kind of responses we going to get.

(Refer Slide Time: 26:38)

Okay so, this is a code which basically simulates your predator-prey model, it is called the
Lotka Volterra because that’s the formal name of this one. So, the basically the code runs like
this, we are going to simulate it for 50 seconds it’s actually a little bit too long we’ll just do
for 30 seconds and the parameters which I’ve taken. So, the growth rate of the deers is 3 ,
the natural death rate of the tigers is 1 ; So, that is a −r l L(t ) and then the interaction
parameters for the deers is 3 and for the tigers is 1 ok.

Now, these have been chosen not with any specific ecological model in mind, it’s more a
mathematical simulation, but if you have actual data from say any forest system you can
actually substitute these appropriate values and the equilibrium point is defined over here it

rl rh
was and . Okay so, that’s over here, now let us say the initial condition, okay. So,
a2 a1
this is the number of deers and this is the number of tigers and the initial condition we usually
assume that the number of deers is much larger than the number of tigers right otherwise all
the deers will die eventually tigers will starve they will also die, but that really does not
happen in most ecological conditions.

999
So, number of deers is more than number of tigers as an initial condition, and then we use this
function from MATLAB called the ode 45 . So, ode 45 calls a function called deq 1
which I have defined in the previous line, line number sixteen over here and it says that we
need to simulate it for about 30 seconds with these initial conditions. For those of you for
whom it’s not clear what is this ode 45 ? what is this deq 1 ? The best thing to do is to
just type help ode 45 in MATLAB and it gives an excellent introduction to what you need
to do and it in fact, gives a very specific example as well ok.

So, you either in this or you can go to Google and check any of the examples for ode 45 .
Let us see how the simulations look like. So, what I’m doing here is to first in line number
20, I am plotting the equilibrium point with as a star, and then when I start my simulation. So,
you can see that in this particular line over here I have computed the solution of my
dynamical system right. So, the solution basically is H (t ) and L(t) and for H ( t ) in
MATLAB I have called it as x (1) , for L(t) I have called it as x (2) . So, x (2) are
the number of tigers, and x (1) are the number of deers.

So, the solution basically will have 2 columns, the first column will be the number of tigers
the number of deers as a function of time, second column will be the number of tigers as a
function of time, there will be the solution. t is, of course, the time vector which is
returned by this ode 45 , it is an internal MATLAB function. So, now, this small piece of
code it takes each value of my solution, each population of tiger, each population of deer
plots it on the graph, then it holds the graph moves to the next value of the tiger population,
deer population again plots it and so on. It’s basically an iterative process this final thing over
here it plots the time series of x and y , like what I told you before in the slide. So, this
will basically part plot the population of tigers as a function of time on the x-axis, and
population of deers as a function of time ok.

So, we’ll see how all of this looks like; now the interesting thing about this particular piece of
code is that it allows me to actually see how this plot actually rotates or evolves. So, let’s go
and click on run and please remember this code will be available for all of you; you can see
that this is how the dynamics is actually evolving from the particular initial condition which
we have chosen of 10 and 3 , which is somewhere over here right we started at this
point and then it evolves like this.

1000
(Refer Slide Time: 31:28)

So, this is the closed orbit which I was talking about initial condition was somewhere at this
point and then it just goes around in this.

Now, we can see in black I have plotted the number of deer, in blue, I have plotted the
number of tigers and always noted there is a slight fish difference between the 2, and that is
basically we can explain that as follows. So, let’s start at this particular point on the time axis
say around 6 , it could be months or years around 6 and we can see that the number of
deers keeps increasing very fast because there are very few tigers left in the jungle right. So,
the number of deers keeps on increasing, and at some particular point the tigers have now
access to a lot of deers, and it is very easy for them to hunt.

So, then the tiger population also starts to increase alright and when the tiger population starts
to increase and reaches a certain critical number, now the tigers are able to hunt the deers
very well, and the deer population starts to fall over here right. So, this cyclical process keeps
on repeating. For those of you who may actually be a bit skeptical about why these
mathematical models will actually represent the true condition, it has been proven in ecology
that if you sample data for they have done an experiment with hares and foxes, and they have
sampled the data for almost 100 years or even 150 years, the behavior over a large timescale
actually looks like this oscillatory behavior with a slight phase shift alright. Of course, when
you actually look at shorter time intervals the behavior will be fairly different, but in the
larger time scale, they will actually look very similar to this ok.

1001
Now, we started at one particular initial condition 10 and 3 , let us see what happens at
different initial conditions. So, let me just say I have got 6, 7 deers and 5 tigers and
you simulate this you are going to get exactly the same kind of behavior alright. So, this goes
around and round the equilibrium point, this is a stable center. And once the plot is complete
we will get this graph as well, Now, we can see the numbers have changed because the initial
condition of the tigers and properly and the deers were different.

(Refer Slide Time: 34:14)

So, this behavior which is a numerical simulation of a non-linear equation actually agrees
with what we expected to get from our linearized model, where we predicted that we are
going to get a stable center and the simulation of the non-linear dynamics has shown exactly
the same behavior.

1002
(Refer Slide Time: 34:26)

So, in the previous example, we have seen a couple non-linear differential equation at the
predator-prey model, and we saw an example of what is called a stable center which is a
particular type of equilibrium point. We also showed that with respect to the equilibrium
point the predictions which we make true or also we have shown that in the simulations we
get the same prediction of the stable center alright.

So, in that particular case we were it was nice that the predictions did with the linearized
model did agree with the non-linear case around the equilibrium point. So, let’s move on to
another example of the van der pol oscillator, and this is basically an example of an oscillator
with non-linear damping which you can see in this particular term over here right. So, if you
think of this as a mass-spring-damper or a RLC circuit. So, you would see that roughly

md 2 x
speaking this would look something like this we’re talking of the differential
dt2

dx
equation let’s say +b +kx =0 right. So, this would be the example of a mass-spring-
dt
damper, in this particular case we have just assume m=1 , k =1 and this damping term
is of a specific kind it’s of this alright it is a non-linear damping term and there have been
various applications of the van der pol oscillator.

So, people in biology when they are studying the behavior of neurons in brain under certain
task-related conditions, they have actually shown that you can actually do the modeling with

1003
the van der pol oscillator it reasonably agrees with the experimental data as well. And the
classical engineering example for the van der pol oscillator is a tunnel diode, this is basically
a diode which allows us to do very high speed switching in electrical circuits. So, now, let’s
take this particular non-linear differential equation, we see that this is where we get the non-
linear term we like will again repeat the same process and see what kind of behavior we can
extract out of this one ok.

(Refer Slide Time: 36:49)

So, before we do that let’s simplify the problem one step before and what we will do is to
assume that this parameter μ=0 right. So, when we assume that μ=0 this entire non-
linear damping term actually goes away, and you basically get this expression

d2 x
+ x=0 , and we will now show very quickly the, this is basically the dynamics of a
d t2
harmonic oscillator. So, let us model this as a state-space system. So, what we are going to do
is to introduce 2 variables as what we did before as well it’s a second-order system. So, we
will introduce 2 variables, we will call the first one as x 1 and I will just denote that to be

dx
x itself, we will call the second as x 2 and we will call it as alright.
dt

d2 x d2 x
So, from this again as we did before we see that x́ 1=x 2 and x́ 2= . And
d t2 d t2
now we can write by looking at this expression over here as basically nothing, but −x and

1004
what is x ? we use this expression over here and x=x 1 . So, if I put this now in a state-

d x1 d x2
space form we actually get , is, of course, the states on the right-hand side
dt dt

and the interconnection matrix will be nothing, but [ 0 1

−1 0 ] .

So, if you actually do the modeling of a mass-spring system or if you do the modeling of LC
circuit, okay so no resistance in the LC circuit, no damping in the mass-spring system you
will actually get exactly the state-space equations, and it is very easy to show that the
behavior of the state-space system is an oscillatory behavior. So, if you plot any one of the
states as a function of time, you will actually get a perfectly oscillatory behavior alright. In
fact, a small exercise for all of you, please compute the Eigenvalues of this system, and you
will see that Eigenvalues will lie on the imaginary axis. And when the Eigenvalues lie on an
imaginary axis we have seen before it produces the behavior of a centre right, a closed orbit.

So in fact, if you plot the phase space and remember the phase space basically is a plot of the
states, it’s very nice to plot this because it actually gives very interesting insight into the
problem you plot the phase space of this and it’s going to look exactly like this assuming we
start at some initial condition over here. So, at a different initial condition and you are going
to get this kind of response so on and so forth right. So, what I just did in this slide was to
consider the van der pol oscillator, we remove the non-linear damping term over here by
assuming that μ=0 , and we get this expression over here this expression is basically the
dynamics either for a simple mass-spring system or LC circuit. In the state-space form it
looks like this, not really very relevant in this particular discussion, but this the structure of
this matrix this is actually called a skew-symmetric matrix.

The structure of this matrix is very interesting and it leads to really beautiful insight into the
analysis of such systems, we won’t be needing it here, but if anyone of you is interested just
look at skew-symmetric and property of energy conservation alright. And you see that indeed

this structure the [ 0 1

−1 0 ] actually leads to really interesting properties. So, we’re done

with a harmonic oscillator which is a special case of the van der pol oscillator.

1005
(Refer Slide Time: 41:19)

Now let’s not assume μ=0 , let us assume that μ=1>0 . Specifically, μ=1 ; and
when we assume μ=1 and we compute from the differential equation we compute the
state-space model same way as we did before. So, we will since it’s a second-order system we
will introduce 2 new variables x 1 and x 2 , and we will relate these to each other through
the interconnection matrix, and when you do that you are going to get this state-space model.

d x1 d x2
Now, let’s do, ¿ x2 , is this one let’s see what we can do with this guy.
dt dt

(Refer Slide Time: 42:03)

1006
So, we have the model here which I am recalling in these 2 expressions and for this, we can
compute the Jacobian. Now because it is a second-order state-space system we will have a
2 X2 Jacobian, just to help you calculate this it will be a 2 X2 Jacobian and you will
essentially just compute the partial derivatives as follows right. For f1 we take the partial
derivative with respect to x1 and x 2 , for f2 we take the partial derivatives with
respect again to x1 and x 2 ; and when you do that you are actually going to get this
expression.

So, this is my f 1 ( x 1 , x 2 ) , in this case, x1 term doesn’t exist, this is f 2(x 1 , x 2 ) . So,
when you take the partial derivative of this expression with respect to x 1 since x 1 term

doesn’t exist this is nothing, but 0 and I take it with respect to x 2 you simply get 1 .
Similarly, when you take the partial derivative of f2 this expression with respect to x1

you would get −1−2 x 1 x 2 ; because you take a derivative of this which is basically 2 x1
then you have the negative sign and finally, when you take it with respect to x 2 , you just
2
have this term over here this comes out and you get 1−x 1 . So, this is my Jacobian which
we also have over here.

(Refer Slide Time: 43:48)

1007
Now, of course, we have the Jacobian and in order to go ahead and linearize the system, we
would also need the equilibrium points. In order to get the equilibrium points, we do the same

d x1
trick as we did before it’s not really a trick it’s a procedure we just set ¿ 0 , and we
dt

d x2
set ¿0 like I set each one of these equal to 0 and then when you compute the
dt
Jacobian the equilibrium point you will see that this is the solution of the above 2 equations.
There is no other solution in this particular case x 1 has to be equal to 0 , x 2 has to be
equal to 0 then you take this equilibrium condition substituted back in the Jacobian and
you are going to get your wonderful linear system over here in this particular case I have
replaced the variables x 1 with x́ and x 2 with y ok.

So, now let’s analyze the property of this system, and the way we are going to do that is again
as we always do we compute the Eigenvalues, remember in transfer function how we used to
compute the poles that’s really the first step you would always do similarly compute the
Eigenvalues of the system and you will see that the Eigenvalues are on here.

(Refer slide time: 45.22)

1008
So, if you look at the s plane or the pole-zero map. You will find that the Eigenvalues are
over here, okay and what do we learn what have we learned from basic controls that if the
poles are in the right half-plane you actually have an unstable system.

So, what we have basically done is to consider a non-linear system, linearize it around an
equilibrium point get this linear system and we have shown that the linear system here is
unstable. Now it is not unstable in any random way it is unstable specifically around the
equilibrium point that we have considered, and this is an interesting distinction because what
I mean by unstable around the equilibrium point is not the same as unstable at the equilibrium
point. So, you will see very quickly that let us say that the equilibrium point was (0,0)

right. So, (x , y ) equilibrium point. So, I will call it ( x 0 , y 0 ) if I start at (0,0) as the
initial condition as per my definition I should be stuck there forever and you will be stuck
there forever, there is absolutely no way that you are going to escape from the equilibrium
point.

So, it is not unstable at the equilibrium point, it is merely unstable around the equilibrium
point. So, if I start anywhere else around that equilibrium point, we are going to see what
actually happens and actually, you will see some very interesting behavior that although your
linear system claims it is unstable, and you will see very weird responses when you look at
the transfer function, which I will be doing in a couple of minutes. It is not unstable in the
sense that this is going to blow up right it is not like your oscillator is going to heat and blow
up or cause some catastrophe, it is actually going to do a very interesting it has a very
interesting behavior. Start anywhere around the equilibrium point and you will see that it goes

1009
into what is called as a limit cycle. Start anywhere around the equilibrium point and it will
always reach the limit cycle and. In fact, in this specific example, we will see that you start
anywhere else in your phase space any initial condition, it will always converge to this limit
cycle and we will see this in a simulation.

So, the limit cycle is a very interesting behavior it is actually telling us that, wherever you
start in the system anywhere you will always converge on to this limit cycle except if you
start at the equilibrium point because if you start at the equilibrium point you will be stuck
there forever.

(Refer Slide Time: 48:25)

So we will see indeed that this behavior which plotted over here, this is actually what is
called as a limit cycle and you start at any initial condition wherever you want except at the
equilibrium, you will always tend to the limit cycle ok.

So, let’s see how this actually looks like in MATLAB. What we saw with the van der pol
oscillator. So, we saw that the Eigenvalues are in the right half-plane and which means the
equilibrium point is unstable. Now I have got a small MATLAB code again like the predator-
prey problem we will be sharing the code with all of you, let me show one specific thing that
let us actually start the simulation at the initial condition. So, that I will show you that if you
started at (0,0) the initial condition, you will always stay there only because this is the

1010
equilibrium point; like we did with the predator-prey mob model we again use this ode 45
the same type as we did before, what will change, of course, is the model of the system good.

So, if I start the simulation by assuming the initial condition is exactly (0,0) the
equilibrium point, you will see that there is really no dynamics right everything is exactly at
(0,0) . Perfectly there it never moves anywhere else now, we have said that this is an
unstable equilibrium point. So, let me just perturb a little bit the initial condition along x-
direction, I can do it along y it has the same behavior. So, when I perturb it along the x along
the x-direction by this factor of 10 power minus 6 or 7 whatever this is and then I do the
simulation because it is an unstable equilibrium point, it is not supposed to stay there right
and that is exactly what happens. So, let me. So, we started off slightly away from the
equilibrium point at this particular place, where my mouse is pointing, and because it is an
unstable equilibrium point it actually spirals out of that.

Now, the reason it is spiraling out of that is because it needs to spiral and go towards this
larger object over here, this larger object is a closed orbit this is the limit cycle, and it will
always go to the limit cycle. So, this is this basically proves that your equilibrium point is not
stable. What happens if we start at different initial conditions right? So, for example, let me
start at 3 point whatever that is right, maybe I can make this as 2 and we will see what
happens. Will it go to the equilibrium point? It will not, we know it’s unstable, where will it
go? Well, it again tends to the, to the limit cycle alright. So, I started over here, and then it
tends to the limit cycle. In the previous case, we started almost close to the equilibrium point
and even that went into the limit cycle. Let’s try this for a lot of initial conditions and we will
actually see a family of solutions for each of the initial condition ok.

So, what I have done here is to take 20 possible choices of initial conditions and from each of
this values 0, 1, -1, 3, -4, and 4. I am actually choosing a value for the initial condition. So, in
my case an initial condition could, for example, be (0, 4) it could be, for example,
(0, 4) right, it could also be (−1,−2) , it could be (−1, 4 ) , it could be (0, 3) so
on and so forth. For each of these initial conditions when you actually compute the solution
and plot it you are actually going to get this.

So, this beautiful looking graph basically tells me the following wherever you start from
whether this is the initial condition, this is the initial condition or this is the initial condition
and so on, wherever you start from the final simulation will always tend to the limit cycle

1011
every single time. This is true as long as μ>0 if μ<0 , then the if μ=0 we see that it
will just be the; a perfect circle which is a perfect oscillation because it is a harmonic
oscillator. So, we also see that as long as the values of x and y are larger than 0 ,
either x or y are larger than 0 it will always tend to the limit cycle, if both x
and y are equal to 0 it will get stuck at the equilibrium point ok.

So, there is actually one specific case where we got x=0 , y=0 initial condition and it
is stuck at this equilibrium point this brown color plot. Otherwise, you see this beautiful
looking graph ok. So, we have seen the example simulation in MATLAB that we obtain what
are called limit cycles, as we see in this plot as well we have seen the notion of a stable limit
cycle, we have also seen the notion of a unstable equilibrium point. So, that concludes lecture
part 2 of lecture 2 which was focused on non-linear systems and the linearization of non-
linear systems ok.

So, the linearization of non-linear systems we have taken 2 specific examples the predator-
prey problem, and this van der pol oscillator both are non-linear systems and we have shown
how we can linearize them and what interesting behavior they actually get. So, thank you
very much and if you want additional references they are over here alright.

Thank you and good luck.

1012
 Control Engineering
Dr. Ramkrishna Pasumarthy
Department of Electrical Engineering
Indian Institute of Technology, Madras

Module - 12
Tutorial - 1
Lecture - 52
State Space Canonical Forms

Hello everyone. In the previous lectures, we were looking at State Space Models, we were
introduced to them.

(Refer Slide Time: 00:25)

So, we looked at the standard form of the state space model as 𝒙(𝒕̇ ) = 𝑨𝒙(𝒕) + 𝑩𝒖(𝒕);
and output 𝒚(𝒕) = 𝑪𝒙(𝒕) + 𝑫𝒖(𝒕). So, here 𝒙 is the state vector, and 𝒖 is the input vector,
𝒚 is the output vector. And 𝑨 is the dynamics matrix; 𝑩 is the input matrix; 𝑪 is the output
matrix; and 𝑫 is the feed forward matrix. So, these are the standard names for these
matrices.

And so there are two things that we can do. So, given a state space model, we can go back
to the transfer function model; and given a transfer function we can go to the state space
model. So, earlier we looked at how can we go from the state space model to the transfer
𝒀(𝒔)
function model using this formula. So, = 𝑪(𝒔𝑰 − 𝑨)−𝟏 𝑩 + 𝑫.
𝑼(𝒔)

1013
So, given a state space model with the matrices 𝑨, 𝑩, 𝑪 and 𝑫, this is how you can
transform it into a transfer function. And this transformation is unique. So, any state space
model given you can only determine one transfer function, so that is the case of going from
state space to transfer function

Now, what about the other way? So, when we go from transfer function to state space, the
transformation is not unique. So, that means, when you are given one state space model
you can find out multiple transfer function models. So, instead of finding some kind of
random transfer, some random state space model, we will try to find out some standardized
state space models which we call canonical forms.

And so today we will discuss about three canonical forms. First one is controllable
canonical form, the second one is the observable canonical form, and the third one is the
diagonal canonical form. So, while we discuss about each of these forms, we will see why
these names come.

So, first one will be the controllable canonical form. So, to derive this, we’ll start with a
strictly proper transfer function on this form, sorry. So, this is strictly proper transfer
function, because the numerator has a degree of 𝒏 − 𝟏, and the denominator has a degree
of 𝒏, that is the number of poles are greater than the number of zeros. So, what happens in
this transfer function is 𝒕 = 𝟎 ok.

So, when we are given this transfer function, how do you find out a controllable canonical
form of the state space model? So, in case you have a transfer function which is not strictly
proper something like you can have here say 𝒃𝟎 𝑺𝒏 . So, in that case, what you can do is
you can simply perform polynomial division, and take this 𝒃𝟎 out. So, that it will come
something like this, and rest again remains in the same form, we can again apply the same
transformation that we are doing now.

So, for now we will stick to a strictly proper transfer function and go ahead with it. So,
given this strictly proper transfer function, how do we convert it into a controllable
canonical form? So, what we’ll do is, we’ll deal with the numerator and the denominator
𝒀(𝒔)
separately. So, for that, I will take this as is equals to I will introduce a new polynomial
𝑼(𝒔)
𝒀(𝒔) 𝒁(𝒔) 𝒀(𝒔)
called 𝒁(𝒔). So, can be written as .
𝑼(𝒔) 𝑼(𝒔) 𝒁(𝒔)

1014
So, I am just multiplying and dividing by a new polynomial 𝒁(𝒔). And this will be taken
as the denominator part, and this part will be the numerator. So, I will just write it down.
𝒁(𝒔) 𝟏 𝒁(𝒔) 𝒀(𝒔)
So, = , so that will be . And then will be just the
𝑼(𝒔) 𝑺𝒏 +𝒂𝟏 𝑺𝒏−𝟏 +....+ 𝒂𝒏 𝑼(𝒔) 𝒁(𝒔)

numerator. So, these are not any transfer function. We are just introducing them and divide
what you say dividing them into two parts just for the sake of calculation convenience.

(Refer Slide Time: 06:59)

So, now, first we will take the first part, and we will do some algebraic manipulations as
follows. So, I can take 𝒁(𝒔) times I can do cross multiplication on this, and then I will get
𝒁(𝒔)[𝑺𝒏 + 𝒂𝟏 𝑺𝒏−𝟏 +. . . . + 𝒂𝒏 ] = 𝑼(𝒔).

So, now what I will do is, I will take inverse Laplace transform. So, when I do inverse
Laplace transforms, 𝒁(𝒔) = 𝒛. And so and it will be ok, I will just first multiply it
𝑺𝒏 𝒁(𝒔) + 𝒂𝟏 𝑺𝒏−𝟏 𝒁(𝒔)+. . . . + 𝒂𝒏 𝒁(𝒔) = 𝑼(𝒔).

Now, when we apply Laplace inverse, 𝑺𝒏 = 𝒛(𝒏) (nth derivative of z). So, I can write it as
𝒛(𝒏) in the brackets. So, this is a notation to denote that we are taking the derivative of
𝒛(𝒏) times, and so we will just follow the same thing 𝒂𝟏 𝒛(𝒏−𝟏) derivative of 𝒛 and so on up
to 𝒂𝒏 𝒛 is equals to 𝒖 (i.e., 𝒛(𝒏) + 𝒂𝟏 𝒛(𝒏−𝟏) + ⋯ + 𝒂𝒏 𝒛 = 𝒖). So, when we take the Laplace
inverse, we are representing everything in small letters. So, 𝒁(𝒔) becomes small 𝒛, and
𝑼(𝒔) becomes small 𝒖.

1015
Now, on this differential equation, we will be defining our state variables. So, here is how
we define? So, I will define my 𝒙𝟏 I need to define 𝒏 state variables. So, 𝒙𝟏 I will take it
as 𝒛, and 𝒙𝟐 I will take it as 𝒙𝟏̇ will which will be equal to 𝒛̇ , so that’s the first derivative

of 𝒛. And then 𝒙𝟑 will be 𝒙𝟐̇ which will be equal to 𝒛̈ , and I will go on up till 𝒙𝒏 is equals

̇ , which will be equals to (𝑛 − 1)th derivative of 𝒛. And 𝒙𝒏 , I will write here 𝒙𝒏̇ is
to 𝒙𝒏−𝟏

equals to the 𝒛(𝒏).

So, this 𝒛(𝒏) , I will take it from here. I will take all these terms and send them to the other
side, and write it here. So, that will be 𝒛(𝒏) = 𝒖 − 𝒂𝟏 𝒛(𝒏−𝟏) … − 𝒂𝒏 𝒛. Now, you can see
that we have 𝒛(𝒏−𝟏), 𝒛(𝒏−𝟐) , … 𝒛, and all these variables are already defined in terms of
the state variables. So, we will just substitute them here, and my 𝒙𝒏̇ can be written as
𝒖 − 𝒂𝟏 𝒙𝒏 − 𝒂𝟐 𝒙𝒏−𝟏 … − 𝒂𝒏 𝒙𝟏 .

So, now you can see that we have represented all the state variables in terms of 𝒙 now. So,
now, we have a set of equations. So, what are those equations? The first one is 𝒙𝟏̇ = 𝒙𝟐 ,
𝒙𝟐̇ = 𝒙𝟑 so on 𝒙𝒏−𝟏
̇ = 𝒙𝒏 , and 𝒙𝒏̇ is this expression. So, using all these equations, we
can write the state space model.

(Refer Slide Time: 11:28)

1016
 𝒙𝟏̇
𝒙̇
So, the state space model will be 𝒙̇ = [ 𝟐 ]. So, what is 𝒙𝟏̇ ? It is simply 𝒙𝟐 . So, I can write
⋮
𝒙𝒏̇
𝒙𝟏
𝒙𝟐
it 0 1 ok, I will just have to get a matrix here times [ ⋮ ].
𝒙𝒏

Now, you can see that this will be 0, this will be 1 and rest all again will be 0’s, because
𝒙𝟏̇ = 𝒙𝟐 . Now, 𝒙𝟐̇ = 𝒙𝟑 . Now, you will get a 1 here and rest all will be 0’s. And similarly,
finally, 𝒙𝒏̇ will be 𝑎𝑛 minus 𝑎𝑛 𝑥1 . So, here it will be −𝒂𝒏 −𝒂𝒏−𝟏 … . − 𝒂𝟏 , so none of
these first (𝑛 − 1) state variables have input. So, this will be all 0’s except at the end times
𝒖.

Now, if you simply multiply these matrices and observe, you will get those set of equations
exactly same ok. So, this is what we call the controllable canonical form. So, and this will
be 𝑨; this will be 𝑩. And you can see output 𝒚, so to get the output 𝒚 we have to use the
other set of equations.

So, I can take this and say 𝒀(𝒔) = 𝒃𝟏 𝒔𝒏−𝟏 𝒁(𝒔) + ⋯ … + 𝒃𝒏 𝒁(𝒔). So, here again I can take
the Laplace inverse, and just use those equations. Then my 𝒀(𝒔) will simply be, when we
take the Laplace inverse, it will be just 𝒚 or 𝒚(𝒕) = 𝒃𝟏 𝒙𝒏 + 𝒃𝟐 𝒙𝒏−𝟏 + ⋯ . . +𝒃𝒏 𝒙𝟏. So, this
will be the equation for 𝒚(𝒕).

And in the matrix form I can just write it as 𝒃𝒏 , sorry, this is 𝒏, 𝒃𝒏 . So, in the matrix form,
I can write it as 𝒚 = [𝒃𝒏 𝒃𝒏−𝟏 … . . 𝒃𝟏 ]𝒙. So, 𝒙 is the state vector. So, this matrix will
become my output matrix 𝑪. So, now you have 𝑨, 𝑩, 𝑪, and 𝑫 is equals to 0. So, when we
use these matrices, the state space model that we get is the controllable canonical form.

1017
(Refer Slide Time: 15:33)

So, why is it called the controllable canonical form? I will draw a small diagram, and you
will be able to see. So, we will draw a block, small block diagram to represent this. So, I
will start with state 𝒙𝟏 and so here I put an integrator block. So, this will be 𝒙𝟏̇ , because
when I integrate 𝒙𝟏̇ , I will get 𝒙𝟏 . And this 𝒙𝟏̇ is nothing but 𝒙𝟐 .

So, similarly I can put a series of integrator blocks and get up to 𝒙𝒏 , and then I add another
integrator block to get 𝒙𝒏̇ . And to get the value of 𝒙𝒏̇ , so as you can see 𝒙𝒏̇ is summing up
or over all the state variables with certain coefficients. So, we need to put a summer here
and with the input, because it is 𝒖 plus all this. So, 𝒖 − 𝒂𝟏 𝒙𝒏 − ⋯ − 𝒂𝒏 𝒙𝟏 .

So, my 𝒙𝒏 is here I can just take a feedback from here and put −𝒂𝟏 . And similarly it will
be another feedback from 𝒙𝟐 which will have a coefficient −𝒂𝒏−𝟏 . And similarly from 𝒙𝟏 ,
we will have another feedback with the coefficient −𝒂𝒏 .

So, just forgetting the output for a while we can see that the input 𝒖 is passing through all
the state variables 𝒙𝒏 , 𝒙𝒏−𝟏 , 𝒙𝒏−𝟐 , ……, 𝒙𝟐 , 𝒙𝟏 . So, output 𝒚 will actually come out here
somewhere. So, it can be observed that input has a control over all the states in the system
which is clearly observed to this block diagram, and that can be clearly seen in this matrix.
So, this is the reason why we call the controllable canonical form, because the input has
control over all possible, all existing states and that can be clearly seen.

1018
So, if you want the output, you can further extend the block diagram by adding 𝒃𝒏 here
and so output 𝒚 sums over 𝒃𝒏 𝒙𝟏 and so it will be 𝒃𝒏−𝟏 𝒙𝟐 and so on up to 𝒃𝟏 𝒙𝒏, so that is
the complete block diagram of the state space model. And you can clearly see why it is a
controllable canonical form.

(Refer Slide Time: 18:48)

So, now we look at the observable canonical form ok. Now, coming to the observable
canonical form. This actually becomes very simple once you know the controllable
canonical form. So, once say 𝑨, 𝑩, 𝑪, and 𝑫 are the matrices pertaining to the controllable
canonical form. So, the matrices pertaining to the observable canonical form can be written
̅, 𝑩
as follows. I will call them as 𝑨 ̅, 𝑪
̅ , and 𝑫
̅.

̅ is nothing but 𝑨𝑻 ; 𝑩
So, 𝑨 ̅ is nothing but 𝑪𝑻 ; 𝑪
̅ is nothing but 𝑩𝑻 ; and 𝑫
̅ is nothing but

𝑫𝑻 . So, this will be the observable canonical form you. So, you can just find out the
controllable canonical form, and take the transforms of those matrices, and arrange them
in this manner to get the observable canonical form ok.

1019
(Refer Slide Time: 20:21)

So, now we look at the diagonal canonical form. So, again we will start with the strictly
perfect transfer function. So, what we will do is, we will take this denominator and
factorize them, factorize the 𝒏 roots by writing it in this form. So, 𝑷𝟏 , 𝑷𝟐 , 𝑷𝟑 , … . , 𝑷𝒏 are
the roots of the denominator polynomial. And so we can write it where we can write the
denominator in this form (𝒔 + 𝑷𝟏 )(𝒔 + 𝑷𝟐 ) … . . (𝒔 + 𝑷𝒏 ).

Now, what we’ll do is, we’ll apply partial fractions. And I can write it in this form
𝑪𝟏 𝑪𝟐 𝑪𝒏
+ + ⋯…+ . Now, what we will do is we will send 𝑼(𝒔) to the other side,
𝒔+𝑷𝟏 𝒔+𝑷𝟐 𝒔+𝑷𝒏
𝑼(𝒔)𝑪𝟏 𝑼(𝒔)𝑪𝟐 𝑼(𝒔)𝑪𝒏
and say + + ⋯…+ ok. So, now I’ll define each of these parts,
𝒔+𝑷𝟏 𝒔+𝑷𝟐 𝒔+𝑷𝒏

leaving the 𝑪’s, I’ll define that as 𝑿𝟏 (𝒔).

1020
(Refer Slide Time: 23:21)

𝑼(𝒔)
So, 𝑿𝟏 (𝒔) = 𝒔+𝑷 . So, I will cross multiply and now I take the inverse Laplace. So, when
𝟏

we take the inverse Laplace of 𝒔𝑿𝟏 (𝒔), it will be just 𝒙𝟏 . So, earlier also and here also,
when we do the inverse Laplace, we assume that all the initial conditions are 0. So, you
should remember that when we apply this inverse Laplace we are assuming that all the
initial conditions are 0. So, 𝒔𝑿𝟏 (𝒔) when take in the inverse it becomes just
𝒙𝟏̇ + 𝑷𝟏 𝒙𝟏 = 𝒖. So, these are all again time based variables. So, I can just write
𝒙𝟏̇ = 𝒖 − 𝑷𝟏 𝒙𝟏.

So, similarly I will take each of these as 𝑿𝟐 (𝒔), 𝑿𝟑 (𝒔),…. 𝑿𝒏 (𝒔). And just repeat the same
process, and I will get 𝒙𝟐̇ = 𝒖 − 𝑷𝟐 𝒙𝟐 ....... 𝒙𝒏̇ = 𝒖 − 𝑷𝒏 𝒙𝒏 . So, now, I have all the state
variables and their derivatives Now, I can write the canonical form I will just write the
matrices directly. So, 𝑨 will be 𝑷𝟏 , 𝑷𝟐 , … . , 𝑷𝒏 along the diagonal and all the terms will be
0’s. So, this is the diagonal matrix, so that is the reason why we have the name diagonal
canonical form.

And then 𝑩 will be all 1’s, because in every 𝒙𝟏̇ , 𝒙𝟐̇ , ….., 𝒙𝒏̇ we have a 𝒖. So, it will be all
just 1’s times ok, 𝒖 won’t come here, it’s just 𝑩. And 𝑪 will be, so you can see why this
𝑿𝟏 (𝒔)𝑪𝟏 + 𝑿𝟐 (𝒔)𝑪𝟐 … . +𝑿𝒏 (𝒔)𝑪𝒏 . So, C n, 𝑪 will be just 𝑪 = [𝑪𝟏 , 𝑪𝟐 , … 𝑪𝒏 ], and 𝑫 will
be anyway 0. So, these are the matrices pertaining to the diagonal canonical form. So, now
we’ll try to look at an example. So, we will just solve one example and try to derive all the
three forms.

1021
(Refer Slide Time: 26:11)

𝟏
So, the transfer function is 𝑮(𝒔) = ok. So, this is the transfer function that
𝒔𝟑 +𝟔𝒔𝟐 +𝟏𝟏𝒔+𝟔
we take. And so when we are deriving the canonical forms, there are two ways of going at
it. One way is actually deriving all the state variables, and other way is remembering what
we got the matrices 𝑨, 𝑩, 𝑪, 𝑫, and just trying to substitute the coefficients of these
polynomials into them. So, you can do it either way.

But if you try to remember the formula, it might become bit complicated, because you
need to know which coefficient pertains to which of those. Because even in the textbooks
some people use it in a different way, some people use it from 𝒃𝟎 to 𝒃𝒏 , and some people
do it from 𝒃𝒏 to 𝒃𝟎 ; and also in the denominator people also use it in the reverse.

So, when you remember the formulas, it might become a bit difficult. So, what I will do is
I will try to derive the state variables using the way that we did earlier. So, I will just take
𝒀(𝒔)
𝒀(𝒔). So, this I will take it as . And so I can write 𝒀(𝒔)(𝒔𝟑 + 𝟔𝒔𝟐 + 𝟏𝟏𝒔 + 𝟔) = 𝑼(𝒔).
𝑼(𝒔)

So, I just did cross multiplication.

Now, I will apply Laplace inverse assuming 0 initial condition. So, it will be
⃛ + 𝟔𝒚̈ + 𝟏𝟏𝒚̇ + 𝟔𝒚 = 𝒖. So, now, I got the differential equation. Now, I can define my
𝒚
state variables. I will define 𝒙𝟏 = 𝒚; 𝒙𝟏̇ = 𝒙𝟐 = 𝒚̇ ; 𝒙𝟐̇ = 𝒙𝟑 = 𝒚̈ ; and 𝒙𝟑̇ = 𝒚
⃛ . So, we just
have three state variables 𝒙𝟏 , 𝒙𝟐 and 𝒙𝟑 .

1022
Now, we can substitute them here? So, actually we already know 𝒙𝟏̇ and 𝒙𝟐̇ . And 𝒙𝟑̇ = 𝒚
⃛
which I will find out from this equation; get it as 𝒖 − 𝟔𝒚̈ − 𝟏𝟏𝒚̇ − 𝟔𝒚. And again 𝒚̈ , 𝒚̇ , and
𝒚, I can substitute it from these. So, I will get 𝒖 − 𝟔𝒙𝟑 − 𝟏𝟏𝒙𝟐 − 𝟔𝒙𝟏. So, now I have the
values of 𝒙𝟏̇ , 𝒙𝟐̇ , and 𝒙𝟑̇ .

(Refer Slide Time: 29:33)

𝟎 𝟏 𝟎 𝟎
[
Now, you can simply write the model as 𝒙 = 𝟎 𝟎 𝟏 𝒙 + 𝟎] 𝒖. And 𝒚 is just 𝒙𝟏
] [
−𝟔 −𝟏𝟏 −𝟔 𝟏
in this case, so it will be just 𝒚 = [𝟏 𝟎 𝟎]𝒙. So, this is the controllable canonical form.
As you can see it is very easy if you can derive it out directly instead of remembering the
formulas. So, now we have the controllable canonical form. So, I will say I will call this
ok, this is the controllable form.

So, what about the observable form? Observable form will be simply 𝒙 is equals to so as
𝟎 𝟎 −𝟔 𝟏
𝑨 = 𝑨𝑻 , I just had to take the transpose of this, so it will be 𝒙 = [𝟏 𝟎 −𝟏𝟏] 𝒙 + [𝟎] 𝒖
𝟎 𝟏 −𝟔 𝟎
𝟏
plus 𝑩 = 𝑪𝑻 = [𝟎] 𝒖. And 𝑪 = 𝑩𝑻 , so it will be 𝑦 = [𝟎 𝟎 𝟏]𝒙. So, this is the observable
𝟎
form.

1023
(Refer Slide Time: 31:20)

So, finally, the diagonal form. So, to get the diagonal form, we need to find the roots of
the denominator polynomial. So, I am just I already found them to be −𝟏, −𝟐, −𝟑, you can
𝒀(𝒔) 𝟏
verify. So, I can write my = . And this I will when I apply partial
𝑼(𝒔) (𝒔+𝟏)(𝒔+𝟐)(𝒔+𝟑)
𝟏⁄ (−𝟏) 𝟏⁄
𝟐 𝟐
fractions, I will get + + . So, this also you can verify yourselves.
𝒔+𝟏 𝒔+𝟐 𝒔+𝟑

And so, now, we have this partial fractions my 𝑨 is nothing but these three coefficients
sorry these three roots coming in the across the denominator, so it will be
−𝟏 𝟎 𝟎 1
𝑨 = [ 𝟎 −𝟐 𝟎 ]. And then 𝑩 is all just 1’s (i.e., 𝐵 = [1]). And 𝑪 is the numerators of
𝟎 𝟎 −𝟑 1
the partial fraction, so it is 𝑪 = [𝟏⁄𝟐 −𝟏 𝟏⁄𝟐]. So, that is this is the diagonal canonical
form.

So, we just took one example and we try to derive all the three canonical forms. So, there
might be some special cases where you might not be able to derive the diagonal form,
because there can be repeated roots and not unique roots as in this case. So, that is when
you go to something called a Jordan canonical form which we are not doing as of now.

So, these are the canonical forms which we discuss. So, I will try to put up some more
𝒔+𝟑 𝒔𝟐 +𝟑𝒔+𝟑
problems which you can try out as an exercise (ii) , and (iii) . So, you can
𝒔𝟐 +𝟑𝒔+𝟐 𝒔𝟐 +𝟐𝒔+𝟏
just try out these two examples and you can see that this last example actually has a non-

1024
zero 𝑫. So, you need to perform polynomial division first, and then apply whatever we did
until now. So, you can just try out those, and maybe we will try to give them as assignment
problems.

Thank you.

1025
 Control Engineering
Dr. Ramkrishna Pasumarthy
Department of Electrical Engineering
Indian Institute of Technology, Madras

Module - 12
Tutorial - 2
Lecture - 53
State Space Solution and Matrix Exponential

Hi everyone, today we’ll be looking at how to solve for the states in State Space System.
And in the process we’ll also introduce what is called Matrix Exponential and we’ll see
how to solve for the Matrix Exponential.

(Refer Slide Time: 00:37)

So, let’s start with a simple scalar system 𝒙̇ = 𝒂𝒙. Now, if we want to get a closed form
expression for 𝒙(𝒕) then what we do is just integrate this. So, we have a
𝒙(𝒕) 𝒅𝒙 𝒕
∫𝒙(𝟎) = 𝒂 ∫𝟎 𝒅𝒕. So, you solve this and you get 𝒙(𝒕) = 𝒙(𝟎)𝒆𝒂𝒕 .
𝒙

So, this is simple because it’s a scalar homogeneous system without any inputs. So, given
a general LTI system 𝒙̇ = 𝑨𝒙 + 𝑩𝒖 we want to arrive at a closed form expression for 𝒙(𝒕)
as some function of the initial states 𝒙(𝟎) and the input 𝒖(𝒕) and of course, this system
matrices.

1026
Now, how do we go about doing this is to take the Laplace transform of the state space
representation. So, we get 𝒔𝑿(𝒔) − 𝒙(𝟎) = 𝑨𝑿(𝒔) + 𝑩𝑼(𝒔). And we all know we can solve
for 𝑿(𝒔) as (𝒔𝑰 − 𝑨)−𝟏 𝒙(𝟎) + (𝒔𝑰 − 𝑨)−𝟏 𝑩𝑼(𝒔).

Now, to take this back to the time domain we take the inverse Laplace transform. In order
to do that we need to know what the inverse Laplace transform of this function here is. So,
we use this expression that. Now, you can check for yourselves that this is true because if
you pre multiply by (𝒔𝑰 − 𝑨) then you get identity. So, using this
𝑨𝟐 𝒕 𝟐
𝑳−𝟏 ((𝒔𝑰 − 𝑨)−𝟏 ) = 𝑰 + 𝑨𝒕 + +....
𝟐!

𝒂𝟐 𝒕𝟐
Now, this is similar to the scalar exponential expansion which is 𝒆𝒂𝒕 = 𝟏 + 𝒂𝒕 + +.....
𝟐!
So, this is what we call the matrix exponential 𝒆𝑨𝒕 . Note that 𝑨 is a matrix. So, using that
𝒕
and from here we say that 𝒙(𝒕) = 𝒆𝑨𝒕 𝒙(𝟎) + ∫𝟎 𝒆𝑨(𝒕−𝝉) 𝑩𝑼(𝝉)𝒅𝝉.

Now, this 𝒙 this the second part of this expression we get by using the property that
multiplication in the s domain corresponds to convolution in the time domain. So, this is
the closed form expression for your state 𝒙(𝒕) as a function of your input, system matrices
and the initial conditions. So, the question now is how do you compute this exponential,
matrix exponential 𝒆𝑨𝒕 ; because it’s an infinite sum of matrices.

Now, the first method is quite obvious from the way we saw here that 𝒆𝑨𝒕 is just the
𝑳−𝟏 ((𝒔𝑰 − 𝑨)−𝟏 ). So, given a matrix 𝑨 you can just compute (𝒔𝑰 − 𝑨)−𝟏 and take the inverse
Laplace transforms of the elements of the matrix and you get the matrix exponential that
is one method.

The second method is by using a diagonalization. So, here we assume that the matrix 𝑨
has a distinct Eigen values in which case it can be written as 𝑨 = 𝑷𝑫𝑷−𝟏; where your 𝑫 is
a diagonal matrix with entries (𝝀𝟏 , 𝝀𝟐 , … . , 𝝀𝒏 ), where these are the Eigen values of the 𝑨
matrix. So, once you why we do this diagonalization is because we can see that if we
express 𝒆𝑨𝒕 in terms of as something times 𝒆𝑫𝒕 then your 𝒆𝑫𝒕 is easy to compute because
it’s a diagonal matrix.

1027
[(Refer Slide Time: 06:46)

𝑨𝟐 𝒕 𝟐
So, we see how we do that. You know that 𝒆𝑨𝒕 = 𝑰 + 𝑨𝒕 + +.... and now we know
𝟐!
𝑨 = 𝑷𝑫𝑷−𝟏. So, by the way 𝑷 is any invertible matrix that accomplishes this
diagonalization of the matrix 𝑨. You can choose 𝑷 to be the matrix of eigenvectors for a
distinct eigen values yeah. So, 𝑨 = 𝑷𝑫𝑷−𝟏, 𝑨𝟐 = 𝑨. 𝑨 = (𝑷𝑫𝑷−𝟏 )(𝑷𝑫𝑷−𝟏 ) = 𝑷𝑫𝟐 𝑷−𝟏.

Similarly, 𝑨𝟑 = 𝑨𝟐 . 𝑨 = (𝑷𝑫𝟐 𝑷−𝟏 )(𝑷𝑫𝑷−𝟏 ) = 𝑷𝑫𝟑 𝑷−𝟏. So, in general we can say that
𝑨𝒌 = 𝑷𝑫𝒌 𝑷−𝟏. So, going back to the original matrix exponential 𝒆𝑨𝒕 is 𝑰 and we express
(𝑷𝑫𝟐 𝑷−𝟏 ) 𝒕𝟐
the identity matrix as 𝑷. 𝑷−𝟏 plus 𝑨𝒕 and 𝑨 = 𝑷𝑫𝑷−𝟏 into 𝒕 plus 𝑨𝟐 which is
𝟐!
𝟐 −𝟏 𝟐
(𝑷𝑫 𝑷 )𝒕
and so on (i.e., 𝒆𝑨𝒕 = 𝑷𝑷−𝟏 + (𝑷𝑫𝑷−𝟏 )𝒕 + +…).
𝟐!

So, we just do some jugglery here. So, take 𝑷 and 𝑷−𝟏 on either side outside. So, you have
𝑫𝟐 𝒕𝟐
𝒆𝑨𝒕 = 𝑷 (𝑰 + 𝑫𝒕 + 𝟐! + ⋯ ) 𝑷−𝟏. So, we basically have that 𝒆𝑨𝒕 = 𝑷 𝒆𝑫𝒕 𝑷−𝟏 . So,

since these are diagonal matrix we can use this way to compute the matrix exponential 𝒆𝑨𝒕 .
We’ll now look at the third method.

1028
(Refer Slide Time: 09:25)

So, the third method for computing the matrix exponential is using the minimal
polynomial. Now, given a matrix 𝑨 from Cayley Hamilton theorem we know that the
matrix satisfies it’s characteristic equation. So, let’s say 𝒇(𝝀) = |𝝀𝑰 − 𝑨| = 𝟎 is the
characteristic equation of the matrix 𝑨. So, we know that 𝒇(𝑨) = 𝟎. Now, for an 𝑛 𝑋 𝑛
matrix 𝑨 this polynomial is of order 𝑛, but there could be a polynomial of lower order for
which the matrix satisfies.

So, let’s say there is another polynomial 𝝋(𝝀) = 𝝀𝒎 + 𝒂𝟏 𝝀𝒎−𝟏 + ⋯ + 𝒂𝒎−𝟏 𝝀 + 𝒂𝒎 . So,
this 𝒎 ≤ 𝒏 and this polynomial is such that 𝝋(𝑨) = 𝟎. So, it’s called a minimal
polynomial; if the for the smallest 𝒎 for which the matrix satisfies this equation 𝝋(𝑨) = 𝟎

1029
(Refer Slide Time: 11:19)

Now, for all computational purposes we can just use this 𝝋(𝝀) equals |𝝀𝑰 − 𝑨| which is
|𝝀𝑰−𝑨|
nothing but the characteristic equation divided by some 𝒅(𝝀) (i.e., 𝝋(𝝀) = ). Now,
𝒅(𝝀)

this 𝒅(𝝀) is the greatest common divisor of 𝒂𝒅𝒋(𝝀𝑰 − 𝑨).

Now, why do we use this to find exponential of 𝑨𝒕 is we use this formula called the
Sylvester interpolation formula which is which basically says that, 𝝋(𝝀) which is the
minimal polynomial of a matrix 𝑨 assume that it is in this form. And let’s say its roots are
(𝝀1 , 𝝀2 , … 𝝀𝑚 ). So, in that case your matrix exponential can be expressed using this
formula. So, basically this whole thing is a matrix. You find the determinant equate it to 0
and one of the entries is 𝒆𝑨𝒕 . So, we get 𝒆𝑨𝒕 in terms of all the lower powers of 𝑨 until
(𝒎 − 𝟏).

So, we’ll see how these 3 methods work with an example. So, we’ll start with the first
method which is just using inverse Laplace transforms. So, 𝒔𝑰 − 𝑨 and minus 𝑨 to find the
inverse, (𝒔𝑰 − 𝑨)−𝟏 you first calculate the determinant which turns out to be just the
product of diagonal terms |𝒔𝑰 − 𝑨| = (𝒔 − 𝟐)𝟐 (𝒔 − 𝟏 ). So, and then your
𝒂𝒅𝒋(𝒔𝑰−𝑨)
(𝒔𝑰 − 𝑨)−𝟏 = ok.
|𝒔𝑰−𝑨|

1030
(Refer Slide Time: 15:08)

𝒂𝒅𝒋(𝒔𝑰−𝑨)
So, 𝒂𝒅𝒋(𝒔𝑰 − 𝑨) and then you get (𝒔𝑰 − 𝑨)−𝟏 = |𝒔𝑰−𝑨|
which is. And 𝒆𝑨𝒕 is the inverse
𝟏
Laplace transform of this. So, 𝑳−𝟏 (𝒔−𝟐) = 𝒆𝟐𝒕 . Similarly, you take an inverse Laplace

transform of every element in the matrix. We know how to compute the inverse Laplace
transform of this by expanding it in terms of partial fractions. So, that will just be
𝟑(𝒆𝟐𝒕 − 𝒆𝒕 ). So, that’s the matrix exponential computed using inverse Laplace transforms.

We will solve it using the minimal polynomial way as well. So, to do that like I had
|𝝀𝑰−𝑨|
mentioned you find the minimal polynomial first which is . Now, so we know that
𝒅(𝝀)

𝒂𝒅𝒋(𝒔𝑰 − 𝑨) is this. And clearly (𝒔 − 𝟐) is a common factor of this adjoint matrix. So, your
𝒅(𝝀) = 𝒔 − 𝟐 and we know |𝝀𝑰 − 𝑨| = (𝒔 − 𝟐)𝟐 (𝒔 − 𝟏 ).

So, we have 𝝋(𝝀) = (𝝀 − 𝟏)(𝝀 − 𝟐). So, we know our 𝝀’s here 𝝀𝟏 = 𝟏, 𝝀𝟐 = 𝟐 and we use
this formula to calculate 𝒆𝑨𝒕 . So, you can just substitute here. So, here you are 𝒎 = 𝟐
because 𝝋(𝝀) is of order 𝟐. So, you have 1 1.

So, you can solve this determinant equation the way you usually solve for determinants
and get 𝒆𝑨𝒕 in terms of 𝑨 and these exponentials. You can cross check that you get the
same answer as this. Now, going to method 2 which is diagonalization ok.

1031
(Refer Slide Time: 19:40)

So, method 2 including diagonalization the matrix. So, firstly we find out the eigen values
of the matrix 𝑨. We have |𝒔𝑰 − 𝑨| here. The eigen values are 1, 2 and 2. There is a repeated
eigen value which is 2. Now, the matrix 𝑷 which is used for diagonalization is a matrix of
eigen vectors 𝒗𝟏 , 𝒗𝟐 , 𝒗𝟑 . So, we solve for these Eigen vectors for each Eigen value say for
𝝀 = 𝟐: 𝑨𝒗 = 𝟐𝒗 this is just 𝝀. 𝑨𝒗 = 𝝀𝒗 will give you. So, when you solve this you get
2𝒗𝟏 = 2𝒗𝟏 .

𝒗𝟏
So, here we see that our Eigen vector 𝒗 is of this form [ 𝒗𝟐 ]. So, there are no constraints
3𝒗𝟐
on the quantity on the first and the second coordinates. So, you see that the null space for
this eigen value which is 𝝀 = 𝟐 has dimension 2. So, basically what I am saying is that for
eigen value 2 you will have 2 eigen vectors, because 2 quantities are not constrained. So,
you can just put in any values for 𝒗𝟏 and 𝒗𝟐 . I will just choose. So, these are the two
eigenvectors corresponding to the repeated eigen value 2.

1032
(Refer Slide Time: 22:10)

Now, we compute for 𝝀 = 𝟏. So, solving this equation you get that 𝒗𝟏 = 𝟎, 𝒗𝟐 = 𝟎 and
𝟎
𝒗𝟑 ∈ 𝑹. So, we choose our Eigen vector to be [𝟎]. So, your matrix 𝑷 finally, looks like
𝟏
𝟏 𝟏 𝟎
[𝟏 𝟐 𝟎].
𝟑 𝟔 𝟏

So, from here you can compute 𝑷−𝟏 as the regular adjoint by determinant. I will give you
the numbers ok. So, once you have 𝑷 and 𝑷−𝟏 inverse you can your 𝑫 is a nothing but a
diagonal matrix with these Eigen values. So, note that the eigen values need to be in the
same order as your eigen vectors. So, these two eigenvectors correspond to 𝝀 = 𝟐

So, 2 and 2 and this is for 𝝀 = 𝟏 and the remaining entries are all 0. So, once you have
again here you can cross check that 𝑷𝑫𝑷−𝟏 will give you 𝑨 and from here your
𝒆𝟐𝒕 𝟎 𝟎
𝟎 ]. So, that’s just the individual exponentials along the diagonal.
𝑫𝒕
𝒆 =[ 𝟎 𝒆𝟐𝒕
𝟎 𝟎 𝒆𝒕

So, once you have 𝒆𝑫𝒕 you compute 𝒆𝑨𝒕 = 𝑷𝒆𝑫𝒕 𝑷−𝟏 . And we had obtained our 𝒆𝑨𝒕 earlier
on using the other two methods before. So, you can again check that it matches with the
value that you get here. So, yeah so today we learnt about why we need matrix exponentials
and how do you compute them using an example.

Thank you.

1033
 Control Engineering
Dr. Ramkrishna Pasumarthy
Department of Electrical Engineering
Indian Institute of Technology, Madras

Module - 12
Lecture - 03
Controllability and pole placement

Hey guys. Welcome back to this lecture series on Control Engineering. And, so today we’ll
keep continuing with the state space analysis. So, far what we have seen is given a system
how to write it down in the state space form, given a non-linear system how would you
linearize that around an equilibrium point and analyze stability.

So, today we will do further or go a little further deep into this topic and essentially deal
with what is called as a controllability, and what does controllability mean, is a system
always controllable? what if it is not always controllable? and if it is controllable how do
we control it? So far we have seen control methods or design methods using the root locus
where we take the help of a root locus plot to achieve the desired closed-loop specification.

Similarly, we did some design problems with the help of bode plots. So, let’s see what all
that means, in the context of a state-space analysis, ok. So, this let us quickly write down
what have we seen so far. So, far we have had state-space representations. And what we
saw, yes in the other lectures was three different forms.

(Refer Slide Time: 01:31)

1034
So, one was the controllable canonical form. Similarly, we had the observable form and
also the diagonal form. And we also saw how to convert a given system into its diagonal
form. There are several other details of the diagonal form in the terms of the Jordan form,
but I will try to skip that, ok.

The first is now how to convert system from a controllable to observable or observable to
a diagonal or whatever. So, if I say I have a transfer function and I convert it into a state-
space, now is the state space representation unique? and how are these three related to each
other? So, let’s do it in a little abstract way.

Let’s say or whenever I am converting. So, let’s say I have a system which is like
𝒙̇ = 𝑨𝒙 + 𝑩𝒖 and I say how do I convert it from one form to the other. And we had earlier
seen that we could actually do a transformation of 𝑨 matrix to something which looks like
𝑷−𝟏 𝑨𝑷, ok. So, where does this come from? So, let’s say I start with 𝒙 I define a new
variable 𝒛 and say 𝒙 is related to 𝒛 via this matrix 𝑷, ok. So, this matrix 𝑷 is invertible. So,
𝑷 is, ok. So, what happens now?

So, if I write down this in terms of the 𝒛 coordinates, I will have 𝒁 = 𝑷−𝟏 𝑿 which means
that 𝒛̇ = 𝑷−𝟏 𝒙̇ . What is 𝒙̇ ? 𝒙̇ = 𝑨𝒙 + 𝑩𝒖. And what is 𝒙? 𝒙 = 𝑷−𝟏 𝒛. So, I have if I write
it down again, so 𝒙 in terms of 𝒛 would be (𝑷−𝟏 𝑨𝑷)𝒛 + 𝑷−𝟏 𝑩𝒖.

̃𝒛 + 𝑩
So, I have a system in the new coordinate 𝒛 which looks like let me call this say 𝑨 ̃ 𝒖,
̃ and 𝑷−𝟏 𝑩 is the 𝑩
where this guy 𝑷−𝟏 𝑨𝑷 is the 𝑨 ̃ , ok.

So, the invertibility is needed because well we need the matrix to be invertible here and
you can always go back from the 𝒙 to 𝒛 and 𝒛 to 𝒙 and vice versa all the time, ok. So, given
a system in a certain way say it’s in the observable form I could go to the controllable form
via some transformation 𝑷, ok. What is this transformation is what we will see a little later,
which kind of 𝑷 to use to get from a standard form to a controllable form, standard form
to an observable form, and so on, right, ok.

Now, what has changed here? Now, can I say that the system in the 𝒙 coordinates is the
same as the one in the 𝒛 coordinates? Well, let’s see. So, what we are interested typically
is in the pole locations, ok. So, we’ll compare the characteristic equation of this guy to the
characteristic equation in 𝒛, characteristic equation of the system written down in the 𝒙

1035
coordinates to the characteristic equation of the system written down in the 𝒛 coordinates,
ok.

So, let’s see. So, if I write down the characteristic equation for the system in the 𝒛
coordinates I’ll have |𝝀𝑰 − 𝑷−𝟏 𝑨𝑷|. So, this will be I can write this down as
|𝝀𝑷−𝟏 𝑷 − 𝑷−𝟏 𝑨𝑷|, I am just multiplying by 𝑷 and its inverse.

So, this will be |𝑷−𝟏 (𝝀𝑰 − 𝑨)𝑷|, where 𝑰 is identity matrix times 𝑷. I am just using some
basic properties of matrix, the determinant just come out like this, (𝝀𝑰 − 𝑨)𝑷. So, what am
I left with say this guy cancels with this guy and I am just left with |𝝀𝑰 − 𝑨|.

So, if I equate this to 0, I get the poles. So, which means that if I write this down in
summary |𝝀𝑰 − 𝑷−𝟏 𝑨𝑷| is |𝝀𝑰 − 𝑨| = 𝟎. So, the characteristic equations are the same which
means that the poles here and the poles here would exactly be the same.

So, almost like nothing changes, right. So, I am exactly dealing with the same system. I
am just doing a little bit of coordinate transformations. And we all would have done
coordinate transformation in some math course- vector calculus, where we transform from
a rectangular to a polar coordinate, cylindrical coordinates and nothing changes in the
system. Something very similar is happening here, right.

So, each and each of these forms the controllable, observable, and diagonal are useful in
their own way. A bit of the analysis which we will do as we progress through this lecture,
ok.

1036
(Refer Slide Time: 07:27)

So, the next thing is controllability. So, let’s again look at it with the help of an example,
ok.

(Refer Slide Time: 07:41)

So, let’s say I have say I am moving in a plane, just for simplicity, this is just for illustration
and say I have a simplified model of a car and say all I could do is I can just control, but I
can just move in this direction, right. I can just go forward I can go back, but steering is
not allowed. So, this movement is not allowed, ok.

1037
Now, with the application of input 𝒖 in this direction I can move forward, right in this
way, application of 𝒖 here I can move backwards. Now, can I say that this simplified
version of a car, the picture is not drawn very well, but can we see that this is actually
controllable? Right. But I am just allowed to move this way or this way.

So, if you look at it from a normal car objective, whenever we drive a car what would we
want is we want to drive across the entire domain that we have to go this way, you want
to go, this way, this way, and so on, right. So, what should also be allowed is and I start
from that position and I end up here, ok. But this control action u will not allow me to go
to this position essentially because the steering is not allowed, ok.

Now, how do we qualitatively analyze this or when is a system controllable? The system
here, with just the two 𝒖’s which either goes in this direction of 𝒙𝟐 -positive direction or in
the negative direction of 𝒙𝟐 , is this controllable or for complete controllability I also need
the steering, right. So, that I can go from this position to actually this position, ok.

So, let’s first understand this with the help of an example, right. So, I have a system well
in the this is a very beautifully written down in the diagonal form 𝒙𝟏̇ = −𝒙𝟏 + 𝒖,
𝒙𝟐̇ = −𝒙𝟐 + 𝒖, ok.

So, these are essentially differential equations if I just try to solve those I get something
like this that 𝒙(𝒕) starting from initial conditions we’ll have a solution like this, which
essentially means I will ask myself a question given initial states 𝒙𝟏 (𝟎) and 𝒙𝟐 (𝟎) what
are the states 𝒙𝟏 and 𝒙𝟐 , I can reach after time 𝒕, ok.

1038
(Refer Slide Time: 10:43)

So, if I again were to draw a graph of this, so let’s say I am moving again in 𝑹𝟐 . Let’s say
𝒙𝟏 and say 𝒙𝟐 , ok. And say I am starting, for simplicity say I am starting from the origin,
right. So, I have a system in 𝑹𝟐 again of the form 𝒙̇ = 𝑨𝒙 + 𝑩𝒖.

Now, I would ask myself a question with any given u which is arbitrary unconstrained, but
bounded what are the directions can I move? So, let’s say I can move this direction, this
direction, this direction, this direction, all 𝟑𝟔𝟎°, right, right. So, this would mean that
starting from origin I can reach any other point in the 𝒙𝟏 , 𝒙𝟐 plane.

So, if I ask a question what are the points one can reach starting from the origin? In the
application with appropriate control. If I say that I can reach all of this space in the plane
any point, then I can say that I can reach entire of 𝑹𝟐 or the reachable subspace is 𝑹𝟐 , the
entire 𝒙𝟏 , 𝒙𝟐 plane on both directions, ok.

Similarly, I can also say well starting from any other points or say what are the set of initial
conditions that can steer the system to the origin. So, if I say if I start from here, so this is
my 𝒙𝟏 , 𝒙𝟐 plane by the application of some control can I go back to the origin? Starting
from here can I go back? Right and then, if I find the set of all points that will give me the
controllable subspace, ok.

1039
What is holistic about this origin? Nothing really. I can even say what are the set of initial
conditions which I can reach this point or this point there is nothing really holistic about
this or even this origin here.

So, the first thing reachability will tell me what are the points I can reach starting from a
given point which in this case is the origin. Controllability would mean what are the points
or what are the initial conditions that will steer me to the origin by application of proper
control input. So, in the linear time-invariant case it turns out that these two are similar,
they are exactly similar. Reachability would mean controllability and vice versa. If I could
go from this point to this point, I can actually come back from this point back to the origin,
ok.

So, if I go back to the example here if I say well if I am not allowing the steering position
the only directions I can move are this one that I can move in 𝒙𝟐 direction on both ways,
but I cannot move in the 𝒙𝟏 direction. So, this movement is not allowed or it’s not possible
because I cannot steer. I cannot turn, it this way or even this way, ok. So, let’s see more
mathematically what these things mean.

So, if I say that a system like this starting from any arbitrary initial conditions what are the
points it can reach in time 𝒕. Well, if I also additionally impose the condition that well I
just want to reach the origin like I did here, right. In this example here, my control problem
was to find out the set of points sorry the set of points which will steer me to the origin.

So, let’s say here that 𝒙(𝒕) is actually the origin, ok. So, what will happen? So, I have a
𝒙𝟏 (𝟎)
zero here 𝒆−𝒕 [ ] and this entire guy. So, it turns out if I just rearrange this terms
𝒙𝟐 (𝟎)

properly that the set of initial conditions 𝒙(𝟎) is a set which looks like this −𝜶(𝒕) [𝟏].
𝟏
What is 𝜶? 𝜶 is this thing, ok.

So, in this thing what is unknown? I know the initial condition 𝒙𝟏 (𝟎), 𝒙𝟐 (𝟎), I know the
final condition which is the origin, I can compute these guys, only unknown is the input.
So, does there exist an input, right which can steer the states to the origin, or what are the
states if I just plot this down in on a piece of paper. So, it looks like this, right.

1040
(Refer Slide Time: 16:29)

The points which can take me to the origin are only in the green line. For example, if I
have start at a point say in the fourth quadrant or the second quadrant or any other point
which is outside this green line I will not be able to reach the origin, ok.

Now, let’s see some other example or in this example, the controls controllable subspace
is of dimension one. So, I guess cannot travel all points in the state space it is something
similar to what was happening here, right. I was not allowing steering of the wheels, so I
can only move in the positive or negative 𝒙𝟐 directions, ok.

1041
(Refer Slide Time: 17:12)

So, let’s see some other example. I just have again the system in the diagonal form with
𝟏
entries [−𝟏 𝟎
], control as [ ] 𝒖, ok. Now, if I just write down again the solution of this
𝟎 −𝟐 𝟐
say given certain initial conditions that I want to go to the origin, the equation would
transform to something like this.

(Refer Slide Time: 17:35)

And if I plot what is the controllable subspace or what are the points starting which I can
reach the origin it turns out to be the entire subspace. So, I can start any point in the first
quadrant, second, third, or fourth, all will steer me to the origin or steer the system to the

1042
origin, right. So, any point in 𝑹𝟐 can be steered to the origin, ok. So, this leads us nicely
to define the concept of controllability.

Now, given any system would I keep computing these solutions all the time right; if I have
a system which is of dimension 5, it may really be difficult for me to compute all the
solutions, right. So, what we will see is given the properties of system which is
𝒙̇ = 𝑨𝒙 + 𝑩𝒖. The properties of the system are the 𝑨 matrix and the 𝑩 matrix. So, given
this information on the 𝑨 matrix and the 𝑩 matrix can I say if the system is controllable or
not?.

So, what is the definition of controllability first? So, based on what we had argued so far.

(Refer Slide Time: 18:37)

A system is said to be controllable at t naught, right, if it is possible to transfer the system

from any initial state 𝒙(𝒕𝟎 ) to any initial state in finite amount of time. So, the key here is
I should be able to reach from point 𝑨 to point 𝑩 in some finite amount of time. Doesn’t
really help me much if I can say that I could travel from Bombay to Delhi in like an infinite
amount of time, right. So, you want to travel or your car should be designed in such a way
that you travel in finite amount of time, ok.

So, before I solve any control problem it is important for me to check if the system is
controllable or not, ok. So, let’s see how we do that. So, I am given a system in the state
space form 𝒙̇ = 𝑨𝒙 + 𝑩𝒖, 𝒚 = 𝑪𝒙 + 𝑫𝒖, ok. So, it will turn out that we will just shortly

1043
derive this result quickly that the system is controllable only if this matrix
[𝑩|𝑨𝑩| … . |𝑨𝒏−𝟏 𝑩|] it will be of dimension 𝒏 𝑋 𝒏𝒎, if this rank is 𝒏 then the system is
controllable, right and this is both a necessary and sufficient condition, ok.

(Refer Slide Time: 19:56)

So, how do we derive this? We will quickly do this we won’t really try to memorize a
derivation we will just, but we will try to make use of this relation, the rank condition for
controllability, ok.

So, I start with this equation 𝒙̇ = 𝑨𝒙 + 𝑩𝒖, the outputs really do not matter towards at
the moment. So, we can just forget about this for a moment. So, 𝒙̇ = 𝑨𝒙 + 𝑩𝒖, the solution
would look something like this. So, by now you also know how to compute 𝒆𝑨𝒕 , ok.

So, given any initial condition assume that I want to reach the origin in some time 𝒕𝟏, I
will ask when does the solution exist. What does it mean by the solution? Again, if I look
at this expression I know the initial condition 𝒙(𝟎), I know the final condition 𝒕(𝟎), I just
want to find out, I also know the matrix 𝑨, I also know the matrix 𝑩, all I want to find out
is can I reach the origin starting from any arbitrary initial condition by application of some
control. So, this is the unknown.

So, we will just use some properties of how to compute the 𝒆𝑨𝒕 , you can refer to the earlier
lectures which we had on this. So, 𝒆𝑨𝒕 can be written as some series in powers of 𝑨 in this

1044
way, ok. So, what I do is I just plug this in over here. Again, in this expression I am just
finding out if there exists a 𝒖, the only unknown in this equation is the 𝒖.

So, if I substitute for 𝒆−𝑨𝝉 over here what I end up is equation which looks like this 𝒙(𝟎)
is summation of this 𝒙 of this summation term in the powers of 𝑨 times some numbers 𝜷𝒌 .
𝒕
What is this 𝜷𝒌 ? So, whatever is in the ∫𝟎 𝟏 𝜶𝒌 (𝝉)𝒖(𝝉)𝒅𝝉, I just called them 𝜷𝒌 . I can
always compute this, right this is always some number, ok.

So, it will turn out that if I, you know I can also write this as 𝒙(𝟎) = 𝑩𝜷𝟎 . If I just take
for 𝒌 = 𝟎 what I will have is 𝑨𝟎 then which is identity. So, it will be 𝑩𝜷𝟎 here. For 𝒌 = 𝟏
I will have 𝑨𝟏 𝑩𝜷𝟏 , this is 𝑨𝑩𝜷𝟏 , and so on, until I reach 𝒏 − 𝟏, ok.

Now, where does the unknown sit in here? The unknown sits in here in this 𝜷s, right. The
𝒖’s appear in 𝜷. So, when does the solution exists? A solution will exist if and only if this
matrix is invertible, right. [𝑩|𝑨𝑩| … . |𝑨𝒏−𝟏 𝑩|] is invertible, ok. So, this is exactly the rank
condition which we are trying to derive here that the system is controllable if and only if
this matrix is invertible, sorry, not invertible, if this matrix is of rank 𝒏, ok. Similarly, a
sorry not about the invertibility, but this matrix should be of rank 𝒏, ok.

So, this is the condition which we will use. We may not necessarily you know remember
this proof all the time, but what is important is to remember the condition that the rank of
this matrix should be 𝒏, ok.

So, once I know that the system is controllable, what do I do with it? How do I then control
the system? So, when I was doing the design with the root locus or with the bode plot I
always knew that well I can use either a lead compensator a lag compensator, or even more
simply the gain adjustment. If the gain adjustment doesn’t work, then I go to one of these
lead or lag compensators depending upon the specifications of the problem. So, how does
that translate in this case?

1045
(Refer Slide Time: 24:00)

So, let’s come to a problem. Now, let’s say that for a given system which is of this form
𝟎 𝟏
𝒙̇ = [ ], can we design a controller 𝒖 which I call as a state feedback controller
−𝟔 −𝟓
which has the following specifications that the damping coefficient is 0.707 and the
response the peak response time is under 3.14 seconds, ok.

So, there are a couple of methods which we could use in this. So, let’s just write try to
write down what the problem statement is for us.

(Refer Slide Time: 24:51)

1046
So, first is the peak time is given to us the 𝒕𝒑 , peak time should be less than 3.14 or 𝝅, ok.
And it’s also said that the 𝛇 the damping coefficient should be 0.707, ok.

So, now based on this I need to compute what is 𝝎𝒅 and what is 𝝎𝒏 . So, if you remember
𝝅
the formulas for the peak time, so I think 𝒕𝒑 = and this turns out that 𝝎𝒅 = 𝟏. And 𝝎𝒅
𝝎𝒅

was related to 𝝎𝒏 ; why are the damping coefficient with this formula, 𝟏 − 𝜻𝟐 which will
give us that 𝝎𝒏 = 𝟏. 𝟒𝟏𝟒, ok.

Now, what does this mean? Right. So, first is well, it says design a state feedback
controller. So, a state feedback controller typically is of the form 𝒖 = −𝑲𝒙, or in this case
well, so the system is of dimension 2, I will have a 𝑲𝟏 𝑲𝟐 ; 𝒙𝟏 and 𝒙𝟐 , ok. Now, how does;
what is the original system?

𝟎 𝟏 𝟎
Original system is of the form 𝒙̇ = [ ] 𝒙 + [ ] 𝒖; ok. So, I could just quickly check
−𝟔 −𝟓 𝟏
what would be the 𝛇 and 𝝎𝒏 for the open-loop system, right. When 𝒖 = 𝟎 I just would say
check it with the help of this characteristic equation |𝒔𝑰 − 𝑨| = 𝟎, right. And it is easy to
check that the open-loop specifications are much further away from what is a desired
closed-loop specification and therefore, we need this control, ok.

So, a first exercise which you could do is to check if the system is controllable. It means
just quickly find the rank of 𝑩 and 𝑨𝑩, and this should definitely be equal to 𝟐 and you
can easily find out that the system is controllable. Also, it’s easier to check because it is
exactly in the controllable canonical form, ok.

So, let’s try solving this problem, right. So, what happens? So, if I have a 𝒖 of this form
𝟎 𝟏 𝒙𝟏 𝟎
my closed-loop system looks like this [ ] [ ] + [ ] 𝒖 , 𝒖 = −𝑲𝒙. So, this thing
−𝟔 −𝟓 𝒙𝟐 𝟏
𝒙𝟏
here will be a −[𝑲𝟏 𝑲𝟐 ] [𝒙 ] which will look something like this,
𝟐
𝟎 𝟏 𝒙𝟏
[ ] [ ], ok.
−𝟔 − 𝑲𝟏 −𝟓 − 𝑲𝟐 𝒙𝟐

So, now the closed-loop characteristic equation should be such that these poles or the
eigenvalues should satisfy this condition, ok. Now, when does a second-order system
satisfy this condition? Right. So, how should which mean, ok; this is what is desired in the
closed-loop case, ok. So, how does 𝒔𝟐 + 𝟐𝜻𝝎𝒏 + 𝝎𝒏 𝟐 terms look like?

1047
So, based on these two specifications of 𝝎𝒏 = 𝟏. 𝟒, 𝜻 = 𝟎. 𝟕, this should look something
like this 𝒔𝟐 + 𝟐𝒔 + 𝟐 = 𝟎, ok. Exactly I am just putting in the values of 𝜻 and 𝝎𝒏 as
desired, ok. Now, this is the desired closed-loop characteristic equation which will ensure
that the closed-loop has a peak time of 3.14 or less and 𝜻 the damping coefficient of 0.707.

Now, what would I do? If this is desired now I compute this guy, |𝒔𝑰 − 𝑨 + 𝑩𝒌| as I would
call it, where does this come from if I just say in a standard system 𝒙̇ = 𝑨𝒙 + 𝑩𝒖 with
𝒖 = −𝑲𝒙. I can rewrite this as 𝒙̇ = 𝑨𝒙 − 𝑩𝑲𝒙. So, the characteristic equation of the
closed-loop system would look like this. So, I would compare this |𝒔𝑰 − 𝑨 + 𝑩𝒌| with this
guy, ok.

So, now I have an equation with two unknowns, ok. So, let’s quickly write down what is
the characteristic equation. 𝒔𝟐 + (𝟓 + 𝑲𝟐 )𝒔 + (𝟔 + 𝑲𝟏 ) = 𝟎. So, I have to solve for 𝑲𝟏 and
𝑲𝟐 . How should this look like? Well, this should exactly look like this, ok.

So, this equation, this characteristic equation should be equal to this characteristic
equation, right. So, therefore, I can just compare the coefficients, right which is
2 = 𝟓 + 𝑲𝟐 and 2 = 𝟔 + 𝑲𝟏 which means 𝑲𝟐 = −𝟑 and 𝑲𝟏 = −𝟒, ok. Let’s assume. So,
how do we do this, take or just check what are the closed-loop specifications, ok?

This will translate to an appropriate characteristic equation of the closed-loop; this is a

desired one. On the right-hand side, I need to find out what are the gains 𝑲𝟏 and 𝑲𝟐 that
will achieve this desired characteristic equation here, ok. And then I just write this down,
I compare the characteristic equation on the right to the desired characteristic equation on
the left, I just equate the coefficients, ok.

So, in some problems it’s also the problem specifications could also be not in terms of 𝜻
and 𝝎𝒏 , but directly in terms of the desired closed-loop poles, ok.

1048
(Refer Slide Time: 32:41)

So, if you say that, given a system, 𝒙̇ = 𝑨𝒙 + 𝑩𝒖, this could be 𝒙 is say in some 𝑹𝒏 , ok.
Now, the problem could be of find a control law again 𝒖 = −𝑲𝒙 which places the closed-
loop poles at say some values 𝝁𝟏 till 𝝁𝒏 .

So, earlier we had the specification in terms of some system performance as a peak time,
it also be in a term of times, in terms of the settling time or the rise time, damping
coefficient and several other things. So, here I have just have specifications directly in
terms of where the poles of the closed-loop system be, ok.

So, in this case, what I will do? Well, what is a desired characteristic equation? The desired
characteristic equation would be I just look at this 𝒔 − 𝝁𝟏 , 𝒔 − 𝝁𝟐 …. 𝒔 − 𝝁𝒏 = 𝟎, ok. This
is I know all these numbers, right, I know all the 𝝁 ones, so I have a polynomial in 𝒔𝒏 with
all the coefficients which are known to me. On the right-hand side, I would have unknowns
in the terms of 𝑲𝟏 till 𝑲𝒏 .

So, I just equate the characteristic equation on this one which is in powers of 𝒔𝒏 plus, ok.
We will see how to write this. And on the right-hand side, I should just equate this to the
characteristic equation which I get by |𝒔𝑰 − 𝑨 + 𝑩𝒌|, ok. Let’s write this down little more
formally now, ok.

1049
(Refer Slide Time: 34:50)

So, this is really what. So, I have the desired poles which are 𝝁𝟏 till 𝝁𝒏 and I will have the
characteristic equation as a polynomial in 𝒔𝒏 . On the right-hand side, I know 𝒔, I know 𝑨,
I know 𝑩, I just need to find out 𝑲. So, I will have again a polynomial here of power 𝒏 I
just equate these two, as I did earlier, ok. So, well this is just straightforward, right, this is
exactly what we did here.

(Refer Slide Time: 35:21)

So, I will just quickly look at this. So, pole placement, so this exact this procedure which
we derived or which we or where we design the controller, a state feedback controller

1050
𝒖 = −𝑲𝒙 to achieve certain performance specifications is called the pole placement
technique, ok.

So, this is required to have the system desired behavior a necessary and sufficient condition
is that the system should be completely controllable, ok. So, what do I do? Well, I know
this 𝝁𝟏 till 𝝁𝒏 , and I just equate these two I compare the coefficients and I and I arrive at
the desired control law.

There is also other method from the controllable canonical form, ok. Let’s first write down
and then and then see what it means, ok.

(Refer Slide Time: 36:15)

So, let’s say I have the system in the controller canonical form, ok. I am just being a little
lazy to write. So, let’s for the purpose of derivation let just say that 𝒙 is in 𝑹𝟑 , ok. So, if in
𝟎 𝟏 𝟎
𝑹𝟑 my 𝑨 matrix would look like this, it will be [ 𝟎 𝟎 𝟏 ]. The 𝑩 matrix will have
−𝒂𝟑 −𝒂𝟐 −𝒂𝟏
𝟎
entries [𝟎], ok, ok.
𝟏

So, let’s first say that the desired closed-loop poles are at 𝝁𝟏 , 𝝁𝟐 , and 𝝁𝟑 , ok. So, my
characteristic equation the desired one would be like this. So, let just write this down as
say 𝒔𝟑 , and let’s denote the coefficient in terms of some 𝜶, say this is like

1051
𝒔𝟑 + 𝜶𝟏 𝒔𝟐 + 𝜶𝟐 𝒔 + 𝜶𝟑 = 𝟎. So, this is the desired in such a way that 𝝁𝟏 , 𝝁𝟐 , 𝝁𝟑 , and in
turn 𝜶𝟏 , 𝜶𝟐 , 𝜶𝟑 are known.

So, this is this system is completely controllable, ok. So, just because it is just in the
controllable canonical form. Now, I say I want to appropriately place the poles of the
closed-loop system at 𝝁𝟏 , 𝝁𝟐 , and 𝝁𝟑 . So, what is the control law
𝒙𝟏
𝑩 = − [𝒙𝟐 ] [𝑲𝟏 𝑲𝟐 𝑲𝟑 ] that will achieve this configuration in the closed-loop. So, the
𝒙𝟑
unknowns here are 𝑲𝟏 , 𝑲𝟐 , and 𝑲𝟑 , ok.

So, what is a characteristic equation? Right. The characteristic equation is |𝒔𝑰 − 𝑨 + 𝑩𝒌|.
So, this would simply be 𝒔𝑰 minus, ok. What is 𝑨?
𝟎 𝟏 𝟎 𝟎
[ 𝟎 𝟎 𝟏 ] + [𝟎] [𝑲𝟏 𝑲𝟐 𝑲𝟑 ]; and this minus will show up here, ok.
−𝒂𝟑 −𝒂𝟐 −𝒂𝟏 𝟏

So, if I just write this down I will get the following 𝒔𝑰, ok, I will just write it completely
in the matrix form. So, this matrix would be like this. We will have a 𝒔 here. So, this is the
entire, 𝑨, ok. So, this should actually be plus because I am looking at |𝒔𝑰 − 𝑨 + 𝑩𝒌|. So,
this minus and minus will cancel out, ok.

𝒔 −𝟏 𝟎
So, I have [ 𝟎 𝒔 −𝟏 ] = 𝟎. Or in other words, this means that the
𝒂𝟑 + 𝑲𝟏 𝒂𝟐 + 𝑲𝟐 𝒔 + 𝒂𝟏 + 𝑲𝟑
characteristic equation now is 𝒔𝟑 + (𝒂𝟏 + 𝑲𝟑 )𝒔𝟐 + (𝒂𝟐 + 𝑲𝟐 )𝒔 + (𝒂𝟑 + 𝑲𝟏 ) = 𝟎.

Now, I want to compare this with this. Which essentially means that 𝒂𝟏 + 𝑲𝟑 = 𝜶𝟏 or in
other words, the unknown 𝑲𝟑 can be determined as 𝜶𝟏 − 𝒂𝟏 . This the 𝒂𝟏 is known to us.
𝜶𝟏 is also known to us. 𝒂𝟏 is from the given system matrix. 𝜶𝟏 comes as a result of the
desired closed-loop poles. Similarly, then I can write 𝑲𝟐 = 𝜶𝟐 − 𝒂𝟐 , 𝑲𝟏 = 𝜶𝟑 − 𝒂𝟑 , ok.

1052
(Refer Slide Time: 42:08)

So, we will just write this down little formally now, ok. So, I have the characteristic
polynomial of this form and if the system is not already in the controllable canonical form
I will use the transformation, right the same thing 𝑷−𝟏 𝑨𝑷 or in this case I call it 𝑻, it will
be 𝑻−𝟏 𝑨𝑻, where I compute the transformation 𝑻 in the following way. 𝑻 = 𝑴𝑾, where
𝑴 and 𝑾 are completely. 𝑴 is just the controllable the controllability matrix, 𝑾 comes
from the entries of the 𝑨 matrix in the following way.

I will not tell you how to do this, but these are very standard texts. So, once you have them
in the controllable canonical form then 𝑲’s are simply computed this way, right 𝜶’s and
you can just compute. So, 𝜶’s are known to us, 𝒂’s are unknown to us, so I can directly
compute what is the 𝑲 matrix. So, what we have seen in this edition is the definition of
controllability, to check if the system is completely controllable or not and if the system is
completely controllable how do I exactly do the control law.

So, the only control technique which we have learnt here is what is called as the pole
placement. There are other formulas is typically called the Ackermann’s formula. I will
skip to those details, but what is important here is to know the concept. And once you
know the concept things the even the Ackermann’s formula is kind of quite easy to derive.

And you could also do the design directly by MATLAB. But before you do ask MATLAB
to compute the gain matrix, the 𝑲 matrix for you, it’s very advantageous for you to know
what is the; what is the procedure to compute the gain 𝑲, right, ok.

1053
So, in the next lecture we will see about what to do if the system is not controllable, is
there any hope, and also to learn the concept of observability because so far we have been
using things which says that the controller is of the form 𝒖 = −𝑲𝒙. Who tells us what the
𝒙 is? Is it really trivial? If 𝒙 is not known to us completely say there are 𝒏 states I can only
measure two states out of those 𝒏, is there any hope? So, these are the two things which
we will quickly run through in the next lecture.

Thank you.

1054
 Control Engineering
Dr. Ramkrishna Pasumarthy
Department of Electrical Engineering
Indian Institute of Technology, Madras

Module - 12
Lecture - 04
Controllable Decomposition and Observability

So in this little lecture, we will first talk about Controllable Decomposition. So, why do
we need to do this? So, in the previous lecture, we saw basics about controllability that the
system is controllable, if the rank of a certain matrix is 𝒏, and if that rank is 𝒏, then I could
actually do something called the pole placement to solve control problems in the state
space ok.

So, now we did not answer the question what if the system is not completely controllable,
what if the rank is not 𝒏, but it is some number 𝒎 which is less than 𝒏 ok. So, let’s quickly
check what we can do with those kind of systems ok.

(Refer Slide Time: 01:15)

So, let’s say that I start with the system 𝒙̇ = 𝑨𝒙 + 𝑩𝒖. And suppose that the rank of this
matrix [𝑨: 𝑨𝑩: … … 𝑨𝒏−𝟏 𝑩] is some number is some number 𝒎 which is less than 𝒏 ok.
Let’s to keep it simple; let’s do it with the help of a little example. Where
𝟎 −𝟏 𝟏 𝟏 𝟎
𝑨 = [𝟏 −𝟐 𝟏 ]; 𝑩 = [𝟏 𝟏] ok. So, the controllability matrix 𝑪, which I call is 𝑩, just
𝟎 𝟏 −𝟏 𝟏 𝟐

1055
 𝟏 𝟎 𝟎 𝟏 𝟎 −𝟏 0 1 0 −1
be 𝑪 = [𝟏 𝟏 𝟎 𝟎 𝟎 𝟎 ]; 𝑨𝑩 would be [0 0 ], and 𝑨𝟐 𝑩 would be [0 0 ] ok.
𝟏 𝟐 𝟎 −𝟏 𝟎 𝟏 0 −1 0 1
And we can easily check here that the rank is 𝟐 ok.

Now, what do we do when the rank is 𝟐? I cannot completely control the system because
what I learnt ok, even though we didn’t do the proof is that for to place all the poles, we
need the system to be completely controllable ok. So, let’s do a bit of a magic here. Let’s
say I want to transform the system into something else right. And we knew that, so that
𝑷−𝟏 𝑨𝑷 is a good transform which transform the system from some 𝒙 coordinates to certain
𝒛 coordinates ok.

So, let’s construct this 𝑷 in a nice way now. Let say 𝑷 is I take two independent columns
from this controllability matrix. So, the obvious ones are directly from the 𝑩 matrix. Now,
can I add another column here such that the rank of this 𝑷 is 𝟑 right. So, I am giving you
two independent vectors in 𝑹𝟑 , and I am asking you to construct something else which is
again independent of the first two right. A simple choice would be something like this ok.
Now, this is an invertible matrix this 𝑷. So, I can always do this transformation.

̃ be? 𝑨
So, what will the new 𝑨 ̃ = 𝑷−𝟏 𝑨𝑷 would turn out to be something like this,
𝟎 𝟏 𝟏 𝟏 𝟎
[𝟎 ̃? 𝑩
−𝟏 𝟎 ]. And what is 𝑩 ̃ = [𝟎 𝟏] ok. So, let’s carefully look at this ok. So, let’s
𝟎 𝟎 −𝟐 𝟎 𝟎
𝒙𝟏̇ = 𝒖
say if I write the system like this right, [ ], it can be easily checked that this
𝒙𝟐̇ = −𝒙𝟐
system is not completely controllable. This system which means which because you know
you can see there is no direct control which is entering the 𝒙𝟐 coordinate ok.

So, here I can clearly see which of the two states is not controllable, you can easily say
that 𝒙𝟏 is controllable, and 𝒙𝟐 is not ok. Now, what happens when I do a transformation
like this? If I look at this thing here, if I say I have this new coordinates in the 𝒁𝟏̇ , 𝒁𝟐̇ , 𝒁𝟑̇
is this new 𝑨 𝑋 𝒁 𝑋 𝒖. That 𝒁𝟑 has no influence of the control, because the entries of the
𝑩 are 𝟎. Whereas, 𝒁𝟏 and 𝒁𝟐 are actually controllable, because there is some, say some
numbers coming in here.

So, what I have done with this transformation is I have split the system into a controllable
part and an uncontrollable part. So, this is my controllable part, and this thing is my
uncontrollable part ok. So, what did we do here? We, well, what was given to us is, were

1056
the system which was not completely controllable, in such a way that the rank of this
controllability matrix was some number 𝒎.

So, based on this 𝒎, we pulled out 𝒎 -independent vectors from the controllability matrix.
I add remaining 𝒏 − 𝒎 vectors such that the rank becomes 𝒏, and I do this kind of
similarity transformation. So, while doing this, I have a system where I can clearly see
which part is controllable and which part is uncontrollable. And this is what is referred to
as the controllable decomposition ok.

(Refer Slide Time: 07:02)

So, there exist a similarity transformation we exactly saw ok, I called it 𝑷 there, but it’s, 𝑻
here, doesn’t really matter. So, the 𝑻−𝟏 𝑨𝑻 is something like this. So, you have the
controllable part; you have the uncontrollable part right. And even the 𝑩 splits very nicely
into the controllable and uncontrollable part, 𝑪 will have some numbers; this is not really
important ok.

So, what is interesting here right? So, I have a system now which is explicitly written in
terms of 𝒙𝒄 which is the controllable states, 𝒙𝒖 which are the uncontrollable states ok.
What I want is the transfer function of this, well, I can simply write down the formula for
the state space to the transfer function form [𝑪][𝒔𝑰 − 𝑨]−𝟏 [𝑩] + [𝑫] ok. So, I will skip
these computations.

1057
(Refer Slide Time: 08:04)

But what is interesting to see is that the transfer function will have numbers or the entries
which are only corresponding to the controllable part 𝑨𝒄 is the one which is corresponding
to the controllable part; 𝑩𝒄 is also the controllable part; 𝑪𝒄 is also the controllable part.

So, what is important here is that the transfer function is only the transfer function of the
controllable part. Therefore, whenever I give you any transfer function, it should, I could
do anything with that transfer function if and only if I know that this is a transfer function
of a completely controllable system. Otherwise, you see that some poles of the system go
missing, or there is some kind of inherent pole-zero cancellation whenever the system is
not completely controllable ok. So, what can we do with this kind of systems right?

1058
(Refer Slide Time: 08:58)

𝒙𝟏̇ 𝟎 𝟎 𝒙𝟏 𝟏
So, say I will just take the other example which where [ ]=[ ] [𝒙 ] + [ ] 𝒖 ok,
𝒙𝟐 ̇ 𝟎 −𝟏 𝟐 𝟎
which can be written again as 𝒙𝟏̇ = 𝒖, 𝒙𝟐̇ = −𝒙𝟐 ok. What can I do with this system?
Well, you can see that I can actually control the first state provided that the second state is
stable.

What does it mean when I say this? Take instead a system which is like this
𝟎 𝟎 𝒙𝟏 𝟏
[ ] [𝒙 ] + [ ] 𝒖, which means this is the same system 𝒙𝟏̇ = 𝒖, 𝒙𝟐̇ = +𝒙𝟐 .
𝟎 𝟏 𝟐 𝟎

And let’s see how the uncontrollable part of the system behaves here and here. The
uncontrollable part here is actually stable, because I can just directly solve for 𝒙𝟐̇ = 𝒙𝟐 .
This is unstable, so this is bad news right. So, I can do anything or I can there is some hope
when the system is not completely controllable if and only if the uncontrollable part is
stable by itself.

Here I could do nothing with the system, because whatever smart kind of input I choose
whatever expensive, whatever you know the best kind of 𝒖, my 𝒙𝟐 is still unstable, which
means that the overall system is unstable. So, there is nothing I could do with these kind
of systems right.

And therefore, we have now a weaker notion of controllability that is stabilizability. For
the system 𝑨 𝑩, if the uncontrollable modes are stable, then the system is said to be

1059
stabilizable. I can do something with the system only if the uncontrollable modes are stable
ok. And then as a very small extension to the definition of controllability. So, far we have
talked about the state going from point 𝑨 to point 𝑩 in some finite amount of time, and we
just relaxing at the point 𝑩 is actually the origin ok.

So, what do I say, well, I want the output to be controllable not the states, because typically
I have 𝒙𝟏̇ = 𝑨𝒙 + 𝑩𝒖, why is some 𝑪𝒙 sometimes some 𝑫𝒖. So, the definition translates
can I transfer the output with some initial value 𝒚(𝒕𝟎 ) to some value 𝒚(𝒕𝟏 )in finite time
and with application of some control.

This again I will just leave for you to derive that the corresponding rank condition would
just be in terms of the output matrix also [|𝑪𝑩|| 𝑪𝑨𝑩| … . |𝑪𝑨𝒏−𝟏 𝑩| 𝑫 ]should be 𝒑. And
this is 𝒑 is like that the number of outputs here ok.

So, what have we learned here is, what do we do if the system is not completely
controllable. If the system is completely controllable, I can just use 𝒖 = −𝑲𝒙 to achieve
the desired performance in terms of placing the poles at the appropriate locations.

Even in root locus, what we were doing, we are just shifting the poles right, via the root
locus shifting the dominant poles to the left depending on if you wanted a faster response
right, and also poles corresponding to when we were doing the lag compensation right. So,
essentially playing around with the poles, this also is something similar ok.

(Refer Slide Time: 12:59)

1060
So, what do we assume when we say 𝒖 = −𝑲𝒙, we always assume that 𝒙 is measurable
that somebody is actually giving me 𝒙. Sometimes now what I am; what am I; what am I
measuring here? I’m measuring the outputs 𝒚 = 𝑪𝒙, let’s just ignore 𝑫 for the moment
right.

So, 𝒙, so given the measurements 𝒚, when can I get all of 𝒙? This is possible if and only
if the output matrix 𝑪 is invertible ok, which means that the number of states and the
number of outputs are the same, or this matrix 𝑪 should be invertible. But this is not
possible all the time, because sometimes we will have maybe only 5 outputs measured
when the complete total number of states is 10 ok.

So, in that case, what do we do, is there again somehow? So, can we use the input we know
the 𝒖 all the time and the output measurements over a time interval to reconstruct the states,
now reconstruct 𝒙(𝟎) for example.

(Refer Slide Time: 14:09)

So, this is what leads to the definition of observability. I’ll not go into the details of this,
but we will just understand what it means. So, I will just stop by just giving a definition
and how to check the conditions of on observability ok.

So, what is a definition? The definition says that the system is observable at time 𝒕𝟎 if this
𝒙(𝒕𝟎 ) can be determined by the observation of the outputs over the time interval 𝒕𝟎 to 𝒕𝟏
ok. Again these are important because all states may not be measurable. So, a simple

1061
computation like earlier would suggest that a system is completely observable if and only
if the rank of this observability matrix is 𝒏 ok. So, it is of dimension whatever 𝒏𝒑 𝑋 𝒏
matrix which is this one, this is called the observability matrix.

So, again given the properties of the system which essentially come from 𝑪 and 𝑨, we can
determine if the system is completely observable or not. And also note that the matrix 𝑩
does not really play any important role in this ok. So, this is where we will end. We will
try to post some problems related to you know the various concepts which we have derived
in state space analysis. And also possibly related to the observer or the observability
matrix.

And this is actually this covers a you know good amount of stuff related to the state space
analysis. And now I will not go deeper into how to design observers and so on, or how to
design a observers and controllers simultaneously for a system that we will leave for some
advanced course in control. But for the moment this is all what I thought would be useful
to you to give you some insights on the theory on state space.

So, this is the last lecture of the entire course. And I hope you have you enjoyed the course,
and you had a good time, and we actually had a good time preparing the course content,
interacting with you while we were doing through the discussion forums. And well, good
luck for your final exam if you are taking those.

And thanks very much.

1062
 THIS BOOK IS
NOT FOR SALE
NOR COMMERCIAL USE

(044) 2257 5905/08 nptel.ac.in swayam.gov.in