Lecture Notes in

Control and
Information Sciences
Edited by A.V. Balakrishnan and M.Thoma

12

lan Postlethwaite
Alistair G. J. MacFarlane

A Complex Variable Approach

to the Analysis of Linear
Multivariable Feedback Systems

Springer-Verlag
Berlin Heidelberg New York 1979
Series Editors
A.V. Balakrishnan M. Thoma

Advisory Board
A. G. J. MacFarlane H. Kwakernaak Ya. Z. Tsypkin

Authors
Dr. I. Postlethwaite,
Research Fellow, Trinity Hall, Cambridge, and
SRC Postdoctoral Research Fellow,
Engineering Department, University of Cambridge.

Professor A. G..I. MacFarlane,

Engineering Department, University of Cambridge,
Control and Management Systems Division,
Mill Laqe,
Cambridge C B 2 1RX.

ISBN 3-540-09340-0 Springer-Verlag Berlin Heidelberg NewYork

ISBN 0-387-09340-0 Springer-Verlag NewYork Heidelberg Berlin

This work is subject to copyright. All rights are reserved, whether the whole
or part of the material is concerned, specifically those of translation, re-
printing, re-use of illustrations, broadcasting, reproduction by photocopying
machine or similar means, and storage in data banks.
Under 54 of the German Copyright Law where copies are made for other
than private use, a fee is payable to the publisher, the amount of the fee to
be determined by agreement with the publisher.
by Springer-VerlagBerlin Heidelberg 1979
Printed in Germany
Printing and binding: Beltz Offsetdruck, Hemsbach/Bergstr.
2061/3020-543210
 Contents

CHAPTER 1 Introducti0n 1
References 6

CHAPTER 2 Preliminaries 8
2.1 System description 8
2.2 Feedback configuration ii
2.3 Stability 13
2.3-1 Free systems 13
2.3-2 Forced systems 16
2.4 Relationship between open- and closed-loop
characteristic polynomials for the general
feedback configuration 17
References 20

CHAPTER 3 Characteristic gain functions and

characteristic f r e q u e n c y 'functions 22
3.1 Duality between open-loop gain and
closed-loop frequency 22
3.2 Algebraic functions: characteristic
gain functions and characteristic
frequency functions 25
3.3 Characteristic gain functions 27
3.3-1 Poles and zeros of a characteristic
gain function 29
3.3-2 Algebraic definition of poles and zeros
for a transfer function matrix 32
3.3-3 Relationship between algebraically defined
poles/zeros of the open-loop gain matrix
G(s) and the poles/zeros of the corresponding
set of characteristic gain functions 36
3.3-4 Riemann surface of a characteristic
gain function 39
3.3-5 Generalized root locus diagrams 46
3.3-6 Example of frequency surface and
characteristic frequency loci 47
3.4 Characteristic frequency functions 49
3.4-1 Generalized Nyquist diagram 50
3.4-2 Example of gain surface and
characteristic gain loci 51
References 57

CHAPTER 4 A ~eneralized Nyquist stability criterion 58

4.1 Generai izea Nyquist stability crfterion 58
4.2 Proof of the generalized Nyquist stability
criterion 60
4.3 Example 72
References 75

CHAPTER 5 A generalized inverse Nyquist stability

criterion 77
5.1 Inverse characteristic gain functions 77
5.2 Pole-zero relationships 78
 N

5.3 Inverse characteristic gain loci -

generalized inverse Nyquist diagrams 81
5.4 Generalized inverse Nyquist criterion 81
5.5 Proof of generalized inverse Nyquist
stability criterion 86
5.6 Example 97
References 99

CHAPTER 6 Multivariable root loci iOO

6.1 Theoretical background IOO
6.2 Asymptotic behaviour 104
6.2-1 Butterworth patterns 109
6.3 Angles of departure and approach 113
6.4 Example 1 115
6.5 Asymptotic behaviour of optimal closed-
loop poles 123
6.6 Example 2 126
References 130

CHAPTER 7 On parametricstability and future

research 132
7.1 Characteristic frequency and
characteristic parameter functions 132
7.2 Gain and phase margins 135
7.3 Example 135
7.4 Future research 138
References 140

Appendix 1 Definition of an algebraic functipn 143

Appendix 2 A reduction to the irreducible rational

canonical form .......................... 143

Appendix 3 The discriminant 147

Appendix 4 .method for constructing the Riemann

surface domains of the algebraic functions
correspondins~to an open-lo.op gaip
matrix G(s) 150

Appendix 5 Extended Principle of the. Argume.nt 157

Appendix 6 Multivariable pivots from the

ch'araceristic equation A'('g,s).=O 166

Appendix 7 Association between branch, point.s

and stat.ionary points on the gain
and frequency surf.aces 168

References 169

Bibliography 170

Index 174
 1. Introduction

The great success of the optimal control and optimal

filtering techniques developed for aerospace work during the

late 1950's and early 1960's n a t u r a l l y led to attempts to

apply these techniques to a wide range of e a r t h - b o u n d multi-

variable industrial processes. In many situations this was

less than immediately successful, particularly in cases where

the available plant models were not s u f f i c i e n t l y accurate or

where the p e r f o r m a n c e indices required to stipulate the controlled

plant b e h a v i o u r were m u c h less obvious in form than in the aero-

space context. Moreover, the controller w h i c h results from a

direct a p p l i c a t i o n of optimal control and optimal filtering

synthesis techniques is in general a c o m p l i c a t e d one; in fact,

if it incorporates a full K a l m a n - B u c y filter it has a

dynamical complexity equal to that of the plant which it is

controlling, since the filter e s s e n t i a l l y consists of a plant

model w i t h feedback around it. In contrast, what was needed

for many m u l t i v a r i a b l e process control problems was a relatively

simple controller w h i c h w o u l d both stabilize, about an operating

point, a plant for w h i c h only a very a p p r o x i m a t e model might be

available, and also mitigate the effect of low-frequency

disturbances by i n c o r p o r a t i n g integral action. To industrial

engineers brought up on f r e q u e n c y - r e s p o n s e ideas the s o p h i s t i c a t e d

optimal control methods seemed difficult to use; these engineers

e s s e n t i a l l y relied on a m i x t u r e of physical insight and straight-

forward techniques, such as the use of derivative and integral

action, to solve their problems. It became obvious that a

huge gap in techniques existed between the classical single-

loop f r e q u e n c y - r e s p o n s e methods, based on the work of Nyquist [i],

 2

Bode [2] and Evans [3],which were still in use for many

industrial applications, and the elegant and powerful multi-

variable time-response methods developed for aerospace

applications.

For these reasons an interest in frequency-response

methods slowly began to revive during the mid-1960's. An

important first step towards closing the yawning gap between

an optimal control approach and the classical frequency-response

approach was taken by Kalman [4] , who studied the frequency-

domain characterization of optimality. A systematic attack

on the whole problem of developing a frequency-response analysis

and design theory for multivariable systems was begun in a

pioneering paper by Rosenbrock [5] which ushered in a decade

of increasing interest in a rejuvenated frequency-response

approach. Prior to this new point-of-departure some fairly

straightforward attacks had been made on the multivariable

control problem. Boksenbom and Hood [6] put forward the

idea of a non-interacting controller. Their procedure consisted

simply of choosing a cascaded compensator such that the overall

transfer function matrix of the compensated system had a

diagonal form. If such a compensator could be found then the

controller design could be finished off using standard single-

loop design techniques. The required compensating matrix which

usually results from such a procedure is necessarily a complicated

one, and the most succinct objection to this approach is simply

that it is not essential to go to such drastic lengths merely

to reduce interaction. A natural further step in this initial

approach to multivariable control was to see what could be

achieved by way of standard matrix calculations using rational

matrices; papers studying the problem in this way were produced

 3

by Golomb and Usdin [7] , Raymond [8] ; Kavanagh [9] , [iO],[ii] ,

and Freeman [12] , [13] Rosenbrock [14],[15], however, opened up

a completely new line of d e v e l o p m e n t by seeking to reduce a multi-
variable p r o b l e m to one amenable to classical techniques in a more
sophisticated way. In his Inverse Nyquist Array Method [14] , [15]
the aim was to reduce interaction to an amount which would then
enable single-loop techniques to be employed, rather than to eliminate
interaction completely. The Rosenbrock approach was based upon a
careful use of a specific criterion of partial interaction - the
diagonal dominance concept. The success of this m e t h o d led other
investigators to develop ways of seeking to reduce a m u l t i v a r i a b l e
control p r o b l e m to a succession of single-loop problems, as in the
Sequential Return Difference approach of Mayne [163.

In the non-interacting, or partially non-interacting,

approach to m u l t i v a r i a b l e control the m o t i v a t i o n was the eventual
deployment of classical single-loop frequency-response techniques
during the final stages of a design study. An alternative approach,
however, is to investigate the t r a n s f e r - f u n c t i o n matrix representation

as a single object in its own right and to ask : how can the key basic
concepts of the classical single-loop frequency-response approach

be suitably extended? What are the relevant g e n e r a l i z a t i o n s to the

multivariable case of the specific concepts of pole, zero, Nyquist
diagram and root locus diagram? It is to questions of this sort
that the work p r e s e n t e d here is addressed, and it is shown that
complex-variable ideas have an important role to play in the study
of m u l t i v a r i a b l e feedback systems. An early attempt to extend

Nyquist diagram ideas to the m u l t i v a r i a b l e p r o b l e m was made by Bohn

[17] , [18] A generalization of the Nyquist stability criterion
was put forward by MacFarlane [19] and, following that heuristic
treatment, complex-variable based proofs were supplied by Barman
and K a t z e n e l s o n [20] and MacFarlane and Postlethwaite [21] . This

generalization of the Nyquist stability criterion to the m u l t i v a r i a b l e

situation was soon followed by complementary generalizations of the

root locus technique E21], E22], ~23], E24].

The aim of the work p r e s e n t e d in this text is to extend

the concepts underlying the techniques of Nyquist, Bode and Evans

to m u l t i v a r i a b l e systems. In the two classical approaches to

linear feedback system design the N y q u i s t - B o d e approach studies

gain as a function of frequency and the Evans' approach studies

frequency as a function of gain. In Chapter 3 it is shown how

the ideas of studying complex gain as a function of complex

frequency and complex frequency as a function of complex gain can

be extended to the m u l t i v a r i a b l e case by a s s o c i a t i n g with transfer

function matrices (having the same number of rows and columns) a

pair of analytic functions : a characteristic gain function and a

characteristic frequency function. These are algebraic functions

[253 and each is defined on an a p p r o p r i a t e Riemann surface [26].

Chapter 2 deals With a number of essential p r e l i m i n a r i e s such as

a d e s c r i p t i o n of the type of m u l t i v a r i a b l e feedback system being

considered; w i t h basic definitions of stability and related

theorems; and w i t h a fundamental r e l a t i o n s h i p between open- and

c l o s e d - l o o p behaviour based on the r e t u r n - d i f f e r e n c e operator.

Chapter 3 also contains a comprehensive discussion of the b a c k g r o u n d

to the g e n e r a l i z e d Nyquist stability criterion for m u l t i v a r i a b l e

feedback systems w h i c h is p r e s e n t e d in Chapter 4. The proof of

this criterion is based on the Principle of the Argument applied

to an algebraic function defined on an appropriate Riemann surface.

In Chapter 5 a generalization of the inverse Nyquist stability

criterion to the multivariable case is d e v e l o p e d which is complement-

ary to the exposition of the g e n e r a l i z e d Nyquist criterion given in

the previous chapter. Using the material developed in Chapter 3,

the Evans' root locus approach is extended to m u l t i v a r i a b l e systems

in Chapter 6; this uses well established results in algebraic

function theory. It is also shown how an algebraic-function

based approach can be used to find the asymptotic behaviour of

the closed-loop poles of a multivariable time-invariant optimal

linear regulator as the weight on the input terms of a quadratic

performance index approaches zero.

As the work presented progresses it becomes evident that

the gain variable used can be considered as a parameter of the
system, and consequently that the techniques developed are not
only applicable to gain and frequency but to any parameter and
frequency. In Chapter 7 the effect of parameter variations
on a multivariable feedback system is considered by the intro-
duction of the concepts of 'parametric' root loci and 'parametric'
Nyquist loci. This chapter concludes with a few tentative
proposals and suggestions for future research.

Information of secondary importance which would unnecessar-

ily break the flow of the text has been placed in appendices.
References are listed at the end of each chapter in which they
are cited, and also at the end of the text where a bibliography
is provided.

References

[i] H.Nyquist, "Regeneration theory", Bell Syst. Tech. J.,

ii, 126-147, 1932.

[23 H.W.Bode, "Network analysis and feedback amplifier design",

Van Nostrand, Princeton, N.J., 1945.

[3] W.R.Evans, "Graphical analysis of control systems", Trans.

AIEE, 67, 547-551, 1948.

[4] R.E.Kalman, "When is a linear control system optimal?",

Trans. ASME J.Basic Eng., Series D., 86, 51-60, 1964.

[5] H.H.Rosenbrock, "On the design of linear multivariable

control systems", Proc. Third IFAC Congress London, I,
1-16, 1966.
[6] A.S.Boksenbom and R.Hood, "General algebraic method
applied to control analysis of complex engine types",
National Advisory Committee for Aeronautics, Report
NCA-TR-980, Washington D.C., 1949.

[7] M.Golomb and E.Usdin, "A theory of multidimensional

servo systems", J.Franklin Inst., 253(1), 28-57, 1952.

[8] F.H.Raymond, "Introduction a l'~tude des asservissements

multiples simultanes", Bull. Soc. Fran. des Mecaniciens,
7, 18-25, 1953.

[9] R.J.Kavanagh, "Noninteraction in linear multivariable

systems", Trans. AIEE, 76, 95-100, 1957.

[i0 ] R.J.Kavanagh, "The application of matrix methods to

multivariable control systems", J.Franklin Inst., 262,
349-367, 1957.

[li ] R.J.Kavanagh, "Multivariable control system synthesis",

Trans. AIEE, Part 2, 77, 425-429, 1958.

[12 ] H.Freeman, "A synthesis method for multipole control

systems", Trans. AIEE, 76, 28-31, 1957.

[13 ] H.Freeman, "Stability and physical realizability con-

siderations in the synthesis of multipole control systems",
Trans.AIEE, Part 2, 77, 1-15, 1958.

[14 ] H.H.Rosenbrock, "Design of multivariable control systems

using the inverse Nyquist array, Proc. IEE, 116, 1929-1936,
1969.

[15 ] H.H.Rosenbrock, "Computer-aided control system design",

Academic Press, London, 1974.

[16 ] D.Q.Mayne, "The design of linear multivariable systems",

Automatica, 9, 201-207, 1973.

[17 ] E.V.Bohn, "Design and synthesis methods for a class of

multivariable feedback control systems based on single
variable methods", Trans.AIEE, 81, Part 2, 109--115, 1962.

[18 ] E.V.Bohn and T.Kasvand, "Use of matrix transformations and

system eigenvalues in the design of linear multivariable
control systems", Proc. IEE, iiO, 989-997, 1963.

[19 ] A.G.J.MacFarlane, "Return-difference and return-ratio

matrices and their use in the analysis and design of
multivariable feedback control systems", Proc. IEE, 117,
2037-2049, 1970.

[20 ] J.F.Barman and J.Katzenelson, "A generalized Nyquist-

type stability criterion for multivariable feedback
systems", Int.J.Control, 20, 593-622, 1974.
[21] A.G.J.MacFarlane and I. Postlethwaite, "The generalized
Nyquist stability criterion and multivariable root loci",
Int. J. Control, 25, 81-127, 1977.

[22] B.Kouvaritakis and U.Shaked, "Asymptotic behaviour of

root loci of linear multivariable systems", Int. J.
Control, 23, 297-340, 1976.

[23] I. Postlethwaite, " The asymptotic behaviour, the

angles of departure, and the angles of approach of
the characteristic frequency loci", Int. J. Control,
25, 677-695, 1977.

[24] A.G.J.MacFarlane, B.Kouvaritakis and ~ E d m u n d s ,

"Complex variable methods for multivariable feedback
systems analysis and design", Alternatives for Linear
Multivariable Control, National Engineering Consortium,
Chicago, 189-228, 1977.

[25] G.A.Bliss, "Algebraic functions", Dover, New York,

1966 (Reprint of 1933 original).

[26] G. Springer, "Introduction to Riemann surfaces",

Addison-Wesley, Reading, Mass., 1957.
 2. Preliminaries

This text considers the g e n e r a l i z a t i o n of the classical

techniques of Nyquist and Evans to a linear time-invariant

dynamical feedback system which consists of several multi-

input, m u l t i - o u t p u t subsystems connected in series. In this

chapter a d e s c r i p t i o n of the m u l t i v a r i a b l e feedback system

under c o n s i d e r a t i o n is given. The chapter also includes

basic definitions of stability, some associated theorems,

and a fundamental r e l a t i o n s h i p between open- and closed-loop

b e h a v i o u r based on the return-difference operator.

2.1 System description

The basic description of a linear t i m e - i n v a r i a n t dynamical

system is taken to be the state-space model

x(t) = Ax(t) + Bu(t)

(2.1.1)
y(t) = Cx(t) + Du(t)

where x(t) is the state vector, y(t) the output vector,

i

u(t) the input vector; x(t) denotes the derivative of

x(t) w i t h respect to time; A,B,C, and D are constant real

matrices. For c o n v e n i e n c e the model will be d e n o t e d by

S(A,B,C,D) or S when the m e a n i n g is obvious, and r e p r e s e n t e d

diagramatically as shown in figure i.

In general S(A,B,C,D) will be considered as being the

state-space representation of several subsystems

Si (Ai,Bi,Ci,Di) : xi(t) = A . x . (t) + B.u. (t)

1 i i i (2.1.2)

i=1,2, ...... h Yi(t) = C.x.l

l(t) + Diui(t)

connected in series, as i l l u s t r a t e d in figure 2. If, for

example,S consists of two subsystems S 1 and S 2 then the

 "I D ,I
+1
u(t ' t)
I
1
g''~~ fA 'c I
I
J S I
I J

Figure 1. Stote-spctce model

r" "~ 1

u{t}=u~(t} I
llYh(t)=
y(t)
I .~
I
I
'
I
S J

Figure 2. Series connection

of subsystems
 10

state-space description of S is g i v e n by e q u a t i o n s (2.1.1) with

x(t) = Ix l(t)] , the c o m b i n e d states of both s u b s y s t e m s ,

!
[x2(t) ]
u(t) = ul(t) , the input to S 1 ,

y(t) = Y2(t) , the o u t p u t of S 2, (2.1.3)

B2c I A2 tB2DlJ

C = [ D2C 1 C2 ] , and D = D2D 1.

The s t a t e - s p a c e model for an i n t e r c o n n e c t i o n of s e v e r a l

subsystems can be d e r i v e d from the above formula by s u c c e s s i v e

application.

The s t a t e - s p a c e model of a s y s t e m is o f t e n referred

to as an i n t e r n a l description since it r e t a i n s a knowledge

of the s y s t e m ' s internal dynamical structure. An e x t e r n a l

or i n p u t - o u t p u t description is o b t a i n e d if in e q u a t i o n s

(2.1.1) single-sided Laplace transforms [i] are taken, to give

s~(s) - x(o) = A~(s) + B~(s)

(2.1.4)
~(s) = c~(s) + D~(s)
where ~(s) denotes the L a p l a c e transform of x(t). If the

initial conditions at time t=O are all zero so that x(o)=O,

then the input and o u t p u t transform vectors are r e l a t e d by

~(s) = G(s) ~(s) (2.1.5)

where

G(s) = C(SIn-A)-IB + D (2.1.6)

I is a unit m a t r i x of o r d e r n and ( [idenotes the i n v e r s e

n
of a matrix. G(s) is a m a t r i x - v a l u e d rational function
 11

of the c o m p l e x variable s, and is c a l l e d the t r a n s f e r

function m a t r i x for the set of i n p u t - o u t p u t transforms, or

the o p e n - l o o p ~ain matrix. The t r a n s f e r function matrix

G(s) can be r e g a r d e d as d e s c r i b i n g a system's response to

an e x p o n e n t i a l input w i t h e x p o n e n t s [2], and t h e r e f o r e

the c o m p l e x variable s can be c o n s i d e r e d a complex

frequency variable as in the s i n g l e - i n p u t , single-output case.

When S(A,B,C,D) consists of h s u b s y s t e m s

{Si(Ai,Bi,Ci,Di): i=i,2, .... ,h} , as shown in figure 2,

each s u b s y s t e m has a t r a n s f e r function matrix

Gi(s ) = Ci(SIn.-Ai)
1
-1Bi + Di (2.1.7)

and the i n p u t - o u t p u t transform vectors of S are r e l a t e d by

~(s) =Sh (s) Gh_l (s) .... Gl(S) ~(s) (2.1.8)

with the o b v i o u s relationship for the o p e n - l o o p gain m a t r i x

G(s) = G h ( S ) G h _ l ( S ) .... Gl(S) (2.1.9)

For the p u r p o s e of c o n n e c t i n g outputs back to inputs to

form a f e e d b a c k loop G(s) is a s s u m e d to be a square m a t r i x

of o r d e r m.

2.2 Feedback confi@uration

The g e n e r a l feedback configuration that w i l l be c o n s i d e r e d

is shown in figure 3. The output of the f e e d b a c k system

is shown as that of the hth s u b s y s t e m but in p r a c t i c e it

m a y be the o u t p u t from an e a r l i e r subsystem in w h i c h case

the later subsystems can be t h o u g h t of as b e i n g feedback

compensators. The p a r a m e t e r k is a real gain c o n t r o l

variable c o m m o n to all the loops. The s y s t e m ' s input and

 12

output are r e l a t e d to the r e f e r e n c e input r(t) by the e q u a t i o n s

e(t) = r(t) - y(t)

(2.2.1)
u(t) = ke(t)

and c o m b i n i n g these w i t h equations (2.1.1) the f o l l o w i n g

closed-loop state-space equations are obtained:

x(t) = AcX (t) + Bcr (t)

(2.2.2)
y(t) = CcX (t) + D c r (t)

where

Ac = A-B(k-IIm+D)-Ic
B c = k B - k B (k-iIm+D)-iD

Cc = (Im+kD) -i c

Dc = (k-iim+D) -I D

y(t)
~l 7"'ml -11 1, "7"hi

Figure 3. Feedback configuration

 13

2.3 Stability

Stability is the most important single requirement

of a feedback system and for general time-dependent

nonlinear systems it poses very complex problems. The

stability problem for linear time-invariant dynamical

systems, however, is much simpler than in the general case.

This is because:

(i) all stability properties are constant with respect

to time, and

(ii) all stability properties are global, since any solution

for the state of the system is proportional to the state

at time zero; see equation (2.3.2).

There are many definitions of stability in the literature

and broadly speaking these can be divided into two classes.

The first class of definitions concerns stability of free

systems i.e. those in which there is no input; the second

class of definitions concerns the behaviour of forced

systems ioe. those in which there is a given input. Both

types of stability are discussed below, the definitions and

associated theorems following very closely those given by

Willems [3] .
2.3-1 Free systems

Let us consider the closed-loop dynamical system of

figure 3 described by equations (2.2.2) with r(t)=O and

Cc=I. Then the stability problem reduces to that of

considering the free system

x(t) = AcX(t) (2.3.1)

The equilibrium state for equation (2.3.1) is clearly the

 14

origin (assuming A is non-singular), and therefore a

c
solution of (2.3.1) which passes through the origin at some

time remains there for all subsequent times; this solution

is called the null solution. The stability of the origin

equilibrium state is characterized using the following

definitions

Definition i. The origin of the free system (2.3.1) is

called stable if when the system is perturbed from the

origin all subsequent motions remain in a correspondingly

small neighbourhood of the origin

Definition 2. The origin of the free system (2.3.1) is

called asymptotically stable if when the system is perturbed

slightly from the origin all subsequent motions return to

the origin.

Definition 3. The origin of the free system (2.3.1) is

called asymptotically stable in the large, or globally

asymptotically stable, if it is stable, and if every

motion converges to the origin as t+~.

The general solution of equation (2.3.1) is [3]

x(t ;x(t O),tO) = exp [ Ac(t-t O) ].x(t O) (2.3.2)

which shows clearly that if the free system is asymptotically

stable it is also asymptotically stable in the large. If J

is the Jordan canonical form [4] of A c such that

A = TJT -I (2.3.3)
c
with

J = Jl
J2

Jk

where each Jordan block J has the form

l
 IS

J. = I. 1
l l

~i ".

hi 1

and li is an eigenvalue of Ac, then it can be shown [ 3] , that

exp[ Ac(t-to) ] = T exp[J(t-to)].T-i (2.3.4)
with

exp [ J(t-t O) ] = exp[J 1 (t-to) ]

exp [ J2 (t-to)]

exp[ Jk (t-to)]

and

exp [ Ji (t-to)] = "i t t2/2 ' .... tr-I/(r-l) : exp[~ i (t-to) ]

0 1 t tr-2/(r-2)'

000 1
where ~ = (t-t O )
and r is the order of the Jordan block Ji" The general

solution of the free system can therefore be expressed as

x(t; x(t O),t O ) = T exp[J(t-to)].T-iX(to ) (2.3.5)

and from this the following theorems can be derived; see

[31 for proofs.

Theorem i. The null solution of system (2.3.1) is asymptotically

stable if and only if all eigenvalues of the matrix A c have

negative real parts.
Theorem 2. The null solution of system (2.3.1) is stable
if and only if the matrix A c has no eigenvalues with positive

real parts, and if the eigenvalues with zero real parts

correspond to Jordan blocks of order i.

 16

2.3-2 Forced s~stems

Let us consider the closed-loop dynamical system
(2.2.2) which has the general solution [3] ,
x(t; x(t o),t o ) = exp (Act).X(to)+/texp[Ac(t-T)].Bcr (T)dT
o
(2.3.6)
To study the stability properties of this system we need to
introduce the concept of input-output stability.
Definition 4. A dynamical system is called input-output
stable if for any bounded input a bounded output results
regardless of the initial state.
By theorem i, asymptotic stability of the unforced
system (2.3.1) implies that all the eigenvalues of A c have
negative real parts, in which case there exist positive numbers
P and a such that
llexp(Act) ll < P exp (-at) ~t~O (2.3.7)
where II. II denotes the Euclidean norm of a matrix or
vector [3]. From equations (2.2.2), (2.3.6) and (2.3.7)

we then have

II y(t)ll -< H Ccx(t)II + II Dcr(t)II

4 II Dcr(t) ll + II Ccexp(Act)'x(to)II

+Cfotllexp[Ac (t- T)] IIllBcr(t)lld~

d +c llx(to)li+ cbMP/a

where b=llBcll, c=llCcl], d=llDcl], and ]]r(t)l]4M ~ t >~ O.

This result is summarized in the following theorem.

Theorem 3. If the null solution of the unforced system
(2.3.1) is asymptotically stable, then the forced system (2.2.2)
is input-output stable.
Note that input -output stability implies asymptotic stability
 17

of the equilibrium state at the origin only if the system

(2.2.2) is state controllable and state observable; or if

all unobservable and/or uncontrollable modes have negative

real parts. In the remainder of this book system stabilit ~

is understood as meaning input-output stability coupled with

asymptotic stability of the equilibrium state at the origin.

Theorem 3 is important because it tells us that the

stability of a linear time-invariant system can be determined

solely from a knowledge of the eigenvalues of the system "A"

matrix. The stability conscious eigenvalues corresponding

to the closed-loop dynamical system (2.2.2) are values of 1

which satisfy the equation

det[IIn-Ac] = O (2.3.8)

The left-hand side of equation (2.3.8) is called the

closed-loop characteristi ~ polynomial, abbreviated as CLCP(1)

so that

CLCP(1)~ det [ IIn-Ac] (2.3.9)

Similarly for the open-loop system S(A,B,C,D) an open-loop

characteristic polynomial, OLCP(1), is defined as

OLCP(1) ~ det[IIn-A ]

= detflInl-Al]det [ IIn2-A 2] . . . .

.... det[iInh-Ah] (2.3.10)

In the next section it is shown how the open- and

closed-loop characteristic polynomials are related via the

return-difference operator [ 5 ] .

2.4. Relationship between open- and closed-loop characteristic

polynomials for the general feedback configuration

Let us suppose that all the feedback loops of the general

 18

~(s)

(a)

P(s) 6(s) r--'q ~(S) r--q 9(s)

- ~ k [m~----~Gl(S)k ..........-,,pl~S)1

i 0 0 ....

(b)
Figure 4. Feedback configuration
(a) dosed-loop
(b) open-toop
 19

closed-loop configuration are broken and that the subsystems

are represented by their transfer function matrices; see

figure 4. The corresponding return-difference matrix

[5] for this break point is

F(s) I + L(s) (2.4.1)

m
where

L(s) = kGh(S)Gh_l(S) .... Gl(S)

= kG(s) (2.4.2)

is called the system return-ratio matrix [5] A return-

difference operator generates the difference between injected

and returned signal transforms from the injected signal

transform. It plays a major role in feedback theory since

the essence of forging a feedback link is making two sets

of signals identically equal, thus making the difference

between them identically zero. Both F(s) and L(s) are

matrix-valued rational functions of a complex variable and

the key concepts in this text revolve around the properties

of such matrices. The importance of the return-difference

matrix is emphasized in the relationship between open- and

closed-loop characteristic polynomials which is now derived

for the general feedback configuration.

If we take determinants of equation (2.4.1) and represent

G(s) by its state-space model, we obtain

detF(s) = det[Im+kC(SIn-A)-iB+kD] (2.4.3)

which using Schur's formula [6] for the evaluation of

partitioned determinants can be rewritten as

detF(s) = det[sI -A i B l+det[SIn-A ] (2.4.4)

which is equivalent to
 20

detF(s) = det
-I
% letlSn l+etSn
: Is L--~C--J, I S k
= det IsIn-A+B (k-IIm+D) -IC :--im+~
0 ]-- det[SIn-A ]
[----ic ...... -,
= det [SIn-A+B(k-iIm+D)-iC]det[Im+k~ (2.4.5)
det [SIn-A ]
Now from equations (2.2.2) we have
A c = A-B(k-IIm+D)-Ic
and it is obvious from equation (2.4.3) that
detF(~) = det[Im+k ~
and therefore under the assumption that det F(~)~O we have
from equation (2.4.5) the following relationship

detF(s) = det[SIn-Ac] = d e t [,,,,s, In-, ,A,,,,,,,,c,, ] ~ CLCP(s)

detF (~) det [Sin_ A ] det [SInh_Ah] . . ..det
..... [slnl_A~ OLCP (s)

(2.4.6)
The zeros of the open- and closed-loop characteristic
polynomials, OLCP(s) and CLCP(s), are known as the open-
and closed-loop poles or characteristic frequencies respectively.
Relationship (2.4.6) shows how the matrix-valued
rational transfer functions F(s) and G(s) are intimately
related to the stability of a dynamical feedback system.
The study of such matrices and their eigenvalues opens the way
to suitable extensions of the classical techniques of Nyquist
[7] and Evans [8;9] ; the results of such a study are given
in Chapter 3.
References

[I] R. Bracewell, "The Fourier Transform and Its Applications",

McGraw-Hill, New York, 1965.
 21

[2] A.G.J. MacFarlane and N. Karcanias, "Poles and zeros of

linear multivariable systems: a survey of the algebraic,
geometric and complex variable theory", Int. J. Control, 24,
33-74, 1976.

[3] J.L. Willems, "Stability Theory of Dynamical Systems",

Nelson, London, 1970.

[4] P.M. Cohn, "Algebra", Vol. i, Wiley, London, 1974.

5] A.G.J. MacFarlane, '~eturn-difference and return-ratio matrices

and their use in analysis and design of multivariable
feedback control systems", Proc. IEE, 117, 2037-2049, 1970.

[6] F.R. Gantmacher, "Theory of M a t r i c e ~ Vol. I, Chelsea, New

York, 1959.

[7] H. Nyquist, "The Regeneration Theory", Bell System Tech. J.,

ii, 126-147, 1932.

[8] W.R. Evans, "Graphical Analysis of Control Systems", Trans.

AIEE, 67, 547-551, 1948.

[9] W.R. Evans, "Control System Synthesis by Root Locus Method",

Trans. AIEE, 69, 1-4, 1950.
 3. Characteristic gain functions and
characteristic frequency functions

In the analysis and design of linear single-loop

feedback systems the two classical approaches use complex
functions to study open-loop gain as a function of imposed
frequency (theNyquist-Bode approach), and to study closed-
loop frequency as a function of imposed gain (the Evans
root locus approach). The primary purpose of this chapter
is to show how these techniques can be extended to the multi-
variable case by associating with appropriate matrix-
valued rational functions of a complex variable characteristic

gain functions and characteristic frequency function s.

3.1 Duality between open-loop vai n and closed-loop frequency

For the general feedback configuration of figure 4 we
have from section 2.4 the fundamental relationship

detF (s) det [Sln-Ac ]

detF(-~ = det[s~n, A ] (3.1.1)

where the return-difference matrix F(s) is given as

F(s) =
I + kG(s) (3.1.2)
m
If we substitute for F(s) in equation (3.1.1) we obtain

det[ SIn-Ac] _ det[ Im+kG(s) ]

det[ SIn-A ] det[ Im+kD ]

det[k-iIm+G(s~
= (3.1.3)
det[ k-iIm+D]

and substituting for the gain variable k using the expression

g= -I (3.1.4)

where g is allowed to be complex i.e. g e ~ (the complex

plane), we have
 23

det[ SIn-Ac ] det[gIm-G (S)]

(3.1 .5)
det[ Sin-A ] det[ qIm-D ]
The closed-loop system matrix A c is given in equations
(2.2.2) as
A c = A - B(k-IIm+D)-Ic (3.1.6)

and substituting for k from equation (3.1.4) we have

A c = A + B(gIm-D)-Ic
S(g) (3.1.7)
The expression (3.1.5) can therefore be rewritten as

det[ SIn-S (g)] det[gIm-G(s) ]

det[ Sin-A ] det[gIm-D ]
or (3.1.8)

det[ SIn-S (g)] det[ gIm-G (s)]

det[ SIn-S (~ ~ det[ gIm-G (~)]

The form of this relationship shows a striking 'duality'

between the complex frequency variable s and the complex
gain variable g via their 'parent' matrices S(g) and G(s)
respectively. This duality between the roles of frequency
and gain forms the basis on which the classical complex
variable methods are generalized to the multivariable case.
S(g) is called the closed-loop frequency matrix; its eigen-
values are the closed-loop characteristic frequencies and
are clearly dependent on the gain variable g. The eigenvalues
of the open-loop 9ain matrix G(s) are called open-loop
characteristic ~ains and are clearly dependent on the frequency
variable s. The similarity between G(s) and S(g) is
stressed if one examines their state-space structures:
G(s) = C(SIn-A)-IB + D (3.1.9)
S(g) = B(gIm-D)-Ic + A (3.1.10)
 24

In figure 5 the feedback configuration of figure 4a is

redrawn with zero reference input, the state-space representation

for G(s),and the substitution (3.1.4) for k in order to

illustrate explicity the duality between the closed-loop

characteristic frequency variable s and the open-loop

characteristic gain variable g.

~ I n _ ] ......

Figure 5. Feedback configuration illustrating

the duality between s and g

The importance of relationship (3.1.8) is that it shows,

for values of s ~ ~ (A) and values of g ~ ~(D) (this condition

is equivalent to det F(~)~O which has already been assumed),

where a(A) denotes the spectrum of A , that

I det[SIn-S(g)] = O ~ det[gIm-G(s)]=O~ (3.1.11)

 25

This tells us that a knowledge of the open-loop characteristic

gain as a function of frequency is equivalent to a knowledge

of closed-loop characteristic frequency as a function of

gain. The inference from this is that it ought to be possible

to determine the stability of a feedback system from a

knowledge of the characteristic gain spectrum of G(s).

Note that from equation (3.1.8) we have that

CLCP(s) = det[gIm-G(s)] . OLCP(s)

(3.1.12)
det[gI m -D]

and such an expression makes it intuitively obvious that

there should be a generalization of Nyquist's stability theorem

to loci of the characteristic gains of G(s) as a function of

frequency.

3.2 A19ebraic functions: characteristic 9ain functions and

characteristic ' frequency function s.

The characteristic equations for G(s) and S(g) i.e.

d(g,s) ~ det[gI m - G(s)] = 0 (3.2.1)

and
?(s,g) ~ det[sI n - S(g)] = 0 (3.2.2)

are algebraic equations relating the complex v a r i a b l ~ s and g.

Each equation can be considered as a polynomial in g or s

with coefficients which are rational functions in s or g

respectively, and if irreducible over the field of rational

functions each equation defines a pair of algebraic functions

[i; appendix i]:

(i) a characteristic 9 a i n function g(s) which gives open-

loop characteristic gain as a function of frequency, and

(ii) a characteristic frequency function s(9) which gives

 26

closed-loopcharacteristic frequency as a function of gain.

In general equations (3.2.1) and (3.2.2) will not be

irreducible and each equation will define a set of characteri-

stic gain and characteristic frequency functions. For

simplicity of exposition and because this is in any case the

usual situation for G(s) and S(g) arising from p r a c t i c a l

situations, it will normally be assumed that equations

(3.2.1) and (3.2.2) are irreducible over the field of rational

functions.

A l t h o u g h b o t h equation (3.2.1) and equation (3.2.2)

define the same functions g(s) and s(g), equation (3.2.2)

will in general contain more information about the system.

It is possible under certain c i r c u m s t a n c e s that ?(s,g) will

contain factors of s independent of g w h i c h are not present

in ~(g,s). These factors occur in the following situations:-

(i) When the A - m a t r i x of the open-loop system S(A,B,C,D)

has eigenvalues which correspond to modes of the system

which are u n o b s e r v a b l e and/or u n c o n t r o l l a b l e from the point

of view of considering the input as that of the first sub-

system and the output as that of the hth sybsystem. Note

that if output m e a s u r e m e n t s for earlier subsystems are

available then in practice some of the u n o b s e r v a b l e modes

of S(A,B,C,D) m a y in fact be observable.

(2) When the poles and zeros of the o p e n - l o o p gain matrix

G(s) are different from the poles and zeros of the characteristic

gain function g(s); see section 3.3-3.

These two conditions, under w h i c h equations (3.2.1) and

(3.2.2) differ, clearly p r e s e n t problems to the d e v e l o p m e n t

 27

of a Nyquist-like stability criterion in terms-of loci of

the characteristic gains of G(s). However, by relating the

poles and zeros of g(s) to the poles and zeros of G(s), and

by careful consideration of the unobservable and uncontrollable

modes these problems can be overcome; a generalized Nyquist

stability criterion is developed in chapter 4.

In the next section a detailed study of the characteristic

gain function is given which results in a generalization

of the root locus diagram. In section 3.4 a similar study

of the characteristic frequency function results in a

generalized Nyquist diagram.

3.3 Characteristic @ain functions

The natural way to define the characteristic gain functiong(s)

is via the characteristic equation

A(g,s) ~ det[gIm-G(s) ] = O (3.3.1)

In general ~(g,s) will be reducible to the form

A(g,s) = Al(g,s)~2(g,s) ...... ~(g,s) (3.3.2)

where the factors {Ai(g,s) :i=1,2 .... ,} are polynomials in

g which are irreducible over the field of rational functions

in s. Let the irreducible factors Ai(g,s) have the form

ti ti-i
Ai(g's) = gi +ail(s)gi + ...... +ait. (s)=O
1 (3.3.3)

where t i is the degree of the ith irreducible polynomial

and the coefficients {aij(s) :i=l,2,...,;j=l,2,...,t i} are

rational functions in s. Then if b (s) is the least

lO
common denominator of the coefficients {aij(s) :j=l,2,...,t i}

equation (3.3.3) can be put in the form

 28

t. t.-i
bio(S)g i l+bil(s)g i i + ...... +biti(s) = O (3.3.4)

i=i,2,...,

where the coefficients {bij(s) li=l,2,...,;j=l,2, .... ,t i}

are polynomials in s. The function of a complex variable

gi(s) defined by equation (3.3.4) is called an algebraic

function [i; appendix i]. Thus associated with an open-

loop gain matrix G(s) is a set of algebraic functions

{gi(s): i=l,2,...,Z} which are directly related to the

eigenvalues of G(s). The characteristic gain functions

of G(s) are defined to be the set of algebraic functions

{gi(s):i=l,2,...,~}.

The problem of finding the irreducible polynomials

{A i(g,s):i=l,2,...,} from which the characteristic gain

functions are defined is closely linked to the problem of

finding an appropriate canonical form of G(s). If A(g,s)

was reducible to factors linear in g then G(s) could be

put into Jordan form [2]. In general this will not be the

case and a suitable canonical form is defined as follows.

Let

O O . . O -aiti(s)

10 . . . O -ai,ti_l(S
(3.3.5)
C(Ai)~ O 1 . . . O -ai,ti_2(s)

O 1 -ail (s)

for t.>l with

1

c (A i) =A - a l l (s) if ti=l (3.3.6)

 29

then a transformation matrix E(s) exists such that

-i
G(s) = E(s) Q(s) E(s) (3.3.7)

where Q(s) is a unique block diagonal matrix, which is

called the irreducible rational canonical form of G(s)

and is given by

Q(s) ~ diag[C(A I),C(A 2) ...... ,C(Az) ] (3.3.8)

It is clear that given Q(s) the irreducible factors Ai(g,s)

can easily be obtained. A proposed method for finding Q(s)

for any given G(s) is presented in appendix 2.

3.3-1 Poles and zeros of a characteristic vain function

Consider the defining equation for a characteristic gain

function g(s) :

~(g,s)~bo(s)g t + bl(S)g t-I + ... + bt(s) = O

(3.3.9)
We will take both

bo(S) ~ O and bt(s) ~ 0

since, if either or both of these polynomial coefficients

were to vanish, we could

find a reduced-order equation such
and
that both the coefficients of the highest/zeroth powers of

g(s) were non-zero; this reduced-order equation would then

be taken as defining an appropriate new algebraic function

for whose defining equation the supposition would be true.

It m a y happen however that bo(S) and bt(s ) share a

common factor and thus both vanish together at some specific

set of values of s . Before looking at the effect of this,

consider the situation when bo(S) and bt(s) do not share

a common factor. The algebraic function will obviously be

zero when

bt(s) = O (3.3.10)
 30

and w i l l t e n d to i n f i n i t y as

bo(S) + O (3.3.11)

For this r e a s o n those v a l u e s of s which satisfy equation

(3.3.10) are d e f i n e d to be the zeros of the a l g e b r a i c function

g(s), and those v a l u e s of s which satisfy the e q u a t i o n

bo(S) = O (3.3.12)

are d e f i n e d to be the p o l e s of the a l g e b r a i c function g(s).

Unless stated otherwise the t e r m i n o l o g y 'poles and zeros'

should be t a k e n as r e f e r r i n g o n l y to finite poles and zeros.

The p o i n t s =~ r e q u i r e s special attention and is dealt w i t h

at the end of this sub-section.

In o r d e r to be able to take e q u a t i o n s (3.3.1o) and

(3.3.12) as d e f i n i n g the zeros and p o l e s of g(s) in the

general case, we m u s t show that they r e m a i n appropriate when

bo(s) and bt(s) share a common factor. Let us first d i s p o s e

of the t r i v i a l case w h e n all the c o e f f i c i e n t s {hi(s) : i=O,2,...,t}

share a c o m m o n factor by saying that such a common factor

would simply be d i v i d e d out to get a n e w d e f i n i n g equation

for an a p p r o p r i a t e algebraic function. Suppose t h e n that

bo(S) and bt(s) have a c o m m o n factor, but that some non-

empty set of c o e f f i c i e n t s {bu(S),bu+l(S),...,bv(S)} do

not share this c o m m o n factor. Then d i v i d i n g through the

left-hand side of e q u a t i o n (3.3.9) by bo(S) we get

t bl(S) gt-i b (s) bv(S) bt(s)
g + b----~ + "'" + u gt-U+, t-v+
-o" " bo(S ) "'+bo---~ "''+bo (s) = O
(3.3.13)

Then, as s + s w h e r e s is a zero of the c o m m o n factor of bo(S)

and bt(s), the m o d u l i of the c o e f f i c i e n t set

 31

bu(S) bv(S)

all become arbitrarily large, and it is obvious that g(s)

will have a pole at s = s .

Again, suppose that bo(S) and bt(s) have a common

factor but that some non-empty set of coefficients

{bj(s), .... bm(S)} do not. Then as s ~ s where s is

a zero of the common factor, the algebraic equation (3.3.9)

may be replaced by

bj(s)gt-j(s)+ ... + bm(~)gt-m(~) = O (3.3.14)

where

bj(s) ~ O . . . . . bm(S) O

so that we must have

g(s) = 0

showing that s is indeed a zero of the algebraic function g(s).

We thus conclude that equations (3.3.10) and (3.3.12)

may be taken as defining the finite zeros and finite poles

of the algebraic function g(s), and that use of these

definitions enables us to cope with the existence of coincident

poles and zeros. The pole and zero polynomials of g(s),

denoted by p (s) and z (s), are defined as

g g
pg(S) ~ bo/ (s)

and (3.3.15)

Zg(S) ~ btl (s)

where b~(s) and bt[ (s) are the monic polynomials obtained

from bo(S) and bt(s) respectively, by dividing each polynomial

by its leading coefficient.

 32

For the purpose of considering g(s) at the point s=~

we put
-i
s=z (3.3.16)

so that

~ (g,s) =~ (g,z-l)=z -q ~ (g,z) (3.3.17)

where q is the number of finite poles of g(s). In any

neighbourhood of the value z=O (the point z=O itself being

excluded from it) the equation ~(g,s)=O is equivalent to the

equation ~(g,z)=O. Therefore if we consider the equation

&
~(g,z)=Co(Z)gt+cl(z)gt-l+ .... +ct(z)=O (3.3.18)

it follows that:

(i) s =~ is a pole of the characteristic gain function g(s)

if and only if Co(O)=O

(ii) s =~ is a zero of the characteristic gain function g(s)

if and only if ct(o)=O

For an open-loop gain matrix G(s) describing a physically

realizable system, which by definition (see section 2.1) we

are considering here, it is not possible for g(s) to have

poles at infinity. In fact it is easy to show that for s=~

the values of the characteristic gain function g(s) are

simply the eigenvalues of D.

3.3-2 A!~ebraic definition of poles and zeros for a transfer

function matrix

Let T(s) be an mx rational m a t r i x - v a l u e d function of

the complex variable s. Then there exists a canonical

form for T(s), the Smith-McMillan form [3 ] MCs) , such

that

T(s) = H(s)M(s)J(s) (3.3.19)

 33

where the mm matrix H(s) and the matrix J(s) are both
unimodular (that is having a constant value for their determinants,
independent of s ). If r is the normal rank of T(s)
(that is T(s) has rank r for almost all values of s ) then
M(s) has the form
]
M(s) = [ M*(s)rr Or, Z-r I (3.3.20)
I
r,r Om-r,m-rJ
with

Cl(S ) e2(s) er(S) l

M* (S) -- diag ~--~-~ , ~2(s ) , . . . . ~--~-~j(3.3.21)

where:
(i) each ei(s) divides all ei+ j (s) and

(ii) each ~i(s) divides all ~i-j (s).

With an appropriate partitioning of H(s),M(s) and J(s) we

therefore have

= HI (s)M* (S) Jl (s) (3.3.22)

where M*(s) is as defined in equation (3.3.21).
Thus T(S) may be expressed in the form

T(s) = Hi(s)
Idiag{ ei (s) 7
~i-~-s~}j Jl (s)

r ci(s) t
= (3.3.23)
h i(s) ~ 3i(s)
i=l
where :
(i) {hi(s) : i = 1,2 .... ,r} are the columns of the
matrix H l(s) ;
 34

(ii) {j~(s) : i = 1,2,...,r} are the rows of the

matrix Jl(S)

We know that

r ~ min(,m)

and that H(s) and J(s) are unimodular matrices of full

rank m and Z respectively for all s. Suppose T(s) is the

transfer function matrix for a system with input transform

vector %(s) and output transform vector ~(s). Then any

A
input vector ~(s) is turned into an output vector y(s)

by

(3.3.24)

For the single-input single-output case where

y(s) = ke(s) ~(s)

~(s)

with k a constant, the transfer function

ks(s)
g(s) =
~(s)
is defined as having zeros at those values of s where (s)

vanishes and poles at those values of s where ~(s) vanishes.

Thus for a non-zero ~(s) the modulus of ~(s) vanishes

when s is a zero of g(s), and becomes arbitrarily large when s

is a pole of g(s). A natural way therefore to characterize

the zeros and poles of T(s) is in terms of those values of

s for which II ~(s)II becomes zero for non-zero II ~(s)ll '

and arbitrarily large for finite II ~(s) II ' where II "II

denotes the standard vector norm. 9his natural extension

 35

of scalar case ideas leads directly to definitions of zeros

and poles of T(s) in terms of the Smith-McMillan form

quantities

E(s)

because of the following pair of simple results.

Zero lemma: II 9 (s)II vanishes for II ~ (s)II ~ O and s finite

if and only if some e. (s) is zero.

l

Pole lemma: II 9 (s)II+ ~ for II~(s)II < ~ if and only if

some ~i (s) + O.

These considerations lead naturally to the following

definitions [3].

Poles of T(s): The poIes of T(s) are defined to be the set

of all zeros of the set of polynomials {~i(s) : i = 1,2,...,r}.

In what follows we will usually denote the poles of T(s) by

{pl,P2,...,p n} and put

PT(S) = (s-Pl) (s-P2) ... (S-Pn) (3.3.25)

where PT(S) is conveniently referred to as the pole polynomial

of T(s) and is given by

r
PT(S) = ~ ~i(s) (3.3.26)
i=l

Zeros of T(s): The zeros of T(s) are defined to be the set

of all zeros of the set of polynomials {si(s) : i = 1,2,...,r}.

We will normally denote the zeros of T(s) by {Zl,Z 2 .... ,z }

and put

ZT(S) = (S-Zl) (s-z 2) ... (s-z) (3.3.27)

where ZT(S) is conveniently referred to as the zero polynomial

 38

of T(s) and is given by

r
ZT(S) = ~ g. (s) (3.3.28)
i=l 1

It is important to remember that ZT(S) and PT(S) are not

necessarily relatively prime; for this reason it is wrong

to simply define ZT(S) and PT(S) for a square matrix T(s)

as the numerator and demominator polynomials of det T(s).

Rules for calculatin9 pole polyDomials and z e r o p o l y n o m i a l s

The route via the Smith-McMillan form is not always

convenient for the determination of the poles and zeros of

T(s), particularly if the calculation is being done by hand.

The following rules [4] can be shown to give the same results

as the Smith-McMillan definitions.

Pole polynomial rule: PT(S) is the monic polynomial obtained

from the least common denominator of all non-zero minors of

all orders of T(s).

Zero polynomial rule: ZT(S) is the monic polynomial obtained

from the greatest common divisor of the numerators of all

minors of T(s) of order r (r being the normal rank of T(s))

which minors hav e a l l been adjusted to have PT(S) as thei r

common denominator

3.3-3 Relationship between algebraically defined poles/zeros

of the open-loop 9ain matrix G(s) and the p~es/zer0 s of the

correspondin 9 set of characteristic @ain functions

As a key step in the establishment of a generalized

Nyquist stability criterion, it is crucially important to

relate the poles and zeros defined by algebraic means to

complex variable theory, and thus to the poles and zeros of

 37

the set of characteristic gain functions.

The coefficients ai(s) in the expansion

det [gIm-G(s) ] = gm + al(s)gm-i + a2(s)gm-2 + ... +am(S)

(3.3.29)

are all appropriate sums of minors of Q(s) since it is well

known that:

det [gIm-G (s) ]

= g m - [trace G(s)]g m-1 + [~principal minors of G(s) of order 2]g m-2

- ... + (-l)mdet G(s) (3.3.30)

and thus the pcle polynomial bo;(S) is the monic polynomial

obtained from the least common denominator of all non-zero

principal minors of all orders of G(s).

Now the pole polynomial p~s) of a square matrix G(s)

is the monic polynomial obtained from the least common

denominator of all non-zero minors of all orders of G(s).

Therefore, if eG(s) is the monic polynomial obtained from the

least common denominator of all non-zero non-principa!

minors, with all factors common to bo;(S) removed, we have

that

PG (s) = eG(S)bo/(S) (3.3.31)

Furthermore since

det G(s) = am(S) = bm(S) (3.3.32)

5--(s)
o

and since from the Smith-McMillan form for G(s)

det G(s) = ~. ZG(S ) (3.3.33)

PG(S)

where a is a scalar quantity independent of s, we must have

that
 38

ZG(S ) = eG (s) bml(S) (3.3.34)

In many cases the least common denominator of the non-

zero non-principal minors of G(s) will divide bo(S) ,

in which case eG(s) will be unity and the p01e and zero

polynomials for G(s) will be b~(s) and b i(s) respectively.

m
In general a square-matrix-valued function of a complex

variable G(s) will have a set of irreducible characteristic

gain functions in the form specified by equation (3.3.3) and

the general form for thepole and zero polynomials can be

written as

PG(S) = eG(s) H b/lo(s) (3.3.35)

i=l

and

ZG(S) = eG(s) ~ b i'ti(s)

i=l I (3.3.36)

where the pole and zero polynomials for the jth characteristic

gain function gj(s) are b!o(S) and b/ (s) respectively.

3 3,tj

Example demonstrating th 9 p o l e V z e r o r e l a t i o n s h i p s

Let
(s-l) (s+2) O
1
G(s) z

(s+l) (s+2) (s-l)

i s+l
o] (s+l) (s+2) (s-l) (s+l)

-i 1
s-i s+2

The pole polynomial for G(s) is obviously

PG(S) = (s+l)(s+2) (s-l)

and consequently the zero polynomial is

ZG(S) = (s-l)
 39

The c h a r a c t e r i s t i c equation for G(s) is

i 1
det LgI-G(s)J = (g s+l ) (g - s - ~ ) = 0

so that the i r r e d u c i b l e characteristic equations are

1
Al(g,s) = g - ~-~ = O

and
1
A 2 (g,s) = g - s+---2 = O

which may be w r i t t e n as

(s+l)g - 1 = O

and

(s+2) g - 1 = 0

Therefore the p o ~ and zero p o l y n o m i a l s for the c h a r a c t e r i s t i c

gain functions gl(S) and g2(s) are

pgl(S) = blo(s) = (s+l) Zg I (s) = bll (s) = 1

-1
Pg2 (s) = b 2 0 (s) = (s+2) Zg 2(s) = b21(s) = 1
-i

Now for G(s) the m o n i c p o ~ n o m i a l obtained from the least

common denominator of all n o n - z e r o non-principal minors with

I I
all factors common to bJ(s)~ (=b~'o(S)b20(s))a removed is given by

e G (s) = (s-l)
which v e r i f i e s the r e l a t i o n s h i p s
2
PG(S) = eG(s) H b.; (s)
i= 1 lO

and
2
ZG(S) = eG(s) i=iH bilJ (s)

3.3-4 Riemann surface of a c h a r a c t e r i s t i c gain function

A characteristic gain function g(s) is d e f i n e d by an

irreducible equation of the form

 40

bo(S)g t + bl(S)g t-I + ... + bt(s) = O (3.3.37)

having in general t distinct finite roots. An exception

occurs only if

(a) bo(S) = O, because the degree of the equation is then

lowered, and as b (s)O one or more of the roots becomes

o
infinite; or if

(b) the equation has multiple roots.

This last situation can occur for finite values of s if, and

only if, an exPression , called the discriminant of the

equation, vanishes. The discriminant [5] is an entire rational

function of the equation coefficients; it will be denoted

by Dg(S) , and is discussed in appendix 3.

Ordinary points of the characteristic gain function

An ordinary point [1;6] of the characteristic gain

function g(s) is any finite point of the complex plane such

that bo(S) ~ O and Dg(S) ~ O.

Critical points of t h e characteristic gain function

A critical point [i; 6] of g(s) is any point of the

complex plane at which either

b (s) = O or D (s) = O,
o g
or both, plus the point s =~.

Branch points of the characteristic function

Solutions of

D (s) = O
g
are called finite branch points of the characteristic gain

function. The point at infinity is a branch point if the

discriminant Dg(Z) of equation (3.3.18) satisfies Dg(O) = O

At every ordinary p o i n t t h e equation (3.3.37) defining the

 41

characteristic gain function has t distinct roots, since

the discriminant does not vanish. The theory of algebraic

functions [i] then shows that in a simply connected region

of the complex plane punctured by the exclusion of the critical

points the values of the characteristic gain function g(s)

form a set of analytic functions; each of these analytic

functions is called a branch of the c h a r a c t e r i s t i c gain

function g(s). Arguments based on standard techniques of

analytic continuation, together w i t h the properties of

algebraic equations, show that the various branches can be

organized into a single entity: the corresponding algebraic

function. This is summarized in the following basic theorem

of algebraic function theory: an irreducible algebraic

equation of the form (3.3.37) defines precisely one t-valued

regular function g(s) in the punctured plane [7 ] .

Functions defined in this w a y are called algebraic

functions, and can be regarded as natural g e n e r a l i z a t i o n s

of the familiar e l e m e n t a r y functions of a complex variable.

An elementary function of a complex variable has the set of

complex numbers C as both its domain and its range. An

algebraic function has the complex number set C as its range

but has a new and a p p r o p r i a t e l y defined domain R which is

called its Riemann Surface [~ Since the Riemann

surface of an algebraic function plays a crucial role in this

work it is important to have an intuitive grasp of the ideas

underlying its definition and formation, which is therefore

now briefly considered.

 42

Figure 6. Anotytic continuation

S u p p o s e we h a v e a r e p r e s e n t a t i o n of p a r t of one b r a n c h of

an a l g e b r a i c function in the f o r m of a p o w e r series; such

a representation is u s u a l l y c a l l e d a f u n c t i o n a l e l e m e n t .

I m a g i n e its c i r c l e of c o n v e r g e n c e to be cut out of p a p e r and

t h a t the i n d i v i d u a l p o i n t s of t h e p a p e r d i s c are m a d e b e a r e r s

of the u n i q u e f u n c t i o n a l v a l u e s of the e l e m e n t s . If n o w this

initital element is a n a l y t i c a l l y c o n t i n u e d by m e a n s of a

second power series, another c i r c l e of c o n v e r g e n c e can be

t h o u g h t of as b e i n g cut out and p a s t e d p a r t l y o v e r the first,

as i l l u s t r a t e d by f i g u r e 6. The p a r t s p a s t e d t o g e t h e r are

m a d e b e a r e r s of the same f u n c t i o n a l v a l u e s and are a c c o r d i n g l y

t r e a t e d as a s i n g l e r e g i o n c o v e r e d o n c e w i t h values. If a

further analytic continuation is c a r r i e d out, a further disc

is s i m i l a r l y p a s t e d on to the p r e c e d i n g one. Now suppose

that, after repeated analytic continuations, one of the d i s c s

lies o v e r a n o t h e r disc, not a s s o c i a t e d w i t h an i m m e d i a t e l y

 43

preceding analytic continuation, as shown in figure 7.

Such an o v e r l a p p i n g disc is pasted together with the one

it overlaps if and only if both are bearers of the same

functional values. If,however,they bear different functional

values they are allowed to overlap but remain disconnected.

Thus two sheets, which are bearers of different functional

values, become superimposed on this part of the complex plane.

1
Figure 7 Repeated anatytic continuation

Continuing this process for as long as possible, a

surface-like configuration is obtained covering t "sheets"

of the complex plane, where t is the degree of the algebraic

function. To form the Riemann surface these sheets can be

joined together in the m o s t varied of ways. This m a y

involve connecting together two sheets which are separated

by several other sheets lying between them. Although such

a construction cannot be carried out in a t h r e e - d i m e n s i o n a l

space it is not difficult to give a p e r f e c t l y satisfactory

topological d e s c r i p t i o n of the process required. This

surface-like configuration is called the Riemann surface

of the m u l t i p l e - v a l u e d algebraic function. On the Riemann

 44

surface the entire domain of values of the algebraic

function is spread out in a completely single-valued manner

so that, on every one of the t copies of the complex plane

involved, every point is the bearer of one and only one

value of the function.

A m e t h o d for building Riemann surfaces is given in

appendix 4. This involves the use of cuts in the complex

plane and it may be helpful to say a w o r d about them at this

point. Let an algebraic function g(s) have r critical

points { a l , a 2 , . . , a r } . Suppose them to be joined to one

another and then to the point at infinity by a line L .

Any line joining critical points will be called a cut. Let

L denote the set of complex numbers defined by the line L.

We then have that the solutions of equation (3.3.37) define

a set of t "distinct" analytic functions {gl(s),~2(s) .... ,gt(s)}

in the cut plane C -i . Each of these functions can be

analytically continued, by standard procedures, across the

cut L . NOW it fellows from the fundamental principles

of analytic c o n t i n u a t i o n that if an analytic function

satisfies an algebraic equation in one part of its domain

of definition, it must satisfy that equation in every region

into w h i c h it is a n a l y t i c a l l y continued. We must therefore

have that:

(i) there are only t "distinct" analytic functions

which satisfy the defining algebraic equation in the cut plane

C-i,
(ii) each analytic continuation of any of these

analytic functions {~i(s) : i = 1,2, .... t) gives rise to an

 45

analytic function which also satisfies the defining algebraic

equation. It follows from this that the set of analytic

functions associated w i t h one side of the cut L must be a

simple p e r m u t a t i o n of the set of analytic functions

associated with the other side of the cut. Therefore by

identifying and suitably matching up c o r r e s p o n d i n g analytic

functions (via their sets of computed values) on opposite

sides of the cut L , one can produce an appropriate domain

on which a single analytic function may be specified which

defines a continuous single-valued m a p p i n g from this domain

into the complex plane. This function is of course the

algebraic function, conceived of as a single entity, and

the domain so constructed is its Riemann surface.

It is sufficient for the purposes of understanding this

book for the reader to know that a Riemann surface can be

constructed for any given algebraic function, on which its

values form a s i n g l e - v a l u e d function of positio n. Many

standard relationships and properties of analytic function

theory generalize, using the Riemann surface concept, to the

algebraic function case and, in p a r t i c u l a r the Principle of

the Argument holds on the Riemann surface; an extension of

the Principle of the Argument is developed in appendix 5.

The Riemann surface which is the domain of the

characteristic gain function g(s) will be called the

frequency surface or s-surface. When the o p e n - l o o p gain

matrix G(s) is m x m and has a corresponding characteristic

equation which is irreducible (i.e. the usual case in practice)

the frequency surface is formed out of m copies of the complex

 46

frequency plane or s-plane.

3.3-5 G e n e r a l i z e d root locus diagrams

The characteristic gain function g(s) is a function of

a complex variable whose poles and zeros are located on the

frequency surface domain. It is convenient to exhibit the

nature of g(s) by drawing constant phase and constant

magnitude contours of g(s) on the frequency surface. If

the computational method outlined in appendix 4 is used to

construct the surface then the superposition of constant

phase and m a g n i t u d e contours is clearly a simple process.

The frequency surface can be thought of as the set of all

possible closed-loop characteristic frequencies associated

with all possible values of the complex gain parameter g.

When the surface is c h a r a c t e r i z e d by constant phase and

magnitude contours of g(s) we have a direct correspondence

between a closed-loop characteristic frequency and an open-

loop gain, and since the surface is constructed from m

copies of the complex frequency plane, for each value of s

there are m corresponding characteristic gains.

From equation (3.1.4) we have

g(s) = -1 (3.3.38)

so that the variation of the c l o s e d - l o o p poles (characteristic

frequencies) with the real control variable k traces out

loci which are equivalent to the 180 phase contours of g(s).

Equation (3.3.38) is a direct generalization of the defining

equation for the single-loop root locus diagram. The 180

phase contours of g(s) are the m u l t i v a r i a b l e root loci i.e.

the variation of the closed-loop poles with the gain control

variable k. The fact that m u l t i v a r i a b l e root loci 'live'

 47

on a Riemann surface explains their c o m p l i c a t e d b e h a v i o u r

[9] as compared with the single-input, single-output case

where the root loci lie on a simple complex plane (a trivial,

i.e. one sheeted, Riemann surface). The m u l t i v a r i a b l e

root loci will sometimes be referred to as the c h a r a c t e r i s t i c

frequency loci.

Recall that in section 3.2 it was pointed out that the

characteristic equations for G(s) and S(g) are in general

different in that the equation for S(g) may contain factors

of s which are independent of g. These factors therefore

correspond to closed-loop poles which are independent of g,

or equivalently independent of the gain control variable k;

and, from the root locus point of view, these factors

correspond to degenerate loci each consisting of a single

point. The degenerate loci are therefore not p i c k e d out by

the 180 phase contours of g(s) on the frequency surface.

In practice the c h a r a c t e r i s t i c frequency loci are g e n e r a t e d

as the set of loci in a single copy of the complex frequency

plane traced out by the eigenvalues of S(g) as g traverses

the negative real axis in the gain plane. This approach

automatically picks out the d e g e n e r a t e loci. In common

with the classical root locus approach of Evans the characteristic

frequency loci are usually calibrated in terms of the gain

-i
control variable k=-g

3.3-6 Example of frg~uency surface and characteristic

frequency loci

As an i l l u s t r a t i v e example consider the general multi-

variable feedback configuration of figure 3 with a corresponding

open-loop gain m a t r i x
 48

~,.':':~-- Root Ioc

Cut

Figure 8. Sheet 1 of the frequency surface

.... Root loci

Cut

-I

-3

Figure 9. Sheet 2 of the frequency surface

 49

G(s) = 1 [s-i s]
1.25(s+i) (s+2) -6 s-2

The matrix is of order two and therefore the appropriate

surface will be constructed from two sheets of the complex

s-plane. The two sheets are shown characterized by

constant phase and magnitude contours of g(s) in figures

8 and 9. The cuts,identifiable by discontinuities in the

contours, are represented by thick black lines; and the

characteristic frequency loci, which are the 180 phase

contours of g(s), are identified by a diamond symbol.

The characteristic frequency loci indicate that variation

of the gain control parameter k, upwards from zero, causes

the system to experience stability, instability and stability

again. This phenomenon is clearly linked with the presence

of a branch point in the right half-plane (at s=~4).

Note that since we have completely characterized the

feedback configuration by its open-loop gain matrix there

are no unobservable or uncontrollable modes.

3.4 Characteristic frequency functions

The natural way to define the characteristic frequency

function s(g) is via the characteristic equation

V(s,g) ~ det[SIn-S(g) ] = 0 (3.4.1)

It is an algebraic function and the detailed study of the

characteristic gain function presented in the previous

section can be applied directly to it with the roles of s

and g reversed.

The Riemann surface which is the domain of the characteristic

frequency function will be called the ~ain surface or

@-surface. It is formed out of n copies of the complex

 50

gain plane or g-plane since there are n values of closed-

loop characteristic frequency (closed-loop poles) for every

value of g. The gain surface can be thought of as the

set of all possible open-loop characteristic gains of the

open-loop gain matrix G(s) associated with all possible

closed-loop characteristic frequencies. In a similar

fashion to the gain function g(s) it is convenient to

exhibit the behaviour of s(g) on the gain surface by

superimposing constant phase and magnitude contours of

s(g) onto the surface. Like g(s) the frequency function

s(g) has poles and zeros but their significance is quite

different.

3.4-1 Generalized Nyquist diagram

Each 'sheet' of a gain surface characterized by constant

phase and magnitude contours of s(g)is divided into regions

corresponding to left half-plane and right half-plane closed

-loop characteristic frequencies. Therefore given such a

calibrated surface one can see at a glance which values of

g (or equivalently k) correspond to stable closed-loop

poles. The boundary between stable and unstable regions

is clearly the ~90 phase contours of s(g). The ~90 phase

contours of s(g) are a natural generalization of the single-

loop Nyquist diagram and are called characteristic gain loci.

In practice the characteristic gain loci are generated

as the loci in the complex gain plane traced out by the

eigenvalues of G(s) as s traverses the so called Nyquist

D-contour in the s-plane. Suppose that we consider a

portion of the imaginary axis. We can then compute a

set of loci corresponding to the eigenvalues gl(j~),...,gm(j~)

( where in this context j = /--2i-- )

 51

in the following way:

(i) Select a value of angular frequency, say a

(ii) Compute the complex matrix G(j~ a)

(iii) Use a standard computer algorithm to compute the

eigenvalues of G(j~ a) , which are a set of complex

numbers denoted by {gi(J~a)} .

(iv) Plot the numbers {gi(J~a)} in the complex plane.

(v) Repeat with further values of angular frequency ~b,~c,

... etc., and join the resulting plots up into

continuous loci using a sorting routine based on the

continuity of the various branches of the characteristic

functions involved.

For the purpose of developing a g e n e r a l i z e d Nyquist

stability criterion in chapter 4 the Nyquist D - c o n t o u r is

traversed in the standard clockwise direction.

3.4-2 Example of ~ain surface and characteristic ~ain loci

As an illustrative example consider the open-loop

gain matrix considered in subsection 3.3-5 w h i c h has a

minimal state-space realization

=- 0.6

1 O.5

The system has two states and therefore the appropriate

gain surface will be c o n s t r u c t e d from two sheets of the

complex g-plane. The two sheets are shown c h a r a c t e r i z e d

by constant phase and m a g n i t u d e contours of s(g) in figures

i0 and ii. The characteristic gain loci, w h i c h are the

~9~phase contours of s(g), are denoted by a series of

crosses.
 52

Charocteristic
.... gain loci
mmmm Cut

Figure 10. Sheet 1 of the gain surface

Characteristic
gain loci

b C~%

Figure11. Sheet 2 of the gain surface

 53

~ Right. half
plane region
Left. half
plane region
. . . . . . . . Choract,erbtic
..... gain loci
Cut. between
branch points.

"l
-0"8~.533 19-2

Figure12. Sketch of figure 10 emphasizing right

half and left hatf-ptane regions

L
533 1<3,2

Figure13. Sketch of figure 11 emphasizing right

half and left half-ptane regions
 54

In figures 12 and 13 sketches of figures iO and ii are

made to emphasize the right and left half-plane regions of

closed-loop poles. From these sketches it is easy to

infer bounds on the gain control parameter k for stability.

Since g=-k -I, as we increase k positively from zero the

critical value of g moves from -~ along the real axis towards

the origin on each sheet. On sheet i, g is in a right half-

plane region for 1.25<k<2.5 while on sheet 2, for positive

k, g never moves into a right half-plane region. Therefore

the closed-loop system is stable for 0~k<1.25 and 2.5<k<~.

If k is increased negatively from zero the critical value

of g moves from ~ along the real axis towards the origin

on each sheet. On both sheets the value of g is in right

half-plane regions for -~<k~-1.875. Therefore for negative

k, which corresponds to positive feedback, the closed-

loop system is stable for -1.875<k~0.

If the calibrated gain surface is projected onto the

complex gain plane ~ , we have the normal representation

of the characteristic gain loci, plus the superposition of

contours representing both right half-plane and left half-

plane regions. Stability can now be predicted by considering

the right half-plane regions in relation to a single critical

1
point (-~+30). However, this presentation will in general

be difficult to comprehend because of the overlapping of

contours. Therefore, although the counting of encirclements

in the generalized Nyquist stability criterion [chapter 4]

is not fundamental to system stability as pointed out by

Saeks [i0~, it does afford the simplest method of predicting

 55

closed-loop stability in the gain plane ~ .

From a gain surface plot it is possible to determine

the closed-loop poles a n d hence the relative stability of

a closed-loop system. This is now i l l u s t r a t e d by finding

the dominant c l o s e d - l o o p poles for the example under

consideration with unity k. It is convenient for this

purpose to have the gain surface characterized by constant

real and imaginary contours as shown in figures 14 and 15.

Let the dominant closed-loop poles be

sd = ezj8
Then ~ is the smallest (in magnitude) negative real contour

that passes through any one of the critical, -I, values of

g, and 8 is the c o r r e s p o n d i n g imaginary contour. From

figures 14 and 15 we have

~- -0.05 and 8 = O

so that

sd ~ -0.05

By hand calculation the dominant c l o s e d - l o o p pole is

sd = -0.0528

Also from a gain surface it is possible to determine the

characteristic frequency loci. Apart from possible single-

point loci, the root loci are simply the values of s at which

the characteristic gain loci has a phase of 180 . Therefore

from a gain surface plot the root loci are determined by

the values of the constant contours as they cross the negative

real axis on each sheet. Similarly, from a calibrated

frequency surface, it is possible to determine the c h a r a c t e r i s t i c

gain loci from the values of constant contours as they cross

the imaginary axes.

 U}
v ~
c~
6 E
r~ .Q
cs
O~ c-
O
~ql L
0 m.
QJ E
U~ v
L
LL
 57

I ~ a l contours
labll~:l thus :

-0.9 -I'O -I'l -1'2 -#'3 -1'4 -f'5 Irnoginary contours

-oe 1.5 labelled thus : 0.5
Characteristic .......
gain |oci ........

-0.7 I'0- Cut

-06

-0.5
-0,4 0-5-
-0.3

4
-i.s ~ -,.o I -o:s..~f-o' ,!o 1.5

:2/ (lll :
0.9 ,'0 I"11~2 ~:3 ~.415
,o

-I.5-

Figure 15. Sheet 2 of the gain surface

References

Ill G.A. Bliss, "Algebraic Functions",

(reprint of 1933 original).
Dover, New York, 1966

[2] P.M. Cohn, "Algebra", vol. l, Wiley, London 1974.

[3] H.H. Rosenbrock, "State Space and Multivariable Theory",

Nelson, London, 1970.

[4] T. Kontakos, Ph.D. Thesis, University of Manchester, 1973.

[2] S. Barnett, "Matrices in Control Theory", Van Nostrand-

Reinhold, London, 1971.

[0] E. Hille, "Analytic Function Theory", Vol. 2, Ginn and Co.,

U.S.A., 1962.

[7] K. Knopp "Theory of Functions", Part 2, Dover, New York,

1947.

[8] G. Springer, "Introduction to Riemann Surfaces", Addison-

Wesley, Reading, Mass., 1957.

[9] B. Kouvaritakis and U Shaked, "Asymptotic behaviour of

root-loci of multivariable systems", Int. J. Control,
23, 297-340, 1977.

iO] R. Saeks, "On the Encirclement Condition and Its Generalization",

IEEE Trans. on Circuits and Systems, 22, 780-785, 1975.
 58

4. A generalized Nyquist stability

criterion

The Nyquist stability criterion [i] is one of the most

fundamental results in the theory of linear feedback systems

and its generalization to the multivariable case is of

great interest. Such a generalization was put forward

by MacFarlane[2] and u s e d as part of a technique called

the Characteristic Locus Method [3] for feedback systems

analysis and design, but no satisfactory proof was supplied.

The proof of a generalized Nyquist stability theorem was

undertaken by Barman and Katzenelson [4;5] but their

approach ignored certain key algebraic properties of the

quantities involved; it also leaned very heavily on the

use of cuts in the complex plane; this made their treatment

technically complicated, and obscured the essential simplicity

of the result. The purpose of this chapter is to give

a rigorous proof of a generalized Nyquist-like stability

criterion for the general feedback configuration based on

a fundamental result in complex variable theory: the Principle

of the Argument applied to an algebraic function defined on

an appropriate Riemann surface [ appendix 5].

Two Nyquist-like stability tests are in fact stated and

proved, the usefulness of each depending on how the subsystems

are characterized. The two statements of the criterion

are given in section 4.1 and proved in section 4.2.

4.1 Generalized Nyquist stability criterion

If the subsystems of the general feedback configuration

of figure 3 are each characterized by a state-space model

then the following statement of the criterion is applicable.

 59

Statement I. The general feedback configuration is closed-

loop stable if and only if:

(i) the net sum of anti-clockwise encirclements of the

critical point (-~+jO) by the set of characteristic gain

loci is equal to the number of right half-plane poles of G(s);

(2) the characteristic gain loci do not pass through the

critical point (-~+jO);

(3) the number of branches of the characteristic gain loci

passing through infinity is equal to the number of poles

of G(s) on the imaginary axis; and

(4) the eigenvalues of the A-matrix, of the open-loop

system S(A,B,C,D), which correspond to modes of the system

which are unobservable and/or uncontrollable from the point

of view of considering the input as that of the first sub-

system and the output as that of the hth subsystem, are

all in the left half-plane.

If the subsystems are completely characterized by

their transfer function matrices, or if it is known that

for each subsystem there are no unobservable and/or

uncontrollable modes in the right half-plane including the

imaginary axis, then the following statement of the criterion

applies.

Statement 2. The general feedback configuration is closed-

loop stable if and only if:

(i) bhe net sum of anti-clockwise encirclements of the

critical point (-~+jO) by the set of characteristic gain

loci is equal to the total number of right half-plane poles

of Gl(S),G2(s) .... , and Gh(S) ;

 60

(2) the c h a r a c t e r i s t i c gain loci do not pass through the

critical point (-kl--+jO); and

(3) t h e number of branches of the characteristic gain

loci passing through infinity is equal to the number of

poles of G(s) on the imaginary axis.

Note that if condition (2) and/or condition (3) do not

hold the closed-loop system has one or more poles on the

imaginary axis and is therefore not input-output stable

although the e q u i l i b r i u m state at the origin may be stable.

4.2 Proof of the ~ e n e r a l i z e d Nyquist stability criterion

In section 2.4 it was shown how the return-difference

operator corresponding to the break point shoWn in figure 4

is related to the open- and c l o s e d - l o o p characteristic

polynomials by the following expression.

detF(s) = CLCP(s) (4.2.1)

detF (~) OLCP (s-sT

This fundamental r e l a t i o n s h i p is the foundation on which

the proof of the g e n e r a l i z e d Nyquist stability criterion

is based.

The first stage in the proof is to consider the

eigenvalue equation of the r e t u r n - d i f f e r e n c e matrix F(s),

that is

det[fI m - F(s)] = O (4.2.2)

which in general, as for the characteristic equation of

G(s), can be expressed as a product of irreducible algebraic

equations of the form

t. t.-i
dio (s)f i i+dil(s)f'i i + .... + dit (s) = O
i
i=i,2,...,~ (4.2.3)

defining a set of algebraic functions {fi(s) : i=l,2,,..,i}

 61

Therefore, as in sub-section 3.3-3 where the p o l ~ and zeros

of G(s) are related to the poles and zeros of the characteristic

gain functions, the pole and zero polynomials of F(s) can

be related to the Ixle and zero polynomials of the algebraic

functions {fi(s):i=l,2,..,Z} as follows

PF(S) = eF(s) ~ d io
! (s)
i=l

and (4.2.4)

ZF(S) = eF(s) K di;tl (s)

i=l

By definition the open-loop characteristic polynomial

of the general feedback configuration is given by

OLCP(s)~det[SIn-A ]

=det[Sin~Al]det[sin2-A2] ... det[sIL n. -Ah ]

(4~2.5)
and it is easily shown [6] that

det[SIn[Ai] = PG. (S)Pd. (s) (4.2.6)

l l
where pG (s) is the pole polynomial for the transfer function
l
matrix Gi(s) and the monic polynomial Pdi(S) has as its
zeros the decoupling zeros of the ith subsystem associated

with that set of characteristic frequencies (eigenvalues)

of A i which correspond to modes of the ith subsystem which

are uncontrollable and/or unobservable. The open-loop

characteristic polynomial can therefore be expressed as

OLCP(s) = pG l(s)pG 2 ~s)...PGh(S)Pdl(S)Pd2(S)...pd (s)~,

(4.2.7)

The pole polynomial PG(S) for the open-loop gain matrix

G(s) is related to the pole polynomials of the subsystem

transfer functions through the relationship
 62

PG(s)Px(S) = PGl(S)PG2(s)...p~(s) (4.2.8)

where Px(S) has as its zeros those poles of GI(S),G 2 (s), ....

d Gh(S) which are lost when G(s) is formed [ 7] . The

zeros of Px(S) are in fact a subset of the unobservable and

uncontrollable modes of the system S(A,B,C,D) where the

input is that of the first subsystem and the output is that

of the hth subsystem. The complete set of unobservable

and uncontrollable modes of S(A,B,C,D) is the set of zeros

of the polynomial Pd(S) where

Pd(S) = Px(s)Pdl (S)Pd2 (s)...Pdh (s) (4.2.9)

The open-loop characteristic polynomial can therefore be

rewritten as

OIX2P(s) = PG(S)Pd(S) (4.2. iO)

or

OLCP(s) = P G ( S ) P x (s)pd I ( s ) p d 2 ( s ) ' ' ' p ~ ( s ) (4.2.11)

and if we combine expression (4.2.10 with the fundamental

relationship (4.2.1) we obtain

CLCP (s) = pd (s) pG (s) detF (s) (4.2.12)
detF (~)

N o w from the Smith-McMillan canonical form for F(s) we

have that
detF(s) = ~. ZF(S) (4.2.13)
PF (s)
where 8 is a scalar quantity independent of s; and from

the structure of F(s), equation (3.1.2), it is clear that

the monic polynomials ZF(S) and PF(S) will be of the same

order and hence that

detF(~) = 8 (4.2.14)

Therefore combining equations (4.2.12-14) the closed-loop

characteristic polynomial is given by

 83

ZF(S)
CLCP (s) = Pd(S)PG(S) (4.2.15)
P F (s)

If we now substitute from equations (4.2.4) and (3.3.35)

this expression can be rewritten as follows

CLCP(s) = Pd(S)eG(s) H b/lo(S) H d / (s)

i=l i=l iti
(4.2.16)

~d~o (s)
i=l

But using the eigenvalue shift theorem [ 8] , on equation

(3.1.2), we have that the characteristic functions

{fi(s) : i=i,2,... ,} of F(s) are related to the characteristic

gain functions {gi(s) : i=1,2...,} by

fi(s) = l+kg i(s) i=1,2 .... , (4.2.17)

and hence that the p o ~ p o l y n o m i a l s for both sets of

algebraic functions are identical that is

dio/(s) = ~ b~o(S) (4.2.18)
i=l i=l

and therefore from (4.2.16) we have

CLCP (s) = Pd(S)eG(s) H dit.l (s) (4.2.19)
i=l

Note that the polynomial H d I (s) is dependent on the

i=l it.I

gain control variable k.

The relationship (4.2.19) implies [ see section 2.3]

that the following conditions are necessary and sufficient

for closed-loop stability:

(a) eG(s) ~ O has only left half-plane roots;

(b) H dilt (s) = O has only left half-plane roots; and

i=l i

(c) Pd(S) = O has only left half-plane roots.

 64

The next step in the proof is to show how condition

(b) can be replaced by an encirclement condition similar

to that in the classical Nyquist criterion.

For the set of irreducible characteristic equations

associated with the return-difference operator F(s) there

is a corresponding set of Riemann surfaces on which the

appropriate characteristic algebraic functions {fi(s) : i=i,2,...

become single-valued, and mappings from these surfaces on

to a corresponding fi - plane are one-to-one and continuous.

Let us consider the jth equation of the set defined by

equations (4.2.3). The degree of the equation is tj and

therefore the corresponding Riemann surface ~ f . is formed

3
by piecing together t. copies of the complex s-plane, C
3
Suppose now that a Nyquist D-contour, as shown in figure 16,

is drawn on each of the tj copies of C before they are

pieced together to f o r m ~ f
. Then when the surface is
3
formed the set of Nyquist D-contours combine to form a set

of closed Jordan contours [ 9] enclosing right half-plane

regions o f ~ f . . The extended Principle of the Argument

3
[appendix 5] can then be applied to each right half-plane

region on ~ f j . Therefore for a particular right half-

plane region, not necessarily simply connected but with a

boundary made up from Nyquist D-contours, we have that the

difference between the number of zeros and poles of the

algebraic function fi(s) in the region, is equal to the

number of clockwise encirclements of the origin in (the

complex f-plane) by the image of the boundary curves, under

fj(s) for that particular region. If we therefore consider

 65

all the right half-plane regions o n ~ f and apply the

3
extended Argument Principle to each, we have that

N(fj , O) = Zf. - Pf. (4.2.20)

3 3
where:
(i) N(f@,O) is the net sum of clockwise encirclements

of the origin in C by the set of image curves, under fi (s) I

of the set of curves formed on ~ f . by piecing together
3
the appropriate set of Nyquist D-contours when f o r m i n g ~ f 3.;
(ii) Zf. is the number of right half-plane zeros of fj (s); and
3
(iii) Pf. is the number of right half-plane poles of fj (s).
3

Trn (s)

Radius r where
S -plane ) ..,......~ r-,-oobut#00
')///4/ / _
f
Re (s)

x Poles
O Zeros
z~ Branch points

Figure 16. Theoretical Nyquist D-contour

 66

Note that the indentatiorson the D-contour are

necessary only from a theoretical viewpoint in the development

of the Extended Principle of the Argument [ appendix 5] and

in practice the D-contour is taken as the imaginary axis.

Condition (b) for closed-loop stability is equivalent

to saying that there are no zeros of

{fi(s) : i=1,2,...,~} (4.2.21)

in the right half-plane or on the imaginary axis, and can

therefore be replaced by

Zf. = 0 i=1,2,...,~
l
or (4.2.22)

N(fi,O) = -Pf. i=1,2,...,

l
plus the condition that the algebraic functions (4.2.21)

have no zeros on the imaginary axis. Conditions (4.2.22)

imply and are implied by

N(fi,O) = - ~ P (4.2.23)
i=l i=l

so that the necessary and sufficient conditions for closed-

loop stability can be rewritten as

(a') eG(S)=O has no right half-plane roots;

(b') e (s)=O has no roots on the imaginary axis;

G

(c') i=iZN(fi'O) =-i~l P % ;

(d ) {fi(s): i=l,2,..,Z } have no zeros on the imaginary

axis; and

(e') Pd(S) has only left half-plane zeros.

Now,as shown by equations (3.3.35) and (4.2.18),e(s)

together with the pole polynomials for the set of characteristic

functions {fi(s): i=1,2,...,} make up the pole polynomial

for G(s). This leads us to consider combining conditions

 67

(a') and (c r) into the single equivalent condition

N ( f i ' O ) = -PG (4.2.24)

i=l
where PG is the number of right half-plane poles of G(s).

The n e c e s s i t y and s u f f i c i e n c y of c o n d i t i o n (4.2.24) for

closed-loop stability when conditions (b'),(d I) and (e')

are satisfied is p r o v e d as follows.

From e q u a t i o n s (3.3.35) and (4.2.18) we have that

PG = e + E Pfi (4.2.25)
i=l

where e is the n u m b e r of right half-plane zeros of eG(s),

and we also know that

Z N(fi,O) = Z Zf - Z Pf (4.2.26)
i=l i=l ii=l i

so that c o m b i n i n g these two e v p r e s s i o n s we have

Z N(fi,O) = ~iZfi + e -- PG (4.2.27)
i=l i

where i=ZlZfi'e and PG are all p o s i t i v e integers, or zero.

To e s t a b l i s h the n e c e s s i t y of c o n d i t i o n (4.2.24)

suppose that

N(fi'O) ~ -PG
i=l

then from e q u a t i o n (4.2.27) this implies that

~Zf ~O and/or e ~ 0
i=l i

and hence we conclude that condition (4.2.24) is n e c e s s a r y

for closed-loop stability.

For s u f f i c i e n c y suppose that

Z N(fi,O) = -PG
i=l

then from e q u a t i o n (4.2.27) we m u s t have that

iZl= Zf i = e = O
 68

and hence the system is closed-loop stable. Thus the

sufficiency of condition (4.2.24) is established.

We have therefore shown that the following conditions

are necessary and sufficient for closed-loop stability:

(a") Z N(fi'O) = - PG ;
i=l
(b") [fi(s) : i=1,2 .... ,} have no zeros on the imaginary

axis;

(c") eG(s) has no zeros on the imaginary axis; and

(d") Pd(S) has only left half-plane zeros.

From equation (4.2.17) the image curve sets in ~ of

the Nyquist D-contour set mapped under fi(s) and kgi(s) are

simply related by a unit shift in C The stability

condition (a") can consequently be replaced by

Z
N (kgi, -i) = -PG (4.2.28)
i=l

where N(kgi,-l) is the net sum of clockwise encirclements

of the point (-l+j~ in C by the characteristic gain loci

scaled by k. As in the classical Nyquist criterion the

scaling by k is avoided by counting the number of encirclements

that the characteristic gain loci make of the critical point

1
(-~ + jO), and hence replacing condition (4.2.28) by

1
N(gi' -k) = -PG (4.2.29)
i=l

From equations, (4.2.17) we also have that condition

(b") is equivalent to
1
gi(Jw) ~-~ , i=i,2, .... i (4.2.30)

and in practice this corresponds to the characteristic

gain loci not passing through the critical point (-~+jO).

Condition (c") can also be replaced by a more practical

 69

test. The p o l e p o l y n o m i a l for G(s) is given, equation

(3.3.35), as

H b/ (s) (4.2.31)
PG(S) = eG(s) i=l Io

where b I (s) is the pole p o l y m o m i a l for the jth c h a r a c t e r i s t i c

]o
gain f u n c t i o n gj(s). Therefore it is clear that eG(s)

will have no zeros on the i m a g i n a r y axis if, and only if,

the number of i n f i n i t e branches of the c h a r a c t e r i s t i c gain

loci is equal to the n u m b e r of poles of G(s) on the i m a g i n a r y

axis. Note that for a b r a n c h of the loci going off to

infinity there will also be a b r a n c h returning from i n f i n i t y

and care should be taken not to count this as two i n f i n i t e

branches.

The n e c e s s a r y and s u f f i c i e n t conditions for c l o s e d -

loop s t a b i l i t y can t h e r e f o r e be r e d u c e d to:

1
(i) Z N(gi' -k ) = - PG;
i=l

(2) the c h a r a c t e r i s t i c gain loci do not pass through the

critical p o i n t (-kl---+jO);
(3) the n u m b e r of b r a n c h e s of the c h a r a c t e r i s t i c gain loci

passing t h r o u g h infinity is equal to the n u m b e r of poles of

G(s) on the i m a g i n a r y axis; and

(4) Pd(S) has only left h a l f - p l a n e zeros.

This c o m p l e t e s the p r o o f of s t a t e m e n t 1 of the g e n e r a l i z e d

Nyquist s t a b i l i t y criterion.
 70

If the s u b s y s t e m s are c o m p l e t e l y characterized by

their t r a n s f e r function matrices then t h e i r state-space

descriptions will each be o b s e r v a b l e and c o n t r o l l a b l e so

that

p d I (s) = p d~ (s) = .... = p d h(s) = i (4.2.32)

and h e n c e

Pd(S) = Px(S) (4.2.32)

Condition (4) of s t a t e m e n t 1 of the N y q u i s t criterion is

then e q u i v a l e n t to Px(S) having only left h a l f - p l a n e zeros.

This situation also arises if the subsystems are c h a r a c t e r i z e d

by their state-space descriptions and we have the a d d i t i o n a l

information that each subsystem has no u n o b s e r v a b l e or

uncontrollable modes in the right h a l f - p l a n e or on the

imaginary axis. In these situations conditions (i) and

(4) can be combined to give a c r i t e r i o n which needs no

information about decoupling zeros.

To s h o w this we w i l l first of all p r o v e that when

(i) Pd(S) = Px(S) (4.2.34)

or

(ii) Pd(S) = Px(S)P~ (S)Pd 2( s ) .... p d h (s)

(4.2.35)

where the zeros of {pd (s) : i=1,2, .... h} are

all in the left h a l f - p l a n e ,

conditions (a'), (c I) and (e I ) are e q u i v a l e n t to

h
N(fi,O) = -Z PG (4.2.36)
i=l i=l
i
where PG~ is the n u m b e r of r i g h t h a l f - p l a n e p o l e s of Gi(s).

The n e c e s s i t y and s u f f i c i e n c y of c o n d i t i o n (4.2.3 6 ) for

 7~

closed-loop stability when conditions (b') and (d') are

satisfied is proved as follows.

From equations (4.2.9),(4.2.18) and (4.2.31) we have

that

h
PG = e + Z Pfi + P (4.2.37)
i=l ~ i=l x

where Px is the number of right half-plane zeros of Px(S),

and combining this with equation (4.2.26) we have that

h
i=iZN(fi,O) = i~l Zfi + e + Px - i~iPGi (4.2.38)

Note also that

Pd(S) = Px(S)Pd] (s) .... Pdh(S)

so that if Pd denotes the number of right half-plane

zeros of Pd(S) we have that

Pd = Px (4.2.39)

and equation (4.2.38) becomes

h
Z N ( f i , O ) = ~ l Z f i + e + Pd - 2 PGi (4.2.40)
i=l i i=l

To establish the necessity of condition (4.2.36)

suppose that
h
7 N(fi,O ) ~ -
i=l il~--1PGi
then from equation (4.2.40) this implies that

Z Zfi ~ O, or e ~ O, or Pd ~ 0
i=l

or any combination of these and thus that the system will

be closed-loop unstable. Hence we conclude that condition

(4.2.36) is necessary for closed-loop stability.

For sufficiency suppose that

h
Z N(fi,O ) = - __ZIPGi
i=l i
 72

then from e q u a t i o n (4.2.40) we m u s t have that

i~iZfi = e = Pd = O

and hence the s y s t e m is c l o s e d - l o o p stable. Thus the

sufficiency of c o n d i t i o n (4.2.36) is e s t a b l i s h e d .

We have t h e r e f o r e shown that w h e n e i t h e r condition

(4.2.34) or c o n d i t i o n (4.2.35) is s a t i s f i e d the f o l l o w i n g

are n e c e s s a r y and s u f f i c i e n t for c l o s e d - l o o p stability.

h
(a"~ Z N(f i O) = - Z PG ;
i=l ' i=l

(b"') {fi(s) : i = 1 , 2 , . . . , } have no zeros on the i m a g i n a r y

axis; and

(c i'') eG(s) has no zeros on the i m a g i n a r y axis.

But we have already shown, in p r o v i n g statement 1 of the

s t a b i l i t y criterion, that

Z N(fi,O) ~ Z N(gi,-!)
k
i=l i=l

that c o n d i t i o n (b"') is e q u i v a l e n t to the c h a r a c t e r i s t i c

gain loci not p a s s i n g through the c r i t i c a l point (-~+jO);

and that c o n d i t i o n (c"') is e q u i v a l e n t to the n u m b e r of

infinite branches of the c h a r a c t e r i s t i c gain loci b e i n g

equal to the n u m b e r of poles of G(s) on the i m a g i n a r y axis.

This t h e r e f o r e completes the p r o o f of s t a t m e n t 2 of the

generalized Nyquist stability criterion.

4.3 Example

To i l l u s t r a t e the s t a b i l i t y criterion consider the

example already u s e d in s u b - s e c t i o n s 3.3-5 and 3.4-2, where

the g e n e r a l feedback configuration is c h a r a c t e r i z e d by its

open-loop gain m a t r i x
 73

0s l
1.25(s+i) (s+2)
i s i-6s s-2
1
By d e f i n i t i o n the s y s t e m has no u n o b s e r v a b l e or u n c o n t r o l l a b l e

modes and by v i r t u e of the p r o b l e m statement we can c o n s i d e r

the c o n f i g u r a t i o n as c o n s i s t i n g of just one subsystem Gl(S)=G(s).

Therefore e i t h e r statement 1 or s t a t e m e n t 2 of the s t a b i l i t y

criterion is d i r e c t l y applicable.

The pole p o l y n o m i a l for G(s) is

PG(S) = (s+l) (s+2)

which has no zeros in the right h a l f - p l a n e or on the i m a g i n a r y

axis and t h e r e f o r e closed-loop stability is e n s u r e d if the

net sum of a n t i - c l o c k w i s e encirclements of the c r i t i c a l

point (-~+jO) by the c h a r a c t e r i s t i c gain loci is zero, and

if the c h a r a c t e r i s t i c gain loci do not pass through the

critical point and have no infinite branches. The c h a r a c t e r i s t i c

gain loci are shown in figure 17 from w h i c h the f o l l o w i n g

stability c o n d i t i o n s are obtained.

1
(i) For -~< -~ < -0.8 there are no e n c i r c l e m e n t s of, or

passage through, the c r i t i c a l point and thus the c l o s e d -

loop system is stable for O ~k<1.25.

1
(ii) For -~ = -0.8 the c h a r a c t e r i s t i c gain loci pass t h r o u g h

the c r i t i c a l p o i n t and t h e r e f o r e the system is c l o s e d -

loop u n s t a b l e for k = 1 . 2 5

(iii) For - O . 8 < - ~ < - O . 4 there is one c l o c k w i s e encirclement

of the c r i t i c a l point and t h e r e f o r e the s y s t e m is c l o s e d -

loop u n s t a b l e for 1 . 2 5 < k < 2 . 5

 74
Im
60 __
",=0-5
jl.O0. ~=1-0

]o.5o

~=2.0.
S ---,,-.- 00

~=0.05. f 60
60

-I.00 -0-50 ~= 20'0 IRe

~=-0.05 60 ~=1-0
~ =-2.,

-]o.5o

~ j l O=_o.s
" 0 = -I.0

Figure17. C h a r a c t e r i s t i c gain loci

 75

1
(iv) For -~ = -0.4 the characteristic gain loci pass through

the critical point and therefore the system is closed-loop

unstable for k=2.5.

(v) For -0.4<-~<0 there are no encirclements of, or passage

through, the critical point and thus the closed-loop system

is stable for 2.5<k<~.

1
(vi) For -~=O the characteristic gain loci pass through

the critical point and therefore the system is closed-loop

unstable for k =~.

(vii) For 0<-~<O.533 there are two clockwise encirclements

of the critical point and therefore the system is closed-

loop unstable for -~<k<-1.875.

(viii) For -ki--=O.533 the characteristic gain loci pass through

the critical point and therefore the system is closed-loop

unstable for k=-1.875.

(ix) For O . 5 3 3 < - ~ there are no encirclements of, or

passage through, the critical point and thus the closed-

loop system is stable for -1.875<k~0.

Note that the conditions where k is negative correspond

to positive feedback.

References

i] H. Nyquist, "The Regeneration Theory", Bell System Tech. J.,

ii, 126-147, 1932.

2] A.G.J. MacFarlane, "Return-difference and return-ratio matrices

and their use in analysis and design of multivariable
feedback control systems", Proc. IEE, 117, 2037-2049, 1970.

~3] A.G.J. MacFarlane and J.J. Belletrutti, "The characteristic

Locus Design Method", Automatica, 9, 575-588, 1973.
 76

4] J.F. Barman and J. Katzenelson, Memorandum ERL-383, Electronics

Research Laboratory, College of Engineering, Univ. of California r
Berkeley, 1973.

5] J.F. Barman and J. Katzenelson, "A generalized Nyquist-type

stability criterion for multivariable feedback systems",
Int. J. Control, 20, 593-622, 1974.

[6] A.G.J. MacFarlane and N. Karcanias, "Poles and zeros of linear

multivariable systems: a survey of the algebraic, geometric
and complex variable theory", Int. J. Control, 24, 33-74,
1976.

7] C.A. Desoer and W.S. Chan, "The Feedback Interconnection of

Lumped Linear Time-invariant Systems", J. Franklin Inst.,
300, 335-351, 1975.

8] A.G.J. MacFarlane, "Dynamical System Models", Harrap,

London, 1970.

~9] E. Hille, "Analytic Function Theory", Vol. i, Ginn and Co.,

U.S.A., 1959.
 77

5. A generalized inverse Nyquis t stability criterion

In this chapter a g e n e r a l i z a t i o n of the inverse Nyquist

stability criterion [i] for single-input single-output

feedback systems is developed for the general feedback

c o n f i g u r a t i o n which is complementary to the exposition

of the g e n e r a l i z e d Nyquist stability criterion p r e s e n t e d

in chapter 4. The development is based on an association

of the o p e n - l o o p gain m a t r i x G(s) with a set of inverse

characteristic gain functions, and a corresponding set of

inverse Nyquist diagrams which will be termed the inverse

characteristic gain loci.

5.1 Inverse c h a r a c t e r i s t i c , ~ain functions

A main feature of the g e n e r a l i z a t i o n of N y q u i s t ' s

stability criterion, given in chapter 4, is the association

of a set of algebraic functions - the characteristic gain

functions - with a square transfer function m a t r i x G(s)

by means of the characteristic equation

A(g,s) = d e t [ gl m - G(s) ] = O (5.1.1)

-i
If G(s) has normal rank m then its inverse function G(s)

exists, and has a corresponding characteristic equation

given by

A(g,s) = det[gI m G(s) -I] = O (5.1.2)

If A(g,s) is regarded as a polynomial in g with coefficients

which are rational functions of s then equation (5.1.2)

defines a set of algebraic functions {gi(s) : i=1,2,...,}

w h i c h will be called the inverse characteristic gain functions.

If A(g,s) is also irreducible over the field of rational

 78

functions in s, then the inverse characteristic gain function

g(s), like g(s),has as its domain an appropriate Riemann

surface formed out of m copies of the complex s-plane

suitably joined together. In fact, because the eigenvalues

of a matrix are reciprocal to the eigenvalues of the matrix

inverse

g(s) = 1 (5.1.3)
g(s)

and therefore g(s) and g(s) have the same branch points,and

hence the same Riemann surface domains.

The inverse characteristic gain function is the foundation

on which the generalized inverse Nyquist stability criterion

is based.

5.2 Pole-zero relationships

In this section a number of pole-zero relationships

are derived which will be used in the proof of the generalized

inverse Nyquist stability criterion, section 5.4.

Because the eigenvalues of a matrix are the reciprocal

of the eigenvalues of the matrix inverse the poles of the

set of inverse characteristic gain functions are simply

the zeros of the characteristic gain function set, and vice

versa. The pole and zero polynomials of gi(s), denoted

by pg*(s) and Zg*(s), are therefore expressible as

pg.* (s) = z~=.(s) (5.1.4)

1 1

and

~ (s) = pg (s) (5.1.5)

1 1
i=l,2,...,i

From sub-section 3.3-2 it is clear that for an open

-loop gain matrix G(s) there exists a canonical form, the

 79

Smith-McMillan form [2 ] , such that

G(s) = HG(S)MG(S)JG(S) (5.1.6)

where HG(S) and JG(S) are both mxm unimodular matrices and

the Smith-McMillan form MG(S) is given by

MG(S) = diag [ ~l
(s)1(s) ~2
(s)e2(s)-' ~m--~
)em(s) ] (5.1.7)

Consequently the poles and zeros of G(s) are defined as

m
PG(S) = H ~i(s) (5.1.8)
i=l

and
m
ZG(S) = H ei(s) (5.1.9)
i=l

Now if G(s) is of normal rank m equation (5.1.6) can be

inverted to give
G(s) -1 = JG(S)-IMG(S)-IHG(S) -I (5. i. lO)

which using the elementary transformation matrix E, of order

m, and given by
O O ... 0 1

O O ... 1 O

E = : . = E
-i (5.1.11)

O 1 ... O O

1 O ... O 0

can be rewritten as
-i -i (5.1.12)
G(s) = JG (s) -IEEM G (s) -IEEHG (s)

or
G(s) -I = H~(s)M6(s)J~(s) (5.1.13)
 80

where
H~(s) = JG(S)-iE (5.1.14)
-i
J~Cs) = EHG(S) (5.1.15)

and
M~(s) = EMG(S)-IE

[ ~m (s) ~m-i (s) . ~I (s) ] (5.1.16)

= diag e~-~) ' em_l(S) ' " "' el(S )

which is the Smith-McMillan form for G(s) -I. The pole-

zero definitions for a transfer function matrix [sub-
section 3.3-2] can now be applied to G-I(s) with the result
that
m
p~(s) -- i=iH s.1(s) = zG(s) (5.1.17)

and
m
z~(s) = H ~i(s) = PG(S) (5.1.18)
i=l
where p*(s)G and z~(s) denote the pole and zero polynomials

of G(s) -i respectively.
In sub-section 3.3-3 it is shown that the pole and
zero polymomials of G(s) are related to the poles and zeros
of the characteristic gain function set {gi(s) :i=1,2,...,} by

PG(S) = eG(s)i=l
K b ti(s) = eG(s)i~iPgi (s) (5.1.19)

and

ZG(S) = eG(s) K b ! (s) = eG(s) ~ z (s) (5.1.20)

i=l iti i=l gi

If the pole and zero polynomials of the complete characteristic

gain function set are denoted by pg(S) and Zg(S) respectively,
and similarly for the inverse characteristic gain function set
 81

using pg*(s) and z*(s),

g then relationships (5.1.19) and

(5.1.20) can be combined with relationships (5.1.17) and

(5.1.18) to give

PG(S) = z~(s) = eG(S)pg(S)=eO(s)z~(s) (5.1.21)

and

ZG(S) = P*(s)=eG(S)Zg(S)G = eG(s)p~(s) (5.1.22)

These pole-zero relationships will be used in the proof of

the inverse Nyquist stability criterion presented later.

5.3 Inverse characteristic ~ain loci - generalized

inverse Nyquist diagrams

The inverse characteristic gain loci are the loci in

the complex plane traced out by the reciprocal of each of

the eigenvalues of the open-loop gain matrix G(s) as s

traverses the Nyquist D-contour in the standard (clockwise)

direction. The algorithm for computing the loci follows

that given for the characteristic gain loci, in sub-section

3.4-1, with the modification that at step (iv) the reciprocal

of the numbers gi(jwa) are plotted in the complex plane.

We are now in a position to state and prove the

generalized inverse Nyquist stability criterion.

5.4 Generalized i n v e r s ~ Nyquist stability criterion

If the subsystems of the general feedback configuration

of figure 3 are each characterized by a state-space model

then the following statement of the criterion is applicable.

Statement i. The general feedback configuration is closed-

loop stable if and only if:

(la) the net sum of anti-clockwise encirclements, of the

critical point (-k+jo), by the set of inverse characteristic

 82

gain loci, minus the net sum of anti-clockwise encirclements,

of the origin, by the set of inverse characteristic gain loci,

is equal to the number of right half-plane poles of G(s);

(2) the inverse c h a r a c t e r i s t i c gain loci do not pass

through the critical point (-k+jo) ; and

(3a) the number of branches of the inverse characteristic

gain loci that pass t h r o u g h the origin is equal to the number

of poles of G ( s ) o n the imaginary axis; and

(4) the eigenvalues of the A- matrix, of the open-loop

system S(A,B,C,D), which correspond to modes of the system

which are unobservable and/or uncontrollable from the point

of v i e w of considering the input as that of the first subsystem

and the output as that of the hth subsystem, are all in the

left half-plane.

Alternatively conditions (la) and (3a) m a y be replaced

by:

(ib) the net sum of anti-clockwise encirclements of the critical

point (-k+jo) by the set of inverse characteristic gain loci

is equal to the n u m b e r of right half-plane zeros of G(s);

(3b) the number of branches of the inverse c h a r a c t e r i s t i c

gain loci passing through infinity is equal to the number of

zeros of G(s) on the imaginary axis.

If the subsystems are completely c h a r a c t e r i z e d by their

transfer function matrices, or if it is known that for each

subsystem there are no unobservable and/or uncontrollable modes

in the right h a l f - p l a n e including the imaginary axis, then

the following statement of the criterion applies.

Statement 2 The general feedback configuration is closed-loop stable

 83

if and only if:

(la) the net sum of anti-clockwise encirclements of the

critical point (-k+jo), by the set of inverse charactezistic

gain loci, minus the net sum of anti-clockwise encirclements,

of the origin, by the set of inverse c h a r a c t e r i s t i c gain loci,

is equal to the total number of right half-plane poles of

Gl(S),G2(s), .... , and Gh(S);

(2) the inverse characteristic gain loci do not pass through

the critical point (-k+jo); and

(3a) the number of branches of the inverse characteristic gain

loci that pass through the origin is equal to the number of

poles of G(s) on the imaginary axis.

Alternatively condition (3a) m a y be replaced by condition

(3b), as in statement 1 of the criterion, and if the subsystems

are square, i.e. have the same number of outputs as inputs,

then condition (la) m a y be replaced by:

(ib) the net sum of anti-clockwise encirclements of the

critical point (-k+jo) by the set of inverse characteristic

gain loci is equal to the total number of right half-plane

zeros of Gl(S), G2(s) . . . . , and Gh(S).

Note that if condition (2) and/or condition (3a)/(3b) do

not hold then the closed-loop system has one or m o r e poles

on the imaginary axis and is therefore not input-output

stable although the e q u i l i b r i u m state at the origin may be

stable,

For strictly proper s y s t e m ~ t h a t is, ones in which the

system D-matrix is zero, the inverse c h a r a c t e r i s t i c gain loci

 84

approach infinity as s approaches infinity, and s=~ is

a pole of the inverse function. Therefore, in practice, it

is n e c e s s a r y to traverse the whole of the D-contour, not

just the imaginary axis, in order to obtain closed curves.

In this way the net sum of encirclements of the critical

point and the origin by the inverse characteristic gain loci

can be obtained. As an example, the inverse characteristic

gain loci for

G(s) = 1 (5.4.1)
3
(s+l)

are shown in figure 18.

The traversal of the D - c o n t o u r off the imaginary axis,

however, is not n e c e s s a r y if we use conditions (la), rather

than (ib), in both statements of the stability criterion.

This is because the extra loci, corresponding to the circular

part of the D-contour, encircle the origin and the critical

point the same number of times, and thereby cancel each other.

A useful rule, when using the (la) conditions, is therefore

to join up the corresponding loose ends of the loci via a

large semi circle, and to forget any extra encirclements that

m a y actually exist. W h e t h e r the ends are joined via the

right half-plane or left half-plane is fixed by the fact that

the mapping under g(s), from the set of Nyquist curves formed

on the Riemann surface of g(s) (by piecing together an appropriate

w
set of Nyquist D-contours) to the g-plane, is conformal. This

means that if we imagine two people, X and Y, X walking along

the domain curve, and Y walking along the corresponding image

curve, such that at each step X defines the corresponding

image point, then if x turns to his right, Y will turn to hi__~

right.
 85

*I~h
20~

~j_ plane "~x~=.2. 4 I0

$=0
J
- 3;0 - 2~0

-20.

(o)

I.O

/ -~
,o( o,:: f.,~ = 6 o
-0.5 0.5 l.~l
~ I ~ ,03

(b)

Figure18. Inverse charact.eris!,ic gain

Loci for G(s)= 1 / (s.l) 3
((:1)small area around origin
(b) large a r e a
 86

5.5 Proof of gene!alize d inverse Nyquist stabilit Y criterion

For the general feedback configuration of figure 3 the

closed-loop transfer function matrix R(s) is given by the

relationship

R(s) = kG(s) [ Im4kG(s)]-i (5.5.1)

which after inverting becomes

-i -1 (5.5.2)
R(s) = 1 G(s) + I
m

If we now denote the set of algebraic functions defining the

eigenvalues of R(s) -I by {~i(s) :i=l,2,...,Z} we can apply

the eigenvalue shift theorem[3] to equation (5.5.2) with

the result that

~. (s) = 1 gi(s)+ 1 (5.5.3)

i=I,2, . . . ,

Then if the poleand zero polynomials for the complete set of

algebraic functions {ri(s):i=l,2,...,Z} are denoted by the

monic polynomials Pr* (s) and *

Zr(S) respectively, we have that

~r(S) = ~g(s) (5.5.4)

Now post-multiplying equation (5.5.2) by kG(s) we have

MR (s)-lG (s) = Im+kG (s)

(5.5.5)
= r(s)

the return difference matrix [4] , and taking determinants of

this we obtain
det F(s) = k m det[ s -l]
det[G(s) -I]

=y z~(s) * (5.5.6)
g( )
 87

where y is a scalar constant independent of s. But

det F(s) = S ~(s) = 8 zf(s) (5.5.7)

PF{S) pf ('s)

where pf(s) and zf(s) are respectively the monic pole and

zero polynor~als for the set of algebraic functions {fi(s):i=1,2,...,}

and therefore

(5.5.8)

and

z f(s) = z*(s)
r @
p*(s)
)
pf(s) (5.5.9)

Now from equations (4.2.18) and (5.1.21)

pf(s) = pg(S) = z*(S)g (5.5.10)

and hence combining this with equations (5.5.4) and (5.5.9)

we obtain

Izf(s) = zr *(s)I (5.5.11)

In chapter 4, equation (4.2.19), it was shown that the

closed-loop characteristic polynomial can be given by

CLCP(s) = Pd(S)e~s) R d ! (s) (5.5.12)
i=l iti

or equivalently, u s i n g the notation presented in this chapter,

CLCP (s) = pd (s) eG(s) zf(s) (5.5.13)

and therefore combining this with equation (5.5.11) we have

CLCP(s) = Pd(s)es(S) Zr(S) ] (5.5.14)

This implies that the following conditions are necessary and

sufficient for closed-loop stability:

(a) eG(s) = 0 has only left half-plane roots;

(b) Zr(S) = 0 has only left half-plane roots; and

(c) Pd(S) = 0 has only left half-plane roots.

 88

Suppose now that a Nyquist D-contour, as shown in figure

16, is drawn on m copies of the complex plane before they are

pieced together t o f o r m t h e Riemann s u r f a c e s {I~.:i=1,2,...,}

1
on which the {ri(s):i=l,2,...,~} are defined. Let us consider

the jth s u r f a c e ~ j corresponding to ~j(s). When the surface

is formed the set of Nyquist D-contours combine to form a set

of closed Jordan contours [5] enclosing right half-plane

regions o n e * . The extended Principle of the Argument

r.]

[appendix 5] can be applied to each right half-plane region

On~r* Therefore for a particular right half-plane region,

3
not necessarily simply connected but with a boundary made up

from Nyquist D-contours, we have that the difference between

the number of zeros and poles of the algebraic function rj(s)

in the region, is equal to the number of clockwise encirclements

of the origin in ~ (the complex r-plane) by t h e image o f t h e

boundary curves, under rj(s), for that particular region.

I f we t h e r e f o r e consider all the right half-plane regions and

apply the extended Argument Principle to each, we have that

N(rj, O) = Z~j - P* (5.5.15)

rj

where:

(i) N(r~,O) is the net sum of clockwise encirclements of the

J
origin in C by the set of image curves, under ~j(s~ of the

set of curves formed on ~ * by p i e c i n g together the appropriate

rj
set of Nyquist D-contours w h e n f o r m i n g ~ . ~
3
w
(ii) Z~. is the number of right half-plane zeros of rj (s);
3
 89

and

(iii) p* is the number of right half-plane poles of r.] (s).

r.
3
If we now consider the corresponding inverse characteristic

gain function gj(s) in the same way, we find that

N(gj , O) = Z~j - P~j (5.5.16)

where:
(i) N(gj,O) is the net sum of clockwise encirelements of

the o r i g i n in C by the inverse characteristic gain loci

corresponding to gj(s);

(ii) ~j is the number of right half-plane zeros of gj(s); and

(iii) ~ j is the number of right half-plane poles of gj(s).

Equations (5.5.15) and (5.5.16) can be combined to give

N(rj,O)-N(gj,O) = Z, - P, - Z, + P, (5.5.17)
rj rj gj gj

which using equation (5.5.4) becomes

N(rj,O) - N(gj,O) = Z~j - Z~j (5.5.18)

and if we consider the complete set of algebraic functions

* {*
{ri(s):i=l,2,...,} and the set gi(s) :i=l,2,...,Z}

we have
Z , ,
Z N(ri,O)- Z N(gi,O) = Z Z* - Z Z* (5.5.19)
i=l i=l i=l ri i=l gi

Now condition (b) for closed-loop stability is equivalent

to saying that there are no zeros of {ri(s):i=l,2,...,}

in the right half-plane or on the imaginary axis, and can

therefore be replaced by

N(ri,O) - N(gi,O) = - ~ Z* (5.5.20)

i=l i=l i=l gi
 90

plus the c o n d i t i o n t h a t {ri(s):i=l,2,...,} have no zeros

on t h e imaginary axis. The necessary and sufficient

conditions for c l o s e d - l o o p stability can t h e r e f o r e be

rewritten as

(ai ) eG(s) = O has no right h a l f - p l a n e roots;

(b! ) eG(s) = O has no roots on the i m a g i n a r y axis;

. Z ,
(c ! ) Z N(ri,O) - Z N(gi,O) = - Z Z* ;
i=l i=l i=l gi

(d I ) {r* (s):i=i,2,...,} have no zeros on the i m a g i n a r y

i
axis; and

(e I ) Pd(S) has only left h a l f - p l a n e zeros.

The p o l e - z e r o relationship (5.1.21) n o w leads us to c o n s i d e r

combining c o n d i t i o n s (a')and(c') into the s i n g l e e q u i v a l e n t condition

. ,
N(ri,O) - Z N(gi,O) = -PG (5.5.21)
i=l i=l

where PG is the n u m b e r of right h a l f - p l a n e poles of G(s).

The n e c e s s i t y and s u f f i c i e n c y of c o n d i t i o n (5.5.21) for _

closed-loop stability when conditions (bl), (d') and (e I )

are s a t i s f i e d is p r o v e d as follows

From relationship (5.1.21) we have that

PG = e + Z Z* (5.5.22)
i=l gi

where e is the n u m b e r of right h a l f - p l a n e poles of eG(s), and

combining this w i t h e q u a t i o n (5.5.19) gives

. Z .
Z N(ri,O) - Z N(gi,O) = Z Z* + e-P G (5.5.23)
i=l i=l i=l ri

To e s t a b l i s h the n e c e s s i t y of c o n d i t i o n (5.5.21)

s u p p o s e that
, ,
N(ri,O) - 2 N(gi,O) fl -PG
i=l i=l
 91

then from e q u a t i o n (5.5.23) this implies that

7. Z* ~ O and/or e ~ 0
i=l r i

and h e n c e we c o n c l u d e that c o n d i t i o n (5.5.21) is n e c e s s a r y

for c l o s e d - l o o p stability.

Fors u f f i c i e n c y suppose that

, ,
N(ri'O) - 7. N(gi'O) =-PG
i=l i=l

then from e q u a t i o n (5.5.23) this implies that

7. Z, = e = 0
i=l r.
l

and h e n c e the s y s t e m is c l o s e d - l o o p stable. Thus the sufficiency

of c o n d i t i o n (5.5.21) is e s t a b l i s h e d .

We have t h e r e f o r e shown that the f o l l o w i n g conditions

are n e c e s s a r y and s u f f i c i e n t for c l o s e d - l o o p stability:

, ~ ,
(a") ~ N(ri'O) - 7. N(gi'O) = -PG ;
i=l i=l
,
(b") {ri(s): i=i,2, .... } have no zeros on the i m a g i n a r y axis;

(c") eG(S)=O has no roots on the i m a g i n a r y axis; and

(d") Pd(S) has only left h a l f - p l a n e zeros.

F r o m e q u a t i o n (5.5.3) it is clear that

, ~ ,
7. N ( r i , O ) = 7 N(gi, -k) (5.5.24)
i=l i=l

and c o n s e q u e n t l y conditon (a") is e q u i v a l e n t to c o n d i t i o n (la)

in s t a t e m e n t 1 of the s t a b i l i t y criterion.

Also from e q u a t i o n (5.5.3) we have that {ri(s): i=i,2,...,}

have zeros on the i m a g i n a r y axis if, and only if the inverse

characteristic gain loci pass t h r o u g h the c r i t i c a l point

(-k+jO); hence condition (b") is e q u i v a l e n t to c o n d i t i o n 2

in s t a t e m e n t 1 of the s t a b i l i t y criterion.

F r o m the p o l e - z e r o relationships, (5.1.21) and (5.1.22),

 92

condition (c") is c l e a r l y equivalent to c o n d i t i o n (3a)

and also c o n d i t i o n (3b) in e i t h e r statement of the s t a b i l i t y

criterion.

Therefore to c o m p l e t e the p r o o f of s t a t e m e n t 1 of the

stability criterion all that is left is to show that c o n d i t i o n

(ib) is e q u i v a l e n t to c o n d i t i o n (a"). To do this we will

s h o w that w h e n c o n d i t i o n s (b"), (c") and (d") are s a t i s f i e d ,

the c o n d i t i o n
, ,
E N(r i, O) = Z N(gi,-k) = - Z G, (5.5.25)
i=l i=l

where Z G is the n u m b e r of r i g h t h a l f - p l a n e zeros of G(s),

is n e c e s s a r y and s u f f i c i e n t for c l o s e d - l o o p stability.

From relationship (5.1.22) we have that

ZG = e + Z P, (5.5.26)
i=l gi

where P*~ is the n u m b e r of right h a l f - p l a n e p o l e s of gj(s),

gj , ,
but from e q u a t i o n (5.5.3) the p o l e s of ri(s) and gi(s) are

the same, and h e n c e

Z
ZG = e + Z P*r. (5.5.27)
i=l l

If we now combine equation (5.5.27) with equation (5.5.15),

which is v a l i d for { j = 1 , 2 , . . . , } , we have

Z N(ri,O)~. = Z Z*r. + e - ZG (5.5.28)

i=l i=l 1

To e s t a b l i s h the n e c e s s i t y of c o n d i t i o n (5.5.25) suppose

that
,
F N(ri,O ) ~ -Z G
i=l

then from e q u a t i o n (5.5.28) this implies that

i
iZ#l~ ~ O and/or e ~ O
i
 93

and hence we conclude that c o n d i t i o n (5.5.28) is n e c e s s a r y

for c l o s e d - l o o p stability.

For s u f f i c i e n c y suppose that

,
N(ri,O) = -Z G
i=l

then from e q u a t i o n (5.5.28) we have that

Z* = e = O
i=l ri

and hence that the system is c l o s e d - l o o p stable, providing

conditions (b"), (c") and (d") are satisfied. Thus the

sufficiency of c o n d i t i o n (5.5.25) is established.

This completes the p r o o f of s t a t e m e n t 1 of the g e n e r a l i z e d

inverse Nyquist stability criterion.

Statement 2 of the s t a b i l i t y criterion applies to systems

see chapter 4] in w h i c h either

(i) Pd(S) = Px(S) (5.5.29)

or

(ii) Pd(S) = Px(S) (s) (s). (s)

Pd I Pd 2 "''Pd h (5.5.30)

where the zeros of {Pdi(S) :i=i,2,...,} are

all in the left half-plane.

In these situations it is possible to combine conditions (la)

and (4), of s t a t e m e n t i, into the single e q u i v a l e n t condition

* i w h
N(g i, -k) - Z N(gi,O) = - Z PG (5.5.31)
i=l i=l i=l i
 94

The n e c e s s i t y and s u f f i c i e n c y of c o n d i t i o n (5.5.31) for c l o s e d -

loop s t a b i l i t y when conditions (2) and (3) are s a t i s f i e d

is p r o v e d as follows.

F r o m e q u a t i o n s (4.2.37) and (4.2.39) we have that

h
PG = e + Z P + Pd (5.5.32)
i=l i i=l f
i

But the poles of {fi(s): i~l,2,...,Z} are the same as the

poles of {gi(s): i=1,2,...,} which in turn are the same as

the zeros of {gi(s) : i = 1 , 2 , . . . , } , and t h e r e f o r e equation

(5.5.32) can be r e w r i t t e n as
h
~iPGi = e + Z Z* + Pd (5.5.33)
i i=l gi

Equation (5.5.33) can now be c o m b i n e d w i t h e q u a t i o n (5.5.19)

to give
, , h
Z N(ri,O) - Z N(gi,O) = Z Z* + e + Pd - Z=IPGi
i=l i=l i=l ri i

(5.5.34)
w h i c h u s i n g e q u a t i o n (5.5.24) b e c o m e s
, i , h
Z N(gf-k) - 2 N(gi,O) = Z Z* + e +Pd - __ZIPGi
i=l i=l i=l ri i

(5.5.35)

To e s t a b l i s h the n e c e s s i t y of c o n d i t i o n (5.5.31)

s u p p o s e that
, , h
N(gi,-k) - Z N(g,O) ~ -Z PG.
i=l i=l i=l l

then from e q u a t i o n (5.5.35) this implies that

Z Z* ~ O, or e ~ O, or Pd ~ O
i=l rl

or any c o m b i n a t i o n of these, and thus that the s y s t e m w i l l be

closed-loop unstable. H e n c e we c o n c l u d e that c o n d i t i o n (5.5.31)

is n e c e s s a r y for c l o s e d - l o o p stability.
 95

For s u f f i c i e n c y s u p p o s e t h a t
, , h
N(gi,-k) - Z N(gi,O) = - Z PG
i=l i=l i=l i

then f r o m e q u a t i o n (5.5.35) we m u s t h a v e t h a t

Z Z* = e = Pd = 0
i=l r i

and h e n c e the s y s t e m is s t a b l e p r o v i d i n g conditions (2)

and (3) are s a t i s f i e d . T h u s the s u f f i c i e n c y of c o n d i t i o n

(5.5.31) is e s t a b l i s h e d .

Therefore to c o m p l e t e the p r o o f of s t a t e m e n t 2 of the

stability c r i t e r i o n all t h a t is left is to s h o w the e q u i v a l e n c e

between conditions (la) a n d (ib). To do t h i s we w i l l show

that w h e n c o n d i t i o n s (2) and (3) are s a t i s f i e d , condition

(ib) i.e.
i , , h
Z N(ri,O) = ~ N(gi,-k) = - ~ ZG. (5.5.36)
i=l i=l i=l l

where zG is the n u m b e r of r i g h t h a l f - p l a n e p o l e s of G. (s), is

i 1
n e c e s s a r y and s u f f i c i e n t for c l o s e d - l o o p s t a b i l i t y .

From equation (5.5.15), w h i c h is v a l i d for {j=1,2,...,},

Z N(ri,O) = Z N(gi,-k) = Z Z* - Z P*
i=l i=l i=l ri i=l ri

(5.5.37)

and since the p o l e s of {r. (s) : i=1,2,...,} are the same as

1
the p o l e s of {gi(s): i=1,2,...,~} equation (5.5.37) can be

rewritten as

Z N(gi,-k) = Z Z* - Z P* (5.5.38)
i=l i=l ri i=l gi

If the s u b s y s t e m s {Gi(s): i=l,2,...,h} are e a c h s q u a r e

t h e n P x ( = P d ) is the n u m b e r of r i g h t h a l f - p l a n e zeros (or

poles) w h i c h are lost t h r o u g h p o l e - z e r o cancellations when

G(s) is formed. Therefore if the n u m b e r of r i g h t h a l f - p l a n e

 96

zeros of G(s) is g i v e n [ e q u a t i o n (3.3.36)] by

9,
ZG = e + Z Z (5.5.39)
i=l gi

t h e n we m u s t h a v e
h
ZG. = e + ~ Z + Pd (5.5.40)
i=l l i=l 9i

or, since the p o l e s of gi(s) are i d e n t i c a l l y e q u a l to the

zeros, of gi(s),
h
ZG = e + Z P* + Pd (5.5.41)
i=l i i=l g i

Consequently equations (5.5.38) and (5.5.41) can be c o m b i n e d

to g i v e
, Z h
Z N(gi,-k) = ~ Z* + e + Pd - ~iZGi
i=l i=l ri i
(5.5.42)

To e s t a b l i s h the n e c e s s i t y of c o n d i t i o n (5.5.36) suppose

that
Z . h
Z N(gi,-k) ~ - ~ ZG
i=l i=l i

then f r o m e q u a t i o n (5.5.42) t h i s i m p l i e s that

E Z* ~ O, or e ~ O, or Pd ~ O
i=l r i

or any c o m b i n a t i o n of t h e s e and t h u s that the s y s t e m w i l l be

closed-loop unstable. H e n c e we c o n c l u d e t h a t c o n d i t i o n

(5.5.42) is n e c e s s a r y for c l o s e d - l o o p stability.

For s u f f i c i e n c y s u p p o s e that
, h
N(gi,-k) = - Z ZG
i=l i=l l

then f r o m e q u a t i o n (5.5.42) we m u s t h a v e t h a t

Z* = e = Pd = O
i=l r i

and h e n c e the s y s t e m is c l o s e d - l o o p stable, providing

conditions (2) and (3) are s a t i s f i e d . T h u s the s u f f i c i e n c y

 97

of c o n d i t i o n (5.5.36) is e s t a b l i s h e d .

This c o m p l e t e s the p r o o f of s t a t e m e n t 2 of the g e n e r a l i z e d

inverse N y q u i s t stability criterion.

5.6 Example

To i l l u s t r a t e the stability criteria consider the system

example used in c h a p t e r s 3 and 4 w h e r e

G(s) = Gl(S)= 1 [ s-i s ]

1.25(s+i) (s+2) -6 s-2

The p o l e and zero p o l y n o m i a l s for the o p e n - l o o p gain m a t r i x

G(s) are

PG(S) = (s+l) (s+2) and zG(s) = 1

so that

PG = O and Z G = O.

The i n v e r s e characteristic gain loci are shown in figure 19.

There are no u n o b s e r v a b l e and/or uncontrollable modes

and t h e r e f o r e either statement 1 or s t a t e m e n t 2 of the c r i t e r i o n

is d i r e c t l y applicable. In fact, b e c a u s e we have e f f e c t i v e l y

only one subsystem, there is no d i f f e r e n c e between the two

statements.

Testing the c o n d i t i o n s for s t a b i l i t y as o u t l i n e d by the

generalized inverse N y q u i s t stability criterion we find that

the c l o s e d - l o o p system is stable for

-1.875 < k ~ 0

O < k< 1.25

and 2.5 < k <

These bounds on the g a i n c o n t r o l variable k agree w i t h those

obtained in s e c t i o n 4.3 using the g e n e r a l i z e d Nyquist

stability criterion.

It is i n t e r e s t i n g to n o t e that an i n v e r s e N y q u i s t - l i k e
 98

~.,, "%.
g-plane
ca.~ 5.o~

\ ..... . /

Figure 19. Inverse characteristic gain loci

 99

stability criterion has recently been used by Mees and Rapp

[6] to establish stability criteria for multiple-loop nonlinear

feedback systems.

References

[1] A.L. Whiteley, "Fundamental Principles of Automatic Regulators

and Servo Mechanisms", J. IEE, 5-22, 1947.

[2] H.H. Rosenbrock, "State Space and Multivariable Theory",

Nelson, London, 1970.

[3] A.G.J. MacFarlane, "Dynamical System Models", Harrap, London,

1970.

[4] A.G.J. MacFarlane, "Return-difference and return-ratio matrices

and their use in analysis and design of multivariable
feedback control systems", Proc. IEE, 117, 2037-2049, 1970.

Is] E. Hille, "Analytic Function Theory", Vol. i, Ginn and Go.,

U.S.A., 1959.

[6] A.I. Mees and P.E. Rapp, "Stability criteria for multiple-
loop nonlinear feedback systems", Proc. IFAC Fourth Multivariable
Technological Systems Symposium, Fredericton, Canada, 1977.
6. Multivariable root loci

The Evans' root locus approach [1;2;3] is a well

established graphical technique used in the analysis and

design of linear time-invariant single-input single-output

feedback systems for estimating the system closed-loop poles

as a function of the gain control variable. Its generaliza%ion

to the multivariable case has caused considerable interest and

steps towards this end have been made with regard to the

asymptotic behaviour of the characteristic frequency loci

[4; 5] , and the angles of departure and approach of the

characteristic frequency loci [6] ; both from a state-space

point of view. In this chapter a method is developed, using

well established results in algebraic function theory [7;8;9],

which allows the asymptotic behaviour and the angles of

departure and approach of the characteristic frequency loci

to be determined from a Laplace transfer function matrix

description of the system.

It is also shown how the method can be utilised to find

the asymptotic behaviour of the optimal closed-loop poles of

a multivariable time-invariant linear regulator as the weight

on the input in the performance criterion approaches zero [ io].

The method seems particularly useful for systems with

small numbers of inputs since the calculations are then simple

enough to be carried out by hand.

6.1 Theoretical background

For the general feedback configuration of figure 3 it

has been shown in sub-section 3.3-4 that the characteristic

 I01

frequency loci (multivariable root loci) are the 180 phase

contours of the characteristic gain function g(s), where g(s)

is defined via the equation

~(g,s) ~ det [gI m - G ( s ) ] = O (6.1.1)

or the equation

~(g,s) = bo(S)~(g,s) = O (6.1.2)

where we have multiplied through by the least common denominator

of the rational coefficients in s.

For simplicity of exposition, and because this is in any case

the usual situation for transfer function matrices arising

from practical situations, it is assumed that ~(g,s) is

irreducible over the field of rational functions in s. The

characteristic gain function g(s) is related to the gain

control variable k via the expression

1 (6.1.3)

so that substituting for g in equation (6.1.2) we have

~(-k-l,s) = k -m F(k,s) = O (6.1.4)

and solutions of

F(k,s) = O (6.1.5)

for s in terms of positive real k, determine the dependence

of the closed-loop poles on the gain control variable k.

Therefore, apart from possible single-point loci, the graphical

description of this dependence constitutes the characteristic

frequency loci of the given system. Note that if there are

no single-point loci then equation (6.1.5) is directly

equivalent to

CLCP(s)~det[SIn-Ac]= det[SIn-S(-k-l)] = O
(6.1.6)
 I02

and the c h a r a c t e r i s t i c frequency loci have n branches.

E a c h of the n b r a n c h e s of the c h a r a c t e r i s t i c frequency

loci can in t h e o r y be r e p r e s e n t e d in the form

si(k) = ui(k) + Jvi(k) i=l,2, .... n (6.1.7)

where j= ~ and the s u b s c r i p t s i are labels for the v a r i o u s

branches. The t a n g e n t to the rth b r a n c h of the loci at a

point So=Sr(ko) is d e f i n e d as the l i m i t i n g position of the

straight line t h r o u g h s o and a n o t h e r p o i n t Sl=Sr(ko+~k) as

s I approaches s o along the branch, that is as gk+O. NOW

the c o m p l e x n u m b e r Sl-S can be r e p r e s e n t e d by the v e c t o r

from s O to s I (figure 20), and the v e c t o r corresponding to

(Sl-So)./~k, w h e r e 8 k > O , has the same d i r e c t i o n as that vector.

Si= 5 (ko+6k)

i\soo
o
Figure 20. To derive a brmuta for the tangent to
the characteristic frequency loci

It follows therefore that the v e c t o r corresponding to

st(k) = dkdSr(k)Jk = ~kOlim sl~_k-so

o
 103

= lim Sr(ko+~k) - Sr(k O) (6.1.8)

6kO Sk

is tangent to the branch at s and the angle e between this

o
vector and the positive real axis is given by

% = argument {Sr(ko) } (6.1.9)

Formula (6.1.9) is fundamental to the determination of

the asymptotic behaviour of the c h a r a c t e r i s t i c frequency

loci, and the angles of departure and approach of the loci.

For k=O, equation (6.1.3) implies that the c h a r a c t e r i s t i c

frequency loci start at poles of the open-loop system, and

therefore the angle of departure of a branch of the loci from

a pole is given by formula (6.1.9) w i t h ko=O. For k=~,

equation (6.1.3) implies that the c h a r a c t e r i s t i c frequency

loci terminate at system zeros. Therefore for a branch

of the loci t e r m i n a t i n g at a finite zero, formula (6.1.9),

with ko=~, gives the angle of approach of the branch to the

zero. If a branch terminates at an infinite zero then formula

(6.1.9), w i t h k =~, gives the angle that the asymptote to

o
the branch makes with the positive real axis.

Formula (6.1.9) does not at first seem to be useful,

because expressions for the separate branches of the characteristic

frequency loci cannot in general be found explicitly. However,

in algebraic function theory [7;8;9], there exists a method

for the p r a c t i c a l construction of series representations

for the branches of an algebraic function in the n e i g h b o u r h o o d

of a given point. The m e t h o d consists of repeatedly using

a "Newton diagram" to find the next most significant terms

in the series. Therefore, by using the Newton diagram just

once, approximations can be obtained for the branches of the

 104

characteristic frequency loci in the v i c i n i t y of a p o l e

or zero (finite or i n f i n i t e ) , of the f o r m

s (k) ~ a + bk ~ (6.1.10)
r
where a and b are c o m p l e x n u m b e r s , ~ is a r a t i o n a l r e a l

number, and w h e n e is f r a c t i o n a l the p r i n c i p a l root of k is

understood. For the n e g a t i v e feedback configuration under

c o n s i d e r a t i o n k is real and p o s i t i v e , ~k e-I w i l l a l w a y s be

real, and t h e r e f o r e applying formula (6.1.9) to the a p p r o x i m a t i o n

(6.1.10) we h a v e

0 = argument {b} + ~180 (6.1.11)

where

[ = O when e > O
T
= 1 when ~ < 0

as a r e s u l t of the d i f f e r e n t i a t i o n with respect to k.

In the f o l l o w i n g sections it is s h o w n h o w to o b t a i n

approximations of the forn% shown in e q u a t i o n (6.1.10) in

the v i c i n i t y of a p o l e or zero, and h e n c e to d e t e r m i n e the

asymptotic b e h a v i o u r of the c h a r a c t e r i s t i c frequency loci

and a l s o the a n g l e s of d e p a r t u r e and a p p r o a c h of the loci.

6.2 Asymptoti q behaviour

The N e w t o n d i a g r a m is a g r a p h i c a l construction which

can be u s e d to find e a c h of the m o s t significant t e r m s in the

series representations for the p a r t i c u l a r b r a n c h e s of an

algebraic f u n c t i o n q(v), in the v i c i n i t y of the origin,

which approach zero as v a p p r o a c h e s zero. Therefore, whether

f i n d i n g the a s y m p t o t i c b e h a v i o u r or the a n g l e s of d e p a r t u r e

and a p p r o a c h of the loci, our f i r s t aim is a l w a y s to r e d u c e

the p r o b l e m , by a c h a n g e of v a r i a b l e s in the c h a r a c t e r i s t i c
 105

equation, to one of f i n d i n g a p p r o x i m a t i o n s for b r a n c h e s of

an a l g e b r a i c function which approach z e r o as the i n d e p e n d e n t

variable approaches zero.

To d e t e r m i n e the a s y m p t o t i c b e h a v i o u r of the c h a r a c t e r i s t i c

frequency loci we n e e d to o b t a i n an a p p r o x i m a t i o n to the b r a n c h

(or b r a n c h e s ) of the loci a b o u t the p o i n t s =~, as k a p p r o a c h e s

infinity. F o r this p u r p o s e we w i l l put

-i
s=z (6.2.1)

in the c h a r a c t e r i s t i c equation (6.1.2) to o b t a i n

(g,s) = ~ (g,z -I) = z -q ~(g,z) (6.2.2)

where q $ n is the n u m b e r of p o l e s of g(s), so that in any

neighbourhood of the v a l u e z = 0 (the p o i n t z = 0 itself is

excluded f r o m the region) the e q u a t i o n # ( g , s ) = O is e q u i v a l e n t

to the e q u a t i o n

~(g,z) = E~xygXzy = O (6.2.3)

For a s t r i c t l y p r o p e r system, that is one w h e r e D is zero,

or if D is s i n g u l a r

oo = 0 (6.2.4)

o t h e r w i s e C o o is n o n - z e r o which, as we shall see later f r o m

the N e w t o n d i a g r a m , corresponds to t h e r e b e i n g no a s y m p t o t i c

behaviour. N o t e t h a t if D is n o n - s i n g u l a r g(s) has t h e same

n u m b e r of f i n i t e z e r o s as p o l e s a n d t h e r e f o r e we w o u l d not

expect any c l o s e d - l o o p p o l e s to a p p r o a c h i n f i n i t y .

The n e x t step is to c o n s t r u c t the N e w t o n d i a g r a m for

~(g,z) , from which approximations of the f o r m

z ~cg ~ (6.2.5)

can be o b t a i n e d , where c is a c o m p l e x n u m b e r , and ~ a r a t i o n a l

real n u m b e r . The p r o c e d u r e for c o n s t r u c t i n g the a p p r o p r i a t e

N e w t o n d i a g r a m is as f o l l o w s [8].

In a ( g , z ) - p l a n e we p l o t the p o i n t s (x,y) for w h i c h

 106

~xy ~ 0 in equation (6.2.3). As an example, the points for

(l-z+2z2-25z3+29z4)g 2 + (z-22z2+199z3-21Oz4)g

-(33z3-594z 4) = O (6.2.6)

are shown in figure 21. A straight line is then made to

coincide with the horizontal axis and rotated clockwise about

the smallest g-axis point P (the point (2,0) on figure 21)

o
until this line passes through another point P1 of our net.

A straight line is then drawn between the points Po and PI"

A horizontal line through P 1 ' and pointing away from P1

towazds the vertical z-axis, is then rotated clockwise about

the point P1 until it passes through another point P2 of

our net. A straight line is then drawn between the points

P1 and P2" The procedure is repeated until the vertical

z-axis is reached. The complete Newton diagram for equation

(6.2.6) is shown in figure 22.

The tangent of the acuteangle which the straight line

PoP1 makes with the vertical z-axis determines the first

possible value of ~ , and the tangent of the acute angle which

the straight line PIP2 makes with the verticle z-axis determines

the second possible value of p , etc. For the Newton diagram

of figure 22 we have U 1 = 1 and ~2 = "

For a particular exponent Pt there may be several approxi-

mations of the form

~t
zit-citg (6.2.7)

Note that if some root of g is implied by ~t i.e. Pt is a

fraction, then it is understood that the principal root is

being considered. To determine the coefficients cit it

 107

Zl
4,

3:

C i 2 a
tJ

Figure 21. Points of the Newton diagram

for equation (6.2.6)
z
4

3P,

0
2 a

Figure 22. Complete Newton diagram

for equation (6.2.6)
 108

~t
is n e c e s s a r y to s u b s t i t u t e z = cg in the terms of

equation (6.2.3) corresponding to the p o i n t s of our net

lying on the link Pt-i Pt ' to e q u a t e to zero the result of

the s u b s t i t u t i o n and to solve the r e s u l t i n g equation. Let

the sum of the relevant terms of e q u a t i o n (6.2.3) have the

form

x y~ gXlzY 1
x g z + ..... (6.2.8)
oy~ "+ ~xlY 1

w h e r e y > ......... >YI' and ~ is a p o s i t i v e integer less than

the n u m b e r of i n f i n i t e zeros of the system. Then b e c a u s e

the terms correspond to the same link we have

x~-x I Xo_l-Xl _x 2 - x 1
Yc-Yl - YO-I-Yl ........ Y2 - Yl ~t

Therefore all the terms of the e x p r e s s i o n (6.2.8) become

Ut
similar as a r e s u l t of the s u b s t i t u t i o n z = cg , and h e n c e

for the d e t e r m i n a t i o n of the c o e f f i c i e n t cit we o b t a i n an

equation of t h ~ a f o r m
Yl
~xgy c + ........ + ~XlY 1 c = O (6.2.9)

Yl
or d i v i d i n g by c (c ~ O) we o b t a i n
Y~-Yl
~xay c + .... +~XlY 1 = 0 (6.2.10)

w h i c h has y a - y I solutions. Note that Yo-Yl is the

difference between the o r d i n a t e s of the p o i n t s Pt and Pt-l"

Consequently it is equal to the p r o j e c t i o n of the link P t _ i P t

on to the o r d i n a t e axis. This is a u s e f u l fact since it

enables us to see at a g l a n c e how m a n y coefficients there are

corresponding to the e x p o n e n t ~t" A l s o the o r d i n a t e associated

w i t h the p o i n t w h e r e the final link reaches the v e r t i c a l axis

 109

must give the total number of coefficients for all the

approximations considering all exponents. This vertical

axis point on the final link is therefore the number of

infinite zeros of the system.

Having obtained the approximations in the form of equation

(6.2.7) it is a simple matter of substitution from equations

(6.1.3) and (6.2.1) to obtain the required form

~t
sit - bit k (6.2.11)

representing approximations to the characteristic frequency

loci about s =~, as k~. If we now apply formula (6.1.9) we

obtain the angles which the asymptotes make w i t h the positive

real axis.

The asymptotes, as will be shown in sub-section 6.2-1,

group into B u t t e r w o r t h configurations. Each pattern has

a common intercept which has been termed the "multivariable

pivot" by Kouvaritakis and Shaked [4] who also gave a m e t h o d

for calculating the pivot given the state space description

of the system. In appendix 6 it is shown how the m u l t i v a r i a b l e

pivots can be derived from the characteristic equation

~(g,s) = O.

6.2-1 B u t t e r w o r t h Patterns

In this sub-section it is proved that the closed-loop

poles that go off to infinity do so along asymptotes which

group into several B u t t e r w o r t h configurations. The proof

is based on well e s t a b l i s h e d results in algebraic function

theory [8]

Let us consider the closed-loop c h a r a c t e r i s t i c polynomial

defined by equation (6.1.5) i.e.

 110

F (k,s) = O (6.2.12)

in w h i c h we a l l o w k to be complex. Also suppose that at a

point k c the a l g e b r a i c function s(k) defined by e q u a t i o n

(6.2.12) has a p- fold root So, and let the n branches of s(k)

in a n e i g h b o u r h o o d N of k be {s.(k):i=l,2,...,p,...,n}
c l

where {si(kc)=Sc:i=l,2,...,p}.

If we a n a l y t i c a l l y continue any one of the b r a n c h e s

{si(k) : i=p+l,...,n} around k c in N, then after one e n c i r c l e m e n t

we return to the o r i g i n a l function. If, however, we a n a l y t i c a l l y

continue Sl(k) around the critical point k c then after one

encirclement we obtain one of the functions {si(k) : i=2,3,...,p} ,

s2(k) say, and after another encirclement one of the functions

{si(k ) : i=3,4,...p}, s3(k ) say. After a finite number ~i' of

encirclements we return to the original function Sl(k).

In this case we shall say that the functions

Sl(k) t . . . . , S~l (k) (6.2.13)

constitute a cyclical system of b r a n c h e s of the a l g e b r a i c

function s(k) (in the n e i g h b o u r h o o d N of kc).

If 91<p we take the function s l+l(k) and a n a l y t i c a l l y

continue it around k in N. Repeating the p r e v i o u s arguments

o
we obtain a second cyclical system of branches, namely

S~l+l(k)' .... ' s~2(k) (6.2.14)

Finally all the functions Sl(k), ..... Sp(k) will be sorted into

cyclical systems.

The functions of each cyclical system pass consecutively

into one another according to the cyclical law when k goes

 111

round the m u l t i p l e p o i n t k c. Hence every cyclical system

of b r a n c h e s defines in the n e i g h b o u r h o o d N of k an analytic

c
function. The n u m b e r of v a l u e s g i v e n by this a n a l y t i c

function at a p o i n t in the n e i g h b o u r h o o d N is equal to the

n u m b e r of b r a n c h e s comprising the c y c l i c a l s y s t e m being

considered. When a single b r a n c h comprises a cyclical system,

the c o r r e s p o n d i n g analytic f u n c t i o n w i l l be s i n g l e - v a l u e d

and i d e n t i c a l w i t h this branch.

Suppose the n e i g h b o u r h o o d N is r e p l a c e d by a p i l e of

n such regions, one for each of the f u n c t i o n s {si(k) : i=l,2,...,n}

and let a p o i n t on the ith sheet r e p r e s e n t the p a i r [k,si(k) ] .

Also suppose that each n e i g h b o u r h o o d N is c i r c u l a r and has a

radial line cut from the c r i t i c a l point k to its p e r i p h e r y ,

c
as shown in figure 23. Then the n e i g h b o u r h o o d s can be j o i n e d

together along the edges of the cuts so that if we a n a l y t i c a l l y

continue the jth b r a n c h (j<p) around k c on the jth sheet,

we move after one e n c i r c l e m e n t across the cut into the (j+l)th

b r a n c h on the (j+l)th sheet, etc. eventually returning to

the jth branch. The pile of n e i g h b o u r h o o d s obviously

falls into cycles containing one or m o r e sheets as shown in

figure 24, and each cycle is a d o m a i n for one of the a n a l y t i c

functions defined in the n e i g h b o u r h o o d N of k .

c

To d e n o t e an a n a l y t i c function determined by a c y c l i c a l

system we w i l l use a Roman numerical subscript. For e x a m p l e

the a n a l y t i c function corresponding to the c y c l i c a l system

Sl(k) , .... , s l(k) w i l l be d e n o t e d by the symbol si(k).

When we c o n s i d e r the a s y m p t o t i c behaviour of the

 112

L cut

Figure 23. Neighbourhood N of the

critical point k c

kc

Figure 24. Cyctes of neighbourhoods for

the critical point k c

characteristic frequency loci, we are concerned with the

branches of the algebraic function s(k) which approach

infinity, as k approaches infinity. Suppose we have p

infinite branches for k =~, then by the previous arguments

the p branches will be arranged into several cyclical

systems. Each cyclical system is defined by an analytic

function, the form of which is given by the f o l l o w i n g

theorem [8, page 39]

Theorem 4

The infinite branches of the algebraic function s(k),

constituting a cyclical system in the neighbourhood of its

critical point kc=~, in their totality constitute an analytic

 113

function given by a series of the form

1 -i -2
s I = k 9 (bO + b I k V +b2k ~ +...) (6.2.15)

where v is the number of branches in the cycle, and I is a

positive integer.

Consequently the infinite branches corresponding to a

particular cyclical system in the n e i g h b o u r h o o d of k =~, can

be a p p r o x i m a t e d by an analytic function of the form

1
m

s I (k) ~bok~ (6.2.16)

The characteristic frequency loci are values of the

algebraic function s(k) for k going from zero to infinity

along the positive real axis. Therefore the infinite

branches of the frequency loci will occur in cyclical

systems which can be a p p r o x i m a t e d by equations of the form

(6.2.16). Taking the ~ roots of k in equation (6.2.16)

we find that the asymptotes corresponding to a cyclical

system arrange themselves into a B u t t e r w o r t h configuration

of order ~ ; that is,we have ~ asymptotes equally spaced

.360.
by angles of (--~-), in the complex plane.

In general, as already discussed, there will be several

cyclical systems of branches in the n e i g h b o u r h o o d of k = ~,

and therefore the asymptotes of the characteristic frequency

loci will arrange themselves into several B u t t e r w o r t h

configurations.

6.3 Angles of d e p a r t u r e and approach

To determine the angles of departure and approach of

the characteristic frequency loci we need to obtain approximations

to the branches of the loci at the poles and finite zeros

 114

of the system. We w i l l first consider the angles of

departure f r o m the o p e n - l o o p poles of the system.

Suppose for the characteristic equation

~(g,s ) = O (6.3.1)

we have a pole (or m u l t i p l e pole) at s = ~. We w i l l m a k e

s'=s-8 the n e w independent variable for e q u a t i o n (6.3.1)

and also make the substitution g = d -1, so t h a t equation

(6.3.1) becomes

~(g,s) = ~(d-l,s'+B) = d - m H ( d , s ') = O (6.3.2)

The situation we are e x a m i n i n g now reduces to the case w h e r e

d = s' = O, and in the n e i g h b o u r h o o d of s = 8 (excluding

itself) equation #(g,s) = O is e q u i v a l e n t to the e q u a t i o n

H(d,s') = Z ~ x y d X s 'y = O , to o = O (6.3.3)

If w e construct the N e w t o n diagram for the e q u a t i o n

Z(d,s') = O an a p p r o x i m a t i o n (or a p p r o x i m a t i o n s in the case of

multiple poles) of the form

s' ~ ed ~ (6.3.4)

is o b t a i n e d , where e is a c o m p l e x number and m a r a t i o n a l

real number. From equation (6.1.3), the change of v a r i a b l e ,

a n d the substitution g = d -I , we therefore arrive at an

approximation to the b r a n c h (or b r a n c h e s ) of the c h a r a c t e r i s t i c

frequency loci departing from the p o l e B , in the form

Hd
s - B + bdk (6.3.5)

If w e n o w a p p l y formula (6.1.9) the angle of d e p a r t u r e 8d is

given as

ed = argument {b d} (6.3.6)

We w i l l now consider the angles of approach to the finite

 115

zeros of the system. Suppose for the characteristic

equation (6.3.1) we h a v e a zero (or m u l t i p l e zero) at s = y .

We w i l l take s' = s-y as the n e w i n d e p e n d e n t variable so

that the e q u a t i o n becomes

~(g,s) = ~(g,s'+y) = x(g,s') = O (6.3.7)

The s i t u a t i o n reduces to the oase w h e r e g = s' = O, and in

the n e i g h b o u r h o o d of s = 7 (excluding y itself) equation

~(~s) = O is e q u i v a l e n t to the equation

gXs'Y= O , = 0 (6.3.8)
x(g's') = ~xy oo

If we construct the N e w t o n diagram for x(g,s') an

approximation of the form

s' ~ pg~ (6.3.9)

is o b t a i n e d where p is a c o m p l e x number a n d n is a r a t i o n a l

real n u m b e r . From equation (6.1.3) a n d the c h a n g e of v a r i a b l e

we therefore arrive at an a p p r o x i m a t i o n to the b r a n c h (or

branches) af the characteristic frequency loci approaching

the zero s = y , in the form

s - y + b ak~a (6.3.10)

If w e n o w a p p l y formula (6.1.9) the angle of a p p r o a c h 8 a is

given as

ea = argument {b a} ~ 180 (6.3.11)

Note that Ua w i l l always be n e g a t i v e and h e n c e the presence

of the 180 t e r m in f o r m u l a (6.3.11). Also note that this

definition for the angle of a p p r o a c h differs by 180 f r o m the

usual definition which is the d i r e c t i o n you would look, when

positioned at the zero, to see the locus arrive.

6.4 Example 1

Consider an o p e n - l o o p gain matrix

 116

0s l [3s3+s21s 8s22s+l 1
s4+5s3-2s2-44s+40 s3+79s2+44s-868 -4s3-4s2+4Os-32

w h i c h has a c h a r a c t e r i s t i c e q u a t i o n

~(g,s) = ( s 4 + 5 s 3 - 2 s 2 - 4 4 s + 4 0 ) g2 + (s3+l16s-432)g

-12 (s2-2s+2) = O

(a) T_o d e t e r m i n e the a s y m p t o t i c b e h a v i o u r

-i
Putting s = z in the c h a r a c t e r i s t i c e q u a t i o n we o b t a i n

~(g,z) = (l+5z-2z2-44z3+4Oz4)g 2 + (z+l16z3-432z4)g

-12 (z2-2z3+2z 4) = O

The N e w t o n d i a g r a m for ~(g,z) is s h o w n in f i g u r e 25 from w h i c h

we o b t a i n ~ = 1 , and h e n c e the a p p r o x i m a t i o n

z - cg

The c o e f f i c i e n t c is c a l c u l a t e d (using the m e t h o d described

in s e c t i o n 6.2) to h a v e the v a l u e s -71 and ~1 . Therefore,

resubstituting for z and g = - k -I , we h a v e the f o l l o w i n g

approximations for the c h a r a c t e r i s t i c frequency loci

s - 4k and s - -3k as k + ~

Two b r a n c h e s of the c h a r a c t e r i s t i c frequency loci t h e r e f o r e

m o v e off to i n f i n i t y at a n g l e s of O and 180 to the p o s i t i v e

real axis.

o ~
\ S

Figure 25. Newton diagram for ~z(g,z)

 117

(b) To find the an@les of departure

The system has four open-loop poles

s = i, s = 2 , s = -4+2j , and s = -4-~

Pole at s = 1
-i
Putting s' = s - i and g = d in t h e characteristic

equation we obtain

~l(d,s') = (S,4+gs,3+19S'2-29S ') + ( s ' 3 + 3 s ' 2 + l l g s ' - 3 1 5 ) d

- 12(s'2+l)d 2 = O

The Newton diagram for Hl(d,s') is shown in figure 26 from

which we obtain e I = 1 , and hence the approximation

s' ~ e l d

The coefficient e I is calculated (using the method described

in s e c t i o n 6.2) to h a v e the value -10.86, resulting in t h e

following approximation to t h e characteristic frequency loci

s - 1 + 10.86 k

about the pole s = i. Therefore the angle of departure from

the p o l e s = 1 is 0 .

Pole at s = 2

Putting s' = s-2 and g= d -I in the characteristic

equation we obtain

E2(d,s' ) = (s,4+13s,3+52s,2+4Os ,) + (s'3+6s'2+128s'-192)d

- 12(s'2+2s'+2)d 2 = O

The N e w t o n diagram for E2(d,s') is shown in figure 27 from

which we obtain ~2 = 1 , and hence the approximation

s' ~ e 2 d

The coefficient e 2 is c a l c u l a t e d to have the value 4.8,

resulting in t h e following approximation to t h e characteristic

frequency loci
 118

s-2-4.Sk

about the pole s = 2 . Therefore the angle of d e p a r t u r e from

the pole s = 2 is 1 8 0 .

Pole a t s = -4+2j

Putting s' = s+4-2j a n d g = d -I in the c h a r a c t e r i s t i c

equation we obtain

H3(d's') = [s'4+(-ll+j8) s'3+(lO-j60)s'2+(88+j104)s']

+ [s' 3+ (_12+96) s' 2+ (152-j 48) s ' + (-912+9 320)] d

-12 I s ' 2+ (-lO+j 4) s ' + ( 2 2 - 9 2 0 ) ] d 2 = 0

The Newton diagram for 23(d,s') is s h o w n in f i g u r e 28 f r o m

which we o b t a i n ~3 = i, a n d h e n c e the approximation

s' ~- e 3d

The coefficient e 3 is c a l c u l a t e d to h a v e the value

912 - j 3 2 0
88 + j l O 4

resulting in the following approximation to the characteristic

frequency loci
(-912+j320) k
s - -4+2j + (88+9104)

about the pole s = -4+2j. Therefore the angle of d e p a r t u r e

from the pole s = -4+29 is

-912+j32Q~ o
argument { 88+9104 j = 110.9

Pole at s = - 4 - 2 j

By symmetry the angle of d e p a r t u r e from the pole

s = -4-2j is - 1 1 0 . 9 .

(c) To f i n d the angles of a p p r o a c h

The system has two finite zeros

s = l+j and s = l-j

 119

4i

2,

o
L
\
2 d ( I 2 =d

Figure 26. Newton Figure 27 Newton

diagram for ~l(d,s') diagram for E2(d,s')
$'
$'
4'

,\
O ---
0 I 2 "d I 2
8
Figure 28. Newton Figure 29. Newton
diagram for E3(d,s') diagram for X(g,s')

3-0

2.0

I I
6'0 7"0

2/
Figure 30. Complete characteristic frequency loci
 120

Zero at s = i ~

Putting s' = s-l-j in the c h a r a c t e r i s t i c e q u a t i o n we

obtain

(g,s') = [ S , 4 + ( 9 + j 4 ) s , 3 + ( 1 3 + j 2 9 ) S , 2 + ( _ 5 8 + j 4 6 ) S , + ( _ 1 8 _ j 3 8 ) ] g2

+[s'3+(3+j3)s'2+(l16+j8)s'+(-318+j118)] g

12[s2j2s '] = 0
The N e w t o n d i a g r a m for x(g,s') is shown in f i g u r e 29 f r o m

w h i c h we o b t a i n n = 1 , and h e n c e the a p p r o x i m a t i o n

s' - pg

The c o e f f i c i e n t p is c a l c u l a t e d to h a v e the v a l u e

-318 + jl18
j24

resulting in the f o l l o w i n g a p p r o x i m a t i o n to the c h a r a c t e r i s t i c

f r e q u e n c y loci
(318-jI18) k -i
s - 1 + j + j24

a b o u t the z e r o s = l+j. T h e r e f o r e the a n g l e of a p p r o a c h

to the zero s = l+j is

f318-jl18% _
argument L j24 ~ + 180 E 69.64

Zero at s = l-j

By s y m m e t r y the a n g l e of a p p r o a c h to the zero at

s = l-j is - 6 9 . 6 4 .

To c h e c k the r e s u l t s the c o m p l e t e characteristic

frequency loci are shown, projected onto a single complex

f r e q u e n c y plane, in f i g u r e 30. Branches of the loci c o i n c i d e

a b o u t the p o l e s = 1 m a k i n g the m o v e m e n t of the c l o s e d - l o o p

poles difficult to c o m p r e h e n d . The frequency surface

c h a r a c t e r i z e d by c o n s t a n t p h a s e and c o n s t a n t m a g n i t u d e

c o n t o u r s of g(s) is t h e r e f o r e shown in f i g u r e s 31(a), (b)

 121

ChQracterJstic
:" ~: frequency loci

(a) I Cutjoining
branch points

(b)

Figure 31.Frequency surface (a)sheetl (b)sheet2

 122

2.1

f 1.95 I ]
4-05 4'0 -3"95 -3"9

Figure 32. Angle of departure from pole s=-42j

iI

J
1.0

0.98

0.9~

0-94

0"9~
0"90

0.90

Figure 33. Angle of approach to zero s=lj

 123

to give a clearer description of the characteristic frequency

loci. Small regions in the neighbourhood of the pole

s = -4+2j and the zero s = l+j are shown in figures 32 and

33 to verify the calculated angles of departure and approach.

6.5 Asymptotic behaviour of optimal, closed-loop poles

In this section the "Newton diagram" approach is used

to determine the asymptotic behaviour of the optimal closed-

loop poles of a multivariable time-invariant linear regulator,

as the weight on the input in the performance criterion

approaches zero. The method is based on an association of

the optimal characteristic frequency loci with the branches

of an appropriate algebraic function. Although the procedure

is equivalent to that given by K w a k e r n a ~ [Ii] the essential

simplicity of the approach is emphasized in the setting of

algebraic function theory.

Consider the stabilizable and detectable time-invariant

linear system

dx(t) _ A x(t) + Bu(t) (6.5.1)

dt

y(t) = C x(t) (6.5.2)

and the performance criterion

V(~) = /o~T(t)Qy(t)+puT(t) Ru(t) ] at (6.5.3)

where Q and R are positive definite symmetric matrices,

and the superscript T denotes the transpose of a matrix or

vector. Then it is well known (e.g. [12]) that the optimal

control action is given by

u(t) = - ! R-IB T Px(t) (6.5.4)

P
where P is the unique positive semi-definite solution of
 124

the steady state matrix Riccati equation

-PA -ATp + ~PBR-IBTp = cTQc (6.5.5)

P
Kwakernaak [ii, equation (16)] has related the closed-

loop characteristic polynomial (denoted by ~c(S)) for the

optimal regulator to the open-loop characteristic polynomial

(denoted by ~o(S)) as follows

~c(S)~c(-S) = ~o(S)~o(-S) det [I m + IR-IGT

P
(-s) QG (s) ]
(6.5.6)

where I is a unit matrix of order m, the number of system

m
inputs, and

G(s) = C(sI - A)-IB (6.5.7)

is the open-loop transfer function or gain matrix of the system.

It is obvious from (6.5.6) that the poles of the optimal closed-

loop system, dependent on the input weighing, are values of

s in the left half-plane which satisfy

det [ Im + ~H(s) ] = O (6.5.8)

P

where
H(s) ~ R-IGT (-s) QG (s) (6.5.9)

Consider now the characteristic equation defining the

eigenvalues of H(s), that is

&(~,s) ~ d e t [ q I s -H(s)] = 0 (6.5.10)

or by expanding the determinant

~(D,s) = D m + al(s2)n m-I + . . . . . +am(S2) O
(6.5.11)

where the coefficients {ai(s2) ; i=i,2, .... ,m} are rational

2 2
functions in s . The coefficients are functions in s because
 125

~ (~,s) -- det [ nI m - R-IG T(-s) QG(s)] (6.5.12)

= det [ HI m - R-GT(-s)QG(S)R -]

and R-GT(-s)QG(s)R - is para-Hermitian implying [ll,

appendix A] that ~(n,s) =A(q,-s) which can only be the case

2
if the coefficients are rational functions in s

If b (s 2) is the least common denominator of the coefficients

o
{ai(s2) ; i=l,2,...,m} then from (6.5.11)

bo(S2)~(~,s)=bo(S2)~ m + bl(S2)nm-l+ ..... +bm(S 2) = O

(6.5.13)

where the coefficients {bi(s 2) : i =1,2, .... ,m} are polynomials

in s 2. The function of a complex variable n(s), defined by

(6.5.13) is an algebraic function. In general, (6.5.11)

and (6.5.13) will be reducible into several irreducible

equations over the field of rational functions in s, thereby

defining a set of algebraic functions. However, for simplicity

of exposition it will be assumed that equations (6.5.11) and

(6.5.13) are irreducible over the field of rational functions

in s.

If we compare equations (6.5.8) and (6.5.10) we see

that the optimal closed-loop poles can be defined as the left

half-plane solutions of

n (s) = -p (6.5.14)

i.e. the optimal characteristic frequency loci are the 180

phase contours of the algebraic function H(s) in the left

half-plane. If we now consider p as a complex variable and

substitute for ~(s) in (6.5.13) we obtain

bo(S 2) (_p)m + bl(S2 ) (_p)m-l+ .... +bm (s2) = 0

(6.5.15)
 126

which is an algebraic equation defining the algebraic

functions s(p) and p(s). The left half-plane branches of

the algebraic function s(p), for p real and positive, are

the optimal characteristic frequency loci. (Note that equation

(6.5.15) can be obtained directly from equation (6.5.8)).

To determine the asymptotic behaviour of the optimal

characteristic frequency loci we need approximations to the

branches of the algebraic function s(p) in the neighbourhood

of s = ~, as p approaches zero. These can be obtained as

the first terms in the series expansions for the branches of

s(p) about the point s =~, as p approaches zero. For this

--1 .
purpose we put s = z in equation (6.5.15) and procede as in

section 6.2.

6.6 Example 2

To demonstrate the procedure the asymptotic behaviour of

the optimal closed-loop poles will be calculated for

s + 2 6],
1
S(s) (s+l) (s+2) (s+3) (s+4)
s + 3 1

Q = I

and

pR = pI

From this data we have

H(s) ~ R-IGT (-s)QG (s)

-2s 2 + 13 -7s + 151

(s2-1) (s2-2) (s2-3) (s2-4)

7s + 15 37

SO that
 127

A(~,s) A= det[ni 2 _ H(s)]

2 ( -2s 2 + 50) n + (-25s 2 + 256)

(s2 i) (s2 2) (s2_3) (s2_4) (s2_l) 2(s2_2)2(s2_3)2(s2_4)

= O

If we now substitute ~ = -P, and multiply throughout by the

least common denominator of the coefficients, we obtain

(s2-1) 2(s2-2)2(s2-3)2(s2-4)2p2+(s2-1) (s2-2) (s2-3) (s2-4) (-2s2+50) p

+ (-25s 2 + 256) = O.
-i
The substitution s = z now gives
(l-z 2) 2(l-2z2)2(l-3z~2(l-4z2)2p 2 + z6(l-z2) (I-2z2) (l-3z2)l-4z 2)

(-2+5Oz2)p
+ z14(-25 + 256z 2) = 0

The Newton diagram for this last equation is shown in figure 34,

from which we obtain

1 1 1
~i =~ ' ~2 = 8 ' and hence the approximationsz-elP 6 and
1 1
z=e2P~_ . To find the values of e I we substitute z = e.p 6 i n t o

I i.e. the terms corresponding to the

p2 _ 2z6p = 0 points on the first link of the Newton

diagram equated to zero,

giving e I = 6/~-~ = O.891 exp (j2kn), k=O,1,2,3,4,5.

6
1
If we substitute z = e2P 8 into
i.e. the terms corresponding to the

-25z 14 - 2z6p = 0 t points on the second link of the Newton

diagram equated to zero,

we obtain

e 2 = 8/-0.08 = 0.729 exp j(~+2k~.), k=O,l,2, .... ,7

8
~o
~Q
m~
~D O N
T~
m7

c~

Q.
~o

LQ

3
 129

Im (s)

. . . . J~ . . . . . . J .....

-40 -20 20 4 0 "Re'(S)

a)

I I
-40 -20 20 40 R (sl

(b)
Figure 35. Asymptotic behaviour of the optimal characteristic
frequency loci (plus right half-plane image)displayed
on the Riemann surface domain of q(s)
(a) sheet 1 of the surface
(b)sheet 2 of the surface
 130

If we now substitute back for s, and take only left half-

plane solutions, we obtain the following expressions for the

asymptotic behaviour of the optimal characteristic frequency

loci
-i
s~l.12 [ e x p
-
( j -~
k~ )]j p ~ ,k = 2,3,4

and
-I
s~ 1.37 [ e x p j (w+2kz)] p 8 , k = 2,3,4,5.
8

Note that using Kwakernaak's notation [ii], these

define Butterworth patterns of orders 3 and 4. The optimal

characteristic frequency loci and their right half--plane

image are shown in figure 35. Note that they have been computed

as the 180 phase contours of the algebraic function ~(s)

and displayed on its Riemann surface domain.

References

i] W.R. Evans, "Graphical Analysis of Control Systems", Trans.

AIEE, 67, 547-551, 1948.

2] W.R. Evans, "Control System Synthesis by Root Locus Method",

Trans. AIEE, 69, 1-4, 1950.

[3] W.R. Evans, "Control System Dynamics", McGraw-Hill, New

York, 1954.

[4] B. Kouvaritakis, and U. Shaked, "Asymptotic behaviour of

root-loci of multivariable systems", Int. J. Control, 23,
297-340, 1976.

[5] D.H. Owens, "A note o n series expansions for multivariable

root-loci", Int. J. Control, 26, 549-557, 1977.
6] U. Shaked, "The angles of departure and approach of the root-
loci in linear multivariable systems", Int. J. Control, 23,
445-457, 1976.

[7] G.A. Bliss, "Algebraic Functions", Dover, New York, 1966

(reprint of 1933 original).
 131

[8] B.A.Fuchs, and V.I.Levin, "Functions of a Complex Variable",

International Series of Monographs in Pure and Applied
Mathematics, Pergamon Press, 1961 (translation of 1951
Russian original).

[9] E.Hille, "Analytic Function Theory", Vol. 2, Ginn and Co.,

U.S.A., 1962.

[i0] I.Postlethwaite, "A note on the characteristic frequency

loci of multivariable linear optimal regulators", IEEE
Trans. Automatic Control, 23, 757-760, 1978.

[ii] H. Kwakernaak, "Asymptotic Root Loci of Multivariable Linear

Optimal Regulators", IEEE Transo Automatic Control, 21,
378-382, 1976.

[12] A.G.J.MacFarlane, "Dual-system methods in dynamical analysis

Pt. 2 - Optimal regulators and optimal servo-mechanisms",
Proc. IEE, 1458-1462, 1969.
7. On parametric stability and future research

A feedback system is said to be stable if all of its

closed-loop poles are in the left half-plane. The stability

of a control system is therefore dependent on its associated

parameters. Sometimes in a control system the value of a

parameter is uncertain perhaps due to ageing, deterioration,

or damage; in other instances it may be desirable, for

economic reasons, to change a p a r a m e t e r value. In both

these cases a technique which predicts the relative stability

of a system w i t h respect to a given p a r a m e t e r would be extremely

useful.

A dominant theme in the p r e c e d i n g chapters has been the

association of a system with two sets of algebraic functions:

characteristic gain functions and c h a r a c t e r i s t i c frequency

functions. In this Chapter c h a r a c t e r i s t i c parameter functions

are introduced, and used to develop the ideas of 'parametric'

root loci and 'parametric' Nyquist loci from w h i c h the relative

stability of a system, w i t h respect to a single parameter,

can be determined.

In the final section of this chapter a few tentative

proposals and suggestions for future research are made.

7.1 Characteristic freqUency and c h a r a c t e r i s t i c parameter

functions

The feedback c o n f i g u r a t i o n considered is shown in figure

36, where A(k2,k3,...,kq), B(k2,k3,...,kq), C ( k 2 , k 3 , . . . k q)

and D(k2,k3,...,k q) are state-space matrices w h i c h are

dependent on (q-l) real, t i m e - i n v a r i a n t parameters and k I is

 133

a scalar, time-invariant gain parameter common to all the loops.

-~ID(k2.....kq),l

Figure 36. Feedback configuration

for parameter anatysis

The closed-loop poles for this configuration are solutions

of

det [ si n - S(k) ] = O (7.1. i)

where

S (k) ~A (k 2 .... ,kq) -B (k 2 .... kq)~l-llm+D(k2 .... kq) ] -Ic (k2, ,kq)

is the closed-loop frequency matrix [see section 3.1]. If

numerical values for all the parameters except one, kj say,

are substituted into equation (7.1.i), and kj considered as

a complex variable, then the resulting algebraic equation

 1:34

(which for simplicity of exposition will be regarded as

irreducible) defines a pair of algebraic functions [ i],

s(kj) and kj(s). The algebraic function s(kj) is called the

characteristic frequency function with respect to kj, and the

algebraic function kj(s) is called the characteristic parameter

function for k . (Note that the characteristic frequency

3
function s(g) and the characteristic gain function g(s),

introduced in chapter 3, are equivalent to s(-kl-l) and

-kl(S)-i respectively).
The branches of s (kj ), for k 3 real, clearly define the

variation of the closed-loop poles with respect to kj, and

as such are termed parametric root loci. Alternatively,

the parametric root loci can be viewed as the O phase

contours of k. (s) on the frequency surface domain for k. (s).

3 ]
Dual to the parametric root loci are the parametric

Nyquist loci or characteristic parameter loci which are the

branches of k.(s) as s traverses the imaginary axes.

]
Alternatively, the characteristic parameter loci can be

viewed as the ~90 phase contours of s(kj) on the Riemann

surface domain for s(kj) which will be called the parameter

surface for k..

3
If, for a particular system, we have a set of nominal

values for the system parameters we can determine which,

if any, are sensitive with respect to stability by looking

at the set of parameter surfaces. To help in such an assessment

the following generalizations of gain and phase margin are

introduced.
 135

?.2 G a i n and phase margins

The ~ 90 phase contours of s(kj) on the p a r a m e t e r surface

for k. trace out the boundary between stable and unstable

3
closed-loop poles and therefore we can define parameter gain
o
and phase margins for k. about a stable operating point k.
3 3
which give a measure of the relative stability of the system

with respect to ko.

3
Parameter gain margi n . Parameter
gain margin is defined with
o
respect to a stable operating point k. as the smallest change
3
o
in parameter gain about kj needed to drive the system into

instability. Let d i be the shortest distahce along the real

o
axis from a stable operating point k to the stability boundary
3
(characteristic p a r a m e t e r loci) on the ith sheet of the parameter

surface for kj. Then the p a r a m e t e r gain margin is defined as

min{d.: i=1,2 ..... n}.

i l

Parameter phase margin. On each of the n sheets of the

parameter surface
for k. imagine that an arc is drawn,
3
o
centre the origin, from a stable operating point kj until

it reaches the stability boundary (characteristic parameter

loci). Let ~i be the angle subtended at the origin by

the corresponding arc on the ith sheet. Then the p a r a m e t e r

phase margin is defined as min{~i: i=1,2 .... ,n}

l

7.3 Example

In this section an inverted p e n d u l u m p o s i t i o n i n g system

(see figure 37) is considered and its stability analysed

with respect to one of its parameters, namely the m a s s of the

carriage.
 136

ndutum

PlV~ll iage

kJ k2
/I/I////
Figure 37. Inverted pendulum
positioning system

This s y s t e m has also been used by K w a k e r n a a k and S i v a n

[2], Cannon[3], and E l g e r d [ 4 ] . The s y s t e m can be m o d e l l e d

by the f o l l o w i n g linearized state d i f f e r e n t i a l equation[2]

x(t) = 0 1 o o x(t)+ ro]u(t )

0 -~F o o I~ I (7.3.1)
I |
o o o l:I
-~ o ~ o
where u(t) is a force e x e r t e d on the carriage by a s m a l l motor;

M is the mass of the carriage; F is the f r i c t i o n coefficient

associated with the m o v e m e n t of the carriage; and L' is given

by

L' = J + mL 2 (7.3.2)
mL
 137

w h e r e m is the m a s s of the p e n d u l u m ; L is the d i s t a n c e from

the p i v o t to the c e n t r e of g r a v i t y of the p e n d u l u m ; and J

is the m o m e n t of i n e r t i a of the p e n d u l u m w i t h r e s p e c t to

the c e n t e of g r a v i t y .

The s y s t e m is s t a b i l i z a b l e using state f e e d b a c k of the

form

u(t) = -Kx(t) (7.3.3)

and u s n g the n u m e r % c a l v a l u e s

F -i
- 1 s
M

1 - 1 kg -I
M
(7.3.4)
-2
= 11.65 s
L'

L' = O.842m

it can be f o u n d [2] t h a t

= [86.81, n.21, -118.4, - 3 3 4 4 ] (7.3.s)

stabilizes the l i n e a r i z e d s y s t e m p l a c i n g the c l o s e d - l o o p

poles at - 4 . 7 0 6 ~ j 1.382 and - 1 . 9 0 2 ~ j 3 . 4 2 0 .

We w i l l now look at the p a r a m e t e r s u r f a c e for M to see

how v a r i a t i o n s in the carriag~ mass, a b o u t an o p e r a t i n g p o i n t

of ikg, affect the s t a b i l i t y of the system. The four s h e e t s

of the m a s s surface, c h a r a c t e r i z e d by c o n s t a n t p h a s e and

magnitude c o n t o u r s of s(M), are s h o w n in f i g u r e s 38-41,

from w h i c h the f o l l o w i n g s t a b i l i t y m a r g i n s are o b t a i n e d :

parameter (mass) g a i n m a r g i n = 1 kg

parameter (mass) p h a s e m a r g i n = 60

The g a i n m a r g i n of ikg c o r r e s p o n d s to r e d u c i n g the c a r r i a g e

mass to zero b e f o r e i n s t a b i l i t y occurs. The p a r a m e t e r

 138

~. :~(M) ,.,~

Figure 38. Sheet 1 of parameter (mass) surface

-II)O u ! "l'r~(l~l)l ~L - - ] 0 ~ -~0

4,

-ifD

--70
,z ReC~

t I0

"-~C

5"d~

-4

Figure 39. Sheet 2 of parameter (mass) surface

 139

J..~cr43

M~

Figure 40. Sheet 3 of parameter (mass) surface

~cM)

~.C

IOaB

] 18o"
-4. o ~ o

1 1"70-

...~c

-,4o~

Figure Z~l.Sheet 4 of parameter (mass) surface

 140

surface also shows that there is a m a x i m u m limit to the carriage

mass, for stability, of 2.125 kg. The phase margin of 60

indicates adequate damping of the closed-loop system.

7.4 Future research

It is thought that parameter surfaces may prove to be

useful in the design of p a r a m e t e r dependent controllers for

systems in which a p a r t i c u l a r p a r a m e t e r suffers large variations

during normal operation. For example, the controller of an

aircraft engine needs to operate satisfactorily over a wide

range of altitudes. A possible design scheme could be

(i) to design real constant controllers at a number of altitudes,

(ii) to obtain an altitude dependent controller by "matrix

interpolation", and finally

(iii) to analyse the stability of the system over the whole

w o r k i n g range using the "altitude surface".

As indicated in this chapter the methods put forward in

this work are not only applicable to gain and frequency, but

any single system p a r a m e t e r and frequency. To obtain stability

tests in terms of more than one p a r a m e t e r variation is a c o m p l i c a t e d

problem, but one with great practical significance. It is felt

that v a l u a b l e insight into this p r o b l e m may stem from a study of

functions of several complex variables.

In recent years stability tests have been developed for

two-dimensional and m u l t i - d i m e n s i o n a l digital filters [5]. Two-

dimensional digital filters are used w i d e l y in many fields, such

as image processing, and geophysics for the p r o c e s s i n g of seismic,

gravity and magnetic data. It is expected that higher

dimensional filters will also find applications; for example,

three-dimensional filters in holography. The stability tests

which have emerged are of a N y q u i s t - t y p e and it is thought that

 141

a deeper understanding of t h e s e d e v e l o p m e n t s by c o n t r o l e n g i n e e r s

may lead to s u i t a b l e adaption for u s e in the c o n t r o l field;

possibly in the s t u d y of s t a b i l i t y u n d e r m u l t i - p a r a m e t e r

variations.

A constant theme throughout this w o r k has b e e n the

a s s o c i a t i o n of a m u l t i v a r i a b l e s y s t e m w i t h one or p o s s i b l y

more algebraic f u n c t i o n s e a c h of w h i c h is d e f i n e d on an

appropriate Riemann surface. A Riemann surface for an

algebraic function is t o p o l o g i c a l l y equivalent to a s p h e r e

with "handles", and the n u m b e r of h a n d l e s is k n o w n as the

g e n u s n u m b e r of the surface [6] . The g e n u s n u m b e r m a y p r o v e

to be an i m p o r t a n t c h a r a c t e r i s t i c of a s y s t e m and it is felt

that it m a y be r e l a t e d to " d e c o u p l i n g " , that is, the t r a n s f o r m a t i o n

of a m u l t i v a r i a b l e s y s t e m to a s y s t e m w h i c h is e f f e c t i v e l y a set

of s i n g l e - i n p u t , single-output systems. A single-input, single-

output s y s t e m has a c o r r e s p o n d i n g trivial (one sheeted) frequency

s u r f a c e of genus zero, so that to d e c o u p l e a multivariahle

s y s t e m w o u l d be e q u i v a l e n t to r e d u c i n g an m - s h e e t e d Riemann

surface, w i t h a g e n u s n u m b e r g r e a t e r t h a n or e q u a l to zero,

to a set of m trivial Riemann surfaces e a c h of zero genus.

Consequently the e s t a b l i s h m e n t of r e l a t i o n s h i p s between branch-

point singularities (whose n a t u r e a n d l o c a t i o n d e t e r m i n e such

topological properties as the g e n u s number) and decoupling

c o u l d be a f r u i t f u l line of r e s e a r c h . In a p p e n d i x 7 an

interesting relationship is d e v e l o p e d w h i c h shows that a b r a n c h

p o i n t on a g a i n s u r f a c e c o r r e s p o n d s to a b r a n c h p o i n t or

s t a t i o n a r y p o i n t on the c o r r e s p o n d i n g frequency surface and

v i c e versa.
 142

References

[i] G.A.Bliss, "Algebraic Functions", Dover, New York,1966

(reprint of 1933 original).

[2] H. Kwakernaak, and R. Sivan, "Linear Optimal Control

Systems", Wiley, New York, 1972.

[3] R.H. Cannon, Jr, "Dynamics of Physical Systems",

McGraw-Hill, New York, 1967.

[4] O.I.Elgerd, "Control Systems Theory", McGraw-Hill,

New York, 1967.

[ 53 E. I. Jury, "Inners and Stability of Dynamical Systems",

Wiley, New York, 1974.

[ 6] G. Springer, "Introduction to Riemann Surfaces",

Addison-Wesley, Reading, Mass., 1957.
 142

References

[i] G.A.Bliss, "Algebraic Functions", Dover, New York,1966

(reprint of 1933 original).

[2] H. Kwakernaak, and R. Sivan, "Linear Optimal Control

Systems", Wiley, New York, 1972.

[3] R.H. Cannon, Jr, "Dynamics of Physical Systems",

McGraw-Hill, New York, 1967.

[4] O.I.Elgerd, "Control Systems Theory", McGraw-Hill,

New York, 1967.

[ 53 E. I. Jury, "Inners and Stability of Dynamical Systems",

Wiley, New York, 1974.

[ 6] G. Springer, "Introduction to Riemann Surfaces",

Addison-Wesley, Reading, Mass., 1957.
 143

Appendix i. Definition of an a l g e b r a i c function

Let A(q,v) be a p o l y n o m i a l in q of the form

A(q,v) = fo(V)q m + f l ( v ) q m - l + .... +fm(V) (Al.1)

where each c o e f f i c i e n t {fi(v): i=l,2,...,m} is itself a

polynomial in v w i t h coefficients in the d o m a i n of c o m p l e x

numbers. T h e n an a l g e b r a i c function is a f u n c t i o n q(v)

defined for v a l u e s of v in the c o m p l e x v- p l a n e by an e q u a t i o n

of the form

A(q,v) = O (AI.2)

The p o l y n o m i a l A(q,v) can be r e w r i t t e n as a p o l y n o m i a l in v

with coefficients which are t h e m s e l v e s polynomials in q, and

when considered in this w a y e q u a t i o n (Ai.2) defines an a l g e b r a i c

function v(q).

For a fixed v a l u e of v, v O say, equation (AI.2) has

m solutions which are called branches of q(v), and in the

neighbourhood of v the b r a n c h e s are r e p r e s e n t a b l e by p o w e r

o
series expansions [i].

It is a s s u m e d in the above d e f i n i t i o n that A(q,v) is

an i r r e d u c i b l e polynomial in (q,v), that is, that A(q,v) is

not the p r o d u c t of two or more p o l y n o m i a l s in (q,v). If

A(q,v) was e x p r e s s i b l e as a p r o d u c t of t p o l y n o m i a l s in (q,v),

equation (AI.2) would then d e f i n e t algebraic functions

of the form q(v), or t a l g e b r a i c functions of the form v(q).

A p p e n d i x 2. A reduction to the i r r e d u c i b l e rational canonical form

In this a p p e n d i x a proposed method is g i v e n for

reducing any m - s q u a r e m a t r i x to its i r r e d u c i b l e rational

canonical form. The m e t h o d u s e d is a v a r i a t i o n on that g i v e n

 144

P ~
by A y r e s [2J for finding the r a t i o n a l canonical form of a

square matrix. The r a t i o n a l canonical form of a s q u a r e

matrix G is s i m i l a r to the i r r e d u c i b l e form e x c e p t that its

diagonal blocks correspond to the i n v a r i a n t factors of gI-G,

rather than the i r r e d u c i b l e factors. The p r o c e d u r e given

b y A y r e s [2] for f i n d i n g the r a t i o n a l canonical form is

outlined b e l o w along w i t h some n e c e s s a r y definitions.

Definitions: If the v e c t o r s

X,GX,G2X,...,Gt-Ix (A2.1)
are l i n e a r l y independent but
2
X,GX,G X,...,Gt-Ix,GtX (A2.2)
are not then (A2.1) is c a l l e d a c h a i n of length t having X

as its leader.

Procedure: For a g i v e n m - s q u a r e matrix G over any f i e l d F :

(i) let X m be the leader of a chain C m of m a x i m u m length

for all m - v e c t o r s over F ;

(ii) let Xm_ 1 be the leader of a c h a i n Cm_ 1 of m a x i m u m

length (any m e m b e r of w h i c h is l i n e a r l y independent of the

preceding members and those of Cm) for all m - v e c t o r s over F

which are l i n e a r l y independent of the v e c t o r s

; of C
m
(iii) let Xm_ 2 be the l e a d e r of a chain Cm_ 2 of m a x i m u m

length (any m e m b e r of w h i c h is l i n e a r l y independent of the

preceding members and those of C m and Cm_ I) for all m-

vectors over F which are l i n e a r l y independent of the v e c t o r s

of C m and Cm_ 1 ;
and so on. Then, for
t.-I tj+l-iXj+l ;
E = [ Xj,GXj .... G 3 ~ j + l , G X j + 1 .... ,G

t -i
... ; X m , G X m ..... G m Xm ]
 145

we have that E-IGE is the rational canonical form of G.

In this approach the chains of m a x i m u m length are used

to pick out the invariant factors. Now the invariant factors

are made up from products of the irreducible characteristic

equations which are required in the irreducible rational

canonical form. Therefore, if instead of using chains of

maximum length those of m i n i m u m length are found, the

t r a n s f o r m a t i o n m a t r i x E so formed is that required to give

the irreducible rational canonical form Q.

The problem with both of these methods is that no

indication is given w h i c h enables one to know when a chain of

m a x i m u m or m i n i m u m length has been obtained, except that in

the contrary case there will appear a chain of longer or shorter

length.

It is interesting to note that the chains of m a x i m u m

and m i n i m u m length form bases for the G - i n v a r i a n t subspaces

of m a x i m u m and minimum dimensions respectively. Therefore

the p r o b l e m of finding the chains of m i n i m u m length is

equivalent to that of finding the invariant suhspaces of

minimum dimension. It also happens that just as the l-

dimensional invariant subspace (eigenvector) picks out an

eigenvalue, a t-dimensional invariant subspace (from the set

of those of m i n i m u m dimensions) picks out an irreducible

equation of degree t, defining t eigenvalues.

Example: To find the irreducible rational canonical form of

S(s) = (s+3) (s+2) 2 O

(s+l) 2 (s+l)

(s-3) -2 0
2 2
(s+l) (s+l)

-i - (s+2) 1
2 - - 2 s+l
(s+l) 2 (s+l)
 146

Let X = ( O 0 1 )t, where "t" denotes the transpose,

1
then GX = (s+l) X

and X is a chain of minimum length.

t
Let Y = ( O 1 0 )

then Gy = [ s+2 -2 -(s+2) I t

(s+l) 2 (s+l) 2 2 (s+l) 2

and G2y = [ s+2 s-2 -(s+2) 1 t 1 s

(s+l)3 (s+]93 2(s+i)3 - (s-~ GY + --(s+l)3 Y

No chain of smaller length can be found and therefore Y,GY

completes the set of minimum length chains. The transformation

matrix E(S) is therefore given by

s+2
-[x Y ='o o 2
(s+l)
-2
O 1 2
(s+l)
-(s+2)
1 O 2
2 (s+l)

and

Q(s) = E-I(s)G(s)E (s)

= s+l O O

0 0 s
(s+l) 3
1
O 1 (s+l)

which implies that the irreducible characteristic equations

are
1
(g,s) = g s+l
2 1
~(g,s) = g (s+l) g (s+l)

Note that these are also obvious from the dependence relations

obtained when finding the chains of minimum length.

 'r47

A~pendix 3. The d i s c r i m i n a n t

In this a p p e n d i x two m e t h o d s are g i v e n for f i n d i n g

the d i s c r i m i n a n t of an e q u a t i o n of the f o r m

~(g,s) = bo(S)gt + bl(s)gt-l+ ... +bt(s) = O

Method 1 (Barnett [ 3 ] )

The r e s u l t a n t R[ a(g),c(g)] of two p o l y n o m i a l s

a(g) and c ( g ) , g i v e n by
n-l+
a(g) = a o g n + alg ... + a n
m m-i
c(g) = Cog + clg + ... + c m

where ai,c i e C ,is the d e t e r m i n a n t

aO aI a2 ... an 0

0 a aI ... an_ 1 an

m
rows
a -i a
n n

R[a(g),c(g)] = .oo
co c I .. Cm_ 1 cm

. , o C
o cI c 2 0 . cm O

ooo n
rows
co C1 c2 ..-

The p o l y n o m i a l s a(g) a n d c(g) have a common factor (of

d e g r e e g r e a t e r t h a n zero) if and o n l y if the (n+m) -order

determinant R [ a(g),c(g) ] is zero, provided that a

and c o are not b o t h zero.

L e t the d e r i v a t i v e w i t h r e s p e c t to g of the p o l y n o m i a l

a(g) be d e n o t e d by a' (g). T h e n the d i s c r i m i n a n t of the

polymonial a(g) is the d e t e r m i n a n t D g ( a o , a l , . . . , a n) d e f i n e d

 148

by
_ 1 R [ a(g) a' (g) ]
Dg(ao'''''an) a
0

The p o l y n o m i a l a(g) has a r e p e a t e d factor if and only if

the d i s c r i m i n a n t D g ( a o , . . . , a n) is zero.

Now consider a polynomial of the type

@(g,s) = b o ( S ) g t + b l ( S ) g t-I +...+bt(s) , t > O

where the coefficients {bi(s) :i=l,2,...,t} are all

polynomials in s. The d i s c r i m i n a n t of this p o l y n o m i a l in

g is found from D g ( b o , b l , . . . , b t) as d e f i n e d above by r e p l a c i n g

b o , . . . , b t by bo(S),...,bt(s) respectively. Thus there is a

function D (s) again called the d i s c r i m i n a n t and defined by

g
_ 1
Dg(S) b (s) R[~(g,s),~'(g,s)]
O

where #'(g,s) is the d e r i v a t i v e with respect to g

of ~(g,s)

Method 2 (Sansone and G e r r e t s e n [4])

Consider the p o l y n o m i a l
n algn-i n > O
a(g) = dog + + ... + a n ,

where ai e ~ ; then the d i s c r i m i n a n t of a(g) is given by

the e x p r e s s i o n

Dg(ao'''''an ) = ao2n-2p

where P is a d e t e r m i n a n t given by

P = So ~I "" On-1

... (~
G1 o2 n

an-i gn --- a2n_ 2

and the elements {~ : i = 1,2,...,2n-2} are functions of

i
 149

the coefficients {a. : i = O,l,...,n} , and

l

= n
o

The elements o1,...,an_ 1 can be found from

aI + ao~ 1 = 0

2a 2 + al(l 1 + ao~ 2 = 0

(n-l)an_ 1 + an_2~ 1 ... + aoan_ 1 = 0

and On,...,a2n_2 from

anO o + an_l$1 + ... + aoa n = O

ana I + an_lO 2 + ... + aoOn+ 1 = 0

+ + = 0
anO m a n_ i m + l + .. aoOn+ m

Consider now the polynomial ~(g,s) given by

~(g,s) = bo(s)g t + bl(S)g t-I + ... + bt(s)

The discriminant of this polynomial in g is found from

D g ( b o , . . . , b t) as defined above by replacing bo,...,b t

by bo(S),...,bt(s) respectively. The polynomial ~(g,s)

therefore has repeated factors if a n d only if

Dg(S) = . 2t-2
DO (S) P

is zero, where P is the determinant of a matrix whose

elements are functions of the coefficients {bi(s) : i = O,l,2,...,t}

 150

A p p e n d i x 4. A m e t h o d for c o n s t r u c t i n g the R i e m a n n surface

domains of the algebraic functions c o r r e s p o n d i n 9

to an o p e n - l o o p gain m a t r i x G(s)

In s u b - s e c t i o n 3.3-4 it was shown that for a c h a r a c t e r i s t i c

gain f u n c t i o n of degree t the c o r r e s p o n d i n g R i e m a n n surface

is made up from t sheets of the complex s-plane stitched

together along cuts made b e t w e e n b r a n c h points and infinity.

A l t h o u g h the cuts are in some sense a r b i t r a r y (i.e. there is

no unique set of cuts), it is still a p r o b l e m to choose a

set that is consistent, and then to be able to identify in w h a t

order the sheets are connected together. In this a p p e n d i x

a systematic m e t h o d is given for solving this problem. The

m e t h o d is quite elegant in that the r e s u l t i n g cuts are symmetrical

about the real axis and are always p a r a l l e l to the i m a g i n a r y

axis except for possible cuts along the real axis. Also,

the a p p r o a c h uses the o p e n - l o o p gain m a t r i x directly, so

that there is no need to find the c h a r a c t e r i s t i c equations,

and the resulting n o n - c o n n e c t e d sets of c o n n e c t e d sheets

r e p r e s e n t the R i e m a n n surfaces for the irreducible c h a r a c t e r i s t i c

equations.

The m e t h o d is based on finding the e i g e n v a l u e s of the

m a t r i x for a grid of values covering the s-plane and then

sorting the e i g e n v a l u e s in a continuous form along certain

lines of the grid. If the transfer function is of order mxm,

m arrays which will e f f e c t i v e l y r e p r e s e n t the m sheets of

the R i e m a n n surfaces are n e e d e d to store the eigenvalues.

The p ~ o c e s s is analogous to the analytic c o n t i n u a t i o n

p r o c e d u r e d e s c r i b e d in s u b - s e c t i o n 3.3-4 where i n d i v i d u a l
 151

points of a circular disc are made bearers of unique functional

values. Here i n d i v i d u a l points of the s-plane sheets

(arrays) are being made the bearers of functional values.

The first line along w h i c h the eigenvalues are c a l c u l a t e d

and sorted is the real axis, and then the c a l c u l a t i o n and

sorting is carried out along lines p a r a l l e l to the imaginary

axis and e m a n a t i n g from the real axis, as shown in figure 42.

The c a l c u l a t i o n is only necessary in the upper h a l f - p l a n e

since the e i g e n v a l u e s in the lower h a l f - p l a n e are the complex

c o n j u g a t e of those in the upper half. C o n t i n u i n g this process

the s-plane is covered, and m arrays of e i g e n v a l u e s are

o b t a i n e d which are continuous along the real axis and along

lines parallel to the i m a g i n a r y axis but not n e c e s s a r i l y

crossing the real axis.

I m (s)
5- plane

indicates continuity
between eigenvalues

Re (s)

Figure 42. Lines atong which eigenvatues

ore catcutated and sorted

By o b s e r v a t i o n of each array or sheet the n e c e s s a r y cuts

are obvious. If p a r a l l e l to the imaginary axis a line of

e i g e n v a l u e s is not continuous w i t h an adjacent line then

these m u s t be separated by a cut starting at a b r a n c h point

and e n d i n g at a branch point or infinity as shown in

 152

figures 43 (a) and (b). Eigenvalues corresponding to v a l u e s

of s in the upper h a l f - p l a n e are c o m p l e x conjugate to those

in the lower h a l f - p l a n e and t h e r e f o r e continuity of e i g e n v a l u e s

across the real axis is i m p o s s i b l e if the e i g e n v a l u e s situated

on the real axis are complex. Therefore a cut a l o n g the

real axis is n e c e s s a r y w h e n e v e r the real axis e i g e n v a l u e s

are complex. Again all the cuts start at a b r a n c h point

and end at a b r a n c h p o i n t or infinity, as s h o w n in f i g u r e 44.

X x indicates continuity

between eigenvalues |to00

i
I
i Cut
~--~ Cut

BrQnch points
/- i

~T--~-
i
< I--" Branch point
t-- --~--

(a) (b)
Figure 43. Cuts parallel to the imaginary axis
(a) finite (b) infinite

Cut. Im (s) Cut

R = Real invalue

Figure 44. Cuts along the real axis

 153

The stitching together of the sheets also becomes

obvious by matching eigenvalues on one side of a cut on one

sheet to those on another. When the matching process is

complete, in general, there will be sets of c o n n e c t e d sheets

each set defining a Riemann surface.

To facilitate the i d e n t i f i c a t i o n of cuts and the

matching of sheets it is very useful to draw the constant

phase and constant m a g n i t u d e contours on each sheet; this

is d e m o n s t r a t e d in the examples given in the main text.

Computationally sorting the e i g e n v a l u e s along the real

axis can be difficult because of the likelihood of real

axis poles. This p r o b l e m is overcome if the first line of

c a l c u l a t i o n and sorting is changed to be a line parallel

to, and a "large" distance away from, the real axis. The

rest of the calculations and sorting are then carried out

along lines parallel to %he imaginary axis starting at this

new line and finishing at the real axis; as shown in figure

45. With this new approach the only possible cuts are

either along the real axis, as before (see figure 44) or b e t w e e n

complex conjugate branch points as shown in figure 46.

A Riemann surface c o n s t r u c t i o n for

G(s) = 1 5s-2 2s-l]

1.25(s+i) (s+2) !
3s-18 s-8J

is shown using the original m e t h o d in figures 47 and 48, and

using the m o d i f i e d m e t h o d in figures 49 and 50, and illustrates

the arbitrariness of the cuts.

 154

Ira(s}

,Y, X X X X

~K X
S -plane

) :K
I

z z :_
Re is)

Figure 45. Lines orang which eigenvaiues are calculated

and sorted (modified method)

Figure 46. Cut parattet to the imaginary axis

(modified method)
 155

..... Root loci

i Cut

Figure 47. Sheet 1 of the Riemann surface

Cut

3_o

Figure 48. Sheet 2 of the Riemann surface

 156

,,*~.~ Root loci

Cut
e

Figure 49. Sheet 1 of the Riemann surface

(modified method of construction)

;~;w---.~'~': R o o t loci

Cut

Figure 50. Sheet 2 of the Riemann surface

(modified method of construction)
 157

Appendix 5. Extended Principle of the Argument

The required extension of the Argument Principle does

not seem to be readily available in the literature, although

a suitable statement of the Principle for a general m u l t i p l e -

valued analytic function has recently appeared in a text by

Evgrafov [5, page 98]. Therefore, to justify its use in

chapters 4 and 5, an appropriately extended Principle of the

Argument is developed here.

A5.1 Introduction

The extension required is non-trivial, with two main

problems to be overcome. The first of these arises from

the fact that, in general, the Riemann surface of an algebraic

function will be multiply connected. For an example of this

source of difficulty consider a c h a r a c t e r i s t i c gain function

g(s) associated with a 2 2 open-loop gain m a t r i x G(s).

Suppose that g(s) has four branch points, and that each

branch point is associated with a cycle of two sheets [6];

then the genus number [6; 7] of the a s s o c i a t e d Riemann surface

is one. Further suppose that these branch points are disposed

in the frequency plane (s-plane) in the way shown in figure 51.

Then, taking two copies of the complex number sphere (Riemann

number sphere), making cuts between the branch points, and

forming the topological equivalent of the Riemann surface

in the usual way [7] one obtains a torus, as illustrated in

figures 52 and 53. The region ~, shown shaded on this torus

and having a boundary ~, corresponds to the interiors of the

pair of Nyquist D-contours (shown in figure 52) for the

original complex number spheres out of w h i c h the torus was

constructed. To cope with such a situation we must ensure

that the extended version of the Argument Principle holds for

 158

[s)~ ~-=Bronchpoint

Figure 51. Frequency ptane

~ branch
_

-
Cut between
points

Figure 52. Two copies of the comp{ex number sphere

Figure 53. Riemann surface shown topologically

equivalent to a torus
 159

a region of a Riemann surface which is n o n - s i m p l y connected

and whose boundary ~ consists of several distinct closed

Jordan contours on the surface. This basic extension of the

principle is achieved via a suitable g e n e r a l i z a t i o n of the Cauchy

Residue Theorem and is covered in section A5.2.

The second obstacle to be overcome in a derivation of a

suitably extended Argument Principle is more directly

associated with the branch points of an algebraic function; in

this case with their effect on the calculation of residues

via contour integrals taken round the b o u n d a r y ~ of a region

having branch points in its interior. This p r o b l e m is

overcome by a change of variable and is treated in section A5.3.

Finally, in section A5.4 the results of sections A5.2 and

.3 are combined to give the required extended Principle of the

Argument.

A5.2 Generalized form of Cauchy's Residue Theorem

The basic requirement is a g e n e r a l i z a t i o n of Cauchy's

Residue Theorem to complex functions defined on the Riemann

surface of some known algebraic function, where the region

of integration may be non-simply connected and have a boundary

consisting of one or more closed Jordan contours. Such a

generalization may be found in Bliss [ 6 ] ; the main theorem

is given below with slight rephrasing and change of notation

for the context of this work Two observations may be useful

in helping the reader u n f a m i l i a r with Riemann surface theory to

understand this theorem.

(i) Each n o n - s i n g u l a r place (or point) on the Riemann

surface of an algebraic function f(s) is uniquely defined by a

 160

pair of values (f,s) which satisfy a known algebraic equation

A(f,s) = 0

Thus at each point on the Riemann surface the value of a

complex function T(f,s) of the pair of complex variables s

and f is defined.

(ii) The Riemann surface for an algebraic function is

an orientable surface. Thus one can u n a m b i g u o u s l y define a

positive sense in which a boundary ~ can be traversed so as

to "enclose" some region ~ . The simplest way in which to

visualize this procedure is to imagine someone walking round

the boundary ~ on a t w o - d i m e n s i o n a l manifold in which

lies (embedded in a familiar space of three dimensions).

The positive sense of traversal may then be taken as that in

which someone walking forward round any p o r t i o n o f ~ will

always have ~ on his right-hand side. In figure 53 the boundary

is shown with a positive orientation.

G e n e r a l i z e d Residue T h e o r e m

Let R be the Riemann surface of an algebraic function

f(s) and ~ a portion o f ~ not n e c e s s a r i l y simply connected but

having a boundary ~ consisting of one or m o r e closed Jordan

contours not passing through any singular place o n ~ .

Then if a function ~(f,s) of the places o n ~ is analytic on

and ~ except possibly for a finite number of singular points

in ~, we have

2~j" ~(f,s)ds = Z residues of ~(f,s) in a

~ (A5.2. i)

where the boundary ~ is traversed in the positive sense with

respect to ~.
 161

To derive the e x t e n d e d A r g u m e n t Principle we w i l l first

apply the G e n e r a l i z e d Residue Theorem to the f u n c t i o n f' (s)/f(s),

where f' (s) d e n o t e s the d e r i v a t i v e of f(s) with respect to s,

to give

1 /~ f' (s) ds = Z residues of f'(s)

f(s) in
2~j ~ f(s)

Note that f(s) m u s t have no poles or b r a n c h points (either of

which are r e f e r r e d to as s i n g u l a r places) on the b o u n d a r y ~.

For the p u r p o s e s of d e r i v i n g an A r g U m e n t Principle we w i l l n o w

also e x c l u d e any zeros of f(s) from ~. The next step is to

calculate the r e s i d u e s of f' (s)/f(s) in ~.

A5.3 Calculation of R e s i d u e s

The poles, zeros and b r a n c h p o i n t s of f(s) are s i n g u l a r i t i e s

f'(s)
of f(s) , a fact w h i c h w i l l b e c o m e clearer as this section

progresses. We w i l l c o n s i d e r all the b r a n c h p o i n t s of f(s)

but it turns out that only those w h i c h are also poles or

zeros of f(s) are relevant.

In the n e i g h b o u r h o o d of a pole or a zero (not i n c l u d i n g

those w h i c h are a s s o c i a t e d with a b r a n c h point) f(s) can be

represented by a series w h i c h defines a single-valued function.

For a pole Pi of o r d e r t p i t h e algebraic function can be

represented by

f(s) = E a n ( S - P i )n, a_t ~ O (A5.3.1)

n=-t Pi
Pi

~(s) t (A5.3.2)
i.e. f(s) = (s-Pi) Pi

where ~(pi ) ~ ~ and ~(s) is a n a l y t i c in the n e i g h b o u r h o o d

 162

of Pi" This gives

-t
f'(s) Pi
+ ~'(s) (A5.3.3)
f(s) ~pi ) ~(s)
!
<s)
and since is a n a l y t i c in the n e i g h b o u r h o o d of D.
"l
~(s)

f' (s) has a simple pole of r e s i d u e -t at

the f u n c t i o n
f(s) Pi

s = Pi"
For a zero s. of o r d e r t the a l g e b r a i c function can
i zi
be r e p r e s e n t e d by the series
co

f(s) = Z Cn (s-zi)n, ct O (A5.3.4)

n=t z.
Z. l
1
t
i.e. f(s) = (s - z i) zi e(s) (A5.3.5)

where e(z i) ~ 0 and 8(s) is a n a l y t i c in the n e i g h b o u r h o o d

of z.. This gives

l

t
f'(s) = zi S' (s)
+ (A5.3.6)
f(s) (s - z i) e(s)
f'(s)
and we see that f(s) has a s i m p l e ~ l e of r e s i d u e

t at s = z..
z. 1
1

To d e t e r m i n e the r e s i d u e s at a b r a n c h p o i n t it is

necessary to m a k e a change of v a r i a b l e [6, page 81]. The

procedure is as follows.

In the n e i g h b o u r h o o d of a finite b r a n c h p o i n t b.
l

the a l g e b r a i c function can be r e p r e s e n t e d by a series of the

form [8],
 163

oo

f(s) = Z dn(r - s/ ~ i )n (A5.3.7)

n___-o~

where r is the number of sheets which form a cycle at the

branch point. If the branch point is a pole of order

tb and consists of a cycle of r sheets we have the series
Pi

expansion

f(s) = n~_t~id n(r s - ~ i ) n , d_tbpi ~ O (A5.3.8)

Similarly if the branch point is a zero of order tb we

z,
1

have that
o0

f(s) = Z b dn(r s - ~ i ) n , dtb ~ O (A5.3.9)

n=t z. z.l
1
otherwise
0o

f(s) =
dn(r s _ ~ i ) n , do ~ O (AS. 3. iO)
n=O

These expansions are multivalued and therefore not

suitable for determining residues. If, however, we make

the substitution
s = b. + x r (A5.3.11)
l

we find that the series expansions (A5.3.8, 9 and i0)

become
f(x) = Z dnxn , d_tb ~ 0 (A5.3.12)
n=-t~i Pi

f(x) = n~tb dnxn' dtb O (A5.3.13)

Z,
Zi 1

and
 164

co

f(x) = Z darn, d O ~ O (A5.3.14)

n=O

respectively, where it is to be understood that

f(x) ~ f(b i + x r) (A5.3.15)

The substitution has therefore mapped the algebraic

function onto the x-plane where it can be represented by a

"single-valued" series expansion.

By definition [6] the residue at b I of (____{s)

f' analytic
f(s)

in a n e i g h b o u r h o o d of b. except possibly at b. itself is

1 1

A 1 Sf'(s) ds
(residue)bl = 249 - cf(~ (A5.3.16)

where C is a closed positively oriented Jordan contour o n e

bounding a n e i g h b o u r h o o d N of b i in which f'(s)

f(s) is

analytic except possibly at b i. If we substitute for s

from (A5.3.11) we find that

A 1 / f ' (s)
(residue)b. = 2~j I f ( s ) ds
i C

_ 1 / f' (x) dx ds
27 f(x) as
c
X

1 f f' (x) (A5.3.17)

dx
2~j , / f(x)
C
x

where C is a closed p o s i t i v e l y o~ented Jordan contour on the

X

x-plane bounding a n e i g h b o u r h o o d of the origin. The residue

f'( (s)
of s----~
f at a branch point b.1 is therefore equal to the

residue of f' (x) at the origin.

f(x) The residue of the function
 165

f'(x)
f(x) is e a s i l y o b t a i n e d f r o m its s e r i e s e x p a n s i o n . F o l l o w i n g

the same p r o c e d u r e as b e f o r e we n o t e t h a t f' (x) in the

f(x----[

neighbourhood of the o r i g i n has:

(i) a simple pole with residue -t b , if the b r a n c h

Pi
p o i n t is a p o l e of o r d e r t b ; or
Pi
(ii) a s i m p l e p o l e w i t h r e s i d u e t; , if the b r a n c h

p o i n t is a zero of o r d e r t b l;
zi

otherwise f'(x) is a n a l y t i c .
f(x)

A5.4 Extended Principle of the A r g u m e n t .

Combining the r e s u l t s of s e c t i o n A 5 . 3 w i t h that of

equation (A5.2.2) we h a v e t h a t

2~j f(s) i i Z Pi i Pi
~n
= WZ -~P (A5.4.1)

where # Z and ~ P are r e s p e c t i v e l y the total n u m b e r of p o l e s

and zeros of f(s) in ~ e n c l o s e d by the b o u n d a r y %~ t r a v e r s e d

in the p o s i t i v e sense w i t h r e s p e c t to ~, and w h e r e ~ lies

in the R i e m a n n s u r f a c e of the a l g e b r a i c function f(s). Note

that multiple poles and zeros m u s t be c o u n t e d as m a n y times

as t h e i r o r d e r s indicate.

The l e f t - h a n d side of e q u a t i o n (4.1) is e q u i v a l e n t to

the net sum of the c l o c k w i s e e n c i r c l e m e n t s of a c u r v e set F

about the o r i g i n of the f-plane w h e r e F is the image of ~ under

f(s). D e n o t i n g this n e t sum of c l o c k w i s e e n c i r c l e m e n t s by

N(F,O), we f i n a l l y o b t a i n the r e q u i r e d e x t e n d e d P r i n c i p l e of

the A r g u m e n t in the f o r m

N(F,O) = # Z - ~ p (A5.4.2)
 166

Appendix 6. Multivariable pivots from the c h a r a c t e r i s t i c

equation ~(g,s)=O

C o n s i d e r the c h a r a c t e r i s t i c equation

~(g,s) = O (A6. i)

and a s s u m e we h a v e f o u n d an a p p r o x i m a t i o n for a b r a n c h of the

characteristic frequency loci a b o u t s =m, as k ~m, of the f o r m

s - bk ~ (A6.2)

If p is the p i v o t corresponding to this a s y m p t o t e then for k =~

s = p + bk ~ (A6.3)
or 1
k : (A64)

The d e p e n d e n c e of the c l o s e d - l o o p p o l e s on k is g i v e n by

equation (6.1.5), i.e.

F(k,s) = O (A6.5)

Therefore substituting for k or s in e q u a t i o n (A6.5) we

obtain a relationship between s and p or a r e l a t i o n s h i p b e t w e e n

k and p. S i n c e the o r d e r of s w i l l in g e n e r a l be m u c h

g r e a t e r t h a n the o r d e r of k in e q u a t i o n (A6.5), we w i l l c o n s i d e r

the case of s u b s t i t u t i n g for k. Substituting for k from

equation (A6.4) into equation (A6.5) we o b t a i n the e q u a t i o n

Z(p,s) = 0 (A6.6)
-i
we r e q u i r e p for s =~, and so w e put s=z in e q u a t i o n (A6.6)

to g i v e

Z(p,s) = Z ( p , z -I) = z-tn(p,z) = o (A6.8)

w h e r e t is the m a x i m u m o r d e r of s in (A6.6). The e q u a t i o n

~(p,O) = O (A6.8)
 167

then gives the multivariable pivot, p.

Example
Consider the open-loop gain matrix

G(s)= 1 ~-s3-11s2-29s+92 -2Os2+35s+70

s4-s3+2s2-25s+29
[ 41s2-s-91 33s2-170s+l18
which has a characteristic equation
A(g,s)=(s4-s3+2s2-25s+29)g2-(-s3+22s2-199s+210)g +(-33s+594)=O

Using the method described in section 6.2 we find that

as k+~ the infinite branches of the characteristic frequency loci
are given by the following approximations
s~k and s~+/(-33k)
To find the pivot corresponding to the second-order

Butterworth pattern j/(33k) we will put

s=p+/(-33k)
giving
2
k=(S-P)
-33
From the characteristic equation ~(g,s)=O we obtain
F(s,k)=~4-s3+2s2-25s+29)+(-s3+22s2-199s+210)k+(-33s+594)k2=O
and substituting for k in this we obtain
Z(p,s)=332(s4_s3+2s2_25s+29)_33(_s3+22s2_199s+210) (_2ps +p2)
-33s2(22s2-199s+210)+(-33s+594) (-4s3p+6s2p2-4Sp3+p4)
+594s4=0
Putting s=z -1 we find
~(p,z)=332(l-z+2z2-25z3+29z 4)

-33(-l+22z-199z2+210z3) (-2p+p2z)-33(22-199z+210z 2)
+(-33+594z) (-40+6zp2-4z2p3+z3p4)+594=O
 188

Therefore

(p, O) = 3 3 2 - 6 6 p - 3 3 x 2 2 + 3 3 4 p + 5 9 4 = O

f r o m w h i c h we find

p=-14.5

Appendix 7. Association between branch points an d stationary

points on the v a i n and f r e q u e n c y s u r f a c e s

Let us s u p p o s e t h a t f r o m an o p e n - l o o p g a i n m a t r i x G(s),

we h a v e o b t a i n e d an a l g e b r a i c e q u a t i o n in t e r m s of the c o m p l e x

gain variable g, and the c o m p l e x frequency variable s, g i v e n b y

F ~ f(g,s) = O (A7.1)

Let us also c o n s i d e r the following equations i n v o l v i n g the

derivatives of F:

and

t ~F ~
<T J -- o CA7.3
g

The v a l u e s of s s i m u l a t a n e o u s l y satisfying equation

(A7.1 and .2) are the b r a n c h p o i n t s on the f r e q u e n c y surface;

the v a l u e s of g s i m u l t a n e o u s l y satisfying equations (A7.1and .3)

b e i n g the b r a n c h p o i n t s on the g a i n surface.

If we c o n s i d e r g as a f u n c t i o n of s, the t o t a l d e r i v a t i v e

of F w i t h r e s p e c t to s is

dF ~F ~F dg (A7.4)
--ds + ds
g s

Alternatively we can c o n s i d e r s as a f u n c t i o n of g, a n d o b t a i n

dF ~F SF ds
s g

N o w from e q u a t i o n (A7.1), F is i d e n t i c a l l y zero and therefore

 169

its total d e r i v a t i v e s m u s t also be zero, so that from e q u a t i o n s

(A7.4 and .5) we have the following:

8F DE dg (A7.6)
= ds
g s

and

-- ds
d-~ (A7.7)
s g

If we have a b r a n c h p o i n t on the gain surface equation

(A7.3) is s a t i s f i e d and h e n c e from e q u a t i o n (A7.6) we have

that

(~_FF~ = O or dg = 0
%~gJ s ds

which imply that on the f r e q u e n c y s u r f a c e we have e i t h e r a

b r a n c h p o i n t or a s t a t i o n a r y point corresponding to the gain

surface branch point.

Alternatively if we have a b r a n c h p o i n t on the f r e q u e n c y

surface equation (A7.2) is s a t i s f i e d and h e n c e from equation

(A7.7) we have that

(~~F
)s g = O or ds
dg o

which i m p l y that on the gain surface we have e i t h e r a b r a n c h

p o i n t or a s t a t i o n a r y point corresponding to the f r e q u e n c y

surface b r a n c h point.

References

i] B.A. Fuchs, and V.I. Levin, " F u n c t i o n s of a C o m p l e x V a r i a b l e " ,

I n t e r n a t i o n a l S e r i e s of M o n o g r a p h s in Pure and A p p l i e d M a t h e m a t i c s ,
P e r g a m o n Press, 1961 (translation of 1951 R u s s i a n original).

2] F.M. Ayres, " T h e o r y and P r o b l e m s of Matrices", Schaum,

N e w York, 1962.
 170

[3] S.Barnett, "Matrices in Control Theory", Van Nostrand-

Reinhold, London, 1971.

[4] G. Sansone, and J. Gerretsen, "Lectures on the Theory of

Functions of a Complex Variable", Vol. 2, Wolters-
Nordhoff, Groningen, 1969.

[53 M.A.Evgrafov, "Analytic Functions", Dover, New York, 1978.

[6~ G.A. Bliss, "Algebraic Functions", Dover, New York, 1966

(reprint of 1933 original).

[7] G. Springer, "Introduction to Riemann Surfaces" Addison-

Wesley, Reading, Mass., 1957.

[8] K. Knopp, "Theory of Functions", Part 2, Dover, New York,

1947.

Bibliography

AYRES, F.M., "Theory and Problems of Matrices", Schaum,

New York, 1962.

BARMAN, J.F., and KATZENELSON, J., Memorandum ERL-383,

Electronics Research Laboratory, College of Engineering,
Univ. of California, Berkeley, 1973.

BARMAN, J.F., and KATZENELSON, J., "A qeneralized Nyquist-

type stability criterion for multivariable feedback systems",
Int. J. Control, 20, 593-622, 1974.

BARNETT, S., "Matrices in Control Theory", Van Nostrand-

Reinhold, London, 1971.

BLISS, G.A., "Algebraic Functions", Dover, New York, 1966,

(reprint of 1933 original).

BODE, H.W., "Network analysis and feedback amplifier design",

Van Nostrand, Princeton, N.J., 1945.

BOHN, E.V. "Design and synthesis methods for a class of

multivariable feedback control systems based on single
variable methods", Trans.AIEE, 81, Part 2, 109-115, 1962.

BOHN, E.V. and KASVAND, T, " Use of matrix transformations

and system eigenvalues in the design of linear multivariable
control systems", Proc. IEE, llO, 989-997, 1963.

BOKSENBOM A.S. and HOOD, R, "General algebraic method applied

to control analysis of complex engine types", National Advisory
Committee for Aeronautics, Report NCA-TR-980, Washington D.C.,
1949.

BRACEWELL, R., "The Fourier Transform and Its Applications",

McGraw-Hill, New York, 1965.
 171

CANNON, R.H., Jr., "Dynamics of Physical Systems", McGraw-

Hill, New York, 1967.

COHN, P.M., "Algebra", Vol. i, Wiley, London, 1974.

DESOER, C.A., and CHAN, W.S., "The Feedback Interconnection

of Lumped Linear Time-invariant Systems", J. Franklin Inst.,
300, 335-351, 1975.

ELGERD, O.I., "Control System Theory", McGraw-Hill, New York,

1967.

EVANS, W.R., "Graphical Analysis of Control Systems", Trans.

AIEE, 67, 547-551, 1948.

EVANS, W.R., "Control System Synthesis by Root Locus Method",

Trans. AIEE, 69, 1-4, 1950.

EVANS, W.R., "Control System Dynamics", McGraw-Hill, New York,

1954.

FREEMAN, H., "A synthesis method for multipole control systems",

Trans. AIEE, 76, 28-31, 1957.

FREEMAN, H., "Stability and physical realizability considerations

in the synthesis of multipole control systems", Trans.AIEE, Part 2,
77, 1-15, 1958.

FUCHS, B.A., and LEVIN, V.I., "Functions of a Complex Variable",

International Series of Monographs in Pure and Applied Mathematics,
Pergamon Press, 1961 (translation of 1951 Russian original).

GANTMACHER, F.R., "Theory of Matrices", Vol. l, Chelsea,

New York, 1959.

GOLOMB, M.,and USDIN, E., "A theory of multidimensional servo

systems", J. Franklin Inst., 253(i) i 28-57, 1952.

HILLE, E., "Analytic Function Theory", Vol. i, Ginn and Co.,

U.S.A., 1959.

HILLE, E., "Analytic Function Theory", Vol. 2, Ginn and Co.,

U.S.A. 1962.

JURY, E.I., "Inners and Stability of Dynamical Systems",

Wiley, New York, 1974.

KALMAN, R.E., "When is a linear control system optimal?",

Trans.ASME J.Basic Eng., Series D., 86, 51-60, 1964.

KAVANAGH, R.J., "Noninteraction in linear multivariable

systems", Trans. AIEE, 76, 95-1OO, 1957.

KAVANAGH, R.J., "The application of matrix methods to multi-

variable control systems", J.Franklin Inst., 262, 349-367, 1957.

KAVANAGH, R.J., "Multivariable control system synthesis",

Trans. AIEE, Part 2, 77, 425-429, 1958.
 172

KONTAKOS, T., Ph.D. Thesis, University of Manchester, 1973.

KNOPP, K., "Theory of Functions", Part 2, Dover, New York, 1947.

KOUVARITAKIS, B., and SHAKED, U., "Asymptotic behaviour of root-

loci of multivariable systems", Int. J. Control, 23, 297-340, 1977.

KWAKERNAAK, H., "Asymptotic Root LOCi of Multivariable Linear

Optimal Regulators", IEEE Trans. Automatic Control, 21, 378-382,
1976.

KWAKERNAAK, H., and SIVAN,R., "Linear Optimal Control Systems",

Wiley, New York, 1972.

MACFARLANE, A.G.J., "Dual system methods in dynamical analysis

Pt. 2 - Optimal regulators and optimal servo-mechanisms",
Proc. IEE, 116, 1458-1462, 1969.

MACFARLANE, A.G.J., "Dynamical System Models", Harrap, London, 1970.

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, 2037-2049, 1970.

MACFARLANE A.G.J., and BELLETRUTTI, J.J., "The Characteristic

Locus Design Method", Automatica, 9, 575-588, 1973.

MACFARLANE, A.G.J., and KARCANIAS, N., "Poles and zeros of

linear multivariable systems: a survey of the algebraic, geometric
and complex variable theory", Int. J. Control, 24, 33-74, 1976.

MACFARLANE, A.G.J. and KOUVARITAKIS, B., "A design technique for

linear multivariable feedback systems", Int. J.Control, 25,
837-874, 1977.

MACFARLANE, A.G.J., KOUVARITAKIS, B., and EDMUNDS, J.M.,"Complex

variable methods for multivariable feedback systems analysis and
design", Alternatives for Linear Multivariable Control, National
Engineering Consortium, Chicago, 189-228, 1977.

MACFARLANE, A.G.J., and POSTLETHWAITE, I., "The generalized

Nyquist stability criterion and multivariable root loci",
Int. J. Control, 25, 81-127, 1977.
MACFARLANE, A.G.J.,and POSTLETHWAITE, I., "Characteristic
frequency functions and characteristic gain functions", Int.
J. Control, 26, 265-278, 1977.

MACFARLANE, A.G.J. and POSTLETHWAITE, I., "Extended Principle

of the Argument", Int. J. Control, 27, 49-55, 1978.

MAYNE, D.Q., "The Design of Linear Multivariable Systems",

Automatica, 9, 201-207, 1973.

MEES, A.I., and RAPP, P.E., "Stability criteria for multiple-

loop nonlinear feedback systems", Proc. IFAC Fourth Multivariable
Technological Systems Symposium, Fredericton, Canada, 1977.
 173

NYQUIST, H., "The Regeneration Theory", Bell System Tech. J.,

ii, 126-147, 1932.

OWENS, D.H., "A note on series expansions for multivariable

root-loci", Int. J. Control, 26, 549-557, 1977.

POSTLETHWAITE, I., "The asymptotic behaviour, the angles of

departure, and the angles of approach of the characteristic
frequency loci", Int. J. Control, 25, 677-695, 1977.

POSTLETHWAITE, I., "A generalized inverse Nyquist stability

criterion", Int., J. Control, 26, 325-340, 1977.

POSTLETHWAITE, I., "A note on the characteristic frequency loci

of multivariable linear optimal regulators", IEEE
Trans. Automatic Control, 23, 757-76o, 1978.

RAYMOND, F.H., "Introduction a l'4tude des asservissements

multiples simultanes", Bull. Soc. Fran. des Mecaniciens, 7,
18-25, 1953.

ROSENBROCK, H.H., "On the design of linear multivariable

control systems", Proc. Third IFAC Congress London, l, 1-16,
1966.

ROSENBROCK, H.H., "Design of multivariable control systems using

the inverse Nyquist array", P r o c ~ I E E , 116, 1929-1936, 1969.

ROSENBROCK, H.H., "State Space and Multivariable Theory", Nelson,

London, 1970.

ROSENBROCK, H.H., "Computer-aided control system design",

Academic Press, London, 1974.

SAEKS, R., "On the Encirclement Condition and Its Generalization",

IEEE Trans. on Circuits and Systems, 22, 780-785, 1975.

SANSONE, G., and GERRETSEN J., "Lectures on the Theory of Functions

of a Complex Variable", Vol. 2, Wolters-Nordhoff, Groningen, 1969.

SHAKED, U., "The angles of departure and approach of the root-

loci in linear multivariable systems", Int. J. Control, 23,
445-457, 1976.

SPRINGER, G., "Introduction to Riemann Surfaces", Addison-

Wesley, Reading, Mass., 1957.

WHITELEY, A.L., "Fundamental Principles of Automatic Regulators

and Servo Mechanisms", J. IEE, 94, Part IIA, 5-22, 1947.

WILLEMS, J.L., "Stability Theory of Dynamical Systems", Nelson,

London, 1970.