De Silva Clarence W Sensors and Actuators Control Systems Instrumentation

Control engineering series 20
Design of
MODERN
CONTROL
SYSTEMS
Edited by
D J Bell PA Cook N Munro
l
l

Peter Peregrinus Ltd. on behfalf of the Institution of Electrical Engineers

IEE CONTROL ENGINEERING SERIES 20
Series Editors: Prof. D. P. Atherton

Dr. K. Warwick
CONTROL
SYSTEMS
Other volumes in this series
Volume 1 Multivariable control theory J. M. Layton
Volume 2 Elevator traffic analysis, design and control G. C. Barney and
S. M. Dos Santos
Volume 3 Transducers in digital systems G. A. Woolvet
Volume 4 Supervisory remote control systems R. E. Young
Volume 5 Structure of interconnected systems H. Nicholson
Volume 6 Power system control M. J. H. Sterling
Volume 7 Feedback and multivariable systems D. H. Owens
Volume 8 A history of control engineering, 1800-1930 S. Bennett
Volume 9 Modern approaches to control system design N. Munro (Editor)
Volume 10 Control of time delay systems J. E. Marshall
Volume 11 Biological systems, modelling and control D. A. Linkens
Volume 12 Modelling of dynamical systems—1 H. Nicholson (Editor)
Volume 13 Modelling of dynamical systems—2 H. Nicholson (Editor)
Volume 14 Optimal relay and saturating control system synthesis E. P. Ryan
Volume 15 Self-tuning and adaptive control: theory and application C. J. Harris
and S. A. Billings (Editors)
Volume 16 Systems modelling and optimisation P. Nash
Volume 17 Control in hazardous environments R. E. Young
Volume 18 Applied control theory J. R. Leigh
Volume 19 Stepping motors: a guide to modern theory and practice P. P. Acarnley
Volume 20 Design of modern control systems D. J. Bell, P. A. Cook
and N. Munro (Editors)
Volume 21 Computer control of industrial processes S. Bennett and
D. A. Linkens (Editors)
Volume 22 Digital signal processing N. B. Jones (Editor)
Volume 23 Robotic technology A. Pugh (Editor)
Volume 24 Real-time computer control S. Bennett and D. A. Linkens (Editors)
Volume 25 Nonlinear system design S. A. Billings, J. O. Gray and
D. H. Owens (Editors)
Volume 26 Measurement and instrumentation for control M. G. Mylroi and
G. Calvert (Editors)
Volume 27 Process dynamics estimation and control A. Johnson
Volume 28 Robots and automated manufacture J. Billingsley (Editor)
Volume 29 Industrial digital control systems K. Warwick and D. Rees (Editors)
Volume 30 Electromagnetic suspension—dynamics and control P. K. Sinha
Volume 31 Modelling and control of fermentation processes J. R. Leigh (Editor)
Volume 32 Multivariable control for industrial applications J. O'Reilly (Editor)
Volume 33 Temperature measurement and control J. R. Leigh
Volume 34 Singular perturbation methodology in control systems D. S. Naidu
Volume 35 Implementation of self-tuning controllers K. Warwick (Editor)
Volume 36 Robot control: theory and applications K. Warwick and A. Pugh
(Editors)
CONTROL
SYSTEMS
D J Bell • PA Cook*T\J Munro

Published by: Peter Peregrinus Ltd., London, United Kingdom
© 1982 Peter Peregrinus Ltd.
Reprinted 1988
All rights reserved. No part of this publication may be reproduced, stored in

a retrieval system or transmitted in any form or by any means—electronic,
mechanical, photocopying, recording or otherwise—without the prior written
permission of the publisher.
While the author and the publishers believe that the information and guidance
given in this work are correct, all parties must rely upon their own skill and judgment
when making use of them. Neither the author nor the publishers assume any liability
to anyone for any loss or damage caused by any error or omission in
the work, whether such error or omission is the result of negligence or any other
cause. Any and all such liability is disclaimed.
British Library Cataloguing in Publication Data.
Bell, D. J.
Design of modern control systems.
(IEE control engineering series; v. 20)
1. Control theory
I. Title II. Cook, P. A. III. Munro, N.
ISBN 0 906048 74 5
Printed in England by Antony Rowe Ltd.

Contents
Preface
1 State-space theory 1
1.1 Introduction 1
1.2 Controlability and observability 2
1.3 Standard forms
1.4 Calculation of transfer function
matrix 9
1.5 Realizations 11
1.6 Inverse Systems 14
References 15
Problems 17
2 Complex variable methods in feedback

Systems analysis and design 18
2.1 Introduction 18
2.2 Generalized Nyquist and root-locus
diagrams 19
2.2.1 Characteristic frequencies and
characteristic gains 23
2.2.2 Generalized Nyquist diagrams and
generalized Nyquist stability
criterion 24
2.2.3 Multivariable root loci 25
2.2.4 Conformal nature of mapping between
frequency and gain 25
2.3 Zeros 27
2.3.1 Restriction of operator in domain
and range 29
2.3.2 Spectral characterisation of Zeros 30
2.3.3 Finite zeros when D is non-zero 30
2.4 Bi-linear transformation of
frequency and gain variables 31
2.5 Geometric theory of root locus and
Nyquist diagrams 32
2.6 Angles of arrival at zeros and
angles of departure from poles 35
2.7 Properties of Nyquist and root locus
diagrams for optimal feedback systems 37
2.8 Design techniques 38
2.9 Conclusions 40
References 41
Robustness in variable control systems design 46
3.1 Introduction 46
3.2 Sensitivity of characteristic gain loci 48
3.2.1 Sensitivity indices 49
3.2.2. Analysis 49
3.3 Uncertainty in a feed back system 51
3.3.1 Normality 53
3.4 Relative stability matrices 53
3.5 Multivariable gain and phase margins 55
3.6 Conclusion 60
References 60
Problems 62
A design study using the characteristic locus
method 64
4.1 Introduction 64
4.2 Two-bed reactor model 67
4.3 Control system design for Bed II 69
4.4 Compensator design 71
4.5 Design of the 2-bed system 76
4.6 System performance and properties 77
References 82
The inverse Nyquist array design method 83
5.1 Introduction 83
5.2 The multi variable design problem 84
5.3 Stability 88
5.3.1 Diagonal dominance 91
5.3.2 Further stability theorems 92
5.3.3 Graphical criteria for stability 93
5.4 Design technique 95
5.5 Conclusions 100
References 101
Appendix 102
Problems 104
Analysis and design of a nuclear boiler control
scheme 106
6.1 Introduction
6.2 Drum boiler plant and its control requirements 107
6.3 Multivariable frequency response methods 109
6.4 Analysis and design approach adopted 111
6.4.1 Transfer function model 112
6.4.2 Analysis of existing control scheme 113
6.4.3 Alternative control scheme design 116
6.5 Performance assessment of improved scheme 120
6.5.1 Three drum version of improved scheme 121
6.6 System integrity 122
6.7 Conclusions 124
References 125
Optimal control 126

7.1 The calculus of variations: classical theory 12 7
7.2 The optimal control problem 129
7.3 Singular control problems 133
7.3.1 The generalised Legendre-Clebsch
condition: a transformation approach 134
7.3.2 Jacobson's necessary condition 136
7.4 Dynamic programming 137
7.5 The Hamilton-Jacobi approach 140
References 142
Problems 14 3
8 Control system design via mathematical programming 14 4
8.1 Introduction 144
8.2 Linear multivariable systems 146
8.3 Non-linear systems
8.4 Semi-infinite programming 152
8.4.1 Feasible-directions algorithms 153
8.4.2 Outer approximations algorithms 155
8.4.3 Non-differentiability 155
8.4.4 Cut-map algorithms 156
8.5 Conclusion 156
References 157
9 Optimisation in multivariable design 159
9.1.1 The allocation problem 159
9.1.2 The scaling problem 159
9.1.3 Precompensation design problem 160
9.2 Problem formulation 160
9.2.1 Decomposition 161
9.2.2 Choice of dominance measure 161
9.2.3 General problem 162
9.3 Allocation problem 16 3
9.4 Sealing problem 16 5
9.5 Compensation design 16 7
9.6 Design example 169
References 17 3
Problems 175
10 Pole assignment 177
10.2 State-feedback algorithms 179
10.2.1 Dyadic designs 179
10.2.2 Full-rank designs 183
10.3 Output-feedback algorithms 187
10.3.1 Dyadic designs 189
10.3.2 Full-rank designs 192
10.3.3 Example 194
10.4 Concluding remarks 196
References 19 7
Appendix 198
Problems 200
11 Nonlinear systems 202
11.1 Nonlinear behaviour 202
11.2 Fourier series 203
11.3 The describing function 206
11.4 Prediction of limit cycles 210
11.4.1 Frequency response of nonlinear
systems 213
11.5 Nonlinear state-space equations 213
11.6 Lyapunov's method 214
11.6.1 Domains of attraction 215
11.6.2 Construction of Lyapunov's
functions 216
11.7 Absolute stability criteria 218
References 220
Problems 222
12 Some DDC system design procedures
12.2 Choice of sampling frequency 225
12.3 Frequency domain compensation method 227
12.4 The compensation of Class 0
(Regulator) systems 2 30
12.5 Noisy input or output signals 232
12.6 Structural resonances and digital notch
networks 232
12.7 Coefficient quantisation in a discrete
controller 234
12.8 Arithmetic roundoff-noise in a discrete
controller 237
12.9 Multivate and substrate Controllers 2 38
12.10 Comparison of time domain synthesis and
frequency domain compensation techniques 240
References 244
13 Robust controller design 246
13.2 General problem to be considered 24 7
13.3 The servomechanism problem - structural
results 248
13.3.1 Problem description 248
13.3.2 Existence results 251
13.3.3 Robust servomechanism controller
structure 259
13.3.4 Some properties of the robust
controller 261
13.3.5 Various classes of stabilizing
compensators 261
13.4 Extensions of previous results
13.4.1 Control of unknown systems
(Multivariable timing regulators) 263
13.4.2 Multivariable error constants 264
13.5 Robust controller design 267
13.5.1 Design example 268
13.6 Conclusions 270
References 2 71
14 Control of distributed parameter systems 274
14.2 Controllability 278
14.3 Observability 281
14.4 Optimal control 281
14.5 Optimal estimation (Kalman filter) 2 82
References 284
Problems 285
15 Decentralised control 286
15.1 Introduction 2 86
15.2 Non-classical information patterns 286
15.2.1 Partially nested information
structures 2 88
15.3 Decentralised stabilisation and pole
placement 2 89
15.4 Design technique of decentralised control 294
15.4.1 The algorithm of Geromel and
Bernussou 294
15.4.2 The model following method 29 8
15.5 Concluding remarks 301
References 302
Solutions
Chapter 1 304
Chapter 3 306
Chapter 5 309
Chapter 7 313
Chapter 10 316
Chapter 11 324
Chapter 14 32 6
Index 331
Preface
Over the last seven years the Science and Engineering

Research Council (previously the Science Research Council)
have been instrumental in the organization and financing of
a series of vacation schools. These have been aimed prim-
arily at recipients of SFFC postgraduate awards but some
non-SFRC students, together with a few workers from industry
(in the UK and elsewhere) have also taken part. Two of
these vacation schools have been on the subject of analysis
and design of control systems and have been held at the
Control Systems Centre in the University of Manchester
Institute of Science and Technology (UMIST). The chapters
of this present book form the lectures to be given at the
third UMIST vacation school during the week 2 9 March
3 April, 1982. The title of the vacation school is the
same as that for this book.
It is inevitable and indeed desirable that a subject such

as engineering design of multivariable systems will develop
over a period of six years. This becomes apparent if one
compares the contents of the 1976 UMIST vacation school as
given in "Modern approaches to control system design"
(N. Munro, Fd., IFE Control Engineering Series 9, Peter
Peregrinus, 1979) with the contents of the volume before you.
Some of the 1976 topics no longer appear in this 1982 ver-
sion, not because they are now obsolete or research is no
longer taking place in those areas but because we have felt
it right and proper to make way for either new subjects
(such as large scale systems) or interesting new applications
(e.g. the nuclear boiler in Chapter 6 ) . Even in the case
where 1976 topics are retained we have arranged for some
different aspects to be presented in this new publication.
It will be noticed that no fundamental background material
on single-input single-output systems and linear algebra is
presented in the present volume, unlike in the reference
given above. This is because we feel that such material is
readily available from many sources and in particular the
quoted reference may be consulted.
The organization of a vacation school and at the same time

the preparation of a book to be published before the school
takes place is a daunting task. We certainly would not have
succeeded without the full cooperation of all the lecturers
involved in the school, the typist who prepared the camera-
ready copy and the staff of the publishers. Our grateful
thanks therefore go to all those who honoured deadlines and
delivered the drafts of the chapters which follow this pre-
face, to Vera Butterworth who typed the whole volume under
exceptionally trying circumstances, and to Pauline Maliphant
of Peter Peregrinus Ltd. who waited so patiently for the
chapters to arrive at Stevenage from Manchester. This book
would not have appeared but for the support given by the
Science and Engineering Research Council and our sincere
thanks go to staff of the SERC and to the other members of
the planning panel which organized the vacation school.
D.J. Bell
P.A. Cook
N. Munro
Manchester, 1982.
Chapter 1
State-space theory
Dr. P. A. COOK
Synopsis
A brief review is given of the use of state-space methods

in the study of linear time-invariant systems. Topics cov-
ered include modal decomposition, controllability and observ-
ability, standard forms, calculation of transfer function
matrices, construction of state-space realizations and system
inversion.
1.1 Introduction
We shall be concerned with linear stationary dynamical

systems described by equations of the general form
x = Ax + Bu (1.1)
y = Cx + Du (1.2)
where the vectors u (inputs), x (states) and y (outputs)

have respective dimensions m,n,p, and A,B,C,D are constant
matrices. Usually we shall have D = 0.
Solving equation (1.1) with initial conditions
x(0) = xQ (1.3)
and substituting into equation (1.2), we obtain
y(t) =
- * r.
C exp(At)x n + I C exp {A (t—r) }Bu (T )dx + Du(t)
(1.4)
which, after Laplace transformation, gives

y(s) = C(sl-A)~ 1 x o + G(s)u(s)
where
G(s) = C(sI-A)" 1 B + D (1.5)
is the transfer function matrix.
Clearly, G(s) is unaltered by coordinate transformation

in the state space, i.e.
A -> R" 1 AR, B -* R" 1 B, C -> CR, D -> D
where R is a nonsingular constant matrix. In particular,

if A has distinct eigenvalues (A,,..., A ) , with corres-
ponding eigenvectors (w,,...,w ) we can take
R = W = [w1,...,wn]
so that
w"'1AW = diag (A1,...,An) (1.6)
and G(s) has the modal decomposition

T
G(S) = +D (1#7)
J i^T
T
where A^ is the ith column of CW and 3. is the ith row
X
of W" B.
Although we are specifically considering continuous-time

systems, the same algebraic structure applies to the discrete
time case, where we have difference equations instead of
differential equations. Thus, if the system is described by
x(t+l) = Ax(t) + Bu(t) (1.8)
y(t) = Cx(t) + Du(t) (1.9)
and we define the z-transforms by
x(z) = I x(t)z" t , etc.,

t=0
then it follows that
y(z) = C(zI-A)~1zx(0) + G(z)u(z)
where G(z) is given by substitution of z for s in equa-

tion (1.5) . The solution of equations (1.8) and (1.9) gives
for t > 0,
y(t) = Du(t) + CBu(t-l) + ... + CAt""1Bu(0) + CA t x(0)
1.2 Controllability and Observability
A system is said to be controllable if, for any initial

state, it is possible to find an input which will transfer it
to the zero state in a finite time. Solving (1.1) subject
to (1.3). and setting
x(t) = 0
for some specified t f 0, we get
exp(-AT)Bu(i)dT
0
which can be solved for U ( T ) , for any given XQ, if and

only if the rows of the matrix
exp(-AT)B
are linearly independent. Since the matrix exponential can

be expanded in terms of (I,A,...,An~ ) by the Cayley-
Hamilton theorem, a necessary and sufficient condition for
controllability is
rank $ = n (1.10)
where
$c = [B / AB / ...,A n " 1 B] (1.11)
This is also a necessary and sufficient condition for any

4
state to be reachable from any other, and it plays the same

role in the discrete-time case, as may be seen by solving
(1.8) .
Observability means that the state of the system at a given

time can be ascertained from a knowledge of its input and out-
put over a subsequent finite time interval. From (1.4), XQ
can be determined, given the vector functions y(t) and u(t)
over an interval, if and only if the columns of the matrix
C exp (At)
are linearly independent. Hence, a necessary and sufficient

condition for the system to be observable is
rank $ = n (1.12)
o
where
$ = [C ,A C ,...,(A ) n ~ C ] (1.13)
with the superscript T denoting transposition.
Again, the same condition applies to a discrete-time system

governed by (1.8) and (1.9) .
The algebraic conditions (1.10) and (1.12) are clearly very

similar, in view of (1.11) and (1.13) . It can be seen that
neither involves the matrix D, which may be ignored for pres-
ent purposes, and that a system described by the matrices
(A,B,C) is controllable if and only if the "dual" system des-
cribed by (A ,C ,B ) is observable.
Both controllability and observability are evidently un-

affected by similarity transformations in the state space. In
particular, if the eigenvalues of A are all distinct, we can
perform the transformation (1.6), and it follows that the
rows of exp (At)B are linearly independent, i.e. the system
is controllable if and only if all the vectors 3^ in (1.7)
are non-vanishing. Similarly, it is observable if and only
if all the Yi / 0 in (1.7), and so a controllable and
observable system is one whose transfer function matrix con-
tains contributions from all its modes.
In general, a system can be decomposed into four parts,
characterised by their controllability and observability pro-
perties . By means of a suitable coordinate transformation,
the matrices (A,B,C) can be brought to the forms
X X X X
0 A 0 X B
CO c
0 0 B =
A
uu X 0
0 0 0 A J 0
uo
[0 x ]
where the crosses denote non-vanishing blocks and the sub-

scripts indicate controllability, observability, uncontroll-
ability or unobservability. Only the controllable and
observable part contributes to the transfer function matrix,
which is now expressed as
G(S) = C C (SI-A C O )" 1 B O
1.3 Standard Forms
Let us suppose that the system described by (1.1) is con-

trollable, and that the inputs enter independently of each
other into the state equations, i.e.
rank B = m (1.14)
Since the matrix $ in (1.11) has full rank, we can con-

struct a nonsingular square matrix by examining the columns
of $ in order and retaining only those which are linearly
independent of their predecessors, The resulting matrix can
then be rearranged in the form
H = b ] (1.15)
m
where B = [b,
l ,...,b m ] and the integers (y,i , . . .mf ii_.) are
known as the controllability indices of the system. Because
of (1.14), they are all strictly positive and, since H has
n columns,
Since they are determined from linear dependence relations,

the controllability indices are unaltered by state-space co-
ordinate transformations. Also, because of their relation
to the ranks of the matrices B,[B,AB],..., they are unaff-
ected, except for rearrangement among themselves, by non-
singular transformations on the input space, i.e. recombina-
tion of dinputs, and also by state-variable feedback . The
quantity
v = max y .
is thus invariant under all these transformations. It is the

smallest integer v such that
rank [B,AB,...,AV~1B] = n
and is sometimes called the "controllability index".
We now perform the coordinate transformation
A -* H~ 1 AH, B + H" 1 B
with H given by (1.15) .
The results have the following form: H~" B is a matrix

with unity in positions (1,1), (yx+l,2),...,(yx+...+y +l,m)
and zeros elsewhere, while
"•• A lm
H~ 1 AH
A
ml mm J
where the blocks have the structure

0 ... 0 a 0 ... 0 a. j 0
iiO
A =
3i
v \ 0
0. . .0 '1 a 0 ... 0 a .
l
and the entries in the final columns are the coefficients in

the relations
min(yi,y.-l) min(y.,y.
A b.±
I + I \
k=0 j=i k=0
(1.16)
It follows from (1.16) that the number of nonvanishing co-

efficients depends upon the controllability indices. The
maximum possible number is nm, and this can only be achieved
in the case where the first n columns of $ are linearly
independent, so that the indices become
= v = v
c ' y r+l c ~
where r is an integer satisfying
n = m(v -1) + r
This is sometimes referred to as the "generic" case, since

any system can be converted to it by an arbitrarily small
perturbation. Consequently, it is the case which can norm-
ally be expected to occur in practice, although it does not
always do so .
There is also another type of standard form, which can be

obtained as follows. Let the rows numbered y,fy, + y9,...,n,
-1 T T
in H be denoted by %i'•••'%m' respectively, and con-
struct the matrix
(1.17)
Q =
8
This matrix is nonsingular 2 , and we can use it to perform

another coordinate transformation.
We obtain the following structure:
T T —1
Q T A(Q T )
T
Q B
where the crosses denote nonzero elements. By further simil-

arity transformations, the number of nonvanishing entries can
be reduced to the same as that in the previous standard form.
The block sizes are also the same as in the previous form.
In view of the analogy between (1.11) and (1.13), we can

T T
obtain similar standard forms for A and C in the case
of an observable system. The block sizes are then given by a
set of integers (IT,,...,TT ) , obtained from the columns of
<J> and called observability indices. These are strictly
positive provided that the outputs in (1.2) are independent
functions of the states, i.e.
rank C = p
and they satisfy
The maximum among them,
max
is sometimes called the "observability index".
1.4 Calculation of the Transfer Function Matrix
We now consider the problem of computing G(s), as given

by (1.5), from the matrices (A,B,C,D). If a modal decompos-
ition is possible, the result can be obtained by summing the
terms in (1.7), but this is a roundabout method because of
the need to calculate the eigenvalues and eigenvectors of A.
A more direct method is to compute the resolvent matrix

(sI-A) and then multiply it by B and C. An algorithm
which is often used is variously attributed to Souriau, Frame
and Faddeev , although its essential features were discovered
by Leverrier . It calculates the coefficients in the expan-
sions
adj (sI-A) = R sn-1 + ... + R,s + R Q
and
det (sI-A) = sn + ^ ^ s 1 1 " 1 + ... + a x s + a Q
as follows: since
(sI-A) adj (sI-A) = I det (sI-A)
we have, by equating coefficients,

10
R Z
n-1
-1 =
D — ID 4- a T
R + a
n-2 " n-1 n-lx
! (1.18)
R = AR + a
0 1 lI
0 = AR Q + a Q I
whence, taking traces,
tr R o = tr
(1.19)
tr R Q = tr (ARX) + n a x
O = tr (ARQ) + n a Q
while, from the identity
~g det (sI-A) = tr adj (sI-A)
we get
(n-1) an—l, =, tr Rn—z
o
i
a tr R
l = 0
which, together with (1.19), yields
an—i-. ",— *" tr (AKn—xi /

!
, (1.20)
a = tr (AR
l "" n^T 1}
ao = - I tr (ARQ)
The sets of equations (1.18) and (1.20) provide an iterative

computational scheme, with the last equation of (1.18) serv-
ing as a check on the procedure. This is desirable, since
the method has been found to be numerically inaccurate in
some cases .
The number of multiplications required by this algorithm is

4
O(n ) . Several methods have been proposed which can be com-
Q
putationally more efficient than this . In general, they
11
involve the conversion of A, by a similarity transformation,

into a companion form
0 0 -a r 0. , ...
or
0 0 1 -a -a r -a_
from which det (sI-A) and adj (sI-A) may readily be ob-
tained. The transformation can be performed with O(n )
multiplications and divisions, but it is also subject to
numerical difficulties and there does not seem to be any
convenient way of checking the results.
1.5 Realizations
Next, we consider the converse problem of constructing a

state-space realization, i.e. a set of matrices (A,B,C) which
give rise, through (1.5), to a specified transfer function
matrix. We assume that G(s) is strictly proper, i.e.
G(s) -> 0 as s + °°, so that D = 0. There are many ways of
setting up a realization; one convenient systematic proced-
ure is the following.
First, express each column of G(s) as a polynomial vector

divided by a common denominator. Specifically, let the least
common denominator of the ith column be
n
i
d
s io
and the corresponding numerator vector be
n ± -l
3
Then, a realization can be immediately be written down in the

form
12
xo
B
l
A = , B = , c = [<:,
0 ^ A 0
where
0 1 0 0
X
I V
B
i =
-d.
iO
and d
i0 J
This realization is controllable, by construction, but not

necessarily observable. However, it can be reduced to an
observable form, without losing controllability, by perform-
ing a similarity transformation, constructed out of element-
ary row and column operations . An alternative realization
could, of course, be generated by starting with the rows,
instead of the columns, of G(s). It would then automatic-
ally be observable but might need to be reduced to controll-
able form.
In the discrete-time case, it may be convenient to start

from the "Markov parameters"
which appear in the solution of (1.8) and (1.9), and which

are also the coefficient matrices in the Laurent expansion
G(z) = D + -i
K.z
We define the matrices

13
K
2 ••' K
n .. . K
K K =
. . Ko
n 2n-l -1
and then, from (1.11) and (1.13),
_. T _ ~
(1.21)
K = 0 $ , K =
so that, if the matrices (A,B,C) give a controllable and

observable realization,
rank K = n
and hence two nonsingular matrices, M and N, can be found

such that
MKN = (1.22)
0
10
by using, for example, Gauss reduction . From (1.21) and
(1.22), it follows that, for some nonsingular matrix L,
N = x ]
where the crosses denote irrelevant blocks, and hence
LB x CL" 1 x
MK = KN =
and
LAL" 1 x
MKN =
from which the controllable and observable realization

(LAL~ ,LB,CL ) can immediately be extracted
14
1.6 Inverse Systems
From a state-space viewpoint, the problem of inverting a

rational function matrix G(s) is equivalent to finding a
realization of a transfer function matrix F(s) such that
F(s)G(s) = s" q I
for a left inverse, or
G(s)F(s) = s" g I
for a right inverse, where q is a non-negative integer,

assuming that F(s) and G(s) are both proper, i.e. bounded
as s •> ». It is useful to define the system matrix
sI-A B
P(s) =
-C D
since a left or right inverse exists if and only if the nor-

12
mal rank of P(s) is (m+n) or (p+n), respectively . We
are assuming that the realization (A,B,C) is controllable
and observable since we are concerned only with properties
which affect G(s). The values of s for which P(s) has
less than full rank are sometimes known as the transmission
zeros 1 of G(s).
If an inverse exists, it may be computed in various ways.

12
One method is to perform repeated manipulations on the
system until the D matrix becomes of full rank, when the
inverse can be constructed directly. If D is square and non-
singular, in fact, a realization of the unique inverse G~ (s)
can immediately be obtained from (1.1) and (1.2) in the form
x = (A - BD~ 1 C)x + y (1,.23)
1
u = -D~ Cx + D ^ y (1..24)
with u regarded as output and y as input. For a square

15
system, i.e. p = m, the inverse is necessarily unique and

two-sided if it exists at all, but (1.23) and (1.24) cannot be
be used to construct it unless D is nonsingular, whereas nor-
mally, on the contrary, D is null.
However, the inverse of a square system can also be

expressed as
l
= [0 l m ]P (s) (1.25)
14
and, by suitable elementary transformations , P(s) can be
reduced to the form
where J is a Jordan matrix with all its eigenvalues equal

to zero. The inverse matrix (1.25) then consists of two
parts, one of which depends on (sI-A) and is strictly
proper, while the other contains (I + sJ) -1 and is a poly-
nomial matrix.
References
KALMAN, R.E. : 'Canonical structure of linear dynamical

systems', Proc. Nat. Acad. Sci.,48, (1962), 596-600.
LUENBERGER, D.G. : 'Canonical forms for linear multi-
variable systems', IEEE Trans. Autom. Control, AC-12,
(1967) , 290-293.
BRUNOVSKY, P. 'A classification of linear controll-
able systems', Kybernetica, 3, (1970), 173-187.
POPOV, V.M. : 'Invariant description of linear time-
invariant systems', SIAM J. Control, 10, (1972) , 252-264.
COOK, P.A. : 'On some questions concerning controll-
ability and observability indices', Proc. 7th IFAC
Triennial World Congress (Helsinki, 1978), pp.1699-1705.
FADDEEV, D.K. and V.N. FADDEEVA : Computational
Methods of Linear Algebra , (Freeman 1963).
BARNETT, S. : Introduction to Mathematical Control
Theory , (Oxford University Press 1975).
BOSLEY, M.J. H.W. KROPHOLLER, F.P. LEES and R.M. NEALE,
'The determination of transfer functions from state-
16
variable models 1 , Automatica, 8, (1972), 213-218.

9. DALY, K.C., : 'A computational procedure for transfer-
function evaluation 1 , Int. J. Control, 20, (1974),
569-576.
10. ROSENBROCK, H.H. : State-Space and Multivariable
Theory , (Nelson-Wiley 1970).
11. HO, B.L. and R.E. KALMAN, : 'Effective construction
of linear state-variable models from input-output
functions 1 , Regelungstechnik, 14, (1966), 545-548.
12. MOYLAN, P.J. : 'Stable inversion of linear systems',
IEEE Trans. Autom. Control, AC-22, (1977), 74-78.
13. DAVISON, E.J. and S.K. WANG : 'Properties and cal-
culation of transmission zeros of linear multivariable
systems', Automatica, 10, (1974), 643-658.
14. VAN DER WEIDEN, A.J.J. : 'Inversions of rational
matrices', Int. J. Control, 25, (1977), 393-402.
17
Problems
P.I.I A system is described by the matrices
2 1 1 0 1 0
1 0 0 -1 0 1
A = B =
1 0 -1 -1 0 1
0 1 1 -1 1 0
0 -1 -1 0
C =
1 0 0 1
Convert the system to standard form with the help of the

matrix Q given by equation (1.17) .
P.1.2 For the above system obtain the resolvent matrix

(sI-A)" 1 from the recurrence relations (1.18) and (1.20),
and hence calculate G(s) .
Chapter 2
Complex variable methods in feedback

systems analysis and design
Professor A. G. J. MACFARLANE
Synopsis
Generalized Nyquist and generalized Root-Locus diagrams are

algebraic curves derived from the spectral analysis of appro-
priate matrix-valued functions of a complex variable. Their
key properties can be comprehensively analysed in terms of a
state-space model of the feedback loop which is being studied
Such an analysis is essentially geometrical in nature and is
concerned with the way in which certain subspaces, defined
via the various operators involved, sit with respect to one
another, and with ways of assigning complex frequencies to
these subspaces. The current status of work on the' design of
feedback systems using these results is briefly discussed,
and possible future developments are noted.
2.1 Introduction
The original Nyquist stability criterion owed its central

importance in classical feedback control work to the fact
that it tested the stability of a proposed feedback loop in
terms of a directly measured transmission characteristic.
This radical departure of inferring stability from measured
signal characteristics instead of from computed dynamical-
model characteristics placed complex-function methods in a
centrally-important position in the development of control
analysis and design techniques up to the late 194 0s. The
root-locus technique developed in the 1950s added further
weight to the arguments for using complex-variable approaches
in the analysis and design of feedback systems. At the time
19
of the introduction of the root-locus method, the close link

with the Nyquist technique was not appreciated; one of the
virtues of the algebraic function approach used here to gen-
eralize these techniques to the multivariable case is that
it shows how both these classical techniques are essentially
different ways of representing curves derived from a single
algebraic function relating complex frequency and complex
gain.
Following the widespread introduction of state-space meth-

ods in the 1960s, interest in complex-variable approaches to
feedback systems analysis and design declined until the advent
of computer-driven interactive graphic displays reawakened
interest in "picture-generating" techniques. The work pre-
sented here arose out of a desire to do two things:
(i) to generalize both the Nyquist-Bode and Root-Locus

techniques to the multivariable case; and
(ii) to relate, these generalizations to state-space descrip-
tions.
As such it is directed towards a fusing of state-space and

classical frequency-response ideas, and to an exploration of
the way in which the emergence of computer graphics can be
best exploited in the further development of multivariable
feedback system design techniques.
2.2 Generalized Nyquist and Root-Locus Diagrams
The dual role of complex numbers is familiar from elemen-

tary complex number theory; they can be used to represent
both objects (e.g. vectors in a plane) and operations on ob-
jects (e.g. rotation and stretching of vectors in a plane).
This dual role is important in classical single-input single
output frequency-response methods, where complex numbers are
used to represent frequencies (complex numbers) and gains
(operations on complex vectors). The two classical approaches
to the complex-variable-based analysis of feedback systems:
(i) study open-loop gain as a function of frequency (the

20
Nyquist-Bode approach); or
(ii) study closed-loop frequency as a function of gain (the
Evans Root-Locus approach).
Our first objective is to show that there is an intimate

relationship between the generalization of these two approa-
ches to the multivariable (many-input many-output) case and
the standard state-space model
£ : x = Ax + Bu
(2.1)
y = Cx + Du
In most of what follows the matrix D will be taken to be

zero for simplicity of exposition. However, its inclusion
at this point gives completeness and symmetry to the devel-
opment of the general case. [A is nxn; B is nx£; C is mxn;
D is mx£. When vector feedback loops are being discussed
l = m]. Suppose we sever m feedback connections to an
arbitrary linear dynamical system as shown in Fig. 2.1, and
that the transmittance between the injection point a and
the return a' is given by a state-space model of the form
shown in equation (2.1). Then on remaking the feedback con-
nections, we see that the closed-loop characteristic frequen-
cies of this system will be given by the spectrum (set of
eigenvalues) of the matrix
S = A + B(I m - D)~ 1 C (2.2)
and that the open-loop transfer function matrix relating vec-

tor exponential signals injected at a to the vector expon-
ential signals returned at a' is
G(s) = D + C(sl n -A)" 1 B (2.3)
(Im and In are unit matrices of orders m and n respec-

tively.) If one compares the right-hand sides of (2.2) and
(2.3) they are seen to have a suggestive structural similarity.
21
linear m inputs
dynamical and
system m outputs
signal
7TFT
signal
vector vector
injection return
Fig 2.1
This becomes an exact structural equivalence if we introduce

a variable g and change (2.2) to the form
S(g) = A + B(gI m -D)~ 1 C (2.4)
where the role of the variable g will emerge shortly. Our

aim is to put, in a state-space-model context, the roles of
open-loop gain (the key Nyquist-Bode concept) and closed-loop
frequency (the key Evans Root-Locus concept) on an exactly
equal footing as complex variables. With this in mind we
look for a system giving the closed-loop frequency matrix of
equation (2.4) and find that it is as shown in Fig.. 2.2,
where g is now interpreted as a gain parameter. On redraw-
ing Fig. 2.2 in the form shown in Fig. 2.3 we now see that
the variables s and g have indeed been placed in a strik-
ingly symmetrical relationship.
In Fig. 2.3 s is to be interpreted as a complex frequency

variable and g as a complex gain variable. For the related
vector feedback loop (with m inputs and m outputs) the open-
loop matrix
G(s) = C(sI n -A)" 1 B
describes open-loop gain as a function of imposed frequency

22
Fig 2.2
Fig 2.3
s; and the closed-loop frequency matrix
S(g) = A
describes the closed-loop frequency as a function of the

imposed gain parameter g.
i—
G(5)
Fig. 2.4
I
g
For the negative feedback arrangement of Fig. 2.4, if
F(s) = I m + g""1G(s)
is the return difference matrix for the set of feedback loops

broken at a~a' then one can show that
det[sIn-S(g)] det[gIm-G(s)] __ detF(s) _ CLCP(s)

det[sIn-S(«>)] det[gIm-G(-)] detF(co) OLCP(s)
(2.5)
where CLCP(s) and OLCP(s) are the system's closed-loop and
open-loop characteristic polynomials respectively. The
importance of equation (2.5) is that it tells us that in this
archetypal negative feedback situation we may equally well
23
study the effect of closing feedback loops on stability in

terms of an open-loop gain description or a closed-loop fre-
quency description. Specifically it shows that, for values
of s not in the spectrum of A and values of g not in
the spectrum of D
det[sIn-S(g)] = 0 <=> det[gIm-G(s)] = 0
This in effect says that a knowledge of how the characteris-

tic values of G(s) vary as a function of the frequency par-
ameter s is equivalent (for the purposes of making infer-
ences about closed-loop stability) to a knowledge of how the
characteristic values of S(g) vary as a function of the
gain parameter g.
2.2.1 Characteristic frequencies and characteristic gains
For a given value of the gain parameter g the eigenvalues

{s. : i = l,2,...,n} of S(g) are the corresponding set of
closed-loop characteristic frequencies, obtained from
det[sl n - S(g)] = 0
For a given value of the frequency parameter s the eigen-

values {g i : i = l,2,...,m} of G(s) may be called the
corresponding set of open-loop characteristic gains, obtained
from
det[gl m - G(s)] = 0
The closed-loop characteristic frequencies {s^} for a given

value of g are, as is well known, associated with invariant
subspaces in the state space. The open-loop characteristic
gains {g.} for a given value of s are associated with
invariant subspaces in an input-output space.
From the characteristic equations for G(s) and S(g),

namely
A(g,s) =A det[gl m - G(s)] =? 0

24
and V(s,g) 4 det[sl n - S (g)] = 0
one obtains a pair of algebraic equations relating the com-

plex variables s and g. These define a pair of algebraic
functions :
(i) a characteristic gain function g(s) which gives open-

loop characteristic gain as a function of frequency;
and
(ii) a characteristic frequency function s(g) which gives
closed-loop characteristic frequency as a function of
gain.
The importance of these two algebraic functions lies in the

fact that they are the natural means of generalizing the con-
cepts of Nyquist diagram and Root-Locus diagram to the multi-
variable case .
2.2.2 Generalized Nyquist Diagrams and the

Generalized Nyquist Stability Criterion
The characteristic gain loci (generalized Nyquist diagrams)

for the m-vector feedback loop with transmittance matrix G(s)
are the loci in the complex gain plane which are traced out
by the eigenvalues of G(s) as s transverses the so-called
Nyquist D-contour in the complex frequency plane (s-plane).
They can also be defined as the +90° constant-phase contours
on the g-plane of the algebraic function s(g). Their util-
ity lies in their role in the following sort of generaliza-
1 2
tion of the Nyquist Stability Criterion ' .
Multivariable Nyquist Stability Criterion; The multivariable

Nyquist stability criterion relates the closed-loop stability
of the configuration of Fig. 2.5 to the characteristic gain
loci for the loop-gain matrix G(s)H(s). Suppose the system
of Fig. 2.5 has no uncontrollable and/or unobservable modes
whose corresponding characteristic frequencies lie in the
right-half s-plane. Then this feedback configuration will be
closed-loop stable if and only if the net sum of anticlockwise
25
G(s) —
Fig 2.5 "j
H(s) —
encirclements of the critical point (-1+jO) by the set of

characteristic gain loci of G(s)H(s) is equal to the total
number of right-half-plane poles of G(s) and H(s).
Proofs of multivariable versions of the Nyquist Stability

Criterion have been given by Barman and Katzenelson and
MacFarlane and Postlethwaite , and Postlethwaite and MacFar-
lane following earlier heuristic approaches by Bohn and
4 5
Kasavand and MacFarlane . More general versions have been
considered by Harris and Valenca and by Desoer and Wang
2.2.3 Multivariable Root Loci
The loci of open-loop characteristic gain with frequency

are generalizations of classical Nyquist diagrams in the
complex gain plane; there is a corresponding generalization
of the classical Evans 1 Root Locus diagram in the frequency
plane. These are the characteristic frequency loci which
are traced out by the eigenvalues of S(g) as g traverses
the negative real axis in the gain plane. They can also be
regarded as the 180° phase contours of g(s) on the frequ-
ency plane. The theory behind the multivariable root locus
has been given by Postlethwaite and MacFarlane and their
use discussed by Retallack , Kouvaritakis and Shaked ,
Q
MacFarlane, Kouvaritakis and Edmunds , Kouvaritakis and

Edmunds , Kouvaritakis and Owens . Many examples of gen-
eralized Nyquist and Root-Locus diagrams are given in ref.12;
a complete review of the development of frequency-response
methods is given in ref. 13.
2.2.4 Conformal Nature of Mapping Between

Frequency and Gain
The algebraic functions g(s), giving characteristic gain

26
as a function of frequency, and s(g), giving characteristic

frequency as a function of gain contain essentially the same
information. One is simply a "re-packaged" version of the
other. One would thus expect there to be an exact structural
relationship between the (generalized) Nyquist and (general-
ized) Root Locus diagrams for a given feedback loop. These
structural relationships also reflect the fact that the map-
pings between s-plane and g-plane and vice-versa are confor-
mal, so that appropriate angular relationships are preserved.
In correlating the main features of the two forms of diagram
the following rule has been found useful.
Locus-Crossing Rule: Let D be a path in the s-plane and

1
let D be its image under g(s) in the g-plane. Also let
E be a path in the g-plane and let E1 be its image under
s(g) in the s-plane. Then the following relationships hold.
(i) Each crossing of E by D* corresponds to a related

crossing of D by E*, and vice-versa.
(ii) If E* crossed D from left to right with respect
to a given orientation for D, then E will cross D*
from left to right with respect to the induced orien-
tation for D*, and vice-versa. (The given orienta-
tions for D and E are taken to induce appropriate
orientations in D1 and E* : that is a traversal of
D by s in a positive orientation, makes g(s)
i
traverse D in a positive orientation, with a sim-

ilar convention for E and E*.)
(iii) The angle at which E1 crosses D is the same as the
angle at which D* crosses E.
This rule is illustrated in Fig. 2.6 taking D as the pos-

itive imaginary axis in the s-plane and E as the negative
real axis in the g-plane. This shows that a crossing of the
negative real axis in the gain plane (at g1 say) by a por-
tion of the generalized Nyquist diagram corresponds to the
migration of a closed-loop characteristic frequency from left
to right (under negative feedback of amount (g1)" applied
equally in all loops).
27
E'. portion of
s-plane Voot locus
D traced from uorO to uu =
g-plane
D1
D1, portion of
nyquist locus
E, traced from
g = oo to g =
F i g . 2.6
2.3 Zeros
In physical terms zeros are associated with the vanishing
of vector gain; that is with the existence of non-zero in-
put exponential signal vectors which result in zero output .
The discussion of zeros in the literature is confusing be-
cause one can talk about zeros in connection with various
objects: £(A,B,C,D), or G(s), or g(s) for example.
The object £(A,B,C,D) has associated with it a larger set
of zeros than the object G(s) because in passing from the
representation £(A,B,C,D) to the representation G(s) one
discards phenomena (uncontrollability and unobservability)
which can be discussed in terms of zeros (the so-called de-
coupling zeros 1 ) . In turn the object G(s) has a larger
set of zeros than the object g(s) because vanishing of all
characteristic gains does not necessarily imply the vanish-
ing of the vector gain for G(s). For the purposes of this
paper we will take the system representations being discussed
to be completely controllable and completely observable, and
the zeros to be what are usually called the transmission
zeros . A useful physical interpretation of a zero then
comes from the following result.
Transmission-blocking theorem: For a system I(A,B,C,D)

having a number of inputs less than or equal to the number
28
of outputs, necessary and sufficient conditions for an input

of the form
u(t) = 1 (t)e Zt u z
to result in a state-space trajectory
x(t) = 1 (t)e zt x 2 for t > 0
and an output
y(t) = 0 for t > 0
are that a complex frequency z and a complex vector

exist such that
zl - A -B
= 0 (2.6)
- C -D
(Here 1 (t) = 0 for t <_ 0 and 1 (t) = 1 for t > 0.)

The corresponding vectors u and x are called the zero
21
directions in the input and state spaces respectively
Zero Pencil; From (2.6) we have that, for the case when
D = 0,
= Bu (2.7)
Cx z = 0
Zero directions in the state space naturally lie in the ker-

nel of C so that
= Mv z (2.8)
where M is a basis matrix representation of Ker C and v_

z
is an appropriate vector of constants. Substitution of (2.8)
into (2.7) and premultiplication by the full-rank transform-
29
ation matrix fN
B+
where N is a full-rank left annihilator of B (i.e.

NB = 0) and B+ is a left inverse of B (i.e. B+B =
yields
(z NM-NAM)v = 0
uz = B + (zI-A)Mv z
The object (sNM-NAM) is called the zero pencil23'2* for the

system £(A,B,C). The zeros are the roots of the invariant
factors of the zero pencil, and this fact can be used to ex-
hibit the geometric characterisation of zeros in a very
succinct way. To do this it is useful to use Sastry and
Desoer's notion of restricting an operator in domain and
range
2.3.1 Restriction of an operator in domain and range
Given a linear map A from <Zn to <Zn where <Zn is the

space of n-dimensional complex vectors, and two subspaces
S , S2 °f ^nr then the restriction of A to S^ in the
n
domain and £2 ^ the range is defined to be the linear map
which associates with x.^ e S^ the orthogonal projection
of Ax, on to #2• This restricted linear map (and also
its matrix representation) will be denoted by
If the columns of S1 and S2 respectively form orthonormal

bases for £, and S2 then the matrix representation of the
restricted operator is given by
v *
A = S 2 A S1
where * denotes complex conjugate transposition. With a

mild abuse of language one then calls the roots of
30
det[S*S 1 - S*AS1] = 0
the spectrum of the restricted operator A
2.3.2 Spectral characterisation of zeros
Suppose we choose N and M so that their columns are

orthonormal bases for Ker B and Ker C respectively.
Then
NM = I
Ker O K e r B
and
NAM = A
Ker C -* Ker B
and one may say that: the finite zeros of £(A,B,C) are
the spectrum of the restriction of A to Ker C in domain
and Ker B in range.
2.3.3 Finite zeros when is non-zero

Suppose the direct-coupling operator D for £(A,B,C,D)
having m inputs and m outputs has nullity dQ . Then
J(A,B,C,D) has at most (n-d ) finite zeros and at least
dQ root-locus asymtotes going to infinity (to infinite
zeros) . Let D have a characteristic decomposition
D = Mo]
rp rp rp
where [U M ] and [V N ] are the eigenvector and dual
o o
eigenvector matrices for D and J is the Jordan block
associated with the non-zero eigenvalues of D. Further,
suppose that there is a complete set of eigenvectors spanning
the null space of D. Define A ,B ,C by
J
^ o o' o
= A -
31
BQ = BM
C N
o = oC
Then the finite zeros and root locus asymptotes of £(A,B,C,D)

are those of I(AQ, B Q , CQ).
In case D has full rank one has that there are n finite
zeros of £(A,B,C,D) which are given by a(A-BD" 1 C), the
spectrum of the matrix (A-BD~ C ) .
2.4 Bilinear Transformation of Frequency and Gain Variables
If £(A,B,C,D) having m inputs and m outputs has a D

which is non-singular then, as noted above, it will have n
finite zeros given by a(A-BD~ C ) .
Given a bilinear transformation
S = — —E-
cp - a
28
on the complex frequency it has been shown by Edmunds
that it is possible to find a correspondingly transformed
system £(A,B,C,D) such that
G(p) = G(s)
Thus, if the original system has a zero at z the trans-

formed system will have a zero at a location given by sub-
stituting z in (2.9). One can therefore choose an
appropriate bilinear transformation to get a D of full rank,
calculate the zeros as a(A-BD C) and transform back to
find the complete set of zeros (finite and infinite) of the
original system. The bilinear mapping is a conformal map
between the s-plane and the p-plane for all points except
p = a/c which gets sent to infinity on the s-plane and
s = -d/c which gets sent to infinity on the p-plane. It
is assumed that (ab - be) ? 0 and that the original system
has no pole at -d/c. The set of transformed system mat-
32
rices is given by:
l
A = (cA + dl) (aA + bl)
B = (cA + dI)""1B
C = (ad - bc)C(cA + d l ) " 1
D = D - cC(cA + dl)"1B
A similar bilinear transformation can be carried out on

the complex gain variable g
a g + b dq + b
Q = 2 Q = — —
^ cg + d ^ c q - a
The correspondingly transformed system J(A',B',C ,D ) is

such that
S'(q) = S(g)
and has state-space parameters given by
A1 = A - cB(cD + dI)~1C
B* = (ab - bc)B(cD + d l ) " 1
C1 = ( c D + dI)"1C
D1 = (cD + dl)""1 (aD + bl)
2.5 Geometric Theory of Root Locus and Nyquist Diagrams
It has been noted that the finite zeros of a vector feed-

back loop transmittance can be characterized in terms of the
spectrum of the restriction of its A-matrix in domain and
range. The idea of using the spectrum of a restricted
operator in connection with the Markov Parameters of a loop
transfer function matrix leads to a geometric treatment of
the high-gain asymptotic behaviour of generalized Root-Locus
diagrams and the high-frequency asymptotic behaviour of gen-
eralized Nyquist diagrams.
Let the transfer function matrix for a strictly proper

33
(i.e. D = 0) feedback loop be expanded in a Taylor series

about s = oo as
G G9
G(s) = -^ + -=- +
S c,
where (G^ : i = 1,2,....) are the Markov Parameters
G = CB G = CAk
l k ~lB k
= 2'3'
The use of Markov Parameters to relate the main features

of generalized Root-Locus diagrams to state-space model para-
meters has been discussed by a number of investigators ' .
The approach and notation used here follows Sastry and
Desoer , who give a good discussion of the assumptions in-
volved in a detailed analysis of this sort.
Let a sequence of restricted operators be defined as foll-

ows :
v
G = G v v
k kK *
Ker G -Ker (G )
V
and take G. = G, .
v
Let d± be the nullity (rank defect) of G±. Then the
high-gain asymptotic behaviour of the various branches of
the generalized root-locus diagram can be determined as foll-
ows.
1st Order branches:
•ifl - " ^ 1- 1,2,...(m-d.)
V
where
± 1 z a[G ]\{0} and g -> 0
v
The collection of symbols co[G.]\{0}
is to be read as
1
v
"belongs
2nd Order to the non-zeros 22 spectrum
branches: of^-tl G^±".
s o = = 1,2 , . . . (d 9 -d, )
v i,2 g z i
where A. 0 E Q [ G ] \ { 0 } and g -• 0
i/2 2
34
and for the kth order branches
s
i,k
V
where A. , ea[G,]\{0} and g + 0
1 ,K K
On invoking the correspondence between the high-frequency

asymptotes for the generalized Nyquist diagrams and the high-
gain asymptotes for the generalized Root-Locus diagrams one
then gets the following description of the high-frequency
branches of the generalized Nyquist diagrams.
1st Order branches: g. , - - —r-^—

1,1 Jb)
2nd Order branches: g. = - -—=-
A .
i
kth Order branches: g. , - -
1,K
v
where A. , ea[G,]\{0} as before, and
1 / .K. K.
For example, if the first Markov Parameter CB has full

rank m then it can be shown that the state space is the dir-
ect sum of the image of the input map B and the kernel of
the output map C :
= Im B © Ker C
and that the zero pencil then becomes
(Sl n _ m - NAK)
There are then (n-m) finite zeros, given by a(NAM), and

there are m first-order asymptotic branches of the general-
ized Root-Locus and Nyquist diagrams given by:
V
s « -c.. (CB)/g g •> 0

i = 1,2,...,m
35
and
ai(CB)
g i /.I " JOJ
where (ai(CB) : i = l,2,...,m} are the eigenvalues of the

first Markov parameter CB.
The numbers of the various types of asymptotic behaviour

are summarized in the following table (where v is the first
integer for which the nullity d^ vanishes):
Order Number of Number of Root

Nyquist Asymptotes Locus Asymptotes
1 m - d m - d.
i
2 d d 2(d1 - d 2 )
i - 2
3 d d 3(d2 - d 3 )
2 " 3
V d vd
v-l v-l
Total v-1
Number of m m + I d.
Asymptotes i=l
Adding up the total number of root-locus asymptotes and

subtracting the result from the dynamical order of the feed-
back loop we find that
v-1
Number of finite zeros = n - m - £ d.
A careful study of the implications of the above relation-

ships gives a vivid geometrical insight into the role played
by the basic state-space model operators A, B and C in
generating the main structural features of the asymptotic
behaviour of generalized Root-Locus and Nyquist diagrams.
2.6 Angles of Arrival at Zeros and Angles of Departure

From Poles
The ways in which branches of the root locus depart from
36
poles and arrive at zeros, together with the relationships of

other important structural features of these diagrams to the
basic state-space model parameters has been considered by a
7 9 33 34 section we follow
number of investigators ' ' '
34 For the output-
the approach of Thompson, Stein and Laub
feedback system
x Ax + Bu
y Cx
u = - - Ky
the closed-loop frequency matrix will be
S K (g) = A - | BKC
Let s., x. and y. be closed-loop characteristic frequen-

cies with their associated right and left eigenvectors so
that for 0 < g < «>
[SR(g) - s i l]x i = 0
i = 1,2,...,n
y*[SK(g) - s.,1] = 0
then s., x., and y. can be obtained by solving the gener-

alized eigenvalue problems specified by
A-s ± I x.
-1
-C -gK w.
i = 1,2,
A-s ± I
= 0
-C -gK -1
where role of integer p is discussed below. Then, in terms

of these quantities it has been shown by Thompson et al
that:
The angles of the root locus for 0 _<_ g <_ » and for dis-
tinct frequencies si are given by
37
-y.BKC x.
arg(ds ± ) = arg - - 0 < g <_ °° (2.10)
i = 1,2, ... ,p
or * -l
n.K w
arg(ds ± ) = arg — ~ 0 £ g < «> (2.11)
y x
i i
i = 1,2,. ..,p
The angles of departure from poles are found using (2.10)
with g = «, and the angles of arrival at zeros are found
using (2.11) with g = 0. Note that for g > 0 we will
have p = n and for g = 0 we will have 0 <_ p <_ n - m,
where p is the number of finite zeros and m is the number
of inputs and outputs of the vector feedback loop.
2.7 Properties of Nyquist and Root Locus Diagrams

For Optimal Feedback Systems
Several investigators have studied the asymptotic Nyquist

diagram behaviour (for high frequencies), the asymptotic
Root-Locus diagram behaviour (as explained below) and the
location of the finite zeros for the standard optimal state
feedback control problem involved in minimizing the cost
function
J = [ (xTQx + pu T Ru)dt (2.12)

J
0
where R is positive definite symmetric, Q is positive

semi-definite symmetric and p a finite positive real con-
stant 10,34,35,36,37,38,51_
It is known as a result of these studies that:
(i) All the finite zeros of the optimal feedback loop lie
in the left-half complex plane,
(ii) All the generalized Nyquist diagrams for the optimal
feedback loop have infinite gain margin at least 60°
phase margin,
(iii) When p in the performance index (2.12) is °° the n
branches of the optimal root locus (i.e. the locus of
38
closed-loop characteristic frequencies traced out with varia-

tion of p) start on a set of poles which are the stable
poles of the original plant together with the mirror images
of the imaginary axis of the unstable poles of the original
plant (assumed to have no purely imaginary poles). As p
varies all the branches of the optimal root locus remain in
the left-half of the frequency plane. For p tending to
zero a number p of the branches (where 0 £ p _< n-m) stay
finite and approach a set of loop transmission zeros. The
remaining (m-p) branches approach infinity in a set of so-
called Butterworth patterns. A kth order Butterworth pat-
tern has k asymptotes each of which radiates from the origin
through the left-half plane solutions of
2.8 Design Techniques
Despite the great efforts which have been expended on the

problems involved over the past decade, much remains to be
done in formulating a definitive design technique for linear
multivariable feedback systems of high (say having up to 100
state variables) dynamic order. A fairly comprehensive
attempt to extend the philosophy and techniques of the
classical Nyquist-Bode-Evans design approach to the multi-
variable case is given in ref. 12, and this line of attack
can be compared with a variety of other approaches to a
common design problem in ref. 39. A fairly complete review
of the development of the classical frequency-response meth-
ods can be found in ref. 13, together with a discussion of
various forms of extension to the multivariable case. A com-
puter-aided interactive-design package has been developed
for use in the implementation of the techniques described in
ref. 12. Experience of its use has shown that an interac-
tive design method of this sort can be useful to industrial
designers. There is thus a considerable incentive to develop
further those forms of design technique which are based on a
combination of the attractive graphical features of complex-
variable methods and the geometrical way in which the main
39
structural features of the algebraic curves used are related

to state-space model parameters.
Two ways in which extensions of classical frequency-

response approaches to the multivariable case are being
developed currently can be summarised as follows.
(i) The characteristic gain (Nyquist) loci and/or the char-

acteristic frequency (Root-Locus) loci and their associated
characteristic vector structures are manipulated using a set
of ad-hoc rules (guided of course by engineering insight and
using some computer-aided synthesis methods) until a set of
engineering design criteria for relative stability and closed
loop performance are satisfied. This approach has been des-
cribed in ref. 12 and has been successfully deployed on a
12 39
number of problems ' . Although it requires engineering
insight and manipulative skill, it has been shown to be a
powerful tool in the hands of a skilled designer. A fairly
extensive computer-aided design package has been developed
for use in this way 4 U ' .
42
(ii) An alternative approach has been developed by Edmunds
which transfers the burden of the designer's choice to the
selection of a closed-loop performance specification and a
controller structure. This approach uses Nyquist and Bode
diagram arrays of the open-loop and closed-loop behaviour as
a means of assessing performance and interaction. Following
a preliminary analysis of the problems (using the techniques
of (i) above) the designer chooses a suitable specification
for the closed-loop behaviour of the feedback system together
with a stipulated structure for the feedback controller to
be used. The controller structure selected has a number of
variable parameters and the design is completed by a computer
optimization (using say a standard weighted least squares
algorithm) of the difference between the specified and actual
closed-loop performance. As one would expect, this method,
when properly used, produces better results than (i) above.
It has the further advantage of being extendable to deal with
variations in the system model parameters. The CAD package
mentioned above ' incorporates the routines for dealing
40
.with this approach. However, much further work remains to be

done for a variety of reasons, including the following ones.
(i) Generalized Nyquist and generalized Root-Locus diagrams

give only a partial picture of feedback system behaviour -
they are primarily concerned with the relative stability of
the closed-loop system.
(ii) Although the picture of the closed-loop system behav-

iour which is given by the use of characteristic gains can
be extended by the use of an appropriate set of characteris-
tic directions (eigenvectors of transfer-function matrices
like G(s) for example), this still does not give an ade-
quate description of system closed-loop performance because,
as already noted, characteristic values (eigenvalues) do not
fully describe the gain behaviour of an operator.
It seems that for multivariable feedback systems it is nec-

essary to work simultaneously with two forms of operator de-
composition to get an adequate description for design pur-
poses - a characteristic (eigenvalue) decomposition for
stability assessment and a principal (singular) value decom-
position for gain assessment.
(iii) The design specification used must be extended to

include the effects of plant uncertainty, noisy disturbances
and noisy sensors.
(iv) The precise roles played by feedback, feedforward and

prefiltering must be defined in the context of a prescriptive
design procedure, and these roles must be coherently related
to an appropriate interpretation of the overall performance
specification.
2.9 Conclusions
In the wider context of the multivariable feedback stabil-

ity problem, Nyquist diagrams and Root-Locus diagrams are
seen to be algebraic curves derived from appropriate matrix-
valued functions of a complex variable. From this point of
41
view, the stability and design approaches of classical feed-

back theory can be thought of as simple exploitations of the
theory of algebraic curves. It is thus possible, and perhaps
even likely, that there will be considerable further develop-
ments of complex-function-based approaches to linear feedback
systems of increasing dynamical complication which exploit
appropriate properties of algebraic curves. Such an approach
would have considerable intuitive appeal, and would be well
suited for use with the increasingly powerful interactive-
graphic terminals which are becoming available for use by de-
sign engineers. The utility and acceptability of such a
development would be greatly enhanced if it were closely
linked to the state-space models on which so much of current
control theory is based. In the forms of generalization pre-
sented here, there is clear evidence of useful and intuit-
ively-appealing links between the algebraic curve theory and
the parameters of the state-space model; this is a most
hopeful augury for the future.
References
1. Postlethwaite, I. and MacFarlane, A.G.J. A Complex

Variable Approach to the Analysis of Linear Multivariable
Feedback Systems, Lecture Notes in Control and Informa-
tion Sciences, Vol. 12, Springer-Verlag, Berlin, 1979.
2. MacFarlane, A.G.J. and Postlethwaite, I. "The general-
ized Nyquist Stability Criterion and Multivariable Root
Loci", International Journal of Control, Vol. 25, 1977
pp.581-622.
3. Barman, J.F. and Katzenelson, J. "A Generalized Nyquist

type Stability Criterion for Multivariable Feedback Sys-
tems", International Journal of Control, Vol. 20, 1974,
pp. 593-622.
4. Bohn, E.V. and Kasvand, I. "Use of Matrix Transforma-
tions and System Eigenvalues in the Design of Linear
Multivariable Control Systems", Proceedings of the
Institution of Electrical Engineers, Vol. 110, 1963,
pp. 989-996.
5. MacFarlane, A.G.J. "The Return-difference and Return-
ratio Matrices and their Use in the Analysis and Design
of Linear Multivariable Feedback Control Systems",
Proc. of the Inst. of Elec. Engrs., Vol. 117, 1970,
pp. 2037-2049.
42
6. Retallack, D.G. "Extended Root-Locus Technique for

Design of Linear Multivariable Feedback Systems",
Proc. Instn. of Elec. Engrs., Vol. 117, 1970, pp.618-622.
7. Kouvaritakis, B. and Shaked, U. "Asymptotic Behaviour
of Root Loci of Linear Multivariable Systems", Internl.
J. of Control, Vol. 23, 1976, pp. 297-340.
8. MacFarlane, A.G.J., Kouvaritakis, B. and Edmunds, J.M.
"Complex Variable Methods for Multivariable Feedback
Systems Analysis and Design", Proc. of Symp. on Alter-
natives for Linear Multivariable Control, National Engin-
eering Consortium, Chicago, 1978, pp. 189-228.
9. Kouvaritakis, B. and Edmunds, J.M. The Characteristic
Frequency and Characteristic Gain Design Method for
Multivariable Feedback Systems", Proc. of Symp. on
Alternatives for Linear Multivariable Control, National
Engineering Consortium, Chicago, 1978, pp. 229-246.
10. Kouvaritakis, B. "The Optimal Root Loci of Linear
Multivariable Systems", Internl. J. of Control, Vol.28,
1978, pp. 33-62.
11. Owens, D.H. Feedback and Multivariable Systems,
Peter Peregrinus, Hitchin, England, 1978.
12. MacFarlane, A.G.J., Ed., Complex Variable Methods for
Linear Multivariable Feedback Systems, Taylor and
Francis, London, 1980.
13. MacFarlane, A.G.J., Ed., Frequency Response Methods in
Control Systems, IEEE Reprint Series, IEEE, New York,
1979.
14. MacFarlane, A.G.J. and Karcanias, N. "Relationships
Between State-Space and Frequency-Response Concepts",
Preprints of Seventh IFAC World Congress, Vol. 3, 1978,
pp. 1771-1779.
15. Posenbrock, H.H. State Space and Multivariable Theory,
Nelson, London, 1970.
16. Moore, B.C. "Singular Value Analysis of Linear Systems,
Parts I and II", Systems Control Reports 7801, 7802,
July 1978, University of Toronto.
17. Klema, V.C. and Laub, A.J. "The Singular Value Decom-
position : Its Computation and Some Applications",
Laboratory for Information and Decision Systems Report
LIDS-R-892, M.I.T., Cambridge MA. See also IEEE Trans.
on Auto. Control, Vol. 25, April 1980.
18. Laub, A.J. "Linear Multivariable Control. Numerical
Considerations", Laboratory for Information and Decis-
ion Systems, Report ESL-P-833, July 1978, M.I.T.,
Cambridge, MA.
19. MacFarlane, A.G.J. and Scott-Jones, D.F.A. "Vector Gain"
International Journal of Control, Vol. 29, 1979,pp.65-91.
20. Postlethwaite, I., Edmunds, J.M. and MacFarlane, A.G.J.
"Principal Gains and Principal Phases in the Analysis of
Linear Multivariable Feedback Systems", Cambridge
43
University Engineering Dept., Report CUED/F-CAMS/TR201,

1980, Cambridge, England.
21. MacFarlane, A.G.J. and Karcanias, N. "Poles and Zeros
of Linear Multivariable Systems : A Survey of Algebraic
Geometric and Complex-Variable Theory", Int. J. Con-
trol, Vol. 24, 1976, pp. 33-74.
22. Daniel, R.W. "Rank-deficient Feedback and Disturbance
Rejection", Int. J. Control, Vol. 31, 1980, pp.547-554.
23. Desoer, C.A. and Schulman, J.D,. "Zeros and Poles of
Matrix Transfer Functions and Their Dynamical Inter-
pretation", IEEE Trans, on Circuits and Systems, Vol.
CAS-21, 1974, pp. 3-8.
24. Kouvaritakis, B. and MacFarlane, A.G.J. "Geometric
Approach to Analysis and Synthesis of System Zeros
Part I : Square Systems, and Part 2 : Non-square Sys-
tems", Int. J. Control, Vol. 23, 1976, pp. 149-166
and pp. 167-181.
25. Karcanias, N. and Kouvaritakis, B. "The output and
Zeroing Problem and its relationship to the Invariant
Zero Structure : A Matrix Pencil Approach", Int. J.
Control, Vol. 30, 1979, pp. 395-415.
26. Sastry, S.S. and Desoer, C.A. "Asymototic Unbounded
Root Loci by the Singular Value Decomposition", Elec-
tronics Research Laboratory, College of Engineering,
Memorandum UCB/ERL M79/63 August 1979, University of
California, Berkeley, CA.
27. Postlethwaite, I. "The Asymptotic Behaviour, the
Angles of Departure and the Angles of Approach of the
Characteristic Frequency Loci", Int. J. Control,
Vol. 25, 1977, pp. 677-695.
28. Edmunds, J.M. "Calculation of Root Loci and Zeros
Using a Bilinear Transformation", Colloquium on
Numerical Algorithms in Control, Institution of Elec.
Engrs., London, 1979.
29. Kouvaritakis, B. and Shaked, U. "Asymptotic Behaviour
of Root Loci of Linear Multivariable Systems",
Int. J. Control, Vol. 23, 1976, pp. 297-340.
30. Kouvaritakis, B. and Edmunds, J.M. "Multivariable Root
Loci : A Unified Approach to Finite and Infinite Zeros",
31. Owens, D.H. "Dynamic Transformations and the Calcula-
tion of Multivariable Root Loci", Int. J. Control,
Vol. 28, 1978, pp. 333-343.
32. Owens, D.H. "A Note on Series Expansions for Multi-
variable Root Loci", Int. J. Control, Vol. 25, 1977,
pp. 819-820.
33. Shaked, U. "The Angles of Departure and Approach of
the Root Loci in Linear Multivariable Systems",
44
34. Thompson, P.M., Stein, G. and Laub, A.J. "Analysis

Techniques for Multivariable P.oot Loci", Laboratory for
Information and Decision Systems, Report LIDS-P-965,
M.I.T., Cambridge, MA.
35. Kwakernaak, H. "Asymptotic Root Loci of Multivariable
Linear Optimal Regulators", IEEE Trans, on Automatic
Control, Vol. AC-21, 1976, pp. 373-382.
36. Shaked, U. and Kouvaritakis, B. "The Zeros of Linear
Optimal Control Systems and their Role in High Feedback
Gain Stability Design", IEEE Trans, on Auto. Control,
Vol. AC-22, 1977, pp. 597-599.
37. Kalman, R.E. "When is a Linear Control System Optimal?",
Trans, of ASME Journal of Basic Eng. Series D, Vol. 86,
1964, pp. 51-60.
38. Safonov, M.G. and Athans, M. "Gain and Phase Margins
for Multiloop LQG Regulators", IEEE Trans. Auto.
Control, Vol. AC-22, 1977 pp. 173-179.
39. Sain, M., Peczkowski, J.L. and Melsa, J.I., Eds.,
Alternatives for Linear Multivariable Control, National
Engineering Consortium, Chicago, 1978.
40. Edmunds, J.M. "Cambridge Linear Analysis and Design
Programs", Cambridge University Engineering Department,
Report CUED/F-CAMS/TR 198, 1980, Cambridge, England.
41. Edmunds, J.M. "Cambridge Linear Analysis and Design
Programs", Proc. of IFAC Symp. on Computer-Aided
Control System Design, Zurich, Switzerland, 1979.
42. Edmunds, J.M. "Control System Design and Analysis
Using Closed-Loop Nyquist and Bode Arrays", Int. J.
Control, Vol. 30, 1979, pp. 773-802.
43. Wonham, W.M. Linear Multivariable Control : A Geometric
Approach, 2nd ed., Springer-Verlag, Berlin, 1979.
44. Wonham, W.M. "Geometric State-Space Theory in Linear
Multivariable Control : A Status Report", Automatica,
Vol. 15, 1979, pp. 5-13.
45. Moore, B.C. "On the Flexibility Offered by State
Feedback in Multivariable Systems Beyond Closed-Loop
Eigenvalue Assignment", IEEE Trans. Automatic Control,
Vol. AC-21, 1976, pp. 689-691.
46. Harvey, C.A. and Stein, G. "Quadratic Weights for
Asymptotic Regulator Properties", IEEE Trans, on
Automatic Control, Vol. AC-23, 1978, pp. 378-387.
47. Stein, G. "Generalized Quadratic Weights for Asymptotic
Regulator Properties", IEEE Trans, on Automatic Control
Vol. AC-24, 1979, pp. 559-566.
48. Grimble, M.J. "Design of Optimal Output Regulators
Using Multivariable Root Loci", Sheffield City Poly-
technic, Report EEE/37, April 1979, Sheffield, England.
49. Safonov, M.G. "Choice of Quadratic Cost and Noise
Matrices and the Feedback Properties of Multiloop LOG
45
Regulators", Proc. Asilomar Conf. on Circuits, Systems

and Computers, Pacific Grove CA, November, 1979.
50. Doyle, J.C. "Robustness of Multiloop Linear Feedback
Systems", Proceedings of 17th IEEE Conf. on Decision
and Control, San Diego, CA, January, 1979.
51. Postlethwaite, I. "A Note on the Characteristic
Frequency Loci of Multivariable Linear Optimal Regula-
tors", IEEE Trans. Automatic Control, Vol. AC-23,
1978, 757-760.
52. Valenca, J.M.E. and Harris, C.J. "Nyquist criterion
for input/output stability of multivariable systems",
Int. J. Control, Vol. 31, 1980, 917-935.
53. Desoer, C A . and Wang, Y.-T. "On the Generalized
Nyquist Stability Criterion", IEEE Trans. Automatic
Control/ Vol. AC-25, 1980, 187-196.
Chapter 3
Robustness in multivariable control

system design
Dr. IAN POSTLETHWAITE
Synopsis
In this lecture, techniques are described for assessing

the relative stability of a multivariable control system
design. An interesting and important feature in multivari-
able feedback systems is that they have different stability
margins at different points in their configuration.
3.1 Introduction
In the generalized Nyquist stability criterion put forward

by MacFarlane , and subsequently proved by several research-
12 7
ers ' ' , the stability of a linear, time-invariant, multi-
variable, feedback system, such as is shown in Fig. 3.1, is
determined from the characteristic gain loci which are a
plot of the eigenvalues of the return-ratio matrix G(s)K(s)
(or equivalently the eigenvalues of the return-ratio matrix
K(s)G(s)) as the Laplace transform variable s traverses
the standard Nyquist D-contour. The criterion can be stated
as follows:
Generalized Nyquist Stability Criterion. The feedback sys-

tem is stable if, and only if, the number of anticlockwise
encirclements of the critical point -1 by the characteristic
gain loci, is equal to the number of open-loop unstable
poles.
If a gain parameter k is introduced into each of the

loops then closed-loop stability can be checked on the same
plot for a family of parameter values k, by counting
47
plant
Fig. 3.1 Feedback system.
encirclements of -1/k. Also, in analogy with the scalar

situation, gain and phase margins can be defined from the
characteristic gain loci which indicate the limiting values
of the modulus and phase that k can attain before instab-
ility occurs. Although these features are very attractive
the characteristic gain loci do not by themselves give reli-
able information concerning the robustness of the closed-
loop stability property. For example, the characteristic
gain loci can have infinite gain margins with respect to the
gain parameter k, and only a small change in some other
system parameter may make the closed-loop system unstable.
In search of a reliable means of assessing the robustness

of the closed-loop stability property, Doyle and Stein
introduced a gain margin concept defined in terms of the
maximum singular value of an appropriate frequency-response
matrix. They showed that the closed-loop system remains
stable when subjected to a stable matrix perturbation I+P(s),
at any point in the configuration, providing the maximum
singular value of P(ju)) multiplied by the maximum singular
value of the frequency-response matrix [I+T(ju)) ] is less
than 1, at all frequencies, where T(s) is the return-ratio
matrix for the point in question. Note that since the return
ratio matrix is generally different at different points in
the configuration it follows that the robustness of the stab-
ility property will also vary around the system, and in par-
ticular it will vary before and after the plant where
uncertainties are most likely to occur. Following the work
48
7 8
of Doyle and Stein , Postlethwaite, Edmunds and MacFarlane
introduced a multivariable phase margin which in conjunction
with Doyle and Stein's gain margin, allows a larger class of
perturbations to be identified for which the feedback system
remains stable. These results are discussed in the latter
half of this lecture.
A system which has a large degree of stability in terms of
these multivariable gain and phase margins will have charac-
teristic gain loci which although not necessarily insensitive
to small perturbations will nevertheless be insensitive
enough for the number of encirclements of the critical point
to remain unchanged. In the first half of this lecture a
direct study is made of the sensitivity of the characteristic
gain loci to small perturbations. Sensitivity indices are
defined in terms of the left-hand and right-hand eigenvectors
of the return-ratio matrix corresponding to the point at
which the system is perturbed. It follows that the charac-
teristic gain loci are at their least sensitive when the
corresponding return-ratio matrix is normal.
3.2 Sensitivity of the Characteristic Gain Loci
In this section the perturbations of the eigenvalues of a

return-ratio matrix T(jco) are considered as the matrix is
perturbed to T(ja)) + eE(jw), where e is a small, posit-
ive, real number. The analysis is based on two theorems due
to Gerschgorin , and follows exactly the treatment given by
Wilkinson in his excellent monograph on the numerical
aspects of the algebraic eigenvalue problem. Because of
this and in order to concentrate on the main ideas, some
details and proofs will be omitted; the interested reader
is strongly encouraged to study Wilkinson's book.
For convenience, the jw's are sometimes omitted, the

notation ti- will be used for the ijth element of the mxm
matrix T, and 11^. | will denote the modulus of t. ..
Theorem 1 (Gerschgorin)
Each eigenvalue of T lies in at least one of the discs

49
with centres t. . and radii I |t. . | .

11 13
i*j
Theorem 2 (Gerschgorin)5
If n of the circular discs of Theorem 1 form a connected

domain which is isolated from the other discs, then there are
precisely n eigenvalues of T within this connected domain.
3.2.1 Sensitivity indices
In the analysis that follows a set of quantities will be

used repeatedly. They are of major importance and so will
be introduced first. They are defined by the relations
s
i = Zi 2Si i = 1,2,...,m
where T
^. is y. transposed, and y.,x. are (respectively)
left-hand and right-hand eigenvectors of T, normalized so
that the sum of the squares of the moduli of their elements
is equal to unity.
As will be seen later the m numbers |s^| give a direct

indication of the sensitivity of the m eigenvalues of T and
will therefore be referred to as the sensitivity indices of T.
It is clear that the sensitivity indices range between 0 and
1 .
3.2.2 Analysis
In the generic situation where T has linear elementary

divisors there exists a non-singular matrix W such that
W""1TW = diag(A ± )
where the A. are the eigenvalues of T, and diag(A.) is a

diagonal matrix of the A^. It is immediately apparent that
the columns of W are a candidate set of right-hand eigen-
vectors for T, and also that the rows of W~ are a candi-
date set of left-hand eigenvectors for T. Further, if the
column vectors of W are normalized to unity then the ith
column of W can be taken to be x., and the ith row of W
50
taken to be Z^/Sj_ •
With this definition for W it follows that
e 11/s 1 e 12/s 1 e lm /s l
e 21 /s 2 e 22 /s 2 ... e 2m / s 2
W -1 (T+eE)W = diag(X i ) + e
eml /sm em2 /sm emnr m
and the application of Theorem 1 to this shows that the per-

turbed eigenvalues lie in discs with centres (A.+ee^./s^)
and radii e £ |e../s.|. If all the eigenvalues are simple
then for a sufficiently small e all the discs are isolated

and by Theorem 2 each disc contains precisely one eigenvalue.
Alternatively, if there are some multiple eigenvalues, for
example A. = and A. = Ac/ then there will be three
Gerschgorin discs with centres A + ee
i i / s i ' i = -1'2'3 a n d
l
ee s = 4
two with centres A 4 + ±j_/ i' i > 5 , all with radii of
order e. For sufficently small e the group of three discs
will be isolated from the group of two, but, in general, no
claim of isolation can be made about the individual discs in
a group.
From the above analysis it is clear that the sensitivity of

the eigenvalues is directly related to the sensitivity indices;
the larger s is, the less sensitive is to small per-
turbations in T.
It is now demonstrated, by example, how, as T approaches

a matrix having non-linear elementary divisors, the corres-
ponding eigenvalues become increasingly sensitive. To facil-
itate this let T be given by
T =
which approaches a simple Jordan submatrix as approaches

The eigenvalues of an<
are 3 the and
, .
corresponding normalized right-hand and left-hand eigenvectors
are given by
51
1
1
[l+(t ll"t2 t
22
<
t
l l "t22 0
=
V
1 > 1
where for convenience t, , and t22 have been assumed real.

The sensitivity indices are therefore given by
and it is seen that both indices tend to zero as t

11
approaches t22. Hence from the previous analysis both simple
eigenvalues become increasingly sensitive as T tends to the
simple Jordan submatrix.
3.3 Uncertainty in a Feedback System
This section considers the implications of the sensitivity

results of Section 3.2 to the design of feedback systems.
For the feedback configuration of Fig. 3.1 the characteris-

tic gain loci are the eigenvalues of the return-ratio matrix
G (jw)K (jco) , or alternatively the eigenvalues of the return-
ratio matrix K(jw)G(jw). Although the eigenvalues of the
two return-ratio matrices are the same, the corresponding
eigenvectors will generally be different, and hence the char-
acteristic gain loci may be sensitive to perturbations in
G(ju))K(ja>) while insensitive to perturbations in K (ju>) G( jw) .
To see the significance of this consider the feedback config-
uration shown in Fig. 3.2, where the return-ratio matrix
G(ju))K(j(u) is assumed to have been perturbed to G(jw)K(jw) +
eE(jo))- Insofar as stability is concerned, this configura-
tion is equivalent to that shown in Fig. 3.3, where
Eo(ja)) = eE(jw) [ (G(joj)K(jw) ) ]'1

52
€E(ju
K(jou) (joa)
-I
Fig. 3.2 Perturbation of return-ratio matrix

G(ja))K(jaO
Therefore, the sensitivities of the eigenvalues of G( ja>)K (ju>)

give information concerning stability with respect to small
perturbations (uncertainty) associated with the plant output.
Note that at frequencies for which [G(jw)K(jw)]" becomes
very large, the moduli of the characteristic gain loci become
very small, and hence their sensitivity becomes unimportant.
(jui)
Fig. 3.3 Perturbation associated with plant output.
In an analogous fashion it can be shown that the sensitiv-

ities of the eigenvalues of K(jw)G(jw) give information
concerning stability with respect to small perturbations (un-
certainty) associated with the plant input. For more compli-
cated feedback configurations a return-ratio matrix can be
defined at any particular point at which uncertainty is felt
to be a problem and the sensitivity of the characteristic
gain loci checked via the corresponding sensitivity indices.
The return-ratio matrices defined for break points before and
after the plant will normally be the most important, and will
therefore be referred to as the input and output return-ratio
matrices, denoted by T.(s) and T (s) respectively.
53
3.3.1 Normality
The sensitivity indices are numbers between 0 and 1, and

when they are all 1 it is clear that the characteristic gain
loci are at their least sensitive. A class of matrices for
which this is always the case is the set of normal matrices.
Diagonal matrices are normal and therefore design techniques

which aim to decouple the outputs from the inputs using a
precompensator are well founded from the point of view of
achieving robustness with respect to small perturbations at
the plant output. It does not follow, however, that such a
design would necessarily be robust with respect to small per-
turbations at the plant input.
3.4 Relative Stability Matrices
In order to study the robustness of the closed-loop stabil-

ity property with respect to large and possibly dynamic per-
turbations consider the feedback configuration of Fig. 3.4,
where I+AG(S) represents a multiplicative perturbation of
the plant from G(s) to G(s) + G(s)AG(s). The configura-
tion can be redrawn as shown in Fig. 3.5, where the pertur-
bation AG(s) now appears in series with a transfer function
R±(s), given by
R. (s) A [I+(K(s)G(s))~ 1 ]" 1
R.(s) will be called the relative stability matrix with res-

pect to uncertainty at the plant input, or simply the input
relative stability matrix.
For a single-input, single-output system a perturbation

1+Ag(s) commutes around the feedback loop and the relative
stability function [l+(k(s)g(s) ] is the same at any
point in the configuration. However, in multivariable sys-
tems this is not the case, and thus it becomes important to
examine the effects of perturbations at different points in
the loop. The most likely places for uncertainties to occur
are at the plant input, considered above, and at the plant
output. To analyse the effects of uncertainty at the output
54
K (s) I • AG ( s ) G(s)
Fig. 3.4 Dynamic uncertainty at plant input.
AG (s) G(s) K (s)
Fig. 3.5 A rearrangement of Fig. 3.4 .
consider a multiplicative perturbation, I+AG(s), after the

plant shown in Fig. 3.6 . This configuration can be redrawn
as shown in Fig. 3.7, where AG(s) now appears in series with
a transfer function RQ(s), given by
R Q (s) A
R (s) w i l l b e c a l l e d t h e relative stability matrix with r e s -

pect to uncertainty at the plant output, or simply the output
relative stability matrix. It is interesting to note that the
input and output relative stability matrices are related by
the following expression:
x
R ± (s) = G(s) R Q (s)G(s)
•" K (s) *- G (s) (s)
Fig. 3.6 Dynamic uncertainty at the plant output.
Fig. 3.7 A rearrangement of Fig. 3.6 .

55
In the next section, it is shown how multivariable gain and

phase margins can be defined in terms of the relative stabil-
ity matrices thereby characterizing perturbations for which
the feedback system remains stable.
3.5 Multivariable Gain and Phase Margins
If R(s) is the relative stability matrix corresponding to

a multiplicative perturbation I+AG(s), at either the input
o
or output of the plant, then conditions can be derived for
AG(s), in terms of R(s), for which the perturbed closed-
loop system remains stable. The conditions are summarized in
theorems 5 and 6 after first introducing some definitions and
notation.
The gain and phase information from which the multivariable
gain and phase margins are derived stems from the polar de-
4
composition of a complex matrix which is now defined. Anal-
ogous to the polar form of a complex number, a complex matrix
R can be represented in the forms
R = UH D (3.1)
R = HLU (3.2)
where U is unitary and HR, HL are positive semidefinite

Hermitian matrices. H D , HT are often called right and left
moduli, and are uniquely determined as /R*~R*, /RR* respec-
tively, where * denotes complex conjugate transpose, and U
is uniquely defined via (3.1) or (3.2) when R is nonsingular.
Representations (3.1) and (3.2) are called Polar Decompositions
of R, and are easily determined from the Singular Value De-
composition of R , for which there exists well tested
software. If R has the singular value decomposition
R = XIV ; X,V unitary, and £ diagonal

and > 0
then R = (XV*) (VZV*) = UH O
and R = (XIX*) (XV*) = HTU

J
56
Prom the polar decompositions the following key quantities

are defined:
(i) The eigenvalues of R are called characteristic gains.

(ii) The eigenvalues of the Hermitian part in either polar
decomposition of R are called principal gains, or
singular values,
(iii) The arguments of the eigenvalues of the unitary part
of the polar decompositions of R are called principal
phases.
These three quantities are related by two important theorems.
Theorem 3 The magnitudes of the characteristic gains are

bounded above and below by the maximum and minimum principal
gains.
Theorem 4 If the principal phases have a spread of less

than IT, then the arguments of the characteristic gains are
bounded above and below by the maximum and minimum principal
phases.
Based on these theorems Postelthwaite, Edmunds and

o
MacFarlane derived the reobustness results discussed in the
remainder of this section.
Let R(jaj) have principal gains and principal phases
and
respectively; it is assumed that the {0^ (<*>)) have a spread

of less than n. Similarly, let the principal gains and
phases of AG(JUJ) be
(co) _< 6 2 (w) <_ ... _< 6 (a))

and
respectively. Also, let the condition numbers of R(jw) and

57
AG(joj), using the ^-induced norm, be c,(u)) and c2(a0

respectively, so that
c
l
and
C
2 (w
and also define
(a)) 4 tan
which will be referred to as a phase modifier. Then the

following theorems can be stated.
Theorem 5 (Small gain theorem) The perturbed closed-loop

system remains stable, if
(i) AG(s) is stable, and

<s a) (X aj < x for
(ii) m( ) m( ) '
Theorem 6 (Small phase theorem) The perturbed closed-loop

system remains stable, if
(i) A G ( S ) is stable,
(ii) {Q^(w)+e•(w) : i,j = 1,...,m} have a
spread of less than TT , for all w,
(iii) [c 1 (u))-1]C2 (w) < 1/ for all u>,
(iv) £;L (a)) + 6 1 (w) ~^ m (w) > -TT, for all w, and
e +G +l a) < 7Ti for
(v) m^) m^^ ^m^ '
Corollary 1. By symmetry theorem 6 can be restated with

c.(w) and C2(w) interchanged.
Corollary 2. As a consequence of theorems 5 and 6, the

perturbed closed-loop system remains stable,
if, for some frequency w, ,
(i) the conditions of theorem 6 are satisfied

58
in the frequency range [0,0)^3, and

(ii) the conditions of theorem 5 are satisfied
in the frequency range [w^, 00 ].
Remark 1. Theorem 5 is exactly the small gain theorem

introduced by Zames for a more general class of systems.
It is also the robustness result used in the LQG-based
3 9
designs of Doyle and Stein and Safonov, Laub and Hartmann .
Remark 2. Although it might be argued that theorem 5 gives

an adequate characterization of the robustness of the stab-
ility property, the combination of theorems 5 and 6 allows a
larger set of perturbations to be identified for which the
closed-loop system remains stable.
Remark 3. If in corollary 1, AG(s) is simply assumed to

be a phase change in each of the feedback loops, then
AG(j(u) = diag{exp (j e^) } and c2(u>) = 1 , so that condition
(iii) is always satisfied, and the phase modification ty (OJ)
is always zero.
Remark 4. Typically a perturbation A G ( S ) representing

small parameter variations and unmodelled high frequency
dynamics will, at low frequencies, have a condition number
of approximately 1, and small principal gains. At high
frequencies its maximum principal gain will increase drasti-
cally, and the minimum principal phase will exceed 180
degrees lag. Consequently, when predicting perturbations
for which the closed-loop systems remain stable, a combina-
tion of theorems 5 and 6, as in corollary 2, is useful.
Remark 5. Theorem 5 indicates that, at frequencies for

which the maximum principal gain of the perturbation is
large, the maximum principal gain of R(joo) should be de-
signed to be small; this will normally be the case at high
frequencies. At low frequencies, theorem 6 indicates that
a perturbation with large gains can be tolerated providing
) is designed to have a condition number close to 1 and
59
a small spread of principal phases. This situation will

normally be feasible at low frequencies but at high frequen-
cies when the phase lag inevitably exceeds 180 degrees the
maximum principal gain of R(jw) has necessarily to be made
small.
From the preceding remarks it is clear that useful inform-

ation concerning the robustness of the closed-loop stability
property, is readily obtained from Bode plots of the princi-
pal gains and phases of the input and output relative stab-
ility matrices. Loosely speaking, in going from the scalar
to multivariable case, analysis moves from studying a single
Bode magnitude and phase plot to studying, for each point of
uncertainty, a band of Bode plots (see Figs. 3.8 and 3.9),
defined by the maximum and minimum principal gains and phases
of the appropriate relative stability matrix. Also, as an
indicator of the robustness of the stability property, gain
and phase margins can be defined in terms of the maximum
principal gain and the minimum principal phase (maximum
phase lag) by considering them to be standard Bode magnitude
and phase plots for a scalar system.
1.0
0.01
Fig. 3.8 Principal gains for a 3-input, 3-output

system; band of Bode magnitude plots.
0.01
Fig. 3.9 Principal phases for a 3-input, 3-output

system; band of Bode phase plots.
60
3.6 Conclusion
The sensitivity indices corresponding to the input and out-

put return-ratio matrices indicate the sensitivity of the
characteristic gain loci to small perturbations (uncertain-
ties) at the input and output of the plant, respectively.
The indices therefore, give crucial information concerning
the possibility of closed-loop instability.
For large and possibly dynamic perturbations (uncertain-

ties) the robustness of the closed-loop stability property
is characterized by the principal gains and principal phases
of the input and output relative stability matrices. From
Bode-type plots of the principal values, multivariable gain
and phase margins can be defined which are analogous to the
classical single-loop stability margins.
REFERENCES
1. BARMAN, J.F. and J. KATZENELSON : A generalizedNyquist

type stability criterion for multivariable feedback
systems. Int. J. Control, 20, (1974), pp. 593-622.
2. DESOER, C.A. and Y.T. WANG : On the generalized
Nyquist stability criterion. IEEE Trans. Autom. Con-
trol, 25, (1980), pp. 187-196.
3. DOYLE, J.C. and G. STEIN : Multivariable feedback
design : concepts for a classical/modern synthesis.
IEEE Trans. Autom. Control, 26, (1981), pp. 4-16.
4. GANTMACHER, F.R. : Theory of Matrices, Vol. 1, (1959)
Chelsea, New York.
5. GERSCHGORIN, S. : Uber die Abgrenzung der Eigenwerte
einer Matrix. Iz.Akad.Nauk SSSR, (1931), 6, pp.749-754.
6. MACFARLANE, A.G.J. : Return-difference and return-
ratio matrices and their use in analysis and design of
multivariable feedback control systems. Proc. IEE, 117,
(1970), pp. 2037-2049.
7. MACFARLANE, A.G.J. and I. POSTLETHWAITE : The general-
ized Nyquist stability criterion and multivariable root
loci. Int. J. Control, 25, (1977), pp. 81-127.
8. POSTLETHWAITE, I., J.M. EDMUNDS and A.G.J. MACFARLANE :
Principal gains and principal phases in the analysis of
linear multivariable feedback systems. IEEE Trans.
Autom Control, 26, (1981), pp. 32-46.
9. SAFONOV^ M.G., A.J. LAUB and G.L. HARTMAN : Feedback
properties of multivariable systems : the role and use
of the return difference matrix. IEEE Trans. Autom.
Control, 26, (1981), pp. 47-65.
61
10. STEWART, W.G. : Introduction to Matrix Computations.

Academic, New York, (1973).
11. WILKINSON, J.H. : The Algebraic Eigenvalue Problem.

Clarendon, Oxford, (1965).
12. ZAMES, G. : On the Input-Output Stability of Time

Varying Nonlinear Feedback Systems. IEEE Trans.
Autom. Control, 11, (1966), pp. 228-238.
62
PROBLEMS
P.3.1 In the feedback configuration of Fig. 3.1, the plant

and controller are given by
-47s+2 56s
G(s)
" (s+1)(s+2)
-42s 50s+2
0
K(s) =
0 k.
where nominally k, = k~ = 1.
The plant has been contrived so that it can be dia-

gonalized at all frequencies using a constant matrix
W. For the nominal controller parameters find W
and hence the sensitivity indices, and plot the
characteristic gain loci. Discuss the robustness
of the closed-loop stability property with respect
to variations in k, and
P.3.2 Figure 3.10 shows the principal gains and phases of

the output relative stability matrix for an automo-
tive gas turbine design study. Discuss briefly the
relative stability of the design with respect to a
dynamic perturbation I+AG(s) after the plant model.
63
0.10 10.0
0.31
0.10
0.03
-50-
-100-
-150
0.10 1.0 10.0
Fig. 3.10 Principal gains and phases of the

output relative stability.
Chapter 4
A design study using the characteristic

locus method
Dr. J. M. EDMUNDS
Synopsis
A 3-input multivariable control system is developed for the

regulation of the product concentration and temperature of a
2-bed exothermic catalytic reactor. Two of the inputs are
the flow rate and temperature of a quench stream injected be-
tween the beds; the third input is the feed temperature.
Through the use of the characteristic locus method of con-

trol system analysis, a cascade of dynamic compensators is
developed by which interaction among variables is suppressed
and the effects of feed concentration disturbances are con-
trolled. In this development, an analysis is made of the
potential improvements in performance accruing from the use
of internal bed temperature measurements. The system also
has the feature of resetting the quench flow rate to its
nominal value upon sustained system disturbance.
4-i\ Introduction
A multivariable control system for a two-bed catalytic re-

actor is synthesized here by a method that tailors the frequ-
ency-dependent eigen-properties of a transfer-function matrix
to achieve control of both concentration and temperature of
the reactor product stream. Such a control objective extends
earlier investigations of single-output control, and the
application of such a design technique complements earlier
applications of the Linear-Quadratic-Gaussian technique to
8 10
this reactor control problem ' . The best possible control
is sought, of course, with either design method, but here,
instead of seeking an optimum of a quadratic criterion, the
suppression of interaction between the two controlled vari-
ables is addressed directly.
To achieve control of both concentration and temperature

65
upon sustained disturbances in feed concentration, a control

system configuration such as that shown in Figure 4.1 is
proposed. An exothermic chemical reaction whose rate depends
upon both temperature and concentration takes place in both
beds of the reactor shown in this figure. It is this coup-
ling of variables through the reaction rate that suggests
the control configuration shown. Measurements of the c o n -
trolled variables C Q . and T Q are assumed available,
and measurements of fluid temperatures at three points inter-
nal to each bed may also be considered available for use in
I main feed
mixing quench
chamber stream
Fig. 4.1 Diagram of two-bed reactor and proposed control

configuration for the regulation of C Q u t and T Qut .
66
the control system should they prove beneficial. In indust-

rial circumstances, measurement of C o u t may be impossible,
in which case a reconstruction of this variable from tempera-
ture measurements may be attempted as demonstrated by Silva
and Wallman. Here, for purposes of system analysis, concen-
tration is considered measurable. Three manipulatable inputs
(main feed temperature TV, quench temperature T~, and
quench flow rate Q ) are to be used. Such a control con-
figuration is applicable in the frequently encountered indus-
trial circumstance where conditions only at the entrance of
each bed can be adjusted. With the exception of the input
Tf, such a configuration has been investigated in the recent
work of Silva and Wallman.
Following the results of that work, it is proposed to

employ the quench flow rate only for short-term corrective
action, forcing the quench flow rate to return to its nominal
value as quickly as practicable after the transient effects
of the disturbance have been suppressed. Such a stipulation
leaves inputs T^ and T~ to eliminate steady state off-
sets in C and T . Temperature inputs are best
suited for long-term corrections owing to their relatively
slow rate of propagation through the reactor and their rela-
tively high sensitivity. It can be argued on physical
grounds that T^ and T determine steady-state concentra-
tion and temperature at the entrance of Bed II and thus are
sufficient to eliminate steady-state offsets in the variables
T
CQut and out provided that the ranges of the manipulated
inputs are adequate for the size of the feed disturbance.
However, it is not so evident without some analysis that this
3-input configuration will be sufficient for the control of
dynamic conditions in the reactor. There are questions of
the relative effect that the inputs have on the outputs over
the important frequency range and the benefits accruing from
the creation and placement of zeros through the use of tem-
perature measurements at the three internal points in each
bed.
The synthesis of dynamic compensators for this control sys-

67
tern configuration is carried out here in two stages. First,

because Bed II is obviously the important dynamic element
of this configuration, a 2-input 2-output control system for
this bed is synthesized and analyzed. Bed I and the associ-
ated third input Tf is then added serially to controlled
Bed II with the additional requirement that the quench flow
rate is to be driven by integral action to its nominal value
at long times. All of these design stages are carried out
with the use of locally linear 7th-order state-space models
of the reactor beds of the form
x = Ax + Bu
y = Cx + Du
The approach to the design, however, was not rigidly pros-

cribed by these models but was influenced by the knowledge
that high-order process modes were lacking in the models.
To investigate the robustness of the designs, simulations
of closed-loop performance were made using 14th-order linear
models of the reactor, thus introducing the potentially
troubling phase lag of unmodelled modes. The design tech-
nique used is the characteristic locus method . The calcul-
ations were conveniently carried out on a minicomputer with
an interactive program specially developed for linear system
analysis .
4.2 Two-Bed Reactor Mo<3el
The reactor model used in these design calculations derives

from a set of nonlinear partial differential equations repre-
senting one-dimensional material and heat balances of the
packed bed. The equations and methods of their reduction to
o
linear state-space form have been detailed by Silva et al .
and Michelson et al , and are not repeated here. Rather it
is sufficient for purposes of this study to state the A,B,C
and D matrices for each bed and the coupling between the beds.
A sketch showing the input-output configuration of the 2-
bed system is given in Figure 4.2 . The models associated
68
with each block of this figure are as follows.
distu rbance^
C0
TI
'out Tm
TF feed L—- bed I r l mixing Cm bed II
T C
heater o X1 out chamber Q X"
input
III
T, T2 T.
n II 1
T 5 T 6 T7
," 'a,
control
inputs
Fig. 4.2 Reactor system components and input-output
variables. Notation: T = temperature,
C = concentration, Q = flow rate, X = state
vector.
Feed heater
-1 T + -1 T _
T o T F
Bed 1 + B1
+ D1
3
I
out
I
out
Mixing chamber
Q
= M I
out
I
"out
Bed 11
II
69
T
5
T
6
11
T
7
T
out
C
out '
The relations for the mixing chamber are linearizations of

the steady-state material and heat balances at that point.
Reactor models of two different orders are used here. The

design is carried out using a 7th-order model for each bed.
These models were obtained from I4th-order models by a reduc-
tion technique in which the fast modes are discarded but the
Q
steady state gains are preserved ; catalyst temperatures

were retained as the state variables. The 14th-order model
is used here in closed-loop simulations to test the perform-
ance of the control system designed on the basis of the lower
order model. The 14th-order model was not used for the de-
sign calculations since the software available at the time
the design was being carried out could not handle all 28
states.
Elements of the matrices of the 7th-order reactor model and
the mixing matrix are given in the appendix. The numerical
values are identical to those determined for the experimental
conditions of Silva and represent reactor dynamic behaviour
in the vicinity of the nominal steady state of those experi-
ments. Matrices for the 14th-order model are given by
Silva 7 .
4.3 Control System Design for Bed II
The control system for Bed II will employ the inputs Q

and T Q shown in Figure 4.2 to regulate effluent variables
T . and C . . One of the first matters to be investigated
in such a multivariable system is the sensitivity of the out-
puts to the inputs. To achieve independent control over
and C out' one would prefer the transfer function
"out
70
matrix G(s) relating outputs to inputs to be well condi-

T and c
tioned over all frequencies. Regulation of out out
to the setpoint, for example, requires that G(s=0) to be

invertible. The situation for this reactor transfer func-
tion matrix is reflected in the open-loop Nyquist array,
the characteristic loci, and the eigenvectors of G(s).
The first two of these are displayed in Figures 4.3 and

4.4 and reveal that we begin the design under inauspicious
circumstances. The close similarity of the columns of the
Nyquist array (Figure 4.3) and the large disparity of the
eigenvalues of G(s) (Figure 4.4) would both seem to imply
a near linear dependence of the inputs Q and TQ for this
set of outputs. However, the determinant of the steady
state gain matrix G(0) is only 10 to 20 times smaller than
the individual elements and thus suggests that control to
setpoint is feasible. A further indicator of the feasi-
bility of this 2x2 control configuration for Bed II is the
angle between the two eigenvectors of G(2). Significant
deviation from orthogonality of the eigenvectors implies
Q
-2-
Fig. 4.3 Myquist array for the second bed of the reactor.
Outputs ( T o u t , c o u t ) '' Inputs (Q, T Q ) .
71
Fig. 4.4 Characteristic loci of Bed II.
the need for inordinately large control inputs. The angle

between the two vectors for Fed II was found to be about
70° in the range of frequencies 0.1 to 3, the range abpve
which model accuracy cannot be assured. The situation seems
to be favourable; angles at the intermediate and high fre-
quencies appear significantly large, and it is these that
are relevant because Q is not used for low frequency
corrective action.
4.4 Compensator Design
The addition of compensators for purposes of forging the

characteristic loci into desired shapes proceeds through
four distinguishable stages: (1) suppression of interaction
at high frequency, (2) attenuation of the loci in the very
high frequency range for which the model is poor, (3) boo-
sting of gain at intermediate frequencies, and (4) addition
of the integral action to enhance steady state accuracy and
to suppress low frequency interaction.
Suppression of interaction at high frequencies is accom-

plished as described above by introducing a constant pre-
compensator K, that approximately aligns the eigenvectors
of G(s)K, with the unit vectors at a selected frequency,
which should be approximately the desired cut-off frequency.
The matrix obtained for the alignment at CD = 2 is

72
-1.16 1.04
(4.1)
-3.35 2.34
and the characteristic loci of the reactor with this pre-

compensator are given in Figure 4.5 . Upon comparison with
the uncompensated system (Figure 4.4), there is seen to be
a considerable reduction in the disparity between the eigen-
values of the system.
Fig. 4.5 Characteristic Loci of reactor with constant

pre-compensator K, to suppress high frequency
interaction.
The larger locus in Figure 4.5 exhibits appreciable mag-

nitude at frequencies of 3 and above, the range of discarded
and neglected modes in the reactor model. Because of the
uncertainties of the characteristic loci in this frequency
range, one is not justified in designing on the basis of
the loci at these frequencies. We therefore propose to
attenuate and lag the large locus at frequencies in the
region of a> = 2 to bring its magnitude and phase into
closer correspondence with the smaller locus.
Both the loci were lagged slightly in order to decrease

the bandwidth to be within the range where the 7th order
and 14th order models were similar. The compensators were
introduced using the approximate commutative controller
method calculated at a frequency w = 2,
73
0.852 0.620 (O.ls+0.5)(0.5s+0.5)

(s+0.5)(s+0.5
-0.414 0.923 0
0.896 -0.571
(4.2)
0.382 0.785
The choice of the lag frequency 0.5 is influenced by the

desire to decrease appreciably the magnitude of the locus at
a) = 2 but to maintain gains as high as possible at low and
intermediate frequencies. The loci resulting from the intro-
duction of K> (s) as a precompensator are shown in Figure
4.6 . The phases of the loci are now seen to be in much
closer correspondence. The loci magnitudes are also more
nearly equal.
0.4r
-0.6L
Figure 4.6 Characteristic loci with approximate

commutative controller at = 2 .
K, and K,(s) given by equations (4.1) and (4.2)
Compensation is now introduced to further modify these

loci for the purpose of boosting the magnitudes of each at
intermediate and low frequencies without affecting the phase
at high frequencies. Lags at w = 0.2 and 0.05, each with a
74
10-fold frequency range, are applied to the large and small

loci respectively. These lags are introduced through an
approximate commutative controller calculated at OJ = 0.3,
the approximate centre of the frequency range of the two nev/
lags. This precompensator is
1.223 0.810 1.495 1.858

K 2 (s)
-0.446 -0.651 -1.023 -2.807
s+0.05
The resulting loci are shown in Figure 4.7 . These loci

seem balanced well enough to attempt closure of the loop.
20r
0.0
-20
Fig. 4.7 Characteristic loci with K, (s) and K 2 ( s ) .
A scalar proportional-integral controller is introduced

at a frequency of 0.05; this produces high gains in both
loci at low frequency with only a few degrees of phase shift
in the slowest locus near the critical point. This compen-
sator is taken to be
K 3 (s) = 4 s+O* 1
75
The gain of 4 was selected to give a 45 phase margin to the

locus having the largest phase lag near the critical point.
To test the performance of this cascade of compensators,

transient responses to step changes in set point were cal-
culated using both the 7th- and 14-th order reactor models.
It was essential that the control system be tested on the
higher order model because the reactor modes neglected in
the design model could easily lead to instability if care
had not been exercised in the design sequence. The closed-
loop response of these two reactor models to a step increase
in the setpoint C . is shown in Figure 4.8 .
1.5
1.0
0.5
-0.5
-1.0
Fig. 4.8 Response of system to step changes in the

setpoint for C . .
7th order model with controllers K^,
K ^ s ) , K 2 (s) , K 3 (s)
14th order model with controllers K^
K x (s) , K 2 (s) , K 3 (s)
... 14th order model with controllers K, ,
K x (s) , K 2 (s) , K 3 (s) , K 4 (s) .
The responses display an acceptable degree of stability;

76
however, there is considerable interaction present in the

mid frequencies.
4.5 Design of the 2-Bed System
The incorporation of Bed I into the multivariable design

just described for Bed II is accomplished in a simple way
with simple objectives. Feed temperature IV is the new
8
input variable associated with Bed I and is known to have
little potential for regulating dynamic conditions in the
reactor. Its sole function in this design is therefore
relegated to driving the quench flow rate to its nominal
value after the quench has accomplished the major portion
of its short-term corrective action. The control system
configuration can therefore be viewed as sketched in Figure
4.9 . The Bed II compensator just developed is the product
of the individual compensators K,, K , K 2 and K~, and is
employed to regulate the two variables T . and C .;
the quench flow rate Q is considered the third output
variable and is linked to input Tf by a single loop.
bed II
compensation •out
K3(s)K2(s)K1(s)Kh TQ ^
2-bed -out
reactor
single-loop
PI controller
Fig. 4.9 Control System Configuration for 2-bed Reactor
The design of this simple loop between Tf and Q is

straight forward. The Nyquist locus, which is shown in
Figure 4.10, has a high-order character. The proportional-
integral controller
77
s+0.2
K 4 (s) = 0.07
when introduced into the loop yields a gain margin of about

3 and an integration rate that would reset Q in about 10
time units. Three time units correspond approximately to
one "time constant" of the uncontrolled 2-bed reactor. In
the course of the investigation of the performance of the
complete system, it was found that some additional tuning
could be made of the scalar PI controller Ko(s) of bed
II. The integration rate was doubled to give the modified
scalar controller
s+0.2
K 3 (s) = 4
O.Ar
072
0.47-
0.06
0.30-
016
Fig. 4.10 Nyquist locus of Q(s)/T f (s) of Fig. 4.9,

(bottom loop open, top loop closed).
4.£ System Performance and Properties
System performance was evaluated by simulation of the

reactor under the controls devised above upon a step distur-
bance in the feed concentration CQ . Fourteenth-order
models of both beds were used in these simulations.
A comparison of the controlled and uncontrolled transients

in and C is given in Figure 4.11 Both temper-
out out
78
ature and concentration are seen to be held by the controls

to small values relative to the uncontrolled case, and both
are returned to the vicinity of the set point with reason-
able speed. The control action required to achieve this
performance is shown in Figure 4.12 . The response to a
set point change is shown by the dotted line in Figure 4.8 .
Mid-frequency interaction is evident in both set point and
disturbance responses.
/ uncontrolled T ou(
i.o-
/ controlled T ou j
time
^Controlled Cout
\
-1.0- V uncontrolled C,
Fig. 4.11 Transient behaviour for a unit step distur-

bance on C using both beds of the 14 th-
order model.
6. Or
-2.0
Fig. 4.12 Control action while suppressing a unit step

disturbance on C Q for the 14th-order model.
79
The magnitude of the control inputs is an important matter

in multivariable systems; it is certain to be larger than
that of a single-output system because of the extra demands
on the controller. For example, in this case independent
control has been demanded over output temperature and con-
centration. In practice, the range of the inputs is limited
and thus an equivalent concern is the magnitude of the dis-
turbances that can be accommodated without saturation of the
control inputs. The results of Figure 4.12 imply, for ex-
ample, that this control system could accommodate feed con-
centration steps of up to 2.5% before saturation would occur
on the quench flow rate of the experimental reactor used in
earlier studies. For larger disturbances the transient
interaction would increase, but steady-state regulation
could still be achieved until one of the heaters saturates.
In the single-output systems of Silva and Wallman concentra-
tion disturbances of 10 to 20% could be handled before the
quench flow rate reached its limits.
A further design study using a computer to tune the con-

trol scheme reduced the use of Q by a factor of 3 while
also reducing the interaction considerably. This would allow
disturbances of up to 7.5% before Q saturates, which is
much more comparable with the single output results.
The three characteristic loci of the final 3x3 system are

shown in Figure 4.13 • One of the three is considerably
slower than the other two and is undoubtedly associated with
the more or less independent PI controller linking Q to
Tf . The two faster loci exhibit gain and phase character-
istics near the critical point that would contribute to some
slight overshoot of the controlled variables, and indeed
this is observed in the response of T and C t shown
in Figure 4.11 . Quench flow rate Q, however, responds
in an overdamped manner, as would be suggested by the shape
of the slower locus.
It is noticed that the seven poles introduced through the

compensators K^ (s) , K^(s), Ko(s) and K, (s), are clustered
in the region 0 to -0.5, about four times slower than the
80
reactor poles. Further, there is an excess of poles over

zeros introduced in that region. This compensator pole-zero
configuration is typical of that usually used for high-order
cascaded processes.
-2.00
Fig. 4.13 Characteristic Loci for Final Design.
The implementation of the set of compensators would be

straightforward in either analogue or digital realizations.
In a digital implementation, a 7th-order state-space realiz-
ation of the compensator transfer functions would be re-
quired, a size that is easily accommodated in a process con-
trol computer. More details of this design study can be
found in reference 4.
Appendix
Matrices for the Reactor System Model
Time unit = 87.5 sec

Concentration unit = 1.0 mole %
Temperature unit = 167.4 °C
Flow rate unit = 13.5 £/min.
Feed heater ~ = 0.2956

81
Bed 1
2.3832 -0.1717 0.1562 -0.1606 0.1802 -0.2527 0.1550

0.9280 -1.8688 0.3027 0.2642 -0.2818 0.3885 -0.2375
0.5462 1.8614 -1 4986 -0.4074 0.3689 -0.4846 0.2938
0.1813 0.4976 2.3399 -1.4191 -0.4404 0.4899 -0.2892
-0 3797 -0.7642 0.6631 2.5293 -1.7368 -0.3369 0.1716
0.0805 -0.8628 0.6172 1.7163 2.0183 -2.2067 -0.1147
0.0265 -0.7599 0.8771 1.4123 2.0361 0.8079 -2.7725
2.5936 0.0179
1.5082 0.0645 0.1760 0.0000
-0.1333 0.0946 -0.1088 0.0000
1
B = -0.5473 0.0902 D = 0.0450 0.0000
0.2840 0.1007 0.0027 0.0000
0.6519 0.1258 0.0000 0.7789
0.6540 0.1356
-0.3755 0.8200 0.4834 -0.1698 0.1229 -0.1427 0.0840

1 0.2118 -0.2439 0.5766 0.7193 -0.2745 0.2758 -0.1585
c = -0.0852 0.0860 -0.1199 0.3097 0.9534 -0.4021 0.2112
-0.0050 0.0048 -0.0059 0.0098 -0.0259 0.2917 0.7260
-0.0846 -0.1912 -0.2804 -0.3298 -0.3012 -0.1621 0.0000
Mixing Chamber
-0.3373 0.2370 0.7630 0.0000

M = -0.7226 0.0000 0.0000 0.7630
1.0000 0.0000 0.0000 0.0000
Bed 11
-2.5108 -0.1285 0.1029 -0.1277 0.1581 -0.2207 0.1346

0.7675 -2.1256 -0.2003 0.2004 -0.2335 0.3194 -0.1940
0.8449 1.7995 -1.8809 -0.2849 0.2694 -0.3464 0.2080
0.3177 1.5039 2.4101 -1.8352 -0.2749 0.2818 -0.1626
-0.6847 -0.2184 6907 2.2016 -2.0525 -0.2373 0.1205
-1.0197 -1.4581 0.3344 1.9009 1.4282 -2.3437 -0.1116
1-1.0121 -1.6634 -0.0508 1.7249 1.4403 0.6234 -2.8285
2.8314 0.0679 -0.0266 0.0471 0.0000 -0.0029

2.3639 0.2812 -0.1235 -0.0941 0.0000 -0.0039
0.8489 0.5514 -0.2026 0.0609 0.0000 -0.0035
-1.4594 0.5891 -0.0781 0.0026 0.0000 -0.0019
-1.5940 0.3488 0.2194 0.0000 0.2433 0.2759
-0.1837 0.2054 0.3809
0.2538 0.1872 0.4044
-0.1104 1.0098 0.0706 -0.0319 0.0244 -0.0289 0.0171

0.1903 -0.2400 0.8659 0.3817 -0.1912 0.2017 -0.1168
CJ1= -0.1188 0.1262 -0.1865 0.6320 0.6820 -0.4327 0.2346
-0.0049 0.0049 -0.0060 0.0100 -0.0263 0.3416 0.6759
-0.0465 -0.1135 -0.1909 -0.2619 -0.2634 -0.1422 -0.0002
82
References
1. EDMUNDS, J.M. : "The Cambridge Linear Analysis and

Design Program", CUED/F-CAMS/TP 170, Eng. Dept.,
University of Cambridge, (1978).
2. EDMUNDS, J.M. and KOUVARITAKIS, B. : "Multivariable
root loci : a unified approach to finite and infinite
zeros", Int. J. Control, vol. 29 No. 3 (1979).
3. EDMUNDS, J.M. : "Control system design and analysis
using closed-loop Nyquist and Bode arrays", Int. J.
Control, vol. 30, No. 5, (1979), 773.
4. FOSS, A.S., EDMUNDS, J.M. and KOUVARITAKIS, B.
"Multivariable control system for two-bed reactors by
the characteristic locus method", Ind. Eng. Chem. Fund.
19, (1980), 109.
5. MACFARLANE, A-G.J. and KOUVARITAKIS, B. : "A design
technique for linear multivariable feedback systems",
Int. J. Control, vol. 25, No. 6, (1977), 837.
6. MICHELSEN, M.L., VAKIL, H.B., FOSS, A.S., Ind. Engr.
Chem. Fund., 12 (1973), 323.
7. SILVA, J.M. : Ph.D. Thesis, University of California,
Berkeley, (1978).
8. SILVA, J.M., WALLMAN, P.H., FOSS, A.S. : Ind. Eng.
Chem. Fund, 18, (1979), 383.
9. V7ALLMAN, P.H. : Ph.D. Thesis, University of Calif.,
Berkeley, (1977).
10. WALLMAN, P.H., SILVA, J.M., FOSS, A.S. : Ind. Eng.
Chem Fund, 18, (1979), 392.
Chapter 5
The inverseNyquist array design method

Professor N. MUNRO
Synopsis
The general problem of the design of multivariable control

systems is considered and the stability of multivariable
feedback systems is examined. The concept of 'diagonal dom-
inance1 is introduced/ and Rosenbrock's Inverse Nyquist Array
Design Method is developed. Methods of achieving diagonal
dominance are discussed and illustrated in terms of practical
problems.
5.1 Introduction
The design of control systems for single-input single-out-

put plant using the classical frequency response methods of
Bode, Nyquist and Nichols is well established. However, the
frequency response approach of Nyquist has been extended by
Rosenbrock1 to deal with multi-input multi-output plant where
significant interaction is present.
During the last decade, interactive computing facilities

have developed rapidly and it is now possible to communicate
with a digital computer in a variety of ways; e.g. graphic
display systems with cursors, joysticks, and light-pens.
Equally, the digital computer can present information to the
user in the form of graphs on a display terminal or as hard-
copy on a digital plotter. The classical frequency-response
methods for single-input single-output systems rely heavily
on graphical representations, and Rosenbrock's 'inverse'
Nyquist array' design method for multivariable systems suit-
ably exploits the graphical output capabilities of the modern
digital computer system. Also, the increased complexity of
multivariable systems has made it necessary to employ inter-
active computer-aided design facilities, such as those 2
developed at the Control Systems Centre, UMIST, Manchester ,
in order to establish an effective dialogue with the user.
84
5.2 The lViultivariable Design Problem
In general, a multivariable control system will have m in-

puts and £ outputs and the system can be described by an &*m
transfer function matrix G(s). Since we are interested in
feedback control we almost always have £ = m.
The uncontrolled plant is assumed to te described by an

rational transfer function matrix G(s) and we wish to
determine a controller matrix K(s) such that when we close
the feedback loops through the feedback matrix F, as in
Figure 5.1, the system is stable and has suitably fast res-
ponses.
The matrix F is assumed diagonal and independent of sf

i.e.
diag {f } (5.1)
F represents loop gains which will usually be implemented

in the forward path, but which it is convenient to move into
the return path. In addition, the design will have high
integrity if the system remains stable as the gain in each
loop is reduced.
e(s) u(s)
r(s)- j———- Ms) G(s) y(s)
F(s)
Fig. 5.1
Let us consider the controller matrix K(s) to consist of

the product of two matrices K (s) and K-,, i.e.
K(s) = K (s) K d (5.2)

85
where Kd is a diagonal matrix independent of s; i.e.
= diag i = 1, . • . ,m (5.3)
Then the system of Figure 5.1 can be re-arranged as shown in

Figure 5.2 where the stability of the closed-loop system is
unaffected by the matrix K, outside the loop. For conven-
ience, we rename the matrix K^F as F. All combinations
of gains and of open or closed-loops can now be obtained by
a suitable choice of the f., and we shall want the system
to remain stable for all values of the from zero up to
their design values.
Fig. 5.2
When = 0 the first loop is open, and all gains in the

first loop up to the design value can be achieved by
"Id
increasing f, .
The elements f. of F diag {f^} can be represented by

points in an m-dimensional space which can be called the gain
space. That part of the gain space in which f^ > 0, i =
l,...,m corresponds to negative feedback in all loops, and
is the region of most practical interest. The point
{f.,f 2 ,...,f } belongs to the asymptotically stable region
in the gain space if and only if the system is asymptotically
stable with F = diag { f i ) .
Let
Q(s) G(s)K(s) (5.4)
86
then the closed-loop system transfer function matrix H(s)

is given by
H(s) = = Q(s) (5.5)
Consider the open-loop system
1 2
s+1 i+3
Q.(s) = 1 (5.6)
1
s+1 s+1
Then if 10 and f 2 = 0, H(s) has all of its poles

in the open left-half plane and the system is asymptotically
stable. However, for f 1 = 10 and f 2 = 10, the closed-
loop system H(s) is unstable, as shown in Figure 5.3 .
20
10 unstable
asymptotically
stable
10 20
Fig. 5.3
This situation could have been predicted by examining the

McMillan form of Q(s) which is
87
(s+1)(s+3)
M(s) = (5.7)
0
where the poles of the McMillan form are referred to as the

"poles of the system" and the zeros of the McMillan form are
referred to as the "zeros of the system". If any of the
"zeros" of Q(s) lie in the closed right-half plane, then
it will not be possible to set up m high gain control loops
around this system.
Consider now the gain space for the system described by
s-1 5s+l
(s+1)2 (s+1) 2
Q(s) = (5.8)
-1 s-1
(s+1)' (s+1)'
which is shown in Figure 5.4 Now, up to a certain point,

increase of gain in one loop allows increase of gain in the
other loop without instability. The McMillan form of this
latter system Q(s) is
M(s) = (5.9)
s+2
s+1
which implies that, despite the non-minimum phase terms in

the diagonal elements of Q ( s ) , no non-minimum phase behav-
iour will be obtained with the feedback loops closed. It is
also interesting to note that this system is stable with a
small amount of positive feedback. However, this is not a
high integrity system, since failure of any one loop may put
the system into an unstable region of operation.
88
10
asymptotically
stable
-A -2 8 10 f1
Fig. 5.4
5.3 Stability
K(s) G(s) L(s)
F(s)
Fig. 5.5
Consider the system shown in Figure 5.5, in which all mat-

rices are mxm; this last condition is easily relaxed (see
Reference 1, p. 131 ff). Let Q = LGK, and suppose that
G arises from the system matrix
T G (s) O G (s)
P G (s) = (5.10)
-V G (s)
89
and similarly for K(s), L(s), F(s). For generality, we do

not require any of these system matrices to have least order.
The equations of the closed-loop system can then be written
as
T
K U
K 0 0 0 0 0 o • 0 0
V W 0 0 0 0 0 ' 0 -e 0
K K
0 0 T
G U
G
0 0 0 0 ' 0 iG 0
0 0 -vG W
G 0 -Im 0 0 f
0 -u 0
0 0 0 0 T
L U
L 0 o • 0 = 0 (5.11)
0 0 0 0 -vL W
L
0
~Im '
0 -y 0
0 0 0 0 0 0 T U 0 0
F F '
0 0 0 0 0 ~VF W
F f
-Jm
-z 0
0 0 0 -v -z
Here the matrix on the left-hand side of equation (5.11) is

a system matrix for the closed-loop system shown in Figure
5.5, which we can also write as
T
K
(5.12)
-VT. w,H J
The closed-loop system poles are the zeros of |T H (s)|, and
Rosenbrock has shown that
|TH(s) |lm+Q(s)F(s) | |TL(s) | |TQ(s) | |TR(s) | |Tp(s)

(5.13)
As the zeros of |T ||T G ||T K ||T p | are the open-loop poles,

we need only the following information to investigate stab-
ility,
(i) The rational function |I m +Q(s)F(s)|

(ii) The locations of any open-loop poles in the
90
closed right half-plane (crhp).
Notice that this result holds whether or not the subsystems

have least order. If we apply Cauchy's theorem, (ii) can be
further reduced: all we need is the number pQ of open-
loop poles in the crhp.
From equation (5.5)
|lm+QF| = |Q(s)|/|H(s)| (5.14)
The Nyquist criterion depends on encirclements of a critical

point by the frequency response locus of the system to indi-
cate stability. Let D be the usual Nyquist stability con-
tour in the s-plane consisting of the imaginary axis from
-jR to +jR, together with a semi-circle of radius R in the
right half plane. The contour D is supposed large enough to
enclose all finite poles and zeros of |Q(s)| and |H(s)|,
lying in the closed right half plane.
Let |Q(s)| map D into rQ/ while |H(s)| maps D into

r^,. As s goes once clockwise around D, let r~ encircle
« Q
the origin N^ times clockwise, and let ru encircle the
origin NH times clockwise. Then, if the open-loop system
characteristic polynomial has pQ zeros in the closed right
half plane, the closed-loop system is asymptotically stable
if and only if
NH - NQ = pQ (5.15)
This form of the stability theorem is directly analogous to

the form used with single-input single-output systems, but
is difficult to use since |H(S)| is a complicated function
of Q (s), namely
= |Q(s)|/|I+QF| (5.16)
Equation (5.5),
H.(s) = (I+Q(s)F)"1Q(s)
shows that the relationship between the open-loop system

91
Q(s) and the closed-loop system H(s) is not simple. How-

ever, if Q" (S) exists then
H" 1 (s) = F + Q" 1 (s) (5.17)
which is simpler to deal with. Instead of H""1 (s) and

1 1
Q" (s) we shall write H(s) = H " (s) and Q(s) = Q* 1 (s) .
Then, the h..(s) are the diagonal elements of H" 1 (s). In
" -1 -1
general, n ii^ s ^ ^ h i i ( s ) , where h^.fs) is the inverse of
the diagonal element h..(s) of H(s).
We shall, in what follows, develop the required stability
theorems in terms of the inverse system matrices. Also, we
note that if K(s) has been chosen such that Q(s) =
G(s)K(s) is diagonal and if F is diagonal, then we have m
single loops. However, several objections to this approach
can be made. In particular, it is an unnecessary extreme.
Instead of diagonalising the system, we shall consider the
much looser criterion of diagonal dominance.
5.3.1 Diagonal dominance
A rational mxm matrix Q(s) is row diagonal dominant on D

if
|qii<s>| > I |q ± i (s)| (5.18a)

11 13
j
for i = l,...,m and all s on D. Column diagonal domin-

ance is defined similarly by
ni
|q ± i (s)| > I |q i ± (s)| (5.18b)
The dominance of a rational matrix Q(s) can be determined

by a simple graphical construction. Let qi:L(s) map D
into f- as in Figure 5.6 . This will look like an
inverse Nyquist plot, but does not represent anything dir-
ectly measurable on the physical system. For each s on D
92
draw a circle of radius
(5.19)
centred on the appropriate point of q..(s), as in Figure

5.6 . Do the same for the other diagonal elements of Q(s).
If each of the bands so produced excludes the origin, for
i = l,...,m, then Q(s) is row dominant on D. A similar
test for column dominance can be defined by using circles of
radius
m
d i (s) = I |q-(s) (5.20)
I
\
L
-@-
\
\
)
\ /
V if
\ /
\ d
\
Fig. 5.6
5.3.2 Further stability theorems
If Q(s) is row (or column) dominant on D, having on it

no zero of 10 (s) | and no pole of q.^s), for i = l,...,m
then let ^(s) map D into r^ and |6(s)| map D into
?„. If f ^ encircles the origin N^ times and fQ encir-
cles the origin N^ times, all encirclements being clock-
wise, then Rosenbrock has shown that
93
NQ = I N± (5.21)
A proof of this result is given in the Appendix.
Let Q(s) and H(s) be dominant on D, let q i ± (s) map

n s ma D
D into r and let ii( ) P into t Let these
encircle the origin N . and N, . times respectively. Then
with pQ defined as in (5.15), the closed-loop system is
asymptotically stable if, and only if,
m A m A
J i Nqi " J i Nhi = Po (5 22)

'
This expression represents a generalised form of Nyquist's
stability criterion, applicable to multivariable systems
which are diagonal dominant.
For |Q|, NQ = - I Nq± (5.23)
and since we are considering inverse polar plots stability

is determined using
NR - N = - pQ + pc (5.24)
where p = 0 for closed-loop stability.
Hence, Nn - Nu = p^ (5.25)
y n o
replaces the relationship used with direct polar plots, given
by equation (5.15).
5.3.3 Graphical criteria for stability
If for each diagonal element q i ; L (s), the band swept out

by its circles does not include the origin or the critical
point (-f.^,0), and if this is true for i = l,2,...,m,
then the generalised form of the inverse Nyquist stability
criterion, defined by (5.25), is satisfied. In general,
94
the q..(s)
do not represent anything directly measurable
11
3
on the system. However, using a theorem due to Ostrowski
h^(s) = h ^ ( s ) - f± (5.26)
is contained within the band swept out by the circles cen-

tred on q..(s), and this remains true for all values of
gain fi in each other loop j between zero and f-d .
Note that h7"f(s) is the inverse transfer function seen
between input i and output i with all loops closed. The
transfer function h.(s) is that seen in the ith loop when
this is open, but the other loops are closed. It is this
transfer function for which we must design a single-loop
controller for the ith loop.
The theorems above tell us that as the gain in each other

loop changes as long as dominance is maintained in the other
loops, h7.(s) will also change but always remains inside
—1
the ith Gershgorin band. The band within which h.^ (s) lies
can be further narrowed. If Q and H are dominant, and if
d. (s)
<i>. (s) = max ^ (5.27)
3 |f i +q i i (s)|
3 3D
then h^ (s) lies within a band based on q.^(s) and def-

ined by circles of radius
r ± (s) = • i (s)d i (s) (5.28)
Thus, once the closed-loop system gains have been chosen

such that stability is achieved in terms of the larger bands,
then a measure of the gain margin for each loop can be deter-
mined by drawing the smaller bands, using the 'shrinking
factors1 < >
$ ±{s) defined by (5.27), with circles of radius
r^. These narrower bands, known as Ostrowski bands, also
reduce the region of uncertainty as to the actual location
of the inverse transfer function h7.(s) for each loop.
95
5.4 Design Technique
Using the ideas developed in the previous section, the

frequency domain design method proposed by Rosenbrock con-
sists essentially of determing a matrix K (s) such that
the product [G(s)K (s)]" is diagonal dominant. When this
K
condition has been achieved then the diagonal matrix ^(s)
can be used to implement single-loop compensators as required
to meet the overall design specification. Since the design
is carried out using the inverse transfer function matrix
then we are essentially trying to determine an inverse pre-
compensator K (s) such that Q(s) = K (s)G(s) is diagonal
dominant. The method is well suited to interactive graph-
ical use of a computer.
One method of determining K (s) is to build up the re-

quired matrix out of elementary row operations using a
graphical display of all of the elements of Q(s) as a
guide. This approach has proven successful in practice and
has, in most cases considered to-date, resulted in K (s)
being a simple matrix of real constants which can be readily
realized.
Another approach which has proved useful is to choose

K = G(o), if |G(o)| is nonsingular. Here again K (s)
is a matrix of real constants which simply diagonalizes the
plant at zero frequency.
For example, Figure 5.7 shows the inverse Nyquist array

(INA) of an uncompensated system with 2 inputs and 2 outputs.
-i-
Fig. 5.7
96
An inspection of this diagram shows that element 1,2 is

larger than element 1,1 at the maximum frequency considered,
and similarly element 2,1 is very much larger than element
2,2 over a wide range of frequencies . Thus, the open-loop
system is not row diagonal dominant, nor is it column dia-
gonal dominant. However, by choosing K = G(o) the result-
ing INA of Q(jw) is shown in Figure 5.8 with the Gershgorin
circles for row dominance superimposed. Both q,1 and q22
are made diagonal dominant with this simple operation. The
dominance of element q22 of this array can be further
improved by another simple row operation of the form
R = R + aR
2 2 l (5.29)
where a - 0.5, in this case.
Fig. 5.8
A further approach, which is perhaps more systematic than

those mentioned above, is to determine K as the 'best 1 ,
in a least-mean-squares sense, wholly real matrix which most
nearly diagonalizes the system Q at some frequency s = ju>,
(Rosenbrock and Hawkins ) . This choice of K can be con-
P
sidered as the best matrix of real constants which makes the
sum of the moduli of the off-diagonal elements in each row
97
of Q as small as possible compared with the modulus of the

diagonal element at some frequency s = ju>. The development
of the two forms of this latter approach is given in the
following paragraphs.
Consider the elements q.^ in some row j of Q(jw) =

KG(jw), i.e.
(5.30)
(5.31)
Now choose rkjm so that
m
(5.32)
I
k=l
is made as small as possible subject to the constraint that
m
(5.33)
ji
Using a Lagrange multiplier, we minimize
m
k=l
(5.34)
m m 2 fm 2
l I k .1 .a l- J ic..31
li=l -*1 1
1
k=l i=l -* i=l -
(5.35)
and taking partial derivatives of 4> . with respect to k_.

(i.e. the elements of the row vector k . ) , we get, on
setting these equal to zero,
—Y> — =
I
k^3
2
fm
'i-l
^ji a ik a
)
• ^ k
j
- 2Xk..
(5.36)
= 0 for SI =
98
Now writing
m
(ct (5.37)
J.
k=l il
ik 6 *k
and (5.38)
the minimization becomes
m m
A.k. - xk^ = 0 (5.39)
since I |q 1k <j«)
k=l
= X (5.40)
Thus, the design problem becomes an eigenvalue/eigenvector

problem where the row vector k., which pseudodiagonalizes
row j of Q at some frequency ju>, is the eigenvector of
the symmetric positive semi-definite (or definite) matrix A.
corresponding to the smallest eigenvalue of A. .
Figure 5.9 shows the INA of a 4-input 4-output system

over the frequency range 0 -> 1 rad/sec. Although the Gersh-
gorin circles superimposed on the diagonal elements show
that the basic system is diagonal dominant, the size of
these circles at the 1 rad/sec end of the frequency range
indicate that the interaction in the system may be unaccept-
able during transient changes. Using the pseudodiagonalisa-
tion algorithm described above, at a frequency of 0.9 rad/
sec in each row, a simple wholly real compensator K can
be determined which yields the INA shown in Figure 5.10 .
Here, we can see that the size of the Gershgorin discs has
in fact been considerably reduced over all of the bandwidth
of interest.
99
F i g . 5.9
+ +
+ +
F i g . 5.10
100
However, in general, we may choose a different frequency

a) for each row of K, and it is also possible to pseudo-
diagonalize each row of Q at a weighted sum of frequencies.
The formulation of this latter problem again results in an
eigenvalue/eigenvector problem (see Rosenbrock ) .
Although this form of pseudodiagonalization frequently

produces useful results, the constraint that the control
vector k. should have unit norm does not prevent the dia-
gonal term q.. from becoming very small, although the row
is diagonal dominant, or vanishing altogether.
So, if instead of the constraint given by (5.33), we sub-

stitute the alternative constraint that
|qj:JO)| = 1 (5.41)
then a similar analysis leads to
AjkT - AEjkT = 0 (5.42)
where A. is as defined by equation (5.37) and E. is the

symmetric positive semidefinite matrix
= [a. .a .+6. .3 .] (5.43)
Equation (5.42) now represents a generalized eigenvalue pro-

blem, since E. can be a singular matr:
matrix, and must be solved
using the appropriate numerical method.
5.5 Conclusions
The inverse Nyquist array method offers a systematic way

of achieving a number of simultaneous objectives, while still
leaving considerable freedom to the designer. It is easy to
learn and use, and has the virtues of all frequency response
methods, namely insensitivity to modelling errors including
nonlinearity, insensitivity to the order of the system, and
101
visual insight. All the results for inverse plots apply with
suitable changes to direct Nyquist plots (see Reference 1,
pp. 174-179). We can also obtain multivariable generalisa-
tions of the circle theorem for systems with nonlinear, time-
dependent, sector-bounded gains ' . Not only do the Gersh-
gorin bands give a stability criterion; they also set bounds
on the transfer function hi:L(s) seen in the ith loop as the
gains in the other loops are varied.
Several further ways of determining K (s) such that

Q(s) is diagonal dominant are the subject of current re-
search (see Leininger ) . However, several industrial multi-
variable control problems have already been solved using this
design method 7 ' 8 ' 9 ' 1 0 .
References
1. ROSENBROCK, H.H. : 'Computer Aided Control System

Design 1 , Academic Press, London, (1974).
2. MUNRO, N. and B.J. BOWLAND : 'The UMIST Computer-
Aided Control System Design Suite 1 , User's Guide,
Control Systems Centre, UMIST, Manchester, England, 1980.
3. OSTROWSKI, A.M. : 'Note on Bounds for Determinants
with Dominant Principal Diagonal', Proc. Am. Math. Soc.
3, (1952), pp. 26-30.
4. HAWKINS, D.J. : 'Pseudodiagonalisation and the Inverse
Nyquist Array Methods', Proc. IEE, vol. 119, (1972),
pp. 337-342.
5. COOK, P.A. : 'Modified Multivariable Circle Theorems',
in Recent Mathematical Developments in Control (Ed.
D.J. Bell), Academic Press, (1973), pp. 367-372.
6. LEININGER, G.G. : 'Diagonal Dominance for Multivariable
Nyquist Array Methods Using Function Minimization',
Automatica, vol. 15, (1979), pp. 339-345.
7. MUNRO, N. : 'Design of Controller for an Open-Loop
Unstable Multivariable System Using the Inverse Nyquist
Array', Proc. IEE, vol. 118, No. 9, (1972),
pp. 1377-1382.
8. WINTERBONE, D.E., N. MUNRO and P.M.G. LOURTIE
'Design of a Multivariable Controller for an Automotive
Gas Turbine', ASME Gas Turbine Conference, Washington,
Paper No. 73-GT-14, 1973.
9. MUNRO, N. and S. ENGELL : 'Regulator Design for the
F100 Turbofan Engine', IEEE Int. Conf. on Control and
its Applications, Warwick University, England, (1981),
pp. 380-387.
102
10. KIDD, P.T., N. MUNRO and D.E. V7INTERB0NE : 'Multi-

variable Control of s Ship Propulsion System 1 , Proc. of
Sixth Ship Control Systems Symposium, Ottowa, Canada,
October, 1981.
APPENDIX
To prove the result given by equation (5.21), we note that

by (5.18a and b) and the conditions of the theorem there is
no pole of q.. on D, i,j = l,2,...,m, nor is there any
zero of q.^ on D, i = l,2,...,m, because D is a com-
pact set and q.. > 0 on D. Moreover, by Gershgorin's
theorem, there is no zero of |o| on D. Let Q(a,s) be
the matrix having
(5.44)
q . (a,s) = aq ± . (s) , j/i
where q..(s), q..(s) are the elements of Q(s) and

0 JL a 1 *• Then every element of Q(a,s) is finite on D,
and so therefore is |6(a,s)|. Consider the function
(Of8) = |Q(«,s)[ (5.45)

n q..(s)
1X
which is finite for 0 <_ a <_ 1 and all s on D, and which

satisfies 3(0,s) = 1. Let the image of D under 6(1**)
be r. Let the image of (0,1) under $(',s) be y
s
Then y is a continuous curve joining the point $(0,s) = 1
to the point 3(l,s) on r. As s goes once round D, y
sweeps out a region in the complex plane and returns at last
to its original position.
Suppose, contrary to what is to be proved, that r encir-
cles the origin. Then the region swept out by y as s
s
goes once round D must include the origin. That is, there
is some a (0,1) and some s on D for which B(a,s) = 0.
But the q.^ are all finite on D, so by (5.45) |Q(a,s)|=0.
103
By Gershgorin's theorem this is impossible. Hence the num-

ber of encirclements of the origin by r is, from (5.45),
m
0 = N - I N± (5.46)
±
which is (5.21).
104
The Inverse Nyquist Array Design Method - Problems
P.I A plant described by
1
s+1
G(s) = 1
s-1
has arisen from a system with
|TG| = (s-Dts+i)'
If feedback is to be applied to this system through a feed-

back matrix F = diag{l,2}, determine whether or not the
resulting closed-loop system is stable.
Comment on cancellations occurring in the formulation of

G(s), and on the encirclements obtained in the resulting
INAs.
{Acknowledgement for this problem and its solution are

hereby made to Professor H.H. Rosenbrock of the Control
Systems Centre, UMIST.}
P.2 Given a system described by the transfer-function matrix
1 1
i+T s+1
G(s) =
l 1
s+2 I+Z
(i) State one reason why the inverse systems G~ (s) and
H (s) are used in Rosenbrock's inverse Nyquist array
design method,
(ii) Sketch the inverse Nyquist array for G~ (s), and
comment on the diagonal dominance of the uncompensated
system,
(iii) Determine a wholly real forward path compensator K
such that the closed-loop system H(s), with unity
feedback, is decoupled.
105
(iv) Introduce integral action into both control loops and

sketch the resulting root-locus diagrams for the final
loop-tuning of the compensated system.
Chapter 6
Analysis and design of a nuclear boiler

control scheme
Dr. F. M. HUGHES
Synopsis
The steam raising plant of a British designed nuclear power

station for which the installed control scheme has given rise
to performance and stability problems is considered. The lec-
ture is based on studies carried out for the plant which ill-
ustrate the way in which multivariable frequency response
methods can be used to analyse and identify the source of
control problems and further enable alternative control
schemes, having improved performance, to be designed.
6.1 Introduction
The role that multivariable frequency response methods can

play in the analysis and design of control schemes for power
plant is demonstrated. The particular plant dealt with com-
prises the drum boiler steam generating units of a nuclear
power station for which the existing conventional feed con-
trol scheme has suffered from performance and stability pro-
blems. Although the boiler plant concerned is non-linear in
nature, a great deal of useful information regarding dynamic
behaviour can be obtained from the study of a linearised
model. Under closed-loop control conditions, where the ex-
cursions in plant variables are held within tight bounds, a
small variation, linearised model can give a realistic repre-
sentation of controlled behaviour at a particular operating
condition.
Using a suitable transfer function matrix model, the exis-
ting scheme is analysed and the reasons for its poor stabil-
ity characteristics are quickly established via the multi-
variable frequency response approach. The design methods
based on this approach are further used to devise an altern-
107
ative control scheme possessing improved stability character-

istics and performance capabilities. Although one particular
power plant is dealt with, the analysis and design approaches
adopted are applicable to power station drum boiler control
in general.
6.2 Drum Boiler Plant and its Control Requirements
steam generator unit
Fig. 6.1 Schematic Diagram of Plant.
A schematic block diagram of a nuclear station drum boiler

is shown in Fig. 6.1 . The feedpump supplies water to the
drum via the feedwater regulating valves. From the drum
water enters the evaporator loop and returns as 'wet1 steam
to the drum. Then steam from the drum passes through the
superheater, emerging as superheated (dry) steam for the
main turbine supply. The heat energy input to the steam gen-
108
erators is obtained from the circulating coolant which trans-

fers heat energy from the reactor. Pith the reactor tempera-
ture levels controlled the heat energy input can be varied
by changing the coolant flow rate through adjustment of the
coolant pump speed (n ) . Feed flow into the drum can be
changed both by adjustment of the position of the feedwater
regulating values (y^ ) and by adjustment of the steam in-
put into the feedpump turbine (y^ ) . The output of steam to
the turbine, which determines the turbine generator power
output, is controlled by adjustment of the position of the
turbine governor valves (y ).
It is important to be able to control the level of water in

the drum since disastrous consequences can result from the
drum overfilling or running dry. The drum water level will
remain steady provided a balance exists between feed supply
and evaporation rate. Obviously when the station power out-
put changes, in order to maintain the drum level within app-
ropriate bounds, the feed flow must be adjusted to match the
new steam delivery rate to the turbine. Under transient
conditions, however, feed adjustment alone may not provide
adequate drum level control. This is because a substantial
fraction of the water in the drum and evaporator loop is at
saturation conditions, and can be readily flashed off to
steam during a fall in steam pressure. Evaporation rate
and therefore drum level variations are therefore strongly
influenced by transient pressure variations. Control over
steam pressure can therefore greatly assist the effective
control of drum water level. In the plant considered the
main variables which need to be controlled are therefore the
drum water level and the steam pressure .
Although not shown in Figure 6.1, the power plant comprises

3 steam generating units, each with its own drum, evaporator,
superheater and separately controlled coolant flow. The feed
to the drums is supplied by a common feedpump through indi-
vidual feed control valves. Also, the three superheaters
feed into a common steam header to provide a single supply
to the main steam turbine.
109
6.3 Multivariable Frequency Response Methods
A linear multivariable control scheme can be represented

most conveniently for frequency response analyses in terms
of the block diagram of Figure 6.2 . The plant is modelled
by the transfer function matrix G(s) which relates its con-
trollable inputs to its measurable outputs and the plant
control scheme is represented by the precontroller matrix
K(s). The system open loop transfer function matrix is given
as Q(s) = G(s)K(s) .
Yrcf(s) E(s) U(s) Y(s)

K(s) G(s)
Fig. 6.2 Plant and controller representation.
If G(s) happens to be diagonal then single loop stability

assessment and design methods can be applied to each diagonal
element in turn and the appropriate individual loop compen-
sation needed to satisfy the specified performance thereby
determined. In general, however, G(s) will not be diagonal
and the off diagonal elements which give rise to the inter-
action between the loops make analysis on an individual loop
basis most unreliable. Alternative means are therefore re-
quired .
The system shown in Figure 6.2 has the closed-loop system

transfer function R(s) given by:
adj(I+Q(s)).Q(s) (6.1)
det(I+Q(s))
The poles of the closed-loop system, which determine stab-

ility, are the values of s for which |R(s) | -*- » and
110
therefore the values of s for which det(I+Q(s))= 0.
If the det(I+Q(s)) is mapped as s traverses the infinite

radius semicircular contour in the right half s plane/ then
the Nyquist stability criterion can he applied to determine
closed-loop stability from the number of encirclements that
det(I+Q(s)) makes about the origin. Although overall system
stability is assessed information regarding the stability
margins of individual loops is not provided.
An alternative approach to stability assessment is provided

by re-expressing Q(s) for any complex value of s, in terms
of its eigenvalues and eigenvectors':
Q(s) = U(s)A(s)u"1(s) (6.2)
where A(s) is a diagonal matrix of characteristic transfer

functions (diagonal elements A.(s), i = 1,2,...,m for an
m-input m-output control system) and U(s) is a matrix of
characteristic direction vectors.
It can readily be shown that:
det(I+Q(s)) = det(I+A(s)) (6.3)
= (1+A1(s)) (1+X2(s) (...) (1+A^(s)) (6.4)
Hence overall stability can be assessed by applying the

Nyquist stability criterion to each characteristic transfer
function. If the net encirclements of the characteristic
loci (i.e. the mappings of the characteristic transfer func-
tions as s traverses the infinite semicircular contour) ind-
icate stability then the overall closed loop system will be
stable.
If the off-diagonal elements of Q(s) are relatively small

then the characteristic transfer functions are strongly rela-
ted to the diagonal elements of Q(s) and the individual
loop transferences. The characteristic loci then give a use-
ful indication of individual loop compensation requirements.
Since a desirable control scheme characteristic is usually

Ill
that it provides low loop interaction, an attempt to reduce

interaction (i.e. the reduction of the off-diagonal elements
of 0(s)) is normally the initial step in the design of the
precontroller matrix K(s). Matrix K(s) performs column
manipulations on G(s) and the manipulations required can be
deduced from the frequency response array G(jw). Provided
interaction has been made sufficiently small the individual
control loop compensation requirements can be assessed and
appropriate compensators designed.
However, in many systems it may prove very difficult to

make all off diagonal elements of Q(s) (or Q~ (s) if the
inverse Nyquist array method is being used) small, and its
attainment may result in a more complex control system than
is strictly necessary for satisfactory control. Also, under
some circumstances a certain level of interaction may be quite
tolerable, so that all the off diagonal elements of Q(s) may
not need to be small. The Q(s) matrix may well have large
off diagonal elements, and the diagonal elements still approx-
imate closely to the system characteristic transfer functions.
This is obviously true when Q(s) is a triangular matrix. In
such circumstances individual loop stability margins and com-
pensation requirements can be assessed from the diagonal ele-
ments of Q(s) despite interaction levels being high for cer-
tain disturbances.
6.4 Analysis and Design Approach Adopted
In the drum boiler studies reported, the stability charac-

teristics of the installed control scheme were assessed from
its characteristic loci, and alternative schemes designed
using the direct Nyquist array approach. The design aim for
the alternative scheme was to determine a simple and effective
control structure which provided good overall performance and
improved stability characteristics. Simplicity of structure
was ensured by restricting the precontroller for interaction
reduction to a matrix of constants. Also, in the light of
previous experience, rather than attempt to drastically re-
duce all off diagonal elements of Q ( s ) , the function of the
112
precontroller was made that of reducing enough off diagonal

elements to ensure that the system characteristic loci were
closely approximated by the diagonal elements of the result-
ing open loop frequency response array.
6.4.1 Transfer function model

The most convenient and flexible form of plant representa-
tion for use with multi-input multi-output frequency response
methods is a matrix transfer function model, which relates
controllable inputs and disturbances to measurable outputs.
To simplify the plant representation for this analysis of
fundamental dynamic characteristics and control requirements,
station operation with symmetrical boiler conditions was
considered. With all three steam generators having identical
operating conditions, a single equivalent steam generator can
be modelled. Using a comprehensive non-linear model of the
plant the measurable system output; responses were predicted
for step changes in the measurable inputs.
The plant inputs considered were:
(i) Coolant pump speed n
(ii) Feedwater regulating value position y f
(iii) Feedpump turbine steam input control y^
(iv) Turbine governor valve position y
The measurable outputs were:

(i) Steam pressure upstream of turbine stop valves p.
(ii) Drum water level £ ,
wci
(iii) Feedvalve pressure drop Ap
Having obtained step responses the equivalent plant trans-
fer functions were derived using an interactive identifica-
tion package with graphical display facilities. The package
incorporated a model adjustment technique and minimised an
integral of error squared criterion to match equivalent
transfer function responses to the given plant response data.
The transfer function model obtained was of the following
113
form:
g
ll (s) g
l2 (s) g
13 (s) n s (s) g 1 4 (s)
g
21 (s) g
22 (s) g
23 (s) g24(s) (6,5)
g
31 (s) g
32 (s) g
33 (s) '34'
Variables ng, yf and yf are the controllable inputs to
the plant, whilst y. represents a disturbance input. The
response of the plant following a disturbance in y is
representative of its performance following a change in
station power output demand.
6.4.2 Analysis of existing control scheme

The existing control scheme, in terms of a single equiva-
lent steam generating unit, is shown in Figure 6.3, and can
be seen to consist of separate loops, (i.e. no interconnect-
ing signals are employed). As a consequence of this and the
loop philosophy adopted, the feed regulating valves had to
be chosen to control drum level so that individual drum level
control is provided under asymmetric operating conditions.
The actual steam raising plant consists of three drums each
with its associated feedvalve for individual feed flow adjust-
ment. However, if realistic valve openings are to be main-
tained, adjustment of feedvalve position alone has little
effect on feed flow since an uncontrolled boiler feedpump
acts as an almost constant flow device. In order to provide
the substantial changes in feed flow needed for effective
drum level control the feedvalve positional changes need to
be backed up by changes in feedpump pressure. The required
back up control is provided by a control loop which monitors
changes in pressure drop across the feed valves and adjusts
the feedpump to maintain at least a minimum differential
pressure across each of the three feed regulator valves.
Constant feed valve differential pressure makes feed flow
approximately proportional to feed valve position.
Good control over drum level therefore requires a fast
114
acting feedpump control loop in addition to feedvalve adjust-

ment. In the existing scheme the feedpump is controlled by
adjusting the actuator which alters steam inlet flow to the
feedpump turbine. This actuator is positioned by an open
loop stepping motor which effectively introduces an integral
term into the control loop.
Ptref J
5—- Ms)
Vfv
*v$—- k 2 (s) G(s) l
WD
A pv ref
> Att Att
&—-
t
k 3 (s)
80s
Fig. 6.3 Existing Feed Control Scheme.
The heat input to the steam generating unit can be modi-

fied by adjustment of the coolant flow via the coolant pump
speed, and is used to ensure control over the steam pressure
The open loop frequency response of the existing system,
expressed in the form of a Nyquist Array is shown in Figure
6.4 . The individual loop compensators used were:
ki (S, = (6.6)
k 2 (s) (6.7)
= .114 (6.8)
It is immediately obvious from the array that the system is

not diagonally dominant, and in particular diagonal element
q22 is very small, being dwarfed by several of the off dia-
gonal terms. This feature was to be expected since, as
previously discussed, the alteration of feed valve position-
115
alone has little effect on the boiler feed flow. The over-
all system stability characteristics cannot readily be ass-
essed from an observation of this array; the off-diagonal
elements are far too large for the diagonal elements to pro-
vide a useful guide to stability and the characteristic loci
were calculated explicitly.
20
*- *
M22 H
-2 23
-15
30
q32 ^33
Fig.6.4 Nyquist Array of Existing Control Scheme.
The characteristic loci studies indicated that the para-

meter which had the most pronounced effect on system stab-
ility was the gain k~. Two of the characteristic loci
inferred wide stability margins and changed only slightly
with variations to k_. The third locus, however, changed
drastically in both size and shape, and Figure 6.5 shows how
system stability is impaired as is increased, with insta-
bility predicted at k 3 = 0.4 . Instability of the oscill-
atory nature indicated has also been experienced on the plant
itself. In addition the detrimental interaction between the
loops which makes system stability sensitive to k^ will
also make stability and performance sensitive to changes in
the operating conditions.
116
A3
k =0.114
Fig. 6.5 Characteristic Locus Variation with Gain k-.
A further feature which aggravates the stability of the

existing scheme is the integral term introduced into the
third loop by the boiler feedpump turbine actuator. This in-
curs an additional 90° of phase lag, obviously reducing the
loop stability margin and adding to the difficulty of provid-
ing the stable, fast acting control loop needed to back up
feed valve adjustments. This problem can be alleviated by
introducing positional feedback into the actuator drive thus
converting the actuator characteristic from 1 / T S to 1 / ( 1 + T S ) .
This lag transfer function characteristic provides a well
defined, finite steady state gain and the faster the actuator
drive is made, the lower is the effective time constant x
and the phase shift introduced.
6.4.3 Alternative control scheme design
As inferred in the previous section, the problems with the

existing scheme are two-fold. The problem caused by the
integral action characteristic of the stepper motor drive of
the boiler feedpump turbine actuator can be alleviated by
incorporating positional feedback. In the actuator position
control loop so formed, a gain giving rise to a transfer
117
function of l/l+10s was considered realistic and used in

the design studies. This change on its own, however, would
still leave the problem of interaction between the loops un-
solved, making it difficult to achieve the required response
characteristics and stability margins over a wide operating
range by means of individual loop compensation.
The design studies took as starting point the plant trans-

fer function model with the modified feedpump turbine actua-
tor characteristic, giving rise to the plant open loop fre-
quency response array shown in Figure 6.6.
1000
-100
-500
plant
compensated
921 923
IFôo
-1000
9
31 932
-400
Fig. 6.6 Plant open loop frequency response array.
Various schemes were designed of varying complexity and

performance characteristics , but only one particular scheme
will be dealt with here. This scheme is considered to pro-
vide good stability and performance capabilities and poss-
esses a simplicity which makes it readily extendible to the
three drum version needed for implementation on the plant
itself.
In designing an alternative control scheme the initial aim

was to reduce interaction levels and the adopted criterion
118
was that the plant input which most influenced a given output
signal should be used for its control. In the open loop fre-
quency response array of Figure 6.6, the largest element of
row 2 is 9o3' implying that drum level control is influenced
most strongly by the boiler feedpump turbine (BFPT) actuator
adjustment. Thus the most effective way of controlling drum
level is via the feedpump turbine, rather than the feedwater
regulator valves.
The required control configuration can be achieved by a

precontroller matrix which swaps over column 2 with column 3,
so that the large element g23 becomes a diagonal element in
the open loop transfer function matrix.
The off diagonal q~2 of the resulting array has a sub-

stantial steady state value, but this can be reduced by add-
ing a factor of 1.2 times column 3 to column 2. It is also
possible to reduce element q ? by adding a factor of column
1 to column 2. However, the reduction achieved is mainly
effective at low frequencies and since integral control was
to be employed the improvement in transient performance was
not considered sufficient to justify the increased complexity
involved. The resulting frequency response array is also
shown in Figure 6.6, where it can be seen that the elements
q 2 -, q 3 1 # q.,3 and q 2 ~ are very small compared with the re-
maining elements. If these small elements are denoted by e
then the determinant of the return difference matrix which
determines stability is given by
Si12
det(I+Q(s)) = e 1+q
32 e (6.9)
q32 i+q 3 3
i.e.
det(I+Q(s)) - (l+q1]L) (l+q 22 ) (l+q 33 ) for e small (6.10)
Although the open loop frequency response array does have

substantial off diagonal elements in q 2 and 332' these
have negligible effect on closed-loop stability. The system
characteristic loci are approximated closely by the diagonal
elements of Q(jw) and individual loop stability character-
119
istics and compensation requirements can be deduced from

these.
The shape of the diagonal elements of the frequency res-

ponse array indicated that the following individual loop con-
trollers would provide good stability margins.
kl(s) = O.OKI (6.11)
k.2(s) = .05(1 (6.12)
k 3 (s) = .04(1+ (6.13)
k,(s)
l
WD
G(s)
Fig. 6,7 Improved Control Scheme Configuration.
The resulting overall control scheme of Figure 6.7 has the

characteristic loci shown in Figure 6.8, where it is obvi-
ous that much higher loop gains than those proposed could
have been employed and reasonable stability margins still
retained. On a linear system basis these higher gains would
provide faster system responses. However, the system control
actuators have physical limitations with respect to the amp-
litude and rate of change of output variations. Consequently,
the relatively modest gains proposed were chosen to ensure
that the demands made of the actuators are not unrealistic-
120
ally high.
f 20 30 40
-20
-30
10 15
-10L
-2
-2
Fig, 6.8 Characteristic Loci of Improved Control Scheme.
6.5 Performance Assessment of Improved Scheme
The improved control scheme was tested by simulation using

the linear transfer function model, and provided stable, ad-
equately fast and well damped performance following changes
in the reference levels of the control loops. A more realis-
tic disturbance in practice is provided by a change in the
position of the turbine governor control valves y . Such
disturbances occur during operation when the turbine governor
gear responds to changes in the electricity system grid fre-
quency. Responses of the designed scheme following a step
change of 0.10 per unit (PU) in y were computed, and the
variation in steam pressure p., drum level 1 ,, and dif-
ferential pressure drop Ap are maintained well within
acceptable bounds as indicated by the responses of Fig. 6.9.
The non-minimum phase water level response due to steam flash-
ing is well illustrated. The magnitudes of the actuator
121
demands were examined and observed to be acceptable.
200
0.6
0.3
Fig. 6.9 System Response Following Step Change in

Turbine Governor Valve position y t g = o.i p.u.
6.5.1 Three drum version of improved scheme
In the design of the improved control scheme it was assumed

that all three generating units possessed identical operating
conditions enabling a single drum equivalent model to be used.
Whilst this is reasonable for assessing fundamental dynamic
characteristics and basic control requirements, any scheme
122
devised must be capable of extension to the three drum case

and provide good control in situations where drum operating
conditions are not balanced .
The designed scheme used the feedpump turbine actuator as

the major influence over drum level variations. In the three
drum version of the scheme the control signal to the feedpump
turbine actuator is derived from the average of the levels in
the drums. The interconnecting signal (of gain .2) which was
introduced from interaction and performance considerations
also serves to provide the necessary facility for individual
drum level control. The drum level signal of each drum pro-
vides a control signal to its associated feed regulating
valve. Transient variations in drum level are controlled
mainly by the feedpump turbine with the feed valves provid-
ing the required trimming action to ensure the correct level
in each drum.
The control over the pressure drop across the feed valves
is not critical^ the main requirement being that reasonable
feed valve openings are maintained. Control merely needs to
be such that the pressure drop across all of the feed regula-
ting valves is above a specified minimum.
6.6 System Integrity
For a system design to be acceptable it should be easy to

commission and resilient to system failures. Specifically/
it should remain stable and capable of being manually con-
trolled, when control element failures occur. This aspect
of system integrity, which is important during commissioning
and in auto/manual changeover, can readily be checked in
terms of the frequency response design approach.
System stability in the event of a controller channel not

functioning can be assessed via the schematic block diagram
of Figure 6.2 by introducing diagonal failure mode matrices
between the controller and plant (F ) and in the feedback
a
path (F,.)
L.
. F^
a
represents the actuator failure mode matrix
and F^ the feedback failure mode matrix. The system open
loop transfer function matrix Q^ is then given by
123
Q i (s) = F f G(s)F a K(s) (6.14)
Under normal operation with the control scheme fully func-

tioning, F = F = I, so that
•L a
Q i (s) = G(s)K(s) = Q(s) (6.15)
If a controller element is disabled, the normal control

signal is not transmitted and this situation can be represen-
ted by setting the appropriate diagonal element in the corr-
esponding failure mode matrix to zero.
Commissioning situations where only certain control loops

are closed can be represented by appropriate choice of Ff,
and manual control situations where the operator adjusts the
actuator rather than the normal control signal can be repre-
sented by appropriate choice of F . If f for example, the
a
feedback channel of loop 1 is open for a 3-input 3-output
system, then
Ff = 0
0 0
1 0
0 (6.16)
0 0 1
System stability is then assessed from either
det(I+Q i (s)) = det(I+F f G(s)F a K(s)) (6.17)
or from the characteristic loci of the open-loop system tran-

sfer function matrix Q.(s).
Such analysis has shown that the designed scheme can with-
stand all combinations of nonfunctioning controller elements
without loss of stability. Consequently, any combination of
automatic and manual control of loops will provide stable
operation and complete freedom of choice over the order of
loop closure is permitted when commissioning.
When a control system element fails during operation its

associated output may change suddenly to its maximum or min-
124
imum value, resulting in a severe disturbance to the system.

Wide and unacceptable variations in plant variables will then
ensue unless appropriate signal freezing is carried out on
detection of the fault. Such emergency action, however, ob-
viously can only be effective if the resulting system is
stable, and this fact can be determined from the integrity
analysis approach outlined.
6.7 Conclusions
When analysing the existing control scheme the multivari-

able frequency response methods quickly enabled the sources
of the control problems to be located and further provided
the means for deriving an alternative scheme having improved
stability and performance capabilities.
In the design of the alternative scheme, strict adherence

to a particular multivariable frequency response method was
not followed, neither was the normal objective of aiming for
a low overall level of interaction. By taking advantage of
the structural properties of the plant matrix model and the
performance requirement of the plant, a simple control scheme
having good stability characteristics was produced. Although
an appreciable, but tolerable, level of interaction exists
in the scheme, the system stability characteristics are dic-
tated by the diagonal elements of the open loop transfer
function matrix which can be used to determine individual
loop stability margins and compensation requirements.
The simple structure of the control scheme, designed using

a single drum equivalent plant model, enables it to be read-
ily extended to the three drum version required for implemen-
tation on the plant itself.
The scheme designed also provides resilience to failures

in that any combination of loops can be on automatic or
manual control at any time, and complete freedom of choice
over the loop closure sequence during commissioning is pro-
vided.
Attention has been confined to control scheme design and

125
analysis via a linear system model, but further simulation

studies have been carried out on a comprehensive non-linear
plant model. In these studies close agreement between lin-
ear and non-linear model performance predictions was obser-
ved, demonstrating that a linearised system representation
can provide a good basis for the design of the basic struc-
ture of a control scheme.
References
1. COLLINS, G.B. and F.M. HUGHES : "The analysis and

design of improved control systems for recirculation
boilers", 8th IFAC World Congress, Kyoto, Japan, 1981.
2. MACFARLANE, A.G.J. : "A survey of some recent results
in linear multivariable feedback theory", Automatica,
vol. 8, No. 4, (1972), pp. 455-492.
3. BELLETRUTTI, J.J. and A.G.J. MACFARLANE :
"Characteristic loci techniques in multivariable control
system design", Proc. IEE, No. 118, (1971),
pp. 1291-1297.
4. ROSENBROCK, H.H. : "Design of multivariable control
systems using the inverse Nyquist array", Proc. IEE,
vol. 117, (1969), pp. 1929-1936.
5. HUGHES, F.M. and G.B. COLLINS : "Application of multi-
variable frequency response methods to boiler control
system design", Proc. 2nd International Conference on
Boiler Dynamics and Control in Nuclear Power stations,
pp. 411-418, British Nuclear Energy Society, London,
1980.
Acknowledgment
The contribution of the UKAEA (Winfrith) to the work

described is gratefully acknowledged, particularly that
of Mr G.B. Collins with whom much of the work was
carried out.
Chapter 7
Optimal control
Dr. D. J. BELL
Synopsis
An overview of dynamic optimization is presented which

commences with a statement of the important conditions from
classical theory including the Weierstrass E-function and
the second variation. The classical results are then re-
formulated in modern optimal control notation which leads
immediately to a statement of Pontryagin's Minimum Principle.
Necessary conditions for optimality of singular trajectories
are deduced from the theory of the second variation. A des-
cription of Dynamic Programming is followed by a study of
the Hamilton-Jacobi approach for the linear system/quadratic
cost problem together with the associated matrix Riccati
equation.
7.1 The Calculus of Variations; Classical Theory
The classical problem of Bolza is that of minimizing the

functional
J = G(t ,x(t ) r t l f x(t 1 )) + f(t,x,x

x)dt
J
o
(where x and x are n-vectors), subject to the differen-
tial constraints
<j>3(t,x,x) = 0 (g = 1,2,...,k < n)
and end conditions
ijjq(to,x(to) ,t 1 ,x(t 1 )) = 0 (q = 1,2,..., <_ 2n+2)
Defining a fundamental function F = f+A (t)<j> (note the

double suffix summation is used throughout this chapter),
where {A3} is a sequence of time dependent Laqranqe
127
multipliers, a number of necessary conditions for a station-

ary value of J can be deduced from the first variation of
J . These are (i) the Euler-Lagrange equations
3F
(7.1)
= 1
'2 n)
(ii) the transversality conditions
3F_
F - x -*-F- dt + dx. + dG = 0
i
3x. 3x.
(7.2)
(iii) the Weierstrass-Erdmann corner conditions
3_F_ 3F
and F - x. (i = 1,2,...,n) (7.3)
3x. ax.
must be continuous across a corner of the extremal (i.e. the
trajectory satisfying the necessary conditions).
(iv) the Weierstrass E-function condition, valid for all

t e (tQ,tl),
E = F(t,x,X,A) - F(t,x,x,A) - (X ± -x i ) 3F (t,x,x,A) > 0

ax.
(7.4)
where a strong variation has been employed such that the
derivative X on the variational curve may differ from the
corresponding derivative x on the extremal by an arbitrary
amount.
Note: If the fundamental function F does not contain the
time t explicitly then a first integral of the Euler-
Lagrange equations (7.1) exists and is
F - x. ^-- = constant J t e (t r
ax
The second variation of J is given by

128
2
+ [(F. - x. F v )dt: + 2F v dx. dt]
U- I X . X, 1
1 1
{F 6x
k.kv i6xk 2F
x.xv 6x
i6xk
1 K IK
i k
which must be non-negative for all admissible variations

6x.,6x.,dt ,dt ,dx.(t ) and dx.Ct^). This suggests a further
minimization problem with the second variation as the func-
tional and the variations 6x. and differentials dt ,dt.
as the unknowns, subject to the differential equations of
constraint,
6x
6x
Yx i
3x X
i
Writing
6x
i6xk 6x
i6xk F
x.xv 6x
i6xk
84) .
IK
6x 6x
-•
the variations 6x which minimize the second variation must

satisfy the Euler-Lagrange type equations
d_
= 0
dt
These equations are called the accessory Euler-Lagrange

equations and the new minimization problem the accessory min-
imum problem.
By considering the second variation it can be shown that

a further necessary condition for optimality is that
x.x v
1 JC
subject to the constraints
*6 •
—•-- 6 x . = 0
3X. >•
129
7.2 The Optimal Control Problem
The general optimal control problem may be formulated as

follows. To find, in a class of control variables u.(t)
(j = l,2,...,m) and a class of state variables x.(t)
(i = l,2,...,n) which satisfy differential equations
= f±(x,u,t)
and end conditions

x(t ) = xQ , t specified
> (x(t1),t1) = 0 (q = 1,2,..., <_ n+1)
(t may or may not be specified), that particular control

T
vector u = (u,1
u~,...,u
z m
) and corresponding state vector
ip
x = (x1 X2/--./X ) which minimizes a cost functional
J = G(x(t ),t,) + [ L(x,u,t)dt

j
t
o
The control vector u e U where U is a control set which
may be open or closed.
The essential differences between this problem and the

classical one discussed earlier lies with the control vector
u(t). From practical considerations we must allow the con-
trol to be discontinuous. By raising the control vector u
to the status of a derivative z we may treat z as an
additional state vector and use the classical theory without
further modification. A second difference lies in the fact
that some or all of the control variables u.(t) arising in
practical problems will be constrained by inequalities of the
form
a. < u. < b.
3 3 3
where a., b. are usually known constants. To use the

classical theory for a problem involving such control con-
straints the technique of Valentine can be used. This is to
introduce additional control variables a . (t) such that
130
(Uj-dj)(bj-Uj) - a* = 0
These new equality constraints can then be treated in the

same way as the set of system equations
f (f f f
* = = l 2 n)T
Let us apply the classical theory to the control problem

formulated above. The fundamental function
F = L(x,u,t) + X i (t)f(f i (x / u f t) - x ± )] (7.5)
(i = 1,2,...,n)
With z. = u. the Euler-Lagrange equations are
d =
H = H"' ^T- constant (j = l,2,...,m)
9x
3x.J i 3z.J
1 3 (7.6)
The transversality conditions are
F — x. —— dt + - 3 — dx. + - — dz.) + dG
X D
O 1
3x, 3x.
l
dz.3 It
;
o
.
= 0 (7.7)
d*q = 0
The functions G and ij> do not involve the control var-

iables explicitly. Since the variables z. are not speci-
fied at either end point the transversality conditions yield
^T- = 0 (j = 1,2,...,m)
3zj
at t and t1 . Then, because of the second set of Euler-

Lagrange equations (7.6), we have
te (7 8)
~r- = w: - ° ^''i' -
dz. D
The remaining Euler-Lagrange equations (7.6) are equivalent

131
to
" A. = |-- (L + X±f±) (7.9)
Now define the Hamiltonian
H = L(x,u,t) + A i (t)f ± (x,u,t) (7.10)
Equations (7.8)-(7.9) then take the form of the Pontryagin

canonical equations
|~- = 0 (j = 1,2,...,m) (7.11)

j
= - ix- U - 1,2,...,n)
Note that F - x. -^— = F + A.x^^ = F + A.f. = H

9x .
The transversality conditions (7.7) then reduce to
|H dt - A i dx. dG = 0 dip = 0
o
where dG = ff- dt,

3 if,
and dip = - ^ dt x + -g^ ( £ — dx ± (t 1 ) (q = 1,2,...,
In particular, if x(t ) and t are specified but x(t^)

and t, are free, then the transversality condition yields
A
ivul; " 3x ± (t 1 )
If, on the other hand, x(t ) and t1 are specified and

I(<n) values specified at ^ (with t1 unspecified),
e.g.
x k (t x ) (k = 1,2,...,£) specified,
then the transversality conditions yield

132
H(t.) = - -|f- and X_(t,) = ff-f-f-T

1 3t. Si s 1
(s = £+1,1+2,...,n)
In the optimal control problem the Weierstrass-Erdmann

corner conditions (7.3) are that H and x i (i = l,2,...,n)
must be continuous at points of discontinuity in xi and
u. (j = 1,2,...,m).
The Weierstrass necessary condition (7.4) may now be re-

written as
E = L(x,Z,t) + X.(f.(x,Z,t) - Xi)
- L(x,z,t) - Xi(fi(x,z,t) - xi)
+ (X.-x.)A. - (Z.-z.) | (L+X.f.-X.x.) > 0

i l l j j " ™ 1 1 1 1 —
j
On u s i n g e q u a t i o n s ( 7 . 1 0 ) - (7.11) , t h i s i n e q u a l i t y r e d u c e s t o
L(x,Z,t) + X.f.(x,Z,t) - L(x,z,t)
- Xifi(x,z,t) >_ 0
Writing Z, the control vector associated with a strong

variation, as u and z, the control vector associated with
the extremal, as u we are led to the inequality
H(x,u,X,t) <_ H(x,u,X,t) / t e (t Q ,t 1 )
Thus, the Hamiltonian H must assume its minimum value with

respect to the control vector u at all points on the extre-
mal. This is Pontryagin 's Minimum Principle.
In our new notation, if the fundamental function F of

(7.5) does not depend explicitly on the time t then a
first integral of the Euler-Lagrange equations exists as
mentioned in the Note above. This first integral is then
L + x^ + xixi ~ constant > t e (t ft^

133
i.e. H = constant e <t o , t l )
Consider a one-parameter family of control arcs u(t,e)

with the optimal control vector given by arc u(t,o). The
corresponding state vector x will also be a function of t
and e with e = 0 along the optimal trajectory. Now int-
roduce the notation for the state and control variations:
(i = 1,2, ,n) 8. (j = 1,2,

de e=0
(7.12)
and, in the case where tl is unspecified so that t =t, (e),

the end-time variation
(7.13)
The second variation must again be non-negative for a min-

imum value of J and is given by
((H -x.H )C + 2H :.£ + n.) ) + d G

1 1
z. l x ± t=t,
H
.u^ u.uh638h)dt
subject to = 0
3f
i i
+
\ Fu~
7.3 Singular Control Problems 3 ' 4 ' 5
In the special case where the Hamiltonian H is linear in

the control vector u it may be possible for the coefficient
of this linear term to vanish identically over a finite in-
terval of time. The Pontryagin Principle does not furnish
any information on the control during such an interval and
134
additional necessary conditions must be used to examine such

extremals, known as singular extremals.
Definition
Let u, be an element of the control vector u on the

interval (t 2 ,t 3 ) <^ (t , t ^ which appears linearly in the
Hamiltonian. Let the (2q)th time derivative of Hu be the
lowest order total derivative in which u, appears explic-
itly with a coefficient which is not identically zero on
(t-jt..). Then the integer q is called the order of the
singular arc. The control variable u, is referred to as
a singular control.
Two necessary conditions will now be given for singular

optimal control.
7.3.1 The generalized Legendre-Clebsch condition:

a transformation approach
The generalized Legendre-Clebsch condition may be derived

in an indirect way by first transforming the singular prob-
lem into a nonsingular one and then applying the conventional
Legendre-Clebsch necessary condition. This condition for the
transformed problem is the generalized Legendre-Clebsch con-
dition for the original singular problem. We shall give
Goh's approach here which retains the full dimensionality
of the original problem and is simpler in the case of vector
controls. Kelley's transformation reduces the dimension of
the state space.
We have the second variation for a singular arc as
i2 = -j (n Gxx n)
tf +
f'4 • z
\- 6 ux
H n)dt
(7.14)
Jo
subject to the equations of variation
n = f n + f,3/ n(o) = 0 (7.15)
Write 3 = C and, without loss of generality, put c(o) = 0.

135
Now write n = a + fu C in order to eliminate h from the

equations of variation. These equations become
= f
- x" + (f f
x u" V '
The second variation becomes
T ( G +
2 = 1 ^ xx « «T G
xx f
u * +
\ ? f
u G xx f
u *>\
Integrate by parts the first term to eliminate z, from

This gives
tf
+ Hui v fu )?) + (-9- C R C+C Q a+ 4 a X P a) dt
t J 0
We can now update c to the status of a derivative since c

has been eliminated completely from I2 and the equations of
variation. The conventional Legendre-Clebsch condition then
yields R >_ 0.
2H f
Now, P.* = f^H x x f u - ^- H u x f u - Ux* xfu~fu*
Thus we arrive at the condition
r\2 \
—.K- H > 0 t e (t ,t~) (7.16)
2 U f
dt J * °
This inequality is a generalized Legendre-Clebsch necessary
condition in that it is similar in form to the classical
Legendre-Clebsch necessary condition
136
i~ H > 0 t e (t ,tf)
9u u - or
If (7.16) is met with equality for all t e ô^f^ then
the functional I2 is totally singular and another transfor-
mation must be made. In this way the generalized necessary
conditions are obtained:
^- 1 ) g fe£-5q H u > 0 t e (to,tf)

(q = 1,2, )
It can also be shown that
H t€
k £P u =° J W
p an odd integer
7.3.2 Jacobson's necessary condition
First adjoin the equations of variation (7.15) to the ex-

pression for the second variation (7.14) by a Lagrange
multiplier vector of the form Q(t)n, where Q is a nxn
time-varying, symmetric matrix. Then
t
+ 3 T H u x n)dt
Integrate by parts the last term of the integrand to give

tf
Now choose Q so that Q = -H - f Q - Of

XX X X
with Q(t f ) = G x x (x(t f ),t f ) (7.17)
tf
Then I?
z
== II u
3T(H
x
+ f^Q)n dt
u
)
137
with n = fxn + f u B , nCo) = 0
Now choose a special variation $ = 8 , a constant for

t e (t 1 / t 1 +AT) r 0 < t x < t x + AT < t f , and B = 0 else-
where. Then
1
t n +AT
-I. s\
T
(H
UX
+f
U
Q)n dt, n (t )
X
= 0 (n continuous)
By the mean value theorem,
I2 = AT B*T((H + f* Q) n ) (0 < e < 1)
Expanding in a Taylor's series:
i2 = AT
0(AT 2 ))
which, since n(t 1 ) = 0, gives
1"2 = AT3* (0AT(H +f Q) (f n+f B) + 0 (AT ))
= 9AT 2 6* T (H iY +f^Q) f ( t )6* + 0 (AT3)

UA U • li. X
For an optimal solution, ^ > 0. Thus,
(H ux + f^Q)fu >. 0 t e (0,tf) (7.18)
Equations (7.17) and inequality (7.18) constitute Jacobson's

condition.
7.4 Dynamic Programming '
Consider the discrete system governed by the difference

equations
x(i+l) = f 1 (x(i),u(i)), x(o) specified,
(i = 0,1,2,...,N-1)
138
in which it is required to find the sequence of decisions

(control vectors) u(o),u(1),...,u(N-l) which minimizes a
cost function <f)[x(N)].
The method of dynamic programming for solving such an opt-

imal control problem is based on Bellman's Principle of
Optimality. This states that an optimal policy has the pro-
perty that whatever the initial state and initial decision
are, the remaining decisions must constitute an optimal pol-
icy with regard to the state resulting from the first dec-
ision. In other words, having chosen u(o) and determined
x(l), the remaining decisions, u(1),u(2),...,u(N-l) must
be chosen so that $[x(N)] is minimized for that particu-
lar x(l). Similarly, having chosen u(o),u(1),...,u(N-2)
and thus determined x(N-l), the remaining decision u(N-l)
must be chosen so that <j>[x(N)] is a minimum for that
x(N-l). Proof is by contradiction.
The principle of optimality leads immediately to an inter-

esting computational algorithm. The basic idea is to start
at the final end point t = Nh, where h is the step-length
in time. Suppose that somehow we have determined x(N-l).
The choice of the final decision u(N-l) simply involves the
search over all values of u(N-l) to minimize <|>(x(N)) =
N 1
<|>(f " (x(N-l),u(N-l))) .
Suppose we denote the minimum value of <|>(x(N)), reached

from x(N-l), by S(x (N-l),(N-l)h)f i.e.
S(x(N-l), (N-l)h) = min <f> (fN~\x(N-l) ,u(N-l)))

u(N-l)
Similarly, we let
S (x(N-2) , (N-2)h) = min min <|>(x(N))

u(N-2) u(N-l)
= min S(x(N-l) , (N-l)h)
u(N-2)
= min S(fN~2(x(N-2),u(N-2)),(N-l)h)
u(N-2)
In general, we are led to the recurrence relation

139
S(x(n),nh) = min S(f n (x(n) ,u(n) ) , (n+l)h) (7.19)

u(n)
Using this recurrence relation and starting from n = N-l

the whole process is repeated stage by stage backwards to
the initial point where x(o) is specified. A sequence of
optimal controls u°(N-l),u°(N-2),...,u°(o) are thus gener-
T
ated. *7ith the completion of this process it is possible to
work forward in time from the specified point x(o) since
u°(o) will be known for that initial point. T-Tork forward
to x(1),x(2),...., using the optimal controls u°(o),u°(l),
...., until the final end-point is reached with the final
optimal control u°(N-l).
Clearly,
S(x(N),Nh) = <f)(x(N)) (7.20)
and this is a boundary condition on S to be used in the

solution of the difference equation (7.19) for S. In solv-
ing (7.19) the minimization process is carried out first and
then, having found the optimal u°(n), S may be treated as
a function depending explicitly upon x(n). By its defini-
tion the optimum cost function must satisfy equation (7.19)
and, furthermore, when S has been found then, by definition,
we have found the optimum cost function. Therefore, the
solution of (7.19) with boundary condition (7.20) is a nec-
essary and sufficient condition for an optimum solution.
Equation (7.19) is called the Hamilton-Jacobi-Bellman equa-
tion since it is closely related to the Hamilton-Jacobi
equation of classical mechanics.
As the interval h is made smaller so the discrete prob-

lem usually goes over naturally into the continuous problem.
The general form of the recurrence relation (7.19), to first
order terms in h, then gives rise to the equation
S(x,t) = min (S(x,t) + \ 9|§~

x
hf1 (x,u) + h 9|f
t
)
u i i
Since S is not a function of u this last equation reduces

to
140
i? 7 (IS. f i (x ' u)
3S
u must be chosen to minimize the sum f. -r— resulting in a
1 oX:
control law u(x,t). The partial differential equation
(7.21) may then be solved for S(x,t) subject to the bound-
ary condition
S(x(t 1 ),t 1 ) = <|>(x(t1)) (7.22)
7.5 The Hamilton-Jacobi Approach '
This approach to the problem of optimal control is an alt-

ernative to Pontryagin's Minimum Principle. It has two ad-
vantages: (i) it eliminates the difficult two-point boundary
value problem associated with the Minimum Principle, (ii) it
yields a closed-loop solution in the form of an optimal con-
trol law u (x,t). A disadvantage is that it gives rise to
a nonlinear partial differential equation, the Hamilton-
Jacobi equation, and a closed form solution has been obtained
for only a few special cases. This difficulty arises since
only one set of initial conditions was considered in the
Pontryagin approach whereas here, by stipulating a closed-
loop solution, we are seeking the solution for a whole class
of initial conditions.
Let the function S(x,t) of (7.21) be given as the opti-

mal cost function
S(x,t) = [ L ( X ( T ) , U ° ( X , T ) , T )dT (7.23)

Jt
i.e. S(x,t) is the value of the cost function evaluated

along an optimal trajectory which begins at a general time
t and associated state x(t). It can be shown that the
partial derivatives 8S/3x. of equation (7.21) have a time
behaviour identical with the Lagrange multipliers X. in
the Pontryagin approach. Thus, with the cost function given
in (7.23), equation (7.21) may be written as
H ( X , vs(x,t),t) + as/at = o (7.24)

141
This equation is known as the Hamilton-Jacobi equation. The

boundary condition (7.22) becomes
S(x(t 1 ),t 1 ) = 0
In the case of a linear, time-invariant system
x = Ax + Bu
with a quadratic cost function
(xTQx + u T Ru)dt
o
(Q and R constant, symmetric matrices with R positive
definite) the solution of the Hamilton-Jacobi equation can
be obtained in closed form. The Hamiltonian is
H = x T Qx + u T Ru + VS T (Ax+Bu)
Assuming u is unbounded,
1-3 = 2Ru + B T VS = 0
oU
whence u° = - -• R " 1 B T VS(x,t) (7.25)
Substituting this result into equation (7.24) yields the

Hamilton-Jacobi equation
X
(VS) AX - ^(VS) BR B 1 VS + X X Qx + ~ = 0
T
It can be shown that S(x,t) = x P(t)x (P > 0) is a sol-
ution to this equation provided that the matrix P satisfies
the Matrix Biooati equation
P + Q - PBR~ 1 B T P + PA + A T P = 0 (7.26)
with boundary condition

P(ti) = 0 (7.27)
142
A relatively straightforward method for solving (7.26) sub-

ject to (7.27) is a backward integration on a digital com-
puter. Once the matrix P is known the optimal control law
is found from (7.25) as
u°(x,t) = -R""1BTPx
= -K T (t)x say.
The elements of the matrix K(t) are known as the feedback

coefficients.
In the case where the final time is infinite the matrix P

is constant. We then only have to solve a set of algebraic
equations (P = 0 in (7.26))
Q - PBR~ 1 B T P + PA + A T P = 0
the reduced or degenerate Riccati equation. Of the several

possible solutions to this nonlinear equation in the elements
of the matrix P the one required is that yielding a posit-
ive definite P.
References
1. BLISS, G.A. : Lectures on the Calculus of Variations.

University of Chicago Press, 1961.
2. BRYSON, A.E. and Y.C. HO : Applied Optimal Control,
Blaisdell, 2nd edition, 1975.
3. BELL, D.J. and D.H. JACOBSON : Singular Optimal Con-
trol Problems, Academic Press, 1975.
4. CLEMENTS, D.J. and ANDERSON, B.D.O. : Singular Optimal
Control : The Linear-Quadratic Problem. Springer-Verlag,
1978.
5. KELLEY, H.J., R.E. KOPP and H.G. MOYER : Singular
Extremals in "Topics in Optimization" (G. Leitmann, ed.)
Academic Press, 1967.
6. BURLEY, D.M. : Studies in Optimization. Intertext Books
1974.
7. DENN, M.M. : Optimization by Variational Methods,
McGraw-Hill, 1969.
143
Problems
P.7.1 Apply Pontryagin's principle to the problem of minim-

ising the time taken for a system to transfer from the
T
given state (-1 2) at time zero to the origin at
time t. when the system equations are
Show that the phase trajectories are circles and that

the phase point makes a complete revolution in time
2TT and in a clockwise direction. Describe the phase
trajectory to the origin assuming there is only one
switching point for the control u. Explain why the
total time for the phase trajectory must be less than
3TT/2.
V.I.2 A system is governed by the equations
xx = cos x 3 , x 1 (0) = 0, x x (t f ) = /2
x3 = u , x 3 (0) = 0, x 3 (t f ) free,|u|<l
It is required to minimise the final time tf . Show that

there is a partially singular control commencing with u = 1
t e (0, TT/4) which yields a candidate trajectory satisfying
all the constraints. Verify that the GLC condition is satis-
fied along the singular subarc.
Chapter 8
Control system design via mathematical

programming
Professor D. Q. MAYNE
Synopsis
Control design requires the choice both of structure of the

controller and parameters for a given structure; the first
task is guided by system theory; the latter is aided by
mathematical programming. For a given control structure de-
sign objectives can often be specified as satisfying a set of
inequalities or minimizing a function subject to these in-
equalities. It is shown that many objectives can be expressed
as infinite dimensional constraints (cj)(z,a) f. 0 for all a e
A). Algorithms for the resultant semi-infinite programming
problems are briefly outlined.
8.1 Introduction
Control design is a fascinating mixture of control theory,

numerical techniques and the experience and intuition of the
designer. Control theory yields important structural
results such as the structure of a controller necessary to
achieve asymptotic tracking and regulation for all desired
outputs (reference signals) and disturbances lying in speci-
fied classes; it also yields synthesis techniques, that is,
solutions to well-defined problems such as optimal control
(linear system, quadratic cost case) and pole allocation.
The problems solved in this way are often too simple (being
well defined) to represent practical design situations yet
the resulting synthesis techniques may often be useful in
solving recurring sub-problems in the design process. Given
the structure of the controller, parameters have to be chosen.
Theoretical results are often of little help at this stage.
For example, multivariable root locus theory may yield the
asymptotic root loci (and aid the choice of a controller
structure) but not suitable values for (finite) controller
parameters; a numerical procedure (e.g. in the form of a
root locus plot) is necessary. The complexity of most design
problems requires the constant interaction of a designer,
specifying and changing some of the many objectives and con-
straints, choosing suitable structures and intervening where
necessary in the many complex calculations.
Given the active intervention of the designer, it appears
145
that many, if not most, of the design specifications can be

expressed as inequalities. It is, of course, possible that
some a -priori constraints are unrealistic and must therefore
be altered during the design process; a design may also pro-
ceed by initially satisfying easily attained constraints and
then tightening these constraints. A computer aided design
package can provide these interactive facilities.
Many design specifications can be expressed as conventional

inequalities. Thus the requirement of stability of a linear
multivariable system can be expressed as Rx 1 (z) <_ 0, i =
1
l,...,n where x (z) denotes the ith eigenvalue of the
closed loop system given that the controller parameter is z.
Nevertheless many design constraints are considerably more
complex and have the form <$> (z,a) <_ 0 for all a e A where
m
A is an interval or, even, a subset of R , m > 1. These
infinite dimensional constraints have only recently been con-
sidered in the mathematical programming literature; never-
theless algorithms for satisfying such constraints or for
minimizing a function subject to these constraints are now
available and provide useful new design aids.
It is interesting to note that infinite dimensional inequal-

ities arise in a wide variety of applications. One example
is the design of a filter such that the output ty to a given
input pulse s satisfies an envelope constraint (\JJ (t) e
[a(t), b(t)] for all t e [0, t ^ ) . This problem is rele-
vant to pulse compression in radar systems, waveform equaliz-
ation, channel equalization for communications and deconvol-
ution of seismic and medical ultrasonic data. Another example
is the design of framed structures such that the displacement
of the structural elements in response to a specified distur-
bance is limited at every time in a given interval; the
differential equations describing the motion of the structure
are non-linear. In these two examples A is an interval.
Consider, however, a circuit design problem where the para-
meters of the manufactured circuit differ from the (nominal)
design value z e Rr by a tolerance error t e Rr known to
r
lie in a tolerance set A c R . If the constraints must be
146
met by all possible manufactured circuits then constraints of

the form <j> (z + t) <_ 0 for all t e A must be satisfied. A
similar problem occurs, for example, in the design of chemical
plant with uncertain parameters. If the uncertainty or tol-
erance error occurs in parameters other than the design
parameter then the constraint becomes <f>(z,a) <_ 0 for all
a e A, A a subset of Rm say. This can be expressed as
max{cj> (z,a) | a e A} <_ 0.
Even more complex constraints can occur. In many circum-

stances post manufacture tuning (or trimming) of the design
parameter z is permitted. If z + r(q) denotes the value
of the tuned parameter, then a design z is acceptable if
for all a e A there exists a q e Q such that <f> (z+r (q) ,a)
_< 0 . This can be expressed as:
max min <j> (z+r (q) , a) <_ 0

aeA qeQ
Further complexities arise when the properties of the function

<j> are considered. In many cases <f> is a differentiable
function (of z and a ) ; in other cases, however, $ is not.
An important example of this is discussed in the next section;
many control design objectives can be specified as constraints
on singular values of transfer function matrices and singular
values are not differentiable functions.
This paper shows how infinite dimensional constraints arise

naturally when design objectives for linear and non-linear
control systems are formulated. The design of linear and non-
linear multivariable systems is discussed in Sections 8.2 and
8.3 respectively. Algorithms for solving the resulting feas-
ibility and optimization problems are discussed in Section
8.4 .
8.2 Linear Multivariable Systems
Design objectives for linear multivariable systems are dis-

cussed in Reference 1 where several current approaches to
control system design are presented. If Q(z,s) denotes the
loop transfer function when the controller parameter is z
147
and T(z,s) A I + Q(z,s) the matrix return difference, then

det[T(z,s)] = P C ( Z / S ) / P O ( Z , S ) where PO(Z,S) and PC(Z,S)
are the open and closed loop characteristic polynomials so
that stability can be assessed in the usual way from the
Nyquist plot of 6(z,w) 4 det[T(z,jw)]. Thus, if the system
is open loop stable, the closed loop system is stable if
<f>(Z/O)) .1 0 f ° r all a) e [O,00) where 4>(z,w) 4 [I6(z,w)] -
2
a[R6(z,w)] +3 (R and I denote the "real part" and "imagin-
ary part" respectively) and a and B are chosen so that
the origin does not lie in the set {s e C |s = u + jv,
v - au + 8 ^_ 0} (a suitable choice is a = 1, $ = 0.2) .
In practice, due to the fact that Q(z,ju>) + 0 as w->°°,
the semi infinite set [0,w) can be replaced by [0,(^1 so
that the stability constraint can be expressed as the infin-
ite dimensional constraint:
<(>(zruj) <_ 0 for all OJ e [0,0^] (8.1)
Although a conventional inequality (<|> (z) <_ 0 where $ (z)

1 1
4 maxiRX (z)!i = l,...,n} - e, A (z) being the ith eigen-
value of the closed system) may be used to ensure stability
when the system is finite dimensional, constraint (8.1) may
be employed for infinite systems.
Performance (e.g. small tracking error) can be expressed as

an upper bound on j | T (z,j w )|| 9 for all w e [0,u)-,] say
-1
(T is the transfer function from the reference to the
error). This can be expressed as in (8.1) if <}>(z,to) 4
a( w ) - £(z, w ) where o_(z,u)) is the smallest singular value
of T(z,jw) and a(w) is a (large) positive function of w;
the constraint <f>(z,o)) <_ 0 is equivalent to £(z,uO >_ a(w),
i.e. to a large loop gain. On the other hand, the transfer
function from the sensor noise to the output is
Q(z,s)T" 1 (z,s) = T"1(z,s)Q(z,s) = [I+Q" 1 (Z,S)l" 1 ;
a constraint on the magnitude of the component of the output

due to sensor noise can be expressed as a[I+Q~ (z,j<u)~ ] ^_
b(w) or £[I + Q (z,jw)] >_ c(w) for all w. This can also be
148
expressed in the form of (8.1) . The performance constraint

requires ||T || to be small and the output noise constraint
requires | |T"* 1 Q| | to be small; since T""1 + T""1Q =
1
[I+Q]~ [I+Q]. = I, these constraints conflict . Similar
expressions can be employed to impose constraints on the con-
trol action in response to reference signals, disturbances
and sensor noise .
While existing frequency response design methods can be

employed to ensure integrity (maintenance of stability) in
the face of large variations in loop gains (due e.g. to actu-
ator or sensor failure) they cannot necessarily cope with
large variations in plant parameters. The reason for this
is that it is possible for small variations in the plant to
cause large variations in det[T(z,s)] (or in the character-
istic loci), so that even though the Nyquist plot of
det[T(z,jw)] has large gain and phase margins, the result
2
design is not necessarily robust. It is argued that the
percentage variation in the model of the plant (| | AG(juj)G~ (jw)||)
is likely to be higher at high frequencies. It is also shown
that stability can be ensured for all plant variations satis-
fying a(AG(jo))G~ (j 03) 3 <_ £ (ai) if a_[Q (z , J<D)T~ (z,jw)] <
)) for all a) c [0,°°); £(w) commonly increases with w
and may exceed unity. This stability constraint may also be
expressed in the form of (8.1) and also conflicts with the
performance constraint; in practice a pass band must be
chosen so the performance (and stability) constraints are
satisfied for GO in this band while the stability constraints
dominate outside (where l/£(w) is small).
The variations in the plant transfer function G(s) con-

sidered above are completely unstructured (all variations
AG(s) satisfying a[AG(jco)G (jw) ] <_ I (w) are permitted).
If more precise information is available (e.g. certain plant
parameters p always lie in a given set P) then the
robustness can be ensured by the modified stability constraint
<f>(z,p,aj) <_ 0 for all a) e [0,0^], all p e P (8.2)

149
where
6(z,p,u)) 4 [I6(z,p,u))] - a[R6 (z,p,aj) ] 2 + 6
and
9(z,p,u>) 4 det[T(z,p,jw)]
the return difference being expressed as a function of the

controller parameter z, the plant parameter p and the
complex frequency s. If q denotes certain plant or con-
troller parameters that can be tuned (after the controller
has been implemented) and Q the tuning range then constraint
(8.2) is replaced by:
for all a) e [0,0^] , all p e P there exists a q e Q

such that <j>(z,p,q,w) <_ 0
or, equivalently:
max min <f> (z ,p,q,u)) ^_ 0

qeQ
where ft 4 [0,0)^] and <J> is defined as above with T(z,p,s)

replaced by T(z/p/s).
It is, of course, easy to express constraints on time res-

ponses of outputs, states, controls or control rates in the
same form . Thus constraints in overshoot, rise time and
settling time on the unit step response of an output y may
be expressed as:
y(z,t) <_ 1.15 for all t e [0,t2]
y(z,t) >_ 0.5 for all t e [t 1# t 2 l | (8.3)
|y(z,t) - l| <_ 0.02 for all t e [t 2 ,t 3 ] J
If robustness to plant variations is required then y(z,t)

should be replaced by y(z,p,t) and constraints* (8.3) be re-
quired to hold for all p e P.
It is thus apparent that many, if not all, of the design

objectives may be expressed as infinite dimensional con-
straints. How can these objectives be achieved? If compen-
150
pensatic-n is employed to yield a diagonally dominant transfer

function the controllers for each loop can be independently
designed to satisfy many design objectives (but not necessar-
ily robustness to plant variations). The characteristic loci,
on the other hand, do not correspond to separate loops; how-
ever, the associated design method employs an alignment tech-
nique so that (at least at selected frequencies) the loci
appear to the designer to be associated with distinct loops.
Again, many design objectives (but not necessarily robustness
with respect to plant variation) can be achieved with this
method.
Since pre-compensation cannot be used to isolate, for exam-

ple, singular values (and would not in any case be desirable)
more sophisticated adjustment techniques, such as those pro-
vided by mathematical programming, are required to choose
suitable parameters.
8.3 Non-linear Systems
Much less effort has been devoted to the problem of design-

ing non-linear multivariable controllers because of the over-
whelming complexity of the task. Of course frequency response
criteria for systems of special structure (e.g. linear dyna-
mics systems in tandem with static non-linearities) and
heuristic criteria, (e.g. describing functions for assessing
stability) have been employed for design. Lyapunov methods
can be employed for systems whose state dimension does not
exceed three or four. In practice, simulations of the non-
linear are often employed to assess stability and perform-
ance. Theoretically this requires simulations over the
semi-infinite interval [0,°°) for every initial state of
interest. We show in this section how stability and per-
formance constraints can be expressed as infinite dimensional
inequalities whose evaluation is considerably less expensive
computationally than simple simulation.
Suppose the controlled system is described by:
x(t) = f(x(t),z) (8.4)

151
where, as before, z denotes the controller parameter, and

f is continuously differentiable. Suppose that the set of
initial states of interest is a compact subset Z of Rn .
Let V : R -* R be a continuous function with the following
properties:
(i) V ( x ) >_ 0 for all x e Rn,
(ii) V(ax) = a V(x) for all a e [0,«>) and all x e Rn,
(iii) V(x) = 0 if and only if x = 0.
T h
An example of such a function is x •> (x Px) where P is
positive definite. For all x in Rn let the set B(x) be
defined by:
B(x) A {x 1 e Rn|V(x') < V(x)}
Clearly B(0) = { 0 } , B(ax) = a B ( x ) , B(ax) ^ B(x) if a < 1

and B(x') « B(x) if and only if V ( x f ) <_ V ( x ) . Choose x
so that x C B 4 B ( x ) ; let V^VCx) . Let x(t,z,xQ)
denote the solution of equation (8.4) with x(0) = x Q . If
z can be chosen so that V(x,z) A V (x)f(x,z) < 0 for all
x e B , x j- 0, then the closed loop system is asymptotically
stable with a region of attraction B (V(«,z) is a Lyapunov
function). This is seldom possible; however, it is possible
to relax the requirement V(x,z) < 0 for all x e B as
4
shown in the following result :
Theorem_8._l . Let a e ( 0 , 1 ) , c e (1,°°), T e (0,°°) . Let

8 : B -> [0,1] satisfy 6 (x) <_ m a x { a , 1 - V(x)/V} <_ 1 and
:
let y B -> [l,c] . If there exists a z such that:
(i) x(t,z,xQ) e y(xo)B(xo) for all x Q e B , all t e [0,T]
(ii) x(T,z,xQ) e B(xQ)B(x0) for all x Q e B,
then:
(a) x ( t , z , x Q ) e cB for all x Q e B , all t e [0,oo)f
(b) x ( t , z , x Q ) -* 0 as t •-»-«> for all xQ £ B .

152
Comment: The hypotheses of Theorem 8.1 are satisfied if the

function 6 is replaced by a constant in the interval (0,1)
and the function y is replaced by the constant c. The re-
sult follows from the fact that
x e Y
i i-l Y i-2 ••' ^ O B ( X O }
where x ± 4 x(iT,z,x Q ) and y± 4 y ( x ^ , i = 0,1,2,
It follows from Theorem 8.1 that stability is ensured if

two infinite dimensional inequalities (hypotheses (i) and (ii))
are satisfied.
Performance has to be assessed in the time domain. Suppose

r a e A
that the set of inputs of interest is ^ al ^ (e.g.
r = a + an
a(t) n a^t) & that the instantaneous tracking error
due to an input r^ is e(t,z,XQ,a). Then a possible-per-
formance constraint is | e (t ,z ,XQ ,a) | <_ a(t,XQ,a) for all
t e [0/t1]/ all X Q e X, all a e A. This is clearly an
infinite dimensional inequality.
8.4 Semi-infinite Programming
It is thus apparent that the task of choosing suitable para-

meters for a controller (as contrasted with the task of choos-
ing a controller structure) may in many cases be expressed as:
PI Determine a z such that:
g j (z) <_ 0, j = 1,. . . ,m (8.5)
<^(z,a) <_ 0 for all a e A, j = l,...,p (8.6)

or
P2 Determine a z which minimizes f(z) subject to con-

straints (8.5) and (8.6).
In P2 g-^(z) <_ 0 is a conventional constraint and f re-

presents some criterion (e.g. a measure of tracking error)
which it is desired to minimize. If tuning is permitted then
PI is replaced by:
153
P3 Determine a z such that (8.5) is satisfied and such

that for all a e A there exists a q e Q such that:
4>D(z + r(q), a) <_ 0, j = l,...,p
Problem P2 is similarly modified if tuning is permitted.

Several classes of algorithms have recently been developed
for these problems. We give a brief exposition of these al-
gorithms for the case when there are no conventional con-
straints and one infinite dimensional case.
8.4.1 Feasible-directions algorithms '
The constraint that we consider is
4>(z,a) <_ 0 for all a e A (8.7)
The feasible directions algorithms are restricted to the case

when A is an interval [<*«,, a ] of the real line (e.g. time
or frequency intervals) . For all z let ty : R •> R and
be
i|> (•) + defined by:
ty ( z ) 4 max { <>
f (z , a ) | a £ A}
and
ty ( z ) 4 max { 0 , i> ( z ) }
Let the feasible set (the set of z satisfying (8.7)) be de-

noted by F. Then:
F = {z e R n | xp ( z ) <_ 0}
= {ze R 11|> ( z ) , = 0}
All semi-infinite mathematical programmes solve an (infinite)

sequence of finite dimensional subproblems. The feasible
directions algorithms accomplish this as follows. For all
£ > 0, all z let A (z) and A^(z) be defined by:
A (z ) 4 {ct£A|4>(z,a) >_ ty (z) + - £ }

154
and
r
A (z) A_ {a A (z) | a is a local minimizer of
$ (z, •) on A}
The reasonable assumptions that <j> is continuously differ-

entiable and that the set A (z) is finite for all z and
all e >_ 0 are made. For PI a search direction is chosen
which is a descent direction for <fr(z,a) at every a e A £ (z) .
Specifically the search direction s (z) solves:
e (z) = min max{<g,s>|s e S, g e d i//(z)}

e e
s g
or
6 (z) = min{ | | s | | |s e convex hull of 9 i|>(z) }
where
S 4 (s I IsI I < 1}
and
3 i^ (z) A {V A (z,a) | a e A (z) }
£ ^ £
and £ is chosen so that:
(z)
The step length is chosen, using an Armijo procedure, to

reduce \\>. For P2 the search direction solves:
(z) = min max{<Vf(z), s> - y

seS
<g,s>, g e 3
and the step length is chosen to reduce ij; if z is not

feasible and to reduce f subject to the constraint if z is
feasible. The positive constant y ensures that the search
direction is a descent direction for ty when z is infeas-
j.ble (I|J(Z)+ > 0) and a feasible direction otherwise. In
the implementable versions of these algorithms ip(z) and
A (z) (the set of £-most-active local maximizers) are only
approximately evaluated, the accuracy being increased automa-
tically to ensure convergence.
155
A quadratically convergent version of the algorithm is des-

ribed elsewhere ;
cribed it us(
uses the extension of Newton's methods
for solving inequalities.
8.4.2 Outer approximations algorithms
In these algorithms the semi-infinite program PI and P2 are

replaced by an infinite sequence of conventional programs in
which the infinite dimensional set A (which is now not nec-
essarily an interval) is replaced by an infinite sequence
{A., i = 0,1,2,...} of suitably chosen finite subsets. The
corresponding feasible sets {Fa } are defined for all i by:
i
F
A A {z|<j)(z,a) £ 0 for all a e A^}
Since F is a subset of FA , FA is called an outer appro-

i i
ximation. At iteration FA i,
is formed bv adding to
i+1
F the a which (approximately) solves the global maxim-
i
ization problem max(4> (z, a) | a e A } and discards other
elements of FA judged to be unnecessary . To solve PI
i
the outer approximation algorithm determines a sequence
{z^} such that z. approximately solves the problem of det-
ermining a z in F
; to solve P2 s. approximately solves
1
i
the problem of minimizing f(z) subject to the constraint
that z lies in F .
l
8.4.3 Non-differentiability 9 ' 1 °' Xl
In some problems (e.g. those with singular value constraints)

4) is not dif ferentiable with respect to z (for given a) ;
9 10
algorithms have been developed ' for the case when <j> (and
therefore ty) is Lipschitz continuous. These algorithms are
similar
similar to
to those
those described
described above with
above wiJ 9 \\> now defined to be
the smeared generalised derivative:
d i) (z) A U{3i|;(z) |z'-z|
where the generalized derivative dip (z) is defined to the

convex hull of the limit of all sequences {8^(zi)/3z} where
156
z. -> z is such that 3i|/(z.)/3z exists {\\> is dif ferentiable

almost everywhere). Since a i^(z) is a set and not comput-
able, it has to be approximated by the convex hull of a
finite number of its elements. An adequate approximation
can be achieved with a finite number of operations ' '
To solve PI the search direction is -h (z) where h £ (z) is
that point in 3 i|/(z)- (or its finite approximation) closest
>to the origin. Corresponding extensions to solve P2 together
with an application to satisfying constraints on singular
values of transfer functions over an interval of frequencies
is described in Reference 10. Further non-differentiable
optimization algorithms are described in Reference 11.
12 13
8.4.4 Cut-map algorithms '
The cut map algorithms are outer approximation algorithms

in which the outer approximations are particularly simple.
Suppose there exists a (lower semi) continuous function 6 :
F c -v R such that 6(z) > 0 and B(z, 6(z)) 4 { z ' ||z'-z||
< 6(z)} does not intersect F for all z in Fc. Then
c
B(z, 6(z)) is an outer approximation to F and the map
c 12
z + B(z, <S(z)) is a cut map. The basic cut map algorithm
for solving PI computes at iteration i any z. not lying in
B(z., 6(z.)), j = 0,l,...,i-l; in fact this set may be re-
placed by B(z., 6(z.)), j e J(i) c {0,1,... 1-1} where
12
J(i) may be a small subset of {0,1,...,i-l}. The algor-
ithm has been applied to the tolerancing problem (determine
a z such that <f> (z+a) <_ 0 for all a e A) and to the tun-
ing problem (determine a z such that for all a e A there
exists a q e Q such that <J> (z+a+r (q) ) £ 0) in circuit
design and may well be applicable to the more general prob-
lems arising in control system design.
8.5 Conclusion
Many control design problems decompose into two phases:

choosing a control structure and choosing an optimal set of
parameters given the control structure. A design process may
pass repeatedly through these two phases. The parameters are
157
chosen to satisfy a set of inequalities specifying design

objectives or to minimize a criterion subject to those
inequalities. The inequalities are often infinite dimen-
sional giving rise to semi infinite programming problems.
Typical design objectives giving rise to such inequalities
and methods for solving the resultant mathematical programs
have been briefly presented. To apply the algorithm it is,
of course, necessary to have an interactive proqram with
the ability to evaluate <J> (z, a) and V <J>(z, a) . How
this may be done for constraints on time and frequency
responses is indicated elsewhere . The approaches
outlined here may be the only practical path towards satisfy-
ing hard constraints on time and frequency response (such as
singular values).
References
1. MUNRO, N. (ed.) : 'Modern approaches to control system

design 1 , Peter Peregrinus Ltd., 1979.
2. 'Special Issue on Linear Multivariable Control Systems',
IEEE Trans. AC-26, Feb. 1981.
3. ZAKIAN, V. and U. AL-NAIB : 'Design of dynamical and
control systems by a method of inequalities', Proc. IEE,
120, (1973), pp. 1421-1427.
4. MAYNE, D.Q. and E. POLAK : 'Design of nonlinear feed-
back controllers', IEEE Trans., AC-26, (1981), pp.730-
732. Also Proc. 19th Conf. on Decision and Control,
Albuquerque, New Mexico, pp. 1100-1103 (1980).
5. MAYNE, D.Q. and E. POLAK : 'An algorithm for optimiza-
tion problems with functional inequality constraints',
IEEE Trans. Auto. Control, AC-21, (1976), p. 184.
6. MAYNE, D.Q., E. POLAK and R. TRAHAN : 'Combined phase I,
phase II methods of feasible directions', Mathematical
Programming, Vol. 17, (1979), pp. 61-73.
7. MAYNE, D.Q. and E. POLAK : 'A quadratically convergent
algorithm for solving infinite dimension inequalities',
submitted to Applied Mathematics and Optimization.
8. MAYNE, D.Q., E. POLAK and R. TRAHAN : 'An outer approx-
imations algorithm for computer aided design problems',
J. of Optimization Theory and Applications, 28, (1979),
pp. 231-252.
9. MAYNE, D.Q. and E. POLAK : 'On the solution of singular
value inequalities over a continuum of frequencies',
IEEE Trans., AC-26, (1981), pp. 690-694. Also Proc. 19th
158
Conf. on Decision and Control, Albuquerque, New Mexico,

(1980), pp. 23-28.
10. MAYNE, D.Q. and E. POLAK : 'Algorithms for the design
of control systems subject to singular value inequali-
ties1 , submitted to Mathematical Programming Studies.
11. MAYNE, D.Q., E. POLAK and Y. WARDI : 'Extension of
non-differentiable optimization algorithms for con-
strained optimization 1 , to be submitted to SIAM J. of
Control and Optimization.
12. MAYNE, D.Q., A. VOREADIS : f A cut map algorithm for
design problems with parameter tolerance', accepted by
IEEE Trans, on Circuit Theory.
13. MAYNE, D.Q. and A. VOREADIS : 'A cut map algorithm
for design problems with parameter tolerances and tuning1
submitted to IEEE Trans, on Circuit Theory.
14. MAYNE, D.Q., R.G. BECKER, A.J. HEUNIS : 'Computer
aided design of control systems via optimization',
Proc. IEE, Vol. 126, (1979), pp. 573-578.
Chapter 9
Optimisation in multivariable design

Dr. G. F. BRYANT
Synopsis
Three new basic algorithms to achieve diagonal dominance

will be described in this paper: (i) to choose optimum per-
mutation of input or output variables of the plant; (ii)
the optimal choice of scaling factors, and (iii) to find a
real precompensator to optimise dominance measure of the
compensated plant.
9.1 Introduction
The design of multivariable systems via frequency domain

based procedures has been intensively studied in the last
decade and an array of significant theoretical results and
design procedures have emerged 1 . For non trivial problems
the designer invariably requires the aid of a computer and in
practice CAD facilities in which the designer can 'interact'
with the design steps have played a significant role. As yet
'black and white 1 displays have proved adequate, but it seems
likely that 'colour' displays will soon prove of value and
that formal optimisation procedures will play an increasingly
important role as a design aid.
One of our most widely used and studied design procedures

employs the concept of dominance of the plant transfer func-
tion 1 '^ and we will mainly concentrate on the use of formal
optimisation procedures in this area. Three important tech-
niques can be employed to obtain a successful design. These
are:
9.1.1 The allocation problem
Swap the inputs or outputs of the plant to obtain a better

'pairing' for subsequent precompensator design and design of
single input-single output closed loops. The pairing that
may emerge from physical considerations or during model form-
ulation is certainly not always the best from a control point
of view and we will consider which pairing is 'best' in terms
of dominance. The technique is of general value in design,
and not limited to those based on dominance. We will show
that modest sized problems can be solved 'by eye 1 using
160
simple calculations, and larger problems require graph-theo-

retic methods.
9.1.2 The scaling problem
Scaling of inputs and outputs changes dominance, and domin-

ance based design approaches might be criticised by not being
'scale invariant 1 . As can be shown , 'optimal' scaling yields
'invariant' dominance, which can be calculated using linear
programming based methods. We will describe algorithms for
calculating optimal scaling.
9.1.3 Precompensation design problem
The calculation of 'real1 compensators which are optimum

over the bandwidth of the plant, has proved one of the most
difficult obstacles in the application of dominance^ and a
large number of heuristic solutions are described in the lit-
erature. As the methods we describe usually yield 'global'
optima1 they usually serve in practice to settle the question
of whether dominance is attainable for the given class of
compensator.
9.2 Problem Formulation
A matrix Q e Cnxn is said to be column dominant if
where
n
4 = J Iq-jJ i = 1,2,...,n
and row dominant if
where
J |q±j| i = 1,2,...,n
The matrix Q e Cnxn ±s said to be dominant if either of the

above inequalities hold, and strictly dominant if either hold
strictly. In multivariable design applications it is necess-
ary to consider transfer function matrix Q(s) e C , when
each of the elements are usually rational functions of a com-
plex parameter s e C . if D is a suitable Nyquist contour
in the extended complex s-plane, we say Q is (strictly)
161
dominant on D if Q(s) is (strictly) dominant for all s

e D .
9.2.1 Decomposition
The simplest and most frequently used precompensator design

problem can be stated as follows. Given a square transfer
function matrix G(s) e cnxn find a real precompensator K
nxn
e # such that at frequency s = joo say,
Q (jw) = G (joa) K
is dominant. Selecting the column dominance criteria we

require
|q i± | i «J i = 1,2,...,n (9.1)
Obviously
q =
ij Ix g i p ( j u ) k P j
or in short form q.. = G.*(jw)K*., where G.*(G*.) is the
ith row (column) vector* of G. It is evident that the ith
column of Q is determined by the ith column of K and
hence we may without loss of generality separately consider
the choice of each of the n columns of K when seeking to
satisfy inequality (9.1) . This decomposition of the prob-
lem into n separate problems is of great practical value.
9.2.2 Choice of dominance measure
We first note that both |q.^| and <S? are positive first
order homogeneous functions of K in the sense that for any
A ± 0
|q i:L (XK)| = |x|q ±i (K) (9.2)

and
6*T(AK) = |x|6^(K) i = 1,2,. . . ,n
Hence a suitable and normalised dominance measure widely

adopted in the literature is
(9.3)
162
where for dominance of column i we require J? < 1, (i = 1,
From equation (9.2) it is clear that J. is independent

c n
of X and that J. is constant on any ray in R and in-
determinate at the origin. However, in any design the mag-
nitude of the scaling factor is determined from engineering
considerations, and the sign of X from stability consider-
ations, and hence these indeterminacies are acceptable. To
avoid computational difficulties associated with constancy
of J C (XK) for any X ^ 0 and the singular behaviour for
X = 0; it is usually convenient to introduce a constraint
of the form ||K|| = 1 say, where ||*|| is any convenient
norm; if we also wish to avoid ambiguity in the sign of X
then a stronger constraint such as k.. = 1 is convenient.
9.2.3 General problem
A great interest has been placed on achieving dominance

over a set of frequencies i] 4 {u. ; i = l,2,...,m} usually
I
spread through the bandwidth of the system. We need to de-
cide a suitable form of cost function. Now design difficul-
ties are usually reflected by the 'Worst Case 1 on Chebyshev
measure, i.e.
max J C (a) . )
there generally being little value in obtaining 'tight' dom-

c c
inance at one frequency (J. << 1) if we have J. - 1 at
some other frequencies. Hence, for multiple frequency case
we consider the decomposed problems.
P : min max J (w ) ; r= l,2,...,m
r
° k r
==
u \w ) O.ÔJ J / q • v co )
r i r i r
n
*r> = I
= G(u r )K
163
where q* e Cnxl, K e Rnx and we have, where there is no

ambiguity dropped the suffix i. Although all the algorithms
described have been successfully extended to multiple fre-
4
quency case , we will, for reasons of space, only state the
single frequency algorithms.
9.3 Allocation Problem
Consider a single frequency problem where we seek to per-

mutate the row (column) of a matrix to obtain (in some sense)
best overall column (row) dominance of the resultant matrix.
Essentially we are deciding how best to pair inputs and out-
puts. Although this problem is trivial for the 2x2 problems
which pervade the literature, it rapidly becomes more diffi-
cult as the dimensions increase (for n = 5 there are 120
possible permutations and for n = 10 more than 3.6 million).
Furthermore, studies indicate that correct pairing can signi-
ficantly improve the design problem.
As each row permutation affects two column dominance (sim-

ilarly for column permutation on row dominance), we need -
first to introduce a method of comparing the column dominance
for any permutation. For any Z e cnxn consider a positive
C n
matrix D where i .
P =l
dc.
It is easily shown that this represents the appropriate dom-

inance measure (analogous to J. of equation 9.3 when rows
i and j are interchanged, and hence element j of column i
is moved into the diagonal position i. Following the Cheby-
shev norm approach we require that the optimally permuted
matrix displays the smallest possible maximum dominance mea-
sure. This criteria in principle only fixes one diagonal
element and the dominance measure of one column. Extending
Chebyshev criteria to the choice of a second diagonal element
should then minimise the maximum dominance measure of the
remaining columns; proceeding in this way we see the opti-
mum permutation will select diagonal element so as to satisfy
164
a sequence of n minmax criteria. In short, we are required

to solve a sequence of problems of the form
P, : min max{J?[P,Q]}
Pi1
An equivalent problem to Pj^ is that of choosing a set of
n elements of DC which form a 'perfect matching1 of D C
and additionally minimise the maximum element in the match-
ing. A set of n elements forms a perfect matching of D C e
/? n x n if there is exactly one element of the set in every
row and column of D . A simple heuristic for finding such
a set (which usually suffices for 'by eye 1 solution of prac-
tical problems) is as follows:
Algorithm^ _(to_find P^
_ , _c n nxn
Data: D eR
Step 0: Set P 1 = D° ; l =1
Step 1: 'Colour' the smallest 'uncoloured' element of P
Step 2: As each additional element is coloured, check to
see if the coloured set contains a perfect matching.
If not, return to Step 1.
Step 3: Colour the element in D c corresponding to the last
£
element coloured in P . If £ = n, stop. Other-
0—1 ( Q—1 W (o - "\ )
wise form a new matrix p ^ - L e / ? ^ - L ^ v ) c ± ; b y
removing the row and column of P£ containing the
last coloured element. Set £ = £+1.
Return to Step 2.
The sequence of coloured element in D is the optimal

solution. To see this, consider the first time Step 3 is
executed. The last element coloured is the smallest element
that can be chosen in its row and column, whilst at the same
time ensuring the other rows and columns can be 'covered' by
the set already coloured and comprising elements of D C
smaller than the last element coloured. The same reasoning
applies for the successively reduced matrices P^ each time
we enter Step 3; hence we end up with a perfect cover whose
elements also satisfy the sequence of minmax properties re-
165
quired for optimality. We note that the solution is unique

c
if the elements of D are distinct.
The only step in the algorithm where difficulty can arise

is in checking to see if the coloured elements contain a
matching (Step 2 ) , and usually this is possible 'by eye 1 for
multivariable design problems (n <_ 3?). For larger problems
there is advantage in finding the optimum match automatically,
especially since it may be required several times in a design
(as shown by example later). The graph-theoretic equivalent
of the problem is a minimax 'bottleneck' matching problem for
which powerful algorithms exist. For instance modern bipar-
tite graph theory matching algorithms or simple modification
of the classical 'Hungarian' algorithm may be used .
9.4 Scaling Problem
Given transfer function G e C (s), for a linear square

system with zero initial condition
y(s) = Gu(s)
where y e Cnxl is the output and u e Cnxl is the input

and s is the Laplace Transform. If the outputs y. are
each scaled by factor s. respectively (i = l,2,...,n),
then the transfer function is simply
Z = SG
where Z e Cnxn(s), and S = d i a g l s ^ , s ± > 0. The fact

that dominance is not 'scale invariant' has at least repre-
sented a philosophical criticism of dominance based approaches.
Again the problem may be easy for n = 2, but rapidly be-

comes difficult for n > 2. Since the worst of the column
dominances represents a measure of design difficulty, it is
again natural to adopt a Chebyshev norm optimality criteria
and seek
min max{J?(Z)}
[s] i *
166
For convenience, introducing the matrix M
otherwise
and define n = max{J?(Z)}
(and taking into account the requirement s. > 0, i = 1,2,

, . , n ) , we seek a solution of the problem
P^ '• min(n}
s
subject to: £ m..s. <_ n s. ; i = l,2,...,n
31 3 1
s.>0,
= 1
where the last equality is introduced to remove indetermin-

acy with respect to scaling in S.
A simple procedure for solving the optimization problem is

as follows:
Data: e > 0
Step 0: Set i = 1; n, = max J m.
i jl D
Step 1: Solve min{z }

s
Subject to: (MT-n^

n
I s3 = 1
j=l
Sj >_ 0, j = 1,2, ... ,n
z£ >_ 0
£
Step 2: d = n1/2 ; if d < e then stop.
167
Step 3: If zz = 0 then n £ + 1 = n£ - d
Step 4: £ = £+1 . Go to Step 1.
The scalar e determines the absolute accuracy to which

T) is determined. Gershgorin's theorem ensures n-, > n in
Step 0 {r\, s denote optimal solution to this problem).
Step 1 involves solution of a standard linear program. Step
3 represents a standard binary search procedure on the inter-
val [0,n] in which successive bisection search is employed
to more closely locate n.
9.5 Compensator Design
Given Q e Cnxn, we seek a real compensator K e Rnxn

such that
Z = QK
has a minimum normalised column dominance. We may without

loss of generality, drop the suffix i and letting it equal
to 1, consider n separate problems (P?) of the form
[K]
z = Qk
l|k|| = 1
with z e CnXl, k e Cnxl and if z. = a.+ig.
• j I 2 4 = /a?j+ B?,
p
j j = l,2,...,n and
c n
6"
This is the 'single frequency1 version, the multiple frequen-

cy case can be stated as
: mln
k
where w., i = l,...,r represent selection frequencies
168
covering the system bandwidth.
The single frequency problem can be trivially re-expressed
P-3 : min n
3
[k]
«2 ~ T)\ l Z il I 1 ° (9.4a)
Z = QK (9.4b)
||k|| = 1 (9.4c)
A number of procedures may be used to solve this problem.

A simple approach which has proved effective is
(Conceptual)
Data: e > 0
Step 0: Set £ = 1, n, = Sîk), k i = 1, i = l,...,n
Step 1: If inequalities (9.4) are consistent, set d = r^/2
n = n - d
= n +
otherwise set ô+i c ^
Step 2: If d < e stop, otherwise set £ = £+1.
Return to Step 1.
The initial step ensures that n = n-j^ yields a consistent

set of inequalities (9.4a). Step 1 employs a simple binary
search on [ O , ^ ] . The consistency test in Step 1 is the
only difficult step. It is well known that feasible direc-
tion algorithms may be used to solve sets of inequalities
such as (9.4a), but that such algorithms may be slow . A
recently developed algorithm which employs both a Newton like
step with quadratic convergence together with a first order
step which yields finite convergence to a feasible point 2 has
been utilised with considerable success. If the Newton and the
first order steps fail, we assume the inequalities are incon-
sistent. More reliable behaviour and faster convergence was
obtained when equality constraint (9.4c) was replaced by in-
equality constraints 1 <^ | |K| | <^ L say, where L >>1; this is
admissible in view of the optimality criteria being first
169
order homogeneous in K. Also it is always advantageous to

substitute (9.4b) into (9.4a) and hence reduce the problem to
checking two simple inequalities for consistency.
It should be remarked that the multiple frecmencv case in-

volves a minmax criteria and that powerful methods have also
o
been developed for this general class of design problem .
These methods again utilise linear programming subroutines;
for the multiple frequency case they behave like Newton meth-
ods and display quadratic convergence (whereas in the single
frequency case they are analogous to steepest ascent and dis-
play only linear convergence).
Finally, mention should be made of special procedures tail-
8
ored to the problem . These have proved reliable but are com-
putationally demanding.
9.6 Design Example
Here is an example of applying these new algorithms devel-
oped earlier to Direct Nyquist Method (DNA). The linear jet
engine model chosen is a 5-states, 3-inputs, 3-outputs
simplified model . Only a simplified actuator dynamic is
included, sensor dynamic is neglected and all output vari-
ables are assumed measurable, and the plant transfer matrix
is normalised by the appropriate input/output nominal oper-
ating factors .
Achieving dominance (a multiple frequency case):
Step 1: Choose a set of frequency which will approximately

cover the whole operating frequency range of the plant to be
controlled.
Step 2: With A-l, to compute a permutation matrix

P : (G, = P,G) yields
' 0 0 1
p 0 1 0
l =
1 0 0_
The origin maximum dominance before is:
6C = (51.33, 527.4, 0.325)

max
170
The improved maximum dominance after permuting is:
6
L V = (0.0968, 527.4, 16.06)
Step 3: By A-2, output scaling algorithms with the same

set of frequency: (G2 = SJ^GJ.S^ ) .
The output rescaling matrix is
S = diag{0.14622, 0.9134, 1.9403} ,
improved dominance measure is
L = < 7 - 92 > 0.285, 7.83}
Step 4: Re-permuting output again: (G^ = P? G 2^ gives

1 0 0
P
2 0 0 1
0 1 0
The improved dominance measure is
(7 92 6
L = ' ' ' 85 ' 0-194)
Step 5: Rescaling again with A-2 (G. = S-G^S- ) gives
S2 = d i a g U . 72190, 1.14100, 0.13700}
and improved dominance measure
6C = {1.85, 1.85, 1.85}

max
Step 6: Compute a real pre-compensator by A-3

(G5 = G 4 K 1 ) gives
1 5.12849E-3 -6.2573E-3
K
l 35.131 -1 1.51895
1 -0.13686 1
and the improved dominance measure is

171
(0
= '9'
Up to this stage the dominance level of the jet engine is

acceptable. Columns 1 and 2 are both open and closed-loop
dominant, column 3 is closed-loop dominant with forward gain
less than 1.5. Figure 9.3 shows the DNA plots.
The process of Step 1 to Step 5 can be summarised as

follows:
G^ = U G L
where U = S
2P2S1P1 and L =
with
0 0 2.51990E-2
U = 0. 5707 0 0
- 0 0. 330706 0
" 39 .6841 0 .2035 -0.24831

L = 12 .7578 -0 .3631 0.55160
14 .5903 -1 .9969 14.5903
The structure of the controlled system is as shown in Fig.

9.1 . In view of steady state responses and gain margins, a
combined lead-compensator and proportional-integral controll-
er is used (K c ).
The forward gain factor for stable and good responses is:
K = diag{50, 0.6, 1.8}

a
The closed-loop step responses are shown in Fig. 9.2 .

172
u y
U Kc L G
J
w
Fig. 9.1
060 120 ISO 2 41
i\r\ _ . . .
F i g . 9.2
3.00 .3.00 3.00
F i g . 9.3
173
9.7 Discussion
The algorithms described can all be extended to

4
(i) Multiple frequency case .
(ii) 'Parameterised' controllers Kc of forms such as
KC = K + K I /S
where K , K t i? nxn are multivariable proportional

and integral terms.
(iii) Use of inequalities on K. ., 1 •< ij <_ n.
1
(iv) 'Sparse controllers where we seek to satisfy the
important practical criterion that the controllers
should be of sample 'structure1 and not employ com-
plexly interconnected controllers.
(v) Solution of large sets of equations by 'tearing1.
The 'tear' inputs and outputs can be 'compensated'
using the methods described, and dominance is a key
property for ensuring the convergence of iterated
'torn' equations.
Research is being undertaken in all these areas.
References
1. ROSENBROCK, H.H. : "Computer aided control system

design", Academic Press (1974).
2. MAYNE, D.Q., E. POLAK, A.J. HEUNIS : "Solving Non-
linear Inequalities in a Finite Number of Iterations",
J. of Optimisation Theory and Applications, 33, No. 2,
(1981).
3. KANTOR, J.C. and R.P. ANDRES : "A note on the exten-
sion of Rosenbrock's Nyquist array technique to a larger
class of transfer function matrices", Int. J. Control,
Vol. 30, No. 3, (1979), pp. 387-93.
4. BRYANT, G.F. : "Dominance Optimisation and Multivari-
able Design", Departmental Report, Dept. of Electrical
Engineering, IAG Group, Imperial College (1981).
5. SAIN, M.K., J.C. PECHKOWSKI, J.L. MELSA (Eds.)
"Alternatives for linear multivariable control",
National Engineering Consortium, Inc., Chicago, (1978).
6. LAWLER, G. : "Combinatorial Optimisation",
Holt Reineholt (1976).
174
7. POLAK, E. : "Computational Methods in Optimisation"/

Academic Press, New York, (1971).
8. BRYANT, G.F., L.F. YEUNG : "Dominance Optimisation",

IEE Control and its Applications Conference, Warwick,
England, (1981).
175
PROBLEMS
P.9.1 Given the matrix

3 4 6 2 100
400 30 200 1.0 0
G = 0 1 4 60 10
4 60 4 80 1
0 1 1 30 1
Find the corresponding column dominance matrix D
134.6 23 34 85.5 .12

0.017 2.2 0.075 172 oo
oo 95 52.75 1.88 10.20

100.75 0.60 52.75 1.162 111.0
oo 95 214 4.77 111.0
Show that the optimum permutation input + output is
1 + 3 , 2 + 1 ,3 + 5 ,4 + 2 , 5 + 4
and confirm that the smallest five elements in Dc do not

give an optimal matching. [Hint: A bigger covering set is
needed.]
P.9.2 Given a system matrix G e R nxn where n = 2
1 2
3 4
Find a scaling matrix S = diagis^}, i = l,...,n and

s. >_ 0 such that it solves the following problem:
P : min max{6 < T(S.G ) }

D
s j
[Hint: (i) By showing that
n s |g i .|
176
Then let n = max{J?(S.G)}
and solve the inequalities for n and S
n
n si
^ 1 I T~ m i-i' j = 1
i=l s j 1 D
(ii) Hence by letting n = 2, and s = 1, show that
S
2
Chapter 10
Pole assignment
Professor N. MUNRO
Synopsis
The problem of pole assignment using state feedback or out-

put feedback is examined. An algorithm for the design of full
rank minimum degree output-feedback compensators is presented
and results in a unification of previously obtained results.
10.1 Introduction
The problem of pole assignment has been extensively studied

by many research workers over the last decade 1 " ' rl
~ . It
is simply concerned with moving the poles (or eigenvalues) of
a given time-invariant linear system to a specified set of
locations in the s-plane (subject to complex pairing) by means
of state or output feedback. The state-feedback approach is
well established, however the output-feedback case is still
18
being explored as a research area .
A large number of algorithms exist for the solution of this

problem. Fortunately, these algorithms can be categorised by
the characteristics of the approaches used, as shown in Figure
1. Both the state-feedback methods and the output-feedback
methods result in compensator matrices which are broadly
speaking either dyadic (i.e. have rank equal to one) or have
full rank. Although the dyadic algorithms have considerable
elegance and simplicity, the resulting closed-loop systems
have poor disturbance rejection properties compared with their
19
full-rank counterparts .
Here, we shall review several existing algorithms and then

consider the development of an algorithm which uses the dyadic
178
pole assignment
state ^ output
Feedback feedback
full rank dyadic

observer non observer
I
canonical non-canonical
r dyadic
full rank
I
spectral mapping
constant
I
dynamic
Fig 10.1 Pole Assignment Algorithms
design mechanism to generate full-rank minimum degree output-

feedback compensators/ which allow arbitrary pole assignment.
The approach to be presented is 'consistent1 with all prev-
iously established results and provides a unifying framework.
The case of state-feedback is readily shown to be a special
case of output-feedback.
The problem to be considered is as follows, Given a system

described in state-space form as
Ax + Bu (10.1)
y = Cx (10.2)
where A,B,C are real matrices with dimensions n*n, nxm, £*n,
respectively, or by its corresponding transfer-function matrix
representation
179
G(s) = C(sl-A)~1 B (10.3)
when and how can we apply feedback to this system such that
the resulting closed-loop system has a desired set of arbit-
rary poles, or eigenvalues (subject to complex pairing). First,
we consider some of the well established results, starting
with the state-feedback case.
10.2 State-Feedback Algorithms
Using a feedback law of the form
u = r - K x (10.4)
where r is an mxl vector of reference inputs and K is an

mxn real matrix, when and how can we determine K such that
the resulting closed-loop system
x = (A-BK)x + Br (10.5)
has a desired set of eigenvalues {Yj^ /* i = l,n ?
The question when was answered by Wonham and the require-

ment is that the pair [A,B] is completely controllable.
The question of how can be examined in several ways, as foll-
ows.
10.2.1 Dyadic designs
Under the heading of dyadic designs (see Fig. 1 ) , the spec-

tral approach essentially requires the calculation of the
eigenvalues (X i >, and the associated eigenvectors (wi> and
reciprocal eigenvectors {v.} of the matrix
Significant A.
2-5
contributions have been made in this area by several workers .
Here, the matrix K is considered as the outer product of
two vectors; i.e.
K = fmt (10.6)
where f is mxl and m is l*n.

180
The problem of simultaneously determining the (m+n-1) free

elements in K is nonlinear. However, by pre-assigning
values to the elements of one vector, the determination of the
remaining elements results in a linear problem . Convention-
ally, the vector f; is chosen such that the resulting pseudo
single-input system [A,Bf_] is completely controllable. Then
for a desired closed-loop pole set (y^)/ the vector m is
given by
m1 = I 6. v* (10.7)
1 x
where q = number of poles to be altered, the v. are the

reciprocal eigenvectors associated with those eigenvalues
{X^} to be altered, and the scalars 6i are weighting fac-
tors calculated as
q
n
(10.8)
P n U.-X.)
1 3
j=l
where the scalars p. are defined by the inner product
p = <
i -i'B i > (10.9)
We note that if the \^ or y^ are complex then the result-

ing 6^ will be complex. However, m will be real if
these values correspond to complex-conjugate pole-pairs.
When the open-loop system has any eigenvalues of multipli-

city greater than 1, the above algorithm breaks down, since
the denominator term in equation (10.8) becomes zero for
certain 6^. However, a modification to this algorithm to
cater for this latter situation has been developed and is
described in Reference [4].
In contrast to the above method, the mapping approach (see

Fig. 1) due to Young and Willems does not require the spec-
181
tral decomposition of the open-loop system. Once the desired

closed-loop system poles have been specified, the correspon-
ding closed—loop system characteristic polynomial Â(S)'
where
A d (s) = s n + d 1 s n ~ 1 + ... + d n (10.10)
can be calculated. If the open-loop system characteristic

polynomial is defined as
AQ(S) =
(10.11)
5
then we can define a difference polynomial (s) as
Ms) Ad(s) - A Q ( S )
= sn (d 1 -a 1 )s n " 1 ... (d n -a n ) (10.12)
Now, if the vector f_ is chosen such that [A,I3f] is com-

pletely controllable, the required vector m is determined
as
-1 i
X
X (10.13)
where = [Bf,ABf,...,An~1Bf] (10.14)
0
1
a,
X = (10.15)
a a
n-2 n-l
a a
n-l n-2
and [d 1 -a 1 ,d 2 -a 2 ,...,d n -a n ] t (10.16)
We note that by definition the nxn matrix X is nonsingular,

and so is the controllability matrix *c. Also, these need
182
only be calculated once for a variety of closed-loop system

pole specifications.
Another way of looking at this latter algorithm, which will

set the scene for the results in the later sections, is from
the frequency-domain point of view.
Let r(s) = adj(sI-A) (10.17)
and bf = Bf (10.18)
then, after ^ has been chosen to give complete controllab-

ility in the resulting single-input system, we have
N 2 (s)
(10.19)
N n (s)
where the N.(s) = + ... + (10.20)
and A (s) is as defined by equation (10.11).

A
We also have d (s) defined as
A,(s) = |sI-A+BK|
|sI-A+bfm
(10.21)
which can equally be written as
A d (s) = (s) + m1" r(s) b f (10.22)

n
A
i.e. d(s) = A
o
(s) +
^ m
i N
i^s) (10.23)
Using (10.12) we can consider the pole assignment problem as

183
consisting of solving the equation
n
I m.N.(s) = 6(s) (10.24)
x ±
i=l
for the parameters of the vector m.
As we shall see later, equation (10.24) is the heart of the

dyadic approaches to pole assignment.
Equating coefficients of like powers in s on both sides

of equation (10.24), we obtain
m
l
1 1
1
=
1
(10.25)
1 1
6
nn • n •
i.e. Xm = 6 (10.26)
where the difference vector 6_ is defined by equation (10.16)

and where the nxn matrix X always has full rank; i.e. the
equations are consistent, Therefore, the desired vector m
is given by
= X-1 (10.27)
for any 6_ .
10.2.2 Full-rank designs
The conditions under which a full-rank state-feedback com-

pensator K can be designed are exactly as before; namely,
that the open-loop system [A,B] is completely controllable.
Here, again, we shall consider two distinct approaches.
In the first approach we assume that the original system

matrices A and B have been transformed into their so-
called controllable canonical form yielding A and B
defined as
-1 (10.28)
A = TAT
184
and TBQ (10.29)
where T is a nonsingular transformation matrix formed from

specifically selected and ordered columns of the system con-
trollability matrix [B,AB,...,An~ B ] , and Q is a nonsin-
gular matrix formed from the nonzero rows of the matrix (TB).
Under this transformation, A has the particular block

structure
11
A = (10.30)
I A ml A
mm J
where the diagonal blocks have the companion form
0 1 0
0 0 1
(10.31)
0 0 1
X X X
with dimensions v • *v • ,
where the v. are the controll
ability indices of the system; i.e.
and where the off-diagonal blocks , i have the form
0
(10.32)
X X
with dimensions v^*v>, i,j = l,m, i^j. The matrix B has

the particular block structure
185
B = (10.33)
where
(10.34)
0 ... 0 1 0 ... 0
column i
has dimensions (v.xm), In equations (10.31)- (10.34), the

X's may be nonzero coefficients.
Given a system in this special form, the feedback matrix K

required to achieve a desired closed-loop pole set {y^l is
chosen such that (A-BK) becomes block upper or lower trian-
gular, or both, and such that the resulting diagonal blocks
have nonzero coefficients which satisfy
in _ n
n det[si-Ai:L + (BK)i:L] = n (s- Yi ) (10.35)
The desired full-rank feedback matrix K in the original

system basis is then given by
K = K T (10.36)
Consider now the situation where a full-rank feedback

matrix K is to be determined as a sum of dyads; i.e. let
K - (10.37)
where the f. are mxl column vectors, the m are l*n

row vectors, and y = min{m,n} = m. K can also be expressed
as the matrix product
K = FM (10.38)
186
where F is an mxm matrix whose columns are the f # and

M is an mxn matrix whose rows are the m . We note that
K will have full rank m, if and only if both F and M
have maximum rank.
Consider the following algorithm:
(1) Suppose we choose _f such that it selects input 1

of the given system; i.e. f ^ = [1 0 ... 0 ] t . Then,
since input 1 must influence g, > 1 poles of the system,
we can determine m (1) t o reposition q.^ poles at desired
locations.
Let the resulting closed-loop system
x = (A - Bf (1) rn (1) )x + Br (10.39)
be considered as
x(s) = [r ( 1 ) (s)/d (1) (s)]r(s) (10.40)
where r (1) (s) = adj(A-Bf (1) m (1) ) (10.41)
and d (s) is the new characteristic polynomial.
(2) Since r (1) (s) only has rank one, we can define
s=^
as
f(2) = [fx f 2 0 ... 0)t (10.42)
and can solve the equations
rU)(yi)f(2) = O (10.43)
for f 2 ^ 0 such that at least ( q ^ l ) ; or q1 if q1 = 1;

if the poles assigned in step (1) become uncontrollable. We
also note that due to the system controllability properties,
this choice of £ will influence q 2 _> 1 more poles of
the resulting single-input system given by
187
f(2) andd(1)(s)
(m) Continuing in the above manner, it can be

shown that this procedure allows us to shift all n poles of
the original system to a desired set {y^}; i = l,n; in m
steps.
We note, that at each stage, those poles which have not

been rendered uncontrollable with respect to that stage of
the design procedure, and which are not being moved to de-
sired locations will move in an arbitrary manner. However,
after m stages, all n poles have been moved to the desired
locations.
The matrix F generated in this manner will have full rank

m, and it can also be shown that the corresponding matrix M
will simultaneously have full rank n. Thus, K will have
full rank; a desired objective; and will also provide the
appropriate pole shifting action.
10.3 Output-Feedback Algorithms
Here, we shall adopt the frequency-domain approach from

the start. Thus, given a system described by the £xm trans-
fer-function matrix G(s), and using a feedback law
u(s) = r(s) - F(s) y(s) (10.44)
when and how can the resulting closed-loop system
H(s) = [I+G(s)F(s)]" 1 G(s) (10.45)
be made to have a pre-assigned set of poles {y-}; i = l,n ?

Complete pole assignment is only possible using output-feed-
er
back when G(s) has arisen from a least-order system ; i.e.
the corresponding state-space triple [A,B,C] is completely
controllable and observable.
The following questions arise naturally at this point:
188
(1) What is the minimum degree of the compensator matrix

F(s) which will allow arbitrary pole locations in the
resulting closed-loop system?
(2) What are the parameters of this F(s) ?
(3) What is the rank of F(s) ?
Q
Davison has shown that, using a constant feedback matrix

F, max{m,&} poles can be assigned arbitrarily close to de-
sired values. However, the remaining n-max{m,U poles move
arbitrarily. Nevertheless, this result may prove useful in
situations where only a few of the open-loop system poles
are to be moved to more desirable locations. It is also
known that by the use of a Luenberger observer of order
(n-£), or a Kalman-Bucy filter of order (n+), a good es-
timate x of the system state-vector x can be reconstruc-
ted. This latter approach then allows any of the well known
state-feedback algorithms to be applied using x. Further,
if a dyadic state-feedback approach is to be used then we
need only estimate a = m x, and this can be done under
12 ~~ ~"
favourable conditions using an observer of order r = v - 1 ,
where v Q is the observability index for the system.
For those situations where a dynamic compensator F(s) is

squired, Brasch ar
required, and Pearson have shown that a dynamic com-
pensator of degree
r = min{v Q -l, v c ~l} (10.46)
where vc is the controllability index for the system, is

sufficient (but not necessary) to allow arbitrary pole ass-
ignment. The design of an observer is implicit in this
approach, and the resulting feedback is essentially dyadic.
A similar result has been obtained by Chen and Hsu in the
14
frequency-domain
Another result, obtained by Davison and Wang results in

a compensator of rank two. These authors have shown that if
m + I - 1 > n (10.47)
189
then a constant compensator F exists which allows the

closed-loop system poles to be assigned arbitrarily close to
a desired set {y . }.
The final result to be considered in this section is due

to Munro and Novin Hirbod which states that if m + £ - 1
>_ n a full-rank constant compensator F can be determined,
from a minimal sequence of dyads, to assign the closed-loop
system poles arbitrarily close to a predefined set (Y^.
Further, if m + £ - 1 < n then a full-rank dynamic compen-
sator F(s) with degree r given by
n- (10.48)
max{m,£}
can be constructed from a minimal sequence of dyads, which

will allow arbitrary pole assignment.
In the following, we shall consider the design of minimum

degree dyadic dynamic feedback compensators, and finally the
design of minimum degree full-rank dynamic feedback compen-
sators.
10.3.1 Dyadic designs
Given a strictly-proper £*m transfer-function matrix G(s),

let the required compensator F(s) be considered as the
outer product of two vectors f_ and m (s) ; i.e.
F(s) = f nT(s) (10.49)
As before, the constant vector f is chosen such that
g(s) = G(s) f (10.50)
is completely controllable. Let g(s) be expressed as
N 1 (s)
£(s) = A o (s)
(10.51)
Vs)
190
where _ _ ,
A (s) = s + a.s + ... + a (10.DZ)
is the characteristic polynomial of G(s), and the
N ± (s) = 8 i l s n " 1 + 3 ± 2 s n " 2 + ... + 3 i n (10.53)
Now, let m (s) have the form
^t(s) =
A \s) CM
l(s) '•• * V S ) J (10.54)
where A (s) = s r + y . s ^ 1 + ... + y r (10.55)

C J- XT
is the characteristic polynomial of m (s) , and the
M
i(s) = 6
iosr + G
il sr "" 1 +
""• + e
ir (10-56
14
Then, Chen and Hsu have shown that the resulting closed-
loop system characteristic polynomial A, (s), where
A d (s) = s n + r + d 1 s n + r " 1 + ... + d n + r (10.57)
is given by the relationship
A d (s) = A Q (s)A c (s) + I N i (s)M i (s) (10.58)
which is the heart of the dyadic approach to pole assignment

using output-feedback.
If we now define a difference polynomial 6 (s) as
6r(s) = A d (s) - A Q (s).s r (10.59)
then (10.58) can be written as

£
{ I N i (s)M(s)} + A Q (s)(A c (s)-s r ) = 6r(s) (10.60)
Equating coefficients of like powers of s on both sides

of equation (10.60) yields the vector-matrix equations
191
x
r Er = £r (10.61)
where X is an (n+r)x((r+1)(£+1)-1) matrix constructed from

the coefficients a± of A Q (s) and the Bi- of the Ni(s)/
the vector £r contains the unknown parameters y . and 6..
defining F(s), and the difference vector 6_ contains the
coefficients 6. of 6 (s) as defined by equation (10.59).
The elements of X r , g> , _6_r are shown explicitly in Appendix
1.
The pole assignment problem can now be stated as when can

equation (10.61) be solved for the vector £ ? A necessary
and sufficient condition for equation (10.61) to have a sol-
ution for the U(r+l)+r) vector £r is that
9X
X =
rr ^r -^-r (10.62)
g
l 17
where X is a g -inverse of the matrix Xr defined by
X
rXr r = X
r (10.63)
If equation (10.61) is consistent, the general solution for

£r is
g g
l l
pr = X r X i r + (I - X r X Xr)z^ (10.64)
where z^ is an arbitrary [£(r+l)+r] x l vector. Since £

is arbitrary, it can be chosen as the zero vector yielding
91
£r = X r 1 lr (10.65)
We note that if the consistency condition given by equation

(10.62) is not satisfied for a particular g1~inverse of Xr,
then it will not be satisfied for any other g^-inverse of Xr-
It has already been mentioned that if r is chosen as
rm = min{v -l,vc-l> (10.66)
ben arbitrary pole assignment can be achieved. This result

192
can now be simply explained, since for r defined by

(10.66) we obtain under these special conditions
(10.67)
This is due to the rank of X , for this particular value

of r, being such that a right-inverse of Xr exists, de-
fined by
Xr1 = X^[X r X ^ f 1 (10.68)
It is also interesting to note that for arbitrary pole

assignment, r is bounded as
r nl (10 69)
) i ± -
We can now see that for the case of state-feedback, since
then i = n, equation (10.69) yields r = 0; i.e. F is a
constant matrix.
However, equation (10.69) does not allow for known excep-

tions, where arbitrary pole assignment can be achieved with a
compensator F(s) of order r < r . To allow for these sit-
uations, we state that the necessary and sufficient conditions
on r are that it is chosen such that
r = p(Xr,6_r) (10.70)
where p{*} means rank.
10.3.2 Full-rank designs 1 6

Suppose that the required compensator F, constant or dyn
amic, is constructed from a minimal sequence of dyads as
F = I f(i) m ( i ) (10.71)
i=l
where f^1' is an P I vector, m is a lx£ vector, and
]i = min{m,U (10.72)
193
Then, equation (10.71) can be equally written as
F * Fc F Q (10.73)
where F is an mxm matrix whose columns are the vectors

f_ , and where FQ is an m n matrix whose rows are the
vectors m .
We can ensure that for the case where m < £ the matrix F
has full-rank by its construction, but F will have full
rank if and only if FQ and FQ both have full rank. It is
tedious but straightforward to show that if F is gener-
ated by the method shown in the following algorithm, then the
corresponding matrix F simultaneously has full rank. The
case where m > £ is obtained by duality.
Algorithm
(1) Choose f_ ' such that it picks up input 1, say, of the

system; i.e. let f/ ' = [1 0 ... 0] ; and determine the
corresponding vector f to place as many poles as poss-
ible in desired locations. Since the system is completely
controllable, the pseudo-input vector b = Bf^ will
— —c
influence n, modes of the system where 1 < n, < n. Also, the
system [A, b (1) , C] has i^ linearly independent outputs and
one input and, in fact, only, q, = min{£..,n,} poles can be
assigned to specific locations at this stage.
(2) For f^ ' = [f f 2 0 ... 0 ] t , say, determine ±l and

f2; f 2 i- 0; to render as many as possible of the q1 poles
to
assigned in stage 1 uncontrollable, and determine JLQ
move at least one more pole to a desired location.
Let n0
be the number of modes influenced by b = Bf
(2) ~
where 1 < n o < n - 1 . T h e n , f is determined to move qo
2'n2-^ poles to desired locations by solving the equa-
tion (10.74) :
A ( 2 ) (s) = A ( 1 ) (s) + f < 2 ) ^ 1 )

^ 1 )
« 2 )
194
where
= A Q (s) + f<^1)C[adj(sI-A)]Bf<l1) (10.75)
Note that the elements f1 and of ^ are determined

by solving
r(1>(s)f<2> = o
= [adj (10.76)
at those roots of A (s) which correspond to eigenvalues

to be made uncontrollable.
(m) For f^ = [f11 f 2 ... determine f 1 to f m ;

f to n
™ ^ ®' °ld (m-1) poles in desired locations, and
determine f^ ' to place minU,n-m+l} poles in de-
sired locations.
Provided m+£-l >_ n, this procedure can always be carried

out. Now, if m+£-l < n, the first m-1 stages of this pro-
cedure are carried out as before, and then the final stage
is carried out using a dynamic dyadic compensator of degree
r, as defined by equation (10.48) . The resulting full-
rank compensator F(s) will have degree r.
10.3.3 Example
Consider a system with transfer-function matrix
s2(s+l)
G(s) = ±j (10.77)
s 0 sz(s+l)
which has eigenvalues {A.} = {0, 0, 0, 0}, m = £ 2,

n = 4, and has two transmission zeros at s = -1.
(1) Choose f(1) = [1 OJ1^ then

—"C
(10.78)
195
So, the system seen through input 1 has two poles at the
origin and one zero at -1. Now, since n 1 = 2 and £ = 1,
we get q = 1. Thus, we determine f (1) to move one pole
to a new location at s = - 2 , say. This is achieved by
fi 1} = [4 0]; i.e.
(10.79)
which results in the closed-loop system pole set {A.} =

{0, 0, - 2 , -2}. So, we have not just moved one pole to
s = -2 but have in fact moved two there. This situation is
fully predictable from the resulting closed-loop system root-
locus plot.
(2) Now, consider {A.} = [{y.} of step 1 ] . Since m+£-l =

3 < n, a dynamic compensator of degree r = 1, as predicted
by equation (10.48), is required. Let
f ( 2 ) (s) = [e ,s + '21 (10.80)

S+Y, io n
and determine f' ' = [f f?^ such that one of the poles
moved to s = -2 Cin step -L1 becomes uncontrollable.
After step 1, the resulting closed-loop system is
s (s+l) (s+l)
(10.81)
s2(s+2)2 (s+l)(s+2)'
and r(1)(-2)f<2) = o (10.82)
yields f(2) = [1 -4] 1 and so
(s-2)
r<2) (s) = (s+2) (s + l) (10.83)
(s+2)
To place the remaining poles at say s = - 1 , -3, - 4 , - 5 ,

the required vector f (s) is
(s) = 37)/2 (-97/8)] (10.84)

196
i.e.
(20s+37)/2 (-97/8)
F 2 (s) = (10.85)
-2(20s+37) (97/2)
A root-locus plot of the resulting closed-loop system pole

behaviour during this step of the compensator design shows
that the effect of feedback on the origin poles is stabiliz-
ing, as was also the case with loop 1 determined in the first
step.
The overall full-rank feedback compensator F(s) = F +F 2 (s)

which results in the desired closed-loop system poles is
, , (28s+45)/2 -97/2
F(s) = -—- I (10.86)
1
-2 (20s+37) 97/2 j
10.4 Concluding Remarks
The full rank output-feedback algorithm, described in the

previous section, provides a systematic and unified approach
to the problem of pole assignment using state or output feed-
back. Both dyadic and full-rank compensators can be readily
designed. The computational procedures required are straight-
forward, and advantage can be taken of 'singular-value
decomposition1 techniques to solve the resulting sets of
equations.
It is also interesting to note that the same full-rank out-

put-feedback algorithm can be directly used to design propor-
tional-plus-integral action controllers which allow arbitrary
pole assignment in the resulting closed-loop system . Here,
the original system [A,B,C]-matrices are augmented to intro-
duce the desired integral action into the open-loop system
description (see Seraji ) , and the full-rank pole assignment
algorithm (of Section 10.3) is then applied to the resulting
noncyclic open-loop system.
Finally, it is interesting to note that although a great

deal of effort has been invested in the pole assignment prob-
197
lem little attention has been given to the numerical proper-

ties of these algorithms. However, recently an interest has
been shown in this area by numerical analysts and some new
results are beginning to emerge
References
1. WONHAM, W.M. : 'On pole assignment of multi-input con-

trollable linear systems', IEEE Trans. Aut. Control,
AC-12, 1967, pp. 680-665.
2. SIMON, J.D. : 'Theory and application of modal control',
Systems Research Centre Report, SRC 104-A-67-46, Case
Institute of Technology, U.S.A.
3. MUNRO, N. : 'Computer aided design of multivariable
control systems', PhD Thesis, 1969, UMIST, Manchester,
England.
4. RETALLACK, D.G. and A.G.J. MACFARLANE : 'Pole shifting
techniques for multivariable feedback systems', Proc.
IEE, 1970, Vol. 117, pp. 1037-1038.
5. PORTER, B. and T.R. CROSSLEY : Modal Control : Theory
and Applications (Taylor and Francis, London, 1972).
6. YOUNG, P.C. and J.C. WILLEMS : 'An approach to the
linear multivariable servomechanism problem', Int. J.
Control, 1972, Vol. 15, No.5, pp. 961-979.
7. MUNRO, N. and A.I. VARDULAKIS : 'Pole-shifting using
output feedback 1 , Int. J. Control, Vol. 18, 1973,
pp. 1267-1273.
8. ROSENBROCK, H.H. : State Space and Multivariable
Theory, (Nelson, 1970).
9. DAVISON, E.J. : 'On pole assignment in linear systems
with incomplete state-feedback', IEEE Trans. Aut. Con-
trol, AC-15, 1970, pp. 348-351.
10. LUENBERGER, D.G. : 'An introduction to observers',
IEEE Trans. Aut. Control, AC-16, 1971, pp. 596-602.
11. BUCY, R.S. and P.D. JOSEPH : 'Filtering for stochastic
processes with applications to guidance', Wiley Inter-
science (New York, 1968).
12. MURDOCH, P. : 'Design of degenerate observers', IEEE
Trans. Aut. Control, Vol. AC-19, 1974, pp. 441-442.
13. BRASCH, F.M. and J.B. PEARSON : 'Pole-placement using
dynamic compensators', IEEE Trans. Aut. Control, AC-15,
1970, pp. 34-43.
14. CHEN, C.T. and C.H. HSU : 'Design of dynamic compensa-
tors for multivariable systems', JACC, 1971, pp.893-900.
15. DAVISON, E.J. and S.H. WANG : 'On pole assignment in
linear multivariable systems using output feedback',
IEEE Trans. Aut. Control, Vol. AC-20, 1975, pp. 516-518.
198
16. MUNRO, N. and S. NOVTN-HIRBOD : 'Pole assignment

using full-rank output-feedback compensators 1 ,
Int. J. System Science, Vol. 10, No. 3, 1979, pp.285-306.
17. PRINGLE, R.M. and A.A. RAYNER : Generalized inverse
matrices with applications to statistics, (Griffin,
London, 1971).
18. WILLEMS, J.C. and W.K. HESSELINK : 'Generic proper-
ties of the pole placement problem 1 , 7th World
Congress of IFAC, Helsinki, 1978.
19. DANIEL, R.W. : 'Rank-deficient feedback and distur-
bance rejection 1 , Int. J. Control, Vol. 31, No. 3,
1980, pp. 547-555.
20. STRANG, G. : Linear algebra and its applications,
(Academic Press, 1976).
21. NOVIN-HIRBOD, S. : 'Pole assignment using proportion-
al-plus-integral output feedback control', Int. J.
Control, vol. 29, No. 6, (1979), pp. 103b-1046.
22. SERAJI, H. : 'On output feedback control of linear
multivariable systems', in Control System Design by
Pole-Zero Assignment, ed. by F. Fallside (Academic
Press, 1977).
23. FLETCHER, L.R. : 'An intermediate algorithm for pole
placement by output feedback in linear multivariable
control systems', Int. J. Control, Vol. 31, No. 6,
1980, pp. 1121-1136.
Appendix 1
The parameters in the vector matrix equation
Xr£r = «r
are shown explicitly below for the case where a dynamic

feedback compensator of degree r is required:
199
I -
•<f <=>
^• • • • * CD
I I
I • • "l 2
ca o
-f J
^ o
o ^
"o
O ~ -3 O ef o o o o
o o ;
tl-
200
POLE ASSIGNMENT - Problems
P.I Given a system described by the state-space equations

-2 1 0 0
x = -1 -2 x +
-2 0
with eigenvalues {A.} = {+j, - j , -2}
and reciprocal eigenvectors
5 2-j -6-2j
5 2+j -6+2j
0 - 4 2
Use the spectral approach to determine a suitable state-feed-

back matrix K, such that (A-BK) has eigenvalues {y^} =
{-2, - 2 , -2}.
P.2 Apply Young's algorithm to the system of Problem 1 to

obtain the same closed-loop pole set using state feedback.
P.3 Given a system in controllable canonical form as
0 1 0 0
80/9 26/3 10/3 x +

8/3 0 1/3
write down a full-rank state feedback matrix K such that

(A-BK) has eigenvalues {y^} = {-4, - 4 , - 4 } . Given that the
transformation matrices T and Q required to put the orig-
inal system into controllable canonical form are
0 0 1/3
T = 1 4/3 5/3
0 1 -2/3
1 4/3
and Q =
0 1
evaluate K in the original basis.

201
P.4 Given a system described by the transfer-function

matrix
s +6s+4
G(s) =
s 3 +6s 2 +7s+2 s+2
determine a dyadic output-feedback compensator such that the

closed-loop system plant poles are {Yi} = {-1, - 2 , -4}.
If the closed-loop system plant-pole locations required

are now changed to {y.} = (-1, - 2 , - 3 ) , determine a suit-
able dyadic output-feedback compensator.
P.5 Given
s2+s+l
G(s) = ±j
determine an appropriate full-rank output feedback compen-

sator such that the resulting closed-loop system poles are
{y±} = {-1, - 2 , -5}.
Chapter 11
Nonlinear systems
Dr. P. A. COOK
Synopsis
Some methods for analysing the behaviour of nonlinear dyn-

amical systems are discussed. Attention is concentrated on
the use of describing functions for predicting limit cycles
and Lyapunov's method for investigating stability. A brief
treatment of absolute stability criteria and their use is
also given.
11.1 Nonlinear Behaviour
The design of a control system is usually based upon a

linear model of the plant to be controlled, for the good
reason that the assumption of linearity makes the dynamical
behaviour much easier to analyse. Moreover, the resulting
design can be expected to be satisfactory at least in the
neighbourhood of the operating point around which the model
has been linearised. In practice, however, all systems are
to some extent nonlinear and, as a result, may exhibit forms
of behaviour which are not at all apparent from the study of
the linearised versions. Also, in some types of control
scheme (e.g. adaptive control, variable-structure systems),
nonlinear features are deliberately introduced in order to
improve the system performance. We begin, therefore, by
considering some of the more important effects which non-
linear properties can introduce.
(i) Dependence on operating point. When a nonlinear model

is linearised around a point representing steady-state oper-
ation, the result will in general depend upon the chosen
operating point. Consequently, a control scheme designed
203
for particular operating conditions may not be suitable in

other circumstances. To overcome this problem, we may need
to use a nonlinear controller, adapted to the nonlinearities
of the plant, or, more probably, a set of different linear
controllers to be used in different operating regions. Al-
ternatively, we may try to smooth out the nonlinearities by
some means, e.g. injecting a rapidly oscillating signal
(dither) together with the control input.
(ii) Stability. For a linear system, local stability

(i.e. in the immediate neighbourhood of an equilibrium point)
is equivalent to global stability, but this is not necessar-
ily true in the nonlinear case. Thus, even if a nonlinear
system is stable for small departures from equilibrium, it
may become unstable under larger perturbations, so we need
to develop methods for estimating the region in which stab-
ility is guaranteed (domain of attraction). Also, instead
of a unique equilibrium point, for given operating conditions,
there may be many equilibria with different stability pro-
perties.
(iii) Limit cycles. A typical feature of nonlinear sys-

tem behaviour is the occurrence of persistent oscillations,
called limit cycles. These are normally undesirable in a
control system and usually arise from the application of
excessive gain, so that the local equilibrium point becomes
unstable, although they can also be excited by large dis-
turbances or changes in operating conditions. It is there-
fore necessary to be able to predict when they are likely to
occur and to estimate their amplitudes and periods of oscill-
ation.
(iv) Frequency response. When a linear system is given a

sinusoidal input, its output is a sinusoid of the same fre-
quency. A nonlinear system, however, will generate harmon-
ics (at multiples of the input frequency) and possibly sub-
harmonics (at submultiples of this frequency) or combination
tones, arising from mixing of the input signal with internal
oscillations of the system. It is also possible for the
204
output amplitude to become a multivalued function of the in-

put frequency, so that discontinuous changes (jump phenomena)
may occur as the frequency is varied.
11.2 Fourier Series
If a linear element is subjected to a sinusoidal input,

the output (in the steady state) is also a sinusoid of the
same frequency, and this is an essential simplifying feature
in the study of linear systems. For a nonlinear element,
this no longer holds, and one way of describing the behaviour
of such an element is to analyse the output into its compon-
ent frequencies. In practice, it is often found that one
particular frequency, namely the input frequency, predomin-
ates in the output to such an extent that the other compon-
ents may be neglected, to a good approximationf in analysing
the behaviour of the system. This is the philosophy of the
describing function approach , also called the method of
harmonic balance. It amounts to treating nonlinear systems
as nearly as possible like linear systems, i.e. each fre-
quency is treated independently, but the nonlinear behaviour
shows up through the appearance of amplitude-dependent
coefficients multiplying signals which pass through nonlinear
elements. As a preliminary to the formulation of this method
we first consider the general question of the resolution of
a periodic signal into sinusoidal components (Fourier analy-
sis) .
Suppose a nonlinear element has a sinusoidal input
u(t) = U sin(a)t)
where U and co are positive constants. In the case of a

memoryless (i.e. single-valued) nonlinearity, or the simpler
types of nonlinearity with memory, the output y(t) will
have the same period as the input, but will contain harmon-
ics whose frequencies are multiples of the fundamental fre-
quency 03. It can be represented by the Fourier series
expansion
205
a
0
y(t) = y- + I {a cos (nut) + b sin (nut)}
n n
n=l
where the coefficients a , b are given by
27T
y(t) cos(na)t)d(a)t)
1 f 27T
bn = -TT y(t) sin(nwt)d(a)t)
JQ
as may be checked by substitution.
The constant term ^ Q / 2 arises from a bias, i.e. lack of

symmetry, in the nonlinearity, and will be absent if the non
linearity is symmetrical; that is to say, if the output is
an odd function of the input. In fact, this applies to all
the even harmonics, i.e. those for which n = even integer.
This is because, for a symmetrical nonlinearity, the output
has half-wave symmetry, i.e.
y(t + 1) = - y(t)
just like the input, and hence a and b vanish for even
n, leaving only the odd harmonics (n = odd integer).
Also, if the nonlinearity is memoryless, the output has

the property
y( t) =
^ - y{t)
again like the input, so that an = 0 for odd n and b n = 0

for even n. Hence, for a symmetrical memoryless nonlinear-
ity, the only nonzero coefficients are the bn for odd n.
Example
An ideal relay.
VUth input u, this element has output
y = sgn (u)
206
i.e.
1, u > 0 ;
y = 0, u = 0 ;
-1, u < 0 .
It is thus symmetrical and memoryless.
For the coefficients in the output, we have
ira - f 7T
0
c o s (ne)de
-I
ZTT
7T
cos(n6)de = 0
-1 -I
TT 2TT
Tib s i n (ne)de sin(n6)de
0 IT
2 (1 - COS nir)
n
so that
— , n odd ;
0 , n even.
Thus, the amplitude of the harmonics falls to zero as n+°° .

This is a general feature of Fourier expansions.
11.3 The Describing Function
The idea of this approximation is to pretend that only the

fundamental component in the Fourier series matters and all
higher harmonics can be neglected. V7e shall also assume
that our nonlinearities are symmetrical so that the constant
bias term does not arise, and hence we approximate the sig-
nal y(t) by
y Q (t) = a^^ cos(cot) + b^ sin(a)t)
This is actually the best linear approximation in a certain

sense, since, if we define the mean-square error A by
f 27r 2
A = (y(t) - y o (t) } d(wt)
and minimise this with respect to a1 and b^ (regarded as
207
variable parameters), we find
= 2
f
~ lYnW ~ y(t)> cos(o)t)d(o)t)
i Jo
2{TT a, - y(t) cos (wt) d (wt) }
1
Jo
~~ = 2 I {y o (t) - y(t)}sin(ait)d(a)t)
2TV
= 2{TT b - y(t) sin(ut)d(a)t) }
j
0
so that A has a stationary point (minimum) for the values

of a,, b, given above.
To compare the (approximate) output with the sinusoidal

input to a nonlinearity, we rewrite it as
y Q (t) = Y(U,a>) sin{wt + <J>(U,w)}

so that
ax = Y(U,a>) sin <)> (U,w)
b1 = Y(U,a)) cos 4> (U,a>)
where the amplitude Y and phase <j> are functions of the

input amplitude U and frequency w. VJe now define the
describing function
Y(
N(U,uO = â)) exp{j|(U)} 1 ai
so that
y Q (t) = Im{UN(U,uO
u(t) = Im (Uexp(jwt)}
The describing function thus represents the effective gain

|N| and phase-shift $ of the nonlinear element. For a
memoryless nonlinearity, it is real, since a1 = 0 and hence
there is no phase-shift (sin <j> = 0) .
208
11.3.1 Validity of the approximation
Using the describing function to represent any nonlinear

element, we can make an approximate calculation of the be-
haviour of a nonlinear system, in a way which is very much
simpler than solving the equations which the system actually
obeys. The justification for this approximation lies in the
smallness of the neglected harmonic terms, compared to the
fundamental. It is therefore important that the linear ele-
ments in the system should have the property of attenuating
high-frequency components more than low-frequency ones, i.e.
they should act as low-pass filters. This is commonly the
case, e.g. elements with transfer functions like
A i_ __L_
s ' 2 ' s+1 '
would usually not spoil the describing function predictions

too much. On the other hand, if the system contains an ele-
ment whose gain rises with frequency, such as
s
s+1
or an element with a sharp resonance, like
s2+l
then we must be wary about using the describing function, at

any rate for low frequencies.
It is possible to extend the describing function method so

as to take account of the presence of several components with
different frequencies, whether related or not, in the same
signal . In control system studies, however, this is
usually not necessary, at least in a semi-quantitative app-
roximation. On the other hand, it may be important to allow
for lack of symmetry in the nonlinearities (which is usually
present) by incorporating bias terms in the Fourier represen-
tations. Iê can do this if we replace the input signal by
209
u(t) = U sin(ut) + V
where V is a constant bias, and take the output signal

representation as
a
0
y o (t) = a1cos(a)t) + b1sin(o>t) + y-
with the coefficients given by the formulae in Section 11.2;

this again gives the best mean-square approximation possible
with this parametrisation. Then, by calculating a,, b, and
aQ in terms of U and V, we can obtain a generalisation
of the describing function, which is now replaced by two
functions (one of which may be complex to allow for a phase-
shift) , both depending on U, V and possibly a).
Example
A quadratic nonlinearity,
2
y = u -
V7ith a biased sinusoidal input as above, the output is given

by
y(t) = U 2 sin 2 (wt) + 2UVsin(wt) + V 2
= ~- {1-cos(2wt)} + 2UVsin(wt) + V 2
so we can immediately obtain the appropriate approximation

without computing any integrals, by simply dropping the
second-harmonic term. We thus have
y Q (t) = 2UVsin(wt) + -|(U2 + 2V 2 )

so that
hl = 2UV
aQ = U 2 + 2V 2
and so there is no phase-shift (as expected since the nonlin-

earity is single-valued) but both terms in the output depend
210
on the bias, as well as the sinusoidal amplitude, of the in-

put. In practice, the bias would probably be essentially
determined by the steady-state operating conditions and the
describing function for small oscillations would be
b
l
N(U) = ~- = 2V
which is just the slope of the nonlinear function at the op-

erating point.
11.4 Prediction of Limit Cycles
One of the principal applications of the describing func-

tion is in predicting the existence or non-existence of limit
cycles in nonlinear systems. The idea is to assume that, if
a limit cycle exists, it will correspond essentially to a
sinusoidal oscillation of a particular frequency, so that
other harmonic components can be ignored. This is exactly
in the spirit of the describing function approximation and
depends for its justification on the assumption that the
neglected high-frequency components are effectively supp-
ressed by the linear dynamical elements of the system. Con-
sequently, although the predictions regarding existence and
stability of limit cycles are usually fairly reliable, the
calculations of amplitude and period of oscillation by this
method can only be expected to have approximate validity.
Consider the behaviour of the following nonlinear feedback

system, with zero reference input,
f(u)
nonlinearity •• 9( S H
Fig. 11.1
In the describing function approximation, we assume that

211
u(t) is a pure sinusoid,
u(t) = U sin(a)t)
and represent the nonlinearity by its describing function

N(U,uO. The system is then described by the equation
U = -g(ja))N(U,u))U
and hence, if there is a limit cycle, i.e. a nonvanishing

periodic solution, it must correspond to this equation with
U ? 0, so that
-1
g(juO
N(U,a>)
Taking the real and imaginary parts of this equation, we get

two equations from which we can find U and w , the ampli-
tude and frequency of the limit cycle, in this approximation.
If the equations have no solution, then no limit cycle is
predicted.
Usually the describing function is independent of frequency

and then the limit cycle condition becomes
_,., x _ - 1
^VJUJ; " N T U T
In this form, it is convenient for graphical solution, by
plotting -1/N(U) on the Nyquist diagram of g(s), and
looking for intersections.
Fig. 11.2
212
If the graphs intersect, a limit cycle is predicted, with

amplitude and frequency corresponding to the intersection
point.
If there is no intersection, then some predictions about

asymptotic stability, analogous to those in the linear case,
may be made. Thus, if g(s) is asymptotically stable, and
its Nyquist plot does not encircle or intersect any part of
the locus of -1/N(U), we expect the feedback system to be
globally asymptotically stable. All these predictions are,
of course, subject to the validity of the describing function
approximation, which can be relied on only under suitable
conditions (in particular, the low-pass nature of the linear
element).
If a limit cycle is predicted, the way in which the graphs

intersect also gives us an indication of whether the limit
cycle will be stable or unstable. Thus, suppose the situa-
tion is that the graph of -1/N(U) in Figure 11.2 passes
into the region encircled by the Nyquist plot of g(s), as
U increases. Assuming that g(s) is stable, this indicates
that an increase in U will take the system into an unstable
region, thereby tending to increase U still further. Sim-
ilarly, a decrease in U will make the system stable and
hence cause U to diminish further. Thus, trajectories in
the neighbourhood of the limit cycle will tend to diverge
from it, and so the limit cycle is predicted to be unstable.
On the other hand, suppose the graph of -1/N(U) passes

out of the encircled region, as U increases. In this case,
again assuming g(s) to be stable, an increase in U will
stabilise the system, causing U to decrease, while a dec-
rease will destabilise it, leading to an increase in U .
Hence, perturbations of the system about its limit cycling
mode of operation will tend to die away as the motion returns
towards the limit cycle again. Thus, the limit cycle is pre-
dicted to be stable. This property is important since only
stable limit cycles can actually be observed in practice.
213
11.4.1 Frequency response of nonlinear systems
Another useful application of the describing function is

in estimating the frequency response of systems containing
nonlinear elements. We proceed as for linear systems, except
that each nonlinear element is replaced by its describing
function, in place of a transfer function. The resulting
frequency response thus becomes a nonlinear function of the
input amplitude and this can lead, in some cases, to the
prediction of phenomena, such as jump resonance , which can-
not occur in linear systems. Also, by extending the descri-
bing function method to take account of multiple frequencies,
we can predict other essentially nonlinear phenomena, exam-
ples being subharmonic generation and the quenching of
natural oscillations.
11.5 Nonlinear State-Space Equations
It is often convenient to express the equations describing

a nonlinear system in state-space form, by defining a state-
vector x which satisfies a first-order vector differential
equation
x = f(x,u,t)
where u is the input-vector of the system and f is a non-

linear vector function. The output-vector y of the system
is then given by
y = h(x,u,t)
where h is another nonlinear vector function. If the sys-

tem is time-invariant, t will not appear explicitly in the
equations and, if the system is also free (independent of u)
or if the input is held fixed, we can suppress the depend-
ence on u and write
x = f(x)
The solutions of this equation are called the state trajec-

214
tories of the system.
11.6 Lyapunov's Method
This is a technique for investigating the stability proper-

ties of nonlinear state-space systems, sometimes called the
second method of Lyapunov (the first method being simply lin-
earisation around an equilibrium point). The idea is to
study the behaviour of a scalar function V(x), called a
Lyapunov function, along the state trajectories of the system.
If the state-vector x satisfies
x = f(x)
where f is a vector function with components f., we have

immediately, along the trajectories,
V(x) = I §2-
3x
fX (X)
i i
which is well-defined provided that V has continuous par-

tial derivatives with respect to all the components of x .
We can then draw various conclusions from the properties of
the expressions for V(x) and V(x), as follows.
(i) If V is everywhere non-positive, i.e.
V(x) <_ 0 for all x
then V cannot increase along any trajectory of the system;

consequently if, also, V is radially unbounded, i.e.
V(x) -* °° as |x| -*- °°
where |x| is the length of the vector x, then no trajec-

tory can go off to infinity (or, equivalently, all traject-
ories remain bounded).
(ii) If, together with the above properties, we have

215
V(0) = 0,
V(x) > 0 for all x f 0
and V does not vanish identically along any trajectory of

the system (apart from the degenerate trajectory x = 0),
then V must decrease to zero along every trajectory and
hence all trajectories approach x = 0 as t -> «; the sys-
tem is then said to be globally asymptotically stable.
Example
Suppose the state-space equations are
X = X
l 2
X =
2 ~X1 " YX
2
where y is a non-negative constant, and take
V(x) = x^ + 2x2
X X
so that V(x) = 4(x^x x + 2 2^
It is clear that V <_ 0 everywhere and that V is rad-

ially unbounded and positive everywhere except at x = 0f
where V = 0. Consequently, all trajectories remain bounded
at t increases. Further, if y > 0, V can only vanish
on the line x 2 = 0, which is not a trajectory of the sys-
tem, and so, for positive y, the system is globally asymp-
totically stable.
11.6.1 Domains of attraction
If a system has an asymptotically stable equilibrium point,

then, even if the stability is not global, we should be able
to find a region such that all trajectories which start in-
side it will converge to the equilibrium point. Such a
region is called a domain of attraction and one of the prin-
216
cipal uses of Lyapunov functions lies in the determination

of these domains as follows.
Suppose the equilibrium point is at x = 0 and that we

can find a bounded region containing it, which consists of
all points satisfying
V(x)
where c is a positive constant, such that, inside this

region, V and V have all the properties required in (i)
and (ii) above (except for radial unboundedness, which is
now irrelevant). Then, along any trajectory starting in this
region, V must decrease, which prevents the trajectory from
leaving the region, and hence V -> 0, so that x -»• 0, as
t -> ». The region is therefore a domain of attraction. It
will in general depend on the Lyapunov function used and
different choices of V(x) will lead to different domains
being found; in this way, we may be able to construct a
larger domain of attraction by taking the union of those ob-
tained by using various Lyapunov functions.
11.6.2 Construction of Lyapunov functions
The problem of finding a suitable Lyapunov function for a

given system can be tackled by various methods, none of which
are entirely satisfactory. One approach is to take a func-
tion whose form is suggested by the appearance of the sys-
tem's equations (e.g. an energy integral), possibly contain-
ing some arbitrary parameters, and then adjust the parameters
so that V a/id V have the required properties.
A more systematic procedure, however, is to linearise the

equations around an equilibrium point and then construct a
Lyapunov function for the linearised system; this function
is thus guaranteed to have the correct properties, at least
in some neighbourhood of the equilibrium point. The linear-
ised equations can be written, in vector form,
Ax
217
where the matrix A has all its eigenvalues in the open

left half-plane, since the equilibrium is assumed to be asym-
ptotically stable. We then take the Lyapunov function as a
quadratic form
V(x) = x T Lx
where suffix T denotes transposition and L is a symmetric

matrix, so that
V(x) = xT(LA + A T L)x
and now the correct properties for V and V can be ensured

by choosing an arbitrary positive-definite matrix Q and
obtaining L by solving the Lyapunov matrix equation
LA + A T L = -Q
whose solution is the positive-definite matrix L given by
L = f exp(ATt)Q exp(At)dt
This integral expression, however, is not normally used for

calculating L; usually, it is more convenient to regard
the Lyapunov matrix equation as a set of linear equations
for the components of L, which can be solved by standard
methods of linear algebra.
When a Lyapunov function has been obtained for the linear-
ised system, we can then apply it to the full nonlinear pro-
lem. The expression for V will now be different but we
shall still have
m
V = -xQx
for small |x|, so that there will be a region containing

the equilibrium point in which V is strictly negative
(except at x = 0 ) . Consequently, the minimum value of V
over all points satisfying V = 0 (x ^ 0) will be strictly
218
v it:
positive; denoting this by minf follows that the re-
gion
V(x) < V
min
is a domain of attraction.
11.7 Absolute Stability Criteria
The term 'absolute stability1 (or hyperstability) is

applied to certain types of nonlinear system which are known
to be stable regardless of the exact form of the nonlinearity
provided that its gain always lies within a given range. For
example, consider the feedback system of Figure 11.1 .
Suppose the nonlinear function f(u) satisfies f(0) = 0

and
for all u f 0, where a and 8 are real constants. Then,

various conditions can be shown to be sufficient for global
asymptotic stability, using Lyapunov functions or some other
method; we shall not prove the results here but simply state
them, for which purpose it is convenient to introduce the
concept of 'positive-realness'.
A
A rational transfer function is said to be positive-real
if it has all the following properties: no poles in the
right half-plane; any poles on the imaginary axis are simple
and have real positive residues; the function has a non-
negative real part at every point, apart from poles, on the
imaginary axis. These properties are easily translated into
conditions on the Nyguist plot of the function, enabling the
stability criteria to be given a graphical interpretation.
Assuming that g(s) is strictly proper, we then have the
following results.
The Circle Criterion. If the rational function
l+gg(s)
l+ag(s)
219
is positive-real, then the nonlinear feedback system is

globally asymptotically stable . The graphical interpreta-
tion of this result depends on the signs of a and $ .
For simplicity, suppose that & > a > 0 and that g(s)
is the transfer function of an asymptotically stable system.
Then, the circle criterion requires that the Nyquist plot of
g(jw) should not intersect or encircle the disc on
[-a" ,-3 1 as diameter.
It is clearly a generalisation of the Nyquist criterion

for linear systems and is of theoretical importance but is
usually rather conservative in practice.
The Popov Criterion. If a = 0 and there exists a positive

constant q such that the rational function
(l+qs)g(s) + e""1
is positive-real, then the system is globally asymptotically

stable . Also, provided that the nonlinear function f(u)
is single-valued, this result can be extended to the case
where q is negative. Like the circle criterion, it is
only a sufficient (and not necessary) condition but is less
conservative because of the freedom to choose the value of
q. The graphical interpretation makes use of the so-called
'modified polar plot 1 of u)Img(jto) against Reg(joj); this
plot is simply required to lie entirely to the right of a
straight line (of slope q" ) passing through the point
Results of this kind, employing the positive-realness

concept, play an important part in nonlinear system theory
and also have practical applications. A linear system with
a positive-real transfer function is said to be 'passive1
and this concept is of importance in the theory of model-
A
reference adaptive systems , where the control system para-

meters are made to vary in a way which depends on the error
between the outputs of the system and the model, so as to
make the system's behaviour approach that of the model,
Figure 11.3.
220
model
input
— controller plant
adaptive law
Fig. 11.3
Such systems are essentially nonlinear and it is necessary

to design the adaptation procedure in such a way that global
stability is ensured; in some cases, this can be done by
utilising the properties of certain types of Lyapunov func-
tion which are known to exist for passive systems.
Another area of application for absolute stability criteria

is in the stability of interconnected systems. The idea is
that, if each of the subsystems is 'sufficiently stable1 and
the interconnections are 'weak enough 1 , then the total system
will be stable. In order to obtain quantitative conditions
it is therefore necessary to have a measure of how stable
the subsystems are and how strong are the connections between
them. Such measures can be derived, for example, in terms
of the properties of Lyapunov functions defined for the var-
ious subsystems; alternatively, an 'effective gain' may be
defined for each subsystem, lower gain corresponding essen-
tially to greater stability. Criteria can then be obtained
which give sufficient conditions for global asymptotic stab-
ility of the whole system, the essential requirement being
that the interactions should be sufficiently weak relative
to the stability of the subsystems .
References
1. ATHERTON, D.P. : 'Nonlinear Control Engineering',

(Van Nostrand Reinhold, 1975).
221
2. GELB, A. and W.E. VANDERVELDE : 'Multiple-input Des-

cribing Functions and Nonlinear System Design 1 ,
McGraw-Hill, 1968).
3. WILLEMS, J.L. : 'Stability Theory of Dynamical Systems'

(Nelson, 1970).
4. LANDAU, Y.D. : 'Adaptive Control - The Model Reference
Approach 1 , (Marcel Dekker, 1979).
5. COOK, P.A. : 'On the stability of interconnected sys-

tems', Int. J. Control, 20, (1974), pp. 407-415.
222
NONLINEAR SYSTEMS - Problems
P. 11.1 Consider the feedback system of Fig. 11.1, with

the nonlinearity being an ideal relay,
f (u) =
and g.(s) =
s(s+l)(s+4)
Show that a stable limit cycle is predicted, and

estimate its amplitude and period of oscillation,
using the describing function method.
P. 11.2 For the system described by the state-space

equations
x = X
l 2
find the largest domain of attraction obtainable by

using the Lyapunov function
Chapter 12
Some DDC system design procedures

Dr. J. B. KNOWLES
Synopsis^
As a result of the enormous impact of microprocessors,

electronic engineers, with sometimes only a cursory back-
ground in control theory, are being involved in direct-
digital-control (D.D.C.) system design. There appears to
be a real need for an easily understood and simply imple-
mented comprehensive design technique for single-input
d.d.c. systems. The proposed design technique provides,
first of all, a simple calculation that ensures that the
data sampling rate is consistent with the control system's
accuracy specification or the fatigue life of its actuators.
Pulsed transfer-function design for a plant controller is
based on two simple rules and a few standard frequency-
response curves, which are easily computed once and for all
time. Structural resonances are eliminated by digital notch
filters, the pole-zero locations of which are directly re-
lated to the frequency and bandwidth of an oscillatory mode;
this is exactly as with analogue networks. In addition a
computationally simple formula gives an upper bound on the
amplitude of the control error (deviation) component due to
multiplicative rounding effects in the digital computer;
this thereby enables the selection of a suitable machine
wordlength or machine. A distinct advantage of the proposed
design technique is that its implementation does not necess-
arily involve a complex computer-aided-design facility.
12.1 Introduction
As a result of the enormous impact of microprocessors,
electronic engineers with sometimes only a cursory back-
ground in control theory are being involved in direct dig-
ital control (ddc) system design. There is now a real need
for an easily understood and simply implemented design tech-
nique for single-input DDC systems which is the objective
of this tutorial paper. The proposed design procedure is
shown as a flow chart in Figure 12.1, and it contains appro-
priate references to sections of text that treat the more
important issues in detail. It is hoped that this diagram
will provide a systematic approach to DDC system design, as
well as an appreciation for the organisation of the text.
© United Kingdom Atomic Energy Authority, 193"2

224
entry
select sampling frequency

from equation 2 - section 2
design filters;
references 13, 14 I calculate pulsed frequency
note common- I response of plant, a d . and d a .
mode levels I equation 7 — section 3
sep?rate sampled noise

process side bands by a high
enough sampling frequency;
section 5
use gain-phase curves i n Figs. 8—19

1 to get adequate gain-phase
yes
margins, equation 10 section 3
phase-retard r.c.
network for class 0
system; section 4, fig.4
Fig. 12.1 d.d.c. design scheme

225
The experienced designer will notice the absence of such

topics as:
(i) 'Bumpless transition1 criteria during changes from

automatic to manual control,
(ii) Provision for timed sequential operations to cope with
fault conditions.
These aspects have been deliberately omitted because they

are considered to be separate programming issues once control
in the normal operating regime has been achieved.
Digital realisations of conventional three-term controllers

have the advantages of: wide applicability, theoretical
simplicity and ease of on-line tuning. However, the result-
ing closed-loop performance is generally inferior to that
obtainable with other algorithms of similar numerical com-
plexity. The graphical compensation procedure ' described
in Section 12.3 copes with the design of digital three-term
controllers and these high performance units with equal ease.
Also the technique is readily exploited by virtue of its
simple calculations (amenable even to slide-rule treatment).
Furthermore, it is shown to result in an 'even-tempered1
closed-loop response for all input signals; unlike DDC
systems synthesised by time-domain procedures.
Compared to analogue controllers, digital controllers offer
distinct advantages in terms of: data transmission, inter-
connection, auxiliary data processing capabilities, fault
tolerance, tamper resistance, etc.. However, a digital con-
troller must evidently provide a control performance at
least as good as that of the analogue controller it replaces.
In this respect, it is suggested that sampling and wordlength
effects are designed to be normally negligible relative to
the control accuracy specification. When this objective is
frustrated by computer performance constraints these degen-
erate effects can be evaluated from formulae given in the
text.
12.2 Choice of Sampling Frequency
The first step in design of a ddc system is the selection

of an appropriate sampling rate (T). Distortion in the
form of spectral side-bands centred on integral multiples of
the sampling frequency (1/T) is inherently produced by the
periodic sampling of information, and the perfect (imprac-
tical) recovery of the signal requires the complete elimina-
tion of these harmonics . A suitable practical choice of
sampling frequency limits the distortion (or aliasing) by
imposing a large enough frequency separation between the
side-bands and the original unsampled signal spectrum for
the low-pass plant elements to effect an adequate attenua-
226
tion. Where comprehensive plant records for an existing

analogue control scheme are available, the sampling period
for a replacement ddc system is sometimes decided on the
basis of requiring a 'small change1 in the time dependent
error or deviation signal during this interval. In the
author's opinion, engineering judgements of this form are
valuable only as confirmation of a fundamentally based and
experimentally confirmed analysis.
For the sinusoidal input:
x(t) = A sin(wt) (12.1)
4
Knowles and Edwards show that the average power of the con
trol error component due to imperfect side-band attenuation
is bounded for a unity-feedback, single-input ddc system by:
If < A2gV(2R+2) M2 w 2 | G ( j w ) | - 2 (12.2)

R "™" S p
where G(s) denotes the transfer function of the plant ex-

cluding the zero-order hold and:
G(s) = M = Closed-loop peak

sR
magnification
wg = 2TT/T (12.3)
Equation (12.2) reveals that the high frequency components

of the continuous input spectrum generate relatively the
greatest distortion and loss of control accuracy. Thus, a
simple design criterion for selecting the sampling frequency
is to ensure that the right-hand-side of equation (12.2)
represents an admissible loss of control accuracy for the
largest permissible amplitude sinusoid at the highest likely
signal frequency. This calculation is not only easily per-
formed, but it is independent of subsequent stabilisation
calculations. However, if computer loading demands the low-
est possible consistent sampling frequency, it is then nec-
essary to follow an interactive procedure which involves the
227
digitally compensated system and the formula 4 :
4 = TT g w,/* |H (jw)/P (jw) T sin Z (wT/2U (w)dw
(12.4)
*
where H (jw) defines the overall pulsed frequency response
of the stabilised closed-loop system, 0 (w) is the input
x
*
signal power spectrum, and P (jw) denotes the pulsed trans-
fer function of the plant and zero-order hold.
12.3 Frequency Domain Compensation Method
For most practical applications, necessary and sufficient

conditions for a ddc system to be stable are that the open-
loop pulsed frequency response:
* *
X K(z)P(z) 4 XK (jf)P (jf) (12.5)
z=exp( j2irfT)
does not encircle the (-1, 0) point. Furthermore, in many

cases the polar diagram of xK (jf)P (jf) is such that an
adequately damped time domain response is produced when its
Gain and Phase Margins are greater than about 12dB and 5 0°
respectively. It is evidently important to confirm that
these stability margins are maintained over the full-range of
plant parameter variations or uncertainties. Denoting the
Laplace transfer function of the plant, A/D, D/A and data
smoothing elements by P(s) then the corresponding pulsed
frequency response is obtained from references 3 and 5 as:
P*(jf) = i I P(j27Tf-j27rnf s ) (12.6)
Data sampling frequencies selected according to the criter-

ion in Section 12.2 allow the above expression to be closely
approximated by:
p s
P*(jf) = k
1
( ) (12.7)
S=j27Tf
over the range of frequencies that determine closed-loop sys-

228
tern stability. If the data smoothing element consists of a

zero-order hold, and the combined gain of the A/D and D/A
converters is xT/ the required pulsed frequency response
P (jf) is readily calculable from:
P (jf) = x T exp(-J7rfT) . s i n .G(s) (12.8)

TTfT
s=j2irf
For ease of nomenclature in the following discussion, the

number of pure integrations in the plant transfer function
G(s) is termed the 'class1 of the system. In general, it is
recommended that one stage of phase-lead compensation is in-
cluded for each pure integration in the plant. Thus, for a
class 1 system, the discrete compensator takes the form:
z-a
X K(z) = x (r—) with 0 < a 9 < a, < 1 (12.9)
where the pole at z = a^ is necessary for physical realis-

ability. Observe that digital realisations of P+I and P+D
controllers can be represented in the form of equation (12.9).
Graphs like Figures 12.2 and 12.3 (references 2,17), which
show the gain and phase of [exp (j 2 TrfT) -c] to base of norm-
alised frequency (fT) for a limited number of real and
complex pairs of values for c are the basis of the proposed
design technique. The value of c corresponds to a con-
troller zero or pole. As with analogue networks, increasing
amounts of phase-lead are seen to be associated with increas-
ingly severe rising gain-frequency characteristics. If the
maximum phase-lead per stage is much above the recommended
value of 55 , the rising gain-frequency response of the dis-
crete controller causes the locus xK (jf)P (jf) to bulge
close to the (-1, 0) point which aggravates transient oscill-
ations.
Pursuing a maximum bandwidth control system design, the

*
measured or calculated values of P (jf) are first used to
determine^ the frequency (f_) for which:
/P*(jf R ) + 55° x Number of Phase-lead Stages = -130°
^ (12.10)
fA Bode or Nyquist diagram is helpful in this respect.
229
1.5
0.001 0.005 0.01 0.05 0.1 0.5 1.0

normalised frequency fT
Fig. 12.2 Gain/normalised-frequency characteristics,

|exp(j27rfT)-a|
0.001 0.005 0.01 0.05 0.1 0.5 1.0

normalised frequency fT
Fig. 12.3 Phase/normalised-frequency characteristics,

/{exp{j27rfT)-a}
230
This procedure with an appropriate choice of the scalar gain

constant (x) is clearly orientated towards establishing an
adequate Phase Margin. By means of Figure 12.3, which shows
/exp(j2frfT) - a to base of fT, real controller zero(s)
are selected to give on average no more than 7 0° phase-lead
per stage at fgT. In the case of a class 1 system, the
controller pole is chosen from the same graph to produce
around 15-20° phase-lag at fg T - For a
class 2 system, com-
plex controller poles have a definite advantage over their
real counterparts, because references 2 and 17 show that rel-
atively less phase-lag is generated at servo frequencies.
Consequently, complex pole pairs may be placed relatively
closer to the (1, 0) point thereby enhancing the zero-
frequency gain of the discrete controller, and the load dis-
turbance rejection capability of the control system. Finally,
by means of a Bode or Nyquist diagram, the scalar gain factor
(x) is set so as to achieve a closed-loop peak magnification
of less than 1.3 and a Gain-Margin of about 12dB. Although
this maximum bandwidth design procedure may not be ideal for
all applications it is nevertheless adaptable. For example,
the pulsed transfer function of a conventional P+I controller
may be written as:
; a =
KD(z) = <V*l><i=T> K ^ <12-U>
so that the gain and phase curves may be applied to select

suitable values for 'a1 and Kp+K-j., which uniquely specify
the controller. Again, the design curves may be used to
match the frequency response function of a digital controller
to that of an existing analogue controller for the range of
frequencies affecting plant stabilisation.
12.4 The Compensation of Class 0 (Regulator) Systems
The performance and stability of class 0 systems are gener-

ally improved by the inclusion of phase-retard compensators,
whose analogue transfer function has the general form:
K(s) = x (\tlaTT) (12.12)

231
where the attenuation constant a is less than unity but

greater than 1/12 for practical applications. A/D converters
are frequently time-division multiplexed between several con-
trollers, but each loop normally has exclusive use of its
own D/A converter which can serve useful dual roles. A
series resistor and capacitor connected across the feedback
resistor of a D/A realises equation (12.12) and it is pro-
posed for the phase-retard compensation of class 0 ddc sys-
tems. Apart from the benefits accruing from savings in com-
puting time, the use of analogue networks in this instance
is superior to digital because experience shows that an
appreciable phase lag then occurs over a markedly narrower
range of frequencies for the same attenuation constant. The
importance of this fact will become apparent after the foll-
owing description of phase-retard network design.
The first step in the compensation process is to determine

the frequency a)g rads/sec for which the phase of the frequ-
ency response function P (jw) is -130°. If the Gain and
Phase Margins of the uncompensated regulator are satisfac-
tory, but the problem is to reduce the zero-frequency con-
£
trol error to DC' then set
10/cex = u>B (12.13)

and
DC *
a = (x _ U^ )P (0)
DC
(12.14a)
X = 1A
With these parameters the gain and phase of K(s) for

frequencies above wB are effectively unity and zero res-
pectively. Hence, inclusion of the network increases the
zero-frequency gain of the open-loop response without alter-
ing the original satisfactory stability margins. Alternat-
ively, if the zero-frequency control accuracy is acceptable,
but improvements in stability margins are desired, then
select parameters according to equation (12.13) and
X = 1
* (12.14b)
a = 1/|P (jw R )| for a 50° Phase Margin
232
With these values, the gain and phase of K(s) for frequen-
cies above u>_ are a and zero respectively. Hence,
hi
inclusion of the network improves the Gain Margin by the
factor 20 log l Q (I/a), without altering the original sat-
isfactory open-loop zero-frequency gain. Both the above
design techniques for phase-retard networks are based on
sacrificing phase in a relatively unimportant portion of
the Nyquist diagram for gain in a separate region that mark-
edly influences control accuracy or stability. As the
frequency range over which the network produces an appreci-
able phase-lag widens, this simple design procedure becomes
evidently complicated by the progressive overlap of these
regions of gain and phase changes.
12.5 Noisy Input or Output Signals
The operational amplifier associated with a D/A converter

can also be used for additional signal conditioning. Some-
times the input or output signal of a control system is
heavily contaminated by relatively wide-bandwidth noise. If
it is impractical to low-pass filter these signals to prevent
13 14
aliasing ' , then an irrevocable loss of control perform-
ance can only be prevented by employing a sampling frequency
which largely separates the side-bands of the sampled noise
process. Unnecessary actuator operations, and therefore
wear and tear, are sometimes avoided in this situation by
including an averaging calculation in the digital compen-
sator. In these circumstances, the side-band attenuation
achieved by the plant is rendered less effective if the out-
put sampling rate of the controller is reduced to an integral
sub-multiple of its input rate. As an alternative solution
to this noisy signal problem, it is proposed that a suitable
capacitor is placed in parallel with the feedback resistor
of the D/A converter.
12.6 Structural Resonances and Digital Notch Networks
A servo-motor is coupled to its load by inherently resil-

ient steel drive shafts, and the combination possesses very
233
selective resonances because the normal dissipative forces

are engineered to be relatively small. In high power equip-
ment (radio telescopes, gun mountings, etc.) the frequency
of such structural resonances can be less than three times
the required servo-bandwidth, and consequently their pres-
ence markedly complicates the compensation procedure. Viewed
on a Nyquist diagram, the resonances cause the locus to loop-
out again from the origin to encircle the (-1, 0) point.
Analogue controllers obtain system stability under these con-
ditions by processing the nodal error signal with tuned
notch filters (e.g. Bridged-Tee) which block the excitation
of each oscillatory mode. For a resonance peak at w rad/
sec, the same effect is provided by the following digital
filter:
(z-zQ) (z-z0)
N(Z) = K (12.15)
(Z r
-r 0 V " 0V
where
exp( jo)QT)
(12.16)
= exp(-jo)QT) 0 < r < 1
*0
and unity gain at zero-frequency is achieved by setting:

2
"(TO (12.17)
K.N 1-z 0
In order to null a mechanical resonance, its bandwidth and

centre-frequency must both be matched by the digital filter.
As may be appreciated from the pole-zero pattern of the fil-
ter, no larger notch bandwidth than necessary should be used
in order to minimise the phase-lag incurred at servo-frequ-
encies. It remains therefore to relate the notch bandwidth
(B ) to the parameter rQ.
The gain-frequency response function of the proposed dig-

ital filter is given by:
(exp(j w T)-z 0 ) (exp(ja)T)-z0)
|N = K, (12.18)
(exp(ja)T)-rozQ) (exp(ja)T)-rQz0)
234
Defining the frequency deviation variable:
6 = oj - IDQ (12.19)
then a power series expansion of exponential functions of 6

yields the first order approximation:
|N*(j6)| = — for 6T << 1 (12.20)
Where:
1 - exP(-j2o>nT)
U (12.21)
- r Q exp(-j2a)0T)
Equation (12.20) evidently has complete similarity with the

gain-frequency characteristic of an analogue notch filter in
the vicinity of its anti-resonance. Accordingly, its band-
w
width is defined as the frequency increment about 0 within
which the attenuation is greater than 3dB:
B = ± { U r
N0 " 0)/T} (12.22)
Calculated or measured plant frequency responses may

therefore be used in conjunction with equations (12.16) and
(12.22) to specify suitable values for WQ and rQ in
equation (12.15), and the resulting filter is incorporated
as an algebraic factor in the pulse function of the discrete
controller.
12,7 Coefficient Quantisation in a Discrete Controller
The coefficients in a pulse transfer function usually

require rounding in order to be accommodated in the finite
wordlength format of a microprocessor or minicomputer.
Because generous stability margins are normally employed,
coefficient rounding is normally unimportant as regards the
practical realisation of phase-lead or phase-retard pulse
transfer functions. However, it is well-known that increas-
ingly severe digital filter specifications (cut-off rate,
bandwidth, etc.) accentuate computer wordlength require-
8 9
ments ' . Hence it is generally prudent to examine the word-
235
length necessary to counter acute structural resonances by

the proposed form of digital notch filter.
By writing equation (12.15) as:
z
N(z) = K ~aZ+1 7 (12.23)
N
z - rbz+r
coefficient rounding on {a, rb, r } is seen to modify both
the centre-frequency and the bandwidth of a notch filter
design. Ideally with no rounding the coefficients in equa-
tion (12.23) are specified by:
a = aQ; (rb) = (r o b Q ); a Q = b Q = 2 cos(a)QT);

2 . (12.24)
(r2) = (r2)
and defining for the practical situation the variables
a = 2 cos(u)1T); b = 2 cos(u>2T) (12.25)

(12.26)
one obtains the gain-frequency response function of the real-
isations as:
exp(j6)-l * " e X ? [ -3
exp(j6+je)-r - r exp[-j
(12.27)
Thus the centre-frequency of the filter is not at o^, and

the shift due to the finite computer wordlength is evaluated
directly from equation (12.25) as:
6a)o = i[cos" 1 (a/2) - cos""1 (aQ/2)] rad/sec (12.28)
It is interesting to note that the change in centre-frequency

cannot be obtained by differentiating equation (12.25) be-
cause a second-order approximation is required if the coeff-
icient aQ approaches - 2 . In the vicinity of the anti-
resonance and with fine quantisation:
236
6T << 1 and eT << 1 (12.29)
and under these conditions a power series expansion of the

exponential terms in equation (12.27) yields:
N (12.30)
|N
((l-r)/6T)
where
K = K (12.31)
N N 1 - r exp (-J2(JO
Thus the bandwidth (BN) of the realisation satisfies:
= 2 (12.32)
B
NT
(12.33)
N
For small enough perturbations about the ideal situation

defined in equation (12.24), it follows that:
3B
N
6B
< + Sr (12.34)
NO 3r
0 0
From equations (12.24), (12.26) and (12.33) one obtains:
6e < 2 fid) <5r £ q/ 2 r 0

0 '
(12.35)
B
N
= 1 ; = 1/T
0 8r 0
where the width of quantisation is defined by:
_ 2 "(wordlength - 1) (12.36)
Hence, the change in the notch bandwidth of the filter due

to coefficient rounding is bounded by:
(12.37)
NO
As a design example, consider the realisation of a notch

237
filter with: centre frequency 12 rad/sec, bandwidth +1 rad/

sec and a sampling period of 0.1 sec. Equation (12.24) de-
fines the ideal numerator coefficient as:
aQ = 0.724716 (12.38)
VJith an 8-bit machine format, the actual numerator coeffic-

ient realised is:
a = 0.718750 (0.1011100) (12.39)
so that from equation (12.28):
6u)0 = 0.032 rad/sec (12.40)
and the change in filter bandwidth evaluates from equation

(12.37) as:
6BN0 £ 0.11 rad/sec (12.41)
Thus, in practical terms, an 8-bit microprocessor is about

sufficient for realising this particular specification.
12.8 Arithmetic Roundoff-Noise in a Discrete Controller
Arithmetic multiplications implemented by a digital com-

puter are subject to rounding errors due to its finite word-
length. As a result, random noise is generated within the
closed loop and the control accuracy is degenerated. In
extreme cases, the control system can even become grossly
unstable. It is therefore important to relate the loss of
control accuracy to the wordlength of the discrete controller.
The analysis in Reference 10 provides an easily calculable
upper-bound for the amplitude of the noise process present
on the error signal due to arithmetic rounding in a discrete
controller. With a directly programmed compensator for ex-
ample, this upper-bound is given by:
I I < ( Number of Significant Multiplications in xK (z)j

|e :
Q' - \ 2x| I Numerator Coef f icients | /q
V
(12.42)
238
where a 'significant1 multiplication does not involve zero

or a positive integral power of 2 including 2 . During a
normally iterative design procedure, equation (12.42) is
valuable as a means of comparing the multiplicative rounding
error generated by various possible controllers. In prelim-
inary calculations, the pessimism of this upper-bound is of
little consequence. However a more accurate estimate of
performance degeneration due to multiplicative rounding
errors can be economically important for the finalised design
when it is realised on a bit-conscious microprocessor. For
this purpose, the analysis in References 11 and 12 is recom-
mended because one straightforward calculation quantifies
the loss of control accuracy as a function of controller
wordlength.
As described earlier, the sampling frequency of a DDC sys-

tem must be high enough to prevent an unacceptable loss of
control accuracy or fatigue damage to plant actuators. How-
ever, strangely enough, it is possible in practical terms
for the sampling frequency to be made too high. Reference
to Figure 12.3 shows that by increasing the sampling frequ-
ency, the phase-lead required at a particular frequency can
only be sustained by moving the compensator zero(s) closer
to unity. To preserve load disturbance rejection, the zero-
frequency gain of the controller (xK(z)| Q) must be main-
tained by then moving the pole(s) also closer to unity. As
a result, the algebraic sum of the compensator's numerator
coefficients is reduced, and multiplicative round-off noise
on the error signal is increased according to equation
(12.42). Thus the choice of an unnecessarily high sampling
frequency can entail unjustified expense in meeting long
wordlength requirements.
12.9 Multirate and Subrate Controllers
A subrate digital controller has its output sampler opera-

ting an integral number of times slower than its input. In
*The formula over-estimated by about 2 bits in the example

considered in Reference 10.
239
a multi-rate controller, the output sampler operates an in-

tegral number of times faster than its input. Section 12.5
describes how subrate systems prejudice ripple spectrum
attenuation, and in practice excessive actuator wear seems
best prevented for noisy input signals by the use of a
single rate controller and a smoothing capacitor across the
D/A unit. It is now pertinent to question if multi-rate
systems afford any advantages over their more easily designed
single-rate counter-parts.
As a result of the higher pulse rate exciting the plant,
Kranc contends that the ripple performance of a multi-rate
system is superior to that of a single rate system, even
though both cases have the same input sampling rate. However,
it should be noted that the input sampler of the multi-rate
sampler still generates spectral side-bands centred on mul-
tiples of ± w Q f an(* t n a t t n e s e a r e n o t eliminated in further
modulation by the faster output sampler. Hence, it is by no
means self-evident that increasing the output sampling rate
alone in a digital compensator effects an improvement in
ripple attenuation. To clarify this issue and other aspects
of performance, a comparison of single and multi-rate unity
feedback systems is implemented in Reference 16 for the plant
and zero-order hold transfer function:
- s T M -sT
sT
P(s) = -—Y^£
— YJ ^ £ L
(12.43)
s (s+l)
with computer sampling periods of:
T. = T t = 0.1 sec Single-rate system )

T. = 3T . = 0.1 sec Multi-rate system
Both closed-loop systems are compensated to have virtually

identical frequency and transient response characteristics.
For two typical input spectra, calculations show that the
output ripple power for the multi-rate system is at least a
decade greater than that for the single rate system. An
intuitive explanation of this result may be based on the
240
fact that in both systems the compensator and zero-order

hold are attempting to predict some function of the error
over the input sampling period. As each prediction inevit-
ably incurs error, a multi-rate system makes more errors per
unit time and thus gives the larger ripple signal. It may
be argued that multi-rate controllers give potentially the
faster transient response. However, the conclusion reached
in Reference 16 is that improvements in transient response
leads to a further degradation in ripple performance.
12.10 Comparison of Time Domain Synthesis and

Frequency Domain Compensation Techniques
A convenient starting point is Figure 12.4, which shows

the step and ramp responses of a time domain synthesised
dead-beat system whose continuous part consists of a D/A
converter and the plant:
P(s) = — — ~ (12.45)
with a sampling period of 1 second. Though the ramp response

of this example (which is taken from Reference 3) is good,
the 100% overshoot in the step response is unacceptable for
most purposes. It is emphasised that this 'highly-tuned'
behaviour is typical of ddc systems which have been synthe-
sised in the time domain by the dead-beat or minimum variance
techniques ' . The example is therefore considered unbiased,
and a useful vehicle for comparing time domain methods with
the proposed frequency domain compensation procedure. For the
Class 1 system in equation (12.45) the algorithm in Section
12.3 specifies a series compensating element of the form:
XK(Z) = X-i~^r (12.46)
whose maximum phase-lead must not exceed about 55 . In order

to achieve an adequate phase-margin (- 50°), the frequency
(fB) at which the phase lag of P*(jw) is 180° indicates
the maximum bandwidth likely to be achieved. It is readily
seen from equation (12.46) that in this case:
241
1 2 3 A 5 6 7
time s
1 2 3 4 5 6 7
time s
3.0r
2.0
1.0
0.1 0.3 0.5 1.0 3.0

frequency,rod/s
Figure 12.4 Response Characteristics of Dead-Beat System

242
f T = 0.16 (12.47)
B
However, Figure 12.3 shows that the phase difference between

numerator and denominator of a discrete compensator is then
well below the maximum which occurs around:
fT = 0.03 (12.48)
More effective compensation can therefore be realised by de-

creasing the sampling period to 0.25 seconds . Calculations
of the pulsed frequency response P (ju)) via equation (12.8)
differ from those computed using the z-transformation in gain
and phase by less than 0.01%; so vindicating the approxima-
tion. As fgT now equals 0.035, a digital compensator which
provides 55° lead at this frequency is selected by means of
Figure 12.3 as:
*K(Z) = * T K T I T (12 49
->
which with a gain constant (x) of 0.4 realises a gain-
margin of approximately lOdB. Figure 12.5 shows that the
step and ramp function responses are now both adequately
damped, and the additional control error due to multiplica-
tive rounding errors is derived from equation (12.42) as:
|e Q | 1 37.5q (12.50)
which indicates that double-precision arithmetic would be

necessary on an 8-bit JJ-processor. Apart from an 'even-
tempered1 response and wordlength requirements, the proposed
design procedure also provides: a simpler digital compensa-
tor, the stability margins available to cope with plant vari-
ations or uncertainties, and an assessment of the rms ripple
error. For the purpose of designing a practical ddc system
it provides necessary information so evidently lacking in the
*Note the rms ripple error at 1 rad/s even with a sampling

period of Is is satisfactory; equation (12.2) evaluates
as 1.3*10""3.
243
<b
Fig. 12.5 Response Characteristics of Frequency Domain

Compensated System.
244
time domain techniques. Furthermore, the design rules are

demonstrably simple and the calculations involve about half
a dozen standard gain and phase curves ' / which are easily
computed once and for all time.
REFERENCES
1. IEE Coloquium: Design of Discrete Controllers,
London, December 1977.
2. KNOWLES ,J.B. : A Contribution to Direct Digital

Control, PhD Thesis, University of Manchester, 1962.
3. RAGAZZINI, J.R. and G. FRANKLIN : Sampled Data Control

Systems. McGraw-Hill Publishers Inc., 1958.
4. KNOWLES, J.B. and R. EDWARDS : Ripple Performance and

Choice of Sampling Frequency for a Direct Digital Control
System. Proc. IEE, 113,(11), (1966), p. 1885.
5. JURY, E.I. : Sampled-data Control Systems.

John Wiley Publishers Inc., 1958.
6. MATTERA, L. : Data Converters Latch onto Microprocess-

ors. Electronics, 81, September 1977.
7. KNOWLES, J.B. and R. EDWARDS : Aspects of Subrate

Digital Control Systems. Proc. IEE, 113, (11), 1966,
p. 1893.
8. AGARWAL, R.C. and C.S. BURRUS : New Recursive Digital

Filter Structures have a Very Low Sensitivity and Round
off Error Noise. IEE Trans, on Circuits and Systems,
22, No. 12, 1975, p. 921.
9. WEINSTEIN, C.J. : Quantization Effects in Digital

Filters. Lincoln Laboratory MIT Rept. 468, Nov. 1969.
10. KNOWLES, J.B. and R. EDWARDS : Computational Error

Effects in a Direct Digital Control System.
Automatica, 4, 1966, 7.
11. KNOWLES, J.B. and R. EDWARDS : Effect of a Finite

Wordlength Computer in a Sampled-Data Feedback System.
Proc. IEE, 112, 1965, p. 1197.
12. KNOWLES, J.B. and R. EDWARDS : Finite Wordlength

Effects in Multi-rate Direct Digital Control Systems.
Proc. IEE, 112, 1965, p. 2377.
13. WONG, Y.T. and W.E. OTT, (of Burr-Brown Inc.)

Function Circuits - Design and Applications.
Published by McGraw Hill, 1976.
245
14. STEWART, R.M. : Statistical Design and Evaluation of

Filters for the Restoration of Sampled Data.
Proc. IREf 44, 1956, 253.
15. KRANC, G.M. : Compensation of an Error-Sampled

System by a Multi-rate Controller. Trans AIEE, 76,
1957, p. 149.
16. KNOWLES, J.B. and R. EDWARDS : Critical Comparison

of Multi-rate and Single-rate DDC System Performance.
JACC Computational Methods Session II.
17. KNOWLES, J.B. : Comprehensive, yet Computationally

Simple Direct Digital Control-system Design Technique.
Proc. IEE, 125, December 1978, p. 1383.
Chapter 13
Robust controller design

Professor E. J. DAVISON
Synopsis
An overview of some recent results in the controller de-

sign of multivariable linear time-invariant systems, in
which uncertainty/perturbation in plant behaviour occurs is
made. In particular, the servomechanism problem is consid-
ered for various cases varying from the extreme case when
the plant is known and fixed, to the opposite extreme when
the plant is completely unknown and is allowed to be pertur-
bed. In addition, the chapter includes the case of time lag
systems, decentralized systems and the control of systems
subject to sensor/actuator failure. The emphasis throughout
is to characterize structural results for the servomechanism
problem, e.g. to determine the existence of a solution to
the problem and to determine the controller structure requ-
ired to solve the problem under different conditions. The
chapter concludes with a proposed design method for obtain-
ing realistic robust controller design.
13.1 Introduction
This chapter shall attempt to look at some aspects of con-
troller synthesis for the multivariable servomechanism pro-
blem when the plant to be controlled is subject to uncertain-
ty - in this case, a controller is to be found so that sat-
isfactory regulation/tracking occurs in spite of the fact
that the parameters of the plant may be allowed to vary (by
arbitrary large amounts), subject only to the condition that
the resultant controlled perturbed system remains stable.
This type of controller design will be called robust con-
troller design in this chapter. It is to be noted that
other definitions of "robust controller design" may be found
in the literature.
The topic of robust control is an extremely active area of
research at present, and so it is impossible to include all
results/approaches obtained to date. Some recent literature
dealing with various aspects of robust control design by us-
ing a number of different approaches may be found, e.g. in
Sain 1 7 ; Spang, Gerhart 19 . No attempt shall be made in this
short chapter to survey all such approaches. Instead, the
chapter shall concentrate only on some of the structural
247
results recently obtained re the servomechanism problem (in

the time domain) and discuss one possible approach for obtain-
ing effective "robust controller design".
The chapter is divided as follows: part 2 describes the
general problem, part 3 describes some structural results ob-
tained for the servomechanism problem under different con-
ditions, part 4 describes some extensions of the previous
results and part 5 describes a proposed design method for ob-
taining realistic robust controller design.
13.2 General Problem to be Considered

A plant has inputs u which can be manipulated, outputs y
which are to be regulated, outputs y which can be measured,
disturbances GO which, in general, cannot be measured, and
reference input signals y ^ which it is desired that the
outputs "track". The general servomechanism problem consists
of finding a controller (either centralized or decentralized)
which has inputs y , Yref and outputs u for the plant so
that:
1. The resultant closed loop system is stable } Stability

2. Asymptotic tracking occurs, independent of
>_* \ Regulation
any disturbances appearing in the system
3. Fast response occurs Transient
4. "Low interaction" occurs behaviour
5. (l)-(4) occur independent of any variations | Robust
in the plant parameters of the system (in- > behaviour
eluding changes of plant model order) *

6. "Non-explosive" behaviour occurs in the
closed-loop system under sensor/actuator } Integrity
failure
7. Under sensor/actuator failure, the
remaining parts of the controlled system [ Reliability
still satisfactorily track/regulate
In addition, depending on the particular physical problem
many other constraints may be imposed on the problem state-
ment, e.g. damping factor constraint, controller gain con-
straint, etc.
248
13.3 The Servomechanism Problem - Structural Results
13.3.1 Problem description
The plant to be controlled is assumed to be described by

the following linear time-invariant model:
X = Ax + Bu + Ew
y = Cx + Du + Fa) (13.1)
+
m = Hh D
mu V , e = y
" y ref
where xeR n is the state, ueR m is the output, yeRr is

m
the output to be regulated, y m eR are the measurable out-
l
puts of the system weR are the disturbances in the system,
v eR
re f are the reference inputs and eeR r is the error in
the system.
It is assumed that the disturbances w arise from the

following system:
n
n A
l
l " lnl ' n
i£R (13.2)
and that the reference inputs yref arise from the system:
n.
n2
a (13.3)
y = Ga
ref
It is assumed for nontrivality that. sp(A ) c C + ,

sp(A 2 ) c C where sp (•) denotes the eigenvalues of (.)
and C denotes the closed right half complex plane. It is
also assumed with no loss of generality, that (C ,A ) ,
(C 2 ,A 2 ) are observable and that rank C=r, rank B=m, rank
C = r rank = rank C = Q rank G
m m' (FJ l ' = ra nk C 2 = dim(a).
The following specific types of servomechanism problems are

now posed for the plant (13.1):
249
I. Feedforward Servomechanism Problem:

Find a controller for (13.1) (with input ym, y r e f and
output u) so that:
(a) The resultant closed-loop system is stable.

(b) Asymptotic tracking occurs, i.e.
lim e(t) = 0, V x(0) £ R n , V n i1<0)eR 1r
t-« n
V n 2 (0)eR %
and for all controller initial conditions.
II. Perfect Control Problem

Find a controller for (13.1) so that (a), (b) above hold
and in addition for any 6 > 0:
(c) || e T Qed T || < 6, V bounded x(0)eR n , n i (0)eR , l
n 2 (0)eR ^
and all bounded controller initial conditions.
III. Feedforward Control Problem with Feedback

Gain Perturbations
Assume the following controller is applied to (13.1):
u = + K y + K y +
V l m 2 ref' * = V A ^ + A 2 y ref
then its desired to find a controller for (13.1), so that

(a), (b) above hold and in addition:
(d) Condition (b) holds for any variations in K Q , K 1 , K 2
which do not cause the resultant system to become
unstable.
IV. Weak Robust Servomechanism Problem

Find a controller for (13.1) so that conditions (a), (b),
(d) above hold and in addition:
(e) Condition (b) holds for any variations in the elements
of (A,B) which do not cause the resultant system to
become unstable.
250
V. Robust Servomechanism Problem

(d) above hold and in addition:
(f) Condition (b) holds for any variations in the plant
model (including changes of model order) which do not
cause the resultant system to become unstable.
VI. Strong Robust Servomechanism Problem

(d), (f) hold and in addition:
(g) Approximate error regulation occurs for any variations
in the parameters of the controller used, the approxi-
mation becoming arbitrarily close to the exact solution
as the controller parameter variation becomes smaller.
VII. Perfect Control Robust Servomechanism Problem

(c), (d) and (f) hold.
VIII. Robust Decentralized Servomechanism Problem

(d), (f) hold and in addition:
Q
(h) The resultant controller is decentralized .
IX. Sequentially Reliable Servomechanism Problem
Find a decentralized controller for (13.1) so that condi-

tions (a), (b), (d), (f), (h) hold and in addition:
(i) If a failure occurs in the i control agent's con-

troller, (1 <_ i £ v ) , conditions (a), (b) , (d) , (f) ,
(h) still apply to the remaining part of the system
consisting of control agents l,2,...,i-l.
X. Robust Servomechanism Problem for Plants with Time Lags
Given that the plant (13.1) is now modelled by

251
X = AQ0X + A l x ( t - A ) + Bu + EGO
C
mx(t~A> (13.4)
y = cQx ^xCt-A) + Du + Fu)
find a controller so that conditions (a), (b), (d), (f) hold.
13.3.2 Existence Results
The following definitions are required in the development

to follow.
Definition
Consider the system (13.1), (13.2), (13.3); then the sys-

tem (C,A,B,D,E,F) is said to be steady-state invertible
with respect to the class of inputs (C,,A ;c o ,A 2 ,G) if and
only if the following two conditions hold:
I ®A-At 0 1
n. In
rank
1
I
n
l
®C
\ *D j
i 1
= rank
I ®C I ®D (I ®F)C-
I 0B
n
2
rank
I ®C I ®D
n n
2
= rank
I ®C ®G)C
~29
n 2
2
Definition
Consider the system (13.1), (13.2) and assume that A, A x

have disjoint eigenvalues (note that if (Cm,A,B) is stab-
ilizable and detectable, this can always be achieved by
applying static output feedback from y m ) ; then the system
252
(C,A,F,F:C ,F ) is said to be output-detectable with respect

to the class of inputs if and only if:
FC +CT
F
mCl+CmT 1
(F
mCl+CmT)Al = rank
rank
n,-! <FmC mT)A 1
(F C +C T)A n^-1
m l m l
(FC T)A
where T is the solution of AT-A X T = -E
Definition
Let the minimal polynomial of A , A 2 be denoted by

(s), A ( s ) respectively, and let the least common multiple
of A (s), (s) be given by A(s). Let the zeros of A(s)
(multiplicities included) be given by A ,A~# Then
A ,\j,...,\ are called the disturbance/reference signal
poles of (13. 2), (13.3).
The following results are obtained re the existence of a

solution to the servomechanism problems defined previously.
I. Feedforward Servomechanism Problem
Theorem I 5 ' 6
There exists a solution to the feedforward servomechanism

problem for (13.1) if and only if the following conditions
all hold:
(i) (C ,A,B) is stabilizable and detectable.

(ii) (C,A,B,E,F) is steady-state invertible with respect
to the class of inputs (C.,A.;C 2 ,A 2 G).
(iii) (C,A,E,F;C ,F ) is output detectable with respect to
the class of input (C.,A 1 ).
The following is a sufficient condition which guarantees

that theorem 1 always holds.
253
Corollary 1
There exists a solution to the feedforward servomechanism

problem for (13.1) if the following conditions all hold:
(i) (C #A,B,) is stabilizable and detectable.

(ii) m >_ r.
2
(iii) the transmission zeros of (C,A,B,D) do not coincide
with X±, i = 1,2,...,p.
(iv) either y = y or co = 0 or r >_ Q and the trans-
mission zeros of (C .A,E,F ) do not coincide with
any eigenvalues of A.. .
II. Perfect Control Problem
Theorem 2 1 0 ' 1 3
There exists a solution to the perfect control problem for

(13.1) if and only if the following conditions all hold:
(i) (C,A,B) is stabilizable and detectable.

(ii) m > r.
(iii) the transmission zeros of (C,A,B,D) are contained in
the open left part of the complex plane (i.e. the
plant is minimum phase).
are
(iv) the states x, n-,/ n 2 available for measurement.
The following result follows directly from theorem 2 and

shows the importance of using to full advantage the total
number of manipulated inputs which a plant may have.
Theorem 3 (Control Inequality Principle) '
Given the system (13.1), then:
(i) If m > r, there exists a solution to the perfect con-

trol problem for "almost all" systems (13.1).
(ii) If m = r, there may or may not exist a solution to
the perfect control problem depending on whether the
plant is minimum or non-minimum phase.
(iii) If m < r, there never exists a solution to the perfect
control problem.
254
III. Feedforward Control Problem with Feedback

Gain Perturbations
Theorem 4
There exists a solution to the feedforward control problem

with feedback gain perturbations if and only if the follow-
ing conditions all hold:
(i) there exists a solution to the feedforward control

problem, i.e. theorem 1 holds.
(ii) the output y must contain either "partial errors"

or "equivalent errors" or both (i.e. the output y
must contain some of the outputs y ) .
IV. T%7eak Robust Servomechanism Problem
Theorem 5 7 ' 3
There exists a solution to the weak robust servomechanism

problem for (13.1) if and only if the followincr conditions
all hold:
(i) (C ,A,B) is stabilizable and detectable
(ii) m > r

with X i , i = l,2,...,i
C
m
(iv) rank C_ = rank ,
V. Robust Servomechism Problem
Theorem 6 7 ' 3
There exists a solution to the robust servomechanism prob-

lem for (13.1) if and only if the following conditions all
255
hold:
(i)
(Cm,A,B) is stabilizable and detectable
(ii) m _> r

with \±, i = 1,2,... ,p
(iv) the outputs of y are physically measurable.
VI. Strong Robust Servomechanism Problem
Theorem 7
There exists a solution to the strong robust servomechan-
ism problem for (13.1) if and only if the following condi-
tions both hold:
(i) there exists a solution to the robust servomechanism
problem, i.e. theorem 6 holds.
(ii) the class of disturbance and reference input signals
is bounded, i.e. the disturbance, reference siqnals
must have the property that Re(X i ) = 0,
i = l,2,...,p and that A1,A2/...,A are all
distinct.
256
VII. Perfect Control Robust Servomechanism Problem
Theorem 8 (Robust Control Limitation Theorem) '
There exists no solution to the perfect control robust

servomechanism problem.
VIII. Robust Decentralized Servomechanism Problem
Assume that the plant (13.1) consists of v control

agents modelled by:
x = Ax+ B.u. + Eo)
<*
i = 1,2,...,v (13.5)
y = C
i ix D
iui F
iw
where y^ c R is the output to be regulated of the ith

control station, y1 are the measurable outputs of the ith
control station, ui are the control inputs of the ith con-
trol station and e. is the error in the ith control sta-
sta-
tion. It is assumed that the disturbance signals to satisfy
(13.2) and that the reference signals ref
'ref 4
satisfy (13.3).
Definition
Let
k , D 4 block diag (D ,D2,...,D ) ,
C . c1 >
m
B 4 (B1#B2#...fBv), C
m A c2
m
cV
m '
and let
257
0 0 .. 0
0 .. 0
r
l
0 0 .. 0
0 X 0
r2 '•
0 0 .. 0
0 0 .. I
o
Theorem 9
There exists a solution to the robust decentralized servo-
mechanism problem for (13.5) if and only if the following
conditions all hold:
(i) (C ,A,B) has no unstable decentralized fixed modes.
(ii) The decentralized fixed modes of the p systems

C
m'<A j = 1,2,...,p
do not contain A., j = , ..,p, respectively,
(iii) y1 contains the output i = l,2,...,v, respec-

tively.
IX. Sequentially Reliable Robust Servomechanism Problem
Definition
Let 0
P+l
Theorem 10 14
Assume that the plant (13.1) is open loop asymptotically
stable; then there exists a solution to the robust decen-
tralized servomechanism problem for (13.5) with sequential
reliability if and only if the following conditions all hold:
(i) there exists a solution to the robust centralized

servomechanism problem for (13.1) with control agent 1
for the class of disturbance/reference signal poles
258
A
l ' X 2 / # * * / X p' X p+l"
(ii) there exists a solution to the robust decentralized
servomechanism problem for (13.1) with control agents
1, 2 for the class of disturbance/reference signal
X X
poles 1' 2 ' * * *' A p' X p+l*
(v) there exists a solution to the robust decentralized

servomechanism problem for (13.1) with control agents
l,2,...,v for the class of disturbance/reference sig-
x x A A #
nal poles 1i 2'•••' p' p+l
X. Robust Servomechanism Problem for Plants with Time Lags
Definition
Given (13.4), let A be defined as follows: given the

system x = AQx+A;Lx (t-A) +Bu, it may be written in the form:
xt = A x t + Bu
where xfc e M and where:
, dx t
de , e e [-A ,0)
Ax.
A Q x t (0)+A ± x t (-A) , e =0
0 [-A ,0)
CD
u)
Bu
Bu 0
CD
II
The following result is obtained:
Theorem 11
There exists a solution to the robust servomechanism pro-

blem for (13.4) if and only if the following conditions all
hold :
(i) rank[A 0 +A 1 e~ X A -AI / B] = n, V X e sp (A) with Re(A) _> 0

259
(ii)
rank = n, V A e sp(A) with Re(A)>0
(iii) B
rank -A, A = n+r, i = 1,2,...,p
D
m
(iv) y is measurable.
1_3 . 3.3 Robust Servomechanism Controller Structure
The following definition of a stabilizing compensator and

servocompensator are required in the development to follow.
Definition (Stabilizing Compensator)
Given the stabilizable-detectable system (C ,A,B,D ) rep-

resented by
X = Ax + Bu
C x
y
m = m "h D
mu
a stabilizing compensator: u = •îin+^2ym' n =

^l n+ ^2^m ^S
defined to be a controller which will stabilize the result-
ant closed loop system such that "desired" transient behav-
iour occurs. It is not a unique device and may be designed
by using a number of different techniques, e.g. observer
theory, LQR theory, output pole assignment methods, modal
design methods, frequency design methods, etc. The ability
to achieve a "desired" transient behaviour depends on the
location of the transmission zeros of (C ,A r B,D), the fixed
modes of (Cm,A,B) etc.
Consider now the system (13.1) and assume that a solution

to the robust servomechanism problem exists, i.e. theorem 6
holds. The following definition is needed:
Definition
Given the disturbance/reference poles A^, i = l,2,...,p,

the matrix C e Rpxp and vector y e Rp is defined by:
260
c 4
pJ
where the coefficients ±, i = l,2,...,p are given by the
coefficients of the polynomial (X-X i ) # i.e.
A P + 6 A*
n
The following compensator, called a servo-aompensator is
of fundamental importance in the design of robust controllers
for (13.1):
Definition (Servo-Compensator)
Consider the system (13.1), (13.2), (13.3); then a servo-

compensator for (13.1) is a controller with input e c R and
output £ e R r p given by:
* (13.7)
= C
B e
where
C 4 block diag(C,C,...,C), B A. block diag(y/ ,..,y)
r r
where Cry are given by (13.6). The servo-compensator is

unique within the class of coordinate transformations and
nonsingular input transformations.
Robust Servomechanism Controller
Consider the system (13.1); then any robust controller

for (13.1) consists of the following structure:
u = KQx + K (13.8)
where e R ^ is the output of the servo-compensator

(13.7) and where x is the output of a stabilizing compen-
v
sator S*,(with inputs refr
where S , KQ, K
are found to stabilize and give desired transient behaviour
261
to the following augmented system:
0
(13.9)
EC C B D
C
m
where t h e fixed modes (if any) o f j f c 0|,| A 0 1,f B

[o ij [B*C c*) [B*D
are equal to the fixed modes of (C ,A,B).
In the above robust controller, the error input e must be

obtained by using the physical outputs corresponding to y.
It is to be noted that the robust controller always has order
1 rp.
13.3.4 Some Properties of the Robust Controller
1. In the above robust controller, it is only necessary to

know the disturbance/reference poles i.e.
it is not necessary to to know E,F,G,C]L ,C 2 ,A1 ,A2 .
V
The robust controller is robust for any changes (i.e. not
just small perturbations) in the plant parameters (C,A,B,
D;C ,D ;E;F;F ) , feedback gain parameters K Q , K, stab-
ilizing compensator parameters of S , and for any changes
in the order of the assumed mathematical model describing
the plant, provided only that the resultant closed loop
system remains stable.
A robust controller exists for "almost all" plants (13.1)
provided that (i) m > r and (ii) the outputs to be regul-
ated can be physically measured; if either (i) or (ii)
fail to hold then there never is a solution to the robust
servomechanism problem.
13.3.5 Various Classes of Stabi1iziiig_Compensators
Various special classes of stabilizing compensators

which can be used in the robust servomechanism controller
(13.8) are given as follows:
262
1. Observer-Stabilizing Compensator 12 : Assume that theorem

6 holds; then an observer stabilizing compensator S for
(13.8) consists of the following device: x in (13.8) is the
output of an observer of order n (either full order n < n
or reduced order n £ n-r) with input y , u which estimates
the state of the detectable system: x = Ax+Bu, y = C x+D u,
and where KQ, K in (13.8) are found, using standard methods,
to stabilize the following stabilizable system:
u, u = (KQ,K)
* B D
B C
(13.10)
In case the plant (13.1) is open loop asymptotically stable

and that Re(A.) = 0, i = 1,2,...,p, the following general-
ized three-term controller can be used for S .
2. Generalized Three-Term Controller: Consider the system

(13.1); then the following compensator
= (13.11)
V K
lym
where y 4 y -D u, y 4 y m ~C Bu, where is the output of

the serve-compensator (13.7) is called a generalized three-
term controller for (13.1). This compensator has the follow-
ing property.
12
Theorem 12
Consider the system (13.1) and assume that A is asymptot-

ically stable, Re(A i ) = 0, i = l,2,..,,p and that theorem 6
holds; then there exists K Q , K.., K 2 so that the resultant
controlled system obtained by applying the generalized three
term controller (13.11) to (13.1) is asymptotically stable.
Remaxk_JL: In this case, arbitrary pole assignment cannot

necessarily be obtained using (13.11) unlike the case of the
observer stabilizing compensator. However, it follows from
Davison, Wang (1975) that (m+2r+rp-l) symmetric poles can be
assigned to the closed-loop system for "almost all" (Cm,A,B)
systems. This implies that if n is not too large, i.e. if
263
n £ (m+2r-l) that complete pole assignment can be carried

out using (13.11) (even if A is unstable).
13.4 Extension of Previous Results
13.4.1 Control of Unknown Systems (Multivariable

Tuning Regulators)
The previous results assumed that the plant model (13.1)

is known, and based on a knowledge of (13.1), conditions for
the existence of a robust controller and the construction of
a controller were obtained. The question may now be asked -
to what extent can one determine the existence of a solution
and obtain a robust controller, when the plant model (13.1)
is completely unknown? The following result shows that by
carrying out some simple experiments on the plant (actual),
the existence of a solution and the construction of a robust
controller can be obtained. In this case, asymptotic regu-
lation/tracking can be achieved, but the ability to achieve
arbitrary fast speed of response, as was obtained when the
plant is known, cannot now be necessarily carried out.
The following assumptions are made:
Assumptions
(i) The plant (actual) can be described by a linear time-
invariant model and is open loop asymptotically stable.
(ii) The disturbance/tracking poles have the property that
Re(X ± ) = 0 , i = 1,2,...,p.
(iii) The outputs y can be measured.
The following definition is now made:
Definition
The tracking steady-state gain parameters of (13.1) are
defined as follows:
T ± 4 D + C(A i I - A)" 1 B , i = 1,2,...,p
Remarks
2. The tracking steady-state gain parameters of a system
can be easily identified by sequentially applying m linearly
264
independent control signals to the plant (actual) and by

measuring the steady-state values of the outputs y for
each of the input signals. In this case, the input signals
applied to the plant are of the same class as the class of
disturbance/tracking signals for which the controller is de-
signed.
3. It is to be emphasized that no other assumptions re

(13.1), e.g. the order of the plant, minimum phase conditions,
controllability-observability, etc., are required in this
problem statement. This differs from the well known self-
tuning regulator theory of Astrom (1973).
Existence Result
Assume that T^, i = l,2,...,p have been experimentally

obtained for the plant. The following existence result- can
then be applied:
o
Theorem 13
There exists a solution to the robust servomechanism pro-
blem for (13.1) if and only if
rank T^ = r i=l,2,...,p
Controller Structure
Assuming that T^, i = l,2,...,p has been determined and

that theorem 13 holds, a robust controller can be obtained
for the plant by carrying out a sequence of one-dimensional
on line experiments on the plant with the servo-compensator
(13.7) connected. This process of adjusting a single gain
parameter is called "tuning the controller on line"; in this
case, it can be guaranteed that the resultant closed-loop
system always remains asymptotically stable. For further
details see reference 9.
13.4.2 Multivariable Error Constants
In classical control, the position, velocity and accelera-

265
tion errors can be used as an indication of "how well" a con-

troller which is designed to give zero asymptotic error for
a given class of signals, will behave when subject to a
larger class of disturbance/reference signals than the de-
sign signals. The following generalization of these defini-
tions can be used in the robust servomechanism problem to
give an indication of "how well" the robust controller will
behave when subject to a larger class of signals than the
design signals. The following special case of error con-
stants is obtained from reference 18.
Given the plant (13.1), assume that theorem 6 holds for

the case of polynomial disturbance/reference signals (i.e.
with A^ = 0, i = 1,2,...,p), and assume the following con-
trollers have been applied to (13.1) so that the resultant
system is asymptotically stable.
Controller 0
u = K(yret^-y) + {K,
l + Ku
A (sI-A n )
u A,}y
l •* m, y m Ar" (ym-D mu)
Controller 1
u = -( " " . . - ! - -
Controller 2
u =
1 2 2
Here, K/s, K +K s/s is a servo-compensator for (13.1) for
the case of constant, ramp signals respectively, and
K , + K Q ( S I - A Q ) "* A.^ is a stabilizing compensator for the sys-
tem.
Let e A, y ~-y ; then the following steady-state errors
are obtained for the controlled system when reference sig-
nals y yty1 t 1 , i = 0,1,2
r; = where y is a constant, and
_ i _
disturbance signals u = wt , i = 0,1,2 where w is a
constant, are applied to the plant controlled by controllers
0,1,2 (see Tables 13.1, 13.2).
266
TABLE 13.1 Generalized Steady-State Errors for

Various Reference Signals
=
y v ^ref=? t2
Controller 0 G y 00 00
t
0
Controller 1 0
V 00
Controller 2 0 0
where:
G.r A {I+C[A+B(K -K n A~ X A )Cni]'"1BK}"1 Tracking Position
0 l u u I m Error Gain
Tracking Velocity
Error Gain
. A " 11 A 1 )C ]~
2{C[A+B(K 1 -K n A" ~ 11BBK
K 11}}"
" 11 Tracking Accelera-
2
'2 tion Error G a m
where:
C A C + D(K1-KQAQ A1)C.
m
TABLE 13.2 Generalized Steady-State Errors for
Various Disturbance Signals
-4.2 1
Controller 0 00
; Controller 1 0 G j 0) oo
d
l
Controller 2 0 0 G
d ^
where
e, A -G. {C[A+B(K 1 -K n A n 1 A 1 )C m ]" 1 E+F} Disturbance Posi-

"d
d fc
0 "" 0 1 0 0 1 m t±on Error Gain
e, A - G . {C[A+B(K 1 -K n A" 1 A 1 )C m ]" 1 E+F} Disturbance Velo-
d := fc
l l 1 0 0 1 m c±ty Error Ga±n
e, A - G . {C[A+B(K,-K n A" 1 A 1 )C l"1!:^?} Disturbance Accel-
d
2 "" t 2 l u u i m eration Error Gain
where:
E A, E + B(K 1 -K 0 A- 1 A 1 )F m
F 4 F + D(K 1 -K 0 A- 1 A 1 )F m
267
It is seen therefore that if a robust controller is design-

ed to give asymptotic regulation for the class of constant
signals say (Controller 1 ) , it is clearly desirable to have
the velocity error gains as "small as possible" (subject to
other constraints in the controller design).
13.5 Robust Controller Design
For a given servo-compensator and stabilizing-compensator

structure, the question of determining the "optimal" feed-
back gains and parameters of the stabilizing compensator has
not been addressed. Generally it is desirable to determine
these parameters so as to stabilize the overall closed-loop
system, maximize speed of response in the system, minimize
interaction between the reference signals and outputs to be
regulated, subject to various constraints on the system such
as integrity in the case of sensor/actuator failure, damping
factor constraints, etc.
This is a topic of immense activity of which a number of

different design approaches have been suggested. One such
approach that can deal with constraints in a natural manner
is to pose the problem as a parameter optimization problem,
i.e. the gain parameters and stabilizing compensator para-
meters are found so as to minimize a performance index which
reflects the speed of response and interaction of the closed
loop system, subject to various constraints on the closed
loop system. The following gives a brief review of such a
design approach '
Problem Statement
Given the plant (13.1), assume A., i = l,2,...,p is given,

that theorem 6 holds and that the stabilizing compensator
u = K Q y m + K i n + K£ , n = AQn + A ^ (13.12)
is to be used where dim(n) is specified; then the closed

loop system is described by (13.1), (13.7) and (13.12).
268
Definition
Let z A D£ where £ is the output of the servo-compensa-
tor (13.7) where
D A block diag{(1,0,...,0),(1,0,...,0),..., (1,0,...,0) },

r
and let z^, u^ denote the steady-state value of z, u for
a given disturbance/reference-input signal. Let Au A. u-u^,
AZ A Z-Z^.
Performance Index
Let
r °°
I nil *P
J £ E (AzTOAz+yAu RAu)dx , Q > 0 , R > 0 ,
J 0 (13.13)
where p -> 0, be the performance index for the system,

where the expected value is taken with respect to all initial
conditions x(0), £(0) , n(0) , n x ( 0 ) , n 2 (°) uniformly distri-
buted on a unit sphere.
Then the following parameter optimization problem is to be

solved:
Parameter Optimization Problem
min J subject to various constraints imposed

K 0 ,K 1 ,K,A Q ,A 1 on the problem formulation
For further details see references 11, 12.
13.5.1 Design Example
The following example gives an illustration of the pro-

posed design procedure. The parameter optimization was
carried out for the case Q = I , R = I , y = 10" .
Example 13.1 (Unstable batch reactor)
An unstable batch reactor (with open loop eigenvalues

1.99, 0.064, -5.057, -8.67) described by:
269
1.38 -0.2077 6.715 -5.676 0 0

-0.5814 -4.29 0 0.675 5.679 0
x= 1.067 4.273 -6.654 5.893 x+ 1.136 -3.146
0.048 4.273 1.343 -2.104 1.136 0
1 0 1-1
y =
0 1 0 1
is to be regulated against constant disturbances such that

the output tracks constant set-points y f, subject to the
integrity constraint that the system's open loop unstable
eigenvalues (i.e. Re(A) <_ 1.99x1.l = 2.19) for a sensor
failure in output y1 or y2.
In this case, since the system has transmission zeros

(-1.192, -5.039) and no fixed modes, there exists a solution
to the robust servomechanism problem from theorem 6. The
following robust controller is obtained from (13.8)
u = KQy (y-y ref )dx

0
and it is desired to determine KQ, K1 so as to minimize

J given by (13.13) subject to the above integrity constraint.
The following result is obtained:
-2
10
opt
eigenvalue^ <- 1 ' 19 ' " 5 - 0 4 ' ~^±^^' -168±jl69>

0.95 -58.4 100 -1.0x10"
(K
0' K) optimal 79.3 -0.72 1.0x10 -100
In this case the constraints are satisfied; in particular

the active constraint occurs when the output sensor for y-j^
fails, and the resulting eigenvalues obtained in this case
are given by (2.15, 0, -1.01, -8.66, -168+J169). It is to
be noted that with this controller a "pole-zero" cancella-
tion occurs with respect to the eigenvalues -1.19, -5.04 .
Some typical responses of the closed-loop system are given
in Figure 13.1 for the case of set-point change in Yref
and for the case of an arbitrary constant disturbance
270
E = (1 1 1 I ) ' for zero initial conditions. It is seen

that excellent tracking and regulation occur.
0.01
10
' ^0.02 0.04
/ -0.01 t.s
0
0.02 0.04
t.s a
-0.005
0 "~" 0.02
t.s
0.04
b
-f 0.02 0.04
t.s
0005 • 0.005 -
0.02 0.04 0.02 0.04

t s t.s
Fig. 13.1 Response of controlled system for example 1

using proposed design method.
(a
> *ref a) = 0, (b) = 0,
! ) •
(c)
*ref ) , E(D = (1 1 1 1)
13.6 Conclusions
A brief overview of some recent results obtained re the

robust control servomechanism problem has been outlined in
this chapter. In particular, structural results for the
following classes of servomechanism problems have been ex-
amined: the feedforward control servomechanism problem
(theorem 1 ) ; the perfect control problem (theorem 2 ) ; the
feedforward control problem with feedback gain perturbations
(theorem 4 ) ; the weak robust servomechanism problem (theor-
em 5 ) ; the robust servomechanism problem (theorem 6 ) ; the
strong robust servomechanism problem (theorem 7 ) ; the rob-
ust control limitation theorem (theorem 8 ) ; the robust
decentralized servomechanism problem (theorem 9 ) ; the
sequentially reliable decentralized servomechanism problem
(theorem 10); the time lag robust servomechanism problem
(theorem 11); tuning regulations for unknown systems
271
(theorem 13); and generalized error constants for multi-

variable systems (Section 13.4.2).
The chapter concludes with a brief description of a robust

controller design method using parameter optimization methods.
An example of the method is applied to design a robust con-
troller for an unstable 4th order 2-input, 2-output system
with an integrity constraint.
References
1. ASTROM, K.J., G. VJITTENMARK : "On Self Tuning Regula-

tors", Automatica, vol. 9, (1973), pp. 185-200.
2. DAVISON, E.J., S.H. V7ANG : "Properties and Calculation
of Transmission Zeros of Linear Multivariable Time-
Invariant Systems", Automatica, vol. 10, (1974),
pp. 643-658.
3. DAVISON, E.J., A. GOLDENBERG : "Robust Control of a
General Servomechanism Problem. : the Servo Compensator",
Automatica, vol. 11, (1975), pp. 461-471.
4. DAVISON, E.J., S.H. WANG : "On Pole Assignment in
Linear Multivariable Systems Using Output Feedback",
IEEE Trans, on Automatic Control, vol. AC-20, no. 4,
(1975), pp. 516-518.
5. DAVISON, E.J. : "The Steady-State Invertibility and
Feedforward Control of Linear Time-Invariant Systems",
IEEE Trans, on Automatic Control, vol. AC-21, no. 4,
(1976 ) , pp. 529-534.
6. DAVISON, E.J. : "Output Detectability, Steady-State
Invertibility and the General Servomechanism Problem",
IEEE Conference on Decision and Control, Dec. 1976,
pp. 1250-1257, (1976 ) .
7. DAVISON, E.J. : "The Robust Control of a Servomechan-
ism Problem for Linear Time-Invariant Multivariable
Systems", IEEE Trans, on Automatic Control, vol. AC-21,
no. 1, (1976 ) , pp. 25-34.
8. DAVISON, E.J. : "The Robust Decentralized Control of
a General Servomechanism Problem", IEEE Trans, on
Automatic Control, vol. AC-21, no. 1, (1976 ) , pp.14-24.
9. DAVISON, E.J. : "Multivariable Tuning Regulators: the
Feedforward and Robust Control of a General Servomecha-
nism Problem", IEEE Trans, on Automatic Control, vol.
AC-21, no. 1, (1976 ) , pp. 35-47.
10. DAVISON, E.J., S.G. CHO™ : "Perfect Control in Linear
Time-Invariant Multivariable Systems: The Control
Inequality Principle", in Control System Design by
Pole-Zero Assignment (editor: F. Fallside), Academic
Press, 1977.
272
11. DAVISON, FERGUSON, I.J. : "Design of Controllers for

the Multivariable Robust Servomechanism Problem Using
Parameter Optimization Methods - Some Case Studies",
2nd IFAC Workship on Control Applications of Nonlinear
Programming and Optimization, Munich, Germany, 15-17
September, 1980.
12. DAVISON, E.J., I.J. FERGUSON : "The Design of Con-
trollers for the Multivariable Robust Servomechanism
Problem Using Parameter Optimization Methods", IEEE
Trans, on Automatic Control, vol. AC-20, no. 1, (1981),
pp. 93-110.
13. DAVISON, E.J. : "Optimal Feedforward-Robust Controll-
ers for the Multivariable Servomechanism Problem",
Dept. of Electrical Engineering, University of Toronto,
Systems Control Report No. 8105, July (1981 ) .
14. DAVISON, E.J. : "Reliability of the Robust Servo-
mechanism Controller for Decentralized Systems", 8th
IFAC World Congress on Automatic Control, Kyoto, Japan,
Aug. 24-28, (1981 ) , to appear.
15. GOLDENBERG, A., E.J. DAVISON : "The Solvability of
the Robust Parameter Feedback Servomechanism Problem",
1976 Conference on Information Sciences and Systems,
April (1976) , John Hopkins University, Baltimore.
16. GOLDENBERG, A., E.J. DAVISON : "The Robust Control of
a Servomechanism Problem with Time Delay", 7th IFAC
World Congress on Automatic Control, Helsinki, Finland,
June, 1978, pp. 2095-2102.
17. SAIN, M.K., (Editor) : IEEE Trans, on Automatic Con-

trol, Special Issue on "Linear Multivariable Control
Systems", vol. AC-26, no. 1, Feb. 1981.
18. SCHERZINGER, B., E.J. DAVISON : "Generalized Error

Coefficients for the Multivariable Servomechanism
Problem", Dept. of Electrical Engineering, University
of Toronto, Systems Control Report No. 8103, June 1981,
IEEE Control and Decision Conference, Dec. 1981,
to appear.
19. SPANG, A., L. GERHART : GE-RPI-NSF Workshop on

Control Design, Schenectady, N.Y., May 20-22, 1981.
Chapter 14
Control of distributed parameter systems

Dr. A. J. PRITCHARD
Synopsis
Much of control theory is concerned with systems which are

modelled by ordinary differential equations, so-called lumped
parameter systems. In this chapter it is shown how the con-
cepts of controllability, observability, optimal control and
estimation may be investigated for system models based upon
partial differential equations, so-called distributed para-
meter systems. Such a system with a single input and a
single output is used to illustrate the theory presented.
14.1 Introduction
If the state of a system at each time t cannot be descri-

bed by a finite set of real numbers then we will call the
system a distributed parameter system. Very important
classes of distributed parameter systems are those described
by partial differential equations and differential-delay
equations. Since physicists and applied mathematicians have
shown that many complex physical phenomena may be modelled
quite accurately by distributed parameter systems it is nec-
essary to consider how they may be controlled. Obviously it
will not be possible to examine all the control theoretic
topics ' in this chapter but it is hoped to show that these
systems pose new problems which are not insurmountable.
Throughout the chapter reference will be made to the follow-
ing simple example.
Example 1
Consider a metal bar in a furnace. The bar is heated by

jets whose distribution b(x) is fixed but whose magnitude
274
can be varied by a control u(t). For one dimensional flow,

heat balance considerations yield
4r (x,t) dx = a ~ (x,t) + b(x)u(t)dx

3t [dX j J
JX- x- x+
where 0(x,t) is the temperature in the bar at time t, and
distance x from one end. a is the thermometric conduct-
ivity. Differentiation yields
2
| | (x,t) = a i - | (x,t) + b(x)u(t)
Let us assume that the temperature at each end is kept at a

constant value T, and the initial temperature distribution
in the bar is 0Q(x), so that
0(0,t) = 0U,t) = T
0(x,O) = 0Q(X)
where Z is the length of the bar. Set
0(x,t) = T(x,t) + T
e Q (x) = T Q (x) + T
then
-^ (X/t) = a i-i. (X/t) + b(x)u(t)

3t
9x 2
(14.1)
T(0,t) = TU,t) = 0
T(x,0) = T Q (x)
Finally assume that it is possible to sense the temperature

at a given point and this is accomplished by taking a weighted
average of the temperature at nearby points. Then the output
of the sensor can be modelled by
h(t) = f c(x)6(x,t)dx
J
0
If h(t) = y(t) + T
- [[ c(x)dx,
i
275
then
y(t) = c(x) T(x,t) dx (14.2)
Jo
Equations (14.1), (14.2) represent a single input-single out-
put distributed parameter system.
The results to be described are more easily formulated by
abstracting the system. To do this let us ask what proper-
ties one would expect of the solution of a dynamical system
on a Hilbert space, H. If it is assumed that the dynamics
which govern the evolution from the initial state z to
z(t) are linear and autonomous, then for each time t, we
can define a linear operator T., such that
TQ = I (14.3)
z(t) = T t zQ (14.4)
Moreover, if z(t+s) is the same point in H as the point

reached by allowing the dynamics to evolve from z(s) for a
time t, we have
T z = T
z(t+s) = t+s o tz(s) = T
tTszo
T = T
t+s t Ts (14
*5)
If we also assume that the solution is continuous in time,

then the operator T. with properties (14.3), (14.5) is
called a strongly continuous semigroup.
Now let us define an operator A by
(T -I)
Az = lim — \h z- , z e D (A)
n
h->0+
where D(A) is the set of points for which the limit exists.
Then it can be shown that (14.4) is the solution of the
differential equation
z(t) = Az(t), z(o) = ZQeD(A)

276
Example 1 may be cast as a control problem
z(t) = Az(t) + bu(t), z(o) = z

o
(14.6)
y(t) = <c, z(t)>
with solution
z(t) = Ttz + T bu(p)dp (14.7)

0
If the initial state is zero, then the input-output map is
of the form
ft
y(t) = T.
<c, T b> ul
t _ p b> u(p)dp
By taking the Laplace transform we obtain
y(s) = <c, R(s,A) b> u(s)
where R(s,A) is the resolvent (sI-A)"" . So the transfer

function is
<c, R(s,A) b>
In order to calculate this for Example 1 we need to iden-

tify the Hilbert space H as L 2 (0,£), and
Az = —
ax
3z
D(A) = (zeH : *-% eH, z (o) = z (£) = 0)
Then we need to solve
(sI-A)z = Z]L
for z. e H, z e D(A). This is the ordinary differential

equation
277
d2z
sz(x) - cA-~(x) = z. (x)
Z 1
z(0) = zU) = 0
The solution is
z(x) = _1 [sinh(s/a)^x f 8lnh(8/o)%(£.a) (o)da
sinh(s/a)^(x-a)z (a)da
So the transfer function is
c(x) I
sinh(s/a)^(£.a)b(a)da
fx 3k 1
sinh(s/a) Mx-a)b(a)da dx
J
0 JJ
It is evident that this is not a rational function of s
so that new theories for irrational transfer functions need
to be developed for distributed parameter systems. An al-
ternative approach is to show that the solution is equiva-
lent to the solution of an infinite number of ordinary
differential equations. Let us suppose that the spectrum
of A consists entirely of distinct real points A^ ,
i = 1,2,.... where
= \± <$>± , $± e D(A) (14.8)
If the adjoint vectors are ty. , so that
e D(A*)
* (14.9)
then <<j> . , i|>. > = 0 for i f j .
A solution of (14.6) is sought of the form

278
z(t) = I a n (t)* n
then
a bu
z(t) = I a (t)4 = A I (t)<f> + < fc )
n=l n n=l
n=l
I Vn(t)*n + bu(t)
If we assume
b c =
= n=l
^Vn ' J Vn
n=l
°° 2 °° 2
where £ b < «> , £ c <»
n=l n=l
then by taking inner products with $ we obtain the un-
coupled system of ordinary differential equations
an(t) = * n a n (t) + bnu(t) n = 1,2,
Jl
For Example 1 we have 4 (x) = (>
j (x) = /ijl sin
0 0 0 n n
X = n Z ir Z a/£
14.2 Controllability
Let us consider a question of controllability for Example
1. If the initial state is zero, we have
ve
-f "f
(t-s)
n
an(t) = b u(s)ds
n
J0
Jn
Suppose we wish to steer the system to a temperature distri-
bution T.(x) at time t., where
279
then t. - .
1 x (t
n l~s)
e
b n u(s)ds
Using the Schwartz inequality
9 ^1 2X (t.
(t.-s)
b e ds
u 2 (s)ds
nl J j
for n = 1,2,...
or
2|tJ2 r .
\ | < f u i(s)ds for A < 0
n
*" In
But if the control energy is to be finite this requires
for some constants K, K. Thus only those states with

sufficiently small Fourier coefficients can be achieved.
Definition of Exact Controllability on (0,t ) :
We say that the system (14.7) or (A,b) is exactly con-

trollable on (0,t ) if for all z , z in the Hilbert
space H there exists a control u e L (0, t ) such that
tx
T. ^ bu(s)ds
It can be shown that this is the case if and only if

there exists a constant y > 0, such that
tx
f
f
J
|b*
| T* S
z|2 ds i Y||Z||2 (14.10)
00
where b z = <b,z>.
Inequality (14.10) is equivalent to

280
1
T g b b* T g z> ds >• Y| | Z | I 2
We have already seen that Example 1 cannot be exactly con-

trollable and so this may be too strong a definition for
distributed parameter systems.
Definition of Approximate Controllability on (0, t, ) :
System (14.7) or (A,b) is approximately controllable on

3, t.) if for all z Q and z in H and
exists a control u e 1/(0, t ) such that
where z(t.) = T. z + T. bu(s)ds
It can be shown that (A,b) is approximately controllable

on (0, t x ) if
b* T* z = 0 on (0,t1) (14.11)
implies z = 0
For Example 1
b* T* z = I b exp(-an2Tr2t/£2)a
r
n=l
where z = } a <j>
n=l n n
So Example 1 is approximately controllable on (0,t ) if
b ¥ 0 for all n = 1,2,..., where
n
281
14.3 Observability
Corresponding to those given in Section 14.2, dual concepts

of observability can be defined.
Definition of Initial Observability on (0, t ) :
We say that (14.6) or (A,c) is observable on (0, t ) if
<c, T t zQ> = 0 for t e ( 0 , tx) (14.12)
implies z = 0
We see immediately that this is the dual concept to approx-

imate controllability on (0, t^). In particular, Example 1
will be initially observable on (0, t ) if cn i 0 where
cn = f£ c(x)cj)n (x)dx n = 1,2,

jn
Although the initial observability implies that the map

from the output to the initial state is 1-1 it may not be
bounded, and this will lead to computational errors. A
stronger definition is as follows:
Definition of Continuous Observability on (0, t . ) :
We say that (14.6) or (A,c) is continuously observable

on (0, t-) if there exists a constant y > 0 such that
|<c, T z>| 2 ds > Y||Z||2

s
o
So continuous observability is the dual concept to exact
controllability and is a very strong requirement for distri-
buted parameter systems and is certainly not satisfied by
Example 1.
14.4 Optimal Control
Let us now pose the linear quadratic regulator problem,

where an optimal control is sought to minimize the functional
282
J(u) = <z(t x ) , Gz(t1)> ,Wz(s)> + Ru (s))ds

(14.13)
where z satisfies (14.7). We assume G >. 0, W > 0, R > 0.
The optimal control u* is given by
u*(t) = -R*"1 <b, P(t) z(t)>
where P( # ) satisfies the Riccati equation
~- <h, P(t) k> + <Ah, P(t) k> + <P(t)h, Ak>

+ <h, Wk>
-1 (14.14)
= R * <h, P(t) b> <b, P(t) k>
= G
So that the only difference between this and the finite

dimensional case is the inner product structure of the
Riccati equation. If we assume (14.8) and (14.9) hold, we
may seek solutions for P in the form
P(t)z = I I P.,(t) <*,,z>*v
Substituting in (14.14) and setting h = <J> , k = <J> , we

obtain
P + (A +A ) P + <d> . W6 >
mn m n' mn m' T n
(14.15)
= R -1
14.5 Optimal Estimation (Kalman Filter)
If the output is corrupted by noise we may pose a state

estimation problem, where the dynamics are
z(t) =
y(t) <c, z (s)> ds + v(t)

283
v(t) is a Weiner process with covariance V, and z is

Gaussian with zero mean and covariance Q . The state es-
timation problem is to find the best estimate of the state,
z(t) at time t based on the observation process y(s)
0 < s < t. The solution is given by the Kalman-Bucy filter
z(t) = V -1 T
t-s Q(s) c dp
where P ( ' ) is the innovations process
p (t) = y(t) - <c, z(s)> ds
and Q(m) satisfies the Riccati equation
— : <Q(t) h,k> - <Q(t)h, A k> - <A h, Q(t)k>
= V , Q(t) c> <c, Q(t) k> h,keD(A )

(14.16)
Q(o) =
If (14.8) and (14.9) hold we may seek solutions of the form
Q(t) =
J
j=l k=l "
Alternatively we may note that the transformations
Q(t)
(14.17)
V R
P G
send equation (14.16) into equation (14.14) and in this

sense the estimation and control problems are dual to each
other.
We have seen that the problems of controllability, obser-

vability, optimal control and estimation can be solved for
distributed parameter systems. The added complexity of
284
dealing with infinite dimensional systems must be balanced

against the small number of physically well defined para-
meters which occur in the partial differential equation
description and the decoupling of the modes which occurs
for many systems.
References
1. BUTKOVSKIY, A.G. : Theory of Optimal Control of

Distributed Parameter Systems, Elsevier, 1969.
2. CURTAIN, R.F. and A.J. PRITCHARD : Infinite Dimension-

al Systems Theory, Lecture Notes in Control and
Information Sciences, vol. 8, Springer-Verlag, 1978.
285
CONTROL OF DISTRIBUTED PARAMETER SYSTEMS - Problems
P. 14.1 If b(x) = /2/I sin TTX/£ in Example 1 of this
chapter, find the optimal control which minimizes
the performance index
[l 2 \tl 2
z
J(u) = J T (x,t_)dx
1 + j0 u (t)dt
J0 1 j
State a dual Kalman filtering problem.
P.14.2 Consider the controlled wave equation
—- (x,t) = — (x,t) + b(x)u(t)
y(t) = f 1 -c(x)w(x,t)dx
j
0
with w(0,t) = w(l,t) = 0
Find the transfer function if the initial state is

zero.
Calculate the semigroup and derive conditions for
the system to be approximately controllable and
initially observable.
Chapter 15
Decentralised control
Professor MADAN G. SINGH
synopsis
In this paper we give a brief overview of some of the re-

cent results in decentralised control. Both the stochastic
decentralised control as well as the deterministic decentra-
lised control literature are reviewed. Some of the decent-
ralised design techniques are also outlined.
15.1 Introduction
Although decentralised controllers have been designed and

used for controlling interconnected dynamical systems for
over two decades, the design was based on ad hoc methods.
For example, one usually assumed that the system comprised
weakly interacting subsystems so that it was plausible to
design the controllers independently for each subsystem.
This is still the basis for most industrial control systems
design. However, with the increased interest in optimisa-
tion techniques in the 1960s, there was an attempt to trans-
late the notions of optimal centralised controller design to
decentralised situations. It is only at this point that the
intrinsic difficulty of decentralised control became appar-
ent. For example, one of the best known and most useful
results in the theory of control for centralised systems is
the separation theorem 1 . This, broadly speaking, states
that for linear dynamical systems subject to Gaussian dis-
turbances, it is possible to design a controller which
minimises the expected value of a quadric cost function by
designing separately an optimal state estimator and an opti-
mal controller. However, the state estimator is finite
dimensional and the optimal controller is linear. These re-
sults fail in the case of decentralised control. To see this
let us examine the information structures associated with
decentralised control.
15.2 Non-classical Information Patterns
Consider a linear stochastic system described by:

287
N
x(t) = Ax(t) + I B.u.(t) + £(t) (15.1)
i=l 1 1
N
yjt) = cx(t) + I D.u. (t) + jKt) (15.2)
i=l
Here equation (15.1) is the state equation for a linear dyn-

amical system being controlled from N control stations,
equation (15.2) is the output equation, ^ ( t ) , ^(t) are ass-
umed to be independent Gaussian White Noise processes, y_(t)
in equation (15.2) corresponds to all the measured outputs
of the system. The partition of these measurements available
to a given controller is defined by the information pattern
of the problem.
For the decentralised control problem, the information

pattern is defined by the set of matrices
H =
where z± (t) = H±(t)v_(t) ; i = 1,2,...N
For the case of the classical information pattern, let us

define
B = Block Diagonal matrix [B.], i = 1,2,... N
D = Block Diagonal matrix [D.], i = 1,2,... N
and H = Block Diagonal matrix [H^], i = 1,2,... N
Then a classical information pattern arises when
H = I
where I is the identity matrix. This implies that all the

controllers have identical information.
Now, let us consider the quadratic performance
J = lim { \ [x?(t)Qx(t) + I u' (t) R.u. (t)dt >

rp-yoo A JQ i = 1
where Q is a positive semi-definite matrix and R = Block

Diag [R.] is a positive definite matrix.
288
A non-classioal information pattern arises if the controls

are restricted to be of the form
u±(t) =
where z_. = (^(T) : 0 < T < t)
i.e. the controls at the ith control station are chosen as

some function
functi y_. of the past and present information avail-
able to it.
Witsenhausen considered a seemingly simple example to show

that for non-classical information patterns, the optimal con-
trol law is not linear and that the optimal estimator may be
infinite dimensional. The reason for this is that due to the
lack of possibility of directly transferring information be-
tween the controllers so as to reduce the overall cost, the
information transfer is attempted through the dynamics of
the system. Thus each controller tries to use its control
action not only to exercise control but also to convey infor-
mation to the other controllers. The interaction between the
control and the signalling effect is necessarily non-linear!
A lot of work has been done in the last few years to seek
out special cases where linear solutions are possible. Ko
and Chu found that the so-called "partially nested informa-
tion structures" lead to decision rules which are linear in
information.
15.2.1 Partially nested information structures
Consider the observations
y = K K + D u
i i i
where D^ needs to satisfy the causality conditions:
If u. acts before u. then [D.]

jth component
Let us consider the information structure
289
k e E.
z± = 1 (15.3)
where E± = { k t D i k ¥ 0}
The information structure equation (15.3) implies that if u^

affects the observation y. (i.e. D.,/0) then u. knows
what u^ knows. This structure is a generalisation of per-
fect recall and defines a partial order of inclusion on what
the various decision makers know. This is called a partially
nested information structure. The reason why linear decision
rules are applicable in this case is because no signalling
is possible.
A lot of the work that has been done on design methods for
stochastic decentralised control ' ' has relied on seeking
fixed structure controllers which are usually constrained to
be linear. Parallel to this, a lot of work has been done on
the deterministic decentralised control problem. Next we
will examine the basis of this work,
15.3 Decentralised Stabilisation and Pole Placement
The deterministic decentralised control problem was

7 8
attacked by Wang and Davison and by Corfmat and Morse
Essentially, they derived the necessary conditions for the
existence of stable decentralised feedback controls.
A fundamental result in control theory, which was derived
q
by Wonham states that the poles of a closed loop system may
be arbitrarily placed using state feedback provided the sys-
tem is controllable. Thus controllable systems can be stab-
ilised using a centralised controller which places the poles
in the left half plane. Brasch and Pearson , amongst others,
have extended Wonham's result to the case of output feedback.
We will next see how this result generalises to the case
where the feedback matrix is constrained to be block diagonal.
The system can be described by:

290
V
x(t) = Ax(t) + I
(15.4)
= C
i ~(
where v is the number of control stations and i = 1 , — N

indexes the input and output variables of -the various con-
trollers.
The set of local feedback laws are assumed to be generated

by the following feedback controllers:
B i y i (t)
(15.5)
u ± (t) = Q i Z i (t) + K ^ U ) + v ± (t)
i = 1,2,... v
m.
wa
Va/ —La •• g J * «& ml~
where z. (t) e R " and v. (t) e R is an external input.
We can rewrite equation (15.5) in a more compact form as:
z(t) = Sz(t) + Rv_(t)

(15.6)
u(t) = Qz(t) + K£(t) + v(t)
where S/Q^R/K are appropriate block diagonal matrices.
If we apply the feedback law (15.6) to the system (15.4) ,

then the closed loop system can be represented by
f
x ] f A+BKC BQ }[ x(t) ] f B
. lx (15.7)
f
z RC S I I z(t) 1 10
where c = (C^ c\.. . C ^ ) T ; B = [ B ^ B 2 , ... B v ]
The deterministic decentralised control problem is to choose

local feedback control laws (15.6) such that the overall sys-
tem (15.7) is stable, i.e. all its poles are in the open left
complex plane.
291
In order to see under what conditions one can solve this

problem, we begin by defining the fixed polynomial of the
triple (C,A,B).
Definition
Let K be the set of block diagonal matrices
K = {K|K = block diag [K^... K J }

m.xp
where K± e R
Then the greatest common divisor (g.c.d.) of the set of

characteristic polynomials of (A+BKC) V K e K is called the
fixed polynomial of the triple (CrA,B) with respect to K and
it is denoted by i[> (A;C,A,B,K) ; i[> (A;C,A,B,K) = g.c.d. {det(I-
A-BKC)} K e K
Comment
It can be shown that there exists a unique monic polynomial

(i.e. a polynomial whose leading coefficient is unity) ip
such that ip divides all the polynomials generated by K .
Such a polynomial is called the greatest common divisor
(g.c.d.) of the set of polynomials generated by K .
Definition
For a given triple (C,A,B) and the given set K of block

diagonal matrices in equation (15.7), the set of fixed modes
of (CfA,B) w.r.t. K is defined as:
A(C,A,B,K) = f[ a (A+BKC)
K e K
where a(A+BKC) denotes the set of eigenvalues of (A+BKC),

i.e. of the matrix whose stability is of interest.
Alternatively, we can define A(C r A r B,K) as
A(C,A,B,K) = {A|A e 0 and y (A;C,A,B,K) = 0}

292
This definition can be interpreted as follows:
Since ip divides all polynomials, it has all the common

eigenvalues shared by these polynomials.
Remark
We next see why A (C,A,B,K) represents the fixed modes of

the system and how these fixed modes can be computed.
Let K e K and let k(i,r,s) denote the (r,sj.)th ele-

ment of K. (i=l,2,...v). Since K contains all block
diagonal matrices, it contains the null matrix. Hence
A(C,A,B,K) c a (A). Figure 15.1 illustrates this relation-
ship (15.8).
For each A^^ e a (A) , det ( A ^ - A - B K C ) is a polynomial in

k(i,r,s). If det(Ail-A-BKC) = 0, then X ± e . A(C,A,B,K).
On the other hand, if det (A^-A-BKC) f 0 then there exists
a K e K (at least one as shown in Figure 15.1) for which
det(A±I-A-BKC) f 0. Hence \± i A(c,A,B,K ) .
This is one way of determining the set of fixed modes of a

system, i.e. we compute all the eigenvalues of A and we
then apply the above test.
The reason for the name fixed modes is now clear, since
these modes do not depend on the values assumed by K c K
293
but depend on the structure of K (block diagonal in our

case).
Suppose we have found the eigenvalues of A that belong

toD A the
and these are X., i = 1/2,... k . Then we can write
the g.c.d. as
iHX,C,A,B,K) =
where \i^ is the multiplicity of X. amongst the eigen-

values of A.
With these preliminary definitions out of the way, we are

now in a position to examine the main result on decentral-
ised stabilisation .
Theorem 15.1
Consider the system (C,A,B). Let K be the set of block

diagonal matrices. Then a necessary and sufficient condition
for the existence of a set of local feedback control laws
(15.6) such that the closed loop system (15.7) is asymptot-
ically stable is
(C,A,B,K) a c"
where C denotes the open left half complex plane.
This result basically says that we can only stabilise by

decentralised control if the system has stable fixed modes
since the latter cannot be moved by decentralised feedback.
Thus the first step in carrying out any decentralised

feedback design is to check if the system has any unstable
fixed modes since, in that case, it would not be possible to
stabilise the system using decentralised feedback. An easy
way of checking this is given by Davison . First we notice
that since K. = 0 , i = l , 2 , . . . v is admissible, the fixed
modes of A are a subset of its eigenvalues. Second, it
can be shown that if K is randomly selected, then with
294
probability one, the fixed modes of A are the common

eigenvalues of A and (A+BKC). Standard algorithms exist
for computing the eigenvalues of fairly high order systems
so the calculation of the fixed modes should not pose any
difficulties.
Finally, we look at design techniques.
15.4 Design Techniques of Decentralised Control
In the case of the stochastic problem where an optimisation

framework is used, due to the difficulties caused by the non-
classical information patterns, all the design techniques
consider fixed structure controllers ' ' . Usually, this
structure comprises linear feedback. Here we will concen-
trate on the deterministic problem. We will begin by exam-
ining the optimal approach of Geromel and Bernussou.
12
15.4.1 The algorithm of Geromel and Bernussou
Consider the large scale system described by the state

equations
N
= A + B
*i * ij£j iHi
=
x, (o) = xio i 1/2,... N
where
the form
I
A r.
x^ c R "*", u^ e R . This system can be rewritten in
where A = {A ± ., i = 1,... N? j = 1,... N} e R n * n ;
B = block diag (B.,B0,... B.,) e R n x r ;
N N
I n^ r = I
we seek a linear static decentralised feedback controller of
295
the type
Hi = - k±x± ; i = 1,2,... N (15.9)
Let Q{S) be the set of all block diagonal gain matrices,

i.e.
fl(S = {K e R r x n | K = block diag(^,k 2 ,... k^) ;
1
k. e R }
Then the problem is to find a decentralised control of the

form given by equation (15.9) so as to minimise a quadratic
cost function, i.e.
min (x Qx + u Ru)dt
J
Keft(s) 0
s.t. x = Ax + Bu
u = - Kx
T
where it is assumed that Q = CC ^ 0 and R > 0 .
Let (j> (t) = exp [A-BK] be the transition matrix associated

with x = (A-BK)x . Then the cost function J can be re-
written as
f00 m m
J = x(o)T <T(t)(Q + KTRK)(j>(t)dt x(o)
0
We can write a new cost function by considering x(o) to be
a random variable which is uniformly distributed over an
n-dimensional sphere of unit radius, i.e.
min J(K) = tr P
KtR(S)
exp([A-BK] T t)(Q+K T BK)exp([A-BK] T t)dt

h
The gradient matrix 8J(K)/8K is given by the system of
equations:
296
H i £ ! = 2(KK-BTP)L
(A-BKTP+P(A-BK)+Q+KTRK = 0 (15.10)
(A-BK)L+L(A-BK)T+I = 0
for those K e fi(S) such that (A-BK) is a matrix with

eigenvalues in the left half plane.
Let us denote by D = block diag(D^,D 2 ,... D N ) e R n x r a

matrix which comprises those entries of 9J(K)/8K which
correspond to the unconstrained entries of K and the other
entries are zero. This is equivalent to projecting the grad-
ient matrix on the set ft(S), since D e ft(S). Then Geromel
and Bernussou propose the following algorithm:
Assume that the initial matrix K e A(S) where A(S) =

{K|K e n(S) and (A-BK) is stable}
Step 1: Using equations (15.10) determine the gradient

matrix 3J(K)/8K and the feasible direction matrix D =
block diag (D 1 # D 2 # ... D N ) .
Step 2: Test for convergence by checking if I{d }. < e

pq l
V p = 1,... r, q = 1,... n.; i = 1,... N; where
{d } is the (p,q)th entry of the matrix D. and e is a
small pre-chosen constant. If the convergence test is sat-
isfied, stop. Otherwise go to step 3.
Step 3: Set K = K - aD where the step size a ^ 0 must

be chosen such that J(K-aD) < J(K) . Go to step 1.
The properties of this algorithm are summarised in the

following lemma:
Lemma 15.1
If J(K) is a function of class C for every K e A(S),

then:
1) at each iteration there exists a step size a > 0 such

297
that J(K-aD) < J(K)
2) provided that (A,C) is completely observable and Q > 0

then for every a which is such that J(K-aD) < J(K)
< oo one has (K-aD) e A(S).
Remarks
1. The step size "a" can be determined using any conven-

tional unidimensional search technique. The authors used
an adaptive step size routine. Thus if £ denotes the
iteration index, then the step size obeys the following rule
= vai if j(K£ - a V ) < J(K)
and i(Kl - a V ) > 0
a = va otherwise
when T T > 1 , 0 < V < 1 and one starts from an arbitrary

initial value of a.
2. We note that it is not feasible to compute 9J(K)/3K

i = 1,... N, i.e. the partial variations of J(K) with
respect to K. directly. We recall from equations (15.10)
that 8J(K)/8K is a full matrix, although we only need the
block diagonal bits of it. It would be attractive if we
could calculate only the block diagonal elements. Unfor-
tunately, this is not feasible.
3. The method needs an initial stable gain matrix K

12
This method has been applied to a number of examples
Finally, we look at a different class of design techniques

where the interactions between the subsystems are modelled
in some way during an off-line study whilst on-line, no in-
formation is passed between the local controllers . Due
to lack of space, we will only look at one of the methods,
13 14
i.e. the model following method of Hassan and Singh '
298
15.4.2 The model following method 1 3 ' 1 4
The basic problem is to minimise the cost function
J
N ,
lfi±l R. }dt
s.t. (15.11)
where x. is the n. dimensional state vector of the ith

subsystem, u. is the mi dimensional control vector and
z.
—i
is the q.1 dimensional interconnection vector. Q.I anc
Ri are respectively positive semi-definite and positive
definite matrices and h^, B., C , L. . are constant mat-
rices of appropriate dimension. | I2
S
The approach requires the introduction of a crude inter-
action model of the form
which can be used to modify the optimisation problem to
N
min J = i I }dt
i-1
R.
s.t. A B (15.12)
i i Hi
where
-i
Then the control is given by

299
1
Hi = Ki *^! (15.13)
It should be recognised that there is no simple and infal-

lible way of choosing the crude model by picking out certain
entries from the A matrix. Another way of choosing A may
z
be to concentrate on the eigenvalues. However, in the i
multivariable situation, the relationship between the modes
and the states is not a clear one.
If we recall the basic notion of the model follower for
centralised control we start with a certain model and then
we minimise the errors between the plant output and the
model output. Thus, the choice of the model is largely gov-
erned by the response that we desire from the plant.
A simple way of specifying this response is in terms of
assignment of the poles of the model in the left half plane.
Thus, we could choose the matrices A to be diagonal with
z
negative elements. i
Let us rewrite the control of u. given in equation

(15.13) as
Then we construct a subsystem model which is forced by the

interconnections, i.e.
«
xi = Aixi + B ^ + C ^ (15.15)
where x. is the state vector of the subsystem model and

jz. is a "desired" interaction input which is consistent
with any a priori knowledge that we have about the system.
Also we substitute for u^ in equations (15.11) and
(15.15) after having replaced z^ by z^ in equations
(15.14)
x± = (A± - B.G il )x i + C ± z ± -
300
or
x± = A.x. + C.z. - E.G. 2 z. (15.16)
A = A
where i i ~ B i G il
x
_\ is the suboptimal state of subsystem i resulting from
—i
using z^. instead of _z. and
= A + C
~i i-i i-i " B i G i2-i (15.17)
Then in order to minimise the errors between subsystem i and

subsystem model, let
z. = z. - z. (15.18)
Subtracting equation (15.17) from equation (15.16) we

obtain
x± = iLx^ + Cjjz^ (15.19)
Write min J. = i (||x.||2 +||z.l|2)dt

1 2 j s
0 H
subject to the constraint (15.19), where H. and S. are

positive semi-definite and positive definite matrices res-
pectively.
The solution of this problem is given by
where K± is a solution of an appropriate Riccati equation.
Using equation (15.18) , we obtain
* * —1 T
z. = z. — S. C.K.x.
Thus, the dynamical equation for the subsystem will be

301
= (A±
x± = A.x. + ( C ^ B ^ ^ ) ^
ii = A zii
and we could represent these equations by Figure 15,2 •
Now further analysis can proceed with the block diagram

in Figure 15.2 .
First, no matter how Az is chosen, the errors will be

corrected to a large extent by the feedback channel. Of
course, the accuracy will depend partly on the numbers in
A . Here, z. produced by the minor loop I acts as a
reference input. It is independent of any controls
that could be applied. The same notion is given by (15.12)
where the left side of A is zero, and the lower part of
z
i
B is also zero. This means that no controls or
other states can effect z_. . For stabilising the minor
loop I, we should choose real negative diagonal elements
in A .If one does not choose a diagonal matrix, at
z
i
least one must ensure that all the eigenvalues of A
i
have negative real parts, and they match the required
responses.
This method has also been tested on a number of examples

including that of the energy cycle of a ship as well as for
13 14
river pollution control '
15.5 Concluding Remarks
In this chapter we have given a brief overview of the

recent research that has been done in the area of Decentra-
lised Control. Essentially, the work can be split into
three broad categories, i.e. stochastic decentralised
control, deterministic stabilisation and pole placement and
302
design techniques. We have outlined some of the more impor-

tant approaches in each of these areas. A more detailed
treatment of the whole field is given in the recent book 19
C:Z:
-B|
Fig. 15.2
References
JOSEPH, P. and J, TOU : "On linear control theory",

Trans. AIEE (Appl. and Industry), vol. 80, (1961),
193-196.
WITSENHAUSEN, S. : "A counter-example in stochastic
optimal control", SIAM J. of Control, vol. 6, (1968) 1.
HO, Y.C. and K.C. CHU : "Team decision theory and in-
formation structures in optimal control problems
Parts I and II", IEEE Trans. AC-17, (1972), 15-28.
SANDELL, N. and M. ATHANS : "Solution of some non-
classical LQG stochastic decision problems", IEEE Trans.
AC-18, (1974), 108-116.
303
5. AOKI, M. : "On decentralised linear stochastic control

problems with quadratic cost", IEEE Trans. AC18, (1973),
243-250.
6. CHONG, C. and M. ATHANS : "On the stochastic control
of linear systems with different information sets",
IEEE Trans AC16, (1971), 423-430.
7. VTANG, S. and E.J. DAVISON : "On the stabilisation of
decentralised control systems", IEEE Trans AC18, (1973),
473-478.
8. CORFMAT, J. and A. MORSE : "Control of linear systems
through specified input channels", FIAM J. Control,
(1976), 14.
9. WONHAM, N. : "On pole assignment in multi-input
controllable linear systems", IEEE Trans AC12, (1967),
660-665.
10. BRASCH, F. and PFARSON, J. : "Pole placement using
dynamic compensators", IEEE Trans AC5, (1970), 34-43.
11. DAVISON, E.J. : "Decentralised stabilisation and
regulation" in "Directions in large scale systems",
Y.C. Ho and S. Mitter, Eds., New York Plennum, 1976.
12. GEROMEL, J. and J. BERNUSSOU : "An algorithm for
optimal decentralised regulation of linear-quadratic
interconnected systems", Automatica 15, (1979), 489-491.
13. HASSAN, M.F. and M.G. Singh : "A decentralised con-
troller with on-line interaction trajectory improvement",
Proc. IEE Series D, Vol. 127, 3, (1980) 142-148.
14. CHEN YULIU and M.G, SINGH : "Certain practical con-
siderations in the model following method of decen-
tralised control", Proc. IEE, D, vol. 128, (1981).
15. HASSAN, M.F. and M.G. SINGH : "A hierarchical struc-
ture for computing near optimal decentralised control",
IEEE Trans. SMC 8, 7, (1978), 575-579.
16. HASSAN, M.F. and M.G. SINGH : "Robust decentralised
controller for linear interconnected dynamical systems",
Proc. IEE, Vol. 125, 5, (1978), 429-432.
17. HASSAN, M.F., M.G. SINGH and A. TITLI : "Near optimal
decentralised control with a fixed degree of stability",
Automatica 15, (1979), 483-488.
18. ARMENTANO, V.A. and M.G. SINGH : "A new approach to
the decentralised controller initialisation problem",
Proc. 8th IFAC World Congress, Kyoto, 1981.
19. SINGH, M.G. : "Decentralised Control", North Holland,
Ams terdam, 1981.
20. MARKLAND, C. : "Optimal model following control system
synthesis technique", Proc. IEE, vol. 117,(1980>,3.
304
Chapter 1: Solutions
S.I.I H = 1 -2 2
0 -2 0
0 -2 -1
1 -1 -2
1 -2 2 2
1 0 0 1
2 1 0 2
0 1 -1 0
r T -T
-1 0
0 -1
0 0 "
1 0
0 0
0 1
305
S.I. 2 1/ a ""• — tr (A) = 4,

3
R2 = A + 41 2 1 1 0 "
' a
2 = - ~tr(AR 2 ) 9,
1 4 0 -1
1 0 3 -1
0 1 1 3 _
AR2+9I 3 2 1 -2 " a = - ~tr(AR 1 )

' l
2 7 -2 -3
1 -2 4 -2
-2
AR1+1OI 1 1 0 - 1 = - £tr(AR0) = 3
-1 5 - 3 -2
0 - 3 3 0
- 1 2 0 1
4s2+8s+3 - (2s3+7s2+7s+2)"
G(s) =
s 4 +4s 3 +9s 2 +10s+3 2s 3 +5s 2 +3s 4s +8s+3
306
S.3.1 For the nominal controller parameters the input and

output return-ratio matrices are both equal to G(s)
which can be diagonalized at all frequencies using a
constant matrix W as follows:
WG(s)W -1
2
s+2
t 8
where W =
6 7
1 -8
-1
and W
-6 •7
The characteristic gain loci are therefore given

explicitly as
and
The columns of W are a candidate set of right-hand

eigenvectors for G(s) , and the rows of W""1 are a
candidate set of left-hand eigenvectors. By normal-
izing the appropriate vectors to unity it follows
that
7
1
/85 -2
6 /ITT
1 7 6
/ITT -8 /85- -7
and therefore that the sensitivity indices are given

(at all frequencies) by
307
s
0.0102
/85 /fl3
s
—=^ 0.0102
/85
The very small indices indicate that the characteris-

tic gain loci are very sensitive to small perturba-
tions either before or after the plant, and indeed it
can be shown that there are small perturbations of
k.^ and k 2 for which the closed-loop system becomes
unstable. Note that for k, = k2 = k the system has
infinite gain margins with respect to k.
S.3.2 The small gain theorem tells us that the feedback

design remains stable if the perturbation AG(s) is
stable and if the maximum principal gain of AG(jw)
is less than the reciprocal of the maximum principal
gain of the output relative stability matrix R (joj) .
At low frequencies this implies that the maximum

principal gain of AG(ja>) is less than 1 for stabil-
ity to be maintained. This limit is reduced to about
0.9 as w increases in the medium frequency range;
but at higher frequencies, as the gains roll off,
clearly much larger perturbations can be accommodated.
The conservativeness of the small gain theorem at low

frequencies can be reduced by considering the small
phase theorem. For example, up to about Irs the
condition number of R (jw) is approximately 1,
i|> (OJ) is approximately 0, and so stability is
maintained (in this frequency range) for large per-
turbation condition numbers providing the minimum
principal phase (maximum phase lag) of the perturba-
tion is greater than say -140°. Note that insofar
as stability is concerned we restrict only the con-
dition number of the perturbation and therefore in
theory the perturbation can have infinitely large
gains.
308
A satisfactory characterization of robustness is obtained

by combining the small gain and phase theorems as in Coroll-
ary 2, where, for example, one might take w, to be 10rs~ .
309
S. 5.1 We have
|TGI|H|/|G|
(s+D+l s+1 0
0 s-1
(s-1)(s+1)
so the closed-loop system is asymptotically stable.
G can arise from

S+1 0 f 1 0
- k *m* ***> .!» J - VÂH
0 s-1 ' 0 1
P o (8) =
-1 0 ' 0 0
0 -1 ' 0 0
which has least order n = 2 and has |T | = (s-1)(s+1).

But G(s) actually arises from a 3rd order system with
|T G | = (s-1)(s+1) . Hence s = -1 a decoupling zero of the
system giving rise to G.
The INA degenerates to the two inverse Nyquist plots, of
gi;L(s) =
The first gives N 0, = 0. The second gives
g2 ' h2 * Also p Q = 1, so
ZN__. - IN = 1 = Pr
hi
whence the system is asymptotically stable.

310
S.5.2
(i) For a conventional feedback system with unity feedback.
Y_(s) H(s) r(s)
where H(s) [I+G(s)K(s) ]-1 G(s)K(s)
i.e. H(s) is a complicated function of G(s) and K(s).

However, if the system is functionally controllable; i.e.
G~ (s) exists; then
H" 1 (s) = K" 1 (s)G" 1 (s) + I
which is simpler to deal with. Thus, Rosenbrock's INA

design method is carried out using the inverse system repre-
sentations.
iir ° 1 -1
G(s) =
1 1
Therefore,
1
1
1
f[-1 M
x)
s+1
0
0
s+2
(s+1) (s+2)
1
2
-(s+1) (s+2)
with corresponding INA for s = jw as shown.

311
0.5 1.0
-0.5
y
22
1.0
INA of given system
An examination of this inverse Nyquist array reveals that
for row diagonal dominance - row 1 is not dominant
row 2 is dominant
and for column diagonal dominance - col 1 is not dominant
col 2 is not dominant
although in the latter case, we are on the margin between
dominance and non-dominance.
In either case, dominance can be determined graphically by
superimposing circles of radius
d ± (s) = j (s) j, i=l,2
on each diagonal element for row dominance, and circles of

radius,
dj(s) = |gi;j(s) | i 7* j , j = 1,2
on each diagonal element for column dominance.
(iii) For this particular system, it is possible to deter-

mine a wholly real compensator K which will diagonalise
the resulting system Q(s) = KG(s).
The required compensator can be readily determined by

inspecting G(s) as
312
-1
K =
1
which results in
s+1 0
Q(s)
0 s+2
1
1
Thus, K = K"1
2 -1 1
(iv) If we now introduce integral action into the two

control loops for the system
1 0
s+1
Q(s) = 1
0 i+2
we can sketch the behaviour of the two decoupled responses

for varying proportional gains as shown below
i jUJ JUJ
-2
Root-locus diagrams for loop tuning

Thus, the final loop turning can be set by the appropriate
choice of k1 and k2 to give the desired time responses
for the two outputs y. and y2*
313
S.7.1 H = ;^x 2 + X2(-x
Canonical equations
h - - \
{
2 =
Then X2 = A sin(t-a) , A >0
Pontryagin's principle yields u = -sgn X 2 = -sgn(sin(t-a))
Thus, control function u(t) alternates between +1 and -1.
For u = 1,
x = X
l 2 ' X2 =
~xl+ 1
dx X
l 2 2 2 2
i.e. gj- = — f whence (xfl) + x2 = R
i.e. Circles with centre (1, 0 ) , radius R.
x
Also, i~1 = ~ R cos(t+Y) and x 2 . = R sin(t+y)
X^ +k22 = R2
Thus, phase point moves clockwise with uniform velocity R.

One revolution therefore takes a time of 2TT.
Similarly for u = -1
x = X X =
l 2 ' 2 "xl ~ 1
2 2 2
Phase trajectories are (x^l) + x 2 = R : Circles centre
(-1, 0)
If A is the point (-1,2) then choose u = -1 and travel

on circle, centre (-1, 0 ) , radius 2 until intersection of
circle, centre (1,0), radius 1, at point B in fourth quadrant.
Then choose u = 1 from B to origin. Time from A to B clearly
less than TT . Time from B to origin less than TT/2. Total
314
time therefore less than 3TT/2.
S.7.2 For scalar control : (-l)q ~ H(2q) > 0 (order q)

dU • U —
For vector control : (-l)q |-v H ( , 2q) ; |- By[p) = 0

dU U. oil u.
(p odd).
s n X + U
H = X, cos x 3 + A 2 i 3 *3
X, = - H X i = 0, X2 = -H^ = 0
X3 = -H = X, sin x.. - X2 cos x^ ,
X 3 (t f )
If X3 > 0 then u = -1
x3 < 0 u = +1
X3 = 0 tan x 3 = A ^ / X ^ (singular arc)
x
Suppose u = +1. Then x- = t, o = l ~ c o s t, x, = sin t.
Suppose X 3 = 0. Then X3 = 0 so tan x 3 = X 2 /X 1 (*)
X 1 =a / X 2 =b, a and b constants
Differentiating (*) : u sec x 3 = 0 so u = 0
Then x 3 = A, x 2 = t sin A + B, x = t cos A + C.
With a junction at t = J , we have A = 4-,B=1--:(1+^-)

/2
C = ±(i - 1) . Also f tan ~ = 1 = -r=- = 5-
q q A a
/2 l
a = b.
H(tJ = - ?2- , G
f 3t f
Then a = b = --.
/2
315
On singular subarc
x3 = \ , x2 = i +£ (t - i - ?-)
x ( t
i - ^
x = anc X 1
l ^f ^ ^ ^ 2 ^f ^ ~ satisfied when
All c o n s t r a i n t s s a t i s f i e d with u = 1, t < -j and
u = C. J < t < 1 + J .
u 3 u 3 1 3 2 3
H = (Xx cos x^ + X2 sin
H = - i ( - + i ) = -1 < 0
u /2 /2 /2
316

S.10.1 We want to shift the two complex eigenvalues
{*L} = {+jf-j> to {y±] = {-2,-2} and to leave * 3 = -2
unaltered; here, q = 2. If we arbitrarily select the
elements of f as
f =
then
bf = Bf =
and , A
0 3-2
1 0 1
1 2 - 2
which has rank 3 = n. Therefore, this choice of f_ results

in [A, b^] being completely controllable.
Before using the relationship given in equation (10.8) to

determine the weighting factors 6., first determine the
scalars p. as given by equation (10.9);
= <v
Pi -i' bf > = -4 - j3
P2 = <2 2 ' bf> = -4 + j3
Now, using (10.8), we obtain
_ -.5(3 + j'
" (3 - J4)
= -0.14 + jO.48
Similarly, = -0.14 - jO.48
Thus, the required measurement vector m is given by

317
m (
"1"4 0#4 3
'6)
and the required state-feedback matrix is then
-1.4 0.4 3.6

K = f m =
-1.4 0.4 3.6
This results in
-2 1 2 ]
A-BK 0.4 -2.4 -1.6
-0.6 -0.4 -1.6
which has the desired eigenvalues ( Y - ) = {-2, - 2 , -2},
S.10.2 Using Young's algorithm on the system of Problem 1,

we obtain
A O (S) IsI-Al = s 3 + 2s 2 + s + 2
and since the desired closed-loop system eigenvalues are
{-2, - 2 , -2}
we have A (S) = + 6s 2 + 12s + 8
Therefore, the difference polynomial is
6(s) = 4s 2 + 11s + 6
Now, for £ = (1 I) 11 we know that [A, b f ] is completely

controllable with $c as in Problem 1. The coefficient
matrix X is
' 1 0 0
X = 2 1 0
1 2 1
and using the relationship given in equation (10.13), we

determine m as
318
m X" 1 6 (-1.4 0.4 3.6)
which is the same result as that achieved using the spectral

approach. Thus, the resulting closed-loop system A-matrix
(A-Bf_ m*) has eigenvalues {y±} = {-2, - 2 , -2}.
=
S.10.3 For (A-BK) to have eigenvalues (YJ_) {~4/ - 4 , -4}
we proceed as follows. The matrix BK has the parametric
form
0 0 0
BK = k k
K k
l 12 *
k.2l
, k
22 k
23
So, by choosing k ^ = 10/9, k ?l = k 22 = ®' t n e m a t r i x

(A-BK) becomes block lower triangular, and its eigenvalues
are determined by the roots of
-80/9+k lx s-26/3+k 12
Therefore, if we express A d (s) as
A d (s) = (s2 + 8s + 16)(s+4)
then the desired result can be achieved with
224/9, k 1 2 = 50/3, 13/3

^23
i.e.
224/9 50/3 10/9
K =
0 0 13/3
which has full rank 2 = m.
The required state-feedback matrix K in the original

basis is then
K Q"1 K T
319
1 4/3 0 0 1/3
K 1 4/3 5/3
0 1 0 1 -2/3
50/3 158/9 1058/27
0 13/3 -26/9
which has rank 2, as expected.
S.10.4 Since m+£-l < n, a compensator of degree r = 1 is

required for arbitrary pole assignment. However, for any
desired pole configuration, the minimum degree of compensat-
or required is given by the smallest r such that
For {Yi} = (-1, - 2 , - 4 } ,
A d (s) s 3 + 7s 2 + 14s + 8
and since A (s) = s + 6s + 7s + 2
we have (for r = 0)
^ = ( 1 7 6 )
Here, 1 0
6 1
4 2
which has rank = 2, and since {X Q , 6^} also has rank = 2

we have that r =0 is both necessary and sufficient for the
desired closed-loop pole set {yi> = (-1/ - 2 , -4}.
Now, a gînverse of XQ is
320
1 0 0
-6 1 0
therefore, p = X x 6
o —o
Thus, the required feedback is F = [1 1 ] .
Now, for {y i } = {-1, - 2 , -3}
A d (s) = s 3 + 6s 2 + 11s + 6
which yields 6 = ( 0 4 4)
However, for this closed-loop pole specification

p{X} j- P{X # and we must increase r from 0 to 1. Since
the resulting closed-loop system has 4 poles, we consider
iy±] = (-1, - 2 , - 3 , -4} with
A d (s) = s 4 + 10s 3 + 35s 2 + 50s + 24
r 4 3 2
Then, since A (s).s = s + 6s + 7s + 2s
o
we obtain (4 28 48 24)
1 0 0 0
6 1 1 0
and
4 6 2 1
0 4 0 2
A gînverse of is
1 0 0 0
0 0 0 0
3 -1 h -i
-9 2 i
-7 2 -1 1
and this results in

•2-1
321
Thus, the required coefficient vector is determined as
Y
• l • 4 '
9
10 0
e
ll 2
X
El = «i
1
|
i
1
2
CD
to
O
4
CD
to
r-
i.e. m t (s) = [2 2(s+2)]
This latter result illustrates the case where = I,
S.10.5 Since m+£-l = 3 = n, a full rank constant output-

feedback compensator F can be designed in two stages such
that the resulting closed-loop system has poles
{y±} = {-1, " 2 , - 5 } .
(1) For f c 1 } = (0 l)t, the pair [A, Bf^ 1 '] is com-

pletely controllable, and
Here, Bf c(1) influences all 3 modes of the system; i.e.

n, = 3; however, there are only 2 linearly independent out-
puts; i.e. JU = 2; therefore, q, = min{n,,£,} = 2.
(1) i l l
So, we determine f to place two poles at s = - 1 , - 2 ,
say, which results in the equations
(s2+3s+2) = s 3 + {f 1 (s 2 +s+l) (s2+s)
Solving for f and ^ we obtain
0 -1 f 3
l
1 -3 f = 2
2
0 -2 e
l
0
322
which yields e1
and f_ = (-6 -7)
0 0
i.e. f (D f (D
—c —o -6 -7
Now, the resulting closed-loop system is
s2+s-6 1
(s+1)(s+2)(s-3)
s +s+6
(2) To hold one of these poles, say s = - 1 , at this

value during the second stage of the design, we let
H ( 1 ) (s) = rU )
(s)/d1(s) and solve
fi 2) r (1) (-D - o t
i.e.
[1 1]
Thus, N(2)(s) .
= [2s(s+l) (s+1)]
Now, to place the two remaining poles at s = - 1 , -5, we

(2)
determine f by solving the equations
2 0 f
l 8
0 1 f
2
i.e.
—c 16
(2)f(2)
4 4
=
Thus, F2 = f
z —c —u 16 16
323
Now , 4 4
10 9
which has rank 2 and generates the desired closed-loop system

poles {y±} = {-1, - 3 , -5}.
324

S.I1.1 From the example in Section 11.2, the fundamental
term in the output of the relay is
- sin(wt)
IT
the describing function is
N(U) = iL
.*. the graph of -1/N(U) lies along the negative real
axis, tending to -« as U -* °°.
The Nyquist plot of g(s) is obtained from
joj (4+5 jco-o) )
which crosses the real axis at -1/20 when a> = 2.
the graphs intersect at the point corresponding to

N(U) = 20, i.e. U = 1/5TT.
Also, as U increases, the graph of -1/N(U) passes
out of the region encircled by the Nyquist plot.
.'. a stable limit cycle is predicted, with amplitude
U = 1/5TT
and p e r i o d
2TT
325
S.I1.2 The time-derivative of the Lyapunov function along

the trajectories of the system is
V =
V £ 0 when x. <_ 1 . Also, in the interior of this
region, V can only vanish when x2 = 0 .*. V does
not vanish identically along any trajectory.
any contour of V lying in the region x 1 <_ 1
encloses a domain of attraction.
the largest domain of attraction obtainable with this
Lyapunov function is
326

S.14.1 By examining the form of the functional J(u) in
eqn. (14.13) we see
W = 0 , G = I, R = l
Also, since b(x) = J~lT/i sin TTX/£ , b. = 1 and b i = 0

i >1
Substituting in eqn. (14.15) yields
p
mn ( t l> = X m = n
= 0 m f n
Clearly pmn = 0 m = n satisfies these equations.

Then if n > 1
= 0
So
P n n (t) = exp(-2-rr2an2/£) (t± - t)
For n = 1
Thus C£/2ir2a\log ( —ii ) = t + constant

P
11+27T a/l
or 2
2
at
— = Ae _
But p l x (tx) = 1, so
327
-t)
The optimal control is
U (t) = -p (t) <(J)_ ( • ) , Z ( " , t) >
f* ,
= -p (t) /2/£ sin TTX/A z(x,t)dx
Since A = A, using equations (14.17) and setting
V=l, c = b , P = I
we obtain the following dual filtering problem,
9t
•V
f
y(t) = I /2/fc sin TTX/£ T(x,s)dxds + v(t)
T(xf0) = TQ(x)
where the covariance of v(t) and T Q ( S ) is 1 and I res-

pectively and the Kalman-Bucy filter is
ft
T ( x£ ,/ t ) =
= I /2/£ e P n ^ , - 3 ) s i n TTX/£ dp ( S )
in
where
f ]
p (t) = y(t) - J2/1 sin TTX/£ z(x,s)dxds
f
328
S.14.2 Taking Laplace transforms we have
2-
s 2 w(x,s) = i~ (x,s) + b(x)u(s) (14.18)
y(s) = c(x)w(x,s)dx
w(O,s) = w(l,s) = 0 (14.19)
Let us consider the function
w(x,s) = ~(esx(a + f e" s p b(p)dp)

Z j
0
+ e" sx (3 - ( e s p b(p)dp) u(s)
Then
s 2 w(x f s) + b(x)5(s)
So this function will satisfy (14.18), (14.19) if

w(0,s) = w(l,s) = 0 and this will be the case if
(a+3 - I e s p b(p)dp) = 0
(a ++
(es(a
J
[ e" s p b(p)dp) + e" s 0) = 0
0
so that
l-e zs
f
J
0
(e sp + e- sp )b(p)d P
J
1-e " 0
This yields the transfer function
i [ c(x) (e sx (a + [ e " s p b(p)dp)
+ e" s x (e - [ e s p b(p)dp) dx
•'x
In order to find the semigroup we have to solve the
329
inhomogeneous equation, but first we write the equation in

abstract form
z = Az + bu
y = <c,z>
z (x,t) = w(x,t) ; zo(x,t) = |~(x,t

b(x) = 0
b(x)
c(x) = (c(x),0)
Az =
and dx
D(A) {z e H J ( O , 1 ) , Z 2 e H 2 (0,l) A H*(0,l)}
In order to solve the inhomogeneous equation we will assume

the solution separates so that
w(x,t) = T(t)X(x)
d 2 X(x)
then (t) =
T(t)" d t 2 X(x)
dx 2
So that
T(t) = A sin At + B cos At
X(x) = C sin Ax + D cos Ax
If the boundary conditions are to be satisfied D = 0 and

sin A = 0. So A = mr, and the complete solution has the
form
w(x,t) = I (A sin rnrt + B cos mrt)sin mrx
n=l
If w(x,0) = w (x) and ~ ( x , 0 ) = w. (x) , we find

O oX, 1
<wQ (x) = /2 sin mrx
mrA = /2
n
330
So
w(x,t) = cos n7Tt > sin
n=l n
But the full state is z(x,t) = w(x,t) so the semi-
§?(*,t>
group is given by
T z = T (<w o , V cos
t o t n 1
sin nirtH
(-mr t+<.w 1 f4>R>
c o s mrt)<|)
From (14.12) we see that the system is initially observable

if
<w (
(<w ,<j> > cos n-rrt + — 1/ f)n> s:i n
- = 0
for all t e [0,t ] implies w = w = 0. Clearly this will

n =
<W ( ) > = <w >
be the case if /• O/ J i 1/2,....
If t = 2 {sin n-rrt, cos mrt} form a basis for L (0,2) so
in this case the system will be initially observable if
c f 0 for all n = 1,2,....
From (14.11) we see that the system will be approximately

controllable if
<w ( ) > s i n n7rt

mr o/ l n + <Ww(() > c o s nirt)b = 0
n=l
f o r t e [0,t ] implies w = w, = 0. Again this will be the
case for = 2 if b f 0 for all n = 1,2,....
Index
Accessary problem, 128 First integral, 127,132

Armi jo, 154 Fixed mode, 291,292
Fourier series, 203
Batch reactor, 268 Frequency response, 109,111,
Bilinear transformation, 31 117,119
Bode plot, 59, 230
Bolza, 126 Gain
Butterworth patterns, 33 parameter, 263
space, 85
Catalytic reactor, 64,67 Gershgorin
Characteristic band, 92
frequency, 24,36 circle, 96,93
gain, 23,48,56 theorem, 48,49,102,
loci, 64,67,70-74,79, 103,167
80,115,116
Chebyshev, 163,165 Hamiltonian, 131,133,134,141
Circle criterion, 218 Harmonic balance, 203
Compensation, 228,231 Hilbert space, 275,276
Compensator, 64,71,79,178, Hyperstability, 218
259-262,268
Conformal mapping, 25 Information
Controllability, 3,278 pattern, 287,288
index, 5,6 structure, 289
Cost functional, 129 Integrity, 122,148,247,269
Inverse Nyquist, 83,91,95
Decomposition, 161,181
Describing function, 203,206 Jacobson condition, 136
Diagonal dominance, 83,91,160
Disturbance, 2 48 Kalman filter, 188,282
Drum boiler, 107
Dynamic programming, 137 LQG
design, 58
Equation, technique, 64
accessory, 128 Laurent expansion, 12
Euler-Lagrange, 127,130,132 Legendre-Clebsch, 134,135
Hamilton-Jacobi, 139,141 Limit cycle, 202,204
Pontryagin, 131 Low-pass filter, 208
Riccati, 141,142,282,283 Luenberger observer, 188
Equivalent error, 254 Lyapunov, 150,202,214,216
Markov parameter, 33,34
Feasible direction, 153,168 Matrix
Feedback coefficients, 142 controller, 84
332
dyadic, 177 Stability, 247

feedback, 84 Strong variation, 127
gain, 295,297 System
Jordan, 15,50,51 discrete, 2
normal, 53 dual, 4
precontroller, 118 interconnected, 220
relative stability, 53,54 inverse, 14
return difference, 22,118, invertible, 251,252
147,149 optimal feedback, 37
return ratio, 52 System performance, 77
transfer function, 9,14,20,
70,84,109,160 Three term controller, 2 62
McMillan form, 87 Time lag, 250
Transient behaviour, 2 47
Notch filter, 234 Transmission zeros, 14,38
Nyquist diagram, 18,19,24,32, 253-255
34,46,70,77,90,110,114,148, Transversality condition,127,
160,212,230 130,131
Observability, 3,2 81 Valentine technique, 129

Operator, 30,275 Weierstrass E-function, 127,
Optimality principle, 138 132
Ostrowski band, 94 Weierstrass-Erdmann, 127,132
Outer approximation, 155 VJeiner process, 283
Output detectable, 2 52
Zero, 27,30,80
Parameter optimization, 268 Zero pencil, 28,34
Partial error, 254
Performance index, 37,129,268
Phase modifier, 57
Polar decomposition, 55
Pole, 35,80,252,257,259
Polynomial, 291
Pontryagin principle, 132,133,
140
Popov criterion, 219
Principal gain, 56
Principal phase, 56
Realization, 11,13
Reference input, 248,255
Regulation, 247
Regulator, 2 63
Reliability, 247
Robustness, 247
Root locus, 21,25,32,34,36,196
Schwartz inequality, 279
Second variation, 128,133-135
Semigroup, 275
Sensitivity index, 49,60
Servomechanism, 246,248
Singular
control, 133
value, 56,147,150
value decomposition, 55
Spectrum, 20
Design of
MODERN CONTROL SYSTEMS
About the book The Editors
The IEE Control Engineering Series of David Bell obtained his B.Sc. in
which this book is a member contains a Mathematics (1955) from the University of
volume (number 9) devoted to the modern Manchester and his M.Sc, Ph.D. degrees
approaches to control system design. The (1962,1970) in Mathematics from the
present text may be considered a University of London. From 1955 to 1961
companion volume which presents new he worked as a research mathematician for
the aircraft firm A. V. Roe and Co. Ltd.,
developments in analysis and design followed by a period as lecturer in
techniques for control systems. Included Mathematics at the Bristol College of
are the exciting new results for feedback Science and Technology (1961-1964), and
systems using complex variable methods the University of Bath (1964-1967). His
as developed by Professor A. G. J. present joint appointment between the
Macfarlane and his coworkers at the Department of Mathematics and the
University of Cambridge, the important Control Systems Centre at UMIST was
concept of robustness in controller design taken up as a lecturer in 1967 and senior
reviewed by Professor E. J. Davison of the lecturer in 1973.
University of Toronto, and the increasingly
important topic of decentralized control for Peter Cook obtained his B.A. in
large scale systems as seen by professor Mathematics (1962) from the University of
M. G. Singh of the Control Systems Cambridge and his Ph.D. (1966) in
Centre, UMIST. These and many other Theoretical Physics from the University of
Manchester. He was then appointed
contributions illustrate the great activity and
lecturer in Applied Mathematics (1968) at
rapid progress which has taken place in the University of Kent. From 1971 to 1974
the subject over the past few years. Only he was a Senior Research Fellow in the
by bringing these contributions together Department of Mathematics at the
under one cover can the practising Loughborough University of Technology.
engineer in industry and indeed the Since 1974 he has been a lecturer in
engineer in university or polytechnic keep Control Engineering at the Control Systems
fully informed on the “state of the art” on a Centre, UMIST.
number of different fronts. Application of
the theoretical developments and practical Neil Munro obtained his B.Sc. in
aspects of the subject are not forgotten; Electrical Engineering (1962) from
analysis and design of a nuclear boiler and Strathclyde University and his M.Sc. (1967)
some direct digital control system design and Ph.D. (1969) in Control Engineering
procedures are but two topics discussed in from the University of Manchester. He was
the present book. Several of the chapters then appointed as a lecturer (1969), senior
lecturer (1973) and Professor of Applied
are followed by problems on the subject
Control Engineering (1979) in the Control
matter and worked solutions to most of Systems Centre, UMIST. He has also spent
these problems are given at the end of the some 10 years in industry as an electronics
book. This aspect will find favour with technician, development engineer and
many readers since such contributions are design engineer working for Barr and
often a great help in the understanding of Stroud Ltd., Ferranti, Parkinson-Cowan
the subject matter. Measurement and ICL.
Peter Peregrinus Ltd.,

Southgate House, Stevenage, Herts., SG1 1HQ, United Kingdom.
ISBN 0 906048 74 5
Printed in the United Kingdom.

De Silva Clarence W Sensors and Actuators Control Systems Instrumentation

Uploaded by

Copyright:

Available Formats

De Silva Clarence W Sensors and Actuators Control Systems Instrumentation

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

De Silva Clarence W Sensors and Actuators Control Systems Instrumentation

Uploaded by

Copyright:

Available Formats

Control engineering series 20

Peter Peregrinus Ltd. on behfalf of the Institution of Electrical Engineers

Series Editors: Prof. D. P. Atherton

D J Bell • PA Cook*T\J Munro

© 1982 Peter Peregrinus Ltd.

All rights reserved. No part of this publication may be reproduced, stored in

British Library Cataloguing in Publication Data.

Printed in England by Antony Rowe Ltd.

2 Complex variable methods in feedback

Optimal control 126

Over the last seven years the Science and Engineering

It is inevitable and indeed desirable that a subject such

The organization of a vacation school and at the same time

A brief review is given of the use of state-space methods

We shall be concerned with linear stationary dynamical

where the vectors u (inputs), x (states) and y (outputs)

Solving equation (1.1) with initial conditions

and substituting into equation (1.2), we obtain

which, after Laplace transformation, gives

is the transfer function matrix.

Clearly, G(s) is unaltered by coordinate transformation

A -> R" 1 AR, B -* R" 1 B, C -> CR, D -> D

where R is a nonsingular constant matrix. In particular,

and G(s) has the modal decomposition

Although we are specifically considering continuous-time

x(t+l) = Ax(t) + Bu(t) (1.8)

y(t) = Cx(t) + Du(t) (1.9)

and we define the z-transforms by

x(z) = I x(t)z" t , etc.,

y(z) = C(zI-A)~1zx(0) + G(z)u(z)

where G(z) is given by substitution of z for s in equa-

y(t) = Du(t) + CBu(t-l) + ... + CAt""1Bu(0) + CA t x(0)

1.2 Controllability and Observability

A system is said to be controllable if, for any initial

for some specified t f 0, we get

which can be solved for U ( T ) , for any given XQ, if and

are linearly independent. Since the matrix exponential can

This is also a necessary and sufficient condition for any

state to be reachable from any other, and it plays the same

Observability means that the state of the system at a given

are linearly independent. Hence, a necessary and sufficient

with the superscript T denoting transposition.

Again, the same condition applies to a discrete-time system

The algebraic conditions (1.10) and (1.12) are clearly very

Both controllability and observability are evidently un-

where the crosses denote non-vanishing blocks and the sub-

G(S) = C C (SI-A C O )" 1 B O

1.3 Standard Forms

Let us suppose that the system described by (1.1) is con-

Since the matrix $ in (1.11) has full rank, we can con-

Since they are determined from linear dependence relations,

is thus invariant under all these transformations. It is the

and is sometimes called the "controllability index".

We now perform the coordinate transformation

with H given by (1.15) .

The results have the following form: H~" B is a matrix

where the blocks have the structure

and the entries in the final columns are the coefficients in

It follows from (1.16) that the number of nonvanishing co-

where r is an integer satisfying