Lecture Notes in Control and Information Sciences: 107 Y. T. Tsay, L.-S. Shieh, S. Barnett

Lecture Notes in
Control and
Information Sciences
Edited by M.Thoma and A.Wyner
107
Y. T. Tsay, L.-S. Shieh,

S. Barnett
Structural Analysis and Design

of Multivariable Control Systems
An Algebraic Approach
Springer-Verlag
Berlin Heidelberg New York
London Paris Tokyo
S e r i e s Editors
M. T h o m a ° A. W y n e r
Advisory Board
L. D. Davisson • A. G. J. MacFarlane • H. Kwakernaak
J. L. Massey • Ya Z. Tsypkin - A. J. Viterbi
Authors
Yih Tsong Tsay
Leang-San Shieh
D e p a r t m e n t of E l e c t r i c a l E n g i n e e r i n g
U n i v e r s i t y of H o u s t o n
Houston, Texas 77004
USA
Stephen Barnett
S c h o o l of M a t h e m a t i c a l S c i e n c e s
U n i v e r s i t y of B r a d f o r d
W e s t Y o r k s h i r e B D 7 1DP
United Kingdom
ISBN 3-540-18916-5 Springer-Verlag Berlin Heidelberg New York

ISBN 0-38?-18916-5 Springer-Verlag N e w Y o r k Berlin Heidelberg
This work is subject to copyright. All rights are reserved, whether the whole or part of the material
is concerned, specifically the rights of translation, reprinting, re-use of illustrations, recitation,
broadcasting, reproduction on microfilms or in other ways, and storage in data banks. Duplication of
this publication or parts thereof is only permitted under the provisions of the German Copyright
Law of September 9,1965, in its version of June 24, 1985, and a copyright fee must always be paid.
Violations fall under the prosecution act of the German Copyright Law.
© Springer-Verlag Berlin, Heidelberg 1988
Printed in Germany
Offsetprinting: Mercedes-Druck, Berlin
Binding: B. Helm, Berlin
2161/3020-543210
Preface
Progress in system theory over the last two decades can be broadly categorized
into two main streams:
(I) Algebraic System Theory - Study of basic notions and fundamental concepts
of both algebra and system theory.
(2) System Analysis and Design Methods - Study of potential design techniques
to analyze the characteristics of systems and to design controllers for
satisfying various specifications and performance criteria.
Thousands of papers have been published in both areas in the last two decades.
Systemic presentations in book form can be found~ for example~ in [1-5] for the
former, in [6-10] for the latter~ and in [11-16] for both. From this
literature9 we find that many elegant theories still cannot be employed to
analyze/design the physical systems with ease. In other words~ work is still
needed to fill the gap between algebraic system theory and practical system
analysis/design techniques. This provides the main motivation for our
monograph.
The development of our work is based upon state-space representations and matrix
fraction descriptions as the mathematical models for physical systems. A
unified approach characterizing the dynamics of a system is presented through
the formulation of the characteristic %-matrix (also known as the matrix
polynomial) of the system. Applications in pole assignment design~ modal
control design for multivariable systems~ parallel realizations~ and cascade
realizations of multiport networks are illustrated. A detailed guide to the
content of the monograph is provided in the last section of Chapter I.

Acknowledgments
The a u t h o r s wish to express their gratitude for the fruitful discussions
with Dr. Jagdlsh Chandra, Director of Mathematics Science, US Acmy R e s e a r c h
Offlce~ Dr. Robert E. Yates, Director of the Guidance and Control Directorate,
US Army Missile Com=~and, and Dr. Norman P. Colemanj Chief of the Automation and
Robotics Group of the Fire Support Laboratory, US Army Armament Research,
D e v e l o p m e n t and E n g i n e e r i n g C e n t e r .
This research monography was f i n a n c i a l l y supported in part by t h e US Army
Research Office under contract DAAL-03-87-K-0001, the US Army M i s s i l e Command
under contract DAAH-01-85-C-Alll, and the NASA-Johnson Space Center under
contract NAGg-211.
TABLE OF CONTENTS
Chapter I. INTRODUCTION
i.I State-space representations and matrix fraction descriptions of

multivariable systems . . . . . . . . . . . . . . . . . . . . . . . . . i
1.2 Fundamental properties of X-matrices . . . . . . . . . . . . . . . . . 5
1.3 Organization of chapters . . . . . . . . . . . . . . . . . . . . . . . 8
CHAPTER 2. CHARACTERISTIC x-MATRICES AND CANONICAL MATRIX FRACTION

DESCRIPTIONS OF MIMO SYSTEMS
2.1 Characteristic X-matrlces of reachable and observable systems ..... 12
2.2 Canonical matrix fraction descriptions (MFDs) of MIMO systems ..... 18
2.3 Minimal nice selections for determining Kronecker and observability

indices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.4 Illustrative examples . . . . . . . . . . . . . . . . . . . . . . . . . 33
CHAPTER 3. SPECTRAL ANALYSIS OF NONSINGULAR X-MATRICES
3.1 Canonical X-matrices and canonical MFDs of MIMO systems ........ 39
3.2 Latent structure of nonsingular l-matrices .............. 46
3.3 Solvents of nonslngular X-matrices . . . . . . . . . . . . . . . . . . 61
3.4 Illustrative examples . . . . . . . . . . . . . . . . . . . . . . . . . 67
CHAPTER 4. DIVISORS AND SPECTRAL FACTORS OF NONSINGULAR k-MATRICES
4.1 Structure theorems for canonical left divisors and complete sets of
canonical left divisors . . . . . . . . . . . . . . . . . . . . . . . . 75
4.2 Structure theorems for canonical right divisors and complete sets of
canonical right divisors . . . . . . . . . . . . . . . . . . . . . . . 85
4.3 Structures theorems for spectral factorization of nonsingular X-

matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
4.4 Computational aspects of divisors and spectral factors for A-matrices 94
4.5 Illustrative examples . . . . . . . . . . . . . . . . . . . . . . . . . i12
CHAPTER 5. FEEDBACK CONTROL OF MULTIVARIABLE SYSTEMS
5.1 Linear state-feedback controls and properties ............. 123
5.2 X-matrix assignment . . . . . . . . . . . . . . . . . . . . . . . . . . 126

Vl
5.3 Left and right latent structure assignment . . . . . . . . . . . . . . 135
5.4 Divisor assignment and decoupling design of MIMO systems ....... 142
CHAPTER 6. STRUCTURAL DECOMPOSITION THEORIES AND APPLICATIONS IN

MULTIVARIABLE CONTROL SYSTEMS
6.1 Parallel decomposition theories and their applications to MIMO systems

and circuits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
6.2 Semi-cascade decomposition theories and their applications to modal

controls of MIMO systems ...... . . . . . . . . . . . . . . . . . 173
6.3 Cascade decomposition theories and their applications to multiport

network synthesis . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
BIBLIOGRAPHY . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200
INDEX . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
CHAPTER I INTRODUCTION
In this introductory chapter~ state-space representations and matrix
fraction descriptions of multivariable linear systems are reviewed in Section
I.I. Some basic definitions on k-matrices~ which are the main mathematical
tools used in our workj are summarized in Section 1.2~ and Section 1.3 gives a
guide to the content of the monograph.
I.i State-Space Representations and Matrix Fraction Descriptions of
Multivariable Systems
An m-input~ p-output linear time-invariant system o can be described by
state equations as follows:
~x(t) = Ax(t) + ~u(t)

(l.la)
(l.lb)
y(t) = cx(t) + Du(t)
where X(t)~x~cnp y(t)EF~C p, u(t)FJJ=_Cm are state, output, and input vectors,
respectively; ~ F~ U are state~ output and input spaces of O~ respectively;
A~BgCgD are matrices of appropriate dimensions. For continuous-time systems~
is a differential operator and tERj while for discrete-time systems~ A is a
forward shift operator and t~2.
Equations (l.la) and (l.lb) are referred to as the state-space
representation of the multivariable system O. A~B,C~ and D can be treated as
linear maps:
System map A: X÷X
Input map B~ U+X
Output map C: X÷F
Forward map D: U ÷ F (1.2)

2
From Eq. (1.2)~ ~ can be described by the following diagram:
X A = X
T
U
D ,~ ;,
(1.3)
which is not commutative.
The diagram in Eq. (1.3) is useful in studying the structural aspects of
the system O.
From Eq. (i.i)~ t h e input-output relationship of the system o can be
represented as
(1.4~)
y(t) = G(1)u(t)
where
G(I) = C(),l -A)-IB+D eCPXm(l) (l.4b)

n
In Eq. (l.4b) cPXm(1) denotes the set of pXm matrices with elements being
rational functions of ~ over the complex field C. G(I) is called the transfer
funct{on matrix of the system O. It has been shown in [i~13] that G(1) can be
represented as the "ratio" of two matrix polynomials:
G(I) = D[I(1)Nz(I) (l.5a)
= Nr(I)D~I(I) (l.5b)
where DI(I) EcP×P[~,], Ng(l) , Nr(l) ecPXm[)~], Dr(%)ecmXm[% ] ; cPxP[)~], Cpxm[)~] and
cmxm[l] are sets of matrix polynomials of ~ with coefficients in cPXP~ CPXm~ and
cmXm~ respectively. Combining Eqs. (1.47 and (1.5)7 yields

y(t) : D~l(X)N~(h)u(t)
(l.6a)
(l.6b)
: Nr(X)D~l(x)u(t)
Equations (l.6a) and (l.6b) are referred to as left matrix fraction descriptions
(LMFD) and right matrix fraction descriptions (RMFD) of the system O~
respectively.
Let TEC n×n be a nonsingular matrix~ and from Eq. (i.I) define
A = TAT -I , B = TB, C = CT-I~ D -- D (l. Ta)
and
~(t) = rx(t) (l.Tb)
Then the state equations for the system O are as follows:
^ A
Ix(t) = ~(t)+Bu(t) (l.Sa)

y(t) = CX(t)+Du(t) (l. Sb)
For the same set of inputs u(t), o in Eq. (i.i) and o in Eq. (1.8) will generate
the same set of outputs y(t) for t~0 if X(O) = TX(0). The difference between
the state vectors X(t) and (t) in the system O and O~ respectively~ is
therefore not apparent if only the input-output relationships are considered.

^
Thus~ we say that o and O are equivalent systems. Formally, we have the
following definition=
Definition I.i The system in Eq. (i.I) and the system in Eq. (1.8) are
equivalent if and only if the states are related by:
x(t) = TX(t)
We will call this equivalence relation similarity equivalence (SE). Q
Let UZ(1)EcPxP[I], and det Uz(1) = KZ which is a nonzero constant (i.e.
UZ(I) is unimodular). Define
D~(),) = Up(I)D~(I) (l.9a)

#%
N£(X) = U9 (),)N9(~) (l.9b)
and
G(I) = ~(~,) (1.9c)
which is an LMFD of a system ~:
y(t) = G(X)u(t) (i. I0)
From Eqs. (l.6a) and (I.i0), ~ in Eq. (l.6a) and O in Eq. (I.i0) will generate
the same set of y(t) for t~0 if the same set of u(t) is used as inputs~ and
^
and ~ both have the same set of initial conditions y(t), t<0. Thus, we say that
and ~ are equivalent systems. Similar reasoning can be applied for RMFDs. We
reach the following definitions:
Definition 1.2 Two systems with LMFDs G(~) = D~I(~)N~(~) and G(~) =
D~-I(%)N~(%) are equivalent if and only if
D~(X) = U~(X)D~(X)
and
5
where U~(%) is unimodular. Similarly, two systems with RMFDs O(~) m Nr(~)Drl(~)
and = are equivalont if and only if
Dr(1) = Dr(X)Ur(X)
and
NrC~) = NrC~)UrC~)
where Ur(% ) is unimodular. We will call this kind of equivalence relations
unimodular equivalence (UE). D
It can easily be verified that both SE and UE satisfy the basic properties
of an equivalence relation: transitivity, symmetry and reflexivity [13]. Since
a system can be represented via state-space equations or matrix fraction
descriptions, we have:
Lemma i.i Denote SE or UE by ~; Ox, Uy and ~z are systems. Then, we have
(i) Transitivity: Ox Uy and Oy O z implies Ox Oz .
(2) Symmetry: o ~ O implies O ~ O .

x y y x
(3) Reflexivity: °x ~ ~x" •
From the idea of equivalent systems, both state-space representations and
matrix fraction descriptions of multivariable systems are non-unique. In
Chapter II, we will develop canonical forms, which are unique for a given
system, for both state-space representations and matrix fraction descriptions.
1.2 Fundamental Properties of h-Matrices
Since the MFD representations of a MIMO (multi-input, multl-output) system
involve the ratio of two %-matrices~ and the results presented in the following
chapters are closely related to h-matrices, it is appropriate to review some

definitions of l-matrices in this section. Further details and properties can
be found, for example, in [2] and [3]. Specifically, we can define l-matrices
as follows [17,18]. Let F be an arbitrary field) and FiX] be the ring of
polynomials over the field F. A l-matrix, denoted by A(X)EFPxm[x] is a pxm
matrix whose elements are in Fill. Let A..(X) be the (i,~)th element of A(X))
i]
then
A(X) -- (Aij(X)) , l~i~p, l&j:m (l.lla)
and k,°
Aij(X) ~ ~J kij-k
k=0 aijkX ) aijk~F (l.llb)
where kij is the degree of the polynomial Aij(X).
Let r = Max(kij ) iKi~p, ISj~m)) then A(X) can be written as
r
A(X) = ~ AkXr-k (l.12a)
k=O
where AkcFpxm , and the (i,j)th element of A k is given by
(Ak)iJ = { aij k if k~kij (l.12b)

0 otherwise
A l-matrix A(X)EFP×P[I] is said to be nonsingular if det(A(X)) ~ 0, and
regular if the matrix coefficient A 0 of the highest degree term (referred to Eq.
(l.12a)) is nonsingular. A regular l-matrix is monie if A 0 is an identity
matrix.
Let A(X) be given by Eq. (I.ii), and define
~i = Max(kij' l~j~m), l~i~p (1.13)

7
Then, ~i' denoted by ~i = Dr. (A(I)), is the row degree [12] of the ith row of
1
A(X). Similarly
. = Max(kij , l~i~p), l~j~m

<3
denoted by Kj : ~c.(A(~)), is the column degree [12] of the jth column of A(X).
J
Define
(1.15a)
Ahr = ((Ahr)ij), IKiKp, IKjKm
where
aiJ~i if kij = ~i
(I. 15b)
(Ahr)ij
L 0 if kij < ~i
Then Ahr is called the leading row matrix of A(X). A(X) called is a row-reduced
X-matrlx if p=m and Ahr is nonslngular [12]. Similarly, defining
(l.16a)
Ahc = ((Ahc)ij) , l~igp, l~j~m
where
I aijKj if kij = ~j
(l.16b)
(Ahc) ij
L 0 if kij < <j
then Ahc is called the leading column matrix of A(X), and if p=m and Ahc is
nonslngular, A(X) is a column-reduced X-matrix [12].
To analyze the structure of ~-matrices, it is convenient to transform a
general l-matrix to certain specific forms whose structures can be easily
handled. The most commonly used transformations are those of equivalence
[12,13] .
Definition 1.3 Two ~-matrices Al(l ) and A2(1) are row equivalent, column
equivalent, or equivalent, iff AI(% ) = UL(I)A2(1) , Al(1) = A2(I)UR(I) , or AI(%)
= UL(~)A2(~)UR(1) , respectively, where UL(~) and UR(1) are unimodular l-
matrices. []
The equivalence of nonsingular l-matrices can be stated as follows:
Lemma 1.2 Any nonsingular h-matrix is row equivalent, column equivalent, or
equivalent to a row-reduced, a column-reduced~ or ~ row- and column-reduced l-
matrix. •
It is well known that equivalent row-reduced or column-reduced h-matrices
of a given nonsingular h-matrix [12] are not unique. According to Definition
1.3, a regular h-matrix is always equivalent to a monic l-matrix, and the
properties and applications of monlc l-matrices have been discussed by many
authors [17-32]. We shall extend some known results on monic h-matrices to row-
reduced or column-reduced l-matrices in the following chapter3.
In the analysis and design of multi-input, multi-output (MI~O~ systems~ M~D
representations of the systems are rational matrices over the co~plex field C.
Therefore, we will set F = C in the follow~ng chapters wheneve~ h-matrices are
involved.
1.3 Organization of Chapters
The material in this monograph can be regarded as being in two parts~ The
first part7 which includes Chapters II, III and IV~ is devoted to exploring the
spectral decomposition theory of h-matrices via the canonical structures of MIMO
systems represented in state space equations and MFDs~ the second part~ which
consists of Chapters V and VI, considers appl~cat~ons of the structure theory
developed in the first part to the design and decomposition of MIMO systems.
Illustrative numerical examples are presented throughout the book.
In Chapter II, the characteristic h-matrices of multivariable control

9
systems are defined. For a reachable system the characteristic X-matrix can be
constructed from the coefficients of the dependence equations for the column
vectors of the teachability test matrix; on the other hand~ for an observable
system~ the left characteristic X-matrix can be constructed from the
coefficients of the dependence equations for the row vectors of the
observability test matrix. The controller and observer canonical state-space
representations for reachable and observable MIMO systems~ respectively~ are
formally defined. The canonical RMFDs and LMFDs for reachable and observable
systems~ respectively~ are defined~ and their properties are discussed based on
the canonical controller and observer state-space representations. The
characteristic %-matrlces~ the canonical state-space forms~ and the canonical
MFDs are highly dependent on the Kronecker or observability indices of the
system. Thus~ we also present a numerical method using an orthogonalized
projection scheme to compute the Kronecker and observability indices of MIMO
system. This numerical algorithm is based on the so-called minimal nice
selections.
Spectral analysis of general nonsingular Z-matrices is presented in Chapter
Ill. Firstly 7 column-reduced and row-reduced canonical X-matrices are defined;
then the equivalent transformations of a nonsingular X-matrices to a column-
reduced or a row-reduced canonical X-matrlx are established. Consequently~ the
latent roots and latent structures of a general nonsingular %-matrix can be
studied in terms of its equivalent column-reduced or row-reduced canonical Z-
matrix. The relationships between the latent structures of nonslngular X-
matrices and the eigenstructures of the system maps in their associated state-
space minimal realization quadruples are investigated. As a result~ the Jordan
chains of nonsingular X-matrices can be easily found from the input and output
maps of their associated Jordan canonical minimal realization quadruples. The
matrix roots~ formally called solvents~ of nonsingular %-matrices are defined
and briefly discussed.

10
Chapter IV is devoted to developing the theory of divisors and spectral
factors of nonsingular l-matrices. The state-space structures of canonical left
and right divisors of nonsingular l-matrices are extensively investigated via
the so-called geometric approaches. Constructive proofs on the existence of the
canonical divisors are provided, and some properties of left/right divisors of
nonsingular l-matrlces are investigated. Also~ the concepts of complete sets of
canonical left/right divisors~ which are extremely important in the applications
to the design and decomposition of MIMO systems~ are presented. For
completeness~ the structures of spectral factorizations used to factor a
nonsingular %-matrix into the product of lower degree canonical l-matrices are
also explored. Finally~ computational algorithms for divisors and spectral
factors based on block triangularization and block diagonallzatlon of square
matrices are discussed. A newly developed matrix sign algorithm is suggested
for effective computation of divisors and spectral factors of nonsingular l-
matrices.
The applications of the theory begin in Chapter V~ where state-feedback
control designs of multivariable systems are studied. Properties of linear
state-feedback controls are discussed first. The invariance property of the
Kronecker indices of MIMO systems under linear state-feedback controls is an
important guide in devising various control schemes. The characteristic %-
matrix and column-reduced l-matrix assignments for the denominators of the
closed-loop MFDs are derived. A study is then made properties of the closed-
loop MFDs. For controlling the latent structure of the characteristic l-
matrices in the closed-loop system, we introduce the left/right latent structure
assignment. For the purposes of closed-loop decomposition, the divisor
assignment and decoupling design, via the notions of divlsors~ are also
presented.
Decomposition theories and their applications to multivariable analysis and
design are developed in Chapter Vl. Parallel decomposition theory is derived

11
based on the complete sets of left/right divisors of the characteristic ~-
matrices. Applications of the parallel decomposition theory to model reduction
problem and multiport network synthesis are discussed. The semi-cascade
decomposition theory as well as applications to the modal control design of MIMO
systems are presented via the notion of spectral factorization of characteristic
),-matrices. Finally~ the cascade decomposition theory for MIMO systems is
considered~ with applications to the cascade realization of multlport networks.

CHAPTER II CHARACTERISTIC k-MATRICES AND CANONICAL MATRIX FRACTION
DESCRIPTIONS OF MIMO SYSTEMS
It is well-known that the dynamics of single-input, single-output (SISO)
systems can be determined from their transfer functions and characteristic
polynomials, which are the denominators of the transfer functions. Many system
design techniques of SIS0 systems can actually be interpreted as altering the
transfer functions or the characteristic polynomials of the closed loop systems
to satisfy some desired dynamic performance criterion. To facilitate the same
concepts for MIMO systems, in this chapter we present formal definitions of the
characteristic k-matrices and the canonical MFDs for MIMO systems, which are the
counterparts of the characteristic polynomials and the transfer functions for
SISO systems, respectively. Properties as well as computational algorithms for
the characteristic k-matrices and the canonical MFDs are given. Relationships
between the state space representations and MFDs of MIMO systems are
established. The results of this chapter provide the foundations for the
spectral analysis of k-matrices and the decompositions of MIMO systems which are
presented in the following chapters.
2.1 Characteristic y-matrices of Reachable and Observable Systems
An m-input, p-output linear time-invariant system can be described by the
state equations as follows:
kX(t) = AX(t)+Bu(t) (2.1a)
y(t) = CX(t)+Du(t) (2.1b)
where X(t) ECn, y(t) E ~ and u(t) ECm are state, output, and input vectors,
respectively; A,B,C and D are matrices of appropriate dimensions. For
continuous-time systems, k is a differential operator and tc~ while, for

13
discrete-time systems, ~ is a forward shift operator and fEZ.
The system described in Eq. (2.1) is reachable if the reachabillty test
matrix [13]
R(A,B) = [B,AB,...,An-IB] (2.2)
is of full rank. The system in Eq. (2.1) is observable if the observability
test matrix [13]
0(A,C) = [cT,ATc T,...,(AT)n-IC T]
I] (2.3)
is of full rank.
Definition 2.1 Let B = [bl,b2,...,bm] ~ bieC nxl. The reachabillty base matrix
is defined by
<i-i ~ -I (2.4a)
P(A~B) =A [bl,Abl~" . o ~ A bl, . . ,bm,Abm~.
. . . ~A m b m ]
where <.'s are the teachability or Kronecker indices [32-37] of (A~B). D

1
Definition 2.2 T T
Let C = [CI~C29...~C ~ ]T CieclXn" The observability base matrix
is defined by
~(A,C) A T T T T-~I -I T C T ATcT ,(AT)~p-Ic~ T (2.4b)

= [CI~A CI~---,(A ) CI,..., p, p'''" ]
where V.'s are the observability indices [32-37] of (A~C). Q

1
14
Obviously, the columns (rows) of P(A,B) ( ~(A, O)) are contained in
~(A,B)(0(A,C)). Thus, p(A,B)(~(ApC)) can be obtained from E(A,B)(0(A~C)). The
selection of the columns (rows) of P(A,B)(Q(A,C)) from R(A,B)(0(A,C)) follows
the sequence proposed by Popov [35]. It is well known [33-36] that when (A~B)
m
is a reachable pair, rank[F(A,B)]=n and ~ <i=n; when.(A~C) is an observable
P i=l
pair, then rank[~(AgC) ] = n and [ 9.=n.
i= I i
Assume that (AgB) is a reachable pair with reachability indices ~j
K.
j=l,2,...,m. Then A 3b. can be uniquely represented as [35-37]
3
I(. j-I K. m min[<.,K.] _ .

AJb. =- ¥ v z 3 _K-t~
arij (<.+I)A 3b i- L k~ 1 arijka Di , if K.>O, l~j~m
-] i=l 3 i=l 3
~.<~.
(2.5a)
and
j-I
(2.5b)
b° -- -
ari'ibi'3 if <. = O, l<jKm
.1 i=l 3
~.>0
I
Note that the input matrix B may contain dependent vectors. As a result~ we can
make a general structural analysis of MIMO systems. Similarly, if (A,C) is an

V.
observable pair with observability indices Vi, i=l,2,...,p, t h e n C.A1 1 can be
uniquely represented as
~. i-I
C.A ~ = - ) ag.iJ(vi+l)CjA vi- ~ m~n[~)i'gJ]a~ijkCjAk-I , if Vi>0 , iKiO
3
15
It has been shown that the Kronecker indices <.l and the set of parameters
{arijk} are invariants of (A~B) under coordinate transformations of states, and
the ~. are also invariant under linear state feedback. Similarly, the
observahility indices ~'x and the set of parameters {a~ijk } are invariants of
(A,C) under coordinate transformations of states and the ~.x are also invariant
for the full state observer construction.
We are now ready to define the characteristic X-matrices of multi-input
multi-output (MIMO) systems [30-32].
Definition 2.3 The right characteristic X-matrix of a reachable system in Eq.
(2.1), or a reachable pair (A,B), is defined by
(2.7a)
Dr(X) ~ [(Dr(X))ij ] ecmXm[x], l~i~m, l~jKm
where the (ipj)th entries of D (~) are defined as

r
A (2.7b)
arii(~.+l ) = I
l
I ~k
if 0~ or <.=0z and i=j
arij(k+l)
(Dr(X))ij
0 if Ki=0 , izj
I Ki if i=j (2.7c)
A Kj if <j, i~j (2.7d)
<ij =
K.-I
l otherwise Q
An illustrative example is given in Section 2.4.
Definition 2.4 The left characteristic % matrix of an observable system in Eq.
(2.1)~ or an observable pair (A,C), is defined by

18
D£(1) ~ [(Dg(I))ij]EcPXP[x], l~iSp, Isj~p (2.8a)
where the (i,j)th entries of Di(I) are defined as
A (2.8b)
a~ii(~.+l ) = I
l
9..
k~ I agij(k+l)l k if 9.>0, or V.=O and i=j

(Dg(1))iJ ~ dgij (l) = { = 3 J (2.8c)
.0 if ~.=0 and i~j
3
(2.8d)
Vij = ~ if 9i<~j, i~j
t gj-I otherwise D
Since {arijk} , {aZijk} are invariant under coordinate transformations of
the states, the right characteristic l-matrix, Dr(1) , and the left
characteristic l-matrix, D~(~)~ are invariant under the coordinate
transformations of states [35-36]. Furthermore,

since <..~<. for l~i~m, the
•j J
column degree of the jth column of D (1) is ~.. Define the leadin~ column
r 3 "
matrix as
A (2.9a)
Drh = [(Drh)ij] , l~iKm, IKjKm
for ~.<~.
j x
A I arij(<j+l) (2.9b)
(Drh)ij = hrij = 0
otherwise
Drh is constructed using the coefficients of the highest powers of I in each
column. Since arii(<.+l ) - i, and arij(<i+l) = 0 for i>j, Drh is an upper

1
triangular matrix with diagonal elements all l's~ so det(Drh) = I. Thus, Dr().)
m
is a column reduced l-matrix and the degree of tier(Dr(1)) is ~ K. = n.
j=l J
17
Similarly, when VijN~ i for lKjNp, the row degree of the ith row of D~(k) is ~i"
Define the leading row matrix
(2.10a)
Dih = [(Dzh)ij] , lKiKp, IKjKp
i a£iJ(~i+l ) for ~i~Vj

(2.10b)
(D£h)ij ~ hlij = O, otherwise
D£h is constructed by the coefficients of the highest powers of ~ in each row.
Since a£ii(~i+l ) ~I, and a%iJ(~i+l ) = 0 for i<j, D£h is a lower triangular
matrix with diagonal elements all l's, so det(Dih) = i. Thus, D~(%) is a row
P
reduced l-matrix and the degree of det(D£(~)) is ~ ~. = n.
i= I i
From the definition of the right characteristic ~-matrix in Eq. (2.7) and
the properties of Drh in Eq. (2.9), we have the following results:
Proposition 2.1 The right characteristic l-matrix D (l) in Eq. (2.7) has the
r
following properties.
(a) d .,(~) ~ O, monic

rll
(b) deg drii(k) > deg drij(k) , i;.j
(c) deg drii(k) > deg drji(k) , j>i
(d) deg drii(X) ~ deg drji(k) , j<i
where drij(k) is the (i,j)th entry of Dr(1). D
Similarly, from Eqs. (2.8) and (2.10), we have:
Proposition 2.2 The left characteristic l-matrix D~(k) in Eq. (2.8) has the
following properties.
18
(a) d£ii(k) z O, monic
(b) deg d~ii(k ) > deg d~ji(k) , izj
(c) deg d£ii(k ) > deg d~ij(k) , j>i
(d) deg d~ii(k ) ~ deg d~ij(k) , j<i
where dEij(k ) is the (i,j)th entry of DE(X). []
2.2 Canonical Matrix Fraction Descriptions (MFDs) of MIMO Systems
As we have men=ioned in Section i.I, the MFD representations of MIMO
systems are nonunique. In order to specify the "standard" MFDs for an MIMO
system, we shall define the canonical right MFD for the reachable systems and
the canonical left I~FD for the observable systems in this section. The
relationships between the left/right characteristic X-matrices and the
left/right MFDs of MIMO systems are also discussed.
To find the RMFD of the reachable system in Eq. (2.1), it is convenient to
transform the state equations in Eq. (2.1) into the canonical controller form
[32-34] using the following similarity transformation:
(2.11a)
Xc(t) = TcX(t)
r ,(pl A ml -I )T ,.. pT J Km-I-T r (2.11b)

Tc = [PI' . . . . ' m''''~Pm A ) ]
where
P. = the o. th row of P-I(A,B), ~.>0 to 11

~,~.~c~
i l l
and
i
(L = ~° (2.11d)
I j=l 3
19
The state equations in the canonical controller form are:
(2.12a)
XXc(t) = AcXc(t)+Bcu(t)
y(t) = CcXc(t)+Dcu(t) (2.12b)
where
Ac ~ TcAT~I =A [(Ac)ij]' IKiKm, l~j~m, ~.<.>01J (2.13a)
r o
K~. XK~.
EC x x
; Acii = [_~riil,... _~rii<.] (2.13b)
1
_0 ] K.x<.
(Ac)ij ~ .... £C i ]; Acij
~
[-arijl'....-arijKj]' ~i~<j
iZj
Acij =i[-a--rijl, .,-ariJKi,0.....0], ~i<<j
(2.13c)
and
[ Acll" "'"'Aclm]" = -D-irh [ arlll,''',arll<l,''',arlml,''',arlm<
AcmI, •,Acmn J
m
armll,''',arml<l~''',armml~'''~armm<m
213d
Bc = TcB = EbcDr ; Ebc = [ebcl~ebc2'''''ebcm] (2.13e)
I
{ en if <.>01
(2.13f)
ebci 0 if <.=0
CT~I T T T T (2.13g)
Cc = [CcI,Cc2,...,Ccp]
20
A (2.13h)
Cci = [Cill,...,CilK.,...,Ciml,...,Cim ~ ]
i m
D ~ D (z.13i)
e
From Eq. (2.12)~ the input-output relationship of the system can be described as
(2.14a)
y(t) = [Cc(lln-Ac)-IBc+Dc]u(t)
Also, from Eq. (2.13), we have
( lln_Ac )-IBc = ~r (%) [Drh6r ( ~)] -I (2.14b)
where
(2.14c)
Qr(~) ~ [~r1(1), ....~rm(~)]
and
0 T ~m
[01xOi_l,l,l,..., l I , ix(n-oi )] , Ki>0,i=l,...
Qri(~)~{ 0nxl' ~.~0

i
(2.14d)
(2.14e)
~r(l) ~ ~rh(l)-Ar~r(~)
~rh(k) ~ diag[% 1,iffil,...,m] (2.14f)
A [A~I,A T .T .T ~a×n (2.14g)

Ar = r2~...,Arm] e~
o. th row of A for K.>0

(2.14h)
Ari =
01× n for K.=0
i
21
From Eq. (2.14a), we have
y(t) = [Nr(k)D~l(X)+Dc]U(r)
(2.15a)
where
Dr(k ) b Drh6r(k) (2.15b)
Nr(X) b CcSr(X5 (2.t5c)
Thus, the RMFD representation of the system in Eq. (2.15 becomes
(2.16)
G(X) in Eq. (2.165 is referred to the canonical RMFD of the system in Eq. (2.1)~
and D (X) is the right characteristic %-matrix of the RMFD. It is well known
r
that, for a reachable systemp Dr(X) and Nr(%) are right coprime. From Eqs.
(2.14c) and (2.15c), any column of Nr(X) corresponding to Ki'0 is a zero column.
~f the system in Eq. (2.15 is observables the state equations can be
transformed into the observer canonical [32-34] form by the similarity
transformation as follows:
X0(t 5 = T;Ix(t5 (2.17a5
~1-1 ~ -1
T0 . [ q l , . . . , A ql'''''qP''''' A p qp] (2.17b)
where
qi = the T.th column of ~-I(A,C), V.>0

(2.17c)
and
T.b j=l
i v.J (2.17d)
22
The state equations in the observable canonical form are as follows:
tx0(t) ffiA0X0(t)+B0u(t) (2.18a)
y(t) = C0X0(t)+D0u(t) (2.18b)
where
A 0 ~ T;IAT0 = [(A0)ij] , l<i~p, l<j~p, Vi•j>O (2.19a)
(A0) £i ~ 0 ,
-IAoIi cC~.x~.
i i = -- -- T (2.19b)
_ , A0i £ [-a0iil, •. •, - a 0 i £ u £ ]
-- -- T
V. xV. [-a0ijl .....-a0iju .] , if Vik~j, izj
(A0)Ij ~ [0 A0i j]~CI j; A0ij m -- J
[-a0ijl,...,-a0iJui,0,...,0lT if Vi<vj,izj
(2.19c)
and
A011'""A01p.
. • ] a~lll,''',aillgl,''',a£pll,''',a£plu p
(2.19d)
A0pl'" 'Aopp J a~Ipl~'''~a£1P~l'''''azppl'''',alpp~
(2.19e)
Bo ~ T~lB ~ [BOI,B02,...,~O~]
A
BOj = [ b l j l , . . . , b l j v i , . . . , b p j l , . . . , b p j ~ p ] T (2.19f)
- A T T T (2.19g)
CO ~ CT0 = D£~Ec0 ; Eco = [ec01,...,ec0p]
T.
l
en , if ~. >0
T ~ ~ (2.19h)
ecoi = 0 , if ~.=0
l
23
DO ~ D (2.191)
From Eq. (2.18), the input-output relationships of the system can be expressed
as
(2.20a)
yCt) = [Co(XIn-Ao)-IB0+Do]U(t)
From Eq. (2.19), we h a v e
CO(lIn_A0)-I = [~9(~)D~h]-l~(~) (2.20b)
where
(2.20c)
*~(~> =~ [~'Lc~:', .... *~,pc~)] '~
and
~.>0
1
ui'0I'A' "'X' z ,01x(n_Ti)],
~i (~) =Af01x-[01xTi-I' (2.20d)
~(~) =A ~h (~)-~( X)A~ (2.20e)
~h(~) =Adieg[~ z,i=l,...,p] (2.20£)
A~ =A [A~I,A~2 , . . . ,A~p] eCnxp (2.20g)
T . t h column o f A0 f o r V.>0
A i l (2.20i)
A£i =
0nX 1 f o r U.=0
24
Prom Eq. (2.20a), we have
y(t) = [D~l(k)NA(k)+V0]u(t) (2.21a)
where
D~(A) ffi ~(k)DEh (2.21b)
N~(X) - SE(k)B 0 (2.21c)
Thus, the LHFD representation of the system in Eq. (2.1) becomes
~(k) -- ,~I(k)N~C~)÷, o - ,il(k)~Ck))~(k) =AN~(k)÷O~(k), ° (2.22)
O(k) in Eq. (2.22) is referred to the canonical LHFD of the system in Eq. (2.1)
and D%(k) is the left characteristic k-matrix of the LMFD. D~(k) and N~(k) are
left coprime if Eq. (2.1) is an observable system. Note t~at avy row of N%(k)
corresponding Co Vi=0 is a zero row.
Prom Eqs. (2.13) and (2.14), the minimal realization of D-I( ~,,, using a pair
r
(A,B) can be formulated as follows.
Lemma 2.1 The quadruple (A)B)Cr)Dr) , where
Cr ~(0), c and Dr ffiA (~m~C0)~<0))o~
is a minimal realization of D l(k)) i.e. V l(k) ffi Cr(kln-A)-iB+Dr.
Proof:
Prom Eqs. (2.13) and (2.14b), we have
Tc( ~n-A)-IB = ~(X)D~t(x) (2.23a)

25
Also, from Eq. ( 2 . 1 4 d ) , we have
(2.23b)
_ m ei)T ,T(0),r(0)
K.=0
1
Thus~
~i=O
Since DrCX) = Drh6(k) , we have D-I(X)r = ~ - I ( A ) D ~ " When < i - 0 , from Eq. (2.14e)
we have {6(X)}i i = I, {6(X)}i j = 0 for i~j, and {6(A)}q i = 0 for q~i. As a
result, when <i=0, we have {6-1(l)}ii - I, {~-l(k)}i j - 0 for izj, and
{6-1(A)}qi = 0 for qzi. Thus
M • • m . .
(2.23d)
--_
[I
=
rh <0>*rC0"O
Ki=0 ~i=0
Rearranging the terms in Eq. (2.23c) and substituting Eq. (2.23d) into the
rearranged equation yields DrI(A) = Cr(AIn-A)-IB+Dr.
It can be easily shown that (AtilT(0)) , is an observable pair or (A~C r) is
an observable pair. Thus, the results of Lemma 2.1 follow. •
From Lem~a 2.1 we observe that, if every ~i>O for l$i~m, then D r = 0m~ both
C r and B are of full rank, and D-I(A) is strictly proper. For representing
r
Dr( A ) without involving the inverse of (Xln-A) , an alternative representation of
Dr( A ) is derived as follows.

28
Lermna 2 •2 D (A) can be represented as

r
Dr(A~ ,, ~rC%~n-A)~(X)÷~',r
A A
~here ~(A) ~ T~l*rCX), c r -~ Dr.E~de~ Dr ~ Drht~m-*~CO~*rCO)l.
Prooft
Using Ebc and A c in Zq. (2.13), *r(A) and A r in Eq. (2.14), and 6rh(A) in
Eq. (2.14f), it can be easily verified that Ar = E~cA c and *rh (A)
E~cASr(A)+[Im-*~(0)$r(0)]. Then, from Eq. (10e) we have
,cA)= ~LA,r(A~÷I~m-,~CO~,rCO,-~L,c,r(A~
T [ X~n_Ae],r(x)+t~m_,~(o),r(o)]
= Ebc
~LTo(A~n~)T~l,rCA~÷t~m ,~(0)~r(0)1
Thus, Dr(A) -- Drh*(%) = Cr(AIn-A)B(A)+D r. •
In a similar fashion, from Eqs. (2.19) and (2.20), the minimal realization
of D~I(%) using a pair (A,C) can be expressed as follows:
Lemma 2.3 Define Bg =
(A,Bg, C,D£) is a minimal realization of D~I(A), or DII(A) = C(AIn-A)-IBg+D E. •
Also, an alternative representation of DE(A) without involving the inverse
of (Xl -A) is as follows.

n
Lemma 2.4 D~(A) can be represented as
D~CA~ = ~(A)C~n A)~÷~ ~

27
where C(X) = ~(X)T;I; B~ " ToEc0D~h; D~ " [Ip-~(0)~(0)]D~h. •
Lemmas 2.3 and 2.4 can be proven following the same reasoning as in Lemmas
2.1 and 2.2, respectively.
2.3 Minimal Nice Selections for Determlnln~ Kronecker and Observability Indices
As we h a v e seen~ the right characteristic ~ - m a t r l x and t h e c a n o n i c a l right
HFD o f a reachable s y s t e m can be d e t e r m i n e d once t h e K r o n e c k e r indices o£ t h e
reachable pair (A,B) are known. To find the Kronecker indices of reachable
systems, we i n t r o d u c e the notions of nice selections and m i n i m a l n i c e sections
for the reachability test matrix R(A,B). An algorithm to determine the
Kronecker indices from t h e m i n i m a l n i c e selection is also presented. The same
procedure is equally applicable for determining the observability indices o f an
observable system.
Definition 2.5 A nice selection 7 is an o r d e r e d subset of the set {k:l~k~nm}
such that
(I) The c a r d i n a l i t y 17[ i s e q u a l t o n .
(2) if i+mjEy, t h e n i + m ( j - l ) c y f o r l ~ i s m and 0 ~ j s n - l . n
A nice selection T c a n be r e p r e s e n t e d by a Young d i a g r a m [39] o f m rows
l a b e l e d from 1 t o m and n columns l a b e l e d from 0 t o n - 1 w i t h a c r o s s a t an ( i , j )
entry iff i+mj~y for l~i~m and O~j~n-1. Due to the property (2) in Definition
2.5, the Young diagram of a nice selection associated with T either has
consecutive crosses o r no c r o s s e s at all i n e a c h row. The Young d i a g r a m w i t h
m=3, n=6 and a nice selection 7 = {1,2,3,4,6,9} is illustrated as follows:

28
J~
0 1 2 3 & 5
1 X X b1
£~ 2 X b2
3 X X X b3
z3 s A2 A3 A4 A
Im denotes an mxm i d e n t i t y matrix.
Definition 2.6 Let (A,B) be a reachable pair and ~j be the number of crosses in
the jth column o f the Young diagram o f a nice selection y associated with (A,B).
If
k-I
(2.24)
j~0 zj " rank(B'AS'''''Ak'lB)' k~n
and (AJbi), V(i,j)¢{(a,8): l~C~m, O~8~n-I and cr~n(B-1)~y} are independent, then
~, defined as ~M' is a minimal nice selection of the reachable pair (A,B). D
It has been shown [37] that minimal nice selections are unique up to
permutations of the rows of the Young diagram. Also, if K. denotes the number
x
of crosses at the ith row of the Young diagram of a minimal nice selection YH'
the set K = {~i,l~i~m} is the Kronecker indices of the reachable pair (A,B)
associated with YM" Note that K i may be zero for some i. The maximum number of
minimal nice selections associated with a reachable pair (A,B) is m!, and each
minimal nice selection results in one controller f o r m and one a s s o c i a t e d RMFD.
All the controller f o r m s and a s s o c i a t e d RHFDs a r e equivalent. To d e t e r m i n e the
canonical controller form and associated canonical RHFD, we define the input
selection sequence as follows.
Definition 2.7 The input selection sequence for a multivariable system with m-
A
input, S = {Sk, l~k~n}, is defined as a permutation of the set {k,l~<Sm}. The
29
A
natural input selectlon sequence, SN~ is de£1ned as SN = { S l ~ S 2 p . * * , s m} =
{l,2,...,m}. The m i n i m a l n i c e selection YM w i t h the input selecelon sequence S
i s d e t e r m i n e d by e x a m i n i n g t h e i n d e p e n d e n t v e c t o r s o f t h e o r d e r e d column v e c t o r s
i n t h e following m a t r i x :
~(A,B) ~ [ b s l ~ b s 2 j . .,bsmtAbsl~Abs2~
. . *~Abs ~ .* ' A n - ~ s l An-2b ,A n - I b s m]
** m ** ~ 82'"
(2.25a)
D
If the set K contains the Kronecker indices o f t h e minimal n i c e selection
~M o f a r e a c h a b l e pair (A,B) w i t h the natural input selectlon s e q u e n c e SNp t h e n
the reachabillty b a s e m a t r i x o f t h e (A,B) due t o Fopov [35] i s d e f i n e d as
~i-I ~ -1 (2.25b)
PCA,B) = [bl,Abl,...,A bl,...,bm,Abm,..., A m bm ]
and the controller form transformation matrix Tc becomes Eq. (2.11b). From
Lemma 2 . 1 , the right characteristic X-~natrix can be d e t e r m i n e d . In addition,
t h e RMFD can be found by u s i n g Eqs. ( 2 . 1 4 ) and ( 2 . 1 5 ) .
The l i n e a r l y independent vectors in a rectangular or square m a t r i x R(A,B)
can o f t e n be determined by a p p l y l n g the Sylvester theorem [13] to the square
synnnetrlc matrix R(A,B)RT(A,B), o r by c a l c u l a t l n g the nonzero elements at the
diagonal entries of two triangular matrices, decomposed from the square
symmetric matrix R(A,B)RT(A,B) [38]. Here~ we propose an orthogonallzed
projection approach to determine the llnearly independent vectors o f R(A,B) in
Eq. ( 2 . 2 5 a ) as f o l l o w s .
Let V = {vi,l~i~N} be
a s e t of v e c t o r s v i ¢ c n × l , N ~ n . The o r t h o g o n a l i z e d
A
projection F(k)~C m~n f o r Vk = { v i , l ~ i ~ k } , l~k~N, can be e v a l u a t e d r e c u r s i v e l y as
follows:
p(O) = I ( t h e u n i t m a t r i x of o r d e r n) (2.26a)
n
30
P(k-l)VkVkP ( k - l )
P(k-1) f o r VkP ( k - l ) VkZO
v~P(k-l)v k
v(k) = (2.26b)
F(k-1) f o r VkP(k-1)v k = 0
where v k d e n o t e s the c o n j u g a t e t r a n s p o s e of v k.
Lemma 2.5 Some i m p o r t a n t p r o p e r t i e s of t h e o r t h o g o n a l l z e d projection [39] are
as f o l l o w s :
(1) P*(k) " P(k) o r P(k) i s s y m m e t r i c .
(2) p2(k) = PCk) or P(k) i s i d e m p o t e n t .
(3) P ( k ) v i = 0nxl, ~ vi~Vk, l~i~'k.

(4) P(k)X = X, ~ XIVk, x£Cn x l , 1 S i ~ k .
From Lemma 2.59 we have t h e following results:
. A
Lemma 2 . 6 V k P ( k - l ) v k = 0 i f f Vk£Vk_l, where Vk_ I = S p a n c [ V k _ l ] .
Proof;
S i n c e P ( k - 1 ) i s i d e m p o t e n t and s y m m e t r i c , we have
VkP(k-1)v k .= VkP ( k - 1 ) P ( k - 1 ) v k .. ( p ( k - 1 ) V k ) ( P ( k - 1 ) v k)
k-1
(a) If VkE{/k_.,,,. t h e n v k = [ a,v.. From p r o p e r t y (3) o f Lemma 2 . 5 we have
i=l I
k-1
p(k-l)v k = [ a i P ( k - 1 ) v i - Onx I
Thus
VkP(k-1)v k = 0
(b) I f V k P ( k - l ) v k = 0j we have P ( k - l ) v k " 0n× I .

31
Let
k-1 ^
" alvi+ k' :k I k-I
11
Then, from p r o p e r t y (4) o f Le~ma 2 . 5 we o b t a i n
P(k-l)v k = P(k-l) V k = Vk = Onxl
k-1
T h u s , vk " ~ a i v i o r VkC k - l "
i 1
A
Lennna 2 . 7 D e f i n e WkCVk; W0 = ~ ;
Wk_lU{Vk) if VkP(k-l)Vk=O for IEk~N
Wk_1 i f V k P ( k - 1 ) v k = 0 f o r l~k~N
Then, WN i s a b a s e f o r VN.
Proof:
Lennna 2 . 7 c a n be p r o v e d d i r e c t l y from Lennna 2 . 6 . •
Lemma 2 . 7 provides a foundation for deriving the recursive algorithm for
selecting the iudepeudent vectors from a s e t of vectors. B a s e d on Le~ma 2 . 7 , we
now d e r i v e an a l g o r i t h m to determine the Kronecker indices o f t h e dynamic s y s t e m
i n Eq. ( 2 . 1 ) via the minimal nice selection as f o l l o w s [40].

32
Al~orlthm 2.1
Given: (A,B) - A r e a c h a b l e p a i r
S - The i n p u t s e l e c t i o n sequence
Find: ~ - The Kronecker i n d i c e s
Algorithm:
{Initialization}
P:=I n {set orthogonallzad projection to identify matrix)
For i:=l to m Do
Begin
K1. : = 0 ; { R e s e t Kronecker i n d i c e s }
Flagi:=True;{Set selection flags true for all inputs}
v i : = b i {Copy b i to v i}
End;
{processing}
Repeat
For i:=l to m Do
Begin
j : = s i ; { S e l e c t Akbs.,k~0}
1
If F l a g j then
Begin
d:=P*v.;
.J
If v.*dzO then
Begin { S e l e c t v.}
.J
P:=P-d*d*/vj*dl{Update P}
v j : = A * v j ; { U p d a t e vj f o r next s e l e c t i o n }
K.:=~.+l{Update Kronecker index}
J J
End
Else
Flagj:=false {Reset selection Flag}
End
End {For loop}
Until all Flag i are false;
Note that the * between two variables in the above algorithm is the product
notation, and the superscript * designates conjugate transpose. •
33
l~hen t h e system in Eq. (2.1) is not reachable, the sum o f

the Kronecker
m
indices ~i found by using the above algorithm is less than n, i.e. ~ ~i<n.
i 1
Thus, the above algorithm can also be used to test the reachabillty of systems
i n state s p a c e form.
As is well-known, teachability and observabillty of a system are dual
concepts. The a b o v e a l g o r i t h m can be modified to flnd the observability indices
~;i by substituting (AT,cT) for (A,B), ~;i for Ki, and treating S as the output
selection s e q u e n c e w h i c h is d e f i n e d in the same way as the input selection
sequence in Definition 2.6.
2.4 Illustrative Examples
Consider the following 3-1nput 2-output system
(2.27a)
;kXCt) = AXCt)+BuCt)
(2.27b)
yCt) - CXCt)+Du(t)
where
-1 0 -2 5 -9
I-3-2]
0 -1 4 -8 14 - 1 1
A = 0 0 1 -4 7 ; B = - 3 5
0 0 0 -1 6 1 4
0 0 0 0 2 0 1
1 -1 4 -5 ; D = 02x 3
Determine various controller forms and associated RM~Ds u s i n g minimal nice
selection YM and v a r i o u s input selection s e q u e n c e s S.
(1) If the input selection sequence, $, is selected as S = {1,2,3}, which is
the natural input selection s e q u e n c e SN, t h e n t h e c a n o n i c a l controller form
and a s s o c i a t e d canonical RHFD c a n b e d e t e r m i n e d as f o l l o w s .

34
Usln E the minimal nice selection algorithm, we o b t a i n the Kronecker
indices KI-2 , K2-3 and K3=0. Note that the input matrix B contains a
dependent vector. The c o r r e s p o n d i n g Young d i a g r a m b e c o m e s
0 1 2 4
X X
X X X b2
b3
$5 A
A4
The s t a t e equations in the canonical controller form o f Eq. (2.27) can be
determined from E q s . (2.11) and ( 2 . 1 2 ) as follows:
Ac
- I
o 11o
2 I 0 0
8"-8"~"o" 1
0 0 iO
o
0
o]
0
o
1
; B<-/O
/ o
roool
LL_o._I_I
0
o
0 I" o- -0-O-
o/
o
o
1
o
1
o
1
" EbcDThI
[0 011 1 -1 LO 1 2j 1
Cc 3 I 0 2 2 ; Dc = 02x3
The r i g h t characteristic l-matrix Dr(l) c a n b e d e t e r m i n e d from Eq. (2.15)

as
Dr(k) = Drh~r(k) = Drh[ ~rh (k)-ArSr( A)]
12-k-2 0 -1 1 (2.28b)
0 13+k2-k-1 -2
0 0 1
where
35
Drh
-
[i0i]I [ioi]
1
0
= 0
0
1
0
-2
I
; ~rh(k) "
[2oo]
0
0
~3
0
0
1
ar =
[20 01 0I 01 -i0] ; Sr(~)
. Ii O~ 0 0 !2IT
0 0 0 0 0 0 0 0
Note t h a t Drh i s an upper triangular matrix with diagonal entries all l's,
and Dr(~ ) is a column-reduced ~--matrlx which satisfies all the conditions
in Proposition 2.1.
The canonical RHFD can be obtained from Eqs. (2.14) and (2.15) as
follows:
y(t) " Nr(A)D:l(k)u(t)

(2.28c)
where
[2),-4 3),2+3),-2 0 ]
~r(~) ,, C~r(~) -
31 2~.2+2~. 0
(2) If the input selection sequence, S, is selected as S = { 2 , 3 , 1 } , then the
K r o n e c k e r i n d i c e s become KI=0 , ~2=3, and K3=2 and t h e Young d i a g r a m i s
0 1 2 3 4
1 b1
2 X X X b2
3 X X b3
15 A A2 A3 A4
The s t a t e equations in a controller form of Eq. ( 2 . 2 7 ) become

36
o ooo
o o 1!o o/ / o o o 1
li°°IE j
0 0 0 _
A
C
i 1 -1Lo o.l~ ~c'LO 1 2 I 0 I = EbcD r
o o, 6=i=I I-~-6-~ -o- -o" o
o o12 lJ L1 o i 0 1
where
Drh =
[, o
0
1
I
0
i]-' =
[0],0
- 0
-
Cc=
[-~ ~ ~ :-4 2]
2 21 0 3
(,.,,a)
Note that Drh is not a triangular matrix.
The corresponding RMFD can be obtained from Eqs. (2.14) and (2.15) as
follows:
(2.29b)
y(t) = Nrl(X)D:~(X)u(t)
where
Drl(X) s
[ 0 2 ~k3+~k2-X-1 -2X2+2X+4
o] ; Nrl(X) =
[
0 3X2+3X-2 2k-4 ]
-1 0 X2-X-2 0 2X2+2X 3X J
Note that Drl(X) is a column-reduced X-matrix but it does not satisfy the
conditions (b) and (c) in Proposition 2.1.
(3) If the input selection sequence, S, is selected as S ffi {3)1,2}, then the
Kronecker indices become glffi2) ~2=0, and K3ffi3, and the Young diagram is
A n
II w
0 fl
!
II I 0 @ oll '~ 0
r~ 0 ,.l~, I
I
0
!
~u ,~'u 0,-, ~,'u
I g. ~°
I I..I
0,~00 I 0
t=--..--J ! m
! e
!
I-P
g,, v ~n
v n I I I I~
','.~ (3o n
~g n , =1
÷ I I, I ! ! 0
1,4 I-'
o o i,,,~
,'t
n 0
M ' I ' I'4
I 0 00|P--' 0 ~'1
I
r~
I | II 0 W
"M
t4
iii I
~-'00lO 0 ~'
L I i
0 , ~
II
i v
! ! I
O00l~.J 0
!
0 0 0100
I
I - I 0 0 i O 0
0 0 l I I
÷ I
÷
! !
I I
N
o
v v
38
Note that in thls case D h i s not a trlaugular matrix, and Dr2(~) is a
column-reduced X-matrlx but i t does not s a t i s f y the c o n d i t i o n s (b) and ( e )
in Proposition 2.1.
cHAPTER IIl SPECTRAL ANALYSIS OF NONSINGULAR k-MATRICES
In this chapter, the latent structures of nonslngular k-matrlces are
investigated. The c o l u m n - r e d u c e d and the row-reduced canonical k-matrlces are
defined as the standard forms of a nonsingular k-matrix. Equivalence
transformations are developed for transforming nonslngular k-matrlces to row-
reduced or column-reduced canonical k-matrices, and for transforming non-
c a n o n i c a l HFDs t o c a n o n i c a l MFDs.
The e x i s t e n c e of the first types of transformations enables us t o a n a l y z e
the latent structures of nonsingular k-matrices in terms of the latent
structures of canonical A-matrices. Due t o the simplicity of the formulation
for minimal realizations of the inverse of the canonical k-matrlces in state-
space, the latent structures of canonical k-matrlces can e a s i l y be i n v e s t i g a t e d
from t h e i r associated state-space minimal realization quadruples.
After the latent structures of nonsingular k-matrices are established, the
"matrix roots", called solvents, of nonslngular k-matrlces are studied via the
latent roots and latent vectors. Solvents of monle k-matrlces have been
discussed in many previous works [17-18,22-30]. Our reasons for considering
solvents in this chapter are: (I) to extend the idea of solvents for
nonsingular k-matrices; (2) to develop the structure of a special class of
divisors, namely monic linear divisors of nonslngular k-matrices. The g e n e r a l
structure of divisors of nonsingular k-matrices is studied in the next chapter.
3.1 Canonical A-matrices and Canonical MFDs of MIMO Systems

In this section, we develop a technique to reduce a nonsingular A-matrix to
a row-reduced or column-reduced canonical A-matrix. The reduction technique is
then extended to convert a general I~FD to one whose denominator is a row-reduced
or column-reduced canonical k-matrlx. The converted MFDs a r e referred to as
canonical MFDs.
40
Definition 3.1 A column-reduced canonical ~-~atrix, Dr(1) , can be d e s c r i b e d as
Dr(k) = D r h 6 r ( k )
where Drh i s an upper triangular square matrix with diagonal elements being all
l*s. 6r(k) is a l-matrix with the (i,i)th entry, (6r(k))i,i, being a monle A
p o l y n o m i a l o f d e g r e e Ki i ~ c i ( D r ( A ) ) and t h e ( i , j ) t h entry, (~r(A))i,j being a A
p o l y n o m i a l of d e g r e e ~ m i n ( K i , ~ j ) - I , i z j .
A row-reduced canonical k-matrlx, D~(k), can be d e s c r i b e d as
D~,(k) " ~,(A)V~.h
where D~h i s a lower triangular square matrix with diagonal elements being all
lls. ~(k) is a k-matrix with the (i,i)th entry, (6~(k)i, i being a monic k
p o l y n o m i a l o£ d e g r e e ~i ffi ~ri(DE ( k ) ) and t h e ( i , j ) t h entry, (~E(k))i, j being a A
polynomial of degree ~ min(gi,~j)-I , izj. []
From D e f i n i t i o n 3.1, we o b s e r v e that the left (right) characteristic k-
matrices o f MIMO s y s t e m s discussed in Chapter II are the row (column) reduced
canonical k-matrices. Their salient feature is that the inverse of these k-
matrices can be d e s c r i b e d by t h e s t a t e - s p a c e realizations in c a n o n i c a l observer
or controller forms. Therefore, the structural properties of these A-Inatrices
can he s y s t e m a t i c a l l y e x p l o r e d i n t e r m s of a s t a t e - s p a c e setting.
As m e n t i o n e d in Lemma 1.2 of Chapter I, a nonslngular l-matrix is row
equivalent, column e q u i v a l e n t or equivalent to a row-reduced, a column-reduced
or a row-and-column-reduced l-matrix. However, since equivalent row-reduced
(column-reduced) k-matrices of a given nonsingular k-matrix are not unique, a
row-reduced (column-reduced) k-matrix is not n e c e s s a r i l y a row-reduced (column-
reduced) canonical k - m a t r i x as d e f i n e d in Deginitlon 3.1. We w i l l investigate a
constructive way to explore the equlvalent trans£ormation of a nonsingular k-

41
m a t r i x to its row-reduced or column-reduced c a n o n i c a l k - m a t r l x as f o l l o w s .
We begin with the state-space minimal realizations of the ~nverse of the
reduced k-matrlces.
L e n a 3.1 Let A(k)eCmxm[k] be a column-reduced Z - m a t r l x w i t h the column d e g r e e s

^
Ki,i=l,...,m. Then A(A) can be r e p r e s e n t e d as
A(k) = Dch [ Dh ( k)-A c r ~ c (k) ]
where Dch i s the l e a d i n g column m a t r i x of A(k) and Dh(k) ~ d i a g [ k Ki ,~'1," ... ,m] ;
K:.-1
+c(~) ^ -~ [~cZ(X),~c2(X),...,~em(X)];^~ci¢X)
. ~)i [O~x o i - 1 , l , X , . . . , x ~ 'o Ix(n-oi)
^ J~T
for I~.>01 and ~c¢(~) =A OnxI for Ki'O ; ai ~ j'l)" K:j, i ' l , . . . , m ; n ~ Ore; Acr ¢Cmxn
can be d e t e r m i n e d from Dch and the c o e f f i c i e n t s of each e n t r y of A ( k ) . •
The minimal r e a l i z a t i o n of A-l(k) can be s t a t e d below•
~o=e 3.2 ~-~c~ - ~cC~n-~ c)-I ;c+~c (3.la)

where
A A
KI Km
^ iI I } ~cl ^& ^ ^
A .A -(A ). ; n=n-KI-~ m
C T .cr 1__ | . . . . . . . . . .
o:j~ . • , o:j
_.=__
OC~=_I~J1 ,, oc;=_l~x:-',, o l o . o
! • • • • • o
g
O_ 'a _" _ ' _ ' _ _ 1
j_"
-(Acr) m
(3.lb)
(Acr) i ~ The ith row of Acr,i'l,.•.,m

(3.1e)
A
e °i i f ~c.>O
Bc = bc - "'''ebcm]; ebcl Onx 1 i f ;i=O
42
~c ~ ~(0) (3.1e)
Dc
Proof:
Lemma 3 . 2 can be e s t a b l l s h e d by d i r e c t verification. •
The q u a d r u p l e ( A c , B c , C c , ; c ) is referred to as a c o n t r o l l e r (not a canonical
oontroller) form r e a l i z a t i o n o f A-ICk). ,n g e n e r a l , we c ~ . de~ine A ~ T:IAc~ c,

~ A A
B ~ T:IBc, C ~ CcTc and D " Dc, where Tc is a non-singular similarity
trans£ormatlen matrix, so t h a t the inverse of a c o l u m n - r e d u c e d X - m a t r i x , A(X),
can be r e a l i z e d by u s i n g a q u a d r u p l e (A,B,CjD)~ or
(3.2)
A-I(A) . C(kln-A)-IB+D
Let Dr(k) be the right characteristic k-matrix of (A,B) in Eq. (3.2).
Using Lemma 2.1, D:l(k) can be realized as follows:
D: l ( k ) ffi Cr(XIn_A)-IB+ Dr (3.3)
The relationship between A(k) in Eq. (3.2) and Dr(k) in Eq. (3.3) is the
following:
Theorem 3 . 1 Let A(X) be a c o l u m n - r e d u c e d X - m a t r i x r e p r e s e n t e d by Eq. ( 3 . 2 ) , and
l e t D (k) be d e f i n e d i n Eq. ( 3 . 3 ) . Then~

r
A(k) ffi D r ( k ) U r ( k ) (3.4)

43
where Ur(k) is a unimodular X-matrix, Ur(k) = [CT:I~r(X)+DDr(X)] -I. T

C
and
St(X) are defined in Eqs. (2.11b) and (2.14c), respectively.
Proof:
From Eqs. (2.6) in Chapter 11 and (3.2)~ we obtain
A-l(k) - C(kln-A)-IB+D - lqr(k)D:l(X)+D - Nr (k)D:l(k)
where Nr(k) ffi cz:l~r(k)¢cmxm[x]| Nr(k) - Nr(X)÷DDr(A) | T c is a transformation

matrix which transforms (A,B) into the canonical controller pair. Since A(k) is
nonsingular; Nr(k) must be nonsingular or A(X) = Dr(X);;l(k). Also, since
O(A(X)) = O(A) = O(Dr(~))~ we obtain det(A(k)) - gcdet(Dr(k)) , where K c is a

non ero c o n s t a n t Therefore, Ass result, i.
a unimodular X-matrlx. •
Theorem 3.1 provides a method for determining a unimodular X-matrlx which
converts a nonsingular column-reduced k-matrix to the column-reduced canonical
X-matrix.
The dual results for row-reduced X-matrices are as follows:
Lena 3.3 Let A(k)ecPxP[k] be a row-reduced X-matrix with row degree ~)i'
iffil,...,p, then A(k) can be represented as
(3.5)
A()k) - [Vh( k)-$0 (k)A0g I V0h
where D0h is the leading row matrix of A(X) and Dh(X) A diag[lV i = l , . . . , m ] ;
~0(k)^ b [~01(t),~02(k) ' . . . . ~0p(k)lT

;^ ^~0£(k)i ~ [ 0 1 x ~ i ' l ' l ' k ' " " 'xVi-l'01xCn-Ti)]
for ~>0~ and ~0i(k) = 01x n for ~i'0; ~i ~ j--ll ~ j , i ' l , ' " , P ; n ~ ~p; A0~cCnxp can
be determined from Doh and the c o e f f i c i e n t s of each entry of A(X). •
44
(3.6a)
Lemma 3.4 A-l(x) . Go(X~n-;o)-l;o+~o
where
G
V]
O~lX(~p-1) ^
0 • I A A ^
AO A -(Ao~) 1 eou
O~xc; _~) ,' -(%pp ;nffin-•l-V
. . . . . . . P . . . .
0
I . . A
0 • •
O~pX(Gl-l) o;~ 2 ; " i

P
(3o6b)
(Aog) i ~ The i t h column of A0g, i ~ l , . . . , p .
~o ~ *o(°) A
o6o)
T. A
0
-~ cO;
DO
Eco
^ ^
~ [ec01, : c02 ,"- . - , :cop' T; ecoi =
{ en • ' i~ ~.,o
^
x (3.6d)
Onx I i f V.=Ox
(3.6e)
;o ~ °o~t:p-*o(°)*o c°)1
A A A ^
(A0,Bo,Co,D 0) is referred to as an observer (not canonical observer) form

resli-atio, of A-I(x). ~n general, it we define A A ~0Ao;;1 B -~ ~0; 0'
C ~0 - land D = DO, where T O is a coordinates trans£ormation matrix, then the
inverse of a row-reduced k-matrix, A(A), can be realized by a quadruple
(A,B,C,D), i.e.
45
(3.7)
A-IcA) . C(kIn-A)-IB+D
Theorem 3 . 2 L e t A(A) b e a r o w - r e d u c e d A - m a t r l x , which can be realized, using a
quadruple (A,B,C,D). Let D~(A) b e the left characteristic A-matrlx o£ ( A , C ) .
Then
(3.8)
ACA) = UECA)D~CA)
where UE(A) is a unimodular A-matrix, UE(A) = [$~(A)T;IB+D~(A)D] -I, T O and ~E(k)

are defined in Eqs. (2.17b) and (2.20c), respectively. •
Combining Lemma 1.2 and Theorems 3.1 and 3.2 yields the results as follows.
Theorem 3 . 3 A nonsingular k-matrix i s row o r column e q u i v a l e n t to a row-reduced
or c o l u m n - r e d u c e d c a n o n i c a l ~matrix, respectively. •
Theorem 3.3 reveals the fact that the structural properties of nonsingular
A-matrlces can be studied from the appropriate row-reduced or column-reduced
canonical k-matrices. In the next chapter~ the structure theorems of divisors
of nonsingular matrices are derived~ using the row-reduced or column-reduced
canonical ~-matrices.
The followlng extension o f Theorem 3 . 3 to t h e RMFD and LHFD c a n b e e a s i l y
proved.
Theorem 3 . 4 An i r r e d u c i b l e RHFD (LMFD) d e s c r i b e d by G ( k ) = N C ~ ) D - I c A ) ,
(G(A) = D-I(~)N(A)), c a n be c o n v e r t e d to a canonical RM~D (LMFD) d e s c r i b e d by
G(~) = NrCX)D~Icx), (G(k) = Dilck)N~(k)), where VCk) and N(A) are coprlme and
Dr(k)(DL(A)) is a column-(row-) reduced canonical k-~natrix. •
Corollary 3.1 An i r r e d u c l b l e proper RMFD o r LMFD c a n b e r e a l i z e d by a c a n o n i c a l
controller-form state-space quadruple (Ac,Bc~Cc,Dc) i n Eq. (2.12) or a canonlcal

46
observer-form state-space quadruple (Ao,B0,C0,D O) in Eq. (2.18), respectively.
P~0of:
Let G(A) = N(X)D-I(x) be an irreducible, proper RI4FD, then from Theorem 3.4
we obtain G(X) = Nr(~)D~I(I), where ~ci(Nr(~)) ~ ~ci(Dr(1)). Thus, G(X) can be
expressed as G(k) = Nr(A)D~I(A)+Dc(A), where ~ci(Nr(k))<~ci(Dr(k)). Since G(I)

iS proper I we have ld,~ G(l)<~.
~im Also Nr(l)Drl(l) is strictly proper, or
~im l(k) and ~i+~m
l÷m Nr(A)D ~ = O, therefore Dc(X)<m ~ or Dc(I) = Dc is a constant
matrix. Thus
G(X) - NrCX)D~I(X)+D c
From Eqs. (2.14) and (2.16)~ we can find the minimal realization quadruple
(Ac,BctCc,D c) in the canonical controller form.
The minimal realization of the irreducible proper LRFD in the canonical
observer ~orm can be proved in a similar way. •
3:2 Latent Structure of Nonsingular l-matrlces
As stated in Theorem 3.3 i n Section 3.1, a nonsingular l-matrlx is row or
column equivalent to a row-reduced or column-reduced canonical k-matrlx~
respectively. We shall first concentrate on the latent structures of canonical
l-matrices and then extend the results to nonslngular l-matrlces.
The latent roots of a nonsingular l-matrlx D(k) are defined as the roots of
det(V(l)) = O.
Lemma 3.5 [17-19] Let the set of latent roots of D(k) be O(D(A)), and the set
of eigenvalues of A, which is the system map of a minimal realization of D-I(x),
be O(A). Then O(D(1)) = O(A), and

47
[o(D(X)[ - [o(A)[ = n •
where I [ denotes the cardinality o f the s e t .
Proof:
Directly from Theorem 3.3 we have
(3. lOs)
D(X) = Dr(X)Ur(X)
where D (k) is a column-reduced canonical k-matrlx~ and Ur(X) is unimodular.

r
From Lemma 2.1~ we have
(3.10b)
D-I(x) = U:I(x)[Cr(XIn-A)-IB+Dr]
Thus the results of Lemma 3.5 follow.
The seneralized latent vectors of a nonsingular X-matrix are defined as
follows:
Definition 3.2 [17-19] Let X. be a latent root of D(X). A left Jordan chain
i
of left generalized latent vectors (left Jordan chain for abbreviation) f o r D(X)
associated with X i is a set of nonzero vectors Pij' 0~J~&i-I such that
(3.11a)
i ~1 (D(k) (Xi))Tpi(j -k) " 0mx 1
k=0
where D(k)(x)~d(k)D(X)/dX k = k-th derivative o£ D(X) with respect to I (3.11b)
The scalar ~i is named the length of the Jordan chain [41]. Similarly, a right
Jordan chain o f D(X) associated with X i is a set o f nonzero vectors qij~

48
0 % J ~ i - 1 such that
i 1 (DCk) (3.11c)
k=0 kT (Ai))qi(j-k) TM Omxl
The v e c t o r s Pi0 and qi0 are r e f e r r e d to as the primary l e f t and r i g h t l a t e n t
vectors, respectively, n
We s h a l l first present the p r o p e r t i e s of l e f t / r l g h t Jordan c h a i n s of the
column-reduced c a n o n i c a l k-matrlces, and the r e l a t i o n s to the e l g e n v e c t o r s of
the system maps i n t h e i r a s s o c i a t e d minimal r e a l i z a t i o n q u a d r u p l e s .
The s a l i e n t p r o p e r t i e s of l e f t and r i g h t Jordan c h a i n s of a column-reduced
c a n o n i c a l k - m a t r i x Dr(k ) are d e s c r i b e d as f o l l o w s .
Len~aa 3.6 Let (qij)~ be the ~-th component of q i j ' then for O~E~m we have
( q i j ) ~ = 0 i f the Kronecker index, of a minimal r e a l i z a t i o n quadruple of D ; I ( A ) ,
satisfies ~ = O, 0 ~ <~.
Froof~
From Eq. (3.11e) we have
D r ( A i ) q l 0 = DrhD(Ai)qio = Ore×1 •
Since Drh i s n o n s l n g u l a r , and i f KEffi0, the Eth column of Dr(A i ) i s e~ and ~th
row of Dr(k i ) i s (e~) T. Thus, the Eth component of qi0 i s z e r o . Also, from
Dr(ki)qil+D~l)(ki)qi 0 = 0m× I
We find that the Eth component of qil is zero. Repeatedly using Eq. (3.11c)
r e s u l t s i n ( q i j ) E = 0 i f KE = O, l ~ m . •
49
A
Lena 3.7 If Plj' O~J~i-l' is a left Jordan c h a i n of Dr(X) and P i j ~ DTr h P i j ,
then the ~ - t h component, (pij)~, of Plj is zero. This i m p l l e s t h a t the £ - t h
column of Drh is orthogonal to Pij if K~ = 0 for 0 ~ m . •

A
Lemma 3 . 7 can be e a s i l y proved by o b s e r v i n g the f a c t t h a t P l j ' 0~J~i-l' is
a r i g h t J o r d a n c h a i n of D~(X) and the E - t h component of P i j i s z e r o i f K~ = 0.
L e t (Ac,Bc,Cc,D c) be a minimal r e a l i z a t i o n quadruple of the inverse of a
column-reduced canonical X-matrix Dr(X) , and let (Ac,B c) be a controller
c a n o n i c a l p a i r as d e f i n e d i n Eqs. ( 2 . 1 3 a ) and ( 2 . 1 3 d ) , i.e.
D:I(X) = Cc(Xln_Ac)-IBc+Dc (3.12)
The relationships between the Jordan chains of A and Jordan cha£ns of Dr(X) can
C
be d e s c r i b e d in the f o l l o w i n g i m p o r t a n t theorems.
Theorem 3.5 If qij' 0£J~£i-1' i s a r i g h t Jordan c h a i n of Dr(k) a s s o c i a t e d with
a latent r o o t Xi, then q c i j ' 0£J££i-I' a right Jordan c h a i n of Ac a s s o c i a t e d
with an eigenvalue li, can be found as
= ! 1 $(k)(ki) q . . . . . (3.13a)
qclj k 0 ~" r I~j-K;
where
~k)(k)~d(k)~r(k)/dkk " n - t h d e r i v a t e of ~ r ( k ) w i t h r e s p e c t to k (3.13b)
and Sr(X) i s d e f i n e d i n Eq. ( 2 . 1 4 c ) .
Proof:
From Eqs. (2.14h) and (2o15b) we have

50
(3.13c)
ScVr(k) = ( k I n - A c ) $ r ( k )
Differentiating b o t h s i d e s of Eq. ( 3 . 1 3 c ) k times y i e l d s
Bc r
vCk)(1) = k$Ck-1)Ck)+Clln-Ac)~k)ck)
-r
, k~l (3.13d)
If qij' 0~J~i-l' is a right Jordan chain of Dr(k) , from Eqs. (3.11c) and
(3.13d) we have
B i ~-TD
1 (rk)(kl)qi," k - = i 1 $(rk-l)(~,i)qi,.k.]
c k=O ~J- ; k=l ~ [ ~J- )
+ ! 1 [(kiln_Ac)~(rk)
k 0~ (Xi)qi(J-k)]
(3.14a)
= Onx 1, l<J~£i-1
Thus, we have
= (3.14b)
(XiIn-Ac)~r(ki)qi0 Onx I
J 1 (k) j-1
(kiln-kc) k~O = Isjs¢.-I
= ~.~ ~r (kl)qi(;-k)
~ +k=[0 ~1¥ r .(k)(ki )qi(j-l-k) 0nxl' i
(3.14c)
(3.15a)
Define qcij i ~"
k=0 1 ~)(rk)(~i)qi(j_k)
Then, Eq. (3.14) becomes
= (3.1Sb)
(liIn-Ac)qci 0 0n× 1
51
(3.15c)
(liTn-Ac)qcij+qei(j.1) = Onx I , l~J~£i-I
Obviously q c i j ' 0~J~Ei-l' is a Jordan chain of Ac.
Theorem 3.6 If qcij' O~J~Ei-l' is a right Jordan chain of A c associated with an
elgenvalue I i , then q l j ' a r i g h t Jordan chain of Dr(k) , can be determined by
qij = ~(O)qclj ' Jffi0'l'""~i'l (3.16)
Proof:
From Eq. (3.15a) we have qci0 ffiSr(kl)qio" So from Eq. (2.23b),
*:(0)qci0 ffi *Tr(O)*r(li)qlo = *T(O)*r(O)qio
The ~-th component of ~ ( 0 ) q c i o can be expressed as
(qio)~ , %>o
(*~(O)qcio)~ 0 , ~=0
From Lemma 3.6 we have qio " ~(O)qci0" Thus, from Eq. (3.15a) we have
iJ 1 ~ k ) ( k i ) q i ( j _ k ) , j>o
k=O
Since
[ ~:(O)~r(O) if k=0
~(o)~Crk)(~i)=
I 0m if k>0
This implies
52
~(0)qcl j " ~(O)~r(O)qlj
The ~-th component of ~(O)qcij is

(qij)~ if K~>O
(~(0)qcij) £ s
o if Kt-o
Thus, we have the r e s u l t in Eq. ( 3 . 1 6 ) .
Theorem 3.7 If Pcij' 0~J~i-l~ is a l e f t Jordan chain of Ac a s s o c i a t e d with an

elgenvalue Xi~ then P i j ' a l e f t Jordan chain of Dr(l) j can be determined by
T (3.17)
P i j = BcPcij '
Proof;
From Eq. (3.13d)~ we have
[D(rk) (~)] T B T - k [ ~ ( r k - l ) ( x ) ] T + [ ¢ ( r k ) ( x ) ] T ( X I n - A c ) T .
Since Pcij~ 0 ~ J ~ i - 1 is a l e f t Jordan chain of Ac~ we have
(kiIn-Ac)Tpcij = -pc£(j_l) ~ 0~i~.-l~ ,
and Pci(-1) ~ Onxl"

Therefore2
J 1 . (k-l)(k)]T
~TI [ D(k)(~)]TB~pci(:rj - k) = k~ 1 ( k - l ) ! tYr i Pci(j-k)
k=0
53
÷
1 [~k)(~i)]T(XiIn-Ac)TPci(
k---f :
j- k)
k-O
k=O v--i" [$ k 0 ~ tVr i J PciCj-k-1)
= -T[
l r.,.CJ)cx i ),T
~r a Pci(-l) " 0mxl
Thus, we conclude t h a t B l P c i j , 0 ~ J ~ t i - I , is a l e f t Jordan chain of Dr(k).
Theorem 3.8 If Pij' O~J~£i-l' i s a l e f t Jordan chain of Dr(1) a s s o c i a t e d with
l a t e n t root Zi' then a l e f t Jordan chain Pclj of Ac can be determined by
J 1 ~(rk)(Xi)Dl (3.lS)
Pcij = Tp k=~O ~ hPi(j_k)
where Tp = {[Tp] i j }; l<j~np l<ign and K.~.xj = 0 ;
[Tp] i i reversed upper t r i a n g u l a r T o e p l i t z m a t r i x [ I I ] with f i r s t

m m
column [ a l i 2 , . . . , a l i ~ . , l ] T ;
l
[Tp] j i r e v e r s e d upper t r i a n g u l a r T o e p l i t z matrix with f i r s t column
[ a i j 2 . . . . . a i j K i , 0 ] T if Ki<%, and [ a l j 2 m . . . , a l j % , 0 , . . . , 0 ] T
i f ~.x >~..
j
Proof:
From Leuuna 3.7, DT
rh P i j ' O~J~i-l' are g e n e r a l i z e d l e f t l a t e n t v e c t o r s of a
l e f t Jordan chain of o(~) = D;~Dr(~). Define
k=O ~ hPi(j-k) ' O~J~li-i

54
It follows from Theorem 3.5 that Pcij' 0~J~i-I is a right Jordan chain of Ac
which is the associated companion form for DT(~), and Ac ~ (Ac)ij for l~i~m and
l~j.~,m is defined according to:
( c)ii = (Ac)IIcR ~
0(~j-1)xki K.xK. ^ [ [aijl,...,-aiJKi] if ~iSKj

CR J ~;Aji = ~ _ if Ki>Kj;izj
A..
jl [-aijl,--.,-aljKj,O,..-,0]
By direct computations, it can be easily shown that A T = T A T -I. Thus, a left

c pcp
Jordan chain of AcT is Pcij = Tp;cij" •
From Theorems 3.5-3.8, the relationships between the Jordan chains of Dr(~)
and the Jordan chains of A, which is a system map of an arbitrary minimal
realization quadruple of D-I(x) p can be formulated as follows:

r
Corollary 3.2 Let Dr(X) be the right characteristic h-matrix of a reachable

pair (ApB); qaij for 0~j~i-I be a right Jordan chain of A associated with an
eigenvalue X.~ of A; qij for 0~j~.-l~ be a right Jordan chain of Dr(~) associated
with the same latent root X£ of Dr(~). Then qa£j and qij can be related by
= T-I i 1 ,(k)
qaij c k= 0 kT ~r (Xi)qi(j-k)' 0~J~£i-I
qij = ~(0)Tcqaij, 0~JS~i'l
where ~r(k) (~) is defined in Eq. (3.13b), and Tc is the transformation matrix
which transforms A to its controller canonical form A c = T cAT -I.
c •
55
Proof:
Corollary 3 . 2 can be p r o v e d in a straightforward fashion from Theorems 3 . 5
and 3.6. In a similar way using Theorems 3.7 and 3.8 we can obtain:
Corollary 3.3 Let Dr(A) be the characteristic A-matrix of a reachable pair
(A,B); Paij for O~j~.-Iz be a left Jordan chain of A a s s o c i a t e d with the
eigenvalue ~i o f A; P i j for 0~J~i-I be a l e f t J o r d a n c h a i n o f Dr(A) a s s o c i a t e d
w i t h t h e same l a t e n t r o o t ~ i o f Dr(A). Then, paid and Pi~ can be r e l a t e d by
i I .(k)(k )DT
Paij = T~Tp k=0 kT ~r i rhPi(j-k) ' 0~J~i-I
and
Pij = BTpcij' O~J~£i-I
where ~b(rk)(A) i s d e f i n e d in Eq. ( 3 . 1 3 b ) and T c i s the t r a n s f o r m a t i o n m a t r i x such
that A ,, T AT- 1 , which i s the c o n t r o l l e r canonical form of A. •

c c c
It is well known t h a t the Jordan chains of a matrix are not unique; the
same is true for the Jordan chains of a A-matrlx. However, from Theorems 3.5
through 3.8, we observe that, given a left/right Jordan chain of a system map
Ac, the corresponding left/right Jordan chain of the right characteristic A-
matrix of a reachable pair (Ac,B c) can be uniquely determined, and vice versa.
Also, from Theorems 3.5 through 3.8, we observe that ~r(A) in Eq. (2o14C) links
the generalized eigenvectors o f Ac t o the generalized latent vectors of D r ( A ) .
M o r e o v e r , t h e i n p u t m a t r i x Bc i n Eq. ( 2 . 1 3 d ) o r B i n Eq. (2.1) relates the left
generalized eigenvectors o f Ac to t h e l e f t generalized latent v e c t o r of D r ( A ) .
'~et . be a mod~l ~atri~ of Ao. ~en, we have Aj ~ M : l A c ~ , where Aj is a
system map in Jordan block canonical form, and M c and (M: 1) contain all the
right and left generalized elgenvectors of A c. Thus, from Theorems 3.6 and 3.7
we can easily derive the following results.

56
Len~a 3.8 (M:IBc)T c o n t a i n s all the left Jordan chains of a column-reduced
c a n o n i c a l A - m a t r l x Dr(A) , and ~ ( 0 ) M c c o n t a i n s all the r i g h t Jordan c h a i n s of
Dr(k)" •
Corollary 3.4 Let (A,B) be a reachable pair, M be a modal matrix of A, and
Dr( k ) be the c h a r a c t e r i s t i c ~ - m a t r l x of (A,B). Then, (M-1B) T and ~ ( 0 ) M c o n t a i n
a l l the l e f t and r i g h t Jordan c h a i n s of Dr(~) , r e s p e c t i v e l y . •
Now, from Lenmas 2.1 and 3.8 we have the f o l l o w i n g c o n c l u s i o n .
Theorem 3.9 Let (Aj,Bj,~j,Dj) be a minimal r e a l i z a C l o n of a column-reduced
c a n o n i c a l A-matrlx Dr(A) with Aj i n Jordan form. Then,
D:I(A) . ~ j ( A I n _ A j ) - I B j ÷ ~J , (3.19)
where BjT i n Eq. (3.19) contains all the left Jordan c h a i n s of Dr(A) , and ~ j
c o n t a i n s a l l the r i g h t Jordan c h a i n s of Dr(A).
Proof:
From Lemma 2.1, (A,B,~,~) is a minimal realization of Dr(At._

_ Thus
D:lcx) = ~(~In-A)-IB+D - ~T(o)(kIn-Ac)-lBc+~
- J
Hence, from L e n a 3.8, BjT and ~j contain all the left and right Jordan chains of
Dr()J. •
In an analogous fashion, we can easily obtain the following results for a
57
l a t e n t s t r u c t u r e of the row-reduced c a n o n i c a l k - m a t r l c e s :
B m
Theorem 3.10 Let (Aj,Bj,Cj,Dj) be a m{uimal reallzatton of a row-reduced
c a n o n i c a l ~ - m a t r l x , D~(k), w i t h Aj i n a Jordan form. Then,
D~I(I) " Cj(~ln-Aj)-l;j+; J (3.20)
=T
where B~ i n Eq. 43.20) contains all the l e f t Jordan c h a i n s of D~(X), and Cj
c o n t a i n s a l l the r i g h t Jordan c h a i n s of D~Ck). •
The latent structures of general nonsingular l-matrlces can be derived
directly from Theorems 3 . 3 , 3 . 9 , and 3.10z
Theorem 3.11 Let D(~) be a n o n s l n g u l a r k - m a t r l x and
DCI) " D (1)U ( 1 )

r r
where Dr(k) i s a column-reduced c a n o n i c a l k - m a t r i x and Ur(1) i s un~modular. Let
i i be a l a t e n t r o o t of Dr(1) , P i j and q l j for 0<J~i-I be an a s s o c i a t e d l e f t and
an a s s o c i a t e d r i g h t Jordan c h a i n of Dr(1) w i t h l e n g t h Ki' r e s p e c t i v e l y . Then
A (3.21a)
and
(3.21b)
~-~ U; ki)qi(j_k ) ,
k-0 x
are respectively a left and a r i 8 h t Jordan c h a i n of D(X) w i t h l e n g t h ~ . . These

I
c o r r e s p o n d t o the l a t e n t r o o t h i o f D(A). u;(k)(kl) d e n o t e s the v a l u e of the
k - t h d e r i v a t i v e of U ~ l ( l ) e v a l u a t e d a t k = ~.
%.
58
Proof:
From Theorem 3.3, we have
D(k) ffi Dr(k)Ur(k)
Since Plj' O~J~ti-l' is a l e f t Jordan chain of Dr(1) , from Definition 3.2

we have
J (3.22)
1 I (v~k)(/i))Tpi(j k) 0mxl' x
k=0
Since D(k) = Dr(A)Ur(A) , we have
D~k)(k) - [D(k)U:I(A)] k
k
kCsD(S)(k)u~(k-s)(k) (3.23a)
I
s=0
where kCs is a binomial coefficient,
A kl (3.23b)
kCs = s l ( k - s ) !
S u b s t i t u t i n g Eq. (3.23) into (3.22) y i e l d s
)~J kT.1 (s=~

Ok k e s [ D ( S ) ( x i ) u -"( k - s )' ( k i ) ur ' ( k - sr) ( x i ) ] T p i ~j-k" ffi Omx I
i
k=O
(3.24a)
Equation (3.24a) can be rearranged as
J I - -(s) T ~ 1 [vCk-a)(Xi)]TpiCj 0~J~ti-I

~T. lUr (li)] (k-s) l -k) = Om×l
s=O k s
(3.24b)
59
O~
°_
s~ 0 ~T. kT (Xi)]Tpi(j-s-k) = 0mxl
(3.24c)
Since u~l(ki ) is nonslngular for all Xi' from Eq. (3.24c) we conclude that
J 1
~
T
[D(k)(/i)] Pi(j-k) = 0mxl
05J5£i_l (3.25)
kffi0
By Definition 3.2, Pij for 05J$£i-i is a left Jordan chain of D(%) corresponding
to the latent root ~i"
Furthermore, since qij for 0KJ~£i-I is a right Jordan chain of Dr(%) , from
Definition 3.2 we have
J[ ~I [D~k)(ki)]qi(j_k) ffi 0mxl 0~J~£i_l (3.26)

k=O
Substituting Eq. (3.23) into (3.26) yields
J i k
) ~-~ [s=~O kCsD(Sl(kilUr(k-sl(ki)]qi(j-k ) 0mxl m
k=O
(3.27a)
Equation (3.27a) can be rearranged as
~. (~i) k'T Ur ~i)qi(j-s-k) = Omxl z

8=0
(3.27b)
By Definition 3.2, we conclude that
6O
qij ~ i k-T
1 U; ( k ) ( X i ) q i ( j _ k ) for O~j~.-I
k=O x
is a right Jordan chain of D(X) corresponding to the latent root A i.
From Theorem 3.11, we can easily deduce the following:
Corollary 3.5 Let D(X) and Dr(l ) be defined in Theorem 3.11, Pij and qij for
0~J$~i-i be a left and a right Jordan chain of D(A) corresponding to a latent
root A.. Then

i
A -- (3.28a)
Pij -
Pij ' o<j ~£.-I
I
and
qij ~ i 1 u(k) -- 0~jS~.-I (3.28b)

k= 0 k'~ r (Xi)qi(j-k) ' i
are a left and a right Jordan chain of Dr(1) corresponding to the same latent
root %..
l
•
Again, for completeness we state the result corresponding to Theorem 3.11
and Corollary 3.5 for the row-reduced case.
Theorem 3.12 Let D(X) be a nonsingular A-matrix, and
D(X) = UE(A)D~(X)
where DE(1) is a row-reduced canonical l-matrix, and U~(X) is unimodular. Let
Pij and qij for O~j~;9~.-It be a left and a right Jordan chain of D~(X) with length
~i' respectively.: They correspond to a latent root X i of Dr(1). Then
?ij =A k=0~L ~-~

1
[O£
(k) Z
(Xi)] PiCj-k) '
0~j<~._l (3.29a)
and
81
- O~j~t.-I (3.29b)
qij ~ qij ' x
are a l e f t and a r i g h t Jordan chain of D(A) with l e n g t h Ei~ r e s p e c t i v e l y . These
correspond to the l a t e n t root A i of D(A). •
C o r o l l a r y 3.6 Let D(A) and DE(A) be d e f i n e d i n Theorem 3.12, P i j and q i j for
0~J~i-i be a l e f t and a r i g h t Jordan c h a i n of D(A) c o r r e s p o n d i n g to a l a t e n t
root Ai, then
(3.30a)
Pij =A i ~I [U(k) ()ki)]T~i(j_k ) , O~j~Ei-I
k=O
and
- (3.30b)
qij ffi qij' 0~j~E.-Ix
are a left and a right Jordan chain of DE(A) corresponding to the same latent
root A.. •
1
Theorems 3.11 and 3.12 with Corollaries 3.2 and 3.3 build up the latent
structures of nonsingular A-matrices based on the latent structures of canonical
A-matrices, whose importance is thereby emphasized.
3.3 Solvents of NonsingularL-matrices
In this section we shall extend the idea of latent roots in nonsingular A-
matrices to "matrix roots" which in the literature [17,18] are usually called
solvents.
Let fzC+C be analytic within, and on a simple closed contour C, and x be
any point interior to C, then the well-known Cauchy's integral theorem [42]
gives
62
f(;k) (3.31)
where C is traversed in the positive (counter-clockwise) sense.
Using this approach, a class of matrix functions can be generated from
scalar function f(%) [43]. Let x ecmXm, the set of mxm complex matrices, and let
Xl,...,Xq, f o r q~n, be the d i s t i n c t eigenvalues of X. Suppose t h a t ~l,...,Xq

are i n t e r i o r to C, then the m a t r i x f u n c t i o n f(X) can be defined [9] as
1 (3.32)
f(x) = ~ -~C f ( ~ ) ( ) ' I m - x ) - l d ~ = ~ ~C (~Im - x ) - l f ( ~ ' ) d k
where I is an m ~ identity matrix.

m
If f(k), an n-th degree complex polynomial with coefficients fk' is
described by
n n
fCX) = >" fkXn-k = ~ xn-kf k (3.33a)
k=O k=O
then, the corresponding matrix function becomes
n n
f(x) ffi i fk xn-k = I xn-kf k (3.33b)
k=O k=O
where X 0 ~ I .
m
An n-th degree, m-th order nonsingular ~-matrix A(~)Ecmxm[~] can be written
as follows:
A(%) = A0%n+Al%n-l+...+An_l%+An (3.34)
• cmxm , k = 0,1,...,n, det(A(%)) ~t 0.

where akE
63
Since A(k) is analytic everywhere in C , we can define the right matrix

polynomial AR:C mxm ÷ Cmxm as
1 (3.35a)
AR(X) ffi~ ~C A()')(IJk-X)-Id~
where xEcmxm has all its eigenvalues interior to the simple closed contour C.
Substituting Eq. (3.34) into Eq. (3.35a) yields
AR(X) = ~ ~C (k=~O Ak%n-k)(Im %-x)-Id~
n
" k~o ~'E~i ~= :-kC~m~-x)-ld~l
n
>~ ~ x n-k (3.35b)
k=O
Similarly, the left matrix polynomial of A(%) is defined as
1 n
AL(X) " 2--~ ~C (Im%-X)-IA(%)d% = [ ~ i PC (Im%-X)-l~n-kd~]Ak
kffiO
n
>~ X'-kAk (3.36)
kffiO
Having these definitions of right (left) matrix polynomials in Eqs. (3.35) and
(3.36), we are ready to define the solvents of %-matrices.
Definition 3.3 Let R be an mxm complex matrix such that
AR(R ) =
AORn+AIRn-I+. " . + A n _ I R + A n = 0m (3.37)
64
Then R is a risht solvent of the A-matrlx A(A), where 0m is an m~n null matrix.
Bimilarlyj if L is an mxm complex matrix such that
(3.38)
AL(L) = LnA0+Ln-IAI+...+LAn_I+A n - 0m
then L is a left solvent of the A-matrix A(A). 0
The solvents of A(A) can be constructed by using the latent roots and
generalized latent vectors of A(A). The structure and existence of solvents
have been extensively studied and r e p o r t e d in the literature (for example, [17-
19, 23-25]). Numerical methods to compute solvents can be found in [24,44-46].
Some important results on the properties of solvents are summarized as follows.
Theorem 3.13 If A(A) has m linearly independent right generalized latent

^ A ^ ,~
vectors ql,...,q m Cleft latent vectors P I " " ' P m ) corresponding to latent roots
Al,...,~m,
A
then QAQ-I(p-IAp) is a right Cleft) solvent, where P = (Pl'''"
^ ;o )T
(Q = (ql,...,qm)) and A ffi diag(Alg...gkm). •
A set of right (left) solvents RkJ k ffi l,...,n (Lk, k - l,...,n) is called
n
a complete set of right (left) solvents of A(X) if O(A(X)) = U O(Rk)(O(A(A)) -

n k=l
O O(Lk)) , where O(A(A)) is the spectrum of A(A) and a(Rk)(O(Lk)) is t h e
k=l
spectrum of the right (left) solvent Rk(Lk). The existence of the complete set
of right (left) solvents has been investigated by Gohberg et. al. [20~21]~ and
Dennis et. al. [18].
A right solvent R of A(A) is regular [17] if
aCR) n oCA(1)(~)) = (3.39a)
where A(1)(A) is given by

85
A(k) = A(1)(t)(Imk-Z) (3.39b)
and ~ denotes the null set. Similarly, the definition of a regular left solvent
is given by
A
oCL) O oCA(1)(1)) = (3.40a)
where A(1)(1) is given by
A(1) = ( I l - L ) A ( 1 ) ( l ) (3.40b)
m
Thus, a complete set of regular right solvents Rk, k = 1,...,n of A(%) is
defined as
k~j and j,k = l,...,n (3.41a)

O(R k) • aCzj) = ~ ,
n
(3.41b)
oCACk)) = v o(~.)
I%
k=l
Similarly, a complete set of regular left solvents Lk, k = l,...,n of A(k) is
given by
o(Lk) n 0~Lj) = 0 , k=j, j , k = 1, . . . . n (3.42a)
n
(3.42b)
oCA()0) = U o(L k)
k=l
An important theorem on regular solvents developed by Markus and Mereuca [23]
is:
68
Theorem 3.14 If {Rk, k " I,...,n}({L k, k = l,...,n}) is a complete set of
regular right (left) solvents of A(1), then the associated block Vandermonde
matrix
I I ... I
m m m
R1
2 2 R2 (3.43a)
V(R1,..., RI R2 " " " n
. . . . .
-1 n-1 R.-1
R2 " "" n
or
2 n-i
Im LI LI •.. LI
Im L2 L2
2 -.. L2-1
(3.43b)
vB(LI,...,Ln ) ...
1 L L2 ... L n-I
M n n n
is nonsingular, where the B in Eq. (3.43b) designates block transpose [18].
From Eq. (3.39b)9 if REC mxm is a right solvent of a nonsingular l-matrix
A(1), then A(I) can be written as
A(I) ffiA(:)(1)(: l-R) (3.4~a)

m
where ACI)(%)~cmxm[%] is a l-matrix with degree lower than that of A(1). As we
shall see in Chapter IV, (Iml-R) is a right divisor of A(%), and A(1)(l) is a
left divisor of A(1). Similarly from Eq. (3.40b) if LEC mxm is a left solvent of
A(k), then A(k) can be decomposed as
A(k) ~ (I k-L)i(1)(k) (3.44b)

m
where A(1)(1)EcmXm[%] is a %-matrix with degree lower than that of A(%); and
67
A
(Imk-L) and A(1)(k) are a l e f t and a r i g h t d i v i s o r of A(k), r e s p e c t l v e l y . The
more general s t r u c t u r e and p r o p e r t i e s of l e f t and r i g h t d i v i s o r s of a nonsiugular
k-matrix are presented in Chapter IV.
Two examples are presented in this section. The first example demonstrates
the r e l a t i o n s h i p s between the l a t e n t structures of canonical k - m a t r l c e s and the
eigenstructures of the system map of the a s s o c i a t e d c o n t r o l l e r canonical minimal
realization. The second example is devoted to verifying the relationships
between the latent structures of nonsingular k-matrlces and the latent
structures of canonical k-matrices. In addition the complete set 0£ solvents of
nonsingular k-matrlces is found.
Example 3.1
Given a column-reduced canonical k-matrix
X3-3k-2 0 -1 1
3k+3 k2+2k+1 -2
vr(k) =
0 0 1
The controller canonical minimal realization of D-l(k) can be found as

r
D;I(I) = Cc(kI4-Ac)-IBc+Dc
where
Ac
=
0o
2
" 0--0-0 i o
-3 -3 0 I-I
0,
3
oI , oI 0
~',o_
-!-I ; Be-t1
FIOO
l-o-
L o
0
o Ol
o
I
07
1.[
2.1
68
"
[iooo o] 0
0 0;0
0 0
0
; Dc"
[ooo]0
0 0
0 0
1
The modal m a t r i x o f A¢ can be found as
0.0 -0.33333 -0.IIIII 0.33333 0.11111
J
0.0 0.33333 -0.22222 -0.33333 0.22222
M~ " 0.0 -0.33333 0.55556 0.33333 0.44444
1.0 0.33333 -0.88889 -0.33333 -0.IIIII
-I.0 0.66667 1.22222 0.33333 -0.22222
The minimal r e a l i z a t i o n o f D : I ( ~ ) i n a J o r d a n form becomes
Aj = MA1AMA = block diag
Bj = H~IB = [ loilT 0 1
-1
0
;
-
~3
,
(: [0001
:)I 0 0 0
1 2 0 0 1
I 0.0 -0.33333 -0.11111 0.33333 0.11111 1

~ j = ~MA = 1.0 0.33333 -0.88889 -0.33333 -0.11111
0.0 0.0 0.0 0.0 0.0
The e l g e n v a l u e s aud the lengths of a s s o c i a t e d J o r d a n c h a i n s a r e {~1=-1, ~2=-1, ~3=2}
and {&l=3, ¢2~1, &3=l}, respectlvely. From Theorem 3.9 we have the l e f t Jordan
chain of Dr(~)
(Plo,Pll,Pl2) ffi 0 1 0 ; (P20) " ; (P30) ~ 0

1 2 1 2 1
Thus, from Theorem 3 . 8 , we have the l e f t generalized eigenvectors of AC as
i 1 ,Ck)(~ )DT
P c l j ffi Tp k'T~r f rhPi(j-k)
k=0
where
69
Zp- Io i:,oi [o oil

-3 0 0 '0
o o oLo 0 [;~r
0 1
- -o--o-I£ - 1 - 1
oolio]
(I)
(X)-
1
2X
o
o
0
0
o
~
(2)
(X)- [o o1
o
2
o
ooo
0 o
0
o
Thus,
PcI0 = TpSr(%I)D~hPI0 = [-2'-I'I'0'0]T
Pc11 = Tp[~r(XI)D~hPII+~I)(XI)D~hPl0 ] = [1'1'0'1'1]T
Pcl2 = Tp[~r(XI)D hPl2 +~ hPll + ~ ~r urhPl01 " [ - I ' - I ' I ' I ' 0 ] T
Pc20 = Tp~r(~2)D~hP20 = [ 3 ' 0 ' 0 ' I ' I ] T
Pc30 = Tp~r(~3)D~hP30 = [ 1 ~ 2 ' I ' 0 ' 1 ] T
If the left generalized elgenvectors Pcij are known, then the left generalized
latent vectors of D (%) can be determined from Theorem 3.7 as

r
T
Pij = BcPcij; 0<JK£i-l' l<iS3 .
From~j, the right Jordan chains of Dr(X) can be determined as
-
[o 033 o11111]
I 0.33333 -0.88889 ; (q20) =
[0.,,,33]
-0.33333 ;
(ql0'q11'ql2) 0 0.0 0,0 0.0
r 0.IIIII-
(q3o) - i-o.~1~ii .
L o.o
Thus, from Theorem 3.5, we have the right generalized eigenvectors of A c as
qcl0 = ~r(Xl)ql0 = [0'O'0'I'-I]T
qcll = Sr(Al)qll+$~l)(Al)ql0 = [-0"33333'0"33333'-0"33333'0"33333'0"66667]T

7O
I (2)(
qcl2 ~ ~r(Al)ql2+~(rl)(Al)qll+ 2 ~r A1)ql0 - [-0.IIIII,-0.22222,0.55556,
0.88889,t.22222] T
-
qc20 " $r(A2)q20 " [ 0 " 3 3 3 3 3 ' - 0 " 3 3 3 3 3 ' 0 " 3 3 3 3 3 ' - 0 " 3 3 3 3 3 ' 0 " 3 3 3 3 3 ] T
qc30 ~ +r(A3)q30 = [0.11111,0.22222,0.44444,-0.11111,-0.22222]T
From Theorem 3.6, we can compute the right generalized latent vectors of Dr(A)
if the right generalized eigenvectors of A c are known. Using the calculated
generalized eigenvectors, i t can be easily verified that
T T T T T .T,
[Pcl2'Pcll'PclO'Pe20'Pe301 tqcl0,qcll,qcl2,qc20,qc30] = 15 •
Example 3.2
Consider a nonsingular A-matrix
D(A), [ A4+9~3+31A2+50A'33A2
7A+12 ~4+IOA3+38)L2"67~k+49
]A2.8~+18
which is not a canonical A-matrlx (check with Definition 3.1). Using column
equivalent transformation (Theorem 3.3), we have
D(A) ~ Dr(A)Ur(A)
where
or ['+72+1717+
11
71
and
l+2 ~+3 ] A+2 -A-3 ]

or(A) -
U:I(A)-
k+l X+2 -l-I l+2
It is easy to verify that Dr(k) is a column-reduced canonlcal k-matrix and Ur(A)
is a unlmodular A-matrlx. Followlng the same a p p r o a c h i n Example 3 . 1 j we h a v e
t h e J o r d a n form m i n i m a l r e a l i z a t i o n of D-I(A):
r
D:I(A)= ~ j ( A I r - A j ) - I B j + D J
where
Aj = [1oo] -I 0 0
O -3 0
0 0 -2
Bj
= - I
-I
-I
1
~ .[o., -o.,~ o.~ -,.o]
3.0 -1.5 0.5 -3.0
~ : [ o o o]
The l a t e n t roots and t h e l e n g t h s of associated J o r d a n c h a i n s o f D (X) a r e

r
{AI=-I, A2=-3, A3=-2} and {~I=2, ~2ffil,' ~3=1}, respectively. From Theorem 3.9,
we h a v e t h e l e f t Jordan chains of D (A):
r
1 -1
(PIO'Pll) ~ [0 1] ' (P20) . [_3] , (P30) = [ _~]
and the right Jordan chains:

72
(qlO,qll) . [0.5
3.0 -0.75]
-1.5 '
..fqio~ = [ 0 . 2 5 1
L0-5 J ' (q30) = [-I.0]_3.0
U s i n g Theorem 3 . 1 1 , we h a v e a l e f t Jordan chains of D(~):
(~10'~11) " [ 0'-'1,I <;=o,-I-i], <~,o>-[-,

,]
and t h e r i g h t Jordan chains:
0.5 -5.5
-qlo " u~ICxl~qio - [0 ' -~] [~.o] - [ ~.o]
~11 = UTI(Xl)qll+UT(1)qlo
1 -i ] r-0. 75 -i -~1r0.~1. [-o.~l

"[o <-~.~ j + [-~ ,,<~.0, ~.o~
- . olro.:51 : [-~.25
q2o " U:1(~'i)q20 ['~ -lJlO.5 J ]
-- -1.0 ]
q30 " UTl(Jt3)q30
". [7 -01][-3.0] " [ 3.0_1.0
From t h e left Jordan chains o f D(A) and Theorem 3 . 1 3 , we h a v e a c o m p l e t e set of
left solvents :
'-1- <,.1o~.,-" [o~1 ~,] <~,o~1~,"
I 1 I]
-1 0
ci
o
II I n | II II
g
| i . iI
' I' I 0
o Q o ! ! 0
i ,~I I I
I I ~=,
o 0
i I
I !
I
i I'
Q ~
Q~, ! u i !
• u
i i ° ~
i i I
! i
i i I I
0
l !
I i i I
m i !
• I ! I I I
l-m l - i
I II N .~I n I 8
I
O.I~" 0 w ~
It
0 ~
~J
I I ~-~
0
I ' I i'
~ O I 0
a I !
0
i I i I
a,
0
(D
E3
I.J
W
.I
¢D
lID
CHAPTER IV DIVISORS AND SPECTRAL FACTORS OF NONSINGULAR k-MATRICES
An essential difficulty in structural decompositions [30,31] of MIMO
systems lies in determining the left/right divisors and spectral factors of a
nonsingular )k-matrix. The state-space mlnlmal realizations of the inverse of a
nonslngular )k-matrix discussed in Chapter III will be used to determine the
left/right divisors [30-31] and spectral factors [51-53]. In this chapterp we
investigate the existence of the left/rlght divisors and the spectral factors of
a nonsingular A-matrix. The so-called geometric approaches [5,47-50j81-82] are
employed to derive the structure decomposition theorems of a nonsingular )k-
matrix.
First, we shall define the left and right divisors of )k--matrices as
follows:
Definition 4.1 Given A()k)ccmxm[)k]. If
A()k) = L(A)R()k) (4.1)
where L(%), R(k) ccmxm[%], then L(A) is a left divisor and R(A) is a right
divisor of A()k), end L()k), R()k) are left and right spectral factors of A()k). If
det(L(A)) z 0, (det(R()k)) ~ 0), then L(A), (R()k)), is a nonsingular left (right)
divisor of A(A). Also, if L()k), (R(A)), is a column-reduced (row-reduced)
canonical )k-matrix, then L(k), (R(k)) is a canonical left (right) divisor of
A()k). []
Definition 4.2 A left (right) divisor L()k) (R()k)) of A()k) is nontrivial if
L()k) (R()k)) is not a unimodular )k-matrix. []
If U(A) is a nnimodular )k-matrix, L()k), (R(A)), is a left (right) divisor
of A(A), then L()k)U()k), (U(A)R(A)), is also a left (right) divisor of A()k).

75
Furthermore, from Theorem 3.3 any nonsingular k-matrix can be reduced to a
column (row) reduced canonical k-matrlx by post multlplylng (premultiplying) a
unimodular x-matrix. Thus, to investigate the structures of left/rlght divisors
of a nonsingular A-matrix, it is sufficient to study the structures of canonical
left (right) divisors of a column (row) reduced canonical A-matrix.
In Section 4.1, we shall discuss the state-space structures of the
canonical left divisors of column-reduced canonical A-matrices; the state-space
structures of the canonical right divisors of row-reduced canonical A-matrlces
are dealt with in Section 4.2. Also, constructive proofs on the existence and
properties of canonical left/right divisors and complete sets o£ canonical
left/right divisors are provided in Sections 4.1 and 4.2. Section 4.3 is devoted
to investigating the state-space structures of the spectral factorization of
nonsingular A-matrices. The computational aspects of the left/right canonical
divisors are presented in Section 4.4. The main theorems used to implement the
computational algorithms are block diagonalization and block triangularization
[54,55] of the system map, which is the minimal realization quadruples of the
inverse of a nonsingular A-matrix. A numerical method to compute the left/rlght
canonical divisors using the matrix sign algorithm [56-60] is presented. Some
illustrative examples are provided in Section 4.5.
4.1 Structure Theorems for Canonical Left Divisors and Complete Sets of
Canonical Left Divisors
In this section, we develop the. structure theorems for canonical left
divisors of a column-reduced canonical %-matrix or a right characteristic %-
matrix of a reachable pair (A,B) in an MIMO system. The main tool used to study
the structural theorems of this section is the reachable state-space realization
[5,47-50] for the inverse of column-reduced canonical A-matrices. The complete
set of canonical left divisors of a column-reduced canonical A-matrices is
defined, together with the structure theorem for complete sets of canonical left
78
divisors.
Consider a reachable pair (A,B). Let X =Ca he t h e a s s o c i a t e d state space,

n
U ffi Cm be the input space, and A: X ÷ ~ B: U÷X. Also, let S- C s be an A-
invariance subspace of X where ns<n , or AScS, and assume that V is the canonical
projection of X on X / ~ or V: X÷X/S. Denote B ffi Im(B), and B L ffi ~+S)/S. Let
A L be the induced map in X/S by A, and B L ffi V'B, then B L - Im(B L) and the
following diagram commutes:
A
B _X ;X
(4.2)
It is well known that V is epic [5], and if (A,B) is reachable, then the induced
n -I
pair (AL,B L) in X/S is also reachable, or X/S ffi <ALIBL> ~ BL+ALBL+...+AL v BL
which is the reachability subspace [5] of (AL,B L) and n v ffi n-n s.
n xn n xm
Theorem 4.1 Let (A,B) and (AL,BL) , where Ace n×n, BEe nxm, AL¢C v v, BLOC v ,
be reachable pairs with right characteristic l-matrices Dr(1) and DLR(1) ,
respectively, and n~nv>0. Then DLr(k) is a canonical left divisor of Dr(1) iff
(AL,B L) is an induced pair of (A,B).
Proof:
(i) Sufficient Part:
Let X be the state space of (A,B), S he an A-invariant subspace of X and V:
X~S be a canonical projection. Then, from Eq. (4.2), we have
ALV ffi VA (4.3a)
BL = VB (4.3b)
77
Alsoj from Eq. (2.14b)~ we obtain
BDr(k) ffi ( k I n - A ) T c l * r ( k ) . (4.4a)
Premultiplying Eq. (4.4a) by V yields
VBVr(k) = V(k~n-A)z~l,r(k )
Or
BLDr (~) ffi (~kln _AL)VTcl,r (%) (4.4b)

v
Since (AL,BL) is an induced p a i r of (A,B), (AL,B L) is reachable and the i n v e r s e

Of the right characteristic h-matrix of (AL,BL) can be realized using Lemma 2.1
as
where
are defined as in Eqs. (2.14e), (2.11b) and (2.9a) for the reachable pair
(AL,BL). Using the results in Eqs. (4.4b) and (4.4c) yields
Since the right-hand side of Eq. (4.5) is a k-matrix, DLr(k) is a column reduced
left divisor of D (k).
r
(2) Necessary Part
Let DLr(k)R(k ) ffiDr(h ) where R(%)¢cmXm[k]. Since

78
BLDLr(~) ffi (~I v -AL)TZ~LR(~), we have
(4.6)
BLDr(~) = (%1n - A L ) T ~ L r ( ~ ) R ( ~ )
V
•
Thus ~ci(Dr(k))>~ci(~Lr(k)R(A)), and ~Lr(A)R(A) can be wri=ten as
n xn
(4.7)
~Lr(X)R(X) ffiW~r(X) , WgC v •
Substituting Eq. (4.7) into Eq. (4.6) gives
Or
(4.ga)
Using the results in Eqs. (2.13), (2.14b) and (2.15b), Eq. (4.8a) becomes
(4.8b)
(ki n -AL)-IB L = T~W[(%In-A)TCI]-IB
V
A I nvxn
Define V = T~cWTcEC~ . Then Eq. (4.8b) can he expressed as
(kin -AL)-IBL = V(AIn-A)-IB

V
or
== i-l_ . -i ~ " " (4.9)

~'L ~L A " VAI-IBk-I
ill iffil
By e q u a t i n g the m a t r i x c o e f f i c i e n t s i n Eq. ( 4 . 9 ) , we get

79
~ B L = VAXB, i>0
or
[ALV-VA] [B,A~,A2B .... ] = [0 n xm,0n ~ , - . . ]

V V
S i n c e (A,B) is reachable, we have
ALV " VA;BL " VB

(4.10)
A l s o , s i n c e (AL,BL) i s r e a c h a b l e and
n -I n -I
[ B L , A L B L , . . . , A v BL] = V [ B , A B , . . . , A v B]
we have
n -I n -I
rank[BL,ALBL,...,AL v BL ] = nv~;min[rank(V),rank(B,AB,...,A v B)] .
n xn
Thus, rank(V)~n v . But vEC v so we o b t a i n rank(V) • nv, ie. V is epic. Thus,
V is a canonical projection V: X÷X/S, where S• Ker(V) and S is an A-invariant
subspace o f X. From t h e r e s u l t s i n Eq. (4.10) we can c o n c l u d e that (AL,B L) i s
an induced p a i r o f (A~B) i n X / S v l a V. •
Some s t r u c t u r e properties of the left divisors of D (~) are described as

r
follows.
Lends 4.1 Let (~1,¢2,...,~m) be the Kronecker indices of (A,B) and
(<LI,<L2,...J~Lm) be the Kronecker i n d i c e s of (AL,BL) , then
KLi~~'.i"Then, we obtain {bLI,ALbLI,...,
-1 -i
ALL1 bLl} = {VbI,VABI,...,VA LI b I} which contains independent vectors.
However, Akbl-- for k><'1 is dependent of AJb I for j " I,...,K i. Thus,
vAkbl(=ALkbLl) for k>~ i is dependent of VAJbl(m~bLl) fOr jffil,...,~i. The
above result contradicts to the assumption, and therefore KLI<K1 .
(2) Let KLi~~ik" Then, we obtain
~LI-I
{bLI,~LbLI,...,
~ bL1,...,ULik,%bLik,...,~Lik-lhLikt -
KLI-I
{Vbl,VAbl,... ,VA bl,...,Vbik, VAbik ,...,vALik-lbik } which contains
independent vectors. However, for k>K..
Ik
• <-i
So that
vAkb ik = ~k I %~jVAJb~
~=I j-i
From the hypothesis KLiKK i for iKik, we have
)_' ak~jVAJb ~ " ~ ak~jVAJb ~

j=l j-l
• ~,.CI
81
This leads to a c o n t r a d i c t i o n , and therefore KLik~Kik.

(3) By mathematical induction, we obtain
KLi~Ki, IKiKm •
Lemma 4.2 Let P(A,B) and P(AL,B L) be the reachability base matrices of (A,B)
and (AL,BL) , and (Kl,...~Km) ' let (KL1,...,gLm) be the Kroneeker indices of
(A,B) and (ALJBL) 9 r e s p e c t i v e l y and l e t V be the c a n o n i c a l p r o j e c t i o n shown in
Eq. ( 4 . 2 ) . Then
(4.11)
VP(A,B)H " P(AL, B L)
where
1 KLI O1+i OI+~L2 Om-I +I Om-l+KLm. nxnv
H = [ e n P . .. ~e n ,e n ,... pe n p.. .,e n p...pe n JEL
i
IKiKm
Oi = j~l= ~J '
Proof:
%-I %-1
Since VP(A,B) = [bLI,ALbL1;...,A L bLI,...,bLm,ALhLm,...,A L bLm] , from
Lemma 4.1 and the definition of H~ the result in Lemma 4.2 follows.
Theorem 4.2 Let V, H, P(A,B), P(AL,BL) be the same as those defined in Lemma
4.2. Then
(4.12)
V " P(AL,BL)HTF(A,B)-I
Proof:
The result follows from Lemma 4.2 and the fact that HTH ffi I
n
v
82
Lemma 4.3 Let T c be the ~ransformation m a t r i x for the c o n t r o l l e r canonical

form. Then T ffi T-1p-I(A,B) where
c cp
T c p = [Tcp]ij; KiWi>0; l~i~m; l~j~m
[Tcp]ii The r e v e r s e d upper t r i a n g u l a r T o e p l i t z m a t r i x with f i r s t row
[arii2,...,ariiKi,l]
[Tcp]ij The reversed upper triangular Toeplitz matrix with first row
[arij2,... 'arijKj,0] if K.<~.

i j or [arlj2' . "'arij<i
. . . .'0' ,0]
if K.<K..
z j
Proof:
Lemma 4.3 can be verified by direct computations.
Corollary 4.1 Let T c and TLC be the similarity transformation matrices which
t r a n s f o r m (A,B) and ( % , B L ) to t h e i r c o n t r o l l e r forms, then
-I (4.13)
V = TLcWTc
where
W ffi TL~pHTTep; T" I = TLcP(AL,B L) ; T c p = p-I(A,B)T~I
Lcp
Proof:
From Lemma 4.3, we obtain P (A,B) = T-IT-It

cp and P(AL,B L) u TLCTLep.-I
-i Thus,
from Theorem 4.2 we have
V = p (AL,BL) HTp-I(A,B ) ffi TL1 TL~pHTTcpTc = TLCr~rrC

-1
83
-I T
where N = TLCPH Tcp. •
From Eq. (4.5), the properties of the right divisor pairing with the
canonical left divisor _DLr(X) of Dr(k) can be stated as follows.
Corollary 4.2 Let DLr(X ) be a canonical left divisor of Dr(h) , which is
defined iu Theorem 4.1, and Dr(X ) = DLr(X)R(k ). Then
Cl) Rex) - ~rC0)W~rCX)+(Zm-~rC0>%rC0))D~6r¢X>
where DRh = D L hDrh is an upper triangular matrix with diagonal elements all
l's.
(2) R(X) is a nonsingular X-matrix.
Proof:
Cl) R(A) = L
~ C ' V ~T : I ~r ( k ) + D ' D r ( ~ )
(2) Since both Dr(h) and DLr(X) are column-reduced canonical h-matrices, R(X)
is nonsingular. •
For some applications (refer to Chapter VI) of the divisor theorems, we
need to have a full set of divisors whose spectra cover the entire spectrum of a
X-matrix. We s h a l l discuss the complete set of left divisors before leaving
this section.
84
Definition 4.3 If L c = {Di(k),i-l,...,k } is a set of left divisors of a

k
nonsingular X-matrix A(k) such that Di(~) are nontrivial and i~ I ~(Di(k)) ffi
~(A(k)), then L c is a complete set of left divisors of A(A); if Di(k) are
canonical left divisors, t h e n Lc i s a complete set of canonical left divisors.
If any complete set of left divisors of Di(k ) contains o n l y one e l e m e n t w h i c h i s
column equivalent to Di(k)j then Lc is an irreducible complete set of left
divisors of A(k); if Vi(k) are canonical left divisors, then Lc is an
irreducible complete set of canonical left divisors. The complete set of
canonical right divisors and irreduclble complete set of canonical right
divisors of A(k) can be defined in the same manner. O
The structure theorems for a complete set of canonical left divisors are
formulated in Theorems 4.3 and 4.4; while the irreducible complete set of left
divisors is discussed in Section 4.4.
Theorem 4.3 Let S I and S 2 be complemented A-invariant subspaces of X(where
$1~ ~ = ~ and X = ~ Q S 2 ) , and (ALi,BLi) , i-1,2, be the induced reachable pairs
of (A,B) in X/Si, i-1,2. Then the followlng diagram comuutes.
AL 1
, Xl$1 :. X/$I
(4.14)
L/ .B =X A ~X
XI$ 2 ;.- XIS2
The set of the right characteristic k-matrices {DLr.(k),i=l,2}, of (ALi,BLi),

I
i=1,2, i s a c o m p l e t e s e t of c a n o n i c a l left divisors of D ( k ) .
r
Proof:
From Theorem 4 . 1 , DLr.,i=l,2 are canonical left divisors of Dr(k). Since

1
ViA = ALiVi, we obtain
85
[vl]A o][vl] (4.15)

V2 0 AL2 V2
Because Ker(V i ) ffi S i and SIOS2 = ~, we have K e r ( V l ) ~ e r ( V 2) - #. Also, S I ~ S 2

ffiX, therefore dim(Ker(Vl))+dlm(Ker(V2)) ffi dlm(X) ffi n. Let ~ ~ [VI,V2
] T T T. Since
Ker(~) = Ker(Vl)~Ker(V 2) = ~, we have dim(V) = n, ie. ~ is nonsingular. As a
result, we obtain take over ~AV-1 I block diag(AL1,AL2) , and
O(DLrlCA))uO(DLr2(A)) - OCALI)UU(AL2) " oCblock diag(ALi,AL2)) - ~(A) -

O(Dr(A)). Thus, {DLr.(l),iffil,2 } is a complete set of canonical left divisors of
1
Dr(k). •
The results of Theorem 4.3 can be easily extended to a more general case as
follows.
Theorem 4.4 Let S i , i f f i l , . . . , k be independent A - i n v a r i a n t subspaces of X such

k
t h a t i__~l S i - X. Define Xi ~ ~ S and ( A L i , B L i ) , i = l , . o . , k be the induced
reachable pairs of (A,B) in X/Xi, and

DLr.(A) , i - l , . . . , k be the r i g h t
1
characteristic A-matrlces of (ALi,BLi),i=l,...,k. Then {DLr.(A),i-l,...,k} is a
complete set of canonical left divisors of Dr(A). •
4.2 Structure Theorems for Canonical Right Divisors and Complete Sets of
Canonical RiKht Divisors
Parallel to the structural analysis of the canonical left divisors for
column-reduced k - m a t r i c e s i n S e c t i o n 4 . 1 , we d e r i v e the s t r u c t u r e theorems for
the c a n o n i c a l r i g h t d i v i s o r s and the complete s e t of c a n o n i c a l r i g h t divisors
for row-reduced A-matrices i n t h i s s e c t i o n .
LetX =Cn be the s t a t e space, Y - C p be the output space, and A: X ÷ ~ C:

n
s
X+~ Let S= C , ns<n be an A-invariant subspace of X, and S be the canonical
injection map s: S*×. Define ~ = A~and C R = C IS. Than the following diagram
86
commutes.
X A
(4.t6)
It £s known that S is monic, and if (A,C) is observable, then the embedded pair
usI Ker(CR~ -I)
(AR, C R) is also observable, or i~ 0.
Dual to Theorem 4.1, we h a v e t h e following results.
Theorem 4.5 Let (A,C) and (AR,CR) be observable pairs with left characteristic
~-matrices D~(~) and DR~(X) , respectively and

n xn pxn s
AcCuxn c~cPxn ARC C s S CR~ C ,n~ns>0. Then DR~(X) is a canonical right
divisor o f D~(k) ire (AR, C R) is an embedded pair o f (A,C), or
s
are defined in Eqs. (2.17b), (2.20c) and (4.16), respectively. •
The structure properties of the right divisors of D~(~) are as follows,
beginning with the analogues of Lemmas 4.1 and 4.2.
Lemma 4.4 Let (~l,V2,...,vp) be the observability indices of (A,C) and
(~RI,VR2, .... ~Rp) be the observability indices of (AR, CR). Then ~RI~Vi, l~i~p. •
Lem~a 4.5 Let Q(AjC) and Q(AR, C R) be ~he observable base matrices of (A,C) and
(AR, CR) , respectively. Let S be the canonical injection map in Eq. (4.16).
Then
~:A,C)S = Q(Az, CR) (4.17)

87
where
^ 1 ~RI TI+I TI+~R2 Tp-I+I T +V_ _ n Xn
H = [en,...,e n ,e n ,...,e n ,...,e n ,...,enP-I ~P]£EC s
i
~i = ~ vj , l<i<p
j-t
(VI,...,~ p) and (VRI,...,VRp) are the observability indices of (A,C) and
(AR, CR) , respectively. •
It then follows directly from Lemma 4 . 5 :
Theorem 4.6 Let S, H, ~(A,C), and ~(AR, C R) be the same as those defined in
Lemma 4.4. Then
S = Q-I(A,c)HT~(AR, CR) (4.18)
Lemma 4 . 6 Let TO be the similarity transformation matrix for the observer
canonical form and Q(A,C) be the observability base matrix. Then
TO = ~ - I ( A , C ) T ; ~ (4.19)
where
V,X~,
TOp = [Top]ijEC ~ J, ISiSp, l~j~p, V.V.>0
i]
[T0p]ii The r e v e r s e d upper t r i a n g u l a r Toeplitz matrix with first row
[a~ii2,...,a£iiV .,I]
%
[Top]ij The r e v e r s e d upper t r i a n g u l a r Toeplitz matrix with first row
[a~.ij2,...,a~iju .,0] f o r VI>V j or [ a ~ . i j 2 , . . . , a ~ i J v i , 0 ,...,0]

J
f o r V.<V.. •
i j
Corresponding to Corollaries 4.1 and 4.2, respectively, we have:

88
C o r o l l a r y 4.3 Let TO and TR0 be the s i m i l a r i t y t r a n s f o r m a t i o n m a t r i c e s which
transform (A,C) and (AK, C R) to their observer forms. Then
^ -I (4.20)
S = T0~rrR0
where = TOp.T;0p,TOp= T;1Q-1C ,c )-1

TR0 P = ~(AR, CR)TR0 •
C o r o l l a r y 4.4 Let Dp~(A) be a c a n o n i c a l r i g h t divisor of D~(X) d e f i n e d i n
Theorem 4.5 and D¢(~) = L(X)DR£(X). Then
(1) L(~) - ~(~) W~ (O)+~9(~)VLh(Im-~R~(O) $~ (0)) -1h i s a low

where DLh = D~hDp~
triangular matrix with diagonal elements all l's.
(2) L(A) is a n o n s i n g u l a r k-matrix. •
The s t r u c t u r e theorems of the complete sets of canonical right divisors
defined in Definition 4.3 can now be given as follows.
Theorem 4.7 Let S 1 and S 2 be complemented A-invariant subspaces of X(where
SIOS 2 = ~ and X = SIQS2), and (ARi,CRi),i=I,2 , be the embedded observable
pairs of (AjC) in Sl,i=l,2. Then the following diagram commutes.
$2 = $2 CR2
x A X ~ V (4.21)
The set of the left characteristic l-matrices, {Dp~.()0,i=l,2}, of (ARI,CRI),

z
i--1,2 is a complete set of canonical right d i v i s o r s o f D~.(X).
89
Proofs
From Theorem 4 . 5 , D R E i , i - I , 2 are c a n o n i c a l r i g h t d i v i s o r s of D~(~). Since

AS i = SIARI , we o b t a i n
A[SI,S 2] = [S1,S2]-block dlag[Azl,AR2] •
Because Im(Si) - 3 i and Sln ~ = ~, we have Im(S1)NIm(S2)= ~. Also, S I ~ S 2 =

I m ( S l ) ~ I m ( S 2) = X, therefore rank(S 1) + rank(S 2) " dim(X) = n,
Since Im([Si,S2] ) = Im(Sl)+Im(S2) = Im(Sl)~Im(S2) = X, or rank([Si,S2]) =
dim(X) = n, and [SI,$2] is nonsingular, therefore, we obtain [SI,S2]-IA[SI,S2] =
block dlag[AR1,A~2]. Thus,
O(DR¢I(~))uoCDR~,2(~)) = OCARI(~))uo(AR2(~)) - oCblock diag(AR1,AR2))
= oCA) = oCD¢(1)) .
Now, we can conclude t h a t {DR¢i(~),i=l,2 } is a complete set of c a n o n i c a l r i g h t
d i v i s o r s of D~(A). •
The r e s u l t s of Theorem 4.7 can be extended to the following more g e n e r a l
cage¢
Theorem 4.8 Let S i , i = l p . . . , k be independent A - i n v a r i a n t aubspaces of X such

k
that X= i=~1 Si, and ( A R i , C R i ) , i = l t . . . , k be the embedded observable p a i r s of
(A,C) in S i , i = l , . . . , k . Also, l e t D R ~ i ( ~ ) , i = l , . . . , k be the l e f t c h a r a c t e r i s t i c ~-
matrices of (ARi,CRi). Then { V R ~ i ( ~ ) , i = l , . . . , k } is a complete set of canonical

r i g h t d i v i s o r s of D~(~). •
4.3 S t r u c t u r e s Theorems for S p e c t r a l F a c t o r l z a t l o n of Nonsiusular X-matrices
We now i n v e s t i g a t e s t r u c t u r e theorems for s p e c t r a l f a c t o r i z a t i o n [50-53] of

90
n o n s i n g u l a r X-matrices.
Let (A,B,G,D) be a minimal r e a l i z a t i o n of A-l(%), where A(%)Ecn~m[%] is a
nonsingular h-matrix. Let XECn, U~Cm and YECm be the s t a t e , input, and output
n
spaces, respectively. SE C s is an A-invarlant subspace of X. (AL,B L) and
(AR, CR) are the induced and embedded p a i r s in X/S and S, v i a the epic map V and
the monic map S, o£ (A,B) and (A,C), r e s p e c t i v e l y . Then the following diagram
CO~Uteg.
~ "L ~ X/$
BL -, IV
(4.22)
s % ~-
CR
The sequence S÷X÷XIS is exact at X, i . e . VS - O(n_ns)Xn s. Also, (AL, BL) and

(AR, CR) are reachable and observable Pairs , respectively.
From the epic map V and the left inverse of the monic map S, we can
construct a transformation T to block-trlangularize the system map A:
Lemma 4.7 Let S + ffi (S*S)-Is * and V + TM V*(VV*) -I be the left and right inverses
of S and V, respectively. Define
(4.23a)
Then
(a) T is nonsingular and T -1 - [$,V T]
(b) Let A T = TAT -I , then

91
. [~'~ S+AV+ nsxm s (n-n s) x(n-n s) (4.23b)

, ~tc , A,~C
L o
(:) o(A) = oCAL)uo(A &) (4.23c)
Proof:
(a) Let T = [S,V*I. Then
TT =
[ ]r "V]
S+S S+V+
VS
W+
= In s
L vs In_ n
S
-I
n
so t h a t T- I = [S,V +]
(b)
AT"
[,] V
A[S,V +1 -
V~ VAV+ J
s i . : e ALV - VA and S ~ " AS, we obtain S÷AS " AS; VAS - VSA R - 0 , rAY + " ~ , o.d
therefore, the r e s u l t in Eq. (4.23b) follows.
(c) The r e s u l t in Eq. ( 4 . 2 3 c ) can be proven d l r e c t l y from Eq. ( 4 . 2 3 b ) . •
From Theorems 4 . 1 , and 4 . 4 , and Lee~a 4 . 7 , we a r r i v e at the followin8
results on the s p e c t r a l f a c t o r i z a t l o n of nonsingular A-matrlces:
Theorem 4 . 9 Suppose (A,B,C,D) is a mlnlmal reallzation quadruple of the
inverse of a nonsingular A-matrlx, A(A) , and the asociated (AEICR) and (AL,B L)
are defined in Eq. (4.22)° Then A(A) can be factored as
(4.24)
A(A) - DE(A)UA(A)DR(A)
where DL(A) and DR(A) a r e the r i g h t characteristic A-matrix of (AL,B L) and the
92
left characteristic l-matrix of (AR, CR) , r e s p e c t i v e l y . UA(A) i s a u n i m o d u l a r k-
matrix.
Proof:
From Theorems 4.1 and 4.5, we have
(4.25)
A(Z) = DL(k)R(Z) = L(k)DR(A)
From Lemma 4.7, there are no shared latent roots between DL(k ) and DR(1). Thus
DR(A) must be a right divisor of E(k), or R(k) = gA(A)'DR(A) and A(A) =
DL(k)UA(k)DR(k). Now, we obtain det(A(A)) = det(DL(A))det(UA(k))det(DR(k)).
Thus from Lemma 4.7 we observe that det(UA(k)) is a nonzero constant, and
therefore UA(k) is a unimodular A-matrix. •
When a column/row-reduced non-canonical k-matrix~ A(A), is given and the
structures of the column/row-reduced, canonical left/right divisors of A(A) are
of interest, we develop the following results.
Theorem 4 . 1 0 Let (AjB,C,D) be a minimal realization of the inverse of a
column-reduced l-matrix, A(k), where AEC n×n, BEC nxm, CE Cain and DEC mxm. Let
n xn n xm
(AL,B L) be a reachable pair, where ALEC v v BL¢C v and nanv>O. Also, let
DLr(k) be the right characteristic l-matrlx of (AL,BL). Then, DLr(k) is a
canonical left divisor of A(k) iff (AL,B L) is an induced reachable pair of
(A,B).
Proof:
Since A(A) = Dr(A)Ur(A) , from Theorem 4.1, the results of Theorem 4.10
follow.
Corollary 4.5 Let (A,B,C,D) be a minimal realization of the inverse of a

93
^
column-reduced k-matrlx, A(k), T c be t h e t r a n s f o r m a t l o 1 1 m a t r i x which t r a n s f o r m s
(A,B) i n t o a controller form (,~,,,." T c A T : 1, Bc = ; c B ) . bl,o, let VLr(k) be t h e
right characteristic k-matrlx o f an induced reachable pair (~,BL) o f (A,B) v i a

11 Xn
the c a n o n i c a l p r o j e c t i o n veC v . Then
(4.26)
A(A) = DLrCA)R(k)
where
o~.lrC~,~ ~,..ck~ ~I.,÷~,,

- n n - rand(V)
V
and ~(k) i s d e f i n e d in Lenma 3 . 1 .
Proof:
From Lemma 3 . 1 and Eq. ( 3 . 2 ) , we o b t a i n
A A
BoA(k)= (k~11 -Ac) ~c (k)

V
or
A
BA(k) = (kln-A)T:l~c(k) .
From Eq. ( 4 . 2 ) , we have
BLA(k~ - (k~ n -AL~V~:l~c(k~ .
Thus
As usual~ the dual results on t h e spectral factorlzatlou of row-reduced X-

94
matrices can be written down at once:
Theorem 4.11 Let (A,B,C,D) be a minimal r e a l i z a t i o n of the i n v e r s e of a row-
reduced A - m a t r l x , A ( k ) , where AECuxn, BeCnap, CcCpxn and DeCpxp. Let (AR, CR) be
n xn pxn
an o b s e r v a b l e p a i r , where ARE C s s, CREC s, n>ns>O" A l s o , l e t DRE(A) be the
left characteristic k-matrlx of (AR, CR). Then, DR~(A) is a canonical right
divisor of A(A) iff (AR, CR) is an embedded pair of (A,C). •
Corollary 4.6 Let (A,B,C,D) be a minimal realization of the inverse of a row-

A
reduced X-matrix, A(k) and let T O be the transformation matrix which transforms
(A,C) into an o b s e r v e r form (A0,C0) , where A0 - T;1AT0, = CT0. Also, let
DR~(A) be the left characteristic k-matrix of an embedded observable pair
(AR, C R) of (A,C) via the canonical injection S. Then,
(4.27)
A(k) = L(k)DRA(k)
where
DR~(k) = CR(AI n -AR)-IBR+DR; n s = rank(S)

a
L(k) = ~0(X)T;lS~R+A(k)~R ,
and ~0(A) is defined in Lemma 3.3. •
Theorem 4.11 and Corollary 4.6 can be easily proved by using the same
a p p r o a c h e s as given in Theorem 4.10 and C o r o l l a r y 4 . 5 .
4.4 Computational Aspects of Divisors and Spectral Factors for k-matrices
In this section, we present schemes for computing left/right divisors and
spectral factors of nonsingular k-matrices, and their associated complete sets
of irreducible left/right divisors and spectral factors. The computational

95
schemes are mainly based on b l o c k triangularization or block diagonalization
[54,55] for the system map of the minimal realization quadruple for the inverse
of a nonsingular X-matrix. At t h e end o f t h i s section, a numerical method for
block triangularization and b l o c k diagonallzation of a square matrix, using the
matrix slgu algorithm [56-60], is presented.
Let (A,B,C,D) be a minimal realization for the inverse of a nonsingular k-
matrix A(A). Assume t h a t A c a n b e u p p e r b l o c k triangularized as
A = I AI2 ; AI¢C i l i=1,2; AI2CC (4.28a)

A2 J
and B,C c a n b e w r i t t e n as
T T T i (4.28b)
B = [BI,B2] ; Bi¢C ; i=1,2
mxni; (4.28c)
C = [Cl,C2]; CicC i=1,2
Then, we o b t a i n the following results.
Theorem 4.12 If (A,B,C,D) is a minimal realization of the inverse of a
nonsingular A-matrix A(A), with A, B and C shown in Eq. (4.28), then
(I) The right characteristic X-matrix of (A2,B2) , defined as Dr2(X) , is a
canonical left divisor o f A(X).
(2) The left characteristic X-matrix of (AI,C1) , defined as DLI(X) , is a
canonical right divisor of A(~).
(3) ACX) = Dr2CA)Ur~CX)D~I(~) w h e r e Ur~CX) i s a u n l m o d u l a r X - m a t r i x .
Proof:
96
From Theorem 3 . 3 , A(X) = Dr(X)Vr(X) = U~(X)D~(X), where Dr(X) and D~(X) a r e
column- and r o w - r e d u c e d c a n o n i c a l X - m a t r l c e s , respectively. Each p a r t of the
theorem i s e s t a b l i s h e d as f o l l o w s :
n2xn
(I) Choose V ~ 1 0 , Z n 2 ! ; C n × n l a n d a p p l y Theorem 4 . I .
(2) Choose S [ I n l J 0 ] EC and a p p l y Theorem 4 . 5 .
(3) From Theorem 4 . 9 and t h e r e s u l t s i n (1) and ( 2 ) .
Similar results can be s t a t e d f o r a lower b l o c k t r l a n g u l a r i z e d system map A
of a minimal r e a l i z a t i o n f o r t h e i n v e r s e of a n o n s l n g u l a r h - m a t r l x as f o l l o w i n g .
Corollary 4.7 If (A,B,C,D) is a mlnlmal realization for the inverse of a
nonsingular h-matrlx, A(X), with A,B and C being partitioned as
A = I AI 0 1 ; Aic Cn.xn.
x x ; i=1,2; A21cC
n2xn I (4.29a)
A21 A2
n.~
T TT
B = [B1,B2] ; Bi~C x ; i=1,2
(4o29b)
mxn i (4.29c)
C = [C1,C2] ~ CicC ; i=1,2
then
(1) The right characteristic X-matrix of (AI,BI) , defined as Drl(X), is a

canonical left d i v i s o r of A(X).
(2) The left characteristic X-matrix of CA2,C2) ~ defined as D~2(X), is a
canonical right divisor of A(X).
(3) A(X) = DrI(X)Ur~(X)D~2(X) where Urn(X) is a unimodular h-matrix. •

97
To find a complete set of canonical left or right divisors, we need to
block-diagonalize the system map of A of the mlnlmal realization quadruple for
the inverse of a nonsingular l-matrlx.
Theorem 4.13 If (A,B,C,D) is a minimal r e a l i z a t i o n of A-l(1) and A is block-
diagonallzed as
n.xn.
A = block diag[Ai,iffil,...,k]| Ale C I l| i=l,..., k (4.30a)
and
n. x m
TT (4.30b)
B = [BI,B2,...,B ]; BieC I ; £=i,..., k
(4.30c)
C = [Cl,C2,...,Ck] ; Ci¢ C 1 i=l,..., k
then
(I) The right characteristic k-matrices of (Ai,Bi) , Dri(k) , for l~i~k,
constitute a complete set of canonical left divisors of ACk).
(2) The left characteristic k-matrices of (Ai,Ci) , D~i(k), for l~i~k,
constitute a complete set of canonical right divisors of A(~).
Proof:
(I) Choose V.z = block dlag[0nl ,...,0nl._l,ln.1 '0n'l+l'''''0nk]" From Theorem
4.12, we o b t a i n D .(k) as a canonical left divisor of A(k) for i=l,...,k.

rl
k
Also, using Lemma 2.1 yields O(DriC1)) = O(AI) , and hence U OCVriC1)) -
k iffil
u oCA) - OCA(1)). Thus, {Dri(k) , i'l,...,k} is a complete set of
i=l
canonical left divisors of A(k).
(2) Similar to the proof in (I).
The computational scheme for the irreducible complete set of canonical

98
left/right divisors for a nonsingular k-matrix A(A) can be formulated by using
the Jordan canonical form minimal realization of A-I(A).
Theorem 4.14 If (A,B,C,D) is a minimal realization of A-I(k) and A is in
Jordan canonical form as
A = block dlag[Aji,i=l,...,k ]
n,xn.
where AjIEC i x is a Jordan block associated with an elgenvalue ~i and a full
Jordan chain as
Aj i = • Xi I • (4.31a)
and B and C are represented as
~.oXm
T T T ~ (4.31b)
B ffi [BjI,...,Bjk ] ; gjiCC ; iffil,...,k
mxn.
(4.31e)
C = [CjI,...,Cjk]; CjiEC l i=l,..., k
then,
(I) The right characteristic k-matrices of (Aji,Bji) , Dri(k) for l~i~k,
constitute an irreducible complete set of canonical left divisors of A(A).
Also, {Dri(A),ifl,...,k} is unique.
(2) The left characteristic k-matrices of (Aji,Cji) , DEi(A) for l~i~k,
c o n s t i t u t e an irreducible complete set o f c a n o n ic al right divisors of A(~).
Also, {D~i(~),i'l,...,k} is unique.

g9
Proofl
(1) From Theorem 4 . 1 3 , {Dri(X),i=l,...,k} is a complete s e t of c a n o n i c a l left
divisors of A(X). Since each A j i contains a Jordan block with a full
Jordan chain, A j i cannot be £ u r t h e r decomposed i n t o more than one b l o c k .
This fact shows t h a t {Dri(~),i-1,...,k } is an i r r e d u c i b l e complete s e t of
canonical left divisors of A(X). Moreover, the Jordan b l o c k d e c o m p o s i t i o n
of a square m a t r i x i s u n i q u e , and t h e r e f o r e the i r r e d u c i b l e complete s e t o f
left divisors, {Dri(~),i=l,...,k}, is unique.
(2) Similar to (I). •
From Theorems 4.12 and 4.13, we observe that the canonical left/right
divisors and complete sets of canonical left/right divisors of nonsingular X-
matrices can be determined from the block trlangularlzatlon/diagonalizatlon of
the system map A of the minimal realization quadruple (A,B,C,D) for the inverse
of the X-matrlces. The block triangularizatlon and dlagonallzation of a square
matrix can be reformulated as an algebraic Riccatl equation problem [54] and
several numerical algorithms are available for s o l v i n g such R i c c a t i equations
[55-56]. Here, we s h a l l p r e s e n t a new a l g o r i t h m f o r b l o c k d l a g o n a l l z a t l o n and
block t r i a n g u l a r l z a t i o n of a square m a t r i x u s i n g the m a t r i x s i g n a l g o r i t h m [60].
Let us d e f i n e the m a t r i x s i g n f u n c t i o n f i r s t .
Definition 4.4 The matrix sign function of AeC nxn with o(A)cC÷UC -, where C
and C- are the open right and left plane of C, respectively, is defined by [56]
Sign(A) = 2 Sign+(A)-In - In-2 Sign-(A) (4.32a)
where
1 ~r+ (~in_A)-Idx (4.32b)

Sign+(A) = 2-~'~
and
100
1 (4.32c)
Sign-(A) = 2"~ ~r (Xln-A)-Id~
r÷(r_) is a simple closed contour in C+(C-), and encloses O(A)O6* (O(A)nC'). [3

From Definition 4.4, we can formulate Slgn(A) using the modal matrix M of
A. Let
J " M-IAM " block diag[J+,J_] (4.33a)
nlxn 1 P
where J+eC and J _ E C 2xn2 (n=nl+n2) are the collectlons of Jordan blocks
with o(J+)cC + and o(J_)cC-~ respectively. Then,
Sign(A) = M[Slgn(J+)QSign(J)]M -I ffi M[Inl~(-In2)lS-1 (4.33b)
Assume that M is partitioned as
nxn I nXn 2 (4.34)

M = [MI,M2] , MIEC , M2EC
where M I and M 2 contain the eigenvectors associated with the eigenvalues of J+

and J_, respectively. Let W be the inverse of M and be partitioned as
nlXn n2Xn
M-I ~ W " [ ]WIW 2 , wIEC , W2cC (4.35)
Then, from Eq. (4.32a), we obtain
(4.36a)
Sign+(A) ffi ~I [Sign(A)+I n] = M[Inl ~ On2]M-1 " M1W1
101
Obviously, rank[Sign+(A)] = n I. Let
s IEcnxn s EC nxl (4.36b)

Sign+(A) = [ S l ' S 2 " " ' n' ' i
and
S ~ Ind[Sign÷(A)]¢C nxnl (4.37)
where S i8 a monic map which c o n t a i n s n I i n d e p e n d e n t column v e c t o r s of Sign+(A).
These independent vectors are selected from the n(>nl) column vectors of
Sign+CA) in Eq. (4.36b).
Thus, we obtain the results as follows.
Lemma 4 . 8 There e x i s t s a n o n s l n g u l a r m a t r i x ~ C n l x n l such t h a t S = MI~ and S+

* ~ * -i *
= ~-iM~j where M 1 is defined in Eq. (4.34); S + ~ (S*S)-Is * and M 1 (HIM I) M 1
are the left inverses of M I and S, respectively.
Proof:
Assume that S contains n I columns of Sign+CA) with column indices as k i for
iffil,2,...,nI. Then from the definition of S we obtain
S = Sign÷(A)Us
[e~ I k2 kn
where Us = ,e n j...,e n ] .
d A
Since Sign+(A) = MIWI, we have S = MIWIUs = MI~ , where ~ = WIUs. By
Sylvester's inequality [61], we o b t a i n
rank(W1)÷rankCUs)-n~rank(~)~minCrank(W1),rank(Us))
and s i n c e rankCW 1) = rankCU s) = n l , we c o n c l u d e t h a t
C4.38)
rank(~) = n1
102
From t h e d e f i n f t l o n of the left inverse o£ S÷~ we h a v e
S+ = ( S * s ) - t s * = [(MI~)*(MI~)]-I(MlC) . ~-IMI+
Two i m p o r t a n t properties of S are g~ven by:
&
Theorem 4 . 1 5 D e f i n e S = I = ( S ) and l e t X b e t h e s t a t e s p a c e and A: X÷X. Then,
(1) S i s an A - i n v a r [ a n t s u b s p a c e of X.
(fi) S is the canonical injection map S : S ~
Proof:
(i) From Lemma 4 . 8 , we o b t a i n
(4.39a)
AS = AMI~ = MIJ+~ = MI~(~-IJ+5) = SAR
where
A~ ~ ~-ta+~ (4.39b)
T h u s , S " Im(S) i s a n A - i n v a r i a n t s u b s p a c e o f X.
(ii) S i n c e S i s monic and S = I m ( S ) , S i s t h e c a n o n i c a l injection map S: S ~ •
Obviously, from Theorems 4 . 5 and 4 . 1 5 , S defined i n Eq. (4.37) can be used
as a canonical injection map for finding the canonical right divisor of
nonslngular ~-mstrlces. Since S is composed of the independent column v e c t o r s
of Sign+(A), the orthogonalized projection algorithm in Section 2.3 can be used
t o f i n d S from S ~ S i g n - l ( A ) :
103
Al~orithm 4.1
nxl
Given: - [SI,S2,...,Sn],Si¢ C , for l~i~n
Find: S - [Skl ,Sk2,...,Skn I], such that Ski¢{Si,l~i~n} for l~ki~n I are
independent.
Algorithm:
{Initialization}
j:=0;{set independent vector index starting at zero}
P:=In;{Set orthogonalized projection to identity matrix}
{Froeessing}
For i=l to n Do
Begin
d:=P*Si;
If S.*d=0
%
Then
Begin
j:=j+l;{increment index j}
Sk~:=Si;{v i is an linearly independent column vector}
P:=P-d*d /Si*d;{Update P}
End{If}
End ;{For loop}
n1:=j;{Total number of linearly independent column vectors which is the
rank of S}
Note: The * between two variables in the above algorithm is the matrix
product notation; and the superscript * designates conjugate
transpose. •
The m a t r i x S d e f i n e d in gq. (4.37) can be used f o r b l o c k d i a g o n a l i z a t i o n of
a system map as follows.

104
Theorem 4.16 Let AcC mxn with Re(~i)Z0 for i=Ip2,...n, where {~i,i=l,2,...,n}
is O(A). Define
nXn-
S I = Ind[Sign+(A)] EC 1 (4.40a)
nxn^
S 2 = Ind[Sign-(A)] EC z (4.40b)
where Sign+(A) and Sign-(A) are defined in Eq. (4.42) and nl+n 2 - n.
Assume that nln 2 = 0~ then
AD = M-IAMss = block diag[ARi,AR2 ] (4.41a)
where M s is a block modal matrix and expressed as
Cnxn (4.41b)
Ms [SI,S 2] ¢
and ARlC; lxnl and AR2ECn2xn2 are defined as
+ nlxn 1
AR 1 . SIASI¢ C (4.41c)
AR2 = S;AS 2C;2xn2 (4.41d)
4- 4- + * -l * 4- *
S I and S 2 are the left inverses of S I and S2, or S 1 = (SIS I) S I and S2 = ($2S 2)
- I s *2' respectively~ where * designates conjugate transpose.
Proof:
Theorem 4.16 can be proved using Theorem 4.15 and Eq. (4.39).
105
l£ more than two blocks are needed for the block diagonalization of A, we
can construct the block modal matrix using generallzed sign matrix functions.
Definition 4.5 Let AcC nxn and Re(O(A))n{rl,r2} = @, where rl<r 2 and rl,r2C~
The ~eneralized matrix sign function of A with respect to the open interval
(rl,r 2) on the real axis is defined as
Sign(rl,r2 ) = 2 S£gn;rl,r2)(A)-I n
- (4.42a)
= I n - 2 Szgn(rl,r2
where
. cA1 . (4.42b)
Szgn(rl9 r2) (A) ~[ S l g n ( r l ) ( A ) - S i g n ( r 2 ) (A) ]
• - A (4.42c)
Stgn( rl, r2 ) (A) = In-S ign;r I' r2 ) (A)
and
(4.42d)
Sign(ri)(A) ~ Sign(A-riln) , for i=1,2 .
rl
Theorem 4.17 Let AECnxn and Re(O(A))O{ri,i=l,..°~k} = @, where r1<r2<...<rk

and r.I ER f o r l~i~'k. Define
nxn.
+ 1 (4.43)
S i = Znd(Sign(ri-l'ri)(A))eC , l~i~k+l
A A
where r0 = -=and rk+ 1 = =. Assume that n.~0 for l~i~k+l.
i
Then
M-1AMss = Block dlag[ARI,AR2,...,AR(k+I) ] (4.44a)

106
where M s is a block modal matrix
Ms ffi [ S l , S 2 J . . . , S k + I ] (4.44b)
and ARi , l ~ i ~ k + l , a r e d e f i n e d as
(4.44c)
ARi = S+AS.
~
where S~l is the left i n v e r s e of S.% for l~i$k+l.
Directly from Theorem 4.15 and Definition 4.5. From Definition 4.5 and
Theorem 4.17, we have the following:
CoTollary 4.8 L e t ARi be d e f i n e d as i n Eq. ( 4 . 4 4 c ) . Then r i _ l < R e ( X i j ) < r i t for
all eigenvalues Aij of ARI , l~j~n i. •
Theorems 4.13 and 4.17, together with Corollary 4.9, allow us to decompose
a nonsingular A-matrix into a complete set of canonical right divisors such that
each divisor has latent roots clustering inside a certain vertical strip of the
complex plane C.
To obtain the canonical projection map from matrix sign function, we define
V = {Ind[(Sign+(A))T]} T (4.45)
which is epic. Note t h a t V c o n t a i n s n 1 i n d e p e n d e n t row v e c t o r s o f Sign+(A) and
these independent vectors are selected from t h e n(>n I) row v e c t o r s o f Sign+(A)
i n Eq. ( 4 . 3 6 c ) . Parallel t o Lemma 4 . 8 j we have=
nlxn 1 ÷
Lemma 4 . 9 There e x i s t s a n o n s i n g u l a r m a t r i x ~EC s u c h t h a t V -- r]W1 and V
107
4.
= W;q -1, where W1 iS defined in Eq. (4.35); V+ &= V * ( W * ) " 1 end W1 &- W~(W1W;)-I
are the right inverse o f V and WI, r e s p e c t i v e l y .
Proof:
From t h e d e f i n i t i o n o f V we c a n d e s c r i b e V as
V ffi VvSign + (A) (4.45)
k
T A k l k2 nl
where Uv ffi [ e n , e n , . . . , e n ].
Since from Eq.

(4.36b) we have Sign+(A) = MIWI, Eq. (4.46) becomes V =
A
UvMIW 1 ffi qWl, where n = UvM I. By applying Sylvester's inequality [61], we have
rank(n) = n I. From the definition o f right inverse, we have
* * * -1 + -1 (4.47)
V + = V*(VV*) -I = (qWl) (nWiWln) = win
Algorithm 4.1 can be utilized to find V from Sign+(A) by replacing
[Sign+(A)] T for ~ and VT for S. Corresponding to Theorem 4.15, the basic
properties of V are given b y :
A
Theorem 4.18 Define S = Ker(V); let X be the state space and A: X÷X. Then
(i) S is an A-invariant subs,ace of X.
(ii) V is the canonical projection map V= X÷X/S. •
Obviously, from Theorems 4.1 and 4.18, V defined in Eq. (4.45) can be used
as the canonical projection map for finding the canonical right divlsors of
nonsingular l-matrices.
Paralleling Theorem 4.17, we can construct the block modal matrix from
canonical projection maps.
Theorem 4 . 1 9 Let AECnxn and R e ( O ( A ) ) N { r i , i = I p . . . , k } = ~, where rl<r2<...<rk

108
and ri~ R for l~i~k. Define
V. = {Ind((Sign ~r I I,rl)(A))T ) }T ¢C nixn , l~i~k+l

(4.48)
1 ° •
where r0 = -~ and rk+ 1 = ~. Assume that ni~0 for l~i~k+l. Then
(4.49a)
MVA~I - ~lock diag[ALI,AL2,...,ALCk+I)]
where 4 1 i s a b l o c k modal m a t r i x
T T T T (4.49b)
MV = [ V I , V 2 , . . . ,Vk+ I]
ALl , l ~ i ~ k + l , a r e d e f i n e d as f o l l o w i n g
A L i = Vi AV~ (4.49e)
V~ is the right inverse of V.. Also, the real parts of a l l the e l g e n v a l u e s o£

1 I
ALi are in (ri_l,ri).
Proof~
Directly from Theorem 4.18 and Definition 4.5. •
Since matrix sign functions can be used t o compute the c a n o n i c a l injection
and canonical projection maps, the block triangularization transformation can be
easily obtained from matrix sign functions of A by Lemma 4.7.
Theorem 4.20 Let AEC nxn and Re(~(A))fl{rl,r2} ffi ~, where rl<r 2 and tiER for
i=I,2. Define
109
+ nXn 1 (4.50a)
S(rl,r2) = Ind(Sign(rl,r2)(A))EC
and
= [Ind((Sign~rl,r2)CA))T)lTcCnxn2 (4.50b)
V ( r l , r 2)
Assume t h a t n I and n 2 are p o s i t i v e . Let
T~ - (4.51)
V ( r l , r 2)
+
where S ( r l , r 2 ) is the l e f t i n v e r s e of S ( r l , r 2 ). Then
(4.52a)
0n2Xn I AL
where
AR ~ S+ (4.52b)
(rl,r2)AS(rltr2)
~L"A~C~, r2>A~r l, r2> <4.52=~
A
and V+( r l , r 2 ) is the r i g h t i n v e r s e of V ( r l , r 2 ). Furthermore, the r e a l p a r t s of

the eigenvalues of AR are inside ( r l , r 2) and Chose of AL are o u t s i d e ( r l , r 2 ) .
Proof:
From Theorem 4.15, S ( r l , r 2 ) is the canonical i n j e c t i o n map, S ( r l , r 2 ) : S÷X,
where S ~ Im(S(rl,r2)) and X is the s t a t e space such t h a t A: X÷X. From the
d e f i n i t i o n of V ( r l , r 2 ) and Lennna 4.9, we have

110
- . +
V(rl,r2 ) = V~(In-Slgn(rl,r2)(A))
where
E1 k2 n2]T
U~ = [e n , e n , . . . , e n
Also from Le---a 4.8, we have
S(rl,r2 ) = Sign[rl,r2)(A)U s
where
k
k I k2
U S s [en ' e n , ' ' * ' e n hI]
From Eq. (4.36a) s it is easy to show that Sign(rl,r2)(A ) is idempotent. Thus,

we obtain
s - • - • - -
rl,r2) (rl,r2) In-S1gn(rl,r2)(A))S1gn(rl,r2)(A)Us 0
since V(rl,r2 ) end S(rl,r2 ) are of full rank. This implles that KerC~Crl ' r2 ) =
-
I m ( S ( r l , r 2 ) = S or V ( r l , r 2 ) i s a canonical projection map. By Lem~a 4 . 7 , the

rest of the r e s u l t s of Theorem 4.20 f o l l o w . •
To compute Slgu(A) or Sign+(A), Roberts [56] proposed a reeursive
algorithm, which is referred to as the m a t r i x sign algorithm [56] and can be
described as f o l l o w s :
Sign(i+l)(A) = {[Sign(i)cx)+(Sign(i)(A))-l]; Sign(0)(A) = A for £>n(4"53a)
and
Sign(A) = i~£m (4.53b)

÷= Sign(1)(A)
The a l g o r l t h m i n Eq. (4.53a) provides quadratic convergence a t the n e i g h b o r h o o d

111
of Sign(A). Recently, a new algorithm to compute the matrix sign functions has
been d e v i s e d [ 6 0 ] . This new a l g o r i t h m f o r Sign(A) i s as f o l l o w s :
Theorem 4.21 The m a t r i x sign function of A can be r e p r e s e n t e d by a m a t r i x
continued f r a c t i o n :
Sign(A) - A[In+(A2-1n)[21n+(A2-1n)[21n+(A2-1n)[...]-I]-I]'I] -I (4.54a)
where AcC nxn and Re[~(A)] ~ 0.
The approximate m a t r i x s i g n f u n c t i o n becomes
Sign(A) ~ [(In+A)J-(In-A)J][(lu*A)J÷(In-A)J] -l ~ Sign(j)(A) (4.54b)
where j is a finite integer.
The recursive algorithm for computing the matrix sign function of A with
Re[O(A)] ~ 0 is
(4.55)
Slgn(nk+l)(A) - SigU(fk) [Sign(nk) (A)]
where $ i g n ( 1 ) ( A ) ~ A; nk+ 1 = f k ' n k and fk>l f o r k~0; n0=l. •
The degree of the convergence rate of the algorithm in Eq. (4.55) at the
neighborhood of Sign(A) is fk if fk is constant for k~0. The reader is referred
to [60] for further details of the new matrix sign algorithm in Eq. (4.55).
Theorems 4.15-4.21 present a set of computational algorithms for block
dlsgonallzation and block triangularizatlon of a square matrix A such that each
block matrix along the diagonal has eigenvalues clustering inside a certain
vertical strip of the complex plane C. These algorithms are suitable for
finding the complete sets of left/right divisors and spectral factors of

112
nonsingular k-matrices occurred in the continuous-time systems (see Chapter VI).
For discrete-time system problems we often need to block-dlagonalize or block-
triangularize a system map A such that each block matrix along the diagonal has
elgenvalues clustering inside a certain annulus centered at origin of the
complex plane C. Therefore, to use Theorems 4.15-4.21 for discrete-tlme
systems, the argument matrix A in the matrix sign functions9 Sign(At, Sign+(A)
and Sign-(A), should be replaced by (A-In)(A+In)-l; Sign(rl,r2)(A),

• ÷ ° --
Sxgn(rl,r2)(A) and Stgn(rl,r2)(A) in Definition 4.5 should be redefined

accordingly:
@
Sign(rl,r2)(A) = 2 Sign(rl,r2)(A)-I n
• - (4.56a)
= In-2 Sxgn(rlpr2)(A)
where
• + 1 (4.56b)
S1gn(rl,r2)(A) ~ ~[Sign(rl)(A)-Sign(r2)(A)]
S1gn(rl,r2)(A ) ~ in_Slgn(rl,r2)(A
. -
• * ) (4.56c)
Sign(ri)(A) ~ Sign((A-In)(A+In)-l-riIn ), for i=1,2 (4.56d)
With the definitions in Eq. (4.56), Sign(rl,r2)(A) is called the generalized
matrix sign function of A with respect to the open annulus centered at origin of
C and bounded with radii (l+rl)l(l-r I) and (l+r2)/(l-r2).

Two examples are illustrated in this section. The first example
demonstrates the computations of the complete sets of canonical left/right
divisors and the spectral factors for a nonslngular k-matrix via the Jordan form
decompositions. The second example illustrates the computations of the complete
sets of canonical left/right divisors and spectral factors for a nonsingular k-
matrix via the matrix sign algorithms.

113
Example 4.1
Consider a reachable pair (A,B) of a contlnuous-tlme system:
-2 2 -5 4
A = 2 -5 I0 -23
I -2 4 -4
0 0 0 2
and 1 3 2
0 -2 -2
B= 0 1 1
2 5 3
1 2 1
The r i g h t characteristic X-matrix is found t o be
X2iX-2 -18X-18 11
Dr(X) = X3+3X2+3X+l -1
0 1
which is a column-reduced canonical X-matrix. The Jordan form minimal
realization quadruple of D-I(x) can be found as

r
-1 1 0 0
Aj = 0 -I 0 0
0 0 -I 0
0 0 0 2
0 1
Bj = 0 1
II
12
Cj -- -I 0 -1
0 0 0
Dj = [ioo] 0
0
0
1
Let Aj and Bj be partitioned as

114
Aj " Bj =
0 A2 B2
where
[
A1 = 00~ I0° -I
] ,~ =
[:o]
2
BI = 0 i 1 B2 =
0 1 1 2 1
The right characteristic l-matrices o£ (AI,BI) and (A2,B 2) are
1+i -I 1 ]
Drl(A) = 0 12+21+1 -I
0 0 1
and
I+4 6 I ]
Dr2(1) = L-3 1-50 -Ii
respectively. From Theorem 4.13, {Drl(1),Dr2(k) } i s a complete set of canonical

left divisors of Dr(1). Similarly, l e t Cj be p a r t i t i o n e d as
Cj = [CI,C2]
where
CI =
[6 513]
I -I 0 C2 =
[i,1311]
-i 0
0 0 0 0 0
The left characteristic k-matrlces of (AI,C I) and (A2,C 2) are
12-41+13 108 O ]
D.l(k) = 25/3 k+70
0 0 1
and
115
D~(~) = [ 0 -17
`1-2
0
,t+1 ~1
0
respectively. From Theorem 4.13, {DRI(k),D~2(Z) } is a complete set of canonical

right divisors of Dr(k).
Let Aj, Bj, Cj be partitioned as
Aj = , Bj- , C~- [~X,~21

o ~2 J ~2
where
~I -~oo] [ o oI X2 = o
0
-1
0
o
2
~12"
[ooo]
1 0 0
[12:] [011]
~1 = ~2 = 1 x 1
0 1 I 2 1
E1 =
[i ~] -i
0
~2 = -1
0
o
0
The right c h a r a c t e r i s t i c `1-matrix of (A2,B 2) is
I A2-X-2 -2X-2 i 1
Dr2(A) = 0 ~+I -i
0 0 1
and the l e f t c h a r a c t e r i s t i c k-matrix of (A1,C 1) is
D~I(A) =
I ~-2
0.5
0
-18
~+4
0
01
0
1
Therefore from Theorem 4.12, we have

116
Dr(),) - D r 2 C ) , ) U r t ( ) , ) D t l ( ) , )
w h e r e U r t ( ) , ) i s u n l m o d u l a r and c a n b e f o u n d as
Urt(),) - -0 5
20]
),-2 0
o 1
Vr(),) c a n a l s o b e f a c t o r e d as
Dr(),) ffi Dr2(),)DEI(),)
where
1 2),+8 01
V£1(),) -. Ur£C),)Dtl(),) - 0 ),2+2),+1 0
0 0 I
which is not a canonical ),-matrix
Example 4 . 2
Consider a reachable pair (A,B) o f a c o n ~ i n u o u s - t l m e system:
-l.O -6.0 11.0 -32.0 63.0 1
I
-1.0 -2.0 3.0 -11.0 24.0
A " 2.0 -4.0 10.0 -18.0 37.0
!.0 -2.0 4.0 - 7.0 12.0
0.O 0.0 -0.5 1.0 -3.0
and
-2.0
1.0 1.0
-2.0 0.01
-3.0
B= 1.0 1.0 3.0
2.0 1.0 4.0
1.0 0.0 1.0
The r i g h t characteristic ) ` - m a t r i x i s f o u n d t o be
1
),3+3.6667),2-0.6111),-4.0556 ),2-0. 3333),-0. 6667 -1.0
Dr(),) ,-
I -0.5556),-9.1667
0.0
),2-0 . 6667),-2 . 0000
0.0
-2.0
1.0
117
From Lemma 2 . 1 , a minimal reallzat~on q u a d r u p l e o f D ; I ( A ) i s ( A , B , C r , D r ) , where
0.29333 -0.64000 1.38667 -2.96000 5.76000

2.06667 -4.42222 9.60000 -18.33333
Cr = / - 1 " 0 4 4 4 4
L o.ooooo 0.00000 0.00000 0.00000 0.00000 J
0.0 0.0 0.0 1
Dr = 0.0 0.0 0.0
0.0 0.0 1.0
Selectlng r0 - -co~ r I " - 3 t r 2 = 0, and r 3 = % we have t h e followlng shifted
matrix sign functions:
Sign(ro)(A) = 15
1.0 ~.0 8.0 -20.0 40.0"

0.0 -1.0 4.0 -10.0 20.0
Sign(r1 ) (A) - 0.0 0.0 1.0 0.0 0.0
0.0 0.0 0.0 1.0 0.0
0.0 0.0 0.0 0.0 1.0
1.0 -4.0 10.0 -24.0 48.0]

0.0 -1.0 0.0 -20.0 4.0|
0.0 0.0 -1.0 4.0 -8.0
SlgnCr2)(A) " I 0.0
0.0
0.0
0.0
0.0
0.0
1.0
0.0
-4.0
-1.0 J
Sign(r3)(A) = -I 5
From Theorem 4 . 1 9 , we have
V 1 = [0.0 2.0 -4.0 10.0 -20.0]
V2 = [ 0 . 0 0.0 -I.0 2.0 -4.0 1

0.0 0.0 0.0 0.0 2.0
Fl.O -2.0 5.0 -12.o 24.0]

£3 = L 0 . 0 0.0 0.0 - 1.0 2.0J
118
O' ne V2
V3
we have the b l o c k d l a g o n a l l z e d minimal r e a l i z a t i o n quadruple of D-I(A):

r
AD = 1 = Block dleg I -4.0), (oo 5) (Io o I]

1.0 -1.0 ' -1.0 3.0
-2.0 2.0 2.0

-~.~ 1.o 1.o
BD = MVB - 1.0 0.0 2.0
-Lo- =2.0- : 3 - 6 -
0.0 -1.0 -2.0
1 r - 0 " 0 2 6 6 7 I 0.18667 -0.08003 ! 0.29333 -0.45333]

cv - C r ~ - -0.01111 ~-o.75556 0.13333 I-I.O4444 1.31111
L 0.00000 a 0.00000 0.00000 ' 0.00000 0.00000 J
DD = D r
From Theorem 4 . 1 3 , we have a complete set of c a n o n i c a l left divisors {D~I(X) ,
DE2(A), D~3CA)}:
X+4.0 1.0 1.01

D~I(A) =
I 0.0 1.0 0.O
0.0 0.0 1.0
I X+1.5 -0.5 -1.01

D~2(X) = 2.5 X-0.5 -2.0
0.0 0.0 1.0
and
X-1.0 0.0 -1.0 1
D£3(X) = -1.0 X-I.0 -2.0 ;
0.0 0.0 1.0
and a complete s e t of c a n o n l c a l r i g h t d i v i s o r s {Drl(~) , Dr2(~), Dr3(~)}:

119
k+4.00000 0.0 0.01

Drl(k) - 0.41667 1.0 0.0
0.00000 0.0 1.0
I" ),-0. 5839b, -0.25000 0.0 1

Dr2(X)- [ 5.69761 h+1.58394 0.0
L o.o 0.0 1.0
and
r x+o. 16667 0.20000 0.0 "1
Dr3(),) = | -6.80556 X-2.16667 0.0
L o.o o.o 1.o J
To p e r f o r a the s p e c t r a l f a c t o r i z a t l o n ~ u s l n 8 Theorem 4 . 2 0 , we have
[L S~2'r3) J1 !--0"--0~0-00
[" 1"°°°°° -2"°°°°°
-0. lee67 4"°°°°° 2"°°°°°
0.33333 o.le6e7 0.00000t
0"00000
T " -- / 0 . 0 0 0 0 0 2.00000 -5.00000 12.00000 -2,,.00000
/~(~2,~3)/ /o.ooooo l.ooooo o.ooooo 1.ooooo-2.00000
LO.OOOOO o.ooooo o.ooooo o.ooooo 2.00000
and a minimal ¢ e a l l , a t l o n q u a d r u p l e (~f,BT,CT,D T) of D;Z(X) with ~ being block-
trlangularized:
FII'O -100.0 t - 8 . 4 6 6 0 0.2652 -7.8314 1
% - TA~ - 1 -
03x 2 AL
- - 770
0.0
0.0
0.0 ~--g.~(Og
0.0 I-1.6000
0.0 ~ 0.2000
-l.6~
-0.8000
0.4000
O.50OO
1.0000
-1.0000
]
B T - TB -
[ ] ,o.ooo .
F 4,0000 11.0000 26.0000 1
0.83333 2.1667 l
i --_5 .-6o-b6- - 3 : o ~ o ~ - ~ 7o~o-o-]
BL | 1. 0000 -1.0000 -1.0000 |
L 2.00O0 0.00000 2.0000 J
I 2.9333 - 3 . 0 6 6 7 ] -0.2471 0.0703 -0.1555

CT - C~-1 - [CR, CL] - -1.0444 11.22221 0.7970 0.2542 0.3517
0.0000 0.0O00 ! 0.0000 0.0000 0.0000 J
and
DT = D
From T h e o r e l 4 . 1 2 , Dr(~) can be f a c t o r e d as

120
Dr(X) = Dr2 ( k)Ur£( X)D¢I(X)
where Dr2(X) i s ehe r i g h t c h a r a c t e r i s t i c X - m a t r l x of (AL,BL)=
I
~2+3.5X+4.25 0.8X+0.7 -1.01
Dr2(X) = 6.25 X+l.5 -2.0
0.00 1.0 1.0
D¢I(X) is the l e f t characteristic X-matrix of (AR, CR):
~+0.1667 0.2000 0.0 ]

D£1(k) = -6.8056 X-2.1667 0.0
0.0000 0.0000 1.0
and UrE(A) i s a u n i m o d u l a r X - m a t r l x :
UrE(X) = z 3
To factor Dr2(k), we find a =inlmal realization quadruple (AL,BL,CL,DL) of
[-3.20 -1.60 0.50]

,',L= / 1 6 0 -O.SO 100 /
L 0.20 0.40 -t.ooJ
r -3. O0 3.00 3 •00 ]

BL = | 1.O0 -I.00 -I.00
L 2.00 0.00 2.00 J
0.04 0.12 0.00~
~L = 0.30 -O.lO 0.50]
o.oo o.oo o.ooj
=D
L r
S e l e c t i n g r 0 = -m, r 1 = -3 and r 2 = ~, we have the f o l l o w i n g s h i f t e d m a t r i x s i g n

121
functions
Sign(ro)(A L) " 13
-0.6 -0.8 0.0 ]

9ign(rl)(A,)~ - -0.8 0.6 0.0
0.0 0.0 1.0
Slgncr2)CAL) " -Z 3
Again, from Theorem 4.20, we have

÷
TL -
[...i,..,] El.O _..o o.oj
- o.o o.o l.O
VL(rl,r2) 0-.8 - "0~4" -0"~0
and a minimal realization quadruple (ALT,BLT,CLT,DLT) of D:~(X) with ALT being

block-triangularized:
ALT = TLATLIffi IA~0R ALRLI -

[ 00 ..20 - 0 -. I2.50 I 0.0
~
0.0]
ALL ~ _ C - - o . o - ~:~.-o
BLT " TLB = " 2.0 0.0 2.0

BLL ~io 2.0 f.~
CLT = CTLI . [CLR CLL] . [0.04 0.0 il]0.25

0.I0 0.50j 1.0
0.00 0.000 0.00
DLT " Dr
Now, from Theorem 4.12 we have

122
~r2(k) = ~r2(k)Or~(k)~1(k)
where ~r2(k) is the right characteristic h-matrix of (ALL,BLL):
k+4.00 1.00 1.00]

~r2(k) = o.oo
0.00
~.oo
0.00 1.O0
o.oo
~l(k) is the left characteristic h-matrix of (ALR,CLR):
k-o.so -0.20 o.oo- I

~.i(k)- 6.25 k+1.5o 0.00|
0.00 0.00 O.OOJ
and Urn(k)£s a unimodular k-matrix~

Urn(k) -
1.0 0.0 0.0 ]
o.o 1.o -2.0
0.0 0.0 1.0
Combining these two parts of spectral factorization, we have
Dr(h) =
[I+4.oo 1.oo 1.oo][1.o
0.00 1.00 0.00| 0.0
o.o
1.0
o.o]
-2.0
0.0o 0.o0 1.00J 0.0 0.0 1.0
k-0.50 -0.20 0.001 rk+0.1667 0.2000 0.0 1
I 6.25
o.oo
1+1.50
o.oo
0.00| |-6.8056
1.ooj L o.ooo
I-2.1667
o.oooo
0.0
1.o J
CRAPTER V FEEDBACK CONTROL OF MULTIVAEIABLE SYSTEMS
Feedback design for multlvarlable systems has attracted many researchers
[5-16,62-78] in the fields of system theory and control engineering. Probably
the most successful design methods are state-feedback control in the time domain
[5,63-71], and the inverse Nyquist array and the root loci design in the
frequency domain [8,72-78]. In this chapter, we shall discuss the feedback
control of multivariable systems employing the notions of the characteristic k-
matrices and their divisors for constructing the state-feedback control laws.
These new approaches permit a deeper insight into some structural aspects of
feedback control. In g e n e r a l , the design schemes presented in this chapter
provide methods f o r c o n s t r u c t i n g s t a t e feedback c o n t r o l l e r s by d e a l i n g with the
assignments of the denominator of the c l o s e d - l o o p m a t r i x fraction description.
Therefore, advantages of both the time domain and the frequency domain d e s i g n
can be achieved via these design methods.
Some properties of linear state-feedback controls are set out in Section
5.1. In Section 5.2j methods using characteristic and non-characteristic k-
matrix assignments to construct the state-feedback controller are developed.
Section 5.3 is devoted to deriving the latent structure assignments which take
the left/right generalized latent vectors of the closed-loop right
characteristic k-matrices as the design guideline. The divisor assignment and
decoupling desigu approaches are discussed in Section 5.4.
5.1 Linear State-Feedback Controls and Properties
Consider an m-input, p-output linear time-invariant system described by
~x(t) = AX(t)+BuCt)
(5.1a)
y(t) = CX(t)+Du(t)
(5.1b)
124
where X(t)~C n, u ( t ) ¢ C m, and y(t)cC p are state, input, and output vectors,
respectively; Aj B, C, and D a r e matrices of appropriate dimensions; ~ is an
operator, and t is an independent variable. For continuous time systems, ~ is a
differential operator with tER. For discrete time systems, k is a forward shift
operator with tEZ. Assume that the system in Eq. (5.1) is reachable. The
linear state-feedback control law c a n b e d e s c r i b e d as~
(5.2)
u(t) = -FX(t)+GFr(t)
where F, GF, and r(t) are the feedback gain, input gain, and reference input~
respectively. The i m p o r t a n t invariance property of t h e K r o n e c k e r indices under
a state-feedback control i s as f o l l o w s .
Lense 5.1 Let the state-feedback control law be Eq. (5.2), and let Drc(~) be
the right characteristic ~-matrix of the closed-loop system. Thenp ~ci[Drc(~)]
= aci[Dr(~) ] = ~i,iffil,...,m, where Dr(~) is the right characteristic ~-matrlx of
the open-loop system, Ki,i=l,...,m are the Kronecker indices, and ~ci[Drc(k)]
denotes the ith column degree of Drc(A).
Proofz
R e f e r t o Popov [ 3 5 ] . •
Le-~a 5.1 gives the structure of all possible right characteristic X-
matrices of the c l o s e d - l o o p s y s t e m s with l i n e a r state-feedback control laws. In
other words, we may o n l y assign the closed-loop right characteristic X-matrix
having the same Kronecker indices as the open-loop right characteristic X-
matrix. However, this structural limitation does not restrict the freedom of
choosing the closed-loop elgenvalues. Thus, the dynamics of the closed-loop
system still can be controlled by a s s i g n i n g appropriate eigenvalues or poles.
From Lemma 5 . 1 , we o b t a i n the results below.

125
Lemma 5.2 Let (A,B,C,D) be a m i n i m a l r e a l l z a t l o n of D : I ( x ) , where Dr(X) i s the
right characteristic X-matrix of the r e a c h a b l e pair (A,B). Lee Drc(X) be the
desired right characteristic X-matrlx of the c l o s e d - l o o p s y s t e m under the s t a t e -
feedback control law ~n Eq. (5.2). Then, (A-BF,BCF,C,DCF) is a minlmal
realization -1
of Drc(~) , where OF = Dr hD-~
r c ,D rug
~_ i s the l e a d i n g - c o l u m n c o e f f i c i e n t
m a t r i x of Drc(X), and Drh i s t h a t of Dr(X).
Proof:
From Lerma 2 . 1 , we have
Drl(X) = ~(X~n-A)-IB+~ - ¢_T(O)(X~ -A )-lS c +~ (5.3)

r n c
where C - ~rT(0)Tc ; D = ( I m - ~ T ( 0 ) ~ r ( 0 ) ) D : h l ; Ac - TcAT:I; Bc - TcB. (Ac,B c) i s

the controller canonical form of (A,B). Let
Fc
FT:I " DrhF c
- (5.4)
we o b t a i n
Tc ( A - B F ) T c- I = Ac-BcFc = A c - E b c F c ~ c
(5.5a)
(5.5b)
TcBGF = BcDrhD;Ic = EbcD:hlc A B c
where Ac and Bc a r e the c o n t r o l l e r canonical forms of the c l o s e d - l o o p system.
From Lem~a 5 . 1 , the g r o n e c k e r indices of (A-BF,BG F) are the same as those of
(A,B), thus we o b t a i n the minimal realization of the inverse of the right
characteristic X-matrix of (A-BF,BGF) as

126
(5.6)
D;~CX) " ~T(0lCklr u -~ c )-Igc +~ c = C(XIn-CA-BFI)-IBGF+DGF
where Dc = (Im-*~(O)*r(O))D;~c " DGF" Thus, (A-BF,BOF,C,DG F) is the minimal
realization of D~(~). •
Lemma 5.2 reveals the state space structure of the closed-loop right
characteristics A-matrix under linear state-feedback controls. From Lemma 5.2,
we have the following results on the structure of the feedback gain when some of
the Kronecker indices are zero.
Corollary 5.1 Let K ~ {~l,...,~m} be the Kronecker indices of (A,B). If Ki=0
for some l~i~m, then the ith rows of Fc defined in Eq. (5.4), can be chosen to
be null, and therefore the ith rows of F and F are also null.
e
Proof:
From Eq. (5.5a), we have Ac = Ac-EbcFc" If <.=0,~ the ith column of Ebc is
null. Thus, the ith row of EbcFc

- is null, and we can choose C~c) i = 0. Since F
= FcTe = DrhFcTc and (Drh)ij = 0 for j~i and ~.=0,

1
the ith rows of F C = Drh ~ C
and F
C
are both null. •
In the following sections of this chapter, we d i s c u s s various schemes f o r
constructing linear state-feedback controls laws for different design purposes.
5.2 ~-matrix Assisnment
In this section, we d i s c u s s methods to construct the linear state-feedback
control laws from t h e d e s i r e d denominator of the closed-loop RMFD. One s i m p l e
way to assign the denominator of the closed-loop RMFD i s to select a column-
reduced canonical A-matrix which has the same K r o n e c k e r i n d i c e s as those of the
open loop right characteristic A-matrix. This approach is referred to as the
right characteristic h-matrix assignment. For this scheme, the numerator ~-
matrix of the closed-loop canonical EMFD is to be the same as that of the open
127
loop c a n o n i c a l RHFD. A s e c o n d way t o approach the problem is to assign the
denominator of the closed-loop RHFD a s a non-canonlcal column-reduced A-matrix
h a v i n g t h e same K r o n e c k e r i n d i c e s a s t h o s e o f t h e o p e n l o o p r i g h t c h a r a c t e r i s t i c A-
matrix, This approach is referred to as the column-reduced A - m a t r i x a s s i g n m e n t .
The important feature of this scheme is that the numerator A-matrix of the
c l o s e d - l o o p c a n o n i c a l RMFD may n o t be t h e same a s t h a t of the open-loop canonical
R~D. This result is different from t h e s i t u a t i o n i n SISO s y s t e m s u s i n g l i n e a r
state-feedback controls.
Theorem 5 . 1 (Right C h a r a c t e r i s t i c A - m a t r i x Assignment)
If the desired closed-loop right characteristic A-matrlx is given by
Drc(A) , and Drh c is the leading-column matrix of Drc(A) ~ then the feedback gain
and input gain become
F = F T (5.7a)
e c
-I (5.7b)
GF = DrhDrh c
where Fc can be d e t e r m i n e d by
(5.7c)
FcSr(A) - GFDrc(A)-Dr(A)
Proof:
From Lemma 5.2~ we o b t a i n
BcDrc(A) = ( A I n - A e ) ~ r ( A ) = ( A I n - A e ÷ E b c F c ) ~ r ( A )
= ( X l n - A c ) ~ r ( A ) + E b c F c ~ r ( A ) = BcDr(A)+EhcFc~r(A)
128
Therefore,
Bc(GFDrc(A)-Dr(k)) = EbcFc~r(k)
Since B c = EbcD r and the ith row of Fc is zero if K'=0'I we have
D:~(GFDrc(k)-Dr(A)) - FcSr(k).
Thus, the result of Eq. (5.7c) follows.
The closed-loop canonical RMFD [79,80] can be found as follows.
Corollary 5.2 Let the desired closed-loop right characteristic k-matrix be
D (X) and F and G F be determined in Theorem 5.1. Then, the closed-loop

rc
canonical RHFD with D = 0 p ~ is given by
Gc(~) - C(nn-(A-BF))-IBUF = Nr(k)Drc1(~) (S.8)
Proof:
Gc(A) = C(AIn-(A-BF))-IBGF = Cc(AIn-Xc)-IBc
= CcSr(X)D;cI(A) = N r(k)D:~(k) •
If arbitrary F and GF are assigned in Eq. (5.2), the closed-loop RRFD,
which may not be the canonical RMFD, can be described as follows.
Theorem 5.2 Let D = 0 and the state-feedback control law be shown i n Eq.
pxm
(5.2) with the ith row of F being null if K.-0 and G F being nonsfngular. Then,
I
the RMFD of the closed-loop system can be expressed as
129
Oc(k) = C( ~n-(A-BF))-IB0 F = Nr( ~);;cl(X) C5.9a)
where ;rc()0 is a column-reduced ~k-matrlx defined by
(5.9b)
; r c (~) = G ; I [ F T : I ~ ( ~ ) ' D r (~)]
Proof:
Gc()t) = C( ~I -(A-BF))-IBGF = Cc(~In-Ac)-iBc
where
X c = Ac-BcFT:I and Bc = TcBGF = EbcDr @F
For the a r b i t r a r y matrices F and GF s p e c i f i e d in Theorem 5 . 2 , Ac and Bc are in
controller forms but not necessary in canonical controller forms. Therefore,

from Len~na 3 . 2 , we have
;:Ioc - , C0 C nZc I o, c
where Dc = (Im-~T(0)~r(0))D;1GF * I t can be easily showu that
BcDrcCX~ = C~In-Ac)~rCX)
whence
A
BcGFDrc(A) = (kln-Ac+BcFT:l)$r(t) = (kln-Ac)~r(A)+BcFT:1~r(k)
= BcDr(~)+BcFT:I~r(~)
Since the i t h row of F i s n u l l i f ~i=O, we have the r e s u l t in Eq. ( 5 . 9 b ) , from
which
130
"thus, the closed-loop RMFD becomes
G (cAr - co( n - "
Remark
Theorem 5.2 reveals the fact that the closed-loop RHFD can be written as
Nr(A);:~(A) , where the EMFD may not be canonical even if Nr(l) is the numerator
of the open-loop canonical RMFD and Drc(A) is column-reduced A-matrix determined
by Dr(A) , F and G F. If the canonical RMFD of the closed-loop system is found to
be Gc(A) = Nrc(A)D:~(A), we can easily prove that NrcCk) = Nr(A)U(A) and
Drc(A) = Drc(A)U(A) where U(A) is a unimodular matrix.
From Theorems 5.1 and 5.2p we observe that the non-canonical column-reduced
A-matrix assignment can be accomplished as follows=
Theorem 5.3 (Column-reduced A-matrix Assignment)
Let Drc(A) be a given column-reduced A-matrix with column degrees
(KIj...jKm) , Which is the set of Kronecker indices of the open loop system in
^ ^
Eq. (5.1). Drh c is the leading-column matrix of Drc(A). We choose the
feedback gain and input gain to be
(5.10a)
F=FT
C C
A
(5.lOb)
G F = DrhDrhlc
where F can be determined by

C
Fc~r(A) = GFDrc(A)-Dr(A) (5. lOc)

131
Then, the c l o s e d - l o o p RHFD w i t h D = 0 pxm i s g i v e n by
Gc(~) = C(~In-(A-BF ))-IBGF . N r

(5.11)
Proof:
This follows directly from Theorems 5.1 and 5.2.
Example 5 . 1
Given a 3-input 2-output c o n t i n u o u s t i m e system
Ax(t) = Ax(t)+Su(t)
y(t) = CX(t)+Du(t)
where
2.2 1.6 4.0 3.0 1.0 1

-6.4 -4.2 -8.0 -6.0 -2.0
A = -1.6 -0.8 -3.0 -2.0 -1.0
3.2 1.6 4.0 3.0 2.0
9.6 4.8 12.0 6.0 2.0
I
-2.0
-I.0 1.0
-2.0 0.0 1
-5.0
B = 3.0 -I.0 1.0
-2.0 3.0 4.0
1.0 -2.0 -3.0
C = [3.65.85.04.0 0.0]
5.42.25.03.0-1.0
D = 02x 3
Using the minimal n i c e s e l e c t i o n a l g o r i t h m in S e c t i o n 2 . 3 , the Kronecker i n d i c e s
are found to be E 1 -- 3, <2 = 2 and ~3 " 0. Note that the input matrix B
contains a dependent vector, but the system is reachable. From Eq. (2.16)~ we
132
have the canonical RF~D
O(k) = Nr(k)D;l(k)
where
-4k2+13k-13 7k-12 0 ]
SrCk) =
-5kX+10k+3 12k+7 0
and k3-2k2-k+2 -k2+k+2 -1 ]

VrCk) = 0 X2+2/+1 -2
0 0 1
The open loop poles are -i,-I,-I,I,2. We shall construct linear state-feedback
control laws using both the right characteristic k-matrix assignment and the
column-reduced k-matrix assignment as follows:
(I) Right characteristic k-matrix assignment
Let the desired closed-loop right characteristic ~-matrix be
Drc(~) " [ Z3+6Z2+22~+30

-10k-22
0
-Z2-7Z-12
k2+7~+12
0
0
0
I
which is a column-reduced canonical A-matrix with the leading-column c o e f f i c i e n t

matrix
I -I 0]
Drh c = 0 I 0
0 0 1
From Theorem 5.1, we have the feedback gain matrix
-12.1 -12.3 -25.5833 -26.4167 -20.58331

F ffi- -25.6 - 6.8 -22.0000 - 1.0000 6.0000
0.0 0.0 0.0000 0.0000 0.0000
133
and the input gain matrix
I.0 0.0 -I.0]

GF = 0.0 1.0 -2.0
0.0 0.0 1.0
The linear state-feedback control law is 8iven by
The closed-loop canonical RMFD can be found as
Gc(~.) = Nr(A)D:~(}O
The closed loop poles are -2~-2,-2,-3,-4.

(2) Column-reduced k-matrix assignment
Let the desired denominator of the closed loop RMFDbe
A3+4k2+8A+6 -k2-7~-12 0 1
rc(),) = 2~2+4~+2 k2+7k+12 0
0 0 1
The leading-column c o e f f i c i e n t matrix is
Drhc 0
which is nonsingular. Thus, Drc(k) is column-reduced with column degrees
~cl(Drc(~)) 1 3, ~c2(Drc (k)) = 2 and ~c3(Drc(X)) = Ol ~ut
- ^ [ X3+6X2+I2X+8 0 0 ]
Dr~cDrc(k) ffi 2k2+4A+2 k2+7A+12 0
0 0 1
134
and from Definition 3.1, ;rc(X) is not a column-reduced canonical X - m a t r i x .

From Theorem 5.3, we have the feedback gain
-10.1 - 8.3 -17.5833 -14.4167 -10.5833 1

F =' - -27.6 -10.8 -30.0000 -13.0000 - 4.0000
0.0 - 0.0 0.0000 0.0000 0.0000
and t h e i n p u t g a i n
I I.0 0.0 -I.0 1

GF = 0.0 1.0 - 2 . 0
0.0 0.0 1.0
The linear state-feedback control law is
.(t) = -FX(t)+GFr(t)
The closed-loop RMFD is given by
Since the denominator o f ¢(X) is not a column-reduced canonical X-matrix, Gc(~)

is not a canonical RRFD. From T h e o r e m 3 . 4 j we h a v e the canonical RMFD o f the
closed-loop system:
ccCx) = .rcCx)D~<x)
where
-4X2-X+ll
7X-12 0 1
Nrc(X) = -SA2-14X-11 12~+7 0
It is easy to check that

135
Nrc(k) = Nr(k)Ur(k)
Drc(~) - Vrc(XIUr(A1
where Ur(X I i s a u n i m o d u l a r k - m a t r i x
Ur(~) - l i0 0]I
0
0
1
The d e n o m i n a t o r of the canonical RHFD o f the closed-loop system in this design
example i s t h e same as t h a t in (11; however, the numerator is different (compare
t h e n u m e r a t o r o f t h e open l o o p c a n o n i c a l RHFD w i t h N r c ( ~ ) ) .
5.3 Left and Right Latent Structure Assignment
In this section, we present an alternative method for constructing the
linear state-feedback control law from t h e d e s i r e d left or right latent structure
of the closed-loop right characteristic A-matrix. This approach is the
counterpart of elgenstructure assignment f o r MIMO s y s t e m d e s i g n , which have been
discussed by many a u t h o r s [64,66]. An a d v a n t a g e of determining the feedback
control law u s i n g t h e l a t e n t structure of the closed-loop right characteristic A-
matrix instead of the eigenstructure of the closed-loop s y s t e m map, is to avoid
dealing with the higher dimensional eigenvectors o f t h e s y s t e m map.
Using the relationships between the latent structures of the right
characteristic ~-matrix of a reachable pair and the elgenstructures of the
s y s t e m map e x p l o r e d in Section 3.2, we d e r i v e the state-feedback control law v i a
latent roots and l a t e n t vectors a s s l g n m e n t as f o l l o w s .
Theorem 5 . 4 (Left Latent Structure Assignment)
A linear state-feedback controller i n Eq. (5.21 is required for controlling
the m - i n p u t p-output reachable s y s t e m i n Eq. (5.11 w i t h D = 0 p×m s u c h that the

136
closed-loop r i g h t c h a r a c t e r i s t i c l - m a t r l x has the assigned l a t e n t r o o t s ,
{[i,i=l,...,k}, and l e f t g e n e r a l i z e d l a t e n t v e c t o r s , { P i j , j = 0 , 1 , . . . ,
~.-1,i=l, .,k} where l i is the length of Jordan chain correspondlng to ~i and
kI °*
i~ l £i=n. Define
Aj ~ block diag{~i,i=lt...~k}cCnxn (5.11a)
where
i ~ .xZ. BJ1
1 [-]
iT.-1
.xm
1 Xl 1 e ~ ~; ~j ~ e ~m;- £
L
1 Bj i
pT
i Bjk io
(5.rib)
Assume that (AjmBj) is a reachable pair and the Kronecker indices of (Aj,Bj) are
equal to those of (A,B). Then, the desired feedback galu F and the input gain
GF can be determined by F - FcT c and G F = DrhDr c' where Tc is the

transformation matrix which transforms A into the canonical controller form A C
TcAT:I , and F c is computed from Fear(k) = GFDre(A)-Dr(A). Drc(k) is the right
characteristic A-matrix of (Aj,Bj) with the leadinE-column matrix Drh c and Dr(k)
is the right characteristic A-matrix of (A,B) with the leading-column matrix
Drh-
Proof:
Theorem 5.4 can be proved using Theorems 5.1 and 3.9.
Corollary 5.3 With the l i n e a r s t a t e - f e e d b a c k c o n t r o l law determined in Theorem

5.4, the l e f t generalized eigenvectors of the closed-loop system map A = A-BF
can be determined by
Paij = TTTp 1 (k) C- )Dz

k=O k'[ ~r ~i rhc iCj-k) , 0~j~Ki-1, i=l,...,k
where
137
[Tp]ii r e v e r s e d upper t r i a n g u l a r T o e p l l t z m a t r i x wlth f i r s t column
[ a i i 2 , • • . , a i i ~ . , 1] T
1
[Tp] j i r e v e r s e d upper t r i a n g u l a r T o e p l l t z m a t r i x with f i r s t column
[aij2,...,alj~i,0] T i f Ki<~j or [ a l j 2 , . . . , a l j K j , O , . . . , 0 ] T if
<.>K..
z J
aijk, k=l,...,min(Ki,Kj) , are the c o e f f i c i e n t s of the ( i , j ) t h entry of ~r(~.),
Proof:
C o r o l l a r y 5.3 can be proved u s i n g Theorem 5.4 and C o r o l l a r y 3 . 3 .
Theorem 5.5 (Right L a t e n t S t r u c t u r e Assignment)
A linear state-feedback controller in Eq. ( 5 . 2 ) i s r e q u i r e d for c o n t r o l l i n g
the r e a c h a b l e system in Eq. ( 5 . 1 ) w i t h D - 0pXm such t h a t the c l o s e d - l o o p r i g h t
characteristic A-matrix has the a s s i g n e d l a t e n t r o o t s , {~i,i=l,...,k}, and r i g h t
generalized latent vectors, {qlj,j=0,1,...,Li-l,i-l,o..,k}, where &. i s the

k 1
length of J o r d a n c h a i n c o r r e s p o n d i n g to ~. and ~ ~.=n. Define Aj ~ b l o c k
~.x~. x i'1
diag[]i,isl,...,k]¢C x x, where J i i s d e f i n e d in Theorem 5 . 4 , and Cj ~
m×&.
[~j1,...,~jk]cd nXn where ~Ji = Eqio''"'qi~-l]¢C ~. Assume that (~j,~j) is
1
an o b s e r v a b l e pair. The o b s e r v a b i l i t y indices of (Aj,Cj) are the Kronecker
i n d i c e s of (A,B), and the l e f t characteristic A-matrlx of ( A j , C j ) i s D~(A) w i t h
a l e a d i n g - r o w m a t r i x I m. Then, the feedback g a i n , P, and the i n p u t g a i n , GF,
can be determined by F - P c T c and GF - DrhD r c where Tc i s the t r a n s f o r m a t i o n
m a t r i x which t r a n s f o r m s A i n t o i t s canonical controller form, A - T AT- 1 . F

C C C C
is computed from the f o l l o w i n g m a t r i x e q u a t i o n :

138
(5.12)
Fc~r(~O = OFDrc(~)-Dr(~)
where Drc(A) = DrhcD~(~) , Drh c i s any square upper t r i a n g u l a r constant matrix
with d i a g o n a l elements a l l l's and Dr(~) i s the r i g h t c h a r a c t e r i s t i c A-matrix of
(A,B).
Proof:
From the assumption and Lerma 2.3, D~I(A) can be realized by
(s.x3)
TOj i s the t r a n s f o r m a t i o n m a t r i x which t r a n s f o r m s Aj i n t o i t s o b s e r v e r c a n o n i c a l
form A0 = TojAjToj-
--I -
Since D~h = Im' the (i,j)th entry of V ~ ( A ) ) i j , has degree m i n ( ~ i , ~ j ) - i ,
where Vi = ~ r i ( V ~ ( A ) ) , and (D~CA))il i s monic. Because Vi = Ki, i = l , . . . , m , where
Ki a r e the Kronecker i n d i c e s of (A,B), thus ~ci(DE(A)) = Ki and DE(A) i s a l s o
column reduced. Therefore, Drc(A) , defined as Drc(A) ~ D~(A)Drhc, has
~ c i ( D r c ( A ) ) = Ki . From the assumption, Drh c i s an upper t r i a n g u l a r m a t r i x with

d i a g o n a l elements a l l l's. Thus, Drc(A) i s a column reduced c a n o n i c a l A-matrix
with column i n d i c e s Bci(Drc(A)) = Ki,i=l,...,m and Drc(k) is an a p p r o p r i a t e

closed-loop right characteristic A-matrlx. From Theorem 3.9 and D:~CA) " ~j~XX n
- A j ) 1BjDrhc+DjDrhc, Cj c o n t a i n s a l l the p r e a s s i g n e d r i g h t l a t e n t v e c t o r s ,
qij,j=0,1,...,~i-l,i=l,...,k, of Drc(A). Therefore, the c o n t r o l law in Theorem
5.5 c o n t a i n s the p r e a s s i g n e d r i g h t l a t e n t s t r u c t u r e . •
C o r o l l a r y 5.4 With the l i n e a r s t a t e - f e e d b a c k c o n t r o l law determined in Theorem
5.5, the right generalized e i g e n v e c t o r of the c l o s e d - l o o p system map A = A-BF
can be determined by
139
-q a i j = T;1 ~ ,(k)(~i)~i ( j-k)' O~j~E--X (5.14)

~1 Yr z
k=O
Proof:
C o r o l l a r y 5.4 can be proved u s i n g C o r o l l a r y 3 . 2 .
Example 5.2
Let us consider the same open loop system described i n Example 5.1. We
shall construct the linear state feedback control laws using left and right
latent structure assignments as follows:
(i) Left latent structure assignment
Let the desired latent roots and left latent vectors of the closed-loop
right characteristic ~-matrix be
Xl =-5 ; ~to = It -2 o] T
X2 = -4 ; P20 = [3 -40] T
i3 = - 3 ; ~3o = [t-lo] z
~4 =-2 ; ~4o = [1 1 o] T
X5 = -2 | P50 = [2 -3 O]
T
From Theorem 5.4, we have
-i -40 00 00 00 1
Aj = o -3 o o
0 0 -2 0
0 0 0 -1
and 1 -2 0 "]
3 -4 0
Bj = i
1
2
-I
1
-3
0
0
0 J
140
It can easily be c h e c k e d t h a t the Kronecker indices of (Aj,Bj) are KI - 3, K2 "
2, K3 = 0 . Using the formulas given i n Theorem 5 . 4 , the feedback gain is found
to b e
[-35.4 -19.2 -46.6667 -29.3333 -17.6667 7

F ffi - | 15.4 - 6.2 -17.5000 - 9.5000 - 3.5000
L 0.0 0.0 0.0000 0.0000 0.0000 J
and t h e i n p u t g a i n i s
I -1 -1 ]
OF " 0 1 -2
0 0 1
The l i n e a r state-feedback control law i s
u(t) = -Fx(t) + GFr(t)
The c l o s e d - l o o p canonical RIdFD i s
GcCk) = N r ( X ) D : ~ ( k )
where D r c ( ~ ) is the closed-loop right characteristic ~-matrix:
X3+7X2÷22k+24 4k+8 0 1
Drc(k) = 6k÷12 k2+gA+14 0
0 0 1
It is straightforward to check the Drc(~) has latent roots ~i and l e f t latent
vectors Pi0 for 1~i~5, as desired.
(2) Right latent structure assignment
Let the desired latent roots and right latent vectors of the closed-loop
right characteristic ~-matrix be

141
~'I = - 5 ; ~10 " [1 o 0]T
~2=-4 ! ~2o" [ 1 2 ° ] T
X3 =-3 ; ~3o= [ 1 ° olT
X4=-2 ; ~4o = [1 o 0IT

T
X5 =-2 ; q50 = [7 6 0]
From Theorem 5.5~ we h a v e
o o0 o0 o]
-4 0
Aj = 0 -3 0 0
0 0 -2 0
0 0 0 -
and
1 0 1 0 1]
~j = o 1 o I -2
0 0 0 0 0
It is easy to check that the observability indices of (AjgCj) are V 1 = 3, V 2 =
2, V 3 = 0 which are the Kronecker indices o£ the open loop system. Choosing the
leading-column coefficient matrix of the closed-loop right characteristic ~-
matrix a s
Drhc
= [11i] 0
0
1
0
with formulas given in Theorem 5.5, the feedback gain is found to be
-47.7 -27.1 -65.1667 -44.8333 -28.66671

F = - -10.6 - 3.8 -11.0000 - 4.0000 0.0000
0.0 0 0.0000 0.0000 0.0000
and t h e i n p u t g a i n i s
142
I I -2 -I 1
GF = 0 I -2
0 0 I
The linear state-feedback control law is
The closed-loop canonlcal RMFD is found as
Gc(k) = Nr()`)D:c1()`)
where Drc()`) i s the c l o s e d - l o o p r i g h t c h a r a c t e r i s t i c l-matrix:
)`3+10A2+31~,+30 2)`2+6.5)`+9 0 1
Drc ()`) =
I 0
0
),2+6A+8
0
0
1
It is straightforward to check that Drc()`) has latent roots ~i and right latent
vectors qi0 for I<i~5 as desired.
5.4 D i v i s o r Assignment and Decouplin~ Deslzn of MIMO Systems

In this section, we s h a l l extend the latent structure assignment scheme
d i s c u s s e d in S e c t i o n 5.3 t o the s i m p l e r s t a t e - f e e d b a c k c o n t r o l l e r d e s i g n method,
namely, divisor assignment. Basically~ the divisor assignment approach is to
construct the feedback law by a s s i g n i n g the desired complete s e t of c a n o n i c a l
left or r i g h t d i v i s o r s f o r the c l o s e d - l o o p r i g h t c h a r a c t e r i s t i c k-matrix. From
Theorems 3.9 and 4 . 4 , divisor assignment is equlvalent to the latent structure
assignment in Section 5.3. However, the notion of divisor assignment enables us

to c o n s t r u c t the s t a t e - f e e d b a c k c o n t r o l l e r s for p a r a l l e l forms ( r e f e r to Chapter
Vl for details), input decoupling, and input-output decoupling [69-71] of

143
closed-loop MIMO systems.
Theorem 5.6 (Left Divisor Assignment)
Let {Drci(X) , i=l,...,k} be a desired complete set of canonical left
d i v i s o r s of the c l o s e d - l o o p r i g h t c h a r a c t e r i s t i c X-matrix. Assume t h a t (Aci ,
Bci,Cci,Dci) is a minimal realization quadruple of D;~i(X) for i-l,...,k.

Define
Ac = Block d i a g [ A c i , i = l , . . . , k ] (5.15a)
and
= -T -T (5.15b)
Bc [Bcl'''''Bck]
If the Kronecker indices of (Ac~Bc) are the same as those of the open loop
system, then the linear state-feedback control law in gq. (5.2) can be
c o n s t r u c t e d as
F = F T (q |f'
"~'-'--a"
C C
GF ffi DrhV;1 c (5.16b)
where F is determined by
C
(5.16c)
Fc~r(X) ffiGFDrc(X)-Dr(X)
Note t h a t Dr(X) i s the r i g h t characteristic k - m a t r i x of the open loop system
d e s c r i b e d in Eq. ( 5 . 1 ) ; Drc(X) i s the r i g h t c h a r a c t e r i s t i c X - m a t r l x of (Ac,Bc);
Drh and Drh c a r e the leading-column coefficient m a t r i c e s of Dr(X) and Vrc(X),
respectively. ~r(X) is defined by Eq. (2.13b). Tc is the matrix which
transforms the system map A of the open loop system to i t s controller canonical
144
form A •
C
Proof=
Similar to the proof of Theorem 5.4.
Theorem 5.7 (Right Divisor Assignment)
Let {D~ci(k) , i=l,...,k} be a desired complete set of canonical right
divisors of the closed-loop right characteristic l-matrix. Assume that (Aoi ,

-i
B0i,Coi,D0i) is a minimal realization quadruple of Sz;i(k) for i=l,...,k.
Define
AO = Block diag[A0i , iffil,...,k] (5.17a)
and
- - (5.17b)
T 0 = [c01,...,C0k]
If the observability indices of (A0,C O) are the same as the Kronecker indices of
the open loop system and the leading-row coefficient matrix of the left
characteristic l-matrix of. (A0~C O) is Im9 then the linear state-feedback control
law in Eq. (5.2) can be constructed as
F = F T (5.18a)
c c
-I (5.18b)
G F = Drh Drh c
where Fc is determined by
(5.18c)
Zc~r(k) = CFDrc(k)-Dr(X)
And, Drh c is an upper triangular matrix with diagonal elements all l's; Drc(k) =
145
Drh c D~c(X) and D~c(X) is the left characteristic h-matrix of (A0,C0). Dr(h) is
the right characteristic h-matrix of the open loop system. ~r(X) is defined by
Eq. (2.13b). T c is the transformation to transform the system map A of the open
loop system to its controller canonical form A C .
Proof:
Similar to t h e p r o o f o f Theorem 5 . 5 .
In practical appllcatlons~ we may c h o o s e the simpler structure for the
c o m p l e t e s e t of c a n o n i c a l left/rlght divisors of the closed loop characteristic X-
matrix such that all the necessary conditions in Theorems 5.6 and 5.7 are
satisfied. In Propositions 5.1-5.4~ various ways to carry out the left/right
divisor assignment are proposed. The i n f o r m a t i o n which i s needed to p e r f o r m t h e
d e s i g n o f an HIM0 s y s t e m u s i n g t h e a p p r o a c h e s in Propositions 5.1-5.4 is the set
of K r o n e c k e r indices of the open loop system. In P r o p o s i t i o n 5.5, we s h a l l
explore the input-output decoupling design problem, using the idea of d i v i s o r
assignment.
k
Proposition 5.1 L e t Vrci(X) = diag{dij(X),j=l,...,m } for i=l,...,k, and
it
deg[dij(X)] = ~j, then {Drci(~),i=l,...,k} is a complete set of left divisors
which can be a s s i g n e d to c o n s t r u c t
D (X). •
k rc
S i n c e Drc(X ) = d l a g { w dij(X),j=l,...,m} , the Kronecker indices of Drc(X)
w i l l be t h e same as t h o s e o f t h e open loop s y s t e m . A p p l y i n g Theorem 5.6 g i v e s
the d e s i r e d f e e d b a c k g a i n s F and i n p u t g a i n GF.
Similarly, for right divisor assignment, we have t h e f o l l o w i n g :
k
Proposition 5.2 Let D&ci(X) = dlag{dij(X),j=l,...,m ) f o r i=l,...,k, and
i=l
deg[dij(X)] - Kj, then {Dg.ci(X),i=l,...,k } is a complete set of right divisors
which can be assigned to construct Drc(X). •

146
k
Since D~c(l) - diag{ w di~(l),j=l,...,m}
, J the observability indices of
i=l
D~c(~) will be the same a s the Kronecker indices of the open loop system.
Selecting a Drh c and applying Theorem 5.7 gives the desired feedback gain F and
input gain GF.
Proposition 5.3 Let
(5.19)
Drci(~) = diag[li_l,di(~),Im_i],i=l,...,m
with deg[di(l) ] = Ki, then {Drci(l),i=l,...,m } is a complete set of left

divisors which can be assigned to construct Drc(~). •
Proposition 5.4 Let
(5.20)
D£ci(l) = diag[li_l,di(1),Im_i],i=l,...,m
with deg[di(~) ] = Ki, then {D£ci(R),i=l,...,m} is a complete set of right
divisors which can be assigned to construct Drc(~). •
The divisors suggested in Proposition 5.3 result in
DEc(1) = diag{di(1),i=l .....m}
and the closed-loop KMFD becomes
m
Gc(~) = C[~In-(A+BF)]-IBG F = ~ N .(~)D-~(~)
i 1 rcl r~
Also, from Eq. (2.15c), we have Nrci(R) = [Ni(~),0pxl,...,0p×l] , where Ni(R) is

a column polynomial vector. The closed-loop transfer function matrix becomes
147
m s. (~,)
(5.21)
y(t) = I~ ri(t)
i=l
The p a r a l l e l decomposed s y s t e m in Eq. ( 5 . 2 1 ) i s an i n p u t d e c o u p l i n g s y s t e m . The
same reasoning can be applied for Proposition 5.4 with Drh c = Im. The technique
for input decoupling can be extended to design an input-output decoupling system
which is d e f i n e d as
n.(X)
I (5.22)
Yi(t) = ~ ri(t) for i'l,...,m and m=p
i
The necessary and sufficient conditions for input-output decoupling using linear
state-feedback have been investigated by many authors [10,68-71]. Wolovich

~im
[10,69] derived the condition that the matrix %÷~ ~G(%)G(%) K G be nonsingular
for m=pj where ~G(%) is the interactor of the open loop systemt namely
f°
~G(~.) = diag{A z,i--l,...,m} for fik0
When K G is nonsingular, the closed-loop characteristic h-matrix can be factored
as
(5.23a)
Drc(X) - Dd(X)N(X)
where N(A) is a right divisor of N r(~), or
(5.23b)
Nr (~) = NdC~)N(~)
Nd(A) = d i a g { n i ( A ) , i = l , . . . , m } ; D d ( A ) " diag{di(k),i'l . . . . ,m}

148
deg[di(A)] = deg[ni(k)]+fl
For practical applications~ N(k) must be stable and invertible. Note that the
m
closed-loop poles associated with w di(A) can be arbltrarily assigned, and
i-i
that ni(k) is the monic common divisor of the ith row component of Nr(A). The
feedback gains for input-output decoupling can be determined as follows.
Proposition 5.5 Assume that di(l) are monic, then the feedback gains for the
input-output decoupling are F = FcTc,G F = K; I, where Pc satisfies
(5.24)
Fc~r(A) " D:hI[K;IDrc(k)-D r(A)]
Proof:
Since
aim ~
l.~o ~G(1)Nr(),)D (1) = KGG F
and
Nr(k)Dr~(k) = diag{ni(1)/di(1),i-l,...,m }
where both ni(k) and di(k) are monic, then KGG F = Im or G F = KG I.
From Eq. (5.7)~ we obtain G = DrhDr h and Drc b = KGDrh.
Thus, Pc~r(A) = (KGDrh)-IDrc(A)-D:hlDrCA) = D:hI[K;IDrc(A)-Dr(A)].
Example 5.3
Consider a 3-input, 2-output s y s t e m
~(t) = D3~(t)+Bu(t)
y(t) " CX(t)÷Du(t)

149
where
I
0
-2
0
4
0
-
[i I
0
4
I
1 -1 4 -5 11 ]
C = 0 1 -1 4 -5 ; D = 02x 3
The Kronecker indices are found to be K I - 3, K 2 = 2 and K 3 = 0. Note that the
input matrix B contains a dependent vector. From Eq. (2.11b) we obtain the
similarity transformation matrix as

I 0 -0.11111 0.55556 -1.55556 3.55556
0 -0.22222 0.11111 -0.11111 0.11111
Tc = 0 0.55556 -0.77778 1.77778 -2.77778
i -2.88889 6.44444 -13.4444 26.4444
-1 3.22222 -7.11111 16.1111 -32.1111
Ac~ Bc and C c in the controller canonical form become
Fo 1o o oI 0 0 0
/ o o 1 o o I 0 0 0
Ac = TcAT:I = i _ 2 _ 3 _ _0_ _0__ _0 3 ; B - z B - I 0 1
/ o o o o 1/ c = 0 0 0
L-3 -3 0 -1 -2 j 0 1 2
Cc CT:I [2 0 2:4 3]
= = 4 4 3 !2 2 ; Dc = 02x3
From Eq. (2.14c) we obtain
I I ~. ~k
2 o OIT 0 o 0]
Sr(X) = 0 0 0 I Ar - -3 0 -1 -2 ;
0 0 0 0 0 0 0 0 0
Drh
= 1
0
2
1
-1
=
[o] 0
0
I
0
-2
1
150
• r ( k
plays an important role in determining the eigenvectors of the system )
matrix Ac and the latent vectors of t h e right characteristic )`-matrix of the
system. The right characteristic )`-matrix of the system can be written as
Dr()`) = Drh[Dh(k)-Ar~r()`)] =
[lo_1][ 3i3 _2 o oj
0 1 -2 3 +3 )`2+2)`+I 0
0 0 1 0 1
=
[ )`3-3)`-2
3).+3
0
0
),2+2)`+1
0
-1 ]
-2
1
Nr(~) in Eq. (2.15c) becomes
[ 2)`2+2 3)`+4 0 ]
~r()`) = Cc~r()`) =
3)`2+4)`+4 2),+2 0
The RHFD of the system can be described as G(~) = Nr(k)D;l()`).
To d e m o n s t r a t e the input decoupling of a MIMO system, using the divisor
a s s i g n m e n t method, we a s s i g n the desired complete s e t of left divisors of the
closed-loop right characteristic )`-matrix a s :
Drcl()`) = 0 )`+2 0 Drc2()`)= 0 )`+4 0

0 0 1 0 0 1
The m i n i m a l r e a l i z a t i o n of t h e c l o s e d - l o o p characteristic )`-matrix i n a Jordan
form can e a s i l y be found from Drcl()`) and Dr¢2(~) a s :
AjC = block diag -2

0
0
-4
, l: -4
; BjC = [lol 1!IT
o°
0
0
I
0
0
0
CjC
[o.33333o.o1.o00o7,_o.5
0.0
oo]0..5 0.0 |l 0.0 -0.5 ; DjC =
[o o o]
0 0 0
0.0 0.0 0.0 I 0.0 0.0 0 0 l
151
Thus, the right characteristic l-matrix of the closed-loop system becomes
[ 13+7~i+141+8 0 0
Vrc(l) = 12+6~+8 0
0 1
From Theorem 5 . 6 , we h a v e t h e s t a t e - f e e d b a c k control law:
u(t) = Fx(t) + Cr(t)
where
0.0 1.0 - 2.0 5.0 -18.0 I
F l
= -3.0 6.3333 -14.66667 24.66667 -45.66667
FeTe 0.0 0.0 0.0 0.0 0.0 J
=
_ [,0
GF DrhDr~ h =
0 0I -
The closed-loop RHFD of the designed system is described by
212+2 31+4 0 13+712+141+8 0 0 -1
312+41+4 21+2 0 0 0
The partial fraction expansion of the closed-loop RMFD yields (refer to Section
6.1)
[ 71+11-1 0][12+!I+4 0 0] -I [~: 4 0][ii2 0 0]-1

Gc(1) = I+2 + I+4 0
71+10 -I 0 0 1 3 0 0 1
Example 5.4
To demonstrate the input-output decoupling of a HIHO system~ we consider
the following 2-input, 2-output system:
IxCt) = AX(t) + Bu(t)
y(t) = CS(t) + Du(t)

152
where
-i -2
-2 24 -5
-9 421 - 1
A = 2 -5 10 -23 ; B= - 3
1 -2 4 -4 1
L0 o o o 2 0
C =[ : - 5 , 7 8
10,-21, 8 8
41 ~. ]; D
- 9 -20 43 -85 = 02x2
The KHFD of the systemcan be computedas
G(X) = I 12+31-40 -X+Il [ 13-31-2 0 ] -1

1-2J L 3x+3 X2+21+1
The interactoris
~o(I> " [ 0
KG 1~
[,
0
The numerator matrix Nr(1) can be decomposed as
Nr(1) " [13+31-40 -I+II

Nd(1)N(1)I-2
=
where
Nd(X) = [ I-I 0 ] I+4 -I ]
0 1-2 ; N(X) = 0 1
The number o f p o l e s w h i c h can be a s s i g n e d is 4. Let the left divisor of the
closed-loop right characteristic l-matrix be

153
Dd(~) =
(X+l)(X+2)
0
o
(~+2) (X+4)
]
Thus, the c l o s e d - l o o p right characteristic X - m a t r i x becomes
Dre(X) = Dd(X)N(X )
= [(~+1)(x+2)(x+4)o -(~+1)(~+z)(x+2)(~+4)
]
The feedback control law can be s o l v e d as
u(t) = OFt(t) + ~ x ( t )
where
o~-,;'-[o ]:I
r-3.0 8.66667 -19.33333 37.33333 -80.33333 ]
F = FcT c ffi L _ 3 . 0 6.33333 -14.66667 24.66667 -45.66667
The closed-loop RMFD of the designed system becomes
Cc(~) =
(~.÷I ) (X+2)
0
0
X-2
(~,+2)(X+4)
]
showing t h a t the d e s i g n e d system i s i n p u t - o u t p u t decoupled.
CHAPTER VI STRUCTURAL DECOMPOSITION THEORIES AND APPLICATIONS IN
MULTIVARIABLE CONTROL SYSTEMS
The main purpose of this chapter is to investigate the ~nterconnection
structures of multlvariable systems described by MFVs. The d i v i s o r and s p e c t r a l
factorization theorems developed in Chapter IV w i l l be utilized as the primary
tools for studying the properties of the basic interconnection structures~ these
being the parallel formsp the cascade forms~ and the semi-cascade forms of
multivarlable systems.
For convenience in explaining these basic interconnection forms, we
consider a class of m-input~ m-output systems~ which can be described by s
proper, 3rd-degree, right matrix fraction description (RMFD) w i t h a monic
denominator ~-matrix [80]. The RMFD c a n be described by the following
fundamental structures:
Y(I) = c (A)u(A) (6. la)

r
Gr(A) ~ N r (A) D: 1 (k) - (NOA3+N 1 A2+N2A+N 3 ) (ImA3+DIA2+D2 A+D 3 )-1 (6. Ib)
= No[(AIm+SI)(IIm+S2)(IIm+S3)] [(lim+~l)()~im+~2)(Xim+~3)]-i (6.1c)

= (N013+NI12+N2 I+N3 ) ( llm+~3 ) -I( llm+~2 ) -i( llm+~l ) -I (6. Id)
= No+K 1 (AIm+LI )-I+K2( XIm+L2)-I+K3 ( AIm+L3 )-1 (6. le)
= NO [(Aim+Z 1) (Aim+p 1)-I] [AIm+Z2 ) (Aim+P2 )-i] [(kim+Z3 ) (klm+P3 )-i]
(6.1f)
^
where NO, N i and Di, i=1,2,3 in Eq. (6.lb) are matrix coefficients| S i and Si'
i=i~2,3 in Eq. (6.1c) are the spectral factors of Nr(A) and Dr(A) , respectively~
-Li, i=1,2~3 in Eq. (6.1e) are the complete set of left solvents [17-18,25-27]
of Dr(k). The matrices Ki, i-I,2,3 in Eq. (6ole) are the block residues [27-31]
155
a s s o c i a t e d with Li, i = 1 , 2 , 3 . Pi and Zi , i = I , 2 , 3 i n Eq. ( 6 . 1 f ) are, in general,
not the spectral factors or s o l v e n t s of Nr(A) and Dr(k) , r e s p e c t l v e l y . Note

A
that, for a s l n g l e v a r l a b l e system, or a completely decoupled system, Si - Zi
and Si = Li = Pi for i=I,2,3. The set of linear A-matrlces, {(~Im+Li) ,
i = I , 2 t 3 } , i n Eq. (6.1e) i s the complete s e t of l e f t d i v i s o r s [31] of Dr(~).
The structure of Eq. (6.1d) is the seml-cascade form in which the
denominator modes with a s s o c i a t e d l a t e n t v e c t o r s of Dr(~ ) are i s o l a t e d from one
another but not from the dynamic modes of the numerator. The seml-cascade form
is a s u i t a b l e structure for modal c o n t r o l of m u l t l v a r l a b l e systems [30]. The
p a r a l l e l s t r u c t u r e of Eq. (6.1e) i s a block p a r t i a l f r a c t i o n form [26] of Gr(k) ,
which is often used because of its satisfactory reliability and sensitivity
p r o p e r t i e s i n p h y s i c a l system r e a l i z a t i o n s . The cascade s t r u c t u r e of Eq. ( 6 . 1 f )
is v a l u a b l e from many a s p e c t s . For example, such a s t r u c t u r e p r o v i d e s advantages
of s i m p l l c l t y i n c o n s t r u c t i o n , sensitivity reduction, maintenance i n p h y s i c a l
system i m p l e m e n t a t i o n and t r o u b l e s h o o t i n g of the implementation hardware.
I n t h i s c h a p t e r , we s h a l l i n v e s t i g a t e the i n t e r c o n u e c t l o n s t r u c t u r e s of an
m-lnput, p - o u t p u t l i n e a r t l m e - l n v a r l a n t system d e s c r i b e d by Eq. ( 2 . 1 ) , repeated
here for c o n v e n i e n c e :
Xx(r) = Ax(t) + Su(t)

(6.2a)
y(t) = CX(t) + Du(t)
(6.2b)
I f the system i n Eq. (6.2) i s r e a c h a b l e , the c a n o n i c a l P.~D of the systems
becomes ( r e f e r to S e c t i o n 2.2)
G(I) = Nr(k)D;I(A)+D
where N r ( ~ ) D ; l ( k ) is a strictly proper r a t i o n a l k - m a t r i x , and Nr(1) and Dr(X)
are r i g h t coprime. I f the system i n Eq. ( 6 . 2 ) i s o b s e r v a b l e , the c a n o n l c a l LHFD

158
of the system i s g i v e n by ( r e f e r to S e c t i o n 2.2)
G(A) = D~I(A)NE(A)+D
where D[I(A)NE(A ) i s a strictly proper rational A-matrlx, and N£(k) and DE(A)
are left co,rime. As p o i n t e d out i n S e c t i o n 2.1, for a general MIMO system,
Dr(A) and DE(A) may not be monic. Instead, they are c a n o n i c a l column-reduced
and c a n o n i c a l row-reduced A - m a t r l c e s , r e s p e c t i v e l y . The p r i n c i p a l t o o l s f o r the
results developed in this chapter are the spectral decomposition/factorization
theorems of nonsingular A-matrices explored in Chapter IV v i a the geometric
approaches [5,81-82].
This chapter is organized as follows. In Section 6.1, we study the
parallel decomposition structure of MIMO systems and its applications to model
reductions and multiport network synthesis. In Section 6.2, the semi-cascade
realizations and their applications to modal control of MIMO systems are
investigated. In Section 6.3~ the minimal cascade realizations and their
applications to multiport network synthesis are discussed.
6.1 Parallel Decomposition T h e o r i e s and T h e i r A p p l i c a t i o n s to MIMO Systems and
Circuits
In t h i s section, a complete s e t of c a n o n i c a l left (right) d i v i s o r s of the
right (left) characteristic A-matrix of a reachable (observable) MIMO system
derived in Section 4 of Chapter IV is employed to study the parallel decomposed
structures [26,27,31] of MIMO systems and to d e v e l o p t h e i r a p p l i c a t i o n s to model
reductions and network synthesis.
6.1.1 Parallel Decomposition T h e o r i e s of MIMO Systems
As i n d i c a t e d in Theorem 4 . 1 3 , i f a MIMO system i s r e a c h a b l e ( o b s e r v a b l e ) , a
complete set of canonical left (right) divisors of the right (left)

157
characteristic l-matrlx exists.
Definition 6.1 Let a system (A,B,C,D) in Eq. (6.2) be reachable and
decomposable i n t o the forms in Eq. ( 4 . 3 0 ) . Also, let (Kil,~i2,...,~im) be the
Kronecker indices of (Ai,Bi), and Tci be the transformation matrix which
transforms ( A i , B i ) i n t o i t s c o n t r o l l e r c a n o n i c a l form ( A c i , B c l ) . Define
n i Xnoi (6.3a)
Acp ~ b l o c k d i a g [ A c i , i f l p . . . ~ k ] ; Ac£¢C
n. Xm
T T T T cC~ (6.3b)
Bcp [ B c l ' B c 2 , ' ' ' , B c k ] ; Bci
Ccp ~ [ C c l , C c 2 , . . . , C c k ] ; CciEcpXni (6.3c)
A (6.3d)
v = D
cp
and Ac~• = T c i. A . T• - ~ ;c l B c i. = TciBi; Cc x . • C.T-~

i cl for i=l,...,k. Then,
(Acp,Bcp,Ccp,Dcp) i s c a l l e d a p a r a l l e l controller canonical form of the system
(A,B,C,D). 0
Theorem 6.1 (The G e n e r a l i z e d Right P a r t i a l F r a c t i o n Expansion Theorem)
Let (Acp,Bcp,Ccp,Dcp) be a p a r a l l e l controller c a n o n i c a l form of the system
(A,B~C,D) i n Eq. ( 6 . 2 ) . Then, the r i g h t m a t r i x f r a c t i o n d e s c r i p t i o n (PJ4FD) can
he decomposed as
(6.4a)
i=I cp
where
SriCl)V~(X) ~ cci<IIn-Aci)-IScl
i
158
Dri(~) i s the r i g h t c h a r a c t e r i s t i c l - m a t r l x of (Aci,Bcl) and
(6.4b)
where
I Kil-I
1 ~ ...
00... ... 00 ... *I
~ri(X) A o o . . .
1 k ... kK ... 0 0 ...
• • tea • .•a • •a• • at• .~
0 0 ... 0 0 0 ... 0 ... 1 ~ ... K m-
(6.4c)
Proof:
From Eq. (6.3) we obtain
G(I) ~ C(XIn-A)-IB+D = Ccp(%In-Acp)-IBcp+Dcp
k k
= I C c i ( k l n - A c i ) - I B c i + D c p = .~ Nri(k)D:~(~k)÷Dcp
i=1 I I 1
Equation (6.4) shows that the RMFD can be described by partial fraction
expansion forms. The interpretation of the system structure in Eq. (6.4) is that
the original system has been decomposed into the parallel connections of
subsystems described by the canonical RMFD NrI(~)D:~(~) , i'l,...,k.
The dual results of the parallel decomposition for observable systems are
d e s c r i b e d as f o l l o w s .
Definition 6.2 If the system (A~B,C,D) in Eq. (6.2) is observable and
decomposable into the forms of Eq. (4.30), let (Vil,Vi2,...,~ip) be the
observability indices of (Ai,C i) and T0i be the transformation matrix which
transforms (Ai,C i) to its observer canonical form (A0itC0i). Define

159
n. Xn.
(6.5a)
Aop block dlag[A0i,i=l,...,k] ; A0£CC I
n. Xm
A T T T T, L (6.5b)
BOp = [B01,B02,...,B0k] " BOiE:C
pXn. (6.5c)
COp [CollC02,...,Cok]; C0£¢C 1
Dop ~ D
and AO£ = T0~AiT0£ ; B0£ = T;~Bi; C0~ = CiT01 for i=l,...,k.

Then, (A0p~B0p,C0p~D0p) is called a parallel observer canonical form of the
system (A,B,C,D). []
Theorem 6.2 (The Generalized Left Partial Fraction Expansion Theorem)
Let (A0p,B0p,Cop,Dop) be a parallel observer canonical £orm of the system
(A,B~C,D) in Eq. (6.2). Then, the LHFD of the system can be decomposed as
k
(6.6a)
i=l
where
D~C X)~i¢X) b C0iCX'n.-A0i)-IB0i

1
D~i(k) is the left characteristic ~-matrix of (Aoi,C0i) and
(6.6b)
where
160
I1 ~, ... Vil-I 0 0 ... ... 0 0 ... "~ ]
,.~b°i(1) . 0 0 . . 0 . 1. ~ . ... . ~ V°2-I 0 0 ...
" " "'" " " "'" " "'" " " "'" p-I
0 0 ... 0 0 0 ... 0 ... 1 k ... X ~-
(6.6c)
Proof:
Similar to Theorem 6.1.
As pointed out in Section 4.4, the matrix sign algorithm can be employed to
compute the transformations for block dlagonalizatlon of the system map A and to
find out the complete set of canonical left/right divisors for the right/left
characteristic ~-matrix of the s y s t e m i n Eq. (6.2). Therefore, the matrix sign
algorithms, together with the minimal nice selection algorithms (Section 2.3)
for determining the Kronecker indices of each parallel decomposed subsystem,
constitute effective numerical methods for performing the left/right matrix
fraction expansions for left/right MFDs o f MIMO s y s t e m s .
Applications of the parallel decomposition f o r MIMO s y s t e m s a r e illustrated
in t h e following subsections.
6.1.2 Model Reduction of MIMO Systems
A low-degree model of a high-degree HFD is often desirable for computer
simulations and c o n t r o l system designs. A model r e d u c t i o n method f o r a c l a s s o£
MIMO s y s t e m s described by MFDs w i t h monic characteristic ~-matrices has been
p r o p o s e d by $ h i e h and Tsay [27], T h i s method i s e x t e n d e d in this subsection for
analysis and d e s i g n [83] o f g e n e r a l MIMO s y s t e m s .
The b a s i c idea for reducing a h i g h o r d e r MIM0 s y s t e m t o a l o w e r o r d e r model
is to decompose the system into dominant and non-dominant subsystems, and to
discard the non-dominant subsystem which has faster dynamic responses. We
define the dominant and non-domlnant modes as follows.
Definition 6.3 The n o n - d o m i n a n t modes o f a c o n t i n u o u s - t i m e system are defined

181
as the modes with Re(kl)<y1, where Y1 is a real number, white the dominant modes
are the modes having Re(ki)>71. If the elgenvalues distribution of the original
system is unknown, the real number YI can be chosen as 71 ffi trace(A)/n, which is
the arithmetic mean of the elgenvalues of A. For a discrete-tlme system, the
non-dominant modes are the modes wlth [~£I<y2, where 72 is a positive real
number, while the dominant modes are the modes having [li[>Y2. If the
elgenvalues d i s t r i b u t i o n of the original system i s unknown, the p o s i t i v e real
number 7 2 can be chosen as 7 2 ffi I n - - I , which i s the geometric mean of the
elgenvalues of Ao O
Two methods, namely the frequency-domain reduction method and the time-
domain reduction method, are to be derived for the model reduction of MIMO
systems based on the idea mentioned above.
The use of Theorem 6.1 yields a frequency-domain model reduction method for
a right MFD as follows.
Theorem 6.3 If an MIMO system can be decomposed into parallel connections of
two subsystems as
(6.7a)
G(A) = NrI(A)D:II(k)+Nr2(A)D:21(A)+Dc p
such that Drl(~) contains latent roots of dominant modes and Dr2(~) contains
non-domlnant modes, then the reduced-degree model of G(k) can be determined by
(6.7b)
CA(~) - SrlC~)V~ll(~)+Vcp -" O(~)
where
A
Dcp ffi Nr2(0)D:21(0)+Dcp ( f o r c o n t l n u o u s - t i m e systems) (6.7c)
and
(6.7d)
Dcp ffi Nr2(1)D:~(1)÷Dcp ( f o r d i s c r e t e - t l m e systems)
^
D
cp i s an pXm c o n s t a n t m a t r i x d e r i v e d for matching the z e r o - t h time-moment of
162
impulse response of G(t), or the steady-state values of the unlt-step responses
o£ t h e o r i g i n a l stable system. •
In a similar manner, by u t i l i z l n g Theorem 6 . 2 , we o b t a i n a frequency-domaln
model r e d u c t i o n method f o r a l e f t MFD as f o l l o w s .
Theorem 6.4 I f an MIMO s y s t e m can be decomposed i n t o p a r a l l e l connections of
two s u b s y s t e m s as
(6.8a)
CCA)" D~(k)N~,1(k)+D~ICk)N&2Ck)+D0p
such that D~I(~ ) and V~2(k) contain latent roots o f dominant modes and non-
dominant modes, respectlvely, then the reduced-degree model of G(X) can be
d e t e r m i n e d by
(6.8b)
GA(A) " DEI( 0p
where
~(o).~2(o)+Dop (for contlnuous-time systems)
A
D0p D (6.8c)
-
and
^ -I
V0p = D~2(1)N£2(1)+Dop ( f o r d i s c r e t e - t i m e systems)
(6.8d)
R
S i n c e D r l ( k ) i n Eq. (6.7b) and D ¢ l ( k ) i n Eq. (6.8b) are obtained from the
induced pair (AL1,BL1) and the embedded pair (ARI,CR1) of (A,B) and (A,C),
respectively, the matrix sign algorithm in Theorem 4 . 2 1 can be employed to
compute t h e f r e q u e n c y - d o m a i n r e d u c t i o n models i n Theorems 6.3 and 6 . 4 .
In some c o n t r o l system problems, state-space reduction models a r e needed.
For i n s t a n c e , the suboptimal control d e s i g n and the a p p r o x i m a t e o b s e r v e r design
are in t h i s category. For t h e s e p u r p o s e s , t h e t l m e - d o m a i n method f o r the model
reduction o f an o b s e r v a b l e HIMO system can be s t a t e d as f o l l o w s .

163
Theorem 6.5 Let (A,B,C,D) be an o b s e r v a b l e c o n t i n u o u s - t i m e system and d e f i n e
nXn 1
SI =A Ind[ Sign + (A-711n)]¢C
nXn2
S 2 ~ Ind[Sign-(A-Ylln)]¢C , nl+n 2 = n
AR1 =A +
SIASI£Cn l x n l , S +I = ( S ; $ I ) - I s ;
A ;IB£Cnl xm -1
BRI = , SI [Inl Onlxn2][S1 S2]
A CsI£CPXnl
Cg1 =
D0
p ^ =AC(_A)-IB+D_CRI(_AR1)-IBRleCP×m
The r e d u c e d - o r d e r model w i t h d o m i n a n t modes of the o r i g i n a l system h a v i n g a n o n -
zero i n i t i a l v e c t o r X(0) becomes
Iz(t) = ARlZ(t)+BRlU(t)
Y s ( t ) = CRlZ(t)+D0pu(t) -" y ( t )
* -1 * ^
z(0) = [CR1CRI] CRI[CX(0)-DopU(0)] i f p~n I
or
• . -I[cx(0)_;
z(0) = CRI[CR1CR1] 0pu(0)] i f pan 1
The c o r r e s p o n d i n g c a n o n i c a l LMFD of (AR1,BRI,CR1,D0p) i s

164
Proof:
Theorem 6.5 can be proved from Theorems 6.4 and 4.16 and C o r o l l a r y 4 . 8 . •
An alternative tlme-domaln method f o r the model r e d u c t i o n of a reachable
MIMO system can be s t a t e d i n a s i m i l a r way:
Theorem 6.6 Let (A,B,C,D) be a r e a c h a b l e c o u t l u u o u s - t l m e system. Define
T nlxn
V I ~ {Ind[(Sign+(A-'(lln))T]} ¢C '
T u2xn n
V 2 ~ {Ind[(Sign-(A-Tlln))T]} ¢C , nl+n 2 "
+ nlxnl + * * -1
ALl ~ VlAVI¢C , V1 - Vl(VlV I)
BLI ~ VIB¢Cnlxm
- pxn I
CL 1 ~ CVI¢ C ,
VI ~ [VlV
T T2] -T [ I n l 0nlxn 2] T
^Dcp ~ C(-A)-IB+D-CLI(-ALI)-IBLIECp×m
Then, t h e r e d u c e d - o r d e r model wLth dominant models of t h e o r i g i n a l system having
a non-zero initlal v e c t o r X(O) becomes
kz(t) = ALlZ(t)+BLIU(t)
Y s ( t ) = CLlZ(t)+Dcpu(t) = y ( t )
165
* -I * ^
Z(0) ffi [CLICLI] CLI[CX(0)-DcpU(0)] if p>n I
or
. .
z(O) ffi CLI[CLICLI] -l[cx(o)_~c p U ( 0 ) ] i f p~n I
The corresponding canonical RHFD of (ALI,BLI,CLI,Dcp) is
Again, the matrix sign algorithm in Theorem 4.21 can be used to compute the
time-domain reduction models in Theorems 6.5 and 6.6.
Example 6.1
Given a 3-input 2-output continuous-time system described by the state-
space equations (6.21 with X(0) = 05xl, where
I 0.5 -3.0 1 0 . 5 -21.5 39.0 1 "-2.0 1.0 0.0

0.5 -2.0 - 3.5 6.5 - 5.0 1.0 -2.0 -3,0
A = -1.0 2.0 - 9.0 17.0 -30.0 ; B " 1.0 1.0 3.0 ;
-0.5 1.0 - 2.5 4.5 -25.0 2.0 1.0 4.0
0.0 0.0 0.0 0.0 - 8.0 1.0 0.0 1.0
C = [I0.0 -15.0 41.0 -65.0 131.0]

5.0 - 2.0 10.0 - 2.0 6.0 ; D " 02x 3
Find: (1) The c o r r e s p o n d i n g canonical LMFD o f t h i s system.
(21 The b l o c k decomposition of t h e MIMO s y s t e m into two s u b s y s t e m s :
one c o n t a i n i n g elgenvalues with real parts greater than
y1 (= trace(A)/n) and the other containing eigenvalues with real
parts less t h a n ~1"
(3) The left block partial fraction expansion and t h e reduced-degree
model of the LMFD.

188
(4) The r e d u c e d - o r d e r model of t h e o r i g i n a l system.
Solution:
Using the minimal nice selection algorithm in Section 2.3, we f i n d t h a t the
Kronecker indices of this system are gl-3, g2-2 and g3=O. Since g3=0, there
exists a dependent vector in the input m a t r i x B. The c o r r e s p o n d i n g canonical
LHFD can be found as
G(A) - D~I(),)NE(X)
where
A3+5.164A2+5.319A÷3.475 -0.208A-I.658 ]
D~(k) -
-0.649k2-6.853k-4.366 )~2+8.836)~+6.688
and
N~(k) - [ 7A2+22"I49A-2"074 16~2+80"527k+55"24 39A2+I83"403A+I08"406]
-4.545A-54.377 6.611A+35.366 8.675A+16.354 J
We s h a l l decompose t h e s y s t e m s into the two s u b s y s t e m s described in (2) above,
where e i g e n v a l u e s w i t h r e a l p a r t s greater t h a n Y1 = t r a c e ( A ) / n = -14/5 = -2.8.

+
First, we have to compute S i g n (A-71I 5) and Sign (A-7IIS) , which can be
e x p r e s s e d as
I
Sign÷(A+2.815) = ~ [Sign(A+2"8Is)+I5]
Sign-(A+2.815) = 15-Sign÷(A+2.815)
The computed m a t r i x s i g n f u n c t i o n s are:
0 1 -2 4 0 0 2 -4
Sign+(A+2.815) = 0 0 0 Sign-(A+2.815) - I -2
O 0 0 -2 0 0 2
0 O 0 0 0 0 0 0 0 I
167
From the ranks of Slgn+(A+2.SI5 ) and Sign-(A+2.SIs) , we have n I = 3 and n 2 = 2,
respectively. Thus~ the system map A can be block-decomposed into two
submatrices by using the block modal matrix M S • To find the block modal matrix
M S i n Eq. (4.41b) for the block-diagonallzatlon of A, we determine the canonical
injection maps S I and S 2 in Eq. (4.40) uslng Algorithm 4.1 in Section 4.4. The
o b t a i n e d m a t r i x MS becomes
-2 -1 -4 "]
M S " [Sl,S 2] -
0
0
0
0
1
0
0
0
4
2
1
O
2
1
0
0
8
4
2
1
]
where S I - Ind[Sign (A+2.815) ] and S2 = Ind[Sign-(A+2.8Is) ].
Let z(t) = M;Ix(t)

nlXl n2xl
,,+here ,'Ct)= [,'~Ct~,,"~Ct~ T, ,,Ct:>,+~,<t, -sCt),,C ,",:Ct~C
Then, the block-decomposed system in z(t)-coordinates becomes
~z(t) = ~z(t)+BDU(t)
y(t) = CvZ(t)+DvU(t)
where
AD " M;1AHS "~ b l o c k diag[ARl+Ap,2] ;
~I"SlASl"
+ [_o., _1.o
-6.0 -2~.S
2.5 ; A~- S~2-
: [,.o,o.o]
-0.5 1.0 4.5 0.0 -8.0
T TT
B D " M;IB = [BRI~B~]
168
BRI = -I.0
0.0
-4.0
1.0
-9.0
2.0
I ; BR2 = [_,.o _,.o]
1.0 0.0 1.0
CD ffi CMS ffi [CRI,CR2] |
CRI = CS1 . [ 10.0 -15.0 -63.0] . [ 1.0 5.0]

5.0 - 2.0 0.0 ; CR2 CS3 = 1.0 6.0
D D ffi D ffi 02x 3
From Theorem 6.2, we obtain the l e f t block p a r t i a l f r a c t i o n expansion of G(~) as
-I k)NELCk)+D~(k)NE2(A)
where
[ ~2+1.154~+0.903 -0.469 ] [ 5~-0.232 17k+13.611 39A+27]
Dg.l(k) = - 0 . 8 1 0 ~ - 0 . 5 6 5 k+0.846 ; NEI(A) = -7.052 4.225 1.398 j
DE2(A) ffi -4 A+8 ; N%2(A) = 3 -I 1
Note that {DEI(~) , D~2(k)} is a complete set of canonical right divisors of

D~(1).
Applying Theorem 6.4, the reduced-degree model obtained by dropping the
subsystem D;~(A)NE2(A), which has elgenvalues l e s s than - 2 . 8 , becomes
where
D0p = D (0)N9.2(0) ffi -0.5 0.25 0.0
-0.625 0.25 -0.125
The reduced-order model obtained from Theorem 6.5 becomes

169
XZs(t) - ARIZs(t)+BRIU(t)
^
YxCt) ffi CRlzs(t)+D0pU(t) -~ y ( t )
* -I * ^
Zs(0) - (CRICRI) CRl[CDZ(0)-D0pU(0)]
1.164 -0.465 2.327 -0.465 1.396 ]

= 0.643 -0.257 1.285 -0.257 0.771 z(0)
-0.539 0.903 -2.390 4.028 -8.084
-0.145 0.058 -0.029 ]

+ -0.080 0.032 -0.016 u(0)
-0.021 -0.011 -0.002
The t i m e - r e s p o n s e curves of the o r l g i n a l system and the reduced-degree model and
r e d u c e d - o r d e r models o b t a i n e d above for the u n i t step i n p u t s of u ( t ) - [ l t 0 , 0 ] T,
u(t) ffi [ 0 , 1 , 0 ] T and u(t) = [ 0 , 0 , 1 ] T are compared in Figs. 6.1-6.6. The
approximation is quite satisfactory. It is interesting to note that the
eigenvalues of the original system are ~1,2 = -0.5±j0.St ~3 = -I.0, ~4 =-4.0 and
k5 = - 8 . The reduced-degree model and the r e d u c e d - o r d e r model o b t a i n e d by u s i n g
Theorems 6.11 and 6.5, r e s p e c t i v e l y , c o n t a i n e i g e n v a l u e s kl,A 2 and k3"
6.1.3 Multiport Network Synthesis
If G(I) is the transfer function matrix of a multiport network to be
realized, then Eqs. (6.4a) and (6.6a) are the parallel decompositions of G(%).
Each subsystem Nri(~)D:~(~) can be r e a l i z e d u s i n g a c t i v e elements or R-C ladder
networks with summers [ 8 4 , 8 5 ] . Thus, a d e s i r a b l e network can be c o n s t r u c t e d by
connecting all subnetworks together with the direct matrix gain Dcp. The
proposed approach can be viewed as the g e n e r a l i z e d F o s t e r r e a l i z a t i o n [86-88]
for m u l t i p o r t network s y n t h e s i s .
170
!
!
S
|
'o:o* am: soo'. rso: Ib.oo 1#',m Jb.oo ~.m ab.m
TIME (IN SEC)
F[O. 8 . 1 RE:SPOHSE OF ¥ l ( t ) TO INPUTS , U ~ - I . U 2 - O . U 3 , , O
OR|G|NN~ SYSTEfl
...... REDUCEDMODEL
I
'o:oo am: |w" Tso: Ib,m ~z'.u Ib.m i)'.m Jb.m
T~nE (~N $E~
FIG. 6.2 R[SPOHSi[ OF Y 2 ( t ) TO INPUTS " U I ' I . U 2 " O . U 3 = O
ORIOTN+U. SrStl~
...... REDUCEDPIOOE~
171
Do a'.so w:oo T'm ~b.oo I~a.m ~k.oo lY.so ab.oo

Till( (|N SEC)
FEO. 6 . 3 RESPONSE OF • l E t ) TO INPUTS * U I - O . U 2 - I . U 3 m O
OR|GXNkL STSTE11
...... I~(ouc(o mOB.
.,
!
oo aim" soo" eso: ~ooo'. Ib.so :k.oo l~.m ab.~

TLrl( (IN SEC)
FIG. 6 . 4 RESPONSE OF Y 2 ( t ) TO INPUTS , U I - O . U 2 - I . U 3 - O
ORIGINAL STSTE31
...... REOuCF.OnOOEL
172
'~° °00
T|ftE (IN SEC)
F|0. S . S RESPONSE OF ¥ 1 ( t ) TO |NPUTS * U I - 0 . U 2 - 0 . U 3 w l
01~ GZHAL STSTI~I
REDUCED hODlel
N aim
: sod
: f'so to'.m la'. N tk.o0 t~.N ab.N
Tilt[ KXR SEC)
FXO. 6 . 6 R(SPOHS[ OF ¥ 2 ( t ) TO INPUTS i U l m O . U 2 - 0 . U 3 - I
ORtGXNAL S~'STEn
...... REDUCED r10O(l,.
173
6.2 Semi-Cascade Decomposition Theories and Their Applications to Modal
Controls of MIMO Systems
In this section, we investigate the realization of MFDs v i a spectral
factorizations of the characteristic l-matrices of MIMO s y s t e m s , and develop
their applications to modal control of MIMO systems [30,89~90].
6.2.1 Semi-Cascade Decomposition Theories
If t h e MIMO system (A,B,C,D) of Eq. (6.2) is reachable and A,B,C can be
decomposed into the forms in Eq. (4.28), then from Theorem 4.12, we obtain the
RMFD as
(6.9)
G(1) = N r ( l ) [ D r 2 ( l ) R ( ~ ) l - l + v = Nr(X)R-I(1)D:~(1)+D
where Dr2(1) is the right characteristic l-matrix of (A2,B 2) and R(1) is a right
divisor of D ( ~ ) .
r
Lemma 6.1 A reachable MIMO system of dimension greater than 1 can always be
decomposed into a semi-cascade realization form as follows.
(6.10a)
IX(t) = AX(t) + Su(t)
y ( t ) = CX(t) + Vu(t) (6.10b)
where
A =
A1 AI2 1; AI¢C
n.xn.
I i
; i=1,2
(6.10c)
0 A2
n.xm
T T T • (6.10d)
B = [BI,B2] ; BI¢C ; i=1,2
174
pxn. (6.10e)
C = [Ci,C2] ; Ci¢C x; i=l,2
V = 0 pXm (6.100
The corresponding semi-cascade RMFD is given by
(6.11a)
G(A) = Nr(A)R-I(A)D:~(A)
where Dr2CA ) is the right characteristic A-matrix of (A2,B2) , and it has
(A2pB2~Cr2~Dr2) as its associated 'minimal realizationp where Cr2 and Dr2 are
defined in Lemma 2.1. Also
Nr(A ) = C T : I ~ r ( A ) (6.11b)
(6.11c)
R(~) " Cr2VT:ISr ()0 ÷Dr2D r (),)
where Dr(A) is the right characteristic A-matrix o f (A,B) and V is the canonical
projectionX~)</S in which (A2,B2) is induced~ where S i s an A-invariant subspace
of Xwhich is the state-space.
Proof:
Lemma 6 . 1 c a n b e p r o v e d by d i r e c t l y u s i n g Theorem 4 . 1 and Theorem 4 . 1 2 . •
Le-,-a 6 . 1 r e v e a l s the state-space structure and t h e RMFD r e p r e s e n t a t i o n for
the semi-cascade realization of a reachable MIMO s y s t e m . From E q s . (6.10) and
(6.11), the seml-cascade realization of a reachable system is, in fact, the
result of spectral factorization of the right characteristic X-matrix of the
system.
The dual results for the semi-cascade realizations of observable systems

175
can be stated as follows.
Lemma 6.2 An observable MIMO system in Eq. (6.2) can always be decomposed into
a semi-cascade realization form as shown in Eq. (6.10) and the corresponding
seml-cascade LMFD is given by
G< x) - D~ll c X)L-lc X).~C X) (6.12a)
where D~I(X) is the left characteristic X-matrix of (AI,C l) and it has
(AI,B~I,CI,D~I) as its minimal realization. Also
N~(X) = ~bE(X)ToIB (6.12b)
(6.12c)
L(X) = ~E( X)T;IsBEI+D~( X)DEI
where DE(X) is the left characteristic X-matrix of (A,C)~ $ is the canonical
injection map from S~X, and (AI~Cl) is the embedded map of (AIC) in S p whereX
and S are defined in Lemma 6.1.
Proof:
Similar to Lemma 6.1.
6.2.2 Modal Control of MIMO Systems
To perform modal control of MIMO systems by assigning the divisors of the
characteristic X-matricesl which contain specific latent roots and associated
latent vectors of the X-matrlces, we study the MIMO system (A,B,C1 D) in Eq.
(6.2) which can be transformed into the following Jordan canonical form, using
the modal matrix M of A:

176
lXj(t) = AjXj(t)+Bju(t) (6.13a)
y(t) = CjXj(t) (6.13b)
where
Xj(t) = M-Ix(t) (6.13c)
n.xn.
Aj = M-IAM = block dlag[Aji,i=l,...,k];Aji¢ C I i (6.13d)
m n.xm
T T _* . ~ 1 (6.13e)
Bj = M-IB = [BjI~Bj2~... ~BJk j ; B j i ¢ C
pxn.
(6.13f)
Cj - CH - [Cjl,Cj2,...,Cjk];CjieC
Each Aji is a full Jordan block with eigenvalue ki' i.e.

~i 1
ii 1 . . n.xn.
I I (6.13g)
Aji " £C
• li
nxn.
The generalized left eigenvectors, pjicC t associated with Aji can be
determined by partitioning the modal matrix M such that the following matrix
equations are satisfied:
(6.13h)
Apj i = PjiAji, i = l , 2 , . . . , k
The goal of modal control is to design a linear state-feedback control law
for the system in Eq. (6.2), or the equivalent system in Eqs. (6.10) and (6.13),
such that some undesirable elgenvalues and associated eigenvectors of the system
177
can be replaced by deslrable ones. Since the system map of Eq. (6.13) can be
viewed as a s p e c i a l c a s e o f the b l o c k t r i a n g u l a r i z e d system map i n Eq. ( 6 . 1 0 ) ~ we
consider t h e more g e n e r a l b l o c k t r l a n g u l a r i z e d system i n Eq. ( 6 . 1 3 ) as
(6.14a)
~XD(t) = ADXD(t)+BDU(t)
(6.14b)
y(t) = CDXD(t)+DDU(t)
Xv(t) " T~Ix(t) (6.14c)
where
- ; BD -
rBvl] [itI ; XD(t) =
o '~2 L BD2 XD2(t)
Cv = [Cvl,Cv2]; DD=0
ADI is the collection of the unimportant system modes not to be controlled~ and
AD2 contains all dynamic modes to be controlled. XDi(t) , BDi and CDi are
matrices of appropriate dimensions. Let the control law be
(6.15a)
u(t) - -FDXD(t)+GFr(t) - -FDT;Ix(t)+GFr(t)
where F D ~ [0,FD2 ]. Then, the closed-loop system becomes
(6.15b)
~XD(t) = ~XD(t)÷BDU(t)
y(t) " CDXv(t) (6.15c)
where
178
= . [ AD1 ADI2-BDIFD2 ] I BDIOF1

BD " " BDGF
AD AD-BDFD O AD2-BD2FD2 BD2GF
The modal control design problem is to find FD2 such that AD2-BD2FD2 has the
prespecifled elgenvalues and eigenvectors. From the structures of ~ and BD' we
observe that the modal control of the entire system is equivalent to the
elgenstructure assignment of t h e partitioned subsystem (AD2,BD2). Using Theorem
5.1, we obtain the control law for the sub-system (AD2,BD2) as follows.
Lemma 6.3 Let Dr2(1) be the right characteristic l-matrix of (AD2,BD2) which
contains the undesirable modes to be controlled. Assume that the desirable
closed-loop right characteristic l-matrix with prespecifled modes is Drc2(1).
Then, the feedback gain FD2 and input gain G F in Eq. (6.15) become FD2 ffi Fc2Tc2
and G F = Drh2Dr~c2 , respectively. Fc2 can be determined from the following

matrix equation:
Fc2~r2(1) = GFDrc2(I)-Dr2(I)
Tc2 is the transformation matrix w h i c h transforms (AD2,BD2) into its canonical
controller form, and Drh 2 and Drhc2 are the leading-column matrices of Dr2(1)
and Drc2(1) , respectively. •

From the structure of the closed-loop state equations in Eq. (6.15) and
Lemons 6.1, we observe that Drc2(1) is the canonical left divisor of the closed-
loop characteristic l-matrix, Drc(1). Therefore, the latent roots and left
generalized latent vectors of Drc2(1) are also part of the latent roots and left
generalized latent vectors of D (1). As a result, we can determine the modal

rc
control law via latent roots and associated latent vectors assignment as follows.
179
Theorem 6.7 (Modal Control Law with Latent Structure Assignment)
Assume that the desired closed-loop latent roots and left latent vectors
corresponding to the modes to be controlled are (~2i,i=l,...,k) and
(P2ij,j=0,1,...,~i-l,i=l,...,k), respectively. Define
n2Xn2
Aj2 = block diag[J2i~i=l~...,k]EC ;n2 = ~. (6.16a)
i= I i
- -T ~T ITECn2 xm (6.16b)
Bj2 = [Bj21~... ~ J2k J
where ~2i and Bj2i are defined in Theorem 5.4.
Assume that (Aj2,Bj2) is a reachable pair and that the Kronecker indices of
(Aj2,Bj2) are equal to those of (AD2,BD2) in Eq. (6.14). Then, the feedback
gain F D and the input gain G F of the modal control law in Eq. (6.15a) can be
determined by FD = [0,FD2], where FD2 = Fc2Tc2 and G F = Drh2Dr c2" Fc2 can be
determined from the following matrix equation=
(6.16c)
Fc2~r2(l) = GFDrc2(1)-Dr2(l)
where Dr2(1) , ~r2(l) are the same as those defined in Lemma 6.3, and Drc2(l) is
the right characteristic l-matrix of (Aj2,Bj2).
Proof:
Theorem 6.7 can be proved from Theorem 5.4 and Lemma 6.3. •
From Theorems 5.2 and 6.3~ the closed-loop RMFD of the system in Eq. (6.10)
with D = 0px m and the modal control law in Eq. (6.15a) can be represented as
G (1) = C(II -i)-IBGF = N (1);-l(1) (6.17a)

c n r rc
where
180
= A_BFDT;I (6.17b)
^Drc(A) ffi G; I[FDT; IT :I~r(A)+Dr(A)] (6.17e)
and Nr(A)D:I(A) = G(A) is the canonical RMFD of the open loop system in Eq.
(6.10), Tc is the transformation matrix which transforms (A,B) into its
controller canonical form. The properties of Drc(A) can be stated as follows.
^
Theorem 6.8 Drc(k) defined in Eq. (6.17c) is column reduced. If the open loop
right characteristic k-matrix of the system in Eq. (6.14) is factored as
(6.18)
Dr(k) = R(k)Dr2(A)
where Dr2(A) is the right characteristic k-matrix of (AD2,BD2) , then Drc(A) can
be factored as
^ (6.19)
D r c (l) = R(A)Drc2(A)
where Drc2(A) is the right characteristic k-matrix of (AD2-BD2FD2,BD2GF).
Proof:
From Eq. (6.17c), the leadlng-column matrix, G;iDrh , is nonsingular because

^
Drh , the leading column matrix of Dr(1) , is nonsingular, and therefore Drc(A) is
column reduced.
If Dr(k) is factored as in Eq. (6.18), then from Theorem 4.1, Lemma 6.1 and
Eq. (6.14), we obtain
Dr2(~ = Cr2(AIn2-AD2)-IBD2+Dr2
181
R(x) = ~ 2V~DIT~I~r(X)+~r2Dr(X)
where V = [0,1n2 ]. From Lemma 5.2, we have
Drc2(k) = Cr2(AIn2-AD2+BD2FD2)-IBD2GF+Dr2GF
Since Drc(X) is column reduced and Drc2(%) is a canonical left divisor of
Drc(X), from Theorem 4.10 and Eq. (6.15), we obtain
D r c (),) = R' ()0Drc2(k)
where
R'c x) = er2VT lT l r< X)+ r2OF rc(X)

Therefore
A
R'( X)-R(~) = Dr2 [GFDrc (),)-Dr(X)]
Also, substituting Eq. (6.17e) into the above equation yields
Since Dr2--[Im-~bTr2(0)~Jr2(0)]Drh12' the ith row of F is null, and (Drh2)ij =
(Drhl2)ij = 0 for j~i if %i = 0, it follows that
Dr2FD2 = (Im-~Tr2(0)~r2(0))Drhl2FD2 =- 0
T h e r e f o r e , R(k) = R ' ( k ) , or D c ( ~ ) = R(~)Drc2(~).
From Eq. (6.17a) and Theorem 6.8 we obtain:

182
Theorem 6.9 The closed-loop RMFD of the modal controlled system can be
represented as the semi-cascade form~
Gc(~) = Nr(X)R-I(X)D~2(~ ) (6.20)
where Nr(~) is the numerator of the open-loop canonical RMFD, R(%) is the right
divisor of the open-loop right characteristic %-matrix associated with the
uncontrolled modes I and Drc2(% ) is the desired closed-loop right characteristic ~-
matrix of the controlled modes.
Theorem 6.9 reveals the important feature of the modal control design that
the left divisor of the denominator of the open-loop RMFD has been changed to the
desired one with prespecified eigenstructures.
Note that Gc(~) in Eq. (6.20) may not be a canonical RMFD. Using Theorem
3.4~ we can transform Eq. (6.20) into a canonical RMFD as follows:
(6.21)
Gc(X) = Nrc(X)Drcl(X)
where Nrc(X) = Nr(%)Ur(~) , Drc(X) = Drc(~)Ur(%) , and Ur(%) is a unimodular %-
matrix. The result in Eq. (6.21) implies that although the zeros of the open-
loop system remain unchanged under modal control laws~ the numerator of the
closed-loop canonical RMFD may be changed.
Example 6.2
Given a 3-input 2-output continuous time system in the form (6.2) where
2.2 1.6 4.0 3.0 1.0 1

-6.4 -4.2 -8.0 -6.0 -2.0
A -- -1.6 -0.8 -3.0 -2.0 -i.0
3.2 1.6 4.0 3.0 2.0
9.6 4.8 12.0 6.0 2.0
183
-2.0 1.0 0.0

-1.0 -2.0 -5.0
B = 3.0 -1.0 1.0
-2.0 3.O 4.O
1.0 -2.0 -3.0
3.6 3.8 5.0 4.0 0.0 ]

C = 5.4 2.2 5.0 3.0 -I.0 ' D = 02x 3
The RMFD of the system is found to be
G(X) = NrfX)Drl(X)
where
Nr(X) = [ -4X2+13~-13
-5 X2+I0 A+3
7%-12
12A+7
0.0 ]
0.0
and
X3-2X2-X+2 -~.2+X+2 -1 ]
D (X) = 0 ~.2+2 k+l -2
r
0 0 1
We shall illustrate the semi-cascade realization and the modal control
design for this system. For these purposesj the system map A has to be block-
triangularized. We shall use Theorem 4.20 to construct the transformation T D to
transform A to a block-trlangular form.
Selecting r0 = -~ and r I = 0 and utilizing the matrix sign algorithm in
Theorem 4.21 yields
S i g n ( r 0 ) ( A ) = 15
184
-2.6 -0.8 -2.0 0.0 0.0

3.2 0.6 4.0 0.0 0.0
Sign(rl)(A ) = 1.6 0.8 1.0 0.0 0.0
-3.2 -1.6 -4.0 -I.0 0.0
6.4 3.2 8.0 4.0 1.0
From Theorem 4.20, we have the canonical injection map S(r0,rl ) and the
canonical projection map V(r0,rl):
1.8 -1.6 -0.8 1.6 -3.2 IT

S(r01rl ) = 0.4 0.2 -0.4 0.8 -1.6
1.0 -2.0 0.0 2.0 -4.0
- = [ -0.8 -0.4 -I.0 0.0 0.0 ]

V(r0~r I) 3.2 1.6 4.0 2.0 1.0
Thus the transformation T D can be constructed as

-I
$Tr0 ' -i F 0.6444 -1.7778 -0.4444 -0.I000 0.2000
rl) |-0.1778 0.9111 -0.2222 0.2000 -0.4000
TD = -i-~.~4_4~ -0.2222 0.4444 0.I000 -0.2000
~(ro,rl) 1-°"8°°° :o-:45o~-:17o~o~ o.oooo o.oooo
t 3.2000 1.6000 4.0000 2.0000 1.oooo
I 1.8000 0.4000 1.0000~ -0.4444 0.0000

-1.6000 0.2000 -2,0000! -0.2222 0.0000
= -0.8000 -0.4000 0.00001 -0.5556 0.0000
1.6000 0.8000 2.0000! 1.6000 0.4000
-3.2000 -1.6000 -4.00001 0.8000 0.2000
Defining
X ( t ) = T D XD(t)
then the state equations of t h e system in X D - coordinates become
~XD(t) = ADXD(t)+BDU(t)
y(~) = CDXD(t)+DDU(t )
185
where
BD=TI__[B02
01]
C D -- CT D = [CDI , Cv2]
and
0.6 0.8 2.0] 2.0889 2.2444 ]

~I = -3.2 -2.6 -4.0 ; L,A~12 = -1.9556 -3.3778
0.0 0.0 -I.0 -0.4889 -0.8444
1.0 1.0 ]
AD2 ffi 0.0 2.0
-2.0444 0.7444 -0.5556 1 -I.0 1.0 1.0 1

ffi -2.0222 -0.3778 -2.7778 ;
BDI 2.0444 0.2556 2.5556 BD2 = 1.0 0.0 1.0
CDI ffi
2.8
10.2
3.4
4.6
4.0
II.0
]; CD2 =
1.1778
-1.6667
1.6000
1.0000
]
D D -- 02x3
(I) Semi-Cascade Realization
From Corollary 4.2~ Dr(X) can be factored as
Dr(X) = Dr2(X)R(X)
186
where Dr2(A) is the right characteristic X-matrix of (AD2,BD2):
A-2 0 -i 1
Dr2(A) = -2 A-I -2
0 0 i
and R(A), which is a right divisor of Dr(A) ~ can be computed by using Eq.
(6.11c):
ALl A1 0 ]
R(A) = 2A+2 A+I 0
0 0 i
Thus, from Lemma 6.1, the seml-cascade RMFD of the entire system is given by:
G(A) = Nr(A)R-I(~,)D;~ (A)
Note that Dr2(X) contains two unstable modes of the system with latent roots 1
and 2, and R(A) contains three stable modes with latent roots -i~-I~ and -I.
(2) Modal Control Law
Since AD2 contains two unstable modes with eigenvalues 1 and 2~ we shall
design a modal controller to stabilize these two modes. Using the left latent
structure assignment in Theorem 5.4~ we select the latent roots and latent
vectors of Drc2(X) in Theorem 6.7 to be
~21 = - 2 P210 = [ I -I O] T
and
T
A22 = -4 P220 " [i I O]
s o we h a v e
187
Aj2 = [ -20 -40 ]
and
- = 1 -I 0 ]
B J2 [ i 1 0
The right characteristic l-matrix of (Aj2,Bj2) is found to be
;%+3 i 0 ]
I ),+3 0
Drc2 ()') = 0 0 I
Thus, from Eq. (6.15a) and Theorem 6.7, we have the modal control law
u(t) = -FX(t)+GFr(t )
where
-18.4 -9.2 -23.0 -12.0 -6.0 ]

F FDTI~ ffi- -19.2 -9.6 -24.0 -14.0 -7.0
0.0 0.0 0.0 0.0 0.0
and
1.0 0.0 -I.0 ]

GF = 0.0 1.0 -2.0
0.0 0.0 1.0
From Theorem 6.8, we have the RMFD of the closed-loop system:
Gc(1) -- Nr(;k)Drcl(X)
where
^
rc Drc2 (
188
13+312+k-I -A2-3k-2 0 ]
= 3k3+8A+5 k2+3~+2 0
0 0 1
^-I
Note that Gc(A) - Nr(A)Dr~(A) is not a canonical RMFD. The canonical RMFD of
the closed loop system can be found from the closed loop state equations:
kX(t) = AX(t)+BGFrCt)
yCt) = CXCt)+Du(t)
where
= A-BF
-19.8 10.4 26.0 13.0 6.0 ]

=
50.4
-17.6
24.2 6 3 . 0 3 4 . 0 18.0
-18.8 -48.0 -24.0 -12.0
- 3 7 . 6
8.8 -22.0 -15.0

-
29,6 -14.8 37.0 22.0

7.0
I0.0
-
]
From Eq. (2.15), we have the canonical RMFD of the closed-loop system:
where
N (~) = [ -412-8l+23 71-12 0 ]

rc -5~2-261-18 12l+7 0
X3+6X2+10X+5 -12-3X-2 0 ]
Drc(k) = -X-I X2+3X+2 0
0 0 i
It is easy to show that

189
A
Drc()%) = Drc(1)Ur()%)
and
Nrc()%) = Nr(l)Ur()% )
where Ur()%) is a unimodular l-matrix|
i 0 0 ]
Ur()%) = -3 1 0
0 0 1
Also, P210 and P220 are left latent vectors of Drc(l).
6.3 Cascade Decomposition Theories and Their Applications to Multiport Network
Synthesis
In this section~ we shall present the factorization of MFDs to achieve the
cascade realizations of MIM0 systems. The minimal factorizations of MFDs are
defined first. Algorithms to perform the minimal factorization [91-98] of RMFDs
and LMFDs are derived. The corresponding state-space cascade realizations are
also discussed. Finally, applications of the cascade system structures to
multiport network realizations are illustrated.
6.3.1 Minimal Factorizations of MFDs
If G( )%)eCmXm[ )%] can be factorized as
(6.22)
G()k) = GI(~,)G2()%)
where G()%), GI()%) and G2()%) are described by MFDs, then Gl(S) and G2()%) are
called the left factor and right factor of G()%), respectively. The
factorization of G()%) in Eq. (6.22) is minimal [94] if

190
Deg(D(X)) = Deg(DI(A))+Deg(D2(X))
(6.23)
where D(A), DI(A) and D2(A) are the left or right characteristic X-matrices of
the systems described by G(A)~ GI(X) and G2(k) ~ respectively. The minimal
factorization of G(X) in Eq. (6.22) can be generalized as
(6.24)
G(A) = GI(X)G2(A)...Gk(A)
k
with Deg(D(~.)) = ~ Deg(Di(A)).
i=l
Definition 6.4 An RMFD~ G(X) = Nr(X)D;I(x)Ecm×m(x), is nonslngular if
D (A)Ecmxm[A] is a column-reduced canonical X-matrlx, and Nr(A)ecm×m[A ] is

r
nonsingular. Similarly~ an LMFD9 G(X) = D;I(A)N~(A)p is nonsingular if

D~(A) e~mxm[A] is a row-reduced canonical X-matrix, and N~(A) Ecm×m[x] is
nonsingular. Q
Theorem 6.10 (RMFD Factorization)
Consider a nonsingular canonical RMFD expressed as G(X) = Nr(X)Dr(X).
Assume that DI(X) and NI(X) are a column-reduced left canonical divisor and a
row-reduced right canonical divisor of Dr(X) and Nr(k) , respectively~ or
Dr(X ) = Dl(D~2(X ) (6.25a)
Nr(X ) = N2(X)~I(k) (6.25b)
Let
(6.25c)
N1 ( X)D;I(X) = HI( X)+;I( ~)~;1(X)
where Hl(k) is a k-matrix and NI(%)D;I(x) is a proper RMFD.

191
Let (AI,BI,CIpD I) he a minimal realization of the RMFD, NI ( D 5 2 1 ( ~ . Then,
(6.26a)
G(I) : 02(1)Gl(1)
where
(6.26b)
Cl(D : N l ( D D [ l (
(6.26c)
G2(~ = N 2 ( ~ D 2 1 ( ~
and
(6.26d)
Nl(X) : N'I( X)+D2( X)HI(X)
(6.26e)
D21()%)NI(X) is the LMFD of (AI,BI,CI,DI).
Proof:
Since
and (AI,BI,CI,DI) is a minimal realization of the RMFD~ NI(X)D21(X), we can
determine the corresponding LNFD, using (A1,B1,CI~D 1) as follows:
Therefore
G(X) = N2(X)D21(X)[NI(X)+D2(X)HI(X)]DII(X) N2(%)D21(%)NI(%)DII(% )
and the result in Eq. (6.26a) follows. •
Theorem 6.10 provides a method for factoring a nonsingular canonical RMFD
into the product of two nonsingular RMFDs. Since D2(%) may not be in a column-
192
reduced canonical form, G2(X) may not be a canonical P~FD. Howeverj using
Theorem 3.4, we can convert G2(X) into a canonical RMFD.
Corollary 6.1 The RMFD factorization in Theorem 6.10 is minimal.
Proof:
Deg(DI(X))+Deg(D2(X) ) = Deg(DI(X))+Deg(D2(X)) = Deg(Dr(X))
Since the left factor GI(X) and right factor G2(X) can be represented as
nonsingular canonical RMFDs, we can repeatedly factorize G(X) to obtain
(6.27)
GCX) = Gk(~)Gk_I(~)...GI(X)
if the conditions in Theorem 6.10 are satisfied.
The main feature of the RMFD factorization in Theorem 6.10 is that the
poles and zeros with associated latent vectors of GICX) and G2(X) can be
preselected by performing spectral factorlzatlons of Nr(~) and Dr(X). This
property is important in multlport network synthesis and factorlzation of a
para-Hermitlan MFD.
The dual results of the LMFD factorizatlon are stated as follows.
Theorem 6.11 (LMFD Factorization)
Given a nonsingular canonical LMFD as
G(X) = D~l(x)N(X) (6.28a)
Let D~(X) = D2(~)DI(X ) and NE(X) = NI(A)N2(X) , where D1CX) and NI(X) a r e a row-
reduced right canonical divisor and a c o l u m n - r e d u c e d l e f t canonlcal divisor of

193
D£(X) and Ng(X), respectively. Let D;I(x) NI(X) - HI(X)+D21(X) NI(X) , where
HI()~) is a l-matrix and D;I(x) NI(X) is proper. Also, let (AI,BI,CI,D I) be a
minimal realization of D21(~k)~l(%). Then
(6.28b)
O(l) = Cl(k)C2(L)
where
el(x) = D~I(x)NI(X); G2(X) = D~ICX)N2(X); Nt(X) = Bt(X)D2(X)
NI(I)D21(I) is the RMFD of (AI,BI,CI,DI). •
Repeatedly applying Theorem 6.11, a nonsingular canonical LMFD can be
factorized into a product of many nonsingular canonical LMFDs as
G(I) -- Ol(1)G2(},)...Ok(1)
(6.2g)
6.3.2 State-space Realization of the Cascade Systems
Let the state-space realizations of GI(I) and G2(l) in Eq. (6.26a) be
(6.30a)
%Xl(t ) = AiXl(t)+BlUl(t )
(6.30h)
Yl(t) ~ ClXl(t)+DlUl(t )
and
(6.30c)
)d(2(t) = A2X2(t)+B2u2(t)
(6.30d)
Y2(t) = C2X2(t)+D2u2(t )
where Ai,Bi,C i and Di, i=I~2 are matrices of appropriate dimensions. For
cascade connections, we let u(t) = Ul(t) , Yl(t) = u2(t) , and y(t) = y2(t), where
u(t) and y(t) are the input and output of the overall system, respectively.
194
Then, t h e state-space realization of t h e overall system becomes
rxict]
[AI01
xict] [i]
= + u(t) (6.31a)
ILX2(t) B2C 1 A 2 JLX2 (t) B2D 1
xl(t) ]
(6.31b)
y(t) = [D2CI,C2] + D2DlU(t)
x2(t)
The above result can be generalized to the case of many subsystems as follows.
Theorem 6.12 Given a nonsingular MFD, G(~)~ which has a minimal factorization
(6.32)
G(I) = Gk(%)Gk_I(I)...GI(I)
and (Ai,Bi,Ci,D i) is the minimal realization quadruples of G.(~), i=l,2,...,k.

l
Then~ the state-space minimal realization quadruple of G(~), (A~B~C,D), can be
expressed as
AI 0 ... 0 0
B2C 1 A2 ... 0 0
B3D2C I B3C 2 ... 0 0

A =
Bk_IDk_2...D2C I Bk_IDk_2...D3C 2 ... Ak_ I 0
BkDk_I'.'D2C I BkDk_I-.-D3C 2 --- BkCk_ I Ak
B = [B~,(B2DI)T,(B3D2DI)T,...,(Bk_IDk_2...DI)T,(BkDk_I...DI)T] T
C = [Dk...D2CI~Dk...D3C2~Dk...D4C3~...,DkCk_I,Ck]
D = [DkDk_I...DI]
195
Proof:
Theorem 6.12 can be proved by using induction on the number of cascade
subsystems. •
6.3.3 Cascade Realizations of Multiport Networks
If the matrix transfer function of a multlport network can be factorized as
G(~,) = GI(1)G2(%)...Gk(1)
the subnetworks G.(%) can be constructed in a way such that each G.(I) contains
1 1
desirable latent roots and associated latent vectors of G(1). Then, we can
realize each subnetwork with each G.(1) and cascade them to construct the
1
complete network for G(1). Since the factorization of G(~) is minimal~ the
realized network uses minimal active elements or reactive elements. The
advantages of cascade structure were noted in the introduction to this chapter.
Example 6.3
Let the canonical right matrix fraction description of a 2-input 2-output
network be
G(1) = Nr(1)Drl(1)
where
Nr(~) = [ 0 t tO
Dr(A) = 1212 + [ 2.0511"8

10.2 ~ + 2.05 9.2
The zeros of G(1) are -5, -I0 and the poles of G(1) are (-l,-l,-10±jl).
196
Find (i) The minimal realizations of D -I in the controller canonical form and
r
in the Jordan form.
(2) The minimal cascade factorizations of G(%).
(3) The cascade realization of G(~) in the state-space representation.
(4) The implementation of the cascade networks.
Solution
Using Lemma 2.1, the minimal realization of D-I(~) in the controller

r
canonical form can be found as
D;I(L) = Cc(%I4-Ac)-IBc+Sc
where
_ 0.00 1.00 i 0.00 0.00
A = -8.75 -n.8o lO.OO 1.8o
c o.oo - ~.~ ~To~ t.oo
-2.05 - 2.05 I-9.20 -10.20
B = i 0
c ~-~
0 1
o o o]
Cc 0 0 I 1 0
Dc 02x 2
The corresponding Jordan form minimal realization of D -I can be obtained by

r
finding the Jordan forms minimal realization quadruple (Aj,Bj,Cj,Dj) from
(Ac,Bc~Cc,D c) as
D;I(1) = Cj(II4-Aj)-IBj+D J
where
Aj
.
lit0 0] E11] r01002s
.
-lJ o
0. 1-10
.
o
.1 ; Bj
o 1
i--0"
-z i-o.I o.15 1
; cj =1-o-.i--_-o-;1- -| ; DJ
=
02×2
0 I-I -i0 I i L 0.0 -0.125 J
197
Note that the latent roots of Dr(1) have been arranged in two groups; {-i)-I}
and {-10±jl} as shown in the system map Aj. Using Theorems 4.9 and 4.12
together with the Jordan form minimal realization quadruple of D-l(1)i yields
r
the spectral factorization of Dr(l ) as follows:
Dr(I ) = DL(1)DoDr(1)
where
DO : 12 ; DR(1) = 121 + [ 0.05

0.8
-0.8 ] II -i
1.2 ; DL(1) = 121 + [ 2 9 ]
In a similar manner) the spectral factorization of N (1) can be determined as

r
Nr(1) = [ 0 10
Thus, G(1) can be represented as
G(X) = Nr(X)IDL(X)DR(X)]-I = Nr(X)DRI(X)DZI(x)
=o,[' ~ ][~ ~+ (~,o °o)] [,~+ ( o::~ -o.~,.~)]-' ['~+ ( ~" -'4-'
Applying Theorem 6.10 gives the cascade factorization of G(X) as
G(X) = G2(X)Gt(X)
where
G2(I) = [ 0 I 1.6569 0.14487J
-I
o,<~<>-[~ ~÷ ,~.ooo,r
o.o~ o.~,~1[~+~.~,,, (~'~ I)}
198
The cascade realization of G(X)~ using Theorem 6.12, can be found as
^ ^
G(I) = G2(A)GI(I) : Cr(ll4-Ar)-lBr+Dr
where
r-~.o ~.o o.o o.o

|- 20 -90 00 00
r !-- 4~-9450 1.3586 -1.8552 0.4414 B r --
L 9.6069 -0.0552 -I.6569 -0.1448
^ [0 01~ ~] ^
Cr 0 0 ! 0 i ; Dr = 02x2
The cascade realization of G(X) using four one-port R-C networks and summers but
without using multiwinding transformers and integrators [84] is shown in Fig.
6.7. The poles and zeros of the first subnetwork~ described by GI(%) , are
{-10±jl} and {-5,-I0}, respectively whilst the poles and zeros of the second
subnetwork, described by G2(X) ~ are {-i,-i} and {~}~ respectively. The
realized cascade multi-port network contains the prespecified poles and zeros in
each subnetwork.
4
In '
.= i (D
Ul i/ "~~ !
I
I
Fig. G,7 Cascade Rea]tzatlon of the Network G(X)=61(x)G2(X)

BIBLIOGRAPHY
[i] R.E. Kalman, P.L. Falb, and M.A. Arbib, Topics in Mathematical System
Theory, McGraw-Hill, New York~ 1969.
[2] S. Barnett, Matrices in Control Theory, Van Nostrand Reinhold Co., New
York, 1971; Second Edition, Krieger, Florida, 1984.
[3] S. Barnett, Polynomials and Linear Control Systems, Marcel Dekker, New
York, 1983.
[4] M.K. Sain, Introduction to Alsebraic System Theory, Academic Press~ New
York, 1981.
[5] W.M. Wonham, Linear Multlvariable Control: A Geometric Approach, Springer

Verlag, New York, 1979.
[6] A.G.J. MacFarlane, Frequency-Response Methods in Control Systems, IEEE

Press~ 1979.
[7] B. Porter and R. Crossley, Modal Control, Taylor & Francis Ltd., London,
1972.
[8] I. Postlethwaite and A.G.J. MacFarlane, A Complex Variable Approach to the

Analysis of Linear Multivariable Feedback Systems, Springer-Verlag, New
York, 1979.
[9] D.D. Siljak m Larse-Scale Dynamics Systems, North-Holland, New York, 1978.
[I0] W.A. Wolovich, Linear Multivariable Systems, New York: Springer-Verlag,

1974.
[ii] C.A. Desoer and M. Vidyasagar, Feedback Systems: Input-Output Properties,

Academic Press, New York, 1975.
[12] L. Pernebo, Alsebraic Control Theory for Linear Multivariable Systems 3 Lund
Institute of Technology, Sweden, 1978.
[13] T. Kailath, Linear Systems, Englewood Cliffs, NJ: Prentice-Hall, 1980.
[14] R.V. Patel and N. Munro, Multivariable System Theory and Design, Pergamon
Press, New York, 1982.
[15] H.H. Rosenbrock, State-Space and Multivariable Theory, Nelson, London,

1970.
[16] D.L. Russell, Mathematics of Finlte-Dimensional Control Systems: Theory

and Design, Marcel Dekker Inc., New York, 1979.
[17] P. Lancaster, Lambda Matrices and Vibrating Systems, Pergamon Press, New
York, 1966.
[18] J.E. Dennis, J.F. Traub, and R.P. Weber, "The Algebraic Theory of Matrix
Polynomials", SlAM J. Numerical Analysis, 13, pp. 831-845, 1976.
[19] P. Lancaster, "A Fundamental Theorem on Lambda-Matrices with Applications-

201
I. Ordinary Differential Equations with Constant Coefficients," Linear

Algebra and Its Applications, Vol. 18, pp. 189-211, 1977.
[20] I. Gohberg, P. Lancaster, L. Rodman, "Spectral Analysis of Matrix

Polynomials I Canonical Forms and Divisors", Linear Algebra and Its
Applications, Vol. 20~ pp. 1-44, 1978.
[21] I. Gohberg, M.A. Kaashoek, and L. Rodman, "Spectral Analysis of Families of

Operator Polynomial and a Generalized Vandermonde Matrox Ip the Finite-
Dimensional Case"9 Academic Press 9 Topics in Functional Analysis, pp. 91-
128, 1978.
[22] I. Gohberg~ P. Lancaster~ and L. Rodman, Matrix Polynomials, Academic

Press, New York, 1982.
[23] A.S. Markus and I.V. Mereuca, "On the Completen n-Tuple of Roots of the
Operator Equation Corresponding to a Polynomial Operator Bundle"9 Math.
USSR Izvestija 9 Vol. 7, pp. 1105-11289 1978.
[24] L.5. Shleh~ Y.T. Tsay~ a n d N.P. Coleman 9 "Algorithms for Solvents and
Spectral Factors of Matrix Polynomials" 9 Int. J. Systems Science, Vol. 12,
pp. 1303-1316, 1981.
[25] L.S. Shieh and Y.T. Tsay 9 "Transformations of Solvents and Spectral Factors
of Matrix Polynomials and their Applications", Int. J. Control 9 Vol. 34,
pp. 813-8239 1981.
[26] Y.T. Tsayy L.S. Shieh 9 R.E. Yates and S. ~arnett, "Block Partial Fraction
Expansion of a Rational Matrix"9 Linear and Multilinear Algebra, Vol. Ii,
pp. 225-241, 1982.
[27] L.S. Shieh and Y.T. Tsay, "Block Modal Matrices and their Applications to
Multivariable Control Systems"9 Proc. lEE, Vol. 129, Pt. D. 9 pp. 41-48j
1982.
[28] L.S. Shieh and Y.T. Tsay 9 "Transformations of a Class of Multivariable

Control Systems to Block Companion Forms", IEEE Trans. Automat. Contr. 9
Vol. AC-279 pp. 199-2039 1982.
[29] Y.T. Tsay and L.S. Shieh 9 "Some Applications of Rational Matrices to
Problems in Systems Theory", Int. J. Systems Science~ Volo 13, pp. 1319-
1337, 1982.
[30] Y.T. Tsay and L.S. Shleh, "Block Decomposition and Block Modal Control of
Multivariable Control Systems", Automatiea, Vol. 199 pp. 29-$0~ 1983.
[31] Y.T. Tsay and L.S. Shieh, "Irreducible Divisors of A-Matrices and their
Applications to Multivariable Control Systems"~ Int. J. Control, Vol. 37,
pp. 17-36, 1983.
[32] L.S. Shleh, Y.T. Tsay and R.E. Yates, "Block Transformations of a Class of
Multivariable Control Systems to Four Basic Block Companion Forms", Int. J.
of Computers and Mathematics with Applications, Vol. 9, pp. 703-714, 1983.
[33] D.G. Luenberger, "Canonical Forms for Linear Multlvariable Systems," IEEE
Trans. Automat. Control, Vol. AC-12, pp. 290-293, 1967.
[34] W.A. Wolovich and P.L. Falb, "On the Structure of Multivariable Systems" 9
202
SlAM J. Contr., Vol. 7, pp. 437-451~ 1969.
[35] V.M. Popov, "Invariant Description of Linear Time-lnvariant Controllable

Systems", SlAM J. Contr.~ Vol. 10p pp. 252-264, 1972.
[36] R.E. Kalman, "Kronecker Invariant and Feedback"9 Ordinary Differential

Equations, Edited by L. Weiss, Academic Press, pp. 459-4719 1974.
[37] A.C. Antoulas~ "On Canonical Forms for Linear Constant Systems"~ Int. J.
Control, Vol. 33, pp. 95-122, 1981.
[38] V. Kucera~ Discrete Linear Control, John Wiley & Sons, New York, 1979.
[39] A.W. Naylor and G.R. Sell, Linear Operator Theory in Engineering and
Science, Springer-Verlag~ N.Y.~ 1982.
[40] L.S. Shieh and Y.T. Tsay, "A Minimal Nice Selection Algorithm for the
Canonical Controller Structures of Multivariable Control System", U.S. Army
Research Report, 1983.
[41] P. Lancaster and P.N. Weber, "Jordan Chains for Lambda Matrices", Linear
Algebra and Its Applications, Vol. I, pp. 563-569, 1968.
[42] F.P. Greenleaf, l~troduction to Complex Variables, W.B. Saunders Co.,

Philadelphia, 1972.
[43] R.F. Rinehart, "The Equivalence of Definitions of a Matrix Function",

American Mathematical Monthly, Vol. 62, pp. 395-414, 1955.
[44] L.S. Shieh and N. Chahin, "A Computer-Aided Method for the Factorization of
Matrix Polynomials", Int. J. Systems Science, Vol. 12, pp. 305-323, 1981.
[45] J.E. Dennis, J.F. Traub and R.P. Weber, "Algorithms for Solvents of Matrix
Polynomials", SLAM, Numer. Anal., Vol. 15~ pp. 523-533, 1978.
[46] E.D. Denman, "Matrix Polynomials, Roots and Spectral Factors", Applied
Mathematics and Computation, Vol. 39 PP. 359-368, 1977.
[47] P.P. Khargonekar and E. Emre, "Further Results on Polynomial

Characterizations of (F,G)-Invariant and Reachabillty Subs,aces", IEEE
Trans. Automat. Control, Vol. AC-27p pp. 352-366, 1982.
[48] R.E. Kalman, "On Partial Realizations, Transfer Functions, and Canonical
Forms", Act Polytechnica Scandinavia, pp. 9-32, 1979.
[49] R. Herman, Linear System Theory and Introductory Algebraic Geometry,

Interdisciplinary Mathematics Vol. VIII, New Jersey, 1974.
[50] P.A. Fuhrmann, "Simulation of Linear Systems and Factorization of Matrix

Polynomials", Int. J. Control, Sol. 28, pp. 689-705, 1978.
[51] P.A. Fuhrmann, "Algebraic System Theory: An Analyst's Point of View",

Journal of the Franklin Institutes, Vol. 301, pp. 521-540, 1978.
[52] P.A. Fuhrmann and J.C. Willems 9 "A Study of (A,B)-Invariant Subs,aces Via
Polynomial Models", Int. J. Control, Vol. 31, pp. 467-494, 1980.
[53] A.C. Antoulas, "A System-Theoretic Approach to the Factorization Theory of

203
Non-Singular Polynomial Matrices", Int. J. Control, Vol. 33, pp. 1005-1026,

1981.
[54] M.S. Mahmoud and M.G. Singh, Large Scale Systems Modellin~, Pergamon Press,
N.Y., 1981.
[55] H. Kwakernaak and R. Sivan, Linear Optimal Control Systems, John Wiley &
Sons, Inc. N.Y., 1972.
[56] J.D. Roberts, "Linear Model Reduction and Solution of the Algebraic Riccati
Equation by Use of the Sign Function", CUED/B-CONTROL/TR 13 Report,
Cambridge University, 1971, Also, Int. J. Control, 1980, Vol. 32, pp. 677-
687.
[57] E.D. Denman and A.N. Beavers, "The Matrix Sign Function and Computations in
Systems", Appl. Math. & Comput., Vol. 2, pp. 63-94~ 1976.
[58] R.L. Mattheys, "Stability Analysis Via the Extended Matrix Sign Function",
Proc. IEE, Vol. 125, pp. 241-243, 1978.
[59] L.A. Balzer, "Accelerated Convergence of the Matrix Sign Function Method of
Solving Lyapunov, Riccati and Other Matrix Equations", Int. J. Control,
1980~ Vol. 32, pp. 1057-1078.
[60] L.S. Shieh, Y.T. Tsay and R.E. Yates, "Some Properties of Matrix Sign
Functions Derived from Continued Fractions", Proc. lEE, Vol. 130, Part D,
pp. 111-118, 1983.
[61] C.T. Chen, Introduction to Linear System Theory, Holt, Rinehart and
Winston s New York, 1970; Second Edition, 1984.
[62] A.J. Fossard, Multivariable System Control, North-Holland, 1972.
[63] B.S. Morgan, "Multivarlable System Synthesis by State-Variable Feedback",

IEEE Trans. Auto. Control, Vol. AC-9, pp. 405-411, 1964.
[64] B.C. Moore, "Eigenvalue and Eigenvector Assignment Using State Feedback",
IEEE Trans. Auto. Control, Vol. AC-21, pp. 689-692, 1976.
[65] E.J. Davlson, "Multlvarlable Systems Subject to Arbitrary Output and State
Feedback", IEEE Trans. Auto. Control, Vol. AC-18, pp. 24-32, 1973.
[66] J.J. D'Azzo and C.H. Houpls, Linear Control System Analysis and Design,
McGraw-Mill~ New York, 1981.
[67] W.A. Wolovich, "The Differential Operator Approach to Linear System

Analysis and Design", J. Franklin Inst., Vol. 301, pp. 27-47, 1976.
[68] W.A. Wolovich and P.L. Falb, "Invariants and Canonical Forms Under Dynamic
Compensation", SlAM J. Contr., Vol. 14, pp. 996-1008, 1976.
[69] P.L. Falb and W.A. Wolovich, "The Roles of the Interactor in Decoupling",
Proc. JACC~ pp. 1456-1460, 1977.
[70] P.L. Falb and W.A. Wolovich, "Decoupling in the Design and Synthesis of
Multlvariable Control Systems", IEEE Trans. Auto. Control, Vol. AC-129 pp.
651-659, 1967.
204
[71] E.G. Gilbert, "The Decoupling of Multivariable Systems by State Feedback"~

SlAM J. Control, Vol. 7, pp. 50-63, 1969.
[72] H.H. Rosenbrock~ "Design of Multivariable Control Systems Using Inverse

Nyquist Array", Proc. lEE, Vol. 116, pp. 1929-1936, 1969.
[73] D.Q. Mayne~ "The Design of Linear Multivariable Systems"~ Automatic, Vol.
9~ pp. 201-207, 1973.
[74] D.H. Owens, "Dyadic Expansion, Characteristic Loci and Multlvariable-

Control System Design", Proc. lEE, Vol. 122, pp. 315-320, 1975.
[75] D.C. Youla, H.A. Jabr and J.J. Bongiorno, "Modern Wiener-Hopf Design of
Optimal Controllers - Part II: The Multivariable Case", IEEE Trans. Autom.
Control, Vol. AC-21, pp. 319-338, 1976.
[76] A.G.J. MacFarlane and I. Postlethwaite, "The Generalized Nyquist Stability

Criterion and Multivariable Root Loci", Int. J. Control, Vol. 25, pp. 81-
127, 1977.
[77] A.G.J. MacFarlane and I. Postlethwaite, "Characteristic Frequency Functions

and Characteristic Gain Functions", Int. J. Control, Vol. 26, pp. 265-278,
1977.
[78] G. Bengtsson, "Output Regulation and Internal Models - A Frequency Domain

Approach", Automatica, Vol. 13, pp. 333-345, 1977.
[79] B.W. Dickinson, T. Kailath and M. Morf, "Canonical Matrix Fraction and
State-space Descriptions for Deterministic and Stochastic Linear Systems",
IEEE Trans. Autom. Control, Vol. AC-19, pp. 656-667, 1974.
[801 J. Fadavi-Ardekanl, S.K. Mitra and B.D.O. Anderson, "Extended State-space

Model of Discrete-time Dynamical Systems. IEEE Trans. Circuits and
Systems", Vol. CAS-29, pp. 547-556, 1982.
[81] J.C. Willems, "Almost Invariant Subspaces: An Approach to High Gain

Feedback Design Part I: Almost Controlled Invariant Subspaces", IEEE
Trans. Autom. Control, Vol. AC-26, pp. 235-252, 1981.
[82] J.C. Willems, "Almost Invarlant Subspaces: An Approach to High Gain

Feedback Design - Part 111 Almost Conditionally Invarlant Subspaces", IEEE
Trans. Autom. Control, Vol. AC-27, pp. 1071-1085, 1982.
[83] S.V. RaG and S.S. Lamba, "Suboptimal Control of Linear Systems Via
Simplified Models of Chidambara", Proe. lEE, Vol. 121, pp. 879-881, 1974.
[84] L.S. Shieh, S. Yeh and H.Y. Zhang, "Realizations of Matrix Transfer
Functions Using RC Ladders and Summers", Proc. lEE, Vol. 128, Part G, pp.
i01-II0, 1981.
[85] J. Tajima, H. Nagase and S.I. Takahashi, "Synthesis Problem of Voltage

Transfer Function Matrix", Electronics and Communications in Japan, Vol.
61-A, pp. 17-25, 1978.
[86] F.F. Kuo~ Network Analysis and Synthesis, John Wiley & Sons~ New York~
1966.
[87] G.C. Temes and J.W. LaPatraj IntEpduction to Circuit Synthesis and Design,
205
McGraw-Hill, New York, 1977.
[88] R.W. Newcomb, Linear Multiport Synthesis, McGraw-Hill, New York, 1966.
[89] B. Porter, Synthesis of Dynamic Systems, Nelson, London, 1969.
[90] Y.T. Tsay and L.S. Shleh, "Block Modal Control Vla Cascade Multivariable
Control Structure", Int. J. Systems Science, Vol. 15, pp. 199-214, 1984.
[91] D.C. Youla, "On the Factorization of Rational Matrices", IRE Trans.
Information Theory, Vol. IT-7, pp. 172-189, 1960.
[92] B.D.O. Anderson, "An Algebraic Solution to the Spectral Factorizatlon

Problems", IEEE Trans. Autom. Control, Vol. AC-12, pp. 410-414, 1967.
[93] B.D.O. Anderson, K.L. Hitz and N.D. Diem, "Recursive Algorithm for Spectral
Factorization", IEEE Trans. Circuits and Systems, Vol. CAS-21, pp. 742-750,
1974.
[94] H. Bart, I. Gohberg and M.A. Kaashoek, Minimal Factorlzations of Matrix and
Operator Functions, Birhauser Verlag, Boston, 1979.
[95] H. Bart, I. Gohberg, M.A. Kaashoek and P. Van Dooren, "Factorizations of

Transfer Functions"~ SIAM J. Control and Optimization, Vol. 18, pp. 675-
696~ 1980.
[96] P. Van Dooren and P. Dewilde, "Minal Cascade Factorization of Real and
Complex Rational Transfer Matrices", IEEE Trans. Circuits and Systems, Vol.
CAS-28, pp. 390-400, 1981.
[97] P. Dewilde and J. Vandewalle, "On the Factorization of a Nonsingular

Rational Matrix", IEEE Trans. Circuits and System, Vol. CAS-22, pp. 637-
645, 1975.
[98] L.S. Shleh, Y.T. Tsay and R.E. Yates, "Cascade Realization of Multiport
Networks Via Factorlzation of Rational Matrix Functions", 7th Int.
Symposium on the Mathematical Theory of Networks and Systems, Stockholm,
Sweden, 1985.
[99] M.K. Saln, "The Growing Algebraic Presence in System Engineering: An

Introduction", Proc. IEEE, Vol. 64, pp. 96-111, 1976.
[I00] J.W° Helton~ "Systems with Infinite-Dimensional State Space: The Hilbert
Space Approach", Proc. IEEEj Vol. 64, pp. 145-160, 1976.
[i01] J.L. Casti, "Recent Developments and Future Perspectives in Non-Linear

System Theory", SIAM Review, Vol. 247 pp. 301-331, 1982.
[102] R.P. Roesser, "A Discrete State-Space Model for Linear Image Processing",
IEEE Trans. Autom. Control~ Vol. AC-20~ pp. i-i0~ 1975.
[103] B.G. Mertzios and P.N. Paraskevopoulos, "Transfer Function Matrix of 2-D
Systems", IEEE Trans. Autom. Control, Vol. AC-26, pp. 722-725, 1981.
104] N.K. Bose, Multidimensional Systems: Theory and Applications~ IEEE Press,
New York, 1979.
[105] S.K. Mitra and M.P. Ekstrom, Two-Dimensional Digital Signal Processing,
Dowden, Hutchinson and Ross, Stroudsburg, 1978.
206
[106] K.J. Astrom~ Introduction to Stochastic Control Theory~ Academic Pressj New
York, 1970.
[107] B.D.O. Anderson and J.B. Moore~ Optimal Filterin~ Prentice-Hall~ Englewood
Cliffs~ New Jersey~ 1979.
[108] G.J. Bierman, Factorizatlon Methods for Discrete Sequential Estimation~

Academic Press~ New York 9 1977.
[109] E.D. Denman~ J. Leyva-Ramos and G.J. Jeon~ "The Algebraic Theory of Latent
Projectors in Lambda Matrices"~ Applied Mathematics and Computation~ Vol.
9~ pp. 273-300, 1981.
[II0] L.S. Shieh~ Y.T. Tsay and R.E. Yates~ "State-Feedback Decomposition of
Multivariable Systems Via Block-Pole Placement", IEEE Trans. Auto.
Control~ Vol. AC-28, pp. 850-852~ 1983.

Lecture Notes in Control and Information Sciences: 107 Y. T. Tsay, L.-S. Shieh, S. Barnett

Uploaded by

Copyright:

Available Formats

Lecture Notes in Control and Information Sciences: 107 Y. T. Tsay, L.-S. Shieh, S. Barnett

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Lecture Notes in Control and Information Sciences: 107 Y. T. Tsay, L.-S. Shieh, S. Barnett

Uploaded by

Copyright:

Available Formats

Lecture Notes in

Y. T. Tsay, L.-S. Shieh,

Structural Analysis and Design

ISBN 3-540-18916-5 Springer-Verlag Berlin Heidelberg New York

into two main streams:

of both algebra and system theory.

to analyze the characteristics of systems and to design controllers for

satisfying various specifications and performance criteria.

literature9 we find that many elegant theories still cannot be employed to

analysis/design techniques. This provides the main motivation for our

fraction descriptions as the mathematical models for physical systems. A

unified approach characterizing the dynamics of a system is presented through

the formulation of the characteristic %-matrix (also known as the matrix

polynomial) of the system. Applications in pole assignment design~ modal

control design for multivariable systems~ parallel realizations~ and cascade

realizations of multiport networks are illustrated. A detailed guide to the

content of the monograph is provided in the last section of Chapter I.

The a u t h o r s wish to express their gratitude for the fruitful discussions

with Dr. Jagdlsh Chandra, Director of Mathematics Science, US Acmy R e s e a r c h

Robotics Group of the Fire Support Laboratory, US Army Armament Research,

This research monography was f i n a n c i a l l y supported in part by t h e US Army

Research Office under contract DAAL-03-87-K-0001, the US Army M i s s i l e Command

under contract DAAH-01-85-C-Alll, and the NASA-Johnson Space Center under

i.I State-space representations and matrix fraction descriptions of

1.2 Fundamental properties of X-matrices . . . . . . . . . . . . . . . . . 5

1.3 Organization of chapters . . . . . . . . . . . . . . . . . . . . . . . 8

CHAPTER 2. CHARACTERISTIC x-MATRICES AND CANONICAL MATRIX FRACTION

2.1 Characteristic X-matrlces of reachable and observable systems ..... 12

2.2 Canonical matrix fraction descriptions (MFDs) of MIMO systems ..... 18

2.3 Minimal nice selections for determining Kronecker and observability

2.4 Illustrative examples . . . . . . . . . . . . . . . . . . . . . . . . . 33

CHAPTER 3. SPECTRAL ANALYSIS OF NONSINGULAR X-MATRICES

3.1 Canonical X-matrices and canonical MFDs of MIMO systems ........ 39

3.2 Latent structure of nonsingular l-matrices .............. 46

3.3 Solvents of nonslngular X-matrices . . . . . . . . . . . . . . . . . . 61

3.4 Illustrative examples . . . . . . . . . . . . . . . . . . . . . . . . . 67

CHAPTER 4. DIVISORS AND SPECTRAL FACTORS OF NONSINGULAR k-MATRICES

4.3 Structures theorems for spectral factorization of nonsingular X-

4.4 Computational aspects of divisors and spectral factors for A-matrices 94

4.5 Illustrative examples . . . . . . . . . . . . . . . . . . . . . . . . . i12

CHAPTER 5. FEEDBACK CONTROL OF MULTIVARIABLE SYSTEMS

5.1 Linear state-feedback controls and properties ............. 123

5.2 X-matrix assignment . . . . . . . . . . . . . . . . . . . . . . . . . . 126

5.3 Left and right latent structure assignment . . . . . . . . . . . . . . 135

CHAPTER 6. STRUCTURAL DECOMPOSITION THEORIES AND APPLICATIONS IN

6.1 Parallel decomposition theories and their applications to MIMO systems

6.2 Semi-cascade decomposition theories and their applications to modal

6.3 Cascade decomposition theories and their applications to multiport

In this introductory chapter~ state-space representations and matrix

fraction descriptions of multivariable linear systems are reviewed in Section

guide to the content of the monograph.

I.i State-Space Representations and Matrix Fraction Descriptions of

An m-input~ p-output linear time-invariant system o can be described by

state equations as follows:

~x(t) = Ax(t) + ~u(t)

respectively; ~ F~ U are state~ output and input spaces of O~ respectively;

A~BgCgD are matrices of appropriate dimensions. For continuous-time systems~

is a differential operator and tERj while for discrete-time systems~ A is a

forward shift operator and t~2.