3/23/2023

Advanced Control
Theory

Prof. Smita Sadhu

Electrical Engineering Department

Jadavpur University

Part I

Optimal and Robust

Control

1
 3/23/2023

Contents/Syllabus (Part –I)

• Concepts of Controllability
• Control Law design for full state feedback.
• Pole placement by state feedback.
• Concepts of Observability, Observer Systems and Design of State Observers.
– Revision exercises
• Optimal Control Systems and Performance Indices.
• Optimal Control of linear systems with Quadratic Performance Index.
• Optimal State Regulator Design through Matrix Ricatti equation.
• Robust Control Systems and System sensitivity.
• Stability of systems with uncertain parameters.
• Kharitonov's methodology.
• Structured and Unstructured uncertainty.
• Stability robustness of Control Systems.
• Integral Control
• Robust Tracking
• H2 and H-infinity control.

Part II

Nonlinear Control

2
 3/23/2023

Contents/Syllabus (Part –II)

• Non-linear systems.
• Describing functions of common Non-linearity.
• Stability Analysis by Describing Function method.
• Phase plane method. Construction of phase Trajectories.
• System Analysis on phase plane.
• Optimal switching in Bang-Bang Control Systems.
• Lyapunov method of Stability Analysis.
• Popov's Circle Criterion.
• Case study in nonlinear Control Performance Analysis of Systems
with Dead-time elements.
• Modelling and Specifications.
• Actuators: Electric, Hydraulic and Pneumatic inertial sensors.
• Gyroscopes and Accelerometers- Modelling and Specifications.
• Case study of an aerospace control system with actuator & inertial
sensor.
5

Part I

Optimal and Robust

Control

3
 3/23/2023

Books and References

(e-books available: softcopy 2b supplied by SS)

1. Feedback Control Theory, Doyle Francis and

Tannenbaum
2. Advanced Control Engineering, Roland S. Burns
3. Linear Optimal Control, Anderson and Moore
4. Robust Control: the parametric approach,
S.P.Bhattacharyya
5. Robust Control Design: An Optimal Control
Approach, Feng Lin
6. Feedback Systems: An Introduction for Scientists
and Engineers, K.J.Astrom and Richard M. Murray

Concept of Controllability
• If an input to a system can be found that
takes every state variable from a desired
initial state to a desired final state, the
system is said to be controllable;
otherwise, the system is uncontrollable

4
 3/23/2023

Test for Controllability

• For an n-th order system described by

• The sufficient condition for complete state controllability

is that the matrix

• Contains n linearly independent row or column vectors

– i.e., is of rank n,
– i.e., is non-singular
– i.e., has a non-zero determinant

Concept of Observability
• The concept of observability relates to
the condition of observing or estimating
the state variables from the output
variables, which are generally measurable

10

5
 3/23/2023

Test for Observability

• For an n-th order system described by

• The sufficient condition for complete state observability

is that the matrix

• Contains n linearly independent row or column vectors

– i.e., is of rank n,
– i.e., is non-singular
– i.e., has a non-zero determinant

11

A Simple Full Order State

Observer

12

6
 3/23/2023

The Luenberger Full-Order State

Observer
x  Ax  Bu
y  Cx
xˆ  Axˆ  Bu  K e y  yˆ 
yˆ  Cxˆ
xˆ  Axˆ  Bu  K Cx  xˆ 
e

xˆ  Axˆ  x  Ax  K eCx  xˆ 

xˆ  x  Axˆ  x   K Cxˆ  x 
e

Let e  x  xˆ
e  Ae  K eCe
e  A  K eCe
13

Observer Equation
x  Ax  Bu
e  A  K eCe
y  Cx I  A  K e C   0
xˆ  Axˆ  Bu  K e y  yˆ 
yˆ  Cxˆ
n - dimensional system
xˆ  Axˆ  Bu  K Cx  xˆ 
e p inputs and r outputs
x  Ax  xˆ  Axˆ  K eCx  xˆ  A nn B n p
x  xˆ  Ax  xˆ   K Cx  xˆ 
e C r n x n1
Let e  x  xˆ y r1 u p1
e  Ae  K eCe
K e nr e n1
e  A  K eCe
14

7
 3/23/2023

Example-1
• Design an observer for the double
integrator system

u x2 x1 y
1/s 1/s

15

Soln to Example-1
• Double integrator system
u x2 x1 y
1/s 1/s

• System equations:  x1  0 1  x1  0

 x   0 0  x   1u
x 2  u  2   2   
x1  x2 x 
y  1 0 1 
y  x1  x2 

16

8
 3/23/2023

Soln to Example-1
• System equations: u x2 x1 y
1/s 1/s
x 2  u
x1  x 2
y  x1

• Matrix notation:  x  0 1   x  0

1 1
 x   0 0   x   1 u
 2   2   
x 
y  1 0 1 
 x2 
17

Soln to Example-1
• Observer equation:  x1  0 1   x1  0
 x   0 0  x   1  u
 2   2   
x  xˆ  Ax  xˆ   K eCx  xˆ 
 x1 
y  1 0 
 x  xˆ  A  K eCx  xˆ   x2 

• What is the dimension of Ke?

18

9
 3/23/2023

Soln to Example-1
• Observer equation:  x1  0 1   x1  0
 x   0 0  x   1  u
 2   2   
x  xˆ  Ax  xˆ   K eCx  xˆ 
 x1 
y  1 0 
 x  xˆ  A  K eCx  xˆ   x2 

• Dimension of Ke is 2X1

K 
• Let K e   e1 
K e2 

19

Soln to Example-1
 x1  0 1   x1  0

Find A  K eC
 x   0 0  x   1  u
 2   2   
 x1 
y  1 0 
 x2 

K 
Where K e   e1 
K e2 

20

10
 3/23/2023

Soln to Example-1
[A  K e C]  x1  0 1   x1  0
 x   0 0  x   1  u
 2   2   
0 1   K e1 0  x1 
y  1 0 
 
 0
 x2 
0 0   K e 2 K 
K e   e1 
  K e1 1  K e2 
 
  K e 2 0 

21

Soln to Example-1
• The characteristic equation is
sI  (A  K eC)  0
 s 0    K e1 1   s  K e1  1
0 s     K 0    K s 
0
   e2   e2
s ( s  K e1 )  K e 2  0
s 2  K e1s  K e 2  0

22

11
 3/23/2023

 s 0   K e1 1  s  K e1  1
0 s     K 
0  K e 2 s 
0
   e2

Soln to Example-1
• Choosing suitable pole-placement at s=-5; -5
s 2  10 s  25  0.
Comparing with
s 2  K e1s  K e 2  0,
We have
K e1  10
K e 2  25
• This is the required observer for the given
system 23

Observer for double integrator system

Ke2 Ke1

• Where Ke1=10 and Ke2=25

0 1  0 
 K e1  A  ; B 
Ke    0 0 1 
K e2  C  1 0
24

12
 3/23/2023

Reduced Order State Observers

Motivation

• A full order observer estimates all the states in a

system, regardless of whether they are
measured or not
• This does not seem to be too bad, except when
e.g., we have a system with ten states and we
can measure eight of them;
• It would be better to estimate two states instead
of all ten states.
• The formalization of this procedure leads to the
reduced order observer.

25

Reduced Order State Observers

Motivation
• Suppose we can measure some of the state variables contained in
x.
• We partition the state vector x into two sets:
x 1 : variables that can be measured directly ,  a11 a12  a1n 

•
x 2 : variables that cannot be measured directly .
The state equations are broken down to
a
A A
  11
21   A12 
  
 
  Aa22nn 
an1 A21

• and the observation equation is

x 
x   1
• where C1 is square and nonsingular matrix. x 2 

26

13
 3/23/2023

Reduced Order State Observers

Motivation

• The full order observer for the states is then

• However it would be redundant to implement the first observer equation for

when we can solve for x1 directly as:

• In this case the observer for those states that cannot be measured directly
becomes

– which is a dynamic system of the same order as the number of state

variables that cannot be measured directly.
• The dynamic behavior of this reduced order observer is governed by
the eigenvalues of A22, a matrix over which the designer has no control.
• Since there is no assurance that the eigenvalues of A22 are suitable, we
need a more general system for the reconstruction of

27

Reduced Order State Observers

Derivation

• In general, we take

• Where

• Define the estimation error

28

14
 3/23/2023

Reduced Order State Observers

Derivation

• and we get

• Since

• We get

29

Reduced Order State Observers

Derivation

• In order for the error to be independent of x1, x2, and u,

the matrices multiplying x1, x2, and u must vanish

• Then

• and for stability the eigenvalues of F must lie in the left

half s–plane.

30

15
 3/23/2023

Selection of Reduced Order Observer Poles

• Therefore, the problem of reduced order

observer is similar to the full order observer with
(A22−LC1A12) playing the role of (A − LC)
• The error in the estimate will be determined by
the eigenvalues of (A22−LC1A12) which we can
• obtain from det[ sI − (A22−LC1A12) ] = 0.
• If (A,C) is observable, we can pick the elements
of L to give the error arbitrary dynamics, similarly
to the control design.

31

Reduced Order State Observer

Block Diagram

32

16
 3/23/2023

Example (ref: Burns)

• A ship roll stabilization system has a forward path transfer function

33

Solution to Example

34

17
 3/23/2023

Revision Exercises on Controllability, Observability and Full-order

Observers

• Ex 1:

• Ex 2:

35

Optimal Control

36

18
 3/23/2023

Approaches for Optimal Control

A historical account
• Variational calculus (Dreyfus, 1962)
– may be employed to obtain a set of differential equations with certain boundary
condition properties known as Euler-Lagrange equations
• Maximum Principle (Pontryagin, 1962)
– can also be applied to provide the same boundary conditions by using a
Hamiltonian function
• Euler-Lagrange and Pontryagin’s equations are applicable to systems with non-linear,
time-varying state equations and non-quadratic time-varying performance criterion

• Dynamic programming method (Bellman, 1957)

– based on the principle of optimality which yields the Hamilton-Jacobi partial
differential equation whose solution results in an optimal control policy
• The Hamilton-Jacobi equation is usually solved for the important and special case of
the linear time invariant plant with quadratic performance criterion called performance
index which takes the form of the matrix Riccati equation 1724.
• This produces an optimal control law as a linear function of the state vector
components which is always stable provided the system is controllable

37

Types of Optimal Control

Problems
• The terminal control problem
• The minimum-time control problem
• The minimum energy control problem
• The regulator control problem
• The tracking control problem

38

19
 3/23/2023

The terminal control problem

• This is used to bring the systems as close
as possible to a given terminal state within
a given period of time. E.g. an automatic
aircraft landing system, whereby the
optimal control policy will focus on
minimizing errors in the state vector at the
point of landing

39

The minimum-time control

problem
• This is used to reach the terminal state in
the shortest possible time period. This
usually results in a bang-bang control
policy whereby the control is set to umax
initially, switching to umin at some specific
time. In the case of a car journey, this is
the equivalent of the driver keeping his
foot flat down on the accelerator for the
entire journey, except at the terminal point,
when he brakes as hard as possible
40

20
 3/23/2023

The minimum energy control

problem
• This is used to transfer the system from an
initial state to a final state with minimum
expenditure of control energy. Used in
satellite control

41

The regulator control problem

• With the system initially displaced from
equilibrium, will return the system to the
equilibrium state in such a manner so as
to minimize a given performance index

42

21
 3/23/2023

The tracking control problem

• This is used to cause the state of a system
to track as close as possible some desired
state time history in such a manner so as
to minimize a given performance index.
This is the generalization of the regulator
control problem.

43

Performance Index

A
44

22
 3/23/2023

Objectives
• Minimizing the distance off-track, or cross-track error
ye(t). Wandering off-track will increase distance travelled
and hence time taken.
• Minimizing course or heading error Ψe(t). It is possible to
have zero heading error but still be off-track.
• Minimizing rudder activity, i.e., actual rudder angle (as
distinct from desired rudder angle) δa(t), and hence
minimizing the expenditure of control energy.
• Minimizing the forward speed loss ue(t). As the vessel
yaws as a result of correcting a track or heading error,
there is an increased angle of attack of the total velocity
vector, which results in increased drag and therefore
increased forward speed loss.

45

Performance Index
• A general performance index for the yacht problem can be written as:

 x1 
x   x2  Q  q11 q22 q33  46
 x3 

23
 3/23/2023

Quadratic Performance Index

• If the state and control variables are squared, then the performance index
becomes quadratic.

47

The Linear Quadratic

Regulator
• The Linear Quadratic Regulator (LQR)
provides an optimal control law for a linear
system with a quadratic performance
index.
– Continuous form
– Discrete form

48

24
 3/23/2023

Continuous form of LQR

• Recall that

• A Hamilton-Jacobi equation may be expressed as

49

Continuous LQR for Linear Time-Invariant Systems

• The state equation becomes

• If the performance index becomes quadratic,

• By substituting in the Hamilton-Jacobi equation, we

have,

50

25
 3/23/2023

Optimum Control Law

• Let us assume

• By substituting these expressions in the Hamilton-Jacobi equation, we have,

51

Optimum Control Law

• To minimize u from the above eqn, we have,

• This eqn can be re-arranged to give the optimal control law

52

26
 3/23/2023

Optimal Control Law Implementation

u
C

-K

53

Matrix Riccati Equation

• Substituting uopt back into the Hamiltonian-Jacobi eqn, we have,

• Since

• Then

• This nonlinear differential equation is known as the Matrix Riccati

Equation

54

27
 3/23/2023

Solution of the Matrix Riccati Equation

55

Example

56

28
 3/23/2023

Solution
• Given

• The matrix Riccati equation is given by

• Thus

57

Soln: …contd.

• Thus the Riccati equation is

• The above equation can be expressed as

• Hence,
• Solving, we have,

58

29
 3/23/2023

Soln: …contd.

• Also

• And

• Hence the Riccati matrix is

• (b) The state feedback matrix is

59

Soln …contd

• (c) The closed loop eigenvalues are given by

60

30
 3/23/2023

Riccati matrix for a 2nd order system

with diagonal matrix Q and scalar R

61

Exercise Anderson&Moore

• Find
– The Matrix Riccati equation
– The optimal Control Law
– The optimal performance index
62

31
 3/23/2023

Ans:
• The Matrix Riccati eqn is

63

Exercise Anderson&Moore

• For a system given by

• and
• Find
– The matrix Riccati eqn
– An optimal control law
– The poles of the closed loop system
64

32
 3/23/2023

Soln:

65

Soln:
• The closed loop system eqn is

66

33
 3/23/2023

Exercise Anderson&Moore

• The system is

• Find
– The optimal control law
– The closed loop system eqn using this control
law
67

Soln:

68

34
 3/23/2023

Exercises Anderson&Moore

69

More Exercises Anderson&Moore

70

35
 3/23/2023

Discrete form of LQR

• Let the discrete state equation of a system be written as

• The discrete performance index is given as

J=

71

Exercise:
• Write down the equations which describe
a discrete linear quadratic regulator.
• Write a flowchart to implement a discrete
LQR using a computer.

72

36
 3/23/2023

Robust Control Systems

An Introduction

• Machines, vehicles, industrial processes, command and control technology

and other dynamic engineering systems operate in changing environments
of
– raw materials,
– fuels, load levels,
– weather conditions, etc.
• It would be impractical and often impossible to describe the dynamics of
such systems by exact mathematical models.
• All control system designs are based on a model of the plant.
• It is assumed that the model of the plant is known precisely.
• However, the model we use is only an approximation of the dynamics of the
system.
• Analysis and design of robust feedback systems is one of the central
problems in modern control theory.
Robust control strives to characterize the uncertainty in the model of the plant
to be controlled and to evaluate the degrees of freedom left to achieve the
control task within specified bounds.

73

Closed loop system structural

diagram

Notations

74

37
 3/23/2023

Steps Involved in Control System Design

• Modeling or identification.
– The system designer develops mathematical models of the relevant aspects of
the system to be controlled. This can be done using knowledge of the system as
well as experimentally observing responses of the system to various excitations.
This is known as system identification.
• Control configuration :
– Selection and placement of sensors and actuators must be done. The designer
decides which signals in the system will be measured or sensed, with what
sensor hardware, and similarly what actuators will be used.
• Control law or controller design.
– The designer decides how the actuators are to be driven by processing the
incoming sensor signals. The control law and the controller describe the signal
processing used by the control processor to generate the actuator signals from
the sensor signals.
• Controller implementation.
– Once the controller is selected, the control processor which implements this law
must be designed.
• Control system testing and validation.

75

Sources of perturbations or
uncertainties in a plant
• The system under control may have been inaccurately
modeled or identified.
• The system under control, or the controller, physically
changes, perhaps due to component tolerances or
temperature coefficients.
• Certain non-linearities may have been ignored in the
design process, but may be significant in the real closed-
loop system.
• The operating point of a nonlinear system changes, so a
small signal linear approximation becomes less accurate.
• Gross failures, such as a sensor or actuator failure may
have occurred.

76

38
 3/23/2023

Uncertainties in a control system

Fictitious
uncertainty

n

1
2
:
n-1
u1
y1
Uncertainty

P11 P12

P21 P22
Practically, each of the
Plant I’s (i=1,...,n) may itself
y2 u2 be a matrix and
represent a different
F(s) kind of physical
uncertainty.
Controller

77

Classification of Uncertainties

• Structured Uncertainty
• Unstructured Uncertainty
–Additive Uncertainty
–Multiplicative Uncertainty

78

39
 3/23/2023

Structured uncertainty
• Structured uncertainty (often called parametric
uncertainty) represents parametric variations in
the plant dynamics,
• For example:
– Uncertainties in certain entries of state space
matrices (A, B, C), e.g., the uncertain variations in
an aircraft’s stability and control derivatives.
– Uncertainties in specific poles and/or zeros of the
plant transfer function.
– Uncertainties in specific loop gains/phases.

79

Structured uncertainty
• A typical example is the variation of the mass of
a car, which changes with the number of
passengers and the weight of the baggage.
• When linearizing a nonlinear system, the
parameters of the linearized model also depend
on the operating condition.
• It is straightforward to investigate effects of
parametric uncertainty simply by evaluating the
performance criteria for a range of parameters.
• Such a calculation directly reveals the
consequences of parameter variations.

80

40
 3/23/2023

Example: Cruise Control (Astrom book)

• The cruise control system of a car is one of the

most common control systems encountered in
everyday life.
• The system attempts to keep the speed of the
car constant in spite of disturbances caused by
– changes in the slope of a road and
– variations in the wind and road surface.
• The controller compensates for these unknowns
by measuring the speed of the car and adjusting
the throttle appropriately.

81

Example: Cruise Control

• Let v be the speed of the car and vr the desired (reference) speed.
• The controller, which is of the proportional-integral (PI) type receives the signals v and vr and generates a control
signal u that is sent to an actuator that controls throttle position.
• The throttle in turn controls the torque T delivered by the engine, which is then transmitted through gears and the
wheels, generating a force F that moves the car.
• There are disturbance forces Fd due to variations in the slope of the road, the effects of rolling resistance and
aerodynamic forces.
• The cruise controller also has a man-machine interface that allows the driver to set and modify the desired speed.
• There are also functions that disconnects cruise control when the brake is touched as well as functions to resume
• cruise control operation.
• The system has many individual components—actuator, engine, transmission, wheels and car body—and a
detailed model can be very complicated.
• In spite of this, the model required to design the cruise controller can be quite simple.
• In essence the model should describe how the car’s speed is influenced by the slope of the road and the control
signal u that drives the throttle actuator.

82

41
 3/23/2023

When are Two Systems Similar?

• A fundamental issue is to determine when
two systems are close.
• This seemingly innocent problem is not as
simple as it may appear.
• A naive idea is to say that two systems are
close if their open loop responses are
close.
• Even if this appears natural, there are
complications
83

Example: Systems which are similar in open loop

but have large differences in closed loop

• Consider the systems with the transfer functions

• with T = 0.025.

84

42
 3/23/2023

Example: Systems which are similar in open loop

but have large differences in closed loop

• These systems have very similar open loop responses

for small values of T.
• The differences between the step responses are barely
noticeable in the figure.

85

Example: Systems which are similar in open loop

but have large differences in closed loop
• The step responses with unit gain error feedback are shown in the
figure.
• Notice that one closed loop system is stable and the other one is
unstable.
• The transfer functions from reference to output are
• T1=100/(s+101); T2= 100/ (0.000625 s^3 + 0.05063 s^2 + 1.05 s + 101)
• Routh Hurwitz criterion shows that the system2 is unstable in closed
loop.

86

43
 3/23/2023

Example: Systems which are similar in open loop

but have large differences in closed loop 1
Step Response

0.8

• Routh array for system2 0.6

Amplitude
• s3: 0.000625 1.05 0.4

• s2: 0.05063 101 0.2

• s1: -0.1968 0
0 0.1 0.2
Time (sec)
0.3 0.4 0.5

Step Response

• s0: 101
3

2.5

1.5

• MATLAB plots:
Amplitude

0.5

-0.5

-1
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Time (sec)

87

Example: Systems which are different in open

loop but similar in closed loop

• Consider the systems

88

44
 3/23/2023

Example: Systems which are different in open

loop but similar in closed loop

• The open loop responses are very

different because P1 is stable and P2 is
unstable, as shown in the Figure.

89

Example: Systems which are different in open

loop but similar in closed loop
• Closing a feedback loop with unit gain around the systems we find
that the closed loop transfer functions are

• which are very close as is also shown in Figure

90

45
 3/23/2023

Example: Systems which are different in

open loop but similar in closed loop
• Bode plots:
100/(s+1) 100/(s-1)
Bode Diagram
Bode Diagram 40
40
30
30

Magnitude (dB)
Magnitude (dB)

20 20

10 10

0
0
0 -90
Phase (deg)

Phase (deg)
-45
-135

-90
-2 -1 0 1 2
10 10 10 10 10
-180
Frequency (rad/sec) -2 -1 0 1 2
10 10 10 10 10
Frequency (rad/sec)

91

Kharitonov’s Theorem
• In 1978 V.L.Kharitonov published a stability theorem for
classes of polynomials defined by letting each coefficient
vary independently in a specified (but arbitrary) interval.
• The effect of uncertain independently varying coefficients
had been examined by Kharitonov.
• Though it is generally assumed that the characteristic
polynomial for a system is known precisely, many
systems have parameters that are uncertain within a
range.
• In such cases, the coefficients of the characteristic
polynomial are unknown except for the bounds.

92

46
 3/23/2023

Kharitonov’s Theorem
• Let us consider a system with the characteristic equation

p( s)  s n  an 1s n1  an 2 s n 2  ...  a0  0

• For all a0, …, an-1 such that

a k  ak  ak , k  0,1,..., n  1
• A polynomial with real coefficients is said to be Hurwitz if
and only if all of its zeros lie in the left-half of the s-plane. A
set of polynomials is said to be Hurwitz if and only if every
member is Hurwitz.
• The remarkable result of Kharitonov states that the whole
class of polynomials is Hurwitz if and only if four special,
well-defined polynomials are Hurwitz.

93

Kharitonov’s Theorem
• Let us define the following polynomials
n
g1 ( s ) : a0  a2 s 2  a4 s 4 ...  j
k  0 ,even
k
. min{ j k .ak , j k ak }.s k

n
g 2 ( s ) : a0  a2 s 2  a4 s 4 ...  j
k  0 ,even
k
. max{ j k .ak , j k ak }.s k

n
h1 ( s ) : a1 s  a3 s 3  a5 s 5 ...  j
k 1, odd
k 1
. min{ j k 1 .ak , j k 1 ak }.s k
n
h2 ( s ) : a1 s  a3 s 3  a5 s 5 ...  j
k 1, odd
k 1
. max{ j k 1 .a k , j k 1 a k }.s k

• where
an : an : 1)

• The Kharitonov polynomials are

k kl ( s ) : g k ( s )  hl ( s ), k , l  1,2
94

47
 3/23/2023

Example 1
• Find out whether the system with the following characteristic
equation will have robust stability

p( s )  s 3  a2 s 2  a1s  a0  0

• where
8  a0  60
12  a1  100
7  a2  25

95

Soln.:
• The four Kharitonov polynomials are
k11 ( s )  s 3  a 2 s 2  a1s  a 0
k11 :
k12 ( s )  s  a 2 s  a1s  a 0
3 2
1 12
k 21 ( s )  s 3  a 2 s 2  a1s  a 0
k ( s)  s  a s  a s  a
22
3
2
2
1 25
0 8
• Substituting the values of the coefficients, we have, 12  25  8

k (s )  s  25s  12s  8
3 2 25
11
8
k12 (s )  s  25s 2  100 s  8
3

k 21 ( s )  s 3  7 s 2  12 s  60
k 22 ( s )  s 3  7 s 2  100 s  60

• By checking these four polynomials using Routh-Hurwitz

criterion, it can be concluded that the system has robust
stability for all uncertain parameters varying within the given
range.
96

48
 3/23/2023

Example 2
• In a position control system of a damped rotating gun turret shown in fig,
the nominal system parameters and their uncertainty ranges are
• a=4+1, T=0.15+0.1, K=2.5+0.5.
• Determine the stability of the closed loop system.

R(s) C(s)
+ K 1
1  sT s s  a 
-

1  aT 1
c2   a
T T
1
c2  3
0.25 97

Soln.:
• Characteristic polynomial is
s3+{(1+aT)/T}s2+(a/T)s+K/T=0
or, c3s3+c2s2+c1s+c0=0
where,
c0 Є[8,60], c1 Є[12,100], c2 Є[7,25], c3=1
The four Kharitonov polynomials are
K11(s)=8+12s+25s2+s3
K12(s)=8+100s+25s2+s3
K21(s)=60+12s+7s2+s3
K22(s)=60+100s+7s2+s3

Performing Routh’s test we conclude that system has robust stability for
all parameter values in the specified ranges.

98

49
 3/23/2023

Simplifications for polynomials of low

degree
• Anderson, Jury and Mansour demonstrated that the Kharitonov
condition could be reduced for polynomials of low degree.
• For polynomials of degree 3, 4, or 5, the Kharitonov test is
equivalent to checking 1, 2, or 3 of the Kharitonov polynomials
respectively.
• For degree greater than 5, all four polynomials must be checked.

• When n=3 and a0>0, the class of polynomials is Hurwitz if and only if
k21(.) is Hurwitz.
• When n=4 and a0>0, the class of polynomials is Hurwitz if and only if
{k21(.),k22(.)} is Hurwitz.
• When n=5 and a0>0, the class of polynomials is Hurwitz if and only if
{ k12(.),k21(.),k22(.)} is Hurwitz.

99

Limitation of Kharitonov's
Theorem
• One of the fundamental limitations
Kharitonov's Theorem is that
– the theorem only applies to polynomials with
independent coefficient perturbations. Note
that uncertainty in the physical parameters of
a linear system typically results in dependent
perturbations in the coefficients of the
characteristic polynomial.

100

50
 3/23/2023

Exercises:
1. Determine if the following systems are robustly
stable for all admissible uncertainties using the
Kharitonov theorem
(a) Ø(s, p) = p0+p1s+p2s2+p3s3+p4s4, where p0 Є[1,3], p1
Є[7,13], p2 Є[2,5], p3 Є[4,14], and p4 Є[3,12].
(b) Ø(s, p) = p0+p1s+p2s2+p3s3+p4s4+p5s5, where p0 Є[1,5], p1
Є[4,7], p2 Є[3,8], p3Є[5,11], p4 Є[80,12] and p5 Є[7,12].
(c) Ø(s, p) = p0+p1s+p2s2+p3s3+p4s4+p5s5+p6s6, where p0
Є[720,725], p1 Є[1448,1458], p2 Є[1213,1227], p3
Є[535,542], p4 Є[131,142], p5 Є[17,19] and p6 Є[1,1] .

101

Prob. 2

102

51
 3/23/2023

Prob 3

103

Additional problems
• Work out problems from D.Roy
Choudhury’s book : Modern Control
Engineering.

104

52
 3/23/2023

Unstructured uncertainty
• It is generally fairly easy to investigate the effects of parametric
variations.
• There are however other uncertainties that also are important.
– The simple model of the cruise control system only captures the
dynamics of the forward motion of the vehicle and the torque
characteristics of the engine and transmission. It does not, for example,
include a detailed model of the engine dynamics (whose combustion
processes are extremely complex), nor the slight delays that can occur
in modern electronically controlled engines (due to the processing time
of the embedded computers).
• These neglected mechanisms are called unmodeled dynamics.
• Unstructured uncertainty refers to that aspect of system uncertainty
associated with unmodelled dynamics, truncation of high frequency
modes, nonlinearities and the effect of linearization, time variation
and randomness in the system.

1 
;
s s  105

Additive Unstructured Uncertainty

When the relation of the uncertainty to the nominal plant is additive.
G  G  A
True plant G

A

F G
G

A

M(s)

106

53
 3/23/2023

Multiplicative Unstructured Uncertainty

• When the relation of the uncertainty to the nominal plant is multiplicative.

G  ( I   M )G

True plant G

M

F G

M(s)

107

System Sensitivity and Robustness

• For small parameter perturbations, the ‘system sensitivity’ may be used as a measure of
robustness.
• System sensitivity is the ratio of the change in system transfer function (or parameter) for a small
incremental change.
T T / T
• SG 
G / G
G ( s)
• For a closed-loop system, the closed loop transfer function is , T ( s ) 
1  G( s ) H ( s)
• When G(s) = P(s) C(s); H(s)=F(s)=1.
• Sensitivity of the feedback system is C P
T T G 1
SG  . 
• G T 1  G ( s ) H ( s )

• System sensitivity may be reduced below that of open loop system by increasing G(s)H(s) over
the frequency range of interest.
• Thus, the sensitivity function of a closed loop system is given by 1
S
1  PC
PC
• And the complementary sensitivity function of a closed loop system is given by T
1  PC
• It may be noted that S+T=1.
G

H 108

54
 3/23/2023

Exercise:
• 1. For a process whose dynamics is a pure time delay,
s   e  s function is
the Ptransfer 1
The ideal delay compensator is a controller with theC s   1  e s

transfer function S s   1  e  s
Show that the sensitivity function is
and that the closed loop system will beT  unstable for
S 
arbitrary small changes in the delay.

s   1
• 2. For the following system, show that

• 3. Examine the sensitivity (w.r.t. gain K) of the second-

order unity feedback system shown below.
109

Internal Stability

110

55
 3/23/2023

Internal Stability

111

Exercise
• 1. Given a system with

Check whether the system is internally stable or not.

• 2.

112

56
 3/23/2023

Exercise
3. A unity feedback system is given by
C(s)=10 and P(s)=1/(s-a), where a is real.
Find the range of a for the feedback
system to be internally stable.

113

Supremum(Least Upper Bound)

• Upper Bound
– An element b is called an upper bound for the set X if
every element in X is less than or equal to b. If such
an upper bound exists, the set X is called bounded
above.
X={1, 2, 3, 4, 5}
Let b=6; b=5; b=10; b=101.56
• Supremum or Least Upper Bound
– An element b in A is called a least upper bound (or supremum) for X if b is an upper bound
for X and there is no other upper bound b' for X that is less than b. We write b = sup(X). By
its definition, if a least upper bound exists, it is unique. The supremum does not have to be
part of the original set.

114

57
 3/23/2023

Supremum(Least Upper Bound)

• Upper Bound
– An element b is called an upper bound for the set X if every element in X is less
than or equal to b. If such an upper bound exists, the set X is called bounded
above.

• Supremum or Least Upper Bound

– An element b in A is called a least upper bound (or
supremum) for X if b is an upper bound for X and
there is no other upper bound b' for X that is less than
b. We write b = sup(X). By its definition, if a least
upper bound exists, it is unique. The supremum does
not have to be part of the original set.
X={all real numbers between 0 and 1}
b=5; b=1; b=7.5

115

Example 1 a)
• Find the supremum for the set S of all rational
numbers strictly between 0 and 1.
– This set has many upper bounds, but only one least upper
bound.
– Five different upper bounds for S are: 1, 10, 100, 42, and Pi.
– Note: An upper bound is therefore not unique. All that is
required is to find number bigger than all other numbers in the
set S.
– To find the least upper bound for S we need to find a number x
such that
• x is an upper bound for S
• there is no other upper bound lower than x
– Clearly, this upper bound is 1. Note that the supremum, or least
upper bound, is unique, but it is not part of the original set S.

116

58
 3/23/2023

Example 1 b)
• Repeat example 1 a) for the case
when the numbers 0 and 1 are
included in the original set S.
• The answers in this case are the same as
that in Ex 1 a) above i.e., the least upper
bound is again 1.
• However, this time the unique supremum
is part of the set S, which may or may not
happen in general.
117

Example 2
• Find the supremum for the set of rational numbers {1,
1.4, 1.41, 1.414, 1.4142, ...} converging to the square
root of 2.
– No number bigger than √2 is the least upper bound (although
each of these numbers is an upper bound), because if x was that
least upper bound, then we can find a rational number between
√2 and x. That rational number would then be an upper bound
smaller than x, which is a contradiction.
– No number less than √2 is the least upper bound, because if x
was that least upper bound, there is some element of the set
between x and √2 . But then x is not an upper bound, which is a
contradiction.
– Hence, the least upper bound of this set is √2 .

118

59
 3/23/2023

Infimum
• Lower bound
– An element b is called a lower bound for the set X if
every element in X is greater than or equal to b. If
such a lower bound exists, the set X is called
bounded below.
• Infimum or greatest lower bound
– An element b in A is called a greatest lower bound (or
infimum) for X if b is a lower bound for X and there is
no other lower bound b' for X that is greater than b.
We write b = inf(X).
– By its definition, if a greatest lower bound exists, it is
unique.
119

Norms
• Norms for Signals

120

60
 3/23/2023

Norms for Systems

• The H2 norm of a system is given by

where g(t) is the impulse response of the controlled

system from input u to output y.

• H∞ norm of a system is given by

G 
 sup G  j 
R

where G(s) is the transfer function of the system

121

Computation of H2 norm of a system

• If G is strictly proper and has no poles on the imaginary axis i.e. its 2-norm is finite,
• We have,
 2
1
 G  j  d
2
G 
2 2

j
1

2j  G s Gs ds
 j

1
 G  s G s ds

2j
• The last integral is a contour integral up the imaginary axis, then around an infinite
semicircle in the left half-plane.
• In case you forgot:
– G(s) is proper if G(j∞) is finite (degree of denominator >= degree of numerator)
– G(s) is strictly proper if G(j∞) = 0 (degree of denominator > degree of numerator)
– G(s) is bi-proper if G(s) and 1/ G(s) are both proper (degree of denominator = degree of
numerator)

122

61
 3/23/2023

Computation of H2 norm of a system

2
• By residue theorem, G 2 equals the sum of the
residues of G(-s)G(s) at its poles in the left half
plane.

• Example: Given G(s) = 1/(Ts+1), T>0. Find G 2

123

Please recollect that

If f(z) has a simple pole at , then

If f(z) has a pole of order 2 at , then

If f(z) has a pole of order 3 at , then

. .
.

If f(z) has a pole of order k at then

124

62
 3/23/2023

Example: contd.

• Given G(s) = 1/(Ts+1), T>0. Find G 2

• Soln:
G(-s) G(s) = 1/{(Ts+1) (-Ts+1)}
The left half-plane pole of G(-s)G(s) is at s=-1/T.
The residue at this pole equals

Hence

125

Example2
• Given G(s)=1/((s+1)2(s+2)). Find ||G||2.

R2=1/((s+1)2(-s+1)2(-s+2)) {@s=-2}=1/36
R1=d/ds(1/((s+2) (-s+1)2(-s+2)){@s=-1}=1/36

G2=sqrt(R1+R2)

126

63
 3/23/2023

Example2 … soln
1
Given G s   . Find G 2 .
s  1 s  2 
2

d  1 
R1  Lt s 1   
 ds  s  2 s  1  s  2 
2

 1  
  
 s  2   s  1  s  2 
2 2

 Lt s 1 
  2   1 
 
  s  2 s  13  s  2   s  2 s  12  s  2 2 
1 2 1
  
4  3 3 8 4  9
662 1
  127
72 36

Example2 … soln contd.

1
Given G s   . Find G 2 .
s  12 s  2
 1 
R2  Lt s 2  
 s  1  s  1  s  2 
2 2

1 1
 2 
3  4 36
1 1 1 1
 G 2  R1  R2    
36 36 18 3 2
128

64
 3/23/2023

Computation of H∞ norm of a system

• Search method

• Computation of derivative

129

Search method
• Search method  Bode Magnitude plot
• ||G||∞ can be computed from the Bode Magnitude plot
• Example:
Consider a system with transfer function given by
as  1
G (s) 
bs  1

with a, b >0. Find ||G||∞

130

65
 3/23/2023

Soln.:
• ||G|| ∞ = a/b, a>b
= 1, a<=b
Bode Diagram Bode Diagram
4 0

3
-1
Magnitude (dB)

Magnitude (dB)
2
-2

1
-3

0
15 -4
0

10
Phase (deg)

-5
Phase (deg)

5
-10

0
-2 -1 0 1
10 10 10 10
-15
Frequency (rad/sec) -2 -1 0 1
10 10 10 10
Frequency (rad/sec)

131

Exercise
• Compute the infinity norm of the following
system:
G(s)=1/(s2+10-6s+1).
Write a flowchart for using a search
algorithm to find the above norm.

132

66
 3/23/2023

H∞ control problem

• The top left figure shows a general

representation of a control problem used in
robust control.
• The input u represents the control signal,
• the input w represents the external influences x
on the system,
• the output z is the generalized error and
• the output y is the measured signal.
• The right figure shows the special case of the
system in Figure where the reference signal is
zero.
• w = (−n, d) and z = (x, v).

133

H∞ control problem

• The closed loop system is represented by two blocks, the process P

and the controller C.
• The process P has two inputs, the control signal u which can be
manipulated by the controller, and the generalized disturbance w,
which represents all external influences, for example command
signals and disturbances.
• The process has two outputs, the generalized error z which
• is a vector of error signals representing the deviation of signals from
their desired values and the measured signal y which can be used
by the controller to compute u.
• For a linear system and a linear controller the closed loop system
can be represented by the linear system
• z = H(P(s),C(s))w
• where w = (−n, d) and z = (x, v).

134

67
 3/23/2023

H2 and H∞ control
• This tells us how the generalized error depends
on the generalized disturbances w.
• The control design problem is to find a controller
C such that the gain of the transfer function H is
small even when the process has uncertainties.
• There are many different ways to specify
uncertainty and gain, giving rise to different
designs.
• The names H2 and H∞ control corresponds to
the corresponding norms |H|2 and |H|∞.

135

H-infinity Control Example

68
 3/23/2023

Soln:
 PC P 
 x  n  1  PC   n 
z     H ( P, C )    1  PC
v  d   C 1   d 

1  PC 1  PC 

• The size (norm) of H would indicate the

effect of the noises and disturbances on
the plant output or controller output

Soln ...contd

 PC P 
 x  n  1  PC   n 
z     H ( P, C )    1  PC
v  d   C 1   d 

1  PC 1  PC 

• The infinity norm would be obtained from

the largest eigenvalue of the matrix

69
 3/23/2023

Steps for calculating largest

eigenvalue of a matrix
• Step 1: Calculate H*H’
• Step 2: Find the eigenvalues of H*H’
• Step 3: Take the largest eigenvalue of
H*H’
• Step 4: Find the square root of the largest
eigenvalue of H*H’ to get the largest
eigenvalue of H

Step 1:
• Calculate H*H’
 PC P 
 1  PC 
H  1  PC
C 1 
 
1  PC 1  PC 
1  PC P   PC C 
H * H ' 2 
1  PC   C 1   P 1 

or , H * H ' 
1 
P 2 C 2  1  PC  1
2

1  PC 2 

 P C 1
2
 C  1 
2

C 2  1 P 2 P
or , H * H ' 
1  PC 2  P 
1

70
 3/23/2023

Step 2:
• Find the eigenvalues of H*H’

I  HH '  0
C 2 1 P2 P
I    0
1  PC 2 P 1
 
P2 C 2 1  
P C 2 1  
  
 1  PC 2 1  PC 2   0
or ,
 P C 12
 

C2 1  
  1  PC 2 1  PC 2 


Step 3:
• Find the largest eigenvalues of H*H’

 1  PC 2
 
 P 2 C 2  1  1  PC   C 2  1  P 2 C 2  1  0
2
  
2

 1  PC    1  PC  C
2 4 2 2

 1   1  PC  P 2 C 2  1
2
 
 
2
 P2 C 2 1  P2 C 2 1  0  
2

2 1  PC 2   P 2  1C 2  1  0

Hence,

  0;
P 2

1 C 2 1 
1  PC 2
142

71
 3/23/2023

Step 4:
• Find the square root of the largest eigenvalue of
H*H’ to get the largest eigenvalue of H
C 2

1 P2 1 
1  PC 2

C 2

1 P2  1 
1  PC 
This is the largest eigenvalue of H

Step 4: … contd.

Hence, H P, C    sup

C  j   1 P j   1
2 2

 1  C  j P  j 

 PC P 
 1  PC 
where H P, C   1  PC
C 1 
 
1  PC 1  PC 

72
 3/23/2023

Steps for H-infinity Controller

Design

H P, C    sup
 C  j   1 P j   1
2 2

 1  C  j P j 

• Choose a value of γ which is small but >0

• Find a controller such that |H(P, C)|∞ < γ
– There are efficient numerical methods to find
a controller such that |H(P, C)|∞ < γ, if such a
controller exists.
• The best controller can then be found by
iterating on γ.

Notes:
• If H P, C  is minimized
P
– > > G yd  1  PC i.e. transmission of load
disturbances to the output is small
C
– >> Gun  1  PC i.e. transmission of measurement
noises to the controller output signal is small
– >>Since the Sensitivity Function S and the
Complementary Sensitivity Function T are also
elements of H(P,C), it is also guaranteed that S
and T are less than 
• Thus this design method balances
performance and robustness

73
 3/23/2023

Notes ... 2
• Minimizing the gain H(P,C)∞ means that the gains of all
individual signal transmissions from disturbances to outputs
are less than γ for all frequencies of the input signals.
• The assumption that the disturbances are equally important
and that all frequencies are also equally important is not very
realistic
– load disturbances typically have low frequencies
– measurement noise is typically dominated by high frequencies.
• Hence the problem may be modified so that disturbances of
different frequencies are given different emphasis by
introducing weighting filters
• E.g., Low-frequency load disturbances may be enhanced by
choosing a low-pass filter as a weighting filter and re-
formulating the problem

Exercises:
1. Given a first order plant with transfer
function P(s)=10/(s+10) and a
proportional controller with C(s)=10. Find
|H(P, C)|∞ .
2. Write a flow chart for design of an H-
infinity controller for a given plant.
3. Given a plant with transfer function
P(s)=100/(s2+14s+100). Use MATLAB to
design an H-infinity controller for the
given plant

74
 3/23/2023

Robust Stability / Stability Robustness

• Stability robustness guarantees closed-

loop stability under a wide range of plant
parameter perturbations and input
disturbances

149

Robust Stability
• Let us assume that the plant transfer function P
belongs to a set P.
• A controller C is robust with respect to some
characteristic if this characteristic holds for every
plant in P.
• E.g. Let us consider internal stability of the plant
as the characteristic of interest. A controller C
provides robust stability if it provides internal
stability for every plant in P.
• If P has an associated size, the maximum size
such that C stabilizes all of P is an useful notion
of stability margin.
150

75
 3/23/2023

Test for robust stability

Kharitonov’s theorem

is
a test for
robust stability
of systems with structured uncertainties.

151

Test for robust stability

• The Nyquist plot gives information about the robust stability.
• The distance from the critical point -1 to the nearest point on the
Nyquist plot of L equals 1/||S||∞. Where L is the loop transfer function
and S is the sensitivity function.

• Thus if ||S|| >>1, the Nyquist plot comes close to the critical point
and the feedback system is nearly unstable.
• However, as a measure of stability margin, this distance is not
entirely adequate because it contains no frequency information

152

76
 3/23/2023

Conditions for permissible process uncertainties

• Consider a feedback system with a process P and a

controller C in the forward path.
• If the process is changed from P to P + ΔP, the loop transfer
function changes from PC to PC + C Δ P.
• If we have a bound on the size of P, then the system
remains stable as long as the process variations never
overlap the −1 point, since this leaves the number of
encirclements of −1 unchanged
L(jw)=P(iw)C(jw)

153

Note
• This condition is conservative:
– it allows for any perturbation that satisfies the
given bounds, while in practice we may have
more information about possible
perturbations.

154

77
 3/23/2023

Calculation of permissible process uncertainties

• The distance from the critical point −1 to the loop

transfer function L is |1 + L|.
• This means that the perturbed Nyquist curve will
not reach the critical point −1 provided that
|C Δ P| < |1 + L|
• which implies

• This condition must be valid for all points on the

Nyquist curve, i.e pointwise for all frequencies.
The condition for stability can thus be written as
155

Calculation of permissible process uncertainties

• This condition allows us to reason about uncertainty without exact

knowledge of the process perturbations. Namely, we can verify
stability for any uncertainty P that satisfies the given bound.
• From an analysis perspective, this gives us a measure of the
robustness of a given design.
• Conversely, if we require robustness of a given level, we can
attempt to choose our controller C such that the desired level of
robustness is available (by asking T to be small)
• The variations can be large for those frequencies where T is small
and smaller variations are allowed for frequencies where T is large.
• A conservative estimate of permissible process variations that will
not cause instability is given by

• where Mt is the largest value of the complementary sensitivity

156

78
 3/23/2023

Example
For a cruise control system, the model of the car in
fourth gear at speed 25 m/s is

and the controller is a PI controller with gains

k = 0.72 and ki = 0.18.
Find
– Mt , i.e., is the largest value of the complementary sensitivity and the
corresponding frequency ωmt.
– the allowable size of the process uncertainty ΔP
– Permissible values of gain and phase variations when a controller is
designed to provide the largest value of the complementary sensitivity
Mt

157

Solution … 1
P(s)C(s) =
0.9936 s + 0.2484
------------------------
s^2 + 0.0142 s
CLTF =
0.9936 s + 0.2484
------------------------------
s^2 + 1.008 s + 0.2484

158

79
 3/23/2023

Solution … 2

• Difficult to
Magnitude (dB)

obtain the
maximum
magnitude
from the Bode
plot
• Hence
method of
Magnitude (abs)

differentiation
should be
used

159
Frequency (rad/s)

Solution … 3
Bode Diagram
1.5
Magnitude (abs)

1 System: CLTF
Frequency (rad/s): 0.353
Magnitude (abs): 1.14
0.5

• The complementary sensitivity has its maximum Mt = 1.14 at ωmt = 0.35.

• This gives the maximum allowable process uncertainty, with
|ΔP/P| < 0.87 or |ΔP| < 3.47
• At low frequencies, T(0) = 1 and so the perturbations can be as large as the
original process (|ΔP/P| < 1).
• At high frequencies T -> 0 and hence the relative error can get very large

• For example, at ω = 5 we have |T(jω)| = 0.195 which means that the stability
requirement is |ΔP/P| < 5.1.
• The analysis clearly indicates that the system has good robustness and that
that the high frequency properties of the transmission system are not
important for the design of the cruise controller
160

80
 3/23/2023

Calculation of permissible values of gain

and phase variations

• If the permissible value of process

uncertainties is known i.e., the largest
value of the complementary
sensitivity function is known,
– The method of M-circles i.e., constant M
loci in the G(jw) plane can be used for
calculation of permissible values of gain
and phase variations
161

Constant M loci in the G(jw) plane

• Graphical method using forward path
transfer function G(jw).
• Method is directly applicable to unity-
feedback systems though with some
modifications, it can also be applied to
non-unity feedback systems.

162

81
 3/23/2023

The method
• Consider that G(s) is a fwd-path TF of a unity
feedback system.
• CLTF is M(s)=G(s)/(1+G(s)).
• When s=jw, G(jw)=Re[G(jw)]+j Im[G(jw)]=x+j y
• where x denotes Re[G(jw)] and y denotes
Im[G(jw)].
• The magnitude of the CLTF is
• |M(jw)|=|G(jw)/(1+G(jw))|=(x2+y2)0.5/((1+x)2+y2)0.5
• Let M denote |M(jw)|.
• Then, M ((1+x)2+y2)0.5= (x2+y2)0.5

163

The method
• Squaring both sides, M2 ((1+x)2+y2) = x2+y2
• Rearranging, we have, (1- M2)x2+ (1- M2) y2-
2M2x =M2.
• Dividing both sides by (1- M2) and adding (M2/
(1- M2)) 2 to both sides, we have,
• (x- M2/ (1- M2)) 2+y2= (M/ (1- M2)) 2 M nt.
eq. 1.
• For a given value of M, this represents a circle
with centre at x= Re[G(jw)] = M2/ (1- M2); y=0.
• The radius of the circle is r= |M/ (1- M2)|;
• This eqn is not valid when M=1. Then, x=-1/2.
164

82
 3/23/2023

Calculation of permissible values of gain

and phase variations
• Mt = 2

-4/3 -2/3
0,0

Phase angle phi

Unit circle
Constant Mt radius 165

Soln.:

• If Mt = 2 then pure gain variations of 50%

or pure phase variations of 30◦ are
permitted without making the closed loop
system unstable

166

83
 3/23/2023

Robust Performance / Performance Robustness

• Often, stability is not the only property of a closed-loop system that

must be robust to perturbations.
• Typically, there are exogenous disturbances acting on the system
(wind gusts, sensor noise) which result in tracking and regulation
errors.
• Under perturbation, the effect that these disturbances have on error
signals can greatly increase.
• In most cases, long before the onset of instability, the closed-loop
performance will degrade to the point of unacceptability, hence the
need for a “robust performance" test.
• Such a test will indicate the worst-case level of performance
degradation associated with a given level of perturbations.
• Performance robustness study strives to produce desired
performance under a wide range of plant parameter perturbations
and input disturbances

167

Performance in the Presence of

Uncertainty
• Disturbance Attenuation
• Command Signal Following

168

84
 3/23/2023

Disturbance Attenuation

• A simple criterion for disturbance attenuation is to compare the output of the

closed loop system in Fig with the output of the corresponding open loop
system.
• If we let the disturbances for the open and closed loop systems be identical,
the output of the closed loop system is then obtained simply by passing the
open loop output through a system with the transfer function S

OL: y/d=P
CL: y/d=P/(1+PC)=PS
169

Disturbance Attenuation and Sensitivity

Function

• The sensitivity function thus tells how the variations in

the output are influenced by feedback.
– Disturbances with frequencies such that |S(jω)| < 1, are
attenuated by feedback
– Disturbances with frequencies such that |S(jω)| > 1 are amplified
by feedback.
• The maximum sensitivity Ms and the sensitivity crossover
frequency ωsc are simple performance measures.
|S|
OL: y/d=P
CL: y/d=P/(1+PC)=PS
w

170

85
 3/23/2023

Effect of transfer function from load disturbances to

process output on disturbance attenuation

• Transfer function from load disturbances to process

output

• Load disturbances typically have low frequencies and it

is therefore important that the transfer function is small
for low frequencies.

171

Ex:
• Find how the transfer function Gyd is
influenced by small variations in the
process transfer function

172

86
 3/23/2023

Soln.:
• Differentiating, we have,

• dGyd/dP=1/(1+PC)^2
• dGyd/dP=(1/(1+PC))*(P/(1+PC))/P

• The response to load disturbances is thus

insensitive to process variations for frequencies
where |S(jω)| is small, i.e. for those frequencies
where load disturbances are important.

173

Effect of measurement noise

• In addition to the load disturbance rejection, it is also important that
the control actions generated by measurement noise are not too
large.
• From Fig, the transfer function Gun from measurement noise to
controller output is given by |C|

• Since measurement noise typically has high frequencies it is w

important that the transfer function Gun is not too large for high
frequencies.
• The loop transfer function PC is typically small for high frequencies,
which implies that Gun ≈ C for large s.
• To avoid injecting too much measurement noise it is therefore
important that C(s) is small for large s.
• This property is called high frequency roll-off.

174

87
 3/23/2023

Effect of parametric variations on Gun

• Differentiating, we have,

• Measurement noise typically has high

frequencies.
• Since the complementary sensitivity function
also is small for high frequencies we find that
process uncertainty has little influence on the
transfer function Gun for frequencies where
measurement are important.

175

Command Signal Following

• The transfer function from reference to output is given by

• which is the complementary sensitivity function.

• To see how variations in P affect the performance of the
system, we diff. eqn. w.r.t. the process transfer function:

176

88
 3/23/2023

Discussions
• Inferences:
– The relative error in the closed loop transfer function thus equals the
product of the sensitivity function and the relative error in the process.
– In particular, it follows from equation that the relative error in the closed
loop transfer function is small when the sensitivity is small.
• This is one of the very useful properties of feedback.
• Assumptions:
– The analysis is limited to small (differential) perturbations.
– The process perturbations dP be stable so that we do not introduce any
new right half plane poles that would require additional encirclements in
the Nyquist criterion.
• Note:
– This condition is conservative: it allows for any perturbation that satisfies
the given bounds, while in practice we have more information about
possible perturbations.

177

Information obtained from the

sensitivity functions

• Disturbance rejection is high

for frequencies where S is
small
• Sensitivity to process
uncertainties is low for
frequencies where S is small
• Tracking performance is
good when T is close to 1.

178

89
 3/23/2023

Useful design goal

S ≈ 0 and T ≈ 1.
w
• Recall that S=1/(1+PC) and T=PC/(1+PC) i. e.,
S+T=1
• L (=PC) is typically large as s → 0 and L → 0 as
s → ∞.
• This means that
 S is typically small as s → 0 and S → 1 as s → ∞.
 T → 1 as s → 0 and T → 0 as s → ∞.

179

Exercise

180

90
 3/23/2023

Exercise

181

Exercise

182

91
 3/23/2023

Limits on the Sensitivities

• A basic problem is to investigate if S can be made small over a large frequency range

Bode Integral
You can move “dirt”
but not eliminate it.

Integral of log
sensitivity

Waterbed effect

Decrease Sensitivity in low frequency and

it will increase at the High frequency end 183

Bode’s integral formula:

• Let S(s) be the sensitivity function for a feedback system and assume that it goes to
zero faster than 1/s for large s.
• If the loop transfer has poles pk in the right half plane then the sensitivity function
satisfies the following integral:

• This implies that

– there are fundamental limitations to what can be achieved by control
– control design can be viewed as a redistribution of disturbance attenuation over different
frequencies.
– if the sensitivity function is made smaller for some frequencies it must increase at other
frequencies.
– This means that if disturbance attenuation is improved in one frequency range it will be
worse in other.
• This is called the waterbed effect.
• It also follows that systems with poles in the right half plane have larger sensitivity.
• For a loop transfer function without poles in the right half plane equation reduces to

184

92
 3/23/2023

Waterbed Effect

WATERBED EFFECT!!

185

Geometric interpretation of Bode’s integral

formula

• When log |S(jω)| is plotted as a function of

ω, the area over the horizontal axis must
be equal to the area under the axis when
frequency is plotted on a linear scale
186

93
 3/23/2023

Analogous formula for

complementary sensitivity function

• where the summation is over all right half plane

zeros.
• Note:
– Small right half plane zeros are worse than large ones
– Large right half plane poles are worse than small
ones.

187

Example: System that admits small sensitivities

• For a closed loop system consisting of a

first order process and a proportional
controller, the loop transfer function is

where parameter k is the controller gain, find

the sensitivity function and the lowest
value of achievable sensitivity.
188

94
 3/23/2023

Soln:
• The sensitivity function is

• and

• This implies that |S(jω)| < 1 for all finite

frequencies and that the sensitivity can be
made arbitrary small for any finite
frequency by making k sufficiently large.
189

Exercises
• For a closed loop system with the
following loop transfer functions, find the
sensitivity function and the lowest value of
achievable sensitivity.
– (i) (s-2)/(s(s-1)(s+3))
– (ii) 5/(s^3+2s^2-s-2)
– (iii) C(s)=2(1+0.5/s);P(s)=2/((s-1)(s-2))

190

95
 3/23/2023

Summary
• Steps involved in Robustness studies:
– Problem formulation
• Nominal performance design
• Robust stability analysis
– Time domain Experiments
• wider tracking range
• better disturbance rejection
– Frequency domain analysis
• Smaller Nominal performance norm
• Smaller Robust stability norm

191

96